xref: /freebsd/sys/net/iflib.c (revision 226fb85d1963ec0159016598d4f908468e49ca44)
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 	uint16_t ifc_nhwrxqs;
170 
171 	iflib_txq_t ifc_txqs;
172 	iflib_rxq_t ifc_rxqs;
173 	uint32_t ifc_if_flags;
174 	uint32_t ifc_flags;
175 	uint32_t ifc_max_fl_buf_size;
176 	int ifc_in_detach;
177 
178 	int ifc_link_state;
179 	int ifc_link_irq;
180 	int ifc_watchdog_events;
181 	struct cdev *ifc_led_dev;
182 	struct resource *ifc_msix_mem;
183 
184 	struct if_irq ifc_legacy_irq;
185 	struct grouptask ifc_admin_task;
186 	struct grouptask ifc_vflr_task;
187 	struct iflib_filter_info ifc_filter_info;
188 	struct ifmedia	ifc_media;
189 
190 	struct sysctl_oid *ifc_sysctl_node;
191 	uint16_t ifc_sysctl_ntxqs;
192 	uint16_t ifc_sysctl_nrxqs;
193 	uint16_t ifc_sysctl_qs_eq_override;
194 	uint16_t ifc_sysctl_rx_budget;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	uint8_t ifc_mac[ETHER_ADDR_LEN];
212 	char ifc_mtx_name[16];
213 };
214 
215 
216 void *
217 iflib_get_softc(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_softc);
221 }
222 
223 device_t
224 iflib_get_dev(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_dev);
228 }
229 
230 if_t
231 iflib_get_ifp(if_ctx_t ctx)
232 {
233 
234 	return (ctx->ifc_ifp);
235 }
236 
237 struct ifmedia *
238 iflib_get_media(if_ctx_t ctx)
239 {
240 
241 	return (&ctx->ifc_media);
242 }
243 
244 void
245 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
246 {
247 
248 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
249 }
250 
251 if_softc_ctx_t
252 iflib_get_softc_ctx(if_ctx_t ctx)
253 {
254 
255 	return (&ctx->ifc_softc_ctx);
256 }
257 
258 if_shared_ctx_t
259 iflib_get_sctx(if_ctx_t ctx)
260 {
261 
262 	return (ctx->ifc_sctx);
263 }
264 
265 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
266 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
267 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
268 
269 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
270 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
271 
272 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
273 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
274 #define RX_SW_DESC_INUSE        (1 << 3)
275 #define TX_SW_DESC_MAPPED       (1 << 4)
276 
277 #define	M_TOOBIG		M_PROTO1
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	uint8_t		*ifsd_flags;
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
289 	uint8_t		*ifsd_flags;
290 } if_txsd_vec_t;
291 
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 /* bnxt supports 64 with hardware LRO enabled */
296 #define IFLIB_MAX_RX_SEGS		64
297 #define IFLIB_RX_COPY_THRESH		128
298 #define IFLIB_MAX_RX_REFRESH		32
299 /* The minimum descriptors per second before we start coalescing */
300 #define IFLIB_MIN_DESC_SEC		16384
301 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
302 #define IFLIB_QUEUE_IDLE		0
303 #define IFLIB_QUEUE_HUNG		1
304 #define IFLIB_QUEUE_WORKING		2
305 /* maximum number of txqs that can share an rx interrupt */
306 #define IFLIB_MAX_TX_SHARED_INTR	4
307 
308 /* this should really scale with ring size - this is a fairly arbitrary value */
309 #define TX_BATCH_SIZE			32
310 
311 #define IFLIB_RESTART_BUDGET		8
312 
313 #define	IFC_LEGACY		0x001
314 #define	IFC_QFLUSH		0x002
315 #define	IFC_MULTISEG		0x004
316 #define	IFC_DMAR		0x008
317 #define	IFC_SC_ALLOCATED	0x010
318 #define	IFC_INIT_DONE		0x020
319 #define	IFC_PREFETCH		0x040
320 #define	IFC_DO_RESET		0x080
321 #define	IFC_CHECK_HUNG		0x100
322 
323 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
324 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
325 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
326 struct iflib_txq {
327 	qidx_t		ift_in_use;
328 	qidx_t		ift_cidx;
329 	qidx_t		ift_cidx_processed;
330 	qidx_t		ift_pidx;
331 	uint8_t		ift_gen;
332 	uint8_t		ift_br_offset;
333 	uint16_t	ift_npending;
334 	uint16_t	ift_db_pending;
335 	uint16_t	ift_rs_pending;
336 	/* implicit pad */
337 	uint8_t		ift_txd_size[8];
338 	uint64_t	ift_processed;
339 	uint64_t	ift_cleaned;
340 	uint64_t	ift_cleaned_prev;
341 #if MEMORY_LOGGING
342 	uint64_t	ift_enqueued;
343 	uint64_t	ift_dequeued;
344 #endif
345 	uint64_t	ift_no_tx_dma_setup;
346 	uint64_t	ift_no_desc_avail;
347 	uint64_t	ift_mbuf_defrag_failed;
348 	uint64_t	ift_mbuf_defrag;
349 	uint64_t	ift_map_failed;
350 	uint64_t	ift_txd_encap_efbig;
351 	uint64_t	ift_pullups;
352 
353 	struct mtx	ift_mtx;
354 	struct mtx	ift_db_mtx;
355 
356 	/* constant values */
357 	if_ctx_t	ift_ctx;
358 	struct ifmp_ring        *ift_br;
359 	struct grouptask	ift_task;
360 	qidx_t		ift_size;
361 	uint16_t	ift_id;
362 	struct callout	ift_timer;
363 
364 	if_txsd_vec_t	ift_sds;
365 	uint8_t		ift_qstatus;
366 	uint8_t		ift_closed;
367 	uint8_t		ift_update_freq;
368 	struct iflib_filter_info ift_filter_info;
369 	bus_dma_tag_t		ift_desc_tag;
370 	bus_dma_tag_t		ift_tso_desc_tag;
371 	iflib_dma_info_t	ift_ifdi;
372 #define MTX_NAME_LEN 16
373 	char                    ift_mtx_name[MTX_NAME_LEN];
374 	char                    ift_db_mtx_name[MTX_NAME_LEN];
375 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
376 #ifdef IFLIB_DIAGNOSTICS
377 	uint64_t ift_cpu_exec_count[256];
378 #endif
379 } __aligned(CACHE_LINE_SIZE);
380 
381 struct iflib_fl {
382 	qidx_t		ifl_cidx;
383 	qidx_t		ifl_pidx;
384 	qidx_t		ifl_credits;
385 	uint8_t		ifl_gen;
386 	uint8_t		ifl_rxd_size;
387 #if MEMORY_LOGGING
388 	uint64_t	ifl_m_enqueued;
389 	uint64_t	ifl_m_dequeued;
390 	uint64_t	ifl_cl_enqueued;
391 	uint64_t	ifl_cl_dequeued;
392 #endif
393 	/* implicit pad */
394 
395 	bitstr_t 	*ifl_rx_bitmap;
396 	qidx_t		ifl_fragidx;
397 	/* constant */
398 	qidx_t		ifl_size;
399 	uint16_t	ifl_buf_size;
400 	uint16_t	ifl_cltype;
401 	uma_zone_t	ifl_zone;
402 	iflib_rxsd_array_t	ifl_sds;
403 	iflib_rxq_t	ifl_rxq;
404 	uint8_t		ifl_id;
405 	bus_dma_tag_t           ifl_desc_tag;
406 	iflib_dma_info_t	ifl_ifdi;
407 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
408 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
409 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
410 }  __aligned(CACHE_LINE_SIZE);
411 
412 static inline qidx_t
413 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
414 {
415 	qidx_t used;
416 
417 	if (pidx > cidx)
418 		used = pidx - cidx;
419 	else if (pidx < cidx)
420 		used = size - cidx + pidx;
421 	else if (gen == 0 && pidx == cidx)
422 		used = 0;
423 	else if (gen == 1 && pidx == cidx)
424 		used = size;
425 	else
426 		panic("bad state");
427 
428 	return (used);
429 }
430 
431 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
432 
433 #define IDXDIFF(head, tail, wrap) \
434 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
435 
436 struct iflib_rxq {
437 	/* If there is a separate completion queue -
438 	 * these are the cq cidx and pidx. Otherwise
439 	 * these are unused.
440 	 */
441 	qidx_t		ifr_size;
442 	qidx_t		ifr_cq_cidx;
443 	qidx_t		ifr_cq_pidx;
444 	uint8_t		ifr_cq_gen;
445 	uint8_t		ifr_fl_offset;
446 
447 	if_ctx_t	ifr_ctx;
448 	iflib_fl_t	ifr_fl;
449 	uint64_t	ifr_rx_irq;
450 	uint16_t	ifr_id;
451 	uint8_t		ifr_lro_enabled;
452 	uint8_t		ifr_nfl;
453 	uint8_t		ifr_ntxqirq;
454 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
455 	struct lro_ctrl			ifr_lc;
456 	struct grouptask        ifr_task;
457 	struct iflib_filter_info ifr_filter_info;
458 	iflib_dma_info_t		ifr_ifdi;
459 
460 	/* dynamically allocate if any drivers need a value substantially larger than this */
461 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
462 #ifdef IFLIB_DIAGNOSTICS
463 	uint64_t ifr_cpu_exec_count[256];
464 #endif
465 }  __aligned(CACHE_LINE_SIZE);
466 
467 typedef struct if_rxsd {
468 	caddr_t *ifsd_cl;
469 	struct mbuf **ifsd_m;
470 	iflib_fl_t ifsd_fl;
471 	qidx_t ifsd_cidx;
472 } *if_rxsd_t;
473 
474 /* multiple of word size */
475 #ifdef __LP64__
476 #define PKT_INFO_SIZE	6
477 #define RXD_INFO_SIZE	5
478 #define PKT_TYPE uint64_t
479 #else
480 #define PKT_INFO_SIZE	11
481 #define RXD_INFO_SIZE	8
482 #define PKT_TYPE uint32_t
483 #endif
484 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
485 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
486 
487 typedef struct if_pkt_info_pad {
488 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
489 } *if_pkt_info_pad_t;
490 typedef struct if_rxd_info_pad {
491 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
492 } *if_rxd_info_pad_t;
493 
494 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
495 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
496 
497 
498 static inline void
499 pkt_info_zero(if_pkt_info_t pi)
500 {
501 	if_pkt_info_pad_t pi_pad;
502 
503 	pi_pad = (if_pkt_info_pad_t)pi;
504 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
505 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
506 #ifndef __LP64__
507 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
508 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
509 #endif
510 }
511 
512 static inline void
513 rxd_info_zero(if_rxd_info_t ri)
514 {
515 	if_rxd_info_pad_t ri_pad;
516 	int i;
517 
518 	ri_pad = (if_rxd_info_pad_t)ri;
519 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
520 		ri_pad->rxd_val[i] = 0;
521 		ri_pad->rxd_val[i+1] = 0;
522 		ri_pad->rxd_val[i+2] = 0;
523 		ri_pad->rxd_val[i+3] = 0;
524 	}
525 #ifdef __LP64__
526 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
527 #endif
528 }
529 
530 /*
531  * Only allow a single packet to take up most 1/nth of the tx ring
532  */
533 #define MAX_SINGLE_PACKET_FRACTION 12
534 #define IF_BAD_DMA (bus_addr_t)-1
535 
536 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
537 
538 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
539 
540 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
541 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
542 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
543 
544 
545 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
546 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
547 
548 
549 /* Our boot-time initialization hook */
550 static int	iflib_module_event_handler(module_t, int, void *);
551 
552 static moduledata_t iflib_moduledata = {
553 	"iflib",
554 	iflib_module_event_handler,
555 	NULL
556 };
557 
558 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
559 MODULE_VERSION(iflib, 1);
560 
561 MODULE_DEPEND(iflib, pci, 1, 1, 1);
562 MODULE_DEPEND(iflib, ether, 1, 1, 1);
563 
564 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
565 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
566 
567 #ifndef IFLIB_DEBUG_COUNTERS
568 #ifdef INVARIANTS
569 #define IFLIB_DEBUG_COUNTERS 1
570 #else
571 #define IFLIB_DEBUG_COUNTERS 0
572 #endif /* !INVARIANTS */
573 #endif
574 
575 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
576                    "iflib driver parameters");
577 
578 /*
579  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
580  */
581 static int iflib_min_tx_latency = 0;
582 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
583 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
584 static int iflib_no_tx_batch = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
586 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
587 
588 
589 #if IFLIB_DEBUG_COUNTERS
590 
591 static int iflib_tx_seen;
592 static int iflib_tx_sent;
593 static int iflib_tx_encap;
594 static int iflib_rx_allocs;
595 static int iflib_fl_refills;
596 static int iflib_fl_refills_large;
597 static int iflib_tx_frees;
598 
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
600 		   &iflib_tx_seen, 0, "# tx mbufs seen");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
602 		   &iflib_tx_sent, 0, "# tx mbufs sent");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
604 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
606 		   &iflib_tx_frees, 0, "# tx frees");
607 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
608 		   &iflib_rx_allocs, 0, "# rx allocations");
609 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
610 		   &iflib_fl_refills, 0, "# refills");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
612 		   &iflib_fl_refills_large, 0, "# large refills");
613 
614 
615 static int iflib_txq_drain_flushing;
616 static int iflib_txq_drain_oactive;
617 static int iflib_txq_drain_notready;
618 static int iflib_txq_drain_encapfail;
619 
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
621 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
623 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
625 		   &iflib_txq_drain_notready, 0, "# drain notready");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
627 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
628 
629 
630 static int iflib_encap_load_mbuf_fail;
631 static int iflib_encap_pad_mbuf_fail;
632 static int iflib_encap_txq_avail_fail;
633 static int iflib_encap_txd_encap_fail;
634 
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
636 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
638 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
640 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
642 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
643 
644 static int iflib_task_fn_rxs;
645 static int iflib_rx_intr_enables;
646 static int iflib_fast_intrs;
647 static int iflib_intr_link;
648 static int iflib_intr_msix;
649 static int iflib_rx_unavail;
650 static int iflib_rx_ctx_inactive;
651 static int iflib_rx_zero_len;
652 static int iflib_rx_if_input;
653 static int iflib_rx_mbuf_null;
654 static int iflib_rxd_flush;
655 
656 static int iflib_verbose_debug;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
659 		   &iflib_intr_link, 0, "# intr link calls");
660 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
661 		   &iflib_intr_msix, 0, "# intr msix calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
663 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
665 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
666 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
667 		   &iflib_fast_intrs, 0, "# fast_intr calls");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
669 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
671 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
673 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
675 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
677 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
678 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
679 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
680 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
681 		   &iflib_verbose_debug, 0, "enable verbose debugging");
682 
683 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
684 static void
685 iflib_debug_reset(void)
686 {
687 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
688 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
689 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
690 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
691 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
692 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
693 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
694 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
695 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
696 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
697 }
698 
699 #else
700 #define DBG_COUNTER_INC(name)
701 static void iflib_debug_reset(void) {}
702 #endif
703 
704 
705 
706 #define IFLIB_DEBUG 0
707 
708 static void iflib_tx_structures_free(if_ctx_t ctx);
709 static void iflib_rx_structures_free(if_ctx_t ctx);
710 static int iflib_queues_alloc(if_ctx_t ctx);
711 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
712 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
713 static int iflib_qset_structures_setup(if_ctx_t ctx);
714 static int iflib_msix_init(if_ctx_t ctx);
715 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
716 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
717 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
718 static int iflib_register(if_ctx_t);
719 static void iflib_init_locked(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
722 static void iflib_ifmp_purge(iflib_txq_t txq);
723 static void _iflib_pre_assert(if_softc_ctx_t scctx);
724 static void iflib_stop(if_ctx_t ctx);
725 static void iflib_if_init_locked(if_ctx_t ctx);
726 #ifndef __NO_STRICT_ALIGNMENT
727 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
728 #endif
729 
730 #ifdef DEV_NETMAP
731 #include <sys/selinfo.h>
732 #include <net/netmap.h>
733 #include <dev/netmap/netmap_kern.h>
734 
735 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
736 
737 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
738 
739 /*
740  * device-specific sysctl variables:
741  *
742  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
743  *	During regular operations the CRC is stripped, but on some
744  *	hardware reception of frames not multiple of 64 is slower,
745  *	so using crcstrip=0 helps in benchmarks.
746  *
747  * iflib_rx_miss, iflib_rx_miss_bufs:
748  *	count packets that might be missed due to lost interrupts.
749  */
750 SYSCTL_DECL(_dev_netmap);
751 /*
752  * The xl driver by default strips CRCs and we do not override it.
753  */
754 
755 int iflib_crcstrip = 1;
756 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
757     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
758 
759 int iflib_rx_miss, iflib_rx_miss_bufs;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
761     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
763     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
764 
765 /*
766  * Register/unregister. We are already under netmap lock.
767  * Only called on the first register or the last unregister.
768  */
769 static int
770 iflib_netmap_register(struct netmap_adapter *na, int onoff)
771 {
772 	struct ifnet *ifp = na->ifp;
773 	if_ctx_t ctx = ifp->if_softc;
774 	int status;
775 
776 	CTX_LOCK(ctx);
777 	IFDI_INTR_DISABLE(ctx);
778 
779 	/* Tell the stack that the interface is no longer active */
780 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
781 
782 	if (!CTX_IS_VF(ctx))
783 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
784 
785 	/* enable or disable flags and callbacks in na and ifp */
786 	if (onoff) {
787 		nm_set_native_flags(na);
788 	} else {
789 		nm_clear_native_flags(na);
790 	}
791 	iflib_stop(ctx);
792 	iflib_init_locked(ctx);
793 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
794 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
795 	if (status)
796 		nm_clear_native_flags(na);
797 	CTX_UNLOCK(ctx);
798 	return (status);
799 }
800 
801 static int
802 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int head = kring->rhead;
807 	struct netmap_ring *ring = kring->ring;
808 	bus_dmamap_t *map;
809 	struct if_rxd_update iru;
810 	if_ctx_t ctx = rxq->ifr_ctx;
811 	iflib_fl_t fl = &rxq->ifr_fl[0];
812 	uint32_t refill_pidx, nic_i;
813 
814 	if (nm_i == head && __predict_true(!init))
815 		return 0;
816 	iru_init(&iru, rxq, 0 /* flid */);
817 	map = fl->ifl_sds.ifsd_map;
818 	refill_pidx = netmap_idx_k2n(kring, nm_i);
819 	/*
820 	 * IMPORTANT: we must leave one free slot in the ring,
821 	 * so move head back by one unit
822 	 */
823 	head = nm_prev(head, lim);
824 	while (nm_i != head) {
825 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
826 			struct netmap_slot *slot = &ring->slot[nm_i];
827 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
828 			uint32_t nic_i_dma = refill_pidx;
829 			nic_i = netmap_idx_k2n(kring, nm_i);
830 
831 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
832 
833 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
834 			        return netmap_ring_reinit(kring);
835 
836 			fl->ifl_vm_addrs[tmp_pidx] = addr;
837 			if (__predict_false(init) && map) {
838 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
839 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
840 				/* buffer has changed, reload map */
841 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
842 			}
843 			slot->flags &= ~NS_BUF_CHANGED;
844 
845 			nm_i = nm_next(nm_i, lim);
846 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
847 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
848 				continue;
849 
850 			iru.iru_pidx = refill_pidx;
851 			iru.iru_count = tmp_pidx+1;
852 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
853 
854 			refill_pidx = nic_i;
855 			if (map == NULL)
856 				continue;
857 
858 			for (int n = 0; n < iru.iru_count; n++) {
859 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
860 						BUS_DMASYNC_PREREAD);
861 				/* XXX - change this to not use the netmap func*/
862 				nic_i_dma = nm_next(nic_i_dma, lim);
863 			}
864 		}
865 	}
866 	kring->nr_hwcur = head;
867 
868 	if (map)
869 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
870 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
871 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
872 	return (0);
873 }
874 
875 /*
876  * Reconcile kernel and user view of the transmit ring.
877  *
878  * All information is in the kring.
879  * Userspace wants to send packets up to the one before kring->rhead,
880  * kernel knows kring->nr_hwcur is the first unsent packet.
881  *
882  * Here we push packets out (as many as possible), and possibly
883  * reclaim buffers from previously completed transmission.
884  *
885  * The caller (netmap) guarantees that there is only one instance
886  * running at any time. Any interference with other driver
887  * methods should be handled by the individual drivers.
888  */
889 static int
890 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
891 {
892 	struct netmap_adapter *na = kring->na;
893 	struct ifnet *ifp = na->ifp;
894 	struct netmap_ring *ring = kring->ring;
895 	u_int nm_i;	/* index into the netmap ring */
896 	u_int nic_i;	/* index into the NIC ring */
897 	u_int n;
898 	u_int const lim = kring->nkr_num_slots - 1;
899 	u_int const head = kring->rhead;
900 	struct if_pkt_info pi;
901 
902 	/*
903 	 * interrupts on every tx packet are expensive so request
904 	 * them every half ring, or where NS_REPORT is set
905 	 */
906 	u_int report_frequency = kring->nkr_num_slots >> 1;
907 	/* device-specific */
908 	if_ctx_t ctx = ifp->if_softc;
909 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
910 
911 	if (txq->ift_sds.ifsd_map)
912 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
913 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
914 
915 
916 	/*
917 	 * First part: process new packets to send.
918 	 * nm_i is the current index in the netmap ring,
919 	 * nic_i is the corresponding index in the NIC ring.
920 	 *
921 	 * If we have packets to send (nm_i != head)
922 	 * iterate over the netmap ring, fetch length and update
923 	 * the corresponding slot in the NIC ring. Some drivers also
924 	 * need to update the buffer's physical address in the NIC slot
925 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
926 	 *
927 	 * The netmap_reload_map() calls is especially expensive,
928 	 * even when (as in this case) the tag is 0, so do only
929 	 * when the buffer has actually changed.
930 	 *
931 	 * If possible do not set the report/intr bit on all slots,
932 	 * but only a few times per ring or when NS_REPORT is set.
933 	 *
934 	 * Finally, on 10G and faster drivers, it might be useful
935 	 * to prefetch the next slot and txr entry.
936 	 */
937 
938 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
939 	pkt_info_zero(&pi);
940 	pi.ipi_segs = txq->ift_segs;
941 	pi.ipi_qsidx = kring->ring_id;
942 	if (nm_i != head) {	/* we have new packets to send */
943 		nic_i = netmap_idx_k2n(kring, nm_i);
944 
945 		__builtin_prefetch(&ring->slot[nm_i]);
946 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
947 		if (txq->ift_sds.ifsd_map)
948 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
949 
950 		for (n = 0; nm_i != head; n++) {
951 			struct netmap_slot *slot = &ring->slot[nm_i];
952 			u_int len = slot->len;
953 			uint64_t paddr;
954 			void *addr = PNMB(na, slot, &paddr);
955 			int flags = (slot->flags & NS_REPORT ||
956 				nic_i == 0 || nic_i == report_frequency) ?
957 				IPI_TX_INTR : 0;
958 
959 			/* device-specific */
960 			pi.ipi_len = len;
961 			pi.ipi_segs[0].ds_addr = paddr;
962 			pi.ipi_segs[0].ds_len = len;
963 			pi.ipi_nsegs = 1;
964 			pi.ipi_ndescs = 0;
965 			pi.ipi_pidx = nic_i;
966 			pi.ipi_flags = flags;
967 
968 			/* Fill the slot in the NIC ring. */
969 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
970 
971 			/* prefetch for next round */
972 			__builtin_prefetch(&ring->slot[nm_i + 1]);
973 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
974 			if (txq->ift_sds.ifsd_map) {
975 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
976 
977 				NM_CHECK_ADDR_LEN(na, addr, len);
978 
979 				if (slot->flags & NS_BUF_CHANGED) {
980 					/* buffer has changed, reload map */
981 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
982 				}
983 				/* make sure changes to the buffer are synced */
984 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
985 						BUS_DMASYNC_PREWRITE);
986 			}
987 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
988 			nm_i = nm_next(nm_i, lim);
989 			nic_i = nm_next(nic_i, lim);
990 		}
991 		kring->nr_hwcur = head;
992 
993 		/* synchronize the NIC ring */
994 		if (txq->ift_sds.ifsd_map)
995 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
996 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
997 
998 		/* (re)start the tx unit up to slot nic_i (excluded) */
999 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1000 	}
1001 
1002 	/*
1003 	 * Second part: reclaim buffers for completed transmissions.
1004 	 */
1005 	if (iflib_tx_credits_update(ctx, txq)) {
1006 		/* some tx completed, increment avail */
1007 		nic_i = txq->ift_cidx_processed;
1008 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1009 	}
1010 	return (0);
1011 }
1012 
1013 /*
1014  * Reconcile kernel and user view of the receive ring.
1015  * Same as for the txsync, this routine must be efficient.
1016  * The caller guarantees a single invocations, but races against
1017  * the rest of the driver should be handled here.
1018  *
1019  * On call, kring->rhead is the first packet that userspace wants
1020  * to keep, and kring->rcur is the wakeup point.
1021  * The kernel has previously reported packets up to kring->rtail.
1022  *
1023  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1024  * of whether or not we received an interrupt.
1025  */
1026 static int
1027 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1028 {
1029 	struct netmap_adapter *na = kring->na;
1030 	struct netmap_ring *ring = kring->ring;
1031 	uint32_t nm_i;	/* index into the netmap ring */
1032 	uint32_t nic_i;	/* index into the NIC ring */
1033 	u_int i, n;
1034 	u_int const lim = kring->nkr_num_slots - 1;
1035 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1036 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1037 	struct if_rxd_info ri;
1038 
1039 	struct ifnet *ifp = na->ifp;
1040 	if_ctx_t ctx = ifp->if_softc;
1041 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1042 	iflib_fl_t fl = rxq->ifr_fl;
1043 	if (head > lim)
1044 		return netmap_ring_reinit(kring);
1045 
1046 	/* XXX check sync modes */
1047 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1048 		if (fl->ifl_sds.ifsd_map == NULL)
1049 			continue;
1050 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1051 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1052 	}
1053 	/*
1054 	 * First part: import newly received packets.
1055 	 *
1056 	 * nm_i is the index of the next free slot in the netmap ring,
1057 	 * nic_i is the index of the next received packet in the NIC ring,
1058 	 * and they may differ in case if_init() has been called while
1059 	 * in netmap mode. For the receive ring we have
1060 	 *
1061 	 *	nic_i = rxr->next_check;
1062 	 *	nm_i = kring->nr_hwtail (previous)
1063 	 * and
1064 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1065 	 *
1066 	 * rxr->next_check is set to 0 on a ring reinit
1067 	 */
1068 	if (netmap_no_pendintr || force_update) {
1069 		int crclen = iflib_crcstrip ? 0 : 4;
1070 		int error, avail;
1071 
1072 		for (i = 0; i < rxq->ifr_nfl; i++) {
1073 			fl = &rxq->ifr_fl[i];
1074 			nic_i = fl->ifl_cidx;
1075 			nm_i = netmap_idx_n2k(kring, nic_i);
1076 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1077 			for (n = 0; avail > 0; n++, avail--) {
1078 				rxd_info_zero(&ri);
1079 				ri.iri_frags = rxq->ifr_frags;
1080 				ri.iri_qsidx = kring->ring_id;
1081 				ri.iri_ifp = ctx->ifc_ifp;
1082 				ri.iri_cidx = nic_i;
1083 
1084 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1085 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1086 				ring->slot[nm_i].flags = 0;
1087 				if (fl->ifl_sds.ifsd_map)
1088 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1089 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1090 				nm_i = nm_next(nm_i, lim);
1091 				nic_i = nm_next(nic_i, lim);
1092 			}
1093 			if (n) { /* update the state variables */
1094 				if (netmap_no_pendintr && !force_update) {
1095 					/* diagnostics */
1096 					iflib_rx_miss ++;
1097 					iflib_rx_miss_bufs += n;
1098 				}
1099 				fl->ifl_cidx = nic_i;
1100 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1101 			}
1102 			kring->nr_kflags &= ~NKR_PENDINTR;
1103 		}
1104 	}
1105 	/*
1106 	 * Second part: skip past packets that userspace has released.
1107 	 * (kring->nr_hwcur to head excluded),
1108 	 * and make the buffers available for reception.
1109 	 * As usual nm_i is the index in the netmap ring,
1110 	 * nic_i is the index in the NIC ring, and
1111 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1112 	 */
1113 	/* XXX not sure how this will work with multiple free lists */
1114 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1115 
1116 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1117 }
1118 
1119 static void
1120 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1121 {
1122 	struct ifnet *ifp = na->ifp;
1123 	if_ctx_t ctx = ifp->if_softc;
1124 
1125 	CTX_LOCK(ctx);
1126 	if (onoff) {
1127 		IFDI_INTR_ENABLE(ctx);
1128 	} else {
1129 		IFDI_INTR_DISABLE(ctx);
1130 	}
1131 	CTX_UNLOCK(ctx);
1132 }
1133 
1134 
1135 static int
1136 iflib_netmap_attach(if_ctx_t ctx)
1137 {
1138 	struct netmap_adapter na;
1139 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1140 
1141 	bzero(&na, sizeof(na));
1142 
1143 	na.ifp = ctx->ifc_ifp;
1144 	na.na_flags = NAF_BDG_MAYSLEEP;
1145 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1146 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1147 
1148 	na.num_tx_desc = scctx->isc_ntxd[0];
1149 	na.num_rx_desc = scctx->isc_nrxd[0];
1150 	na.nm_txsync = iflib_netmap_txsync;
1151 	na.nm_rxsync = iflib_netmap_rxsync;
1152 	na.nm_register = iflib_netmap_register;
1153 	na.nm_intr = iflib_netmap_intr;
1154 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1155 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1156 	return (netmap_attach(&na));
1157 }
1158 
1159 static void
1160 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1161 {
1162 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1163 	struct netmap_slot *slot;
1164 
1165 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1166 	if (slot == NULL)
1167 		return;
1168 	if (txq->ift_sds.ifsd_map == NULL)
1169 		return;
1170 
1171 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1172 
1173 		/*
1174 		 * In netmap mode, set the map for the packet buffer.
1175 		 * NOTE: Some drivers (not this one) also need to set
1176 		 * the physical buffer address in the NIC ring.
1177 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1178 		 * netmap slot index, si
1179 		 */
1180 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1181 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1182 	}
1183 }
1184 
1185 static void
1186 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1187 {
1188 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1189 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1190 	struct netmap_slot *slot;
1191 	uint32_t nm_i;
1192 
1193 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1194 	if (slot == NULL)
1195 		return;
1196 	nm_i = netmap_idx_n2k(kring, 0);
1197 	netmap_fl_refill(rxq, kring, nm_i, true);
1198 }
1199 
1200 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1201 
1202 #else
1203 #define iflib_netmap_txq_init(ctx, txq)
1204 #define iflib_netmap_rxq_init(ctx, rxq)
1205 #define iflib_netmap_detach(ifp)
1206 
1207 #define iflib_netmap_attach(ctx) (0)
1208 #define netmap_rx_irq(ifp, qid, budget) (0)
1209 #define netmap_tx_irq(ifp, qid) do {} while (0)
1210 
1211 #endif
1212 
1213 #if defined(__i386__) || defined(__amd64__)
1214 static __inline void
1215 prefetch(void *x)
1216 {
1217 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1218 }
1219 static __inline void
1220 prefetch2cachelines(void *x)
1221 {
1222 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1223 #if (CACHE_LINE_SIZE < 128)
1224 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1225 #endif
1226 }
1227 #else
1228 #define prefetch(x)
1229 #define prefetch2cachelines(x)
1230 #endif
1231 
1232 static void
1233 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1234 {
1235 	iflib_fl_t fl;
1236 
1237 	fl = &rxq->ifr_fl[flid];
1238 	iru->iru_paddrs = fl->ifl_bus_addrs;
1239 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1240 	iru->iru_idxs = fl->ifl_rxd_idxs;
1241 	iru->iru_qsidx = rxq->ifr_id;
1242 	iru->iru_buf_size = fl->ifl_buf_size;
1243 	iru->iru_flidx = fl->ifl_id;
1244 }
1245 
1246 static void
1247 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1248 {
1249 	if (err)
1250 		return;
1251 	*(bus_addr_t *) arg = segs[0].ds_addr;
1252 }
1253 
1254 int
1255 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1256 {
1257 	int err;
1258 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1259 	device_t dev = ctx->ifc_dev;
1260 
1261 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1262 
1263 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1264 				sctx->isc_q_align, 0,	/* alignment, bounds */
1265 				BUS_SPACE_MAXADDR,	/* lowaddr */
1266 				BUS_SPACE_MAXADDR,	/* highaddr */
1267 				NULL, NULL,		/* filter, filterarg */
1268 				size,			/* maxsize */
1269 				1,			/* nsegments */
1270 				size,			/* maxsegsize */
1271 				BUS_DMA_ALLOCNOW,	/* flags */
1272 				NULL,			/* lockfunc */
1273 				NULL,			/* lockarg */
1274 				&dma->idi_tag);
1275 	if (err) {
1276 		device_printf(dev,
1277 		    "%s: bus_dma_tag_create failed: %d\n",
1278 		    __func__, err);
1279 		goto fail_0;
1280 	}
1281 
1282 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1283 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1284 	if (err) {
1285 		device_printf(dev,
1286 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1287 		    __func__, (uintmax_t)size, err);
1288 		goto fail_1;
1289 	}
1290 
1291 	dma->idi_paddr = IF_BAD_DMA;
1292 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1293 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1294 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1295 		device_printf(dev,
1296 		    "%s: bus_dmamap_load failed: %d\n",
1297 		    __func__, err);
1298 		goto fail_2;
1299 	}
1300 
1301 	dma->idi_size = size;
1302 	return (0);
1303 
1304 fail_2:
1305 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1306 fail_1:
1307 	bus_dma_tag_destroy(dma->idi_tag);
1308 fail_0:
1309 	dma->idi_tag = NULL;
1310 
1311 	return (err);
1312 }
1313 
1314 int
1315 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1316 {
1317 	int i, err;
1318 	iflib_dma_info_t *dmaiter;
1319 
1320 	dmaiter = dmalist;
1321 	for (i = 0; i < count; i++, dmaiter++) {
1322 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1323 			break;
1324 	}
1325 	if (err)
1326 		iflib_dma_free_multi(dmalist, i);
1327 	return (err);
1328 }
1329 
1330 void
1331 iflib_dma_free(iflib_dma_info_t dma)
1332 {
1333 	if (dma->idi_tag == NULL)
1334 		return;
1335 	if (dma->idi_paddr != IF_BAD_DMA) {
1336 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1337 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1338 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1339 		dma->idi_paddr = IF_BAD_DMA;
1340 	}
1341 	if (dma->idi_vaddr != NULL) {
1342 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1343 		dma->idi_vaddr = NULL;
1344 	}
1345 	bus_dma_tag_destroy(dma->idi_tag);
1346 	dma->idi_tag = NULL;
1347 }
1348 
1349 void
1350 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1351 {
1352 	int i;
1353 	iflib_dma_info_t *dmaiter = dmalist;
1354 
1355 	for (i = 0; i < count; i++, dmaiter++)
1356 		iflib_dma_free(*dmaiter);
1357 }
1358 
1359 #ifdef EARLY_AP_STARTUP
1360 static const int iflib_started = 1;
1361 #else
1362 /*
1363  * We used to abuse the smp_started flag to decide if the queues have been
1364  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1365  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1366  * is set.  Run a SYSINIT() strictly after that to just set a usable
1367  * completion flag.
1368  */
1369 
1370 static int iflib_started;
1371 
1372 static void
1373 iflib_record_started(void *arg)
1374 {
1375 	iflib_started = 1;
1376 }
1377 
1378 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1379 	iflib_record_started, NULL);
1380 #endif
1381 
1382 static int
1383 iflib_fast_intr(void *arg)
1384 {
1385 	iflib_filter_info_t info = arg;
1386 	struct grouptask *gtask = info->ifi_task;
1387 	if (!iflib_started)
1388 		return (FILTER_HANDLED);
1389 
1390 	DBG_COUNTER_INC(fast_intrs);
1391 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1392 		return (FILTER_HANDLED);
1393 
1394 	GROUPTASK_ENQUEUE(gtask);
1395 	return (FILTER_HANDLED);
1396 }
1397 
1398 static int
1399 iflib_fast_intr_rxtx(void *arg)
1400 {
1401 	iflib_filter_info_t info = arg;
1402 	struct grouptask *gtask = info->ifi_task;
1403 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1404 	if_ctx_t ctx;
1405 	int i, cidx;
1406 
1407 	if (!iflib_started)
1408 		return (FILTER_HANDLED);
1409 
1410 	DBG_COUNTER_INC(fast_intrs);
1411 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1412 		return (FILTER_HANDLED);
1413 
1414 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1415 		qidx_t txqid = rxq->ifr_txqid[i];
1416 
1417 		ctx = rxq->ifr_ctx;
1418 
1419 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1420 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1421 			continue;
1422 		}
1423 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1424 	}
1425 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1426 		cidx = rxq->ifr_cq_cidx;
1427 	else
1428 		cidx = rxq->ifr_fl[0].ifl_cidx;
1429 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1430 		GROUPTASK_ENQUEUE(gtask);
1431 	else
1432 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1433 	return (FILTER_HANDLED);
1434 }
1435 
1436 
1437 static int
1438 iflib_fast_intr_ctx(void *arg)
1439 {
1440 	iflib_filter_info_t info = arg;
1441 	struct grouptask *gtask = info->ifi_task;
1442 
1443 	if (!iflib_started)
1444 		return (FILTER_HANDLED);
1445 
1446 	DBG_COUNTER_INC(fast_intrs);
1447 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1448 		return (FILTER_HANDLED);
1449 
1450 	GROUPTASK_ENQUEUE(gtask);
1451 	return (FILTER_HANDLED);
1452 }
1453 
1454 static int
1455 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1456 	driver_filter_t filter, driver_intr_t handler, void *arg,
1457 				 char *name)
1458 {
1459 	int rc, flags;
1460 	struct resource *res;
1461 	void *tag = NULL;
1462 	device_t dev = ctx->ifc_dev;
1463 
1464 	flags = RF_ACTIVE;
1465 	if (ctx->ifc_flags & IFC_LEGACY)
1466 		flags |= RF_SHAREABLE;
1467 	MPASS(rid < 512);
1468 	irq->ii_rid = rid;
1469 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1470 	if (res == NULL) {
1471 		device_printf(dev,
1472 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1473 		return (ENOMEM);
1474 	}
1475 	irq->ii_res = res;
1476 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1477 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1478 						filter, handler, arg, &tag);
1479 	if (rc != 0) {
1480 		device_printf(dev,
1481 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1482 					  rid, name ? name : "unknown", rc);
1483 		return (rc);
1484 	} else if (name)
1485 		bus_describe_intr(dev, res, tag, "%s", name);
1486 
1487 	irq->ii_tag = tag;
1488 	return (0);
1489 }
1490 
1491 
1492 /*********************************************************************
1493  *
1494  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1495  *  the information needed to transmit a packet on the wire. This is
1496  *  called only once at attach, setup is done every reset.
1497  *
1498  **********************************************************************/
1499 
1500 static int
1501 iflib_txsd_alloc(iflib_txq_t txq)
1502 {
1503 	if_ctx_t ctx = txq->ift_ctx;
1504 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1505 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1506 	device_t dev = ctx->ifc_dev;
1507 	int err, nsegments, ntsosegments;
1508 
1509 	nsegments = scctx->isc_tx_nsegments;
1510 	ntsosegments = scctx->isc_tx_tso_segments_max;
1511 	MPASS(scctx->isc_ntxd[0] > 0);
1512 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1513 	MPASS(nsegments > 0);
1514 	MPASS(ntsosegments > 0);
1515 	/*
1516 	 * Setup DMA descriptor areas.
1517 	 */
1518 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1519 			       1, 0,			/* alignment, bounds */
1520 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1521 			       BUS_SPACE_MAXADDR,	/* highaddr */
1522 			       NULL, NULL,		/* filter, filterarg */
1523 			       sctx->isc_tx_maxsize,		/* maxsize */
1524 			       nsegments,	/* nsegments */
1525 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1526 			       0,			/* flags */
1527 			       NULL,			/* lockfunc */
1528 			       NULL,			/* lockfuncarg */
1529 			       &txq->ift_desc_tag))) {
1530 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1531 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1532 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1533 		goto fail;
1534 	}
1535 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1536 			       1, 0,			/* alignment, bounds */
1537 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1538 			       BUS_SPACE_MAXADDR,	/* highaddr */
1539 			       NULL, NULL,		/* filter, filterarg */
1540 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1541 			       ntsosegments,	/* nsegments */
1542 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1543 			       0,			/* flags */
1544 			       NULL,			/* lockfunc */
1545 			       NULL,			/* lockfuncarg */
1546 			       &txq->ift_tso_desc_tag))) {
1547 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1548 
1549 		goto fail;
1550 	}
1551 	if (!(txq->ift_sds.ifsd_flags =
1552 	    (uint8_t *) malloc(sizeof(uint8_t) *
1553 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1554 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1555 		err = ENOMEM;
1556 		goto fail;
1557 	}
1558 	if (!(txq->ift_sds.ifsd_m =
1559 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1560 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1561 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1562 		err = ENOMEM;
1563 		goto fail;
1564 	}
1565 
1566         /* Create the descriptor buffer dma maps */
1567 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1568 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1569 		return (0);
1570 
1571 	if (!(txq->ift_sds.ifsd_map =
1572 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1573 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1574 		err = ENOMEM;
1575 		goto fail;
1576 	}
1577 
1578 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1579 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1580 		if (err != 0) {
1581 			device_printf(dev, "Unable to create TX DMA map\n");
1582 			goto fail;
1583 		}
1584 	}
1585 #endif
1586 	return (0);
1587 fail:
1588 	/* We free all, it handles case where we are in the middle */
1589 	iflib_tx_structures_free(ctx);
1590 	return (err);
1591 }
1592 
1593 static void
1594 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1595 {
1596 	bus_dmamap_t map;
1597 
1598 	map = NULL;
1599 	if (txq->ift_sds.ifsd_map != NULL)
1600 		map = txq->ift_sds.ifsd_map[i];
1601 	if (map != NULL) {
1602 		bus_dmamap_unload(txq->ift_desc_tag, map);
1603 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1604 		txq->ift_sds.ifsd_map[i] = NULL;
1605 	}
1606 }
1607 
1608 static void
1609 iflib_txq_destroy(iflib_txq_t txq)
1610 {
1611 	if_ctx_t ctx = txq->ift_ctx;
1612 
1613 	for (int i = 0; i < txq->ift_size; i++)
1614 		iflib_txsd_destroy(ctx, txq, i);
1615 	if (txq->ift_sds.ifsd_map != NULL) {
1616 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1617 		txq->ift_sds.ifsd_map = NULL;
1618 	}
1619 	if (txq->ift_sds.ifsd_m != NULL) {
1620 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1621 		txq->ift_sds.ifsd_m = NULL;
1622 	}
1623 	if (txq->ift_sds.ifsd_flags != NULL) {
1624 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1625 		txq->ift_sds.ifsd_flags = NULL;
1626 	}
1627 	if (txq->ift_desc_tag != NULL) {
1628 		bus_dma_tag_destroy(txq->ift_desc_tag);
1629 		txq->ift_desc_tag = NULL;
1630 	}
1631 	if (txq->ift_tso_desc_tag != NULL) {
1632 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1633 		txq->ift_tso_desc_tag = NULL;
1634 	}
1635 }
1636 
1637 static void
1638 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1639 {
1640 	struct mbuf **mp;
1641 
1642 	mp = &txq->ift_sds.ifsd_m[i];
1643 	if (*mp == NULL)
1644 		return;
1645 
1646 	if (txq->ift_sds.ifsd_map != NULL) {
1647 		bus_dmamap_sync(txq->ift_desc_tag,
1648 				txq->ift_sds.ifsd_map[i],
1649 				BUS_DMASYNC_POSTWRITE);
1650 		bus_dmamap_unload(txq->ift_desc_tag,
1651 				  txq->ift_sds.ifsd_map[i]);
1652 	}
1653 	m_free(*mp);
1654 	DBG_COUNTER_INC(tx_frees);
1655 	*mp = NULL;
1656 }
1657 
1658 static int
1659 iflib_txq_setup(iflib_txq_t txq)
1660 {
1661 	if_ctx_t ctx = txq->ift_ctx;
1662 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1663 	iflib_dma_info_t di;
1664 	int i;
1665 
1666 	/* Set number of descriptors available */
1667 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1668 	/* XXX make configurable */
1669 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1670 
1671 	/* Reset indices */
1672 	txq->ift_cidx_processed = 0;
1673 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1674 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1675 
1676 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1677 		bzero((void *)di->idi_vaddr, di->idi_size);
1678 
1679 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1680 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1681 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1682 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1683 	return (0);
1684 }
1685 
1686 /*********************************************************************
1687  *
1688  *  Allocate memory for rx_buffer structures. Since we use one
1689  *  rx_buffer per received packet, the maximum number of rx_buffer's
1690  *  that we'll need is equal to the number of receive descriptors
1691  *  that we've allocated.
1692  *
1693  **********************************************************************/
1694 static int
1695 iflib_rxsd_alloc(iflib_rxq_t rxq)
1696 {
1697 	if_ctx_t ctx = rxq->ifr_ctx;
1698 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1699 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1700 	device_t dev = ctx->ifc_dev;
1701 	iflib_fl_t fl;
1702 	int			err;
1703 
1704 	MPASS(scctx->isc_nrxd[0] > 0);
1705 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1706 
1707 	fl = rxq->ifr_fl;
1708 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1709 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1710 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1711 					 1, 0,			/* alignment, bounds */
1712 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1713 					 BUS_SPACE_MAXADDR,	/* highaddr */
1714 					 NULL, NULL,		/* filter, filterarg */
1715 					 sctx->isc_rx_maxsize,	/* maxsize */
1716 					 sctx->isc_rx_nsegments,	/* nsegments */
1717 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1718 					 0,			/* flags */
1719 					 NULL,			/* lockfunc */
1720 					 NULL,			/* lockarg */
1721 					 &fl->ifl_desc_tag);
1722 		if (err) {
1723 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1724 				__func__, err);
1725 			goto fail;
1726 		}
1727 		if (!(fl->ifl_sds.ifsd_flags =
1728 		      (uint8_t *) malloc(sizeof(uint8_t) *
1729 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1730 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1731 			err = ENOMEM;
1732 			goto fail;
1733 		}
1734 		if (!(fl->ifl_sds.ifsd_m =
1735 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1736 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1737 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1738 			err = ENOMEM;
1739 			goto fail;
1740 		}
1741 		if (!(fl->ifl_sds.ifsd_cl =
1742 		      (caddr_t *) malloc(sizeof(caddr_t) *
1743 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1744 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1745 			err = ENOMEM;
1746 			goto fail;
1747 		}
1748 
1749 		/* Create the descriptor buffer dma maps */
1750 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1751 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1752 			continue;
1753 
1754 		if (!(fl->ifl_sds.ifsd_map =
1755 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1756 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1757 			err = ENOMEM;
1758 			goto fail;
1759 		}
1760 
1761 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1762 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1763 			if (err != 0) {
1764 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1765 				goto fail;
1766 			}
1767 		}
1768 #endif
1769 	}
1770 	return (0);
1771 
1772 fail:
1773 	iflib_rx_structures_free(ctx);
1774 	return (err);
1775 }
1776 
1777 
1778 /*
1779  * Internal service routines
1780  */
1781 
1782 struct rxq_refill_cb_arg {
1783 	int               error;
1784 	bus_dma_segment_t seg;
1785 	int               nseg;
1786 };
1787 
1788 static void
1789 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1790 {
1791 	struct rxq_refill_cb_arg *cb_arg = arg;
1792 
1793 	cb_arg->error = error;
1794 	cb_arg->seg = segs[0];
1795 	cb_arg->nseg = nseg;
1796 }
1797 
1798 
1799 #ifdef ACPI_DMAR
1800 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1801 #else
1802 #define IS_DMAR(ctx) (0)
1803 #endif
1804 
1805 /**
1806  *	rxq_refill - refill an rxq  free-buffer list
1807  *	@ctx: the iflib context
1808  *	@rxq: the free-list to refill
1809  *	@n: the number of new buffers to allocate
1810  *
1811  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1812  *	The caller must assure that @n does not exceed the queue's capacity.
1813  */
1814 static void
1815 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1816 {
1817 	struct mbuf *m;
1818 	int idx, frag_idx = fl->ifl_fragidx;
1819         int pidx = fl->ifl_pidx;
1820 	caddr_t cl, *sd_cl;
1821 	struct mbuf **sd_m;
1822 	uint8_t *sd_flags;
1823 	struct if_rxd_update iru;
1824 	bus_dmamap_t *sd_map;
1825 	int n, i = 0;
1826 	uint64_t bus_addr;
1827 	int err;
1828 	qidx_t credits;
1829 
1830 	sd_m = fl->ifl_sds.ifsd_m;
1831 	sd_map = fl->ifl_sds.ifsd_map;
1832 	sd_cl = fl->ifl_sds.ifsd_cl;
1833 	sd_flags = fl->ifl_sds.ifsd_flags;
1834 	idx = pidx;
1835 	credits = fl->ifl_credits;
1836 
1837 	n  = count;
1838 	MPASS(n > 0);
1839 	MPASS(credits + n <= fl->ifl_size);
1840 
1841 	if (pidx < fl->ifl_cidx)
1842 		MPASS(pidx + n <= fl->ifl_cidx);
1843 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1844 		MPASS(fl->ifl_gen == 0);
1845 	if (pidx > fl->ifl_cidx)
1846 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1847 
1848 	DBG_COUNTER_INC(fl_refills);
1849 	if (n > 8)
1850 		DBG_COUNTER_INC(fl_refills_large);
1851 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1852 	while (n--) {
1853 		/*
1854 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1855 		 * initialized after rx.
1856 		 *
1857 		 * If the cluster is still set then we know a minimum sized packet was received
1858 		 */
1859 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1860 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1861                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1862 		if ((cl = sd_cl[frag_idx]) == NULL) {
1863                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1864 				break;
1865 #if MEMORY_LOGGING
1866 			fl->ifl_cl_enqueued++;
1867 #endif
1868 		}
1869 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1870 			break;
1871 		}
1872 #if MEMORY_LOGGING
1873 		fl->ifl_m_enqueued++;
1874 #endif
1875 
1876 		DBG_COUNTER_INC(rx_allocs);
1877 #if defined(__i386__) || defined(__amd64__)
1878 		if (!IS_DMAR(ctx)) {
1879 			bus_addr = pmap_kextract((vm_offset_t)cl);
1880 		} else
1881 #endif
1882 		{
1883 			struct rxq_refill_cb_arg cb_arg;
1884 			iflib_rxq_t q;
1885 
1886 			cb_arg.error = 0;
1887 			q = fl->ifl_rxq;
1888 			MPASS(sd_map != NULL);
1889 			MPASS(sd_map[frag_idx] != NULL);
1890 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1891 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1892 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1893 					BUS_DMASYNC_PREREAD);
1894 
1895 			if (err != 0 || cb_arg.error) {
1896 				/*
1897 				 * !zone_pack ?
1898 				 */
1899 				if (fl->ifl_zone == zone_pack)
1900 					uma_zfree(fl->ifl_zone, cl);
1901 				m_free(m);
1902 				n = 0;
1903 				goto done;
1904 			}
1905 			bus_addr = cb_arg.seg.ds_addr;
1906 		}
1907                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1908 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1909 
1910 		MPASS(sd_m[frag_idx] == NULL);
1911 		sd_cl[frag_idx] = cl;
1912 		sd_m[frag_idx] = m;
1913 		fl->ifl_rxd_idxs[i] = frag_idx;
1914 		fl->ifl_bus_addrs[i] = bus_addr;
1915 		fl->ifl_vm_addrs[i] = cl;
1916 		credits++;
1917 		i++;
1918 		MPASS(credits <= fl->ifl_size);
1919 		if (++idx == fl->ifl_size) {
1920 			fl->ifl_gen = 1;
1921 			idx = 0;
1922 		}
1923 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1924 			iru.iru_pidx = pidx;
1925 			iru.iru_count = i;
1926 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1927 			i = 0;
1928 			pidx = idx;
1929 			fl->ifl_pidx = idx;
1930 			fl->ifl_credits = credits;
1931 		}
1932 
1933 	}
1934 done:
1935 	if (i) {
1936 		iru.iru_pidx = pidx;
1937 		iru.iru_count = i;
1938 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1939 		fl->ifl_pidx = idx;
1940 		fl->ifl_credits = credits;
1941 	}
1942 	DBG_COUNTER_INC(rxd_flush);
1943 	if (fl->ifl_pidx == 0)
1944 		pidx = fl->ifl_size - 1;
1945 	else
1946 		pidx = fl->ifl_pidx - 1;
1947 
1948 	if (sd_map)
1949 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1950 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1951 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1952 	fl->ifl_fragidx = frag_idx;
1953 }
1954 
1955 static __inline void
1956 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1957 {
1958 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1959 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1960 #ifdef INVARIANTS
1961 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1962 #endif
1963 
1964 	MPASS(fl->ifl_credits <= fl->ifl_size);
1965 	MPASS(reclaimable == delta);
1966 
1967 	if (reclaimable > 0)
1968 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1969 }
1970 
1971 static void
1972 iflib_fl_bufs_free(iflib_fl_t fl)
1973 {
1974 	iflib_dma_info_t idi = fl->ifl_ifdi;
1975 	uint32_t i;
1976 
1977 	for (i = 0; i < fl->ifl_size; i++) {
1978 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1979 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1980 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1981 
1982 		if (*sd_flags & RX_SW_DESC_INUSE) {
1983 			if (fl->ifl_sds.ifsd_map != NULL) {
1984 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1985 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1986 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
1987 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1988 			}
1989 			if (*sd_m != NULL) {
1990 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1991 				uma_zfree(zone_mbuf, *sd_m);
1992 			}
1993 			if (*sd_cl != NULL)
1994 				uma_zfree(fl->ifl_zone, *sd_cl);
1995 			*sd_flags = 0;
1996 		} else {
1997 			MPASS(*sd_cl == NULL);
1998 			MPASS(*sd_m == NULL);
1999 		}
2000 #if MEMORY_LOGGING
2001 		fl->ifl_m_dequeued++;
2002 		fl->ifl_cl_dequeued++;
2003 #endif
2004 		*sd_cl = NULL;
2005 		*sd_m = NULL;
2006 	}
2007 #ifdef INVARIANTS
2008 	for (i = 0; i < fl->ifl_size; i++) {
2009 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2010 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2011 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2012 	}
2013 #endif
2014 	/*
2015 	 * Reset free list values
2016 	 */
2017 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2018 	bzero(idi->idi_vaddr, idi->idi_size);
2019 }
2020 
2021 /*********************************************************************
2022  *
2023  *  Initialize a receive ring and its buffers.
2024  *
2025  **********************************************************************/
2026 static int
2027 iflib_fl_setup(iflib_fl_t fl)
2028 {
2029 	iflib_rxq_t rxq = fl->ifl_rxq;
2030 	if_ctx_t ctx = rxq->ifr_ctx;
2031 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2032 
2033 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2034 	/*
2035 	** Free current RX buffer structs and their mbufs
2036 	*/
2037 	iflib_fl_bufs_free(fl);
2038 	/* Now replenish the mbufs */
2039 	MPASS(fl->ifl_credits == 0);
2040 	/*
2041 	 * XXX don't set the max_frame_size to larger
2042 	 * than the hardware can handle
2043 	 */
2044 	if (sctx->isc_max_frame_size <= 2048)
2045 		fl->ifl_buf_size = MCLBYTES;
2046 #ifndef CONTIGMALLOC_WORKS
2047 	else
2048 		fl->ifl_buf_size = MJUMPAGESIZE;
2049 #else
2050 	else if (sctx->isc_max_frame_size <= 4096)
2051 		fl->ifl_buf_size = MJUMPAGESIZE;
2052 	else if (sctx->isc_max_frame_size <= 9216)
2053 		fl->ifl_buf_size = MJUM9BYTES;
2054 	else
2055 		fl->ifl_buf_size = MJUM16BYTES;
2056 #endif
2057 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2058 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2059 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2060 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2061 
2062 
2063 	/* avoid pre-allocating zillions of clusters to an idle card
2064 	 * potentially speeding up attach
2065 	 */
2066 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2067 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2068 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2069 		return (ENOBUFS);
2070 	/*
2071 	 * handle failure
2072 	 */
2073 	MPASS(rxq != NULL);
2074 	MPASS(fl->ifl_ifdi != NULL);
2075 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2076 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2077 	return (0);
2078 }
2079 
2080 /*********************************************************************
2081  *
2082  *  Free receive ring data structures
2083  *
2084  **********************************************************************/
2085 static void
2086 iflib_rx_sds_free(iflib_rxq_t rxq)
2087 {
2088 	iflib_fl_t fl;
2089 	int i;
2090 
2091 	if (rxq->ifr_fl != NULL) {
2092 		for (i = 0; i < rxq->ifr_nfl; i++) {
2093 			fl = &rxq->ifr_fl[i];
2094 			if (fl->ifl_desc_tag != NULL) {
2095 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2096 				fl->ifl_desc_tag = NULL;
2097 			}
2098 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2099 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2100 			/* XXX destroy maps first */
2101 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2102 			fl->ifl_sds.ifsd_m = NULL;
2103 			fl->ifl_sds.ifsd_cl = NULL;
2104 			fl->ifl_sds.ifsd_map = NULL;
2105 		}
2106 		free(rxq->ifr_fl, M_IFLIB);
2107 		rxq->ifr_fl = NULL;
2108 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2109 	}
2110 }
2111 
2112 /*
2113  * MI independent logic
2114  *
2115  */
2116 static void
2117 iflib_timer(void *arg)
2118 {
2119 	iflib_txq_t txq = arg;
2120 	if_ctx_t ctx = txq->ift_ctx;
2121 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2122 
2123 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2124 		return;
2125 	/*
2126 	** Check on the state of the TX queue(s), this
2127 	** can be done without the lock because its RO
2128 	** and the HUNG state will be static if set.
2129 	*/
2130 	IFDI_TIMER(ctx, txq->ift_id);
2131 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2132 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2133 	     (sctx->isc_pause_frames == 0)))
2134 		goto hung;
2135 
2136 	if (ifmp_ring_is_stalled(txq->ift_br))
2137 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2138 	txq->ift_cleaned_prev = txq->ift_cleaned;
2139 	/* handle any laggards */
2140 	if (txq->ift_db_pending)
2141 		GROUPTASK_ENQUEUE(&txq->ift_task);
2142 
2143 	sctx->isc_pause_frames = 0;
2144 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2145 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2146 	return;
2147 hung:
2148 	CTX_LOCK(ctx);
2149 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2150 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2151 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2152 
2153 	IFDI_WATCHDOG_RESET(ctx);
2154 	ctx->ifc_watchdog_events++;
2155 
2156 	ctx->ifc_flags |= IFC_DO_RESET;
2157 	iflib_admin_intr_deferred(ctx);
2158 	CTX_UNLOCK(ctx);
2159 }
2160 
2161 static void
2162 iflib_init_locked(if_ctx_t ctx)
2163 {
2164 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2165 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2166 	if_t ifp = ctx->ifc_ifp;
2167 	iflib_fl_t fl;
2168 	iflib_txq_t txq;
2169 	iflib_rxq_t rxq;
2170 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2171 
2172 
2173 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2174 	IFDI_INTR_DISABLE(ctx);
2175 
2176 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2177 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2178 	/* Set hardware offload abilities */
2179 	if_clearhwassist(ifp);
2180 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2181 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2182 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2183 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2184 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2185 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2186 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2187 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2188 
2189 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2190 		CALLOUT_LOCK(txq);
2191 		callout_stop(&txq->ift_timer);
2192 		CALLOUT_UNLOCK(txq);
2193 		iflib_netmap_txq_init(ctx, txq);
2194 	}
2195 #ifdef INVARIANTS
2196 	i = if_getdrvflags(ifp);
2197 #endif
2198 	IFDI_INIT(ctx);
2199 	MPASS(if_getdrvflags(ifp) == i);
2200 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2201 		/* XXX this should really be done on a per-queue basis */
2202 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2203 			MPASS(rxq->ifr_id == i);
2204 			iflib_netmap_rxq_init(ctx, rxq);
2205 			continue;
2206 		}
2207 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2208 			if (iflib_fl_setup(fl)) {
2209 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2210 				goto done;
2211 			}
2212 		}
2213 	}
2214 	done:
2215 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2216 	IFDI_INTR_ENABLE(ctx);
2217 	txq = ctx->ifc_txqs;
2218 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2219 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2220 			txq->ift_timer.c_cpu);
2221 }
2222 
2223 static int
2224 iflib_media_change(if_t ifp)
2225 {
2226 	if_ctx_t ctx = if_getsoftc(ifp);
2227 	int err;
2228 
2229 	CTX_LOCK(ctx);
2230 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2231 		iflib_init_locked(ctx);
2232 	CTX_UNLOCK(ctx);
2233 	return (err);
2234 }
2235 
2236 static void
2237 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2238 {
2239 	if_ctx_t ctx = if_getsoftc(ifp);
2240 
2241 	CTX_LOCK(ctx);
2242 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2243 	IFDI_MEDIA_STATUS(ctx, ifmr);
2244 	CTX_UNLOCK(ctx);
2245 }
2246 
2247 static void
2248 iflib_stop(if_ctx_t ctx)
2249 {
2250 	iflib_txq_t txq = ctx->ifc_txqs;
2251 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2252 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2253 	iflib_dma_info_t di;
2254 	iflib_fl_t fl;
2255 	int i, j;
2256 
2257 	/* Tell the stack that the interface is no longer active */
2258 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2259 
2260 	IFDI_INTR_DISABLE(ctx);
2261 	DELAY(1000);
2262 	IFDI_STOP(ctx);
2263 	DELAY(1000);
2264 
2265 	iflib_debug_reset();
2266 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2267 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2268 		/* make sure all transmitters have completed before proceeding XXX */
2269 
2270 		CALLOUT_LOCK(txq);
2271 		callout_stop(&txq->ift_timer);
2272 		CALLOUT_UNLOCK(txq);
2273 
2274 		/* clean any enqueued buffers */
2275 		iflib_ifmp_purge(txq);
2276 		/* Free any existing tx buffers. */
2277 		for (j = 0; j < txq->ift_size; j++) {
2278 			iflib_txsd_free(ctx, txq, j);
2279 		}
2280 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2281 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2282 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2283 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2284 		txq->ift_pullups = 0;
2285 		ifmp_ring_reset_stats(txq->ift_br);
2286 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2287 			bzero((void *)di->idi_vaddr, di->idi_size);
2288 	}
2289 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2290 		/* make sure all transmitters have completed before proceeding XXX */
2291 
2292 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2293 			bzero((void *)di->idi_vaddr, di->idi_size);
2294 		/* also resets the free lists pidx/cidx */
2295 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2296 			iflib_fl_bufs_free(fl);
2297 	}
2298 }
2299 
2300 static inline caddr_t
2301 calc_next_rxd(iflib_fl_t fl, int cidx)
2302 {
2303 	qidx_t size;
2304 	int nrxd;
2305 	caddr_t start, end, cur, next;
2306 
2307 	nrxd = fl->ifl_size;
2308 	size = fl->ifl_rxd_size;
2309 	start = fl->ifl_ifdi->idi_vaddr;
2310 
2311 	if (__predict_false(size == 0))
2312 		return (start);
2313 	cur = start + size*cidx;
2314 	end = start + size*nrxd;
2315 	next = CACHE_PTR_NEXT(cur);
2316 	return (next < end ? next : start);
2317 }
2318 
2319 static inline void
2320 prefetch_pkts(iflib_fl_t fl, int cidx)
2321 {
2322 	int nextptr;
2323 	int nrxd = fl->ifl_size;
2324 	caddr_t next_rxd;
2325 
2326 
2327 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2328 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2329 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2330 	next_rxd = calc_next_rxd(fl, cidx);
2331 	prefetch(next_rxd);
2332 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2333 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2334 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2335 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2336 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2337 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2338 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2339 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2340 }
2341 
2342 static void
2343 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2344 {
2345 	int flid, cidx;
2346 	bus_dmamap_t map;
2347 	iflib_fl_t fl;
2348 	iflib_dma_info_t di;
2349 	int next;
2350 
2351 	map = NULL;
2352 	flid = irf->irf_flid;
2353 	cidx = irf->irf_idx;
2354 	fl = &rxq->ifr_fl[flid];
2355 	sd->ifsd_fl = fl;
2356 	sd->ifsd_cidx = cidx;
2357 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2358 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2359 	fl->ifl_credits--;
2360 #if MEMORY_LOGGING
2361 	fl->ifl_m_dequeued++;
2362 #endif
2363 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2364 		prefetch_pkts(fl, cidx);
2365 	if (fl->ifl_sds.ifsd_map != NULL) {
2366 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2367 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2368 		map = fl->ifl_sds.ifsd_map[cidx];
2369 		di = fl->ifl_ifdi;
2370 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2371 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2372 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2373 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2374 
2375 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2376 		MPASS(fl->ifl_cidx == cidx);
2377 		if (unload)
2378 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2379 	}
2380 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2381 	if (__predict_false(fl->ifl_cidx == 0))
2382 		fl->ifl_gen = 0;
2383 	if (map != NULL)
2384 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2385 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2386         bit_clear(fl->ifl_rx_bitmap, cidx);
2387 }
2388 
2389 static struct mbuf *
2390 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2391 {
2392 	int i, padlen , flags;
2393 	struct mbuf *m, *mh, *mt;
2394 	caddr_t cl;
2395 
2396 	i = 0;
2397 	mh = NULL;
2398 	do {
2399 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2400 
2401 		MPASS(*sd->ifsd_cl != NULL);
2402 		MPASS(*sd->ifsd_m != NULL);
2403 
2404 		/* Don't include zero-length frags */
2405 		if (ri->iri_frags[i].irf_len == 0) {
2406 			/* XXX we can save the cluster here, but not the mbuf */
2407 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2408 			m_free(*sd->ifsd_m);
2409 			*sd->ifsd_m = NULL;
2410 			continue;
2411 		}
2412 		m = *sd->ifsd_m;
2413 		*sd->ifsd_m = NULL;
2414 		if (mh == NULL) {
2415 			flags = M_PKTHDR|M_EXT;
2416 			mh = mt = m;
2417 			padlen = ri->iri_pad;
2418 		} else {
2419 			flags = M_EXT;
2420 			mt->m_next = m;
2421 			mt = m;
2422 			/* assuming padding is only on the first fragment */
2423 			padlen = 0;
2424 		}
2425 		cl = *sd->ifsd_cl;
2426 		*sd->ifsd_cl = NULL;
2427 
2428 		/* Can these two be made one ? */
2429 		m_init(m, M_NOWAIT, MT_DATA, flags);
2430 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2431 		/*
2432 		 * These must follow m_init and m_cljset
2433 		 */
2434 		m->m_data += padlen;
2435 		ri->iri_len -= padlen;
2436 		m->m_len = ri->iri_frags[i].irf_len;
2437 	} while (++i < ri->iri_nfrags);
2438 
2439 	return (mh);
2440 }
2441 
2442 /*
2443  * Process one software descriptor
2444  */
2445 static struct mbuf *
2446 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2447 {
2448 	struct if_rxsd sd;
2449 	struct mbuf *m;
2450 
2451 	/* should I merge this back in now that the two paths are basically duplicated? */
2452 	if (ri->iri_nfrags == 1 &&
2453 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2454 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2455 		m = *sd.ifsd_m;
2456 		*sd.ifsd_m = NULL;
2457 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2458 #ifndef __NO_STRICT_ALIGNMENT
2459 		if (!IP_ALIGNED(m))
2460 			m->m_data += 2;
2461 #endif
2462 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2463 		m->m_len = ri->iri_frags[0].irf_len;
2464        } else {
2465 		m = assemble_segments(rxq, ri, &sd);
2466 	}
2467 	m->m_pkthdr.len = ri->iri_len;
2468 	m->m_pkthdr.rcvif = ri->iri_ifp;
2469 	m->m_flags |= ri->iri_flags;
2470 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2471 	m->m_pkthdr.flowid = ri->iri_flowid;
2472 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2473 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2474 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2475 	return (m);
2476 }
2477 
2478 #if defined(INET6) || defined(INET)
2479 static void
2480 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2481 {
2482 	CURVNET_SET(lc->ifp->if_vnet);
2483 #if defined(INET6)
2484 	*v6 = VNET(ip6_forwarding);
2485 #endif
2486 #if defined(INET)
2487 	*v4 = VNET(ipforwarding);
2488 #endif
2489 	CURVNET_RESTORE();
2490 }
2491 
2492 /*
2493  * Returns true if it's possible this packet could be LROed.
2494  * if it returns false, it is guaranteed that tcp_lro_rx()
2495  * would not return zero.
2496  */
2497 static bool
2498 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2499 {
2500 	struct ether_header *eh;
2501 	uint16_t eh_type;
2502 
2503 	eh = mtod(m, struct ether_header *);
2504 	eh_type = ntohs(eh->ether_type);
2505 	switch (eh_type) {
2506 #if defined(INET6)
2507 		case ETHERTYPE_IPV6:
2508 			return !v6_forwarding;
2509 #endif
2510 #if defined (INET)
2511 		case ETHERTYPE_IP:
2512 			return !v4_forwarding;
2513 #endif
2514 	}
2515 
2516 	return false;
2517 }
2518 #else
2519 static void
2520 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2521 {
2522 }
2523 #endif
2524 
2525 static bool
2526 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2527 {
2528 	if_ctx_t ctx = rxq->ifr_ctx;
2529 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2530 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2531 	int avail, i;
2532 	qidx_t *cidxp;
2533 	struct if_rxd_info ri;
2534 	int err, budget_left, rx_bytes, rx_pkts;
2535 	iflib_fl_t fl;
2536 	struct ifnet *ifp;
2537 	int lro_enabled;
2538 	bool lro_possible = false;
2539 	bool v4_forwarding, v6_forwarding;
2540 
2541 	/*
2542 	 * XXX early demux data packets so that if_input processing only handles
2543 	 * acks in interrupt context
2544 	 */
2545 	struct mbuf *m, *mh, *mt, *mf;
2546 
2547 	ifp = ctx->ifc_ifp;
2548 	mh = mt = NULL;
2549 	MPASS(budget > 0);
2550 	rx_pkts	= rx_bytes = 0;
2551 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2552 		cidxp = &rxq->ifr_cq_cidx;
2553 	else
2554 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2555 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2556 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2557 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2558 		DBG_COUNTER_INC(rx_unavail);
2559 		return (false);
2560 	}
2561 
2562 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2563 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2564 			DBG_COUNTER_INC(rx_ctx_inactive);
2565 			break;
2566 		}
2567 		/*
2568 		 * Reset client set fields to their default values
2569 		 */
2570 		rxd_info_zero(&ri);
2571 		ri.iri_qsidx = rxq->ifr_id;
2572 		ri.iri_cidx = *cidxp;
2573 		ri.iri_ifp = ifp;
2574 		ri.iri_frags = rxq->ifr_frags;
2575 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2576 
2577 		if (err)
2578 			goto err;
2579 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2580 			*cidxp = ri.iri_cidx;
2581 			/* Update our consumer index */
2582 			/* XXX NB: shurd - check if this is still safe */
2583 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2584 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2585 				rxq->ifr_cq_gen = 0;
2586 			}
2587 			/* was this only a completion queue message? */
2588 			if (__predict_false(ri.iri_nfrags == 0))
2589 				continue;
2590 		}
2591 		MPASS(ri.iri_nfrags != 0);
2592 		MPASS(ri.iri_len != 0);
2593 
2594 		/* will advance the cidx on the corresponding free lists */
2595 		m = iflib_rxd_pkt_get(rxq, &ri);
2596 		if (avail == 0 && budget_left)
2597 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2598 
2599 		if (__predict_false(m == NULL)) {
2600 			DBG_COUNTER_INC(rx_mbuf_null);
2601 			continue;
2602 		}
2603 		/* imm_pkt: -- cxgb */
2604 		if (mh == NULL)
2605 			mh = mt = m;
2606 		else {
2607 			mt->m_nextpkt = m;
2608 			mt = m;
2609 		}
2610 	}
2611 	/* make sure that we can refill faster than drain */
2612 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2613 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2614 
2615 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2616 	if (lro_enabled)
2617 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2618 	mt = mf = NULL;
2619 	while (mh != NULL) {
2620 		m = mh;
2621 		mh = mh->m_nextpkt;
2622 		m->m_nextpkt = NULL;
2623 #ifndef __NO_STRICT_ALIGNMENT
2624 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2625 			continue;
2626 #endif
2627 		rx_bytes += m->m_pkthdr.len;
2628 		rx_pkts++;
2629 #if defined(INET6) || defined(INET)
2630 		if (lro_enabled) {
2631 			if (!lro_possible) {
2632 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2633 				if (lro_possible && mf != NULL) {
2634 					ifp->if_input(ifp, mf);
2635 					DBG_COUNTER_INC(rx_if_input);
2636 					mt = mf = NULL;
2637 				}
2638 			}
2639 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2640 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2641 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2642 					continue;
2643 			}
2644 		}
2645 #endif
2646 		if (lro_possible) {
2647 			ifp->if_input(ifp, m);
2648 			DBG_COUNTER_INC(rx_if_input);
2649 			continue;
2650 		}
2651 
2652 		if (mf == NULL)
2653 			mf = m;
2654 		if (mt != NULL)
2655 			mt->m_nextpkt = m;
2656 		mt = m;
2657 	}
2658 	if (mf != NULL) {
2659 		ifp->if_input(ifp, mf);
2660 		DBG_COUNTER_INC(rx_if_input);
2661 	}
2662 
2663 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2664 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2665 
2666 	/*
2667 	 * Flush any outstanding LRO work
2668 	 */
2669 #if defined(INET6) || defined(INET)
2670 	tcp_lro_flush_all(&rxq->ifr_lc);
2671 #endif
2672 	if (avail)
2673 		return true;
2674 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2675 err:
2676 	CTX_LOCK(ctx);
2677 	ctx->ifc_flags |= IFC_DO_RESET;
2678 	iflib_admin_intr_deferred(ctx);
2679 	CTX_UNLOCK(ctx);
2680 	return (false);
2681 }
2682 
2683 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2684 static inline qidx_t
2685 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2686 {
2687 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2688 	qidx_t minthresh = txq->ift_size / 8;
2689 	if (in_use > 4*minthresh)
2690 		return (notify_count);
2691 	if (in_use > 2*minthresh)
2692 		return (notify_count >> 1);
2693 	if (in_use > minthresh)
2694 		return (notify_count >> 3);
2695 	return (0);
2696 }
2697 
2698 static inline qidx_t
2699 txq_max_rs_deferred(iflib_txq_t txq)
2700 {
2701 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2702 	qidx_t minthresh = txq->ift_size / 8;
2703 	if (txq->ift_in_use > 4*minthresh)
2704 		return (notify_count);
2705 	if (txq->ift_in_use > 2*minthresh)
2706 		return (notify_count >> 1);
2707 	if (txq->ift_in_use > minthresh)
2708 		return (notify_count >> 2);
2709 	return (2);
2710 }
2711 
2712 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2713 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2714 
2715 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2716 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2717 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2718 
2719 /* forward compatibility for cxgb */
2720 #define FIRST_QSET(ctx) 0
2721 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2722 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2723 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2724 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2725 
2726 /* XXX we should be setting this to something other than zero */
2727 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2728 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2729 
2730 static inline bool
2731 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2732 {
2733 	qidx_t dbval, max;
2734 	bool rang;
2735 
2736 	rang = false;
2737 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2738 	if (ring || txq->ift_db_pending >= max) {
2739 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2740 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2741 		txq->ift_db_pending = txq->ift_npending = 0;
2742 		rang = true;
2743 	}
2744 	return (rang);
2745 }
2746 
2747 #ifdef PKT_DEBUG
2748 static void
2749 print_pkt(if_pkt_info_t pi)
2750 {
2751 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2752 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2753 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2754 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2755 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2756 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2757 }
2758 #endif
2759 
2760 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2761 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2762 
2763 static int
2764 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2765 {
2766 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2767 	struct ether_vlan_header *eh;
2768 	struct mbuf *m, *n;
2769 
2770 	n = m = *mp;
2771 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2772 	    M_WRITABLE(m) == 0) {
2773 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2774 			return (ENOMEM);
2775 		} else {
2776 			m_freem(*mp);
2777 			n = *mp = m;
2778 		}
2779 	}
2780 
2781 	/*
2782 	 * Determine where frame payload starts.
2783 	 * Jump over vlan headers if already present,
2784 	 * helpful for QinQ too.
2785 	 */
2786 	if (__predict_false(m->m_len < sizeof(*eh))) {
2787 		txq->ift_pullups++;
2788 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2789 			return (ENOMEM);
2790 	}
2791 	eh = mtod(m, struct ether_vlan_header *);
2792 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2793 		pi->ipi_etype = ntohs(eh->evl_proto);
2794 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2795 	} else {
2796 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2797 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2798 	}
2799 
2800 	switch (pi->ipi_etype) {
2801 #ifdef INET
2802 	case ETHERTYPE_IP:
2803 	{
2804 		struct ip *ip = NULL;
2805 		struct tcphdr *th = NULL;
2806 		int minthlen;
2807 
2808 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2809 		if (__predict_false(m->m_len < minthlen)) {
2810 			/*
2811 			 * if this code bloat is causing too much of a hit
2812 			 * move it to a separate function and mark it noinline
2813 			 */
2814 			if (m->m_len == pi->ipi_ehdrlen) {
2815 				n = m->m_next;
2816 				MPASS(n);
2817 				if (n->m_len >= sizeof(*ip))  {
2818 					ip = (struct ip *)n->m_data;
2819 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2820 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2821 				} else {
2822 					txq->ift_pullups++;
2823 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2824 						return (ENOMEM);
2825 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2826 				}
2827 			} else {
2828 				txq->ift_pullups++;
2829 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2830 					return (ENOMEM);
2831 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2832 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2833 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2834 			}
2835 		} else {
2836 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2837 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2838 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2839 		}
2840 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2841 		pi->ipi_ipproto = ip->ip_p;
2842 		pi->ipi_flags |= IPI_TX_IPV4;
2843 
2844 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2845                        ip->ip_sum = 0;
2846 
2847 		if (IS_TSO4(pi)) {
2848 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2849 				if (__predict_false(th == NULL)) {
2850 					txq->ift_pullups++;
2851 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2852 						return (ENOMEM);
2853 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2854 				}
2855 				pi->ipi_tcp_hflags = th->th_flags;
2856 				pi->ipi_tcp_hlen = th->th_off << 2;
2857 				pi->ipi_tcp_seq = th->th_seq;
2858 			}
2859 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2860 				return (ENXIO);
2861 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2862 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2863 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2864 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2865 				ip->ip_sum = 0;
2866 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2867 			}
2868 		}
2869 		break;
2870 	}
2871 #endif
2872 #ifdef INET6
2873 	case ETHERTYPE_IPV6:
2874 	{
2875 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2876 		struct tcphdr *th;
2877 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2878 
2879 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2880 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2881 				return (ENOMEM);
2882 		}
2883 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2884 
2885 		/* XXX-BZ this will go badly in case of ext hdrs. */
2886 		pi->ipi_ipproto = ip6->ip6_nxt;
2887 		pi->ipi_flags |= IPI_TX_IPV6;
2888 
2889 		if (IS_TSO6(pi)) {
2890 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2891 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2892 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2893 						return (ENOMEM);
2894 				}
2895 				pi->ipi_tcp_hflags = th->th_flags;
2896 				pi->ipi_tcp_hlen = th->th_off << 2;
2897 			}
2898 
2899 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2900 				return (ENXIO);
2901 			/*
2902 			 * The corresponding flag is set by the stack in the IPv4
2903 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2904 			 * So, set it here because the rest of the flow requires it.
2905 			 */
2906 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2907 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2908 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2909 		}
2910 		break;
2911 	}
2912 #endif
2913 	default:
2914 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2915 		pi->ipi_ip_hlen = 0;
2916 		break;
2917 	}
2918 	*mp = m;
2919 
2920 	return (0);
2921 }
2922 
2923 static  __noinline  struct mbuf *
2924 collapse_pkthdr(struct mbuf *m0)
2925 {
2926 	struct mbuf *m, *m_next, *tmp;
2927 
2928 	m = m0;
2929 	m_next = m->m_next;
2930 	while (m_next != NULL && m_next->m_len == 0) {
2931 		m = m_next;
2932 		m->m_next = NULL;
2933 		m_free(m);
2934 		m_next = m_next->m_next;
2935 	}
2936 	m = m0;
2937 	m->m_next = m_next;
2938 	if ((m_next->m_flags & M_EXT) == 0) {
2939 		m = m_defrag(m, M_NOWAIT);
2940 	} else {
2941 		tmp = m_next->m_next;
2942 		memcpy(m_next, m, MPKTHSIZE);
2943 		m = m_next;
2944 		m->m_next = tmp;
2945 	}
2946 	return (m);
2947 }
2948 
2949 /*
2950  * If dodgy hardware rejects the scatter gather chain we've handed it
2951  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2952  * m_defrag'd mbufs
2953  */
2954 static __noinline struct mbuf *
2955 iflib_remove_mbuf(iflib_txq_t txq)
2956 {
2957 	int ntxd, i, pidx;
2958 	struct mbuf *m, *mh, **ifsd_m;
2959 
2960 	pidx = txq->ift_pidx;
2961 	ifsd_m = txq->ift_sds.ifsd_m;
2962 	ntxd = txq->ift_size;
2963 	mh = m = ifsd_m[pidx];
2964 	ifsd_m[pidx] = NULL;
2965 #if MEMORY_LOGGING
2966 	txq->ift_dequeued++;
2967 #endif
2968 	i = 1;
2969 
2970 	while (m) {
2971 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2972 #if MEMORY_LOGGING
2973 		txq->ift_dequeued++;
2974 #endif
2975 		m = m->m_next;
2976 		i++;
2977 	}
2978 	return (mh);
2979 }
2980 
2981 static int
2982 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2983 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2984 			  int max_segs, int flags)
2985 {
2986 	if_ctx_t ctx;
2987 	if_shared_ctx_t		sctx;
2988 	if_softc_ctx_t		scctx;
2989 	int i, next, pidx, err, ntxd, count;
2990 	struct mbuf *m, *tmp, **ifsd_m;
2991 
2992 	m = *m0;
2993 
2994 	/*
2995 	 * Please don't ever do this
2996 	 */
2997 	if (__predict_false(m->m_len == 0))
2998 		*m0 = m = collapse_pkthdr(m);
2999 
3000 	ctx = txq->ift_ctx;
3001 	sctx = ctx->ifc_sctx;
3002 	scctx = &ctx->ifc_softc_ctx;
3003 	ifsd_m = txq->ift_sds.ifsd_m;
3004 	ntxd = txq->ift_size;
3005 	pidx = txq->ift_pidx;
3006 	if (map != NULL) {
3007 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3008 
3009 		err = bus_dmamap_load_mbuf_sg(tag, map,
3010 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3011 		if (err)
3012 			return (err);
3013 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3014 		count = 0;
3015 		m = *m0;
3016 		do {
3017 			if (__predict_false(m->m_len <= 0)) {
3018 				tmp = m;
3019 				m = m->m_next;
3020 				tmp->m_next = NULL;
3021 				m_free(tmp);
3022 				continue;
3023 			}
3024 			m = m->m_next;
3025 			count++;
3026 		} while (m != NULL);
3027 		if (count > *nsegs) {
3028 			ifsd_m[pidx] = *m0;
3029 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3030 			return (0);
3031 		}
3032 		m = *m0;
3033 		count = 0;
3034 		do {
3035 			next = (pidx + count) & (ntxd-1);
3036 			MPASS(ifsd_m[next] == NULL);
3037 			ifsd_m[next] = m;
3038 			count++;
3039 			tmp = m;
3040 			m = m->m_next;
3041 		} while (m != NULL);
3042 	} else {
3043 		int buflen, sgsize, maxsegsz, max_sgsize;
3044 		vm_offset_t vaddr;
3045 		vm_paddr_t curaddr;
3046 
3047 		count = i = 0;
3048 		m = *m0;
3049 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3050 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3051 		else
3052 			maxsegsz = sctx->isc_tx_maxsegsize;
3053 
3054 		do {
3055 			if (__predict_false(m->m_len <= 0)) {
3056 				tmp = m;
3057 				m = m->m_next;
3058 				tmp->m_next = NULL;
3059 				m_free(tmp);
3060 				continue;
3061 			}
3062 			buflen = m->m_len;
3063 			vaddr = (vm_offset_t)m->m_data;
3064 			/*
3065 			 * see if we can't be smarter about physically
3066 			 * contiguous mappings
3067 			 */
3068 			next = (pidx + count) & (ntxd-1);
3069 			MPASS(ifsd_m[next] == NULL);
3070 #if MEMORY_LOGGING
3071 			txq->ift_enqueued++;
3072 #endif
3073 			ifsd_m[next] = m;
3074 			while (buflen > 0) {
3075 				if (i >= max_segs)
3076 					goto err;
3077 				max_sgsize = MIN(buflen, maxsegsz);
3078 				curaddr = pmap_kextract(vaddr);
3079 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3080 				sgsize = MIN(sgsize, max_sgsize);
3081 				segs[i].ds_addr = curaddr;
3082 				segs[i].ds_len = sgsize;
3083 				vaddr += sgsize;
3084 				buflen -= sgsize;
3085 				i++;
3086 			}
3087 			count++;
3088 			tmp = m;
3089 			m = m->m_next;
3090 		} while (m != NULL);
3091 		*nsegs = i;
3092 	}
3093 	return (0);
3094 err:
3095 	*m0 = iflib_remove_mbuf(txq);
3096 	return (EFBIG);
3097 }
3098 
3099 static inline caddr_t
3100 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3101 {
3102 	qidx_t size;
3103 	int ntxd;
3104 	caddr_t start, end, cur, next;
3105 
3106 	ntxd = txq->ift_size;
3107 	size = txq->ift_txd_size[qid];
3108 	start = txq->ift_ifdi[qid].idi_vaddr;
3109 
3110 	if (__predict_false(size == 0))
3111 		return (start);
3112 	cur = start + size*cidx;
3113 	end = start + size*ntxd;
3114 	next = CACHE_PTR_NEXT(cur);
3115 	return (next < end ? next : start);
3116 }
3117 
3118 /*
3119  * Pad an mbuf to ensure a minimum ethernet frame size.
3120  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3121  */
3122 static __noinline int
3123 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3124 {
3125 	/*
3126 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3127 	 * and ARP message is the smallest common payload I can think of
3128 	 */
3129 	static char pad[18];	/* just zeros */
3130 	int n;
3131 	struct mbuf *new_head;
3132 
3133 	if (!M_WRITABLE(*m_head)) {
3134 		new_head = m_dup(*m_head, M_NOWAIT);
3135 		if (new_head == NULL) {
3136 			m_freem(*m_head);
3137 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3138 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3139 			return ENOMEM;
3140 		}
3141 		m_freem(*m_head);
3142 		*m_head = new_head;
3143 	}
3144 
3145 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3146 	     n > 0; n -= sizeof(pad))
3147 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3148 			break;
3149 
3150 	if (n > 0) {
3151 		m_freem(*m_head);
3152 		device_printf(dev, "cannot pad short frame\n");
3153 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3154 		return (ENOBUFS);
3155 	}
3156 
3157 	return 0;
3158 }
3159 
3160 static int
3161 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3162 {
3163 	if_ctx_t		ctx;
3164 	if_shared_ctx_t		sctx;
3165 	if_softc_ctx_t		scctx;
3166 	bus_dma_segment_t	*segs;
3167 	struct mbuf		*m_head;
3168 	void			*next_txd;
3169 	bus_dmamap_t		map;
3170 	struct if_pkt_info	pi;
3171 	int remap = 0;
3172 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3173 	bus_dma_tag_t desc_tag;
3174 
3175 	segs = txq->ift_segs;
3176 	ctx = txq->ift_ctx;
3177 	sctx = ctx->ifc_sctx;
3178 	scctx = &ctx->ifc_softc_ctx;
3179 	segs = txq->ift_segs;
3180 	ntxd = txq->ift_size;
3181 	m_head = *m_headp;
3182 	map = NULL;
3183 
3184 	/*
3185 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3186 	 */
3187 	cidx = txq->ift_cidx;
3188 	pidx = txq->ift_pidx;
3189 	if (ctx->ifc_flags & IFC_PREFETCH) {
3190 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3191 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3192 			next_txd = calc_next_txd(txq, cidx, 0);
3193 			prefetch(next_txd);
3194 		}
3195 
3196 		/* prefetch the next cache line of mbuf pointers and flags */
3197 		prefetch(&txq->ift_sds.ifsd_m[next]);
3198 		if (txq->ift_sds.ifsd_map != NULL) {
3199 			prefetch(&txq->ift_sds.ifsd_map[next]);
3200 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3201 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3202 		}
3203 	} else if (txq->ift_sds.ifsd_map != NULL)
3204 		map = txq->ift_sds.ifsd_map[pidx];
3205 
3206 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3207 		desc_tag = txq->ift_tso_desc_tag;
3208 		max_segs = scctx->isc_tx_tso_segments_max;
3209 	} else {
3210 		desc_tag = txq->ift_desc_tag;
3211 		max_segs = scctx->isc_tx_nsegments;
3212 	}
3213 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3214 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3215 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3216 		if (err)
3217 			return err;
3218 	}
3219 	m_head = *m_headp;
3220 
3221 	pkt_info_zero(&pi);
3222 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3223 	pi.ipi_pidx = pidx;
3224 	pi.ipi_qsidx = txq->ift_id;
3225 	pi.ipi_len = m_head->m_pkthdr.len;
3226 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3227 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3228 
3229 	/* deliberate bitwise OR to make one condition */
3230 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3231 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3232 			return (err);
3233 		m_head = *m_headp;
3234 	}
3235 
3236 retry:
3237 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3238 defrag:
3239 	if (__predict_false(err)) {
3240 		switch (err) {
3241 		case EFBIG:
3242 			/* try collapse once and defrag once */
3243 			if (remap == 0)
3244 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3245 			if (remap == 1)
3246 				m_head = m_defrag(*m_headp, M_NOWAIT);
3247 			remap++;
3248 			if (__predict_false(m_head == NULL))
3249 				goto defrag_failed;
3250 			txq->ift_mbuf_defrag++;
3251 			*m_headp = m_head;
3252 			goto retry;
3253 			break;
3254 		case ENOMEM:
3255 			txq->ift_no_tx_dma_setup++;
3256 			break;
3257 		default:
3258 			txq->ift_no_tx_dma_setup++;
3259 			m_freem(*m_headp);
3260 			DBG_COUNTER_INC(tx_frees);
3261 			*m_headp = NULL;
3262 			break;
3263 		}
3264 		txq->ift_map_failed++;
3265 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3266 		return (err);
3267 	}
3268 
3269 	/*
3270 	 * XXX assumes a 1 to 1 relationship between segments and
3271 	 *        descriptors - this does not hold true on all drivers, e.g.
3272 	 *        cxgb
3273 	 */
3274 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3275 		txq->ift_no_desc_avail++;
3276 		if (map != NULL)
3277 			bus_dmamap_unload(desc_tag, map);
3278 		DBG_COUNTER_INC(encap_txq_avail_fail);
3279 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3280 			GROUPTASK_ENQUEUE(&txq->ift_task);
3281 		return (ENOBUFS);
3282 	}
3283 	/*
3284 	 * On Intel cards we can greatly reduce the number of TX interrupts
3285 	 * we see by only setting report status on every Nth descriptor.
3286 	 * However, this also means that the driver will need to keep track
3287 	 * of the descriptors that RS was set on to check them for the DD bit.
3288 	 */
3289 	txq->ift_rs_pending += nsegs + 1;
3290 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3291 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3292 		pi.ipi_flags |= IPI_TX_INTR;
3293 		txq->ift_rs_pending = 0;
3294 	}
3295 
3296 	pi.ipi_segs = segs;
3297 	pi.ipi_nsegs = nsegs;
3298 
3299 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3300 #ifdef PKT_DEBUG
3301 	print_pkt(&pi);
3302 #endif
3303 	if (map != NULL)
3304 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3305 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3306 		if (map != NULL)
3307 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3308 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3309 		DBG_COUNTER_INC(tx_encap);
3310 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3311 
3312 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3313 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3314 			ndesc += txq->ift_size;
3315 			txq->ift_gen = 1;
3316 		}
3317 		/*
3318 		 * drivers can need as many as
3319 		 * two sentinels
3320 		 */
3321 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3322 		MPASS(pi.ipi_new_pidx != pidx);
3323 		MPASS(ndesc > 0);
3324 		txq->ift_in_use += ndesc;
3325 
3326 		/*
3327 		 * We update the last software descriptor again here because there may
3328 		 * be a sentinel and/or there may be more mbufs than segments
3329 		 */
3330 		txq->ift_pidx = pi.ipi_new_pidx;
3331 		txq->ift_npending += pi.ipi_ndescs;
3332 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3333 		*m_headp = m_head = iflib_remove_mbuf(txq);
3334 		remap = 1;
3335 		txq->ift_txd_encap_efbig++;
3336 		goto defrag;
3337 	} else
3338 		DBG_COUNTER_INC(encap_txd_encap_fail);
3339 	return (err);
3340 
3341 defrag_failed:
3342 	txq->ift_mbuf_defrag_failed++;
3343 	txq->ift_map_failed++;
3344 	m_freem(*m_headp);
3345 	DBG_COUNTER_INC(tx_frees);
3346 	*m_headp = NULL;
3347 	return (ENOMEM);
3348 }
3349 
3350 static void
3351 iflib_tx_desc_free(iflib_txq_t txq, int n)
3352 {
3353 	int hasmap;
3354 	uint32_t qsize, cidx, mask, gen;
3355 	struct mbuf *m, **ifsd_m;
3356 	uint8_t *ifsd_flags;
3357 	bus_dmamap_t *ifsd_map;
3358 	bool do_prefetch;
3359 
3360 	cidx = txq->ift_cidx;
3361 	gen = txq->ift_gen;
3362 	qsize = txq->ift_size;
3363 	mask = qsize-1;
3364 	hasmap = txq->ift_sds.ifsd_map != NULL;
3365 	ifsd_flags = txq->ift_sds.ifsd_flags;
3366 	ifsd_m = txq->ift_sds.ifsd_m;
3367 	ifsd_map = txq->ift_sds.ifsd_map;
3368 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3369 
3370 	while (n--) {
3371 		if (do_prefetch) {
3372 			prefetch(ifsd_m[(cidx + 3) & mask]);
3373 			prefetch(ifsd_m[(cidx + 4) & mask]);
3374 		}
3375 		if (ifsd_m[cidx] != NULL) {
3376 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3377 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3378 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3379 				/*
3380 				 * does it matter if it's not the TSO tag? If so we'll
3381 				 * have to add the type to flags
3382 				 */
3383 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3384 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3385 			}
3386 			if ((m = ifsd_m[cidx]) != NULL) {
3387 				/* XXX we don't support any drivers that batch packets yet */
3388 				MPASS(m->m_nextpkt == NULL);
3389 				/* if the number of clusters exceeds the number of segments
3390 				 * there won't be space on the ring to save a pointer to each
3391 				 * cluster so we simply free the list here
3392 				 */
3393 				if (m->m_flags & M_TOOBIG) {
3394 					m_freem(m);
3395 				} else {
3396 					m_free(m);
3397 				}
3398 				ifsd_m[cidx] = NULL;
3399 #if MEMORY_LOGGING
3400 				txq->ift_dequeued++;
3401 #endif
3402 				DBG_COUNTER_INC(tx_frees);
3403 			}
3404 		}
3405 		if (__predict_false(++cidx == qsize)) {
3406 			cidx = 0;
3407 			gen = 0;
3408 		}
3409 	}
3410 	txq->ift_cidx = cidx;
3411 	txq->ift_gen = gen;
3412 }
3413 
3414 static __inline int
3415 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3416 {
3417 	int reclaim;
3418 	if_ctx_t ctx = txq->ift_ctx;
3419 
3420 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3421 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3422 
3423 	/*
3424 	 * Need a rate-limiting check so that this isn't called every time
3425 	 */
3426 	iflib_tx_credits_update(ctx, txq);
3427 	reclaim = DESC_RECLAIMABLE(txq);
3428 
3429 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3430 #ifdef INVARIANTS
3431 		if (iflib_verbose_debug) {
3432 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3433 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3434 			       reclaim, thresh);
3435 
3436 		}
3437 #endif
3438 		return (0);
3439 	}
3440 	iflib_tx_desc_free(txq, reclaim);
3441 	txq->ift_cleaned += reclaim;
3442 	txq->ift_in_use -= reclaim;
3443 
3444 	return (reclaim);
3445 }
3446 
3447 static struct mbuf **
3448 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3449 {
3450 	int next, size;
3451 	struct mbuf **items;
3452 
3453 	size = r->size;
3454 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3455 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3456 
3457 	prefetch(items[(cidx + offset) & (size-1)]);
3458 	if (remaining > 1) {
3459 		prefetch2cachelines(&items[next]);
3460 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3461 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3462 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3463 	}
3464 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3465 }
3466 
3467 static void
3468 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3469 {
3470 
3471 	ifmp_ring_check_drainage(txq->ift_br, budget);
3472 }
3473 
3474 static uint32_t
3475 iflib_txq_can_drain(struct ifmp_ring *r)
3476 {
3477 	iflib_txq_t txq = r->cookie;
3478 	if_ctx_t ctx = txq->ift_ctx;
3479 
3480 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3481 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3482 }
3483 
3484 static uint32_t
3485 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3486 {
3487 	iflib_txq_t txq = r->cookie;
3488 	if_ctx_t ctx = txq->ift_ctx;
3489 	struct ifnet *ifp = ctx->ifc_ifp;
3490 	struct mbuf **mp, *m;
3491 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3492 	int reclaimed, err, in_use_prev, desc_used;
3493 	bool do_prefetch, ring, rang;
3494 
3495 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3496 			    !LINK_ACTIVE(ctx))) {
3497 		DBG_COUNTER_INC(txq_drain_notready);
3498 		return (0);
3499 	}
3500 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3501 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3502 	avail = IDXDIFF(pidx, cidx, r->size);
3503 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3504 		DBG_COUNTER_INC(txq_drain_flushing);
3505 		for (i = 0; i < avail; i++) {
3506 			m_free(r->items[(cidx + i) & (r->size-1)]);
3507 			r->items[(cidx + i) & (r->size-1)] = NULL;
3508 		}
3509 		return (avail);
3510 	}
3511 
3512 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3513 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3514 		CALLOUT_LOCK(txq);
3515 		callout_stop(&txq->ift_timer);
3516 		CALLOUT_UNLOCK(txq);
3517 		DBG_COUNTER_INC(txq_drain_oactive);
3518 		return (0);
3519 	}
3520 	if (reclaimed)
3521 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3522 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3523 	count = MIN(avail, TX_BATCH_SIZE);
3524 #ifdef INVARIANTS
3525 	if (iflib_verbose_debug)
3526 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3527 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3528 #endif
3529 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3530 	avail = TXQ_AVAIL(txq);
3531 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3532 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3533 
3534 		mp = _ring_peek_one(r, cidx, i, rem);
3535 		MPASS(mp != NULL && *mp != NULL);
3536 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3537 			consumed++;
3538 			reclaimed++;
3539 			continue;
3540 		}
3541 		in_use_prev = txq->ift_in_use;
3542 		pidx_prev = txq->ift_pidx;
3543 		err = iflib_encap(txq, mp);
3544 		if (__predict_false(err)) {
3545 			DBG_COUNTER_INC(txq_drain_encapfail);
3546 			/* no room - bail out */
3547 			if (err == ENOBUFS)
3548 				break;
3549 			consumed++;
3550 			DBG_COUNTER_INC(txq_drain_encapfail);
3551 			/* we can't send this packet - skip it */
3552 			continue;
3553 		}
3554 		consumed++;
3555 		pkt_sent++;
3556 		m = *mp;
3557 		DBG_COUNTER_INC(tx_sent);
3558 		bytes_sent += m->m_pkthdr.len;
3559 		mcast_sent += !!(m->m_flags & M_MCAST);
3560 		avail = TXQ_AVAIL(txq);
3561 
3562 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3563 		desc_used += (txq->ift_in_use - in_use_prev);
3564 		ETHER_BPF_MTAP(ifp, m);
3565 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3566 			break;
3567 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3568 	}
3569 
3570 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3571 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3572 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3573 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3574 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3575 	if (mcast_sent)
3576 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3577 #ifdef INVARIANTS
3578 	if (iflib_verbose_debug)
3579 		printf("consumed=%d\n", consumed);
3580 #endif
3581 	return (consumed);
3582 }
3583 
3584 static uint32_t
3585 iflib_txq_drain_always(struct ifmp_ring *r)
3586 {
3587 	return (1);
3588 }
3589 
3590 static uint32_t
3591 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3592 {
3593 	int i, avail;
3594 	struct mbuf **mp;
3595 	iflib_txq_t txq;
3596 
3597 	txq = r->cookie;
3598 
3599 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3600 	CALLOUT_LOCK(txq);
3601 	callout_stop(&txq->ift_timer);
3602 	CALLOUT_UNLOCK(txq);
3603 
3604 	avail = IDXDIFF(pidx, cidx, r->size);
3605 	for (i = 0; i < avail; i++) {
3606 		mp = _ring_peek_one(r, cidx, i, avail - i);
3607 		if (__predict_false(*mp == (struct mbuf *)txq))
3608 			continue;
3609 		m_freem(*mp);
3610 	}
3611 	MPASS(ifmp_ring_is_stalled(r) == 0);
3612 	return (avail);
3613 }
3614 
3615 static void
3616 iflib_ifmp_purge(iflib_txq_t txq)
3617 {
3618 	struct ifmp_ring *r;
3619 
3620 	r = txq->ift_br;
3621 	r->drain = iflib_txq_drain_free;
3622 	r->can_drain = iflib_txq_drain_always;
3623 
3624 	ifmp_ring_check_drainage(r, r->size);
3625 
3626 	r->drain = iflib_txq_drain;
3627 	r->can_drain = iflib_txq_can_drain;
3628 }
3629 
3630 static void
3631 _task_fn_tx(void *context)
3632 {
3633 	iflib_txq_t txq = context;
3634 	if_ctx_t ctx = txq->ift_ctx;
3635 	struct ifnet *ifp = ctx->ifc_ifp;
3636 	int rc;
3637 
3638 #ifdef IFLIB_DIAGNOSTICS
3639 	txq->ift_cpu_exec_count[curcpu]++;
3640 #endif
3641 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3642 		return;
3643 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3644 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3645 			netmap_tx_irq(ifp, txq->ift_id);
3646 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3647 		return;
3648 	}
3649 	if (txq->ift_db_pending)
3650 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3651 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3652 	if (ctx->ifc_flags & IFC_LEGACY)
3653 		IFDI_INTR_ENABLE(ctx);
3654 	else {
3655 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3656 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3657 	}
3658 }
3659 
3660 static void
3661 _task_fn_rx(void *context)
3662 {
3663 	iflib_rxq_t rxq = context;
3664 	if_ctx_t ctx = rxq->ifr_ctx;
3665 	bool more;
3666 	int rc;
3667 	uint16_t budget;
3668 
3669 #ifdef IFLIB_DIAGNOSTICS
3670 	rxq->ifr_cpu_exec_count[curcpu]++;
3671 #endif
3672 	DBG_COUNTER_INC(task_fn_rxs);
3673 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3674 		return;
3675 	more = true;
3676 #ifdef DEV_NETMAP
3677 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3678 		u_int work = 0;
3679 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3680 			more = false;
3681 		}
3682 	}
3683 #endif
3684 	budget = ctx->ifc_sysctl_rx_budget;
3685 	if (budget == 0)
3686 		budget = 16;	/* XXX */
3687 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3688 		if (ctx->ifc_flags & IFC_LEGACY)
3689 			IFDI_INTR_ENABLE(ctx);
3690 		else {
3691 			DBG_COUNTER_INC(rx_intr_enables);
3692 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3693 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3694 		}
3695 	}
3696 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3697 		return;
3698 	if (more)
3699 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3700 }
3701 
3702 static void
3703 _task_fn_admin(void *context)
3704 {
3705 	if_ctx_t ctx = context;
3706 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3707 	iflib_txq_t txq;
3708 	int i;
3709 
3710 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3711 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3712 			return;
3713 		}
3714 	}
3715 
3716 	CTX_LOCK(ctx);
3717 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3718 		CALLOUT_LOCK(txq);
3719 		callout_stop(&txq->ift_timer);
3720 		CALLOUT_UNLOCK(txq);
3721 	}
3722 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3723 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3724 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3725 	IFDI_LINK_INTR_ENABLE(ctx);
3726 	if (ctx->ifc_flags & IFC_DO_RESET) {
3727 		ctx->ifc_flags &= ~IFC_DO_RESET;
3728 		iflib_if_init_locked(ctx);
3729 	}
3730 	CTX_UNLOCK(ctx);
3731 
3732 	if (LINK_ACTIVE(ctx) == 0)
3733 		return;
3734 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3735 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3736 }
3737 
3738 
3739 static void
3740 _task_fn_iov(void *context)
3741 {
3742 	if_ctx_t ctx = context;
3743 
3744 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3745 		return;
3746 
3747 	CTX_LOCK(ctx);
3748 	IFDI_VFLR_HANDLE(ctx);
3749 	CTX_UNLOCK(ctx);
3750 }
3751 
3752 static int
3753 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3754 {
3755 	int err;
3756 	if_int_delay_info_t info;
3757 	if_ctx_t ctx;
3758 
3759 	info = (if_int_delay_info_t)arg1;
3760 	ctx = info->iidi_ctx;
3761 	info->iidi_req = req;
3762 	info->iidi_oidp = oidp;
3763 	CTX_LOCK(ctx);
3764 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3765 	CTX_UNLOCK(ctx);
3766 	return (err);
3767 }
3768 
3769 /*********************************************************************
3770  *
3771  *  IFNET FUNCTIONS
3772  *
3773  **********************************************************************/
3774 
3775 static void
3776 iflib_if_init_locked(if_ctx_t ctx)
3777 {
3778 	iflib_stop(ctx);
3779 	iflib_init_locked(ctx);
3780 }
3781 
3782 
3783 static void
3784 iflib_if_init(void *arg)
3785 {
3786 	if_ctx_t ctx = arg;
3787 
3788 	CTX_LOCK(ctx);
3789 	iflib_if_init_locked(ctx);
3790 	CTX_UNLOCK(ctx);
3791 }
3792 
3793 static int
3794 iflib_if_transmit(if_t ifp, struct mbuf *m)
3795 {
3796 	if_ctx_t	ctx = if_getsoftc(ifp);
3797 
3798 	iflib_txq_t txq;
3799 	int err, qidx;
3800 
3801 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3802 		DBG_COUNTER_INC(tx_frees);
3803 		m_freem(m);
3804 		return (ENOBUFS);
3805 	}
3806 
3807 	MPASS(m->m_nextpkt == NULL);
3808 	qidx = 0;
3809 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3810 		qidx = QIDX(ctx, m);
3811 	/*
3812 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3813 	 */
3814 	txq = &ctx->ifc_txqs[qidx];
3815 
3816 #ifdef DRIVER_BACKPRESSURE
3817 	if (txq->ift_closed) {
3818 		while (m != NULL) {
3819 			next = m->m_nextpkt;
3820 			m->m_nextpkt = NULL;
3821 			m_freem(m);
3822 			m = next;
3823 		}
3824 		return (ENOBUFS);
3825 	}
3826 #endif
3827 #ifdef notyet
3828 	qidx = count = 0;
3829 	mp = marr;
3830 	next = m;
3831 	do {
3832 		count++;
3833 		next = next->m_nextpkt;
3834 	} while (next != NULL);
3835 
3836 	if (count > nitems(marr))
3837 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3838 			/* XXX check nextpkt */
3839 			m_freem(m);
3840 			/* XXX simplify for now */
3841 			DBG_COUNTER_INC(tx_frees);
3842 			return (ENOBUFS);
3843 		}
3844 	for (next = m, i = 0; next != NULL; i++) {
3845 		mp[i] = next;
3846 		next = next->m_nextpkt;
3847 		mp[i]->m_nextpkt = NULL;
3848 	}
3849 #endif
3850 	DBG_COUNTER_INC(tx_seen);
3851 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3852 
3853 	GROUPTASK_ENQUEUE(&txq->ift_task);
3854 	if (err) {
3855 		/* support forthcoming later */
3856 #ifdef DRIVER_BACKPRESSURE
3857 		txq->ift_closed = TRUE;
3858 #endif
3859 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3860 		m_freem(m);
3861 	}
3862 
3863 	return (err);
3864 }
3865 
3866 static void
3867 iflib_if_qflush(if_t ifp)
3868 {
3869 	if_ctx_t ctx = if_getsoftc(ifp);
3870 	iflib_txq_t txq = ctx->ifc_txqs;
3871 	int i;
3872 
3873 	CTX_LOCK(ctx);
3874 	ctx->ifc_flags |= IFC_QFLUSH;
3875 	CTX_UNLOCK(ctx);
3876 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3877 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3878 			iflib_txq_check_drain(txq, 0);
3879 	CTX_LOCK(ctx);
3880 	ctx->ifc_flags &= ~IFC_QFLUSH;
3881 	CTX_UNLOCK(ctx);
3882 
3883 	if_qflush(ifp);
3884 }
3885 
3886 
3887 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3888 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3889 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3890 
3891 static int
3892 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3893 {
3894 	if_ctx_t ctx = if_getsoftc(ifp);
3895 	struct ifreq	*ifr = (struct ifreq *)data;
3896 #if defined(INET) || defined(INET6)
3897 	struct ifaddr	*ifa = (struct ifaddr *)data;
3898 #endif
3899 	bool		avoid_reset = FALSE;
3900 	int		err = 0, reinit = 0, bits;
3901 
3902 	switch (command) {
3903 	case SIOCSIFADDR:
3904 #ifdef INET
3905 		if (ifa->ifa_addr->sa_family == AF_INET)
3906 			avoid_reset = TRUE;
3907 #endif
3908 #ifdef INET6
3909 		if (ifa->ifa_addr->sa_family == AF_INET6)
3910 			avoid_reset = TRUE;
3911 #endif
3912 		/*
3913 		** Calling init results in link renegotiation,
3914 		** so we avoid doing it when possible.
3915 		*/
3916 		if (avoid_reset) {
3917 			if_setflagbits(ifp, IFF_UP,0);
3918 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3919 				reinit = 1;
3920 #ifdef INET
3921 			if (!(if_getflags(ifp) & IFF_NOARP))
3922 				arp_ifinit(ifp, ifa);
3923 #endif
3924 		} else
3925 			err = ether_ioctl(ifp, command, data);
3926 		break;
3927 	case SIOCSIFMTU:
3928 		CTX_LOCK(ctx);
3929 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3930 			CTX_UNLOCK(ctx);
3931 			break;
3932 		}
3933 		bits = if_getdrvflags(ifp);
3934 		/* stop the driver and free any clusters before proceeding */
3935 		iflib_stop(ctx);
3936 
3937 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3938 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3939 				ctx->ifc_flags |= IFC_MULTISEG;
3940 			else
3941 				ctx->ifc_flags &= ~IFC_MULTISEG;
3942 			err = if_setmtu(ifp, ifr->ifr_mtu);
3943 		}
3944 		iflib_init_locked(ctx);
3945 		if_setdrvflags(ifp, bits);
3946 		CTX_UNLOCK(ctx);
3947 		break;
3948 	case SIOCSIFFLAGS:
3949 		CTX_LOCK(ctx);
3950 		if (if_getflags(ifp) & IFF_UP) {
3951 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3952 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3953 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3954 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3955 				}
3956 			} else
3957 				reinit = 1;
3958 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3959 			iflib_stop(ctx);
3960 		}
3961 		ctx->ifc_if_flags = if_getflags(ifp);
3962 		CTX_UNLOCK(ctx);
3963 		break;
3964 	case SIOCADDMULTI:
3965 	case SIOCDELMULTI:
3966 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3967 			CTX_LOCK(ctx);
3968 			IFDI_INTR_DISABLE(ctx);
3969 			IFDI_MULTI_SET(ctx);
3970 			IFDI_INTR_ENABLE(ctx);
3971 			CTX_UNLOCK(ctx);
3972 		}
3973 		break;
3974 	case SIOCSIFMEDIA:
3975 		CTX_LOCK(ctx);
3976 		IFDI_MEDIA_SET(ctx);
3977 		CTX_UNLOCK(ctx);
3978 		/* falls thru */
3979 	case SIOCGIFMEDIA:
3980 	case SIOCGIFXMEDIA:
3981 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3982 		break;
3983 	case SIOCGI2C:
3984 	{
3985 		struct ifi2creq i2c;
3986 
3987 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3988 		if (err != 0)
3989 			break;
3990 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3991 			err = EINVAL;
3992 			break;
3993 		}
3994 		if (i2c.len > sizeof(i2c.data)) {
3995 			err = EINVAL;
3996 			break;
3997 		}
3998 
3999 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4000 			err = copyout(&i2c, ifr->ifr_data, 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 
4201 	/*
4202 	 * XXX sanity check that ntxd & nrxd are a power of 2
4203 	 */
4204 	if (ctx->ifc_sysctl_ntxqs != 0)
4205 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4206 	if (ctx->ifc_sysctl_nrxqs != 0)
4207 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4208 
4209 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4210 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4211 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4212 		else
4213 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4214 	}
4215 
4216 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4217 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4218 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4219 		else
4220 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4221 	}
4222 
4223 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4224 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4225 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4226 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4227 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4228 		}
4229 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4230 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4231 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4232 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4233 		}
4234 	}
4235 
4236 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4237 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4238 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4239 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4240 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4241 		}
4242 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4243 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4244 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4245 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4246 		}
4247 	}
4248 
4249 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4250 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4251 		return (err);
4252 	}
4253 	_iflib_pre_assert(scctx);
4254 	ctx->ifc_txrx = *scctx->isc_txrx;
4255 
4256 #ifdef INVARIANTS
4257 	MPASS(scctx->isc_capenable);
4258 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4259 		MPASS(scctx->isc_tx_csum_flags);
4260 #endif
4261 
4262 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4263 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4264 
4265 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4266 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4267 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4268 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4269 
4270 #ifdef ACPI_DMAR
4271 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4272 		ctx->ifc_flags |= IFC_DMAR;
4273 #elif !(defined(__i386__) || defined(__amd64__))
4274 	/* set unconditionally for !x86 */
4275 	ctx->ifc_flags |= IFC_DMAR;
4276 #endif
4277 
4278 	msix_bar = scctx->isc_msix_bar;
4279 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4280 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4281 
4282 	/* XXX change for per-queue sizes */
4283 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4284 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4285 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4286 		if (!powerof2(scctx->isc_nrxd[i])) {
4287 			/* round down instead? */
4288 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4289 			err = EINVAL;
4290 			goto fail;
4291 		}
4292 	}
4293 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4294 		if (!powerof2(scctx->isc_ntxd[i])) {
4295 			device_printf(dev,
4296 			    "# tx descriptors must be a power of 2");
4297 			err = EINVAL;
4298 			goto fail;
4299 		}
4300 	}
4301 
4302 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4303 	    MAX_SINGLE_PACKET_FRACTION)
4304 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4305 		    MAX_SINGLE_PACKET_FRACTION);
4306 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4307 	    MAX_SINGLE_PACKET_FRACTION)
4308 		scctx->isc_tx_tso_segments_max = max(1,
4309 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4310 
4311 	/*
4312 	 * Protect the stack against modern hardware
4313 	 */
4314 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4315 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4316 
4317 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4318 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4319 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4320 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4321 	if (scctx->isc_rss_table_size == 0)
4322 		scctx->isc_rss_table_size = 64;
4323 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4324 
4325 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4326 	/* XXX format name */
4327 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4328 
4329 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4330 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4331 		device_printf(dev, "Unable to fetch CPU list\n");
4332 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4333 	}
4334 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4335 
4336 	/*
4337 	** Now setup MSI or MSI/X, should
4338 	** return us the number of supported
4339 	** vectors. (Will be 1 for MSI)
4340 	*/
4341 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4342 		msix = scctx->isc_vectors;
4343 	} else if (scctx->isc_msix_bar != 0)
4344 	       /*
4345 		* The simple fact that isc_msix_bar is not 0 does not mean we
4346 		* we have a good value there that is known to work.
4347 		*/
4348 		msix = iflib_msix_init(ctx);
4349 	else {
4350 		scctx->isc_vectors = 1;
4351 		scctx->isc_ntxqsets = 1;
4352 		scctx->isc_nrxqsets = 1;
4353 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4354 		msix = 0;
4355 	}
4356 	/* Get memory for the station queues */
4357 	if ((err = iflib_queues_alloc(ctx))) {
4358 		device_printf(dev, "Unable to allocate queue memory\n");
4359 		goto fail;
4360 	}
4361 
4362 	if ((err = iflib_qset_structures_setup(ctx))) {
4363 		device_printf(dev, "qset structure setup failed %d\n", err);
4364 		goto fail_queues;
4365 	}
4366 
4367 	/*
4368 	 * Group taskqueues aren't properly set up until SMP is started,
4369 	 * so we disable interrupts until we can handle them post
4370 	 * SI_SUB_SMP.
4371 	 *
4372 	 * XXX: disabling interrupts doesn't actually work, at least for
4373 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4374 	 * we do null handling and depend on this not causing too large an
4375 	 * interrupt storm.
4376 	 */
4377 	IFDI_INTR_DISABLE(ctx);
4378 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4379 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4380 		goto fail_intr_free;
4381 	}
4382 	if (msix <= 1) {
4383 		rid = 0;
4384 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4385 			MPASS(msix == 1);
4386 			rid = 1;
4387 		}
4388 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4389 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4390 			goto fail_intr_free;
4391 		}
4392 	}
4393 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4394 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4395 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4396 		goto fail_detach;
4397 	}
4398 	if ((err = iflib_netmap_attach(ctx))) {
4399 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4400 		goto fail_detach;
4401 	}
4402 	*ctxp = ctx;
4403 
4404 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4405 	iflib_add_device_sysctl_post(ctx);
4406 	ctx->ifc_flags |= IFC_INIT_DONE;
4407 	return (0);
4408 fail_detach:
4409 	ether_ifdetach(ctx->ifc_ifp);
4410 fail_intr_free:
4411 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4412 		pci_release_msi(ctx->ifc_dev);
4413 fail_queues:
4414 	/* XXX free queues */
4415 fail:
4416 	IFDI_DETACH(ctx);
4417 	return (err);
4418 }
4419 
4420 int
4421 iflib_device_attach(device_t dev)
4422 {
4423 	if_ctx_t ctx;
4424 	if_shared_ctx_t sctx;
4425 
4426 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4427 		return (ENOTSUP);
4428 
4429 	pci_enable_busmaster(dev);
4430 
4431 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4432 }
4433 
4434 int
4435 iflib_device_deregister(if_ctx_t ctx)
4436 {
4437 	if_t ifp = ctx->ifc_ifp;
4438 	iflib_txq_t txq;
4439 	iflib_rxq_t rxq;
4440 	device_t dev = ctx->ifc_dev;
4441 	int i, j;
4442 	struct taskqgroup *tqg;
4443 	iflib_fl_t fl;
4444 
4445 	/* Make sure VLANS are not using driver */
4446 	if (if_vlantrunkinuse(ifp)) {
4447 		device_printf(dev,"Vlan in use, detach first\n");
4448 		return (EBUSY);
4449 	}
4450 
4451 	CTX_LOCK(ctx);
4452 	ctx->ifc_in_detach = 1;
4453 	iflib_stop(ctx);
4454 	CTX_UNLOCK(ctx);
4455 
4456 	/* Unregister VLAN events */
4457 	if (ctx->ifc_vlan_attach_event != NULL)
4458 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4459 	if (ctx->ifc_vlan_detach_event != NULL)
4460 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4461 
4462 	iflib_netmap_detach(ifp);
4463 	ether_ifdetach(ifp);
4464 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4465 	CTX_LOCK_DESTROY(ctx);
4466 	if (ctx->ifc_led_dev != NULL)
4467 		led_destroy(ctx->ifc_led_dev);
4468 	/* XXX drain any dependent tasks */
4469 	tqg = qgroup_if_io_tqg;
4470 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4471 		callout_drain(&txq->ift_timer);
4472 		if (txq->ift_task.gt_uniq != NULL)
4473 			taskqgroup_detach(tqg, &txq->ift_task);
4474 	}
4475 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4476 		if (rxq->ifr_task.gt_uniq != NULL)
4477 			taskqgroup_detach(tqg, &rxq->ifr_task);
4478 
4479 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4480 			free(fl->ifl_rx_bitmap, M_IFLIB);
4481 
4482 	}
4483 	tqg = qgroup_if_config_tqg;
4484 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4485 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4486 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4487 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4488 
4489 	IFDI_DETACH(ctx);
4490 	device_set_softc(ctx->ifc_dev, NULL);
4491 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4492 		pci_release_msi(dev);
4493 	}
4494 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4495 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4496 	}
4497 	if (ctx->ifc_msix_mem != NULL) {
4498 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4499 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4500 		ctx->ifc_msix_mem = NULL;
4501 	}
4502 
4503 	bus_generic_detach(dev);
4504 	if_free(ifp);
4505 
4506 	iflib_tx_structures_free(ctx);
4507 	iflib_rx_structures_free(ctx);
4508 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4509 		free(ctx->ifc_softc, M_IFLIB);
4510 	free(ctx, M_IFLIB);
4511 	return (0);
4512 }
4513 
4514 
4515 int
4516 iflib_device_detach(device_t dev)
4517 {
4518 	if_ctx_t ctx = device_get_softc(dev);
4519 
4520 	return (iflib_device_deregister(ctx));
4521 }
4522 
4523 int
4524 iflib_device_suspend(device_t dev)
4525 {
4526 	if_ctx_t ctx = device_get_softc(dev);
4527 
4528 	CTX_LOCK(ctx);
4529 	IFDI_SUSPEND(ctx);
4530 	CTX_UNLOCK(ctx);
4531 
4532 	return bus_generic_suspend(dev);
4533 }
4534 int
4535 iflib_device_shutdown(device_t dev)
4536 {
4537 	if_ctx_t ctx = device_get_softc(dev);
4538 
4539 	CTX_LOCK(ctx);
4540 	IFDI_SHUTDOWN(ctx);
4541 	CTX_UNLOCK(ctx);
4542 
4543 	return bus_generic_suspend(dev);
4544 }
4545 
4546 
4547 int
4548 iflib_device_resume(device_t dev)
4549 {
4550 	if_ctx_t ctx = device_get_softc(dev);
4551 	iflib_txq_t txq = ctx->ifc_txqs;
4552 
4553 	CTX_LOCK(ctx);
4554 	IFDI_RESUME(ctx);
4555 	iflib_init_locked(ctx);
4556 	CTX_UNLOCK(ctx);
4557 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4558 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4559 
4560 	return (bus_generic_resume(dev));
4561 }
4562 
4563 int
4564 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4565 {
4566 	int error;
4567 	if_ctx_t ctx = device_get_softc(dev);
4568 
4569 	CTX_LOCK(ctx);
4570 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4571 	CTX_UNLOCK(ctx);
4572 
4573 	return (error);
4574 }
4575 
4576 void
4577 iflib_device_iov_uninit(device_t dev)
4578 {
4579 	if_ctx_t ctx = device_get_softc(dev);
4580 
4581 	CTX_LOCK(ctx);
4582 	IFDI_IOV_UNINIT(ctx);
4583 	CTX_UNLOCK(ctx);
4584 }
4585 
4586 int
4587 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4588 {
4589 	int error;
4590 	if_ctx_t ctx = device_get_softc(dev);
4591 
4592 	CTX_LOCK(ctx);
4593 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4594 	CTX_UNLOCK(ctx);
4595 
4596 	return (error);
4597 }
4598 
4599 /*********************************************************************
4600  *
4601  *  MODULE FUNCTION DEFINITIONS
4602  *
4603  **********************************************************************/
4604 
4605 /*
4606  * - Start a fast taskqueue thread for each core
4607  * - Start a taskqueue for control operations
4608  */
4609 static int
4610 iflib_module_init(void)
4611 {
4612 	return (0);
4613 }
4614 
4615 static int
4616 iflib_module_event_handler(module_t mod, int what, void *arg)
4617 {
4618 	int err;
4619 
4620 	switch (what) {
4621 	case MOD_LOAD:
4622 		if ((err = iflib_module_init()) != 0)
4623 			return (err);
4624 		break;
4625 	case MOD_UNLOAD:
4626 		return (EBUSY);
4627 	default:
4628 		return (EOPNOTSUPP);
4629 	}
4630 
4631 	return (0);
4632 }
4633 
4634 /*********************************************************************
4635  *
4636  *  PUBLIC FUNCTION DEFINITIONS
4637  *     ordered as in iflib.h
4638  *
4639  **********************************************************************/
4640 
4641 
4642 static void
4643 _iflib_assert(if_shared_ctx_t sctx)
4644 {
4645 	MPASS(sctx->isc_tx_maxsize);
4646 	MPASS(sctx->isc_tx_maxsegsize);
4647 
4648 	MPASS(sctx->isc_rx_maxsize);
4649 	MPASS(sctx->isc_rx_nsegments);
4650 	MPASS(sctx->isc_rx_maxsegsize);
4651 
4652 	MPASS(sctx->isc_nrxd_min[0]);
4653 	MPASS(sctx->isc_nrxd_max[0]);
4654 	MPASS(sctx->isc_nrxd_default[0]);
4655 	MPASS(sctx->isc_ntxd_min[0]);
4656 	MPASS(sctx->isc_ntxd_max[0]);
4657 	MPASS(sctx->isc_ntxd_default[0]);
4658 }
4659 
4660 static void
4661 _iflib_pre_assert(if_softc_ctx_t scctx)
4662 {
4663 
4664 	MPASS(scctx->isc_txrx->ift_txd_encap);
4665 	MPASS(scctx->isc_txrx->ift_txd_flush);
4666 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4667 	MPASS(scctx->isc_txrx->ift_rxd_available);
4668 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4669 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4670 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4671 }
4672 
4673 static int
4674 iflib_register(if_ctx_t ctx)
4675 {
4676 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4677 	driver_t *driver = sctx->isc_driver;
4678 	device_t dev = ctx->ifc_dev;
4679 	if_t ifp;
4680 
4681 	_iflib_assert(sctx);
4682 
4683 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4684 
4685 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4686 	if (ifp == NULL) {
4687 		device_printf(dev, "can not allocate ifnet structure\n");
4688 		return (ENOMEM);
4689 	}
4690 
4691 	/*
4692 	 * Initialize our context's device specific methods
4693 	 */
4694 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4695 	kobj_class_compile((kobj_class_t) driver);
4696 	driver->refs++;
4697 
4698 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4699 	if_setsoftc(ifp, ctx);
4700 	if_setdev(ifp, dev);
4701 	if_setinitfn(ifp, iflib_if_init);
4702 	if_setioctlfn(ifp, iflib_if_ioctl);
4703 	if_settransmitfn(ifp, iflib_if_transmit);
4704 	if_setqflushfn(ifp, iflib_if_qflush);
4705 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4706 
4707 	ctx->ifc_vlan_attach_event =
4708 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4709 							  EVENTHANDLER_PRI_FIRST);
4710 	ctx->ifc_vlan_detach_event =
4711 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4712 							  EVENTHANDLER_PRI_FIRST);
4713 
4714 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4715 					 iflib_media_change, iflib_media_status);
4716 
4717 	return (0);
4718 }
4719 
4720 
4721 static int
4722 iflib_queues_alloc(if_ctx_t ctx)
4723 {
4724 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4725 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4726 	device_t dev = ctx->ifc_dev;
4727 	int nrxqsets = scctx->isc_nrxqsets;
4728 	int ntxqsets = scctx->isc_ntxqsets;
4729 	iflib_txq_t txq;
4730 	iflib_rxq_t rxq;
4731 	iflib_fl_t fl = NULL;
4732 	int i, j, cpu, err, txconf, rxconf;
4733 	iflib_dma_info_t ifdip;
4734 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4735 	uint32_t *txqsizes = scctx->isc_txqsizes;
4736 	uint8_t nrxqs = sctx->isc_nrxqs;
4737 	uint8_t ntxqs = sctx->isc_ntxqs;
4738 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4739 	caddr_t *vaddrs;
4740 	uint64_t *paddrs;
4741 	struct ifmp_ring **brscp;
4742 
4743 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4744 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4745 
4746 	brscp = NULL;
4747 	txq = NULL;
4748 	rxq = NULL;
4749 
4750 /* Allocate the TX ring struct memory */
4751 	if (!(txq =
4752 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4753 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4754 		device_printf(dev, "Unable to allocate TX ring memory\n");
4755 		err = ENOMEM;
4756 		goto fail;
4757 	}
4758 
4759 	/* Now allocate the RX */
4760 	if (!(rxq =
4761 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4762 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4763 		device_printf(dev, "Unable to allocate RX ring memory\n");
4764 		err = ENOMEM;
4765 		goto rx_fail;
4766 	}
4767 
4768 	ctx->ifc_txqs = txq;
4769 	ctx->ifc_rxqs = rxq;
4770 
4771 	/*
4772 	 * XXX handle allocation failure
4773 	 */
4774 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4775 		/* Set up some basics */
4776 
4777 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4778 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4779 			err = ENOMEM;
4780 			goto err_tx_desc;
4781 		}
4782 		txq->ift_ifdi = ifdip;
4783 		for (j = 0; j < ntxqs; j++, ifdip++) {
4784 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4785 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4786 				err = ENOMEM;
4787 				goto err_tx_desc;
4788 			}
4789 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4790 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4791 		}
4792 		txq->ift_ctx = ctx;
4793 		txq->ift_id = i;
4794 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4795 			txq->ift_br_offset = 1;
4796 		} else {
4797 			txq->ift_br_offset = 0;
4798 		}
4799 		/* XXX fix this */
4800 		txq->ift_timer.c_cpu = cpu;
4801 
4802 		if (iflib_txsd_alloc(txq)) {
4803 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4804 			err = ENOMEM;
4805 			goto err_tx_desc;
4806 		}
4807 
4808 		/* Initialize the TX lock */
4809 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4810 		    device_get_nameunit(dev), txq->ift_id);
4811 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4812 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4813 
4814 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4815 			 device_get_nameunit(dev), txq->ift_id);
4816 
4817 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4818 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4819 		if (err) {
4820 			/* XXX free any allocated rings */
4821 			device_printf(dev, "Unable to allocate buf_ring\n");
4822 			goto err_tx_desc;
4823 		}
4824 	}
4825 
4826 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4827 		/* Set up some basics */
4828 
4829 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4830 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4831 			err = ENOMEM;
4832 			goto err_tx_desc;
4833 		}
4834 
4835 		rxq->ifr_ifdi = ifdip;
4836 		/* XXX this needs to be changed if #rx queues != #tx queues */
4837 		rxq->ifr_ntxqirq = 1;
4838 		rxq->ifr_txqid[0] = i;
4839 		for (j = 0; j < nrxqs; j++, ifdip++) {
4840 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4841 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4842 				err = ENOMEM;
4843 				goto err_tx_desc;
4844 			}
4845 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4846 		}
4847 		rxq->ifr_ctx = ctx;
4848 		rxq->ifr_id = i;
4849 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4850 			rxq->ifr_fl_offset = 1;
4851 		} else {
4852 			rxq->ifr_fl_offset = 0;
4853 		}
4854 		rxq->ifr_nfl = nfree_lists;
4855 		if (!(fl =
4856 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4857 			device_printf(dev, "Unable to allocate free list memory\n");
4858 			err = ENOMEM;
4859 			goto err_tx_desc;
4860 		}
4861 		rxq->ifr_fl = fl;
4862 		for (j = 0; j < nfree_lists; j++) {
4863 			fl[j].ifl_rxq = rxq;
4864 			fl[j].ifl_id = j;
4865 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4866 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4867 		}
4868         /* Allocate receive buffers for the ring*/
4869 		if (iflib_rxsd_alloc(rxq)) {
4870 			device_printf(dev,
4871 			    "Critical Failure setting up receive buffers\n");
4872 			err = ENOMEM;
4873 			goto err_rx_desc;
4874 		}
4875 
4876 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4877 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4878 	}
4879 
4880 	/* TXQs */
4881 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4882 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4883 	for (i = 0; i < ntxqsets; i++) {
4884 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4885 
4886 		for (j = 0; j < ntxqs; j++, di++) {
4887 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4888 			paddrs[i*ntxqs + j] = di->idi_paddr;
4889 		}
4890 	}
4891 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4892 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4893 		iflib_tx_structures_free(ctx);
4894 		free(vaddrs, M_IFLIB);
4895 		free(paddrs, M_IFLIB);
4896 		goto err_rx_desc;
4897 	}
4898 	free(vaddrs, M_IFLIB);
4899 	free(paddrs, M_IFLIB);
4900 
4901 	/* RXQs */
4902 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4903 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4904 	for (i = 0; i < nrxqsets; i++) {
4905 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4906 
4907 		for (j = 0; j < nrxqs; j++, di++) {
4908 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4909 			paddrs[i*nrxqs + j] = di->idi_paddr;
4910 		}
4911 	}
4912 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4913 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4914 		iflib_tx_structures_free(ctx);
4915 		free(vaddrs, M_IFLIB);
4916 		free(paddrs, M_IFLIB);
4917 		goto err_rx_desc;
4918 	}
4919 	free(vaddrs, M_IFLIB);
4920 	free(paddrs, M_IFLIB);
4921 
4922 	return (0);
4923 
4924 /* XXX handle allocation failure changes */
4925 err_rx_desc:
4926 err_tx_desc:
4927 	if (ctx->ifc_rxqs != NULL)
4928 		free(ctx->ifc_rxqs, M_IFLIB);
4929 	ctx->ifc_rxqs = NULL;
4930 	if (ctx->ifc_txqs != NULL)
4931 		free(ctx->ifc_txqs, M_IFLIB);
4932 	ctx->ifc_txqs = NULL;
4933 rx_fail:
4934 	if (brscp != NULL)
4935 		free(brscp, M_IFLIB);
4936 	if (rxq != NULL)
4937 		free(rxq, M_IFLIB);
4938 	if (txq != NULL)
4939 		free(txq, M_IFLIB);
4940 fail:
4941 	return (err);
4942 }
4943 
4944 static int
4945 iflib_tx_structures_setup(if_ctx_t ctx)
4946 {
4947 	iflib_txq_t txq = ctx->ifc_txqs;
4948 	int i;
4949 
4950 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4951 		iflib_txq_setup(txq);
4952 
4953 	return (0);
4954 }
4955 
4956 static void
4957 iflib_tx_structures_free(if_ctx_t ctx)
4958 {
4959 	iflib_txq_t txq = ctx->ifc_txqs;
4960 	int i, j;
4961 
4962 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4963 		iflib_txq_destroy(txq);
4964 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4965 			iflib_dma_free(&txq->ift_ifdi[j]);
4966 	}
4967 	free(ctx->ifc_txqs, M_IFLIB);
4968 	ctx->ifc_txqs = NULL;
4969 	IFDI_QUEUES_FREE(ctx);
4970 }
4971 
4972 /*********************************************************************
4973  *
4974  *  Initialize all receive rings.
4975  *
4976  **********************************************************************/
4977 static int
4978 iflib_rx_structures_setup(if_ctx_t ctx)
4979 {
4980 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4981 	int q;
4982 #if defined(INET6) || defined(INET)
4983 	int i, err;
4984 #endif
4985 
4986 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4987 #if defined(INET6) || defined(INET)
4988 		tcp_lro_free(&rxq->ifr_lc);
4989 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4990 		    TCP_LRO_ENTRIES, min(1024,
4991 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4992 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4993 			goto fail;
4994 		}
4995 		rxq->ifr_lro_enabled = TRUE;
4996 #endif
4997 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4998 	}
4999 	return (0);
5000 #if defined(INET6) || defined(INET)
5001 fail:
5002 	/*
5003 	 * Free RX software descriptors allocated so far, we will only handle
5004 	 * the rings that completed, the failing case will have
5005 	 * cleaned up for itself. 'q' failed, so its the terminus.
5006 	 */
5007 	rxq = ctx->ifc_rxqs;
5008 	for (i = 0; i < q; ++i, rxq++) {
5009 		iflib_rx_sds_free(rxq);
5010 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5011 	}
5012 	return (err);
5013 #endif
5014 }
5015 
5016 /*********************************************************************
5017  *
5018  *  Free all receive rings.
5019  *
5020  **********************************************************************/
5021 static void
5022 iflib_rx_structures_free(if_ctx_t ctx)
5023 {
5024 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5025 
5026 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5027 		iflib_rx_sds_free(rxq);
5028 	}
5029 }
5030 
5031 static int
5032 iflib_qset_structures_setup(if_ctx_t ctx)
5033 {
5034 	int err;
5035 
5036 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5037 		return (err);
5038 
5039 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5040 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5041 		iflib_tx_structures_free(ctx);
5042 		iflib_rx_structures_free(ctx);
5043 	}
5044 	return (err);
5045 }
5046 
5047 int
5048 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5049 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5050 {
5051 
5052 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5053 }
5054 
5055 #ifdef SMP
5056 static int
5057 find_nth(if_ctx_t ctx, int qid)
5058 {
5059 	cpuset_t cpus;
5060 	int i, cpuid, eqid, count;
5061 
5062 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5063 	count = CPU_COUNT(&cpus);
5064 	eqid = qid % count;
5065 	/* clear up to the qid'th bit */
5066 	for (i = 0; i < eqid; i++) {
5067 		cpuid = CPU_FFS(&cpus);
5068 		MPASS(cpuid != 0);
5069 		CPU_CLR(cpuid-1, &cpus);
5070 	}
5071 	cpuid = CPU_FFS(&cpus);
5072 	MPASS(cpuid != 0);
5073 	return (cpuid-1);
5074 }
5075 
5076 #ifdef SCHED_ULE
5077 extern struct cpu_group *cpu_top;              /* CPU topology */
5078 
5079 static int
5080 find_child_with_core(int cpu, struct cpu_group *grp)
5081 {
5082 	int i;
5083 
5084 	if (grp->cg_children == 0)
5085 		return -1;
5086 
5087 	MPASS(grp->cg_child);
5088 	for (i = 0; i < grp->cg_children; i++) {
5089 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5090 			return i;
5091 	}
5092 
5093 	return -1;
5094 }
5095 
5096 /*
5097  * Find the nth thread on the specified core
5098  */
5099 static int
5100 find_thread(int cpu, int thread_num)
5101 {
5102 	struct cpu_group *grp;
5103 	int i;
5104 	cpuset_t cs;
5105 
5106 	grp = cpu_top;
5107 	if (grp == NULL)
5108 		return cpu;
5109 	i = 0;
5110 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5111 		/* If the child only has one cpu, don't descend */
5112 		if (grp->cg_child[i].cg_count <= 1)
5113 			break;
5114 		grp = &grp->cg_child[i];
5115 	}
5116 
5117 	/* If they don't share at least an L2 cache, use the same CPU */
5118 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5119 		return cpu;
5120 
5121 	/* Now pick one */
5122 	CPU_COPY(&grp->cg_mask, &cs);
5123 	for (i = thread_num % grp->cg_count; i > 0; i--) {
5124 		MPASS(CPU_FFS(&cs));
5125 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5126 	}
5127 	MPASS(CPU_FFS(&cs));
5128 	return CPU_FFS(&cs) - 1;
5129 }
5130 #else
5131 static int
5132 find_thread(int cpu, int thread_num __unused)
5133 {
5134 	return cpu;
5135 }
5136 #endif
5137 
5138 static int
5139 get_thread_num(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5140 {
5141 	switch (type) {
5142 	case IFLIB_INTR_TX:
5143 		/* TX queues get threads on the same core as the corresponding RX queue */
5144 		/* XXX handle multiple RX threads per core and more than two threads per core */
5145 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5146 	case IFLIB_INTR_RX:
5147 	case IFLIB_INTR_RXTX:
5148 		/* RX queues get the first thread on their core */
5149 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5150 	default:
5151 		return -1;
5152 	}
5153 }
5154 #else
5155 #define get_thread_num(ctx, type, qid)	CPU_FIRST()
5156 #define find_thread(cpuid, tid)		CPU_FIRST()
5157 #define find_nth(ctx, gid)		CPU_FIRST()
5158 #endif
5159 
5160 /* Just to avoid copy/paste */
5161 static inline int
5162 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5163     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5164 {
5165 	int cpuid;
5166 	int err, tid;
5167 
5168 	cpuid = find_nth(ctx, qid);
5169 	tid = get_thread_num(ctx, type, qid);
5170 	MPASS(tid >= 0);
5171 	cpuid = find_thread(cpuid, tid);
5172 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5173 	if (err) {
5174 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5175 		return (err);
5176 	}
5177 #ifdef notyet
5178 	if (cpuid > ctx->ifc_cpuid_highest)
5179 		ctx->ifc_cpuid_highest = cpuid;
5180 #endif
5181 	return 0;
5182 }
5183 
5184 int
5185 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5186 						iflib_intr_type_t type, driver_filter_t *filter,
5187 						void *filter_arg, int qid, char *name)
5188 {
5189 	struct grouptask *gtask;
5190 	struct taskqgroup *tqg;
5191 	iflib_filter_info_t info;
5192 	gtask_fn_t *fn;
5193 	int tqrid, err;
5194 	driver_filter_t *intr_fast;
5195 	void *q;
5196 
5197 	info = &ctx->ifc_filter_info;
5198 	tqrid = rid;
5199 
5200 	switch (type) {
5201 	/* XXX merge tx/rx for netmap? */
5202 	case IFLIB_INTR_TX:
5203 		q = &ctx->ifc_txqs[qid];
5204 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5205 		gtask = &ctx->ifc_txqs[qid].ift_task;
5206 		tqg = qgroup_if_io_tqg;
5207 		fn = _task_fn_tx;
5208 		intr_fast = iflib_fast_intr;
5209 		GROUPTASK_INIT(gtask, 0, fn, q);
5210 		break;
5211 	case IFLIB_INTR_RX:
5212 		q = &ctx->ifc_rxqs[qid];
5213 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5214 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5215 		tqg = qgroup_if_io_tqg;
5216 		fn = _task_fn_rx;
5217 		intr_fast = iflib_fast_intr;
5218 		GROUPTASK_INIT(gtask, 0, fn, q);
5219 		break;
5220 	case IFLIB_INTR_RXTX:
5221 		q = &ctx->ifc_rxqs[qid];
5222 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5223 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5224 		tqg = qgroup_if_io_tqg;
5225 		fn = _task_fn_rx;
5226 		intr_fast = iflib_fast_intr_rxtx;
5227 		GROUPTASK_INIT(gtask, 0, fn, q);
5228 		break;
5229 	case IFLIB_INTR_ADMIN:
5230 		q = ctx;
5231 		tqrid = -1;
5232 		info = &ctx->ifc_filter_info;
5233 		gtask = &ctx->ifc_admin_task;
5234 		tqg = qgroup_if_config_tqg;
5235 		fn = _task_fn_admin;
5236 		intr_fast = iflib_fast_intr_ctx;
5237 		break;
5238 	default:
5239 		panic("unknown net intr type");
5240 	}
5241 
5242 	info->ifi_filter = filter;
5243 	info->ifi_filter_arg = filter_arg;
5244 	info->ifi_task = gtask;
5245 	info->ifi_ctx = q;
5246 
5247 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5248 	if (err != 0) {
5249 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5250 		return (err);
5251 	}
5252 	if (type == IFLIB_INTR_ADMIN)
5253 		return (0);
5254 
5255 	if (tqrid != -1) {
5256 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5257 		if (err)
5258 			return (err);
5259 	} else {
5260 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5261 	}
5262 
5263 	return (0);
5264 }
5265 
5266 void
5267 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5268 {
5269 	struct grouptask *gtask;
5270 	struct taskqgroup *tqg;
5271 	gtask_fn_t *fn;
5272 	void *q;
5273 	int irq_num = -1;
5274 	int err;
5275 
5276 	switch (type) {
5277 	case IFLIB_INTR_TX:
5278 		q = &ctx->ifc_txqs[qid];
5279 		gtask = &ctx->ifc_txqs[qid].ift_task;
5280 		tqg = qgroup_if_io_tqg;
5281 		fn = _task_fn_tx;
5282 		if (irq != NULL)
5283 			irq_num = rman_get_start(irq->ii_res);
5284 		break;
5285 	case IFLIB_INTR_RX:
5286 		q = &ctx->ifc_rxqs[qid];
5287 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5288 		tqg = qgroup_if_io_tqg;
5289 		fn = _task_fn_rx;
5290 		if (irq != NULL)
5291 			irq_num = rman_get_start(irq->ii_res);
5292 		break;
5293 	case IFLIB_INTR_IOV:
5294 		q = ctx;
5295 		gtask = &ctx->ifc_vflr_task;
5296 		tqg = qgroup_if_config_tqg;
5297 		fn = _task_fn_iov;
5298 		break;
5299 	default:
5300 		panic("unknown net intr type");
5301 	}
5302 	GROUPTASK_INIT(gtask, 0, fn, q);
5303 	if (irq_num != -1) {
5304 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5305 		if (err)
5306 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5307 	}
5308 	else {
5309 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5310 	}
5311 }
5312 
5313 void
5314 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5315 {
5316 	if (irq->ii_tag)
5317 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5318 
5319 	if (irq->ii_res)
5320 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5321 }
5322 
5323 static int
5324 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5325 {
5326 	iflib_txq_t txq = ctx->ifc_txqs;
5327 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5328 	if_irq_t irq = &ctx->ifc_legacy_irq;
5329 	iflib_filter_info_t info;
5330 	struct grouptask *gtask;
5331 	struct taskqgroup *tqg;
5332 	gtask_fn_t *fn;
5333 	int tqrid;
5334 	void *q;
5335 	int err;
5336 
5337 	q = &ctx->ifc_rxqs[0];
5338 	info = &rxq[0].ifr_filter_info;
5339 	gtask = &rxq[0].ifr_task;
5340 	tqg = qgroup_if_io_tqg;
5341 	tqrid = irq->ii_rid = *rid;
5342 	fn = _task_fn_rx;
5343 
5344 	ctx->ifc_flags |= IFC_LEGACY;
5345 	info->ifi_filter = filter;
5346 	info->ifi_filter_arg = filter_arg;
5347 	info->ifi_task = gtask;
5348 	info->ifi_ctx = ctx;
5349 
5350 	/* We allocate a single interrupt resource */
5351 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5352 		return (err);
5353 	GROUPTASK_INIT(gtask, 0, fn, q);
5354 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5355 
5356 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5357 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5358 	return (0);
5359 }
5360 
5361 void
5362 iflib_led_create(if_ctx_t ctx)
5363 {
5364 
5365 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5366 	    device_get_nameunit(ctx->ifc_dev));
5367 }
5368 
5369 void
5370 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5371 {
5372 
5373 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5374 }
5375 
5376 void
5377 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5378 {
5379 
5380 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5381 }
5382 
5383 void
5384 iflib_admin_intr_deferred(if_ctx_t ctx)
5385 {
5386 #ifdef INVARIANTS
5387 	struct grouptask *gtask;
5388 
5389 	gtask = &ctx->ifc_admin_task;
5390 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5391 #endif
5392 
5393 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5394 }
5395 
5396 void
5397 iflib_iov_intr_deferred(if_ctx_t ctx)
5398 {
5399 
5400 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5401 }
5402 
5403 void
5404 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5405 {
5406 
5407 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5408 }
5409 
5410 void
5411 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5412 	char *name)
5413 {
5414 
5415 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5416 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5417 }
5418 
5419 void
5420 iflib_config_gtask_deinit(struct grouptask *gtask)
5421 {
5422 
5423 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5424 }
5425 
5426 void
5427 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5428 {
5429 	if_t ifp = ctx->ifc_ifp;
5430 	iflib_txq_t txq = ctx->ifc_txqs;
5431 
5432 	if_setbaudrate(ifp, baudrate);
5433 	if (baudrate >= IF_Gbps(10))
5434 		ctx->ifc_flags |= IFC_PREFETCH;
5435 
5436 	/* If link down, disable watchdog */
5437 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5438 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5439 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5440 	}
5441 	ctx->ifc_link_state = link_state;
5442 	if_link_state_change(ifp, link_state);
5443 }
5444 
5445 static int
5446 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5447 {
5448 	int credits;
5449 #ifdef INVARIANTS
5450 	int credits_pre = txq->ift_cidx_processed;
5451 #endif
5452 
5453 	if (ctx->isc_txd_credits_update == NULL)
5454 		return (0);
5455 
5456 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5457 		return (0);
5458 
5459 	txq->ift_processed += credits;
5460 	txq->ift_cidx_processed += credits;
5461 
5462 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5463 	if (txq->ift_cidx_processed >= txq->ift_size)
5464 		txq->ift_cidx_processed -= txq->ift_size;
5465 	return (credits);
5466 }
5467 
5468 static int
5469 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5470 {
5471 
5472 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5473 	    budget));
5474 }
5475 
5476 void
5477 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5478 	const char *description, if_int_delay_info_t info,
5479 	int offset, int value)
5480 {
5481 	info->iidi_ctx = ctx;
5482 	info->iidi_offset = offset;
5483 	info->iidi_value = value;
5484 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5485 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5486 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5487 	    info, 0, iflib_sysctl_int_delay, "I", description);
5488 }
5489 
5490 struct mtx *
5491 iflib_ctx_lock_get(if_ctx_t ctx)
5492 {
5493 
5494 	return (&ctx->ifc_mtx);
5495 }
5496 
5497 static int
5498 iflib_msix_init(if_ctx_t ctx)
5499 {
5500 	device_t dev = ctx->ifc_dev;
5501 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5502 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5503 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5504 	int iflib_num_tx_queues, iflib_num_rx_queues;
5505 	int err, admincnt, bar;
5506 
5507 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5508 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5509 
5510 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5511 
5512 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5513 	admincnt = sctx->isc_admin_intrcnt;
5514 	/* Override by global tuneable */
5515 	{
5516 		int i;
5517 		size_t len = sizeof(i);
5518 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5519 		if (err == 0) {
5520 			if (i == 0)
5521 				goto msi;
5522 		}
5523 		else {
5524 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5525 		}
5526 	}
5527 	/* Override by tuneable */
5528 	if (scctx->isc_disable_msix)
5529 		goto msi;
5530 
5531 	/*
5532 	** When used in a virtualized environment
5533 	** PCI BUSMASTER capability may not be set
5534 	** so explicity set it here and rewrite
5535 	** the ENABLE in the MSIX control register
5536 	** at this point to cause the host to
5537 	** successfully initialize us.
5538 	*/
5539 	{
5540 		int msix_ctrl, rid;
5541 
5542  		pci_enable_busmaster(dev);
5543 		rid = 0;
5544 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5545 			rid += PCIR_MSIX_CTRL;
5546 			msix_ctrl = pci_read_config(dev, rid, 2);
5547 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5548 			pci_write_config(dev, rid, msix_ctrl, 2);
5549 		} else {
5550 			device_printf(dev, "PCIY_MSIX capability not found; "
5551 			                   "or rid %d == 0.\n", rid);
5552 			goto msi;
5553 		}
5554 	}
5555 
5556 	/*
5557 	 * bar == -1 => "trust me I know what I'm doing"
5558 	 * Some drivers are for hardware that is so shoddily
5559 	 * documented that no one knows which bars are which
5560 	 * so the developer has to map all bars. This hack
5561 	 * allows shoddy garbage to use msix in this framework.
5562 	 */
5563 	if (bar != -1) {
5564 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5565 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5566 		if (ctx->ifc_msix_mem == NULL) {
5567 			/* May not be enabled */
5568 			device_printf(dev, "Unable to map MSIX table \n");
5569 			goto msi;
5570 		}
5571 	}
5572 	/* First try MSI/X */
5573 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5574 		device_printf(dev, "System has MSIX disabled \n");
5575 		bus_release_resource(dev, SYS_RES_MEMORY,
5576 		    bar, ctx->ifc_msix_mem);
5577 		ctx->ifc_msix_mem = NULL;
5578 		goto msi;
5579 	}
5580 #if IFLIB_DEBUG
5581 	/* use only 1 qset in debug mode */
5582 	queuemsgs = min(msgs - admincnt, 1);
5583 #else
5584 	queuemsgs = msgs - admincnt;
5585 #endif
5586 #ifdef RSS
5587 	queues = imin(queuemsgs, rss_getnumbuckets());
5588 #else
5589 	queues = queuemsgs;
5590 #endif
5591 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5592 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5593 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5594 #ifdef  RSS
5595 	/* If we're doing RSS, clamp at the number of RSS buckets */
5596 	if (queues > rss_getnumbuckets())
5597 		queues = rss_getnumbuckets();
5598 #endif
5599 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5600 		rx_queues = iflib_num_rx_queues;
5601 	else
5602 		rx_queues = queues;
5603 
5604 	if (rx_queues > scctx->isc_nrxqsets)
5605 		rx_queues = scctx->isc_nrxqsets;
5606 
5607 	/*
5608 	 * We want this to be all logical CPUs by default
5609 	 */
5610 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5611 		tx_queues = iflib_num_tx_queues;
5612 	else
5613 		tx_queues = mp_ncpus;
5614 
5615 	if (tx_queues > scctx->isc_ntxqsets)
5616 		tx_queues = scctx->isc_ntxqsets;
5617 
5618 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5619 #ifdef INVARIANTS
5620 		if (tx_queues != rx_queues)
5621 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5622 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5623 #endif
5624 		tx_queues = min(rx_queues, tx_queues);
5625 		rx_queues = min(rx_queues, tx_queues);
5626 	}
5627 
5628 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5629 
5630 	vectors = rx_queues + admincnt;
5631 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5632 		device_printf(dev,
5633 					  "Using MSIX interrupts with %d vectors\n", vectors);
5634 		scctx->isc_vectors = vectors;
5635 		scctx->isc_nrxqsets = rx_queues;
5636 		scctx->isc_ntxqsets = tx_queues;
5637 		scctx->isc_intr = IFLIB_INTR_MSIX;
5638 
5639 		return (vectors);
5640 	} else {
5641 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5642 	}
5643 msi:
5644 	vectors = pci_msi_count(dev);
5645 	scctx->isc_nrxqsets = 1;
5646 	scctx->isc_ntxqsets = 1;
5647 	scctx->isc_vectors = vectors;
5648 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5649 		device_printf(dev,"Using an MSI interrupt\n");
5650 		scctx->isc_intr = IFLIB_INTR_MSI;
5651 	} else {
5652 		device_printf(dev,"Using a Legacy interrupt\n");
5653 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5654 	}
5655 
5656 	return (vectors);
5657 }
5658 
5659 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5660 
5661 static int
5662 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5663 {
5664 	int rc;
5665 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5666 	struct sbuf *sb;
5667 	char *ring_state = "UNKNOWN";
5668 
5669 	/* XXX needed ? */
5670 	rc = sysctl_wire_old_buffer(req, 0);
5671 	MPASS(rc == 0);
5672 	if (rc != 0)
5673 		return (rc);
5674 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5675 	MPASS(sb != NULL);
5676 	if (sb == NULL)
5677 		return (ENOMEM);
5678 	if (state[3] <= 3)
5679 		ring_state = ring_states[state[3]];
5680 
5681 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5682 		    state[0], state[1], state[2], ring_state);
5683 	rc = sbuf_finish(sb);
5684 	sbuf_delete(sb);
5685         return(rc);
5686 }
5687 
5688 enum iflib_ndesc_handler {
5689 	IFLIB_NTXD_HANDLER,
5690 	IFLIB_NRXD_HANDLER,
5691 };
5692 
5693 static int
5694 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5695 {
5696 	if_ctx_t ctx = (void *)arg1;
5697 	enum iflib_ndesc_handler type = arg2;
5698 	char buf[256] = {0};
5699 	qidx_t *ndesc;
5700 	char *p, *next;
5701 	int nqs, rc, i;
5702 
5703 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5704 
5705 	nqs = 8;
5706 	switch(type) {
5707 	case IFLIB_NTXD_HANDLER:
5708 		ndesc = ctx->ifc_sysctl_ntxds;
5709 		if (ctx->ifc_sctx)
5710 			nqs = ctx->ifc_sctx->isc_ntxqs;
5711 		break;
5712 	case IFLIB_NRXD_HANDLER:
5713 		ndesc = ctx->ifc_sysctl_nrxds;
5714 		if (ctx->ifc_sctx)
5715 			nqs = ctx->ifc_sctx->isc_nrxqs;
5716 		break;
5717 	}
5718 	if (nqs == 0)
5719 		nqs = 8;
5720 
5721 	for (i=0; i<8; i++) {
5722 		if (i >= nqs)
5723 			break;
5724 		if (i)
5725 			strcat(buf, ",");
5726 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5727 	}
5728 
5729 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5730 	if (rc || req->newptr == NULL)
5731 		return rc;
5732 
5733 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5734 	    i++, p = strsep(&next, " ,")) {
5735 		ndesc[i] = strtoul(p, NULL, 10);
5736 	}
5737 
5738 	return(rc);
5739 }
5740 
5741 #define NAME_BUFLEN 32
5742 static void
5743 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5744 {
5745         device_t dev = iflib_get_dev(ctx);
5746 	struct sysctl_oid_list *child, *oid_list;
5747 	struct sysctl_ctx_list *ctx_list;
5748 	struct sysctl_oid *node;
5749 
5750 	ctx_list = device_get_sysctl_ctx(dev);
5751 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5752 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5753 						      CTLFLAG_RD, NULL, "IFLIB fields");
5754 	oid_list = SYSCTL_CHILDREN(node);
5755 
5756 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5757 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5758 		       "driver version");
5759 
5760 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5761 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5762 			"# of txqs to use, 0 => use default #");
5763 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5764 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5765 			"# of rxqs to use, 0 => use default #");
5766 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5767 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5768                        "permit #txq != #rxq");
5769 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5770                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5771                       "disable MSIX (default 0)");
5772 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5773 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5774                        "set the rx budget");
5775 
5776 	/* XXX change for per-queue sizes */
5777 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5778 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5779                        mp_ndesc_handler, "A",
5780                        "list of # of tx descriptors to use, 0 = use default #");
5781 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5782 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5783                        mp_ndesc_handler, "A",
5784                        "list of # of rx descriptors to use, 0 = use default #");
5785 }
5786 
5787 static void
5788 iflib_add_device_sysctl_post(if_ctx_t ctx)
5789 {
5790 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5791 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5792         device_t dev = iflib_get_dev(ctx);
5793 	struct sysctl_oid_list *child;
5794 	struct sysctl_ctx_list *ctx_list;
5795 	iflib_fl_t fl;
5796 	iflib_txq_t txq;
5797 	iflib_rxq_t rxq;
5798 	int i, j;
5799 	char namebuf[NAME_BUFLEN];
5800 	char *qfmt;
5801 	struct sysctl_oid *queue_node, *fl_node, *node;
5802 	struct sysctl_oid_list *queue_list, *fl_list;
5803 	ctx_list = device_get_sysctl_ctx(dev);
5804 
5805 	node = ctx->ifc_sysctl_node;
5806 	child = SYSCTL_CHILDREN(node);
5807 
5808 	if (scctx->isc_ntxqsets > 100)
5809 		qfmt = "txq%03d";
5810 	else if (scctx->isc_ntxqsets > 10)
5811 		qfmt = "txq%02d";
5812 	else
5813 		qfmt = "txq%d";
5814 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5815 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5816 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5817 					     CTLFLAG_RD, NULL, "Queue Name");
5818 		queue_list = SYSCTL_CHILDREN(queue_node);
5819 #if MEMORY_LOGGING
5820 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5821 				CTLFLAG_RD,
5822 				&txq->ift_dequeued, "total mbufs freed");
5823 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5824 				CTLFLAG_RD,
5825 				&txq->ift_enqueued, "total mbufs enqueued");
5826 #endif
5827 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5828 				   CTLFLAG_RD,
5829 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5830 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5831 				   CTLFLAG_RD,
5832 				   &txq->ift_pullups, "# of times m_pullup was called");
5833 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5834 				   CTLFLAG_RD,
5835 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5836 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5837 				   CTLFLAG_RD,
5838 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5839 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5840 				   CTLFLAG_RD,
5841 				   &txq->ift_map_failed, "# of times dma map failed");
5842 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5843 				   CTLFLAG_RD,
5844 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5845 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5846 				   CTLFLAG_RD,
5847 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5848 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5849 				   CTLFLAG_RD,
5850 				   &txq->ift_pidx, 1, "Producer Index");
5851 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5852 				   CTLFLAG_RD,
5853 				   &txq->ift_cidx, 1, "Consumer Index");
5854 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5855 				   CTLFLAG_RD,
5856 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5857 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5858 				   CTLFLAG_RD,
5859 				   &txq->ift_in_use, 1, "descriptors in use");
5860 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5861 				   CTLFLAG_RD,
5862 				   &txq->ift_processed, "descriptors procesed for clean");
5863 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5864 				   CTLFLAG_RD,
5865 				   &txq->ift_cleaned, "total cleaned");
5866 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5867 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5868 				0, mp_ring_state_handler, "A", "soft ring state");
5869 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5870 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5871 				       "# of enqueues to the mp_ring for this queue");
5872 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5873 				       CTLFLAG_RD, &txq->ift_br->drops,
5874 				       "# of drops in the mp_ring for this queue");
5875 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5876 				       CTLFLAG_RD, &txq->ift_br->starts,
5877 				       "# of normal consumer starts in the mp_ring for this queue");
5878 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5879 				       CTLFLAG_RD, &txq->ift_br->stalls,
5880 					       "# of consumer stalls in the mp_ring for this queue");
5881 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5882 			       CTLFLAG_RD, &txq->ift_br->restarts,
5883 				       "# of consumer restarts in the mp_ring for this queue");
5884 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5885 				       CTLFLAG_RD, &txq->ift_br->abdications,
5886 				       "# of consumer abdications in the mp_ring for this queue");
5887 	}
5888 
5889 	if (scctx->isc_nrxqsets > 100)
5890 		qfmt = "rxq%03d";
5891 	else if (scctx->isc_nrxqsets > 10)
5892 		qfmt = "rxq%02d";
5893 	else
5894 		qfmt = "rxq%d";
5895 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5896 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5897 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5898 					     CTLFLAG_RD, NULL, "Queue Name");
5899 		queue_list = SYSCTL_CHILDREN(queue_node);
5900 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5901 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5902 				       CTLFLAG_RD,
5903 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5904 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5905 				       CTLFLAG_RD,
5906 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5907 		}
5908 
5909 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5910 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5911 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5912 						     CTLFLAG_RD, NULL, "freelist Name");
5913 			fl_list = SYSCTL_CHILDREN(fl_node);
5914 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5915 				       CTLFLAG_RD,
5916 				       &fl->ifl_pidx, 1, "Producer Index");
5917 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5918 				       CTLFLAG_RD,
5919 				       &fl->ifl_cidx, 1, "Consumer Index");
5920 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5921 				       CTLFLAG_RD,
5922 				       &fl->ifl_credits, 1, "credits available");
5923 #if MEMORY_LOGGING
5924 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5925 					CTLFLAG_RD,
5926 					&fl->ifl_m_enqueued, "mbufs allocated");
5927 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5928 					CTLFLAG_RD,
5929 					&fl->ifl_m_dequeued, "mbufs freed");
5930 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5931 					CTLFLAG_RD,
5932 					&fl->ifl_cl_enqueued, "clusters allocated");
5933 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5934 					CTLFLAG_RD,
5935 					&fl->ifl_cl_dequeued, "clusters freed");
5936 #endif
5937 
5938 		}
5939 	}
5940 
5941 }
5942 
5943 #ifndef __NO_STRICT_ALIGNMENT
5944 static struct mbuf *
5945 iflib_fixup_rx(struct mbuf *m)
5946 {
5947 	struct mbuf *n;
5948 
5949 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5950 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5951 		m->m_data += ETHER_HDR_LEN;
5952 		n = m;
5953 	} else {
5954 		MGETHDR(n, M_NOWAIT, MT_DATA);
5955 		if (n == NULL) {
5956 			m_freem(m);
5957 			return (NULL);
5958 		}
5959 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5960 		m->m_data += ETHER_HDR_LEN;
5961 		m->m_len -= ETHER_HDR_LEN;
5962 		n->m_len = ETHER_HDR_LEN;
5963 		M_MOVE_PKTHDR(n, m);
5964 		n->m_next = m;
5965 	}
5966 	return (n);
5967 }
5968 #endif
5969