xref: /freebsd/sys/net/iflib.c (revision f0483545503a78e16e256d46d458a2faae2f07ea)
1 /*-
2  * Copyright (c) 2014-2018, 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/jail.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/md5.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/kobj.h>
47 #include <sys/rman.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
57 
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
62 #include <net/bpf.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
92 
93 #include "ifdi_if.h"
94 
95 #ifdef PCI_IOV
96 #include <dev/pci/pci_iov.h>
97 #endif
98 
99 #include <sys/bitstring.h>
100 /*
101  * enable accounting of every mbuf as it comes in to and goes out of
102  * iflib's software descriptor references
103  */
104 #define MEMORY_LOGGING 0
105 /*
106  * Enable mbuf vectors for compressing long mbuf chains
107  */
108 
109 /*
110  * NB:
111  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
112  *   we prefetch needs to be determined by the time spent in m_free vis a vis
113  *   the cost of a prefetch. This will of course vary based on the workload:
114  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
115  *        is quite expensive, thus suggesting very little prefetch.
116  *      - small packet forwarding which is just returning a single mbuf to
117  *        UMA will typically be very fast vis a vis the cost of a memory
118  *        access.
119  */
120 
121 
122 /*
123  * File organization:
124  *  - private structures
125  *  - iflib private utility functions
126  *  - ifnet functions
127  *  - vlan registry and other exported functions
128  *  - iflib public core functions
129  *
130  *
131  */
132 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
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 sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	iflib_txq_t ifc_txqs;
170 	iflib_rxq_t ifc_rxqs;
171 	uint32_t ifc_if_flags;
172 	uint32_t ifc_flags;
173 	uint32_t ifc_max_fl_buf_size;
174 
175 	int ifc_link_state;
176 	int ifc_link_irq;
177 	int ifc_watchdog_events;
178 	struct cdev *ifc_led_dev;
179 	struct resource *ifc_msix_mem;
180 
181 	struct if_irq ifc_legacy_irq;
182 	struct grouptask ifc_admin_task;
183 	struct grouptask ifc_vflr_task;
184 	struct iflib_filter_info ifc_filter_info;
185 	struct ifmedia	ifc_media;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 
194 	qidx_t ifc_sysctl_ntxds[8];
195 	qidx_t ifc_sysctl_nrxds[8];
196 	struct if_txrx ifc_txrx;
197 #define isc_txd_encap  ifc_txrx.ift_txd_encap
198 #define isc_txd_flush  ifc_txrx.ift_txd_flush
199 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
200 #define isc_rxd_available ifc_txrx.ift_rxd_available
201 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
207 	eventhandler_tag ifc_vlan_attach_event;
208 	eventhandler_tag ifc_vlan_detach_event;
209 	uint8_t ifc_mac[ETHER_ADDR_LEN];
210 	char ifc_mtx_name[16];
211 };
212 
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (&ctx->ifc_media);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 #define IFLIB_RX_COPY_THRESH		128
293 #define IFLIB_MAX_RX_REFRESH		32
294 /* The minimum descriptors per second before we start coalescing */
295 #define IFLIB_MIN_DESC_SEC		16384
296 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
297 #define IFLIB_QUEUE_IDLE		0
298 #define IFLIB_QUEUE_HUNG		1
299 #define IFLIB_QUEUE_WORKING		2
300 /* maximum number of txqs that can share an rx interrupt */
301 #define IFLIB_MAX_TX_SHARED_INTR	4
302 
303 /* this should really scale with ring size - this is a fairly arbitrary value */
304 #define TX_BATCH_SIZE			32
305 
306 #define IFLIB_RESTART_BUDGET		8
307 
308 
309 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
310 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
311 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
312 struct iflib_txq {
313 	qidx_t		ift_in_use;
314 	qidx_t		ift_cidx;
315 	qidx_t		ift_cidx_processed;
316 	qidx_t		ift_pidx;
317 	uint8_t		ift_gen;
318 	uint8_t		ift_br_offset;
319 	uint16_t	ift_npending;
320 	uint16_t	ift_db_pending;
321 	uint16_t	ift_rs_pending;
322 	/* implicit pad */
323 	uint8_t		ift_txd_size[8];
324 	uint64_t	ift_processed;
325 	uint64_t	ift_cleaned;
326 	uint64_t	ift_cleaned_prev;
327 #if MEMORY_LOGGING
328 	uint64_t	ift_enqueued;
329 	uint64_t	ift_dequeued;
330 #endif
331 	uint64_t	ift_no_tx_dma_setup;
332 	uint64_t	ift_no_desc_avail;
333 	uint64_t	ift_mbuf_defrag_failed;
334 	uint64_t	ift_mbuf_defrag;
335 	uint64_t	ift_map_failed;
336 	uint64_t	ift_txd_encap_efbig;
337 	uint64_t	ift_pullups;
338 	uint64_t	ift_last_timer_tick;
339 
340 	struct mtx	ift_mtx;
341 	struct mtx	ift_db_mtx;
342 
343 	/* constant values */
344 	if_ctx_t	ift_ctx;
345 	struct ifmp_ring        *ift_br;
346 	struct grouptask	ift_task;
347 	qidx_t		ift_size;
348 	uint16_t	ift_id;
349 	struct callout	ift_timer;
350 
351 	if_txsd_vec_t	ift_sds;
352 	uint8_t		ift_qstatus;
353 	uint8_t		ift_closed;
354 	uint8_t		ift_update_freq;
355 	struct iflib_filter_info ift_filter_info;
356 	bus_dma_tag_t	ift_buf_tag;
357 	bus_dma_tag_t	ift_tso_buf_tag;
358 	iflib_dma_info_t	ift_ifdi;
359 #define MTX_NAME_LEN 16
360 	char                    ift_mtx_name[MTX_NAME_LEN];
361 	char                    ift_db_mtx_name[MTX_NAME_LEN];
362 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
363 #ifdef IFLIB_DIAGNOSTICS
364 	uint64_t ift_cpu_exec_count[256];
365 #endif
366 } __aligned(CACHE_LINE_SIZE);
367 
368 struct iflib_fl {
369 	qidx_t		ifl_cidx;
370 	qidx_t		ifl_pidx;
371 	qidx_t		ifl_credits;
372 	uint8_t		ifl_gen;
373 	uint8_t		ifl_rxd_size;
374 #if MEMORY_LOGGING
375 	uint64_t	ifl_m_enqueued;
376 	uint64_t	ifl_m_dequeued;
377 	uint64_t	ifl_cl_enqueued;
378 	uint64_t	ifl_cl_dequeued;
379 #endif
380 	/* implicit pad */
381 
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	/* If there is a separate completion queue -
425 	 * these are the cq cidx and pidx. Otherwise
426 	 * these are unused.
427 	 */
428 	qidx_t		ifr_size;
429 	qidx_t		ifr_cq_cidx;
430 	qidx_t		ifr_cq_pidx;
431 	uint8_t		ifr_cq_gen;
432 	uint8_t		ifr_fl_offset;
433 
434 	if_ctx_t	ifr_ctx;
435 	iflib_fl_t	ifr_fl;
436 	uint64_t	ifr_rx_irq;
437 	uint16_t	ifr_id;
438 	uint8_t		ifr_lro_enabled;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	struct lro_ctrl			ifr_lc;
443 	struct grouptask        ifr_task;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	struct mbuf **ifsd_m;
457 	iflib_fl_t ifsd_fl;
458 	qidx_t ifsd_cidx;
459 } *if_rxsd_t;
460 
461 /* multiple of word size */
462 #ifdef __LP64__
463 #define PKT_INFO_SIZE	6
464 #define RXD_INFO_SIZE	5
465 #define PKT_TYPE uint64_t
466 #else
467 #define PKT_INFO_SIZE	11
468 #define RXD_INFO_SIZE	8
469 #define PKT_TYPE uint32_t
470 #endif
471 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
472 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
473 
474 typedef struct if_pkt_info_pad {
475 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
476 } *if_pkt_info_pad_t;
477 typedef struct if_rxd_info_pad {
478 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
479 } *if_rxd_info_pad_t;
480 
481 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
482 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
483 
484 
485 static inline void
486 pkt_info_zero(if_pkt_info_t pi)
487 {
488 	if_pkt_info_pad_t pi_pad;
489 
490 	pi_pad = (if_pkt_info_pad_t)pi;
491 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
492 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
493 #ifndef __LP64__
494 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
495 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
496 #endif
497 }
498 
499 static device_method_t iflib_pseudo_methods[] = {
500 	DEVMETHOD(device_attach, noop_attach),
501 	DEVMETHOD(device_detach, iflib_pseudo_detach),
502 	DEVMETHOD_END
503 };
504 
505 driver_t iflib_pseudodriver = {
506 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
507 };
508 
509 static inline void
510 rxd_info_zero(if_rxd_info_t ri)
511 {
512 	if_rxd_info_pad_t ri_pad;
513 	int i;
514 
515 	ri_pad = (if_rxd_info_pad_t)ri;
516 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
517 		ri_pad->rxd_val[i] = 0;
518 		ri_pad->rxd_val[i+1] = 0;
519 		ri_pad->rxd_val[i+2] = 0;
520 		ri_pad->rxd_val[i+3] = 0;
521 	}
522 #ifdef __LP64__
523 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
524 #endif
525 }
526 
527 /*
528  * Only allow a single packet to take up most 1/nth of the tx ring
529  */
530 #define MAX_SINGLE_PACKET_FRACTION 12
531 #define IF_BAD_DMA (bus_addr_t)-1
532 
533 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
534 
535 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
536 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
537 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
538 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
539 
540 
541 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
542 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
543 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
544 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
545 
546 
547 
548 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
549 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
550 
551 void
552 iflib_set_detach(if_ctx_t ctx)
553 {
554 	STATE_LOCK(ctx);
555 	ctx->ifc_flags |= IFC_IN_DETACH;
556 	STATE_UNLOCK(ctx);
557 }
558 
559 /* Our boot-time initialization hook */
560 static int	iflib_module_event_handler(module_t, int, void *);
561 
562 static moduledata_t iflib_moduledata = {
563 	"iflib",
564 	iflib_module_event_handler,
565 	NULL
566 };
567 
568 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
569 MODULE_VERSION(iflib, 1);
570 
571 MODULE_DEPEND(iflib, pci, 1, 1, 1);
572 MODULE_DEPEND(iflib, ether, 1, 1, 1);
573 
574 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
575 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
576 
577 #ifndef IFLIB_DEBUG_COUNTERS
578 #ifdef INVARIANTS
579 #define IFLIB_DEBUG_COUNTERS 1
580 #else
581 #define IFLIB_DEBUG_COUNTERS 0
582 #endif /* !INVARIANTS */
583 #endif
584 
585 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
586                    "iflib driver parameters");
587 
588 /*
589  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
590  */
591 static int iflib_min_tx_latency = 0;
592 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
593 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
594 static int iflib_no_tx_batch = 0;
595 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
596 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
597 
598 
599 #if IFLIB_DEBUG_COUNTERS
600 
601 static int iflib_tx_seen;
602 static int iflib_tx_sent;
603 static int iflib_tx_encap;
604 static int iflib_rx_allocs;
605 static int iflib_fl_refills;
606 static int iflib_fl_refills_large;
607 static int iflib_tx_frees;
608 
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
610 		   &iflib_tx_seen, 0, "# tx mbufs seen");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
612 		   &iflib_tx_sent, 0, "# tx mbufs sent");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
614 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
616 		   &iflib_tx_frees, 0, "# tx frees");
617 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
618 		   &iflib_rx_allocs, 0, "# rx allocations");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
620 		   &iflib_fl_refills, 0, "# refills");
621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
622 		   &iflib_fl_refills_large, 0, "# large refills");
623 
624 
625 static int iflib_txq_drain_flushing;
626 static int iflib_txq_drain_oactive;
627 static int iflib_txq_drain_notready;
628 
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
630 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
632 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
634 		   &iflib_txq_drain_notready, 0, "# drain notready");
635 
636 
637 static int iflib_encap_load_mbuf_fail;
638 static int iflib_encap_pad_mbuf_fail;
639 static int iflib_encap_txq_avail_fail;
640 static int iflib_encap_txd_encap_fail;
641 
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
643 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
645 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
647 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
649 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
650 
651 static int iflib_task_fn_rxs;
652 static int iflib_rx_intr_enables;
653 static int iflib_fast_intrs;
654 static int iflib_rx_unavail;
655 static int iflib_rx_ctx_inactive;
656 static int iflib_rx_if_input;
657 static int iflib_rx_mbuf_null;
658 static int iflib_rxd_flush;
659 
660 static int iflib_verbose_debug;
661 
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_if_input, CTLFLAG_RD,
673 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
675 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
677 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
678 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
679 		   &iflib_verbose_debug, 0, "enable verbose debugging");
680 
681 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
682 static void
683 iflib_debug_reset(void)
684 {
685 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
686 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
687 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
688 		iflib_txq_drain_notready =
689 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
690 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
691 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
692 		iflib_rx_unavail =
693 		iflib_rx_ctx_inactive = iflib_rx_if_input =
694 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
695 }
696 
697 #else
698 #define DBG_COUNTER_INC(name)
699 static void iflib_debug_reset(void) {}
700 #endif
701 
702 #define IFLIB_DEBUG 0
703 
704 static void iflib_tx_structures_free(if_ctx_t ctx);
705 static void iflib_rx_structures_free(if_ctx_t ctx);
706 static int iflib_queues_alloc(if_ctx_t ctx);
707 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
708 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
709 static int iflib_qset_structures_setup(if_ctx_t ctx);
710 static int iflib_msix_init(if_ctx_t ctx);
711 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
712 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
713 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
714 #ifdef ALTQ
715 static void iflib_altq_if_start(if_t ifp);
716 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
717 #endif
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_if_init_locked(if_ctx_t ctx);
725 static void iflib_free_intr_mem(if_ctx_t ctx);
726 #ifndef __NO_STRICT_ALIGNMENT
727 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
728 #endif
729 
730 NETDUMP_DEFINE(iflib);
731 
732 #ifdef DEV_NETMAP
733 #include <sys/selinfo.h>
734 #include <net/netmap.h>
735 #include <dev/netmap/netmap_kern.h>
736 
737 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
738 
739 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
740 
741 /*
742  * device-specific sysctl variables:
743  *
744  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
745  *	During regular operations the CRC is stripped, but on some
746  *	hardware reception of frames not multiple of 64 is slower,
747  *	so using crcstrip=0 helps in benchmarks.
748  *
749  * iflib_rx_miss, iflib_rx_miss_bufs:
750  *	count packets that might be missed due to lost interrupts.
751  */
752 SYSCTL_DECL(_dev_netmap);
753 /*
754  * The xl driver by default strips CRCs and we do not override it.
755  */
756 
757 int iflib_crcstrip = 1;
758 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
759     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
760 
761 int iflib_rx_miss, iflib_rx_miss_bufs;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
763     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
765     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
766 
767 /*
768  * Register/unregister. We are already under netmap lock.
769  * Only called on the first register or the last unregister.
770  */
771 static int
772 iflib_netmap_register(struct netmap_adapter *na, int onoff)
773 {
774 	struct ifnet *ifp = na->ifp;
775 	if_ctx_t ctx = ifp->if_softc;
776 	int status;
777 
778 	CTX_LOCK(ctx);
779 	IFDI_INTR_DISABLE(ctx);
780 
781 	/* Tell the stack that the interface is no longer active */
782 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
783 
784 	if (!CTX_IS_VF(ctx))
785 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
786 
787 	/* enable or disable flags and callbacks in na and ifp */
788 	if (onoff) {
789 		nm_set_native_flags(na);
790 	} else {
791 		nm_clear_native_flags(na);
792 	}
793 	iflib_stop(ctx);
794 	iflib_init_locked(ctx);
795 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
796 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
797 	if (status)
798 		nm_clear_native_flags(na);
799 	CTX_UNLOCK(ctx);
800 	return (status);
801 }
802 
803 static int
804 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
805 {
806 	struct netmap_adapter *na = kring->na;
807 	u_int const lim = kring->nkr_num_slots - 1;
808 	u_int head = kring->rhead;
809 	struct netmap_ring *ring = kring->ring;
810 	bus_dmamap_t *map;
811 	struct if_rxd_update iru;
812 	if_ctx_t ctx = rxq->ifr_ctx;
813 	iflib_fl_t fl = &rxq->ifr_fl[0];
814 	uint32_t refill_pidx, nic_i;
815 #if IFLIB_DEBUG_COUNTERS
816 	int rf_count = 0;
817 #endif
818 
819 	if (nm_i == head && __predict_true(!init))
820 		return 0;
821 	iru_init(&iru, rxq, 0 /* flid */);
822 	map = fl->ifl_sds.ifsd_map;
823 	refill_pidx = netmap_idx_k2n(kring, nm_i);
824 	/*
825 	 * IMPORTANT: we must leave one free slot in the ring,
826 	 * so move head back by one unit
827 	 */
828 	head = nm_prev(head, lim);
829 	nic_i = UINT_MAX;
830 	DBG_COUNTER_INC(fl_refills);
831 	while (nm_i != head) {
832 #if IFLIB_DEBUG_COUNTERS
833 		if (++rf_count == 9)
834 			DBG_COUNTER_INC(fl_refills_large);
835 #endif
836 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
837 			struct netmap_slot *slot = &ring->slot[nm_i];
838 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
839 			uint32_t nic_i_dma = refill_pidx;
840 			nic_i = netmap_idx_k2n(kring, nm_i);
841 
842 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
843 
844 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
845 			        return netmap_ring_reinit(kring);
846 
847 			fl->ifl_vm_addrs[tmp_pidx] = addr;
848 			if (__predict_false(init) && map) {
849 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
850 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
851 				/* buffer has changed, reload map */
852 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
853 			}
854 			slot->flags &= ~NS_BUF_CHANGED;
855 
856 			nm_i = nm_next(nm_i, lim);
857 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
858 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
859 				continue;
860 
861 			iru.iru_pidx = refill_pidx;
862 			iru.iru_count = tmp_pidx+1;
863 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
864 
865 			refill_pidx = nic_i;
866 			if (map == NULL)
867 				continue;
868 
869 			for (int n = 0; n < iru.iru_count; n++) {
870 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
871 						BUS_DMASYNC_PREREAD);
872 				/* XXX - change this to not use the netmap func*/
873 				nic_i_dma = nm_next(nic_i_dma, lim);
874 			}
875 		}
876 	}
877 	kring->nr_hwcur = head;
878 
879 	if (map)
880 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
881 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
882 	if (__predict_true(nic_i != UINT_MAX)) {
883 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
884 		DBG_COUNTER_INC(rxd_flush);
885 	}
886 	return (0);
887 }
888 
889 /*
890  * Reconcile kernel and user view of the transmit ring.
891  *
892  * All information is in the kring.
893  * Userspace wants to send packets up to the one before kring->rhead,
894  * kernel knows kring->nr_hwcur is the first unsent packet.
895  *
896  * Here we push packets out (as many as possible), and possibly
897  * reclaim buffers from previously completed transmission.
898  *
899  * The caller (netmap) guarantees that there is only one instance
900  * running at any time. Any interference with other driver
901  * methods should be handled by the individual drivers.
902  */
903 static int
904 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
905 {
906 	struct netmap_adapter *na = kring->na;
907 	struct ifnet *ifp = na->ifp;
908 	struct netmap_ring *ring = kring->ring;
909 	u_int nm_i;	/* index into the netmap kring */
910 	u_int nic_i;	/* index into the NIC ring */
911 	u_int n;
912 	u_int const lim = kring->nkr_num_slots - 1;
913 	u_int const head = kring->rhead;
914 	struct if_pkt_info pi;
915 
916 	/*
917 	 * interrupts on every tx packet are expensive so request
918 	 * them every half ring, or where NS_REPORT is set
919 	 */
920 	u_int report_frequency = kring->nkr_num_slots >> 1;
921 	/* device-specific */
922 	if_ctx_t ctx = ifp->if_softc;
923 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
924 
925 	bus_dmamap_sync(txq->ift_buf_tag, txq->ift_ifdi->idi_map,
926 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
927 
928 	/*
929 	 * First part: process new packets to send.
930 	 * nm_i is the current index in the netmap kring,
931 	 * nic_i is the corresponding index in the NIC ring.
932 	 *
933 	 * If we have packets to send (nm_i != head)
934 	 * iterate over the netmap ring, fetch length and update
935 	 * the corresponding slot in the NIC ring. Some drivers also
936 	 * need to update the buffer's physical address in the NIC slot
937 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
938 	 *
939 	 * The netmap_reload_map() calls is especially expensive,
940 	 * even when (as in this case) the tag is 0, so do only
941 	 * when the buffer has actually changed.
942 	 *
943 	 * If possible do not set the report/intr bit on all slots,
944 	 * but only a few times per ring or when NS_REPORT is set.
945 	 *
946 	 * Finally, on 10G and faster drivers, it might be useful
947 	 * to prefetch the next slot and txr entry.
948 	 */
949 
950 	nm_i = kring->nr_hwcur;
951 	if (nm_i != head) {	/* we have new packets to send */
952 		pkt_info_zero(&pi);
953 		pi.ipi_segs = txq->ift_segs;
954 		pi.ipi_qsidx = kring->ring_id;
955 		nic_i = netmap_idx_k2n(kring, nm_i);
956 
957 		__builtin_prefetch(&ring->slot[nm_i]);
958 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
959 		if (txq->ift_sds.ifsd_map)
960 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
961 
962 		for (n = 0; nm_i != head; n++) {
963 			struct netmap_slot *slot = &ring->slot[nm_i];
964 			u_int len = slot->len;
965 			uint64_t paddr;
966 			void *addr = PNMB(na, slot, &paddr);
967 			int flags = (slot->flags & NS_REPORT ||
968 				nic_i == 0 || nic_i == report_frequency) ?
969 				IPI_TX_INTR : 0;
970 
971 			/* device-specific */
972 			pi.ipi_len = len;
973 			pi.ipi_segs[0].ds_addr = paddr;
974 			pi.ipi_segs[0].ds_len = len;
975 			pi.ipi_nsegs = 1;
976 			pi.ipi_ndescs = 0;
977 			pi.ipi_pidx = nic_i;
978 			pi.ipi_flags = flags;
979 
980 			/* Fill the slot in the NIC ring. */
981 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
982 			DBG_COUNTER_INC(tx_encap);
983 
984 			/* prefetch for next round */
985 			__builtin_prefetch(&ring->slot[nm_i + 1]);
986 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
987 			if (txq->ift_sds.ifsd_map) {
988 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
989 
990 				NM_CHECK_ADDR_LEN(na, addr, len);
991 
992 				if (slot->flags & NS_BUF_CHANGED) {
993 					/* buffer has changed, reload map */
994 					netmap_reload_map(na, txq->ift_buf_tag,
995 					    txq->ift_sds.ifsd_map[nic_i], addr);
996 				}
997 				/* make sure changes to the buffer are synced */
998 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
999 						BUS_DMASYNC_PREWRITE);
1000 			}
1001 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1002 			nm_i = nm_next(nm_i, lim);
1003 			nic_i = nm_next(nic_i, lim);
1004 		}
1005 		kring->nr_hwcur = nm_i;
1006 
1007 		/* synchronize the NIC ring */
1008 		bus_dmamap_sync(txq->ift_buf_tag, txq->ift_ifdi->idi_map,
1009 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1010 
1011 		/* (re)start the tx unit up to slot nic_i (excluded) */
1012 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1013 	}
1014 
1015 	/*
1016 	 * Second part: reclaim buffers for completed transmissions.
1017 	 *
1018 	 * If there are unclaimed buffers, attempt to reclaim them.
1019 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1020 	 * minimal delay, then trigger the tx handler which will spin in the
1021 	 * group task queue.
1022 	 */
1023 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1024 		if (iflib_tx_credits_update(ctx, txq)) {
1025 			/* some tx completed, increment avail */
1026 			nic_i = txq->ift_cidx_processed;
1027 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1028 		}
1029 	}
1030 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1031 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1032 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1033 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1034 	}
1035 	return (0);
1036 }
1037 
1038 /*
1039  * Reconcile kernel and user view of the receive ring.
1040  * Same as for the txsync, this routine must be efficient.
1041  * The caller guarantees a single invocations, but races against
1042  * the rest of the driver should be handled here.
1043  *
1044  * On call, kring->rhead is the first packet that userspace wants
1045  * to keep, and kring->rcur is the wakeup point.
1046  * The kernel has previously reported packets up to kring->rtail.
1047  *
1048  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1049  * of whether or not we received an interrupt.
1050  */
1051 static int
1052 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1053 {
1054 	struct netmap_adapter *na = kring->na;
1055 	struct netmap_ring *ring = kring->ring;
1056 	uint32_t nm_i;	/* index into the netmap ring */
1057 	uint32_t nic_i;	/* index into the NIC ring */
1058 	u_int i, n;
1059 	u_int const lim = kring->nkr_num_slots - 1;
1060 	u_int const head = kring->rhead;
1061 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1062 	struct if_rxd_info ri;
1063 
1064 	struct ifnet *ifp = na->ifp;
1065 	if_ctx_t ctx = ifp->if_softc;
1066 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1067 	iflib_fl_t fl = rxq->ifr_fl;
1068 	if (head > lim)
1069 		return netmap_ring_reinit(kring);
1070 
1071 	/* XXX check sync modes */
1072 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1073 		if (fl->ifl_sds.ifsd_map == NULL)
1074 			continue;
1075 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_buf_tag,
1076 		    fl->ifl_ifdi->idi_map,
1077 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1078 	}
1079 	/*
1080 	 * First part: import newly received packets.
1081 	 *
1082 	 * nm_i is the index of the next free slot in the netmap ring,
1083 	 * nic_i is the index of the next received packet in the NIC ring,
1084 	 * and they may differ in case if_init() has been called while
1085 	 * in netmap mode. For the receive ring we have
1086 	 *
1087 	 *	nic_i = rxr->next_check;
1088 	 *	nm_i = kring->nr_hwtail (previous)
1089 	 * and
1090 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1091 	 *
1092 	 * rxr->next_check is set to 0 on a ring reinit
1093 	 */
1094 	if (netmap_no_pendintr || force_update) {
1095 		int crclen = iflib_crcstrip ? 0 : 4;
1096 		int error, avail;
1097 
1098 		for (i = 0; i < rxq->ifr_nfl; i++) {
1099 			fl = &rxq->ifr_fl[i];
1100 			nic_i = fl->ifl_cidx;
1101 			nm_i = netmap_idx_n2k(kring, nic_i);
1102 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1103 			for (n = 0; avail > 0; n++, avail--) {
1104 				rxd_info_zero(&ri);
1105 				ri.iri_frags = rxq->ifr_frags;
1106 				ri.iri_qsidx = kring->ring_id;
1107 				ri.iri_ifp = ctx->ifc_ifp;
1108 				ri.iri_cidx = nic_i;
1109 
1110 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1111 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1112 				ring->slot[nm_i].flags = 0;
1113 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1114 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1115 				nm_i = nm_next(nm_i, lim);
1116 				nic_i = nm_next(nic_i, lim);
1117 			}
1118 			if (n) { /* update the state variables */
1119 				if (netmap_no_pendintr && !force_update) {
1120 					/* diagnostics */
1121 					iflib_rx_miss ++;
1122 					iflib_rx_miss_bufs += n;
1123 				}
1124 				fl->ifl_cidx = nic_i;
1125 				kring->nr_hwtail = nm_i;
1126 			}
1127 			kring->nr_kflags &= ~NKR_PENDINTR;
1128 		}
1129 	}
1130 	/*
1131 	 * Second part: skip past packets that userspace has released.
1132 	 * (kring->nr_hwcur to head excluded),
1133 	 * and make the buffers available for reception.
1134 	 * As usual nm_i is the index in the netmap ring,
1135 	 * nic_i is the index in the NIC ring, and
1136 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1137 	 */
1138 	/* XXX not sure how this will work with multiple free lists */
1139 	nm_i = kring->nr_hwcur;
1140 
1141 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1142 }
1143 
1144 static void
1145 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1146 {
1147 	struct ifnet *ifp = na->ifp;
1148 	if_ctx_t ctx = ifp->if_softc;
1149 
1150 	CTX_LOCK(ctx);
1151 	if (onoff) {
1152 		IFDI_INTR_ENABLE(ctx);
1153 	} else {
1154 		IFDI_INTR_DISABLE(ctx);
1155 	}
1156 	CTX_UNLOCK(ctx);
1157 }
1158 
1159 
1160 static int
1161 iflib_netmap_attach(if_ctx_t ctx)
1162 {
1163 	struct netmap_adapter na;
1164 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1165 
1166 	bzero(&na, sizeof(na));
1167 
1168 	na.ifp = ctx->ifc_ifp;
1169 	na.na_flags = NAF_BDG_MAYSLEEP;
1170 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1171 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1172 
1173 	na.num_tx_desc = scctx->isc_ntxd[0];
1174 	na.num_rx_desc = scctx->isc_nrxd[0];
1175 	na.nm_txsync = iflib_netmap_txsync;
1176 	na.nm_rxsync = iflib_netmap_rxsync;
1177 	na.nm_register = iflib_netmap_register;
1178 	na.nm_intr = iflib_netmap_intr;
1179 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1180 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1181 	return (netmap_attach(&na));
1182 }
1183 
1184 static void
1185 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1186 {
1187 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1188 	struct netmap_slot *slot;
1189 
1190 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1191 	if (slot == NULL)
1192 		return;
1193 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1194 
1195 		/*
1196 		 * In netmap mode, set the map for the packet buffer.
1197 		 * NOTE: Some drivers (not this one) also need to set
1198 		 * the physical buffer address in the NIC ring.
1199 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1200 		 * netmap slot index, si
1201 		 */
1202 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1203 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1204 		    NMB(na, slot + si));
1205 	}
1206 }
1207 
1208 static void
1209 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1210 {
1211 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1212 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1213 	struct netmap_slot *slot;
1214 	uint32_t nm_i;
1215 
1216 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1217 	if (slot == NULL)
1218 		return;
1219 	nm_i = netmap_idx_n2k(kring, 0);
1220 	netmap_fl_refill(rxq, kring, nm_i, true);
1221 }
1222 
1223 static void
1224 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1225 {
1226 	struct netmap_kring *kring;
1227 
1228 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1229 
1230 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1231 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1232 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1233 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1234 			if (hz < 2000)
1235 				*reset_on = 1;
1236 			else
1237 				*reset_on = hz / 1000;
1238 		}
1239 	}
1240 }
1241 
1242 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1243 
1244 #else
1245 #define iflib_netmap_txq_init(ctx, txq)
1246 #define iflib_netmap_rxq_init(ctx, rxq)
1247 #define iflib_netmap_detach(ifp)
1248 
1249 #define iflib_netmap_attach(ctx) (0)
1250 #define netmap_rx_irq(ifp, qid, budget) (0)
1251 #define netmap_tx_irq(ifp, qid) do {} while (0)
1252 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1253 
1254 #endif
1255 
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1258 prefetch(void *x)
1259 {
1260 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1261 }
1262 static __inline void
1263 prefetch2cachelines(void *x)
1264 {
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1268 #endif
1269 }
1270 #else
1271 #define prefetch(x)
1272 #define prefetch2cachelines(x)
1273 #endif
1274 
1275 static void
1276 iflib_gen_mac(if_ctx_t ctx)
1277 {
1278 	struct thread *td;
1279 	MD5_CTX mdctx;
1280 	char uuid[HOSTUUIDLEN+1];
1281 	char buf[HOSTUUIDLEN+16];
1282 	uint8_t *mac;
1283 	unsigned char digest[16];
1284 
1285 	td = curthread;
1286 	mac = ctx->ifc_mac;
1287 	uuid[HOSTUUIDLEN] = 0;
1288 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1289 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1290 	/*
1291 	 * Generate a pseudo-random, deterministic MAC
1292 	 * address based on the UUID and unit number.
1293 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1294 	 */
1295 	MD5Init(&mdctx);
1296 	MD5Update(&mdctx, buf, strlen(buf));
1297 	MD5Final(digest, &mdctx);
1298 
1299 	mac[0] = 0x58;
1300 	mac[1] = 0x9C;
1301 	mac[2] = 0xFC;
1302 	mac[3] = digest[0];
1303 	mac[4] = digest[1];
1304 	mac[5] = digest[2];
1305 }
1306 
1307 static void
1308 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1309 {
1310 	iflib_fl_t fl;
1311 
1312 	fl = &rxq->ifr_fl[flid];
1313 	iru->iru_paddrs = fl->ifl_bus_addrs;
1314 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1315 	iru->iru_idxs = fl->ifl_rxd_idxs;
1316 	iru->iru_qsidx = rxq->ifr_id;
1317 	iru->iru_buf_size = fl->ifl_buf_size;
1318 	iru->iru_flidx = fl->ifl_id;
1319 }
1320 
1321 static void
1322 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1323 {
1324 	if (err)
1325 		return;
1326 	*(bus_addr_t *) arg = segs[0].ds_addr;
1327 }
1328 
1329 int
1330 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1331 {
1332 	int err;
1333 	device_t dev = ctx->ifc_dev;
1334 
1335 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1336 				align, 0,		/* alignment, bounds */
1337 				BUS_SPACE_MAXADDR,	/* lowaddr */
1338 				BUS_SPACE_MAXADDR,	/* highaddr */
1339 				NULL, NULL,		/* filter, filterarg */
1340 				size,			/* maxsize */
1341 				1,			/* nsegments */
1342 				size,			/* maxsegsize */
1343 				BUS_DMA_ALLOCNOW,	/* flags */
1344 				NULL,			/* lockfunc */
1345 				NULL,			/* lockarg */
1346 				&dma->idi_tag);
1347 	if (err) {
1348 		device_printf(dev,
1349 		    "%s: bus_dma_tag_create failed: %d\n",
1350 		    __func__, err);
1351 		goto fail_0;
1352 	}
1353 
1354 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1355 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1356 	if (err) {
1357 		device_printf(dev,
1358 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1359 		    __func__, (uintmax_t)size, err);
1360 		goto fail_1;
1361 	}
1362 
1363 	dma->idi_paddr = IF_BAD_DMA;
1364 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1365 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1366 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1367 		device_printf(dev,
1368 		    "%s: bus_dmamap_load failed: %d\n",
1369 		    __func__, err);
1370 		goto fail_2;
1371 	}
1372 
1373 	dma->idi_size = size;
1374 	return (0);
1375 
1376 fail_2:
1377 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1378 fail_1:
1379 	bus_dma_tag_destroy(dma->idi_tag);
1380 fail_0:
1381 	dma->idi_tag = NULL;
1382 
1383 	return (err);
1384 }
1385 
1386 int
1387 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1388 {
1389 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1390 
1391 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1392 
1393 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1394 }
1395 
1396 int
1397 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1398 {
1399 	int i, err;
1400 	iflib_dma_info_t *dmaiter;
1401 
1402 	dmaiter = dmalist;
1403 	for (i = 0; i < count; i++, dmaiter++) {
1404 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1405 			break;
1406 	}
1407 	if (err)
1408 		iflib_dma_free_multi(dmalist, i);
1409 	return (err);
1410 }
1411 
1412 void
1413 iflib_dma_free(iflib_dma_info_t dma)
1414 {
1415 	if (dma->idi_tag == NULL)
1416 		return;
1417 	if (dma->idi_paddr != IF_BAD_DMA) {
1418 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1419 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1420 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1421 		dma->idi_paddr = IF_BAD_DMA;
1422 	}
1423 	if (dma->idi_vaddr != NULL) {
1424 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1425 		dma->idi_vaddr = NULL;
1426 	}
1427 	bus_dma_tag_destroy(dma->idi_tag);
1428 	dma->idi_tag = NULL;
1429 }
1430 
1431 void
1432 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1433 {
1434 	int i;
1435 	iflib_dma_info_t *dmaiter = dmalist;
1436 
1437 	for (i = 0; i < count; i++, dmaiter++)
1438 		iflib_dma_free(*dmaiter);
1439 }
1440 
1441 #ifdef EARLY_AP_STARTUP
1442 static const int iflib_started = 1;
1443 #else
1444 /*
1445  * We used to abuse the smp_started flag to decide if the queues have been
1446  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1447  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1448  * is set.  Run a SYSINIT() strictly after that to just set a usable
1449  * completion flag.
1450  */
1451 
1452 static int iflib_started;
1453 
1454 static void
1455 iflib_record_started(void *arg)
1456 {
1457 	iflib_started = 1;
1458 }
1459 
1460 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1461 	iflib_record_started, NULL);
1462 #endif
1463 
1464 static int
1465 iflib_fast_intr(void *arg)
1466 {
1467 	iflib_filter_info_t info = arg;
1468 	struct grouptask *gtask = info->ifi_task;
1469 	if (!iflib_started)
1470 		return (FILTER_HANDLED);
1471 
1472 	DBG_COUNTER_INC(fast_intrs);
1473 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1474 		return (FILTER_HANDLED);
1475 
1476 	GROUPTASK_ENQUEUE(gtask);
1477 	return (FILTER_HANDLED);
1478 }
1479 
1480 static int
1481 iflib_fast_intr_rxtx(void *arg)
1482 {
1483 	iflib_filter_info_t info = arg;
1484 	struct grouptask *gtask = info->ifi_task;
1485 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1486 	if_ctx_t ctx = NULL;;
1487 	int i, cidx;
1488 
1489 	if (!iflib_started)
1490 		return (FILTER_HANDLED);
1491 
1492 	DBG_COUNTER_INC(fast_intrs);
1493 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1494 		return (FILTER_HANDLED);
1495 
1496 	MPASS(rxq->ifr_ntxqirq);
1497 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1498 		qidx_t txqid = rxq->ifr_txqid[i];
1499 
1500 		ctx = rxq->ifr_ctx;
1501 
1502 		bus_dmamap_sync(rxq->ifr_ifdi->idi_tag, rxq->ifr_ifdi->idi_map,
1503 		    BUS_DMASYNC_POSTREAD);
1504 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1505 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1506 			continue;
1507 		}
1508 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1509 	}
1510 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1511 		cidx = rxq->ifr_cq_cidx;
1512 	else
1513 		cidx = rxq->ifr_fl[0].ifl_cidx;
1514 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1515 		GROUPTASK_ENQUEUE(gtask);
1516 	else {
1517 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1518 		DBG_COUNTER_INC(rx_intr_enables);
1519 	}
1520 	return (FILTER_HANDLED);
1521 }
1522 
1523 
1524 static int
1525 iflib_fast_intr_ctx(void *arg)
1526 {
1527 	iflib_filter_info_t info = arg;
1528 	struct grouptask *gtask = info->ifi_task;
1529 
1530 	if (!iflib_started)
1531 		return (FILTER_HANDLED);
1532 
1533 	DBG_COUNTER_INC(fast_intrs);
1534 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1535 		return (FILTER_HANDLED);
1536 
1537 	GROUPTASK_ENQUEUE(gtask);
1538 	return (FILTER_HANDLED);
1539 }
1540 
1541 static int
1542 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1543 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1544 		 const char *name)
1545 {
1546 	int rc, flags;
1547 	struct resource *res;
1548 	void *tag = NULL;
1549 	device_t dev = ctx->ifc_dev;
1550 
1551 	flags = RF_ACTIVE;
1552 	if (ctx->ifc_flags & IFC_LEGACY)
1553 		flags |= RF_SHAREABLE;
1554 	MPASS(rid < 512);
1555 	irq->ii_rid = rid;
1556 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1557 	if (res == NULL) {
1558 		device_printf(dev,
1559 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1560 		return (ENOMEM);
1561 	}
1562 	irq->ii_res = res;
1563 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1564 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1565 						filter, handler, arg, &tag);
1566 	if (rc != 0) {
1567 		device_printf(dev,
1568 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1569 					  rid, name ? name : "unknown", rc);
1570 		return (rc);
1571 	} else if (name)
1572 		bus_describe_intr(dev, res, tag, "%s", name);
1573 
1574 	irq->ii_tag = tag;
1575 	return (0);
1576 }
1577 
1578 
1579 /*********************************************************************
1580  *
1581  *  Allocate DMA resources for TX buffers as well as memory for the TX
1582  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1583  *  iflib_sw_tx_desc_array structure, storing all the information that
1584  *  is needed to transmit a packet on the wire.  This is called only
1585  *  once at attach, setup is done every reset.
1586  *
1587  **********************************************************************/
1588 static int
1589 iflib_txsd_alloc(iflib_txq_t txq)
1590 {
1591 	if_ctx_t ctx = txq->ift_ctx;
1592 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1593 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1594 	device_t dev = ctx->ifc_dev;
1595 	bus_size_t tsomaxsize;
1596 	int err, nsegments, ntsosegments;
1597 	bool tso;
1598 
1599 	nsegments = scctx->isc_tx_nsegments;
1600 	ntsosegments = scctx->isc_tx_tso_segments_max;
1601 	tsomaxsize = scctx->isc_tx_tso_size_max;
1602 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1603 		tsomaxsize += sizeof(struct ether_vlan_header);
1604 	MPASS(scctx->isc_ntxd[0] > 0);
1605 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1606 	MPASS(nsegments > 0);
1607 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1608 		MPASS(ntsosegments > 0);
1609 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1610 	}
1611 
1612 	/*
1613 	 * Set up DMA tags for TX buffers.
1614 	 */
1615 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1616 			       1, 0,			/* alignment, bounds */
1617 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1618 			       BUS_SPACE_MAXADDR,	/* highaddr */
1619 			       NULL, NULL,		/* filter, filterarg */
1620 			       sctx->isc_tx_maxsize,		/* maxsize */
1621 			       nsegments,	/* nsegments */
1622 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1623 			       0,			/* flags */
1624 			       NULL,			/* lockfunc */
1625 			       NULL,			/* lockfuncarg */
1626 			       &txq->ift_buf_tag))) {
1627 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1628 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1629 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1630 		goto fail;
1631 	}
1632 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1633 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1634 			       1, 0,			/* alignment, bounds */
1635 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1636 			       BUS_SPACE_MAXADDR,	/* highaddr */
1637 			       NULL, NULL,		/* filter, filterarg */
1638 			       tsomaxsize,		/* maxsize */
1639 			       ntsosegments,	/* nsegments */
1640 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1641 			       0,			/* flags */
1642 			       NULL,			/* lockfunc */
1643 			       NULL,			/* lockfuncarg */
1644 			       &txq->ift_tso_buf_tag))) {
1645 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1646 		    err);
1647 		goto fail;
1648 	}
1649 
1650 	/* Allocate memory for the TX mbuf map. */
1651 	if (!(txq->ift_sds.ifsd_m =
1652 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1653 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1654 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1655 		err = ENOMEM;
1656 		goto fail;
1657 	}
1658 
1659 	/*
1660 	 * Create the DMA maps for TX buffers.
1661 	 */
1662 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1663 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1664 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1665 		device_printf(dev,
1666 		    "Unable to allocate TX buffer DMA map memory\n");
1667 		err = ENOMEM;
1668 		goto fail;
1669 	}
1670 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1671 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1672 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1673 		device_printf(dev,
1674 		    "Unable to allocate TSO TX buffer map memory\n");
1675 		err = ENOMEM;
1676 		goto fail;
1677 	}
1678 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1679 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1680 		    &txq->ift_sds.ifsd_map[i]);
1681 		if (err != 0) {
1682 			device_printf(dev, "Unable to create TX DMA map\n");
1683 			goto fail;
1684 		}
1685 		if (!tso)
1686 			continue;
1687 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1688 		    &txq->ift_sds.ifsd_tso_map[i]);
1689 		if (err != 0) {
1690 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1691 			goto fail;
1692 		}
1693 	}
1694 	return (0);
1695 fail:
1696 	/* We free all, it handles case where we are in the middle */
1697 	iflib_tx_structures_free(ctx);
1698 	return (err);
1699 }
1700 
1701 static void
1702 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1703 {
1704 	bus_dmamap_t map;
1705 
1706 	map = NULL;
1707 	if (txq->ift_sds.ifsd_map != NULL)
1708 		map = txq->ift_sds.ifsd_map[i];
1709 	if (map != NULL) {
1710 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1711 		bus_dmamap_unload(txq->ift_buf_tag, map);
1712 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1713 		txq->ift_sds.ifsd_map[i] = NULL;
1714 	}
1715 
1716 	map = NULL;
1717 	if (txq->ift_sds.ifsd_tso_map != NULL)
1718 		map = txq->ift_sds.ifsd_tso_map[i];
1719 	if (map != NULL) {
1720 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1721 		    BUS_DMASYNC_POSTWRITE);
1722 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1723 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1724 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1725 	}
1726 }
1727 
1728 static void
1729 iflib_txq_destroy(iflib_txq_t txq)
1730 {
1731 	if_ctx_t ctx = txq->ift_ctx;
1732 
1733 	for (int i = 0; i < txq->ift_size; i++)
1734 		iflib_txsd_destroy(ctx, txq, i);
1735 	if (txq->ift_sds.ifsd_map != NULL) {
1736 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1737 		txq->ift_sds.ifsd_map = NULL;
1738 	}
1739 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1740 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1741 		txq->ift_sds.ifsd_tso_map = NULL;
1742 	}
1743 	if (txq->ift_sds.ifsd_m != NULL) {
1744 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1745 		txq->ift_sds.ifsd_m = NULL;
1746 	}
1747 	if (txq->ift_buf_tag != NULL) {
1748 		bus_dma_tag_destroy(txq->ift_buf_tag);
1749 		txq->ift_buf_tag = NULL;
1750 	}
1751 	if (txq->ift_tso_buf_tag != NULL) {
1752 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1753 		txq->ift_tso_buf_tag = NULL;
1754 	}
1755 }
1756 
1757 static void
1758 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1759 {
1760 	struct mbuf **mp;
1761 
1762 	mp = &txq->ift_sds.ifsd_m[i];
1763 	if (*mp == NULL)
1764 		return;
1765 
1766 	if (txq->ift_sds.ifsd_map != NULL) {
1767 		bus_dmamap_sync(txq->ift_buf_tag,
1768 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1769 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1770 	}
1771 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1772 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1773 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1774 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1775 		    txq->ift_sds.ifsd_tso_map[i]);
1776 	}
1777 	m_free(*mp);
1778 	DBG_COUNTER_INC(tx_frees);
1779 	*mp = NULL;
1780 }
1781 
1782 static int
1783 iflib_txq_setup(iflib_txq_t txq)
1784 {
1785 	if_ctx_t ctx = txq->ift_ctx;
1786 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1787 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1788 	iflib_dma_info_t di;
1789 	int i;
1790 
1791 	/* Set number of descriptors available */
1792 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1793 	/* XXX make configurable */
1794 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1795 
1796 	/* Reset indices */
1797 	txq->ift_cidx_processed = 0;
1798 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1799 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1800 
1801 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1802 		bzero((void *)di->idi_vaddr, di->idi_size);
1803 
1804 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1805 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1806 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1807 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1808 	return (0);
1809 }
1810 
1811 /*********************************************************************
1812  *
1813  *  Allocate DMA resources for RX buffers as well as memory for the RX
1814  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1815  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1816  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1817  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1818  *  packet, the maximum number of entries we'll need is equal to the
1819  *  number of hardware receive descriptors that we've allocated.
1820  *
1821  **********************************************************************/
1822 static int
1823 iflib_rxsd_alloc(iflib_rxq_t rxq)
1824 {
1825 	if_ctx_t ctx = rxq->ifr_ctx;
1826 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1827 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1828 	device_t dev = ctx->ifc_dev;
1829 	iflib_fl_t fl;
1830 	int			err;
1831 
1832 	MPASS(scctx->isc_nrxd[0] > 0);
1833 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1834 
1835 	fl = rxq->ifr_fl;
1836 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1837 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1838 		/* Set up DMA tag for RX buffers. */
1839 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1840 					 1, 0,			/* alignment, bounds */
1841 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1842 					 BUS_SPACE_MAXADDR,	/* highaddr */
1843 					 NULL, NULL,		/* filter, filterarg */
1844 					 sctx->isc_rx_maxsize,	/* maxsize */
1845 					 sctx->isc_rx_nsegments,	/* nsegments */
1846 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1847 					 0,			/* flags */
1848 					 NULL,			/* lockfunc */
1849 					 NULL,			/* lockarg */
1850 					 &fl->ifl_buf_tag);
1851 		if (err) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX DMA tag: %d\n", err);
1854 			goto fail;
1855 		}
1856 
1857 		/* Allocate memory for the RX mbuf map. */
1858 		if (!(fl->ifl_sds.ifsd_m =
1859 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1860 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1861 			device_printf(dev,
1862 			    "Unable to allocate RX mbuf map memory\n");
1863 			err = ENOMEM;
1864 			goto fail;
1865 		}
1866 
1867 		/* Allocate memory for the direct RX cluster pointer map. */
1868 		if (!(fl->ifl_sds.ifsd_cl =
1869 		      (caddr_t *) malloc(sizeof(caddr_t) *
1870 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1871 			device_printf(dev,
1872 			    "Unable to allocate RX cluster map memory\n");
1873 			err = ENOMEM;
1874 			goto fail;
1875 		}
1876 
1877 		/* Allocate memory for the RX cluster bus address map. */
1878 		if (!(fl->ifl_sds.ifsd_ba =
1879 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1880 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1881 			device_printf(dev,
1882 			    "Unable to allocate RX bus address map memory\n");
1883 			err = ENOMEM;
1884 			goto fail;
1885 		}
1886 
1887 		/*
1888 		 * Create the DMA maps for RX buffers.
1889 		 */
1890 		if (!(fl->ifl_sds.ifsd_map =
1891 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1892 			device_printf(dev,
1893 			    "Unable to allocate RX buffer DMA map memory\n");
1894 			err = ENOMEM;
1895 			goto fail;
1896 		}
1897 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1898 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1899 			    &fl->ifl_sds.ifsd_map[i]);
1900 			if (err != 0) {
1901 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1902 				goto fail;
1903 			}
1904 		}
1905 	}
1906 	return (0);
1907 
1908 fail:
1909 	iflib_rx_structures_free(ctx);
1910 	return (err);
1911 }
1912 
1913 
1914 /*
1915  * Internal service routines
1916  */
1917 
1918 struct rxq_refill_cb_arg {
1919 	int               error;
1920 	bus_dma_segment_t seg;
1921 	int               nseg;
1922 };
1923 
1924 static void
1925 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1926 {
1927 	struct rxq_refill_cb_arg *cb_arg = arg;
1928 
1929 	cb_arg->error = error;
1930 	cb_arg->seg = segs[0];
1931 	cb_arg->nseg = nseg;
1932 }
1933 
1934 /**
1935  *	rxq_refill - refill an rxq  free-buffer list
1936  *	@ctx: the iflib context
1937  *	@rxq: the free-list to refill
1938  *	@n: the number of new buffers to allocate
1939  *
1940  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1941  *	The caller must assure that @n does not exceed the queue's capacity.
1942  */
1943 static void
1944 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1945 {
1946 	struct if_rxd_update iru;
1947 	struct rxq_refill_cb_arg cb_arg;
1948 	struct mbuf *m;
1949 	caddr_t cl, *sd_cl;
1950 	struct mbuf **sd_m;
1951 	bus_dmamap_t *sd_map;
1952 	bus_addr_t bus_addr, *sd_ba;
1953 	int err, frag_idx, i, idx, n, pidx;
1954 	qidx_t credits;
1955 
1956 	sd_m = fl->ifl_sds.ifsd_m;
1957 	sd_map = fl->ifl_sds.ifsd_map;
1958 	sd_cl = fl->ifl_sds.ifsd_cl;
1959 	sd_ba = fl->ifl_sds.ifsd_ba;
1960 	pidx = fl->ifl_pidx;
1961 	idx = pidx;
1962 	frag_idx = fl->ifl_fragidx;
1963 	credits = fl->ifl_credits;
1964 
1965 	i = 0;
1966 	n = count;
1967 	MPASS(n > 0);
1968 	MPASS(credits + n <= fl->ifl_size);
1969 
1970 	if (pidx < fl->ifl_cidx)
1971 		MPASS(pidx + n <= fl->ifl_cidx);
1972 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1973 		MPASS(fl->ifl_gen == 0);
1974 	if (pidx > fl->ifl_cidx)
1975 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1976 
1977 	DBG_COUNTER_INC(fl_refills);
1978 	if (n > 8)
1979 		DBG_COUNTER_INC(fl_refills_large);
1980 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1981 	while (n--) {
1982 		/*
1983 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1984 		 * initialized after rx.
1985 		 *
1986 		 * If the cluster is still set then we know a minimum sized packet was received
1987 		 */
1988 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1989 		    &frag_idx);
1990 		if (frag_idx < 0)
1991 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1992 		MPASS(frag_idx >= 0);
1993 		if ((cl = sd_cl[frag_idx]) == NULL) {
1994 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1995 				break;
1996 
1997 			cb_arg.error = 0;
1998 			MPASS(sd_map != NULL);
1999 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2000 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2001 			    BUS_DMA_NOWAIT);
2002 			if (err != 0 || cb_arg.error) {
2003 				/*
2004 				 * !zone_pack ?
2005 				 */
2006 				if (fl->ifl_zone == zone_pack)
2007 					uma_zfree(fl->ifl_zone, cl);
2008 				break;
2009 			}
2010 
2011 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2012 			    BUS_DMASYNC_PREREAD);
2013 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2014 			sd_cl[frag_idx] = cl;
2015 #if MEMORY_LOGGING
2016 			fl->ifl_cl_enqueued++;
2017 #endif
2018 		} else {
2019 			bus_addr = sd_ba[frag_idx];
2020 		}
2021 
2022 		MPASS(sd_m[frag_idx] == NULL);
2023 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2024 			break;
2025 		}
2026 		sd_m[frag_idx] = m;
2027 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2028 #if MEMORY_LOGGING
2029 		fl->ifl_m_enqueued++;
2030 #endif
2031 
2032 		DBG_COUNTER_INC(rx_allocs);
2033 		fl->ifl_rxd_idxs[i] = frag_idx;
2034 		fl->ifl_bus_addrs[i] = bus_addr;
2035 		fl->ifl_vm_addrs[i] = cl;
2036 		credits++;
2037 		i++;
2038 		MPASS(credits <= fl->ifl_size);
2039 		if (++idx == fl->ifl_size) {
2040 			fl->ifl_gen = 1;
2041 			idx = 0;
2042 		}
2043 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2044 			iru.iru_pidx = pidx;
2045 			iru.iru_count = i;
2046 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2047 			i = 0;
2048 			pidx = idx;
2049 			fl->ifl_pidx = idx;
2050 			fl->ifl_credits = credits;
2051 		}
2052 	}
2053 
2054 	if (i) {
2055 		iru.iru_pidx = pidx;
2056 		iru.iru_count = i;
2057 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2058 		fl->ifl_pidx = idx;
2059 		fl->ifl_credits = credits;
2060 	}
2061 	DBG_COUNTER_INC(rxd_flush);
2062 	if (fl->ifl_pidx == 0)
2063 		pidx = fl->ifl_size - 1;
2064 	else
2065 		pidx = fl->ifl_pidx - 1;
2066 
2067 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2068 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2069 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2070 	fl->ifl_fragidx = frag_idx;
2071 }
2072 
2073 static __inline void
2074 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2075 {
2076 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2077 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2078 #ifdef INVARIANTS
2079 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2080 #endif
2081 
2082 	MPASS(fl->ifl_credits <= fl->ifl_size);
2083 	MPASS(reclaimable == delta);
2084 
2085 	if (reclaimable > 0)
2086 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2087 }
2088 
2089 uint8_t
2090 iflib_in_detach(if_ctx_t ctx)
2091 {
2092 	bool in_detach;
2093 	STATE_LOCK(ctx);
2094 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2095 	STATE_UNLOCK(ctx);
2096 	return (in_detach);
2097 }
2098 
2099 static void
2100 iflib_fl_bufs_free(iflib_fl_t fl)
2101 {
2102 	iflib_dma_info_t idi = fl->ifl_ifdi;
2103 	bus_dmamap_t sd_map;
2104 	uint32_t i;
2105 
2106 	for (i = 0; i < fl->ifl_size; i++) {
2107 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2108 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2109 
2110 		if (*sd_cl != NULL) {
2111 			sd_map = fl->ifl_sds.ifsd_map[i];
2112 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2113 			    BUS_DMASYNC_POSTREAD);
2114 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2115 			if (*sd_cl != NULL)
2116 				uma_zfree(fl->ifl_zone, *sd_cl);
2117 			// XXX: Should this get moved out?
2118 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2119 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2120 			if (*sd_m != NULL) {
2121 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2122 				uma_zfree(zone_mbuf, *sd_m);
2123 			}
2124 		} else {
2125 			MPASS(*sd_cl == NULL);
2126 			MPASS(*sd_m == NULL);
2127 		}
2128 #if MEMORY_LOGGING
2129 		fl->ifl_m_dequeued++;
2130 		fl->ifl_cl_dequeued++;
2131 #endif
2132 		*sd_cl = NULL;
2133 		*sd_m = NULL;
2134 	}
2135 #ifdef INVARIANTS
2136 	for (i = 0; i < fl->ifl_size; i++) {
2137 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2138 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2139 	}
2140 #endif
2141 	/*
2142 	 * Reset free list values
2143 	 */
2144 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2145 	bzero(idi->idi_vaddr, idi->idi_size);
2146 }
2147 
2148 /*********************************************************************
2149  *
2150  *  Initialize a receive ring and its buffers.
2151  *
2152  **********************************************************************/
2153 static int
2154 iflib_fl_setup(iflib_fl_t fl)
2155 {
2156 	iflib_rxq_t rxq = fl->ifl_rxq;
2157 	if_ctx_t ctx = rxq->ifr_ctx;
2158 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2159 
2160 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2161 	/*
2162 	** Free current RX buffer structs and their mbufs
2163 	*/
2164 	iflib_fl_bufs_free(fl);
2165 	/* Now replenish the mbufs */
2166 	MPASS(fl->ifl_credits == 0);
2167 	/*
2168 	 * XXX don't set the max_frame_size to larger
2169 	 * than the hardware can handle
2170 	 */
2171 	if (sctx->isc_max_frame_size <= 2048)
2172 		fl->ifl_buf_size = MCLBYTES;
2173 #ifndef CONTIGMALLOC_WORKS
2174 	else
2175 		fl->ifl_buf_size = MJUMPAGESIZE;
2176 #else
2177 	else if (sctx->isc_max_frame_size <= 4096)
2178 		fl->ifl_buf_size = MJUMPAGESIZE;
2179 	else if (sctx->isc_max_frame_size <= 9216)
2180 		fl->ifl_buf_size = MJUM9BYTES;
2181 	else
2182 		fl->ifl_buf_size = MJUM16BYTES;
2183 #endif
2184 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2185 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2186 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2187 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2188 
2189 
2190 	/* avoid pre-allocating zillions of clusters to an idle card
2191 	 * potentially speeding up attach
2192 	 */
2193 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2194 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2195 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2196 		return (ENOBUFS);
2197 	/*
2198 	 * handle failure
2199 	 */
2200 	MPASS(rxq != NULL);
2201 	MPASS(fl->ifl_ifdi != NULL);
2202 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2203 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2204 	return (0);
2205 }
2206 
2207 /*********************************************************************
2208  *
2209  *  Free receive ring data structures
2210  *
2211  **********************************************************************/
2212 static void
2213 iflib_rx_sds_free(iflib_rxq_t rxq)
2214 {
2215 	iflib_fl_t fl;
2216 	int i, j;
2217 
2218 	if (rxq->ifr_fl != NULL) {
2219 		for (i = 0; i < rxq->ifr_nfl; i++) {
2220 			fl = &rxq->ifr_fl[i];
2221 			if (fl->ifl_buf_tag != NULL) {
2222 				if (fl->ifl_sds.ifsd_map != NULL) {
2223 					for (j = 0; j < fl->ifl_size; j++) {
2224 						if (fl->ifl_sds.ifsd_map[j] ==
2225 						    NULL)
2226 							continue;
2227 						bus_dmamap_sync(
2228 						    fl->ifl_buf_tag,
2229 						    fl->ifl_sds.ifsd_map[j],
2230 						    BUS_DMASYNC_POSTREAD);
2231 						bus_dmamap_unload(
2232 						    fl->ifl_buf_tag,
2233 						    fl->ifl_sds.ifsd_map[j]);
2234 					}
2235 				}
2236 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2237 				fl->ifl_buf_tag = NULL;
2238 			}
2239 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2240 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2241 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2242 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2243 			fl->ifl_sds.ifsd_m = NULL;
2244 			fl->ifl_sds.ifsd_cl = NULL;
2245 			fl->ifl_sds.ifsd_ba = NULL;
2246 			fl->ifl_sds.ifsd_map = NULL;
2247 		}
2248 		free(rxq->ifr_fl, M_IFLIB);
2249 		rxq->ifr_fl = NULL;
2250 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2251 	}
2252 }
2253 
2254 /*
2255  * MI independent logic
2256  *
2257  */
2258 static void
2259 iflib_timer(void *arg)
2260 {
2261 	iflib_txq_t txq = arg;
2262 	if_ctx_t ctx = txq->ift_ctx;
2263 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2264 	uint64_t this_tick = ticks;
2265 	uint32_t reset_on = hz / 2;
2266 
2267 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2268 		return;
2269 	/*
2270 	** Check on the state of the TX queue(s), this
2271 	** can be done without the lock because its RO
2272 	** and the HUNG state will be static if set.
2273 	*/
2274 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2275 		txq->ift_last_timer_tick = this_tick;
2276 		IFDI_TIMER(ctx, txq->ift_id);
2277 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2278 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2279 		     (sctx->isc_pause_frames == 0)))
2280 			goto hung;
2281 
2282 		if (ifmp_ring_is_stalled(txq->ift_br))
2283 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2284 		txq->ift_cleaned_prev = txq->ift_cleaned;
2285 	}
2286 #ifdef DEV_NETMAP
2287 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2288 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2289 #endif
2290 	/* handle any laggards */
2291 	if (txq->ift_db_pending)
2292 		GROUPTASK_ENQUEUE(&txq->ift_task);
2293 
2294 	sctx->isc_pause_frames = 0;
2295 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2296 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2297 	return;
2298  hung:
2299 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2300 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2301 	STATE_LOCK(ctx);
2302 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2303 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2304 	iflib_admin_intr_deferred(ctx);
2305 	STATE_UNLOCK(ctx);
2306 }
2307 
2308 static void
2309 iflib_init_locked(if_ctx_t ctx)
2310 {
2311 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2312 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2313 	if_t ifp = ctx->ifc_ifp;
2314 	iflib_fl_t fl;
2315 	iflib_txq_t txq;
2316 	iflib_rxq_t rxq;
2317 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2318 
2319 
2320 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2321 	IFDI_INTR_DISABLE(ctx);
2322 
2323 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2324 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2325 	/* Set hardware offload abilities */
2326 	if_clearhwassist(ifp);
2327 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2328 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2329 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2330 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2331 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2332 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2333 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2334 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2335 
2336 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2337 		CALLOUT_LOCK(txq);
2338 		callout_stop(&txq->ift_timer);
2339 		CALLOUT_UNLOCK(txq);
2340 		iflib_netmap_txq_init(ctx, txq);
2341 	}
2342 #ifdef INVARIANTS
2343 	i = if_getdrvflags(ifp);
2344 #endif
2345 	IFDI_INIT(ctx);
2346 	MPASS(if_getdrvflags(ifp) == i);
2347 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2348 		/* XXX this should really be done on a per-queue basis */
2349 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2350 			MPASS(rxq->ifr_id == i);
2351 			iflib_netmap_rxq_init(ctx, rxq);
2352 			continue;
2353 		}
2354 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2355 			if (iflib_fl_setup(fl)) {
2356 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2357 				goto done;
2358 			}
2359 		}
2360 	}
2361 done:
2362 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2363 	IFDI_INTR_ENABLE(ctx);
2364 	txq = ctx->ifc_txqs;
2365 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2366 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2367 			txq->ift_timer.c_cpu);
2368 }
2369 
2370 static int
2371 iflib_media_change(if_t ifp)
2372 {
2373 	if_ctx_t ctx = if_getsoftc(ifp);
2374 	int err;
2375 
2376 	CTX_LOCK(ctx);
2377 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2378 		iflib_init_locked(ctx);
2379 	CTX_UNLOCK(ctx);
2380 	return (err);
2381 }
2382 
2383 static void
2384 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2385 {
2386 	if_ctx_t ctx = if_getsoftc(ifp);
2387 
2388 	CTX_LOCK(ctx);
2389 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2390 	IFDI_MEDIA_STATUS(ctx, ifmr);
2391 	CTX_UNLOCK(ctx);
2392 }
2393 
2394 void
2395 iflib_stop(if_ctx_t ctx)
2396 {
2397 	iflib_txq_t txq = ctx->ifc_txqs;
2398 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2399 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2400 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2401 	iflib_dma_info_t di;
2402 	iflib_fl_t fl;
2403 	int i, j;
2404 
2405 	/* Tell the stack that the interface is no longer active */
2406 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2407 
2408 	IFDI_INTR_DISABLE(ctx);
2409 	DELAY(1000);
2410 	IFDI_STOP(ctx);
2411 	DELAY(1000);
2412 
2413 	iflib_debug_reset();
2414 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2415 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2416 		/* make sure all transmitters have completed before proceeding XXX */
2417 
2418 		CALLOUT_LOCK(txq);
2419 		callout_stop(&txq->ift_timer);
2420 		CALLOUT_UNLOCK(txq);
2421 
2422 		/* clean any enqueued buffers */
2423 		iflib_ifmp_purge(txq);
2424 		/* Free any existing tx buffers. */
2425 		for (j = 0; j < txq->ift_size; j++) {
2426 			iflib_txsd_free(ctx, txq, j);
2427 		}
2428 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2429 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2430 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2431 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2432 		txq->ift_pullups = 0;
2433 		ifmp_ring_reset_stats(txq->ift_br);
2434 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2435 			bzero((void *)di->idi_vaddr, di->idi_size);
2436 	}
2437 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2438 		/* make sure all transmitters have completed before proceeding XXX */
2439 
2440 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2441 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2442 			bzero((void *)di->idi_vaddr, di->idi_size);
2443 		/* also resets the free lists pidx/cidx */
2444 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2445 			iflib_fl_bufs_free(fl);
2446 	}
2447 }
2448 
2449 static inline caddr_t
2450 calc_next_rxd(iflib_fl_t fl, int cidx)
2451 {
2452 	qidx_t size;
2453 	int nrxd;
2454 	caddr_t start, end, cur, next;
2455 
2456 	nrxd = fl->ifl_size;
2457 	size = fl->ifl_rxd_size;
2458 	start = fl->ifl_ifdi->idi_vaddr;
2459 
2460 	if (__predict_false(size == 0))
2461 		return (start);
2462 	cur = start + size*cidx;
2463 	end = start + size*nrxd;
2464 	next = CACHE_PTR_NEXT(cur);
2465 	return (next < end ? next : start);
2466 }
2467 
2468 static inline void
2469 prefetch_pkts(iflib_fl_t fl, int cidx)
2470 {
2471 	int nextptr;
2472 	int nrxd = fl->ifl_size;
2473 	caddr_t next_rxd;
2474 
2475 
2476 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2477 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2478 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2479 	next_rxd = calc_next_rxd(fl, cidx);
2480 	prefetch(next_rxd);
2481 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2482 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2483 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2484 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2485 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2486 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2487 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2488 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2489 }
2490 
2491 static void
2492 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2493 {
2494 	int flid, cidx;
2495 	bus_dmamap_t map;
2496 	iflib_fl_t fl;
2497 	iflib_dma_info_t di;
2498 	int next;
2499 
2500 	map = NULL;
2501 	flid = irf->irf_flid;
2502 	cidx = irf->irf_idx;
2503 	fl = &rxq->ifr_fl[flid];
2504 	sd->ifsd_fl = fl;
2505 	sd->ifsd_cidx = cidx;
2506 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2507 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2508 	fl->ifl_credits--;
2509 #if MEMORY_LOGGING
2510 	fl->ifl_m_dequeued++;
2511 #endif
2512 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2513 		prefetch_pkts(fl, cidx);
2514 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2515 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2516 	map = fl->ifl_sds.ifsd_map[cidx];
2517 	di = fl->ifl_ifdi;
2518 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2519 
2520 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2521 	MPASS(fl->ifl_cidx == cidx);
2522 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2523 	if (unload)
2524 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2525 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2526 	if (__predict_false(fl->ifl_cidx == 0))
2527 		fl->ifl_gen = 0;
2528 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2529 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2530 	bit_clear(fl->ifl_rx_bitmap, cidx);
2531 }
2532 
2533 static struct mbuf *
2534 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2535 {
2536 	int i, padlen , flags;
2537 	struct mbuf *m, *mh, *mt;
2538 	caddr_t cl;
2539 
2540 	i = 0;
2541 	mh = NULL;
2542 	do {
2543 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2544 
2545 		MPASS(*sd->ifsd_cl != NULL);
2546 		MPASS(*sd->ifsd_m != NULL);
2547 
2548 		/* Don't include zero-length frags */
2549 		if (ri->iri_frags[i].irf_len == 0) {
2550 			/* XXX we can save the cluster here, but not the mbuf */
2551 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2552 			m_free(*sd->ifsd_m);
2553 			*sd->ifsd_m = NULL;
2554 			continue;
2555 		}
2556 		m = *sd->ifsd_m;
2557 		*sd->ifsd_m = NULL;
2558 		if (mh == NULL) {
2559 			flags = M_PKTHDR|M_EXT;
2560 			mh = mt = m;
2561 			padlen = ri->iri_pad;
2562 		} else {
2563 			flags = M_EXT;
2564 			mt->m_next = m;
2565 			mt = m;
2566 			/* assuming padding is only on the first fragment */
2567 			padlen = 0;
2568 		}
2569 		cl = *sd->ifsd_cl;
2570 		*sd->ifsd_cl = NULL;
2571 
2572 		/* Can these two be made one ? */
2573 		m_init(m, M_NOWAIT, MT_DATA, flags);
2574 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2575 		/*
2576 		 * These must follow m_init and m_cljset
2577 		 */
2578 		m->m_data += padlen;
2579 		ri->iri_len -= padlen;
2580 		m->m_len = ri->iri_frags[i].irf_len;
2581 	} while (++i < ri->iri_nfrags);
2582 
2583 	return (mh);
2584 }
2585 
2586 /*
2587  * Process one software descriptor
2588  */
2589 static struct mbuf *
2590 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2591 {
2592 	struct if_rxsd sd;
2593 	struct mbuf *m;
2594 
2595 	/* should I merge this back in now that the two paths are basically duplicated? */
2596 	if (ri->iri_nfrags == 1 &&
2597 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2598 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2599 		m = *sd.ifsd_m;
2600 		*sd.ifsd_m = NULL;
2601 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2602 #ifndef __NO_STRICT_ALIGNMENT
2603 		if (!IP_ALIGNED(m))
2604 			m->m_data += 2;
2605 #endif
2606 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2607 		bus_dmamap_sync(rxq->ifr_fl->ifl_buf_tag,
2608 		    rxq->ifr_fl->ifl_sds.ifsd_map[ri->iri_frags[0].irf_idx],
2609 		    BUS_DMASYNC_PREREAD);
2610 		m->m_len = ri->iri_frags[0].irf_len;
2611        } else {
2612 		m = assemble_segments(rxq, ri, &sd);
2613 	}
2614 	m->m_pkthdr.len = ri->iri_len;
2615 	m->m_pkthdr.rcvif = ri->iri_ifp;
2616 	m->m_flags |= ri->iri_flags;
2617 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2618 	m->m_pkthdr.flowid = ri->iri_flowid;
2619 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2620 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2621 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2622 	return (m);
2623 }
2624 
2625 #if defined(INET6) || defined(INET)
2626 static void
2627 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2628 {
2629 	CURVNET_SET(lc->ifp->if_vnet);
2630 #if defined(INET6)
2631 	*v6 = VNET(ip6_forwarding);
2632 #endif
2633 #if defined(INET)
2634 	*v4 = VNET(ipforwarding);
2635 #endif
2636 	CURVNET_RESTORE();
2637 }
2638 
2639 /*
2640  * Returns true if it's possible this packet could be LROed.
2641  * if it returns false, it is guaranteed that tcp_lro_rx()
2642  * would not return zero.
2643  */
2644 static bool
2645 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2646 {
2647 	struct ether_header *eh;
2648 	uint16_t eh_type;
2649 
2650 	eh = mtod(m, struct ether_header *);
2651 	eh_type = ntohs(eh->ether_type);
2652 	switch (eh_type) {
2653 #if defined(INET6)
2654 		case ETHERTYPE_IPV6:
2655 			return !v6_forwarding;
2656 #endif
2657 #if defined (INET)
2658 		case ETHERTYPE_IP:
2659 			return !v4_forwarding;
2660 #endif
2661 	}
2662 
2663 	return false;
2664 }
2665 #else
2666 static void
2667 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2668 {
2669 }
2670 #endif
2671 
2672 static bool
2673 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2674 {
2675 	if_ctx_t ctx = rxq->ifr_ctx;
2676 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2677 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2678 	iflib_dma_info_t di;
2679 	int avail, i;
2680 	qidx_t *cidxp;
2681 	struct if_rxd_info ri;
2682 	int err, budget_left, rx_bytes, rx_pkts;
2683 	iflib_fl_t fl;
2684 	struct ifnet *ifp;
2685 	int lro_enabled;
2686 	bool v4_forwarding, v6_forwarding, lro_possible;
2687 
2688 	/*
2689 	 * XXX early demux data packets so that if_input processing only handles
2690 	 * acks in interrupt context
2691 	 */
2692 	struct mbuf *m, *mh, *mt, *mf;
2693 
2694 	lro_possible = v4_forwarding = v6_forwarding = false;
2695 	ifp = ctx->ifc_ifp;
2696 	mh = mt = NULL;
2697 	MPASS(budget > 0);
2698 	rx_pkts	= rx_bytes = 0;
2699 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2700 		cidxp = &rxq->ifr_cq_cidx;
2701 	else
2702 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2703 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2704 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2705 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2706 		DBG_COUNTER_INC(rx_unavail);
2707 		return (false);
2708 	}
2709 
2710 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2711 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2712 			DBG_COUNTER_INC(rx_ctx_inactive);
2713 			break;
2714 		}
2715 		/*
2716 		 * Reset client set fields to their default values
2717 		 */
2718 		rxd_info_zero(&ri);
2719 		ri.iri_qsidx = rxq->ifr_id;
2720 		ri.iri_cidx = *cidxp;
2721 		ri.iri_ifp = ifp;
2722 		ri.iri_frags = rxq->ifr_frags;
2723 		di = rxq->ifr_fl[rxq->ifr_frags[0].irf_flid].ifl_ifdi;
2724 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2725 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2726 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2727 
2728 		if (err)
2729 			goto err;
2730 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2731 			*cidxp = ri.iri_cidx;
2732 			/* Update our consumer index */
2733 			/* XXX NB: shurd - check if this is still safe */
2734 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2735 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2736 				rxq->ifr_cq_gen = 0;
2737 			}
2738 			/* was this only a completion queue message? */
2739 			if (__predict_false(ri.iri_nfrags == 0))
2740 				continue;
2741 		}
2742 		MPASS(ri.iri_nfrags != 0);
2743 		MPASS(ri.iri_len != 0);
2744 
2745 		/* will advance the cidx on the corresponding free lists */
2746 		m = iflib_rxd_pkt_get(rxq, &ri);
2747 		avail--;
2748 		budget_left--;
2749 		if (avail == 0 && budget_left)
2750 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2751 
2752 		if (__predict_false(m == NULL)) {
2753 			DBG_COUNTER_INC(rx_mbuf_null);
2754 			continue;
2755 		}
2756 		/* imm_pkt: -- cxgb */
2757 		if (mh == NULL)
2758 			mh = mt = m;
2759 		else {
2760 			mt->m_nextpkt = m;
2761 			mt = m;
2762 		}
2763 	}
2764 	/* make sure that we can refill faster than drain */
2765 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2766 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2767 
2768 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2769 	if (lro_enabled)
2770 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2771 	mt = mf = NULL;
2772 	while (mh != NULL) {
2773 		m = mh;
2774 		mh = mh->m_nextpkt;
2775 		m->m_nextpkt = NULL;
2776 #ifndef __NO_STRICT_ALIGNMENT
2777 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2778 			continue;
2779 #endif
2780 		rx_bytes += m->m_pkthdr.len;
2781 		rx_pkts++;
2782 #if defined(INET6) || defined(INET)
2783 		if (lro_enabled) {
2784 			if (!lro_possible) {
2785 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2786 				if (lro_possible && mf != NULL) {
2787 					ifp->if_input(ifp, mf);
2788 					DBG_COUNTER_INC(rx_if_input);
2789 					mt = mf = NULL;
2790 				}
2791 			}
2792 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2793 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2794 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2795 					continue;
2796 			}
2797 		}
2798 #endif
2799 		if (lro_possible) {
2800 			ifp->if_input(ifp, m);
2801 			DBG_COUNTER_INC(rx_if_input);
2802 			continue;
2803 		}
2804 
2805 		if (mf == NULL)
2806 			mf = m;
2807 		if (mt != NULL)
2808 			mt->m_nextpkt = m;
2809 		mt = m;
2810 	}
2811 	if (mf != NULL) {
2812 		ifp->if_input(ifp, mf);
2813 		DBG_COUNTER_INC(rx_if_input);
2814 	}
2815 
2816 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2817 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2818 
2819 	/*
2820 	 * Flush any outstanding LRO work
2821 	 */
2822 #if defined(INET6) || defined(INET)
2823 	tcp_lro_flush_all(&rxq->ifr_lc);
2824 #endif
2825 	if (avail)
2826 		return true;
2827 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2828 err:
2829 	STATE_LOCK(ctx);
2830 	ctx->ifc_flags |= IFC_DO_RESET;
2831 	iflib_admin_intr_deferred(ctx);
2832 	STATE_UNLOCK(ctx);
2833 	return (false);
2834 }
2835 
2836 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2837 static inline qidx_t
2838 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2839 {
2840 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2841 	qidx_t minthresh = txq->ift_size / 8;
2842 	if (in_use > 4*minthresh)
2843 		return (notify_count);
2844 	if (in_use > 2*minthresh)
2845 		return (notify_count >> 1);
2846 	if (in_use > minthresh)
2847 		return (notify_count >> 3);
2848 	return (0);
2849 }
2850 
2851 static inline qidx_t
2852 txq_max_rs_deferred(iflib_txq_t txq)
2853 {
2854 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2855 	qidx_t minthresh = txq->ift_size / 8;
2856 	if (txq->ift_in_use > 4*minthresh)
2857 		return (notify_count);
2858 	if (txq->ift_in_use > 2*minthresh)
2859 		return (notify_count >> 1);
2860 	if (txq->ift_in_use > minthresh)
2861 		return (notify_count >> 2);
2862 	return (2);
2863 }
2864 
2865 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2866 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2867 
2868 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2869 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2870 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2871 
2872 /* forward compatibility for cxgb */
2873 #define FIRST_QSET(ctx) 0
2874 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2875 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2876 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2877 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2878 
2879 /* XXX we should be setting this to something other than zero */
2880 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2881 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2882     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2883 
2884 static inline bool
2885 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2886 {
2887 	qidx_t dbval, max;
2888 	bool rang;
2889 
2890 	rang = false;
2891 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2892 	if (ring || txq->ift_db_pending >= max) {
2893 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2894 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2895 		txq->ift_db_pending = txq->ift_npending = 0;
2896 		rang = true;
2897 	}
2898 	return (rang);
2899 }
2900 
2901 #ifdef PKT_DEBUG
2902 static void
2903 print_pkt(if_pkt_info_t pi)
2904 {
2905 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2906 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2907 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2908 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2909 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2910 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2911 }
2912 #endif
2913 
2914 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2915 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2916 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2917 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2918 
2919 static int
2920 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2921 {
2922 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2923 	struct ether_vlan_header *eh;
2924 	struct mbuf *m;
2925 
2926 	m = *mp;
2927 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2928 	    M_WRITABLE(m) == 0) {
2929 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2930 			return (ENOMEM);
2931 		} else {
2932 			m_freem(*mp);
2933 			DBG_COUNTER_INC(tx_frees);
2934 			*mp = m;
2935 		}
2936 	}
2937 
2938 	/*
2939 	 * Determine where frame payload starts.
2940 	 * Jump over vlan headers if already present,
2941 	 * helpful for QinQ too.
2942 	 */
2943 	if (__predict_false(m->m_len < sizeof(*eh))) {
2944 		txq->ift_pullups++;
2945 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2946 			return (ENOMEM);
2947 	}
2948 	eh = mtod(m, struct ether_vlan_header *);
2949 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2950 		pi->ipi_etype = ntohs(eh->evl_proto);
2951 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2952 	} else {
2953 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2954 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2955 	}
2956 
2957 	switch (pi->ipi_etype) {
2958 #ifdef INET
2959 	case ETHERTYPE_IP:
2960 	{
2961 		struct mbuf *n;
2962 		struct ip *ip = NULL;
2963 		struct tcphdr *th = NULL;
2964 		int minthlen;
2965 
2966 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2967 		if (__predict_false(m->m_len < minthlen)) {
2968 			/*
2969 			 * if this code bloat is causing too much of a hit
2970 			 * move it to a separate function and mark it noinline
2971 			 */
2972 			if (m->m_len == pi->ipi_ehdrlen) {
2973 				n = m->m_next;
2974 				MPASS(n);
2975 				if (n->m_len >= sizeof(*ip))  {
2976 					ip = (struct ip *)n->m_data;
2977 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2978 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2979 				} else {
2980 					txq->ift_pullups++;
2981 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2982 						return (ENOMEM);
2983 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2984 				}
2985 			} else {
2986 				txq->ift_pullups++;
2987 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2988 					return (ENOMEM);
2989 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2990 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2991 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2992 			}
2993 		} else {
2994 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2995 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2996 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2997 		}
2998 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2999 		pi->ipi_ipproto = ip->ip_p;
3000 		pi->ipi_flags |= IPI_TX_IPV4;
3001 
3002 		/* TCP checksum offload may require TCP header length */
3003 		if (IS_TX_OFFLOAD4(pi)) {
3004 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3005 				if (__predict_false(th == NULL)) {
3006 					txq->ift_pullups++;
3007 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3008 						return (ENOMEM);
3009 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3010 				}
3011 				pi->ipi_tcp_hflags = th->th_flags;
3012 				pi->ipi_tcp_hlen = th->th_off << 2;
3013 				pi->ipi_tcp_seq = th->th_seq;
3014 			}
3015 			if (IS_TSO4(pi)) {
3016 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3017 					return (ENXIO);
3018 				/*
3019 				 * TSO always requires hardware checksum offload.
3020 				 */
3021 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3022 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3023 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3024 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3025 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3026 					ip->ip_sum = 0;
3027 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3028 				}
3029 			}
3030 		}
3031 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3032                        ip->ip_sum = 0;
3033 
3034 		break;
3035 	}
3036 #endif
3037 #ifdef INET6
3038 	case ETHERTYPE_IPV6:
3039 	{
3040 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3041 		struct tcphdr *th;
3042 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3043 
3044 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3045 			txq->ift_pullups++;
3046 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3047 				return (ENOMEM);
3048 		}
3049 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3050 
3051 		/* XXX-BZ this will go badly in case of ext hdrs. */
3052 		pi->ipi_ipproto = ip6->ip6_nxt;
3053 		pi->ipi_flags |= IPI_TX_IPV6;
3054 
3055 		/* TCP checksum offload may require TCP header length */
3056 		if (IS_TX_OFFLOAD6(pi)) {
3057 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3058 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3059 					txq->ift_pullups++;
3060 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3061 						return (ENOMEM);
3062 				}
3063 				pi->ipi_tcp_hflags = th->th_flags;
3064 				pi->ipi_tcp_hlen = th->th_off << 2;
3065 				pi->ipi_tcp_seq = th->th_seq;
3066 			}
3067 			if (IS_TSO6(pi)) {
3068 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3069 					return (ENXIO);
3070 				/*
3071 				 * TSO always requires hardware checksum offload.
3072 				 */
3073 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3074 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3075 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3076 			}
3077 		}
3078 		break;
3079 	}
3080 #endif
3081 	default:
3082 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3083 		pi->ipi_ip_hlen = 0;
3084 		break;
3085 	}
3086 	*mp = m;
3087 
3088 	return (0);
3089 }
3090 
3091 /*
3092  * If dodgy hardware rejects the scatter gather chain we've handed it
3093  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3094  * m_defrag'd mbufs
3095  */
3096 static __noinline struct mbuf *
3097 iflib_remove_mbuf(iflib_txq_t txq)
3098 {
3099 	int ntxd, pidx;
3100 	struct mbuf *m, **ifsd_m;
3101 
3102 	ifsd_m = txq->ift_sds.ifsd_m;
3103 	ntxd = txq->ift_size;
3104 	pidx = txq->ift_pidx & (ntxd - 1);
3105 	ifsd_m = txq->ift_sds.ifsd_m;
3106 	m = ifsd_m[pidx];
3107 	ifsd_m[pidx] = NULL;
3108 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3109 	if (txq->ift_sds.ifsd_tso_map != NULL)
3110 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3111 		    txq->ift_sds.ifsd_tso_map[pidx]);
3112 #if MEMORY_LOGGING
3113 	txq->ift_dequeued++;
3114 #endif
3115 	return (m);
3116 }
3117 
3118 static inline caddr_t
3119 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3120 {
3121 	qidx_t size;
3122 	int ntxd;
3123 	caddr_t start, end, cur, next;
3124 
3125 	ntxd = txq->ift_size;
3126 	size = txq->ift_txd_size[qid];
3127 	start = txq->ift_ifdi[qid].idi_vaddr;
3128 
3129 	if (__predict_false(size == 0))
3130 		return (start);
3131 	cur = start + size*cidx;
3132 	end = start + size*ntxd;
3133 	next = CACHE_PTR_NEXT(cur);
3134 	return (next < end ? next : start);
3135 }
3136 
3137 /*
3138  * Pad an mbuf to ensure a minimum ethernet frame size.
3139  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3140  */
3141 static __noinline int
3142 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3143 {
3144 	/*
3145 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3146 	 * and ARP message is the smallest common payload I can think of
3147 	 */
3148 	static char pad[18];	/* just zeros */
3149 	int n;
3150 	struct mbuf *new_head;
3151 
3152 	if (!M_WRITABLE(*m_head)) {
3153 		new_head = m_dup(*m_head, M_NOWAIT);
3154 		if (new_head == NULL) {
3155 			m_freem(*m_head);
3156 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3157 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3158 			DBG_COUNTER_INC(tx_frees);
3159 			return ENOMEM;
3160 		}
3161 		m_freem(*m_head);
3162 		*m_head = new_head;
3163 	}
3164 
3165 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3166 	     n > 0; n -= sizeof(pad))
3167 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3168 			break;
3169 
3170 	if (n > 0) {
3171 		m_freem(*m_head);
3172 		device_printf(dev, "cannot pad short frame\n");
3173 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3174 		DBG_COUNTER_INC(tx_frees);
3175 		return (ENOBUFS);
3176 	}
3177 
3178 	return 0;
3179 }
3180 
3181 static int
3182 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3183 {
3184 	if_ctx_t		ctx;
3185 	if_shared_ctx_t		sctx;
3186 	if_softc_ctx_t		scctx;
3187 	bus_dma_tag_t		buf_tag;
3188 	bus_dma_segment_t	*segs;
3189 	struct mbuf		*m_head, **ifsd_m;
3190 	void			*next_txd;
3191 	bus_dmamap_t		map;
3192 	struct if_pkt_info	pi;
3193 	int remap = 0;
3194 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3195 
3196 	ctx = txq->ift_ctx;
3197 	sctx = ctx->ifc_sctx;
3198 	scctx = &ctx->ifc_softc_ctx;
3199 	segs = txq->ift_segs;
3200 	ntxd = txq->ift_size;
3201 	m_head = *m_headp;
3202 	map = NULL;
3203 
3204 	/*
3205 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3206 	 */
3207 	cidx = txq->ift_cidx;
3208 	pidx = txq->ift_pidx;
3209 	if (ctx->ifc_flags & IFC_PREFETCH) {
3210 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3211 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3212 			next_txd = calc_next_txd(txq, cidx, 0);
3213 			prefetch(next_txd);
3214 		}
3215 
3216 		/* prefetch the next cache line of mbuf pointers and flags */
3217 		prefetch(&txq->ift_sds.ifsd_m[next]);
3218 		prefetch(&txq->ift_sds.ifsd_map[next]);
3219 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3220 	}
3221 	map = txq->ift_sds.ifsd_map[pidx];
3222 	ifsd_m = txq->ift_sds.ifsd_m;
3223 
3224 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3225 		buf_tag = txq->ift_tso_buf_tag;
3226 		max_segs = scctx->isc_tx_tso_segments_max;
3227 		map = txq->ift_sds.ifsd_tso_map[pidx];
3228 		MPASS(buf_tag != NULL);
3229 		MPASS(max_segs > 0);
3230 	} else {
3231 		buf_tag = txq->ift_buf_tag;
3232 		max_segs = scctx->isc_tx_nsegments;
3233 		map = txq->ift_sds.ifsd_map[pidx];
3234 	}
3235 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3236 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3237 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3238 		if (err) {
3239 			DBG_COUNTER_INC(encap_txd_encap_fail);
3240 			return err;
3241 		}
3242 	}
3243 	m_head = *m_headp;
3244 
3245 	pkt_info_zero(&pi);
3246 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3247 	pi.ipi_pidx = pidx;
3248 	pi.ipi_qsidx = txq->ift_id;
3249 	pi.ipi_len = m_head->m_pkthdr.len;
3250 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3251 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3252 
3253 	/* deliberate bitwise OR to make one condition */
3254 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3255 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3256 			DBG_COUNTER_INC(encap_txd_encap_fail);
3257 			return (err);
3258 		}
3259 		m_head = *m_headp;
3260 	}
3261 
3262 retry:
3263 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3264 	    BUS_DMA_NOWAIT);
3265 defrag:
3266 	if (__predict_false(err)) {
3267 		switch (err) {
3268 		case EFBIG:
3269 			/* try collapse once and defrag once */
3270 			if (remap == 0) {
3271 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3272 				/* try defrag if collapsing fails */
3273 				if (m_head == NULL)
3274 					remap++;
3275 			}
3276 			if (remap == 1) {
3277 				txq->ift_mbuf_defrag++;
3278 				m_head = m_defrag(*m_headp, M_NOWAIT);
3279 			}
3280 			remap++;
3281 			if (__predict_false(m_head == NULL))
3282 				goto defrag_failed;
3283 			*m_headp = m_head;
3284 			goto retry;
3285 			break;
3286 		case ENOMEM:
3287 			txq->ift_no_tx_dma_setup++;
3288 			break;
3289 		default:
3290 			txq->ift_no_tx_dma_setup++;
3291 			m_freem(*m_headp);
3292 			DBG_COUNTER_INC(tx_frees);
3293 			*m_headp = NULL;
3294 			break;
3295 		}
3296 		txq->ift_map_failed++;
3297 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3298 		DBG_COUNTER_INC(encap_txd_encap_fail);
3299 		return (err);
3300 	}
3301 	ifsd_m[pidx] = m_head;
3302 	/*
3303 	 * XXX assumes a 1 to 1 relationship between segments and
3304 	 *        descriptors - this does not hold true on all drivers, e.g.
3305 	 *        cxgb
3306 	 */
3307 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3308 		txq->ift_no_desc_avail++;
3309 		bus_dmamap_unload(buf_tag, map);
3310 		DBG_COUNTER_INC(encap_txq_avail_fail);
3311 		DBG_COUNTER_INC(encap_txd_encap_fail);
3312 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3313 			GROUPTASK_ENQUEUE(&txq->ift_task);
3314 		return (ENOBUFS);
3315 	}
3316 	/*
3317 	 * On Intel cards we can greatly reduce the number of TX interrupts
3318 	 * we see by only setting report status on every Nth descriptor.
3319 	 * However, this also means that the driver will need to keep track
3320 	 * of the descriptors that RS was set on to check them for the DD bit.
3321 	 */
3322 	txq->ift_rs_pending += nsegs + 1;
3323 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3324 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3325 		pi.ipi_flags |= IPI_TX_INTR;
3326 		txq->ift_rs_pending = 0;
3327 	}
3328 
3329 	pi.ipi_segs = segs;
3330 	pi.ipi_nsegs = nsegs;
3331 
3332 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3333 #ifdef PKT_DEBUG
3334 	print_pkt(&pi);
3335 #endif
3336 	bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3337 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3338 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3339 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3340 		DBG_COUNTER_INC(tx_encap);
3341 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3342 
3343 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3344 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3345 			ndesc += txq->ift_size;
3346 			txq->ift_gen = 1;
3347 		}
3348 		/*
3349 		 * drivers can need as many as
3350 		 * two sentinels
3351 		 */
3352 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3353 		MPASS(pi.ipi_new_pidx != pidx);
3354 		MPASS(ndesc > 0);
3355 		txq->ift_in_use += ndesc;
3356 
3357 		/*
3358 		 * We update the last software descriptor again here because there may
3359 		 * be a sentinel and/or there may be more mbufs than segments
3360 		 */
3361 		txq->ift_pidx = pi.ipi_new_pidx;
3362 		txq->ift_npending += pi.ipi_ndescs;
3363 	} else {
3364 		*m_headp = m_head = iflib_remove_mbuf(txq);
3365 		if (err == EFBIG) {
3366 			txq->ift_txd_encap_efbig++;
3367 			if (remap < 2) {
3368 				remap = 1;
3369 				goto defrag;
3370 			}
3371 		}
3372 		goto defrag_failed;
3373 	}
3374 	/*
3375 	 * err can't possibly be non-zero here, so we don't neet to test it
3376 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3377 	 */
3378 	return (err);
3379 
3380 defrag_failed:
3381 	txq->ift_mbuf_defrag_failed++;
3382 	txq->ift_map_failed++;
3383 	m_freem(*m_headp);
3384 	DBG_COUNTER_INC(tx_frees);
3385 	*m_headp = NULL;
3386 	DBG_COUNTER_INC(encap_txd_encap_fail);
3387 	return (ENOMEM);
3388 }
3389 
3390 static void
3391 iflib_tx_desc_free(iflib_txq_t txq, int n)
3392 {
3393 	uint32_t qsize, cidx, mask, gen;
3394 	struct mbuf *m, **ifsd_m;
3395 	bool do_prefetch;
3396 
3397 	cidx = txq->ift_cidx;
3398 	gen = txq->ift_gen;
3399 	qsize = txq->ift_size;
3400 	mask = qsize-1;
3401 	ifsd_m = txq->ift_sds.ifsd_m;
3402 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3403 
3404 	while (n-- > 0) {
3405 		if (do_prefetch) {
3406 			prefetch(ifsd_m[(cidx + 3) & mask]);
3407 			prefetch(ifsd_m[(cidx + 4) & mask]);
3408 		}
3409 		if ((m = ifsd_m[cidx]) != NULL) {
3410 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3411 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3412 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3413 				    txq->ift_sds.ifsd_tso_map[cidx],
3414 				    BUS_DMASYNC_POSTWRITE);
3415 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3416 				    txq->ift_sds.ifsd_tso_map[cidx]);
3417 			} else {
3418 				bus_dmamap_sync(txq->ift_buf_tag,
3419 				    txq->ift_sds.ifsd_map[cidx],
3420 				    BUS_DMASYNC_POSTWRITE);
3421 				bus_dmamap_unload(txq->ift_buf_tag,
3422 				    txq->ift_sds.ifsd_map[cidx]);
3423 			}
3424 			/* XXX we don't support any drivers that batch packets yet */
3425 			MPASS(m->m_nextpkt == NULL);
3426 			m_freem(m);
3427 			ifsd_m[cidx] = NULL;
3428 #if MEMORY_LOGGING
3429 			txq->ift_dequeued++;
3430 #endif
3431 			DBG_COUNTER_INC(tx_frees);
3432 		}
3433 		if (__predict_false(++cidx == qsize)) {
3434 			cidx = 0;
3435 			gen = 0;
3436 		}
3437 	}
3438 	txq->ift_cidx = cidx;
3439 	txq->ift_gen = gen;
3440 }
3441 
3442 static __inline int
3443 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3444 {
3445 	int reclaim;
3446 	if_ctx_t ctx = txq->ift_ctx;
3447 
3448 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3449 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3450 
3451 	/*
3452 	 * Need a rate-limiting check so that this isn't called every time
3453 	 */
3454 	iflib_tx_credits_update(ctx, txq);
3455 	reclaim = DESC_RECLAIMABLE(txq);
3456 
3457 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3458 #ifdef INVARIANTS
3459 		if (iflib_verbose_debug) {
3460 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3461 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3462 			       reclaim, thresh);
3463 
3464 		}
3465 #endif
3466 		return (0);
3467 	}
3468 	iflib_tx_desc_free(txq, reclaim);
3469 	txq->ift_cleaned += reclaim;
3470 	txq->ift_in_use -= reclaim;
3471 
3472 	return (reclaim);
3473 }
3474 
3475 static struct mbuf **
3476 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3477 {
3478 	int next, size;
3479 	struct mbuf **items;
3480 
3481 	size = r->size;
3482 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3483 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3484 
3485 	prefetch(items[(cidx + offset) & (size-1)]);
3486 	if (remaining > 1) {
3487 		prefetch2cachelines(&items[next]);
3488 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3489 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3490 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3491 	}
3492 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3493 }
3494 
3495 static void
3496 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3497 {
3498 
3499 	ifmp_ring_check_drainage(txq->ift_br, budget);
3500 }
3501 
3502 static uint32_t
3503 iflib_txq_can_drain(struct ifmp_ring *r)
3504 {
3505 	iflib_txq_t txq = r->cookie;
3506 	if_ctx_t ctx = txq->ift_ctx;
3507 
3508 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3509 	    BUS_DMASYNC_POSTREAD);
3510 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3511 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3512 }
3513 
3514 static uint32_t
3515 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3516 {
3517 	iflib_txq_t txq = r->cookie;
3518 	if_ctx_t ctx = txq->ift_ctx;
3519 	struct ifnet *ifp = ctx->ifc_ifp;
3520 	struct mbuf **mp, *m;
3521 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3522 	int reclaimed, err, in_use_prev, desc_used;
3523 	bool do_prefetch, ring, rang;
3524 
3525 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3526 			    !LINK_ACTIVE(ctx))) {
3527 		DBG_COUNTER_INC(txq_drain_notready);
3528 		return (0);
3529 	}
3530 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3531 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3532 	avail = IDXDIFF(pidx, cidx, r->size);
3533 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3534 		DBG_COUNTER_INC(txq_drain_flushing);
3535 		for (i = 0; i < avail; i++) {
3536 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3537 				m_free(r->items[(cidx + i) & (r->size-1)]);
3538 			r->items[(cidx + i) & (r->size-1)] = NULL;
3539 		}
3540 		return (avail);
3541 	}
3542 
3543 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3544 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3545 		CALLOUT_LOCK(txq);
3546 		callout_stop(&txq->ift_timer);
3547 		CALLOUT_UNLOCK(txq);
3548 		DBG_COUNTER_INC(txq_drain_oactive);
3549 		return (0);
3550 	}
3551 	if (reclaimed)
3552 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3553 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3554 	count = MIN(avail, TX_BATCH_SIZE);
3555 #ifdef INVARIANTS
3556 	if (iflib_verbose_debug)
3557 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3558 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3559 #endif
3560 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3561 	avail = TXQ_AVAIL(txq);
3562 	err = 0;
3563 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3564 		int rem = do_prefetch ? count - i : 0;
3565 
3566 		mp = _ring_peek_one(r, cidx, i, rem);
3567 		MPASS(mp != NULL && *mp != NULL);
3568 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3569 			consumed++;
3570 			reclaimed++;
3571 			continue;
3572 		}
3573 		in_use_prev = txq->ift_in_use;
3574 		err = iflib_encap(txq, mp);
3575 		if (__predict_false(err)) {
3576 			/* no room - bail out */
3577 			if (err == ENOBUFS)
3578 				break;
3579 			consumed++;
3580 			/* we can't send this packet - skip it */
3581 			continue;
3582 		}
3583 		consumed++;
3584 		pkt_sent++;
3585 		m = *mp;
3586 		DBG_COUNTER_INC(tx_sent);
3587 		bytes_sent += m->m_pkthdr.len;
3588 		mcast_sent += !!(m->m_flags & M_MCAST);
3589 		avail = TXQ_AVAIL(txq);
3590 
3591 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3592 		desc_used += (txq->ift_in_use - in_use_prev);
3593 		ETHER_BPF_MTAP(ifp, m);
3594 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3595 			break;
3596 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3597 	}
3598 
3599 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3600 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3601 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3602 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3603 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3604 	if (mcast_sent)
3605 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3606 #ifdef INVARIANTS
3607 	if (iflib_verbose_debug)
3608 		printf("consumed=%d\n", consumed);
3609 #endif
3610 	return (consumed);
3611 }
3612 
3613 static uint32_t
3614 iflib_txq_drain_always(struct ifmp_ring *r)
3615 {
3616 	return (1);
3617 }
3618 
3619 static uint32_t
3620 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3621 {
3622 	int i, avail;
3623 	struct mbuf **mp;
3624 	iflib_txq_t txq;
3625 
3626 	txq = r->cookie;
3627 
3628 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3629 	CALLOUT_LOCK(txq);
3630 	callout_stop(&txq->ift_timer);
3631 	CALLOUT_UNLOCK(txq);
3632 
3633 	avail = IDXDIFF(pidx, cidx, r->size);
3634 	for (i = 0; i < avail; i++) {
3635 		mp = _ring_peek_one(r, cidx, i, avail - i);
3636 		if (__predict_false(*mp == (struct mbuf *)txq))
3637 			continue;
3638 		m_freem(*mp);
3639 		DBG_COUNTER_INC(tx_frees);
3640 	}
3641 	MPASS(ifmp_ring_is_stalled(r) == 0);
3642 	return (avail);
3643 }
3644 
3645 static void
3646 iflib_ifmp_purge(iflib_txq_t txq)
3647 {
3648 	struct ifmp_ring *r;
3649 
3650 	r = txq->ift_br;
3651 	r->drain = iflib_txq_drain_free;
3652 	r->can_drain = iflib_txq_drain_always;
3653 
3654 	ifmp_ring_check_drainage(r, r->size);
3655 
3656 	r->drain = iflib_txq_drain;
3657 	r->can_drain = iflib_txq_can_drain;
3658 }
3659 
3660 static void
3661 _task_fn_tx(void *context)
3662 {
3663 	iflib_txq_t txq = context;
3664 	if_ctx_t ctx = txq->ift_ctx;
3665 	struct ifnet *ifp = ctx->ifc_ifp;
3666 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3667 
3668 #ifdef IFLIB_DIAGNOSTICS
3669 	txq->ift_cpu_exec_count[curcpu]++;
3670 #endif
3671 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3672 		return;
3673 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3674 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3675 		    BUS_DMASYNC_POSTREAD);
3676 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3677 			netmap_tx_irq(ifp, txq->ift_id);
3678 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3679 		return;
3680 	}
3681 #ifdef ALTQ
3682 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3683 		iflib_altq_if_start(ifp);
3684 #endif
3685 	if (txq->ift_db_pending)
3686 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3687 	else if (!abdicate)
3688 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3689 	/*
3690 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3691 	 */
3692 	if (abdicate)
3693 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3694 	if (ctx->ifc_flags & IFC_LEGACY)
3695 		IFDI_INTR_ENABLE(ctx);
3696 	else {
3697 #ifdef INVARIANTS
3698 		int rc =
3699 #endif
3700 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3701 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3702 	}
3703 }
3704 
3705 static void
3706 _task_fn_rx(void *context)
3707 {
3708 	iflib_rxq_t rxq = context;
3709 	if_ctx_t ctx = rxq->ifr_ctx;
3710 	bool more;
3711 	uint16_t budget;
3712 
3713 #ifdef IFLIB_DIAGNOSTICS
3714 	rxq->ifr_cpu_exec_count[curcpu]++;
3715 #endif
3716 	DBG_COUNTER_INC(task_fn_rxs);
3717 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3718 		return;
3719 	more = true;
3720 #ifdef DEV_NETMAP
3721 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3722 		u_int work = 0;
3723 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3724 			more = false;
3725 		}
3726 	}
3727 #endif
3728 	budget = ctx->ifc_sysctl_rx_budget;
3729 	if (budget == 0)
3730 		budget = 16;	/* XXX */
3731 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3732 		if (ctx->ifc_flags & IFC_LEGACY)
3733 			IFDI_INTR_ENABLE(ctx);
3734 		else {
3735 #ifdef INVARIANTS
3736 			int rc =
3737 #endif
3738 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3739 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3740 			DBG_COUNTER_INC(rx_intr_enables);
3741 		}
3742 	}
3743 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3744 		return;
3745 	if (more)
3746 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3747 }
3748 
3749 static void
3750 _task_fn_admin(void *context)
3751 {
3752 	if_ctx_t ctx = context;
3753 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3754 	iflib_txq_t txq;
3755 	int i;
3756 	bool oactive, running, do_reset, do_watchdog, in_detach;
3757 	uint32_t reset_on = hz / 2;
3758 
3759 	STATE_LOCK(ctx);
3760 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3761 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3762 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3763 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3764 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3765 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3766 	STATE_UNLOCK(ctx);
3767 
3768 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3769 		return;
3770 	if (in_detach)
3771 		return;
3772 
3773 	CTX_LOCK(ctx);
3774 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3775 		CALLOUT_LOCK(txq);
3776 		callout_stop(&txq->ift_timer);
3777 		CALLOUT_UNLOCK(txq);
3778 	}
3779 	if (do_watchdog) {
3780 		ctx->ifc_watchdog_events++;
3781 		IFDI_WATCHDOG_RESET(ctx);
3782 	}
3783 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3784 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3785 #ifdef DEV_NETMAP
3786 		reset_on = hz / 2;
3787 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3788 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3789 #endif
3790 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3791 	}
3792 	IFDI_LINK_INTR_ENABLE(ctx);
3793 	if (do_reset)
3794 		iflib_if_init_locked(ctx);
3795 	CTX_UNLOCK(ctx);
3796 
3797 	if (LINK_ACTIVE(ctx) == 0)
3798 		return;
3799 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3800 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3801 }
3802 
3803 
3804 static void
3805 _task_fn_iov(void *context)
3806 {
3807 	if_ctx_t ctx = context;
3808 
3809 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3810 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3811 		return;
3812 
3813 	CTX_LOCK(ctx);
3814 	IFDI_VFLR_HANDLE(ctx);
3815 	CTX_UNLOCK(ctx);
3816 }
3817 
3818 static int
3819 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3820 {
3821 	int err;
3822 	if_int_delay_info_t info;
3823 	if_ctx_t ctx;
3824 
3825 	info = (if_int_delay_info_t)arg1;
3826 	ctx = info->iidi_ctx;
3827 	info->iidi_req = req;
3828 	info->iidi_oidp = oidp;
3829 	CTX_LOCK(ctx);
3830 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3831 	CTX_UNLOCK(ctx);
3832 	return (err);
3833 }
3834 
3835 /*********************************************************************
3836  *
3837  *  IFNET FUNCTIONS
3838  *
3839  **********************************************************************/
3840 
3841 static void
3842 iflib_if_init_locked(if_ctx_t ctx)
3843 {
3844 	iflib_stop(ctx);
3845 	iflib_init_locked(ctx);
3846 }
3847 
3848 
3849 static void
3850 iflib_if_init(void *arg)
3851 {
3852 	if_ctx_t ctx = arg;
3853 
3854 	CTX_LOCK(ctx);
3855 	iflib_if_init_locked(ctx);
3856 	CTX_UNLOCK(ctx);
3857 }
3858 
3859 static int
3860 iflib_if_transmit(if_t ifp, struct mbuf *m)
3861 {
3862 	if_ctx_t	ctx = if_getsoftc(ifp);
3863 
3864 	iflib_txq_t txq;
3865 	int err, qidx;
3866 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3867 
3868 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3869 		DBG_COUNTER_INC(tx_frees);
3870 		m_freem(m);
3871 		return (ENOBUFS);
3872 	}
3873 
3874 	MPASS(m->m_nextpkt == NULL);
3875 	/* ALTQ-enabled interfaces always use queue 0. */
3876 	qidx = 0;
3877 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3878 		qidx = QIDX(ctx, m);
3879 	/*
3880 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3881 	 */
3882 	txq = &ctx->ifc_txqs[qidx];
3883 
3884 #ifdef DRIVER_BACKPRESSURE
3885 	if (txq->ift_closed) {
3886 		while (m != NULL) {
3887 			next = m->m_nextpkt;
3888 			m->m_nextpkt = NULL;
3889 			m_freem(m);
3890 			DBG_COUNTER_INC(tx_frees);
3891 			m = next;
3892 		}
3893 		return (ENOBUFS);
3894 	}
3895 #endif
3896 #ifdef notyet
3897 	qidx = count = 0;
3898 	mp = marr;
3899 	next = m;
3900 	do {
3901 		count++;
3902 		next = next->m_nextpkt;
3903 	} while (next != NULL);
3904 
3905 	if (count > nitems(marr))
3906 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3907 			/* XXX check nextpkt */
3908 			m_freem(m);
3909 			/* XXX simplify for now */
3910 			DBG_COUNTER_INC(tx_frees);
3911 			return (ENOBUFS);
3912 		}
3913 	for (next = m, i = 0; next != NULL; i++) {
3914 		mp[i] = next;
3915 		next = next->m_nextpkt;
3916 		mp[i]->m_nextpkt = NULL;
3917 	}
3918 #endif
3919 	DBG_COUNTER_INC(tx_seen);
3920 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3921 
3922 	if (abdicate)
3923 		GROUPTASK_ENQUEUE(&txq->ift_task);
3924  	if (err) {
3925 		if (!abdicate)
3926 			GROUPTASK_ENQUEUE(&txq->ift_task);
3927 		/* support forthcoming later */
3928 #ifdef DRIVER_BACKPRESSURE
3929 		txq->ift_closed = TRUE;
3930 #endif
3931 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3932 		m_freem(m);
3933 		DBG_COUNTER_INC(tx_frees);
3934 	}
3935 
3936 	return (err);
3937 }
3938 
3939 #ifdef ALTQ
3940 /*
3941  * The overall approach to integrating iflib with ALTQ is to continue to use
3942  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3943  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3944  * is redundant/unnecessary, but doing so minimizes the amount of
3945  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3946  * redundantly queueing to an intermediate mp_ring is swamped by the
3947  * performance limitations inherent in using ALTQ.
3948  *
3949  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3950  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3951  * given interface.  If ALTQ is enabled for an interface, then all
3952  * transmitted packets for that interface will be submitted to the ALTQ
3953  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3954  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3955  * update stats that the iflib machinery handles, and which is sensitve to
3956  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3957  * will be installed as the start routine for use by ALTQ facilities that
3958  * need to trigger queue drains on a scheduled basis.
3959  *
3960  */
3961 static void
3962 iflib_altq_if_start(if_t ifp)
3963 {
3964 	struct ifaltq *ifq = &ifp->if_snd;
3965 	struct mbuf *m;
3966 
3967 	IFQ_LOCK(ifq);
3968 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3969 	while (m != NULL) {
3970 		iflib_if_transmit(ifp, m);
3971 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3972 	}
3973 	IFQ_UNLOCK(ifq);
3974 }
3975 
3976 static int
3977 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3978 {
3979 	int err;
3980 
3981 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3982 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3983 		if (err == 0)
3984 			iflib_altq_if_start(ifp);
3985 	} else
3986 		err = iflib_if_transmit(ifp, m);
3987 
3988 	return (err);
3989 }
3990 #endif /* ALTQ */
3991 
3992 static void
3993 iflib_if_qflush(if_t ifp)
3994 {
3995 	if_ctx_t ctx = if_getsoftc(ifp);
3996 	iflib_txq_t txq = ctx->ifc_txqs;
3997 	int i;
3998 
3999 	STATE_LOCK(ctx);
4000 	ctx->ifc_flags |= IFC_QFLUSH;
4001 	STATE_UNLOCK(ctx);
4002 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4003 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4004 			iflib_txq_check_drain(txq, 0);
4005 	STATE_LOCK(ctx);
4006 	ctx->ifc_flags &= ~IFC_QFLUSH;
4007 	STATE_UNLOCK(ctx);
4008 
4009 	/*
4010 	 * When ALTQ is enabled, this will also take care of purging the
4011 	 * ALTQ queue(s).
4012 	 */
4013 	if_qflush(ifp);
4014 }
4015 
4016 
4017 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4018 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4019 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4020 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4021 
4022 static int
4023 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4024 {
4025 	if_ctx_t ctx = if_getsoftc(ifp);
4026 	struct ifreq	*ifr = (struct ifreq *)data;
4027 #if defined(INET) || defined(INET6)
4028 	struct ifaddr	*ifa = (struct ifaddr *)data;
4029 #endif
4030 	bool		avoid_reset = FALSE;
4031 	int		err = 0, reinit = 0, bits;
4032 
4033 	switch (command) {
4034 	case SIOCSIFADDR:
4035 #ifdef INET
4036 		if (ifa->ifa_addr->sa_family == AF_INET)
4037 			avoid_reset = TRUE;
4038 #endif
4039 #ifdef INET6
4040 		if (ifa->ifa_addr->sa_family == AF_INET6)
4041 			avoid_reset = TRUE;
4042 #endif
4043 		/*
4044 		** Calling init results in link renegotiation,
4045 		** so we avoid doing it when possible.
4046 		*/
4047 		if (avoid_reset) {
4048 			if_setflagbits(ifp, IFF_UP,0);
4049 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4050 				reinit = 1;
4051 #ifdef INET
4052 			if (!(if_getflags(ifp) & IFF_NOARP))
4053 				arp_ifinit(ifp, ifa);
4054 #endif
4055 		} else
4056 			err = ether_ioctl(ifp, command, data);
4057 		break;
4058 	case SIOCSIFMTU:
4059 		CTX_LOCK(ctx);
4060 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4061 			CTX_UNLOCK(ctx);
4062 			break;
4063 		}
4064 		bits = if_getdrvflags(ifp);
4065 		/* stop the driver and free any clusters before proceeding */
4066 		iflib_stop(ctx);
4067 
4068 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4069 			STATE_LOCK(ctx);
4070 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4071 				ctx->ifc_flags |= IFC_MULTISEG;
4072 			else
4073 				ctx->ifc_flags &= ~IFC_MULTISEG;
4074 			STATE_UNLOCK(ctx);
4075 			err = if_setmtu(ifp, ifr->ifr_mtu);
4076 		}
4077 		iflib_init_locked(ctx);
4078 		STATE_LOCK(ctx);
4079 		if_setdrvflags(ifp, bits);
4080 		STATE_UNLOCK(ctx);
4081 		CTX_UNLOCK(ctx);
4082 		break;
4083 	case SIOCSIFFLAGS:
4084 		CTX_LOCK(ctx);
4085 		if (if_getflags(ifp) & IFF_UP) {
4086 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4087 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4088 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4089 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4090 				}
4091 			} else
4092 				reinit = 1;
4093 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4094 			iflib_stop(ctx);
4095 		}
4096 		ctx->ifc_if_flags = if_getflags(ifp);
4097 		CTX_UNLOCK(ctx);
4098 		break;
4099 	case SIOCADDMULTI:
4100 	case SIOCDELMULTI:
4101 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4102 			CTX_LOCK(ctx);
4103 			IFDI_INTR_DISABLE(ctx);
4104 			IFDI_MULTI_SET(ctx);
4105 			IFDI_INTR_ENABLE(ctx);
4106 			CTX_UNLOCK(ctx);
4107 		}
4108 		break;
4109 	case SIOCSIFMEDIA:
4110 		CTX_LOCK(ctx);
4111 		IFDI_MEDIA_SET(ctx);
4112 		CTX_UNLOCK(ctx);
4113 		/* falls thru */
4114 	case SIOCGIFMEDIA:
4115 	case SIOCGIFXMEDIA:
4116 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4117 		break;
4118 	case SIOCGI2C:
4119 	{
4120 		struct ifi2creq i2c;
4121 
4122 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4123 		if (err != 0)
4124 			break;
4125 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4126 			err = EINVAL;
4127 			break;
4128 		}
4129 		if (i2c.len > sizeof(i2c.data)) {
4130 			err = EINVAL;
4131 			break;
4132 		}
4133 
4134 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4135 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4136 			    sizeof(i2c));
4137 		break;
4138 	}
4139 	case SIOCSIFCAP:
4140 	{
4141 		int mask, setmask, oldmask;
4142 
4143 		oldmask = if_getcapenable(ifp);
4144 		mask = ifr->ifr_reqcap ^ oldmask;
4145 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4146 		setmask = 0;
4147 #ifdef TCP_OFFLOAD
4148 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4149 #endif
4150 		setmask |= (mask & IFCAP_FLAGS);
4151 		setmask |= (mask & IFCAP_WOL);
4152 
4153 		/*
4154 		 * If any RX csum has changed, change all the ones that
4155 		 * are supported by the driver.
4156 		 */
4157 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4158 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4159 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4160 		}
4161 
4162 		/*
4163 		 * want to ensure that traffic has stopped before we change any of the flags
4164 		 */
4165 		if (setmask) {
4166 			CTX_LOCK(ctx);
4167 			bits = if_getdrvflags(ifp);
4168 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4169 				iflib_stop(ctx);
4170 			STATE_LOCK(ctx);
4171 			if_togglecapenable(ifp, setmask);
4172 			STATE_UNLOCK(ctx);
4173 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4174 				iflib_init_locked(ctx);
4175 			STATE_LOCK(ctx);
4176 			if_setdrvflags(ifp, bits);
4177 			STATE_UNLOCK(ctx);
4178 			CTX_UNLOCK(ctx);
4179 		}
4180 		if_vlancap(ifp);
4181 		break;
4182 	}
4183 	case SIOCGPRIVATE_0:
4184 	case SIOCSDRVSPEC:
4185 	case SIOCGDRVSPEC:
4186 		CTX_LOCK(ctx);
4187 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4188 		CTX_UNLOCK(ctx);
4189 		break;
4190 	default:
4191 		err = ether_ioctl(ifp, command, data);
4192 		break;
4193 	}
4194 	if (reinit)
4195 		iflib_if_init(ctx);
4196 	return (err);
4197 }
4198 
4199 static uint64_t
4200 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4201 {
4202 	if_ctx_t ctx = if_getsoftc(ifp);
4203 
4204 	return (IFDI_GET_COUNTER(ctx, cnt));
4205 }
4206 
4207 /*********************************************************************
4208  *
4209  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4210  *
4211  **********************************************************************/
4212 
4213 static void
4214 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4215 {
4216 	if_ctx_t ctx = if_getsoftc(ifp);
4217 
4218 	if ((void *)ctx != arg)
4219 		return;
4220 
4221 	if ((vtag == 0) || (vtag > 4095))
4222 		return;
4223 
4224 	CTX_LOCK(ctx);
4225 	IFDI_VLAN_REGISTER(ctx, vtag);
4226 	/* Re-init to load the changes */
4227 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4228 		iflib_if_init_locked(ctx);
4229 	CTX_UNLOCK(ctx);
4230 }
4231 
4232 static void
4233 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4234 {
4235 	if_ctx_t ctx = if_getsoftc(ifp);
4236 
4237 	if ((void *)ctx != arg)
4238 		return;
4239 
4240 	if ((vtag == 0) || (vtag > 4095))
4241 		return;
4242 
4243 	CTX_LOCK(ctx);
4244 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4245 	/* Re-init to load the changes */
4246 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4247 		iflib_if_init_locked(ctx);
4248 	CTX_UNLOCK(ctx);
4249 }
4250 
4251 static void
4252 iflib_led_func(void *arg, int onoff)
4253 {
4254 	if_ctx_t ctx = arg;
4255 
4256 	CTX_LOCK(ctx);
4257 	IFDI_LED_FUNC(ctx, onoff);
4258 	CTX_UNLOCK(ctx);
4259 }
4260 
4261 /*********************************************************************
4262  *
4263  *  BUS FUNCTION DEFINITIONS
4264  *
4265  **********************************************************************/
4266 
4267 int
4268 iflib_device_probe(device_t dev)
4269 {
4270 	pci_vendor_info_t *ent;
4271 
4272 	uint16_t	pci_vendor_id, pci_device_id;
4273 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4274 	uint16_t	pci_rev_id;
4275 	if_shared_ctx_t sctx;
4276 
4277 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4278 		return (ENOTSUP);
4279 
4280 	pci_vendor_id = pci_get_vendor(dev);
4281 	pci_device_id = pci_get_device(dev);
4282 	pci_subvendor_id = pci_get_subvendor(dev);
4283 	pci_subdevice_id = pci_get_subdevice(dev);
4284 	pci_rev_id = pci_get_revid(dev);
4285 	if (sctx->isc_parse_devinfo != NULL)
4286 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4287 
4288 	ent = sctx->isc_vendor_info;
4289 	while (ent->pvi_vendor_id != 0) {
4290 		if (pci_vendor_id != ent->pvi_vendor_id) {
4291 			ent++;
4292 			continue;
4293 		}
4294 		if ((pci_device_id == ent->pvi_device_id) &&
4295 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4296 		     (ent->pvi_subvendor_id == 0)) &&
4297 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4298 		     (ent->pvi_subdevice_id == 0)) &&
4299 		    ((pci_rev_id == ent->pvi_rev_id) ||
4300 		     (ent->pvi_rev_id == 0))) {
4301 
4302 			device_set_desc_copy(dev, ent->pvi_name);
4303 			/* this needs to be changed to zero if the bus probing code
4304 			 * ever stops re-probing on best match because the sctx
4305 			 * may have its values over written by register calls
4306 			 * in subsequent probes
4307 			 */
4308 			return (BUS_PROBE_DEFAULT);
4309 		}
4310 		ent++;
4311 	}
4312 	return (ENXIO);
4313 }
4314 
4315 static void
4316 iflib_reset_qvalues(if_ctx_t ctx)
4317 {
4318 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4319 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4320 	device_t dev = ctx->ifc_dev;
4321 	int i;
4322 
4323 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4324 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4325 	/*
4326 	 * XXX sanity check that ntxd & nrxd are a power of 2
4327 	 */
4328 	if (ctx->ifc_sysctl_ntxqs != 0)
4329 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4330 	if (ctx->ifc_sysctl_nrxqs != 0)
4331 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4332 
4333 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4334 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4335 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4336 		else
4337 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4338 	}
4339 
4340 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4341 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4342 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4343 		else
4344 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4345 	}
4346 
4347 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4348 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4349 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4350 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4351 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4352 		}
4353 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4354 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4355 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4356 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4357 		}
4358 	}
4359 
4360 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4361 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4362 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4363 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4364 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4365 		}
4366 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4367 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4368 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4369 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4370 		}
4371 	}
4372 }
4373 
4374 int
4375 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4376 {
4377 	int err, rid, msix;
4378 	if_ctx_t ctx;
4379 	if_t ifp;
4380 	if_softc_ctx_t scctx;
4381 	int i;
4382 	uint16_t main_txq;
4383 	uint16_t main_rxq;
4384 
4385 
4386 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4387 
4388 	if (sc == NULL) {
4389 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4390 		device_set_softc(dev, ctx);
4391 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4392 	}
4393 
4394 	ctx->ifc_sctx = sctx;
4395 	ctx->ifc_dev = dev;
4396 	ctx->ifc_softc = sc;
4397 
4398 	if ((err = iflib_register(ctx)) != 0) {
4399 		device_printf(dev, "iflib_register failed %d\n", err);
4400 		goto fail_ctx_free;
4401 	}
4402 	iflib_add_device_sysctl_pre(ctx);
4403 
4404 	scctx = &ctx->ifc_softc_ctx;
4405 	ifp = ctx->ifc_ifp;
4406 
4407 	iflib_reset_qvalues(ctx);
4408 	CTX_LOCK(ctx);
4409 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4410 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4411 		goto fail_unlock;
4412 	}
4413 	_iflib_pre_assert(scctx);
4414 	ctx->ifc_txrx = *scctx->isc_txrx;
4415 
4416 #ifdef INVARIANTS
4417 	MPASS(scctx->isc_capabilities);
4418 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4419 		MPASS(scctx->isc_tx_csum_flags);
4420 #endif
4421 
4422 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4423 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4424 
4425 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4426 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4427 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4428 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4429 
4430 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4431 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4432 
4433 	/* XXX change for per-queue sizes */
4434 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4435 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4436 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4437 		if (!powerof2(scctx->isc_nrxd[i])) {
4438 			/* round down instead? */
4439 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4440 			err = EINVAL;
4441 			goto fail_iflib_detach;
4442 		}
4443 	}
4444 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4445 		if (!powerof2(scctx->isc_ntxd[i])) {
4446 			device_printf(dev,
4447 			    "# tx descriptors must be a power of 2");
4448 			err = EINVAL;
4449 			goto fail_iflib_detach;
4450 		}
4451 	}
4452 
4453 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4454 	    MAX_SINGLE_PACKET_FRACTION)
4455 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4456 		    MAX_SINGLE_PACKET_FRACTION);
4457 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4458 	    MAX_SINGLE_PACKET_FRACTION)
4459 		scctx->isc_tx_tso_segments_max = max(1,
4460 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4461 
4462 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4463 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4464 		/*
4465 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4466 		 * but some MACs do.
4467 		 */
4468 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4469 		    IP_MAXPACKET));
4470 		/*
4471 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4472 		 * into account.  In the worst case, each of these calls will
4473 		 * add another mbuf and, thus, the requirement for another DMA
4474 		 * segment.  So for best performance, it doesn't make sense to
4475 		 * advertize a maximum of TSO segments that typically will
4476 		 * require defragmentation in iflib_encap().
4477 		 */
4478 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4479 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4480 	}
4481 	if (scctx->isc_rss_table_size == 0)
4482 		scctx->isc_rss_table_size = 64;
4483 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4484 
4485 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4486 	/* XXX format name */
4487 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4488 
4489 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4490 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4491 		device_printf(dev, "Unable to fetch CPU list\n");
4492 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4493 	}
4494 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4495 
4496 	/*
4497 	** Now set up MSI or MSI-X, should return us the number of supported
4498 	** vectors (will be 1 for a legacy interrupt and MSI).
4499 	*/
4500 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4501 		msix = scctx->isc_vectors;
4502 	} else if (scctx->isc_msix_bar != 0)
4503 	       /*
4504 		* The simple fact that isc_msix_bar is not 0 does not mean we
4505 		* we have a good value there that is known to work.
4506 		*/
4507 		msix = iflib_msix_init(ctx);
4508 	else {
4509 		scctx->isc_vectors = 1;
4510 		scctx->isc_ntxqsets = 1;
4511 		scctx->isc_nrxqsets = 1;
4512 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4513 		msix = 0;
4514 	}
4515 	/* Get memory for the station queues */
4516 	if ((err = iflib_queues_alloc(ctx))) {
4517 		device_printf(dev, "Unable to allocate queue memory\n");
4518 		goto fail_intr_free;
4519 	}
4520 
4521 	if ((err = iflib_qset_structures_setup(ctx)))
4522 		goto fail_queues;
4523 
4524 	/*
4525 	 * Group taskqueues aren't properly set up until SMP is started,
4526 	 * so we disable interrupts until we can handle them post
4527 	 * SI_SUB_SMP.
4528 	 *
4529 	 * XXX: disabling interrupts doesn't actually work, at least for
4530 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4531 	 * we do null handling and depend on this not causing too large an
4532 	 * interrupt storm.
4533 	 */
4534 	IFDI_INTR_DISABLE(ctx);
4535 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4536 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4537 		goto fail_queues;
4538 	}
4539 	if (msix <= 1) {
4540 		rid = 0;
4541 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4542 			MPASS(msix == 1);
4543 			rid = 1;
4544 		}
4545 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4546 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4547 			goto fail_queues;
4548 		}
4549 	}
4550 
4551 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4552 
4553 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4554 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4555 		goto fail_detach;
4556 	}
4557 
4558 	/*
4559 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4560 	 * This must appear after the call to ether_ifattach() because
4561 	 * ether_ifattach() sets if_hdrlen to the default value.
4562 	 */
4563 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4564 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4565 
4566 	if ((err = iflib_netmap_attach(ctx))) {
4567 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4568 		goto fail_detach;
4569 	}
4570 	*ctxp = ctx;
4571 
4572 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4573 
4574 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4575 	iflib_add_device_sysctl_post(ctx);
4576 	ctx->ifc_flags |= IFC_INIT_DONE;
4577 	CTX_UNLOCK(ctx);
4578 	return (0);
4579 
4580 fail_detach:
4581 	ether_ifdetach(ctx->ifc_ifp);
4582 fail_intr_free:
4583 	iflib_free_intr_mem(ctx);
4584 fail_queues:
4585 	iflib_tx_structures_free(ctx);
4586 	iflib_rx_structures_free(ctx);
4587 fail_iflib_detach:
4588 	IFDI_DETACH(ctx);
4589 fail_unlock:
4590 	CTX_UNLOCK(ctx);
4591 fail_ctx_free:
4592         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4593                 free(ctx->ifc_softc, M_IFLIB);
4594         free(ctx, M_IFLIB);
4595 	return (err);
4596 }
4597 
4598 int
4599 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4600 					  struct iflib_cloneattach_ctx *clctx)
4601 {
4602 	int err;
4603 	if_ctx_t ctx;
4604 	if_t ifp;
4605 	if_softc_ctx_t scctx;
4606 	int i;
4607 	void *sc;
4608 	uint16_t main_txq;
4609 	uint16_t main_rxq;
4610 
4611 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4612 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4613 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4614 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4615 		ctx->ifc_flags |= IFC_PSEUDO;
4616 
4617 	ctx->ifc_sctx = sctx;
4618 	ctx->ifc_softc = sc;
4619 	ctx->ifc_dev = dev;
4620 
4621 	if ((err = iflib_register(ctx)) != 0) {
4622 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4623 		goto fail_ctx_free;
4624 	}
4625 	iflib_add_device_sysctl_pre(ctx);
4626 
4627 	scctx = &ctx->ifc_softc_ctx;
4628 	ifp = ctx->ifc_ifp;
4629 
4630 	/*
4631 	 * XXX sanity check that ntxd & nrxd are a power of 2
4632 	 */
4633 	iflib_reset_qvalues(ctx);
4634 
4635 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4636 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4637 		goto fail_ctx_free;
4638 	}
4639 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4640 		iflib_gen_mac(ctx);
4641 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4642 								clctx->cc_params)) != 0) {
4643 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4644 		goto fail_ctx_free;
4645 	}
4646 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4647 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4648 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4649 
4650 #ifdef INVARIANTS
4651 	MPASS(scctx->isc_capabilities);
4652 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4653 		MPASS(scctx->isc_tx_csum_flags);
4654 #endif
4655 
4656 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4657 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4658 
4659 	ifp->if_flags |= IFF_NOGROUP;
4660 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4661 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4662 
4663 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4664 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4665 			goto fail_detach;
4666 		}
4667 		*ctxp = ctx;
4668 
4669 		/*
4670 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4671 		 * This must appear after the call to ether_ifattach() because
4672 		 * ether_ifattach() sets if_hdrlen to the default value.
4673 		 */
4674 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4675 			if_setifheaderlen(ifp,
4676 			    sizeof(struct ether_vlan_header));
4677 
4678 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4679 		iflib_add_device_sysctl_post(ctx);
4680 		ctx->ifc_flags |= IFC_INIT_DONE;
4681 		return (0);
4682 	}
4683 	_iflib_pre_assert(scctx);
4684 	ctx->ifc_txrx = *scctx->isc_txrx;
4685 
4686 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4687 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4688 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4689 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4690 
4691 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4692 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4693 
4694 	/* XXX change for per-queue sizes */
4695 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4696 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4697 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4698 		if (!powerof2(scctx->isc_nrxd[i])) {
4699 			/* round down instead? */
4700 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4701 			err = EINVAL;
4702 			goto fail_iflib_detach;
4703 		}
4704 	}
4705 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4706 		if (!powerof2(scctx->isc_ntxd[i])) {
4707 			device_printf(dev,
4708 			    "# tx descriptors must be a power of 2");
4709 			err = EINVAL;
4710 			goto fail_iflib_detach;
4711 		}
4712 	}
4713 
4714 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4715 	    MAX_SINGLE_PACKET_FRACTION)
4716 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4717 		    MAX_SINGLE_PACKET_FRACTION);
4718 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4719 	    MAX_SINGLE_PACKET_FRACTION)
4720 		scctx->isc_tx_tso_segments_max = max(1,
4721 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4722 
4723 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4724 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4725 		/*
4726 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4727 		 * but some MACs do.
4728 		 */
4729 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4730 		    IP_MAXPACKET));
4731 		/*
4732 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4733 		 * into account.  In the worst case, each of these calls will
4734 		 * add another mbuf and, thus, the requirement for another DMA
4735 		 * segment.  So for best performance, it doesn't make sense to
4736 		 * advertize a maximum of TSO segments that typically will
4737 		 * require defragmentation in iflib_encap().
4738 		 */
4739 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4740 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4741 	}
4742 	if (scctx->isc_rss_table_size == 0)
4743 		scctx->isc_rss_table_size = 64;
4744 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4745 
4746 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4747 	/* XXX format name */
4748 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4749 
4750 	/* XXX --- can support > 1 -- but keep it simple for now */
4751 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4752 
4753 	/* Get memory for the station queues */
4754 	if ((err = iflib_queues_alloc(ctx))) {
4755 		device_printf(dev, "Unable to allocate queue memory\n");
4756 		goto fail_iflib_detach;
4757 	}
4758 
4759 	if ((err = iflib_qset_structures_setup(ctx))) {
4760 		device_printf(dev, "qset structure setup failed %d\n", err);
4761 		goto fail_queues;
4762 	}
4763 
4764 	/*
4765 	 * XXX What if anything do we want to do about interrupts?
4766 	 */
4767 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4768 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4769 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4770 		goto fail_detach;
4771 	}
4772 
4773 	/*
4774 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4775 	 * This must appear after the call to ether_ifattach() because
4776 	 * ether_ifattach() sets if_hdrlen to the default value.
4777 	 */
4778 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4779 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4780 
4781 	/* XXX handle more than one queue */
4782 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4783 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4784 
4785 	*ctxp = ctx;
4786 
4787 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4788 	iflib_add_device_sysctl_post(ctx);
4789 	ctx->ifc_flags |= IFC_INIT_DONE;
4790 	return (0);
4791 fail_detach:
4792 	ether_ifdetach(ctx->ifc_ifp);
4793 fail_queues:
4794 	iflib_tx_structures_free(ctx);
4795 	iflib_rx_structures_free(ctx);
4796 fail_iflib_detach:
4797 	IFDI_DETACH(ctx);
4798 fail_ctx_free:
4799 	free(ctx->ifc_softc, M_IFLIB);
4800 	free(ctx, M_IFLIB);
4801 	return (err);
4802 }
4803 
4804 int
4805 iflib_pseudo_deregister(if_ctx_t ctx)
4806 {
4807 	if_t ifp = ctx->ifc_ifp;
4808 	iflib_txq_t txq;
4809 	iflib_rxq_t rxq;
4810 	int i, j;
4811 	struct taskqgroup *tqg;
4812 	iflib_fl_t fl;
4813 
4814 	/* Unregister VLAN events */
4815 	if (ctx->ifc_vlan_attach_event != NULL)
4816 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4817 	if (ctx->ifc_vlan_detach_event != NULL)
4818 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4819 
4820 	ether_ifdetach(ifp);
4821 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4822 	CTX_LOCK_DESTROY(ctx);
4823 	/* XXX drain any dependent tasks */
4824 	tqg = qgroup_if_io_tqg;
4825 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4826 		callout_drain(&txq->ift_timer);
4827 		if (txq->ift_task.gt_uniq != NULL)
4828 			taskqgroup_detach(tqg, &txq->ift_task);
4829 	}
4830 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4831 		if (rxq->ifr_task.gt_uniq != NULL)
4832 			taskqgroup_detach(tqg, &rxq->ifr_task);
4833 
4834 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4835 			free(fl->ifl_rx_bitmap, M_IFLIB);
4836 	}
4837 	tqg = qgroup_if_config_tqg;
4838 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4839 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4840 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4841 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4842 
4843 	if_free(ifp);
4844 
4845 	iflib_tx_structures_free(ctx);
4846 	iflib_rx_structures_free(ctx);
4847 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4848 		free(ctx->ifc_softc, M_IFLIB);
4849 	free(ctx, M_IFLIB);
4850 	return (0);
4851 }
4852 
4853 int
4854 iflib_device_attach(device_t dev)
4855 {
4856 	if_ctx_t ctx;
4857 	if_shared_ctx_t sctx;
4858 
4859 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4860 		return (ENOTSUP);
4861 
4862 	pci_enable_busmaster(dev);
4863 
4864 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4865 }
4866 
4867 int
4868 iflib_device_deregister(if_ctx_t ctx)
4869 {
4870 	if_t ifp = ctx->ifc_ifp;
4871 	iflib_txq_t txq;
4872 	iflib_rxq_t rxq;
4873 	device_t dev = ctx->ifc_dev;
4874 	int i, j;
4875 	struct taskqgroup *tqg;
4876 	iflib_fl_t fl;
4877 
4878 	/* Make sure VLANS are not using driver */
4879 	if (if_vlantrunkinuse(ifp)) {
4880 		device_printf(dev, "Vlan in use, detach first\n");
4881 		return (EBUSY);
4882 	}
4883 #ifdef PCI_IOV
4884 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4885 		device_printf(dev, "SR-IOV in use; detach first.\n");
4886 		return (EBUSY);
4887 	}
4888 #endif
4889 
4890 	STATE_LOCK(ctx);
4891 	ctx->ifc_flags |= IFC_IN_DETACH;
4892 	STATE_UNLOCK(ctx);
4893 
4894 	CTX_LOCK(ctx);
4895 	iflib_stop(ctx);
4896 	CTX_UNLOCK(ctx);
4897 
4898 	/* Unregister VLAN events */
4899 	if (ctx->ifc_vlan_attach_event != NULL)
4900 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4901 	if (ctx->ifc_vlan_detach_event != NULL)
4902 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4903 
4904 	iflib_netmap_detach(ifp);
4905 	ether_ifdetach(ifp);
4906 	if (ctx->ifc_led_dev != NULL)
4907 		led_destroy(ctx->ifc_led_dev);
4908 	/* XXX drain any dependent tasks */
4909 	tqg = qgroup_if_io_tqg;
4910 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4911 		callout_drain(&txq->ift_timer);
4912 		if (txq->ift_task.gt_uniq != NULL)
4913 			taskqgroup_detach(tqg, &txq->ift_task);
4914 	}
4915 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4916 		if (rxq->ifr_task.gt_uniq != NULL)
4917 			taskqgroup_detach(tqg, &rxq->ifr_task);
4918 
4919 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4920 			free(fl->ifl_rx_bitmap, M_IFLIB);
4921 	}
4922 	tqg = qgroup_if_config_tqg;
4923 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4924 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4925 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4926 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4927 	CTX_LOCK(ctx);
4928 	IFDI_DETACH(ctx);
4929 	CTX_UNLOCK(ctx);
4930 
4931 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4932 	CTX_LOCK_DESTROY(ctx);
4933 	device_set_softc(ctx->ifc_dev, NULL);
4934 	iflib_free_intr_mem(ctx);
4935 
4936 	bus_generic_detach(dev);
4937 	if_free(ifp);
4938 
4939 	iflib_tx_structures_free(ctx);
4940 	iflib_rx_structures_free(ctx);
4941 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4942 		free(ctx->ifc_softc, M_IFLIB);
4943 	STATE_LOCK_DESTROY(ctx);
4944 	free(ctx, M_IFLIB);
4945 	return (0);
4946 }
4947 
4948 static void
4949 iflib_free_intr_mem(if_ctx_t ctx)
4950 {
4951 
4952 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4953 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4954 	}
4955 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4956 		pci_release_msi(ctx->ifc_dev);
4957 	}
4958 	if (ctx->ifc_msix_mem != NULL) {
4959 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4960 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
4961 		ctx->ifc_msix_mem = NULL;
4962 	}
4963 }
4964 
4965 int
4966 iflib_device_detach(device_t dev)
4967 {
4968 	if_ctx_t ctx = device_get_softc(dev);
4969 
4970 	return (iflib_device_deregister(ctx));
4971 }
4972 
4973 int
4974 iflib_device_suspend(device_t dev)
4975 {
4976 	if_ctx_t ctx = device_get_softc(dev);
4977 
4978 	CTX_LOCK(ctx);
4979 	IFDI_SUSPEND(ctx);
4980 	CTX_UNLOCK(ctx);
4981 
4982 	return bus_generic_suspend(dev);
4983 }
4984 int
4985 iflib_device_shutdown(device_t dev)
4986 {
4987 	if_ctx_t ctx = device_get_softc(dev);
4988 
4989 	CTX_LOCK(ctx);
4990 	IFDI_SHUTDOWN(ctx);
4991 	CTX_UNLOCK(ctx);
4992 
4993 	return bus_generic_suspend(dev);
4994 }
4995 
4996 
4997 int
4998 iflib_device_resume(device_t dev)
4999 {
5000 	if_ctx_t ctx = device_get_softc(dev);
5001 	iflib_txq_t txq = ctx->ifc_txqs;
5002 
5003 	CTX_LOCK(ctx);
5004 	IFDI_RESUME(ctx);
5005 	iflib_if_init_locked(ctx);
5006 	CTX_UNLOCK(ctx);
5007 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5008 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5009 
5010 	return (bus_generic_resume(dev));
5011 }
5012 
5013 int
5014 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5015 {
5016 	int error;
5017 	if_ctx_t ctx = device_get_softc(dev);
5018 
5019 	CTX_LOCK(ctx);
5020 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5021 	CTX_UNLOCK(ctx);
5022 
5023 	return (error);
5024 }
5025 
5026 void
5027 iflib_device_iov_uninit(device_t dev)
5028 {
5029 	if_ctx_t ctx = device_get_softc(dev);
5030 
5031 	CTX_LOCK(ctx);
5032 	IFDI_IOV_UNINIT(ctx);
5033 	CTX_UNLOCK(ctx);
5034 }
5035 
5036 int
5037 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5038 {
5039 	int error;
5040 	if_ctx_t ctx = device_get_softc(dev);
5041 
5042 	CTX_LOCK(ctx);
5043 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5044 	CTX_UNLOCK(ctx);
5045 
5046 	return (error);
5047 }
5048 
5049 /*********************************************************************
5050  *
5051  *  MODULE FUNCTION DEFINITIONS
5052  *
5053  **********************************************************************/
5054 
5055 /*
5056  * - Start a fast taskqueue thread for each core
5057  * - Start a taskqueue for control operations
5058  */
5059 static int
5060 iflib_module_init(void)
5061 {
5062 	return (0);
5063 }
5064 
5065 static int
5066 iflib_module_event_handler(module_t mod, int what, void *arg)
5067 {
5068 	int err;
5069 
5070 	switch (what) {
5071 	case MOD_LOAD:
5072 		if ((err = iflib_module_init()) != 0)
5073 			return (err);
5074 		break;
5075 	case MOD_UNLOAD:
5076 		return (EBUSY);
5077 	default:
5078 		return (EOPNOTSUPP);
5079 	}
5080 
5081 	return (0);
5082 }
5083 
5084 /*********************************************************************
5085  *
5086  *  PUBLIC FUNCTION DEFINITIONS
5087  *     ordered as in iflib.h
5088  *
5089  **********************************************************************/
5090 
5091 
5092 static void
5093 _iflib_assert(if_shared_ctx_t sctx)
5094 {
5095 	MPASS(sctx->isc_tx_maxsize);
5096 	MPASS(sctx->isc_tx_maxsegsize);
5097 
5098 	MPASS(sctx->isc_rx_maxsize);
5099 	MPASS(sctx->isc_rx_nsegments);
5100 	MPASS(sctx->isc_rx_maxsegsize);
5101 
5102 	MPASS(sctx->isc_nrxd_min[0]);
5103 	MPASS(sctx->isc_nrxd_max[0]);
5104 	MPASS(sctx->isc_nrxd_default[0]);
5105 	MPASS(sctx->isc_ntxd_min[0]);
5106 	MPASS(sctx->isc_ntxd_max[0]);
5107 	MPASS(sctx->isc_ntxd_default[0]);
5108 }
5109 
5110 static void
5111 _iflib_pre_assert(if_softc_ctx_t scctx)
5112 {
5113 
5114 	MPASS(scctx->isc_txrx->ift_txd_encap);
5115 	MPASS(scctx->isc_txrx->ift_txd_flush);
5116 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5117 	MPASS(scctx->isc_txrx->ift_rxd_available);
5118 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5119 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5120 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5121 }
5122 
5123 static int
5124 iflib_register(if_ctx_t ctx)
5125 {
5126 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5127 	driver_t *driver = sctx->isc_driver;
5128 	device_t dev = ctx->ifc_dev;
5129 	if_t ifp;
5130 
5131 	_iflib_assert(sctx);
5132 
5133 	CTX_LOCK_INIT(ctx);
5134 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5135 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5136 	if (ifp == NULL) {
5137 		device_printf(dev, "can not allocate ifnet structure\n");
5138 		return (ENOMEM);
5139 	}
5140 
5141 	/*
5142 	 * Initialize our context's device specific methods
5143 	 */
5144 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5145 	kobj_class_compile((kobj_class_t) driver);
5146 	driver->refs++;
5147 
5148 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5149 	if_setsoftc(ifp, ctx);
5150 	if_setdev(ifp, dev);
5151 	if_setinitfn(ifp, iflib_if_init);
5152 	if_setioctlfn(ifp, iflib_if_ioctl);
5153 #ifdef ALTQ
5154 	if_setstartfn(ifp, iflib_altq_if_start);
5155 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5156 	if_setsendqready(ifp);
5157 #else
5158 	if_settransmitfn(ifp, iflib_if_transmit);
5159 #endif
5160 	if_setqflushfn(ifp, iflib_if_qflush);
5161 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5162 
5163 	ctx->ifc_vlan_attach_event =
5164 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5165 							  EVENTHANDLER_PRI_FIRST);
5166 	ctx->ifc_vlan_detach_event =
5167 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5168 							  EVENTHANDLER_PRI_FIRST);
5169 
5170 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5171 					 iflib_media_change, iflib_media_status);
5172 
5173 	return (0);
5174 }
5175 
5176 
5177 static int
5178 iflib_queues_alloc(if_ctx_t ctx)
5179 {
5180 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5181 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5182 	device_t dev = ctx->ifc_dev;
5183 	int nrxqsets = scctx->isc_nrxqsets;
5184 	int ntxqsets = scctx->isc_ntxqsets;
5185 	iflib_txq_t txq;
5186 	iflib_rxq_t rxq;
5187 	iflib_fl_t fl = NULL;
5188 	int i, j, cpu, err, txconf, rxconf;
5189 	iflib_dma_info_t ifdip;
5190 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5191 	uint32_t *txqsizes = scctx->isc_txqsizes;
5192 	uint8_t nrxqs = sctx->isc_nrxqs;
5193 	uint8_t ntxqs = sctx->isc_ntxqs;
5194 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5195 	caddr_t *vaddrs;
5196 	uint64_t *paddrs;
5197 
5198 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5199 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5200 
5201 	/* Allocate the TX ring struct memory */
5202 	if (!(ctx->ifc_txqs =
5203 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5204 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5205 		device_printf(dev, "Unable to allocate TX ring memory\n");
5206 		err = ENOMEM;
5207 		goto fail;
5208 	}
5209 
5210 	/* Now allocate the RX */
5211 	if (!(ctx->ifc_rxqs =
5212 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5213 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5214 		device_printf(dev, "Unable to allocate RX ring memory\n");
5215 		err = ENOMEM;
5216 		goto rx_fail;
5217 	}
5218 
5219 	txq = ctx->ifc_txqs;
5220 	rxq = ctx->ifc_rxqs;
5221 
5222 	/*
5223 	 * XXX handle allocation failure
5224 	 */
5225 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5226 		/* Set up some basics */
5227 
5228 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5229 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5230 			device_printf(dev,
5231 			    "Unable to allocate TX DMA info memory\n");
5232 			err = ENOMEM;
5233 			goto err_tx_desc;
5234 		}
5235 		txq->ift_ifdi = ifdip;
5236 		for (j = 0; j < ntxqs; j++, ifdip++) {
5237 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5238 				device_printf(dev,
5239 				    "Unable to allocate TX descriptors\n");
5240 				err = ENOMEM;
5241 				goto err_tx_desc;
5242 			}
5243 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5244 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5245 		}
5246 		txq->ift_ctx = ctx;
5247 		txq->ift_id = i;
5248 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5249 			txq->ift_br_offset = 1;
5250 		} else {
5251 			txq->ift_br_offset = 0;
5252 		}
5253 		/* XXX fix this */
5254 		txq->ift_timer.c_cpu = cpu;
5255 
5256 		if (iflib_txsd_alloc(txq)) {
5257 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5258 			err = ENOMEM;
5259 			goto err_tx_desc;
5260 		}
5261 
5262 		/* Initialize the TX lock */
5263 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5264 		    device_get_nameunit(dev), txq->ift_id);
5265 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5266 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5267 
5268 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5269 			 device_get_nameunit(dev), txq->ift_id);
5270 
5271 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5272 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5273 		if (err) {
5274 			/* XXX free any allocated rings */
5275 			device_printf(dev, "Unable to allocate buf_ring\n");
5276 			goto err_tx_desc;
5277 		}
5278 	}
5279 
5280 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5281 		/* Set up some basics */
5282 
5283 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5284 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5285 			device_printf(dev,
5286 			    "Unable to allocate RX DMA info memory\n");
5287 			err = ENOMEM;
5288 			goto err_tx_desc;
5289 		}
5290 
5291 		rxq->ifr_ifdi = ifdip;
5292 		/* XXX this needs to be changed if #rx queues != #tx queues */
5293 		rxq->ifr_ntxqirq = 1;
5294 		rxq->ifr_txqid[0] = i;
5295 		for (j = 0; j < nrxqs; j++, ifdip++) {
5296 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5297 				device_printf(dev,
5298 				    "Unable to allocate RX descriptors\n");
5299 				err = ENOMEM;
5300 				goto err_tx_desc;
5301 			}
5302 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5303 		}
5304 		rxq->ifr_ctx = ctx;
5305 		rxq->ifr_id = i;
5306 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5307 			rxq->ifr_fl_offset = 1;
5308 		} else {
5309 			rxq->ifr_fl_offset = 0;
5310 		}
5311 		rxq->ifr_nfl = nfree_lists;
5312 		if (!(fl =
5313 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5314 			device_printf(dev, "Unable to allocate free list memory\n");
5315 			err = ENOMEM;
5316 			goto err_tx_desc;
5317 		}
5318 		rxq->ifr_fl = fl;
5319 		for (j = 0; j < nfree_lists; j++) {
5320 			fl[j].ifl_rxq = rxq;
5321 			fl[j].ifl_id = j;
5322 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5323 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5324 		}
5325 		/* Allocate receive buffers for the ring */
5326 		if (iflib_rxsd_alloc(rxq)) {
5327 			device_printf(dev,
5328 			    "Critical Failure setting up receive buffers\n");
5329 			err = ENOMEM;
5330 			goto err_rx_desc;
5331 		}
5332 
5333 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5334 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5335 			    M_WAITOK);
5336 	}
5337 
5338 	/* TXQs */
5339 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5340 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5341 	for (i = 0; i < ntxqsets; i++) {
5342 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5343 
5344 		for (j = 0; j < ntxqs; j++, di++) {
5345 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5346 			paddrs[i*ntxqs + j] = di->idi_paddr;
5347 		}
5348 	}
5349 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5350 		device_printf(ctx->ifc_dev,
5351 		    "Unable to allocate device TX queue\n");
5352 		iflib_tx_structures_free(ctx);
5353 		free(vaddrs, M_IFLIB);
5354 		free(paddrs, M_IFLIB);
5355 		goto err_rx_desc;
5356 	}
5357 	free(vaddrs, M_IFLIB);
5358 	free(paddrs, M_IFLIB);
5359 
5360 	/* RXQs */
5361 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5362 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5363 	for (i = 0; i < nrxqsets; i++) {
5364 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5365 
5366 		for (j = 0; j < nrxqs; j++, di++) {
5367 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5368 			paddrs[i*nrxqs + j] = di->idi_paddr;
5369 		}
5370 	}
5371 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5372 		device_printf(ctx->ifc_dev,
5373 		    "Unable to allocate device RX queue\n");
5374 		iflib_tx_structures_free(ctx);
5375 		free(vaddrs, M_IFLIB);
5376 		free(paddrs, M_IFLIB);
5377 		goto err_rx_desc;
5378 	}
5379 	free(vaddrs, M_IFLIB);
5380 	free(paddrs, M_IFLIB);
5381 
5382 	return (0);
5383 
5384 /* XXX handle allocation failure changes */
5385 err_rx_desc:
5386 err_tx_desc:
5387 rx_fail:
5388 	if (ctx->ifc_rxqs != NULL)
5389 		free(ctx->ifc_rxqs, M_IFLIB);
5390 	ctx->ifc_rxqs = NULL;
5391 	if (ctx->ifc_txqs != NULL)
5392 		free(ctx->ifc_txqs, M_IFLIB);
5393 	ctx->ifc_txqs = NULL;
5394 fail:
5395 	return (err);
5396 }
5397 
5398 static int
5399 iflib_tx_structures_setup(if_ctx_t ctx)
5400 {
5401 	iflib_txq_t txq = ctx->ifc_txqs;
5402 	int i;
5403 
5404 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5405 		iflib_txq_setup(txq);
5406 
5407 	return (0);
5408 }
5409 
5410 static void
5411 iflib_tx_structures_free(if_ctx_t ctx)
5412 {
5413 	iflib_txq_t txq = ctx->ifc_txqs;
5414 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5415 	int i, j;
5416 
5417 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5418 		iflib_txq_destroy(txq);
5419 		for (j = 0; j < sctx->isc_ntxqs; j++)
5420 			iflib_dma_free(&txq->ift_ifdi[j]);
5421 	}
5422 	free(ctx->ifc_txqs, M_IFLIB);
5423 	ctx->ifc_txqs = NULL;
5424 	IFDI_QUEUES_FREE(ctx);
5425 }
5426 
5427 /*********************************************************************
5428  *
5429  *  Initialize all receive rings.
5430  *
5431  **********************************************************************/
5432 static int
5433 iflib_rx_structures_setup(if_ctx_t ctx)
5434 {
5435 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5436 	int q;
5437 #if defined(INET6) || defined(INET)
5438 	int i, err;
5439 #endif
5440 
5441 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5442 #if defined(INET6) || defined(INET)
5443 		tcp_lro_free(&rxq->ifr_lc);
5444 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5445 		    TCP_LRO_ENTRIES, min(1024,
5446 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5447 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5448 			goto fail;
5449 		}
5450 		rxq->ifr_lro_enabled = TRUE;
5451 #endif
5452 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5453 	}
5454 	return (0);
5455 #if defined(INET6) || defined(INET)
5456 fail:
5457 	/*
5458 	 * Free RX software descriptors allocated so far, we will only handle
5459 	 * the rings that completed, the failing case will have
5460 	 * cleaned up for itself. 'q' failed, so its the terminus.
5461 	 */
5462 	rxq = ctx->ifc_rxqs;
5463 	for (i = 0; i < q; ++i, rxq++) {
5464 		iflib_rx_sds_free(rxq);
5465 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5466 	}
5467 	return (err);
5468 #endif
5469 }
5470 
5471 /*********************************************************************
5472  *
5473  *  Free all receive rings.
5474  *
5475  **********************************************************************/
5476 static void
5477 iflib_rx_structures_free(if_ctx_t ctx)
5478 {
5479 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5480 
5481 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5482 		iflib_rx_sds_free(rxq);
5483 	}
5484 	free(ctx->ifc_rxqs, M_IFLIB);
5485 	ctx->ifc_rxqs = NULL;
5486 }
5487 
5488 static int
5489 iflib_qset_structures_setup(if_ctx_t ctx)
5490 {
5491 	int err;
5492 
5493 	/*
5494 	 * It is expected that the caller takes care of freeing queues if this
5495 	 * fails.
5496 	 */
5497 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5498 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5499 		return (err);
5500 	}
5501 
5502 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5503 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5504 
5505 	return (err);
5506 }
5507 
5508 int
5509 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5510 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5511 {
5512 
5513 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5514 }
5515 
5516 #ifdef SMP
5517 static int
5518 find_nth(if_ctx_t ctx, int qid)
5519 {
5520 	cpuset_t cpus;
5521 	int i, cpuid, eqid, count;
5522 
5523 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5524 	count = CPU_COUNT(&cpus);
5525 	eqid = qid % count;
5526 	/* clear up to the qid'th bit */
5527 	for (i = 0; i < eqid; i++) {
5528 		cpuid = CPU_FFS(&cpus);
5529 		MPASS(cpuid != 0);
5530 		CPU_CLR(cpuid-1, &cpus);
5531 	}
5532 	cpuid = CPU_FFS(&cpus);
5533 	MPASS(cpuid != 0);
5534 	return (cpuid-1);
5535 }
5536 
5537 #ifdef SCHED_ULE
5538 extern struct cpu_group *cpu_top;              /* CPU topology */
5539 
5540 static int
5541 find_child_with_core(int cpu, struct cpu_group *grp)
5542 {
5543 	int i;
5544 
5545 	if (grp->cg_children == 0)
5546 		return -1;
5547 
5548 	MPASS(grp->cg_child);
5549 	for (i = 0; i < grp->cg_children; i++) {
5550 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5551 			return i;
5552 	}
5553 
5554 	return -1;
5555 }
5556 
5557 /*
5558  * Find the nth "close" core to the specified core
5559  * "close" is defined as the deepest level that shares
5560  * at least an L2 cache.  With threads, this will be
5561  * threads on the same core.  If the sahred cache is L3
5562  * or higher, simply returns the same core.
5563  */
5564 static int
5565 find_close_core(int cpu, int core_offset)
5566 {
5567 	struct cpu_group *grp;
5568 	int i;
5569 	int fcpu;
5570 	cpuset_t cs;
5571 
5572 	grp = cpu_top;
5573 	if (grp == NULL)
5574 		return cpu;
5575 	i = 0;
5576 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5577 		/* If the child only has one cpu, don't descend */
5578 		if (grp->cg_child[i].cg_count <= 1)
5579 			break;
5580 		grp = &grp->cg_child[i];
5581 	}
5582 
5583 	/* If they don't share at least an L2 cache, use the same CPU */
5584 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5585 		return cpu;
5586 
5587 	/* Now pick one */
5588 	CPU_COPY(&grp->cg_mask, &cs);
5589 
5590 	/* Add the selected CPU offset to core offset. */
5591 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5592 		if (fcpu - 1 == cpu)
5593 			break;
5594 		CPU_CLR(fcpu - 1, &cs);
5595 	}
5596 	MPASS(fcpu);
5597 
5598 	core_offset += i;
5599 
5600 	CPU_COPY(&grp->cg_mask, &cs);
5601 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5602 		MPASS(CPU_FFS(&cs));
5603 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5604 	}
5605 	MPASS(CPU_FFS(&cs));
5606 	return CPU_FFS(&cs) - 1;
5607 }
5608 #else
5609 static int
5610 find_close_core(int cpu, int core_offset __unused)
5611 {
5612 	return cpu;
5613 }
5614 #endif
5615 
5616 static int
5617 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5618 {
5619 	switch (type) {
5620 	case IFLIB_INTR_TX:
5621 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5622 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5623 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5624 	case IFLIB_INTR_RX:
5625 	case IFLIB_INTR_RXTX:
5626 		/* RX queues get the specified core */
5627 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5628 	default:
5629 		return -1;
5630 	}
5631 }
5632 #else
5633 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5634 #define find_close_core(cpuid, tid)	CPU_FIRST()
5635 #define find_nth(ctx, gid)		CPU_FIRST()
5636 #endif
5637 
5638 /* Just to avoid copy/paste */
5639 static inline int
5640 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5641     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5642 {
5643 	int cpuid;
5644 	int err, tid;
5645 
5646 	cpuid = find_nth(ctx, qid);
5647 	tid = get_core_offset(ctx, type, qid);
5648 	MPASS(tid >= 0);
5649 	cpuid = find_close_core(cpuid, tid);
5650 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5651 	if (err) {
5652 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5653 		return (err);
5654 	}
5655 #ifdef notyet
5656 	if (cpuid > ctx->ifc_cpuid_highest)
5657 		ctx->ifc_cpuid_highest = cpuid;
5658 #endif
5659 	return 0;
5660 }
5661 
5662 int
5663 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5664 			iflib_intr_type_t type, driver_filter_t *filter,
5665 			void *filter_arg, int qid, const char *name)
5666 {
5667 	struct grouptask *gtask;
5668 	struct taskqgroup *tqg;
5669 	iflib_filter_info_t info;
5670 	gtask_fn_t *fn;
5671 	int tqrid, err;
5672 	driver_filter_t *intr_fast;
5673 	void *q;
5674 
5675 	info = &ctx->ifc_filter_info;
5676 	tqrid = rid;
5677 
5678 	switch (type) {
5679 	/* XXX merge tx/rx for netmap? */
5680 	case IFLIB_INTR_TX:
5681 		q = &ctx->ifc_txqs[qid];
5682 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5683 		gtask = &ctx->ifc_txqs[qid].ift_task;
5684 		tqg = qgroup_if_io_tqg;
5685 		fn = _task_fn_tx;
5686 		intr_fast = iflib_fast_intr;
5687 		GROUPTASK_INIT(gtask, 0, fn, q);
5688 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5689 		break;
5690 	case IFLIB_INTR_RX:
5691 		q = &ctx->ifc_rxqs[qid];
5692 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5693 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5694 		tqg = qgroup_if_io_tqg;
5695 		fn = _task_fn_rx;
5696 		intr_fast = iflib_fast_intr;
5697 		GROUPTASK_INIT(gtask, 0, fn, q);
5698 		break;
5699 	case IFLIB_INTR_RXTX:
5700 		q = &ctx->ifc_rxqs[qid];
5701 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5702 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5703 		tqg = qgroup_if_io_tqg;
5704 		fn = _task_fn_rx;
5705 		intr_fast = iflib_fast_intr_rxtx;
5706 		GROUPTASK_INIT(gtask, 0, fn, q);
5707 		break;
5708 	case IFLIB_INTR_ADMIN:
5709 		q = ctx;
5710 		tqrid = -1;
5711 		info = &ctx->ifc_filter_info;
5712 		gtask = &ctx->ifc_admin_task;
5713 		tqg = qgroup_if_config_tqg;
5714 		fn = _task_fn_admin;
5715 		intr_fast = iflib_fast_intr_ctx;
5716 		break;
5717 	default:
5718 		panic("unknown net intr type");
5719 	}
5720 
5721 	info->ifi_filter = filter;
5722 	info->ifi_filter_arg = filter_arg;
5723 	info->ifi_task = gtask;
5724 	info->ifi_ctx = q;
5725 
5726 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5727 	if (err != 0) {
5728 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5729 		return (err);
5730 	}
5731 	if (type == IFLIB_INTR_ADMIN)
5732 		return (0);
5733 
5734 	if (tqrid != -1) {
5735 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5736 		if (err)
5737 			return (err);
5738 	} else {
5739 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5740 	}
5741 
5742 	return (0);
5743 }
5744 
5745 void
5746 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5747 {
5748 	struct grouptask *gtask;
5749 	struct taskqgroup *tqg;
5750 	gtask_fn_t *fn;
5751 	void *q;
5752 	int irq_num = -1;
5753 	int err;
5754 
5755 	switch (type) {
5756 	case IFLIB_INTR_TX:
5757 		q = &ctx->ifc_txqs[qid];
5758 		gtask = &ctx->ifc_txqs[qid].ift_task;
5759 		tqg = qgroup_if_io_tqg;
5760 		fn = _task_fn_tx;
5761 		if (irq != NULL)
5762 			irq_num = rman_get_start(irq->ii_res);
5763 		break;
5764 	case IFLIB_INTR_RX:
5765 		q = &ctx->ifc_rxqs[qid];
5766 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5767 		tqg = qgroup_if_io_tqg;
5768 		fn = _task_fn_rx;
5769 		if (irq != NULL)
5770 			irq_num = rman_get_start(irq->ii_res);
5771 		break;
5772 	case IFLIB_INTR_IOV:
5773 		q = ctx;
5774 		gtask = &ctx->ifc_vflr_task;
5775 		tqg = qgroup_if_config_tqg;
5776 		fn = _task_fn_iov;
5777 		break;
5778 	default:
5779 		panic("unknown net intr type");
5780 	}
5781 	GROUPTASK_INIT(gtask, 0, fn, q);
5782 	if (irq_num != -1) {
5783 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5784 		if (err)
5785 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5786 	}
5787 	else {
5788 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5789 	}
5790 }
5791 
5792 void
5793 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5794 {
5795 
5796 	if (irq->ii_tag)
5797 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5798 
5799 	if (irq->ii_res)
5800 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5801 		    rman_get_rid(irq->ii_res), irq->ii_res);
5802 }
5803 
5804 static int
5805 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5806 {
5807 	iflib_txq_t txq = ctx->ifc_txqs;
5808 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5809 	if_irq_t irq = &ctx->ifc_legacy_irq;
5810 	iflib_filter_info_t info;
5811 	struct grouptask *gtask;
5812 	struct taskqgroup *tqg;
5813 	gtask_fn_t *fn;
5814 	int tqrid;
5815 	void *q;
5816 	int err;
5817 
5818 	q = &ctx->ifc_rxqs[0];
5819 	info = &rxq[0].ifr_filter_info;
5820 	gtask = &rxq[0].ifr_task;
5821 	tqg = qgroup_if_io_tqg;
5822 	tqrid = irq->ii_rid = *rid;
5823 	fn = _task_fn_rx;
5824 
5825 	ctx->ifc_flags |= IFC_LEGACY;
5826 	info->ifi_filter = filter;
5827 	info->ifi_filter_arg = filter_arg;
5828 	info->ifi_task = gtask;
5829 	info->ifi_ctx = ctx;
5830 
5831 	/* We allocate a single interrupt resource */
5832 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5833 		return (err);
5834 	GROUPTASK_INIT(gtask, 0, fn, q);
5835 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5836 
5837 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5838 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5839 	return (0);
5840 }
5841 
5842 void
5843 iflib_led_create(if_ctx_t ctx)
5844 {
5845 
5846 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5847 	    device_get_nameunit(ctx->ifc_dev));
5848 }
5849 
5850 void
5851 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5852 {
5853 
5854 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5855 }
5856 
5857 void
5858 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5859 {
5860 
5861 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5862 }
5863 
5864 void
5865 iflib_admin_intr_deferred(if_ctx_t ctx)
5866 {
5867 #ifdef INVARIANTS
5868 	struct grouptask *gtask;
5869 
5870 	gtask = &ctx->ifc_admin_task;
5871 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5872 #endif
5873 
5874 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5875 }
5876 
5877 void
5878 iflib_iov_intr_deferred(if_ctx_t ctx)
5879 {
5880 
5881 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5882 }
5883 
5884 void
5885 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5886 {
5887 
5888 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5889 }
5890 
5891 void
5892 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5893 	const char *name)
5894 {
5895 
5896 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5897 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5898 }
5899 
5900 void
5901 iflib_config_gtask_deinit(struct grouptask *gtask)
5902 {
5903 
5904 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5905 }
5906 
5907 void
5908 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5909 {
5910 	if_t ifp = ctx->ifc_ifp;
5911 	iflib_txq_t txq = ctx->ifc_txqs;
5912 
5913 	if_setbaudrate(ifp, baudrate);
5914 	if (baudrate >= IF_Gbps(10)) {
5915 		STATE_LOCK(ctx);
5916 		ctx->ifc_flags |= IFC_PREFETCH;
5917 		STATE_UNLOCK(ctx);
5918 	}
5919 	/* If link down, disable watchdog */
5920 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5921 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5922 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5923 	}
5924 	ctx->ifc_link_state = link_state;
5925 	if_link_state_change(ifp, link_state);
5926 }
5927 
5928 static int
5929 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5930 {
5931 	int credits;
5932 #ifdef INVARIANTS
5933 	int credits_pre = txq->ift_cidx_processed;
5934 #endif
5935 
5936 	if (ctx->isc_txd_credits_update == NULL)
5937 		return (0);
5938 
5939 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5940 	    BUS_DMASYNC_POSTREAD);
5941 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5942 		return (0);
5943 
5944 	txq->ift_processed += credits;
5945 	txq->ift_cidx_processed += credits;
5946 
5947 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5948 	if (txq->ift_cidx_processed >= txq->ift_size)
5949 		txq->ift_cidx_processed -= txq->ift_size;
5950 	return (credits);
5951 }
5952 
5953 static int
5954 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5955 {
5956 
5957 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5958 	    budget));
5959 }
5960 
5961 void
5962 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5963 	const char *description, if_int_delay_info_t info,
5964 	int offset, int value)
5965 {
5966 	info->iidi_ctx = ctx;
5967 	info->iidi_offset = offset;
5968 	info->iidi_value = value;
5969 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5970 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5971 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5972 	    info, 0, iflib_sysctl_int_delay, "I", description);
5973 }
5974 
5975 struct sx *
5976 iflib_ctx_lock_get(if_ctx_t ctx)
5977 {
5978 
5979 	return (&ctx->ifc_ctx_sx);
5980 }
5981 
5982 static int
5983 iflib_msix_init(if_ctx_t ctx)
5984 {
5985 	device_t dev = ctx->ifc_dev;
5986 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5987 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5988 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5989 	int iflib_num_tx_queues, iflib_num_rx_queues;
5990 	int err, admincnt, bar;
5991 
5992 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5993 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5994 
5995 	if (bootverbose)
5996 		device_printf(dev, "msix_init qsets capped at %d\n",
5997 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5998 
5999 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6000 	admincnt = sctx->isc_admin_intrcnt;
6001 	/* Override by tuneable */
6002 	if (scctx->isc_disable_msix)
6003 		goto msi;
6004 
6005 	/* First try MSI-X */
6006 	if ((msgs = pci_msix_count(dev)) == 0) {
6007 		if (bootverbose)
6008 			device_printf(dev, "MSI-X not supported or disabled\n");
6009 		goto msi;
6010 	}
6011 	/*
6012 	 * bar == -1 => "trust me I know what I'm doing"
6013 	 * Some drivers are for hardware that is so shoddily
6014 	 * documented that no one knows which bars are which
6015 	 * so the developer has to map all bars. This hack
6016 	 * allows shoddy garbage to use MSI-X in this framework.
6017 	 */
6018 	if (bar != -1) {
6019 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6020 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6021 		if (ctx->ifc_msix_mem == NULL) {
6022 			device_printf(dev, "Unable to map MSI-X table\n");
6023 			goto msi;
6024 		}
6025 	}
6026 #if IFLIB_DEBUG
6027 	/* use only 1 qset in debug mode */
6028 	queuemsgs = min(msgs - admincnt, 1);
6029 #else
6030 	queuemsgs = msgs - admincnt;
6031 #endif
6032 #ifdef RSS
6033 	queues = imin(queuemsgs, rss_getnumbuckets());
6034 #else
6035 	queues = queuemsgs;
6036 #endif
6037 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6038 	if (bootverbose)
6039 		device_printf(dev,
6040 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6041 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6042 #ifdef  RSS
6043 	/* If we're doing RSS, clamp at the number of RSS buckets */
6044 	if (queues > rss_getnumbuckets())
6045 		queues = rss_getnumbuckets();
6046 #endif
6047 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6048 		rx_queues = iflib_num_rx_queues;
6049 	else
6050 		rx_queues = queues;
6051 
6052 	if (rx_queues > scctx->isc_nrxqsets)
6053 		rx_queues = scctx->isc_nrxqsets;
6054 
6055 	/*
6056 	 * We want this to be all logical CPUs by default
6057 	 */
6058 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6059 		tx_queues = iflib_num_tx_queues;
6060 	else
6061 		tx_queues = mp_ncpus;
6062 
6063 	if (tx_queues > scctx->isc_ntxqsets)
6064 		tx_queues = scctx->isc_ntxqsets;
6065 
6066 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6067 #ifdef INVARIANTS
6068 		if (tx_queues != rx_queues)
6069 			device_printf(dev,
6070 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6071 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6072 #endif
6073 		tx_queues = min(rx_queues, tx_queues);
6074 		rx_queues = min(rx_queues, tx_queues);
6075 	}
6076 
6077 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6078 	    rx_queues, tx_queues);
6079 
6080 	vectors = rx_queues + admincnt;
6081 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6082 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6083 		    vectors);
6084 		scctx->isc_vectors = vectors;
6085 		scctx->isc_nrxqsets = rx_queues;
6086 		scctx->isc_ntxqsets = tx_queues;
6087 		scctx->isc_intr = IFLIB_INTR_MSIX;
6088 
6089 		return (vectors);
6090 	} else {
6091 		device_printf(dev,
6092 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6093 		    vectors, err);
6094 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6095 		    ctx->ifc_msix_mem);
6096 		ctx->ifc_msix_mem = NULL;
6097 	}
6098 msi:
6099 	vectors = pci_msi_count(dev);
6100 	scctx->isc_nrxqsets = 1;
6101 	scctx->isc_ntxqsets = 1;
6102 	scctx->isc_vectors = vectors;
6103 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6104 		device_printf(dev,"Using an MSI interrupt\n");
6105 		scctx->isc_intr = IFLIB_INTR_MSI;
6106 	} else {
6107 		scctx->isc_vectors = 1;
6108 		device_printf(dev,"Using a Legacy interrupt\n");
6109 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6110 	}
6111 
6112 	return (vectors);
6113 }
6114 
6115 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6116 
6117 static int
6118 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6119 {
6120 	int rc;
6121 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6122 	struct sbuf *sb;
6123 	const char *ring_state = "UNKNOWN";
6124 
6125 	/* XXX needed ? */
6126 	rc = sysctl_wire_old_buffer(req, 0);
6127 	MPASS(rc == 0);
6128 	if (rc != 0)
6129 		return (rc);
6130 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6131 	MPASS(sb != NULL);
6132 	if (sb == NULL)
6133 		return (ENOMEM);
6134 	if (state[3] <= 3)
6135 		ring_state = ring_states[state[3]];
6136 
6137 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6138 		    state[0], state[1], state[2], ring_state);
6139 	rc = sbuf_finish(sb);
6140 	sbuf_delete(sb);
6141         return(rc);
6142 }
6143 
6144 enum iflib_ndesc_handler {
6145 	IFLIB_NTXD_HANDLER,
6146 	IFLIB_NRXD_HANDLER,
6147 };
6148 
6149 static int
6150 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6151 {
6152 	if_ctx_t ctx = (void *)arg1;
6153 	enum iflib_ndesc_handler type = arg2;
6154 	char buf[256] = {0};
6155 	qidx_t *ndesc;
6156 	char *p, *next;
6157 	int nqs, rc, i;
6158 
6159 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6160 
6161 	nqs = 8;
6162 	switch(type) {
6163 	case IFLIB_NTXD_HANDLER:
6164 		ndesc = ctx->ifc_sysctl_ntxds;
6165 		if (ctx->ifc_sctx)
6166 			nqs = ctx->ifc_sctx->isc_ntxqs;
6167 		break;
6168 	case IFLIB_NRXD_HANDLER:
6169 		ndesc = ctx->ifc_sysctl_nrxds;
6170 		if (ctx->ifc_sctx)
6171 			nqs = ctx->ifc_sctx->isc_nrxqs;
6172 		break;
6173 	default:
6174 			panic("unhandled type");
6175 	}
6176 	if (nqs == 0)
6177 		nqs = 8;
6178 
6179 	for (i=0; i<8; i++) {
6180 		if (i >= nqs)
6181 			break;
6182 		if (i)
6183 			strcat(buf, ",");
6184 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6185 	}
6186 
6187 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6188 	if (rc || req->newptr == NULL)
6189 		return rc;
6190 
6191 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6192 	    i++, p = strsep(&next, " ,")) {
6193 		ndesc[i] = strtoul(p, NULL, 10);
6194 	}
6195 
6196 	return(rc);
6197 }
6198 
6199 #define NAME_BUFLEN 32
6200 static void
6201 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6202 {
6203         device_t dev = iflib_get_dev(ctx);
6204 	struct sysctl_oid_list *child, *oid_list;
6205 	struct sysctl_ctx_list *ctx_list;
6206 	struct sysctl_oid *node;
6207 
6208 	ctx_list = device_get_sysctl_ctx(dev);
6209 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6210 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6211 						      CTLFLAG_RD, NULL, "IFLIB fields");
6212 	oid_list = SYSCTL_CHILDREN(node);
6213 
6214 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6215 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6216 		       "driver version");
6217 
6218 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6219 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6220 			"# of txqs to use, 0 => use default #");
6221 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6222 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6223 			"# of rxqs to use, 0 => use default #");
6224 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6225 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6226                        "permit #txq != #rxq");
6227 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6228                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6229                       "disable MSI-X (default 0)");
6230 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6231 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6232                        "set the rx budget");
6233 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6234 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6235 		       "cause tx to abdicate instead of running to completion");
6236 
6237 	/* XXX change for per-queue sizes */
6238 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6239 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6240                        mp_ndesc_handler, "A",
6241                        "list of # of tx descriptors to use, 0 = use default #");
6242 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6243 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6244                        mp_ndesc_handler, "A",
6245                        "list of # of rx descriptors to use, 0 = use default #");
6246 }
6247 
6248 static void
6249 iflib_add_device_sysctl_post(if_ctx_t ctx)
6250 {
6251 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6252 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6253         device_t dev = iflib_get_dev(ctx);
6254 	struct sysctl_oid_list *child;
6255 	struct sysctl_ctx_list *ctx_list;
6256 	iflib_fl_t fl;
6257 	iflib_txq_t txq;
6258 	iflib_rxq_t rxq;
6259 	int i, j;
6260 	char namebuf[NAME_BUFLEN];
6261 	char *qfmt;
6262 	struct sysctl_oid *queue_node, *fl_node, *node;
6263 	struct sysctl_oid_list *queue_list, *fl_list;
6264 	ctx_list = device_get_sysctl_ctx(dev);
6265 
6266 	node = ctx->ifc_sysctl_node;
6267 	child = SYSCTL_CHILDREN(node);
6268 
6269 	if (scctx->isc_ntxqsets > 100)
6270 		qfmt = "txq%03d";
6271 	else if (scctx->isc_ntxqsets > 10)
6272 		qfmt = "txq%02d";
6273 	else
6274 		qfmt = "txq%d";
6275 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6276 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6277 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6278 					     CTLFLAG_RD, NULL, "Queue Name");
6279 		queue_list = SYSCTL_CHILDREN(queue_node);
6280 #if MEMORY_LOGGING
6281 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6282 				CTLFLAG_RD,
6283 				&txq->ift_dequeued, "total mbufs freed");
6284 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6285 				CTLFLAG_RD,
6286 				&txq->ift_enqueued, "total mbufs enqueued");
6287 #endif
6288 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6289 				   CTLFLAG_RD,
6290 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6291 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6292 				   CTLFLAG_RD,
6293 				   &txq->ift_pullups, "# of times m_pullup was called");
6294 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6295 				   CTLFLAG_RD,
6296 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6297 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6298 				   CTLFLAG_RD,
6299 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6300 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6301 				   CTLFLAG_RD,
6302 				   &txq->ift_map_failed, "# of times dma map failed");
6303 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6304 				   CTLFLAG_RD,
6305 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6306 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6307 				   CTLFLAG_RD,
6308 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6309 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6310 				   CTLFLAG_RD,
6311 				   &txq->ift_pidx, 1, "Producer Index");
6312 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6313 				   CTLFLAG_RD,
6314 				   &txq->ift_cidx, 1, "Consumer Index");
6315 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6316 				   CTLFLAG_RD,
6317 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6318 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6319 				   CTLFLAG_RD,
6320 				   &txq->ift_in_use, 1, "descriptors in use");
6321 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6322 				   CTLFLAG_RD,
6323 				   &txq->ift_processed, "descriptors procesed for clean");
6324 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6325 				   CTLFLAG_RD,
6326 				   &txq->ift_cleaned, "total cleaned");
6327 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6328 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6329 				0, mp_ring_state_handler, "A", "soft ring state");
6330 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6331 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6332 				       "# of enqueues to the mp_ring for this queue");
6333 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6334 				       CTLFLAG_RD, &txq->ift_br->drops,
6335 				       "# of drops in the mp_ring for this queue");
6336 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6337 				       CTLFLAG_RD, &txq->ift_br->starts,
6338 				       "# of normal consumer starts in the mp_ring for this queue");
6339 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6340 				       CTLFLAG_RD, &txq->ift_br->stalls,
6341 					       "# of consumer stalls in the mp_ring for this queue");
6342 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6343 			       CTLFLAG_RD, &txq->ift_br->restarts,
6344 				       "# of consumer restarts in the mp_ring for this queue");
6345 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6346 				       CTLFLAG_RD, &txq->ift_br->abdications,
6347 				       "# of consumer abdications in the mp_ring for this queue");
6348 	}
6349 
6350 	if (scctx->isc_nrxqsets > 100)
6351 		qfmt = "rxq%03d";
6352 	else if (scctx->isc_nrxqsets > 10)
6353 		qfmt = "rxq%02d";
6354 	else
6355 		qfmt = "rxq%d";
6356 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6357 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6358 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6359 					     CTLFLAG_RD, NULL, "Queue Name");
6360 		queue_list = SYSCTL_CHILDREN(queue_node);
6361 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6362 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6363 				       CTLFLAG_RD,
6364 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6365 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6366 				       CTLFLAG_RD,
6367 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6368 		}
6369 
6370 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6371 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6372 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6373 						     CTLFLAG_RD, NULL, "freelist Name");
6374 			fl_list = SYSCTL_CHILDREN(fl_node);
6375 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6376 				       CTLFLAG_RD,
6377 				       &fl->ifl_pidx, 1, "Producer Index");
6378 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6379 				       CTLFLAG_RD,
6380 				       &fl->ifl_cidx, 1, "Consumer Index");
6381 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6382 				       CTLFLAG_RD,
6383 				       &fl->ifl_credits, 1, "credits available");
6384 #if MEMORY_LOGGING
6385 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6386 					CTLFLAG_RD,
6387 					&fl->ifl_m_enqueued, "mbufs allocated");
6388 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6389 					CTLFLAG_RD,
6390 					&fl->ifl_m_dequeued, "mbufs freed");
6391 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6392 					CTLFLAG_RD,
6393 					&fl->ifl_cl_enqueued, "clusters allocated");
6394 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6395 					CTLFLAG_RD,
6396 					&fl->ifl_cl_dequeued, "clusters freed");
6397 #endif
6398 
6399 		}
6400 	}
6401 
6402 }
6403 
6404 void
6405 iflib_request_reset(if_ctx_t ctx)
6406 {
6407 
6408 	STATE_LOCK(ctx);
6409 	ctx->ifc_flags |= IFC_DO_RESET;
6410 	STATE_UNLOCK(ctx);
6411 }
6412 
6413 #ifndef __NO_STRICT_ALIGNMENT
6414 static struct mbuf *
6415 iflib_fixup_rx(struct mbuf *m)
6416 {
6417 	struct mbuf *n;
6418 
6419 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6420 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6421 		m->m_data += ETHER_HDR_LEN;
6422 		n = m;
6423 	} else {
6424 		MGETHDR(n, M_NOWAIT, MT_DATA);
6425 		if (n == NULL) {
6426 			m_freem(m);
6427 			return (NULL);
6428 		}
6429 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6430 		m->m_data += ETHER_HDR_LEN;
6431 		m->m_len -= ETHER_HDR_LEN;
6432 		n->m_len = ETHER_HDR_LEN;
6433 		M_MOVE_PKTHDR(n, m);
6434 		n->m_next = m;
6435 	}
6436 	return (n);
6437 }
6438 #endif
6439 
6440 #ifdef NETDUMP
6441 static void
6442 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6443 {
6444 	if_ctx_t ctx;
6445 
6446 	ctx = if_getsoftc(ifp);
6447 	CTX_LOCK(ctx);
6448 	*nrxr = NRXQSETS(ctx);
6449 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6450 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6451 	CTX_UNLOCK(ctx);
6452 }
6453 
6454 static void
6455 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6456 {
6457 	if_ctx_t ctx;
6458 	if_softc_ctx_t scctx;
6459 	iflib_fl_t fl;
6460 	iflib_rxq_t rxq;
6461 	int i, j;
6462 
6463 	ctx = if_getsoftc(ifp);
6464 	scctx = &ctx->ifc_softc_ctx;
6465 
6466 	switch (event) {
6467 	case NETDUMP_START:
6468 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6469 			rxq = &ctx->ifc_rxqs[i];
6470 			for (j = 0; j < rxq->ifr_nfl; j++) {
6471 				fl = rxq->ifr_fl;
6472 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6473 			}
6474 		}
6475 		iflib_no_tx_batch = 1;
6476 		break;
6477 	default:
6478 		break;
6479 	}
6480 }
6481 
6482 static int
6483 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6484 {
6485 	if_ctx_t ctx;
6486 	iflib_txq_t txq;
6487 	int error;
6488 
6489 	ctx = if_getsoftc(ifp);
6490 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6491 	    IFF_DRV_RUNNING)
6492 		return (EBUSY);
6493 
6494 	txq = &ctx->ifc_txqs[0];
6495 	error = iflib_encap(txq, &m);
6496 	if (error == 0)
6497 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6498 	return (error);
6499 }
6500 
6501 static int
6502 iflib_netdump_poll(struct ifnet *ifp, int count)
6503 {
6504 	if_ctx_t ctx;
6505 	if_softc_ctx_t scctx;
6506 	iflib_txq_t txq;
6507 	int i;
6508 
6509 	ctx = if_getsoftc(ifp);
6510 	scctx = &ctx->ifc_softc_ctx;
6511 
6512 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6513 	    IFF_DRV_RUNNING)
6514 		return (EBUSY);
6515 
6516 	txq = &ctx->ifc_txqs[0];
6517 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6518 
6519 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6520 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6521 	return (0);
6522 }
6523 #endif /* NETDUMP */
6524