xref: /freebsd/sys/net/iflib.c (revision 28f4385e45a2681c14bd04b83fe1796eaefe8265)
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_desc_tag;
357 	bus_dma_tag_t		ift_tso_desc_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_desc_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_desc_tag, txq->ift_ifdi->idi_map,
926 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
927 
928 
929 	/*
930 	 * First part: process new packets to send.
931 	 * nm_i is the current index in the netmap kring,
932 	 * nic_i is the corresponding index in the NIC ring.
933 	 *
934 	 * If we have packets to send (nm_i != head)
935 	 * iterate over the netmap ring, fetch length and update
936 	 * the corresponding slot in the NIC ring. Some drivers also
937 	 * need to update the buffer's physical address in the NIC slot
938 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
939 	 *
940 	 * The netmap_reload_map() calls is especially expensive,
941 	 * even when (as in this case) the tag is 0, so do only
942 	 * when the buffer has actually changed.
943 	 *
944 	 * If possible do not set the report/intr bit on all slots,
945 	 * but only a few times per ring or when NS_REPORT is set.
946 	 *
947 	 * Finally, on 10G and faster drivers, it might be useful
948 	 * to prefetch the next slot and txr entry.
949 	 */
950 
951 	nm_i = kring->nr_hwcur;
952 	if (nm_i != head) {	/* we have new packets to send */
953 		pkt_info_zero(&pi);
954 		pi.ipi_segs = txq->ift_segs;
955 		pi.ipi_qsidx = kring->ring_id;
956 		nic_i = netmap_idx_k2n(kring, nm_i);
957 
958 		__builtin_prefetch(&ring->slot[nm_i]);
959 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
960 		if (txq->ift_sds.ifsd_map)
961 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
962 
963 		for (n = 0; nm_i != head; n++) {
964 			struct netmap_slot *slot = &ring->slot[nm_i];
965 			u_int len = slot->len;
966 			uint64_t paddr;
967 			void *addr = PNMB(na, slot, &paddr);
968 			int flags = (slot->flags & NS_REPORT ||
969 				nic_i == 0 || nic_i == report_frequency) ?
970 				IPI_TX_INTR : 0;
971 
972 			/* device-specific */
973 			pi.ipi_len = len;
974 			pi.ipi_segs[0].ds_addr = paddr;
975 			pi.ipi_segs[0].ds_len = len;
976 			pi.ipi_nsegs = 1;
977 			pi.ipi_ndescs = 0;
978 			pi.ipi_pidx = nic_i;
979 			pi.ipi_flags = flags;
980 
981 			/* Fill the slot in the NIC ring. */
982 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
983 			DBG_COUNTER_INC(tx_encap);
984 
985 			/* prefetch for next round */
986 			__builtin_prefetch(&ring->slot[nm_i + 1]);
987 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
988 			if (txq->ift_sds.ifsd_map) {
989 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
990 
991 				NM_CHECK_ADDR_LEN(na, addr, len);
992 
993 				if (slot->flags & NS_BUF_CHANGED) {
994 					/* buffer has changed, reload map */
995 					netmap_reload_map(na, txq->ift_desc_tag, 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_desc_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_desc_tag, fl->ifl_ifdi->idi_map,
1076 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1077 	}
1078 	/*
1079 	 * First part: import newly received packets.
1080 	 *
1081 	 * nm_i is the index of the next free slot in the netmap ring,
1082 	 * nic_i is the index of the next received packet in the NIC ring,
1083 	 * and they may differ in case if_init() has been called while
1084 	 * in netmap mode. For the receive ring we have
1085 	 *
1086 	 *	nic_i = rxr->next_check;
1087 	 *	nm_i = kring->nr_hwtail (previous)
1088 	 * and
1089 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1090 	 *
1091 	 * rxr->next_check is set to 0 on a ring reinit
1092 	 */
1093 	if (netmap_no_pendintr || force_update) {
1094 		int crclen = iflib_crcstrip ? 0 : 4;
1095 		int error, avail;
1096 
1097 		for (i = 0; i < rxq->ifr_nfl; i++) {
1098 			fl = &rxq->ifr_fl[i];
1099 			nic_i = fl->ifl_cidx;
1100 			nm_i = netmap_idx_n2k(kring, nic_i);
1101 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1102 			for (n = 0; avail > 0; n++, avail--) {
1103 				rxd_info_zero(&ri);
1104 				ri.iri_frags = rxq->ifr_frags;
1105 				ri.iri_qsidx = kring->ring_id;
1106 				ri.iri_ifp = ctx->ifc_ifp;
1107 				ri.iri_cidx = nic_i;
1108 
1109 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1110 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1111 				ring->slot[nm_i].flags = 0;
1112 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1113 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1114 				nm_i = nm_next(nm_i, lim);
1115 				nic_i = nm_next(nic_i, lim);
1116 			}
1117 			if (n) { /* update the state variables */
1118 				if (netmap_no_pendintr && !force_update) {
1119 					/* diagnostics */
1120 					iflib_rx_miss ++;
1121 					iflib_rx_miss_bufs += n;
1122 				}
1123 				fl->ifl_cidx = nic_i;
1124 				kring->nr_hwtail = nm_i;
1125 			}
1126 			kring->nr_kflags &= ~NKR_PENDINTR;
1127 		}
1128 	}
1129 	/*
1130 	 * Second part: skip past packets that userspace has released.
1131 	 * (kring->nr_hwcur to head excluded),
1132 	 * and make the buffers available for reception.
1133 	 * As usual nm_i is the index in the netmap ring,
1134 	 * nic_i is the index in the NIC ring, and
1135 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1136 	 */
1137 	/* XXX not sure how this will work with multiple free lists */
1138 	nm_i = kring->nr_hwcur;
1139 
1140 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1141 }
1142 
1143 static void
1144 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1145 {
1146 	struct ifnet *ifp = na->ifp;
1147 	if_ctx_t ctx = ifp->if_softc;
1148 
1149 	CTX_LOCK(ctx);
1150 	if (onoff) {
1151 		IFDI_INTR_ENABLE(ctx);
1152 	} else {
1153 		IFDI_INTR_DISABLE(ctx);
1154 	}
1155 	CTX_UNLOCK(ctx);
1156 }
1157 
1158 
1159 static int
1160 iflib_netmap_attach(if_ctx_t ctx)
1161 {
1162 	struct netmap_adapter na;
1163 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1164 
1165 	bzero(&na, sizeof(na));
1166 
1167 	na.ifp = ctx->ifc_ifp;
1168 	na.na_flags = NAF_BDG_MAYSLEEP;
1169 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1170 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1171 
1172 	na.num_tx_desc = scctx->isc_ntxd[0];
1173 	na.num_rx_desc = scctx->isc_nrxd[0];
1174 	na.nm_txsync = iflib_netmap_txsync;
1175 	na.nm_rxsync = iflib_netmap_rxsync;
1176 	na.nm_register = iflib_netmap_register;
1177 	na.nm_intr = iflib_netmap_intr;
1178 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1179 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1180 	return (netmap_attach(&na));
1181 }
1182 
1183 static void
1184 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1185 {
1186 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1187 	struct netmap_slot *slot;
1188 
1189 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1190 	if (slot == NULL)
1191 		return;
1192 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1193 
1194 		/*
1195 		 * In netmap mode, set the map for the packet buffer.
1196 		 * NOTE: Some drivers (not this one) also need to set
1197 		 * the physical buffer address in the NIC ring.
1198 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1199 		 * netmap slot index, si
1200 		 */
1201 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1202 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1203 	}
1204 }
1205 
1206 static void
1207 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1208 {
1209 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1210 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1211 	struct netmap_slot *slot;
1212 	uint32_t nm_i;
1213 
1214 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1215 	if (slot == NULL)
1216 		return;
1217 	nm_i = netmap_idx_n2k(kring, 0);
1218 	netmap_fl_refill(rxq, kring, nm_i, true);
1219 }
1220 
1221 static void
1222 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1223 {
1224 	struct netmap_kring *kring;
1225 
1226 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1227 
1228 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1229 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1230 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1231 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1232 			if (hz < 2000)
1233 				*reset_on = 1;
1234 			else
1235 				*reset_on = hz / 1000;
1236 		}
1237 	}
1238 }
1239 
1240 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1241 
1242 #else
1243 #define iflib_netmap_txq_init(ctx, txq)
1244 #define iflib_netmap_rxq_init(ctx, rxq)
1245 #define iflib_netmap_detach(ifp)
1246 
1247 #define iflib_netmap_attach(ctx) (0)
1248 #define netmap_rx_irq(ifp, qid, budget) (0)
1249 #define netmap_tx_irq(ifp, qid) do {} while (0)
1250 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1251 
1252 #endif
1253 
1254 #if defined(__i386__) || defined(__amd64__)
1255 static __inline void
1256 prefetch(void *x)
1257 {
1258 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1259 }
1260 static __inline void
1261 prefetch2cachelines(void *x)
1262 {
1263 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1264 #if (CACHE_LINE_SIZE < 128)
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1266 #endif
1267 }
1268 #else
1269 #define prefetch(x)
1270 #define prefetch2cachelines(x)
1271 #endif
1272 
1273 static void
1274 iflib_gen_mac(if_ctx_t ctx)
1275 {
1276 	struct thread *td;
1277 	MD5_CTX mdctx;
1278 	char uuid[HOSTUUIDLEN+1];
1279 	char buf[HOSTUUIDLEN+16];
1280 	uint8_t *mac;
1281 	unsigned char digest[16];
1282 
1283 	td = curthread;
1284 	mac = ctx->ifc_mac;
1285 	uuid[HOSTUUIDLEN] = 0;
1286 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1287 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1288 	/*
1289 	 * Generate a pseudo-random, deterministic MAC
1290 	 * address based on the UUID and unit number.
1291 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1292 	 */
1293 	MD5Init(&mdctx);
1294 	MD5Update(&mdctx, buf, strlen(buf));
1295 	MD5Final(digest, &mdctx);
1296 
1297 	mac[0] = 0x58;
1298 	mac[1] = 0x9C;
1299 	mac[2] = 0xFC;
1300 	mac[3] = digest[0];
1301 	mac[4] = digest[1];
1302 	mac[5] = digest[2];
1303 }
1304 
1305 static void
1306 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1307 {
1308 	iflib_fl_t fl;
1309 
1310 	fl = &rxq->ifr_fl[flid];
1311 	iru->iru_paddrs = fl->ifl_bus_addrs;
1312 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1313 	iru->iru_idxs = fl->ifl_rxd_idxs;
1314 	iru->iru_qsidx = rxq->ifr_id;
1315 	iru->iru_buf_size = fl->ifl_buf_size;
1316 	iru->iru_flidx = fl->ifl_id;
1317 }
1318 
1319 static void
1320 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1321 {
1322 	if (err)
1323 		return;
1324 	*(bus_addr_t *) arg = segs[0].ds_addr;
1325 }
1326 
1327 int
1328 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1329 {
1330 	int err;
1331 	device_t dev = ctx->ifc_dev;
1332 
1333 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1334 				align, 0,		/* alignment, bounds */
1335 				BUS_SPACE_MAXADDR,	/* lowaddr */
1336 				BUS_SPACE_MAXADDR,	/* highaddr */
1337 				NULL, NULL,		/* filter, filterarg */
1338 				size,			/* maxsize */
1339 				1,			/* nsegments */
1340 				size,			/* maxsegsize */
1341 				BUS_DMA_ALLOCNOW,	/* flags */
1342 				NULL,			/* lockfunc */
1343 				NULL,			/* lockarg */
1344 				&dma->idi_tag);
1345 	if (err) {
1346 		device_printf(dev,
1347 		    "%s: bus_dma_tag_create failed: %d\n",
1348 		    __func__, err);
1349 		goto fail_0;
1350 	}
1351 
1352 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1353 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1354 	if (err) {
1355 		device_printf(dev,
1356 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1357 		    __func__, (uintmax_t)size, err);
1358 		goto fail_1;
1359 	}
1360 
1361 	dma->idi_paddr = IF_BAD_DMA;
1362 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1363 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1364 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1365 		device_printf(dev,
1366 		    "%s: bus_dmamap_load failed: %d\n",
1367 		    __func__, err);
1368 		goto fail_2;
1369 	}
1370 
1371 	dma->idi_size = size;
1372 	return (0);
1373 
1374 fail_2:
1375 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1376 fail_1:
1377 	bus_dma_tag_destroy(dma->idi_tag);
1378 fail_0:
1379 	dma->idi_tag = NULL;
1380 
1381 	return (err);
1382 }
1383 
1384 int
1385 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1386 {
1387 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1388 
1389 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1390 
1391 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1392 }
1393 
1394 int
1395 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1396 {
1397 	int i, err;
1398 	iflib_dma_info_t *dmaiter;
1399 
1400 	dmaiter = dmalist;
1401 	for (i = 0; i < count; i++, dmaiter++) {
1402 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1403 			break;
1404 	}
1405 	if (err)
1406 		iflib_dma_free_multi(dmalist, i);
1407 	return (err);
1408 }
1409 
1410 void
1411 iflib_dma_free(iflib_dma_info_t dma)
1412 {
1413 	if (dma->idi_tag == NULL)
1414 		return;
1415 	if (dma->idi_paddr != IF_BAD_DMA) {
1416 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1417 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1418 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1419 		dma->idi_paddr = IF_BAD_DMA;
1420 	}
1421 	if (dma->idi_vaddr != NULL) {
1422 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1423 		dma->idi_vaddr = NULL;
1424 	}
1425 	bus_dma_tag_destroy(dma->idi_tag);
1426 	dma->idi_tag = NULL;
1427 }
1428 
1429 void
1430 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1431 {
1432 	int i;
1433 	iflib_dma_info_t *dmaiter = dmalist;
1434 
1435 	for (i = 0; i < count; i++, dmaiter++)
1436 		iflib_dma_free(*dmaiter);
1437 }
1438 
1439 #ifdef EARLY_AP_STARTUP
1440 static const int iflib_started = 1;
1441 #else
1442 /*
1443  * We used to abuse the smp_started flag to decide if the queues have been
1444  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1445  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1446  * is set.  Run a SYSINIT() strictly after that to just set a usable
1447  * completion flag.
1448  */
1449 
1450 static int iflib_started;
1451 
1452 static void
1453 iflib_record_started(void *arg)
1454 {
1455 	iflib_started = 1;
1456 }
1457 
1458 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1459 	iflib_record_started, NULL);
1460 #endif
1461 
1462 static int
1463 iflib_fast_intr(void *arg)
1464 {
1465 	iflib_filter_info_t info = arg;
1466 	struct grouptask *gtask = info->ifi_task;
1467 	if (!iflib_started)
1468 		return (FILTER_HANDLED);
1469 
1470 	DBG_COUNTER_INC(fast_intrs);
1471 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1472 		return (FILTER_HANDLED);
1473 
1474 	GROUPTASK_ENQUEUE(gtask);
1475 	return (FILTER_HANDLED);
1476 }
1477 
1478 static int
1479 iflib_fast_intr_rxtx(void *arg)
1480 {
1481 	iflib_filter_info_t info = arg;
1482 	struct grouptask *gtask = info->ifi_task;
1483 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1484 	if_ctx_t ctx = NULL;;
1485 	int i, cidx;
1486 
1487 	if (!iflib_started)
1488 		return (FILTER_HANDLED);
1489 
1490 	DBG_COUNTER_INC(fast_intrs);
1491 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1492 		return (FILTER_HANDLED);
1493 
1494 	MPASS(rxq->ifr_ntxqirq);
1495 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1496 		qidx_t txqid = rxq->ifr_txqid[i];
1497 
1498 		ctx = rxq->ifr_ctx;
1499 
1500 		bus_dmamap_sync(rxq->ifr_ifdi->idi_tag, rxq->ifr_ifdi->idi_map,
1501 		    BUS_DMASYNC_POSTREAD);
1502 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1503 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1504 			continue;
1505 		}
1506 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1507 	}
1508 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1509 		cidx = rxq->ifr_cq_cidx;
1510 	else
1511 		cidx = rxq->ifr_fl[0].ifl_cidx;
1512 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1513 		GROUPTASK_ENQUEUE(gtask);
1514 	else {
1515 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1516 		DBG_COUNTER_INC(rx_intr_enables);
1517 	}
1518 	return (FILTER_HANDLED);
1519 }
1520 
1521 
1522 static int
1523 iflib_fast_intr_ctx(void *arg)
1524 {
1525 	iflib_filter_info_t info = arg;
1526 	struct grouptask *gtask = info->ifi_task;
1527 
1528 	if (!iflib_started)
1529 		return (FILTER_HANDLED);
1530 
1531 	DBG_COUNTER_INC(fast_intrs);
1532 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1533 		return (FILTER_HANDLED);
1534 
1535 	GROUPTASK_ENQUEUE(gtask);
1536 	return (FILTER_HANDLED);
1537 }
1538 
1539 static int
1540 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1541 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1542 		 const char *name)
1543 {
1544 	int rc, flags;
1545 	struct resource *res;
1546 	void *tag = NULL;
1547 	device_t dev = ctx->ifc_dev;
1548 
1549 	flags = RF_ACTIVE;
1550 	if (ctx->ifc_flags & IFC_LEGACY)
1551 		flags |= RF_SHAREABLE;
1552 	MPASS(rid < 512);
1553 	irq->ii_rid = rid;
1554 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1555 	if (res == NULL) {
1556 		device_printf(dev,
1557 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1558 		return (ENOMEM);
1559 	}
1560 	irq->ii_res = res;
1561 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1562 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1563 						filter, handler, arg, &tag);
1564 	if (rc != 0) {
1565 		device_printf(dev,
1566 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1567 					  rid, name ? name : "unknown", rc);
1568 		return (rc);
1569 	} else if (name)
1570 		bus_describe_intr(dev, res, tag, "%s", name);
1571 
1572 	irq->ii_tag = tag;
1573 	return (0);
1574 }
1575 
1576 
1577 /*********************************************************************
1578  *
1579  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1580  *  the information needed to transmit a packet on the wire. This is
1581  *  called only once at attach, setup is done every reset.
1582  *
1583  **********************************************************************/
1584 
1585 static int
1586 iflib_txsd_alloc(iflib_txq_t txq)
1587 {
1588 	if_ctx_t ctx = txq->ift_ctx;
1589 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1590 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1591 	device_t dev = ctx->ifc_dev;
1592 	bus_size_t tsomaxsize;
1593 	int err, nsegments, ntsosegments;
1594 	bool tso;
1595 
1596 	nsegments = scctx->isc_tx_nsegments;
1597 	ntsosegments = scctx->isc_tx_tso_segments_max;
1598 	tsomaxsize = scctx->isc_tx_tso_size_max;
1599 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1600 		tsomaxsize += sizeof(struct ether_vlan_header);
1601 	MPASS(scctx->isc_ntxd[0] > 0);
1602 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1603 	MPASS(nsegments > 0);
1604 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1605 		MPASS(ntsosegments > 0);
1606 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1607 	}
1608 
1609 	/*
1610 	 * Setup DMA descriptor areas.
1611 	 */
1612 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1613 			       1, 0,			/* alignment, bounds */
1614 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1615 			       BUS_SPACE_MAXADDR,	/* highaddr */
1616 			       NULL, NULL,		/* filter, filterarg */
1617 			       sctx->isc_tx_maxsize,		/* maxsize */
1618 			       nsegments,	/* nsegments */
1619 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1620 			       0,			/* flags */
1621 			       NULL,			/* lockfunc */
1622 			       NULL,			/* lockfuncarg */
1623 			       &txq->ift_desc_tag))) {
1624 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1625 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1626 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1627 		goto fail;
1628 	}
1629 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1630 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1631 			       1, 0,			/* alignment, bounds */
1632 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1633 			       BUS_SPACE_MAXADDR,	/* highaddr */
1634 			       NULL, NULL,		/* filter, filterarg */
1635 			       tsomaxsize,		/* maxsize */
1636 			       ntsosegments,	/* nsegments */
1637 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1638 			       0,			/* flags */
1639 			       NULL,			/* lockfunc */
1640 			       NULL,			/* lockfuncarg */
1641 			       &txq->ift_tso_desc_tag))) {
1642 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1643 		goto fail;
1644 	}
1645 	if (!(txq->ift_sds.ifsd_m =
1646 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1647 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1648 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1649 		err = ENOMEM;
1650 		goto fail;
1651 	}
1652 
1653         /* Create the descriptor buffer dma maps */
1654 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1655 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1656 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1657 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1658 		err = ENOMEM;
1659 		goto fail;
1660 	}
1661 
1662 	if (tso && (txq->ift_sds.ifsd_tso_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, "Unable to allocate TSO tx_buffer "
1666 		    "map memory\n");
1667 		err = ENOMEM;
1668 		goto fail;
1669 	}
1670 
1671 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1672 		err = bus_dmamap_create(txq->ift_desc_tag, 0,
1673 		    &txq->ift_sds.ifsd_map[i]);
1674 		if (err != 0) {
1675 			device_printf(dev, "Unable to create TX DMA map\n");
1676 			goto fail;
1677 		}
1678 		if (!tso)
1679 			continue;
1680 		err = bus_dmamap_create(txq->ift_tso_desc_tag, 0,
1681 		    &txq->ift_sds.ifsd_tso_map[i]);
1682 		if (err != 0) {
1683 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1684 			goto fail;
1685 		}
1686 	}
1687 	return (0);
1688 fail:
1689 	/* We free all, it handles case where we are in the middle */
1690 	iflib_tx_structures_free(ctx);
1691 	return (err);
1692 }
1693 
1694 static void
1695 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1696 {
1697 	bus_dmamap_t map;
1698 
1699 	map = NULL;
1700 	if (txq->ift_sds.ifsd_map != NULL)
1701 		map = txq->ift_sds.ifsd_map[i];
1702 	if (map != NULL) {
1703 		bus_dmamap_sync(txq->ift_desc_tag, map, BUS_DMASYNC_POSTWRITE);
1704 		bus_dmamap_unload(txq->ift_desc_tag, map);
1705 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1706 		txq->ift_sds.ifsd_map[i] = NULL;
1707 	}
1708 
1709 	map = NULL;
1710 	if (txq->ift_sds.ifsd_tso_map != NULL)
1711 		map = txq->ift_sds.ifsd_tso_map[i];
1712 	if (map != NULL) {
1713 		bus_dmamap_sync(txq->ift_tso_desc_tag, map,
1714 		    BUS_DMASYNC_POSTWRITE);
1715 		bus_dmamap_unload(txq->ift_tso_desc_tag, map);
1716 		bus_dmamap_destroy(txq->ift_tso_desc_tag, map);
1717 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1718 	}
1719 }
1720 
1721 static void
1722 iflib_txq_destroy(iflib_txq_t txq)
1723 {
1724 	if_ctx_t ctx = txq->ift_ctx;
1725 
1726 	for (int i = 0; i < txq->ift_size; i++)
1727 		iflib_txsd_destroy(ctx, txq, i);
1728 	if (txq->ift_sds.ifsd_map != NULL) {
1729 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1730 		txq->ift_sds.ifsd_map = NULL;
1731 	}
1732 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1733 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1734 		txq->ift_sds.ifsd_tso_map = NULL;
1735 	}
1736 	if (txq->ift_sds.ifsd_m != NULL) {
1737 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1738 		txq->ift_sds.ifsd_m = NULL;
1739 	}
1740 	if (txq->ift_desc_tag != NULL) {
1741 		bus_dma_tag_destroy(txq->ift_desc_tag);
1742 		txq->ift_desc_tag = NULL;
1743 	}
1744 	if (txq->ift_tso_desc_tag != NULL) {
1745 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1746 		txq->ift_tso_desc_tag = NULL;
1747 	}
1748 }
1749 
1750 static void
1751 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1752 {
1753 	struct mbuf **mp;
1754 
1755 	mp = &txq->ift_sds.ifsd_m[i];
1756 	if (*mp == NULL)
1757 		return;
1758 
1759 	if (txq->ift_sds.ifsd_map != NULL) {
1760 		bus_dmamap_sync(txq->ift_desc_tag,
1761 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1762 		bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i]);
1763 	}
1764 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1765 		bus_dmamap_sync(txq->ift_tso_desc_tag,
1766 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1767 		bus_dmamap_unload(txq->ift_tso_desc_tag,
1768 		    txq->ift_sds.ifsd_tso_map[i]);
1769 	}
1770 	m_free(*mp);
1771 	DBG_COUNTER_INC(tx_frees);
1772 	*mp = NULL;
1773 }
1774 
1775 static int
1776 iflib_txq_setup(iflib_txq_t txq)
1777 {
1778 	if_ctx_t ctx = txq->ift_ctx;
1779 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1780 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1781 	iflib_dma_info_t di;
1782 	int i;
1783 
1784 	/* Set number of descriptors available */
1785 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1786 	/* XXX make configurable */
1787 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1788 
1789 	/* Reset indices */
1790 	txq->ift_cidx_processed = 0;
1791 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1792 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1793 
1794 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1795 		bzero((void *)di->idi_vaddr, di->idi_size);
1796 
1797 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1798 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1799 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1800 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1801 	return (0);
1802 }
1803 
1804 /*********************************************************************
1805  *
1806  *  Allocate memory for rx_buffer structures. Since we use one
1807  *  rx_buffer per received packet, the maximum number of rx_buffer's
1808  *  that we'll need is equal to the number of receive descriptors
1809  *  that we've allocated.
1810  *
1811  **********************************************************************/
1812 static int
1813 iflib_rxsd_alloc(iflib_rxq_t rxq)
1814 {
1815 	if_ctx_t ctx = rxq->ifr_ctx;
1816 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1817 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1818 	device_t dev = ctx->ifc_dev;
1819 	iflib_fl_t fl;
1820 	int			err;
1821 
1822 	MPASS(scctx->isc_nrxd[0] > 0);
1823 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1824 
1825 	fl = rxq->ifr_fl;
1826 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1827 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1828 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1829 					 1, 0,			/* alignment, bounds */
1830 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1831 					 BUS_SPACE_MAXADDR,	/* highaddr */
1832 					 NULL, NULL,		/* filter, filterarg */
1833 					 sctx->isc_rx_maxsize,	/* maxsize */
1834 					 sctx->isc_rx_nsegments,	/* nsegments */
1835 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1836 					 0,			/* flags */
1837 					 NULL,			/* lockfunc */
1838 					 NULL,			/* lockarg */
1839 					 &fl->ifl_desc_tag);
1840 		if (err) {
1841 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1842 				__func__, err);
1843 			goto fail;
1844 		}
1845 		if (!(fl->ifl_sds.ifsd_m =
1846 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1847 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1848 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1849 			err = ENOMEM;
1850 			goto fail;
1851 		}
1852 		if (!(fl->ifl_sds.ifsd_cl =
1853 		      (caddr_t *) malloc(sizeof(caddr_t) *
1854 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1855 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1856 			err = ENOMEM;
1857 			goto fail;
1858 		}
1859 
1860 		if (!(fl->ifl_sds.ifsd_ba =
1861 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1862 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1863 			device_printf(dev, "Unable to allocate rx bus addr memory\n");
1864 			err = ENOMEM;
1865 			goto fail;
1866 		}
1867 
1868 		/* Create the descriptor buffer dma maps */
1869 		if (!(fl->ifl_sds.ifsd_map =
1870 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1871 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1872 			err = ENOMEM;
1873 			goto fail;
1874 		}
1875 
1876 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1877 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1878 			if (err != 0) {
1879 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1880 				goto fail;
1881 			}
1882 		}
1883 	}
1884 	return (0);
1885 
1886 fail:
1887 	iflib_rx_structures_free(ctx);
1888 	return (err);
1889 }
1890 
1891 
1892 /*
1893  * Internal service routines
1894  */
1895 
1896 struct rxq_refill_cb_arg {
1897 	int               error;
1898 	bus_dma_segment_t seg;
1899 	int               nseg;
1900 };
1901 
1902 static void
1903 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1904 {
1905 	struct rxq_refill_cb_arg *cb_arg = arg;
1906 
1907 	cb_arg->error = error;
1908 	cb_arg->seg = segs[0];
1909 	cb_arg->nseg = nseg;
1910 }
1911 
1912 /**
1913  *	rxq_refill - refill an rxq  free-buffer list
1914  *	@ctx: the iflib context
1915  *	@rxq: the free-list to refill
1916  *	@n: the number of new buffers to allocate
1917  *
1918  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1919  *	The caller must assure that @n does not exceed the queue's capacity.
1920  */
1921 static void
1922 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1923 {
1924 	struct mbuf *m;
1925 	int idx, frag_idx = fl->ifl_fragidx;
1926         int pidx = fl->ifl_pidx;
1927 	caddr_t cl, *sd_cl;
1928 	struct mbuf **sd_m;
1929 	struct if_rxd_update iru;
1930 	struct rxq_refill_cb_arg cb_arg;
1931 	bus_dmamap_t *sd_map;
1932 	int n, i = 0;
1933 	bus_addr_t bus_addr, *sd_ba;
1934 	int err;
1935 	qidx_t credits;
1936 
1937 	sd_m = fl->ifl_sds.ifsd_m;
1938 	sd_map = fl->ifl_sds.ifsd_map;
1939 	sd_cl = fl->ifl_sds.ifsd_cl;
1940 	sd_ba = fl->ifl_sds.ifsd_ba;
1941 	idx = pidx;
1942 	credits = fl->ifl_credits;
1943 
1944 	n  = count;
1945 	MPASS(n > 0);
1946 	MPASS(credits + n <= fl->ifl_size);
1947 
1948 	if (pidx < fl->ifl_cidx)
1949 		MPASS(pidx + n <= fl->ifl_cidx);
1950 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1951 		MPASS(fl->ifl_gen == 0);
1952 	if (pidx > fl->ifl_cidx)
1953 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1954 
1955 	DBG_COUNTER_INC(fl_refills);
1956 	if (n > 8)
1957 		DBG_COUNTER_INC(fl_refills_large);
1958 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1959 	while (n--) {
1960 		/*
1961 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1962 		 * initialized after rx.
1963 		 *
1964 		 * If the cluster is still set then we know a minimum sized packet was received
1965 		 */
1966 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1967 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1968                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1969 		if ((cl = sd_cl[frag_idx]) == NULL) {
1970 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1971 				break;
1972 
1973 			cb_arg.error = 0;
1974 			MPASS(sd_map != NULL);
1975 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1976 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1977 			    BUS_DMA_NOWAIT);
1978 			if (err != 0 || cb_arg.error) {
1979 				/*
1980 				 * !zone_pack ?
1981 				 */
1982 				if (fl->ifl_zone == zone_pack)
1983 					uma_zfree(fl->ifl_zone, cl);
1984 				break;
1985 			}
1986 
1987 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1988 			    BUS_DMASYNC_PREREAD);
1989 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1990 			sd_cl[frag_idx] = cl;
1991 #if MEMORY_LOGGING
1992 			fl->ifl_cl_enqueued++;
1993 #endif
1994 		} else {
1995 			bus_addr = sd_ba[frag_idx];
1996 		}
1997 
1998 		bit_set(fl->ifl_rx_bitmap, frag_idx);
1999 		MPASS(sd_m[frag_idx] == NULL);
2000 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2001 			break;
2002 		}
2003 		sd_m[frag_idx] = m;
2004 #if MEMORY_LOGGING
2005 		fl->ifl_m_enqueued++;
2006 #endif
2007 
2008 		DBG_COUNTER_INC(rx_allocs);
2009 		fl->ifl_rxd_idxs[i] = frag_idx;
2010 		fl->ifl_bus_addrs[i] = bus_addr;
2011 		fl->ifl_vm_addrs[i] = cl;
2012 		credits++;
2013 		i++;
2014 		MPASS(credits <= fl->ifl_size);
2015 		if (++idx == fl->ifl_size) {
2016 			fl->ifl_gen = 1;
2017 			idx = 0;
2018 		}
2019 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2020 			iru.iru_pidx = pidx;
2021 			iru.iru_count = i;
2022 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2023 			i = 0;
2024 			pidx = idx;
2025 			fl->ifl_pidx = idx;
2026 			fl->ifl_credits = credits;
2027 		}
2028 
2029 	}
2030 
2031 	if (i) {
2032 		iru.iru_pidx = pidx;
2033 		iru.iru_count = i;
2034 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2035 		fl->ifl_pidx = idx;
2036 		fl->ifl_credits = credits;
2037 	}
2038 	DBG_COUNTER_INC(rxd_flush);
2039 	if (fl->ifl_pidx == 0)
2040 		pidx = fl->ifl_size - 1;
2041 	else
2042 		pidx = fl->ifl_pidx - 1;
2043 
2044 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2045 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2046 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2047 	fl->ifl_fragidx = frag_idx;
2048 }
2049 
2050 static __inline void
2051 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2052 {
2053 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2054 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2055 #ifdef INVARIANTS
2056 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2057 #endif
2058 
2059 	MPASS(fl->ifl_credits <= fl->ifl_size);
2060 	MPASS(reclaimable == delta);
2061 
2062 	if (reclaimable > 0)
2063 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2064 }
2065 
2066 uint8_t
2067 iflib_in_detach(if_ctx_t ctx)
2068 {
2069 	bool in_detach;
2070 	STATE_LOCK(ctx);
2071 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2072 	STATE_UNLOCK(ctx);
2073 	return (in_detach);
2074 }
2075 
2076 static void
2077 iflib_fl_bufs_free(iflib_fl_t fl)
2078 {
2079 	iflib_dma_info_t idi = fl->ifl_ifdi;
2080 	bus_dmamap_t sd_map;
2081 	uint32_t i;
2082 
2083 	for (i = 0; i < fl->ifl_size; i++) {
2084 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2085 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2086 
2087 		if (*sd_cl != NULL) {
2088 			sd_map = fl->ifl_sds.ifsd_map[i];
2089 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map,
2090 			    BUS_DMASYNC_POSTREAD);
2091 			bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2092 			if (*sd_cl != NULL)
2093 				uma_zfree(fl->ifl_zone, *sd_cl);
2094 			// XXX: Should this get moved out?
2095 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2096 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2097 			if (*sd_m != NULL) {
2098 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2099 				uma_zfree(zone_mbuf, *sd_m);
2100 			}
2101 		} else {
2102 			MPASS(*sd_cl == NULL);
2103 			MPASS(*sd_m == NULL);
2104 		}
2105 #if MEMORY_LOGGING
2106 		fl->ifl_m_dequeued++;
2107 		fl->ifl_cl_dequeued++;
2108 #endif
2109 		*sd_cl = NULL;
2110 		*sd_m = NULL;
2111 	}
2112 #ifdef INVARIANTS
2113 	for (i = 0; i < fl->ifl_size; i++) {
2114 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2115 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2116 	}
2117 #endif
2118 	/*
2119 	 * Reset free list values
2120 	 */
2121 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2122 	bzero(idi->idi_vaddr, idi->idi_size);
2123 }
2124 
2125 /*********************************************************************
2126  *
2127  *  Initialize a receive ring and its buffers.
2128  *
2129  **********************************************************************/
2130 static int
2131 iflib_fl_setup(iflib_fl_t fl)
2132 {
2133 	iflib_rxq_t rxq = fl->ifl_rxq;
2134 	if_ctx_t ctx = rxq->ifr_ctx;
2135 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2136 
2137 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2138 	/*
2139 	** Free current RX buffer structs and their mbufs
2140 	*/
2141 	iflib_fl_bufs_free(fl);
2142 	/* Now replenish the mbufs */
2143 	MPASS(fl->ifl_credits == 0);
2144 	/*
2145 	 * XXX don't set the max_frame_size to larger
2146 	 * than the hardware can handle
2147 	 */
2148 	if (sctx->isc_max_frame_size <= 2048)
2149 		fl->ifl_buf_size = MCLBYTES;
2150 #ifndef CONTIGMALLOC_WORKS
2151 	else
2152 		fl->ifl_buf_size = MJUMPAGESIZE;
2153 #else
2154 	else if (sctx->isc_max_frame_size <= 4096)
2155 		fl->ifl_buf_size = MJUMPAGESIZE;
2156 	else if (sctx->isc_max_frame_size <= 9216)
2157 		fl->ifl_buf_size = MJUM9BYTES;
2158 	else
2159 		fl->ifl_buf_size = MJUM16BYTES;
2160 #endif
2161 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2162 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2163 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2164 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2165 
2166 
2167 	/* avoid pre-allocating zillions of clusters to an idle card
2168 	 * potentially speeding up attach
2169 	 */
2170 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2171 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2172 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2173 		return (ENOBUFS);
2174 	/*
2175 	 * handle failure
2176 	 */
2177 	MPASS(rxq != NULL);
2178 	MPASS(fl->ifl_ifdi != NULL);
2179 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2180 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2181 	return (0);
2182 }
2183 
2184 /*********************************************************************
2185  *
2186  *  Free receive ring data structures
2187  *
2188  **********************************************************************/
2189 static void
2190 iflib_rx_sds_free(iflib_rxq_t rxq)
2191 {
2192 	iflib_fl_t fl;
2193 	int i, j;
2194 
2195 	if (rxq->ifr_fl != NULL) {
2196 		for (i = 0; i < rxq->ifr_nfl; i++) {
2197 			fl = &rxq->ifr_fl[i];
2198 			if (fl->ifl_desc_tag != NULL) {
2199 				if (fl->ifl_sds.ifsd_map != NULL) {
2200 					for (j = 0; j < fl->ifl_size; j++) {
2201 						if (fl->ifl_sds.ifsd_map[j] ==
2202 						    NULL)
2203 							continue;
2204 						bus_dmamap_sync(
2205 						    fl->ifl_desc_tag,
2206 						    fl->ifl_sds.ifsd_map[j],
2207 						    BUS_DMASYNC_POSTREAD);
2208 						bus_dmamap_unload(
2209 						    fl->ifl_desc_tag,
2210 						    fl->ifl_sds.ifsd_map[j]);
2211 					}
2212 				}
2213 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2214 				fl->ifl_desc_tag = NULL;
2215 			}
2216 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2217 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2219 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2220 			fl->ifl_sds.ifsd_m = NULL;
2221 			fl->ifl_sds.ifsd_cl = NULL;
2222 			fl->ifl_sds.ifsd_ba = NULL;
2223 			fl->ifl_sds.ifsd_map = NULL;
2224 		}
2225 		free(rxq->ifr_fl, M_IFLIB);
2226 		rxq->ifr_fl = NULL;
2227 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2228 	}
2229 }
2230 
2231 /*
2232  * MI independent logic
2233  *
2234  */
2235 static void
2236 iflib_timer(void *arg)
2237 {
2238 	iflib_txq_t txq = arg;
2239 	if_ctx_t ctx = txq->ift_ctx;
2240 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2241 	uint64_t this_tick = ticks;
2242 	uint32_t reset_on = hz / 2;
2243 
2244 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2245 		return;
2246 	/*
2247 	** Check on the state of the TX queue(s), this
2248 	** can be done without the lock because its RO
2249 	** and the HUNG state will be static if set.
2250 	*/
2251 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2252 		txq->ift_last_timer_tick = this_tick;
2253 		IFDI_TIMER(ctx, txq->ift_id);
2254 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2255 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2256 		     (sctx->isc_pause_frames == 0)))
2257 			goto hung;
2258 
2259 		if (ifmp_ring_is_stalled(txq->ift_br))
2260 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2261 		txq->ift_cleaned_prev = txq->ift_cleaned;
2262 	}
2263 #ifdef DEV_NETMAP
2264 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2265 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2266 #endif
2267 	/* handle any laggards */
2268 	if (txq->ift_db_pending)
2269 		GROUPTASK_ENQUEUE(&txq->ift_task);
2270 
2271 	sctx->isc_pause_frames = 0;
2272 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2273 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2274 	return;
2275  hung:
2276 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2277 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2278 	STATE_LOCK(ctx);
2279 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2280 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2281 	iflib_admin_intr_deferred(ctx);
2282 	STATE_UNLOCK(ctx);
2283 }
2284 
2285 static void
2286 iflib_init_locked(if_ctx_t ctx)
2287 {
2288 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2289 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2290 	if_t ifp = ctx->ifc_ifp;
2291 	iflib_fl_t fl;
2292 	iflib_txq_t txq;
2293 	iflib_rxq_t rxq;
2294 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2295 
2296 
2297 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2298 	IFDI_INTR_DISABLE(ctx);
2299 
2300 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2301 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2302 	/* Set hardware offload abilities */
2303 	if_clearhwassist(ifp);
2304 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2305 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2306 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2307 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2308 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2309 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2310 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2311 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2312 
2313 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2314 		CALLOUT_LOCK(txq);
2315 		callout_stop(&txq->ift_timer);
2316 		CALLOUT_UNLOCK(txq);
2317 		iflib_netmap_txq_init(ctx, txq);
2318 	}
2319 #ifdef INVARIANTS
2320 	i = if_getdrvflags(ifp);
2321 #endif
2322 	IFDI_INIT(ctx);
2323 	MPASS(if_getdrvflags(ifp) == i);
2324 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2325 		/* XXX this should really be done on a per-queue basis */
2326 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2327 			MPASS(rxq->ifr_id == i);
2328 			iflib_netmap_rxq_init(ctx, rxq);
2329 			continue;
2330 		}
2331 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2332 			if (iflib_fl_setup(fl)) {
2333 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2334 				goto done;
2335 			}
2336 		}
2337 	}
2338 done:
2339 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2340 	IFDI_INTR_ENABLE(ctx);
2341 	txq = ctx->ifc_txqs;
2342 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2343 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2344 			txq->ift_timer.c_cpu);
2345 }
2346 
2347 static int
2348 iflib_media_change(if_t ifp)
2349 {
2350 	if_ctx_t ctx = if_getsoftc(ifp);
2351 	int err;
2352 
2353 	CTX_LOCK(ctx);
2354 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2355 		iflib_init_locked(ctx);
2356 	CTX_UNLOCK(ctx);
2357 	return (err);
2358 }
2359 
2360 static void
2361 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2362 {
2363 	if_ctx_t ctx = if_getsoftc(ifp);
2364 
2365 	CTX_LOCK(ctx);
2366 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2367 	IFDI_MEDIA_STATUS(ctx, ifmr);
2368 	CTX_UNLOCK(ctx);
2369 }
2370 
2371 void
2372 iflib_stop(if_ctx_t ctx)
2373 {
2374 	iflib_txq_t txq = ctx->ifc_txqs;
2375 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2376 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2377 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2378 	iflib_dma_info_t di;
2379 	iflib_fl_t fl;
2380 	int i, j;
2381 
2382 	/* Tell the stack that the interface is no longer active */
2383 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2384 
2385 	IFDI_INTR_DISABLE(ctx);
2386 	DELAY(1000);
2387 	IFDI_STOP(ctx);
2388 	DELAY(1000);
2389 
2390 	iflib_debug_reset();
2391 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2392 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2393 		/* make sure all transmitters have completed before proceeding XXX */
2394 
2395 		CALLOUT_LOCK(txq);
2396 		callout_stop(&txq->ift_timer);
2397 		CALLOUT_UNLOCK(txq);
2398 
2399 		/* clean any enqueued buffers */
2400 		iflib_ifmp_purge(txq);
2401 		/* Free any existing tx buffers. */
2402 		for (j = 0; j < txq->ift_size; j++) {
2403 			iflib_txsd_free(ctx, txq, j);
2404 		}
2405 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2406 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2407 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2408 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2409 		txq->ift_pullups = 0;
2410 		ifmp_ring_reset_stats(txq->ift_br);
2411 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2412 			bzero((void *)di->idi_vaddr, di->idi_size);
2413 	}
2414 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2415 		/* make sure all transmitters have completed before proceeding XXX */
2416 
2417 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2418 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2419 			bzero((void *)di->idi_vaddr, di->idi_size);
2420 		/* also resets the free lists pidx/cidx */
2421 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2422 			iflib_fl_bufs_free(fl);
2423 	}
2424 }
2425 
2426 static inline caddr_t
2427 calc_next_rxd(iflib_fl_t fl, int cidx)
2428 {
2429 	qidx_t size;
2430 	int nrxd;
2431 	caddr_t start, end, cur, next;
2432 
2433 	nrxd = fl->ifl_size;
2434 	size = fl->ifl_rxd_size;
2435 	start = fl->ifl_ifdi->idi_vaddr;
2436 
2437 	if (__predict_false(size == 0))
2438 		return (start);
2439 	cur = start + size*cidx;
2440 	end = start + size*nrxd;
2441 	next = CACHE_PTR_NEXT(cur);
2442 	return (next < end ? next : start);
2443 }
2444 
2445 static inline void
2446 prefetch_pkts(iflib_fl_t fl, int cidx)
2447 {
2448 	int nextptr;
2449 	int nrxd = fl->ifl_size;
2450 	caddr_t next_rxd;
2451 
2452 
2453 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2454 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2455 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2456 	next_rxd = calc_next_rxd(fl, cidx);
2457 	prefetch(next_rxd);
2458 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2459 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2460 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2461 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2462 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2463 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2464 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2465 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2466 }
2467 
2468 static void
2469 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2470 {
2471 	int flid, cidx;
2472 	bus_dmamap_t map;
2473 	iflib_fl_t fl;
2474 	iflib_dma_info_t di;
2475 	int next;
2476 
2477 	map = NULL;
2478 	flid = irf->irf_flid;
2479 	cidx = irf->irf_idx;
2480 	fl = &rxq->ifr_fl[flid];
2481 	sd->ifsd_fl = fl;
2482 	sd->ifsd_cidx = cidx;
2483 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2484 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2485 	fl->ifl_credits--;
2486 #if MEMORY_LOGGING
2487 	fl->ifl_m_dequeued++;
2488 #endif
2489 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2490 		prefetch_pkts(fl, cidx);
2491 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2492 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2493 	map = fl->ifl_sds.ifsd_map[cidx];
2494 	di = fl->ifl_ifdi;
2495 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2496 
2497 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2498 	MPASS(fl->ifl_cidx == cidx);
2499 	bus_dmamap_sync(fl->ifl_desc_tag, map, BUS_DMASYNC_POSTREAD);
2500 	if (unload)
2501 		bus_dmamap_unload(fl->ifl_desc_tag, map);
2502 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2503 	if (__predict_false(fl->ifl_cidx == 0))
2504 		fl->ifl_gen = 0;
2505 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2506 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2507 	bit_clear(fl->ifl_rx_bitmap, cidx);
2508 }
2509 
2510 static struct mbuf *
2511 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2512 {
2513 	int i, padlen , flags;
2514 	struct mbuf *m, *mh, *mt;
2515 	caddr_t cl;
2516 
2517 	i = 0;
2518 	mh = NULL;
2519 	do {
2520 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2521 
2522 		MPASS(*sd->ifsd_cl != NULL);
2523 		MPASS(*sd->ifsd_m != NULL);
2524 
2525 		/* Don't include zero-length frags */
2526 		if (ri->iri_frags[i].irf_len == 0) {
2527 			/* XXX we can save the cluster here, but not the mbuf */
2528 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2529 			m_free(*sd->ifsd_m);
2530 			*sd->ifsd_m = NULL;
2531 			continue;
2532 		}
2533 		m = *sd->ifsd_m;
2534 		*sd->ifsd_m = NULL;
2535 		if (mh == NULL) {
2536 			flags = M_PKTHDR|M_EXT;
2537 			mh = mt = m;
2538 			padlen = ri->iri_pad;
2539 		} else {
2540 			flags = M_EXT;
2541 			mt->m_next = m;
2542 			mt = m;
2543 			/* assuming padding is only on the first fragment */
2544 			padlen = 0;
2545 		}
2546 		cl = *sd->ifsd_cl;
2547 		*sd->ifsd_cl = NULL;
2548 
2549 		/* Can these two be made one ? */
2550 		m_init(m, M_NOWAIT, MT_DATA, flags);
2551 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2552 		/*
2553 		 * These must follow m_init and m_cljset
2554 		 */
2555 		m->m_data += padlen;
2556 		ri->iri_len -= padlen;
2557 		m->m_len = ri->iri_frags[i].irf_len;
2558 	} while (++i < ri->iri_nfrags);
2559 
2560 	return (mh);
2561 }
2562 
2563 /*
2564  * Process one software descriptor
2565  */
2566 static struct mbuf *
2567 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2568 {
2569 	struct if_rxsd sd;
2570 	struct mbuf *m;
2571 
2572 	/* should I merge this back in now that the two paths are basically duplicated? */
2573 	if (ri->iri_nfrags == 1 &&
2574 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2575 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2576 		m = *sd.ifsd_m;
2577 		*sd.ifsd_m = NULL;
2578 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2579 #ifndef __NO_STRICT_ALIGNMENT
2580 		if (!IP_ALIGNED(m))
2581 			m->m_data += 2;
2582 #endif
2583 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2584 		bus_dmamap_sync(rxq->ifr_fl->ifl_desc_tag,
2585 		    rxq->ifr_fl->ifl_sds.ifsd_map[ri->iri_frags[0].irf_idx],
2586 		    BUS_DMASYNC_PREREAD);
2587 		m->m_len = ri->iri_frags[0].irf_len;
2588        } else {
2589 		m = assemble_segments(rxq, ri, &sd);
2590 	}
2591 	m->m_pkthdr.len = ri->iri_len;
2592 	m->m_pkthdr.rcvif = ri->iri_ifp;
2593 	m->m_flags |= ri->iri_flags;
2594 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2595 	m->m_pkthdr.flowid = ri->iri_flowid;
2596 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2597 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2598 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2599 	return (m);
2600 }
2601 
2602 #if defined(INET6) || defined(INET)
2603 static void
2604 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2605 {
2606 	CURVNET_SET(lc->ifp->if_vnet);
2607 #if defined(INET6)
2608 	*v6 = VNET(ip6_forwarding);
2609 #endif
2610 #if defined(INET)
2611 	*v4 = VNET(ipforwarding);
2612 #endif
2613 	CURVNET_RESTORE();
2614 }
2615 
2616 /*
2617  * Returns true if it's possible this packet could be LROed.
2618  * if it returns false, it is guaranteed that tcp_lro_rx()
2619  * would not return zero.
2620  */
2621 static bool
2622 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2623 {
2624 	struct ether_header *eh;
2625 	uint16_t eh_type;
2626 
2627 	eh = mtod(m, struct ether_header *);
2628 	eh_type = ntohs(eh->ether_type);
2629 	switch (eh_type) {
2630 #if defined(INET6)
2631 		case ETHERTYPE_IPV6:
2632 			return !v6_forwarding;
2633 #endif
2634 #if defined (INET)
2635 		case ETHERTYPE_IP:
2636 			return !v4_forwarding;
2637 #endif
2638 	}
2639 
2640 	return false;
2641 }
2642 #else
2643 static void
2644 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2645 {
2646 }
2647 #endif
2648 
2649 static bool
2650 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2651 {
2652 	if_ctx_t ctx = rxq->ifr_ctx;
2653 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2654 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2655 	iflib_dma_info_t di;
2656 	int avail, i;
2657 	qidx_t *cidxp;
2658 	struct if_rxd_info ri;
2659 	int err, budget_left, rx_bytes, rx_pkts;
2660 	iflib_fl_t fl;
2661 	struct ifnet *ifp;
2662 	int lro_enabled;
2663 	bool v4_forwarding, v6_forwarding, lro_possible;
2664 
2665 	/*
2666 	 * XXX early demux data packets so that if_input processing only handles
2667 	 * acks in interrupt context
2668 	 */
2669 	struct mbuf *m, *mh, *mt, *mf;
2670 
2671 	lro_possible = v4_forwarding = v6_forwarding = false;
2672 	ifp = ctx->ifc_ifp;
2673 	mh = mt = NULL;
2674 	MPASS(budget > 0);
2675 	rx_pkts	= rx_bytes = 0;
2676 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2677 		cidxp = &rxq->ifr_cq_cidx;
2678 	else
2679 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2680 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2681 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2682 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2683 		DBG_COUNTER_INC(rx_unavail);
2684 		return (false);
2685 	}
2686 
2687 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2688 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2689 			DBG_COUNTER_INC(rx_ctx_inactive);
2690 			break;
2691 		}
2692 		/*
2693 		 * Reset client set fields to their default values
2694 		 */
2695 		rxd_info_zero(&ri);
2696 		ri.iri_qsidx = rxq->ifr_id;
2697 		ri.iri_cidx = *cidxp;
2698 		ri.iri_ifp = ifp;
2699 		ri.iri_frags = rxq->ifr_frags;
2700 		di = rxq->ifr_fl[rxq->ifr_frags[0].irf_flid].ifl_ifdi;
2701 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2702 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2703 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2704 
2705 		if (err)
2706 			goto err;
2707 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2708 			*cidxp = ri.iri_cidx;
2709 			/* Update our consumer index */
2710 			/* XXX NB: shurd - check if this is still safe */
2711 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2712 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2713 				rxq->ifr_cq_gen = 0;
2714 			}
2715 			/* was this only a completion queue message? */
2716 			if (__predict_false(ri.iri_nfrags == 0))
2717 				continue;
2718 		}
2719 		MPASS(ri.iri_nfrags != 0);
2720 		MPASS(ri.iri_len != 0);
2721 
2722 		/* will advance the cidx on the corresponding free lists */
2723 		m = iflib_rxd_pkt_get(rxq, &ri);
2724 		avail--;
2725 		budget_left--;
2726 		if (avail == 0 && budget_left)
2727 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2728 
2729 		if (__predict_false(m == NULL)) {
2730 			DBG_COUNTER_INC(rx_mbuf_null);
2731 			continue;
2732 		}
2733 		/* imm_pkt: -- cxgb */
2734 		if (mh == NULL)
2735 			mh = mt = m;
2736 		else {
2737 			mt->m_nextpkt = m;
2738 			mt = m;
2739 		}
2740 	}
2741 	/* make sure that we can refill faster than drain */
2742 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2743 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2744 
2745 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2746 	if (lro_enabled)
2747 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2748 	mt = mf = NULL;
2749 	while (mh != NULL) {
2750 		m = mh;
2751 		mh = mh->m_nextpkt;
2752 		m->m_nextpkt = NULL;
2753 #ifndef __NO_STRICT_ALIGNMENT
2754 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2755 			continue;
2756 #endif
2757 		rx_bytes += m->m_pkthdr.len;
2758 		rx_pkts++;
2759 #if defined(INET6) || defined(INET)
2760 		if (lro_enabled) {
2761 			if (!lro_possible) {
2762 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2763 				if (lro_possible && mf != NULL) {
2764 					ifp->if_input(ifp, mf);
2765 					DBG_COUNTER_INC(rx_if_input);
2766 					mt = mf = NULL;
2767 				}
2768 			}
2769 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2770 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2771 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2772 					continue;
2773 			}
2774 		}
2775 #endif
2776 		if (lro_possible) {
2777 			ifp->if_input(ifp, m);
2778 			DBG_COUNTER_INC(rx_if_input);
2779 			continue;
2780 		}
2781 
2782 		if (mf == NULL)
2783 			mf = m;
2784 		if (mt != NULL)
2785 			mt->m_nextpkt = m;
2786 		mt = m;
2787 	}
2788 	if (mf != NULL) {
2789 		ifp->if_input(ifp, mf);
2790 		DBG_COUNTER_INC(rx_if_input);
2791 	}
2792 
2793 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2794 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2795 
2796 	/*
2797 	 * Flush any outstanding LRO work
2798 	 */
2799 #if defined(INET6) || defined(INET)
2800 	tcp_lro_flush_all(&rxq->ifr_lc);
2801 #endif
2802 	if (avail)
2803 		return true;
2804 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2805 err:
2806 	STATE_LOCK(ctx);
2807 	ctx->ifc_flags |= IFC_DO_RESET;
2808 	iflib_admin_intr_deferred(ctx);
2809 	STATE_UNLOCK(ctx);
2810 	return (false);
2811 }
2812 
2813 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2814 static inline qidx_t
2815 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2816 {
2817 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2818 	qidx_t minthresh = txq->ift_size / 8;
2819 	if (in_use > 4*minthresh)
2820 		return (notify_count);
2821 	if (in_use > 2*minthresh)
2822 		return (notify_count >> 1);
2823 	if (in_use > minthresh)
2824 		return (notify_count >> 3);
2825 	return (0);
2826 }
2827 
2828 static inline qidx_t
2829 txq_max_rs_deferred(iflib_txq_t txq)
2830 {
2831 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2832 	qidx_t minthresh = txq->ift_size / 8;
2833 	if (txq->ift_in_use > 4*minthresh)
2834 		return (notify_count);
2835 	if (txq->ift_in_use > 2*minthresh)
2836 		return (notify_count >> 1);
2837 	if (txq->ift_in_use > minthresh)
2838 		return (notify_count >> 2);
2839 	return (2);
2840 }
2841 
2842 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2843 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2844 
2845 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2846 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2847 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2848 
2849 /* forward compatibility for cxgb */
2850 #define FIRST_QSET(ctx) 0
2851 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2852 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2853 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2854 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2855 
2856 /* XXX we should be setting this to something other than zero */
2857 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2858 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2859     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2860 
2861 static inline bool
2862 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2863 {
2864 	qidx_t dbval, max;
2865 	bool rang;
2866 
2867 	rang = false;
2868 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2869 	if (ring || txq->ift_db_pending >= max) {
2870 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2871 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2872 		txq->ift_db_pending = txq->ift_npending = 0;
2873 		rang = true;
2874 	}
2875 	return (rang);
2876 }
2877 
2878 #ifdef PKT_DEBUG
2879 static void
2880 print_pkt(if_pkt_info_t pi)
2881 {
2882 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2883 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2884 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2885 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2886 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2887 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2888 }
2889 #endif
2890 
2891 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2892 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2893 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2894 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2895 
2896 static int
2897 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2898 {
2899 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2900 	struct ether_vlan_header *eh;
2901 	struct mbuf *m;
2902 
2903 	m = *mp;
2904 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2905 	    M_WRITABLE(m) == 0) {
2906 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2907 			return (ENOMEM);
2908 		} else {
2909 			m_freem(*mp);
2910 			DBG_COUNTER_INC(tx_frees);
2911 			*mp = m;
2912 		}
2913 	}
2914 
2915 	/*
2916 	 * Determine where frame payload starts.
2917 	 * Jump over vlan headers if already present,
2918 	 * helpful for QinQ too.
2919 	 */
2920 	if (__predict_false(m->m_len < sizeof(*eh))) {
2921 		txq->ift_pullups++;
2922 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2923 			return (ENOMEM);
2924 	}
2925 	eh = mtod(m, struct ether_vlan_header *);
2926 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2927 		pi->ipi_etype = ntohs(eh->evl_proto);
2928 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2929 	} else {
2930 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2931 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2932 	}
2933 
2934 	switch (pi->ipi_etype) {
2935 #ifdef INET
2936 	case ETHERTYPE_IP:
2937 	{
2938 		struct mbuf *n;
2939 		struct ip *ip = NULL;
2940 		struct tcphdr *th = NULL;
2941 		int minthlen;
2942 
2943 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2944 		if (__predict_false(m->m_len < minthlen)) {
2945 			/*
2946 			 * if this code bloat is causing too much of a hit
2947 			 * move it to a separate function and mark it noinline
2948 			 */
2949 			if (m->m_len == pi->ipi_ehdrlen) {
2950 				n = m->m_next;
2951 				MPASS(n);
2952 				if (n->m_len >= sizeof(*ip))  {
2953 					ip = (struct ip *)n->m_data;
2954 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2955 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2956 				} else {
2957 					txq->ift_pullups++;
2958 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2959 						return (ENOMEM);
2960 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2961 				}
2962 			} else {
2963 				txq->ift_pullups++;
2964 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2965 					return (ENOMEM);
2966 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2967 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2968 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2969 			}
2970 		} else {
2971 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2972 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2973 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2974 		}
2975 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2976 		pi->ipi_ipproto = ip->ip_p;
2977 		pi->ipi_flags |= IPI_TX_IPV4;
2978 
2979 		/* TCP checksum offload may require TCP header length */
2980 		if (IS_TX_OFFLOAD4(pi)) {
2981 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2982 				if (__predict_false(th == NULL)) {
2983 					txq->ift_pullups++;
2984 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2985 						return (ENOMEM);
2986 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2987 				}
2988 				pi->ipi_tcp_hflags = th->th_flags;
2989 				pi->ipi_tcp_hlen = th->th_off << 2;
2990 				pi->ipi_tcp_seq = th->th_seq;
2991 			}
2992 			if (IS_TSO4(pi)) {
2993 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2994 					return (ENXIO);
2995 				/*
2996 				 * TSO always requires hardware checksum offload.
2997 				 */
2998 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
2999 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3000 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3001 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3002 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3003 					ip->ip_sum = 0;
3004 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3005 				}
3006 			}
3007 		}
3008 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3009                        ip->ip_sum = 0;
3010 
3011 		break;
3012 	}
3013 #endif
3014 #ifdef INET6
3015 	case ETHERTYPE_IPV6:
3016 	{
3017 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3018 		struct tcphdr *th;
3019 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3020 
3021 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3022 			txq->ift_pullups++;
3023 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3024 				return (ENOMEM);
3025 		}
3026 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3027 
3028 		/* XXX-BZ this will go badly in case of ext hdrs. */
3029 		pi->ipi_ipproto = ip6->ip6_nxt;
3030 		pi->ipi_flags |= IPI_TX_IPV6;
3031 
3032 		/* TCP checksum offload may require TCP header length */
3033 		if (IS_TX_OFFLOAD6(pi)) {
3034 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3035 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3036 					txq->ift_pullups++;
3037 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3038 						return (ENOMEM);
3039 				}
3040 				pi->ipi_tcp_hflags = th->th_flags;
3041 				pi->ipi_tcp_hlen = th->th_off << 2;
3042 				pi->ipi_tcp_seq = th->th_seq;
3043 			}
3044 			if (IS_TSO6(pi)) {
3045 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3046 					return (ENXIO);
3047 				/*
3048 				 * TSO always requires hardware checksum offload.
3049 				 */
3050 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3051 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3052 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3053 			}
3054 		}
3055 		break;
3056 	}
3057 #endif
3058 	default:
3059 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3060 		pi->ipi_ip_hlen = 0;
3061 		break;
3062 	}
3063 	*mp = m;
3064 
3065 	return (0);
3066 }
3067 
3068 /*
3069  * If dodgy hardware rejects the scatter gather chain we've handed it
3070  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3071  * m_defrag'd mbufs
3072  */
3073 static __noinline struct mbuf *
3074 iflib_remove_mbuf(iflib_txq_t txq)
3075 {
3076 	int ntxd, pidx;
3077 	struct mbuf *m, **ifsd_m;
3078 
3079 	ifsd_m = txq->ift_sds.ifsd_m;
3080 	ntxd = txq->ift_size;
3081 	pidx = txq->ift_pidx & (ntxd - 1);
3082 	ifsd_m = txq->ift_sds.ifsd_m;
3083 	m = ifsd_m[pidx];
3084 	ifsd_m[pidx] = NULL;
3085 	bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[pidx]);
3086 	if (txq->ift_sds.ifsd_tso_map != NULL)
3087 		bus_dmamap_unload(txq->ift_tso_desc_tag,
3088 		    txq->ift_sds.ifsd_tso_map[pidx]);
3089 #if MEMORY_LOGGING
3090 	txq->ift_dequeued++;
3091 #endif
3092 	return (m);
3093 }
3094 
3095 static inline caddr_t
3096 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3097 {
3098 	qidx_t size;
3099 	int ntxd;
3100 	caddr_t start, end, cur, next;
3101 
3102 	ntxd = txq->ift_size;
3103 	size = txq->ift_txd_size[qid];
3104 	start = txq->ift_ifdi[qid].idi_vaddr;
3105 
3106 	if (__predict_false(size == 0))
3107 		return (start);
3108 	cur = start + size*cidx;
3109 	end = start + size*ntxd;
3110 	next = CACHE_PTR_NEXT(cur);
3111 	return (next < end ? next : start);
3112 }
3113 
3114 /*
3115  * Pad an mbuf to ensure a minimum ethernet frame size.
3116  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3117  */
3118 static __noinline int
3119 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3120 {
3121 	/*
3122 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3123 	 * and ARP message is the smallest common payload I can think of
3124 	 */
3125 	static char pad[18];	/* just zeros */
3126 	int n;
3127 	struct mbuf *new_head;
3128 
3129 	if (!M_WRITABLE(*m_head)) {
3130 		new_head = m_dup(*m_head, M_NOWAIT);
3131 		if (new_head == NULL) {
3132 			m_freem(*m_head);
3133 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3134 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3135 			DBG_COUNTER_INC(tx_frees);
3136 			return ENOMEM;
3137 		}
3138 		m_freem(*m_head);
3139 		*m_head = new_head;
3140 	}
3141 
3142 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3143 	     n > 0; n -= sizeof(pad))
3144 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3145 			break;
3146 
3147 	if (n > 0) {
3148 		m_freem(*m_head);
3149 		device_printf(dev, "cannot pad short frame\n");
3150 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3151 		DBG_COUNTER_INC(tx_frees);
3152 		return (ENOBUFS);
3153 	}
3154 
3155 	return 0;
3156 }
3157 
3158 static int
3159 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3160 {
3161 	if_ctx_t		ctx;
3162 	if_shared_ctx_t		sctx;
3163 	if_softc_ctx_t		scctx;
3164 	bus_dma_segment_t	*segs;
3165 	struct mbuf		*m_head, **ifsd_m;
3166 	void			*next_txd;
3167 	bus_dmamap_t		map;
3168 	struct if_pkt_info	pi;
3169 	int remap = 0;
3170 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3171 	bus_dma_tag_t desc_tag;
3172 
3173 	ctx = txq->ift_ctx;
3174 	sctx = ctx->ifc_sctx;
3175 	scctx = &ctx->ifc_softc_ctx;
3176 	segs = txq->ift_segs;
3177 	ntxd = txq->ift_size;
3178 	m_head = *m_headp;
3179 	map = NULL;
3180 
3181 	/*
3182 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3183 	 */
3184 	cidx = txq->ift_cidx;
3185 	pidx = txq->ift_pidx;
3186 	if (ctx->ifc_flags & IFC_PREFETCH) {
3187 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3188 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3189 			next_txd = calc_next_txd(txq, cidx, 0);
3190 			prefetch(next_txd);
3191 		}
3192 
3193 		/* prefetch the next cache line of mbuf pointers and flags */
3194 		prefetch(&txq->ift_sds.ifsd_m[next]);
3195 		prefetch(&txq->ift_sds.ifsd_map[next]);
3196 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3197 	}
3198 	map = txq->ift_sds.ifsd_map[pidx];
3199 	ifsd_m = txq->ift_sds.ifsd_m;
3200 
3201 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3202 		desc_tag = txq->ift_tso_desc_tag;
3203 		max_segs = scctx->isc_tx_tso_segments_max;
3204 		map = txq->ift_sds.ifsd_tso_map[pidx];
3205 		MPASS(desc_tag != NULL);
3206 		MPASS(max_segs > 0);
3207 	} else {
3208 		desc_tag = txq->ift_desc_tag;
3209 		max_segs = scctx->isc_tx_nsegments;
3210 		map = txq->ift_sds.ifsd_map[pidx];
3211 	}
3212 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3213 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3214 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3215 		if (err) {
3216 			DBG_COUNTER_INC(encap_txd_encap_fail);
3217 			return err;
3218 		}
3219 	}
3220 	m_head = *m_headp;
3221 
3222 	pkt_info_zero(&pi);
3223 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3224 	pi.ipi_pidx = pidx;
3225 	pi.ipi_qsidx = txq->ift_id;
3226 	pi.ipi_len = m_head->m_pkthdr.len;
3227 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3228 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3229 
3230 	/* deliberate bitwise OR to make one condition */
3231 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3232 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3233 			DBG_COUNTER_INC(encap_txd_encap_fail);
3234 			return (err);
3235 		}
3236 		m_head = *m_headp;
3237 	}
3238 
3239 retry:
3240 	err = bus_dmamap_load_mbuf_sg(desc_tag, map, m_head, segs, &nsegs,
3241 	    BUS_DMA_NOWAIT);
3242 defrag:
3243 	if (__predict_false(err)) {
3244 		switch (err) {
3245 		case EFBIG:
3246 			/* try collapse once and defrag once */
3247 			if (remap == 0) {
3248 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3249 				/* try defrag if collapsing fails */
3250 				if (m_head == NULL)
3251 					remap++;
3252 			}
3253 			if (remap == 1) {
3254 				txq->ift_mbuf_defrag++;
3255 				m_head = m_defrag(*m_headp, M_NOWAIT);
3256 			}
3257 			remap++;
3258 			if (__predict_false(m_head == NULL))
3259 				goto defrag_failed;
3260 			*m_headp = m_head;
3261 			goto retry;
3262 			break;
3263 		case ENOMEM:
3264 			txq->ift_no_tx_dma_setup++;
3265 			break;
3266 		default:
3267 			txq->ift_no_tx_dma_setup++;
3268 			m_freem(*m_headp);
3269 			DBG_COUNTER_INC(tx_frees);
3270 			*m_headp = NULL;
3271 			break;
3272 		}
3273 		txq->ift_map_failed++;
3274 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3275 		DBG_COUNTER_INC(encap_txd_encap_fail);
3276 		return (err);
3277 	}
3278 	ifsd_m[pidx] = m_head;
3279 	/*
3280 	 * XXX assumes a 1 to 1 relationship between segments and
3281 	 *        descriptors - this does not hold true on all drivers, e.g.
3282 	 *        cxgb
3283 	 */
3284 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3285 		txq->ift_no_desc_avail++;
3286 		bus_dmamap_unload(desc_tag, map);
3287 		DBG_COUNTER_INC(encap_txq_avail_fail);
3288 		DBG_COUNTER_INC(encap_txd_encap_fail);
3289 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3290 			GROUPTASK_ENQUEUE(&txq->ift_task);
3291 		return (ENOBUFS);
3292 	}
3293 	/*
3294 	 * On Intel cards we can greatly reduce the number of TX interrupts
3295 	 * we see by only setting report status on every Nth descriptor.
3296 	 * However, this also means that the driver will need to keep track
3297 	 * of the descriptors that RS was set on to check them for the DD bit.
3298 	 */
3299 	txq->ift_rs_pending += nsegs + 1;
3300 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3301 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3302 		pi.ipi_flags |= IPI_TX_INTR;
3303 		txq->ift_rs_pending = 0;
3304 	}
3305 
3306 	pi.ipi_segs = segs;
3307 	pi.ipi_nsegs = nsegs;
3308 
3309 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3310 #ifdef PKT_DEBUG
3311 	print_pkt(&pi);
3312 #endif
3313 	bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3314 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3315 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3316 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3317 		DBG_COUNTER_INC(tx_encap);
3318 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3319 
3320 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3321 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3322 			ndesc += txq->ift_size;
3323 			txq->ift_gen = 1;
3324 		}
3325 		/*
3326 		 * drivers can need as many as
3327 		 * two sentinels
3328 		 */
3329 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3330 		MPASS(pi.ipi_new_pidx != pidx);
3331 		MPASS(ndesc > 0);
3332 		txq->ift_in_use += ndesc;
3333 
3334 		/*
3335 		 * We update the last software descriptor again here because there may
3336 		 * be a sentinel and/or there may be more mbufs than segments
3337 		 */
3338 		txq->ift_pidx = pi.ipi_new_pidx;
3339 		txq->ift_npending += pi.ipi_ndescs;
3340 	} else {
3341 		*m_headp = m_head = iflib_remove_mbuf(txq);
3342 		if (err == EFBIG) {
3343 			txq->ift_txd_encap_efbig++;
3344 			if (remap < 2) {
3345 				remap = 1;
3346 				goto defrag;
3347 			}
3348 		}
3349 		goto defrag_failed;
3350 	}
3351 	/*
3352 	 * err can't possibly be non-zero here, so we don't neet to test it
3353 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3354 	 */
3355 	return (err);
3356 
3357 defrag_failed:
3358 	txq->ift_mbuf_defrag_failed++;
3359 	txq->ift_map_failed++;
3360 	m_freem(*m_headp);
3361 	DBG_COUNTER_INC(tx_frees);
3362 	*m_headp = NULL;
3363 	DBG_COUNTER_INC(encap_txd_encap_fail);
3364 	return (ENOMEM);
3365 }
3366 
3367 static void
3368 iflib_tx_desc_free(iflib_txq_t txq, int n)
3369 {
3370 	uint32_t qsize, cidx, mask, gen;
3371 	struct mbuf *m, **ifsd_m;
3372 	bool do_prefetch;
3373 
3374 	cidx = txq->ift_cidx;
3375 	gen = txq->ift_gen;
3376 	qsize = txq->ift_size;
3377 	mask = qsize-1;
3378 	ifsd_m = txq->ift_sds.ifsd_m;
3379 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3380 
3381 	while (n-- > 0) {
3382 		if (do_prefetch) {
3383 			prefetch(ifsd_m[(cidx + 3) & mask]);
3384 			prefetch(ifsd_m[(cidx + 4) & mask]);
3385 		}
3386 		if ((m = ifsd_m[cidx]) != NULL) {
3387 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3388 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3389 				bus_dmamap_sync(txq->ift_tso_desc_tag,
3390 				    txq->ift_sds.ifsd_tso_map[cidx],
3391 				    BUS_DMASYNC_POSTWRITE);
3392 				bus_dmamap_unload(txq->ift_tso_desc_tag,
3393 				    txq->ift_sds.ifsd_tso_map[cidx]);
3394 			} else {
3395 				bus_dmamap_sync(txq->ift_desc_tag,
3396 				    txq->ift_sds.ifsd_map[cidx],
3397 				    BUS_DMASYNC_POSTWRITE);
3398 				bus_dmamap_unload(txq->ift_desc_tag,
3399 				    txq->ift_sds.ifsd_map[cidx]);
3400 			}
3401 			/* XXX we don't support any drivers that batch packets yet */
3402 			MPASS(m->m_nextpkt == NULL);
3403 			m_freem(m);
3404 			ifsd_m[cidx] = NULL;
3405 #if MEMORY_LOGGING
3406 			txq->ift_dequeued++;
3407 #endif
3408 			DBG_COUNTER_INC(tx_frees);
3409 		}
3410 		if (__predict_false(++cidx == qsize)) {
3411 			cidx = 0;
3412 			gen = 0;
3413 		}
3414 	}
3415 	txq->ift_cidx = cidx;
3416 	txq->ift_gen = gen;
3417 }
3418 
3419 static __inline int
3420 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3421 {
3422 	int reclaim;
3423 	if_ctx_t ctx = txq->ift_ctx;
3424 
3425 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3426 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3427 
3428 	/*
3429 	 * Need a rate-limiting check so that this isn't called every time
3430 	 */
3431 	iflib_tx_credits_update(ctx, txq);
3432 	reclaim = DESC_RECLAIMABLE(txq);
3433 
3434 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3435 #ifdef INVARIANTS
3436 		if (iflib_verbose_debug) {
3437 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3438 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3439 			       reclaim, thresh);
3440 
3441 		}
3442 #endif
3443 		return (0);
3444 	}
3445 	iflib_tx_desc_free(txq, reclaim);
3446 	txq->ift_cleaned += reclaim;
3447 	txq->ift_in_use -= reclaim;
3448 
3449 	return (reclaim);
3450 }
3451 
3452 static struct mbuf **
3453 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3454 {
3455 	int next, size;
3456 	struct mbuf **items;
3457 
3458 	size = r->size;
3459 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3460 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3461 
3462 	prefetch(items[(cidx + offset) & (size-1)]);
3463 	if (remaining > 1) {
3464 		prefetch2cachelines(&items[next]);
3465 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3466 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3467 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3468 	}
3469 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3470 }
3471 
3472 static void
3473 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3474 {
3475 
3476 	ifmp_ring_check_drainage(txq->ift_br, budget);
3477 }
3478 
3479 static uint32_t
3480 iflib_txq_can_drain(struct ifmp_ring *r)
3481 {
3482 	iflib_txq_t txq = r->cookie;
3483 	if_ctx_t ctx = txq->ift_ctx;
3484 
3485 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3486 	    BUS_DMASYNC_POSTREAD);
3487 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3488 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3489 }
3490 
3491 static uint32_t
3492 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3493 {
3494 	iflib_txq_t txq = r->cookie;
3495 	if_ctx_t ctx = txq->ift_ctx;
3496 	struct ifnet *ifp = ctx->ifc_ifp;
3497 	struct mbuf **mp, *m;
3498 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3499 	int reclaimed, err, in_use_prev, desc_used;
3500 	bool do_prefetch, ring, rang;
3501 
3502 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3503 			    !LINK_ACTIVE(ctx))) {
3504 		DBG_COUNTER_INC(txq_drain_notready);
3505 		return (0);
3506 	}
3507 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3508 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3509 	avail = IDXDIFF(pidx, cidx, r->size);
3510 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3511 		DBG_COUNTER_INC(txq_drain_flushing);
3512 		for (i = 0; i < avail; i++) {
3513 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3514 				m_free(r->items[(cidx + i) & (r->size-1)]);
3515 			r->items[(cidx + i) & (r->size-1)] = NULL;
3516 		}
3517 		return (avail);
3518 	}
3519 
3520 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3521 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3522 		CALLOUT_LOCK(txq);
3523 		callout_stop(&txq->ift_timer);
3524 		CALLOUT_UNLOCK(txq);
3525 		DBG_COUNTER_INC(txq_drain_oactive);
3526 		return (0);
3527 	}
3528 	if (reclaimed)
3529 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3530 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3531 	count = MIN(avail, TX_BATCH_SIZE);
3532 #ifdef INVARIANTS
3533 	if (iflib_verbose_debug)
3534 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3535 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3536 #endif
3537 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3538 	avail = TXQ_AVAIL(txq);
3539 	err = 0;
3540 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3541 		int rem = do_prefetch ? count - i : 0;
3542 
3543 		mp = _ring_peek_one(r, cidx, i, rem);
3544 		MPASS(mp != NULL && *mp != NULL);
3545 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3546 			consumed++;
3547 			reclaimed++;
3548 			continue;
3549 		}
3550 		in_use_prev = txq->ift_in_use;
3551 		err = iflib_encap(txq, mp);
3552 		if (__predict_false(err)) {
3553 			/* no room - bail out */
3554 			if (err == ENOBUFS)
3555 				break;
3556 			consumed++;
3557 			/* we can't send this packet - skip it */
3558 			continue;
3559 		}
3560 		consumed++;
3561 		pkt_sent++;
3562 		m = *mp;
3563 		DBG_COUNTER_INC(tx_sent);
3564 		bytes_sent += m->m_pkthdr.len;
3565 		mcast_sent += !!(m->m_flags & M_MCAST);
3566 		avail = TXQ_AVAIL(txq);
3567 
3568 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3569 		desc_used += (txq->ift_in_use - in_use_prev);
3570 		ETHER_BPF_MTAP(ifp, m);
3571 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3572 			break;
3573 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3574 	}
3575 
3576 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3577 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3578 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3579 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3580 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3581 	if (mcast_sent)
3582 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3583 #ifdef INVARIANTS
3584 	if (iflib_verbose_debug)
3585 		printf("consumed=%d\n", consumed);
3586 #endif
3587 	return (consumed);
3588 }
3589 
3590 static uint32_t
3591 iflib_txq_drain_always(struct ifmp_ring *r)
3592 {
3593 	return (1);
3594 }
3595 
3596 static uint32_t
3597 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3598 {
3599 	int i, avail;
3600 	struct mbuf **mp;
3601 	iflib_txq_t txq;
3602 
3603 	txq = r->cookie;
3604 
3605 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3606 	CALLOUT_LOCK(txq);
3607 	callout_stop(&txq->ift_timer);
3608 	CALLOUT_UNLOCK(txq);
3609 
3610 	avail = IDXDIFF(pidx, cidx, r->size);
3611 	for (i = 0; i < avail; i++) {
3612 		mp = _ring_peek_one(r, cidx, i, avail - i);
3613 		if (__predict_false(*mp == (struct mbuf *)txq))
3614 			continue;
3615 		m_freem(*mp);
3616 		DBG_COUNTER_INC(tx_frees);
3617 	}
3618 	MPASS(ifmp_ring_is_stalled(r) == 0);
3619 	return (avail);
3620 }
3621 
3622 static void
3623 iflib_ifmp_purge(iflib_txq_t txq)
3624 {
3625 	struct ifmp_ring *r;
3626 
3627 	r = txq->ift_br;
3628 	r->drain = iflib_txq_drain_free;
3629 	r->can_drain = iflib_txq_drain_always;
3630 
3631 	ifmp_ring_check_drainage(r, r->size);
3632 
3633 	r->drain = iflib_txq_drain;
3634 	r->can_drain = iflib_txq_can_drain;
3635 }
3636 
3637 static void
3638 _task_fn_tx(void *context)
3639 {
3640 	iflib_txq_t txq = context;
3641 	if_ctx_t ctx = txq->ift_ctx;
3642 	struct ifnet *ifp = ctx->ifc_ifp;
3643 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3644 
3645 #ifdef IFLIB_DIAGNOSTICS
3646 	txq->ift_cpu_exec_count[curcpu]++;
3647 #endif
3648 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3649 		return;
3650 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3651 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3652 		    BUS_DMASYNC_POSTREAD);
3653 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3654 			netmap_tx_irq(ifp, txq->ift_id);
3655 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3656 		return;
3657 	}
3658 #ifdef ALTQ
3659 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3660 		iflib_altq_if_start(ifp);
3661 #endif
3662 	if (txq->ift_db_pending)
3663 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3664 	else if (!abdicate)
3665 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3666 	/*
3667 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3668 	 */
3669 	if (abdicate)
3670 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3671 	if (ctx->ifc_flags & IFC_LEGACY)
3672 		IFDI_INTR_ENABLE(ctx);
3673 	else {
3674 #ifdef INVARIANTS
3675 		int rc =
3676 #endif
3677 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3678 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3679 	}
3680 }
3681 
3682 static void
3683 _task_fn_rx(void *context)
3684 {
3685 	iflib_rxq_t rxq = context;
3686 	if_ctx_t ctx = rxq->ifr_ctx;
3687 	bool more;
3688 	uint16_t budget;
3689 
3690 #ifdef IFLIB_DIAGNOSTICS
3691 	rxq->ifr_cpu_exec_count[curcpu]++;
3692 #endif
3693 	DBG_COUNTER_INC(task_fn_rxs);
3694 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3695 		return;
3696 	more = true;
3697 #ifdef DEV_NETMAP
3698 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3699 		u_int work = 0;
3700 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3701 			more = false;
3702 		}
3703 	}
3704 #endif
3705 	budget = ctx->ifc_sysctl_rx_budget;
3706 	if (budget == 0)
3707 		budget = 16;	/* XXX */
3708 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3709 		if (ctx->ifc_flags & IFC_LEGACY)
3710 			IFDI_INTR_ENABLE(ctx);
3711 		else {
3712 #ifdef INVARIANTS
3713 			int rc =
3714 #endif
3715 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3716 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3717 			DBG_COUNTER_INC(rx_intr_enables);
3718 		}
3719 	}
3720 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3721 		return;
3722 	if (more)
3723 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3724 }
3725 
3726 static void
3727 _task_fn_admin(void *context)
3728 {
3729 	if_ctx_t ctx = context;
3730 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3731 	iflib_txq_t txq;
3732 	int i;
3733 	bool oactive, running, do_reset, do_watchdog, in_detach;
3734 	uint32_t reset_on = hz / 2;
3735 
3736 	STATE_LOCK(ctx);
3737 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3738 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3739 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3740 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3741 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3742 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3743 	STATE_UNLOCK(ctx);
3744 
3745 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3746 		return;
3747 	if (in_detach)
3748 		return;
3749 
3750 	CTX_LOCK(ctx);
3751 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3752 		CALLOUT_LOCK(txq);
3753 		callout_stop(&txq->ift_timer);
3754 		CALLOUT_UNLOCK(txq);
3755 	}
3756 	if (do_watchdog) {
3757 		ctx->ifc_watchdog_events++;
3758 		IFDI_WATCHDOG_RESET(ctx);
3759 	}
3760 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3761 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3762 #ifdef DEV_NETMAP
3763 		reset_on = hz / 2;
3764 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3765 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3766 #endif
3767 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3768 	}
3769 	IFDI_LINK_INTR_ENABLE(ctx);
3770 	if (do_reset)
3771 		iflib_if_init_locked(ctx);
3772 	CTX_UNLOCK(ctx);
3773 
3774 	if (LINK_ACTIVE(ctx) == 0)
3775 		return;
3776 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3777 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3778 }
3779 
3780 
3781 static void
3782 _task_fn_iov(void *context)
3783 {
3784 	if_ctx_t ctx = context;
3785 
3786 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3787 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3788 		return;
3789 
3790 	CTX_LOCK(ctx);
3791 	IFDI_VFLR_HANDLE(ctx);
3792 	CTX_UNLOCK(ctx);
3793 }
3794 
3795 static int
3796 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3797 {
3798 	int err;
3799 	if_int_delay_info_t info;
3800 	if_ctx_t ctx;
3801 
3802 	info = (if_int_delay_info_t)arg1;
3803 	ctx = info->iidi_ctx;
3804 	info->iidi_req = req;
3805 	info->iidi_oidp = oidp;
3806 	CTX_LOCK(ctx);
3807 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3808 	CTX_UNLOCK(ctx);
3809 	return (err);
3810 }
3811 
3812 /*********************************************************************
3813  *
3814  *  IFNET FUNCTIONS
3815  *
3816  **********************************************************************/
3817 
3818 static void
3819 iflib_if_init_locked(if_ctx_t ctx)
3820 {
3821 	iflib_stop(ctx);
3822 	iflib_init_locked(ctx);
3823 }
3824 
3825 
3826 static void
3827 iflib_if_init(void *arg)
3828 {
3829 	if_ctx_t ctx = arg;
3830 
3831 	CTX_LOCK(ctx);
3832 	iflib_if_init_locked(ctx);
3833 	CTX_UNLOCK(ctx);
3834 }
3835 
3836 static int
3837 iflib_if_transmit(if_t ifp, struct mbuf *m)
3838 {
3839 	if_ctx_t	ctx = if_getsoftc(ifp);
3840 
3841 	iflib_txq_t txq;
3842 	int err, qidx;
3843 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3844 
3845 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3846 		DBG_COUNTER_INC(tx_frees);
3847 		m_freem(m);
3848 		return (ENOBUFS);
3849 	}
3850 
3851 	MPASS(m->m_nextpkt == NULL);
3852 	/* ALTQ-enabled interfaces always use queue 0. */
3853 	qidx = 0;
3854 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3855 		qidx = QIDX(ctx, m);
3856 	/*
3857 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3858 	 */
3859 	txq = &ctx->ifc_txqs[qidx];
3860 
3861 #ifdef DRIVER_BACKPRESSURE
3862 	if (txq->ift_closed) {
3863 		while (m != NULL) {
3864 			next = m->m_nextpkt;
3865 			m->m_nextpkt = NULL;
3866 			m_freem(m);
3867 			DBG_COUNTER_INC(tx_frees);
3868 			m = next;
3869 		}
3870 		return (ENOBUFS);
3871 	}
3872 #endif
3873 #ifdef notyet
3874 	qidx = count = 0;
3875 	mp = marr;
3876 	next = m;
3877 	do {
3878 		count++;
3879 		next = next->m_nextpkt;
3880 	} while (next != NULL);
3881 
3882 	if (count > nitems(marr))
3883 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3884 			/* XXX check nextpkt */
3885 			m_freem(m);
3886 			/* XXX simplify for now */
3887 			DBG_COUNTER_INC(tx_frees);
3888 			return (ENOBUFS);
3889 		}
3890 	for (next = m, i = 0; next != NULL; i++) {
3891 		mp[i] = next;
3892 		next = next->m_nextpkt;
3893 		mp[i]->m_nextpkt = NULL;
3894 	}
3895 #endif
3896 	DBG_COUNTER_INC(tx_seen);
3897 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3898 
3899 	if (abdicate)
3900 		GROUPTASK_ENQUEUE(&txq->ift_task);
3901  	if (err) {
3902 		if (!abdicate)
3903 			GROUPTASK_ENQUEUE(&txq->ift_task);
3904 		/* support forthcoming later */
3905 #ifdef DRIVER_BACKPRESSURE
3906 		txq->ift_closed = TRUE;
3907 #endif
3908 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3909 		m_freem(m);
3910 		DBG_COUNTER_INC(tx_frees);
3911 	}
3912 
3913 	return (err);
3914 }
3915 
3916 #ifdef ALTQ
3917 /*
3918  * The overall approach to integrating iflib with ALTQ is to continue to use
3919  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3920  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3921  * is redundant/unnecessary, but doing so minimizes the amount of
3922  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3923  * redundantly queueing to an intermediate mp_ring is swamped by the
3924  * performance limitations inherent in using ALTQ.
3925  *
3926  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3927  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3928  * given interface.  If ALTQ is enabled for an interface, then all
3929  * transmitted packets for that interface will be submitted to the ALTQ
3930  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3931  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3932  * update stats that the iflib machinery handles, and which is sensitve to
3933  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3934  * will be installed as the start routine for use by ALTQ facilities that
3935  * need to trigger queue drains on a scheduled basis.
3936  *
3937  */
3938 static void
3939 iflib_altq_if_start(if_t ifp)
3940 {
3941 	struct ifaltq *ifq = &ifp->if_snd;
3942 	struct mbuf *m;
3943 
3944 	IFQ_LOCK(ifq);
3945 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3946 	while (m != NULL) {
3947 		iflib_if_transmit(ifp, m);
3948 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3949 	}
3950 	IFQ_UNLOCK(ifq);
3951 }
3952 
3953 static int
3954 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3955 {
3956 	int err;
3957 
3958 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3959 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3960 		if (err == 0)
3961 			iflib_altq_if_start(ifp);
3962 	} else
3963 		err = iflib_if_transmit(ifp, m);
3964 
3965 	return (err);
3966 }
3967 #endif /* ALTQ */
3968 
3969 static void
3970 iflib_if_qflush(if_t ifp)
3971 {
3972 	if_ctx_t ctx = if_getsoftc(ifp);
3973 	iflib_txq_t txq = ctx->ifc_txqs;
3974 	int i;
3975 
3976 	STATE_LOCK(ctx);
3977 	ctx->ifc_flags |= IFC_QFLUSH;
3978 	STATE_UNLOCK(ctx);
3979 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3980 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3981 			iflib_txq_check_drain(txq, 0);
3982 	STATE_LOCK(ctx);
3983 	ctx->ifc_flags &= ~IFC_QFLUSH;
3984 	STATE_UNLOCK(ctx);
3985 
3986 	/*
3987 	 * When ALTQ is enabled, this will also take care of purging the
3988 	 * ALTQ queue(s).
3989 	 */
3990 	if_qflush(ifp);
3991 }
3992 
3993 
3994 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3995 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3996 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
3997 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
3998 
3999 static int
4000 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4001 {
4002 	if_ctx_t ctx = if_getsoftc(ifp);
4003 	struct ifreq	*ifr = (struct ifreq *)data;
4004 #if defined(INET) || defined(INET6)
4005 	struct ifaddr	*ifa = (struct ifaddr *)data;
4006 #endif
4007 	bool		avoid_reset = FALSE;
4008 	int		err = 0, reinit = 0, bits;
4009 
4010 	switch (command) {
4011 	case SIOCSIFADDR:
4012 #ifdef INET
4013 		if (ifa->ifa_addr->sa_family == AF_INET)
4014 			avoid_reset = TRUE;
4015 #endif
4016 #ifdef INET6
4017 		if (ifa->ifa_addr->sa_family == AF_INET6)
4018 			avoid_reset = TRUE;
4019 #endif
4020 		/*
4021 		** Calling init results in link renegotiation,
4022 		** so we avoid doing it when possible.
4023 		*/
4024 		if (avoid_reset) {
4025 			if_setflagbits(ifp, IFF_UP,0);
4026 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4027 				reinit = 1;
4028 #ifdef INET
4029 			if (!(if_getflags(ifp) & IFF_NOARP))
4030 				arp_ifinit(ifp, ifa);
4031 #endif
4032 		} else
4033 			err = ether_ioctl(ifp, command, data);
4034 		break;
4035 	case SIOCSIFMTU:
4036 		CTX_LOCK(ctx);
4037 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4038 			CTX_UNLOCK(ctx);
4039 			break;
4040 		}
4041 		bits = if_getdrvflags(ifp);
4042 		/* stop the driver and free any clusters before proceeding */
4043 		iflib_stop(ctx);
4044 
4045 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4046 			STATE_LOCK(ctx);
4047 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4048 				ctx->ifc_flags |= IFC_MULTISEG;
4049 			else
4050 				ctx->ifc_flags &= ~IFC_MULTISEG;
4051 			STATE_UNLOCK(ctx);
4052 			err = if_setmtu(ifp, ifr->ifr_mtu);
4053 		}
4054 		iflib_init_locked(ctx);
4055 		STATE_LOCK(ctx);
4056 		if_setdrvflags(ifp, bits);
4057 		STATE_UNLOCK(ctx);
4058 		CTX_UNLOCK(ctx);
4059 		break;
4060 	case SIOCSIFFLAGS:
4061 		CTX_LOCK(ctx);
4062 		if (if_getflags(ifp) & IFF_UP) {
4063 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4064 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4065 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4066 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4067 				}
4068 			} else
4069 				reinit = 1;
4070 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4071 			iflib_stop(ctx);
4072 		}
4073 		ctx->ifc_if_flags = if_getflags(ifp);
4074 		CTX_UNLOCK(ctx);
4075 		break;
4076 	case SIOCADDMULTI:
4077 	case SIOCDELMULTI:
4078 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4079 			CTX_LOCK(ctx);
4080 			IFDI_INTR_DISABLE(ctx);
4081 			IFDI_MULTI_SET(ctx);
4082 			IFDI_INTR_ENABLE(ctx);
4083 			CTX_UNLOCK(ctx);
4084 		}
4085 		break;
4086 	case SIOCSIFMEDIA:
4087 		CTX_LOCK(ctx);
4088 		IFDI_MEDIA_SET(ctx);
4089 		CTX_UNLOCK(ctx);
4090 		/* falls thru */
4091 	case SIOCGIFMEDIA:
4092 	case SIOCGIFXMEDIA:
4093 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4094 		break;
4095 	case SIOCGI2C:
4096 	{
4097 		struct ifi2creq i2c;
4098 
4099 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4100 		if (err != 0)
4101 			break;
4102 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4103 			err = EINVAL;
4104 			break;
4105 		}
4106 		if (i2c.len > sizeof(i2c.data)) {
4107 			err = EINVAL;
4108 			break;
4109 		}
4110 
4111 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4112 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4113 			    sizeof(i2c));
4114 		break;
4115 	}
4116 	case SIOCSIFCAP:
4117 	{
4118 		int mask, setmask, oldmask;
4119 
4120 		oldmask = if_getcapenable(ifp);
4121 		mask = ifr->ifr_reqcap ^ oldmask;
4122 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4123 		setmask = 0;
4124 #ifdef TCP_OFFLOAD
4125 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4126 #endif
4127 		setmask |= (mask & IFCAP_FLAGS);
4128 		setmask |= (mask & IFCAP_WOL);
4129 
4130 		/*
4131 		 * If any RX csum has changed, change all the ones that
4132 		 * are supported by the driver.
4133 		 */
4134 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4135 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4136 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4137 		}
4138 
4139 		/*
4140 		 * want to ensure that traffic has stopped before we change any of the flags
4141 		 */
4142 		if (setmask) {
4143 			CTX_LOCK(ctx);
4144 			bits = if_getdrvflags(ifp);
4145 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4146 				iflib_stop(ctx);
4147 			STATE_LOCK(ctx);
4148 			if_togglecapenable(ifp, setmask);
4149 			STATE_UNLOCK(ctx);
4150 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4151 				iflib_init_locked(ctx);
4152 			STATE_LOCK(ctx);
4153 			if_setdrvflags(ifp, bits);
4154 			STATE_UNLOCK(ctx);
4155 			CTX_UNLOCK(ctx);
4156 		}
4157 		if_vlancap(ifp);
4158 		break;
4159 	}
4160 	case SIOCGPRIVATE_0:
4161 	case SIOCSDRVSPEC:
4162 	case SIOCGDRVSPEC:
4163 		CTX_LOCK(ctx);
4164 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4165 		CTX_UNLOCK(ctx);
4166 		break;
4167 	default:
4168 		err = ether_ioctl(ifp, command, data);
4169 		break;
4170 	}
4171 	if (reinit)
4172 		iflib_if_init(ctx);
4173 	return (err);
4174 }
4175 
4176 static uint64_t
4177 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4178 {
4179 	if_ctx_t ctx = if_getsoftc(ifp);
4180 
4181 	return (IFDI_GET_COUNTER(ctx, cnt));
4182 }
4183 
4184 /*********************************************************************
4185  *
4186  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4187  *
4188  **********************************************************************/
4189 
4190 static void
4191 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4192 {
4193 	if_ctx_t ctx = if_getsoftc(ifp);
4194 
4195 	if ((void *)ctx != arg)
4196 		return;
4197 
4198 	if ((vtag == 0) || (vtag > 4095))
4199 		return;
4200 
4201 	CTX_LOCK(ctx);
4202 	IFDI_VLAN_REGISTER(ctx, vtag);
4203 	/* Re-init to load the changes */
4204 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4205 		iflib_if_init_locked(ctx);
4206 	CTX_UNLOCK(ctx);
4207 }
4208 
4209 static void
4210 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4211 {
4212 	if_ctx_t ctx = if_getsoftc(ifp);
4213 
4214 	if ((void *)ctx != arg)
4215 		return;
4216 
4217 	if ((vtag == 0) || (vtag > 4095))
4218 		return;
4219 
4220 	CTX_LOCK(ctx);
4221 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4222 	/* Re-init to load the changes */
4223 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4224 		iflib_if_init_locked(ctx);
4225 	CTX_UNLOCK(ctx);
4226 }
4227 
4228 static void
4229 iflib_led_func(void *arg, int onoff)
4230 {
4231 	if_ctx_t ctx = arg;
4232 
4233 	CTX_LOCK(ctx);
4234 	IFDI_LED_FUNC(ctx, onoff);
4235 	CTX_UNLOCK(ctx);
4236 }
4237 
4238 /*********************************************************************
4239  *
4240  *  BUS FUNCTION DEFINITIONS
4241  *
4242  **********************************************************************/
4243 
4244 int
4245 iflib_device_probe(device_t dev)
4246 {
4247 	pci_vendor_info_t *ent;
4248 
4249 	uint16_t	pci_vendor_id, pci_device_id;
4250 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4251 	uint16_t	pci_rev_id;
4252 	if_shared_ctx_t sctx;
4253 
4254 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4255 		return (ENOTSUP);
4256 
4257 	pci_vendor_id = pci_get_vendor(dev);
4258 	pci_device_id = pci_get_device(dev);
4259 	pci_subvendor_id = pci_get_subvendor(dev);
4260 	pci_subdevice_id = pci_get_subdevice(dev);
4261 	pci_rev_id = pci_get_revid(dev);
4262 	if (sctx->isc_parse_devinfo != NULL)
4263 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4264 
4265 	ent = sctx->isc_vendor_info;
4266 	while (ent->pvi_vendor_id != 0) {
4267 		if (pci_vendor_id != ent->pvi_vendor_id) {
4268 			ent++;
4269 			continue;
4270 		}
4271 		if ((pci_device_id == ent->pvi_device_id) &&
4272 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4273 		     (ent->pvi_subvendor_id == 0)) &&
4274 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4275 		     (ent->pvi_subdevice_id == 0)) &&
4276 		    ((pci_rev_id == ent->pvi_rev_id) ||
4277 		     (ent->pvi_rev_id == 0))) {
4278 
4279 			device_set_desc_copy(dev, ent->pvi_name);
4280 			/* this needs to be changed to zero if the bus probing code
4281 			 * ever stops re-probing on best match because the sctx
4282 			 * may have its values over written by register calls
4283 			 * in subsequent probes
4284 			 */
4285 			return (BUS_PROBE_DEFAULT);
4286 		}
4287 		ent++;
4288 	}
4289 	return (ENXIO);
4290 }
4291 
4292 static void
4293 iflib_reset_qvalues(if_ctx_t ctx)
4294 {
4295 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4296 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4297 	device_t dev = ctx->ifc_dev;
4298 	int i;
4299 
4300 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4301 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4302 	/*
4303 	 * XXX sanity check that ntxd & nrxd are a power of 2
4304 	 */
4305 	if (ctx->ifc_sysctl_ntxqs != 0)
4306 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4307 	if (ctx->ifc_sysctl_nrxqs != 0)
4308 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4309 
4310 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4311 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4312 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4313 		else
4314 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4315 	}
4316 
4317 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4318 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4319 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4320 		else
4321 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4322 	}
4323 
4324 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4325 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4326 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4327 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4328 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4329 		}
4330 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4331 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4332 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4333 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4334 		}
4335 	}
4336 
4337 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4338 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4339 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4340 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4341 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4342 		}
4343 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4344 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4345 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4346 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4347 		}
4348 	}
4349 }
4350 
4351 int
4352 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4353 {
4354 	int err, rid, msix;
4355 	if_ctx_t ctx;
4356 	if_t ifp;
4357 	if_softc_ctx_t scctx;
4358 	int i;
4359 	uint16_t main_txq;
4360 	uint16_t main_rxq;
4361 
4362 
4363 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4364 
4365 	if (sc == NULL) {
4366 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4367 		device_set_softc(dev, ctx);
4368 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4369 	}
4370 
4371 	ctx->ifc_sctx = sctx;
4372 	ctx->ifc_dev = dev;
4373 	ctx->ifc_softc = sc;
4374 
4375 	if ((err = iflib_register(ctx)) != 0) {
4376 		device_printf(dev, "iflib_register failed %d\n", err);
4377 		goto fail_ctx_free;
4378 	}
4379 	iflib_add_device_sysctl_pre(ctx);
4380 
4381 	scctx = &ctx->ifc_softc_ctx;
4382 	ifp = ctx->ifc_ifp;
4383 
4384 	iflib_reset_qvalues(ctx);
4385 	CTX_LOCK(ctx);
4386 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4387 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4388 		goto fail_unlock;
4389 	}
4390 	_iflib_pre_assert(scctx);
4391 	ctx->ifc_txrx = *scctx->isc_txrx;
4392 
4393 #ifdef INVARIANTS
4394 	MPASS(scctx->isc_capabilities);
4395 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4396 		MPASS(scctx->isc_tx_csum_flags);
4397 #endif
4398 
4399 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4400 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4401 
4402 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4403 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4404 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4405 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4406 
4407 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4408 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4409 
4410 	/* XXX change for per-queue sizes */
4411 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4412 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4413 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4414 		if (!powerof2(scctx->isc_nrxd[i])) {
4415 			/* round down instead? */
4416 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4417 			err = EINVAL;
4418 			goto fail_iflib_detach;
4419 		}
4420 	}
4421 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4422 		if (!powerof2(scctx->isc_ntxd[i])) {
4423 			device_printf(dev,
4424 			    "# tx descriptors must be a power of 2");
4425 			err = EINVAL;
4426 			goto fail_iflib_detach;
4427 		}
4428 	}
4429 
4430 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4431 	    MAX_SINGLE_PACKET_FRACTION)
4432 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4433 		    MAX_SINGLE_PACKET_FRACTION);
4434 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4435 	    MAX_SINGLE_PACKET_FRACTION)
4436 		scctx->isc_tx_tso_segments_max = max(1,
4437 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4438 
4439 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4440 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4441 		/*
4442 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4443 		 * but some MACs do.
4444 		 */
4445 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4446 		    IP_MAXPACKET));
4447 		/*
4448 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4449 		 * into account.  In the worst case, each of these calls will
4450 		 * add another mbuf and, thus, the requirement for another DMA
4451 		 * segment.  So for best performance, it doesn't make sense to
4452 		 * advertize a maximum of TSO segments that typically will
4453 		 * require defragmentation in iflib_encap().
4454 		 */
4455 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4456 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4457 	}
4458 	if (scctx->isc_rss_table_size == 0)
4459 		scctx->isc_rss_table_size = 64;
4460 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4461 
4462 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4463 	/* XXX format name */
4464 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4465 
4466 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4467 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4468 		device_printf(dev, "Unable to fetch CPU list\n");
4469 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4470 	}
4471 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4472 
4473 	/*
4474 	** Now setup MSI or MSI/X, should
4475 	** return us the number of supported
4476 	** vectors. (Will be 1 for MSI)
4477 	*/
4478 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4479 		msix = scctx->isc_vectors;
4480 	} else if (scctx->isc_msix_bar != 0)
4481 	       /*
4482 		* The simple fact that isc_msix_bar is not 0 does not mean we
4483 		* we have a good value there that is known to work.
4484 		*/
4485 		msix = iflib_msix_init(ctx);
4486 	else {
4487 		scctx->isc_vectors = 1;
4488 		scctx->isc_ntxqsets = 1;
4489 		scctx->isc_nrxqsets = 1;
4490 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4491 		msix = 0;
4492 	}
4493 	/* Get memory for the station queues */
4494 	if ((err = iflib_queues_alloc(ctx))) {
4495 		device_printf(dev, "Unable to allocate queue memory\n");
4496 		goto fail_intr_free;
4497 	}
4498 
4499 	if ((err = iflib_qset_structures_setup(ctx)))
4500 		goto fail_queues;
4501 
4502 	/*
4503 	 * Group taskqueues aren't properly set up until SMP is started,
4504 	 * so we disable interrupts until we can handle them post
4505 	 * SI_SUB_SMP.
4506 	 *
4507 	 * XXX: disabling interrupts doesn't actually work, at least for
4508 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4509 	 * we do null handling and depend on this not causing too large an
4510 	 * interrupt storm.
4511 	 */
4512 	IFDI_INTR_DISABLE(ctx);
4513 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4514 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4515 		goto fail_queues;
4516 	}
4517 	if (msix <= 1) {
4518 		rid = 0;
4519 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4520 			MPASS(msix == 1);
4521 			rid = 1;
4522 		}
4523 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4524 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4525 			goto fail_queues;
4526 		}
4527 	}
4528 
4529 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4530 
4531 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4532 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4533 		goto fail_detach;
4534 	}
4535 
4536 	/*
4537 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4538 	 * This must appear after the call to ether_ifattach() because
4539 	 * ether_ifattach() sets if_hdrlen to the default value.
4540 	 */
4541 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4542 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4543 
4544 	if ((err = iflib_netmap_attach(ctx))) {
4545 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4546 		goto fail_detach;
4547 	}
4548 	*ctxp = ctx;
4549 
4550 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4551 
4552 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4553 	iflib_add_device_sysctl_post(ctx);
4554 	ctx->ifc_flags |= IFC_INIT_DONE;
4555 	CTX_UNLOCK(ctx);
4556 	return (0);
4557 
4558 fail_detach:
4559 	ether_ifdetach(ctx->ifc_ifp);
4560 fail_intr_free:
4561 	iflib_free_intr_mem(ctx);
4562 fail_queues:
4563 	iflib_tx_structures_free(ctx);
4564 	iflib_rx_structures_free(ctx);
4565 fail_iflib_detach:
4566 	IFDI_DETACH(ctx);
4567 fail_unlock:
4568 	CTX_UNLOCK(ctx);
4569 fail_ctx_free:
4570         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4571                 free(ctx->ifc_softc, M_IFLIB);
4572         free(ctx, M_IFLIB);
4573 	return (err);
4574 }
4575 
4576 int
4577 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4578 					  struct iflib_cloneattach_ctx *clctx)
4579 {
4580 	int err;
4581 	if_ctx_t ctx;
4582 	if_t ifp;
4583 	if_softc_ctx_t scctx;
4584 	int i;
4585 	void *sc;
4586 	uint16_t main_txq;
4587 	uint16_t main_rxq;
4588 
4589 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4590 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4591 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4592 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4593 		ctx->ifc_flags |= IFC_PSEUDO;
4594 
4595 	ctx->ifc_sctx = sctx;
4596 	ctx->ifc_softc = sc;
4597 	ctx->ifc_dev = dev;
4598 
4599 	if ((err = iflib_register(ctx)) != 0) {
4600 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4601 		goto fail_ctx_free;
4602 	}
4603 	iflib_add_device_sysctl_pre(ctx);
4604 
4605 	scctx = &ctx->ifc_softc_ctx;
4606 	ifp = ctx->ifc_ifp;
4607 
4608 	/*
4609 	 * XXX sanity check that ntxd & nrxd are a power of 2
4610 	 */
4611 	iflib_reset_qvalues(ctx);
4612 
4613 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4614 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4615 		goto fail_ctx_free;
4616 	}
4617 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4618 		iflib_gen_mac(ctx);
4619 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4620 								clctx->cc_params)) != 0) {
4621 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4622 		goto fail_ctx_free;
4623 	}
4624 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4625 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4626 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4627 
4628 #ifdef INVARIANTS
4629 	MPASS(scctx->isc_capabilities);
4630 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4631 		MPASS(scctx->isc_tx_csum_flags);
4632 #endif
4633 
4634 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4635 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4636 
4637 	ifp->if_flags |= IFF_NOGROUP;
4638 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4639 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4640 
4641 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4642 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4643 			goto fail_detach;
4644 		}
4645 		*ctxp = ctx;
4646 
4647 		/*
4648 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4649 		 * This must appear after the call to ether_ifattach() because
4650 		 * ether_ifattach() sets if_hdrlen to the default value.
4651 		 */
4652 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4653 			if_setifheaderlen(ifp,
4654 			    sizeof(struct ether_vlan_header));
4655 
4656 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4657 		iflib_add_device_sysctl_post(ctx);
4658 		ctx->ifc_flags |= IFC_INIT_DONE;
4659 		return (0);
4660 	}
4661 	_iflib_pre_assert(scctx);
4662 	ctx->ifc_txrx = *scctx->isc_txrx;
4663 
4664 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4665 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4666 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4667 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4668 
4669 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4670 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4671 
4672 	/* XXX change for per-queue sizes */
4673 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4674 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4675 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4676 		if (!powerof2(scctx->isc_nrxd[i])) {
4677 			/* round down instead? */
4678 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4679 			err = EINVAL;
4680 			goto fail_iflib_detach;
4681 		}
4682 	}
4683 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4684 		if (!powerof2(scctx->isc_ntxd[i])) {
4685 			device_printf(dev,
4686 			    "# tx descriptors must be a power of 2");
4687 			err = EINVAL;
4688 			goto fail_iflib_detach;
4689 		}
4690 	}
4691 
4692 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4693 	    MAX_SINGLE_PACKET_FRACTION)
4694 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4695 		    MAX_SINGLE_PACKET_FRACTION);
4696 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4697 	    MAX_SINGLE_PACKET_FRACTION)
4698 		scctx->isc_tx_tso_segments_max = max(1,
4699 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4700 
4701 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4702 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4703 		/*
4704 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4705 		 * but some MACs do.
4706 		 */
4707 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4708 		    IP_MAXPACKET));
4709 		/*
4710 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4711 		 * into account.  In the worst case, each of these calls will
4712 		 * add another mbuf and, thus, the requirement for another DMA
4713 		 * segment.  So for best performance, it doesn't make sense to
4714 		 * advertize a maximum of TSO segments that typically will
4715 		 * require defragmentation in iflib_encap().
4716 		 */
4717 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4718 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4719 	}
4720 	if (scctx->isc_rss_table_size == 0)
4721 		scctx->isc_rss_table_size = 64;
4722 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4723 
4724 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4725 	/* XXX format name */
4726 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4727 
4728 	/* XXX --- can support > 1 -- but keep it simple for now */
4729 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4730 
4731 	/* Get memory for the station queues */
4732 	if ((err = iflib_queues_alloc(ctx))) {
4733 		device_printf(dev, "Unable to allocate queue memory\n");
4734 		goto fail_iflib_detach;
4735 	}
4736 
4737 	if ((err = iflib_qset_structures_setup(ctx))) {
4738 		device_printf(dev, "qset structure setup failed %d\n", err);
4739 		goto fail_queues;
4740 	}
4741 
4742 	/*
4743 	 * XXX What if anything do we want to do about interrupts?
4744 	 */
4745 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4746 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4747 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4748 		goto fail_detach;
4749 	}
4750 
4751 	/*
4752 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4753 	 * This must appear after the call to ether_ifattach() because
4754 	 * ether_ifattach() sets if_hdrlen to the default value.
4755 	 */
4756 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4757 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4758 
4759 	/* XXX handle more than one queue */
4760 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4761 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4762 
4763 	*ctxp = ctx;
4764 
4765 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4766 	iflib_add_device_sysctl_post(ctx);
4767 	ctx->ifc_flags |= IFC_INIT_DONE;
4768 	return (0);
4769 fail_detach:
4770 	ether_ifdetach(ctx->ifc_ifp);
4771 fail_queues:
4772 	iflib_tx_structures_free(ctx);
4773 	iflib_rx_structures_free(ctx);
4774 fail_iflib_detach:
4775 	IFDI_DETACH(ctx);
4776 fail_ctx_free:
4777 	free(ctx->ifc_softc, M_IFLIB);
4778 	free(ctx, M_IFLIB);
4779 	return (err);
4780 }
4781 
4782 int
4783 iflib_pseudo_deregister(if_ctx_t ctx)
4784 {
4785 	if_t ifp = ctx->ifc_ifp;
4786 	iflib_txq_t txq;
4787 	iflib_rxq_t rxq;
4788 	int i, j;
4789 	struct taskqgroup *tqg;
4790 	iflib_fl_t fl;
4791 
4792 	/* Unregister VLAN events */
4793 	if (ctx->ifc_vlan_attach_event != NULL)
4794 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4795 	if (ctx->ifc_vlan_detach_event != NULL)
4796 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4797 
4798 	ether_ifdetach(ifp);
4799 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4800 	CTX_LOCK_DESTROY(ctx);
4801 	/* XXX drain any dependent tasks */
4802 	tqg = qgroup_if_io_tqg;
4803 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4804 		callout_drain(&txq->ift_timer);
4805 		if (txq->ift_task.gt_uniq != NULL)
4806 			taskqgroup_detach(tqg, &txq->ift_task);
4807 	}
4808 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4809 		if (rxq->ifr_task.gt_uniq != NULL)
4810 			taskqgroup_detach(tqg, &rxq->ifr_task);
4811 
4812 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4813 			free(fl->ifl_rx_bitmap, M_IFLIB);
4814 	}
4815 	tqg = qgroup_if_config_tqg;
4816 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4817 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4818 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4819 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4820 
4821 	if_free(ifp);
4822 
4823 	iflib_tx_structures_free(ctx);
4824 	iflib_rx_structures_free(ctx);
4825 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4826 		free(ctx->ifc_softc, M_IFLIB);
4827 	free(ctx, M_IFLIB);
4828 	return (0);
4829 }
4830 
4831 int
4832 iflib_device_attach(device_t dev)
4833 {
4834 	if_ctx_t ctx;
4835 	if_shared_ctx_t sctx;
4836 
4837 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4838 		return (ENOTSUP);
4839 
4840 	pci_enable_busmaster(dev);
4841 
4842 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4843 }
4844 
4845 int
4846 iflib_device_deregister(if_ctx_t ctx)
4847 {
4848 	if_t ifp = ctx->ifc_ifp;
4849 	iflib_txq_t txq;
4850 	iflib_rxq_t rxq;
4851 	device_t dev = ctx->ifc_dev;
4852 	int i, j;
4853 	struct taskqgroup *tqg;
4854 	iflib_fl_t fl;
4855 
4856 	/* Make sure VLANS are not using driver */
4857 	if (if_vlantrunkinuse(ifp)) {
4858 		device_printf(dev, "Vlan in use, detach first\n");
4859 		return (EBUSY);
4860 	}
4861 #ifdef PCI_IOV
4862 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4863 		device_printf(dev, "SR-IOV in use; detach first.\n");
4864 		return (EBUSY);
4865 	}
4866 #endif
4867 
4868 	STATE_LOCK(ctx);
4869 	ctx->ifc_flags |= IFC_IN_DETACH;
4870 	STATE_UNLOCK(ctx);
4871 
4872 	CTX_LOCK(ctx);
4873 	iflib_stop(ctx);
4874 	CTX_UNLOCK(ctx);
4875 
4876 	/* Unregister VLAN events */
4877 	if (ctx->ifc_vlan_attach_event != NULL)
4878 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4879 	if (ctx->ifc_vlan_detach_event != NULL)
4880 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4881 
4882 	iflib_netmap_detach(ifp);
4883 	ether_ifdetach(ifp);
4884 	if (ctx->ifc_led_dev != NULL)
4885 		led_destroy(ctx->ifc_led_dev);
4886 	/* XXX drain any dependent tasks */
4887 	tqg = qgroup_if_io_tqg;
4888 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4889 		callout_drain(&txq->ift_timer);
4890 		if (txq->ift_task.gt_uniq != NULL)
4891 			taskqgroup_detach(tqg, &txq->ift_task);
4892 	}
4893 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4894 		if (rxq->ifr_task.gt_uniq != NULL)
4895 			taskqgroup_detach(tqg, &rxq->ifr_task);
4896 
4897 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4898 			free(fl->ifl_rx_bitmap, M_IFLIB);
4899 
4900 	}
4901 	tqg = qgroup_if_config_tqg;
4902 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4903 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4904 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4905 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4906 	CTX_LOCK(ctx);
4907 	IFDI_DETACH(ctx);
4908 	CTX_UNLOCK(ctx);
4909 
4910 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4911 	CTX_LOCK_DESTROY(ctx);
4912 	device_set_softc(ctx->ifc_dev, NULL);
4913 	iflib_free_intr_mem(ctx);
4914 
4915 	bus_generic_detach(dev);
4916 	if_free(ifp);
4917 
4918 	iflib_tx_structures_free(ctx);
4919 	iflib_rx_structures_free(ctx);
4920 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4921 		free(ctx->ifc_softc, M_IFLIB);
4922 	STATE_LOCK_DESTROY(ctx);
4923 	free(ctx, M_IFLIB);
4924 	return (0);
4925 }
4926 
4927 static void
4928 iflib_free_intr_mem(if_ctx_t ctx)
4929 {
4930 
4931 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4932 		pci_release_msi(ctx->ifc_dev);
4933 	}
4934 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4935 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4936 	}
4937 	if (ctx->ifc_msix_mem != NULL) {
4938 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4939 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4940 		ctx->ifc_msix_mem = NULL;
4941 	}
4942 }
4943 
4944 int
4945 iflib_device_detach(device_t dev)
4946 {
4947 	if_ctx_t ctx = device_get_softc(dev);
4948 
4949 	return (iflib_device_deregister(ctx));
4950 }
4951 
4952 int
4953 iflib_device_suspend(device_t dev)
4954 {
4955 	if_ctx_t ctx = device_get_softc(dev);
4956 
4957 	CTX_LOCK(ctx);
4958 	IFDI_SUSPEND(ctx);
4959 	CTX_UNLOCK(ctx);
4960 
4961 	return bus_generic_suspend(dev);
4962 }
4963 int
4964 iflib_device_shutdown(device_t dev)
4965 {
4966 	if_ctx_t ctx = device_get_softc(dev);
4967 
4968 	CTX_LOCK(ctx);
4969 	IFDI_SHUTDOWN(ctx);
4970 	CTX_UNLOCK(ctx);
4971 
4972 	return bus_generic_suspend(dev);
4973 }
4974 
4975 
4976 int
4977 iflib_device_resume(device_t dev)
4978 {
4979 	if_ctx_t ctx = device_get_softc(dev);
4980 	iflib_txq_t txq = ctx->ifc_txqs;
4981 
4982 	CTX_LOCK(ctx);
4983 	IFDI_RESUME(ctx);
4984 	iflib_if_init_locked(ctx);
4985 	CTX_UNLOCK(ctx);
4986 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4987 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4988 
4989 	return (bus_generic_resume(dev));
4990 }
4991 
4992 int
4993 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4994 {
4995 	int error;
4996 	if_ctx_t ctx = device_get_softc(dev);
4997 
4998 	CTX_LOCK(ctx);
4999 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5000 	CTX_UNLOCK(ctx);
5001 
5002 	return (error);
5003 }
5004 
5005 void
5006 iflib_device_iov_uninit(device_t dev)
5007 {
5008 	if_ctx_t ctx = device_get_softc(dev);
5009 
5010 	CTX_LOCK(ctx);
5011 	IFDI_IOV_UNINIT(ctx);
5012 	CTX_UNLOCK(ctx);
5013 }
5014 
5015 int
5016 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5017 {
5018 	int error;
5019 	if_ctx_t ctx = device_get_softc(dev);
5020 
5021 	CTX_LOCK(ctx);
5022 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5023 	CTX_UNLOCK(ctx);
5024 
5025 	return (error);
5026 }
5027 
5028 /*********************************************************************
5029  *
5030  *  MODULE FUNCTION DEFINITIONS
5031  *
5032  **********************************************************************/
5033 
5034 /*
5035  * - Start a fast taskqueue thread for each core
5036  * - Start a taskqueue for control operations
5037  */
5038 static int
5039 iflib_module_init(void)
5040 {
5041 	return (0);
5042 }
5043 
5044 static int
5045 iflib_module_event_handler(module_t mod, int what, void *arg)
5046 {
5047 	int err;
5048 
5049 	switch (what) {
5050 	case MOD_LOAD:
5051 		if ((err = iflib_module_init()) != 0)
5052 			return (err);
5053 		break;
5054 	case MOD_UNLOAD:
5055 		return (EBUSY);
5056 	default:
5057 		return (EOPNOTSUPP);
5058 	}
5059 
5060 	return (0);
5061 }
5062 
5063 /*********************************************************************
5064  *
5065  *  PUBLIC FUNCTION DEFINITIONS
5066  *     ordered as in iflib.h
5067  *
5068  **********************************************************************/
5069 
5070 
5071 static void
5072 _iflib_assert(if_shared_ctx_t sctx)
5073 {
5074 	MPASS(sctx->isc_tx_maxsize);
5075 	MPASS(sctx->isc_tx_maxsegsize);
5076 
5077 	MPASS(sctx->isc_rx_maxsize);
5078 	MPASS(sctx->isc_rx_nsegments);
5079 	MPASS(sctx->isc_rx_maxsegsize);
5080 
5081 	MPASS(sctx->isc_nrxd_min[0]);
5082 	MPASS(sctx->isc_nrxd_max[0]);
5083 	MPASS(sctx->isc_nrxd_default[0]);
5084 	MPASS(sctx->isc_ntxd_min[0]);
5085 	MPASS(sctx->isc_ntxd_max[0]);
5086 	MPASS(sctx->isc_ntxd_default[0]);
5087 }
5088 
5089 static void
5090 _iflib_pre_assert(if_softc_ctx_t scctx)
5091 {
5092 
5093 	MPASS(scctx->isc_txrx->ift_txd_encap);
5094 	MPASS(scctx->isc_txrx->ift_txd_flush);
5095 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5096 	MPASS(scctx->isc_txrx->ift_rxd_available);
5097 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5098 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5099 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5100 }
5101 
5102 static int
5103 iflib_register(if_ctx_t ctx)
5104 {
5105 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5106 	driver_t *driver = sctx->isc_driver;
5107 	device_t dev = ctx->ifc_dev;
5108 	if_t ifp;
5109 
5110 	_iflib_assert(sctx);
5111 
5112 	CTX_LOCK_INIT(ctx);
5113 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5114 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5115 	if (ifp == NULL) {
5116 		device_printf(dev, "can not allocate ifnet structure\n");
5117 		return (ENOMEM);
5118 	}
5119 
5120 	/*
5121 	 * Initialize our context's device specific methods
5122 	 */
5123 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5124 	kobj_class_compile((kobj_class_t) driver);
5125 	driver->refs++;
5126 
5127 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5128 	if_setsoftc(ifp, ctx);
5129 	if_setdev(ifp, dev);
5130 	if_setinitfn(ifp, iflib_if_init);
5131 	if_setioctlfn(ifp, iflib_if_ioctl);
5132 #ifdef ALTQ
5133 	if_setstartfn(ifp, iflib_altq_if_start);
5134 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5135 	if_setsendqready(ifp);
5136 #else
5137 	if_settransmitfn(ifp, iflib_if_transmit);
5138 #endif
5139 	if_setqflushfn(ifp, iflib_if_qflush);
5140 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5141 
5142 	ctx->ifc_vlan_attach_event =
5143 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5144 							  EVENTHANDLER_PRI_FIRST);
5145 	ctx->ifc_vlan_detach_event =
5146 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5147 							  EVENTHANDLER_PRI_FIRST);
5148 
5149 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5150 					 iflib_media_change, iflib_media_status);
5151 
5152 	return (0);
5153 }
5154 
5155 
5156 static int
5157 iflib_queues_alloc(if_ctx_t ctx)
5158 {
5159 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5160 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5161 	device_t dev = ctx->ifc_dev;
5162 	int nrxqsets = scctx->isc_nrxqsets;
5163 	int ntxqsets = scctx->isc_ntxqsets;
5164 	iflib_txq_t txq;
5165 	iflib_rxq_t rxq;
5166 	iflib_fl_t fl = NULL;
5167 	int i, j, cpu, err, txconf, rxconf;
5168 	iflib_dma_info_t ifdip;
5169 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5170 	uint32_t *txqsizes = scctx->isc_txqsizes;
5171 	uint8_t nrxqs = sctx->isc_nrxqs;
5172 	uint8_t ntxqs = sctx->isc_ntxqs;
5173 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5174 	caddr_t *vaddrs;
5175 	uint64_t *paddrs;
5176 
5177 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5178 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5179 
5180 	/* Allocate the TX ring struct memory */
5181 	if (!(ctx->ifc_txqs =
5182 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5183 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5184 		device_printf(dev, "Unable to allocate TX ring memory\n");
5185 		err = ENOMEM;
5186 		goto fail;
5187 	}
5188 
5189 	/* Now allocate the RX */
5190 	if (!(ctx->ifc_rxqs =
5191 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5192 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5193 		device_printf(dev, "Unable to allocate RX ring memory\n");
5194 		err = ENOMEM;
5195 		goto rx_fail;
5196 	}
5197 
5198 	txq = ctx->ifc_txqs;
5199 	rxq = ctx->ifc_rxqs;
5200 
5201 	/*
5202 	 * XXX handle allocation failure
5203 	 */
5204 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5205 		/* Set up some basics */
5206 
5207 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5208 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5209 			err = ENOMEM;
5210 			goto err_tx_desc;
5211 		}
5212 		txq->ift_ifdi = ifdip;
5213 		for (j = 0; j < ntxqs; j++, ifdip++) {
5214 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5215 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5216 				err = ENOMEM;
5217 				goto err_tx_desc;
5218 			}
5219 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5220 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5221 		}
5222 		txq->ift_ctx = ctx;
5223 		txq->ift_id = i;
5224 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5225 			txq->ift_br_offset = 1;
5226 		} else {
5227 			txq->ift_br_offset = 0;
5228 		}
5229 		/* XXX fix this */
5230 		txq->ift_timer.c_cpu = cpu;
5231 
5232 		if (iflib_txsd_alloc(txq)) {
5233 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5234 			err = ENOMEM;
5235 			goto err_tx_desc;
5236 		}
5237 
5238 		/* Initialize the TX lock */
5239 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5240 		    device_get_nameunit(dev), txq->ift_id);
5241 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5242 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5243 
5244 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5245 			 device_get_nameunit(dev), txq->ift_id);
5246 
5247 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5248 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5249 		if (err) {
5250 			/* XXX free any allocated rings */
5251 			device_printf(dev, "Unable to allocate buf_ring\n");
5252 			goto err_tx_desc;
5253 		}
5254 	}
5255 
5256 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5257 		/* Set up some basics */
5258 
5259 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5260 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5261 			err = ENOMEM;
5262 			goto err_tx_desc;
5263 		}
5264 
5265 		rxq->ifr_ifdi = ifdip;
5266 		/* XXX this needs to be changed if #rx queues != #tx queues */
5267 		rxq->ifr_ntxqirq = 1;
5268 		rxq->ifr_txqid[0] = i;
5269 		for (j = 0; j < nrxqs; j++, ifdip++) {
5270 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5271 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5272 				err = ENOMEM;
5273 				goto err_tx_desc;
5274 			}
5275 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5276 		}
5277 		rxq->ifr_ctx = ctx;
5278 		rxq->ifr_id = i;
5279 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5280 			rxq->ifr_fl_offset = 1;
5281 		} else {
5282 			rxq->ifr_fl_offset = 0;
5283 		}
5284 		rxq->ifr_nfl = nfree_lists;
5285 		if (!(fl =
5286 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5287 			device_printf(dev, "Unable to allocate free list memory\n");
5288 			err = ENOMEM;
5289 			goto err_tx_desc;
5290 		}
5291 		rxq->ifr_fl = fl;
5292 		for (j = 0; j < nfree_lists; j++) {
5293 			fl[j].ifl_rxq = rxq;
5294 			fl[j].ifl_id = j;
5295 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5296 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5297 		}
5298 		/* Allocate receive buffers for the ring */
5299 		if (iflib_rxsd_alloc(rxq)) {
5300 			device_printf(dev,
5301 			    "Critical Failure setting up receive buffers\n");
5302 			err = ENOMEM;
5303 			goto err_rx_desc;
5304 		}
5305 
5306 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5307 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5308 	}
5309 
5310 	/* TXQs */
5311 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5312 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5313 	for (i = 0; i < ntxqsets; i++) {
5314 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5315 
5316 		for (j = 0; j < ntxqs; j++, di++) {
5317 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5318 			paddrs[i*ntxqs + j] = di->idi_paddr;
5319 		}
5320 	}
5321 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5322 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5323 		iflib_tx_structures_free(ctx);
5324 		free(vaddrs, M_IFLIB);
5325 		free(paddrs, M_IFLIB);
5326 		goto err_rx_desc;
5327 	}
5328 	free(vaddrs, M_IFLIB);
5329 	free(paddrs, M_IFLIB);
5330 
5331 	/* RXQs */
5332 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5333 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5334 	for (i = 0; i < nrxqsets; i++) {
5335 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5336 
5337 		for (j = 0; j < nrxqs; j++, di++) {
5338 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5339 			paddrs[i*nrxqs + j] = di->idi_paddr;
5340 		}
5341 	}
5342 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5343 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5344 		iflib_tx_structures_free(ctx);
5345 		free(vaddrs, M_IFLIB);
5346 		free(paddrs, M_IFLIB);
5347 		goto err_rx_desc;
5348 	}
5349 	free(vaddrs, M_IFLIB);
5350 	free(paddrs, M_IFLIB);
5351 
5352 	return (0);
5353 
5354 /* XXX handle allocation failure changes */
5355 err_rx_desc:
5356 err_tx_desc:
5357 rx_fail:
5358 	if (ctx->ifc_rxqs != NULL)
5359 		free(ctx->ifc_rxqs, M_IFLIB);
5360 	ctx->ifc_rxqs = NULL;
5361 	if (ctx->ifc_txqs != NULL)
5362 		free(ctx->ifc_txqs, M_IFLIB);
5363 	ctx->ifc_txqs = NULL;
5364 fail:
5365 	return (err);
5366 }
5367 
5368 static int
5369 iflib_tx_structures_setup(if_ctx_t ctx)
5370 {
5371 	iflib_txq_t txq = ctx->ifc_txqs;
5372 	int i;
5373 
5374 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5375 		iflib_txq_setup(txq);
5376 
5377 	return (0);
5378 }
5379 
5380 static void
5381 iflib_tx_structures_free(if_ctx_t ctx)
5382 {
5383 	iflib_txq_t txq = ctx->ifc_txqs;
5384 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5385 	int i, j;
5386 
5387 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5388 		iflib_txq_destroy(txq);
5389 		for (j = 0; j < sctx->isc_ntxqs; j++)
5390 			iflib_dma_free(&txq->ift_ifdi[j]);
5391 	}
5392 	free(ctx->ifc_txqs, M_IFLIB);
5393 	ctx->ifc_txqs = NULL;
5394 	IFDI_QUEUES_FREE(ctx);
5395 }
5396 
5397 /*********************************************************************
5398  *
5399  *  Initialize all receive rings.
5400  *
5401  **********************************************************************/
5402 static int
5403 iflib_rx_structures_setup(if_ctx_t ctx)
5404 {
5405 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5406 	int q;
5407 #if defined(INET6) || defined(INET)
5408 	int i, err;
5409 #endif
5410 
5411 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5412 #if defined(INET6) || defined(INET)
5413 		tcp_lro_free(&rxq->ifr_lc);
5414 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5415 		    TCP_LRO_ENTRIES, min(1024,
5416 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5417 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5418 			goto fail;
5419 		}
5420 		rxq->ifr_lro_enabled = TRUE;
5421 #endif
5422 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5423 	}
5424 	return (0);
5425 #if defined(INET6) || defined(INET)
5426 fail:
5427 	/*
5428 	 * Free RX software descriptors allocated so far, we will only handle
5429 	 * the rings that completed, the failing case will have
5430 	 * cleaned up for itself. 'q' failed, so its the terminus.
5431 	 */
5432 	rxq = ctx->ifc_rxqs;
5433 	for (i = 0; i < q; ++i, rxq++) {
5434 		iflib_rx_sds_free(rxq);
5435 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5436 	}
5437 	return (err);
5438 #endif
5439 }
5440 
5441 /*********************************************************************
5442  *
5443  *  Free all receive rings.
5444  *
5445  **********************************************************************/
5446 static void
5447 iflib_rx_structures_free(if_ctx_t ctx)
5448 {
5449 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5450 
5451 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5452 		iflib_rx_sds_free(rxq);
5453 	}
5454 	free(ctx->ifc_rxqs, M_IFLIB);
5455 	ctx->ifc_rxqs = NULL;
5456 }
5457 
5458 static int
5459 iflib_qset_structures_setup(if_ctx_t ctx)
5460 {
5461 	int err;
5462 
5463 	/*
5464 	 * It is expected that the caller takes care of freeing queues if this
5465 	 * fails.
5466 	 */
5467 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5468 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5469 		return (err);
5470 	}
5471 
5472 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5473 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5474 
5475 	return (err);
5476 }
5477 
5478 int
5479 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5480 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5481 {
5482 
5483 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5484 }
5485 
5486 #ifdef SMP
5487 static int
5488 find_nth(if_ctx_t ctx, int qid)
5489 {
5490 	cpuset_t cpus;
5491 	int i, cpuid, eqid, count;
5492 
5493 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5494 	count = CPU_COUNT(&cpus);
5495 	eqid = qid % count;
5496 	/* clear up to the qid'th bit */
5497 	for (i = 0; i < eqid; i++) {
5498 		cpuid = CPU_FFS(&cpus);
5499 		MPASS(cpuid != 0);
5500 		CPU_CLR(cpuid-1, &cpus);
5501 	}
5502 	cpuid = CPU_FFS(&cpus);
5503 	MPASS(cpuid != 0);
5504 	return (cpuid-1);
5505 }
5506 
5507 #ifdef SCHED_ULE
5508 extern struct cpu_group *cpu_top;              /* CPU topology */
5509 
5510 static int
5511 find_child_with_core(int cpu, struct cpu_group *grp)
5512 {
5513 	int i;
5514 
5515 	if (grp->cg_children == 0)
5516 		return -1;
5517 
5518 	MPASS(grp->cg_child);
5519 	for (i = 0; i < grp->cg_children; i++) {
5520 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5521 			return i;
5522 	}
5523 
5524 	return -1;
5525 }
5526 
5527 /*
5528  * Find the nth "close" core to the specified core
5529  * "close" is defined as the deepest level that shares
5530  * at least an L2 cache.  With threads, this will be
5531  * threads on the same core.  If the sahred cache is L3
5532  * or higher, simply returns the same core.
5533  */
5534 static int
5535 find_close_core(int cpu, int core_offset)
5536 {
5537 	struct cpu_group *grp;
5538 	int i;
5539 	int fcpu;
5540 	cpuset_t cs;
5541 
5542 	grp = cpu_top;
5543 	if (grp == NULL)
5544 		return cpu;
5545 	i = 0;
5546 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5547 		/* If the child only has one cpu, don't descend */
5548 		if (grp->cg_child[i].cg_count <= 1)
5549 			break;
5550 		grp = &grp->cg_child[i];
5551 	}
5552 
5553 	/* If they don't share at least an L2 cache, use the same CPU */
5554 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5555 		return cpu;
5556 
5557 	/* Now pick one */
5558 	CPU_COPY(&grp->cg_mask, &cs);
5559 
5560 	/* Add the selected CPU offset to core offset. */
5561 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5562 		if (fcpu - 1 == cpu)
5563 			break;
5564 		CPU_CLR(fcpu - 1, &cs);
5565 	}
5566 	MPASS(fcpu);
5567 
5568 	core_offset += i;
5569 
5570 	CPU_COPY(&grp->cg_mask, &cs);
5571 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5572 		MPASS(CPU_FFS(&cs));
5573 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5574 	}
5575 	MPASS(CPU_FFS(&cs));
5576 	return CPU_FFS(&cs) - 1;
5577 }
5578 #else
5579 static int
5580 find_close_core(int cpu, int core_offset __unused)
5581 {
5582 	return cpu;
5583 }
5584 #endif
5585 
5586 static int
5587 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5588 {
5589 	switch (type) {
5590 	case IFLIB_INTR_TX:
5591 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5592 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5593 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5594 	case IFLIB_INTR_RX:
5595 	case IFLIB_INTR_RXTX:
5596 		/* RX queues get the specified core */
5597 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5598 	default:
5599 		return -1;
5600 	}
5601 }
5602 #else
5603 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5604 #define find_close_core(cpuid, tid)	CPU_FIRST()
5605 #define find_nth(ctx, gid)		CPU_FIRST()
5606 #endif
5607 
5608 /* Just to avoid copy/paste */
5609 static inline int
5610 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5611     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5612 {
5613 	int cpuid;
5614 	int err, tid;
5615 
5616 	cpuid = find_nth(ctx, qid);
5617 	tid = get_core_offset(ctx, type, qid);
5618 	MPASS(tid >= 0);
5619 	cpuid = find_close_core(cpuid, tid);
5620 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5621 	if (err) {
5622 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5623 		return (err);
5624 	}
5625 #ifdef notyet
5626 	if (cpuid > ctx->ifc_cpuid_highest)
5627 		ctx->ifc_cpuid_highest = cpuid;
5628 #endif
5629 	return 0;
5630 }
5631 
5632 int
5633 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5634 			iflib_intr_type_t type, driver_filter_t *filter,
5635 			void *filter_arg, int qid, const char *name)
5636 {
5637 	struct grouptask *gtask;
5638 	struct taskqgroup *tqg;
5639 	iflib_filter_info_t info;
5640 	gtask_fn_t *fn;
5641 	int tqrid, err;
5642 	driver_filter_t *intr_fast;
5643 	void *q;
5644 
5645 	info = &ctx->ifc_filter_info;
5646 	tqrid = rid;
5647 
5648 	switch (type) {
5649 	/* XXX merge tx/rx for netmap? */
5650 	case IFLIB_INTR_TX:
5651 		q = &ctx->ifc_txqs[qid];
5652 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5653 		gtask = &ctx->ifc_txqs[qid].ift_task;
5654 		tqg = qgroup_if_io_tqg;
5655 		fn = _task_fn_tx;
5656 		intr_fast = iflib_fast_intr;
5657 		GROUPTASK_INIT(gtask, 0, fn, q);
5658 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5659 		break;
5660 	case IFLIB_INTR_RX:
5661 		q = &ctx->ifc_rxqs[qid];
5662 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5663 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5664 		tqg = qgroup_if_io_tqg;
5665 		fn = _task_fn_rx;
5666 		intr_fast = iflib_fast_intr;
5667 		GROUPTASK_INIT(gtask, 0, fn, q);
5668 		break;
5669 	case IFLIB_INTR_RXTX:
5670 		q = &ctx->ifc_rxqs[qid];
5671 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5672 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5673 		tqg = qgroup_if_io_tqg;
5674 		fn = _task_fn_rx;
5675 		intr_fast = iflib_fast_intr_rxtx;
5676 		GROUPTASK_INIT(gtask, 0, fn, q);
5677 		break;
5678 	case IFLIB_INTR_ADMIN:
5679 		q = ctx;
5680 		tqrid = -1;
5681 		info = &ctx->ifc_filter_info;
5682 		gtask = &ctx->ifc_admin_task;
5683 		tqg = qgroup_if_config_tqg;
5684 		fn = _task_fn_admin;
5685 		intr_fast = iflib_fast_intr_ctx;
5686 		break;
5687 	default:
5688 		panic("unknown net intr type");
5689 	}
5690 
5691 	info->ifi_filter = filter;
5692 	info->ifi_filter_arg = filter_arg;
5693 	info->ifi_task = gtask;
5694 	info->ifi_ctx = q;
5695 
5696 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5697 	if (err != 0) {
5698 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5699 		return (err);
5700 	}
5701 	if (type == IFLIB_INTR_ADMIN)
5702 		return (0);
5703 
5704 	if (tqrid != -1) {
5705 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5706 		if (err)
5707 			return (err);
5708 	} else {
5709 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5710 	}
5711 
5712 	return (0);
5713 }
5714 
5715 void
5716 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5717 {
5718 	struct grouptask *gtask;
5719 	struct taskqgroup *tqg;
5720 	gtask_fn_t *fn;
5721 	void *q;
5722 	int irq_num = -1;
5723 	int err;
5724 
5725 	switch (type) {
5726 	case IFLIB_INTR_TX:
5727 		q = &ctx->ifc_txqs[qid];
5728 		gtask = &ctx->ifc_txqs[qid].ift_task;
5729 		tqg = qgroup_if_io_tqg;
5730 		fn = _task_fn_tx;
5731 		if (irq != NULL)
5732 			irq_num = rman_get_start(irq->ii_res);
5733 		break;
5734 	case IFLIB_INTR_RX:
5735 		q = &ctx->ifc_rxqs[qid];
5736 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5737 		tqg = qgroup_if_io_tqg;
5738 		fn = _task_fn_rx;
5739 		if (irq != NULL)
5740 			irq_num = rman_get_start(irq->ii_res);
5741 		break;
5742 	case IFLIB_INTR_IOV:
5743 		q = ctx;
5744 		gtask = &ctx->ifc_vflr_task;
5745 		tqg = qgroup_if_config_tqg;
5746 		fn = _task_fn_iov;
5747 		break;
5748 	default:
5749 		panic("unknown net intr type");
5750 	}
5751 	GROUPTASK_INIT(gtask, 0, fn, q);
5752 	if (irq_num != -1) {
5753 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5754 		if (err)
5755 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5756 	}
5757 	else {
5758 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5759 	}
5760 }
5761 
5762 void
5763 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5764 {
5765 	if (irq->ii_tag)
5766 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5767 
5768 	if (irq->ii_res)
5769 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5770 }
5771 
5772 static int
5773 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5774 {
5775 	iflib_txq_t txq = ctx->ifc_txqs;
5776 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5777 	if_irq_t irq = &ctx->ifc_legacy_irq;
5778 	iflib_filter_info_t info;
5779 	struct grouptask *gtask;
5780 	struct taskqgroup *tqg;
5781 	gtask_fn_t *fn;
5782 	int tqrid;
5783 	void *q;
5784 	int err;
5785 
5786 	q = &ctx->ifc_rxqs[0];
5787 	info = &rxq[0].ifr_filter_info;
5788 	gtask = &rxq[0].ifr_task;
5789 	tqg = qgroup_if_io_tqg;
5790 	tqrid = irq->ii_rid = *rid;
5791 	fn = _task_fn_rx;
5792 
5793 	ctx->ifc_flags |= IFC_LEGACY;
5794 	info->ifi_filter = filter;
5795 	info->ifi_filter_arg = filter_arg;
5796 	info->ifi_task = gtask;
5797 	info->ifi_ctx = ctx;
5798 
5799 	/* We allocate a single interrupt resource */
5800 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5801 		return (err);
5802 	GROUPTASK_INIT(gtask, 0, fn, q);
5803 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5804 
5805 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5806 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5807 	return (0);
5808 }
5809 
5810 void
5811 iflib_led_create(if_ctx_t ctx)
5812 {
5813 
5814 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5815 	    device_get_nameunit(ctx->ifc_dev));
5816 }
5817 
5818 void
5819 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5820 {
5821 
5822 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5823 }
5824 
5825 void
5826 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5827 {
5828 
5829 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5830 }
5831 
5832 void
5833 iflib_admin_intr_deferred(if_ctx_t ctx)
5834 {
5835 #ifdef INVARIANTS
5836 	struct grouptask *gtask;
5837 
5838 	gtask = &ctx->ifc_admin_task;
5839 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5840 #endif
5841 
5842 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5843 }
5844 
5845 void
5846 iflib_iov_intr_deferred(if_ctx_t ctx)
5847 {
5848 
5849 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5850 }
5851 
5852 void
5853 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5854 {
5855 
5856 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5857 }
5858 
5859 void
5860 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5861 	const char *name)
5862 {
5863 
5864 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5865 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5866 }
5867 
5868 void
5869 iflib_config_gtask_deinit(struct grouptask *gtask)
5870 {
5871 
5872 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5873 }
5874 
5875 void
5876 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5877 {
5878 	if_t ifp = ctx->ifc_ifp;
5879 	iflib_txq_t txq = ctx->ifc_txqs;
5880 
5881 	if_setbaudrate(ifp, baudrate);
5882 	if (baudrate >= IF_Gbps(10)) {
5883 		STATE_LOCK(ctx);
5884 		ctx->ifc_flags |= IFC_PREFETCH;
5885 		STATE_UNLOCK(ctx);
5886 	}
5887 	/* If link down, disable watchdog */
5888 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5889 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5890 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5891 	}
5892 	ctx->ifc_link_state = link_state;
5893 	if_link_state_change(ifp, link_state);
5894 }
5895 
5896 static int
5897 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5898 {
5899 	int credits;
5900 #ifdef INVARIANTS
5901 	int credits_pre = txq->ift_cidx_processed;
5902 #endif
5903 
5904 	if (ctx->isc_txd_credits_update == NULL)
5905 		return (0);
5906 
5907 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5908 	    BUS_DMASYNC_POSTREAD);
5909 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5910 		return (0);
5911 
5912 	txq->ift_processed += credits;
5913 	txq->ift_cidx_processed += credits;
5914 
5915 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5916 	if (txq->ift_cidx_processed >= txq->ift_size)
5917 		txq->ift_cidx_processed -= txq->ift_size;
5918 	return (credits);
5919 }
5920 
5921 static int
5922 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5923 {
5924 
5925 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5926 	    budget));
5927 }
5928 
5929 void
5930 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5931 	const char *description, if_int_delay_info_t info,
5932 	int offset, int value)
5933 {
5934 	info->iidi_ctx = ctx;
5935 	info->iidi_offset = offset;
5936 	info->iidi_value = value;
5937 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5938 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5939 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5940 	    info, 0, iflib_sysctl_int_delay, "I", description);
5941 }
5942 
5943 struct sx *
5944 iflib_ctx_lock_get(if_ctx_t ctx)
5945 {
5946 
5947 	return (&ctx->ifc_ctx_sx);
5948 }
5949 
5950 static int
5951 iflib_msix_init(if_ctx_t ctx)
5952 {
5953 	device_t dev = ctx->ifc_dev;
5954 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5955 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5956 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5957 	int iflib_num_tx_queues, iflib_num_rx_queues;
5958 	int err, admincnt, bar;
5959 
5960 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5961 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5962 
5963 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5964 
5965 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5966 	admincnt = sctx->isc_admin_intrcnt;
5967 	/* Override by tuneable */
5968 	if (scctx->isc_disable_msix)
5969 		goto msi;
5970 
5971 	/*
5972 	 * bar == -1 => "trust me I know what I'm doing"
5973 	 * Some drivers are for hardware that is so shoddily
5974 	 * documented that no one knows which bars are which
5975 	 * so the developer has to map all bars. This hack
5976 	 * allows shoddy garbage to use msix in this framework.
5977 	 */
5978 	if (bar != -1) {
5979 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5980 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5981 		if (ctx->ifc_msix_mem == NULL) {
5982 			/* May not be enabled */
5983 			device_printf(dev, "Unable to map MSIX table \n");
5984 			goto msi;
5985 		}
5986 	}
5987 	/* First try MSI/X */
5988 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5989 		device_printf(dev, "System has MSIX disabled \n");
5990 		bus_release_resource(dev, SYS_RES_MEMORY,
5991 		    bar, ctx->ifc_msix_mem);
5992 		ctx->ifc_msix_mem = NULL;
5993 		goto msi;
5994 	}
5995 #if IFLIB_DEBUG
5996 	/* use only 1 qset in debug mode */
5997 	queuemsgs = min(msgs - admincnt, 1);
5998 #else
5999 	queuemsgs = msgs - admincnt;
6000 #endif
6001 #ifdef RSS
6002 	queues = imin(queuemsgs, rss_getnumbuckets());
6003 #else
6004 	queues = queuemsgs;
6005 #endif
6006 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6007 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6008 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6009 #ifdef  RSS
6010 	/* If we're doing RSS, clamp at the number of RSS buckets */
6011 	if (queues > rss_getnumbuckets())
6012 		queues = rss_getnumbuckets();
6013 #endif
6014 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6015 		rx_queues = iflib_num_rx_queues;
6016 	else
6017 		rx_queues = queues;
6018 
6019 	if (rx_queues > scctx->isc_nrxqsets)
6020 		rx_queues = scctx->isc_nrxqsets;
6021 
6022 	/*
6023 	 * We want this to be all logical CPUs by default
6024 	 */
6025 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6026 		tx_queues = iflib_num_tx_queues;
6027 	else
6028 		tx_queues = mp_ncpus;
6029 
6030 	if (tx_queues > scctx->isc_ntxqsets)
6031 		tx_queues = scctx->isc_ntxqsets;
6032 
6033 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6034 #ifdef INVARIANTS
6035 		if (tx_queues != rx_queues)
6036 			device_printf(dev,
6037 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6038 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6039 #endif
6040 		tx_queues = min(rx_queues, tx_queues);
6041 		rx_queues = min(rx_queues, tx_queues);
6042 	}
6043 
6044 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6045 
6046 	vectors = rx_queues + admincnt;
6047 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6048 		device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors);
6049 		scctx->isc_vectors = vectors;
6050 		scctx->isc_nrxqsets = rx_queues;
6051 		scctx->isc_ntxqsets = tx_queues;
6052 		scctx->isc_intr = IFLIB_INTR_MSIX;
6053 
6054 		return (vectors);
6055 	} else {
6056 		device_printf(dev,
6057 		    "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6058 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6059 		    ctx->ifc_msix_mem);
6060 		ctx->ifc_msix_mem = NULL;
6061 	}
6062 msi:
6063 	vectors = pci_msi_count(dev);
6064 	scctx->isc_nrxqsets = 1;
6065 	scctx->isc_ntxqsets = 1;
6066 	scctx->isc_vectors = vectors;
6067 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6068 		device_printf(dev,"Using an MSI interrupt\n");
6069 		scctx->isc_intr = IFLIB_INTR_MSI;
6070 	} else {
6071 		scctx->isc_vectors = 1;
6072 		device_printf(dev,"Using a Legacy interrupt\n");
6073 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6074 	}
6075 
6076 	return (vectors);
6077 }
6078 
6079 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6080 
6081 static int
6082 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6083 {
6084 	int rc;
6085 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6086 	struct sbuf *sb;
6087 	const char *ring_state = "UNKNOWN";
6088 
6089 	/* XXX needed ? */
6090 	rc = sysctl_wire_old_buffer(req, 0);
6091 	MPASS(rc == 0);
6092 	if (rc != 0)
6093 		return (rc);
6094 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6095 	MPASS(sb != NULL);
6096 	if (sb == NULL)
6097 		return (ENOMEM);
6098 	if (state[3] <= 3)
6099 		ring_state = ring_states[state[3]];
6100 
6101 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6102 		    state[0], state[1], state[2], ring_state);
6103 	rc = sbuf_finish(sb);
6104 	sbuf_delete(sb);
6105         return(rc);
6106 }
6107 
6108 enum iflib_ndesc_handler {
6109 	IFLIB_NTXD_HANDLER,
6110 	IFLIB_NRXD_HANDLER,
6111 };
6112 
6113 static int
6114 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6115 {
6116 	if_ctx_t ctx = (void *)arg1;
6117 	enum iflib_ndesc_handler type = arg2;
6118 	char buf[256] = {0};
6119 	qidx_t *ndesc;
6120 	char *p, *next;
6121 	int nqs, rc, i;
6122 
6123 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6124 
6125 	nqs = 8;
6126 	switch(type) {
6127 	case IFLIB_NTXD_HANDLER:
6128 		ndesc = ctx->ifc_sysctl_ntxds;
6129 		if (ctx->ifc_sctx)
6130 			nqs = ctx->ifc_sctx->isc_ntxqs;
6131 		break;
6132 	case IFLIB_NRXD_HANDLER:
6133 		ndesc = ctx->ifc_sysctl_nrxds;
6134 		if (ctx->ifc_sctx)
6135 			nqs = ctx->ifc_sctx->isc_nrxqs;
6136 		break;
6137 	default:
6138 			panic("unhandled type");
6139 	}
6140 	if (nqs == 0)
6141 		nqs = 8;
6142 
6143 	for (i=0; i<8; i++) {
6144 		if (i >= nqs)
6145 			break;
6146 		if (i)
6147 			strcat(buf, ",");
6148 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6149 	}
6150 
6151 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6152 	if (rc || req->newptr == NULL)
6153 		return rc;
6154 
6155 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6156 	    i++, p = strsep(&next, " ,")) {
6157 		ndesc[i] = strtoul(p, NULL, 10);
6158 	}
6159 
6160 	return(rc);
6161 }
6162 
6163 #define NAME_BUFLEN 32
6164 static void
6165 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6166 {
6167         device_t dev = iflib_get_dev(ctx);
6168 	struct sysctl_oid_list *child, *oid_list;
6169 	struct sysctl_ctx_list *ctx_list;
6170 	struct sysctl_oid *node;
6171 
6172 	ctx_list = device_get_sysctl_ctx(dev);
6173 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6174 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6175 						      CTLFLAG_RD, NULL, "IFLIB fields");
6176 	oid_list = SYSCTL_CHILDREN(node);
6177 
6178 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6179 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6180 		       "driver version");
6181 
6182 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6183 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6184 			"# of txqs to use, 0 => use default #");
6185 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6186 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6187 			"# of rxqs to use, 0 => use default #");
6188 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6189 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6190                        "permit #txq != #rxq");
6191 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6192                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6193                       "disable MSIX (default 0)");
6194 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6195 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6196                        "set the rx budget");
6197 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6198 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6199 		       "cause tx to abdicate instead of running to completion");
6200 
6201 	/* XXX change for per-queue sizes */
6202 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6203 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6204                        mp_ndesc_handler, "A",
6205                        "list of # of tx descriptors to use, 0 = use default #");
6206 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6207 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6208                        mp_ndesc_handler, "A",
6209                        "list of # of rx descriptors to use, 0 = use default #");
6210 }
6211 
6212 static void
6213 iflib_add_device_sysctl_post(if_ctx_t ctx)
6214 {
6215 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6216 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6217         device_t dev = iflib_get_dev(ctx);
6218 	struct sysctl_oid_list *child;
6219 	struct sysctl_ctx_list *ctx_list;
6220 	iflib_fl_t fl;
6221 	iflib_txq_t txq;
6222 	iflib_rxq_t rxq;
6223 	int i, j;
6224 	char namebuf[NAME_BUFLEN];
6225 	char *qfmt;
6226 	struct sysctl_oid *queue_node, *fl_node, *node;
6227 	struct sysctl_oid_list *queue_list, *fl_list;
6228 	ctx_list = device_get_sysctl_ctx(dev);
6229 
6230 	node = ctx->ifc_sysctl_node;
6231 	child = SYSCTL_CHILDREN(node);
6232 
6233 	if (scctx->isc_ntxqsets > 100)
6234 		qfmt = "txq%03d";
6235 	else if (scctx->isc_ntxqsets > 10)
6236 		qfmt = "txq%02d";
6237 	else
6238 		qfmt = "txq%d";
6239 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6240 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6241 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6242 					     CTLFLAG_RD, NULL, "Queue Name");
6243 		queue_list = SYSCTL_CHILDREN(queue_node);
6244 #if MEMORY_LOGGING
6245 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6246 				CTLFLAG_RD,
6247 				&txq->ift_dequeued, "total mbufs freed");
6248 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6249 				CTLFLAG_RD,
6250 				&txq->ift_enqueued, "total mbufs enqueued");
6251 #endif
6252 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6253 				   CTLFLAG_RD,
6254 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6255 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6256 				   CTLFLAG_RD,
6257 				   &txq->ift_pullups, "# of times m_pullup was called");
6258 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6259 				   CTLFLAG_RD,
6260 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6261 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6262 				   CTLFLAG_RD,
6263 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6264 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6265 				   CTLFLAG_RD,
6266 				   &txq->ift_map_failed, "# of times dma map failed");
6267 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6268 				   CTLFLAG_RD,
6269 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6270 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6271 				   CTLFLAG_RD,
6272 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6273 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6274 				   CTLFLAG_RD,
6275 				   &txq->ift_pidx, 1, "Producer Index");
6276 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6277 				   CTLFLAG_RD,
6278 				   &txq->ift_cidx, 1, "Consumer Index");
6279 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6280 				   CTLFLAG_RD,
6281 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6282 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6283 				   CTLFLAG_RD,
6284 				   &txq->ift_in_use, 1, "descriptors in use");
6285 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6286 				   CTLFLAG_RD,
6287 				   &txq->ift_processed, "descriptors procesed for clean");
6288 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6289 				   CTLFLAG_RD,
6290 				   &txq->ift_cleaned, "total cleaned");
6291 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6292 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6293 				0, mp_ring_state_handler, "A", "soft ring state");
6294 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6295 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6296 				       "# of enqueues to the mp_ring for this queue");
6297 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6298 				       CTLFLAG_RD, &txq->ift_br->drops,
6299 				       "# of drops in the mp_ring for this queue");
6300 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6301 				       CTLFLAG_RD, &txq->ift_br->starts,
6302 				       "# of normal consumer starts in the mp_ring for this queue");
6303 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6304 				       CTLFLAG_RD, &txq->ift_br->stalls,
6305 					       "# of consumer stalls in the mp_ring for this queue");
6306 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6307 			       CTLFLAG_RD, &txq->ift_br->restarts,
6308 				       "# of consumer restarts in the mp_ring for this queue");
6309 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6310 				       CTLFLAG_RD, &txq->ift_br->abdications,
6311 				       "# of consumer abdications in the mp_ring for this queue");
6312 	}
6313 
6314 	if (scctx->isc_nrxqsets > 100)
6315 		qfmt = "rxq%03d";
6316 	else if (scctx->isc_nrxqsets > 10)
6317 		qfmt = "rxq%02d";
6318 	else
6319 		qfmt = "rxq%d";
6320 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6321 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6322 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6323 					     CTLFLAG_RD, NULL, "Queue Name");
6324 		queue_list = SYSCTL_CHILDREN(queue_node);
6325 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6326 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6327 				       CTLFLAG_RD,
6328 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6329 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6330 				       CTLFLAG_RD,
6331 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6332 		}
6333 
6334 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6335 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6336 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6337 						     CTLFLAG_RD, NULL, "freelist Name");
6338 			fl_list = SYSCTL_CHILDREN(fl_node);
6339 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6340 				       CTLFLAG_RD,
6341 				       &fl->ifl_pidx, 1, "Producer Index");
6342 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6343 				       CTLFLAG_RD,
6344 				       &fl->ifl_cidx, 1, "Consumer Index");
6345 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6346 				       CTLFLAG_RD,
6347 				       &fl->ifl_credits, 1, "credits available");
6348 #if MEMORY_LOGGING
6349 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6350 					CTLFLAG_RD,
6351 					&fl->ifl_m_enqueued, "mbufs allocated");
6352 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6353 					CTLFLAG_RD,
6354 					&fl->ifl_m_dequeued, "mbufs freed");
6355 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6356 					CTLFLAG_RD,
6357 					&fl->ifl_cl_enqueued, "clusters allocated");
6358 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6359 					CTLFLAG_RD,
6360 					&fl->ifl_cl_dequeued, "clusters freed");
6361 #endif
6362 
6363 		}
6364 	}
6365 
6366 }
6367 
6368 void
6369 iflib_request_reset(if_ctx_t ctx)
6370 {
6371 
6372 	STATE_LOCK(ctx);
6373 	ctx->ifc_flags |= IFC_DO_RESET;
6374 	STATE_UNLOCK(ctx);
6375 }
6376 
6377 #ifndef __NO_STRICT_ALIGNMENT
6378 static struct mbuf *
6379 iflib_fixup_rx(struct mbuf *m)
6380 {
6381 	struct mbuf *n;
6382 
6383 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6384 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6385 		m->m_data += ETHER_HDR_LEN;
6386 		n = m;
6387 	} else {
6388 		MGETHDR(n, M_NOWAIT, MT_DATA);
6389 		if (n == NULL) {
6390 			m_freem(m);
6391 			return (NULL);
6392 		}
6393 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6394 		m->m_data += ETHER_HDR_LEN;
6395 		m->m_len -= ETHER_HDR_LEN;
6396 		n->m_len = ETHER_HDR_LEN;
6397 		M_MOVE_PKTHDR(n, m);
6398 		n->m_next = m;
6399 	}
6400 	return (n);
6401 }
6402 #endif
6403 
6404 #ifdef NETDUMP
6405 static void
6406 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6407 {
6408 	if_ctx_t ctx;
6409 
6410 	ctx = if_getsoftc(ifp);
6411 	CTX_LOCK(ctx);
6412 	*nrxr = NRXQSETS(ctx);
6413 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6414 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6415 	CTX_UNLOCK(ctx);
6416 }
6417 
6418 static void
6419 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6420 {
6421 	if_ctx_t ctx;
6422 	if_softc_ctx_t scctx;
6423 	iflib_fl_t fl;
6424 	iflib_rxq_t rxq;
6425 	int i, j;
6426 
6427 	ctx = if_getsoftc(ifp);
6428 	scctx = &ctx->ifc_softc_ctx;
6429 
6430 	switch (event) {
6431 	case NETDUMP_START:
6432 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6433 			rxq = &ctx->ifc_rxqs[i];
6434 			for (j = 0; j < rxq->ifr_nfl; j++) {
6435 				fl = rxq->ifr_fl;
6436 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6437 			}
6438 		}
6439 		iflib_no_tx_batch = 1;
6440 		break;
6441 	default:
6442 		break;
6443 	}
6444 }
6445 
6446 static int
6447 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6448 {
6449 	if_ctx_t ctx;
6450 	iflib_txq_t txq;
6451 	int error;
6452 
6453 	ctx = if_getsoftc(ifp);
6454 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6455 	    IFF_DRV_RUNNING)
6456 		return (EBUSY);
6457 
6458 	txq = &ctx->ifc_txqs[0];
6459 	error = iflib_encap(txq, &m);
6460 	if (error == 0)
6461 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6462 	return (error);
6463 }
6464 
6465 static int
6466 iflib_netdump_poll(struct ifnet *ifp, int count)
6467 {
6468 	if_ctx_t ctx;
6469 	if_softc_ctx_t scctx;
6470 	iflib_txq_t txq;
6471 	int i;
6472 
6473 	ctx = if_getsoftc(ifp);
6474 	scctx = &ctx->ifc_softc_ctx;
6475 
6476 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6477 	    IFF_DRV_RUNNING)
6478 		return (EBUSY);
6479 
6480 	txq = &ctx->ifc_txqs[0];
6481 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6482 
6483 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6484 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6485 	return (0);
6486 }
6487 #endif /* NETDUMP */
6488