xref: /freebsd/sys/net/iflib.c (revision 66def52613043a86172a2ebe6feab214258fa2fa)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/sockio.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/kobj.h>
46 #include <sys/rman.h>
47 #include <sys/sbuf.h>
48 #include <sys/smp.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #if defined(__i386__) || defined(__amd64__)
92 #include <sys/memdesc.h>
93 #include <machine/bus.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
96 #include <x86/include/busdma_impl.h>
97 #include <x86/iommu/busdma_dmar.h>
98 #endif
99 
100 #include <sys/bitstring.h>
101 /*
102  * enable accounting of every mbuf as it comes in to and goes out of
103  * iflib's software descriptor references
104  */
105 #define MEMORY_LOGGING 0
106 /*
107  * Enable mbuf vectors for compressing long mbuf chains
108  */
109 
110 /*
111  * NB:
112  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
113  *   we prefetch needs to be determined by the time spent in m_free vis a vis
114  *   the cost of a prefetch. This will of course vary based on the workload:
115  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
116  *        is quite expensive, thus suggesting very little prefetch.
117  *      - small packet forwarding which is just returning a single mbuf to
118  *        UMA will typically be very fast vis a vis the cost of a memory
119  *        access.
120  */
121 
122 
123 /*
124  * File organization:
125  *  - private structures
126  *  - iflib private utility functions
127  *  - ifnet functions
128  *  - vlan registry and other exported functions
129  *  - iflib public core functions
130  *
131  *
132  */
133 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155    /*
156    * Pointer to hardware driver's softc
157    */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct mtx ifc_mtx;
167 
168 	uint16_t ifc_nhwtxqs;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	int ifc_in_detach;
176 
177 	int ifc_link_state;
178 	int ifc_link_irq;
179 	int ifc_watchdog_events;
180 	struct cdev *ifc_led_dev;
181 	struct resource *ifc_msix_mem;
182 
183 	struct if_irq ifc_legacy_irq;
184 	struct grouptask ifc_admin_task;
185 	struct grouptask ifc_vflr_task;
186 	struct iflib_filter_info ifc_filter_info;
187 	struct ifmedia	ifc_media;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 
195 	qidx_t ifc_sysctl_ntxds[8];
196 	qidx_t ifc_sysctl_nrxds[8];
197 	struct if_txrx ifc_txrx;
198 #define isc_txd_encap  ifc_txrx.ift_txd_encap
199 #define isc_txd_flush  ifc_txrx.ift_txd_flush
200 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
201 #define isc_rxd_available ifc_txrx.ift_rxd_available
202 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
203 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
204 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
208 	eventhandler_tag ifc_vlan_attach_event;
209 	eventhandler_tag ifc_vlan_detach_event;
210 	uint8_t ifc_mac[ETHER_ADDR_LEN];
211 	char ifc_mtx_name[16];
212 };
213 
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (&ctx->ifc_media);
241 }
242 
243 void
244 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
245 {
246 
247 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
248 }
249 
250 if_softc_ctx_t
251 iflib_get_softc_ctx(if_ctx_t ctx)
252 {
253 
254 	return (&ctx->ifc_softc_ctx);
255 }
256 
257 if_shared_ctx_t
258 iflib_get_sctx(if_ctx_t ctx)
259 {
260 
261 	return (ctx->ifc_sctx);
262 }
263 
264 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
265 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
266 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
267 
268 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
269 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
270 
271 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
272 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
273 #define RX_SW_DESC_INUSE        (1 << 3)
274 #define TX_SW_DESC_MAPPED       (1 << 4)
275 
276 #define	M_TOOBIG		M_PROTO1
277 
278 typedef struct iflib_sw_rx_desc_array {
279 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
280 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
281 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
282 	uint8_t		*ifsd_flags;
283 } iflib_rxsd_array_t;
284 
285 typedef struct iflib_sw_tx_desc_array {
286 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 	uint8_t		*ifsd_flags;
289 } if_txsd_vec_t;
290 
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 /* bnxt supports 64 with hardware LRO enabled */
295 #define IFLIB_MAX_RX_SEGS		64
296 #define IFLIB_RX_COPY_THRESH		128
297 #define IFLIB_MAX_RX_REFRESH		32
298 /* The minimum descriptors per second before we start coalescing */
299 #define IFLIB_MIN_DESC_SEC		16384
300 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
301 #define IFLIB_QUEUE_IDLE		0
302 #define IFLIB_QUEUE_HUNG		1
303 #define IFLIB_QUEUE_WORKING		2
304 /* maximum number of txqs that can share an rx interrupt */
305 #define IFLIB_MAX_TX_SHARED_INTR	4
306 
307 /* this should really scale with ring size - this is a fairly arbitrary value */
308 #define TX_BATCH_SIZE			32
309 
310 #define IFLIB_RESTART_BUDGET		8
311 
312 #define	IFC_LEGACY		0x001
313 #define	IFC_QFLUSH		0x002
314 #define	IFC_MULTISEG		0x004
315 #define	IFC_DMAR		0x008
316 #define	IFC_SC_ALLOCATED	0x010
317 #define	IFC_INIT_DONE		0x020
318 #define	IFC_PREFETCH		0x040
319 #define	IFC_DO_RESET		0x080
320 #define	IFC_CHECK_HUNG		0x100
321 
322 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
323 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
324 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
325 struct iflib_txq {
326 	qidx_t		ift_in_use;
327 	qidx_t		ift_cidx;
328 	qidx_t		ift_cidx_processed;
329 	qidx_t		ift_pidx;
330 	uint8_t		ift_gen;
331 	uint8_t		ift_br_offset;
332 	uint16_t	ift_npending;
333 	uint16_t	ift_db_pending;
334 	uint16_t	ift_rs_pending;
335 	/* implicit pad */
336 	uint8_t		ift_txd_size[8];
337 	uint64_t	ift_processed;
338 	uint64_t	ift_cleaned;
339 	uint64_t	ift_cleaned_prev;
340 #if MEMORY_LOGGING
341 	uint64_t	ift_enqueued;
342 	uint64_t	ift_dequeued;
343 #endif
344 	uint64_t	ift_no_tx_dma_setup;
345 	uint64_t	ift_no_desc_avail;
346 	uint64_t	ift_mbuf_defrag_failed;
347 	uint64_t	ift_mbuf_defrag;
348 	uint64_t	ift_map_failed;
349 	uint64_t	ift_txd_encap_efbig;
350 	uint64_t	ift_pullups;
351 
352 	struct mtx	ift_mtx;
353 	struct mtx	ift_db_mtx;
354 
355 	/* constant values */
356 	if_ctx_t	ift_ctx;
357 	struct ifmp_ring        *ift_br;
358 	struct grouptask	ift_task;
359 	qidx_t		ift_size;
360 	uint16_t	ift_id;
361 	struct callout	ift_timer;
362 
363 	if_txsd_vec_t	ift_sds;
364 	uint8_t		ift_qstatus;
365 	uint8_t		ift_closed;
366 	uint8_t		ift_update_freq;
367 	struct iflib_filter_info ift_filter_info;
368 	bus_dma_tag_t		ift_desc_tag;
369 	bus_dma_tag_t		ift_tso_desc_tag;
370 	iflib_dma_info_t	ift_ifdi;
371 #define MTX_NAME_LEN 16
372 	char                    ift_mtx_name[MTX_NAME_LEN];
373 	char                    ift_db_mtx_name[MTX_NAME_LEN];
374 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
375 #ifdef IFLIB_DIAGNOSTICS
376 	uint64_t ift_cpu_exec_count[256];
377 #endif
378 } __aligned(CACHE_LINE_SIZE);
379 
380 struct iflib_fl {
381 	qidx_t		ifl_cidx;
382 	qidx_t		ifl_pidx;
383 	qidx_t		ifl_credits;
384 	uint8_t		ifl_gen;
385 	uint8_t		ifl_rxd_size;
386 #if MEMORY_LOGGING
387 	uint64_t	ifl_m_enqueued;
388 	uint64_t	ifl_m_dequeued;
389 	uint64_t	ifl_cl_enqueued;
390 	uint64_t	ifl_cl_dequeued;
391 #endif
392 	/* implicit pad */
393 
394 	bitstr_t 	*ifl_rx_bitmap;
395 	qidx_t		ifl_fragidx;
396 	/* constant */
397 	qidx_t		ifl_size;
398 	uint16_t	ifl_buf_size;
399 	uint16_t	ifl_cltype;
400 	uma_zone_t	ifl_zone;
401 	iflib_rxsd_array_t	ifl_sds;
402 	iflib_rxq_t	ifl_rxq;
403 	uint8_t		ifl_id;
404 	bus_dma_tag_t           ifl_desc_tag;
405 	iflib_dma_info_t	ifl_ifdi;
406 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
407 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
408 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
409 }  __aligned(CACHE_LINE_SIZE);
410 
411 static inline qidx_t
412 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
413 {
414 	qidx_t used;
415 
416 	if (pidx > cidx)
417 		used = pidx - cidx;
418 	else if (pidx < cidx)
419 		used = size - cidx + pidx;
420 	else if (gen == 0 && pidx == cidx)
421 		used = 0;
422 	else if (gen == 1 && pidx == cidx)
423 		used = size;
424 	else
425 		panic("bad state");
426 
427 	return (used);
428 }
429 
430 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
431 
432 #define IDXDIFF(head, tail, wrap) \
433 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
434 
435 struct iflib_rxq {
436 	/* If there is a separate completion queue -
437 	 * these are the cq cidx and pidx. Otherwise
438 	 * these are unused.
439 	 */
440 	qidx_t		ifr_size;
441 	qidx_t		ifr_cq_cidx;
442 	qidx_t		ifr_cq_pidx;
443 	uint8_t		ifr_cq_gen;
444 	uint8_t		ifr_fl_offset;
445 
446 	if_ctx_t	ifr_ctx;
447 	iflib_fl_t	ifr_fl;
448 	uint64_t	ifr_rx_irq;
449 	uint16_t	ifr_id;
450 	uint8_t		ifr_lro_enabled;
451 	uint8_t		ifr_nfl;
452 	uint8_t		ifr_ntxqirq;
453 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
454 	struct lro_ctrl			ifr_lc;
455 	struct grouptask        ifr_task;
456 	struct iflib_filter_info ifr_filter_info;
457 	iflib_dma_info_t		ifr_ifdi;
458 
459 	/* dynamically allocate if any drivers need a value substantially larger than this */
460 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
461 #ifdef IFLIB_DIAGNOSTICS
462 	uint64_t ifr_cpu_exec_count[256];
463 #endif
464 }  __aligned(CACHE_LINE_SIZE);
465 
466 typedef struct if_rxsd {
467 	caddr_t *ifsd_cl;
468 	struct mbuf **ifsd_m;
469 	iflib_fl_t ifsd_fl;
470 	qidx_t ifsd_cidx;
471 } *if_rxsd_t;
472 
473 /* multiple of word size */
474 #ifdef __LP64__
475 #define PKT_INFO_SIZE	6
476 #define RXD_INFO_SIZE	5
477 #define PKT_TYPE uint64_t
478 #else
479 #define PKT_INFO_SIZE	11
480 #define RXD_INFO_SIZE	8
481 #define PKT_TYPE uint32_t
482 #endif
483 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
484 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
485 
486 typedef struct if_pkt_info_pad {
487 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
488 } *if_pkt_info_pad_t;
489 typedef struct if_rxd_info_pad {
490 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
491 } *if_rxd_info_pad_t;
492 
493 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
494 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
495 
496 
497 static inline void
498 pkt_info_zero(if_pkt_info_t pi)
499 {
500 	if_pkt_info_pad_t pi_pad;
501 
502 	pi_pad = (if_pkt_info_pad_t)pi;
503 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
504 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
505 #ifndef __LP64__
506 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
507 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
508 #endif
509 }
510 
511 static inline void
512 rxd_info_zero(if_rxd_info_t ri)
513 {
514 	if_rxd_info_pad_t ri_pad;
515 	int i;
516 
517 	ri_pad = (if_rxd_info_pad_t)ri;
518 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
519 		ri_pad->rxd_val[i] = 0;
520 		ri_pad->rxd_val[i+1] = 0;
521 		ri_pad->rxd_val[i+2] = 0;
522 		ri_pad->rxd_val[i+3] = 0;
523 	}
524 #ifdef __LP64__
525 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
526 #endif
527 }
528 
529 /*
530  * Only allow a single packet to take up most 1/nth of the tx ring
531  */
532 #define MAX_SINGLE_PACKET_FRACTION 12
533 #define IF_BAD_DMA (bus_addr_t)-1
534 
535 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
536 
537 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
538 
539 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
540 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
541 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
542 
543 
544 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
545 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
546 
547 
548 /* Our boot-time initialization hook */
549 static int	iflib_module_event_handler(module_t, int, void *);
550 
551 static moduledata_t iflib_moduledata = {
552 	"iflib",
553 	iflib_module_event_handler,
554 	NULL
555 };
556 
557 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
558 MODULE_VERSION(iflib, 1);
559 
560 MODULE_DEPEND(iflib, pci, 1, 1, 1);
561 MODULE_DEPEND(iflib, ether, 1, 1, 1);
562 
563 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
564 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
565 
566 #ifndef IFLIB_DEBUG_COUNTERS
567 #ifdef INVARIANTS
568 #define IFLIB_DEBUG_COUNTERS 1
569 #else
570 #define IFLIB_DEBUG_COUNTERS 0
571 #endif /* !INVARIANTS */
572 #endif
573 
574 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
575                    "iflib driver parameters");
576 
577 /*
578  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
579  */
580 static int iflib_min_tx_latency = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
582 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
583 static int iflib_no_tx_batch = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
585 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
586 
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# tx mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# tx mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# tx frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# rx allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 static int iflib_txq_drain_encapfail;
618 
619 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
620 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
622 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
623 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
624 		   &iflib_txq_drain_notready, 0, "# drain notready");
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
626 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
627 
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_intr_link;
647 static int iflib_intr_msix;
648 static int iflib_rx_unavail;
649 static int iflib_rx_ctx_inactive;
650 static int iflib_rx_zero_len;
651 static int iflib_rx_if_input;
652 static int iflib_rx_mbuf_null;
653 static int iflib_rxd_flush;
654 
655 static int iflib_verbose_debug;
656 
657 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
658 		   &iflib_intr_link, 0, "# intr link calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
660 		   &iflib_intr_msix, 0, "# intr msix calls");
661 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
662 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
664 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
665 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
666 		   &iflib_fast_intrs, 0, "# fast_intr calls");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
668 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
670 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
671 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
672 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
673 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
674 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
675 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
676 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
677 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
678 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
679 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
680 		   &iflib_verbose_debug, 0, "enable verbose debugging");
681 
682 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
683 static void
684 iflib_debug_reset(void)
685 {
686 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
687 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
688 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
689 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
690 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
691 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
692 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
693 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
694 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
695 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
696 }
697 
698 #else
699 #define DBG_COUNTER_INC(name)
700 static void iflib_debug_reset(void) {}
701 #endif
702 
703 
704 
705 #define IFLIB_DEBUG 0
706 
707 static void iflib_tx_structures_free(if_ctx_t ctx);
708 static void iflib_rx_structures_free(if_ctx_t ctx);
709 static int iflib_queues_alloc(if_ctx_t ctx);
710 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
711 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
712 static int iflib_qset_structures_setup(if_ctx_t ctx);
713 static int iflib_msix_init(if_ctx_t ctx);
714 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
715 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
716 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
717 static int iflib_register(if_ctx_t);
718 static void iflib_init_locked(if_ctx_t ctx);
719 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
721 static void iflib_ifmp_purge(iflib_txq_t txq);
722 static void _iflib_pre_assert(if_softc_ctx_t scctx);
723 static void iflib_stop(if_ctx_t ctx);
724 static void iflib_if_init_locked(if_ctx_t ctx);
725 #ifndef __NO_STRICT_ALIGNMENT
726 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
727 #endif
728 
729 #ifdef DEV_NETMAP
730 #include <sys/selinfo.h>
731 #include <net/netmap.h>
732 #include <dev/netmap/netmap_kern.h>
733 
734 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
735 
736 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
737 
738 /*
739  * device-specific sysctl variables:
740  *
741  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
742  *	During regular operations the CRC is stripped, but on some
743  *	hardware reception of frames not multiple of 64 is slower,
744  *	so using crcstrip=0 helps in benchmarks.
745  *
746  * iflib_rx_miss, iflib_rx_miss_bufs:
747  *	count packets that might be missed due to lost interrupts.
748  */
749 SYSCTL_DECL(_dev_netmap);
750 /*
751  * The xl driver by default strips CRCs and we do not override it.
752  */
753 
754 int iflib_crcstrip = 1;
755 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
756     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
757 
758 int iflib_rx_miss, iflib_rx_miss_bufs;
759 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
760     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
761 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
762     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
763 
764 /*
765  * Register/unregister. We are already under netmap lock.
766  * Only called on the first register or the last unregister.
767  */
768 static int
769 iflib_netmap_register(struct netmap_adapter *na, int onoff)
770 {
771 	struct ifnet *ifp = na->ifp;
772 	if_ctx_t ctx = ifp->if_softc;
773 	int status;
774 
775 	CTX_LOCK(ctx);
776 	IFDI_INTR_DISABLE(ctx);
777 
778 	/* Tell the stack that the interface is no longer active */
779 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
780 
781 	if (!CTX_IS_VF(ctx))
782 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
783 
784 	/* enable or disable flags and callbacks in na and ifp */
785 	if (onoff) {
786 		nm_set_native_flags(na);
787 	} else {
788 		nm_clear_native_flags(na);
789 	}
790 	iflib_stop(ctx);
791 	iflib_init_locked(ctx);
792 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
793 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
794 	if (status)
795 		nm_clear_native_flags(na);
796 	CTX_UNLOCK(ctx);
797 	return (status);
798 }
799 
800 static int
801 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
802 {
803 	struct netmap_adapter *na = kring->na;
804 	u_int const lim = kring->nkr_num_slots - 1;
805 	u_int head = kring->rhead;
806 	struct netmap_ring *ring = kring->ring;
807 	bus_dmamap_t *map;
808 	struct if_rxd_update iru;
809 	if_ctx_t ctx = rxq->ifr_ctx;
810 	iflib_fl_t fl = &rxq->ifr_fl[0];
811 	uint32_t refill_pidx, nic_i;
812 
813 	if (nm_i == head && __predict_true(!init))
814 		return 0;
815 	iru_init(&iru, rxq, 0 /* flid */);
816 	map = fl->ifl_sds.ifsd_map;
817 	refill_pidx = netmap_idx_k2n(kring, nm_i);
818 	/*
819 	 * IMPORTANT: we must leave one free slot in the ring,
820 	 * so move head back by one unit
821 	 */
822 	head = nm_prev(head, lim);
823 	while (nm_i != head) {
824 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
825 			struct netmap_slot *slot = &ring->slot[nm_i];
826 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
827 			uint32_t nic_i_dma = refill_pidx;
828 			nic_i = netmap_idx_k2n(kring, nm_i);
829 
830 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
831 
832 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
833 			        return netmap_ring_reinit(kring);
834 
835 			fl->ifl_vm_addrs[tmp_pidx] = addr;
836 			if (__predict_false(init) && map) {
837 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
838 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
839 				/* buffer has changed, reload map */
840 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
841 			}
842 			slot->flags &= ~NS_BUF_CHANGED;
843 
844 			nm_i = nm_next(nm_i, lim);
845 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
846 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
847 				continue;
848 
849 			iru.iru_pidx = refill_pidx;
850 			iru.iru_count = tmp_pidx+1;
851 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
852 
853 			refill_pidx = nic_i;
854 			if (map == NULL)
855 				continue;
856 
857 			for (int n = 0; n < iru.iru_count; n++) {
858 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
859 						BUS_DMASYNC_PREREAD);
860 				/* XXX - change this to not use the netmap func*/
861 				nic_i_dma = nm_next(nic_i_dma, lim);
862 			}
863 		}
864 	}
865 	kring->nr_hwcur = head;
866 
867 	if (map)
868 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
869 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
870 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
871 	return (0);
872 }
873 
874 /*
875  * Reconcile kernel and user view of the transmit ring.
876  *
877  * All information is in the kring.
878  * Userspace wants to send packets up to the one before kring->rhead,
879  * kernel knows kring->nr_hwcur is the first unsent packet.
880  *
881  * Here we push packets out (as many as possible), and possibly
882  * reclaim buffers from previously completed transmission.
883  *
884  * The caller (netmap) guarantees that there is only one instance
885  * running at any time. Any interference with other driver
886  * methods should be handled by the individual drivers.
887  */
888 static int
889 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
890 {
891 	struct netmap_adapter *na = kring->na;
892 	struct ifnet *ifp = na->ifp;
893 	struct netmap_ring *ring = kring->ring;
894 	u_int nm_i;	/* index into the netmap ring */
895 	u_int nic_i;	/* index into the NIC ring */
896 	u_int n;
897 	u_int const lim = kring->nkr_num_slots - 1;
898 	u_int const head = kring->rhead;
899 	struct if_pkt_info pi;
900 
901 	/*
902 	 * interrupts on every tx packet are expensive so request
903 	 * them every half ring, or where NS_REPORT is set
904 	 */
905 	u_int report_frequency = kring->nkr_num_slots >> 1;
906 	/* device-specific */
907 	if_ctx_t ctx = ifp->if_softc;
908 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
909 
910 	if (txq->ift_sds.ifsd_map)
911 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
912 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
913 
914 
915 	/*
916 	 * First part: process new packets to send.
917 	 * nm_i is the current index in the netmap ring,
918 	 * nic_i is the corresponding index in the NIC ring.
919 	 *
920 	 * If we have packets to send (nm_i != head)
921 	 * iterate over the netmap ring, fetch length and update
922 	 * the corresponding slot in the NIC ring. Some drivers also
923 	 * need to update the buffer's physical address in the NIC slot
924 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
925 	 *
926 	 * The netmap_reload_map() calls is especially expensive,
927 	 * even when (as in this case) the tag is 0, so do only
928 	 * when the buffer has actually changed.
929 	 *
930 	 * If possible do not set the report/intr bit on all slots,
931 	 * but only a few times per ring or when NS_REPORT is set.
932 	 *
933 	 * Finally, on 10G and faster drivers, it might be useful
934 	 * to prefetch the next slot and txr entry.
935 	 */
936 
937 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
938 	pkt_info_zero(&pi);
939 	pi.ipi_segs = txq->ift_segs;
940 	pi.ipi_qsidx = kring->ring_id;
941 	if (nm_i != head) {	/* we have new packets to send */
942 		nic_i = netmap_idx_k2n(kring, nm_i);
943 
944 		__builtin_prefetch(&ring->slot[nm_i]);
945 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
946 		if (txq->ift_sds.ifsd_map)
947 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
948 
949 		for (n = 0; nm_i != head; n++) {
950 			struct netmap_slot *slot = &ring->slot[nm_i];
951 			u_int len = slot->len;
952 			uint64_t paddr;
953 			void *addr = PNMB(na, slot, &paddr);
954 			int flags = (slot->flags & NS_REPORT ||
955 				nic_i == 0 || nic_i == report_frequency) ?
956 				IPI_TX_INTR : 0;
957 
958 			/* device-specific */
959 			pi.ipi_len = len;
960 			pi.ipi_segs[0].ds_addr = paddr;
961 			pi.ipi_segs[0].ds_len = len;
962 			pi.ipi_nsegs = 1;
963 			pi.ipi_ndescs = 0;
964 			pi.ipi_pidx = nic_i;
965 			pi.ipi_flags = flags;
966 
967 			/* Fill the slot in the NIC ring. */
968 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
969 
970 			/* prefetch for next round */
971 			__builtin_prefetch(&ring->slot[nm_i + 1]);
972 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
973 			if (txq->ift_sds.ifsd_map) {
974 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
975 
976 				NM_CHECK_ADDR_LEN(na, addr, len);
977 
978 				if (slot->flags & NS_BUF_CHANGED) {
979 					/* buffer has changed, reload map */
980 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
981 				}
982 				/* make sure changes to the buffer are synced */
983 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
984 						BUS_DMASYNC_PREWRITE);
985 			}
986 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
987 			nm_i = nm_next(nm_i, lim);
988 			nic_i = nm_next(nic_i, lim);
989 		}
990 		kring->nr_hwcur = head;
991 
992 		/* synchronize the NIC ring */
993 		if (txq->ift_sds.ifsd_map)
994 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
995 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
996 
997 		/* (re)start the tx unit up to slot nic_i (excluded) */
998 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
999 	}
1000 
1001 	/*
1002 	 * Second part: reclaim buffers for completed transmissions.
1003 	 */
1004 	if (iflib_tx_credits_update(ctx, txq)) {
1005 		/* some tx completed, increment avail */
1006 		nic_i = txq->ift_cidx_processed;
1007 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1008 	}
1009 	return (0);
1010 }
1011 
1012 /*
1013  * Reconcile kernel and user view of the receive ring.
1014  * Same as for the txsync, this routine must be efficient.
1015  * The caller guarantees a single invocations, but races against
1016  * the rest of the driver should be handled here.
1017  *
1018  * On call, kring->rhead is the first packet that userspace wants
1019  * to keep, and kring->rcur is the wakeup point.
1020  * The kernel has previously reported packets up to kring->rtail.
1021  *
1022  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1023  * of whether or not we received an interrupt.
1024  */
1025 static int
1026 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1027 {
1028 	struct netmap_adapter *na = kring->na;
1029 	struct netmap_ring *ring = kring->ring;
1030 	uint32_t nm_i;	/* index into the netmap ring */
1031 	uint32_t nic_i;	/* index into the NIC ring */
1032 	u_int i, n;
1033 	u_int const lim = kring->nkr_num_slots - 1;
1034 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1035 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1036 	struct if_rxd_info ri;
1037 
1038 	struct ifnet *ifp = na->ifp;
1039 	if_ctx_t ctx = ifp->if_softc;
1040 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1041 	iflib_fl_t fl = rxq->ifr_fl;
1042 	if (head > lim)
1043 		return netmap_ring_reinit(kring);
1044 
1045 	/* XXX check sync modes */
1046 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1047 		if (fl->ifl_sds.ifsd_map == NULL)
1048 			continue;
1049 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1050 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1051 	}
1052 	/*
1053 	 * First part: import newly received packets.
1054 	 *
1055 	 * nm_i is the index of the next free slot in the netmap ring,
1056 	 * nic_i is the index of the next received packet in the NIC ring,
1057 	 * and they may differ in case if_init() has been called while
1058 	 * in netmap mode. For the receive ring we have
1059 	 *
1060 	 *	nic_i = rxr->next_check;
1061 	 *	nm_i = kring->nr_hwtail (previous)
1062 	 * and
1063 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1064 	 *
1065 	 * rxr->next_check is set to 0 on a ring reinit
1066 	 */
1067 	if (netmap_no_pendintr || force_update) {
1068 		int crclen = iflib_crcstrip ? 0 : 4;
1069 		int error, avail;
1070 
1071 		for (i = 0; i < rxq->ifr_nfl; i++) {
1072 			fl = &rxq->ifr_fl[i];
1073 			nic_i = fl->ifl_cidx;
1074 			nm_i = netmap_idx_n2k(kring, nic_i);
1075 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1076 			for (n = 0; avail > 0; n++, avail--) {
1077 				rxd_info_zero(&ri);
1078 				ri.iri_frags = rxq->ifr_frags;
1079 				ri.iri_qsidx = kring->ring_id;
1080 				ri.iri_ifp = ctx->ifc_ifp;
1081 				ri.iri_cidx = nic_i;
1082 
1083 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1084 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1085 				ring->slot[nm_i].flags = 0;
1086 				if (fl->ifl_sds.ifsd_map)
1087 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1088 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1089 				nm_i = nm_next(nm_i, lim);
1090 				nic_i = nm_next(nic_i, lim);
1091 			}
1092 			if (n) { /* update the state variables */
1093 				if (netmap_no_pendintr && !force_update) {
1094 					/* diagnostics */
1095 					iflib_rx_miss ++;
1096 					iflib_rx_miss_bufs += n;
1097 				}
1098 				fl->ifl_cidx = nic_i;
1099 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1100 			}
1101 			kring->nr_kflags &= ~NKR_PENDINTR;
1102 		}
1103 	}
1104 	/*
1105 	 * Second part: skip past packets that userspace has released.
1106 	 * (kring->nr_hwcur to head excluded),
1107 	 * and make the buffers available for reception.
1108 	 * As usual nm_i is the index in the netmap ring,
1109 	 * nic_i is the index in the NIC ring, and
1110 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1111 	 */
1112 	/* XXX not sure how this will work with multiple free lists */
1113 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1114 
1115 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1116 }
1117 
1118 static void
1119 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1120 {
1121 	struct ifnet *ifp = na->ifp;
1122 	if_ctx_t ctx = ifp->if_softc;
1123 
1124 	CTX_LOCK(ctx);
1125 	if (onoff) {
1126 		IFDI_INTR_ENABLE(ctx);
1127 	} else {
1128 		IFDI_INTR_DISABLE(ctx);
1129 	}
1130 	CTX_UNLOCK(ctx);
1131 }
1132 
1133 
1134 static int
1135 iflib_netmap_attach(if_ctx_t ctx)
1136 {
1137 	struct netmap_adapter na;
1138 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1139 
1140 	bzero(&na, sizeof(na));
1141 
1142 	na.ifp = ctx->ifc_ifp;
1143 	na.na_flags = NAF_BDG_MAYSLEEP;
1144 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1145 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1146 
1147 	na.num_tx_desc = scctx->isc_ntxd[0];
1148 	na.num_rx_desc = scctx->isc_nrxd[0];
1149 	na.nm_txsync = iflib_netmap_txsync;
1150 	na.nm_rxsync = iflib_netmap_rxsync;
1151 	na.nm_register = iflib_netmap_register;
1152 	na.nm_intr = iflib_netmap_intr;
1153 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1154 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1155 	return (netmap_attach(&na));
1156 }
1157 
1158 static void
1159 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1160 {
1161 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1162 	struct netmap_slot *slot;
1163 
1164 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1165 	if (slot == NULL)
1166 		return;
1167 	if (txq->ift_sds.ifsd_map == NULL)
1168 		return;
1169 
1170 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1171 
1172 		/*
1173 		 * In netmap mode, set the map for the packet buffer.
1174 		 * NOTE: Some drivers (not this one) also need to set
1175 		 * the physical buffer address in the NIC ring.
1176 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1177 		 * netmap slot index, si
1178 		 */
1179 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1180 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1181 	}
1182 }
1183 
1184 static void
1185 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1186 {
1187 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1188 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1189 	struct netmap_slot *slot;
1190 	uint32_t nm_i;
1191 
1192 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1193 	if (slot == NULL)
1194 		return;
1195 	nm_i = netmap_idx_n2k(kring, 0);
1196 	netmap_fl_refill(rxq, kring, nm_i, true);
1197 }
1198 
1199 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1200 
1201 #else
1202 #define iflib_netmap_txq_init(ctx, txq)
1203 #define iflib_netmap_rxq_init(ctx, rxq)
1204 #define iflib_netmap_detach(ifp)
1205 
1206 #define iflib_netmap_attach(ctx) (0)
1207 #define netmap_rx_irq(ifp, qid, budget) (0)
1208 #define netmap_tx_irq(ifp, qid) do {} while (0)
1209 
1210 #endif
1211 
1212 #if defined(__i386__) || defined(__amd64__)
1213 static __inline void
1214 prefetch(void *x)
1215 {
1216 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1217 }
1218 static __inline void
1219 prefetch2cachelines(void *x)
1220 {
1221 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1222 #if (CACHE_LINE_SIZE < 128)
1223 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1224 #endif
1225 }
1226 #else
1227 #define prefetch(x)
1228 #define prefetch2cachelines(x)
1229 #endif
1230 
1231 static void
1232 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1233 {
1234 	iflib_fl_t fl;
1235 
1236 	fl = &rxq->ifr_fl[flid];
1237 	iru->iru_paddrs = fl->ifl_bus_addrs;
1238 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1239 	iru->iru_idxs = fl->ifl_rxd_idxs;
1240 	iru->iru_qsidx = rxq->ifr_id;
1241 	iru->iru_buf_size = fl->ifl_buf_size;
1242 	iru->iru_flidx = fl->ifl_id;
1243 }
1244 
1245 static void
1246 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1247 {
1248 	if (err)
1249 		return;
1250 	*(bus_addr_t *) arg = segs[0].ds_addr;
1251 }
1252 
1253 int
1254 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1255 {
1256 	int err;
1257 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1258 	device_t dev = ctx->ifc_dev;
1259 
1260 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1261 
1262 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1263 				sctx->isc_q_align, 0,	/* alignment, bounds */
1264 				BUS_SPACE_MAXADDR,	/* lowaddr */
1265 				BUS_SPACE_MAXADDR,	/* highaddr */
1266 				NULL, NULL,		/* filter, filterarg */
1267 				size,			/* maxsize */
1268 				1,			/* nsegments */
1269 				size,			/* maxsegsize */
1270 				BUS_DMA_ALLOCNOW,	/* flags */
1271 				NULL,			/* lockfunc */
1272 				NULL,			/* lockarg */
1273 				&dma->idi_tag);
1274 	if (err) {
1275 		device_printf(dev,
1276 		    "%s: bus_dma_tag_create failed: %d\n",
1277 		    __func__, err);
1278 		goto fail_0;
1279 	}
1280 
1281 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1282 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1283 	if (err) {
1284 		device_printf(dev,
1285 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1286 		    __func__, (uintmax_t)size, err);
1287 		goto fail_1;
1288 	}
1289 
1290 	dma->idi_paddr = IF_BAD_DMA;
1291 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1292 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1293 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1294 		device_printf(dev,
1295 		    "%s: bus_dmamap_load failed: %d\n",
1296 		    __func__, err);
1297 		goto fail_2;
1298 	}
1299 
1300 	dma->idi_size = size;
1301 	return (0);
1302 
1303 fail_2:
1304 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1305 fail_1:
1306 	bus_dma_tag_destroy(dma->idi_tag);
1307 fail_0:
1308 	dma->idi_tag = NULL;
1309 
1310 	return (err);
1311 }
1312 
1313 int
1314 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1315 {
1316 	int i, err;
1317 	iflib_dma_info_t *dmaiter;
1318 
1319 	dmaiter = dmalist;
1320 	for (i = 0; i < count; i++, dmaiter++) {
1321 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1322 			break;
1323 	}
1324 	if (err)
1325 		iflib_dma_free_multi(dmalist, i);
1326 	return (err);
1327 }
1328 
1329 void
1330 iflib_dma_free(iflib_dma_info_t dma)
1331 {
1332 	if (dma->idi_tag == NULL)
1333 		return;
1334 	if (dma->idi_paddr != IF_BAD_DMA) {
1335 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1336 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1337 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1338 		dma->idi_paddr = IF_BAD_DMA;
1339 	}
1340 	if (dma->idi_vaddr != NULL) {
1341 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1342 		dma->idi_vaddr = NULL;
1343 	}
1344 	bus_dma_tag_destroy(dma->idi_tag);
1345 	dma->idi_tag = NULL;
1346 }
1347 
1348 void
1349 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1350 {
1351 	int i;
1352 	iflib_dma_info_t *dmaiter = dmalist;
1353 
1354 	for (i = 0; i < count; i++, dmaiter++)
1355 		iflib_dma_free(*dmaiter);
1356 }
1357 
1358 #ifdef EARLY_AP_STARTUP
1359 static const int iflib_started = 1;
1360 #else
1361 /*
1362  * We used to abuse the smp_started flag to decide if the queues have been
1363  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1364  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1365  * is set.  Run a SYSINIT() strictly after that to just set a usable
1366  * completion flag.
1367  */
1368 
1369 static int iflib_started;
1370 
1371 static void
1372 iflib_record_started(void *arg)
1373 {
1374 	iflib_started = 1;
1375 }
1376 
1377 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1378 	iflib_record_started, NULL);
1379 #endif
1380 
1381 static int
1382 iflib_fast_intr(void *arg)
1383 {
1384 	iflib_filter_info_t info = arg;
1385 	struct grouptask *gtask = info->ifi_task;
1386 	if (!iflib_started)
1387 		return (FILTER_HANDLED);
1388 
1389 	DBG_COUNTER_INC(fast_intrs);
1390 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1391 		return (FILTER_HANDLED);
1392 
1393 	GROUPTASK_ENQUEUE(gtask);
1394 	return (FILTER_HANDLED);
1395 }
1396 
1397 static int
1398 iflib_fast_intr_rxtx(void *arg)
1399 {
1400 	iflib_filter_info_t info = arg;
1401 	struct grouptask *gtask = info->ifi_task;
1402 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1403 	if_ctx_t ctx;
1404 	int i, cidx;
1405 
1406 	if (!iflib_started)
1407 		return (FILTER_HANDLED);
1408 
1409 	DBG_COUNTER_INC(fast_intrs);
1410 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1411 		return (FILTER_HANDLED);
1412 
1413 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1414 		qidx_t txqid = rxq->ifr_txqid[i];
1415 
1416 		ctx = rxq->ifr_ctx;
1417 
1418 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1419 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1420 			continue;
1421 		}
1422 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1423 	}
1424 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1425 		cidx = rxq->ifr_cq_cidx;
1426 	else
1427 		cidx = rxq->ifr_fl[0].ifl_cidx;
1428 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1429 		GROUPTASK_ENQUEUE(gtask);
1430 	else
1431 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1432 	return (FILTER_HANDLED);
1433 }
1434 
1435 
1436 static int
1437 iflib_fast_intr_ctx(void *arg)
1438 {
1439 	iflib_filter_info_t info = arg;
1440 	struct grouptask *gtask = info->ifi_task;
1441 
1442 	if (!iflib_started)
1443 		return (FILTER_HANDLED);
1444 
1445 	DBG_COUNTER_INC(fast_intrs);
1446 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1447 		return (FILTER_HANDLED);
1448 
1449 	GROUPTASK_ENQUEUE(gtask);
1450 	return (FILTER_HANDLED);
1451 }
1452 
1453 static int
1454 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1455 	driver_filter_t filter, driver_intr_t handler, void *arg,
1456 				 char *name)
1457 {
1458 	int rc, flags;
1459 	struct resource *res;
1460 	void *tag = NULL;
1461 	device_t dev = ctx->ifc_dev;
1462 
1463 	flags = RF_ACTIVE;
1464 	if (ctx->ifc_flags & IFC_LEGACY)
1465 		flags |= RF_SHAREABLE;
1466 	MPASS(rid < 512);
1467 	irq->ii_rid = rid;
1468 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1469 	if (res == NULL) {
1470 		device_printf(dev,
1471 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1472 		return (ENOMEM);
1473 	}
1474 	irq->ii_res = res;
1475 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1476 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1477 						filter, handler, arg, &tag);
1478 	if (rc != 0) {
1479 		device_printf(dev,
1480 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1481 					  rid, name ? name : "unknown", rc);
1482 		return (rc);
1483 	} else if (name)
1484 		bus_describe_intr(dev, res, tag, "%s", name);
1485 
1486 	irq->ii_tag = tag;
1487 	return (0);
1488 }
1489 
1490 
1491 /*********************************************************************
1492  *
1493  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1494  *  the information needed to transmit a packet on the wire. This is
1495  *  called only once at attach, setup is done every reset.
1496  *
1497  **********************************************************************/
1498 
1499 static int
1500 iflib_txsd_alloc(iflib_txq_t txq)
1501 {
1502 	if_ctx_t ctx = txq->ift_ctx;
1503 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1504 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1505 	device_t dev = ctx->ifc_dev;
1506 	int err, nsegments, ntsosegments;
1507 
1508 	nsegments = scctx->isc_tx_nsegments;
1509 	ntsosegments = scctx->isc_tx_tso_segments_max;
1510 	MPASS(scctx->isc_ntxd[0] > 0);
1511 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1512 	MPASS(nsegments > 0);
1513 	MPASS(ntsosegments > 0);
1514 	/*
1515 	 * Setup DMA descriptor areas.
1516 	 */
1517 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1518 			       1, 0,			/* alignment, bounds */
1519 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1520 			       BUS_SPACE_MAXADDR,	/* highaddr */
1521 			       NULL, NULL,		/* filter, filterarg */
1522 			       sctx->isc_tx_maxsize,		/* maxsize */
1523 			       nsegments,	/* nsegments */
1524 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1525 			       0,			/* flags */
1526 			       NULL,			/* lockfunc */
1527 			       NULL,			/* lockfuncarg */
1528 			       &txq->ift_desc_tag))) {
1529 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1530 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1531 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1532 		goto fail;
1533 	}
1534 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1535 			       1, 0,			/* alignment, bounds */
1536 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1537 			       BUS_SPACE_MAXADDR,	/* highaddr */
1538 			       NULL, NULL,		/* filter, filterarg */
1539 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1540 			       ntsosegments,	/* nsegments */
1541 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1542 			       0,			/* flags */
1543 			       NULL,			/* lockfunc */
1544 			       NULL,			/* lockfuncarg */
1545 			       &txq->ift_tso_desc_tag))) {
1546 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1547 
1548 		goto fail;
1549 	}
1550 	if (!(txq->ift_sds.ifsd_flags =
1551 	    (uint8_t *) malloc(sizeof(uint8_t) *
1552 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1553 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1554 		err = ENOMEM;
1555 		goto fail;
1556 	}
1557 	if (!(txq->ift_sds.ifsd_m =
1558 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1559 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1560 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1561 		err = ENOMEM;
1562 		goto fail;
1563 	}
1564 
1565         /* Create the descriptor buffer dma maps */
1566 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1567 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1568 		return (0);
1569 
1570 	if (!(txq->ift_sds.ifsd_map =
1571 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1572 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1573 		err = ENOMEM;
1574 		goto fail;
1575 	}
1576 
1577 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1578 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1579 		if (err != 0) {
1580 			device_printf(dev, "Unable to create TX DMA map\n");
1581 			goto fail;
1582 		}
1583 	}
1584 #endif
1585 	return (0);
1586 fail:
1587 	/* We free all, it handles case where we are in the middle */
1588 	iflib_tx_structures_free(ctx);
1589 	return (err);
1590 }
1591 
1592 static void
1593 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1594 {
1595 	bus_dmamap_t map;
1596 
1597 	map = NULL;
1598 	if (txq->ift_sds.ifsd_map != NULL)
1599 		map = txq->ift_sds.ifsd_map[i];
1600 	if (map != NULL) {
1601 		bus_dmamap_unload(txq->ift_desc_tag, map);
1602 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1603 		txq->ift_sds.ifsd_map[i] = NULL;
1604 	}
1605 }
1606 
1607 static void
1608 iflib_txq_destroy(iflib_txq_t txq)
1609 {
1610 	if_ctx_t ctx = txq->ift_ctx;
1611 
1612 	for (int i = 0; i < txq->ift_size; i++)
1613 		iflib_txsd_destroy(ctx, txq, i);
1614 	if (txq->ift_sds.ifsd_map != NULL) {
1615 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1616 		txq->ift_sds.ifsd_map = NULL;
1617 	}
1618 	if (txq->ift_sds.ifsd_m != NULL) {
1619 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1620 		txq->ift_sds.ifsd_m = NULL;
1621 	}
1622 	if (txq->ift_sds.ifsd_flags != NULL) {
1623 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1624 		txq->ift_sds.ifsd_flags = NULL;
1625 	}
1626 	if (txq->ift_desc_tag != NULL) {
1627 		bus_dma_tag_destroy(txq->ift_desc_tag);
1628 		txq->ift_desc_tag = NULL;
1629 	}
1630 	if (txq->ift_tso_desc_tag != NULL) {
1631 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1632 		txq->ift_tso_desc_tag = NULL;
1633 	}
1634 }
1635 
1636 static void
1637 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1638 {
1639 	struct mbuf **mp;
1640 
1641 	mp = &txq->ift_sds.ifsd_m[i];
1642 	if (*mp == NULL)
1643 		return;
1644 
1645 	if (txq->ift_sds.ifsd_map != NULL) {
1646 		bus_dmamap_sync(txq->ift_desc_tag,
1647 				txq->ift_sds.ifsd_map[i],
1648 				BUS_DMASYNC_POSTWRITE);
1649 		bus_dmamap_unload(txq->ift_desc_tag,
1650 				  txq->ift_sds.ifsd_map[i]);
1651 	}
1652 	m_free(*mp);
1653 	DBG_COUNTER_INC(tx_frees);
1654 	*mp = NULL;
1655 }
1656 
1657 static int
1658 iflib_txq_setup(iflib_txq_t txq)
1659 {
1660 	if_ctx_t ctx = txq->ift_ctx;
1661 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1662 	iflib_dma_info_t di;
1663 	int i;
1664 
1665 	/* Set number of descriptors available */
1666 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1667 	/* XXX make configurable */
1668 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1669 
1670 	/* Reset indices */
1671 	txq->ift_cidx_processed = 0;
1672 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1673 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1674 
1675 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1676 		bzero((void *)di->idi_vaddr, di->idi_size);
1677 
1678 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1679 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1680 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1681 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1682 	return (0);
1683 }
1684 
1685 /*********************************************************************
1686  *
1687  *  Allocate memory for rx_buffer structures. Since we use one
1688  *  rx_buffer per received packet, the maximum number of rx_buffer's
1689  *  that we'll need is equal to the number of receive descriptors
1690  *  that we've allocated.
1691  *
1692  **********************************************************************/
1693 static int
1694 iflib_rxsd_alloc(iflib_rxq_t rxq)
1695 {
1696 	if_ctx_t ctx = rxq->ifr_ctx;
1697 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1698 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1699 	device_t dev = ctx->ifc_dev;
1700 	iflib_fl_t fl;
1701 	int			err;
1702 
1703 	MPASS(scctx->isc_nrxd[0] > 0);
1704 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1705 
1706 	fl = rxq->ifr_fl;
1707 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1708 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1709 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1710 					 1, 0,			/* alignment, bounds */
1711 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1712 					 BUS_SPACE_MAXADDR,	/* highaddr */
1713 					 NULL, NULL,		/* filter, filterarg */
1714 					 sctx->isc_rx_maxsize,	/* maxsize */
1715 					 sctx->isc_rx_nsegments,	/* nsegments */
1716 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1717 					 0,			/* flags */
1718 					 NULL,			/* lockfunc */
1719 					 NULL,			/* lockarg */
1720 					 &fl->ifl_desc_tag);
1721 		if (err) {
1722 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1723 				__func__, err);
1724 			goto fail;
1725 		}
1726 		if (!(fl->ifl_sds.ifsd_flags =
1727 		      (uint8_t *) malloc(sizeof(uint8_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 		if (!(fl->ifl_sds.ifsd_m =
1734 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1735 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1736 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1737 			err = ENOMEM;
1738 			goto fail;
1739 		}
1740 		if (!(fl->ifl_sds.ifsd_cl =
1741 		      (caddr_t *) malloc(sizeof(caddr_t) *
1742 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1743 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1744 			err = ENOMEM;
1745 			goto fail;
1746 		}
1747 
1748 		/* Create the descriptor buffer dma maps */
1749 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1750 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1751 			continue;
1752 
1753 		if (!(fl->ifl_sds.ifsd_map =
1754 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1755 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1756 			err = ENOMEM;
1757 			goto fail;
1758 		}
1759 
1760 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1761 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1762 			if (err != 0) {
1763 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1764 				goto fail;
1765 			}
1766 		}
1767 #endif
1768 	}
1769 	return (0);
1770 
1771 fail:
1772 	iflib_rx_structures_free(ctx);
1773 	return (err);
1774 }
1775 
1776 
1777 /*
1778  * Internal service routines
1779  */
1780 
1781 struct rxq_refill_cb_arg {
1782 	int               error;
1783 	bus_dma_segment_t seg;
1784 	int               nseg;
1785 };
1786 
1787 static void
1788 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1789 {
1790 	struct rxq_refill_cb_arg *cb_arg = arg;
1791 
1792 	cb_arg->error = error;
1793 	cb_arg->seg = segs[0];
1794 	cb_arg->nseg = nseg;
1795 }
1796 
1797 
1798 #ifdef ACPI_DMAR
1799 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1800 #else
1801 #define IS_DMAR(ctx) (0)
1802 #endif
1803 
1804 /**
1805  *	rxq_refill - refill an rxq  free-buffer list
1806  *	@ctx: the iflib context
1807  *	@rxq: the free-list to refill
1808  *	@n: the number of new buffers to allocate
1809  *
1810  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1811  *	The caller must assure that @n does not exceed the queue's capacity.
1812  */
1813 static void
1814 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1815 {
1816 	struct mbuf *m;
1817 	int idx, frag_idx = fl->ifl_fragidx;
1818         int pidx = fl->ifl_pidx;
1819 	caddr_t cl, *sd_cl;
1820 	struct mbuf **sd_m;
1821 	uint8_t *sd_flags;
1822 	struct if_rxd_update iru;
1823 	bus_dmamap_t *sd_map;
1824 	int n, i = 0;
1825 	uint64_t bus_addr;
1826 	int err;
1827 	qidx_t credits;
1828 
1829 	sd_m = fl->ifl_sds.ifsd_m;
1830 	sd_map = fl->ifl_sds.ifsd_map;
1831 	sd_cl = fl->ifl_sds.ifsd_cl;
1832 	sd_flags = fl->ifl_sds.ifsd_flags;
1833 	idx = pidx;
1834 	credits = fl->ifl_credits;
1835 
1836 	n  = count;
1837 	MPASS(n > 0);
1838 	MPASS(credits + n <= fl->ifl_size);
1839 
1840 	if (pidx < fl->ifl_cidx)
1841 		MPASS(pidx + n <= fl->ifl_cidx);
1842 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1843 		MPASS(fl->ifl_gen == 0);
1844 	if (pidx > fl->ifl_cidx)
1845 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1846 
1847 	DBG_COUNTER_INC(fl_refills);
1848 	if (n > 8)
1849 		DBG_COUNTER_INC(fl_refills_large);
1850 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1851 	while (n--) {
1852 		/*
1853 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1854 		 * initialized after rx.
1855 		 *
1856 		 * If the cluster is still set then we know a minimum sized packet was received
1857 		 */
1858 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1859 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1860                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1861 		if ((cl = sd_cl[frag_idx]) == NULL) {
1862                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1863 				break;
1864 #if MEMORY_LOGGING
1865 			fl->ifl_cl_enqueued++;
1866 #endif
1867 		}
1868 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1869 			break;
1870 		}
1871 #if MEMORY_LOGGING
1872 		fl->ifl_m_enqueued++;
1873 #endif
1874 
1875 		DBG_COUNTER_INC(rx_allocs);
1876 #if defined(__i386__) || defined(__amd64__)
1877 		if (!IS_DMAR(ctx)) {
1878 			bus_addr = pmap_kextract((vm_offset_t)cl);
1879 		} else
1880 #endif
1881 		{
1882 			struct rxq_refill_cb_arg cb_arg;
1883 			iflib_rxq_t q;
1884 
1885 			cb_arg.error = 0;
1886 			q = fl->ifl_rxq;
1887 			MPASS(sd_map != NULL);
1888 			MPASS(sd_map[frag_idx] != NULL);
1889 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1890 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1891 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1892 					BUS_DMASYNC_PREREAD);
1893 
1894 			if (err != 0 || cb_arg.error) {
1895 				/*
1896 				 * !zone_pack ?
1897 				 */
1898 				if (fl->ifl_zone == zone_pack)
1899 					uma_zfree(fl->ifl_zone, cl);
1900 				m_free(m);
1901 				n = 0;
1902 				goto done;
1903 			}
1904 			bus_addr = cb_arg.seg.ds_addr;
1905 		}
1906                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1907 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1908 
1909 		MPASS(sd_m[frag_idx] == NULL);
1910 		sd_cl[frag_idx] = cl;
1911 		sd_m[frag_idx] = m;
1912 		fl->ifl_rxd_idxs[i] = frag_idx;
1913 		fl->ifl_bus_addrs[i] = bus_addr;
1914 		fl->ifl_vm_addrs[i] = cl;
1915 		credits++;
1916 		i++;
1917 		MPASS(credits <= fl->ifl_size);
1918 		if (++idx == fl->ifl_size) {
1919 			fl->ifl_gen = 1;
1920 			idx = 0;
1921 		}
1922 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1923 			iru.iru_pidx = pidx;
1924 			iru.iru_count = i;
1925 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1926 			i = 0;
1927 			pidx = idx;
1928 			fl->ifl_pidx = idx;
1929 			fl->ifl_credits = credits;
1930 		}
1931 
1932 	}
1933 done:
1934 	if (i) {
1935 		iru.iru_pidx = pidx;
1936 		iru.iru_count = i;
1937 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1938 		fl->ifl_pidx = idx;
1939 		fl->ifl_credits = credits;
1940 	}
1941 	DBG_COUNTER_INC(rxd_flush);
1942 	if (fl->ifl_pidx == 0)
1943 		pidx = fl->ifl_size - 1;
1944 	else
1945 		pidx = fl->ifl_pidx - 1;
1946 
1947 	if (sd_map)
1948 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1949 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1950 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1951 	fl->ifl_fragidx = frag_idx;
1952 }
1953 
1954 static __inline void
1955 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1956 {
1957 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1958 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1959 #ifdef INVARIANTS
1960 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1961 #endif
1962 
1963 	MPASS(fl->ifl_credits <= fl->ifl_size);
1964 	MPASS(reclaimable == delta);
1965 
1966 	if (reclaimable > 0)
1967 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1968 }
1969 
1970 static void
1971 iflib_fl_bufs_free(iflib_fl_t fl)
1972 {
1973 	iflib_dma_info_t idi = fl->ifl_ifdi;
1974 	uint32_t i;
1975 
1976 	for (i = 0; i < fl->ifl_size; i++) {
1977 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1978 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1979 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1980 
1981 		if (*sd_flags & RX_SW_DESC_INUSE) {
1982 			if (fl->ifl_sds.ifsd_map != NULL) {
1983 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1984 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1985 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
1986 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1987 			}
1988 			if (*sd_m != NULL) {
1989 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1990 				uma_zfree(zone_mbuf, *sd_m);
1991 			}
1992 			if (*sd_cl != NULL)
1993 				uma_zfree(fl->ifl_zone, *sd_cl);
1994 			*sd_flags = 0;
1995 		} else {
1996 			MPASS(*sd_cl == NULL);
1997 			MPASS(*sd_m == NULL);
1998 		}
1999 #if MEMORY_LOGGING
2000 		fl->ifl_m_dequeued++;
2001 		fl->ifl_cl_dequeued++;
2002 #endif
2003 		*sd_cl = NULL;
2004 		*sd_m = NULL;
2005 	}
2006 #ifdef INVARIANTS
2007 	for (i = 0; i < fl->ifl_size; i++) {
2008 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2009 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2010 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2011 	}
2012 #endif
2013 	/*
2014 	 * Reset free list values
2015 	 */
2016 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2017 	bzero(idi->idi_vaddr, idi->idi_size);
2018 }
2019 
2020 /*********************************************************************
2021  *
2022  *  Initialize a receive ring and its buffers.
2023  *
2024  **********************************************************************/
2025 static int
2026 iflib_fl_setup(iflib_fl_t fl)
2027 {
2028 	iflib_rxq_t rxq = fl->ifl_rxq;
2029 	if_ctx_t ctx = rxq->ifr_ctx;
2030 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2031 
2032 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2033 	/*
2034 	** Free current RX buffer structs and their mbufs
2035 	*/
2036 	iflib_fl_bufs_free(fl);
2037 	/* Now replenish the mbufs */
2038 	MPASS(fl->ifl_credits == 0);
2039 	/*
2040 	 * XXX don't set the max_frame_size to larger
2041 	 * than the hardware can handle
2042 	 */
2043 	if (sctx->isc_max_frame_size <= 2048)
2044 		fl->ifl_buf_size = MCLBYTES;
2045 #ifndef CONTIGMALLOC_WORKS
2046 	else
2047 		fl->ifl_buf_size = MJUMPAGESIZE;
2048 #else
2049 	else if (sctx->isc_max_frame_size <= 4096)
2050 		fl->ifl_buf_size = MJUMPAGESIZE;
2051 	else if (sctx->isc_max_frame_size <= 9216)
2052 		fl->ifl_buf_size = MJUM9BYTES;
2053 	else
2054 		fl->ifl_buf_size = MJUM16BYTES;
2055 #endif
2056 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2057 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2058 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2059 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2060 
2061 
2062 	/* avoid pre-allocating zillions of clusters to an idle card
2063 	 * potentially speeding up attach
2064 	 */
2065 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2066 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2067 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2068 		return (ENOBUFS);
2069 	/*
2070 	 * handle failure
2071 	 */
2072 	MPASS(rxq != NULL);
2073 	MPASS(fl->ifl_ifdi != NULL);
2074 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2075 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2076 	return (0);
2077 }
2078 
2079 /*********************************************************************
2080  *
2081  *  Free receive ring data structures
2082  *
2083  **********************************************************************/
2084 static void
2085 iflib_rx_sds_free(iflib_rxq_t rxq)
2086 {
2087 	iflib_fl_t fl;
2088 	int i;
2089 
2090 	if (rxq->ifr_fl != NULL) {
2091 		for (i = 0; i < rxq->ifr_nfl; i++) {
2092 			fl = &rxq->ifr_fl[i];
2093 			if (fl->ifl_desc_tag != NULL) {
2094 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2095 				fl->ifl_desc_tag = NULL;
2096 			}
2097 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2098 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2099 			/* XXX destroy maps first */
2100 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2101 			fl->ifl_sds.ifsd_m = NULL;
2102 			fl->ifl_sds.ifsd_cl = NULL;
2103 			fl->ifl_sds.ifsd_map = NULL;
2104 		}
2105 		free(rxq->ifr_fl, M_IFLIB);
2106 		rxq->ifr_fl = NULL;
2107 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2108 	}
2109 }
2110 
2111 /*
2112  * MI independent logic
2113  *
2114  */
2115 static void
2116 iflib_timer(void *arg)
2117 {
2118 	iflib_txq_t txq = arg;
2119 	if_ctx_t ctx = txq->ift_ctx;
2120 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2121 
2122 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2123 		return;
2124 	/*
2125 	** Check on the state of the TX queue(s), this
2126 	** can be done without the lock because its RO
2127 	** and the HUNG state will be static if set.
2128 	*/
2129 	IFDI_TIMER(ctx, txq->ift_id);
2130 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2131 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2132 	     (sctx->isc_pause_frames == 0)))
2133 		goto hung;
2134 
2135 	if (ifmp_ring_is_stalled(txq->ift_br))
2136 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2137 	txq->ift_cleaned_prev = txq->ift_cleaned;
2138 	/* handle any laggards */
2139 	if (txq->ift_db_pending)
2140 		GROUPTASK_ENQUEUE(&txq->ift_task);
2141 
2142 	sctx->isc_pause_frames = 0;
2143 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2144 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2145 	return;
2146 hung:
2147 	CTX_LOCK(ctx);
2148 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2149 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2150 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2151 
2152 	IFDI_WATCHDOG_RESET(ctx);
2153 	ctx->ifc_watchdog_events++;
2154 
2155 	ctx->ifc_flags |= IFC_DO_RESET;
2156 	iflib_admin_intr_deferred(ctx);
2157 	CTX_UNLOCK(ctx);
2158 }
2159 
2160 static void
2161 iflib_init_locked(if_ctx_t ctx)
2162 {
2163 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2164 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2165 	if_t ifp = ctx->ifc_ifp;
2166 	iflib_fl_t fl;
2167 	iflib_txq_t txq;
2168 	iflib_rxq_t rxq;
2169 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2170 
2171 
2172 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2173 	IFDI_INTR_DISABLE(ctx);
2174 
2175 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2176 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2177 	/* Set hardware offload abilities */
2178 	if_clearhwassist(ifp);
2179 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2180 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2181 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2182 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2183 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2184 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2185 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2186 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2187 
2188 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2189 		CALLOUT_LOCK(txq);
2190 		callout_stop(&txq->ift_timer);
2191 		CALLOUT_UNLOCK(txq);
2192 		iflib_netmap_txq_init(ctx, txq);
2193 	}
2194 #ifdef INVARIANTS
2195 	i = if_getdrvflags(ifp);
2196 #endif
2197 	IFDI_INIT(ctx);
2198 	MPASS(if_getdrvflags(ifp) == i);
2199 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2200 		/* XXX this should really be done on a per-queue basis */
2201 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2202 			MPASS(rxq->ifr_id == i);
2203 			iflib_netmap_rxq_init(ctx, rxq);
2204 			continue;
2205 		}
2206 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2207 			if (iflib_fl_setup(fl)) {
2208 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2209 				goto done;
2210 			}
2211 		}
2212 	}
2213 	done:
2214 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2215 	IFDI_INTR_ENABLE(ctx);
2216 	txq = ctx->ifc_txqs;
2217 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2218 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2219 			txq->ift_timer.c_cpu);
2220 }
2221 
2222 static int
2223 iflib_media_change(if_t ifp)
2224 {
2225 	if_ctx_t ctx = if_getsoftc(ifp);
2226 	int err;
2227 
2228 	CTX_LOCK(ctx);
2229 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2230 		iflib_init_locked(ctx);
2231 	CTX_UNLOCK(ctx);
2232 	return (err);
2233 }
2234 
2235 static void
2236 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2237 {
2238 	if_ctx_t ctx = if_getsoftc(ifp);
2239 
2240 	CTX_LOCK(ctx);
2241 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2242 	IFDI_MEDIA_STATUS(ctx, ifmr);
2243 	CTX_UNLOCK(ctx);
2244 }
2245 
2246 static void
2247 iflib_stop(if_ctx_t ctx)
2248 {
2249 	iflib_txq_t txq = ctx->ifc_txqs;
2250 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2251 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2252 	iflib_dma_info_t di;
2253 	iflib_fl_t fl;
2254 	int i, j;
2255 
2256 	/* Tell the stack that the interface is no longer active */
2257 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2258 
2259 	IFDI_INTR_DISABLE(ctx);
2260 	DELAY(1000);
2261 	IFDI_STOP(ctx);
2262 	DELAY(1000);
2263 
2264 	iflib_debug_reset();
2265 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2266 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2267 		/* make sure all transmitters have completed before proceeding XXX */
2268 
2269 		CALLOUT_LOCK(txq);
2270 		callout_stop(&txq->ift_timer);
2271 		CALLOUT_UNLOCK(txq);
2272 
2273 		/* clean any enqueued buffers */
2274 		iflib_ifmp_purge(txq);
2275 		/* Free any existing tx buffers. */
2276 		for (j = 0; j < txq->ift_size; j++) {
2277 			iflib_txsd_free(ctx, txq, j);
2278 		}
2279 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2280 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2281 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2282 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2283 		txq->ift_pullups = 0;
2284 		ifmp_ring_reset_stats(txq->ift_br);
2285 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2286 			bzero((void *)di->idi_vaddr, di->idi_size);
2287 	}
2288 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2289 		/* make sure all transmitters have completed before proceeding XXX */
2290 
2291 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2292 			bzero((void *)di->idi_vaddr, di->idi_size);
2293 		/* also resets the free lists pidx/cidx */
2294 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2295 			iflib_fl_bufs_free(fl);
2296 	}
2297 }
2298 
2299 static inline caddr_t
2300 calc_next_rxd(iflib_fl_t fl, int cidx)
2301 {
2302 	qidx_t size;
2303 	int nrxd;
2304 	caddr_t start, end, cur, next;
2305 
2306 	nrxd = fl->ifl_size;
2307 	size = fl->ifl_rxd_size;
2308 	start = fl->ifl_ifdi->idi_vaddr;
2309 
2310 	if (__predict_false(size == 0))
2311 		return (start);
2312 	cur = start + size*cidx;
2313 	end = start + size*nrxd;
2314 	next = CACHE_PTR_NEXT(cur);
2315 	return (next < end ? next : start);
2316 }
2317 
2318 static inline void
2319 prefetch_pkts(iflib_fl_t fl, int cidx)
2320 {
2321 	int nextptr;
2322 	int nrxd = fl->ifl_size;
2323 	caddr_t next_rxd;
2324 
2325 
2326 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2327 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2328 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2329 	next_rxd = calc_next_rxd(fl, cidx);
2330 	prefetch(next_rxd);
2331 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2332 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2333 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2334 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2335 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2336 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2337 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2338 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2339 }
2340 
2341 static void
2342 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2343 {
2344 	int flid, cidx;
2345 	bus_dmamap_t map;
2346 	iflib_fl_t fl;
2347 	iflib_dma_info_t di;
2348 	int next;
2349 
2350 	map = NULL;
2351 	flid = irf->irf_flid;
2352 	cidx = irf->irf_idx;
2353 	fl = &rxq->ifr_fl[flid];
2354 	sd->ifsd_fl = fl;
2355 	sd->ifsd_cidx = cidx;
2356 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2357 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2358 	fl->ifl_credits--;
2359 #if MEMORY_LOGGING
2360 	fl->ifl_m_dequeued++;
2361 #endif
2362 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2363 		prefetch_pkts(fl, cidx);
2364 	if (fl->ifl_sds.ifsd_map != NULL) {
2365 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2366 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2367 		map = fl->ifl_sds.ifsd_map[cidx];
2368 		di = fl->ifl_ifdi;
2369 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2370 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2371 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2372 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2373 
2374 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2375 		MPASS(fl->ifl_cidx == cidx);
2376 		if (unload)
2377 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2378 	}
2379 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2380 	if (__predict_false(fl->ifl_cidx == 0))
2381 		fl->ifl_gen = 0;
2382 	if (map != NULL)
2383 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2384 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2385         bit_clear(fl->ifl_rx_bitmap, cidx);
2386 }
2387 
2388 static struct mbuf *
2389 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2390 {
2391 	int i, padlen , flags;
2392 	struct mbuf *m, *mh, *mt;
2393 	caddr_t cl;
2394 
2395 	i = 0;
2396 	mh = NULL;
2397 	do {
2398 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2399 
2400 		MPASS(*sd->ifsd_cl != NULL);
2401 		MPASS(*sd->ifsd_m != NULL);
2402 
2403 		/* Don't include zero-length frags */
2404 		if (ri->iri_frags[i].irf_len == 0) {
2405 			/* XXX we can save the cluster here, but not the mbuf */
2406 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2407 			m_free(*sd->ifsd_m);
2408 			*sd->ifsd_m = NULL;
2409 			continue;
2410 		}
2411 		m = *sd->ifsd_m;
2412 		*sd->ifsd_m = NULL;
2413 		if (mh == NULL) {
2414 			flags = M_PKTHDR|M_EXT;
2415 			mh = mt = m;
2416 			padlen = ri->iri_pad;
2417 		} else {
2418 			flags = M_EXT;
2419 			mt->m_next = m;
2420 			mt = m;
2421 			/* assuming padding is only on the first fragment */
2422 			padlen = 0;
2423 		}
2424 		cl = *sd->ifsd_cl;
2425 		*sd->ifsd_cl = NULL;
2426 
2427 		/* Can these two be made one ? */
2428 		m_init(m, M_NOWAIT, MT_DATA, flags);
2429 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2430 		/*
2431 		 * These must follow m_init and m_cljset
2432 		 */
2433 		m->m_data += padlen;
2434 		ri->iri_len -= padlen;
2435 		m->m_len = ri->iri_frags[i].irf_len;
2436 	} while (++i < ri->iri_nfrags);
2437 
2438 	return (mh);
2439 }
2440 
2441 /*
2442  * Process one software descriptor
2443  */
2444 static struct mbuf *
2445 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2446 {
2447 	struct if_rxsd sd;
2448 	struct mbuf *m;
2449 
2450 	/* should I merge this back in now that the two paths are basically duplicated? */
2451 	if (ri->iri_nfrags == 1 &&
2452 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2453 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2454 		m = *sd.ifsd_m;
2455 		*sd.ifsd_m = NULL;
2456 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2457 #ifndef __NO_STRICT_ALIGNMENT
2458 		if (!IP_ALIGNED(m))
2459 			m->m_data += 2;
2460 #endif
2461 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2462 		m->m_len = ri->iri_frags[0].irf_len;
2463        } else {
2464 		m = assemble_segments(rxq, ri, &sd);
2465 	}
2466 	m->m_pkthdr.len = ri->iri_len;
2467 	m->m_pkthdr.rcvif = ri->iri_ifp;
2468 	m->m_flags |= ri->iri_flags;
2469 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2470 	m->m_pkthdr.flowid = ri->iri_flowid;
2471 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2472 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2473 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2474 	return (m);
2475 }
2476 
2477 #if defined(INET6) || defined(INET)
2478 static void
2479 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2480 {
2481 	CURVNET_SET(lc->ifp->if_vnet);
2482 #if defined(INET6)
2483 	*v6 = VNET(ip6_forwarding);
2484 #endif
2485 #if defined(INET)
2486 	*v4 = VNET(ipforwarding);
2487 #endif
2488 	CURVNET_RESTORE();
2489 }
2490 
2491 /*
2492  * Returns true if it's possible this packet could be LROed.
2493  * if it returns false, it is guaranteed that tcp_lro_rx()
2494  * would not return zero.
2495  */
2496 static bool
2497 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2498 {
2499 	struct ether_header *eh;
2500 	uint16_t eh_type;
2501 
2502 	eh = mtod(m, struct ether_header *);
2503 	eh_type = ntohs(eh->ether_type);
2504 	switch (eh_type) {
2505 #if defined(INET6)
2506 		case ETHERTYPE_IPV6:
2507 			return !v6_forwarding;
2508 #endif
2509 #if defined (INET)
2510 		case ETHERTYPE_IP:
2511 			return !v4_forwarding;
2512 #endif
2513 	}
2514 
2515 	return false;
2516 }
2517 #else
2518 static void
2519 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2520 {
2521 }
2522 #endif
2523 
2524 static bool
2525 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2526 {
2527 	if_ctx_t ctx = rxq->ifr_ctx;
2528 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2529 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2530 	int avail, i;
2531 	qidx_t *cidxp;
2532 	struct if_rxd_info ri;
2533 	int err, budget_left, rx_bytes, rx_pkts;
2534 	iflib_fl_t fl;
2535 	struct ifnet *ifp;
2536 	int lro_enabled;
2537 	bool lro_possible = false;
2538 	bool v4_forwarding, v6_forwarding;
2539 
2540 	/*
2541 	 * XXX early demux data packets so that if_input processing only handles
2542 	 * acks in interrupt context
2543 	 */
2544 	struct mbuf *m, *mh, *mt, *mf;
2545 
2546 	ifp = ctx->ifc_ifp;
2547 	mh = mt = NULL;
2548 	MPASS(budget > 0);
2549 	rx_pkts	= rx_bytes = 0;
2550 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2551 		cidxp = &rxq->ifr_cq_cidx;
2552 	else
2553 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2554 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2555 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2556 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2557 		DBG_COUNTER_INC(rx_unavail);
2558 		return (false);
2559 	}
2560 
2561 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2562 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2563 			DBG_COUNTER_INC(rx_ctx_inactive);
2564 			break;
2565 		}
2566 		/*
2567 		 * Reset client set fields to their default values
2568 		 */
2569 		rxd_info_zero(&ri);
2570 		ri.iri_qsidx = rxq->ifr_id;
2571 		ri.iri_cidx = *cidxp;
2572 		ri.iri_ifp = ifp;
2573 		ri.iri_frags = rxq->ifr_frags;
2574 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2575 
2576 		if (err)
2577 			goto err;
2578 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2579 			*cidxp = ri.iri_cidx;
2580 			/* Update our consumer index */
2581 			/* XXX NB: shurd - check if this is still safe */
2582 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2583 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2584 				rxq->ifr_cq_gen = 0;
2585 			}
2586 			/* was this only a completion queue message? */
2587 			if (__predict_false(ri.iri_nfrags == 0))
2588 				continue;
2589 		}
2590 		MPASS(ri.iri_nfrags != 0);
2591 		MPASS(ri.iri_len != 0);
2592 
2593 		/* will advance the cidx on the corresponding free lists */
2594 		m = iflib_rxd_pkt_get(rxq, &ri);
2595 		if (avail == 0 && budget_left)
2596 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2597 
2598 		if (__predict_false(m == NULL)) {
2599 			DBG_COUNTER_INC(rx_mbuf_null);
2600 			continue;
2601 		}
2602 		/* imm_pkt: -- cxgb */
2603 		if (mh == NULL)
2604 			mh = mt = m;
2605 		else {
2606 			mt->m_nextpkt = m;
2607 			mt = m;
2608 		}
2609 	}
2610 	/* make sure that we can refill faster than drain */
2611 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2612 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2613 
2614 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2615 	if (lro_enabled)
2616 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2617 	mt = mf = NULL;
2618 	while (mh != NULL) {
2619 		m = mh;
2620 		mh = mh->m_nextpkt;
2621 		m->m_nextpkt = NULL;
2622 #ifndef __NO_STRICT_ALIGNMENT
2623 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2624 			continue;
2625 #endif
2626 		rx_bytes += m->m_pkthdr.len;
2627 		rx_pkts++;
2628 #if defined(INET6) || defined(INET)
2629 		if (lro_enabled) {
2630 			if (!lro_possible) {
2631 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2632 				if (lro_possible && mf != NULL) {
2633 					ifp->if_input(ifp, mf);
2634 					DBG_COUNTER_INC(rx_if_input);
2635 					mt = mf = NULL;
2636 				}
2637 			}
2638 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2639 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2640 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2641 					continue;
2642 			}
2643 		}
2644 #endif
2645 		if (lro_possible) {
2646 			ifp->if_input(ifp, m);
2647 			DBG_COUNTER_INC(rx_if_input);
2648 			continue;
2649 		}
2650 
2651 		if (mf == NULL)
2652 			mf = m;
2653 		if (mt != NULL)
2654 			mt->m_nextpkt = m;
2655 		mt = m;
2656 	}
2657 	if (mf != NULL) {
2658 		ifp->if_input(ifp, mf);
2659 		DBG_COUNTER_INC(rx_if_input);
2660 	}
2661 
2662 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2663 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2664 
2665 	/*
2666 	 * Flush any outstanding LRO work
2667 	 */
2668 #if defined(INET6) || defined(INET)
2669 	tcp_lro_flush_all(&rxq->ifr_lc);
2670 #endif
2671 	if (avail)
2672 		return true;
2673 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2674 err:
2675 	CTX_LOCK(ctx);
2676 	ctx->ifc_flags |= IFC_DO_RESET;
2677 	iflib_admin_intr_deferred(ctx);
2678 	CTX_UNLOCK(ctx);
2679 	return (false);
2680 }
2681 
2682 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2683 static inline qidx_t
2684 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2685 {
2686 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2687 	qidx_t minthresh = txq->ift_size / 8;
2688 	if (in_use > 4*minthresh)
2689 		return (notify_count);
2690 	if (in_use > 2*minthresh)
2691 		return (notify_count >> 1);
2692 	if (in_use > minthresh)
2693 		return (notify_count >> 3);
2694 	return (0);
2695 }
2696 
2697 static inline qidx_t
2698 txq_max_rs_deferred(iflib_txq_t txq)
2699 {
2700 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2701 	qidx_t minthresh = txq->ift_size / 8;
2702 	if (txq->ift_in_use > 4*minthresh)
2703 		return (notify_count);
2704 	if (txq->ift_in_use > 2*minthresh)
2705 		return (notify_count >> 1);
2706 	if (txq->ift_in_use > minthresh)
2707 		return (notify_count >> 2);
2708 	return (2);
2709 }
2710 
2711 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2712 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2713 
2714 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2715 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2716 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2717 
2718 /* forward compatibility for cxgb */
2719 #define FIRST_QSET(ctx) 0
2720 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2721 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2722 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2723 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2724 
2725 /* XXX we should be setting this to something other than zero */
2726 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2727 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2728 
2729 static inline bool
2730 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2731 {
2732 	qidx_t dbval, max;
2733 	bool rang;
2734 
2735 	rang = false;
2736 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2737 	if (ring || txq->ift_db_pending >= max) {
2738 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2739 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2740 		txq->ift_db_pending = txq->ift_npending = 0;
2741 		rang = true;
2742 	}
2743 	return (rang);
2744 }
2745 
2746 #ifdef PKT_DEBUG
2747 static void
2748 print_pkt(if_pkt_info_t pi)
2749 {
2750 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2751 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2752 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2753 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2754 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2755 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2756 }
2757 #endif
2758 
2759 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2760 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2761 
2762 static int
2763 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2764 {
2765 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2766 	struct ether_vlan_header *eh;
2767 	struct mbuf *m, *n;
2768 
2769 	n = m = *mp;
2770 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2771 	    M_WRITABLE(m) == 0) {
2772 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2773 			return (ENOMEM);
2774 		} else {
2775 			m_freem(*mp);
2776 			n = *mp = m;
2777 		}
2778 	}
2779 
2780 	/*
2781 	 * Determine where frame payload starts.
2782 	 * Jump over vlan headers if already present,
2783 	 * helpful for QinQ too.
2784 	 */
2785 	if (__predict_false(m->m_len < sizeof(*eh))) {
2786 		txq->ift_pullups++;
2787 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2788 			return (ENOMEM);
2789 	}
2790 	eh = mtod(m, struct ether_vlan_header *);
2791 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2792 		pi->ipi_etype = ntohs(eh->evl_proto);
2793 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2794 	} else {
2795 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2796 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2797 	}
2798 
2799 	switch (pi->ipi_etype) {
2800 #ifdef INET
2801 	case ETHERTYPE_IP:
2802 	{
2803 		struct ip *ip = NULL;
2804 		struct tcphdr *th = NULL;
2805 		int minthlen;
2806 
2807 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2808 		if (__predict_false(m->m_len < minthlen)) {
2809 			/*
2810 			 * if this code bloat is causing too much of a hit
2811 			 * move it to a separate function and mark it noinline
2812 			 */
2813 			if (m->m_len == pi->ipi_ehdrlen) {
2814 				n = m->m_next;
2815 				MPASS(n);
2816 				if (n->m_len >= sizeof(*ip))  {
2817 					ip = (struct ip *)n->m_data;
2818 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2819 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2820 				} else {
2821 					txq->ift_pullups++;
2822 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2823 						return (ENOMEM);
2824 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2825 				}
2826 			} else {
2827 				txq->ift_pullups++;
2828 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2829 					return (ENOMEM);
2830 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2831 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2832 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2833 			}
2834 		} else {
2835 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2836 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2837 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2838 		}
2839 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2840 		pi->ipi_ipproto = ip->ip_p;
2841 		pi->ipi_flags |= IPI_TX_IPV4;
2842 
2843 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2844                        ip->ip_sum = 0;
2845 
2846 		if (IS_TSO4(pi)) {
2847 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2848 				if (__predict_false(th == NULL)) {
2849 					txq->ift_pullups++;
2850 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2851 						return (ENOMEM);
2852 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2853 				}
2854 				pi->ipi_tcp_hflags = th->th_flags;
2855 				pi->ipi_tcp_hlen = th->th_off << 2;
2856 				pi->ipi_tcp_seq = th->th_seq;
2857 			}
2858 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2859 				return (ENXIO);
2860 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2861 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2862 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2863 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2864 				ip->ip_sum = 0;
2865 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2866 			}
2867 		}
2868 		break;
2869 	}
2870 #endif
2871 #ifdef INET6
2872 	case ETHERTYPE_IPV6:
2873 	{
2874 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2875 		struct tcphdr *th;
2876 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2877 
2878 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2879 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2880 				return (ENOMEM);
2881 		}
2882 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2883 
2884 		/* XXX-BZ this will go badly in case of ext hdrs. */
2885 		pi->ipi_ipproto = ip6->ip6_nxt;
2886 		pi->ipi_flags |= IPI_TX_IPV6;
2887 
2888 		if (IS_TSO6(pi)) {
2889 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2890 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2891 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2892 						return (ENOMEM);
2893 				}
2894 				pi->ipi_tcp_hflags = th->th_flags;
2895 				pi->ipi_tcp_hlen = th->th_off << 2;
2896 			}
2897 
2898 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2899 				return (ENXIO);
2900 			/*
2901 			 * The corresponding flag is set by the stack in the IPv4
2902 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2903 			 * So, set it here because the rest of the flow requires it.
2904 			 */
2905 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2906 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2907 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2908 		}
2909 		break;
2910 	}
2911 #endif
2912 	default:
2913 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2914 		pi->ipi_ip_hlen = 0;
2915 		break;
2916 	}
2917 	*mp = m;
2918 
2919 	return (0);
2920 }
2921 
2922 static  __noinline  struct mbuf *
2923 collapse_pkthdr(struct mbuf *m0)
2924 {
2925 	struct mbuf *m, *m_next, *tmp;
2926 
2927 	m = m0;
2928 	m_next = m->m_next;
2929 	while (m_next != NULL && m_next->m_len == 0) {
2930 		m = m_next;
2931 		m->m_next = NULL;
2932 		m_free(m);
2933 		m_next = m_next->m_next;
2934 	}
2935 	m = m0;
2936 	m->m_next = m_next;
2937 	if ((m_next->m_flags & M_EXT) == 0) {
2938 		m = m_defrag(m, M_NOWAIT);
2939 	} else {
2940 		tmp = m_next->m_next;
2941 		memcpy(m_next, m, MPKTHSIZE);
2942 		m = m_next;
2943 		m->m_next = tmp;
2944 	}
2945 	return (m);
2946 }
2947 
2948 /*
2949  * If dodgy hardware rejects the scatter gather chain we've handed it
2950  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2951  * m_defrag'd mbufs
2952  */
2953 static __noinline struct mbuf *
2954 iflib_remove_mbuf(iflib_txq_t txq)
2955 {
2956 	int ntxd, i, pidx;
2957 	struct mbuf *m, *mh, **ifsd_m;
2958 
2959 	pidx = txq->ift_pidx;
2960 	ifsd_m = txq->ift_sds.ifsd_m;
2961 	ntxd = txq->ift_size;
2962 	mh = m = ifsd_m[pidx];
2963 	ifsd_m[pidx] = NULL;
2964 #if MEMORY_LOGGING
2965 	txq->ift_dequeued++;
2966 #endif
2967 	i = 1;
2968 
2969 	while (m) {
2970 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2971 #if MEMORY_LOGGING
2972 		txq->ift_dequeued++;
2973 #endif
2974 		m = m->m_next;
2975 		i++;
2976 	}
2977 	return (mh);
2978 }
2979 
2980 static int
2981 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2982 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2983 			  int max_segs, int flags)
2984 {
2985 	if_ctx_t ctx;
2986 	if_shared_ctx_t		sctx;
2987 	if_softc_ctx_t		scctx;
2988 	int i, next, pidx, err, ntxd, count;
2989 	struct mbuf *m, *tmp, **ifsd_m;
2990 
2991 	m = *m0;
2992 
2993 	/*
2994 	 * Please don't ever do this
2995 	 */
2996 	if (__predict_false(m->m_len == 0))
2997 		*m0 = m = collapse_pkthdr(m);
2998 
2999 	ctx = txq->ift_ctx;
3000 	sctx = ctx->ifc_sctx;
3001 	scctx = &ctx->ifc_softc_ctx;
3002 	ifsd_m = txq->ift_sds.ifsd_m;
3003 	ntxd = txq->ift_size;
3004 	pidx = txq->ift_pidx;
3005 	if (map != NULL) {
3006 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3007 
3008 		err = bus_dmamap_load_mbuf_sg(tag, map,
3009 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3010 		if (err)
3011 			return (err);
3012 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3013 		count = 0;
3014 		m = *m0;
3015 		do {
3016 			if (__predict_false(m->m_len <= 0)) {
3017 				tmp = m;
3018 				m = m->m_next;
3019 				tmp->m_next = NULL;
3020 				m_free(tmp);
3021 				continue;
3022 			}
3023 			m = m->m_next;
3024 			count++;
3025 		} while (m != NULL);
3026 		if (count > *nsegs) {
3027 			ifsd_m[pidx] = *m0;
3028 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3029 			return (0);
3030 		}
3031 		m = *m0;
3032 		count = 0;
3033 		do {
3034 			next = (pidx + count) & (ntxd-1);
3035 			MPASS(ifsd_m[next] == NULL);
3036 			ifsd_m[next] = m;
3037 			count++;
3038 			tmp = m;
3039 			m = m->m_next;
3040 		} while (m != NULL);
3041 	} else {
3042 		int buflen, sgsize, maxsegsz, max_sgsize;
3043 		vm_offset_t vaddr;
3044 		vm_paddr_t curaddr;
3045 
3046 		count = i = 0;
3047 		m = *m0;
3048 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3049 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3050 		else
3051 			maxsegsz = sctx->isc_tx_maxsegsize;
3052 
3053 		do {
3054 			if (__predict_false(m->m_len <= 0)) {
3055 				tmp = m;
3056 				m = m->m_next;
3057 				tmp->m_next = NULL;
3058 				m_free(tmp);
3059 				continue;
3060 			}
3061 			buflen = m->m_len;
3062 			vaddr = (vm_offset_t)m->m_data;
3063 			/*
3064 			 * see if we can't be smarter about physically
3065 			 * contiguous mappings
3066 			 */
3067 			next = (pidx + count) & (ntxd-1);
3068 			MPASS(ifsd_m[next] == NULL);
3069 #if MEMORY_LOGGING
3070 			txq->ift_enqueued++;
3071 #endif
3072 			ifsd_m[next] = m;
3073 			while (buflen > 0) {
3074 				if (i >= max_segs)
3075 					goto err;
3076 				max_sgsize = MIN(buflen, maxsegsz);
3077 				curaddr = pmap_kextract(vaddr);
3078 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3079 				sgsize = MIN(sgsize, max_sgsize);
3080 				segs[i].ds_addr = curaddr;
3081 				segs[i].ds_len = sgsize;
3082 				vaddr += sgsize;
3083 				buflen -= sgsize;
3084 				i++;
3085 			}
3086 			count++;
3087 			tmp = m;
3088 			m = m->m_next;
3089 		} while (m != NULL);
3090 		*nsegs = i;
3091 	}
3092 	return (0);
3093 err:
3094 	*m0 = iflib_remove_mbuf(txq);
3095 	return (EFBIG);
3096 }
3097 
3098 static inline caddr_t
3099 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3100 {
3101 	qidx_t size;
3102 	int ntxd;
3103 	caddr_t start, end, cur, next;
3104 
3105 	ntxd = txq->ift_size;
3106 	size = txq->ift_txd_size[qid];
3107 	start = txq->ift_ifdi[qid].idi_vaddr;
3108 
3109 	if (__predict_false(size == 0))
3110 		return (start);
3111 	cur = start + size*cidx;
3112 	end = start + size*ntxd;
3113 	next = CACHE_PTR_NEXT(cur);
3114 	return (next < end ? next : start);
3115 }
3116 
3117 /*
3118  * Pad an mbuf to ensure a minimum ethernet frame size.
3119  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3120  */
3121 static __noinline int
3122 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3123 {
3124 	/*
3125 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3126 	 * and ARP message is the smallest common payload I can think of
3127 	 */
3128 	static char pad[18];	/* just zeros */
3129 	int n;
3130 	struct mbuf *new_head;
3131 
3132 	if (!M_WRITABLE(*m_head)) {
3133 		new_head = m_dup(*m_head, M_NOWAIT);
3134 		if (new_head == NULL) {
3135 			m_freem(*m_head);
3136 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3137 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3138 			return ENOMEM;
3139 		}
3140 		m_freem(*m_head);
3141 		*m_head = new_head;
3142 	}
3143 
3144 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3145 	     n > 0; n -= sizeof(pad))
3146 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3147 			break;
3148 
3149 	if (n > 0) {
3150 		m_freem(*m_head);
3151 		device_printf(dev, "cannot pad short frame\n");
3152 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3153 		return (ENOBUFS);
3154 	}
3155 
3156 	return 0;
3157 }
3158 
3159 static int
3160 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3161 {
3162 	if_ctx_t		ctx;
3163 	if_shared_ctx_t		sctx;
3164 	if_softc_ctx_t		scctx;
3165 	bus_dma_segment_t	*segs;
3166 	struct mbuf		*m_head;
3167 	void			*next_txd;
3168 	bus_dmamap_t		map;
3169 	struct if_pkt_info	pi;
3170 	int remap = 0;
3171 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3172 	bus_dma_tag_t desc_tag;
3173 
3174 	segs = txq->ift_segs;
3175 	ctx = txq->ift_ctx;
3176 	sctx = ctx->ifc_sctx;
3177 	scctx = &ctx->ifc_softc_ctx;
3178 	segs = txq->ift_segs;
3179 	ntxd = txq->ift_size;
3180 	m_head = *m_headp;
3181 	map = NULL;
3182 
3183 	/*
3184 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3185 	 */
3186 	cidx = txq->ift_cidx;
3187 	pidx = txq->ift_pidx;
3188 	if (ctx->ifc_flags & IFC_PREFETCH) {
3189 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3190 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3191 			next_txd = calc_next_txd(txq, cidx, 0);
3192 			prefetch(next_txd);
3193 		}
3194 
3195 		/* prefetch the next cache line of mbuf pointers and flags */
3196 		prefetch(&txq->ift_sds.ifsd_m[next]);
3197 		if (txq->ift_sds.ifsd_map != NULL) {
3198 			prefetch(&txq->ift_sds.ifsd_map[next]);
3199 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3200 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3201 		}
3202 	} else if (txq->ift_sds.ifsd_map != NULL)
3203 		map = txq->ift_sds.ifsd_map[pidx];
3204 
3205 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3206 		desc_tag = txq->ift_tso_desc_tag;
3207 		max_segs = scctx->isc_tx_tso_segments_max;
3208 	} else {
3209 		desc_tag = txq->ift_desc_tag;
3210 		max_segs = scctx->isc_tx_nsegments;
3211 	}
3212 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3213 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3214 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3215 		if (err)
3216 			return err;
3217 	}
3218 	m_head = *m_headp;
3219 
3220 	pkt_info_zero(&pi);
3221 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3222 	pi.ipi_pidx = pidx;
3223 	pi.ipi_qsidx = txq->ift_id;
3224 	pi.ipi_len = m_head->m_pkthdr.len;
3225 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3226 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3227 
3228 	/* deliberate bitwise OR to make one condition */
3229 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3230 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3231 			return (err);
3232 		m_head = *m_headp;
3233 	}
3234 
3235 retry:
3236 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3237 defrag:
3238 	if (__predict_false(err)) {
3239 		switch (err) {
3240 		case EFBIG:
3241 			/* try collapse once and defrag once */
3242 			if (remap == 0)
3243 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3244 			if (remap == 1)
3245 				m_head = m_defrag(*m_headp, M_NOWAIT);
3246 			remap++;
3247 			if (__predict_false(m_head == NULL))
3248 				goto defrag_failed;
3249 			txq->ift_mbuf_defrag++;
3250 			*m_headp = m_head;
3251 			goto retry;
3252 			break;
3253 		case ENOMEM:
3254 			txq->ift_no_tx_dma_setup++;
3255 			break;
3256 		default:
3257 			txq->ift_no_tx_dma_setup++;
3258 			m_freem(*m_headp);
3259 			DBG_COUNTER_INC(tx_frees);
3260 			*m_headp = NULL;
3261 			break;
3262 		}
3263 		txq->ift_map_failed++;
3264 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3265 		return (err);
3266 	}
3267 
3268 	/*
3269 	 * XXX assumes a 1 to 1 relationship between segments and
3270 	 *        descriptors - this does not hold true on all drivers, e.g.
3271 	 *        cxgb
3272 	 */
3273 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3274 		txq->ift_no_desc_avail++;
3275 		if (map != NULL)
3276 			bus_dmamap_unload(desc_tag, map);
3277 		DBG_COUNTER_INC(encap_txq_avail_fail);
3278 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3279 			GROUPTASK_ENQUEUE(&txq->ift_task);
3280 		return (ENOBUFS);
3281 	}
3282 	/*
3283 	 * On Intel cards we can greatly reduce the number of TX interrupts
3284 	 * we see by only setting report status on every Nth descriptor.
3285 	 * However, this also means that the driver will need to keep track
3286 	 * of the descriptors that RS was set on to check them for the DD bit.
3287 	 */
3288 	txq->ift_rs_pending += nsegs + 1;
3289 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3290 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3291 		pi.ipi_flags |= IPI_TX_INTR;
3292 		txq->ift_rs_pending = 0;
3293 	}
3294 
3295 	pi.ipi_segs = segs;
3296 	pi.ipi_nsegs = nsegs;
3297 
3298 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3299 #ifdef PKT_DEBUG
3300 	print_pkt(&pi);
3301 #endif
3302 	if (map != NULL)
3303 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3304 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3305 		if (map != NULL)
3306 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3307 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3308 		DBG_COUNTER_INC(tx_encap);
3309 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3310 
3311 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3312 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3313 			ndesc += txq->ift_size;
3314 			txq->ift_gen = 1;
3315 		}
3316 		/*
3317 		 * drivers can need as many as
3318 		 * two sentinels
3319 		 */
3320 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3321 		MPASS(pi.ipi_new_pidx != pidx);
3322 		MPASS(ndesc > 0);
3323 		txq->ift_in_use += ndesc;
3324 
3325 		/*
3326 		 * We update the last software descriptor again here because there may
3327 		 * be a sentinel and/or there may be more mbufs than segments
3328 		 */
3329 		txq->ift_pidx = pi.ipi_new_pidx;
3330 		txq->ift_npending += pi.ipi_ndescs;
3331 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3332 		*m_headp = m_head = iflib_remove_mbuf(txq);
3333 		remap = 1;
3334 		txq->ift_txd_encap_efbig++;
3335 		goto defrag;
3336 	} else
3337 		DBG_COUNTER_INC(encap_txd_encap_fail);
3338 	return (err);
3339 
3340 defrag_failed:
3341 	txq->ift_mbuf_defrag_failed++;
3342 	txq->ift_map_failed++;
3343 	m_freem(*m_headp);
3344 	DBG_COUNTER_INC(tx_frees);
3345 	*m_headp = NULL;
3346 	return (ENOMEM);
3347 }
3348 
3349 static void
3350 iflib_tx_desc_free(iflib_txq_t txq, int n)
3351 {
3352 	int hasmap;
3353 	uint32_t qsize, cidx, mask, gen;
3354 	struct mbuf *m, **ifsd_m;
3355 	uint8_t *ifsd_flags;
3356 	bus_dmamap_t *ifsd_map;
3357 	bool do_prefetch;
3358 
3359 	cidx = txq->ift_cidx;
3360 	gen = txq->ift_gen;
3361 	qsize = txq->ift_size;
3362 	mask = qsize-1;
3363 	hasmap = txq->ift_sds.ifsd_map != NULL;
3364 	ifsd_flags = txq->ift_sds.ifsd_flags;
3365 	ifsd_m = txq->ift_sds.ifsd_m;
3366 	ifsd_map = txq->ift_sds.ifsd_map;
3367 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3368 
3369 	while (n--) {
3370 		if (do_prefetch) {
3371 			prefetch(ifsd_m[(cidx + 3) & mask]);
3372 			prefetch(ifsd_m[(cidx + 4) & mask]);
3373 		}
3374 		if (ifsd_m[cidx] != NULL) {
3375 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3376 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3377 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3378 				/*
3379 				 * does it matter if it's not the TSO tag? If so we'll
3380 				 * have to add the type to flags
3381 				 */
3382 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3383 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3384 			}
3385 			if ((m = ifsd_m[cidx]) != NULL) {
3386 				/* XXX we don't support any drivers that batch packets yet */
3387 				MPASS(m->m_nextpkt == NULL);
3388 				/* if the number of clusters exceeds the number of segments
3389 				 * there won't be space on the ring to save a pointer to each
3390 				 * cluster so we simply free the list here
3391 				 */
3392 				if (m->m_flags & M_TOOBIG) {
3393 					m_freem(m);
3394 				} else {
3395 					m_free(m);
3396 				}
3397 				ifsd_m[cidx] = NULL;
3398 #if MEMORY_LOGGING
3399 				txq->ift_dequeued++;
3400 #endif
3401 				DBG_COUNTER_INC(tx_frees);
3402 			}
3403 		}
3404 		if (__predict_false(++cidx == qsize)) {
3405 			cidx = 0;
3406 			gen = 0;
3407 		}
3408 	}
3409 	txq->ift_cidx = cidx;
3410 	txq->ift_gen = gen;
3411 }
3412 
3413 static __inline int
3414 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3415 {
3416 	int reclaim;
3417 	if_ctx_t ctx = txq->ift_ctx;
3418 
3419 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3420 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3421 
3422 	/*
3423 	 * Need a rate-limiting check so that this isn't called every time
3424 	 */
3425 	iflib_tx_credits_update(ctx, txq);
3426 	reclaim = DESC_RECLAIMABLE(txq);
3427 
3428 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3429 #ifdef INVARIANTS
3430 		if (iflib_verbose_debug) {
3431 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3432 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3433 			       reclaim, thresh);
3434 
3435 		}
3436 #endif
3437 		return (0);
3438 	}
3439 	iflib_tx_desc_free(txq, reclaim);
3440 	txq->ift_cleaned += reclaim;
3441 	txq->ift_in_use -= reclaim;
3442 
3443 	return (reclaim);
3444 }
3445 
3446 static struct mbuf **
3447 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3448 {
3449 	int next, size;
3450 	struct mbuf **items;
3451 
3452 	size = r->size;
3453 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3454 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3455 
3456 	prefetch(items[(cidx + offset) & (size-1)]);
3457 	if (remaining > 1) {
3458 		prefetch2cachelines(&items[next]);
3459 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3460 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3461 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3462 	}
3463 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3464 }
3465 
3466 static void
3467 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3468 {
3469 
3470 	ifmp_ring_check_drainage(txq->ift_br, budget);
3471 }
3472 
3473 static uint32_t
3474 iflib_txq_can_drain(struct ifmp_ring *r)
3475 {
3476 	iflib_txq_t txq = r->cookie;
3477 	if_ctx_t ctx = txq->ift_ctx;
3478 
3479 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3480 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3481 }
3482 
3483 static uint32_t
3484 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3485 {
3486 	iflib_txq_t txq = r->cookie;
3487 	if_ctx_t ctx = txq->ift_ctx;
3488 	struct ifnet *ifp = ctx->ifc_ifp;
3489 	struct mbuf **mp, *m;
3490 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3491 	int reclaimed, err, in_use_prev, desc_used;
3492 	bool do_prefetch, ring, rang;
3493 
3494 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3495 			    !LINK_ACTIVE(ctx))) {
3496 		DBG_COUNTER_INC(txq_drain_notready);
3497 		return (0);
3498 	}
3499 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3500 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3501 	avail = IDXDIFF(pidx, cidx, r->size);
3502 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3503 		DBG_COUNTER_INC(txq_drain_flushing);
3504 		for (i = 0; i < avail; i++) {
3505 			m_free(r->items[(cidx + i) & (r->size-1)]);
3506 			r->items[(cidx + i) & (r->size-1)] = NULL;
3507 		}
3508 		return (avail);
3509 	}
3510 
3511 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3512 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3513 		CALLOUT_LOCK(txq);
3514 		callout_stop(&txq->ift_timer);
3515 		CALLOUT_UNLOCK(txq);
3516 		DBG_COUNTER_INC(txq_drain_oactive);
3517 		return (0);
3518 	}
3519 	if (reclaimed)
3520 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3521 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3522 	count = MIN(avail, TX_BATCH_SIZE);
3523 #ifdef INVARIANTS
3524 	if (iflib_verbose_debug)
3525 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3526 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3527 #endif
3528 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3529 	avail = TXQ_AVAIL(txq);
3530 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3531 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3532 
3533 		mp = _ring_peek_one(r, cidx, i, rem);
3534 		MPASS(mp != NULL && *mp != NULL);
3535 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3536 			consumed++;
3537 			reclaimed++;
3538 			continue;
3539 		}
3540 		in_use_prev = txq->ift_in_use;
3541 		pidx_prev = txq->ift_pidx;
3542 		err = iflib_encap(txq, mp);
3543 		if (__predict_false(err)) {
3544 			DBG_COUNTER_INC(txq_drain_encapfail);
3545 			/* no room - bail out */
3546 			if (err == ENOBUFS)
3547 				break;
3548 			consumed++;
3549 			DBG_COUNTER_INC(txq_drain_encapfail);
3550 			/* we can't send this packet - skip it */
3551 			continue;
3552 		}
3553 		consumed++;
3554 		pkt_sent++;
3555 		m = *mp;
3556 		DBG_COUNTER_INC(tx_sent);
3557 		bytes_sent += m->m_pkthdr.len;
3558 		mcast_sent += !!(m->m_flags & M_MCAST);
3559 		avail = TXQ_AVAIL(txq);
3560 
3561 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3562 		desc_used += (txq->ift_in_use - in_use_prev);
3563 		ETHER_BPF_MTAP(ifp, m);
3564 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3565 			break;
3566 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3567 	}
3568 
3569 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3570 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3571 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3572 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3573 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3574 	if (mcast_sent)
3575 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3576 #ifdef INVARIANTS
3577 	if (iflib_verbose_debug)
3578 		printf("consumed=%d\n", consumed);
3579 #endif
3580 	return (consumed);
3581 }
3582 
3583 static uint32_t
3584 iflib_txq_drain_always(struct ifmp_ring *r)
3585 {
3586 	return (1);
3587 }
3588 
3589 static uint32_t
3590 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3591 {
3592 	int i, avail;
3593 	struct mbuf **mp;
3594 	iflib_txq_t txq;
3595 
3596 	txq = r->cookie;
3597 
3598 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3599 	CALLOUT_LOCK(txq);
3600 	callout_stop(&txq->ift_timer);
3601 	CALLOUT_UNLOCK(txq);
3602 
3603 	avail = IDXDIFF(pidx, cidx, r->size);
3604 	for (i = 0; i < avail; i++) {
3605 		mp = _ring_peek_one(r, cidx, i, avail - i);
3606 		if (__predict_false(*mp == (struct mbuf *)txq))
3607 			continue;
3608 		m_freem(*mp);
3609 	}
3610 	MPASS(ifmp_ring_is_stalled(r) == 0);
3611 	return (avail);
3612 }
3613 
3614 static void
3615 iflib_ifmp_purge(iflib_txq_t txq)
3616 {
3617 	struct ifmp_ring *r;
3618 
3619 	r = txq->ift_br;
3620 	r->drain = iflib_txq_drain_free;
3621 	r->can_drain = iflib_txq_drain_always;
3622 
3623 	ifmp_ring_check_drainage(r, r->size);
3624 
3625 	r->drain = iflib_txq_drain;
3626 	r->can_drain = iflib_txq_can_drain;
3627 }
3628 
3629 static void
3630 _task_fn_tx(void *context)
3631 {
3632 	iflib_txq_t txq = context;
3633 	if_ctx_t ctx = txq->ift_ctx;
3634 	struct ifnet *ifp = ctx->ifc_ifp;
3635 	int rc;
3636 
3637 #ifdef IFLIB_DIAGNOSTICS
3638 	txq->ift_cpu_exec_count[curcpu]++;
3639 #endif
3640 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3641 		return;
3642 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3643 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3644 			netmap_tx_irq(ifp, txq->ift_id);
3645 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3646 		return;
3647 	}
3648 	if (txq->ift_db_pending)
3649 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3650 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3651 	if (ctx->ifc_flags & IFC_LEGACY)
3652 		IFDI_INTR_ENABLE(ctx);
3653 	else {
3654 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3655 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3656 	}
3657 }
3658 
3659 static void
3660 _task_fn_rx(void *context)
3661 {
3662 	iflib_rxq_t rxq = context;
3663 	if_ctx_t ctx = rxq->ifr_ctx;
3664 	bool more;
3665 	int rc;
3666 	uint16_t budget;
3667 
3668 #ifdef IFLIB_DIAGNOSTICS
3669 	rxq->ifr_cpu_exec_count[curcpu]++;
3670 #endif
3671 	DBG_COUNTER_INC(task_fn_rxs);
3672 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3673 		return;
3674 	more = true;
3675 #ifdef DEV_NETMAP
3676 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3677 		u_int work = 0;
3678 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3679 			more = false;
3680 		}
3681 	}
3682 #endif
3683 	budget = ctx->ifc_sysctl_rx_budget;
3684 	if (budget == 0)
3685 		budget = 16;	/* XXX */
3686 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3687 		if (ctx->ifc_flags & IFC_LEGACY)
3688 			IFDI_INTR_ENABLE(ctx);
3689 		else {
3690 			DBG_COUNTER_INC(rx_intr_enables);
3691 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3692 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3693 		}
3694 	}
3695 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3696 		return;
3697 	if (more)
3698 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3699 }
3700 
3701 static void
3702 _task_fn_admin(void *context)
3703 {
3704 	if_ctx_t ctx = context;
3705 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3706 	iflib_txq_t txq;
3707 	int i;
3708 
3709 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3710 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3711 			return;
3712 		}
3713 	}
3714 
3715 	CTX_LOCK(ctx);
3716 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3717 		CALLOUT_LOCK(txq);
3718 		callout_stop(&txq->ift_timer);
3719 		CALLOUT_UNLOCK(txq);
3720 	}
3721 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3722 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3723 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3724 	IFDI_LINK_INTR_ENABLE(ctx);
3725 	if (ctx->ifc_flags & IFC_DO_RESET) {
3726 		ctx->ifc_flags &= ~IFC_DO_RESET;
3727 		iflib_if_init_locked(ctx);
3728 	}
3729 	CTX_UNLOCK(ctx);
3730 
3731 	if (LINK_ACTIVE(ctx) == 0)
3732 		return;
3733 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3734 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3735 }
3736 
3737 
3738 static void
3739 _task_fn_iov(void *context)
3740 {
3741 	if_ctx_t ctx = context;
3742 
3743 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3744 		return;
3745 
3746 	CTX_LOCK(ctx);
3747 	IFDI_VFLR_HANDLE(ctx);
3748 	CTX_UNLOCK(ctx);
3749 }
3750 
3751 static int
3752 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3753 {
3754 	int err;
3755 	if_int_delay_info_t info;
3756 	if_ctx_t ctx;
3757 
3758 	info = (if_int_delay_info_t)arg1;
3759 	ctx = info->iidi_ctx;
3760 	info->iidi_req = req;
3761 	info->iidi_oidp = oidp;
3762 	CTX_LOCK(ctx);
3763 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3764 	CTX_UNLOCK(ctx);
3765 	return (err);
3766 }
3767 
3768 /*********************************************************************
3769  *
3770  *  IFNET FUNCTIONS
3771  *
3772  **********************************************************************/
3773 
3774 static void
3775 iflib_if_init_locked(if_ctx_t ctx)
3776 {
3777 	iflib_stop(ctx);
3778 	iflib_init_locked(ctx);
3779 }
3780 
3781 
3782 static void
3783 iflib_if_init(void *arg)
3784 {
3785 	if_ctx_t ctx = arg;
3786 
3787 	CTX_LOCK(ctx);
3788 	iflib_if_init_locked(ctx);
3789 	CTX_UNLOCK(ctx);
3790 }
3791 
3792 static int
3793 iflib_if_transmit(if_t ifp, struct mbuf *m)
3794 {
3795 	if_ctx_t	ctx = if_getsoftc(ifp);
3796 
3797 	iflib_txq_t txq;
3798 	int err, qidx;
3799 
3800 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3801 		DBG_COUNTER_INC(tx_frees);
3802 		m_freem(m);
3803 		return (ENOBUFS);
3804 	}
3805 
3806 	MPASS(m->m_nextpkt == NULL);
3807 	qidx = 0;
3808 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3809 		qidx = QIDX(ctx, m);
3810 	/*
3811 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3812 	 */
3813 	txq = &ctx->ifc_txqs[qidx];
3814 
3815 #ifdef DRIVER_BACKPRESSURE
3816 	if (txq->ift_closed) {
3817 		while (m != NULL) {
3818 			next = m->m_nextpkt;
3819 			m->m_nextpkt = NULL;
3820 			m_freem(m);
3821 			m = next;
3822 		}
3823 		return (ENOBUFS);
3824 	}
3825 #endif
3826 #ifdef notyet
3827 	qidx = count = 0;
3828 	mp = marr;
3829 	next = m;
3830 	do {
3831 		count++;
3832 		next = next->m_nextpkt;
3833 	} while (next != NULL);
3834 
3835 	if (count > nitems(marr))
3836 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3837 			/* XXX check nextpkt */
3838 			m_freem(m);
3839 			/* XXX simplify for now */
3840 			DBG_COUNTER_INC(tx_frees);
3841 			return (ENOBUFS);
3842 		}
3843 	for (next = m, i = 0; next != NULL; i++) {
3844 		mp[i] = next;
3845 		next = next->m_nextpkt;
3846 		mp[i]->m_nextpkt = NULL;
3847 	}
3848 #endif
3849 	DBG_COUNTER_INC(tx_seen);
3850 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3851 
3852 	GROUPTASK_ENQUEUE(&txq->ift_task);
3853 	if (err) {
3854 		/* support forthcoming later */
3855 #ifdef DRIVER_BACKPRESSURE
3856 		txq->ift_closed = TRUE;
3857 #endif
3858 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3859 		m_freem(m);
3860 	}
3861 
3862 	return (err);
3863 }
3864 
3865 static void
3866 iflib_if_qflush(if_t ifp)
3867 {
3868 	if_ctx_t ctx = if_getsoftc(ifp);
3869 	iflib_txq_t txq = ctx->ifc_txqs;
3870 	int i;
3871 
3872 	CTX_LOCK(ctx);
3873 	ctx->ifc_flags |= IFC_QFLUSH;
3874 	CTX_UNLOCK(ctx);
3875 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3876 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3877 			iflib_txq_check_drain(txq, 0);
3878 	CTX_LOCK(ctx);
3879 	ctx->ifc_flags &= ~IFC_QFLUSH;
3880 	CTX_UNLOCK(ctx);
3881 
3882 	if_qflush(ifp);
3883 }
3884 
3885 
3886 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3887 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3888 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3889 
3890 static int
3891 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3892 {
3893 	if_ctx_t ctx = if_getsoftc(ifp);
3894 	struct ifreq	*ifr = (struct ifreq *)data;
3895 #if defined(INET) || defined(INET6)
3896 	struct ifaddr	*ifa = (struct ifaddr *)data;
3897 #endif
3898 	bool		avoid_reset = FALSE;
3899 	int		err = 0, reinit = 0, bits;
3900 
3901 	switch (command) {
3902 	case SIOCSIFADDR:
3903 #ifdef INET
3904 		if (ifa->ifa_addr->sa_family == AF_INET)
3905 			avoid_reset = TRUE;
3906 #endif
3907 #ifdef INET6
3908 		if (ifa->ifa_addr->sa_family == AF_INET6)
3909 			avoid_reset = TRUE;
3910 #endif
3911 		/*
3912 		** Calling init results in link renegotiation,
3913 		** so we avoid doing it when possible.
3914 		*/
3915 		if (avoid_reset) {
3916 			if_setflagbits(ifp, IFF_UP,0);
3917 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3918 				reinit = 1;
3919 #ifdef INET
3920 			if (!(if_getflags(ifp) & IFF_NOARP))
3921 				arp_ifinit(ifp, ifa);
3922 #endif
3923 		} else
3924 			err = ether_ioctl(ifp, command, data);
3925 		break;
3926 	case SIOCSIFMTU:
3927 		CTX_LOCK(ctx);
3928 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3929 			CTX_UNLOCK(ctx);
3930 			break;
3931 		}
3932 		bits = if_getdrvflags(ifp);
3933 		/* stop the driver and free any clusters before proceeding */
3934 		iflib_stop(ctx);
3935 
3936 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3937 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3938 				ctx->ifc_flags |= IFC_MULTISEG;
3939 			else
3940 				ctx->ifc_flags &= ~IFC_MULTISEG;
3941 			err = if_setmtu(ifp, ifr->ifr_mtu);
3942 		}
3943 		iflib_init_locked(ctx);
3944 		if_setdrvflags(ifp, bits);
3945 		CTX_UNLOCK(ctx);
3946 		break;
3947 	case SIOCSIFFLAGS:
3948 		CTX_LOCK(ctx);
3949 		if (if_getflags(ifp) & IFF_UP) {
3950 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3951 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3952 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3953 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3954 				}
3955 			} else
3956 				reinit = 1;
3957 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3958 			iflib_stop(ctx);
3959 		}
3960 		ctx->ifc_if_flags = if_getflags(ifp);
3961 		CTX_UNLOCK(ctx);
3962 		break;
3963 	case SIOCADDMULTI:
3964 	case SIOCDELMULTI:
3965 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3966 			CTX_LOCK(ctx);
3967 			IFDI_INTR_DISABLE(ctx);
3968 			IFDI_MULTI_SET(ctx);
3969 			IFDI_INTR_ENABLE(ctx);
3970 			CTX_UNLOCK(ctx);
3971 		}
3972 		break;
3973 	case SIOCSIFMEDIA:
3974 		CTX_LOCK(ctx);
3975 		IFDI_MEDIA_SET(ctx);
3976 		CTX_UNLOCK(ctx);
3977 		/* falls thru */
3978 	case SIOCGIFMEDIA:
3979 	case SIOCGIFXMEDIA:
3980 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3981 		break;
3982 	case SIOCGI2C:
3983 	{
3984 		struct ifi2creq i2c;
3985 
3986 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
3987 		if (err != 0)
3988 			break;
3989 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3990 			err = EINVAL;
3991 			break;
3992 		}
3993 		if (i2c.len > sizeof(i2c.data)) {
3994 			err = EINVAL;
3995 			break;
3996 		}
3997 
3998 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3999 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4000 			    sizeof(i2c));
4001 		break;
4002 	}
4003 	case SIOCSIFCAP:
4004 	{
4005 		int mask, setmask;
4006 
4007 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4008 		setmask = 0;
4009 #ifdef TCP_OFFLOAD
4010 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4011 #endif
4012 		setmask |= (mask & IFCAP_FLAGS);
4013 
4014 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4015 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4016 		if ((mask & IFCAP_WOL) &&
4017 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4018 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4019 		if_vlancap(ifp);
4020 		/*
4021 		 * want to ensure that traffic has stopped before we change any of the flags
4022 		 */
4023 		if (setmask) {
4024 			CTX_LOCK(ctx);
4025 			bits = if_getdrvflags(ifp);
4026 			if (bits & IFF_DRV_RUNNING)
4027 				iflib_stop(ctx);
4028 			if_togglecapenable(ifp, setmask);
4029 			if (bits & IFF_DRV_RUNNING)
4030 				iflib_init_locked(ctx);
4031 			if_setdrvflags(ifp, bits);
4032 			CTX_UNLOCK(ctx);
4033 		}
4034 		break;
4035 	    }
4036 	case SIOCGPRIVATE_0:
4037 	case SIOCSDRVSPEC:
4038 	case SIOCGDRVSPEC:
4039 		CTX_LOCK(ctx);
4040 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4041 		CTX_UNLOCK(ctx);
4042 		break;
4043 	default:
4044 		err = ether_ioctl(ifp, command, data);
4045 		break;
4046 	}
4047 	if (reinit)
4048 		iflib_if_init(ctx);
4049 	return (err);
4050 }
4051 
4052 static uint64_t
4053 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4054 {
4055 	if_ctx_t ctx = if_getsoftc(ifp);
4056 
4057 	return (IFDI_GET_COUNTER(ctx, cnt));
4058 }
4059 
4060 /*********************************************************************
4061  *
4062  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4063  *
4064  **********************************************************************/
4065 
4066 static void
4067 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4068 {
4069 	if_ctx_t ctx = if_getsoftc(ifp);
4070 
4071 	if ((void *)ctx != arg)
4072 		return;
4073 
4074 	if ((vtag == 0) || (vtag > 4095))
4075 		return;
4076 
4077 	CTX_LOCK(ctx);
4078 	IFDI_VLAN_REGISTER(ctx, vtag);
4079 	/* Re-init to load the changes */
4080 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4081 		iflib_if_init_locked(ctx);
4082 	CTX_UNLOCK(ctx);
4083 }
4084 
4085 static void
4086 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4087 {
4088 	if_ctx_t ctx = if_getsoftc(ifp);
4089 
4090 	if ((void *)ctx != arg)
4091 		return;
4092 
4093 	if ((vtag == 0) || (vtag > 4095))
4094 		return;
4095 
4096 	CTX_LOCK(ctx);
4097 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4098 	/* Re-init to load the changes */
4099 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4100 		iflib_if_init_locked(ctx);
4101 	CTX_UNLOCK(ctx);
4102 }
4103 
4104 static void
4105 iflib_led_func(void *arg, int onoff)
4106 {
4107 	if_ctx_t ctx = arg;
4108 
4109 	CTX_LOCK(ctx);
4110 	IFDI_LED_FUNC(ctx, onoff);
4111 	CTX_UNLOCK(ctx);
4112 }
4113 
4114 /*********************************************************************
4115  *
4116  *  BUS FUNCTION DEFINITIONS
4117  *
4118  **********************************************************************/
4119 
4120 int
4121 iflib_device_probe(device_t dev)
4122 {
4123 	pci_vendor_info_t *ent;
4124 
4125 	uint16_t	pci_vendor_id, pci_device_id;
4126 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4127 	uint16_t	pci_rev_id;
4128 	if_shared_ctx_t sctx;
4129 
4130 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4131 		return (ENOTSUP);
4132 
4133 	pci_vendor_id = pci_get_vendor(dev);
4134 	pci_device_id = pci_get_device(dev);
4135 	pci_subvendor_id = pci_get_subvendor(dev);
4136 	pci_subdevice_id = pci_get_subdevice(dev);
4137 	pci_rev_id = pci_get_revid(dev);
4138 	if (sctx->isc_parse_devinfo != NULL)
4139 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4140 
4141 	ent = sctx->isc_vendor_info;
4142 	while (ent->pvi_vendor_id != 0) {
4143 		if (pci_vendor_id != ent->pvi_vendor_id) {
4144 			ent++;
4145 			continue;
4146 		}
4147 		if ((pci_device_id == ent->pvi_device_id) &&
4148 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4149 		     (ent->pvi_subvendor_id == 0)) &&
4150 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4151 		     (ent->pvi_subdevice_id == 0)) &&
4152 		    ((pci_rev_id == ent->pvi_rev_id) ||
4153 		     (ent->pvi_rev_id == 0))) {
4154 
4155 			device_set_desc_copy(dev, ent->pvi_name);
4156 			/* this needs to be changed to zero if the bus probing code
4157 			 * ever stops re-probing on best match because the sctx
4158 			 * may have its values over written by register calls
4159 			 * in subsequent probes
4160 			 */
4161 			return (BUS_PROBE_DEFAULT);
4162 		}
4163 		ent++;
4164 	}
4165 	return (ENXIO);
4166 }
4167 
4168 int
4169 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4170 {
4171 	int err, rid, msix, msix_bar;
4172 	if_ctx_t ctx;
4173 	if_t ifp;
4174 	if_softc_ctx_t scctx;
4175 	int i;
4176 	uint16_t main_txq;
4177 	uint16_t main_rxq;
4178 
4179 
4180 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4181 
4182 	if (sc == NULL) {
4183 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4184 		device_set_softc(dev, ctx);
4185 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4186 	}
4187 
4188 	ctx->ifc_sctx = sctx;
4189 	ctx->ifc_dev = dev;
4190 	ctx->ifc_softc = sc;
4191 
4192 	if ((err = iflib_register(ctx)) != 0) {
4193 		device_printf(dev, "iflib_register failed %d\n", err);
4194 		return (err);
4195 	}
4196 	iflib_add_device_sysctl_pre(ctx);
4197 
4198 	scctx = &ctx->ifc_softc_ctx;
4199 	ifp = ctx->ifc_ifp;
4200 	ctx->ifc_nhwtxqs = sctx->isc_ntxqs;
4201 
4202 	/*
4203 	 * XXX sanity check that ntxd & nrxd are a power of 2
4204 	 */
4205 	if (ctx->ifc_sysctl_ntxqs != 0)
4206 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4207 	if (ctx->ifc_sysctl_nrxqs != 0)
4208 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4209 
4210 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4211 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4212 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4213 		else
4214 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4215 	}
4216 
4217 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4218 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4219 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4220 		else
4221 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4222 	}
4223 
4224 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4225 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4226 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4227 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4228 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4229 		}
4230 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4231 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4232 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4233 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4234 		}
4235 	}
4236 
4237 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4238 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4239 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4240 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4241 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4242 		}
4243 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4244 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4245 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4246 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4247 		}
4248 	}
4249 
4250 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4251 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4252 		return (err);
4253 	}
4254 	_iflib_pre_assert(scctx);
4255 	ctx->ifc_txrx = *scctx->isc_txrx;
4256 
4257 #ifdef INVARIANTS
4258 	MPASS(scctx->isc_capenable);
4259 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4260 		MPASS(scctx->isc_tx_csum_flags);
4261 #endif
4262 
4263 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4264 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4265 
4266 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4267 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4268 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4269 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4270 
4271 #ifdef ACPI_DMAR
4272 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4273 		ctx->ifc_flags |= IFC_DMAR;
4274 #elif !(defined(__i386__) || defined(__amd64__))
4275 	/* set unconditionally for !x86 */
4276 	ctx->ifc_flags |= IFC_DMAR;
4277 #endif
4278 
4279 	msix_bar = scctx->isc_msix_bar;
4280 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4281 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4282 
4283 	/* XXX change for per-queue sizes */
4284 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4285 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4286 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4287 		if (!powerof2(scctx->isc_nrxd[i])) {
4288 			/* round down instead? */
4289 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4290 			err = EINVAL;
4291 			goto fail;
4292 		}
4293 	}
4294 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4295 		if (!powerof2(scctx->isc_ntxd[i])) {
4296 			device_printf(dev,
4297 			    "# tx descriptors must be a power of 2");
4298 			err = EINVAL;
4299 			goto fail;
4300 		}
4301 	}
4302 
4303 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4304 	    MAX_SINGLE_PACKET_FRACTION)
4305 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4306 		    MAX_SINGLE_PACKET_FRACTION);
4307 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4308 	    MAX_SINGLE_PACKET_FRACTION)
4309 		scctx->isc_tx_tso_segments_max = max(1,
4310 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4311 
4312 	/*
4313 	 * Protect the stack against modern hardware
4314 	 */
4315 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4316 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4317 
4318 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4319 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4320 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4321 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4322 	if (scctx->isc_rss_table_size == 0)
4323 		scctx->isc_rss_table_size = 64;
4324 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4325 
4326 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4327 	/* XXX format name */
4328 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4329 
4330 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4331 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4332 		device_printf(dev, "Unable to fetch CPU list\n");
4333 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4334 	}
4335 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4336 
4337 	/*
4338 	** Now setup MSI or MSI/X, should
4339 	** return us the number of supported
4340 	** vectors. (Will be 1 for MSI)
4341 	*/
4342 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4343 		msix = scctx->isc_vectors;
4344 	} else if (scctx->isc_msix_bar != 0)
4345 	       /*
4346 		* The simple fact that isc_msix_bar is not 0 does not mean we
4347 		* we have a good value there that is known to work.
4348 		*/
4349 		msix = iflib_msix_init(ctx);
4350 	else {
4351 		scctx->isc_vectors = 1;
4352 		scctx->isc_ntxqsets = 1;
4353 		scctx->isc_nrxqsets = 1;
4354 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4355 		msix = 0;
4356 	}
4357 	/* Get memory for the station queues */
4358 	if ((err = iflib_queues_alloc(ctx))) {
4359 		device_printf(dev, "Unable to allocate queue memory\n");
4360 		goto fail;
4361 	}
4362 
4363 	if ((err = iflib_qset_structures_setup(ctx))) {
4364 		device_printf(dev, "qset structure setup failed %d\n", err);
4365 		goto fail_queues;
4366 	}
4367 
4368 	/*
4369 	 * Group taskqueues aren't properly set up until SMP is started,
4370 	 * so we disable interrupts until we can handle them post
4371 	 * SI_SUB_SMP.
4372 	 *
4373 	 * XXX: disabling interrupts doesn't actually work, at least for
4374 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4375 	 * we do null handling and depend on this not causing too large an
4376 	 * interrupt storm.
4377 	 */
4378 	IFDI_INTR_DISABLE(ctx);
4379 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4380 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4381 		goto fail_intr_free;
4382 	}
4383 	if (msix <= 1) {
4384 		rid = 0;
4385 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4386 			MPASS(msix == 1);
4387 			rid = 1;
4388 		}
4389 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4390 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4391 			goto fail_intr_free;
4392 		}
4393 	}
4394 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4395 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4396 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4397 		goto fail_detach;
4398 	}
4399 	if ((err = iflib_netmap_attach(ctx))) {
4400 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4401 		goto fail_detach;
4402 	}
4403 	*ctxp = ctx;
4404 
4405 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4406 	iflib_add_device_sysctl_post(ctx);
4407 	ctx->ifc_flags |= IFC_INIT_DONE;
4408 	return (0);
4409 fail_detach:
4410 	ether_ifdetach(ctx->ifc_ifp);
4411 fail_intr_free:
4412 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4413 		pci_release_msi(ctx->ifc_dev);
4414 fail_queues:
4415 	/* XXX free queues */
4416 fail:
4417 	IFDI_DETACH(ctx);
4418 	return (err);
4419 }
4420 
4421 int
4422 iflib_device_attach(device_t dev)
4423 {
4424 	if_ctx_t ctx;
4425 	if_shared_ctx_t sctx;
4426 
4427 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4428 		return (ENOTSUP);
4429 
4430 	pci_enable_busmaster(dev);
4431 
4432 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4433 }
4434 
4435 int
4436 iflib_device_deregister(if_ctx_t ctx)
4437 {
4438 	if_t ifp = ctx->ifc_ifp;
4439 	iflib_txq_t txq;
4440 	iflib_rxq_t rxq;
4441 	device_t dev = ctx->ifc_dev;
4442 	int i, j;
4443 	struct taskqgroup *tqg;
4444 	iflib_fl_t fl;
4445 
4446 	/* Make sure VLANS are not using driver */
4447 	if (if_vlantrunkinuse(ifp)) {
4448 		device_printf(dev,"Vlan in use, detach first\n");
4449 		return (EBUSY);
4450 	}
4451 
4452 	CTX_LOCK(ctx);
4453 	ctx->ifc_in_detach = 1;
4454 	iflib_stop(ctx);
4455 	CTX_UNLOCK(ctx);
4456 
4457 	/* Unregister VLAN events */
4458 	if (ctx->ifc_vlan_attach_event != NULL)
4459 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4460 	if (ctx->ifc_vlan_detach_event != NULL)
4461 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4462 
4463 	iflib_netmap_detach(ifp);
4464 	ether_ifdetach(ifp);
4465 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4466 	CTX_LOCK_DESTROY(ctx);
4467 	if (ctx->ifc_led_dev != NULL)
4468 		led_destroy(ctx->ifc_led_dev);
4469 	/* XXX drain any dependent tasks */
4470 	tqg = qgroup_if_io_tqg;
4471 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4472 		callout_drain(&txq->ift_timer);
4473 		if (txq->ift_task.gt_uniq != NULL)
4474 			taskqgroup_detach(tqg, &txq->ift_task);
4475 	}
4476 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4477 		if (rxq->ifr_task.gt_uniq != NULL)
4478 			taskqgroup_detach(tqg, &rxq->ifr_task);
4479 
4480 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4481 			free(fl->ifl_rx_bitmap, M_IFLIB);
4482 
4483 	}
4484 	tqg = qgroup_if_config_tqg;
4485 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4486 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4487 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4488 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4489 
4490 	IFDI_DETACH(ctx);
4491 	device_set_softc(ctx->ifc_dev, NULL);
4492 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4493 		pci_release_msi(dev);
4494 	}
4495 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4496 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4497 	}
4498 	if (ctx->ifc_msix_mem != NULL) {
4499 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4500 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4501 		ctx->ifc_msix_mem = NULL;
4502 	}
4503 
4504 	bus_generic_detach(dev);
4505 	if_free(ifp);
4506 
4507 	iflib_tx_structures_free(ctx);
4508 	iflib_rx_structures_free(ctx);
4509 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4510 		free(ctx->ifc_softc, M_IFLIB);
4511 	free(ctx, M_IFLIB);
4512 	return (0);
4513 }
4514 
4515 
4516 int
4517 iflib_device_detach(device_t dev)
4518 {
4519 	if_ctx_t ctx = device_get_softc(dev);
4520 
4521 	return (iflib_device_deregister(ctx));
4522 }
4523 
4524 int
4525 iflib_device_suspend(device_t dev)
4526 {
4527 	if_ctx_t ctx = device_get_softc(dev);
4528 
4529 	CTX_LOCK(ctx);
4530 	IFDI_SUSPEND(ctx);
4531 	CTX_UNLOCK(ctx);
4532 
4533 	return bus_generic_suspend(dev);
4534 }
4535 int
4536 iflib_device_shutdown(device_t dev)
4537 {
4538 	if_ctx_t ctx = device_get_softc(dev);
4539 
4540 	CTX_LOCK(ctx);
4541 	IFDI_SHUTDOWN(ctx);
4542 	CTX_UNLOCK(ctx);
4543 
4544 	return bus_generic_suspend(dev);
4545 }
4546 
4547 
4548 int
4549 iflib_device_resume(device_t dev)
4550 {
4551 	if_ctx_t ctx = device_get_softc(dev);
4552 	iflib_txq_t txq = ctx->ifc_txqs;
4553 
4554 	CTX_LOCK(ctx);
4555 	IFDI_RESUME(ctx);
4556 	iflib_init_locked(ctx);
4557 	CTX_UNLOCK(ctx);
4558 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4559 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4560 
4561 	return (bus_generic_resume(dev));
4562 }
4563 
4564 int
4565 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4566 {
4567 	int error;
4568 	if_ctx_t ctx = device_get_softc(dev);
4569 
4570 	CTX_LOCK(ctx);
4571 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4572 	CTX_UNLOCK(ctx);
4573 
4574 	return (error);
4575 }
4576 
4577 void
4578 iflib_device_iov_uninit(device_t dev)
4579 {
4580 	if_ctx_t ctx = device_get_softc(dev);
4581 
4582 	CTX_LOCK(ctx);
4583 	IFDI_IOV_UNINIT(ctx);
4584 	CTX_UNLOCK(ctx);
4585 }
4586 
4587 int
4588 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4589 {
4590 	int error;
4591 	if_ctx_t ctx = device_get_softc(dev);
4592 
4593 	CTX_LOCK(ctx);
4594 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4595 	CTX_UNLOCK(ctx);
4596 
4597 	return (error);
4598 }
4599 
4600 /*********************************************************************
4601  *
4602  *  MODULE FUNCTION DEFINITIONS
4603  *
4604  **********************************************************************/
4605 
4606 /*
4607  * - Start a fast taskqueue thread for each core
4608  * - Start a taskqueue for control operations
4609  */
4610 static int
4611 iflib_module_init(void)
4612 {
4613 	return (0);
4614 }
4615 
4616 static int
4617 iflib_module_event_handler(module_t mod, int what, void *arg)
4618 {
4619 	int err;
4620 
4621 	switch (what) {
4622 	case MOD_LOAD:
4623 		if ((err = iflib_module_init()) != 0)
4624 			return (err);
4625 		break;
4626 	case MOD_UNLOAD:
4627 		return (EBUSY);
4628 	default:
4629 		return (EOPNOTSUPP);
4630 	}
4631 
4632 	return (0);
4633 }
4634 
4635 /*********************************************************************
4636  *
4637  *  PUBLIC FUNCTION DEFINITIONS
4638  *     ordered as in iflib.h
4639  *
4640  **********************************************************************/
4641 
4642 
4643 static void
4644 _iflib_assert(if_shared_ctx_t sctx)
4645 {
4646 	MPASS(sctx->isc_tx_maxsize);
4647 	MPASS(sctx->isc_tx_maxsegsize);
4648 
4649 	MPASS(sctx->isc_rx_maxsize);
4650 	MPASS(sctx->isc_rx_nsegments);
4651 	MPASS(sctx->isc_rx_maxsegsize);
4652 
4653 	MPASS(sctx->isc_nrxd_min[0]);
4654 	MPASS(sctx->isc_nrxd_max[0]);
4655 	MPASS(sctx->isc_nrxd_default[0]);
4656 	MPASS(sctx->isc_ntxd_min[0]);
4657 	MPASS(sctx->isc_ntxd_max[0]);
4658 	MPASS(sctx->isc_ntxd_default[0]);
4659 }
4660 
4661 static void
4662 _iflib_pre_assert(if_softc_ctx_t scctx)
4663 {
4664 
4665 	MPASS(scctx->isc_txrx->ift_txd_encap);
4666 	MPASS(scctx->isc_txrx->ift_txd_flush);
4667 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4668 	MPASS(scctx->isc_txrx->ift_rxd_available);
4669 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4670 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4671 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4672 }
4673 
4674 static int
4675 iflib_register(if_ctx_t ctx)
4676 {
4677 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4678 	driver_t *driver = sctx->isc_driver;
4679 	device_t dev = ctx->ifc_dev;
4680 	if_t ifp;
4681 
4682 	_iflib_assert(sctx);
4683 
4684 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4685 
4686 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4687 	if (ifp == NULL) {
4688 		device_printf(dev, "can not allocate ifnet structure\n");
4689 		return (ENOMEM);
4690 	}
4691 
4692 	/*
4693 	 * Initialize our context's device specific methods
4694 	 */
4695 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4696 	kobj_class_compile((kobj_class_t) driver);
4697 	driver->refs++;
4698 
4699 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4700 	if_setsoftc(ifp, ctx);
4701 	if_setdev(ifp, dev);
4702 	if_setinitfn(ifp, iflib_if_init);
4703 	if_setioctlfn(ifp, iflib_if_ioctl);
4704 	if_settransmitfn(ifp, iflib_if_transmit);
4705 	if_setqflushfn(ifp, iflib_if_qflush);
4706 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4707 
4708 	ctx->ifc_vlan_attach_event =
4709 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4710 							  EVENTHANDLER_PRI_FIRST);
4711 	ctx->ifc_vlan_detach_event =
4712 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4713 							  EVENTHANDLER_PRI_FIRST);
4714 
4715 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4716 					 iflib_media_change, iflib_media_status);
4717 
4718 	return (0);
4719 }
4720 
4721 
4722 static int
4723 iflib_queues_alloc(if_ctx_t ctx)
4724 {
4725 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4726 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4727 	device_t dev = ctx->ifc_dev;
4728 	int nrxqsets = scctx->isc_nrxqsets;
4729 	int ntxqsets = scctx->isc_ntxqsets;
4730 	iflib_txq_t txq;
4731 	iflib_rxq_t rxq;
4732 	iflib_fl_t fl = NULL;
4733 	int i, j, cpu, err, txconf, rxconf;
4734 	iflib_dma_info_t ifdip;
4735 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4736 	uint32_t *txqsizes = scctx->isc_txqsizes;
4737 	uint8_t nrxqs = sctx->isc_nrxqs;
4738 	uint8_t ntxqs = sctx->isc_ntxqs;
4739 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4740 	caddr_t *vaddrs;
4741 	uint64_t *paddrs;
4742 	struct ifmp_ring **brscp;
4743 
4744 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4745 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4746 
4747 	brscp = NULL;
4748 	txq = NULL;
4749 	rxq = NULL;
4750 
4751 /* Allocate the TX ring struct memory */
4752 	if (!(txq =
4753 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4754 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4755 		device_printf(dev, "Unable to allocate TX ring memory\n");
4756 		err = ENOMEM;
4757 		goto fail;
4758 	}
4759 
4760 	/* Now allocate the RX */
4761 	if (!(rxq =
4762 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4763 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4764 		device_printf(dev, "Unable to allocate RX ring memory\n");
4765 		err = ENOMEM;
4766 		goto rx_fail;
4767 	}
4768 
4769 	ctx->ifc_txqs = txq;
4770 	ctx->ifc_rxqs = rxq;
4771 
4772 	/*
4773 	 * XXX handle allocation failure
4774 	 */
4775 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4776 		/* Set up some basics */
4777 
4778 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4779 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4780 			err = ENOMEM;
4781 			goto err_tx_desc;
4782 		}
4783 		txq->ift_ifdi = ifdip;
4784 		for (j = 0; j < ntxqs; j++, ifdip++) {
4785 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4786 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4787 				err = ENOMEM;
4788 				goto err_tx_desc;
4789 			}
4790 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4791 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4792 		}
4793 		txq->ift_ctx = ctx;
4794 		txq->ift_id = i;
4795 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4796 			txq->ift_br_offset = 1;
4797 		} else {
4798 			txq->ift_br_offset = 0;
4799 		}
4800 		/* XXX fix this */
4801 		txq->ift_timer.c_cpu = cpu;
4802 
4803 		if (iflib_txsd_alloc(txq)) {
4804 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4805 			err = ENOMEM;
4806 			goto err_tx_desc;
4807 		}
4808 
4809 		/* Initialize the TX lock */
4810 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4811 		    device_get_nameunit(dev), txq->ift_id);
4812 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4813 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4814 
4815 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4816 			 device_get_nameunit(dev), txq->ift_id);
4817 
4818 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4819 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4820 		if (err) {
4821 			/* XXX free any allocated rings */
4822 			device_printf(dev, "Unable to allocate buf_ring\n");
4823 			goto err_tx_desc;
4824 		}
4825 	}
4826 
4827 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4828 		/* Set up some basics */
4829 
4830 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4831 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4832 			err = ENOMEM;
4833 			goto err_tx_desc;
4834 		}
4835 
4836 		rxq->ifr_ifdi = ifdip;
4837 		/* XXX this needs to be changed if #rx queues != #tx queues */
4838 		rxq->ifr_ntxqirq = 1;
4839 		rxq->ifr_txqid[0] = i;
4840 		for (j = 0; j < nrxqs; j++, ifdip++) {
4841 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4842 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4843 				err = ENOMEM;
4844 				goto err_tx_desc;
4845 			}
4846 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4847 		}
4848 		rxq->ifr_ctx = ctx;
4849 		rxq->ifr_id = i;
4850 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4851 			rxq->ifr_fl_offset = 1;
4852 		} else {
4853 			rxq->ifr_fl_offset = 0;
4854 		}
4855 		rxq->ifr_nfl = nfree_lists;
4856 		if (!(fl =
4857 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4858 			device_printf(dev, "Unable to allocate free list memory\n");
4859 			err = ENOMEM;
4860 			goto err_tx_desc;
4861 		}
4862 		rxq->ifr_fl = fl;
4863 		for (j = 0; j < nfree_lists; j++) {
4864 			fl[j].ifl_rxq = rxq;
4865 			fl[j].ifl_id = j;
4866 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4867 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4868 		}
4869         /* Allocate receive buffers for the ring*/
4870 		if (iflib_rxsd_alloc(rxq)) {
4871 			device_printf(dev,
4872 			    "Critical Failure setting up receive buffers\n");
4873 			err = ENOMEM;
4874 			goto err_rx_desc;
4875 		}
4876 
4877 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4878 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4879 	}
4880 
4881 	/* TXQs */
4882 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4883 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4884 	for (i = 0; i < ntxqsets; i++) {
4885 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4886 
4887 		for (j = 0; j < ntxqs; j++, di++) {
4888 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4889 			paddrs[i*ntxqs + j] = di->idi_paddr;
4890 		}
4891 	}
4892 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4893 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4894 		iflib_tx_structures_free(ctx);
4895 		free(vaddrs, M_IFLIB);
4896 		free(paddrs, M_IFLIB);
4897 		goto err_rx_desc;
4898 	}
4899 	free(vaddrs, M_IFLIB);
4900 	free(paddrs, M_IFLIB);
4901 
4902 	/* RXQs */
4903 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4904 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4905 	for (i = 0; i < nrxqsets; i++) {
4906 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4907 
4908 		for (j = 0; j < nrxqs; j++, di++) {
4909 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4910 			paddrs[i*nrxqs + j] = di->idi_paddr;
4911 		}
4912 	}
4913 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4914 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4915 		iflib_tx_structures_free(ctx);
4916 		free(vaddrs, M_IFLIB);
4917 		free(paddrs, M_IFLIB);
4918 		goto err_rx_desc;
4919 	}
4920 	free(vaddrs, M_IFLIB);
4921 	free(paddrs, M_IFLIB);
4922 
4923 	return (0);
4924 
4925 /* XXX handle allocation failure changes */
4926 err_rx_desc:
4927 err_tx_desc:
4928 	if (ctx->ifc_rxqs != NULL)
4929 		free(ctx->ifc_rxqs, M_IFLIB);
4930 	ctx->ifc_rxqs = NULL;
4931 	if (ctx->ifc_txqs != NULL)
4932 		free(ctx->ifc_txqs, M_IFLIB);
4933 	ctx->ifc_txqs = NULL;
4934 rx_fail:
4935 	if (brscp != NULL)
4936 		free(brscp, M_IFLIB);
4937 	if (rxq != NULL)
4938 		free(rxq, M_IFLIB);
4939 	if (txq != NULL)
4940 		free(txq, M_IFLIB);
4941 fail:
4942 	return (err);
4943 }
4944 
4945 static int
4946 iflib_tx_structures_setup(if_ctx_t ctx)
4947 {
4948 	iflib_txq_t txq = ctx->ifc_txqs;
4949 	int i;
4950 
4951 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4952 		iflib_txq_setup(txq);
4953 
4954 	return (0);
4955 }
4956 
4957 static void
4958 iflib_tx_structures_free(if_ctx_t ctx)
4959 {
4960 	iflib_txq_t txq = ctx->ifc_txqs;
4961 	int i, j;
4962 
4963 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4964 		iflib_txq_destroy(txq);
4965 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4966 			iflib_dma_free(&txq->ift_ifdi[j]);
4967 	}
4968 	free(ctx->ifc_txqs, M_IFLIB);
4969 	ctx->ifc_txqs = NULL;
4970 	IFDI_QUEUES_FREE(ctx);
4971 }
4972 
4973 /*********************************************************************
4974  *
4975  *  Initialize all receive rings.
4976  *
4977  **********************************************************************/
4978 static int
4979 iflib_rx_structures_setup(if_ctx_t ctx)
4980 {
4981 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4982 	int q;
4983 #if defined(INET6) || defined(INET)
4984 	int i, err;
4985 #endif
4986 
4987 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4988 #if defined(INET6) || defined(INET)
4989 		tcp_lro_free(&rxq->ifr_lc);
4990 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4991 		    TCP_LRO_ENTRIES, min(1024,
4992 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4993 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4994 			goto fail;
4995 		}
4996 		rxq->ifr_lro_enabled = TRUE;
4997 #endif
4998 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4999 	}
5000 	return (0);
5001 #if defined(INET6) || defined(INET)
5002 fail:
5003 	/*
5004 	 * Free RX software descriptors allocated so far, we will only handle
5005 	 * the rings that completed, the failing case will have
5006 	 * cleaned up for itself. 'q' failed, so its the terminus.
5007 	 */
5008 	rxq = ctx->ifc_rxqs;
5009 	for (i = 0; i < q; ++i, rxq++) {
5010 		iflib_rx_sds_free(rxq);
5011 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5012 	}
5013 	return (err);
5014 #endif
5015 }
5016 
5017 /*********************************************************************
5018  *
5019  *  Free all receive rings.
5020  *
5021  **********************************************************************/
5022 static void
5023 iflib_rx_structures_free(if_ctx_t ctx)
5024 {
5025 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5026 
5027 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5028 		iflib_rx_sds_free(rxq);
5029 	}
5030 }
5031 
5032 static int
5033 iflib_qset_structures_setup(if_ctx_t ctx)
5034 {
5035 	int err;
5036 
5037 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5038 		return (err);
5039 
5040 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5041 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5042 		iflib_tx_structures_free(ctx);
5043 		iflib_rx_structures_free(ctx);
5044 	}
5045 	return (err);
5046 }
5047 
5048 int
5049 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5050 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5051 {
5052 
5053 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5054 }
5055 
5056 #ifdef SMP
5057 static int
5058 find_nth(if_ctx_t ctx, int qid)
5059 {
5060 	cpuset_t cpus;
5061 	int i, cpuid, eqid, count;
5062 
5063 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5064 	count = CPU_COUNT(&cpus);
5065 	eqid = qid % count;
5066 	/* clear up to the qid'th bit */
5067 	for (i = 0; i < eqid; i++) {
5068 		cpuid = CPU_FFS(&cpus);
5069 		MPASS(cpuid != 0);
5070 		CPU_CLR(cpuid-1, &cpus);
5071 	}
5072 	cpuid = CPU_FFS(&cpus);
5073 	MPASS(cpuid != 0);
5074 	return (cpuid-1);
5075 }
5076 
5077 #ifdef SCHED_ULE
5078 extern struct cpu_group *cpu_top;              /* CPU topology */
5079 
5080 static int
5081 find_child_with_core(int cpu, struct cpu_group *grp)
5082 {
5083 	int i;
5084 
5085 	if (grp->cg_children == 0)
5086 		return -1;
5087 
5088 	MPASS(grp->cg_child);
5089 	for (i = 0; i < grp->cg_children; i++) {
5090 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5091 			return i;
5092 	}
5093 
5094 	return -1;
5095 }
5096 
5097 /*
5098  * Find the nth thread on the specified core
5099  */
5100 static int
5101 find_thread(int cpu, int thread_num)
5102 {
5103 	struct cpu_group *grp;
5104 	int i;
5105 	cpuset_t cs;
5106 
5107 	grp = cpu_top;
5108 	if (grp == NULL)
5109 		return cpu;
5110 	i = 0;
5111 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5112 		/* If the child only has one cpu, don't descend */
5113 		if (grp->cg_child[i].cg_count <= 1)
5114 			break;
5115 		grp = &grp->cg_child[i];
5116 	}
5117 
5118 	/* If they don't share at least an L2 cache, use the same CPU */
5119 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5120 		return cpu;
5121 
5122 	/* Now pick one */
5123 	CPU_COPY(&grp->cg_mask, &cs);
5124 	for (i = thread_num % grp->cg_count; i > 0; i--) {
5125 		MPASS(CPU_FFS(&cs));
5126 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5127 	}
5128 	MPASS(CPU_FFS(&cs));
5129 	return CPU_FFS(&cs) - 1;
5130 }
5131 #else
5132 static int
5133 find_thread(int cpu, int thread_num __unused)
5134 {
5135 	return cpu;
5136 }
5137 #endif
5138 
5139 static int
5140 get_thread_num(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5141 {
5142 	switch (type) {
5143 	case IFLIB_INTR_TX:
5144 		/* TX queues get threads on the same core as the corresponding RX queue */
5145 		/* XXX handle multiple RX threads per core and more than two threads per core */
5146 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5147 	case IFLIB_INTR_RX:
5148 	case IFLIB_INTR_RXTX:
5149 		/* RX queues get the first thread on their core */
5150 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5151 	default:
5152 		return -1;
5153 	}
5154 }
5155 #else
5156 #define get_thread_num(ctx, type, qid)	CPU_FIRST()
5157 #define find_thread(cpuid, tid)		CPU_FIRST()
5158 #define find_nth(ctx, gid)		CPU_FIRST()
5159 #endif
5160 
5161 /* Just to avoid copy/paste */
5162 static inline int
5163 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5164     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5165 {
5166 	int cpuid;
5167 	int err, tid;
5168 
5169 	cpuid = find_nth(ctx, qid);
5170 	tid = get_thread_num(ctx, type, qid);
5171 	MPASS(tid >= 0);
5172 	cpuid = find_thread(cpuid, tid);
5173 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5174 	if (err) {
5175 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5176 		return (err);
5177 	}
5178 #ifdef notyet
5179 	if (cpuid > ctx->ifc_cpuid_highest)
5180 		ctx->ifc_cpuid_highest = cpuid;
5181 #endif
5182 	return 0;
5183 }
5184 
5185 int
5186 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5187 						iflib_intr_type_t type, driver_filter_t *filter,
5188 						void *filter_arg, int qid, char *name)
5189 {
5190 	struct grouptask *gtask;
5191 	struct taskqgroup *tqg;
5192 	iflib_filter_info_t info;
5193 	gtask_fn_t *fn;
5194 	int tqrid, err;
5195 	driver_filter_t *intr_fast;
5196 	void *q;
5197 
5198 	info = &ctx->ifc_filter_info;
5199 	tqrid = rid;
5200 
5201 	switch (type) {
5202 	/* XXX merge tx/rx for netmap? */
5203 	case IFLIB_INTR_TX:
5204 		q = &ctx->ifc_txqs[qid];
5205 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5206 		gtask = &ctx->ifc_txqs[qid].ift_task;
5207 		tqg = qgroup_if_io_tqg;
5208 		fn = _task_fn_tx;
5209 		intr_fast = iflib_fast_intr;
5210 		GROUPTASK_INIT(gtask, 0, fn, q);
5211 		break;
5212 	case IFLIB_INTR_RX:
5213 		q = &ctx->ifc_rxqs[qid];
5214 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5215 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5216 		tqg = qgroup_if_io_tqg;
5217 		fn = _task_fn_rx;
5218 		intr_fast = iflib_fast_intr;
5219 		GROUPTASK_INIT(gtask, 0, fn, q);
5220 		break;
5221 	case IFLIB_INTR_RXTX:
5222 		q = &ctx->ifc_rxqs[qid];
5223 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5224 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5225 		tqg = qgroup_if_io_tqg;
5226 		fn = _task_fn_rx;
5227 		intr_fast = iflib_fast_intr_rxtx;
5228 		GROUPTASK_INIT(gtask, 0, fn, q);
5229 		break;
5230 	case IFLIB_INTR_ADMIN:
5231 		q = ctx;
5232 		tqrid = -1;
5233 		info = &ctx->ifc_filter_info;
5234 		gtask = &ctx->ifc_admin_task;
5235 		tqg = qgroup_if_config_tqg;
5236 		fn = _task_fn_admin;
5237 		intr_fast = iflib_fast_intr_ctx;
5238 		break;
5239 	default:
5240 		panic("unknown net intr type");
5241 	}
5242 
5243 	info->ifi_filter = filter;
5244 	info->ifi_filter_arg = filter_arg;
5245 	info->ifi_task = gtask;
5246 	info->ifi_ctx = q;
5247 
5248 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5249 	if (err != 0) {
5250 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5251 		return (err);
5252 	}
5253 	if (type == IFLIB_INTR_ADMIN)
5254 		return (0);
5255 
5256 	if (tqrid != -1) {
5257 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5258 		if (err)
5259 			return (err);
5260 	} else {
5261 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5262 	}
5263 
5264 	return (0);
5265 }
5266 
5267 void
5268 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5269 {
5270 	struct grouptask *gtask;
5271 	struct taskqgroup *tqg;
5272 	gtask_fn_t *fn;
5273 	void *q;
5274 	int irq_num = -1;
5275 	int err;
5276 
5277 	switch (type) {
5278 	case IFLIB_INTR_TX:
5279 		q = &ctx->ifc_txqs[qid];
5280 		gtask = &ctx->ifc_txqs[qid].ift_task;
5281 		tqg = qgroup_if_io_tqg;
5282 		fn = _task_fn_tx;
5283 		if (irq != NULL)
5284 			irq_num = rman_get_start(irq->ii_res);
5285 		break;
5286 	case IFLIB_INTR_RX:
5287 		q = &ctx->ifc_rxqs[qid];
5288 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5289 		tqg = qgroup_if_io_tqg;
5290 		fn = _task_fn_rx;
5291 		if (irq != NULL)
5292 			irq_num = rman_get_start(irq->ii_res);
5293 		break;
5294 	case IFLIB_INTR_IOV:
5295 		q = ctx;
5296 		gtask = &ctx->ifc_vflr_task;
5297 		tqg = qgroup_if_config_tqg;
5298 		fn = _task_fn_iov;
5299 		break;
5300 	default:
5301 		panic("unknown net intr type");
5302 	}
5303 	GROUPTASK_INIT(gtask, 0, fn, q);
5304 	if (irq_num != -1) {
5305 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5306 		if (err)
5307 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5308 	}
5309 	else {
5310 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5311 	}
5312 }
5313 
5314 void
5315 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5316 {
5317 	if (irq->ii_tag)
5318 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5319 
5320 	if (irq->ii_res)
5321 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5322 }
5323 
5324 static int
5325 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5326 {
5327 	iflib_txq_t txq = ctx->ifc_txqs;
5328 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5329 	if_irq_t irq = &ctx->ifc_legacy_irq;
5330 	iflib_filter_info_t info;
5331 	struct grouptask *gtask;
5332 	struct taskqgroup *tqg;
5333 	gtask_fn_t *fn;
5334 	int tqrid;
5335 	void *q;
5336 	int err;
5337 
5338 	q = &ctx->ifc_rxqs[0];
5339 	info = &rxq[0].ifr_filter_info;
5340 	gtask = &rxq[0].ifr_task;
5341 	tqg = qgroup_if_io_tqg;
5342 	tqrid = irq->ii_rid = *rid;
5343 	fn = _task_fn_rx;
5344 
5345 	ctx->ifc_flags |= IFC_LEGACY;
5346 	info->ifi_filter = filter;
5347 	info->ifi_filter_arg = filter_arg;
5348 	info->ifi_task = gtask;
5349 	info->ifi_ctx = ctx;
5350 
5351 	/* We allocate a single interrupt resource */
5352 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5353 		return (err);
5354 	GROUPTASK_INIT(gtask, 0, fn, q);
5355 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5356 
5357 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5358 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5359 	return (0);
5360 }
5361 
5362 void
5363 iflib_led_create(if_ctx_t ctx)
5364 {
5365 
5366 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5367 	    device_get_nameunit(ctx->ifc_dev));
5368 }
5369 
5370 void
5371 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5372 {
5373 
5374 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5375 }
5376 
5377 void
5378 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5379 {
5380 
5381 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5382 }
5383 
5384 void
5385 iflib_admin_intr_deferred(if_ctx_t ctx)
5386 {
5387 #ifdef INVARIANTS
5388 	struct grouptask *gtask;
5389 
5390 	gtask = &ctx->ifc_admin_task;
5391 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5392 #endif
5393 
5394 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5395 }
5396 
5397 void
5398 iflib_iov_intr_deferred(if_ctx_t ctx)
5399 {
5400 
5401 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5402 }
5403 
5404 void
5405 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5406 {
5407 
5408 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5409 }
5410 
5411 void
5412 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5413 	char *name)
5414 {
5415 
5416 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5417 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5418 }
5419 
5420 void
5421 iflib_config_gtask_deinit(struct grouptask *gtask)
5422 {
5423 
5424 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5425 }
5426 
5427 void
5428 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5429 {
5430 	if_t ifp = ctx->ifc_ifp;
5431 	iflib_txq_t txq = ctx->ifc_txqs;
5432 
5433 	if_setbaudrate(ifp, baudrate);
5434 	if (baudrate >= IF_Gbps(10))
5435 		ctx->ifc_flags |= IFC_PREFETCH;
5436 
5437 	/* If link down, disable watchdog */
5438 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5439 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5440 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5441 	}
5442 	ctx->ifc_link_state = link_state;
5443 	if_link_state_change(ifp, link_state);
5444 }
5445 
5446 static int
5447 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5448 {
5449 	int credits;
5450 #ifdef INVARIANTS
5451 	int credits_pre = txq->ift_cidx_processed;
5452 #endif
5453 
5454 	if (ctx->isc_txd_credits_update == NULL)
5455 		return (0);
5456 
5457 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5458 		return (0);
5459 
5460 	txq->ift_processed += credits;
5461 	txq->ift_cidx_processed += credits;
5462 
5463 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5464 	if (txq->ift_cidx_processed >= txq->ift_size)
5465 		txq->ift_cidx_processed -= txq->ift_size;
5466 	return (credits);
5467 }
5468 
5469 static int
5470 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5471 {
5472 
5473 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5474 	    budget));
5475 }
5476 
5477 void
5478 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5479 	const char *description, if_int_delay_info_t info,
5480 	int offset, int value)
5481 {
5482 	info->iidi_ctx = ctx;
5483 	info->iidi_offset = offset;
5484 	info->iidi_value = value;
5485 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5486 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5487 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5488 	    info, 0, iflib_sysctl_int_delay, "I", description);
5489 }
5490 
5491 struct mtx *
5492 iflib_ctx_lock_get(if_ctx_t ctx)
5493 {
5494 
5495 	return (&ctx->ifc_mtx);
5496 }
5497 
5498 static int
5499 iflib_msix_init(if_ctx_t ctx)
5500 {
5501 	device_t dev = ctx->ifc_dev;
5502 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5503 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5504 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5505 	int iflib_num_tx_queues, iflib_num_rx_queues;
5506 	int err, admincnt, bar;
5507 
5508 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5509 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5510 
5511 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5512 
5513 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5514 	admincnt = sctx->isc_admin_intrcnt;
5515 	/* Override by global tuneable */
5516 	{
5517 		int i;
5518 		size_t len = sizeof(i);
5519 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5520 		if (err == 0) {
5521 			if (i == 0)
5522 				goto msi;
5523 		}
5524 		else {
5525 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5526 		}
5527 	}
5528 	/* Override by tuneable */
5529 	if (scctx->isc_disable_msix)
5530 		goto msi;
5531 
5532 	/*
5533 	** When used in a virtualized environment
5534 	** PCI BUSMASTER capability may not be set
5535 	** so explicity set it here and rewrite
5536 	** the ENABLE in the MSIX control register
5537 	** at this point to cause the host to
5538 	** successfully initialize us.
5539 	*/
5540 	{
5541 		int msix_ctrl, rid;
5542 
5543  		pci_enable_busmaster(dev);
5544 		rid = 0;
5545 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5546 			rid += PCIR_MSIX_CTRL;
5547 			msix_ctrl = pci_read_config(dev, rid, 2);
5548 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5549 			pci_write_config(dev, rid, msix_ctrl, 2);
5550 		} else {
5551 			device_printf(dev, "PCIY_MSIX capability not found; "
5552 			                   "or rid %d == 0.\n", rid);
5553 			goto msi;
5554 		}
5555 	}
5556 
5557 	/*
5558 	 * bar == -1 => "trust me I know what I'm doing"
5559 	 * Some drivers are for hardware that is so shoddily
5560 	 * documented that no one knows which bars are which
5561 	 * so the developer has to map all bars. This hack
5562 	 * allows shoddy garbage to use msix in this framework.
5563 	 */
5564 	if (bar != -1) {
5565 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5566 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5567 		if (ctx->ifc_msix_mem == NULL) {
5568 			/* May not be enabled */
5569 			device_printf(dev, "Unable to map MSIX table \n");
5570 			goto msi;
5571 		}
5572 	}
5573 	/* First try MSI/X */
5574 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5575 		device_printf(dev, "System has MSIX disabled \n");
5576 		bus_release_resource(dev, SYS_RES_MEMORY,
5577 		    bar, ctx->ifc_msix_mem);
5578 		ctx->ifc_msix_mem = NULL;
5579 		goto msi;
5580 	}
5581 #if IFLIB_DEBUG
5582 	/* use only 1 qset in debug mode */
5583 	queuemsgs = min(msgs - admincnt, 1);
5584 #else
5585 	queuemsgs = msgs - admincnt;
5586 #endif
5587 #ifdef RSS
5588 	queues = imin(queuemsgs, rss_getnumbuckets());
5589 #else
5590 	queues = queuemsgs;
5591 #endif
5592 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5593 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5594 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5595 #ifdef  RSS
5596 	/* If we're doing RSS, clamp at the number of RSS buckets */
5597 	if (queues > rss_getnumbuckets())
5598 		queues = rss_getnumbuckets();
5599 #endif
5600 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5601 		rx_queues = iflib_num_rx_queues;
5602 	else
5603 		rx_queues = queues;
5604 
5605 	if (rx_queues > scctx->isc_nrxqsets)
5606 		rx_queues = scctx->isc_nrxqsets;
5607 
5608 	/*
5609 	 * We want this to be all logical CPUs by default
5610 	 */
5611 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5612 		tx_queues = iflib_num_tx_queues;
5613 	else
5614 		tx_queues = mp_ncpus;
5615 
5616 	if (tx_queues > scctx->isc_ntxqsets)
5617 		tx_queues = scctx->isc_ntxqsets;
5618 
5619 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5620 #ifdef INVARIANTS
5621 		if (tx_queues != rx_queues)
5622 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5623 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5624 #endif
5625 		tx_queues = min(rx_queues, tx_queues);
5626 		rx_queues = min(rx_queues, tx_queues);
5627 	}
5628 
5629 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5630 
5631 	vectors = rx_queues + admincnt;
5632 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5633 		device_printf(dev,
5634 					  "Using MSIX interrupts with %d vectors\n", vectors);
5635 		scctx->isc_vectors = vectors;
5636 		scctx->isc_nrxqsets = rx_queues;
5637 		scctx->isc_ntxqsets = tx_queues;
5638 		scctx->isc_intr = IFLIB_INTR_MSIX;
5639 
5640 		return (vectors);
5641 	} else {
5642 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5643 	}
5644 msi:
5645 	vectors = pci_msi_count(dev);
5646 	scctx->isc_nrxqsets = 1;
5647 	scctx->isc_ntxqsets = 1;
5648 	scctx->isc_vectors = vectors;
5649 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5650 		device_printf(dev,"Using an MSI interrupt\n");
5651 		scctx->isc_intr = IFLIB_INTR_MSI;
5652 	} else {
5653 		device_printf(dev,"Using a Legacy interrupt\n");
5654 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5655 	}
5656 
5657 	return (vectors);
5658 }
5659 
5660 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5661 
5662 static int
5663 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5664 {
5665 	int rc;
5666 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5667 	struct sbuf *sb;
5668 	char *ring_state = "UNKNOWN";
5669 
5670 	/* XXX needed ? */
5671 	rc = sysctl_wire_old_buffer(req, 0);
5672 	MPASS(rc == 0);
5673 	if (rc != 0)
5674 		return (rc);
5675 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5676 	MPASS(sb != NULL);
5677 	if (sb == NULL)
5678 		return (ENOMEM);
5679 	if (state[3] <= 3)
5680 		ring_state = ring_states[state[3]];
5681 
5682 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5683 		    state[0], state[1], state[2], ring_state);
5684 	rc = sbuf_finish(sb);
5685 	sbuf_delete(sb);
5686         return(rc);
5687 }
5688 
5689 enum iflib_ndesc_handler {
5690 	IFLIB_NTXD_HANDLER,
5691 	IFLIB_NRXD_HANDLER,
5692 };
5693 
5694 static int
5695 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5696 {
5697 	if_ctx_t ctx = (void *)arg1;
5698 	enum iflib_ndesc_handler type = arg2;
5699 	char buf[256] = {0};
5700 	qidx_t *ndesc;
5701 	char *p, *next;
5702 	int nqs, rc, i;
5703 
5704 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5705 
5706 	nqs = 8;
5707 	switch(type) {
5708 	case IFLIB_NTXD_HANDLER:
5709 		ndesc = ctx->ifc_sysctl_ntxds;
5710 		if (ctx->ifc_sctx)
5711 			nqs = ctx->ifc_sctx->isc_ntxqs;
5712 		break;
5713 	case IFLIB_NRXD_HANDLER:
5714 		ndesc = ctx->ifc_sysctl_nrxds;
5715 		if (ctx->ifc_sctx)
5716 			nqs = ctx->ifc_sctx->isc_nrxqs;
5717 		break;
5718 	}
5719 	if (nqs == 0)
5720 		nqs = 8;
5721 
5722 	for (i=0; i<8; i++) {
5723 		if (i >= nqs)
5724 			break;
5725 		if (i)
5726 			strcat(buf, ",");
5727 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5728 	}
5729 
5730 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5731 	if (rc || req->newptr == NULL)
5732 		return rc;
5733 
5734 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5735 	    i++, p = strsep(&next, " ,")) {
5736 		ndesc[i] = strtoul(p, NULL, 10);
5737 	}
5738 
5739 	return(rc);
5740 }
5741 
5742 #define NAME_BUFLEN 32
5743 static void
5744 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5745 {
5746         device_t dev = iflib_get_dev(ctx);
5747 	struct sysctl_oid_list *child, *oid_list;
5748 	struct sysctl_ctx_list *ctx_list;
5749 	struct sysctl_oid *node;
5750 
5751 	ctx_list = device_get_sysctl_ctx(dev);
5752 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5753 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5754 						      CTLFLAG_RD, NULL, "IFLIB fields");
5755 	oid_list = SYSCTL_CHILDREN(node);
5756 
5757 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5758 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5759 		       "driver version");
5760 
5761 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5762 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5763 			"# of txqs to use, 0 => use default #");
5764 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5765 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5766 			"# of rxqs to use, 0 => use default #");
5767 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5768 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5769                        "permit #txq != #rxq");
5770 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5771                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5772                       "disable MSIX (default 0)");
5773 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5774 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5775                        "set the rx budget");
5776 
5777 	/* XXX change for per-queue sizes */
5778 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5779 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5780                        mp_ndesc_handler, "A",
5781                        "list of # of tx descriptors to use, 0 = use default #");
5782 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5783 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5784                        mp_ndesc_handler, "A",
5785                        "list of # of rx descriptors to use, 0 = use default #");
5786 }
5787 
5788 static void
5789 iflib_add_device_sysctl_post(if_ctx_t ctx)
5790 {
5791 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5792 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5793         device_t dev = iflib_get_dev(ctx);
5794 	struct sysctl_oid_list *child;
5795 	struct sysctl_ctx_list *ctx_list;
5796 	iflib_fl_t fl;
5797 	iflib_txq_t txq;
5798 	iflib_rxq_t rxq;
5799 	int i, j;
5800 	char namebuf[NAME_BUFLEN];
5801 	char *qfmt;
5802 	struct sysctl_oid *queue_node, *fl_node, *node;
5803 	struct sysctl_oid_list *queue_list, *fl_list;
5804 	ctx_list = device_get_sysctl_ctx(dev);
5805 
5806 	node = ctx->ifc_sysctl_node;
5807 	child = SYSCTL_CHILDREN(node);
5808 
5809 	if (scctx->isc_ntxqsets > 100)
5810 		qfmt = "txq%03d";
5811 	else if (scctx->isc_ntxqsets > 10)
5812 		qfmt = "txq%02d";
5813 	else
5814 		qfmt = "txq%d";
5815 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5816 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5817 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5818 					     CTLFLAG_RD, NULL, "Queue Name");
5819 		queue_list = SYSCTL_CHILDREN(queue_node);
5820 #if MEMORY_LOGGING
5821 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5822 				CTLFLAG_RD,
5823 				&txq->ift_dequeued, "total mbufs freed");
5824 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5825 				CTLFLAG_RD,
5826 				&txq->ift_enqueued, "total mbufs enqueued");
5827 #endif
5828 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5829 				   CTLFLAG_RD,
5830 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5831 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5832 				   CTLFLAG_RD,
5833 				   &txq->ift_pullups, "# of times m_pullup was called");
5834 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5835 				   CTLFLAG_RD,
5836 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5837 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5838 				   CTLFLAG_RD,
5839 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5840 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5841 				   CTLFLAG_RD,
5842 				   &txq->ift_map_failed, "# of times dma map failed");
5843 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5844 				   CTLFLAG_RD,
5845 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5846 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5847 				   CTLFLAG_RD,
5848 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5849 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5850 				   CTLFLAG_RD,
5851 				   &txq->ift_pidx, 1, "Producer Index");
5852 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5853 				   CTLFLAG_RD,
5854 				   &txq->ift_cidx, 1, "Consumer Index");
5855 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5856 				   CTLFLAG_RD,
5857 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5858 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5859 				   CTLFLAG_RD,
5860 				   &txq->ift_in_use, 1, "descriptors in use");
5861 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5862 				   CTLFLAG_RD,
5863 				   &txq->ift_processed, "descriptors procesed for clean");
5864 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5865 				   CTLFLAG_RD,
5866 				   &txq->ift_cleaned, "total cleaned");
5867 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5868 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5869 				0, mp_ring_state_handler, "A", "soft ring state");
5870 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5871 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5872 				       "# of enqueues to the mp_ring for this queue");
5873 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5874 				       CTLFLAG_RD, &txq->ift_br->drops,
5875 				       "# of drops in the mp_ring for this queue");
5876 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5877 				       CTLFLAG_RD, &txq->ift_br->starts,
5878 				       "# of normal consumer starts in the mp_ring for this queue");
5879 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5880 				       CTLFLAG_RD, &txq->ift_br->stalls,
5881 					       "# of consumer stalls in the mp_ring for this queue");
5882 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5883 			       CTLFLAG_RD, &txq->ift_br->restarts,
5884 				       "# of consumer restarts in the mp_ring for this queue");
5885 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5886 				       CTLFLAG_RD, &txq->ift_br->abdications,
5887 				       "# of consumer abdications in the mp_ring for this queue");
5888 	}
5889 
5890 	if (scctx->isc_nrxqsets > 100)
5891 		qfmt = "rxq%03d";
5892 	else if (scctx->isc_nrxqsets > 10)
5893 		qfmt = "rxq%02d";
5894 	else
5895 		qfmt = "rxq%d";
5896 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5897 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5898 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5899 					     CTLFLAG_RD, NULL, "Queue Name");
5900 		queue_list = SYSCTL_CHILDREN(queue_node);
5901 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5902 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5903 				       CTLFLAG_RD,
5904 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5905 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5906 				       CTLFLAG_RD,
5907 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5908 		}
5909 
5910 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5911 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5912 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5913 						     CTLFLAG_RD, NULL, "freelist Name");
5914 			fl_list = SYSCTL_CHILDREN(fl_node);
5915 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5916 				       CTLFLAG_RD,
5917 				       &fl->ifl_pidx, 1, "Producer Index");
5918 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5919 				       CTLFLAG_RD,
5920 				       &fl->ifl_cidx, 1, "Consumer Index");
5921 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5922 				       CTLFLAG_RD,
5923 				       &fl->ifl_credits, 1, "credits available");
5924 #if MEMORY_LOGGING
5925 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5926 					CTLFLAG_RD,
5927 					&fl->ifl_m_enqueued, "mbufs allocated");
5928 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5929 					CTLFLAG_RD,
5930 					&fl->ifl_m_dequeued, "mbufs freed");
5931 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5932 					CTLFLAG_RD,
5933 					&fl->ifl_cl_enqueued, "clusters allocated");
5934 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5935 					CTLFLAG_RD,
5936 					&fl->ifl_cl_dequeued, "clusters freed");
5937 #endif
5938 
5939 		}
5940 	}
5941 
5942 }
5943 
5944 #ifndef __NO_STRICT_ALIGNMENT
5945 static struct mbuf *
5946 iflib_fixup_rx(struct mbuf *m)
5947 {
5948 	struct mbuf *n;
5949 
5950 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5951 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5952 		m->m_data += ETHER_HDR_LEN;
5953 		n = m;
5954 	} else {
5955 		MGETHDR(n, M_NOWAIT, MT_DATA);
5956 		if (n == NULL) {
5957 			m_freem(m);
5958 			return (NULL);
5959 		}
5960 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5961 		m->m_data += ETHER_HDR_LEN;
5962 		m->m_len -= ETHER_HDR_LEN;
5963 		n->m_len = ETHER_HDR_LEN;
5964 		M_MOVE_PKTHDR(n, m);
5965 		n->m_next = m;
5966 	}
5967 	return (n);
5968 }
5969 #endif
5970