xref: /freebsd/sys/net/iflib.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
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 		uint16_t slot_flags = kring->nkr_slot_flags;
1072 
1073 		for (i = 0; i < rxq->ifr_nfl; i++) {
1074 			fl = &rxq->ifr_fl[i];
1075 			nic_i = fl->ifl_cidx;
1076 			nm_i = netmap_idx_n2k(kring, nic_i);
1077 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1078 			for (n = 0; avail > 0; n++, avail--) {
1079 				rxd_info_zero(&ri);
1080 				ri.iri_frags = rxq->ifr_frags;
1081 				ri.iri_qsidx = kring->ring_id;
1082 				ri.iri_ifp = ctx->ifc_ifp;
1083 				ri.iri_cidx = nic_i;
1084 
1085 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1086 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1087 				ring->slot[nm_i].flags = slot_flags;
1088 				if (fl->ifl_sds.ifsd_map)
1089 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1090 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1091 				nm_i = nm_next(nm_i, lim);
1092 				nic_i = nm_next(nic_i, lim);
1093 			}
1094 			if (n) { /* update the state variables */
1095 				if (netmap_no_pendintr && !force_update) {
1096 					/* diagnostics */
1097 					iflib_rx_miss ++;
1098 					iflib_rx_miss_bufs += n;
1099 				}
1100 				fl->ifl_cidx = nic_i;
1101 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1102 			}
1103 			kring->nr_kflags &= ~NKR_PENDINTR;
1104 		}
1105 	}
1106 	/*
1107 	 * Second part: skip past packets that userspace has released.
1108 	 * (kring->nr_hwcur to head excluded),
1109 	 * and make the buffers available for reception.
1110 	 * As usual nm_i is the index in the netmap ring,
1111 	 * nic_i is the index in the NIC ring, and
1112 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1113 	 */
1114 	/* XXX not sure how this will work with multiple free lists */
1115 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1116 
1117 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1118 }
1119 
1120 static void
1121 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1122 {
1123 	struct ifnet *ifp = na->ifp;
1124 	if_ctx_t ctx = ifp->if_softc;
1125 
1126 	CTX_LOCK(ctx);
1127 	if (onoff) {
1128 		IFDI_INTR_ENABLE(ctx);
1129 	} else {
1130 		IFDI_INTR_DISABLE(ctx);
1131 	}
1132 	CTX_UNLOCK(ctx);
1133 }
1134 
1135 
1136 static int
1137 iflib_netmap_attach(if_ctx_t ctx)
1138 {
1139 	struct netmap_adapter na;
1140 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1141 
1142 	bzero(&na, sizeof(na));
1143 
1144 	na.ifp = ctx->ifc_ifp;
1145 	na.na_flags = NAF_BDG_MAYSLEEP;
1146 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1147 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1148 
1149 	na.num_tx_desc = scctx->isc_ntxd[0];
1150 	na.num_rx_desc = scctx->isc_nrxd[0];
1151 	na.nm_txsync = iflib_netmap_txsync;
1152 	na.nm_rxsync = iflib_netmap_rxsync;
1153 	na.nm_register = iflib_netmap_register;
1154 	na.nm_intr = iflib_netmap_intr;
1155 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1156 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1157 	return (netmap_attach(&na));
1158 }
1159 
1160 static void
1161 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1162 {
1163 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1164 	struct netmap_slot *slot;
1165 
1166 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1167 	if (slot == NULL)
1168 		return;
1169 	if (txq->ift_sds.ifsd_map == NULL)
1170 		return;
1171 
1172 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1173 
1174 		/*
1175 		 * In netmap mode, set the map for the packet buffer.
1176 		 * NOTE: Some drivers (not this one) also need to set
1177 		 * the physical buffer address in the NIC ring.
1178 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1179 		 * netmap slot index, si
1180 		 */
1181 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1182 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1183 	}
1184 }
1185 
1186 static void
1187 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1188 {
1189 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1190 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1191 	struct netmap_slot *slot;
1192 	uint32_t nm_i;
1193 
1194 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1195 	if (slot == NULL)
1196 		return;
1197 	nm_i = netmap_idx_n2k(kring, 0);
1198 	netmap_fl_refill(rxq, kring, nm_i, true);
1199 }
1200 
1201 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1202 
1203 #else
1204 #define iflib_netmap_txq_init(ctx, txq)
1205 #define iflib_netmap_rxq_init(ctx, rxq)
1206 #define iflib_netmap_detach(ifp)
1207 
1208 #define iflib_netmap_attach(ctx) (0)
1209 #define netmap_rx_irq(ifp, qid, budget) (0)
1210 #define netmap_tx_irq(ifp, qid) do {} while (0)
1211 
1212 #endif
1213 
1214 #if defined(__i386__) || defined(__amd64__)
1215 static __inline void
1216 prefetch(void *x)
1217 {
1218 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1219 }
1220 static __inline void
1221 prefetch2cachelines(void *x)
1222 {
1223 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1224 #if (CACHE_LINE_SIZE < 128)
1225 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1226 #endif
1227 }
1228 #else
1229 #define prefetch(x)
1230 #define prefetch2cachelines(x)
1231 #endif
1232 
1233 static void
1234 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1235 {
1236 	iflib_fl_t fl;
1237 
1238 	fl = &rxq->ifr_fl[flid];
1239 	iru->iru_paddrs = fl->ifl_bus_addrs;
1240 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1241 	iru->iru_idxs = fl->ifl_rxd_idxs;
1242 	iru->iru_qsidx = rxq->ifr_id;
1243 	iru->iru_buf_size = fl->ifl_buf_size;
1244 	iru->iru_flidx = fl->ifl_id;
1245 }
1246 
1247 static void
1248 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1249 {
1250 	if (err)
1251 		return;
1252 	*(bus_addr_t *) arg = segs[0].ds_addr;
1253 }
1254 
1255 int
1256 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1257 {
1258 	int err;
1259 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1260 	device_t dev = ctx->ifc_dev;
1261 
1262 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1263 
1264 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1265 				sctx->isc_q_align, 0,	/* alignment, bounds */
1266 				BUS_SPACE_MAXADDR,	/* lowaddr */
1267 				BUS_SPACE_MAXADDR,	/* highaddr */
1268 				NULL, NULL,		/* filter, filterarg */
1269 				size,			/* maxsize */
1270 				1,			/* nsegments */
1271 				size,			/* maxsegsize */
1272 				BUS_DMA_ALLOCNOW,	/* flags */
1273 				NULL,			/* lockfunc */
1274 				NULL,			/* lockarg */
1275 				&dma->idi_tag);
1276 	if (err) {
1277 		device_printf(dev,
1278 		    "%s: bus_dma_tag_create failed: %d\n",
1279 		    __func__, err);
1280 		goto fail_0;
1281 	}
1282 
1283 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1284 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1285 	if (err) {
1286 		device_printf(dev,
1287 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1288 		    __func__, (uintmax_t)size, err);
1289 		goto fail_1;
1290 	}
1291 
1292 	dma->idi_paddr = IF_BAD_DMA;
1293 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1294 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1295 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1296 		device_printf(dev,
1297 		    "%s: bus_dmamap_load failed: %d\n",
1298 		    __func__, err);
1299 		goto fail_2;
1300 	}
1301 
1302 	dma->idi_size = size;
1303 	return (0);
1304 
1305 fail_2:
1306 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1307 fail_1:
1308 	bus_dma_tag_destroy(dma->idi_tag);
1309 fail_0:
1310 	dma->idi_tag = NULL;
1311 
1312 	return (err);
1313 }
1314 
1315 int
1316 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1317 {
1318 	int i, err;
1319 	iflib_dma_info_t *dmaiter;
1320 
1321 	dmaiter = dmalist;
1322 	for (i = 0; i < count; i++, dmaiter++) {
1323 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1324 			break;
1325 	}
1326 	if (err)
1327 		iflib_dma_free_multi(dmalist, i);
1328 	return (err);
1329 }
1330 
1331 void
1332 iflib_dma_free(iflib_dma_info_t dma)
1333 {
1334 	if (dma->idi_tag == NULL)
1335 		return;
1336 	if (dma->idi_paddr != IF_BAD_DMA) {
1337 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1338 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1339 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1340 		dma->idi_paddr = IF_BAD_DMA;
1341 	}
1342 	if (dma->idi_vaddr != NULL) {
1343 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1344 		dma->idi_vaddr = NULL;
1345 	}
1346 	bus_dma_tag_destroy(dma->idi_tag);
1347 	dma->idi_tag = NULL;
1348 }
1349 
1350 void
1351 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1352 {
1353 	int i;
1354 	iflib_dma_info_t *dmaiter = dmalist;
1355 
1356 	for (i = 0; i < count; i++, dmaiter++)
1357 		iflib_dma_free(*dmaiter);
1358 }
1359 
1360 #ifdef EARLY_AP_STARTUP
1361 static const int iflib_started = 1;
1362 #else
1363 /*
1364  * We used to abuse the smp_started flag to decide if the queues have been
1365  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1366  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1367  * is set.  Run a SYSINIT() strictly after that to just set a usable
1368  * completion flag.
1369  */
1370 
1371 static int iflib_started;
1372 
1373 static void
1374 iflib_record_started(void *arg)
1375 {
1376 	iflib_started = 1;
1377 }
1378 
1379 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1380 	iflib_record_started, NULL);
1381 #endif
1382 
1383 static int
1384 iflib_fast_intr(void *arg)
1385 {
1386 	iflib_filter_info_t info = arg;
1387 	struct grouptask *gtask = info->ifi_task;
1388 	if (!iflib_started)
1389 		return (FILTER_HANDLED);
1390 
1391 	DBG_COUNTER_INC(fast_intrs);
1392 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1393 		return (FILTER_HANDLED);
1394 
1395 	GROUPTASK_ENQUEUE(gtask);
1396 	return (FILTER_HANDLED);
1397 }
1398 
1399 static int
1400 iflib_fast_intr_rxtx(void *arg)
1401 {
1402 	iflib_filter_info_t info = arg;
1403 	struct grouptask *gtask = info->ifi_task;
1404 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1405 	if_ctx_t ctx;
1406 	int i, cidx;
1407 
1408 	if (!iflib_started)
1409 		return (FILTER_HANDLED);
1410 
1411 	DBG_COUNTER_INC(fast_intrs);
1412 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1413 		return (FILTER_HANDLED);
1414 
1415 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1416 		qidx_t txqid = rxq->ifr_txqid[i];
1417 
1418 		ctx = rxq->ifr_ctx;
1419 
1420 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1421 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1422 			continue;
1423 		}
1424 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1425 	}
1426 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1427 		cidx = rxq->ifr_cq_cidx;
1428 	else
1429 		cidx = rxq->ifr_fl[0].ifl_cidx;
1430 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1431 		GROUPTASK_ENQUEUE(gtask);
1432 	else
1433 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1434 	return (FILTER_HANDLED);
1435 }
1436 
1437 
1438 static int
1439 iflib_fast_intr_ctx(void *arg)
1440 {
1441 	iflib_filter_info_t info = arg;
1442 	struct grouptask *gtask = info->ifi_task;
1443 
1444 	if (!iflib_started)
1445 		return (FILTER_HANDLED);
1446 
1447 	DBG_COUNTER_INC(fast_intrs);
1448 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1449 		return (FILTER_HANDLED);
1450 
1451 	GROUPTASK_ENQUEUE(gtask);
1452 	return (FILTER_HANDLED);
1453 }
1454 
1455 static int
1456 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1457 	driver_filter_t filter, driver_intr_t handler, void *arg,
1458 				 char *name)
1459 {
1460 	int rc, flags;
1461 	struct resource *res;
1462 	void *tag = NULL;
1463 	device_t dev = ctx->ifc_dev;
1464 
1465 	flags = RF_ACTIVE;
1466 	if (ctx->ifc_flags & IFC_LEGACY)
1467 		flags |= RF_SHAREABLE;
1468 	MPASS(rid < 512);
1469 	irq->ii_rid = rid;
1470 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1471 	if (res == NULL) {
1472 		device_printf(dev,
1473 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1474 		return (ENOMEM);
1475 	}
1476 	irq->ii_res = res;
1477 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1478 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1479 						filter, handler, arg, &tag);
1480 	if (rc != 0) {
1481 		device_printf(dev,
1482 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1483 					  rid, name ? name : "unknown", rc);
1484 		return (rc);
1485 	} else if (name)
1486 		bus_describe_intr(dev, res, tag, "%s", name);
1487 
1488 	irq->ii_tag = tag;
1489 	return (0);
1490 }
1491 
1492 
1493 /*********************************************************************
1494  *
1495  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1496  *  the information needed to transmit a packet on the wire. This is
1497  *  called only once at attach, setup is done every reset.
1498  *
1499  **********************************************************************/
1500 
1501 static int
1502 iflib_txsd_alloc(iflib_txq_t txq)
1503 {
1504 	if_ctx_t ctx = txq->ift_ctx;
1505 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1506 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1507 	device_t dev = ctx->ifc_dev;
1508 	int err, nsegments, ntsosegments;
1509 
1510 	nsegments = scctx->isc_tx_nsegments;
1511 	ntsosegments = scctx->isc_tx_tso_segments_max;
1512 	MPASS(scctx->isc_ntxd[0] > 0);
1513 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1514 	MPASS(nsegments > 0);
1515 	MPASS(ntsosegments > 0);
1516 	/*
1517 	 * Setup DMA descriptor areas.
1518 	 */
1519 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1520 			       1, 0,			/* alignment, bounds */
1521 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1522 			       BUS_SPACE_MAXADDR,	/* highaddr */
1523 			       NULL, NULL,		/* filter, filterarg */
1524 			       sctx->isc_tx_maxsize,		/* maxsize */
1525 			       nsegments,	/* nsegments */
1526 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1527 			       0,			/* flags */
1528 			       NULL,			/* lockfunc */
1529 			       NULL,			/* lockfuncarg */
1530 			       &txq->ift_desc_tag))) {
1531 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1532 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1533 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1534 		goto fail;
1535 	}
1536 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1537 			       1, 0,			/* alignment, bounds */
1538 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1539 			       BUS_SPACE_MAXADDR,	/* highaddr */
1540 			       NULL, NULL,		/* filter, filterarg */
1541 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1542 			       ntsosegments,	/* nsegments */
1543 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1544 			       0,			/* flags */
1545 			       NULL,			/* lockfunc */
1546 			       NULL,			/* lockfuncarg */
1547 			       &txq->ift_tso_desc_tag))) {
1548 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1549 
1550 		goto fail;
1551 	}
1552 	if (!(txq->ift_sds.ifsd_flags =
1553 	    (uint8_t *) malloc(sizeof(uint8_t) *
1554 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1555 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1556 		err = ENOMEM;
1557 		goto fail;
1558 	}
1559 	if (!(txq->ift_sds.ifsd_m =
1560 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1561 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1562 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1563 		err = ENOMEM;
1564 		goto fail;
1565 	}
1566 
1567         /* Create the descriptor buffer dma maps */
1568 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1569 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1570 		return (0);
1571 
1572 	if (!(txq->ift_sds.ifsd_map =
1573 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1574 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1575 		err = ENOMEM;
1576 		goto fail;
1577 	}
1578 
1579 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1580 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1581 		if (err != 0) {
1582 			device_printf(dev, "Unable to create TX DMA map\n");
1583 			goto fail;
1584 		}
1585 	}
1586 #endif
1587 	return (0);
1588 fail:
1589 	/* We free all, it handles case where we are in the middle */
1590 	iflib_tx_structures_free(ctx);
1591 	return (err);
1592 }
1593 
1594 static void
1595 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1596 {
1597 	bus_dmamap_t map;
1598 
1599 	map = NULL;
1600 	if (txq->ift_sds.ifsd_map != NULL)
1601 		map = txq->ift_sds.ifsd_map[i];
1602 	if (map != NULL) {
1603 		bus_dmamap_unload(txq->ift_desc_tag, map);
1604 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1605 		txq->ift_sds.ifsd_map[i] = NULL;
1606 	}
1607 }
1608 
1609 static void
1610 iflib_txq_destroy(iflib_txq_t txq)
1611 {
1612 	if_ctx_t ctx = txq->ift_ctx;
1613 
1614 	for (int i = 0; i < txq->ift_size; i++)
1615 		iflib_txsd_destroy(ctx, txq, i);
1616 	if (txq->ift_sds.ifsd_map != NULL) {
1617 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1618 		txq->ift_sds.ifsd_map = NULL;
1619 	}
1620 	if (txq->ift_sds.ifsd_m != NULL) {
1621 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1622 		txq->ift_sds.ifsd_m = NULL;
1623 	}
1624 	if (txq->ift_sds.ifsd_flags != NULL) {
1625 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1626 		txq->ift_sds.ifsd_flags = NULL;
1627 	}
1628 	if (txq->ift_desc_tag != NULL) {
1629 		bus_dma_tag_destroy(txq->ift_desc_tag);
1630 		txq->ift_desc_tag = NULL;
1631 	}
1632 	if (txq->ift_tso_desc_tag != NULL) {
1633 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1634 		txq->ift_tso_desc_tag = NULL;
1635 	}
1636 }
1637 
1638 static void
1639 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1640 {
1641 	struct mbuf **mp;
1642 
1643 	mp = &txq->ift_sds.ifsd_m[i];
1644 	if (*mp == NULL)
1645 		return;
1646 
1647 	if (txq->ift_sds.ifsd_map != NULL) {
1648 		bus_dmamap_sync(txq->ift_desc_tag,
1649 				txq->ift_sds.ifsd_map[i],
1650 				BUS_DMASYNC_POSTWRITE);
1651 		bus_dmamap_unload(txq->ift_desc_tag,
1652 				  txq->ift_sds.ifsd_map[i]);
1653 	}
1654 	m_free(*mp);
1655 	DBG_COUNTER_INC(tx_frees);
1656 	*mp = NULL;
1657 }
1658 
1659 static int
1660 iflib_txq_setup(iflib_txq_t txq)
1661 {
1662 	if_ctx_t ctx = txq->ift_ctx;
1663 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1664 	iflib_dma_info_t di;
1665 	int i;
1666 
1667 	/* Set number of descriptors available */
1668 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1669 	/* XXX make configurable */
1670 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1671 
1672 	/* Reset indices */
1673 	txq->ift_cidx_processed = 0;
1674 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1675 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1676 
1677 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1678 		bzero((void *)di->idi_vaddr, di->idi_size);
1679 
1680 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1681 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1682 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1683 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1684 	return (0);
1685 }
1686 
1687 /*********************************************************************
1688  *
1689  *  Allocate memory for rx_buffer structures. Since we use one
1690  *  rx_buffer per received packet, the maximum number of rx_buffer's
1691  *  that we'll need is equal to the number of receive descriptors
1692  *  that we've allocated.
1693  *
1694  **********************************************************************/
1695 static int
1696 iflib_rxsd_alloc(iflib_rxq_t rxq)
1697 {
1698 	if_ctx_t ctx = rxq->ifr_ctx;
1699 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1700 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1701 	device_t dev = ctx->ifc_dev;
1702 	iflib_fl_t fl;
1703 	int			err;
1704 
1705 	MPASS(scctx->isc_nrxd[0] > 0);
1706 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1707 
1708 	fl = rxq->ifr_fl;
1709 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1710 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1711 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1712 					 1, 0,			/* alignment, bounds */
1713 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1714 					 BUS_SPACE_MAXADDR,	/* highaddr */
1715 					 NULL, NULL,		/* filter, filterarg */
1716 					 sctx->isc_rx_maxsize,	/* maxsize */
1717 					 sctx->isc_rx_nsegments,	/* nsegments */
1718 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1719 					 0,			/* flags */
1720 					 NULL,			/* lockfunc */
1721 					 NULL,			/* lockarg */
1722 					 &fl->ifl_desc_tag);
1723 		if (err) {
1724 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1725 				__func__, err);
1726 			goto fail;
1727 		}
1728 		if (!(fl->ifl_sds.ifsd_flags =
1729 		      (uint8_t *) malloc(sizeof(uint8_t) *
1730 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1731 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1732 			err = ENOMEM;
1733 			goto fail;
1734 		}
1735 		if (!(fl->ifl_sds.ifsd_m =
1736 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1737 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1738 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1739 			err = ENOMEM;
1740 			goto fail;
1741 		}
1742 		if (!(fl->ifl_sds.ifsd_cl =
1743 		      (caddr_t *) malloc(sizeof(caddr_t) *
1744 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1745 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1746 			err = ENOMEM;
1747 			goto fail;
1748 		}
1749 
1750 		/* Create the descriptor buffer dma maps */
1751 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1752 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1753 			continue;
1754 
1755 		if (!(fl->ifl_sds.ifsd_map =
1756 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1757 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1758 			err = ENOMEM;
1759 			goto fail;
1760 		}
1761 
1762 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1763 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1764 			if (err != 0) {
1765 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1766 				goto fail;
1767 			}
1768 		}
1769 #endif
1770 	}
1771 	return (0);
1772 
1773 fail:
1774 	iflib_rx_structures_free(ctx);
1775 	return (err);
1776 }
1777 
1778 
1779 /*
1780  * Internal service routines
1781  */
1782 
1783 struct rxq_refill_cb_arg {
1784 	int               error;
1785 	bus_dma_segment_t seg;
1786 	int               nseg;
1787 };
1788 
1789 static void
1790 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1791 {
1792 	struct rxq_refill_cb_arg *cb_arg = arg;
1793 
1794 	cb_arg->error = error;
1795 	cb_arg->seg = segs[0];
1796 	cb_arg->nseg = nseg;
1797 }
1798 
1799 
1800 #ifdef ACPI_DMAR
1801 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1802 #else
1803 #define IS_DMAR(ctx) (0)
1804 #endif
1805 
1806 /**
1807  *	rxq_refill - refill an rxq  free-buffer list
1808  *	@ctx: the iflib context
1809  *	@rxq: the free-list to refill
1810  *	@n: the number of new buffers to allocate
1811  *
1812  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1813  *	The caller must assure that @n does not exceed the queue's capacity.
1814  */
1815 static void
1816 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1817 {
1818 	struct mbuf *m;
1819 	int idx, frag_idx = fl->ifl_fragidx;
1820         int pidx = fl->ifl_pidx;
1821 	caddr_t cl, *sd_cl;
1822 	struct mbuf **sd_m;
1823 	uint8_t *sd_flags;
1824 	struct if_rxd_update iru;
1825 	bus_dmamap_t *sd_map;
1826 	int n, i = 0;
1827 	uint64_t bus_addr;
1828 	int err;
1829 	qidx_t credits;
1830 
1831 	sd_m = fl->ifl_sds.ifsd_m;
1832 	sd_map = fl->ifl_sds.ifsd_map;
1833 	sd_cl = fl->ifl_sds.ifsd_cl;
1834 	sd_flags = fl->ifl_sds.ifsd_flags;
1835 	idx = pidx;
1836 	credits = fl->ifl_credits;
1837 
1838 	n  = count;
1839 	MPASS(n > 0);
1840 	MPASS(credits + n <= fl->ifl_size);
1841 
1842 	if (pidx < fl->ifl_cidx)
1843 		MPASS(pidx + n <= fl->ifl_cidx);
1844 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1845 		MPASS(fl->ifl_gen == 0);
1846 	if (pidx > fl->ifl_cidx)
1847 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1848 
1849 	DBG_COUNTER_INC(fl_refills);
1850 	if (n > 8)
1851 		DBG_COUNTER_INC(fl_refills_large);
1852 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1853 	while (n--) {
1854 		/*
1855 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1856 		 * initialized after rx.
1857 		 *
1858 		 * If the cluster is still set then we know a minimum sized packet was received
1859 		 */
1860 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1861 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1862                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1863 		if ((cl = sd_cl[frag_idx]) == NULL) {
1864                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1865 				break;
1866 #if MEMORY_LOGGING
1867 			fl->ifl_cl_enqueued++;
1868 #endif
1869 		}
1870 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1871 			break;
1872 		}
1873 #if MEMORY_LOGGING
1874 		fl->ifl_m_enqueued++;
1875 #endif
1876 
1877 		DBG_COUNTER_INC(rx_allocs);
1878 #if defined(__i386__) || defined(__amd64__)
1879 		if (!IS_DMAR(ctx)) {
1880 			bus_addr = pmap_kextract((vm_offset_t)cl);
1881 		} else
1882 #endif
1883 		{
1884 			struct rxq_refill_cb_arg cb_arg;
1885 			iflib_rxq_t q;
1886 
1887 			cb_arg.error = 0;
1888 			q = fl->ifl_rxq;
1889 			MPASS(sd_map != NULL);
1890 			MPASS(sd_map[frag_idx] != NULL);
1891 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1892 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1893 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1894 					BUS_DMASYNC_PREREAD);
1895 
1896 			if (err != 0 || cb_arg.error) {
1897 				/*
1898 				 * !zone_pack ?
1899 				 */
1900 				if (fl->ifl_zone == zone_pack)
1901 					uma_zfree(fl->ifl_zone, cl);
1902 				m_free(m);
1903 				n = 0;
1904 				goto done;
1905 			}
1906 			bus_addr = cb_arg.seg.ds_addr;
1907 		}
1908                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1909 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1910 
1911 		MPASS(sd_m[frag_idx] == NULL);
1912 		sd_cl[frag_idx] = cl;
1913 		sd_m[frag_idx] = m;
1914 		fl->ifl_rxd_idxs[i] = frag_idx;
1915 		fl->ifl_bus_addrs[i] = bus_addr;
1916 		fl->ifl_vm_addrs[i] = cl;
1917 		credits++;
1918 		i++;
1919 		MPASS(credits <= fl->ifl_size);
1920 		if (++idx == fl->ifl_size) {
1921 			fl->ifl_gen = 1;
1922 			idx = 0;
1923 		}
1924 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1925 			iru.iru_pidx = pidx;
1926 			iru.iru_count = i;
1927 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1928 			i = 0;
1929 			pidx = idx;
1930 			fl->ifl_pidx = idx;
1931 			fl->ifl_credits = credits;
1932 		}
1933 
1934 	}
1935 done:
1936 	if (i) {
1937 		iru.iru_pidx = pidx;
1938 		iru.iru_count = i;
1939 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1940 		fl->ifl_pidx = idx;
1941 		fl->ifl_credits = credits;
1942 	}
1943 	DBG_COUNTER_INC(rxd_flush);
1944 	if (fl->ifl_pidx == 0)
1945 		pidx = fl->ifl_size - 1;
1946 	else
1947 		pidx = fl->ifl_pidx - 1;
1948 
1949 	if (sd_map)
1950 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1951 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1952 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1953 	fl->ifl_fragidx = frag_idx;
1954 }
1955 
1956 static __inline void
1957 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1958 {
1959 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1960 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1961 #ifdef INVARIANTS
1962 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1963 #endif
1964 
1965 	MPASS(fl->ifl_credits <= fl->ifl_size);
1966 	MPASS(reclaimable == delta);
1967 
1968 	if (reclaimable > 0)
1969 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1970 }
1971 
1972 static void
1973 iflib_fl_bufs_free(iflib_fl_t fl)
1974 {
1975 	iflib_dma_info_t idi = fl->ifl_ifdi;
1976 	uint32_t i;
1977 
1978 	for (i = 0; i < fl->ifl_size; i++) {
1979 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1980 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1981 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1982 
1983 		if (*sd_flags & RX_SW_DESC_INUSE) {
1984 			if (fl->ifl_sds.ifsd_map != NULL) {
1985 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1986 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
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 		/* clean any enqueued buffers */
2271 		iflib_ifmp_purge(txq);
2272 		/* Free any existing tx buffers. */
2273 		for (j = 0; j < txq->ift_size; j++) {
2274 			iflib_txsd_free(ctx, txq, j);
2275 		}
2276 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2277 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2278 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2279 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2280 		txq->ift_pullups = 0;
2281 		ifmp_ring_reset_stats(txq->ift_br);
2282 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2283 			bzero((void *)di->idi_vaddr, di->idi_size);
2284 	}
2285 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2286 		/* make sure all transmitters have completed before proceeding XXX */
2287 
2288 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2289 			bzero((void *)di->idi_vaddr, di->idi_size);
2290 		/* also resets the free lists pidx/cidx */
2291 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2292 			iflib_fl_bufs_free(fl);
2293 	}
2294 }
2295 
2296 static inline caddr_t
2297 calc_next_rxd(iflib_fl_t fl, int cidx)
2298 {
2299 	qidx_t size;
2300 	int nrxd;
2301 	caddr_t start, end, cur, next;
2302 
2303 	nrxd = fl->ifl_size;
2304 	size = fl->ifl_rxd_size;
2305 	start = fl->ifl_ifdi->idi_vaddr;
2306 
2307 	if (__predict_false(size == 0))
2308 		return (start);
2309 	cur = start + size*cidx;
2310 	end = start + size*nrxd;
2311 	next = CACHE_PTR_NEXT(cur);
2312 	return (next < end ? next : start);
2313 }
2314 
2315 static inline void
2316 prefetch_pkts(iflib_fl_t fl, int cidx)
2317 {
2318 	int nextptr;
2319 	int nrxd = fl->ifl_size;
2320 	caddr_t next_rxd;
2321 
2322 
2323 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2324 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2325 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2326 	next_rxd = calc_next_rxd(fl, cidx);
2327 	prefetch(next_rxd);
2328 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2329 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2330 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2331 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2332 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2333 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2334 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2335 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2336 }
2337 
2338 static void
2339 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2340 {
2341 	int flid, cidx;
2342 	bus_dmamap_t map;
2343 	iflib_fl_t fl;
2344 	iflib_dma_info_t di;
2345 	int next;
2346 
2347 	map = NULL;
2348 	flid = irf->irf_flid;
2349 	cidx = irf->irf_idx;
2350 	fl = &rxq->ifr_fl[flid];
2351 	sd->ifsd_fl = fl;
2352 	sd->ifsd_cidx = cidx;
2353 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2354 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2355 	fl->ifl_credits--;
2356 #if MEMORY_LOGGING
2357 	fl->ifl_m_dequeued++;
2358 #endif
2359 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2360 		prefetch_pkts(fl, cidx);
2361 	if (fl->ifl_sds.ifsd_map != NULL) {
2362 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2363 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2364 		map = fl->ifl_sds.ifsd_map[cidx];
2365 		di = fl->ifl_ifdi;
2366 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2367 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2368 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2369 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2370 
2371 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2372 		MPASS(fl->ifl_cidx == cidx);
2373 		if (unload)
2374 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2375 	}
2376 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2377 	if (__predict_false(fl->ifl_cidx == 0))
2378 		fl->ifl_gen = 0;
2379 	if (map != NULL)
2380 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2381 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2382         bit_clear(fl->ifl_rx_bitmap, cidx);
2383 }
2384 
2385 static struct mbuf *
2386 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2387 {
2388 	int i, padlen , flags;
2389 	struct mbuf *m, *mh, *mt;
2390 	caddr_t cl;
2391 
2392 	i = 0;
2393 	mh = NULL;
2394 	do {
2395 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2396 
2397 		MPASS(*sd->ifsd_cl != NULL);
2398 		MPASS(*sd->ifsd_m != NULL);
2399 
2400 		/* Don't include zero-length frags */
2401 		if (ri->iri_frags[i].irf_len == 0) {
2402 			/* XXX we can save the cluster here, but not the mbuf */
2403 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2404 			m_free(*sd->ifsd_m);
2405 			*sd->ifsd_m = NULL;
2406 			continue;
2407 		}
2408 		m = *sd->ifsd_m;
2409 		*sd->ifsd_m = NULL;
2410 		if (mh == NULL) {
2411 			flags = M_PKTHDR|M_EXT;
2412 			mh = mt = m;
2413 			padlen = ri->iri_pad;
2414 		} else {
2415 			flags = M_EXT;
2416 			mt->m_next = m;
2417 			mt = m;
2418 			/* assuming padding is only on the first fragment */
2419 			padlen = 0;
2420 		}
2421 		cl = *sd->ifsd_cl;
2422 		*sd->ifsd_cl = NULL;
2423 
2424 		/* Can these two be made one ? */
2425 		m_init(m, M_NOWAIT, MT_DATA, flags);
2426 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2427 		/*
2428 		 * These must follow m_init and m_cljset
2429 		 */
2430 		m->m_data += padlen;
2431 		ri->iri_len -= padlen;
2432 		m->m_len = ri->iri_frags[i].irf_len;
2433 	} while (++i < ri->iri_nfrags);
2434 
2435 	return (mh);
2436 }
2437 
2438 /*
2439  * Process one software descriptor
2440  */
2441 static struct mbuf *
2442 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2443 {
2444 	struct if_rxsd sd;
2445 	struct mbuf *m;
2446 
2447 	/* should I merge this back in now that the two paths are basically duplicated? */
2448 	if (ri->iri_nfrags == 1 &&
2449 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2450 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2451 		m = *sd.ifsd_m;
2452 		*sd.ifsd_m = NULL;
2453 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2454 #ifndef __NO_STRICT_ALIGNMENT
2455 		if (!IP_ALIGNED(m))
2456 			m->m_data += 2;
2457 #endif
2458 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2459 		m->m_len = ri->iri_frags[0].irf_len;
2460        } else {
2461 		m = assemble_segments(rxq, ri, &sd);
2462 	}
2463 	m->m_pkthdr.len = ri->iri_len;
2464 	m->m_pkthdr.rcvif = ri->iri_ifp;
2465 	m->m_flags |= ri->iri_flags;
2466 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2467 	m->m_pkthdr.flowid = ri->iri_flowid;
2468 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2469 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2470 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2471 	return (m);
2472 }
2473 
2474 #if defined(INET6) || defined(INET)
2475 static void
2476 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2477 {
2478 	CURVNET_SET(lc->ifp->if_vnet);
2479 #if defined(INET6)
2480 	*v6 = VNET(ip6_forwarding);
2481 #endif
2482 #if defined(INET)
2483 	*v4 = VNET(ipforwarding);
2484 #endif
2485 	CURVNET_RESTORE();
2486 }
2487 
2488 /*
2489  * Returns true if it's possible this packet could be LROed.
2490  * if it returns false, it is guaranteed that tcp_lro_rx()
2491  * would not return zero.
2492  */
2493 static bool
2494 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2495 {
2496 	struct ether_header *eh;
2497 	uint16_t eh_type;
2498 
2499 	eh = mtod(m, struct ether_header *);
2500 	eh_type = ntohs(eh->ether_type);
2501 	switch (eh_type) {
2502 #if defined(INET6)
2503 		case ETHERTYPE_IPV6:
2504 			return !v6_forwarding;
2505 #endif
2506 #if defined (INET)
2507 		case ETHERTYPE_IP:
2508 			return !v4_forwarding;
2509 #endif
2510 	}
2511 
2512 	return false;
2513 }
2514 #else
2515 static void
2516 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2517 {
2518 }
2519 #endif
2520 
2521 static bool
2522 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2523 {
2524 	if_ctx_t ctx = rxq->ifr_ctx;
2525 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2526 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2527 	int avail, i;
2528 	qidx_t *cidxp;
2529 	struct if_rxd_info ri;
2530 	int err, budget_left, rx_bytes, rx_pkts;
2531 	iflib_fl_t fl;
2532 	struct ifnet *ifp;
2533 	int lro_enabled;
2534 	bool lro_possible = false;
2535 	bool v4_forwarding, v6_forwarding;
2536 
2537 	/*
2538 	 * XXX early demux data packets so that if_input processing only handles
2539 	 * acks in interrupt context
2540 	 */
2541 	struct mbuf *m, *mh, *mt, *mf;
2542 
2543 	ifp = ctx->ifc_ifp;
2544 	mh = mt = NULL;
2545 	MPASS(budget > 0);
2546 	rx_pkts	= rx_bytes = 0;
2547 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2548 		cidxp = &rxq->ifr_cq_cidx;
2549 	else
2550 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2551 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2552 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2553 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2554 		DBG_COUNTER_INC(rx_unavail);
2555 		return (false);
2556 	}
2557 
2558 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2559 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2560 			DBG_COUNTER_INC(rx_ctx_inactive);
2561 			break;
2562 		}
2563 		/*
2564 		 * Reset client set fields to their default values
2565 		 */
2566 		rxd_info_zero(&ri);
2567 		ri.iri_qsidx = rxq->ifr_id;
2568 		ri.iri_cidx = *cidxp;
2569 		ri.iri_ifp = ifp;
2570 		ri.iri_frags = rxq->ifr_frags;
2571 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2572 
2573 		if (err)
2574 			goto err;
2575 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2576 			*cidxp = ri.iri_cidx;
2577 			/* Update our consumer index */
2578 			/* XXX NB: shurd - check if this is still safe */
2579 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2580 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2581 				rxq->ifr_cq_gen = 0;
2582 			}
2583 			/* was this only a completion queue message? */
2584 			if (__predict_false(ri.iri_nfrags == 0))
2585 				continue;
2586 		}
2587 		MPASS(ri.iri_nfrags != 0);
2588 		MPASS(ri.iri_len != 0);
2589 
2590 		/* will advance the cidx on the corresponding free lists */
2591 		m = iflib_rxd_pkt_get(rxq, &ri);
2592 		if (avail == 0 && budget_left)
2593 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2594 
2595 		if (__predict_false(m == NULL)) {
2596 			DBG_COUNTER_INC(rx_mbuf_null);
2597 			continue;
2598 		}
2599 		/* imm_pkt: -- cxgb */
2600 		if (mh == NULL)
2601 			mh = mt = m;
2602 		else {
2603 			mt->m_nextpkt = m;
2604 			mt = m;
2605 		}
2606 	}
2607 	/* make sure that we can refill faster than drain */
2608 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2609 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2610 
2611 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2612 	if (lro_enabled)
2613 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2614 	mt = mf = NULL;
2615 	while (mh != NULL) {
2616 		m = mh;
2617 		mh = mh->m_nextpkt;
2618 		m->m_nextpkt = NULL;
2619 #ifndef __NO_STRICT_ALIGNMENT
2620 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2621 			continue;
2622 #endif
2623 		rx_bytes += m->m_pkthdr.len;
2624 		rx_pkts++;
2625 #if defined(INET6) || defined(INET)
2626 		if (lro_enabled) {
2627 			if (!lro_possible) {
2628 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2629 				if (lro_possible && mf != NULL) {
2630 					ifp->if_input(ifp, mf);
2631 					DBG_COUNTER_INC(rx_if_input);
2632 					mt = mf = NULL;
2633 				}
2634 			}
2635 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2636 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2637 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2638 					continue;
2639 			}
2640 		}
2641 #endif
2642 		if (lro_possible) {
2643 			ifp->if_input(ifp, m);
2644 			DBG_COUNTER_INC(rx_if_input);
2645 			continue;
2646 		}
2647 
2648 		if (mf == NULL)
2649 			mf = m;
2650 		if (mt != NULL)
2651 			mt->m_nextpkt = m;
2652 		mt = m;
2653 	}
2654 	if (mf != NULL) {
2655 		ifp->if_input(ifp, mf);
2656 		DBG_COUNTER_INC(rx_if_input);
2657 	}
2658 
2659 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2660 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2661 
2662 	/*
2663 	 * Flush any outstanding LRO work
2664 	 */
2665 #if defined(INET6) || defined(INET)
2666 	tcp_lro_flush_all(&rxq->ifr_lc);
2667 #endif
2668 	if (avail)
2669 		return true;
2670 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2671 err:
2672 	CTX_LOCK(ctx);
2673 	ctx->ifc_flags |= IFC_DO_RESET;
2674 	iflib_admin_intr_deferred(ctx);
2675 	CTX_UNLOCK(ctx);
2676 	return (false);
2677 }
2678 
2679 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2680 static inline qidx_t
2681 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2682 {
2683 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2684 	qidx_t minthresh = txq->ift_size / 8;
2685 	if (in_use > 4*minthresh)
2686 		return (notify_count);
2687 	if (in_use > 2*minthresh)
2688 		return (notify_count >> 1);
2689 	if (in_use > minthresh)
2690 		return (notify_count >> 3);
2691 	return (0);
2692 }
2693 
2694 static inline qidx_t
2695 txq_max_rs_deferred(iflib_txq_t txq)
2696 {
2697 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2698 	qidx_t minthresh = txq->ift_size / 8;
2699 	if (txq->ift_in_use > 4*minthresh)
2700 		return (notify_count);
2701 	if (txq->ift_in_use > 2*minthresh)
2702 		return (notify_count >> 1);
2703 	if (txq->ift_in_use > minthresh)
2704 		return (notify_count >> 2);
2705 	return (2);
2706 }
2707 
2708 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2709 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2710 
2711 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2712 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2713 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2714 
2715 /* forward compatibility for cxgb */
2716 #define FIRST_QSET(ctx) 0
2717 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2718 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2719 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2720 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2721 
2722 /* XXX we should be setting this to something other than zero */
2723 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2724 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2725 
2726 static inline bool
2727 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2728 {
2729 	qidx_t dbval, max;
2730 	bool rang;
2731 
2732 	rang = false;
2733 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2734 	if (ring || txq->ift_db_pending >= max) {
2735 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2736 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2737 		txq->ift_db_pending = txq->ift_npending = 0;
2738 		rang = true;
2739 	}
2740 	return (rang);
2741 }
2742 
2743 #ifdef PKT_DEBUG
2744 static void
2745 print_pkt(if_pkt_info_t pi)
2746 {
2747 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2748 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2749 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2750 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2751 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2752 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2753 }
2754 #endif
2755 
2756 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2757 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2758 
2759 static int
2760 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2761 {
2762 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2763 	struct ether_vlan_header *eh;
2764 	struct mbuf *m, *n;
2765 
2766 	n = m = *mp;
2767 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2768 	    M_WRITABLE(m) == 0) {
2769 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2770 			return (ENOMEM);
2771 		} else {
2772 			m_freem(*mp);
2773 			n = *mp = m;
2774 		}
2775 	}
2776 
2777 	/*
2778 	 * Determine where frame payload starts.
2779 	 * Jump over vlan headers if already present,
2780 	 * helpful for QinQ too.
2781 	 */
2782 	if (__predict_false(m->m_len < sizeof(*eh))) {
2783 		txq->ift_pullups++;
2784 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2785 			return (ENOMEM);
2786 	}
2787 	eh = mtod(m, struct ether_vlan_header *);
2788 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2789 		pi->ipi_etype = ntohs(eh->evl_proto);
2790 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2791 	} else {
2792 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2793 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2794 	}
2795 
2796 	switch (pi->ipi_etype) {
2797 #ifdef INET
2798 	case ETHERTYPE_IP:
2799 	{
2800 		struct ip *ip = NULL;
2801 		struct tcphdr *th = NULL;
2802 		int minthlen;
2803 
2804 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2805 		if (__predict_false(m->m_len < minthlen)) {
2806 			/*
2807 			 * if this code bloat is causing too much of a hit
2808 			 * move it to a separate function and mark it noinline
2809 			 */
2810 			if (m->m_len == pi->ipi_ehdrlen) {
2811 				n = m->m_next;
2812 				MPASS(n);
2813 				if (n->m_len >= sizeof(*ip))  {
2814 					ip = (struct ip *)n->m_data;
2815 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2816 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2817 				} else {
2818 					txq->ift_pullups++;
2819 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2820 						return (ENOMEM);
2821 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2822 				}
2823 			} else {
2824 				txq->ift_pullups++;
2825 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2826 					return (ENOMEM);
2827 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2828 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2829 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2830 			}
2831 		} else {
2832 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2833 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2834 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2835 		}
2836 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2837 		pi->ipi_ipproto = ip->ip_p;
2838 		pi->ipi_flags |= IPI_TX_IPV4;
2839 
2840 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2841                        ip->ip_sum = 0;
2842 
2843 		if (IS_TSO4(pi)) {
2844 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2845 				if (__predict_false(th == NULL)) {
2846 					txq->ift_pullups++;
2847 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2848 						return (ENOMEM);
2849 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2850 				}
2851 				pi->ipi_tcp_hflags = th->th_flags;
2852 				pi->ipi_tcp_hlen = th->th_off << 2;
2853 				pi->ipi_tcp_seq = th->th_seq;
2854 			}
2855 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2856 				return (ENXIO);
2857 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2858 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2859 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2860 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2861 				ip->ip_sum = 0;
2862 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2863 			}
2864 		}
2865 		break;
2866 	}
2867 #endif
2868 #ifdef INET6
2869 	case ETHERTYPE_IPV6:
2870 	{
2871 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2872 		struct tcphdr *th;
2873 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2874 
2875 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2876 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2877 				return (ENOMEM);
2878 		}
2879 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2880 
2881 		/* XXX-BZ this will go badly in case of ext hdrs. */
2882 		pi->ipi_ipproto = ip6->ip6_nxt;
2883 		pi->ipi_flags |= IPI_TX_IPV6;
2884 
2885 		if (IS_TSO6(pi)) {
2886 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2887 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2888 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2889 						return (ENOMEM);
2890 				}
2891 				pi->ipi_tcp_hflags = th->th_flags;
2892 				pi->ipi_tcp_hlen = th->th_off << 2;
2893 			}
2894 
2895 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2896 				return (ENXIO);
2897 			/*
2898 			 * The corresponding flag is set by the stack in the IPv4
2899 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2900 			 * So, set it here because the rest of the flow requires it.
2901 			 */
2902 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2903 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2904 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2905 		}
2906 		break;
2907 	}
2908 #endif
2909 	default:
2910 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2911 		pi->ipi_ip_hlen = 0;
2912 		break;
2913 	}
2914 	*mp = m;
2915 
2916 	return (0);
2917 }
2918 
2919 static  __noinline  struct mbuf *
2920 collapse_pkthdr(struct mbuf *m0)
2921 {
2922 	struct mbuf *m, *m_next, *tmp;
2923 
2924 	m = m0;
2925 	m_next = m->m_next;
2926 	while (m_next != NULL && m_next->m_len == 0) {
2927 		m = m_next;
2928 		m->m_next = NULL;
2929 		m_free(m);
2930 		m_next = m_next->m_next;
2931 	}
2932 	m = m0;
2933 	m->m_next = m_next;
2934 	if ((m_next->m_flags & M_EXT) == 0) {
2935 		m = m_defrag(m, M_NOWAIT);
2936 	} else {
2937 		tmp = m_next->m_next;
2938 		memcpy(m_next, m, MPKTHSIZE);
2939 		m = m_next;
2940 		m->m_next = tmp;
2941 	}
2942 	return (m);
2943 }
2944 
2945 /*
2946  * If dodgy hardware rejects the scatter gather chain we've handed it
2947  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2948  * m_defrag'd mbufs
2949  */
2950 static __noinline struct mbuf *
2951 iflib_remove_mbuf(iflib_txq_t txq)
2952 {
2953 	int ntxd, i, pidx;
2954 	struct mbuf *m, *mh, **ifsd_m;
2955 
2956 	pidx = txq->ift_pidx;
2957 	ifsd_m = txq->ift_sds.ifsd_m;
2958 	ntxd = txq->ift_size;
2959 	mh = m = ifsd_m[pidx];
2960 	ifsd_m[pidx] = NULL;
2961 #if MEMORY_LOGGING
2962 	txq->ift_dequeued++;
2963 #endif
2964 	i = 1;
2965 
2966 	while (m) {
2967 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2968 #if MEMORY_LOGGING
2969 		txq->ift_dequeued++;
2970 #endif
2971 		m = m->m_next;
2972 		i++;
2973 	}
2974 	return (mh);
2975 }
2976 
2977 static int
2978 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2979 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2980 			  int max_segs, int flags)
2981 {
2982 	if_ctx_t ctx;
2983 	if_shared_ctx_t		sctx;
2984 	if_softc_ctx_t		scctx;
2985 	int i, next, pidx, err, ntxd, count;
2986 	struct mbuf *m, *tmp, **ifsd_m;
2987 
2988 	m = *m0;
2989 
2990 	/*
2991 	 * Please don't ever do this
2992 	 */
2993 	if (__predict_false(m->m_len == 0))
2994 		*m0 = m = collapse_pkthdr(m);
2995 
2996 	ctx = txq->ift_ctx;
2997 	sctx = ctx->ifc_sctx;
2998 	scctx = &ctx->ifc_softc_ctx;
2999 	ifsd_m = txq->ift_sds.ifsd_m;
3000 	ntxd = txq->ift_size;
3001 	pidx = txq->ift_pidx;
3002 	if (map != NULL) {
3003 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3004 
3005 		err = bus_dmamap_load_mbuf_sg(tag, map,
3006 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3007 		if (err)
3008 			return (err);
3009 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3010 		count = 0;
3011 		m = *m0;
3012 		do {
3013 			if (__predict_false(m->m_len <= 0)) {
3014 				tmp = m;
3015 				m = m->m_next;
3016 				tmp->m_next = NULL;
3017 				m_free(tmp);
3018 				continue;
3019 			}
3020 			m = m->m_next;
3021 			count++;
3022 		} while (m != NULL);
3023 		if (count > *nsegs) {
3024 			ifsd_m[pidx] = *m0;
3025 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3026 			return (0);
3027 		}
3028 		m = *m0;
3029 		count = 0;
3030 		do {
3031 			next = (pidx + count) & (ntxd-1);
3032 			MPASS(ifsd_m[next] == NULL);
3033 			ifsd_m[next] = m;
3034 			count++;
3035 			tmp = m;
3036 			m = m->m_next;
3037 		} while (m != NULL);
3038 	} else {
3039 		int buflen, sgsize, maxsegsz, max_sgsize;
3040 		vm_offset_t vaddr;
3041 		vm_paddr_t curaddr;
3042 
3043 		count = i = 0;
3044 		m = *m0;
3045 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3046 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3047 		else
3048 			maxsegsz = sctx->isc_tx_maxsegsize;
3049 
3050 		do {
3051 			if (__predict_false(m->m_len <= 0)) {
3052 				tmp = m;
3053 				m = m->m_next;
3054 				tmp->m_next = NULL;
3055 				m_free(tmp);
3056 				continue;
3057 			}
3058 			buflen = m->m_len;
3059 			vaddr = (vm_offset_t)m->m_data;
3060 			/*
3061 			 * see if we can't be smarter about physically
3062 			 * contiguous mappings
3063 			 */
3064 			next = (pidx + count) & (ntxd-1);
3065 			MPASS(ifsd_m[next] == NULL);
3066 #if MEMORY_LOGGING
3067 			txq->ift_enqueued++;
3068 #endif
3069 			ifsd_m[next] = m;
3070 			while (buflen > 0) {
3071 				if (i >= max_segs)
3072 					goto err;
3073 				max_sgsize = MIN(buflen, maxsegsz);
3074 				curaddr = pmap_kextract(vaddr);
3075 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3076 				sgsize = MIN(sgsize, max_sgsize);
3077 				segs[i].ds_addr = curaddr;
3078 				segs[i].ds_len = sgsize;
3079 				vaddr += sgsize;
3080 				buflen -= sgsize;
3081 				i++;
3082 			}
3083 			count++;
3084 			tmp = m;
3085 			m = m->m_next;
3086 		} while (m != NULL);
3087 		*nsegs = i;
3088 	}
3089 	return (0);
3090 err:
3091 	*m0 = iflib_remove_mbuf(txq);
3092 	return (EFBIG);
3093 }
3094 
3095 static inline caddr_t
3096 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3097 {
3098 	qidx_t size;
3099 	int ntxd;
3100 	caddr_t start, end, cur, next;
3101 
3102 	ntxd = txq->ift_size;
3103 	size = txq->ift_txd_size[qid];
3104 	start = txq->ift_ifdi[qid].idi_vaddr;
3105 
3106 	if (__predict_false(size == 0))
3107 		return (start);
3108 	cur = start + size*cidx;
3109 	end = start + size*ntxd;
3110 	next = CACHE_PTR_NEXT(cur);
3111 	return (next < end ? next : start);
3112 }
3113 
3114 /*
3115  * Pad an mbuf to ensure a minimum ethernet frame size.
3116  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3117  */
3118 static __noinline int
3119 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3120 {
3121 	/*
3122 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3123 	 * and ARP message is the smallest common payload I can think of
3124 	 */
3125 	static char pad[18];	/* just zeros */
3126 	int n;
3127 	struct mbuf *new_head;
3128 
3129 	if (!M_WRITABLE(*m_head)) {
3130 		new_head = m_dup(*m_head, M_NOWAIT);
3131 		if (new_head == NULL) {
3132 			m_freem(*m_head);
3133 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3134 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3135 			return ENOMEM;
3136 		}
3137 		m_freem(*m_head);
3138 		*m_head = new_head;
3139 	}
3140 
3141 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3142 	     n > 0; n -= sizeof(pad))
3143 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3144 			break;
3145 
3146 	if (n > 0) {
3147 		m_freem(*m_head);
3148 		device_printf(dev, "cannot pad short frame\n");
3149 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3150 		return (ENOBUFS);
3151 	}
3152 
3153 	return 0;
3154 }
3155 
3156 static int
3157 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3158 {
3159 	if_ctx_t		ctx;
3160 	if_shared_ctx_t		sctx;
3161 	if_softc_ctx_t		scctx;
3162 	bus_dma_segment_t	*segs;
3163 	struct mbuf		*m_head;
3164 	void			*next_txd;
3165 	bus_dmamap_t		map;
3166 	struct if_pkt_info	pi;
3167 	int remap = 0;
3168 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3169 	bus_dma_tag_t desc_tag;
3170 
3171 	segs = txq->ift_segs;
3172 	ctx = txq->ift_ctx;
3173 	sctx = ctx->ifc_sctx;
3174 	scctx = &ctx->ifc_softc_ctx;
3175 	segs = txq->ift_segs;
3176 	ntxd = txq->ift_size;
3177 	m_head = *m_headp;
3178 	map = NULL;
3179 
3180 	/*
3181 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3182 	 */
3183 	cidx = txq->ift_cidx;
3184 	pidx = txq->ift_pidx;
3185 	if (ctx->ifc_flags & IFC_PREFETCH) {
3186 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3187 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3188 			next_txd = calc_next_txd(txq, cidx, 0);
3189 			prefetch(next_txd);
3190 		}
3191 
3192 		/* prefetch the next cache line of mbuf pointers and flags */
3193 		prefetch(&txq->ift_sds.ifsd_m[next]);
3194 		if (txq->ift_sds.ifsd_map != NULL) {
3195 			prefetch(&txq->ift_sds.ifsd_map[next]);
3196 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3197 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3198 		}
3199 	} else if (txq->ift_sds.ifsd_map != NULL)
3200 		map = txq->ift_sds.ifsd_map[pidx];
3201 
3202 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3203 		desc_tag = txq->ift_tso_desc_tag;
3204 		max_segs = scctx->isc_tx_tso_segments_max;
3205 	} else {
3206 		desc_tag = txq->ift_desc_tag;
3207 		max_segs = scctx->isc_tx_nsegments;
3208 	}
3209 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3210 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3211 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3212 		if (err)
3213 			return err;
3214 	}
3215 	m_head = *m_headp;
3216 
3217 	pkt_info_zero(&pi);
3218 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3219 	pi.ipi_pidx = pidx;
3220 	pi.ipi_qsidx = txq->ift_id;
3221 	pi.ipi_len = m_head->m_pkthdr.len;
3222 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3223 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3224 
3225 	/* deliberate bitwise OR to make one condition */
3226 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3227 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3228 			return (err);
3229 		m_head = *m_headp;
3230 	}
3231 
3232 retry:
3233 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3234 defrag:
3235 	if (__predict_false(err)) {
3236 		switch (err) {
3237 		case EFBIG:
3238 			/* try collapse once and defrag once */
3239 			if (remap == 0)
3240 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3241 			if (remap == 1)
3242 				m_head = m_defrag(*m_headp, M_NOWAIT);
3243 			remap++;
3244 			if (__predict_false(m_head == NULL))
3245 				goto defrag_failed;
3246 			txq->ift_mbuf_defrag++;
3247 			*m_headp = m_head;
3248 			goto retry;
3249 			break;
3250 		case ENOMEM:
3251 			txq->ift_no_tx_dma_setup++;
3252 			break;
3253 		default:
3254 			txq->ift_no_tx_dma_setup++;
3255 			m_freem(*m_headp);
3256 			DBG_COUNTER_INC(tx_frees);
3257 			*m_headp = NULL;
3258 			break;
3259 		}
3260 		txq->ift_map_failed++;
3261 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3262 		return (err);
3263 	}
3264 
3265 	/*
3266 	 * XXX assumes a 1 to 1 relationship between segments and
3267 	 *        descriptors - this does not hold true on all drivers, e.g.
3268 	 *        cxgb
3269 	 */
3270 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3271 		txq->ift_no_desc_avail++;
3272 		if (map != NULL)
3273 			bus_dmamap_unload(desc_tag, map);
3274 		DBG_COUNTER_INC(encap_txq_avail_fail);
3275 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3276 			GROUPTASK_ENQUEUE(&txq->ift_task);
3277 		return (ENOBUFS);
3278 	}
3279 	/*
3280 	 * On Intel cards we can greatly reduce the number of TX interrupts
3281 	 * we see by only setting report status on every Nth descriptor.
3282 	 * However, this also means that the driver will need to keep track
3283 	 * of the descriptors that RS was set on to check them for the DD bit.
3284 	 */
3285 	txq->ift_rs_pending += nsegs + 1;
3286 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3287 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3288 		pi.ipi_flags |= IPI_TX_INTR;
3289 		txq->ift_rs_pending = 0;
3290 	}
3291 
3292 	pi.ipi_segs = segs;
3293 	pi.ipi_nsegs = nsegs;
3294 
3295 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3296 #ifdef PKT_DEBUG
3297 	print_pkt(&pi);
3298 #endif
3299 	if (map != NULL)
3300 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3301 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3302 		if (map != NULL)
3303 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3304 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3305 		DBG_COUNTER_INC(tx_encap);
3306 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3307 
3308 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3309 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3310 			ndesc += txq->ift_size;
3311 			txq->ift_gen = 1;
3312 		}
3313 		/*
3314 		 * drivers can need as many as
3315 		 * two sentinels
3316 		 */
3317 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3318 		MPASS(pi.ipi_new_pidx != pidx);
3319 		MPASS(ndesc > 0);
3320 		txq->ift_in_use += ndesc;
3321 
3322 		/*
3323 		 * We update the last software descriptor again here because there may
3324 		 * be a sentinel and/or there may be more mbufs than segments
3325 		 */
3326 		txq->ift_pidx = pi.ipi_new_pidx;
3327 		txq->ift_npending += pi.ipi_ndescs;
3328 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3329 		*m_headp = m_head = iflib_remove_mbuf(txq);
3330 		remap = 1;
3331 		txq->ift_txd_encap_efbig++;
3332 		goto defrag;
3333 	} else
3334 		DBG_COUNTER_INC(encap_txd_encap_fail);
3335 	return (err);
3336 
3337 defrag_failed:
3338 	txq->ift_mbuf_defrag_failed++;
3339 	txq->ift_map_failed++;
3340 	m_freem(*m_headp);
3341 	DBG_COUNTER_INC(tx_frees);
3342 	*m_headp = NULL;
3343 	return (ENOMEM);
3344 }
3345 
3346 static void
3347 iflib_tx_desc_free(iflib_txq_t txq, int n)
3348 {
3349 	int hasmap;
3350 	uint32_t qsize, cidx, mask, gen;
3351 	struct mbuf *m, **ifsd_m;
3352 	uint8_t *ifsd_flags;
3353 	bus_dmamap_t *ifsd_map;
3354 	bool do_prefetch;
3355 
3356 	cidx = txq->ift_cidx;
3357 	gen = txq->ift_gen;
3358 	qsize = txq->ift_size;
3359 	mask = qsize-1;
3360 	hasmap = txq->ift_sds.ifsd_map != NULL;
3361 	ifsd_flags = txq->ift_sds.ifsd_flags;
3362 	ifsd_m = txq->ift_sds.ifsd_m;
3363 	ifsd_map = txq->ift_sds.ifsd_map;
3364 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3365 
3366 	while (n--) {
3367 		if (do_prefetch) {
3368 			prefetch(ifsd_m[(cidx + 3) & mask]);
3369 			prefetch(ifsd_m[(cidx + 4) & mask]);
3370 		}
3371 		if (ifsd_m[cidx] != NULL) {
3372 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3373 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3374 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3375 				/*
3376 				 * does it matter if it's not the TSO tag? If so we'll
3377 				 * have to add the type to flags
3378 				 */
3379 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3380 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3381 			}
3382 			if ((m = ifsd_m[cidx]) != NULL) {
3383 				/* XXX we don't support any drivers that batch packets yet */
3384 				MPASS(m->m_nextpkt == NULL);
3385 				/* if the number of clusters exceeds the number of segments
3386 				 * there won't be space on the ring to save a pointer to each
3387 				 * cluster so we simply free the list here
3388 				 */
3389 				if (m->m_flags & M_TOOBIG) {
3390 					m_freem(m);
3391 				} else {
3392 					m_free(m);
3393 				}
3394 				ifsd_m[cidx] = NULL;
3395 #if MEMORY_LOGGING
3396 				txq->ift_dequeued++;
3397 #endif
3398 				DBG_COUNTER_INC(tx_frees);
3399 			}
3400 		}
3401 		if (__predict_false(++cidx == qsize)) {
3402 			cidx = 0;
3403 			gen = 0;
3404 		}
3405 	}
3406 	txq->ift_cidx = cidx;
3407 	txq->ift_gen = gen;
3408 }
3409 
3410 static __inline int
3411 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3412 {
3413 	int reclaim;
3414 	if_ctx_t ctx = txq->ift_ctx;
3415 
3416 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3417 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3418 
3419 	/*
3420 	 * Need a rate-limiting check so that this isn't called every time
3421 	 */
3422 	iflib_tx_credits_update(ctx, txq);
3423 	reclaim = DESC_RECLAIMABLE(txq);
3424 
3425 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3426 #ifdef INVARIANTS
3427 		if (iflib_verbose_debug) {
3428 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3429 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3430 			       reclaim, thresh);
3431 
3432 		}
3433 #endif
3434 		return (0);
3435 	}
3436 	iflib_tx_desc_free(txq, reclaim);
3437 	txq->ift_cleaned += reclaim;
3438 	txq->ift_in_use -= reclaim;
3439 
3440 	return (reclaim);
3441 }
3442 
3443 static struct mbuf **
3444 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3445 {
3446 	int next, size;
3447 	struct mbuf **items;
3448 
3449 	size = r->size;
3450 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3451 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3452 
3453 	prefetch(items[(cidx + offset) & (size-1)]);
3454 	if (remaining > 1) {
3455 		prefetch2cachelines(&items[next]);
3456 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3457 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3458 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3459 	}
3460 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3461 }
3462 
3463 static void
3464 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3465 {
3466 
3467 	ifmp_ring_check_drainage(txq->ift_br, budget);
3468 }
3469 
3470 static uint32_t
3471 iflib_txq_can_drain(struct ifmp_ring *r)
3472 {
3473 	iflib_txq_t txq = r->cookie;
3474 	if_ctx_t ctx = txq->ift_ctx;
3475 
3476 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3477 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3478 }
3479 
3480 static uint32_t
3481 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3482 {
3483 	iflib_txq_t txq = r->cookie;
3484 	if_ctx_t ctx = txq->ift_ctx;
3485 	struct ifnet *ifp = ctx->ifc_ifp;
3486 	struct mbuf **mp, *m;
3487 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3488 	int reclaimed, err, in_use_prev, desc_used;
3489 	bool do_prefetch, ring, rang;
3490 
3491 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3492 			    !LINK_ACTIVE(ctx))) {
3493 		DBG_COUNTER_INC(txq_drain_notready);
3494 		return (0);
3495 	}
3496 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3497 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3498 	avail = IDXDIFF(pidx, cidx, r->size);
3499 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3500 		DBG_COUNTER_INC(txq_drain_flushing);
3501 		for (i = 0; i < avail; i++) {
3502 			m_free(r->items[(cidx + i) & (r->size-1)]);
3503 			r->items[(cidx + i) & (r->size-1)] = NULL;
3504 		}
3505 		return (avail);
3506 	}
3507 
3508 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3509 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3510 		CALLOUT_LOCK(txq);
3511 		callout_stop(&txq->ift_timer);
3512 		CALLOUT_UNLOCK(txq);
3513 		DBG_COUNTER_INC(txq_drain_oactive);
3514 		return (0);
3515 	}
3516 	if (reclaimed)
3517 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3518 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3519 	count = MIN(avail, TX_BATCH_SIZE);
3520 #ifdef INVARIANTS
3521 	if (iflib_verbose_debug)
3522 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3523 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3524 #endif
3525 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3526 	avail = TXQ_AVAIL(txq);
3527 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3528 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3529 
3530 		mp = _ring_peek_one(r, cidx, i, rem);
3531 		MPASS(mp != NULL && *mp != NULL);
3532 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3533 			consumed++;
3534 			reclaimed++;
3535 			continue;
3536 		}
3537 		in_use_prev = txq->ift_in_use;
3538 		pidx_prev = txq->ift_pidx;
3539 		err = iflib_encap(txq, mp);
3540 		if (__predict_false(err)) {
3541 			DBG_COUNTER_INC(txq_drain_encapfail);
3542 			/* no room - bail out */
3543 			if (err == ENOBUFS)
3544 				break;
3545 			consumed++;
3546 			DBG_COUNTER_INC(txq_drain_encapfail);
3547 			/* we can't send this packet - skip it */
3548 			continue;
3549 		}
3550 		consumed++;
3551 		pkt_sent++;
3552 		m = *mp;
3553 		DBG_COUNTER_INC(tx_sent);
3554 		bytes_sent += m->m_pkthdr.len;
3555 		mcast_sent += !!(m->m_flags & M_MCAST);
3556 		avail = TXQ_AVAIL(txq);
3557 
3558 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3559 		desc_used += (txq->ift_in_use - in_use_prev);
3560 		ETHER_BPF_MTAP(ifp, m);
3561 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3562 			break;
3563 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3564 	}
3565 
3566 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3567 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3568 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3569 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3570 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3571 	if (mcast_sent)
3572 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3573 #ifdef INVARIANTS
3574 	if (iflib_verbose_debug)
3575 		printf("consumed=%d\n", consumed);
3576 #endif
3577 	return (consumed);
3578 }
3579 
3580 static uint32_t
3581 iflib_txq_drain_always(struct ifmp_ring *r)
3582 {
3583 	return (1);
3584 }
3585 
3586 static uint32_t
3587 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3588 {
3589 	int i, avail;
3590 	struct mbuf **mp;
3591 	iflib_txq_t txq;
3592 
3593 	txq = r->cookie;
3594 
3595 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3596 	CALLOUT_LOCK(txq);
3597 	callout_stop(&txq->ift_timer);
3598 	CALLOUT_UNLOCK(txq);
3599 
3600 	avail = IDXDIFF(pidx, cidx, r->size);
3601 	for (i = 0; i < avail; i++) {
3602 		mp = _ring_peek_one(r, cidx, i, avail - i);
3603 		if (__predict_false(*mp == (struct mbuf *)txq))
3604 			continue;
3605 		m_freem(*mp);
3606 	}
3607 	MPASS(ifmp_ring_is_stalled(r) == 0);
3608 	return (avail);
3609 }
3610 
3611 static void
3612 iflib_ifmp_purge(iflib_txq_t txq)
3613 {
3614 	struct ifmp_ring *r;
3615 
3616 	r = txq->ift_br;
3617 	r->drain = iflib_txq_drain_free;
3618 	r->can_drain = iflib_txq_drain_always;
3619 
3620 	ifmp_ring_check_drainage(r, r->size);
3621 
3622 	r->drain = iflib_txq_drain;
3623 	r->can_drain = iflib_txq_can_drain;
3624 }
3625 
3626 static void
3627 _task_fn_tx(void *context)
3628 {
3629 	iflib_txq_t txq = context;
3630 	if_ctx_t ctx = txq->ift_ctx;
3631 	struct ifnet *ifp = ctx->ifc_ifp;
3632 	int rc;
3633 
3634 #ifdef IFLIB_DIAGNOSTICS
3635 	txq->ift_cpu_exec_count[curcpu]++;
3636 #endif
3637 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3638 		return;
3639 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3640 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3641 			netmap_tx_irq(ifp, txq->ift_id);
3642 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3643 		return;
3644 	}
3645 	if (txq->ift_db_pending)
3646 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3647 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3648 	if (ctx->ifc_flags & IFC_LEGACY)
3649 		IFDI_INTR_ENABLE(ctx);
3650 	else {
3651 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3652 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3653 	}
3654 }
3655 
3656 static void
3657 _task_fn_rx(void *context)
3658 {
3659 	iflib_rxq_t rxq = context;
3660 	if_ctx_t ctx = rxq->ifr_ctx;
3661 	bool more;
3662 	int rc;
3663 	uint16_t budget;
3664 
3665 #ifdef IFLIB_DIAGNOSTICS
3666 	rxq->ifr_cpu_exec_count[curcpu]++;
3667 #endif
3668 	DBG_COUNTER_INC(task_fn_rxs);
3669 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3670 		return;
3671 	more = true;
3672 #ifdef DEV_NETMAP
3673 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3674 		u_int work = 0;
3675 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3676 			more = false;
3677 		}
3678 	}
3679 #endif
3680 	budget = ctx->ifc_sysctl_rx_budget;
3681 	if (budget == 0)
3682 		budget = 16;	/* XXX */
3683 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3684 		if (ctx->ifc_flags & IFC_LEGACY)
3685 			IFDI_INTR_ENABLE(ctx);
3686 		else {
3687 			DBG_COUNTER_INC(rx_intr_enables);
3688 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3689 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3690 		}
3691 	}
3692 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3693 		return;
3694 	if (more)
3695 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3696 }
3697 
3698 static void
3699 _task_fn_admin(void *context)
3700 {
3701 	if_ctx_t ctx = context;
3702 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3703 	iflib_txq_t txq;
3704 	int i;
3705 
3706 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3707 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3708 			return;
3709 		}
3710 	}
3711 
3712 	CTX_LOCK(ctx);
3713 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3714 		CALLOUT_LOCK(txq);
3715 		callout_stop(&txq->ift_timer);
3716 		CALLOUT_UNLOCK(txq);
3717 	}
3718 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3719 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3720 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3721 	IFDI_LINK_INTR_ENABLE(ctx);
3722 	if (ctx->ifc_flags & IFC_DO_RESET) {
3723 		ctx->ifc_flags &= ~IFC_DO_RESET;
3724 		iflib_if_init_locked(ctx);
3725 	}
3726 	CTX_UNLOCK(ctx);
3727 
3728 	if (LINK_ACTIVE(ctx) == 0)
3729 		return;
3730 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3731 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3732 }
3733 
3734 
3735 static void
3736 _task_fn_iov(void *context)
3737 {
3738 	if_ctx_t ctx = context;
3739 
3740 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3741 		return;
3742 
3743 	CTX_LOCK(ctx);
3744 	IFDI_VFLR_HANDLE(ctx);
3745 	CTX_UNLOCK(ctx);
3746 }
3747 
3748 static int
3749 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3750 {
3751 	int err;
3752 	if_int_delay_info_t info;
3753 	if_ctx_t ctx;
3754 
3755 	info = (if_int_delay_info_t)arg1;
3756 	ctx = info->iidi_ctx;
3757 	info->iidi_req = req;
3758 	info->iidi_oidp = oidp;
3759 	CTX_LOCK(ctx);
3760 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3761 	CTX_UNLOCK(ctx);
3762 	return (err);
3763 }
3764 
3765 /*********************************************************************
3766  *
3767  *  IFNET FUNCTIONS
3768  *
3769  **********************************************************************/
3770 
3771 static void
3772 iflib_if_init_locked(if_ctx_t ctx)
3773 {
3774 	iflib_stop(ctx);
3775 	iflib_init_locked(ctx);
3776 }
3777 
3778 
3779 static void
3780 iflib_if_init(void *arg)
3781 {
3782 	if_ctx_t ctx = arg;
3783 
3784 	CTX_LOCK(ctx);
3785 	iflib_if_init_locked(ctx);
3786 	CTX_UNLOCK(ctx);
3787 }
3788 
3789 static int
3790 iflib_if_transmit(if_t ifp, struct mbuf *m)
3791 {
3792 	if_ctx_t	ctx = if_getsoftc(ifp);
3793 
3794 	iflib_txq_t txq;
3795 	int err, qidx;
3796 
3797 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3798 		DBG_COUNTER_INC(tx_frees);
3799 		m_freem(m);
3800 		return (ENOBUFS);
3801 	}
3802 
3803 	MPASS(m->m_nextpkt == NULL);
3804 	qidx = 0;
3805 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3806 		qidx = QIDX(ctx, m);
3807 	/*
3808 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3809 	 */
3810 	txq = &ctx->ifc_txqs[qidx];
3811 
3812 #ifdef DRIVER_BACKPRESSURE
3813 	if (txq->ift_closed) {
3814 		while (m != NULL) {
3815 			next = m->m_nextpkt;
3816 			m->m_nextpkt = NULL;
3817 			m_freem(m);
3818 			m = next;
3819 		}
3820 		return (ENOBUFS);
3821 	}
3822 #endif
3823 #ifdef notyet
3824 	qidx = count = 0;
3825 	mp = marr;
3826 	next = m;
3827 	do {
3828 		count++;
3829 		next = next->m_nextpkt;
3830 	} while (next != NULL);
3831 
3832 	if (count > nitems(marr))
3833 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3834 			/* XXX check nextpkt */
3835 			m_freem(m);
3836 			/* XXX simplify for now */
3837 			DBG_COUNTER_INC(tx_frees);
3838 			return (ENOBUFS);
3839 		}
3840 	for (next = m, i = 0; next != NULL; i++) {
3841 		mp[i] = next;
3842 		next = next->m_nextpkt;
3843 		mp[i]->m_nextpkt = NULL;
3844 	}
3845 #endif
3846 	DBG_COUNTER_INC(tx_seen);
3847 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3848 
3849 	GROUPTASK_ENQUEUE(&txq->ift_task);
3850 	if (err) {
3851 		/* support forthcoming later */
3852 #ifdef DRIVER_BACKPRESSURE
3853 		txq->ift_closed = TRUE;
3854 #endif
3855 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3856 		m_freem(m);
3857 	}
3858 
3859 	return (err);
3860 }
3861 
3862 static void
3863 iflib_if_qflush(if_t ifp)
3864 {
3865 	if_ctx_t ctx = if_getsoftc(ifp);
3866 	iflib_txq_t txq = ctx->ifc_txqs;
3867 	int i;
3868 
3869 	CTX_LOCK(ctx);
3870 	ctx->ifc_flags |= IFC_QFLUSH;
3871 	CTX_UNLOCK(ctx);
3872 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3873 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3874 			iflib_txq_check_drain(txq, 0);
3875 	CTX_LOCK(ctx);
3876 	ctx->ifc_flags &= ~IFC_QFLUSH;
3877 	CTX_UNLOCK(ctx);
3878 
3879 	if_qflush(ifp);
3880 }
3881 
3882 
3883 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3884 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3885 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3886 
3887 static int
3888 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3889 {
3890 	if_ctx_t ctx = if_getsoftc(ifp);
3891 	struct ifreq	*ifr = (struct ifreq *)data;
3892 #if defined(INET) || defined(INET6)
3893 	struct ifaddr	*ifa = (struct ifaddr *)data;
3894 #endif
3895 	bool		avoid_reset = FALSE;
3896 	int		err = 0, reinit = 0, bits;
3897 
3898 	switch (command) {
3899 	case SIOCSIFADDR:
3900 #ifdef INET
3901 		if (ifa->ifa_addr->sa_family == AF_INET)
3902 			avoid_reset = TRUE;
3903 #endif
3904 #ifdef INET6
3905 		if (ifa->ifa_addr->sa_family == AF_INET6)
3906 			avoid_reset = TRUE;
3907 #endif
3908 		/*
3909 		** Calling init results in link renegotiation,
3910 		** so we avoid doing it when possible.
3911 		*/
3912 		if (avoid_reset) {
3913 			if_setflagbits(ifp, IFF_UP,0);
3914 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3915 				reinit = 1;
3916 #ifdef INET
3917 			if (!(if_getflags(ifp) & IFF_NOARP))
3918 				arp_ifinit(ifp, ifa);
3919 #endif
3920 		} else
3921 			err = ether_ioctl(ifp, command, data);
3922 		break;
3923 	case SIOCSIFMTU:
3924 		CTX_LOCK(ctx);
3925 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3926 			CTX_UNLOCK(ctx);
3927 			break;
3928 		}
3929 		bits = if_getdrvflags(ifp);
3930 		/* stop the driver and free any clusters before proceeding */
3931 		iflib_stop(ctx);
3932 
3933 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3934 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3935 				ctx->ifc_flags |= IFC_MULTISEG;
3936 			else
3937 				ctx->ifc_flags &= ~IFC_MULTISEG;
3938 			err = if_setmtu(ifp, ifr->ifr_mtu);
3939 		}
3940 		iflib_init_locked(ctx);
3941 		if_setdrvflags(ifp, bits);
3942 		CTX_UNLOCK(ctx);
3943 		break;
3944 	case SIOCSIFFLAGS:
3945 		CTX_LOCK(ctx);
3946 		if (if_getflags(ifp) & IFF_UP) {
3947 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3948 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3949 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3950 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3951 				}
3952 			} else
3953 				reinit = 1;
3954 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3955 			iflib_stop(ctx);
3956 		}
3957 		ctx->ifc_if_flags = if_getflags(ifp);
3958 		CTX_UNLOCK(ctx);
3959 		break;
3960 	case SIOCADDMULTI:
3961 	case SIOCDELMULTI:
3962 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3963 			CTX_LOCK(ctx);
3964 			IFDI_INTR_DISABLE(ctx);
3965 			IFDI_MULTI_SET(ctx);
3966 			IFDI_INTR_ENABLE(ctx);
3967 			CTX_UNLOCK(ctx);
3968 		}
3969 		break;
3970 	case SIOCSIFMEDIA:
3971 		CTX_LOCK(ctx);
3972 		IFDI_MEDIA_SET(ctx);
3973 		CTX_UNLOCK(ctx);
3974 		/* falls thru */
3975 	case SIOCGIFMEDIA:
3976 	case SIOCGIFXMEDIA:
3977 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3978 		break;
3979 	case SIOCGI2C:
3980 	{
3981 		struct ifi2creq i2c;
3982 
3983 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3984 		if (err != 0)
3985 			break;
3986 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3987 			err = EINVAL;
3988 			break;
3989 		}
3990 		if (i2c.len > sizeof(i2c.data)) {
3991 			err = EINVAL;
3992 			break;
3993 		}
3994 
3995 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3996 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3997 		break;
3998 	}
3999 	case SIOCSIFCAP:
4000 	{
4001 		int mask, setmask;
4002 
4003 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4004 		setmask = 0;
4005 #ifdef TCP_OFFLOAD
4006 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4007 #endif
4008 		setmask |= (mask & IFCAP_FLAGS);
4009 
4010 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4011 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4012 		if ((mask & IFCAP_WOL) &&
4013 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4014 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4015 		if_vlancap(ifp);
4016 		/*
4017 		 * want to ensure that traffic has stopped before we change any of the flags
4018 		 */
4019 		if (setmask) {
4020 			CTX_LOCK(ctx);
4021 			bits = if_getdrvflags(ifp);
4022 			if (bits & IFF_DRV_RUNNING)
4023 				iflib_stop(ctx);
4024 			if_togglecapenable(ifp, setmask);
4025 			if (bits & IFF_DRV_RUNNING)
4026 				iflib_init_locked(ctx);
4027 			if_setdrvflags(ifp, bits);
4028 			CTX_UNLOCK(ctx);
4029 		}
4030 		break;
4031 	    }
4032 	case SIOCGPRIVATE_0:
4033 	case SIOCSDRVSPEC:
4034 	case SIOCGDRVSPEC:
4035 		CTX_LOCK(ctx);
4036 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4037 		CTX_UNLOCK(ctx);
4038 		break;
4039 	default:
4040 		err = ether_ioctl(ifp, command, data);
4041 		break;
4042 	}
4043 	if (reinit)
4044 		iflib_if_init(ctx);
4045 	return (err);
4046 }
4047 
4048 static uint64_t
4049 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4050 {
4051 	if_ctx_t ctx = if_getsoftc(ifp);
4052 
4053 	return (IFDI_GET_COUNTER(ctx, cnt));
4054 }
4055 
4056 /*********************************************************************
4057  *
4058  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4059  *
4060  **********************************************************************/
4061 
4062 static void
4063 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4064 {
4065 	if_ctx_t ctx = if_getsoftc(ifp);
4066 
4067 	if ((void *)ctx != arg)
4068 		return;
4069 
4070 	if ((vtag == 0) || (vtag > 4095))
4071 		return;
4072 
4073 	CTX_LOCK(ctx);
4074 	IFDI_VLAN_REGISTER(ctx, vtag);
4075 	/* Re-init to load the changes */
4076 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4077 		iflib_if_init_locked(ctx);
4078 	CTX_UNLOCK(ctx);
4079 }
4080 
4081 static void
4082 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4083 {
4084 	if_ctx_t ctx = if_getsoftc(ifp);
4085 
4086 	if ((void *)ctx != arg)
4087 		return;
4088 
4089 	if ((vtag == 0) || (vtag > 4095))
4090 		return;
4091 
4092 	CTX_LOCK(ctx);
4093 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4094 	/* Re-init to load the changes */
4095 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4096 		iflib_if_init_locked(ctx);
4097 	CTX_UNLOCK(ctx);
4098 }
4099 
4100 static void
4101 iflib_led_func(void *arg, int onoff)
4102 {
4103 	if_ctx_t ctx = arg;
4104 
4105 	CTX_LOCK(ctx);
4106 	IFDI_LED_FUNC(ctx, onoff);
4107 	CTX_UNLOCK(ctx);
4108 }
4109 
4110 /*********************************************************************
4111  *
4112  *  BUS FUNCTION DEFINITIONS
4113  *
4114  **********************************************************************/
4115 
4116 int
4117 iflib_device_probe(device_t dev)
4118 {
4119 	pci_vendor_info_t *ent;
4120 
4121 	uint16_t	pci_vendor_id, pci_device_id;
4122 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4123 	uint16_t	pci_rev_id;
4124 	if_shared_ctx_t sctx;
4125 
4126 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4127 		return (ENOTSUP);
4128 
4129 	pci_vendor_id = pci_get_vendor(dev);
4130 	pci_device_id = pci_get_device(dev);
4131 	pci_subvendor_id = pci_get_subvendor(dev);
4132 	pci_subdevice_id = pci_get_subdevice(dev);
4133 	pci_rev_id = pci_get_revid(dev);
4134 	if (sctx->isc_parse_devinfo != NULL)
4135 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4136 
4137 	ent = sctx->isc_vendor_info;
4138 	while (ent->pvi_vendor_id != 0) {
4139 		if (pci_vendor_id != ent->pvi_vendor_id) {
4140 			ent++;
4141 			continue;
4142 		}
4143 		if ((pci_device_id == ent->pvi_device_id) &&
4144 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4145 		     (ent->pvi_subvendor_id == 0)) &&
4146 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4147 		     (ent->pvi_subdevice_id == 0)) &&
4148 		    ((pci_rev_id == ent->pvi_rev_id) ||
4149 		     (ent->pvi_rev_id == 0))) {
4150 
4151 			device_set_desc_copy(dev, ent->pvi_name);
4152 			/* this needs to be changed to zero if the bus probing code
4153 			 * ever stops re-probing on best match because the sctx
4154 			 * may have its values over written by register calls
4155 			 * in subsequent probes
4156 			 */
4157 			return (BUS_PROBE_DEFAULT);
4158 		}
4159 		ent++;
4160 	}
4161 	return (ENXIO);
4162 }
4163 
4164 int
4165 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4166 {
4167 	int err, rid, msix, msix_bar;
4168 	if_ctx_t ctx;
4169 	if_t ifp;
4170 	if_softc_ctx_t scctx;
4171 	int i;
4172 	uint16_t main_txq;
4173 	uint16_t main_rxq;
4174 
4175 
4176 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4177 
4178 	if (sc == NULL) {
4179 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4180 		device_set_softc(dev, ctx);
4181 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4182 	}
4183 
4184 	ctx->ifc_sctx = sctx;
4185 	ctx->ifc_dev = dev;
4186 	ctx->ifc_softc = sc;
4187 
4188 	if ((err = iflib_register(ctx)) != 0) {
4189 		device_printf(dev, "iflib_register failed %d\n", err);
4190 		return (err);
4191 	}
4192 	iflib_add_device_sysctl_pre(ctx);
4193 
4194 	scctx = &ctx->ifc_softc_ctx;
4195 	ifp = ctx->ifc_ifp;
4196 
4197 	/*
4198 	 * XXX sanity check that ntxd & nrxd are a power of 2
4199 	 */
4200 	if (ctx->ifc_sysctl_ntxqs != 0)
4201 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4202 	if (ctx->ifc_sysctl_nrxqs != 0)
4203 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4204 
4205 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4206 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4207 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4208 		else
4209 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4210 	}
4211 
4212 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4213 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4214 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4215 		else
4216 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4217 	}
4218 
4219 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4220 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4221 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4222 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4223 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4224 		}
4225 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4226 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4227 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4228 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4229 		}
4230 	}
4231 
4232 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4233 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4234 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4235 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4236 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4237 		}
4238 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4239 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4240 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4241 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4242 		}
4243 	}
4244 
4245 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4246 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4247 		return (err);
4248 	}
4249 	_iflib_pre_assert(scctx);
4250 	ctx->ifc_txrx = *scctx->isc_txrx;
4251 
4252 #ifdef INVARIANTS
4253 	MPASS(scctx->isc_capenable);
4254 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4255 		MPASS(scctx->isc_tx_csum_flags);
4256 #endif
4257 
4258 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4259 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4260 
4261 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4262 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4263 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4264 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4265 
4266 #ifdef ACPI_DMAR
4267 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4268 		ctx->ifc_flags |= IFC_DMAR;
4269 #elif !(defined(__i386__) || defined(__amd64__))
4270 	/* set unconditionally for !x86 */
4271 	ctx->ifc_flags |= IFC_DMAR;
4272 #endif
4273 
4274 	msix_bar = scctx->isc_msix_bar;
4275 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4276 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4277 
4278 	/* XXX change for per-queue sizes */
4279 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4280 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4281 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4282 		if (!powerof2(scctx->isc_nrxd[i])) {
4283 			/* round down instead? */
4284 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4285 			err = EINVAL;
4286 			goto fail;
4287 		}
4288 	}
4289 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4290 		if (!powerof2(scctx->isc_ntxd[i])) {
4291 			device_printf(dev,
4292 			    "# tx descriptors must be a power of 2");
4293 			err = EINVAL;
4294 			goto fail;
4295 		}
4296 	}
4297 
4298 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4299 	    MAX_SINGLE_PACKET_FRACTION)
4300 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4301 		    MAX_SINGLE_PACKET_FRACTION);
4302 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4303 	    MAX_SINGLE_PACKET_FRACTION)
4304 		scctx->isc_tx_tso_segments_max = max(1,
4305 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4306 
4307 	/*
4308 	 * Protect the stack against modern hardware
4309 	 */
4310 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4311 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4312 
4313 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4314 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4315 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4316 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4317 	if (scctx->isc_rss_table_size == 0)
4318 		scctx->isc_rss_table_size = 64;
4319 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4320 
4321 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4322 	/* XXX format name */
4323 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4324 
4325 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4326 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4327 		device_printf(dev, "Unable to fetch CPU list\n");
4328 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4329 	}
4330 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4331 
4332 	/*
4333 	** Now setup MSI or MSI/X, should
4334 	** return us the number of supported
4335 	** vectors. (Will be 1 for MSI)
4336 	*/
4337 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4338 		msix = scctx->isc_vectors;
4339 	} else if (scctx->isc_msix_bar != 0)
4340 	       /*
4341 		* The simple fact that isc_msix_bar is not 0 does not mean we
4342 		* we have a good value there that is known to work.
4343 		*/
4344 		msix = iflib_msix_init(ctx);
4345 	else {
4346 		scctx->isc_vectors = 1;
4347 		scctx->isc_ntxqsets = 1;
4348 		scctx->isc_nrxqsets = 1;
4349 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4350 		msix = 0;
4351 	}
4352 	/* Get memory for the station queues */
4353 	if ((err = iflib_queues_alloc(ctx))) {
4354 		device_printf(dev, "Unable to allocate queue memory\n");
4355 		goto fail;
4356 	}
4357 
4358 	if ((err = iflib_qset_structures_setup(ctx))) {
4359 		device_printf(dev, "qset structure setup failed %d\n", err);
4360 		goto fail_queues;
4361 	}
4362 
4363 	/*
4364 	 * Group taskqueues aren't properly set up until SMP is started,
4365 	 * so we disable interrupts until we can handle them post
4366 	 * SI_SUB_SMP.
4367 	 *
4368 	 * XXX: disabling interrupts doesn't actually work, at least for
4369 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4370 	 * we do null handling and depend on this not causing too large an
4371 	 * interrupt storm.
4372 	 */
4373 	IFDI_INTR_DISABLE(ctx);
4374 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4375 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4376 		goto fail_intr_free;
4377 	}
4378 	if (msix <= 1) {
4379 		rid = 0;
4380 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4381 			MPASS(msix == 1);
4382 			rid = 1;
4383 		}
4384 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4385 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4386 			goto fail_intr_free;
4387 		}
4388 	}
4389 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4390 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4391 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4392 		goto fail_detach;
4393 	}
4394 	if ((err = iflib_netmap_attach(ctx))) {
4395 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4396 		goto fail_detach;
4397 	}
4398 	*ctxp = ctx;
4399 
4400 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4401 	iflib_add_device_sysctl_post(ctx);
4402 	ctx->ifc_flags |= IFC_INIT_DONE;
4403 	return (0);
4404 fail_detach:
4405 	ether_ifdetach(ctx->ifc_ifp);
4406 fail_intr_free:
4407 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4408 		pci_release_msi(ctx->ifc_dev);
4409 fail_queues:
4410 	/* XXX free queues */
4411 fail:
4412 	IFDI_DETACH(ctx);
4413 	return (err);
4414 }
4415 
4416 int
4417 iflib_device_attach(device_t dev)
4418 {
4419 	if_ctx_t ctx;
4420 	if_shared_ctx_t sctx;
4421 
4422 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4423 		return (ENOTSUP);
4424 
4425 	pci_enable_busmaster(dev);
4426 
4427 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4428 }
4429 
4430 int
4431 iflib_device_deregister(if_ctx_t ctx)
4432 {
4433 	if_t ifp = ctx->ifc_ifp;
4434 	iflib_txq_t txq;
4435 	iflib_rxq_t rxq;
4436 	device_t dev = ctx->ifc_dev;
4437 	int i, j;
4438 	struct taskqgroup *tqg;
4439 	iflib_fl_t fl;
4440 
4441 	/* Make sure VLANS are not using driver */
4442 	if (if_vlantrunkinuse(ifp)) {
4443 		device_printf(dev,"Vlan in use, detach first\n");
4444 		return (EBUSY);
4445 	}
4446 
4447 	CTX_LOCK(ctx);
4448 	ctx->ifc_in_detach = 1;
4449 	iflib_stop(ctx);
4450 	CTX_UNLOCK(ctx);
4451 
4452 	/* Unregister VLAN events */
4453 	if (ctx->ifc_vlan_attach_event != NULL)
4454 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4455 	if (ctx->ifc_vlan_detach_event != NULL)
4456 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4457 
4458 	iflib_netmap_detach(ifp);
4459 	ether_ifdetach(ifp);
4460 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4461 	CTX_LOCK_DESTROY(ctx);
4462 	if (ctx->ifc_led_dev != NULL)
4463 		led_destroy(ctx->ifc_led_dev);
4464 	/* XXX drain any dependent tasks */
4465 	tqg = qgroup_if_io_tqg;
4466 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4467 		callout_drain(&txq->ift_timer);
4468 		if (txq->ift_task.gt_uniq != NULL)
4469 			taskqgroup_detach(tqg, &txq->ift_task);
4470 	}
4471 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4472 		if (rxq->ifr_task.gt_uniq != NULL)
4473 			taskqgroup_detach(tqg, &rxq->ifr_task);
4474 
4475 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4476 			free(fl->ifl_rx_bitmap, M_IFLIB);
4477 
4478 	}
4479 	tqg = qgroup_if_config_tqg;
4480 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4481 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4482 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4483 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4484 
4485 	IFDI_DETACH(ctx);
4486 	device_set_softc(ctx->ifc_dev, NULL);
4487 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4488 		pci_release_msi(dev);
4489 	}
4490 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4491 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4492 	}
4493 	if (ctx->ifc_msix_mem != NULL) {
4494 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4495 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4496 		ctx->ifc_msix_mem = NULL;
4497 	}
4498 
4499 	bus_generic_detach(dev);
4500 	if_free(ifp);
4501 
4502 	iflib_tx_structures_free(ctx);
4503 	iflib_rx_structures_free(ctx);
4504 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4505 		free(ctx->ifc_softc, M_IFLIB);
4506 	free(ctx, M_IFLIB);
4507 	return (0);
4508 }
4509 
4510 
4511 int
4512 iflib_device_detach(device_t dev)
4513 {
4514 	if_ctx_t ctx = device_get_softc(dev);
4515 
4516 	return (iflib_device_deregister(ctx));
4517 }
4518 
4519 int
4520 iflib_device_suspend(device_t dev)
4521 {
4522 	if_ctx_t ctx = device_get_softc(dev);
4523 
4524 	CTX_LOCK(ctx);
4525 	IFDI_SUSPEND(ctx);
4526 	CTX_UNLOCK(ctx);
4527 
4528 	return bus_generic_suspend(dev);
4529 }
4530 int
4531 iflib_device_shutdown(device_t dev)
4532 {
4533 	if_ctx_t ctx = device_get_softc(dev);
4534 
4535 	CTX_LOCK(ctx);
4536 	IFDI_SHUTDOWN(ctx);
4537 	CTX_UNLOCK(ctx);
4538 
4539 	return bus_generic_suspend(dev);
4540 }
4541 
4542 
4543 int
4544 iflib_device_resume(device_t dev)
4545 {
4546 	if_ctx_t ctx = device_get_softc(dev);
4547 	iflib_txq_t txq = ctx->ifc_txqs;
4548 
4549 	CTX_LOCK(ctx);
4550 	IFDI_RESUME(ctx);
4551 	iflib_init_locked(ctx);
4552 	CTX_UNLOCK(ctx);
4553 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4554 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4555 
4556 	return (bus_generic_resume(dev));
4557 }
4558 
4559 int
4560 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4561 {
4562 	int error;
4563 	if_ctx_t ctx = device_get_softc(dev);
4564 
4565 	CTX_LOCK(ctx);
4566 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4567 	CTX_UNLOCK(ctx);
4568 
4569 	return (error);
4570 }
4571 
4572 void
4573 iflib_device_iov_uninit(device_t dev)
4574 {
4575 	if_ctx_t ctx = device_get_softc(dev);
4576 
4577 	CTX_LOCK(ctx);
4578 	IFDI_IOV_UNINIT(ctx);
4579 	CTX_UNLOCK(ctx);
4580 }
4581 
4582 int
4583 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4584 {
4585 	int error;
4586 	if_ctx_t ctx = device_get_softc(dev);
4587 
4588 	CTX_LOCK(ctx);
4589 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4590 	CTX_UNLOCK(ctx);
4591 
4592 	return (error);
4593 }
4594 
4595 /*********************************************************************
4596  *
4597  *  MODULE FUNCTION DEFINITIONS
4598  *
4599  **********************************************************************/
4600 
4601 /*
4602  * - Start a fast taskqueue thread for each core
4603  * - Start a taskqueue for control operations
4604  */
4605 static int
4606 iflib_module_init(void)
4607 {
4608 	return (0);
4609 }
4610 
4611 static int
4612 iflib_module_event_handler(module_t mod, int what, void *arg)
4613 {
4614 	int err;
4615 
4616 	switch (what) {
4617 	case MOD_LOAD:
4618 		if ((err = iflib_module_init()) != 0)
4619 			return (err);
4620 		break;
4621 	case MOD_UNLOAD:
4622 		return (EBUSY);
4623 	default:
4624 		return (EOPNOTSUPP);
4625 	}
4626 
4627 	return (0);
4628 }
4629 
4630 /*********************************************************************
4631  *
4632  *  PUBLIC FUNCTION DEFINITIONS
4633  *     ordered as in iflib.h
4634  *
4635  **********************************************************************/
4636 
4637 
4638 static void
4639 _iflib_assert(if_shared_ctx_t sctx)
4640 {
4641 	MPASS(sctx->isc_tx_maxsize);
4642 	MPASS(sctx->isc_tx_maxsegsize);
4643 
4644 	MPASS(sctx->isc_rx_maxsize);
4645 	MPASS(sctx->isc_rx_nsegments);
4646 	MPASS(sctx->isc_rx_maxsegsize);
4647 
4648 	MPASS(sctx->isc_nrxd_min[0]);
4649 	MPASS(sctx->isc_nrxd_max[0]);
4650 	MPASS(sctx->isc_nrxd_default[0]);
4651 	MPASS(sctx->isc_ntxd_min[0]);
4652 	MPASS(sctx->isc_ntxd_max[0]);
4653 	MPASS(sctx->isc_ntxd_default[0]);
4654 }
4655 
4656 static void
4657 _iflib_pre_assert(if_softc_ctx_t scctx)
4658 {
4659 
4660 	MPASS(scctx->isc_txrx->ift_txd_encap);
4661 	MPASS(scctx->isc_txrx->ift_txd_flush);
4662 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4663 	MPASS(scctx->isc_txrx->ift_rxd_available);
4664 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4665 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4666 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4667 }
4668 
4669 static int
4670 iflib_register(if_ctx_t ctx)
4671 {
4672 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4673 	driver_t *driver = sctx->isc_driver;
4674 	device_t dev = ctx->ifc_dev;
4675 	if_t ifp;
4676 
4677 	_iflib_assert(sctx);
4678 
4679 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4680 
4681 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4682 	if (ifp == NULL) {
4683 		device_printf(dev, "can not allocate ifnet structure\n");
4684 		return (ENOMEM);
4685 	}
4686 
4687 	/*
4688 	 * Initialize our context's device specific methods
4689 	 */
4690 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4691 	kobj_class_compile((kobj_class_t) driver);
4692 	driver->refs++;
4693 
4694 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4695 	if_setsoftc(ifp, ctx);
4696 	if_setdev(ifp, dev);
4697 	if_setinitfn(ifp, iflib_if_init);
4698 	if_setioctlfn(ifp, iflib_if_ioctl);
4699 	if_settransmitfn(ifp, iflib_if_transmit);
4700 	if_setqflushfn(ifp, iflib_if_qflush);
4701 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4702 
4703 	ctx->ifc_vlan_attach_event =
4704 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4705 							  EVENTHANDLER_PRI_FIRST);
4706 	ctx->ifc_vlan_detach_event =
4707 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4708 							  EVENTHANDLER_PRI_FIRST);
4709 
4710 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4711 					 iflib_media_change, iflib_media_status);
4712 
4713 	return (0);
4714 }
4715 
4716 
4717 static int
4718 iflib_queues_alloc(if_ctx_t ctx)
4719 {
4720 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4721 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4722 	device_t dev = ctx->ifc_dev;
4723 	int nrxqsets = scctx->isc_nrxqsets;
4724 	int ntxqsets = scctx->isc_ntxqsets;
4725 	iflib_txq_t txq;
4726 	iflib_rxq_t rxq;
4727 	iflib_fl_t fl = NULL;
4728 	int i, j, cpu, err, txconf, rxconf;
4729 	iflib_dma_info_t ifdip;
4730 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4731 	uint32_t *txqsizes = scctx->isc_txqsizes;
4732 	uint8_t nrxqs = sctx->isc_nrxqs;
4733 	uint8_t ntxqs = sctx->isc_ntxqs;
4734 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4735 	caddr_t *vaddrs;
4736 	uint64_t *paddrs;
4737 	struct ifmp_ring **brscp;
4738 
4739 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4740 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4741 
4742 	brscp = NULL;
4743 	txq = NULL;
4744 	rxq = NULL;
4745 
4746 /* Allocate the TX ring struct memory */
4747 	if (!(txq =
4748 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4749 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4750 		device_printf(dev, "Unable to allocate TX ring memory\n");
4751 		err = ENOMEM;
4752 		goto fail;
4753 	}
4754 
4755 	/* Now allocate the RX */
4756 	if (!(rxq =
4757 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4758 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4759 		device_printf(dev, "Unable to allocate RX ring memory\n");
4760 		err = ENOMEM;
4761 		goto rx_fail;
4762 	}
4763 
4764 	ctx->ifc_txqs = txq;
4765 	ctx->ifc_rxqs = rxq;
4766 
4767 	/*
4768 	 * XXX handle allocation failure
4769 	 */
4770 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4771 		/* Set up some basics */
4772 
4773 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4774 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4775 			err = ENOMEM;
4776 			goto err_tx_desc;
4777 		}
4778 		txq->ift_ifdi = ifdip;
4779 		for (j = 0; j < ntxqs; j++, ifdip++) {
4780 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4781 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4782 				err = ENOMEM;
4783 				goto err_tx_desc;
4784 			}
4785 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4786 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4787 		}
4788 		txq->ift_ctx = ctx;
4789 		txq->ift_id = i;
4790 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4791 			txq->ift_br_offset = 1;
4792 		} else {
4793 			txq->ift_br_offset = 0;
4794 		}
4795 		/* XXX fix this */
4796 		txq->ift_timer.c_cpu = cpu;
4797 
4798 		if (iflib_txsd_alloc(txq)) {
4799 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4800 			err = ENOMEM;
4801 			goto err_tx_desc;
4802 		}
4803 
4804 		/* Initialize the TX lock */
4805 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4806 		    device_get_nameunit(dev), txq->ift_id);
4807 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4808 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4809 
4810 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4811 			 device_get_nameunit(dev), txq->ift_id);
4812 
4813 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4814 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4815 		if (err) {
4816 			/* XXX free any allocated rings */
4817 			device_printf(dev, "Unable to allocate buf_ring\n");
4818 			goto err_tx_desc;
4819 		}
4820 	}
4821 
4822 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4823 		/* Set up some basics */
4824 
4825 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4826 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4827 			err = ENOMEM;
4828 			goto err_tx_desc;
4829 		}
4830 
4831 		rxq->ifr_ifdi = ifdip;
4832 		/* XXX this needs to be changed if #rx queues != #tx queues */
4833 		rxq->ifr_ntxqirq = 1;
4834 		rxq->ifr_txqid[0] = i;
4835 		for (j = 0; j < nrxqs; j++, ifdip++) {
4836 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4837 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4838 				err = ENOMEM;
4839 				goto err_tx_desc;
4840 			}
4841 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4842 		}
4843 		rxq->ifr_ctx = ctx;
4844 		rxq->ifr_id = i;
4845 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4846 			rxq->ifr_fl_offset = 1;
4847 		} else {
4848 			rxq->ifr_fl_offset = 0;
4849 		}
4850 		rxq->ifr_nfl = nfree_lists;
4851 		if (!(fl =
4852 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4853 			device_printf(dev, "Unable to allocate free list memory\n");
4854 			err = ENOMEM;
4855 			goto err_tx_desc;
4856 		}
4857 		rxq->ifr_fl = fl;
4858 		for (j = 0; j < nfree_lists; j++) {
4859 			fl[j].ifl_rxq = rxq;
4860 			fl[j].ifl_id = j;
4861 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4862 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4863 		}
4864         /* Allocate receive buffers for the ring*/
4865 		if (iflib_rxsd_alloc(rxq)) {
4866 			device_printf(dev,
4867 			    "Critical Failure setting up receive buffers\n");
4868 			err = ENOMEM;
4869 			goto err_rx_desc;
4870 		}
4871 
4872 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4873 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4874 	}
4875 
4876 	/* TXQs */
4877 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4878 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4879 	for (i = 0; i < ntxqsets; i++) {
4880 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4881 
4882 		for (j = 0; j < ntxqs; j++, di++) {
4883 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4884 			paddrs[i*ntxqs + j] = di->idi_paddr;
4885 		}
4886 	}
4887 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4888 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4889 		iflib_tx_structures_free(ctx);
4890 		free(vaddrs, M_IFLIB);
4891 		free(paddrs, M_IFLIB);
4892 		goto err_rx_desc;
4893 	}
4894 	free(vaddrs, M_IFLIB);
4895 	free(paddrs, M_IFLIB);
4896 
4897 	/* RXQs */
4898 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4899 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4900 	for (i = 0; i < nrxqsets; i++) {
4901 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4902 
4903 		for (j = 0; j < nrxqs; j++, di++) {
4904 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4905 			paddrs[i*nrxqs + j] = di->idi_paddr;
4906 		}
4907 	}
4908 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4909 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4910 		iflib_tx_structures_free(ctx);
4911 		free(vaddrs, M_IFLIB);
4912 		free(paddrs, M_IFLIB);
4913 		goto err_rx_desc;
4914 	}
4915 	free(vaddrs, M_IFLIB);
4916 	free(paddrs, M_IFLIB);
4917 
4918 	return (0);
4919 
4920 /* XXX handle allocation failure changes */
4921 err_rx_desc:
4922 err_tx_desc:
4923 	if (ctx->ifc_rxqs != NULL)
4924 		free(ctx->ifc_rxqs, M_IFLIB);
4925 	ctx->ifc_rxqs = NULL;
4926 	if (ctx->ifc_txqs != NULL)
4927 		free(ctx->ifc_txqs, M_IFLIB);
4928 	ctx->ifc_txqs = NULL;
4929 rx_fail:
4930 	if (brscp != NULL)
4931 		free(brscp, M_IFLIB);
4932 	if (rxq != NULL)
4933 		free(rxq, M_IFLIB);
4934 	if (txq != NULL)
4935 		free(txq, M_IFLIB);
4936 fail:
4937 	return (err);
4938 }
4939 
4940 static int
4941 iflib_tx_structures_setup(if_ctx_t ctx)
4942 {
4943 	iflib_txq_t txq = ctx->ifc_txqs;
4944 	int i;
4945 
4946 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4947 		iflib_txq_setup(txq);
4948 
4949 	return (0);
4950 }
4951 
4952 static void
4953 iflib_tx_structures_free(if_ctx_t ctx)
4954 {
4955 	iflib_txq_t txq = ctx->ifc_txqs;
4956 	int i, j;
4957 
4958 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4959 		iflib_txq_destroy(txq);
4960 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4961 			iflib_dma_free(&txq->ift_ifdi[j]);
4962 	}
4963 	free(ctx->ifc_txqs, M_IFLIB);
4964 	ctx->ifc_txqs = NULL;
4965 	IFDI_QUEUES_FREE(ctx);
4966 }
4967 
4968 /*********************************************************************
4969  *
4970  *  Initialize all receive rings.
4971  *
4972  **********************************************************************/
4973 static int
4974 iflib_rx_structures_setup(if_ctx_t ctx)
4975 {
4976 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4977 	int q;
4978 #if defined(INET6) || defined(INET)
4979 	int i, err;
4980 #endif
4981 
4982 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4983 #if defined(INET6) || defined(INET)
4984 		tcp_lro_free(&rxq->ifr_lc);
4985 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4986 		    TCP_LRO_ENTRIES, min(1024,
4987 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4988 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4989 			goto fail;
4990 		}
4991 		rxq->ifr_lro_enabled = TRUE;
4992 #endif
4993 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4994 	}
4995 	return (0);
4996 #if defined(INET6) || defined(INET)
4997 fail:
4998 	/*
4999 	 * Free RX software descriptors allocated so far, we will only handle
5000 	 * the rings that completed, the failing case will have
5001 	 * cleaned up for itself. 'q' failed, so its the terminus.
5002 	 */
5003 	rxq = ctx->ifc_rxqs;
5004 	for (i = 0; i < q; ++i, rxq++) {
5005 		iflib_rx_sds_free(rxq);
5006 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5007 	}
5008 	return (err);
5009 #endif
5010 }
5011 
5012 /*********************************************************************
5013  *
5014  *  Free all receive rings.
5015  *
5016  **********************************************************************/
5017 static void
5018 iflib_rx_structures_free(if_ctx_t ctx)
5019 {
5020 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5021 
5022 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5023 		iflib_rx_sds_free(rxq);
5024 	}
5025 }
5026 
5027 static int
5028 iflib_qset_structures_setup(if_ctx_t ctx)
5029 {
5030 	int err;
5031 
5032 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5033 		return (err);
5034 
5035 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5036 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5037 		iflib_tx_structures_free(ctx);
5038 		iflib_rx_structures_free(ctx);
5039 	}
5040 	return (err);
5041 }
5042 
5043 int
5044 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5045 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5046 {
5047 
5048 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5049 }
5050 
5051 #ifdef SMP
5052 static int
5053 find_nth(if_ctx_t ctx, int qid)
5054 {
5055 	cpuset_t cpus;
5056 	int i, cpuid, eqid, count;
5057 
5058 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5059 	count = CPU_COUNT(&cpus);
5060 	eqid = qid % count;
5061 	/* clear up to the qid'th bit */
5062 	for (i = 0; i < eqid; i++) {
5063 		cpuid = CPU_FFS(&cpus);
5064 		MPASS(cpuid != 0);
5065 		CPU_CLR(cpuid-1, &cpus);
5066 	}
5067 	cpuid = CPU_FFS(&cpus);
5068 	MPASS(cpuid != 0);
5069 	return (cpuid-1);
5070 }
5071 
5072 #ifdef SCHED_ULE
5073 extern struct cpu_group *cpu_top;              /* CPU topology */
5074 
5075 static int
5076 find_child_with_core(int cpu, struct cpu_group *grp)
5077 {
5078 	int i;
5079 
5080 	if (grp->cg_children == 0)
5081 		return -1;
5082 
5083 	MPASS(grp->cg_child);
5084 	for (i = 0; i < grp->cg_children; i++) {
5085 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5086 			return i;
5087 	}
5088 
5089 	return -1;
5090 }
5091 
5092 /*
5093  * Find the nth thread on the specified core
5094  */
5095 static int
5096 find_thread(int cpu, int thread_num)
5097 {
5098 	struct cpu_group *grp;
5099 	int i;
5100 	cpuset_t cs;
5101 
5102 	grp = cpu_top;
5103 	if (grp == NULL)
5104 		return cpu;
5105 	i = 0;
5106 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5107 		/* If the child only has one cpu, don't descend */
5108 		if (grp->cg_child[i].cg_count <= 1)
5109 			break;
5110 		grp = &grp->cg_child[i];
5111 	}
5112 
5113 	/* If they don't share at least an L2 cache, use the same CPU */
5114 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5115 		return cpu;
5116 
5117 	/* Now pick one */
5118 	CPU_COPY(&grp->cg_mask, &cs);
5119 	for (i = thread_num % grp->cg_count; i > 0; i--) {
5120 		MPASS(CPU_FFS(&cs));
5121 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5122 	}
5123 	MPASS(CPU_FFS(&cs));
5124 	return CPU_FFS(&cs) - 1;
5125 }
5126 #else
5127 static int
5128 find_thread(int cpu, int thread_num __unused)
5129 {
5130 	return cpu;
5131 }
5132 #endif
5133 
5134 static int
5135 get_thread_num(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5136 {
5137 	switch (type) {
5138 	case IFLIB_INTR_TX:
5139 		/* TX queues get threads on the same core as the corresponding RX queue */
5140 		/* XXX handle multiple RX threads per core and more than two threads per core */
5141 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5142 	case IFLIB_INTR_RX:
5143 	case IFLIB_INTR_RXTX:
5144 		/* RX queues get the first thread on their core */
5145 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5146 	default:
5147 		return -1;
5148 	}
5149 }
5150 #else
5151 #define get_thread_num(ctx, type, qid)	CPU_FIRST()
5152 #define find_thread(cpuid, tid)		CPU_FIRST()
5153 #define find_nth(ctx, gid)		CPU_FIRST()
5154 #endif
5155 
5156 /* Just to avoid copy/paste */
5157 static inline int
5158 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5159     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5160 {
5161 	int cpuid;
5162 	int err, tid;
5163 
5164 	cpuid = find_nth(ctx, qid);
5165 	tid = get_thread_num(ctx, type, qid);
5166 	MPASS(tid >= 0);
5167 	cpuid = find_thread(cpuid, tid);
5168 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5169 	if (err) {
5170 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5171 		return (err);
5172 	}
5173 #ifdef notyet
5174 	if (cpuid > ctx->ifc_cpuid_highest)
5175 		ctx->ifc_cpuid_highest = cpuid;
5176 #endif
5177 	return 0;
5178 }
5179 
5180 int
5181 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5182 						iflib_intr_type_t type, driver_filter_t *filter,
5183 						void *filter_arg, int qid, char *name)
5184 {
5185 	struct grouptask *gtask;
5186 	struct taskqgroup *tqg;
5187 	iflib_filter_info_t info;
5188 	gtask_fn_t *fn;
5189 	int tqrid, err;
5190 	driver_filter_t *intr_fast;
5191 	void *q;
5192 
5193 	info = &ctx->ifc_filter_info;
5194 	tqrid = rid;
5195 
5196 	switch (type) {
5197 	/* XXX merge tx/rx for netmap? */
5198 	case IFLIB_INTR_TX:
5199 		q = &ctx->ifc_txqs[qid];
5200 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5201 		gtask = &ctx->ifc_txqs[qid].ift_task;
5202 		tqg = qgroup_if_io_tqg;
5203 		fn = _task_fn_tx;
5204 		intr_fast = iflib_fast_intr;
5205 		GROUPTASK_INIT(gtask, 0, fn, q);
5206 		break;
5207 	case IFLIB_INTR_RX:
5208 		q = &ctx->ifc_rxqs[qid];
5209 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5210 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5211 		tqg = qgroup_if_io_tqg;
5212 		fn = _task_fn_rx;
5213 		intr_fast = iflib_fast_intr;
5214 		GROUPTASK_INIT(gtask, 0, fn, q);
5215 		break;
5216 	case IFLIB_INTR_RXTX:
5217 		q = &ctx->ifc_rxqs[qid];
5218 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5219 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5220 		tqg = qgroup_if_io_tqg;
5221 		fn = _task_fn_rx;
5222 		intr_fast = iflib_fast_intr_rxtx;
5223 		GROUPTASK_INIT(gtask, 0, fn, q);
5224 		break;
5225 	case IFLIB_INTR_ADMIN:
5226 		q = ctx;
5227 		tqrid = -1;
5228 		info = &ctx->ifc_filter_info;
5229 		gtask = &ctx->ifc_admin_task;
5230 		tqg = qgroup_if_config_tqg;
5231 		fn = _task_fn_admin;
5232 		intr_fast = iflib_fast_intr_ctx;
5233 		break;
5234 	default:
5235 		panic("unknown net intr type");
5236 	}
5237 
5238 	info->ifi_filter = filter;
5239 	info->ifi_filter_arg = filter_arg;
5240 	info->ifi_task = gtask;
5241 	info->ifi_ctx = q;
5242 
5243 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5244 	if (err != 0) {
5245 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5246 		return (err);
5247 	}
5248 	if (type == IFLIB_INTR_ADMIN)
5249 		return (0);
5250 
5251 	if (tqrid != -1) {
5252 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5253 		if (err)
5254 			return (err);
5255 	} else {
5256 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5257 	}
5258 
5259 	return (0);
5260 }
5261 
5262 void
5263 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5264 {
5265 	struct grouptask *gtask;
5266 	struct taskqgroup *tqg;
5267 	gtask_fn_t *fn;
5268 	void *q;
5269 	int irq_num = -1;
5270 	int err;
5271 
5272 	switch (type) {
5273 	case IFLIB_INTR_TX:
5274 		q = &ctx->ifc_txqs[qid];
5275 		gtask = &ctx->ifc_txqs[qid].ift_task;
5276 		tqg = qgroup_if_io_tqg;
5277 		fn = _task_fn_tx;
5278 		if (irq != NULL)
5279 			irq_num = rman_get_start(irq->ii_res);
5280 		break;
5281 	case IFLIB_INTR_RX:
5282 		q = &ctx->ifc_rxqs[qid];
5283 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5284 		tqg = qgroup_if_io_tqg;
5285 		fn = _task_fn_rx;
5286 		if (irq != NULL)
5287 			irq_num = rman_get_start(irq->ii_res);
5288 		break;
5289 	case IFLIB_INTR_IOV:
5290 		q = ctx;
5291 		gtask = &ctx->ifc_vflr_task;
5292 		tqg = qgroup_if_config_tqg;
5293 		fn = _task_fn_iov;
5294 		break;
5295 	default:
5296 		panic("unknown net intr type");
5297 	}
5298 	GROUPTASK_INIT(gtask, 0, fn, q);
5299 	if (irq_num != -1) {
5300 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5301 		if (err)
5302 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5303 	}
5304 	else {
5305 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5306 	}
5307 }
5308 
5309 void
5310 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5311 {
5312 	if (irq->ii_tag)
5313 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5314 
5315 	if (irq->ii_res)
5316 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5317 }
5318 
5319 static int
5320 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5321 {
5322 	iflib_txq_t txq = ctx->ifc_txqs;
5323 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5324 	if_irq_t irq = &ctx->ifc_legacy_irq;
5325 	iflib_filter_info_t info;
5326 	struct grouptask *gtask;
5327 	struct taskqgroup *tqg;
5328 	gtask_fn_t *fn;
5329 	int tqrid;
5330 	void *q;
5331 	int err;
5332 
5333 	q = &ctx->ifc_rxqs[0];
5334 	info = &rxq[0].ifr_filter_info;
5335 	gtask = &rxq[0].ifr_task;
5336 	tqg = qgroup_if_io_tqg;
5337 	tqrid = irq->ii_rid = *rid;
5338 	fn = _task_fn_rx;
5339 
5340 	ctx->ifc_flags |= IFC_LEGACY;
5341 	info->ifi_filter = filter;
5342 	info->ifi_filter_arg = filter_arg;
5343 	info->ifi_task = gtask;
5344 	info->ifi_ctx = ctx;
5345 
5346 	/* We allocate a single interrupt resource */
5347 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5348 		return (err);
5349 	GROUPTASK_INIT(gtask, 0, fn, q);
5350 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5351 
5352 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5353 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5354 	return (0);
5355 }
5356 
5357 void
5358 iflib_led_create(if_ctx_t ctx)
5359 {
5360 
5361 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5362 	    device_get_nameunit(ctx->ifc_dev));
5363 }
5364 
5365 void
5366 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5367 {
5368 
5369 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5370 }
5371 
5372 void
5373 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5374 {
5375 
5376 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5377 }
5378 
5379 void
5380 iflib_admin_intr_deferred(if_ctx_t ctx)
5381 {
5382 #ifdef INVARIANTS
5383 	struct grouptask *gtask;
5384 
5385 	gtask = &ctx->ifc_admin_task;
5386 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5387 #endif
5388 
5389 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5390 }
5391 
5392 void
5393 iflib_iov_intr_deferred(if_ctx_t ctx)
5394 {
5395 
5396 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5397 }
5398 
5399 void
5400 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5401 {
5402 
5403 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5404 }
5405 
5406 void
5407 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5408 	char *name)
5409 {
5410 
5411 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5412 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5413 }
5414 
5415 void
5416 iflib_config_gtask_deinit(struct grouptask *gtask)
5417 {
5418 
5419 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5420 }
5421 
5422 void
5423 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5424 {
5425 	if_t ifp = ctx->ifc_ifp;
5426 	iflib_txq_t txq = ctx->ifc_txqs;
5427 
5428 	if_setbaudrate(ifp, baudrate);
5429 	if (baudrate >= IF_Gbps(10))
5430 		ctx->ifc_flags |= IFC_PREFETCH;
5431 
5432 	/* If link down, disable watchdog */
5433 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5434 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5435 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5436 	}
5437 	ctx->ifc_link_state = link_state;
5438 	if_link_state_change(ifp, link_state);
5439 }
5440 
5441 static int
5442 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5443 {
5444 	int credits;
5445 #ifdef INVARIANTS
5446 	int credits_pre = txq->ift_cidx_processed;
5447 #endif
5448 
5449 	if (ctx->isc_txd_credits_update == NULL)
5450 		return (0);
5451 
5452 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5453 		return (0);
5454 
5455 	txq->ift_processed += credits;
5456 	txq->ift_cidx_processed += credits;
5457 
5458 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5459 	if (txq->ift_cidx_processed >= txq->ift_size)
5460 		txq->ift_cidx_processed -= txq->ift_size;
5461 	return (credits);
5462 }
5463 
5464 static int
5465 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5466 {
5467 
5468 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5469 	    budget));
5470 }
5471 
5472 void
5473 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5474 	const char *description, if_int_delay_info_t info,
5475 	int offset, int value)
5476 {
5477 	info->iidi_ctx = ctx;
5478 	info->iidi_offset = offset;
5479 	info->iidi_value = value;
5480 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5481 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5482 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5483 	    info, 0, iflib_sysctl_int_delay, "I", description);
5484 }
5485 
5486 struct mtx *
5487 iflib_ctx_lock_get(if_ctx_t ctx)
5488 {
5489 
5490 	return (&ctx->ifc_mtx);
5491 }
5492 
5493 static int
5494 iflib_msix_init(if_ctx_t ctx)
5495 {
5496 	device_t dev = ctx->ifc_dev;
5497 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5498 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5499 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5500 	int iflib_num_tx_queues, iflib_num_rx_queues;
5501 	int err, admincnt, bar;
5502 
5503 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5504 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5505 
5506 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5507 
5508 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5509 	admincnt = sctx->isc_admin_intrcnt;
5510 	/* Override by global tuneable */
5511 	{
5512 		int i;
5513 		size_t len = sizeof(i);
5514 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5515 		if (err == 0) {
5516 			if (i == 0)
5517 				goto msi;
5518 		}
5519 		else {
5520 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5521 		}
5522 	}
5523 	/* Override by tuneable */
5524 	if (scctx->isc_disable_msix)
5525 		goto msi;
5526 
5527 	/*
5528 	** When used in a virtualized environment
5529 	** PCI BUSMASTER capability may not be set
5530 	** so explicity set it here and rewrite
5531 	** the ENABLE in the MSIX control register
5532 	** at this point to cause the host to
5533 	** successfully initialize us.
5534 	*/
5535 	{
5536 		int msix_ctrl, rid;
5537 
5538  		pci_enable_busmaster(dev);
5539 		rid = 0;
5540 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5541 			rid += PCIR_MSIX_CTRL;
5542 			msix_ctrl = pci_read_config(dev, rid, 2);
5543 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5544 			pci_write_config(dev, rid, msix_ctrl, 2);
5545 		} else {
5546 			device_printf(dev, "PCIY_MSIX capability not found; "
5547 			                   "or rid %d == 0.\n", rid);
5548 			goto msi;
5549 		}
5550 	}
5551 
5552 	/*
5553 	 * bar == -1 => "trust me I know what I'm doing"
5554 	 * Some drivers are for hardware that is so shoddily
5555 	 * documented that no one knows which bars are which
5556 	 * so the developer has to map all bars. This hack
5557 	 * allows shoddy garbage to use msix in this framework.
5558 	 */
5559 	if (bar != -1) {
5560 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5561 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5562 		if (ctx->ifc_msix_mem == NULL) {
5563 			/* May not be enabled */
5564 			device_printf(dev, "Unable to map MSIX table \n");
5565 			goto msi;
5566 		}
5567 	}
5568 	/* First try MSI/X */
5569 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5570 		device_printf(dev, "System has MSIX disabled \n");
5571 		bus_release_resource(dev, SYS_RES_MEMORY,
5572 		    bar, ctx->ifc_msix_mem);
5573 		ctx->ifc_msix_mem = NULL;
5574 		goto msi;
5575 	}
5576 #if IFLIB_DEBUG
5577 	/* use only 1 qset in debug mode */
5578 	queuemsgs = min(msgs - admincnt, 1);
5579 #else
5580 	queuemsgs = msgs - admincnt;
5581 #endif
5582 #ifdef RSS
5583 	queues = imin(queuemsgs, rss_getnumbuckets());
5584 #else
5585 	queues = queuemsgs;
5586 #endif
5587 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5588 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5589 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5590 #ifdef  RSS
5591 	/* If we're doing RSS, clamp at the number of RSS buckets */
5592 	if (queues > rss_getnumbuckets())
5593 		queues = rss_getnumbuckets();
5594 #endif
5595 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5596 		rx_queues = iflib_num_rx_queues;
5597 	else
5598 		rx_queues = queues;
5599 
5600 	if (rx_queues > scctx->isc_nrxqsets)
5601 		rx_queues = scctx->isc_nrxqsets;
5602 
5603 	/*
5604 	 * We want this to be all logical CPUs by default
5605 	 */
5606 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5607 		tx_queues = iflib_num_tx_queues;
5608 	else
5609 		tx_queues = mp_ncpus;
5610 
5611 	if (tx_queues > scctx->isc_ntxqsets)
5612 		tx_queues = scctx->isc_ntxqsets;
5613 
5614 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5615 #ifdef INVARIANTS
5616 		if (tx_queues != rx_queues)
5617 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5618 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5619 #endif
5620 		tx_queues = min(rx_queues, tx_queues);
5621 		rx_queues = min(rx_queues, tx_queues);
5622 	}
5623 
5624 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5625 
5626 	vectors = rx_queues + admincnt;
5627 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5628 		device_printf(dev,
5629 					  "Using MSIX interrupts with %d vectors\n", vectors);
5630 		scctx->isc_vectors = vectors;
5631 		scctx->isc_nrxqsets = rx_queues;
5632 		scctx->isc_ntxqsets = tx_queues;
5633 		scctx->isc_intr = IFLIB_INTR_MSIX;
5634 
5635 		return (vectors);
5636 	} else {
5637 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5638 	}
5639 msi:
5640 	vectors = pci_msi_count(dev);
5641 	scctx->isc_nrxqsets = 1;
5642 	scctx->isc_ntxqsets = 1;
5643 	scctx->isc_vectors = vectors;
5644 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5645 		device_printf(dev,"Using an MSI interrupt\n");
5646 		scctx->isc_intr = IFLIB_INTR_MSI;
5647 	} else {
5648 		device_printf(dev,"Using a Legacy interrupt\n");
5649 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5650 	}
5651 
5652 	return (vectors);
5653 }
5654 
5655 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5656 
5657 static int
5658 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5659 {
5660 	int rc;
5661 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5662 	struct sbuf *sb;
5663 	char *ring_state = "UNKNOWN";
5664 
5665 	/* XXX needed ? */
5666 	rc = sysctl_wire_old_buffer(req, 0);
5667 	MPASS(rc == 0);
5668 	if (rc != 0)
5669 		return (rc);
5670 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5671 	MPASS(sb != NULL);
5672 	if (sb == NULL)
5673 		return (ENOMEM);
5674 	if (state[3] <= 3)
5675 		ring_state = ring_states[state[3]];
5676 
5677 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5678 		    state[0], state[1], state[2], ring_state);
5679 	rc = sbuf_finish(sb);
5680 	sbuf_delete(sb);
5681         return(rc);
5682 }
5683 
5684 enum iflib_ndesc_handler {
5685 	IFLIB_NTXD_HANDLER,
5686 	IFLIB_NRXD_HANDLER,
5687 };
5688 
5689 static int
5690 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5691 {
5692 	if_ctx_t ctx = (void *)arg1;
5693 	enum iflib_ndesc_handler type = arg2;
5694 	char buf[256] = {0};
5695 	qidx_t *ndesc;
5696 	char *p, *next;
5697 	int nqs, rc, i;
5698 
5699 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5700 
5701 	nqs = 8;
5702 	switch(type) {
5703 	case IFLIB_NTXD_HANDLER:
5704 		ndesc = ctx->ifc_sysctl_ntxds;
5705 		if (ctx->ifc_sctx)
5706 			nqs = ctx->ifc_sctx->isc_ntxqs;
5707 		break;
5708 	case IFLIB_NRXD_HANDLER:
5709 		ndesc = ctx->ifc_sysctl_nrxds;
5710 		if (ctx->ifc_sctx)
5711 			nqs = ctx->ifc_sctx->isc_nrxqs;
5712 		break;
5713 	}
5714 	if (nqs == 0)
5715 		nqs = 8;
5716 
5717 	for (i=0; i<8; i++) {
5718 		if (i >= nqs)
5719 			break;
5720 		if (i)
5721 			strcat(buf, ",");
5722 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5723 	}
5724 
5725 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5726 	if (rc || req->newptr == NULL)
5727 		return rc;
5728 
5729 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5730 	    i++, p = strsep(&next, " ,")) {
5731 		ndesc[i] = strtoul(p, NULL, 10);
5732 	}
5733 
5734 	return(rc);
5735 }
5736 
5737 #define NAME_BUFLEN 32
5738 static void
5739 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5740 {
5741         device_t dev = iflib_get_dev(ctx);
5742 	struct sysctl_oid_list *child, *oid_list;
5743 	struct sysctl_ctx_list *ctx_list;
5744 	struct sysctl_oid *node;
5745 
5746 	ctx_list = device_get_sysctl_ctx(dev);
5747 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5748 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5749 						      CTLFLAG_RD, NULL, "IFLIB fields");
5750 	oid_list = SYSCTL_CHILDREN(node);
5751 
5752 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5753 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5754 		       "driver version");
5755 
5756 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5757 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5758 			"# of txqs to use, 0 => use default #");
5759 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5760 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5761 			"# of rxqs to use, 0 => use default #");
5762 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5763 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5764                        "permit #txq != #rxq");
5765 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5766                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5767                       "disable MSIX (default 0)");
5768 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5769 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5770                        "set the rx budget");
5771 
5772 	/* XXX change for per-queue sizes */
5773 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5774 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5775                        mp_ndesc_handler, "A",
5776                        "list of # of tx descriptors to use, 0 = use default #");
5777 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5778 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5779                        mp_ndesc_handler, "A",
5780                        "list of # of rx descriptors to use, 0 = use default #");
5781 }
5782 
5783 static void
5784 iflib_add_device_sysctl_post(if_ctx_t ctx)
5785 {
5786 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5787 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5788         device_t dev = iflib_get_dev(ctx);
5789 	struct sysctl_oid_list *child;
5790 	struct sysctl_ctx_list *ctx_list;
5791 	iflib_fl_t fl;
5792 	iflib_txq_t txq;
5793 	iflib_rxq_t rxq;
5794 	int i, j;
5795 	char namebuf[NAME_BUFLEN];
5796 	char *qfmt;
5797 	struct sysctl_oid *queue_node, *fl_node, *node;
5798 	struct sysctl_oid_list *queue_list, *fl_list;
5799 	ctx_list = device_get_sysctl_ctx(dev);
5800 
5801 	node = ctx->ifc_sysctl_node;
5802 	child = SYSCTL_CHILDREN(node);
5803 
5804 	if (scctx->isc_ntxqsets > 100)
5805 		qfmt = "txq%03d";
5806 	else if (scctx->isc_ntxqsets > 10)
5807 		qfmt = "txq%02d";
5808 	else
5809 		qfmt = "txq%d";
5810 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5811 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5812 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5813 					     CTLFLAG_RD, NULL, "Queue Name");
5814 		queue_list = SYSCTL_CHILDREN(queue_node);
5815 #if MEMORY_LOGGING
5816 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5817 				CTLFLAG_RD,
5818 				&txq->ift_dequeued, "total mbufs freed");
5819 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5820 				CTLFLAG_RD,
5821 				&txq->ift_enqueued, "total mbufs enqueued");
5822 #endif
5823 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5824 				   CTLFLAG_RD,
5825 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5826 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5827 				   CTLFLAG_RD,
5828 				   &txq->ift_pullups, "# of times m_pullup was called");
5829 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5830 				   CTLFLAG_RD,
5831 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5832 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5833 				   CTLFLAG_RD,
5834 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5835 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5836 				   CTLFLAG_RD,
5837 				   &txq->ift_map_failed, "# of times dma map failed");
5838 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5839 				   CTLFLAG_RD,
5840 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5841 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5842 				   CTLFLAG_RD,
5843 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5844 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5845 				   CTLFLAG_RD,
5846 				   &txq->ift_pidx, 1, "Producer Index");
5847 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5848 				   CTLFLAG_RD,
5849 				   &txq->ift_cidx, 1, "Consumer Index");
5850 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5851 				   CTLFLAG_RD,
5852 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5853 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5854 				   CTLFLAG_RD,
5855 				   &txq->ift_in_use, 1, "descriptors in use");
5856 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5857 				   CTLFLAG_RD,
5858 				   &txq->ift_processed, "descriptors procesed for clean");
5859 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5860 				   CTLFLAG_RD,
5861 				   &txq->ift_cleaned, "total cleaned");
5862 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5863 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5864 				0, mp_ring_state_handler, "A", "soft ring state");
5865 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5866 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5867 				       "# of enqueues to the mp_ring for this queue");
5868 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5869 				       CTLFLAG_RD, &txq->ift_br->drops,
5870 				       "# of drops in the mp_ring for this queue");
5871 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5872 				       CTLFLAG_RD, &txq->ift_br->starts,
5873 				       "# of normal consumer starts in the mp_ring for this queue");
5874 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5875 				       CTLFLAG_RD, &txq->ift_br->stalls,
5876 					       "# of consumer stalls in the mp_ring for this queue");
5877 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5878 			       CTLFLAG_RD, &txq->ift_br->restarts,
5879 				       "# of consumer restarts in the mp_ring for this queue");
5880 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5881 				       CTLFLAG_RD, &txq->ift_br->abdications,
5882 				       "# of consumer abdications in the mp_ring for this queue");
5883 	}
5884 
5885 	if (scctx->isc_nrxqsets > 100)
5886 		qfmt = "rxq%03d";
5887 	else if (scctx->isc_nrxqsets > 10)
5888 		qfmt = "rxq%02d";
5889 	else
5890 		qfmt = "rxq%d";
5891 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5892 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5893 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5894 					     CTLFLAG_RD, NULL, "Queue Name");
5895 		queue_list = SYSCTL_CHILDREN(queue_node);
5896 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5897 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5898 				       CTLFLAG_RD,
5899 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5900 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5901 				       CTLFLAG_RD,
5902 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5903 		}
5904 
5905 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5906 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5907 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5908 						     CTLFLAG_RD, NULL, "freelist Name");
5909 			fl_list = SYSCTL_CHILDREN(fl_node);
5910 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5911 				       CTLFLAG_RD,
5912 				       &fl->ifl_pidx, 1, "Producer Index");
5913 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5914 				       CTLFLAG_RD,
5915 				       &fl->ifl_cidx, 1, "Consumer Index");
5916 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5917 				       CTLFLAG_RD,
5918 				       &fl->ifl_credits, 1, "credits available");
5919 #if MEMORY_LOGGING
5920 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5921 					CTLFLAG_RD,
5922 					&fl->ifl_m_enqueued, "mbufs allocated");
5923 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5924 					CTLFLAG_RD,
5925 					&fl->ifl_m_dequeued, "mbufs freed");
5926 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5927 					CTLFLAG_RD,
5928 					&fl->ifl_cl_enqueued, "clusters allocated");
5929 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5930 					CTLFLAG_RD,
5931 					&fl->ifl_cl_dequeued, "clusters freed");
5932 #endif
5933 
5934 		}
5935 	}
5936 
5937 }
5938 
5939 #ifndef __NO_STRICT_ALIGNMENT
5940 static struct mbuf *
5941 iflib_fixup_rx(struct mbuf *m)
5942 {
5943 	struct mbuf *n;
5944 
5945 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5946 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5947 		m->m_data += ETHER_HDR_LEN;
5948 		n = m;
5949 	} else {
5950 		MGETHDR(n, M_NOWAIT, MT_DATA);
5951 		if (n == NULL) {
5952 			m_freem(m);
5953 			return (NULL);
5954 		}
5955 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5956 		m->m_data += ETHER_HDR_LEN;
5957 		m->m_len -= ETHER_HDR_LEN;
5958 		n->m_len = ETHER_HDR_LEN;
5959 		M_MOVE_PKTHDR(n, m);
5960 		n->m_next = m;
5961 	}
5962 	return (n);
5963 }
5964 #endif
5965