xref: /freebsd/sys/net/iflib.c (revision 7d958cc9aa0aa5303daef28fbba829197b580d5e)
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 #if defined(__i386__) || defined(__amd64__)
96 #include <sys/memdesc.h>
97 #include <machine/bus.h>
98 #include <machine/md_var.h>
99 #include <machine/specialreg.h>
100 #include <x86/include/busdma_impl.h>
101 #include <x86/iommu/busdma_dmar.h>
102 #endif
103 
104 #include <sys/bitstring.h>
105 /*
106  * enable accounting of every mbuf as it comes in to and goes out of
107  * iflib's software descriptor references
108  */
109 #define MEMORY_LOGGING 0
110 /*
111  * Enable mbuf vectors for compressing long mbuf chains
112  */
113 
114 /*
115  * NB:
116  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
117  *   we prefetch needs to be determined by the time spent in m_free vis a vis
118  *   the cost of a prefetch. This will of course vary based on the workload:
119  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
120  *        is quite expensive, thus suggesting very little prefetch.
121  *      - small packet forwarding which is just returning a single mbuf to
122  *        UMA will typically be very fast vis a vis the cost of a memory
123  *        access.
124  */
125 
126 
127 /*
128  * File organization:
129  *  - private structures
130  *  - iflib private utility functions
131  *  - ifnet functions
132  *  - vlan registry and other exported functions
133  *  - iflib public core functions
134  *
135  *
136  */
137 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
138 
139 struct iflib_txq;
140 typedef struct iflib_txq *iflib_txq_t;
141 struct iflib_rxq;
142 typedef struct iflib_rxq *iflib_rxq_t;
143 struct iflib_fl;
144 typedef struct iflib_fl *iflib_fl_t;
145 
146 struct iflib_ctx;
147 
148 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
149 
150 typedef struct iflib_filter_info {
151 	driver_filter_t *ifi_filter;
152 	void *ifi_filter_arg;
153 	struct grouptask *ifi_task;
154 	void *ifi_ctx;
155 } *iflib_filter_info_t;
156 
157 struct iflib_ctx {
158 	KOBJ_FIELDS;
159    /*
160    * Pointer to hardware driver's softc
161    */
162 	void *ifc_softc;
163 	device_t ifc_dev;
164 	if_t ifc_ifp;
165 
166 	cpuset_t ifc_cpus;
167 	if_shared_ctx_t ifc_sctx;
168 	struct if_softc_ctx ifc_softc_ctx;
169 
170 	struct sx ifc_ctx_sx;
171 	struct mtx ifc_state_mtx;
172 
173 	uint16_t ifc_nhwtxqs;
174 
175 	iflib_txq_t ifc_txqs;
176 	iflib_rxq_t ifc_rxqs;
177 	uint32_t ifc_if_flags;
178 	uint32_t ifc_flags;
179 	uint32_t ifc_max_fl_buf_size;
180 	int ifc_in_detach;
181 
182 	int ifc_link_state;
183 	int ifc_link_irq;
184 	int ifc_watchdog_events;
185 	struct cdev *ifc_led_dev;
186 	struct resource *ifc_msix_mem;
187 
188 	struct if_irq ifc_legacy_irq;
189 	struct grouptask ifc_admin_task;
190 	struct grouptask ifc_vflr_task;
191 	struct iflib_filter_info ifc_filter_info;
192 	struct ifmedia	ifc_media;
193 
194 	struct sysctl_oid *ifc_sysctl_node;
195 	uint16_t ifc_sysctl_ntxqs;
196 	uint16_t ifc_sysctl_nrxqs;
197 	uint16_t ifc_sysctl_qs_eq_override;
198 	uint16_t ifc_sysctl_rx_budget;
199 
200 	qidx_t ifc_sysctl_ntxds[8];
201 	qidx_t ifc_sysctl_nrxds[8];
202 	struct if_txrx ifc_txrx;
203 #define isc_txd_encap  ifc_txrx.ift_txd_encap
204 #define isc_txd_flush  ifc_txrx.ift_txd_flush
205 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
206 #define isc_rxd_available ifc_txrx.ift_rxd_available
207 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
208 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
209 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
212 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
213 	eventhandler_tag ifc_vlan_attach_event;
214 	eventhandler_tag ifc_vlan_detach_event;
215 	uint8_t ifc_mac[ETHER_ADDR_LEN];
216 	char ifc_mtx_name[16];
217 };
218 
219 
220 void *
221 iflib_get_softc(if_ctx_t ctx)
222 {
223 
224 	return (ctx->ifc_softc);
225 }
226 
227 device_t
228 iflib_get_dev(if_ctx_t ctx)
229 {
230 
231 	return (ctx->ifc_dev);
232 }
233 
234 if_t
235 iflib_get_ifp(if_ctx_t ctx)
236 {
237 
238 	return (ctx->ifc_ifp);
239 }
240 
241 struct ifmedia *
242 iflib_get_media(if_ctx_t ctx)
243 {
244 
245 	return (&ctx->ifc_media);
246 }
247 
248 uint32_t
249 iflib_get_flags(if_ctx_t ctx)
250 {
251 	return (ctx->ifc_flags);
252 }
253 
254 void
255 iflib_set_detach(if_ctx_t ctx)
256 {
257 	ctx->ifc_in_detach = 1;
258 }
259 
260 void
261 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
262 {
263 
264 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
265 }
266 
267 if_softc_ctx_t
268 iflib_get_softc_ctx(if_ctx_t ctx)
269 {
270 
271 	return (&ctx->ifc_softc_ctx);
272 }
273 
274 if_shared_ctx_t
275 iflib_get_sctx(if_ctx_t ctx)
276 {
277 
278 	return (ctx->ifc_sctx);
279 }
280 
281 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
282 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
283 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
284 
285 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
286 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
287 
288 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
289 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
290 #define RX_SW_DESC_INUSE        (1 << 3)
291 #define TX_SW_DESC_MAPPED       (1 << 4)
292 
293 #define	M_TOOBIG		M_PROTO1
294 
295 typedef struct iflib_sw_rx_desc_array {
296 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
297 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
298 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
299 	uint8_t		*ifsd_flags;
300 } iflib_rxsd_array_t;
301 
302 typedef struct iflib_sw_tx_desc_array {
303 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
304 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
305 	uint8_t		*ifsd_flags;
306 } if_txsd_vec_t;
307 
308 
309 /* magic number that should be high enough for any hardware */
310 #define IFLIB_MAX_TX_SEGS		128
311 /* bnxt supports 64 with hardware LRO enabled */
312 #define IFLIB_MAX_RX_SEGS		64
313 #define IFLIB_RX_COPY_THRESH		128
314 #define IFLIB_MAX_RX_REFRESH		32
315 /* The minimum descriptors per second before we start coalescing */
316 #define IFLIB_MIN_DESC_SEC		16384
317 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
318 #define IFLIB_QUEUE_IDLE		0
319 #define IFLIB_QUEUE_HUNG		1
320 #define IFLIB_QUEUE_WORKING		2
321 /* maximum number of txqs that can share an rx interrupt */
322 #define IFLIB_MAX_TX_SHARED_INTR	4
323 
324 /* this should really scale with ring size - this is a fairly arbitrary value */
325 #define TX_BATCH_SIZE			32
326 
327 #define IFLIB_RESTART_BUDGET		8
328 
329 
330 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
331 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
332 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
333 struct iflib_txq {
334 	qidx_t		ift_in_use;
335 	qidx_t		ift_cidx;
336 	qidx_t		ift_cidx_processed;
337 	qidx_t		ift_pidx;
338 	uint8_t		ift_gen;
339 	uint8_t		ift_br_offset;
340 	uint16_t	ift_npending;
341 	uint16_t	ift_db_pending;
342 	uint16_t	ift_rs_pending;
343 	/* implicit pad */
344 	uint8_t		ift_txd_size[8];
345 	uint64_t	ift_processed;
346 	uint64_t	ift_cleaned;
347 	uint64_t	ift_cleaned_prev;
348 #if MEMORY_LOGGING
349 	uint64_t	ift_enqueued;
350 	uint64_t	ift_dequeued;
351 #endif
352 	uint64_t	ift_no_tx_dma_setup;
353 	uint64_t	ift_no_desc_avail;
354 	uint64_t	ift_mbuf_defrag_failed;
355 	uint64_t	ift_mbuf_defrag;
356 	uint64_t	ift_map_failed;
357 	uint64_t	ift_txd_encap_efbig;
358 	uint64_t	ift_pullups;
359 
360 	struct mtx	ift_mtx;
361 	struct mtx	ift_db_mtx;
362 
363 	/* constant values */
364 	if_ctx_t	ift_ctx;
365 	struct ifmp_ring        *ift_br;
366 	struct grouptask	ift_task;
367 	qidx_t		ift_size;
368 	uint16_t	ift_id;
369 	struct callout	ift_timer;
370 
371 	if_txsd_vec_t	ift_sds;
372 	uint8_t		ift_qstatus;
373 	uint8_t		ift_closed;
374 	uint8_t		ift_update_freq;
375 	struct iflib_filter_info ift_filter_info;
376 	bus_dma_tag_t		ift_desc_tag;
377 	bus_dma_tag_t		ift_tso_desc_tag;
378 	iflib_dma_info_t	ift_ifdi;
379 #define MTX_NAME_LEN 16
380 	char                    ift_mtx_name[MTX_NAME_LEN];
381 	char                    ift_db_mtx_name[MTX_NAME_LEN];
382 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
383 #ifdef IFLIB_DIAGNOSTICS
384 	uint64_t ift_cpu_exec_count[256];
385 #endif
386 } __aligned(CACHE_LINE_SIZE);
387 
388 struct iflib_fl {
389 	qidx_t		ifl_cidx;
390 	qidx_t		ifl_pidx;
391 	qidx_t		ifl_credits;
392 	uint8_t		ifl_gen;
393 	uint8_t		ifl_rxd_size;
394 #if MEMORY_LOGGING
395 	uint64_t	ifl_m_enqueued;
396 	uint64_t	ifl_m_dequeued;
397 	uint64_t	ifl_cl_enqueued;
398 	uint64_t	ifl_cl_dequeued;
399 #endif
400 	/* implicit pad */
401 
402 	bitstr_t 	*ifl_rx_bitmap;
403 	qidx_t		ifl_fragidx;
404 	/* constant */
405 	qidx_t		ifl_size;
406 	uint16_t	ifl_buf_size;
407 	uint16_t	ifl_cltype;
408 	uma_zone_t	ifl_zone;
409 	iflib_rxsd_array_t	ifl_sds;
410 	iflib_rxq_t	ifl_rxq;
411 	uint8_t		ifl_id;
412 	bus_dma_tag_t           ifl_desc_tag;
413 	iflib_dma_info_t	ifl_ifdi;
414 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
415 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
416 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
417 }  __aligned(CACHE_LINE_SIZE);
418 
419 static inline qidx_t
420 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
421 {
422 	qidx_t used;
423 
424 	if (pidx > cidx)
425 		used = pidx - cidx;
426 	else if (pidx < cidx)
427 		used = size - cidx + pidx;
428 	else if (gen == 0 && pidx == cidx)
429 		used = 0;
430 	else if (gen == 1 && pidx == cidx)
431 		used = size;
432 	else
433 		panic("bad state");
434 
435 	return (used);
436 }
437 
438 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
439 
440 #define IDXDIFF(head, tail, wrap) \
441 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
442 
443 struct iflib_rxq {
444 	/* If there is a separate completion queue -
445 	 * these are the cq cidx and pidx. Otherwise
446 	 * these are unused.
447 	 */
448 	qidx_t		ifr_size;
449 	qidx_t		ifr_cq_cidx;
450 	qidx_t		ifr_cq_pidx;
451 	uint8_t		ifr_cq_gen;
452 	uint8_t		ifr_fl_offset;
453 
454 	if_ctx_t	ifr_ctx;
455 	iflib_fl_t	ifr_fl;
456 	uint64_t	ifr_rx_irq;
457 	uint16_t	ifr_id;
458 	uint8_t		ifr_lro_enabled;
459 	uint8_t		ifr_nfl;
460 	uint8_t		ifr_ntxqirq;
461 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
462 	struct lro_ctrl			ifr_lc;
463 	struct grouptask        ifr_task;
464 	struct iflib_filter_info ifr_filter_info;
465 	iflib_dma_info_t		ifr_ifdi;
466 
467 	/* dynamically allocate if any drivers need a value substantially larger than this */
468 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
469 #ifdef IFLIB_DIAGNOSTICS
470 	uint64_t ifr_cpu_exec_count[256];
471 #endif
472 }  __aligned(CACHE_LINE_SIZE);
473 
474 typedef struct if_rxsd {
475 	caddr_t *ifsd_cl;
476 	struct mbuf **ifsd_m;
477 	iflib_fl_t ifsd_fl;
478 	qidx_t ifsd_cidx;
479 } *if_rxsd_t;
480 
481 /* multiple of word size */
482 #ifdef __LP64__
483 #define PKT_INFO_SIZE	6
484 #define RXD_INFO_SIZE	5
485 #define PKT_TYPE uint64_t
486 #else
487 #define PKT_INFO_SIZE	11
488 #define RXD_INFO_SIZE	8
489 #define PKT_TYPE uint32_t
490 #endif
491 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
492 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
493 
494 typedef struct if_pkt_info_pad {
495 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
496 } *if_pkt_info_pad_t;
497 typedef struct if_rxd_info_pad {
498 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
499 } *if_rxd_info_pad_t;
500 
501 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
502 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
503 
504 
505 static inline void
506 pkt_info_zero(if_pkt_info_t pi)
507 {
508 	if_pkt_info_pad_t pi_pad;
509 
510 	pi_pad = (if_pkt_info_pad_t)pi;
511 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
512 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
513 #ifndef __LP64__
514 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
515 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
516 #endif
517 }
518 
519 static device_method_t iflib_pseudo_methods[] = {
520 	DEVMETHOD(device_attach, noop_attach),
521 	DEVMETHOD(device_detach, iflib_pseudo_detach),
522 	DEVMETHOD_END
523 };
524 
525 driver_t iflib_pseudodriver = {
526 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
527 };
528 
529 static inline void
530 rxd_info_zero(if_rxd_info_t ri)
531 {
532 	if_rxd_info_pad_t ri_pad;
533 	int i;
534 
535 	ri_pad = (if_rxd_info_pad_t)ri;
536 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
537 		ri_pad->rxd_val[i] = 0;
538 		ri_pad->rxd_val[i+1] = 0;
539 		ri_pad->rxd_val[i+2] = 0;
540 		ri_pad->rxd_val[i+3] = 0;
541 	}
542 #ifdef __LP64__
543 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
544 #endif
545 }
546 
547 /*
548  * Only allow a single packet to take up most 1/nth of the tx ring
549  */
550 #define MAX_SINGLE_PACKET_FRACTION 12
551 #define IF_BAD_DMA (bus_addr_t)-1
552 
553 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
554 
555 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
556 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
557 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
558 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
559 
560 
561 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
562 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
563 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
564 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
565 
566 
567 
568 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
569 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
570 
571 
572 /* Our boot-time initialization hook */
573 static int	iflib_module_event_handler(module_t, int, void *);
574 
575 static moduledata_t iflib_moduledata = {
576 	"iflib",
577 	iflib_module_event_handler,
578 	NULL
579 };
580 
581 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
582 MODULE_VERSION(iflib, 1);
583 
584 MODULE_DEPEND(iflib, pci, 1, 1, 1);
585 MODULE_DEPEND(iflib, ether, 1, 1, 1);
586 
587 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
588 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
589 
590 #ifndef IFLIB_DEBUG_COUNTERS
591 #ifdef INVARIANTS
592 #define IFLIB_DEBUG_COUNTERS 1
593 #else
594 #define IFLIB_DEBUG_COUNTERS 0
595 #endif /* !INVARIANTS */
596 #endif
597 
598 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
599                    "iflib driver parameters");
600 
601 /*
602  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
603  */
604 static int iflib_min_tx_latency = 0;
605 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
606 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
607 static int iflib_no_tx_batch = 0;
608 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
609 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
610 
611 
612 #if IFLIB_DEBUG_COUNTERS
613 
614 static int iflib_tx_seen;
615 static int iflib_tx_sent;
616 static int iflib_tx_encap;
617 static int iflib_rx_allocs;
618 static int iflib_fl_refills;
619 static int iflib_fl_refills_large;
620 static int iflib_tx_frees;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
623 		   &iflib_tx_seen, 0, "# tx mbufs seen");
624 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
625 		   &iflib_tx_sent, 0, "# tx mbufs sent");
626 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
627 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
628 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
629 		   &iflib_tx_frees, 0, "# tx frees");
630 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
631 		   &iflib_rx_allocs, 0, "# rx allocations");
632 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
633 		   &iflib_fl_refills, 0, "# refills");
634 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
635 		   &iflib_fl_refills_large, 0, "# large refills");
636 
637 
638 static int iflib_txq_drain_flushing;
639 static int iflib_txq_drain_oactive;
640 static int iflib_txq_drain_notready;
641 static int iflib_txq_drain_encapfail;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
644 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
645 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
646 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
647 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
648 		   &iflib_txq_drain_notready, 0, "# drain notready");
649 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
650 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
651 
652 
653 static int iflib_encap_load_mbuf_fail;
654 static int iflib_encap_pad_mbuf_fail;
655 static int iflib_encap_txq_avail_fail;
656 static int iflib_encap_txd_encap_fail;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
659 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
660 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
661 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
662 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
663 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
664 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
665 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
666 
667 static int iflib_task_fn_rxs;
668 static int iflib_rx_intr_enables;
669 static int iflib_fast_intrs;
670 static int iflib_intr_link;
671 static int iflib_intr_msix;
672 static int iflib_rx_unavail;
673 static int iflib_rx_ctx_inactive;
674 static int iflib_rx_zero_len;
675 static int iflib_rx_if_input;
676 static int iflib_rx_mbuf_null;
677 static int iflib_rxd_flush;
678 
679 static int iflib_verbose_debug;
680 
681 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
682 		   &iflib_intr_link, 0, "# intr link calls");
683 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
684 		   &iflib_intr_msix, 0, "# intr msix calls");
685 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
686 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
687 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
688 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
689 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
690 		   &iflib_fast_intrs, 0, "# fast_intr calls");
691 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
692 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
693 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
694 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
695 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
696 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
697 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
698 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
699 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
700 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
701 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
702 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
703 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
704 		   &iflib_verbose_debug, 0, "enable verbose debugging");
705 
706 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
707 static void
708 iflib_debug_reset(void)
709 {
710 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
711 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
712 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
713 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
714 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
715 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
716 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
717 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
718 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
719 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
720 }
721 
722 #else
723 #define DBG_COUNTER_INC(name)
724 static void iflib_debug_reset(void) {}
725 #endif
726 
727 #define IFLIB_DEBUG 0
728 
729 static void iflib_tx_structures_free(if_ctx_t ctx);
730 static void iflib_rx_structures_free(if_ctx_t ctx);
731 static int iflib_queues_alloc(if_ctx_t ctx);
732 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
733 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
734 static int iflib_qset_structures_setup(if_ctx_t ctx);
735 static int iflib_msix_init(if_ctx_t ctx);
736 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
737 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
738 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
739 static int iflib_register(if_ctx_t);
740 static void iflib_init_locked(if_ctx_t ctx);
741 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
742 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
743 static void iflib_ifmp_purge(iflib_txq_t txq);
744 static void _iflib_pre_assert(if_softc_ctx_t scctx);
745 static void iflib_if_init_locked(if_ctx_t ctx);
746 #ifndef __NO_STRICT_ALIGNMENT
747 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
748 #endif
749 
750 NETDUMP_DEFINE(iflib);
751 
752 #ifdef DEV_NETMAP
753 #include <sys/selinfo.h>
754 #include <net/netmap.h>
755 #include <dev/netmap/netmap_kern.h>
756 
757 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
758 
759 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
760 
761 /*
762  * device-specific sysctl variables:
763  *
764  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
765  *	During regular operations the CRC is stripped, but on some
766  *	hardware reception of frames not multiple of 64 is slower,
767  *	so using crcstrip=0 helps in benchmarks.
768  *
769  * iflib_rx_miss, iflib_rx_miss_bufs:
770  *	count packets that might be missed due to lost interrupts.
771  */
772 SYSCTL_DECL(_dev_netmap);
773 /*
774  * The xl driver by default strips CRCs and we do not override it.
775  */
776 
777 int iflib_crcstrip = 1;
778 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
779     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
780 
781 int iflib_rx_miss, iflib_rx_miss_bufs;
782 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
783     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
785     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
786 
787 /*
788  * Register/unregister. We are already under netmap lock.
789  * Only called on the first register or the last unregister.
790  */
791 static int
792 iflib_netmap_register(struct netmap_adapter *na, int onoff)
793 {
794 	struct ifnet *ifp = na->ifp;
795 	if_ctx_t ctx = ifp->if_softc;
796 	int status;
797 
798 	CTX_LOCK(ctx);
799 	IFDI_INTR_DISABLE(ctx);
800 
801 	/* Tell the stack that the interface is no longer active */
802 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
803 
804 	if (!CTX_IS_VF(ctx))
805 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
806 
807 	/* enable or disable flags and callbacks in na and ifp */
808 	if (onoff) {
809 		nm_set_native_flags(na);
810 	} else {
811 		nm_clear_native_flags(na);
812 	}
813 	iflib_stop(ctx);
814 	iflib_init_locked(ctx);
815 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
816 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
817 	if (status)
818 		nm_clear_native_flags(na);
819 	CTX_UNLOCK(ctx);
820 	return (status);
821 }
822 
823 static int
824 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
825 {
826 	struct netmap_adapter *na = kring->na;
827 	u_int const lim = kring->nkr_num_slots - 1;
828 	u_int head = kring->rhead;
829 	struct netmap_ring *ring = kring->ring;
830 	bus_dmamap_t *map;
831 	struct if_rxd_update iru;
832 	if_ctx_t ctx = rxq->ifr_ctx;
833 	iflib_fl_t fl = &rxq->ifr_fl[0];
834 	uint32_t refill_pidx, nic_i;
835 
836 	if (nm_i == head && __predict_true(!init))
837 		return 0;
838 	iru_init(&iru, rxq, 0 /* flid */);
839 	map = fl->ifl_sds.ifsd_map;
840 	refill_pidx = netmap_idx_k2n(kring, nm_i);
841 	/*
842 	 * IMPORTANT: we must leave one free slot in the ring,
843 	 * so move head back by one unit
844 	 */
845 	head = nm_prev(head, lim);
846 	nic_i = UINT_MAX;
847 	while (nm_i != head) {
848 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
849 			struct netmap_slot *slot = &ring->slot[nm_i];
850 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
851 			uint32_t nic_i_dma = refill_pidx;
852 			nic_i = netmap_idx_k2n(kring, nm_i);
853 
854 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
855 
856 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
857 			        return netmap_ring_reinit(kring);
858 
859 			fl->ifl_vm_addrs[tmp_pidx] = addr;
860 			if (__predict_false(init) && map) {
861 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
862 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
863 				/* buffer has changed, reload map */
864 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
865 			}
866 			slot->flags &= ~NS_BUF_CHANGED;
867 
868 			nm_i = nm_next(nm_i, lim);
869 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
870 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
871 				continue;
872 
873 			iru.iru_pidx = refill_pidx;
874 			iru.iru_count = tmp_pidx+1;
875 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
876 
877 			refill_pidx = nic_i;
878 			if (map == NULL)
879 				continue;
880 
881 			for (int n = 0; n < iru.iru_count; n++) {
882 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
883 						BUS_DMASYNC_PREREAD);
884 				/* XXX - change this to not use the netmap func*/
885 				nic_i_dma = nm_next(nic_i_dma, lim);
886 			}
887 		}
888 	}
889 	kring->nr_hwcur = head;
890 
891 	if (map)
892 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
893 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
894 	if (__predict_true(nic_i != UINT_MAX))
895 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
896 	return (0);
897 }
898 
899 /*
900  * Reconcile kernel and user view of the transmit ring.
901  *
902  * All information is in the kring.
903  * Userspace wants to send packets up to the one before kring->rhead,
904  * kernel knows kring->nr_hwcur is the first unsent packet.
905  *
906  * Here we push packets out (as many as possible), and possibly
907  * reclaim buffers from previously completed transmission.
908  *
909  * The caller (netmap) guarantees that there is only one instance
910  * running at any time. Any interference with other driver
911  * methods should be handled by the individual drivers.
912  */
913 static int
914 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
915 {
916 	struct netmap_adapter *na = kring->na;
917 	struct ifnet *ifp = na->ifp;
918 	struct netmap_ring *ring = kring->ring;
919 	u_int nm_i;	/* index into the netmap ring */
920 	u_int nic_i;	/* index into the NIC ring */
921 	u_int n;
922 	u_int const lim = kring->nkr_num_slots - 1;
923 	u_int const head = kring->rhead;
924 	struct if_pkt_info pi;
925 
926 	/*
927 	 * interrupts on every tx packet are expensive so request
928 	 * them every half ring, or where NS_REPORT is set
929 	 */
930 	u_int report_frequency = kring->nkr_num_slots >> 1;
931 	/* device-specific */
932 	if_ctx_t ctx = ifp->if_softc;
933 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
934 
935 	if (txq->ift_sds.ifsd_map)
936 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
937 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
938 
939 
940 	/*
941 	 * First part: process new packets to send.
942 	 * nm_i is the current index in the netmap ring,
943 	 * nic_i is the corresponding index in the NIC ring.
944 	 *
945 	 * If we have packets to send (nm_i != head)
946 	 * iterate over the netmap ring, fetch length and update
947 	 * the corresponding slot in the NIC ring. Some drivers also
948 	 * need to update the buffer's physical address in the NIC slot
949 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
950 	 *
951 	 * The netmap_reload_map() calls is especially expensive,
952 	 * even when (as in this case) the tag is 0, so do only
953 	 * when the buffer has actually changed.
954 	 *
955 	 * If possible do not set the report/intr bit on all slots,
956 	 * but only a few times per ring or when NS_REPORT is set.
957 	 *
958 	 * Finally, on 10G and faster drivers, it might be useful
959 	 * to prefetch the next slot and txr entry.
960 	 */
961 
962 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
963 	if (nm_i != head) {	/* we have new packets to send */
964 		pkt_info_zero(&pi);
965 		pi.ipi_segs = txq->ift_segs;
966 		pi.ipi_qsidx = kring->ring_id;
967 		nic_i = netmap_idx_k2n(kring, nm_i);
968 
969 		__builtin_prefetch(&ring->slot[nm_i]);
970 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
971 		if (txq->ift_sds.ifsd_map)
972 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
973 
974 		for (n = 0; nm_i != head; n++) {
975 			struct netmap_slot *slot = &ring->slot[nm_i];
976 			u_int len = slot->len;
977 			uint64_t paddr;
978 			void *addr = PNMB(na, slot, &paddr);
979 			int flags = (slot->flags & NS_REPORT ||
980 				nic_i == 0 || nic_i == report_frequency) ?
981 				IPI_TX_INTR : 0;
982 
983 			/* device-specific */
984 			pi.ipi_len = len;
985 			pi.ipi_segs[0].ds_addr = paddr;
986 			pi.ipi_segs[0].ds_len = len;
987 			pi.ipi_nsegs = 1;
988 			pi.ipi_ndescs = 0;
989 			pi.ipi_pidx = nic_i;
990 			pi.ipi_flags = flags;
991 
992 			/* Fill the slot in the NIC ring. */
993 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
994 
995 			/* prefetch for next round */
996 			__builtin_prefetch(&ring->slot[nm_i + 1]);
997 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
998 			if (txq->ift_sds.ifsd_map) {
999 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1000 
1001 				NM_CHECK_ADDR_LEN(na, addr, len);
1002 
1003 				if (slot->flags & NS_BUF_CHANGED) {
1004 					/* buffer has changed, reload map */
1005 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1006 				}
1007 				/* make sure changes to the buffer are synced */
1008 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1009 						BUS_DMASYNC_PREWRITE);
1010 			}
1011 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1012 			nm_i = nm_next(nm_i, lim);
1013 			nic_i = nm_next(nic_i, lim);
1014 		}
1015 		kring->nr_hwcur = head;
1016 
1017 		/* synchronize the NIC ring */
1018 		if (txq->ift_sds.ifsd_map)
1019 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1020 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1021 
1022 		/* (re)start the tx unit up to slot nic_i (excluded) */
1023 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1024 	}
1025 
1026 	/*
1027 	 * Second part: reclaim buffers for completed transmissions.
1028 	 *
1029 	 * If there are unclaimed buffers, attempt to reclaim them.
1030 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1031 	 * minimal delay, then trigger the tx handler which will spin in the
1032 	 * group task queue.
1033 	 */
1034 	if (kring->nr_hwtail != nm_prev(head, lim)) {
1035 		if (iflib_tx_credits_update(ctx, txq)) {
1036 			/* some tx completed, increment avail */
1037 			nic_i = txq->ift_cidx_processed;
1038 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1039 		}
1040 		else {
1041 			if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1042 				DELAY(1);
1043 				GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txq->ift_id].ift_task);
1044 			}
1045 		}
1046 	}
1047 	return (0);
1048 }
1049 
1050 /*
1051  * Reconcile kernel and user view of the receive ring.
1052  * Same as for the txsync, this routine must be efficient.
1053  * The caller guarantees a single invocations, but races against
1054  * the rest of the driver should be handled here.
1055  *
1056  * On call, kring->rhead is the first packet that userspace wants
1057  * to keep, and kring->rcur is the wakeup point.
1058  * The kernel has previously reported packets up to kring->rtail.
1059  *
1060  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1061  * of whether or not we received an interrupt.
1062  */
1063 static int
1064 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1065 {
1066 	struct netmap_adapter *na = kring->na;
1067 	struct netmap_ring *ring = kring->ring;
1068 	uint32_t nm_i;	/* index into the netmap ring */
1069 	uint32_t nic_i;	/* index into the NIC ring */
1070 	u_int i, n;
1071 	u_int const lim = kring->nkr_num_slots - 1;
1072 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1073 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1074 	struct if_rxd_info ri;
1075 
1076 	struct ifnet *ifp = na->ifp;
1077 	if_ctx_t ctx = ifp->if_softc;
1078 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1079 	iflib_fl_t fl = rxq->ifr_fl;
1080 	if (head > lim)
1081 		return netmap_ring_reinit(kring);
1082 
1083 	/* XXX check sync modes */
1084 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1085 		if (fl->ifl_sds.ifsd_map == NULL)
1086 			continue;
1087 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1088 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1089 	}
1090 	/*
1091 	 * First part: import newly received packets.
1092 	 *
1093 	 * nm_i is the index of the next free slot in the netmap ring,
1094 	 * nic_i is the index of the next received packet in the NIC ring,
1095 	 * and they may differ in case if_init() has been called while
1096 	 * in netmap mode. For the receive ring we have
1097 	 *
1098 	 *	nic_i = rxr->next_check;
1099 	 *	nm_i = kring->nr_hwtail (previous)
1100 	 * and
1101 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1102 	 *
1103 	 * rxr->next_check is set to 0 on a ring reinit
1104 	 */
1105 	if (netmap_no_pendintr || force_update) {
1106 		int crclen = iflib_crcstrip ? 0 : 4;
1107 		int error, avail;
1108 
1109 		for (i = 0; i < rxq->ifr_nfl; i++) {
1110 			fl = &rxq->ifr_fl[i];
1111 			nic_i = fl->ifl_cidx;
1112 			nm_i = netmap_idx_n2k(kring, nic_i);
1113 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1114 			for (n = 0; avail > 0; n++, avail--) {
1115 				rxd_info_zero(&ri);
1116 				ri.iri_frags = rxq->ifr_frags;
1117 				ri.iri_qsidx = kring->ring_id;
1118 				ri.iri_ifp = ctx->ifc_ifp;
1119 				ri.iri_cidx = nic_i;
1120 
1121 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1122 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1123 				ring->slot[nm_i].flags = 0;
1124 				if (fl->ifl_sds.ifsd_map)
1125 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1126 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1127 				nm_i = nm_next(nm_i, lim);
1128 				nic_i = nm_next(nic_i, lim);
1129 			}
1130 			if (n) { /* update the state variables */
1131 				if (netmap_no_pendintr && !force_update) {
1132 					/* diagnostics */
1133 					iflib_rx_miss ++;
1134 					iflib_rx_miss_bufs += n;
1135 				}
1136 				fl->ifl_cidx = nic_i;
1137 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1138 			}
1139 			kring->nr_kflags &= ~NKR_PENDINTR;
1140 		}
1141 	}
1142 	/*
1143 	 * Second part: skip past packets that userspace has released.
1144 	 * (kring->nr_hwcur to head excluded),
1145 	 * and make the buffers available for reception.
1146 	 * As usual nm_i is the index in the netmap ring,
1147 	 * nic_i is the index in the NIC ring, and
1148 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1149 	 */
1150 	/* XXX not sure how this will work with multiple free lists */
1151 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1152 
1153 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1154 }
1155 
1156 static void
1157 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1158 {
1159 	struct ifnet *ifp = na->ifp;
1160 	if_ctx_t ctx = ifp->if_softc;
1161 
1162 	CTX_LOCK(ctx);
1163 	if (onoff) {
1164 		IFDI_INTR_ENABLE(ctx);
1165 	} else {
1166 		IFDI_INTR_DISABLE(ctx);
1167 	}
1168 	CTX_UNLOCK(ctx);
1169 }
1170 
1171 
1172 static int
1173 iflib_netmap_attach(if_ctx_t ctx)
1174 {
1175 	struct netmap_adapter na;
1176 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1177 
1178 	bzero(&na, sizeof(na));
1179 
1180 	na.ifp = ctx->ifc_ifp;
1181 	na.na_flags = NAF_BDG_MAYSLEEP;
1182 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1183 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1184 
1185 	na.num_tx_desc = scctx->isc_ntxd[0];
1186 	na.num_rx_desc = scctx->isc_nrxd[0];
1187 	na.nm_txsync = iflib_netmap_txsync;
1188 	na.nm_rxsync = iflib_netmap_rxsync;
1189 	na.nm_register = iflib_netmap_register;
1190 	na.nm_intr = iflib_netmap_intr;
1191 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1192 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1193 	return (netmap_attach(&na));
1194 }
1195 
1196 static void
1197 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1198 {
1199 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1200 	struct netmap_slot *slot;
1201 
1202 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1203 	if (slot == NULL)
1204 		return;
1205 	if (txq->ift_sds.ifsd_map == NULL)
1206 		return;
1207 
1208 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1209 
1210 		/*
1211 		 * In netmap mode, set the map for the packet buffer.
1212 		 * NOTE: Some drivers (not this one) also need to set
1213 		 * the physical buffer address in the NIC ring.
1214 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1215 		 * netmap slot index, si
1216 		 */
1217 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1218 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1219 	}
1220 }
1221 
1222 static void
1223 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1224 {
1225 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1226 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1227 	struct netmap_slot *slot;
1228 	uint32_t nm_i;
1229 
1230 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1231 	if (slot == NULL)
1232 		return;
1233 	nm_i = netmap_idx_n2k(kring, 0);
1234 	netmap_fl_refill(rxq, kring, nm_i, true);
1235 }
1236 
1237 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1238 
1239 #else
1240 #define iflib_netmap_txq_init(ctx, txq)
1241 #define iflib_netmap_rxq_init(ctx, rxq)
1242 #define iflib_netmap_detach(ifp)
1243 
1244 #define iflib_netmap_attach(ctx) (0)
1245 #define netmap_rx_irq(ifp, qid, budget) (0)
1246 #define netmap_tx_irq(ifp, qid) do {} while (0)
1247 
1248 #endif
1249 
1250 #if defined(__i386__) || defined(__amd64__)
1251 static __inline void
1252 prefetch(void *x)
1253 {
1254 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1255 }
1256 static __inline void
1257 prefetch2cachelines(void *x)
1258 {
1259 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1260 #if (CACHE_LINE_SIZE < 128)
1261 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1262 #endif
1263 }
1264 #else
1265 #define prefetch(x)
1266 #define prefetch2cachelines(x)
1267 #endif
1268 
1269 static void
1270 iflib_gen_mac(if_ctx_t ctx)
1271 {
1272 	struct thread *td;
1273 	MD5_CTX mdctx;
1274 	char uuid[HOSTUUIDLEN+1];
1275 	char buf[HOSTUUIDLEN+16];
1276 	uint8_t *mac;
1277 	unsigned char digest[16];
1278 
1279 	td = curthread;
1280 	mac = ctx->ifc_mac;
1281 	uuid[HOSTUUIDLEN] = 0;
1282 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1283 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1284 	/*
1285 	 * Generate a pseudo-random, deterministic MAC
1286 	 * address based on the UUID and unit number.
1287 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1288 	 */
1289 	MD5Init(&mdctx);
1290 	MD5Update(&mdctx, buf, strlen(buf));
1291 	MD5Final(digest, &mdctx);
1292 
1293 	mac[0] = 0x58;
1294 	mac[1] = 0x9C;
1295 	mac[2] = 0xFC;
1296 	mac[3] = digest[0];
1297 	mac[4] = digest[1];
1298 	mac[5] = digest[2];
1299 }
1300 
1301 static void
1302 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1303 {
1304 	iflib_fl_t fl;
1305 
1306 	fl = &rxq->ifr_fl[flid];
1307 	iru->iru_paddrs = fl->ifl_bus_addrs;
1308 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1309 	iru->iru_idxs = fl->ifl_rxd_idxs;
1310 	iru->iru_qsidx = rxq->ifr_id;
1311 	iru->iru_buf_size = fl->ifl_buf_size;
1312 	iru->iru_flidx = fl->ifl_id;
1313 }
1314 
1315 static void
1316 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1317 {
1318 	if (err)
1319 		return;
1320 	*(bus_addr_t *) arg = segs[0].ds_addr;
1321 }
1322 
1323 int
1324 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1325 {
1326 	int err;
1327 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1328 	device_t dev = ctx->ifc_dev;
1329 
1330 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1331 
1332 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1333 				sctx->isc_q_align, 0,	/* alignment, bounds */
1334 				BUS_SPACE_MAXADDR,	/* lowaddr */
1335 				BUS_SPACE_MAXADDR,	/* highaddr */
1336 				NULL, NULL,		/* filter, filterarg */
1337 				size,			/* maxsize */
1338 				1,			/* nsegments */
1339 				size,			/* maxsegsize */
1340 				BUS_DMA_ALLOCNOW,	/* flags */
1341 				NULL,			/* lockfunc */
1342 				NULL,			/* lockarg */
1343 				&dma->idi_tag);
1344 	if (err) {
1345 		device_printf(dev,
1346 		    "%s: bus_dma_tag_create failed: %d\n",
1347 		    __func__, err);
1348 		goto fail_0;
1349 	}
1350 
1351 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1352 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1353 	if (err) {
1354 		device_printf(dev,
1355 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1356 		    __func__, (uintmax_t)size, err);
1357 		goto fail_1;
1358 	}
1359 
1360 	dma->idi_paddr = IF_BAD_DMA;
1361 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1362 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1363 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1364 		device_printf(dev,
1365 		    "%s: bus_dmamap_load failed: %d\n",
1366 		    __func__, err);
1367 		goto fail_2;
1368 	}
1369 
1370 	dma->idi_size = size;
1371 	return (0);
1372 
1373 fail_2:
1374 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1375 fail_1:
1376 	bus_dma_tag_destroy(dma->idi_tag);
1377 fail_0:
1378 	dma->idi_tag = NULL;
1379 
1380 	return (err);
1381 }
1382 
1383 int
1384 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1385 {
1386 	int i, err;
1387 	iflib_dma_info_t *dmaiter;
1388 
1389 	dmaiter = dmalist;
1390 	for (i = 0; i < count; i++, dmaiter++) {
1391 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1392 			break;
1393 	}
1394 	if (err)
1395 		iflib_dma_free_multi(dmalist, i);
1396 	return (err);
1397 }
1398 
1399 void
1400 iflib_dma_free(iflib_dma_info_t dma)
1401 {
1402 	if (dma->idi_tag == NULL)
1403 		return;
1404 	if (dma->idi_paddr != IF_BAD_DMA) {
1405 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1406 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1407 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1408 		dma->idi_paddr = IF_BAD_DMA;
1409 	}
1410 	if (dma->idi_vaddr != NULL) {
1411 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1412 		dma->idi_vaddr = NULL;
1413 	}
1414 	bus_dma_tag_destroy(dma->idi_tag);
1415 	dma->idi_tag = NULL;
1416 }
1417 
1418 void
1419 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1420 {
1421 	int i;
1422 	iflib_dma_info_t *dmaiter = dmalist;
1423 
1424 	for (i = 0; i < count; i++, dmaiter++)
1425 		iflib_dma_free(*dmaiter);
1426 }
1427 
1428 #ifdef EARLY_AP_STARTUP
1429 static const int iflib_started = 1;
1430 #else
1431 /*
1432  * We used to abuse the smp_started flag to decide if the queues have been
1433  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1434  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1435  * is set.  Run a SYSINIT() strictly after that to just set a usable
1436  * completion flag.
1437  */
1438 
1439 static int iflib_started;
1440 
1441 static void
1442 iflib_record_started(void *arg)
1443 {
1444 	iflib_started = 1;
1445 }
1446 
1447 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1448 	iflib_record_started, NULL);
1449 #endif
1450 
1451 static int
1452 iflib_fast_intr(void *arg)
1453 {
1454 	iflib_filter_info_t info = arg;
1455 	struct grouptask *gtask = info->ifi_task;
1456 	if (!iflib_started)
1457 		return (FILTER_HANDLED);
1458 
1459 	DBG_COUNTER_INC(fast_intrs);
1460 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1461 		return (FILTER_HANDLED);
1462 
1463 	GROUPTASK_ENQUEUE(gtask);
1464 	return (FILTER_HANDLED);
1465 }
1466 
1467 static int
1468 iflib_fast_intr_rxtx(void *arg)
1469 {
1470 	iflib_filter_info_t info = arg;
1471 	struct grouptask *gtask = info->ifi_task;
1472 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1473 	if_ctx_t ctx = NULL;;
1474 	int i, cidx;
1475 
1476 	if (!iflib_started)
1477 		return (FILTER_HANDLED);
1478 
1479 	DBG_COUNTER_INC(fast_intrs);
1480 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1481 		return (FILTER_HANDLED);
1482 
1483 	MPASS(rxq->ifr_ntxqirq);
1484 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1485 		qidx_t txqid = rxq->ifr_txqid[i];
1486 
1487 		ctx = rxq->ifr_ctx;
1488 
1489 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1490 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1491 			continue;
1492 		}
1493 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1494 	}
1495 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1496 		cidx = rxq->ifr_cq_cidx;
1497 	else
1498 		cidx = rxq->ifr_fl[0].ifl_cidx;
1499 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1500 		GROUPTASK_ENQUEUE(gtask);
1501 	else
1502 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1503 	return (FILTER_HANDLED);
1504 }
1505 
1506 
1507 static int
1508 iflib_fast_intr_ctx(void *arg)
1509 {
1510 	iflib_filter_info_t info = arg;
1511 	struct grouptask *gtask = info->ifi_task;
1512 
1513 	if (!iflib_started)
1514 		return (FILTER_HANDLED);
1515 
1516 	DBG_COUNTER_INC(fast_intrs);
1517 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1518 		return (FILTER_HANDLED);
1519 
1520 	GROUPTASK_ENQUEUE(gtask);
1521 	return (FILTER_HANDLED);
1522 }
1523 
1524 static int
1525 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1526 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1527 		 const char *name)
1528 {
1529 	int rc, flags;
1530 	struct resource *res;
1531 	void *tag = NULL;
1532 	device_t dev = ctx->ifc_dev;
1533 
1534 	flags = RF_ACTIVE;
1535 	if (ctx->ifc_flags & IFC_LEGACY)
1536 		flags |= RF_SHAREABLE;
1537 	MPASS(rid < 512);
1538 	irq->ii_rid = rid;
1539 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1540 	if (res == NULL) {
1541 		device_printf(dev,
1542 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1543 		return (ENOMEM);
1544 	}
1545 	irq->ii_res = res;
1546 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1547 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1548 						filter, handler, arg, &tag);
1549 	if (rc != 0) {
1550 		device_printf(dev,
1551 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1552 					  rid, name ? name : "unknown", rc);
1553 		return (rc);
1554 	} else if (name)
1555 		bus_describe_intr(dev, res, tag, "%s", name);
1556 
1557 	irq->ii_tag = tag;
1558 	return (0);
1559 }
1560 
1561 
1562 /*********************************************************************
1563  *
1564  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1565  *  the information needed to transmit a packet on the wire. This is
1566  *  called only once at attach, setup is done every reset.
1567  *
1568  **********************************************************************/
1569 
1570 static int
1571 iflib_txsd_alloc(iflib_txq_t txq)
1572 {
1573 	if_ctx_t ctx = txq->ift_ctx;
1574 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1575 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1576 	device_t dev = ctx->ifc_dev;
1577 	int err, nsegments, ntsosegments;
1578 
1579 	nsegments = scctx->isc_tx_nsegments;
1580 	ntsosegments = scctx->isc_tx_tso_segments_max;
1581 	MPASS(scctx->isc_ntxd[0] > 0);
1582 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1583 	MPASS(nsegments > 0);
1584 	MPASS(ntsosegments > 0);
1585 	/*
1586 	 * Setup DMA descriptor areas.
1587 	 */
1588 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1589 			       1, 0,			/* alignment, bounds */
1590 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1591 			       BUS_SPACE_MAXADDR,	/* highaddr */
1592 			       NULL, NULL,		/* filter, filterarg */
1593 			       sctx->isc_tx_maxsize,		/* maxsize */
1594 			       nsegments,	/* nsegments */
1595 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1596 			       0,			/* flags */
1597 			       NULL,			/* lockfunc */
1598 			       NULL,			/* lockfuncarg */
1599 			       &txq->ift_desc_tag))) {
1600 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1601 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1602 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1603 		goto fail;
1604 	}
1605 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1606 			       1, 0,			/* alignment, bounds */
1607 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1608 			       BUS_SPACE_MAXADDR,	/* highaddr */
1609 			       NULL, NULL,		/* filter, filterarg */
1610 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1611 			       ntsosegments,	/* nsegments */
1612 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1613 			       0,			/* flags */
1614 			       NULL,			/* lockfunc */
1615 			       NULL,			/* lockfuncarg */
1616 			       &txq->ift_tso_desc_tag))) {
1617 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1618 
1619 		goto fail;
1620 	}
1621 	if (!(txq->ift_sds.ifsd_flags =
1622 	    (uint8_t *) malloc(sizeof(uint8_t) *
1623 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1624 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1625 		err = ENOMEM;
1626 		goto fail;
1627 	}
1628 	if (!(txq->ift_sds.ifsd_m =
1629 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1630 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1631 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1632 		err = ENOMEM;
1633 		goto fail;
1634 	}
1635 
1636         /* Create the descriptor buffer dma maps */
1637 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1638 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1639 		return (0);
1640 
1641 	if (!(txq->ift_sds.ifsd_map =
1642 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1643 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1644 		err = ENOMEM;
1645 		goto fail;
1646 	}
1647 
1648 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1649 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1650 		if (err != 0) {
1651 			device_printf(dev, "Unable to create TX DMA map\n");
1652 			goto fail;
1653 		}
1654 	}
1655 #endif
1656 	return (0);
1657 fail:
1658 	/* We free all, it handles case where we are in the middle */
1659 	iflib_tx_structures_free(ctx);
1660 	return (err);
1661 }
1662 
1663 static void
1664 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1665 {
1666 	bus_dmamap_t map;
1667 
1668 	map = NULL;
1669 	if (txq->ift_sds.ifsd_map != NULL)
1670 		map = txq->ift_sds.ifsd_map[i];
1671 	if (map != NULL) {
1672 		bus_dmamap_unload(txq->ift_desc_tag, map);
1673 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1674 		txq->ift_sds.ifsd_map[i] = NULL;
1675 	}
1676 }
1677 
1678 static void
1679 iflib_txq_destroy(iflib_txq_t txq)
1680 {
1681 	if_ctx_t ctx = txq->ift_ctx;
1682 
1683 	for (int i = 0; i < txq->ift_size; i++)
1684 		iflib_txsd_destroy(ctx, txq, i);
1685 	if (txq->ift_sds.ifsd_map != NULL) {
1686 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1687 		txq->ift_sds.ifsd_map = NULL;
1688 	}
1689 	if (txq->ift_sds.ifsd_m != NULL) {
1690 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1691 		txq->ift_sds.ifsd_m = NULL;
1692 	}
1693 	if (txq->ift_sds.ifsd_flags != NULL) {
1694 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1695 		txq->ift_sds.ifsd_flags = NULL;
1696 	}
1697 	if (txq->ift_desc_tag != NULL) {
1698 		bus_dma_tag_destroy(txq->ift_desc_tag);
1699 		txq->ift_desc_tag = NULL;
1700 	}
1701 	if (txq->ift_tso_desc_tag != NULL) {
1702 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1703 		txq->ift_tso_desc_tag = NULL;
1704 	}
1705 }
1706 
1707 static void
1708 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1709 {
1710 	struct mbuf **mp;
1711 
1712 	mp = &txq->ift_sds.ifsd_m[i];
1713 	if (*mp == NULL)
1714 		return;
1715 
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		bus_dmamap_sync(txq->ift_desc_tag,
1718 				txq->ift_sds.ifsd_map[i],
1719 				BUS_DMASYNC_POSTWRITE);
1720 		bus_dmamap_unload(txq->ift_desc_tag,
1721 				  txq->ift_sds.ifsd_map[i]);
1722 	}
1723 	m_free(*mp);
1724 	DBG_COUNTER_INC(tx_frees);
1725 	*mp = NULL;
1726 }
1727 
1728 static int
1729 iflib_txq_setup(iflib_txq_t txq)
1730 {
1731 	if_ctx_t ctx = txq->ift_ctx;
1732 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1733 	iflib_dma_info_t di;
1734 	int i;
1735 
1736 	/* Set number of descriptors available */
1737 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1738 	/* XXX make configurable */
1739 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1740 
1741 	/* Reset indices */
1742 	txq->ift_cidx_processed = 0;
1743 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1744 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1745 
1746 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1747 		bzero((void *)di->idi_vaddr, di->idi_size);
1748 
1749 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1750 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1751 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1752 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1753 	return (0);
1754 }
1755 
1756 /*********************************************************************
1757  *
1758  *  Allocate memory for rx_buffer structures. Since we use one
1759  *  rx_buffer per received packet, the maximum number of rx_buffer's
1760  *  that we'll need is equal to the number of receive descriptors
1761  *  that we've allocated.
1762  *
1763  **********************************************************************/
1764 static int
1765 iflib_rxsd_alloc(iflib_rxq_t rxq)
1766 {
1767 	if_ctx_t ctx = rxq->ifr_ctx;
1768 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1769 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1770 	device_t dev = ctx->ifc_dev;
1771 	iflib_fl_t fl;
1772 	int			err;
1773 
1774 	MPASS(scctx->isc_nrxd[0] > 0);
1775 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1776 
1777 	fl = rxq->ifr_fl;
1778 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1779 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1780 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1781 					 1, 0,			/* alignment, bounds */
1782 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1783 					 BUS_SPACE_MAXADDR,	/* highaddr */
1784 					 NULL, NULL,		/* filter, filterarg */
1785 					 sctx->isc_rx_maxsize,	/* maxsize */
1786 					 sctx->isc_rx_nsegments,	/* nsegments */
1787 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1788 					 0,			/* flags */
1789 					 NULL,			/* lockfunc */
1790 					 NULL,			/* lockarg */
1791 					 &fl->ifl_desc_tag);
1792 		if (err) {
1793 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1794 				__func__, err);
1795 			goto fail;
1796 		}
1797 		if (!(fl->ifl_sds.ifsd_flags =
1798 		      (uint8_t *) malloc(sizeof(uint8_t) *
1799 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1800 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1801 			err = ENOMEM;
1802 			goto fail;
1803 		}
1804 		if (!(fl->ifl_sds.ifsd_m =
1805 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1806 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1807 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1808 			err = ENOMEM;
1809 			goto fail;
1810 		}
1811 		if (!(fl->ifl_sds.ifsd_cl =
1812 		      (caddr_t *) malloc(sizeof(caddr_t) *
1813 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1814 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1815 			err = ENOMEM;
1816 			goto fail;
1817 		}
1818 
1819 		/* Create the descriptor buffer dma maps */
1820 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1821 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1822 			continue;
1823 
1824 		if (!(fl->ifl_sds.ifsd_map =
1825 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1826 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1827 			err = ENOMEM;
1828 			goto fail;
1829 		}
1830 
1831 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1832 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1833 			if (err != 0) {
1834 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1835 				goto fail;
1836 			}
1837 		}
1838 #endif
1839 	}
1840 	return (0);
1841 
1842 fail:
1843 	iflib_rx_structures_free(ctx);
1844 	return (err);
1845 }
1846 
1847 
1848 /*
1849  * Internal service routines
1850  */
1851 
1852 struct rxq_refill_cb_arg {
1853 	int               error;
1854 	bus_dma_segment_t seg;
1855 	int               nseg;
1856 };
1857 
1858 static void
1859 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1860 {
1861 	struct rxq_refill_cb_arg *cb_arg = arg;
1862 
1863 	cb_arg->error = error;
1864 	cb_arg->seg = segs[0];
1865 	cb_arg->nseg = nseg;
1866 }
1867 
1868 
1869 #ifdef ACPI_DMAR
1870 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1871 #else
1872 #define IS_DMAR(ctx) (0)
1873 #endif
1874 
1875 /**
1876  *	rxq_refill - refill an rxq  free-buffer list
1877  *	@ctx: the iflib context
1878  *	@rxq: the free-list to refill
1879  *	@n: the number of new buffers to allocate
1880  *
1881  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1882  *	The caller must assure that @n does not exceed the queue's capacity.
1883  */
1884 static void
1885 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1886 {
1887 	struct mbuf *m;
1888 	int idx, frag_idx = fl->ifl_fragidx;
1889         int pidx = fl->ifl_pidx;
1890 	caddr_t cl, *sd_cl;
1891 	struct mbuf **sd_m;
1892 	uint8_t *sd_flags;
1893 	struct if_rxd_update iru;
1894 	bus_dmamap_t *sd_map;
1895 	int n, i = 0;
1896 	uint64_t bus_addr;
1897 	int err;
1898 	qidx_t credits;
1899 
1900 	sd_m = fl->ifl_sds.ifsd_m;
1901 	sd_map = fl->ifl_sds.ifsd_map;
1902 	sd_cl = fl->ifl_sds.ifsd_cl;
1903 	sd_flags = fl->ifl_sds.ifsd_flags;
1904 	idx = pidx;
1905 	credits = fl->ifl_credits;
1906 
1907 	n  = count;
1908 	MPASS(n > 0);
1909 	MPASS(credits + n <= fl->ifl_size);
1910 
1911 	if (pidx < fl->ifl_cidx)
1912 		MPASS(pidx + n <= fl->ifl_cidx);
1913 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1914 		MPASS(fl->ifl_gen == 0);
1915 	if (pidx > fl->ifl_cidx)
1916 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1917 
1918 	DBG_COUNTER_INC(fl_refills);
1919 	if (n > 8)
1920 		DBG_COUNTER_INC(fl_refills_large);
1921 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1922 	while (n--) {
1923 		/*
1924 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1925 		 * initialized after rx.
1926 		 *
1927 		 * If the cluster is still set then we know a minimum sized packet was received
1928 		 */
1929 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1930 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1931                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1932 		if ((cl = sd_cl[frag_idx]) == NULL) {
1933                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1934 				break;
1935 #if MEMORY_LOGGING
1936 			fl->ifl_cl_enqueued++;
1937 #endif
1938 		}
1939 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1940 			break;
1941 		}
1942 #if MEMORY_LOGGING
1943 		fl->ifl_m_enqueued++;
1944 #endif
1945 
1946 		DBG_COUNTER_INC(rx_allocs);
1947 #if defined(__i386__) || defined(__amd64__)
1948 		if (!IS_DMAR(ctx)) {
1949 			bus_addr = pmap_kextract((vm_offset_t)cl);
1950 		} else
1951 #endif
1952 		{
1953 			struct rxq_refill_cb_arg cb_arg;
1954 
1955 			cb_arg.error = 0;
1956 			MPASS(sd_map != NULL);
1957 			MPASS(sd_map[frag_idx] != NULL);
1958 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1959 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1960 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1961 					BUS_DMASYNC_PREREAD);
1962 
1963 			if (err != 0 || cb_arg.error) {
1964 				/*
1965 				 * !zone_pack ?
1966 				 */
1967 				if (fl->ifl_zone == zone_pack)
1968 					uma_zfree(fl->ifl_zone, cl);
1969 				m_free(m);
1970 				n = 0;
1971 				goto done;
1972 			}
1973 			bus_addr = cb_arg.seg.ds_addr;
1974 		}
1975                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1976 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1977 
1978 		MPASS(sd_m[frag_idx] == NULL);
1979 		sd_cl[frag_idx] = cl;
1980 		sd_m[frag_idx] = m;
1981 		fl->ifl_rxd_idxs[i] = frag_idx;
1982 		fl->ifl_bus_addrs[i] = bus_addr;
1983 		fl->ifl_vm_addrs[i] = cl;
1984 		credits++;
1985 		i++;
1986 		MPASS(credits <= fl->ifl_size);
1987 		if (++idx == fl->ifl_size) {
1988 			fl->ifl_gen = 1;
1989 			idx = 0;
1990 		}
1991 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1992 			iru.iru_pidx = pidx;
1993 			iru.iru_count = i;
1994 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1995 			i = 0;
1996 			pidx = idx;
1997 			fl->ifl_pidx = idx;
1998 			fl->ifl_credits = credits;
1999 		}
2000 
2001 	}
2002 done:
2003 	if (i) {
2004 		iru.iru_pidx = pidx;
2005 		iru.iru_count = i;
2006 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2007 		fl->ifl_pidx = idx;
2008 		fl->ifl_credits = credits;
2009 	}
2010 	DBG_COUNTER_INC(rxd_flush);
2011 	if (fl->ifl_pidx == 0)
2012 		pidx = fl->ifl_size - 1;
2013 	else
2014 		pidx = fl->ifl_pidx - 1;
2015 
2016 	if (sd_map)
2017 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2018 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2019 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2020 	fl->ifl_fragidx = frag_idx;
2021 }
2022 
2023 static __inline void
2024 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2025 {
2026 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2027 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2028 #ifdef INVARIANTS
2029 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2030 #endif
2031 
2032 	MPASS(fl->ifl_credits <= fl->ifl_size);
2033 	MPASS(reclaimable == delta);
2034 
2035 	if (reclaimable > 0)
2036 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2037 }
2038 
2039 static void
2040 iflib_fl_bufs_free(iflib_fl_t fl)
2041 {
2042 	iflib_dma_info_t idi = fl->ifl_ifdi;
2043 	uint32_t i;
2044 
2045 	for (i = 0; i < fl->ifl_size; i++) {
2046 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2047 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2048 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2049 
2050 		if (*sd_flags & RX_SW_DESC_INUSE) {
2051 			if (fl->ifl_sds.ifsd_map != NULL) {
2052 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2053 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2054 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2055 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2056 			}
2057 			if (*sd_m != NULL) {
2058 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2059 				uma_zfree(zone_mbuf, *sd_m);
2060 			}
2061 			if (*sd_cl != NULL)
2062 				uma_zfree(fl->ifl_zone, *sd_cl);
2063 			*sd_flags = 0;
2064 		} else {
2065 			MPASS(*sd_cl == NULL);
2066 			MPASS(*sd_m == NULL);
2067 		}
2068 #if MEMORY_LOGGING
2069 		fl->ifl_m_dequeued++;
2070 		fl->ifl_cl_dequeued++;
2071 #endif
2072 		*sd_cl = NULL;
2073 		*sd_m = NULL;
2074 	}
2075 #ifdef INVARIANTS
2076 	for (i = 0; i < fl->ifl_size; i++) {
2077 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2078 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2079 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2080 	}
2081 #endif
2082 	/*
2083 	 * Reset free list values
2084 	 */
2085 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2086 	bzero(idi->idi_vaddr, idi->idi_size);
2087 }
2088 
2089 /*********************************************************************
2090  *
2091  *  Initialize a receive ring and its buffers.
2092  *
2093  **********************************************************************/
2094 static int
2095 iflib_fl_setup(iflib_fl_t fl)
2096 {
2097 	iflib_rxq_t rxq = fl->ifl_rxq;
2098 	if_ctx_t ctx = rxq->ifr_ctx;
2099 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2100 
2101 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2102 	/*
2103 	** Free current RX buffer structs and their mbufs
2104 	*/
2105 	iflib_fl_bufs_free(fl);
2106 	/* Now replenish the mbufs */
2107 	MPASS(fl->ifl_credits == 0);
2108 	/*
2109 	 * XXX don't set the max_frame_size to larger
2110 	 * than the hardware can handle
2111 	 */
2112 	if (sctx->isc_max_frame_size <= 2048)
2113 		fl->ifl_buf_size = MCLBYTES;
2114 #ifndef CONTIGMALLOC_WORKS
2115 	else
2116 		fl->ifl_buf_size = MJUMPAGESIZE;
2117 #else
2118 	else if (sctx->isc_max_frame_size <= 4096)
2119 		fl->ifl_buf_size = MJUMPAGESIZE;
2120 	else if (sctx->isc_max_frame_size <= 9216)
2121 		fl->ifl_buf_size = MJUM9BYTES;
2122 	else
2123 		fl->ifl_buf_size = MJUM16BYTES;
2124 #endif
2125 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2126 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2127 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2128 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2129 
2130 
2131 	/* avoid pre-allocating zillions of clusters to an idle card
2132 	 * potentially speeding up attach
2133 	 */
2134 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2135 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2136 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2137 		return (ENOBUFS);
2138 	/*
2139 	 * handle failure
2140 	 */
2141 	MPASS(rxq != NULL);
2142 	MPASS(fl->ifl_ifdi != NULL);
2143 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2144 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2145 	return (0);
2146 }
2147 
2148 /*********************************************************************
2149  *
2150  *  Free receive ring data structures
2151  *
2152  **********************************************************************/
2153 static void
2154 iflib_rx_sds_free(iflib_rxq_t rxq)
2155 {
2156 	iflib_fl_t fl;
2157 	int i;
2158 
2159 	if (rxq->ifr_fl != NULL) {
2160 		for (i = 0; i < rxq->ifr_nfl; i++) {
2161 			fl = &rxq->ifr_fl[i];
2162 			if (fl->ifl_desc_tag != NULL) {
2163 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2164 				fl->ifl_desc_tag = NULL;
2165 			}
2166 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2167 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2168 			/* XXX destroy maps first */
2169 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2170 			fl->ifl_sds.ifsd_m = NULL;
2171 			fl->ifl_sds.ifsd_cl = NULL;
2172 			fl->ifl_sds.ifsd_map = NULL;
2173 		}
2174 		free(rxq->ifr_fl, M_IFLIB);
2175 		rxq->ifr_fl = NULL;
2176 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2177 	}
2178 }
2179 
2180 /*
2181  * MI independent logic
2182  *
2183  */
2184 static void
2185 iflib_timer(void *arg)
2186 {
2187 	iflib_txq_t txq = arg;
2188 	if_ctx_t ctx = txq->ift_ctx;
2189 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2190 
2191 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2192 		return;
2193 	/*
2194 	** Check on the state of the TX queue(s), this
2195 	** can be done without the lock because its RO
2196 	** and the HUNG state will be static if set.
2197 	*/
2198 	IFDI_TIMER(ctx, txq->ift_id);
2199 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2200 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2201 	     (sctx->isc_pause_frames == 0)))
2202 		goto hung;
2203 
2204 	if (ifmp_ring_is_stalled(txq->ift_br))
2205 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2206 	txq->ift_cleaned_prev = txq->ift_cleaned;
2207 	/* handle any laggards */
2208 	if (txq->ift_db_pending)
2209 		GROUPTASK_ENQUEUE(&txq->ift_task);
2210 
2211 	sctx->isc_pause_frames = 0;
2212 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2213 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2214 	return;
2215  hung:
2216 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2217 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2218 	STATE_LOCK(ctx);
2219 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2220 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2221 	iflib_admin_intr_deferred(ctx);
2222 	STATE_UNLOCK(ctx);
2223 }
2224 
2225 static void
2226 iflib_init_locked(if_ctx_t ctx)
2227 {
2228 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2229 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2230 	if_t ifp = ctx->ifc_ifp;
2231 	iflib_fl_t fl;
2232 	iflib_txq_t txq;
2233 	iflib_rxq_t rxq;
2234 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2235 
2236 
2237 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2238 	IFDI_INTR_DISABLE(ctx);
2239 
2240 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2241 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2242 	/* Set hardware offload abilities */
2243 	if_clearhwassist(ifp);
2244 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2245 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2246 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2247 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2248 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2249 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2250 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2251 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2252 
2253 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2254 		CALLOUT_LOCK(txq);
2255 		callout_stop(&txq->ift_timer);
2256 		CALLOUT_UNLOCK(txq);
2257 		iflib_netmap_txq_init(ctx, txq);
2258 	}
2259 #ifdef INVARIANTS
2260 	i = if_getdrvflags(ifp);
2261 #endif
2262 	IFDI_INIT(ctx);
2263 	MPASS(if_getdrvflags(ifp) == i);
2264 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2265 		/* XXX this should really be done on a per-queue basis */
2266 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2267 			MPASS(rxq->ifr_id == i);
2268 			iflib_netmap_rxq_init(ctx, rxq);
2269 			continue;
2270 		}
2271 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2272 			if (iflib_fl_setup(fl)) {
2273 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2274 				goto done;
2275 			}
2276 		}
2277 	}
2278 	done:
2279 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2280 	IFDI_INTR_ENABLE(ctx);
2281 	txq = ctx->ifc_txqs;
2282 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2283 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2284 			txq->ift_timer.c_cpu);
2285 }
2286 
2287 static int
2288 iflib_media_change(if_t ifp)
2289 {
2290 	if_ctx_t ctx = if_getsoftc(ifp);
2291 	int err;
2292 
2293 	CTX_LOCK(ctx);
2294 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2295 		iflib_init_locked(ctx);
2296 	CTX_UNLOCK(ctx);
2297 	return (err);
2298 }
2299 
2300 static void
2301 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2302 {
2303 	if_ctx_t ctx = if_getsoftc(ifp);
2304 
2305 	CTX_LOCK(ctx);
2306 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2307 	IFDI_MEDIA_STATUS(ctx, ifmr);
2308 	CTX_UNLOCK(ctx);
2309 }
2310 
2311 void
2312 iflib_stop(if_ctx_t ctx)
2313 {
2314 	iflib_txq_t txq = ctx->ifc_txqs;
2315 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2316 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2317 	iflib_dma_info_t di;
2318 	iflib_fl_t fl;
2319 	int i, j;
2320 
2321 	/* Tell the stack that the interface is no longer active */
2322 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2323 
2324 	IFDI_INTR_DISABLE(ctx);
2325 	DELAY(1000);
2326 	IFDI_STOP(ctx);
2327 	DELAY(1000);
2328 
2329 	iflib_debug_reset();
2330 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2331 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2332 		/* make sure all transmitters have completed before proceeding XXX */
2333 
2334 		CALLOUT_LOCK(txq);
2335 		callout_stop(&txq->ift_timer);
2336 		CALLOUT_UNLOCK(txq);
2337 
2338 		/* clean any enqueued buffers */
2339 		iflib_ifmp_purge(txq);
2340 		/* Free any existing tx buffers. */
2341 		for (j = 0; j < txq->ift_size; j++) {
2342 			iflib_txsd_free(ctx, txq, j);
2343 		}
2344 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2345 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2346 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2347 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2348 		txq->ift_pullups = 0;
2349 		ifmp_ring_reset_stats(txq->ift_br);
2350 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2351 			bzero((void *)di->idi_vaddr, di->idi_size);
2352 	}
2353 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2354 		/* make sure all transmitters have completed before proceeding XXX */
2355 
2356 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2357 			bzero((void *)di->idi_vaddr, di->idi_size);
2358 		/* also resets the free lists pidx/cidx */
2359 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2360 			iflib_fl_bufs_free(fl);
2361 	}
2362 }
2363 
2364 static inline caddr_t
2365 calc_next_rxd(iflib_fl_t fl, int cidx)
2366 {
2367 	qidx_t size;
2368 	int nrxd;
2369 	caddr_t start, end, cur, next;
2370 
2371 	nrxd = fl->ifl_size;
2372 	size = fl->ifl_rxd_size;
2373 	start = fl->ifl_ifdi->idi_vaddr;
2374 
2375 	if (__predict_false(size == 0))
2376 		return (start);
2377 	cur = start + size*cidx;
2378 	end = start + size*nrxd;
2379 	next = CACHE_PTR_NEXT(cur);
2380 	return (next < end ? next : start);
2381 }
2382 
2383 static inline void
2384 prefetch_pkts(iflib_fl_t fl, int cidx)
2385 {
2386 	int nextptr;
2387 	int nrxd = fl->ifl_size;
2388 	caddr_t next_rxd;
2389 
2390 
2391 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2392 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2393 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2394 	next_rxd = calc_next_rxd(fl, cidx);
2395 	prefetch(next_rxd);
2396 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2397 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2398 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2399 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2400 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2401 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2402 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2403 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2404 }
2405 
2406 static void
2407 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2408 {
2409 	int flid, cidx;
2410 	bus_dmamap_t map;
2411 	iflib_fl_t fl;
2412 	iflib_dma_info_t di;
2413 	int next;
2414 
2415 	map = NULL;
2416 	flid = irf->irf_flid;
2417 	cidx = irf->irf_idx;
2418 	fl = &rxq->ifr_fl[flid];
2419 	sd->ifsd_fl = fl;
2420 	sd->ifsd_cidx = cidx;
2421 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2422 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2423 	fl->ifl_credits--;
2424 #if MEMORY_LOGGING
2425 	fl->ifl_m_dequeued++;
2426 #endif
2427 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2428 		prefetch_pkts(fl, cidx);
2429 	if (fl->ifl_sds.ifsd_map != NULL) {
2430 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2431 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2432 		map = fl->ifl_sds.ifsd_map[cidx];
2433 		di = fl->ifl_ifdi;
2434 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2435 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2436 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2437 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2438 
2439 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2440 		MPASS(fl->ifl_cidx == cidx);
2441 		if (unload)
2442 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2443 	}
2444 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2445 	if (__predict_false(fl->ifl_cidx == 0))
2446 		fl->ifl_gen = 0;
2447 	if (map != NULL)
2448 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2449 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2450         bit_clear(fl->ifl_rx_bitmap, cidx);
2451 }
2452 
2453 static struct mbuf *
2454 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2455 {
2456 	int i, padlen , flags;
2457 	struct mbuf *m, *mh, *mt;
2458 	caddr_t cl;
2459 
2460 	i = 0;
2461 	mh = NULL;
2462 	do {
2463 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2464 
2465 		MPASS(*sd->ifsd_cl != NULL);
2466 		MPASS(*sd->ifsd_m != NULL);
2467 
2468 		/* Don't include zero-length frags */
2469 		if (ri->iri_frags[i].irf_len == 0) {
2470 			/* XXX we can save the cluster here, but not the mbuf */
2471 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2472 			m_free(*sd->ifsd_m);
2473 			*sd->ifsd_m = NULL;
2474 			continue;
2475 		}
2476 		m = *sd->ifsd_m;
2477 		*sd->ifsd_m = NULL;
2478 		if (mh == NULL) {
2479 			flags = M_PKTHDR|M_EXT;
2480 			mh = mt = m;
2481 			padlen = ri->iri_pad;
2482 		} else {
2483 			flags = M_EXT;
2484 			mt->m_next = m;
2485 			mt = m;
2486 			/* assuming padding is only on the first fragment */
2487 			padlen = 0;
2488 		}
2489 		cl = *sd->ifsd_cl;
2490 		*sd->ifsd_cl = NULL;
2491 
2492 		/* Can these two be made one ? */
2493 		m_init(m, M_NOWAIT, MT_DATA, flags);
2494 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2495 		/*
2496 		 * These must follow m_init and m_cljset
2497 		 */
2498 		m->m_data += padlen;
2499 		ri->iri_len -= padlen;
2500 		m->m_len = ri->iri_frags[i].irf_len;
2501 	} while (++i < ri->iri_nfrags);
2502 
2503 	return (mh);
2504 }
2505 
2506 /*
2507  * Process one software descriptor
2508  */
2509 static struct mbuf *
2510 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2511 {
2512 	struct if_rxsd sd;
2513 	struct mbuf *m;
2514 
2515 	/* should I merge this back in now that the two paths are basically duplicated? */
2516 	if (ri->iri_nfrags == 1 &&
2517 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2518 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2519 		m = *sd.ifsd_m;
2520 		*sd.ifsd_m = NULL;
2521 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2522 #ifndef __NO_STRICT_ALIGNMENT
2523 		if (!IP_ALIGNED(m))
2524 			m->m_data += 2;
2525 #endif
2526 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2527 		m->m_len = ri->iri_frags[0].irf_len;
2528        } else {
2529 		m = assemble_segments(rxq, ri, &sd);
2530 	}
2531 	m->m_pkthdr.len = ri->iri_len;
2532 	m->m_pkthdr.rcvif = ri->iri_ifp;
2533 	m->m_flags |= ri->iri_flags;
2534 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2535 	m->m_pkthdr.flowid = ri->iri_flowid;
2536 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2537 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2538 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2539 	return (m);
2540 }
2541 
2542 #if defined(INET6) || defined(INET)
2543 static void
2544 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2545 {
2546 	CURVNET_SET(lc->ifp->if_vnet);
2547 #if defined(INET6)
2548 	*v6 = VNET(ip6_forwarding);
2549 #endif
2550 #if defined(INET)
2551 	*v4 = VNET(ipforwarding);
2552 #endif
2553 	CURVNET_RESTORE();
2554 }
2555 
2556 /*
2557  * Returns true if it's possible this packet could be LROed.
2558  * if it returns false, it is guaranteed that tcp_lro_rx()
2559  * would not return zero.
2560  */
2561 static bool
2562 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2563 {
2564 	struct ether_header *eh;
2565 	uint16_t eh_type;
2566 
2567 	eh = mtod(m, struct ether_header *);
2568 	eh_type = ntohs(eh->ether_type);
2569 	switch (eh_type) {
2570 #if defined(INET6)
2571 		case ETHERTYPE_IPV6:
2572 			return !v6_forwarding;
2573 #endif
2574 #if defined (INET)
2575 		case ETHERTYPE_IP:
2576 			return !v4_forwarding;
2577 #endif
2578 	}
2579 
2580 	return false;
2581 }
2582 #else
2583 static void
2584 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2585 {
2586 }
2587 #endif
2588 
2589 static bool
2590 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2591 {
2592 	if_ctx_t ctx = rxq->ifr_ctx;
2593 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2594 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2595 	int avail, i;
2596 	qidx_t *cidxp;
2597 	struct if_rxd_info ri;
2598 	int err, budget_left, rx_bytes, rx_pkts;
2599 	iflib_fl_t fl;
2600 	struct ifnet *ifp;
2601 	int lro_enabled;
2602 	bool v4_forwarding, v6_forwarding, lro_possible;
2603 
2604 	/*
2605 	 * XXX early demux data packets so that if_input processing only handles
2606 	 * acks in interrupt context
2607 	 */
2608 	struct mbuf *m, *mh, *mt, *mf;
2609 
2610 	lro_possible = v4_forwarding = v6_forwarding = false;
2611 	ifp = ctx->ifc_ifp;
2612 	mh = mt = NULL;
2613 	MPASS(budget > 0);
2614 	rx_pkts	= rx_bytes = 0;
2615 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2616 		cidxp = &rxq->ifr_cq_cidx;
2617 	else
2618 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2619 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2620 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2621 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2622 		DBG_COUNTER_INC(rx_unavail);
2623 		return (false);
2624 	}
2625 
2626 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2627 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2628 			DBG_COUNTER_INC(rx_ctx_inactive);
2629 			break;
2630 		}
2631 		/*
2632 		 * Reset client set fields to their default values
2633 		 */
2634 		rxd_info_zero(&ri);
2635 		ri.iri_qsidx = rxq->ifr_id;
2636 		ri.iri_cidx = *cidxp;
2637 		ri.iri_ifp = ifp;
2638 		ri.iri_frags = rxq->ifr_frags;
2639 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2640 
2641 		if (err)
2642 			goto err;
2643 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2644 			*cidxp = ri.iri_cidx;
2645 			/* Update our consumer index */
2646 			/* XXX NB: shurd - check if this is still safe */
2647 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2648 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2649 				rxq->ifr_cq_gen = 0;
2650 			}
2651 			/* was this only a completion queue message? */
2652 			if (__predict_false(ri.iri_nfrags == 0))
2653 				continue;
2654 		}
2655 		MPASS(ri.iri_nfrags != 0);
2656 		MPASS(ri.iri_len != 0);
2657 
2658 		/* will advance the cidx on the corresponding free lists */
2659 		m = iflib_rxd_pkt_get(rxq, &ri);
2660 		if (avail == 0 && budget_left)
2661 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2662 
2663 		if (__predict_false(m == NULL)) {
2664 			DBG_COUNTER_INC(rx_mbuf_null);
2665 			continue;
2666 		}
2667 		/* imm_pkt: -- cxgb */
2668 		if (mh == NULL)
2669 			mh = mt = m;
2670 		else {
2671 			mt->m_nextpkt = m;
2672 			mt = m;
2673 		}
2674 	}
2675 	/* make sure that we can refill faster than drain */
2676 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2677 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2678 
2679 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2680 	if (lro_enabled)
2681 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2682 	mt = mf = NULL;
2683 	while (mh != NULL) {
2684 		m = mh;
2685 		mh = mh->m_nextpkt;
2686 		m->m_nextpkt = NULL;
2687 #ifndef __NO_STRICT_ALIGNMENT
2688 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2689 			continue;
2690 #endif
2691 		rx_bytes += m->m_pkthdr.len;
2692 		rx_pkts++;
2693 #if defined(INET6) || defined(INET)
2694 		if (lro_enabled) {
2695 			if (!lro_possible) {
2696 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2697 				if (lro_possible && mf != NULL) {
2698 					ifp->if_input(ifp, mf);
2699 					DBG_COUNTER_INC(rx_if_input);
2700 					mt = mf = NULL;
2701 				}
2702 			}
2703 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2704 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2705 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2706 					continue;
2707 			}
2708 		}
2709 #endif
2710 		if (lro_possible) {
2711 			ifp->if_input(ifp, m);
2712 			DBG_COUNTER_INC(rx_if_input);
2713 			continue;
2714 		}
2715 
2716 		if (mf == NULL)
2717 			mf = m;
2718 		if (mt != NULL)
2719 			mt->m_nextpkt = m;
2720 		mt = m;
2721 	}
2722 	if (mf != NULL) {
2723 		ifp->if_input(ifp, mf);
2724 		DBG_COUNTER_INC(rx_if_input);
2725 	}
2726 
2727 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2728 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2729 
2730 	/*
2731 	 * Flush any outstanding LRO work
2732 	 */
2733 #if defined(INET6) || defined(INET)
2734 	tcp_lro_flush_all(&rxq->ifr_lc);
2735 #endif
2736 	if (avail)
2737 		return true;
2738 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2739 err:
2740 	STATE_LOCK(ctx);
2741 	ctx->ifc_flags |= IFC_DO_RESET;
2742 	iflib_admin_intr_deferred(ctx);
2743 	STATE_UNLOCK(ctx);
2744 	return (false);
2745 }
2746 
2747 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2748 static inline qidx_t
2749 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2750 {
2751 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2752 	qidx_t minthresh = txq->ift_size / 8;
2753 	if (in_use > 4*minthresh)
2754 		return (notify_count);
2755 	if (in_use > 2*minthresh)
2756 		return (notify_count >> 1);
2757 	if (in_use > minthresh)
2758 		return (notify_count >> 3);
2759 	return (0);
2760 }
2761 
2762 static inline qidx_t
2763 txq_max_rs_deferred(iflib_txq_t txq)
2764 {
2765 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2766 	qidx_t minthresh = txq->ift_size / 8;
2767 	if (txq->ift_in_use > 4*minthresh)
2768 		return (notify_count);
2769 	if (txq->ift_in_use > 2*minthresh)
2770 		return (notify_count >> 1);
2771 	if (txq->ift_in_use > minthresh)
2772 		return (notify_count >> 2);
2773 	return (2);
2774 }
2775 
2776 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2777 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2778 
2779 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2780 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2781 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2782 
2783 /* forward compatibility for cxgb */
2784 #define FIRST_QSET(ctx) 0
2785 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2786 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2787 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2788 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2789 
2790 /* XXX we should be setting this to something other than zero */
2791 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2792 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2793 
2794 static inline bool
2795 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2796 {
2797 	qidx_t dbval, max;
2798 	bool rang;
2799 
2800 	rang = false;
2801 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2802 	if (ring || txq->ift_db_pending >= max) {
2803 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2804 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2805 		txq->ift_db_pending = txq->ift_npending = 0;
2806 		rang = true;
2807 	}
2808 	return (rang);
2809 }
2810 
2811 #ifdef PKT_DEBUG
2812 static void
2813 print_pkt(if_pkt_info_t pi)
2814 {
2815 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2816 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2817 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2818 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2819 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2820 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2821 }
2822 #endif
2823 
2824 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2825 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2826 
2827 static int
2828 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2829 {
2830 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2831 	struct ether_vlan_header *eh;
2832 	struct mbuf *m, *n;
2833 
2834 	n = m = *mp;
2835 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2836 	    M_WRITABLE(m) == 0) {
2837 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2838 			return (ENOMEM);
2839 		} else {
2840 			m_freem(*mp);
2841 			n = *mp = m;
2842 		}
2843 	}
2844 
2845 	/*
2846 	 * Determine where frame payload starts.
2847 	 * Jump over vlan headers if already present,
2848 	 * helpful for QinQ too.
2849 	 */
2850 	if (__predict_false(m->m_len < sizeof(*eh))) {
2851 		txq->ift_pullups++;
2852 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2853 			return (ENOMEM);
2854 	}
2855 	eh = mtod(m, struct ether_vlan_header *);
2856 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2857 		pi->ipi_etype = ntohs(eh->evl_proto);
2858 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2859 	} else {
2860 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2861 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2862 	}
2863 
2864 	switch (pi->ipi_etype) {
2865 #ifdef INET
2866 	case ETHERTYPE_IP:
2867 	{
2868 		struct ip *ip = NULL;
2869 		struct tcphdr *th = NULL;
2870 		int minthlen;
2871 
2872 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2873 		if (__predict_false(m->m_len < minthlen)) {
2874 			/*
2875 			 * if this code bloat is causing too much of a hit
2876 			 * move it to a separate function and mark it noinline
2877 			 */
2878 			if (m->m_len == pi->ipi_ehdrlen) {
2879 				n = m->m_next;
2880 				MPASS(n);
2881 				if (n->m_len >= sizeof(*ip))  {
2882 					ip = (struct ip *)n->m_data;
2883 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2884 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2885 				} else {
2886 					txq->ift_pullups++;
2887 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2888 						return (ENOMEM);
2889 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2890 				}
2891 			} else {
2892 				txq->ift_pullups++;
2893 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2894 					return (ENOMEM);
2895 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2896 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2897 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2898 			}
2899 		} else {
2900 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2901 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2902 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2903 		}
2904 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2905 		pi->ipi_ipproto = ip->ip_p;
2906 		pi->ipi_flags |= IPI_TX_IPV4;
2907 
2908 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2909                        ip->ip_sum = 0;
2910 
2911 		if (IS_TSO4(pi)) {
2912 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2913 				if (__predict_false(th == NULL)) {
2914 					txq->ift_pullups++;
2915 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2916 						return (ENOMEM);
2917 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2918 				}
2919 				pi->ipi_tcp_hflags = th->th_flags;
2920 				pi->ipi_tcp_hlen = th->th_off << 2;
2921 				pi->ipi_tcp_seq = th->th_seq;
2922 			}
2923 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2924 				return (ENXIO);
2925 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2926 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2927 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2928 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2929 				ip->ip_sum = 0;
2930 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2931 			}
2932 		}
2933 		break;
2934 	}
2935 #endif
2936 #ifdef INET6
2937 	case ETHERTYPE_IPV6:
2938 	{
2939 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2940 		struct tcphdr *th;
2941 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2942 
2943 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2944 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2945 				return (ENOMEM);
2946 		}
2947 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2948 
2949 		/* XXX-BZ this will go badly in case of ext hdrs. */
2950 		pi->ipi_ipproto = ip6->ip6_nxt;
2951 		pi->ipi_flags |= IPI_TX_IPV6;
2952 
2953 		if (IS_TSO6(pi)) {
2954 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2955 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2956 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2957 						return (ENOMEM);
2958 				}
2959 				pi->ipi_tcp_hflags = th->th_flags;
2960 				pi->ipi_tcp_hlen = th->th_off << 2;
2961 			}
2962 
2963 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2964 				return (ENXIO);
2965 			/*
2966 			 * The corresponding flag is set by the stack in the IPv4
2967 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2968 			 * So, set it here because the rest of the flow requires it.
2969 			 */
2970 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2971 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2972 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2973 		}
2974 		break;
2975 	}
2976 #endif
2977 	default:
2978 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2979 		pi->ipi_ip_hlen = 0;
2980 		break;
2981 	}
2982 	*mp = m;
2983 
2984 	return (0);
2985 }
2986 
2987 static  __noinline  struct mbuf *
2988 collapse_pkthdr(struct mbuf *m0)
2989 {
2990 	struct mbuf *m, *m_next, *tmp;
2991 
2992 	m = m0;
2993 	m_next = m->m_next;
2994 	while (m_next != NULL && m_next->m_len == 0) {
2995 		m = m_next;
2996 		m->m_next = NULL;
2997 		m_free(m);
2998 		m_next = m_next->m_next;
2999 	}
3000 	m = m0;
3001 	m->m_next = m_next;
3002 	if ((m_next->m_flags & M_EXT) == 0) {
3003 		m = m_defrag(m, M_NOWAIT);
3004 	} else {
3005 		tmp = m_next->m_next;
3006 		memcpy(m_next, m, MPKTHSIZE);
3007 		m = m_next;
3008 		m->m_next = tmp;
3009 	}
3010 	return (m);
3011 }
3012 
3013 /*
3014  * If dodgy hardware rejects the scatter gather chain we've handed it
3015  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3016  * m_defrag'd mbufs
3017  */
3018 static __noinline struct mbuf *
3019 iflib_remove_mbuf(iflib_txq_t txq)
3020 {
3021 	int ntxd, i, pidx;
3022 	struct mbuf *m, *mh, **ifsd_m;
3023 
3024 	pidx = txq->ift_pidx;
3025 	ifsd_m = txq->ift_sds.ifsd_m;
3026 	ntxd = txq->ift_size;
3027 	mh = m = ifsd_m[pidx];
3028 	ifsd_m[pidx] = NULL;
3029 #if MEMORY_LOGGING
3030 	txq->ift_dequeued++;
3031 #endif
3032 	i = 1;
3033 
3034 	while (m) {
3035 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3036 #if MEMORY_LOGGING
3037 		txq->ift_dequeued++;
3038 #endif
3039 		m = m->m_next;
3040 		i++;
3041 	}
3042 	return (mh);
3043 }
3044 
3045 static int
3046 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3047 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3048 			  int max_segs, int flags)
3049 {
3050 	if_ctx_t ctx;
3051 	if_shared_ctx_t		sctx;
3052 	if_softc_ctx_t		scctx;
3053 	int i, next, pidx, err, ntxd, count;
3054 	struct mbuf *m, *tmp, **ifsd_m;
3055 
3056 	m = *m0;
3057 
3058 	/*
3059 	 * Please don't ever do this
3060 	 */
3061 	if (__predict_false(m->m_len == 0))
3062 		*m0 = m = collapse_pkthdr(m);
3063 
3064 	ctx = txq->ift_ctx;
3065 	sctx = ctx->ifc_sctx;
3066 	scctx = &ctx->ifc_softc_ctx;
3067 	ifsd_m = txq->ift_sds.ifsd_m;
3068 	ntxd = txq->ift_size;
3069 	pidx = txq->ift_pidx;
3070 	if (map != NULL) {
3071 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3072 
3073 		err = bus_dmamap_load_mbuf_sg(tag, map,
3074 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3075 		if (err)
3076 			return (err);
3077 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3078 		count = 0;
3079 		m = *m0;
3080 		do {
3081 			if (__predict_false(m->m_len <= 0)) {
3082 				tmp = m;
3083 				m = m->m_next;
3084 				tmp->m_next = NULL;
3085 				m_free(tmp);
3086 				continue;
3087 			}
3088 			m = m->m_next;
3089 			count++;
3090 		} while (m != NULL);
3091 		if (count > *nsegs) {
3092 			ifsd_m[pidx] = *m0;
3093 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3094 			return (0);
3095 		}
3096 		m = *m0;
3097 		count = 0;
3098 		do {
3099 			next = (pidx + count) & (ntxd-1);
3100 			MPASS(ifsd_m[next] == NULL);
3101 			ifsd_m[next] = m;
3102 			count++;
3103 			tmp = m;
3104 			m = m->m_next;
3105 		} while (m != NULL);
3106 	} else {
3107 		int buflen, sgsize, maxsegsz, max_sgsize;
3108 		vm_offset_t vaddr;
3109 		vm_paddr_t curaddr;
3110 
3111 		count = i = 0;
3112 		m = *m0;
3113 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3114 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3115 		else
3116 			maxsegsz = sctx->isc_tx_maxsegsize;
3117 
3118 		do {
3119 			if (__predict_false(m->m_len <= 0)) {
3120 				tmp = m;
3121 				m = m->m_next;
3122 				tmp->m_next = NULL;
3123 				m_free(tmp);
3124 				continue;
3125 			}
3126 			buflen = m->m_len;
3127 			vaddr = (vm_offset_t)m->m_data;
3128 			/*
3129 			 * see if we can't be smarter about physically
3130 			 * contiguous mappings
3131 			 */
3132 			next = (pidx + count) & (ntxd-1);
3133 			MPASS(ifsd_m[next] == NULL);
3134 #if MEMORY_LOGGING
3135 			txq->ift_enqueued++;
3136 #endif
3137 			ifsd_m[next] = m;
3138 			while (buflen > 0) {
3139 				if (i >= max_segs)
3140 					goto err;
3141 				max_sgsize = MIN(buflen, maxsegsz);
3142 				curaddr = pmap_kextract(vaddr);
3143 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3144 				sgsize = MIN(sgsize, max_sgsize);
3145 				segs[i].ds_addr = curaddr;
3146 				segs[i].ds_len = sgsize;
3147 				vaddr += sgsize;
3148 				buflen -= sgsize;
3149 				i++;
3150 			}
3151 			count++;
3152 			tmp = m;
3153 			m = m->m_next;
3154 		} while (m != NULL);
3155 		*nsegs = i;
3156 	}
3157 	return (0);
3158 err:
3159 	*m0 = iflib_remove_mbuf(txq);
3160 	return (EFBIG);
3161 }
3162 
3163 static inline caddr_t
3164 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3165 {
3166 	qidx_t size;
3167 	int ntxd;
3168 	caddr_t start, end, cur, next;
3169 
3170 	ntxd = txq->ift_size;
3171 	size = txq->ift_txd_size[qid];
3172 	start = txq->ift_ifdi[qid].idi_vaddr;
3173 
3174 	if (__predict_false(size == 0))
3175 		return (start);
3176 	cur = start + size*cidx;
3177 	end = start + size*ntxd;
3178 	next = CACHE_PTR_NEXT(cur);
3179 	return (next < end ? next : start);
3180 }
3181 
3182 /*
3183  * Pad an mbuf to ensure a minimum ethernet frame size.
3184  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3185  */
3186 static __noinline int
3187 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3188 {
3189 	/*
3190 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3191 	 * and ARP message is the smallest common payload I can think of
3192 	 */
3193 	static char pad[18];	/* just zeros */
3194 	int n;
3195 	struct mbuf *new_head;
3196 
3197 	if (!M_WRITABLE(*m_head)) {
3198 		new_head = m_dup(*m_head, M_NOWAIT);
3199 		if (new_head == NULL) {
3200 			m_freem(*m_head);
3201 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3202 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3203 			return ENOMEM;
3204 		}
3205 		m_freem(*m_head);
3206 		*m_head = new_head;
3207 	}
3208 
3209 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3210 	     n > 0; n -= sizeof(pad))
3211 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3212 			break;
3213 
3214 	if (n > 0) {
3215 		m_freem(*m_head);
3216 		device_printf(dev, "cannot pad short frame\n");
3217 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3218 		return (ENOBUFS);
3219 	}
3220 
3221 	return 0;
3222 }
3223 
3224 static int
3225 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3226 {
3227 	if_ctx_t		ctx;
3228 	if_shared_ctx_t		sctx;
3229 	if_softc_ctx_t		scctx;
3230 	bus_dma_segment_t	*segs;
3231 	struct mbuf		*m_head;
3232 	void			*next_txd;
3233 	bus_dmamap_t		map;
3234 	struct if_pkt_info	pi;
3235 	int remap = 0;
3236 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3237 	bus_dma_tag_t desc_tag;
3238 
3239 	segs = txq->ift_segs;
3240 	ctx = txq->ift_ctx;
3241 	sctx = ctx->ifc_sctx;
3242 	scctx = &ctx->ifc_softc_ctx;
3243 	segs = txq->ift_segs;
3244 	ntxd = txq->ift_size;
3245 	m_head = *m_headp;
3246 	map = NULL;
3247 
3248 	/*
3249 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3250 	 */
3251 	cidx = txq->ift_cidx;
3252 	pidx = txq->ift_pidx;
3253 	if (ctx->ifc_flags & IFC_PREFETCH) {
3254 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3255 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3256 			next_txd = calc_next_txd(txq, cidx, 0);
3257 			prefetch(next_txd);
3258 		}
3259 
3260 		/* prefetch the next cache line of mbuf pointers and flags */
3261 		prefetch(&txq->ift_sds.ifsd_m[next]);
3262 		if (txq->ift_sds.ifsd_map != NULL) {
3263 			prefetch(&txq->ift_sds.ifsd_map[next]);
3264 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3265 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3266 		}
3267 	} else if (txq->ift_sds.ifsd_map != NULL)
3268 		map = txq->ift_sds.ifsd_map[pidx];
3269 
3270 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3271 		desc_tag = txq->ift_tso_desc_tag;
3272 		max_segs = scctx->isc_tx_tso_segments_max;
3273 	} else {
3274 		desc_tag = txq->ift_desc_tag;
3275 		max_segs = scctx->isc_tx_nsegments;
3276 	}
3277 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3278 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3279 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3280 		if (err)
3281 			return err;
3282 	}
3283 	m_head = *m_headp;
3284 
3285 	pkt_info_zero(&pi);
3286 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3287 	pi.ipi_pidx = pidx;
3288 	pi.ipi_qsidx = txq->ift_id;
3289 	pi.ipi_len = m_head->m_pkthdr.len;
3290 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3291 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3292 
3293 	/* deliberate bitwise OR to make one condition */
3294 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3295 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3296 			return (err);
3297 		m_head = *m_headp;
3298 	}
3299 
3300 retry:
3301 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3302 defrag:
3303 	if (__predict_false(err)) {
3304 		switch (err) {
3305 		case EFBIG:
3306 			/* try collapse once and defrag once */
3307 			if (remap == 0) {
3308 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3309 				/* try defrag if collapsing fails */
3310 				if (m_head == NULL)
3311 					remap++;
3312 			}
3313 			if (remap == 1)
3314 				m_head = m_defrag(*m_headp, M_NOWAIT);
3315 			remap++;
3316 			if (__predict_false(m_head == NULL))
3317 				goto defrag_failed;
3318 			txq->ift_mbuf_defrag++;
3319 			*m_headp = m_head;
3320 			goto retry;
3321 			break;
3322 		case ENOMEM:
3323 			txq->ift_no_tx_dma_setup++;
3324 			break;
3325 		default:
3326 			txq->ift_no_tx_dma_setup++;
3327 			m_freem(*m_headp);
3328 			DBG_COUNTER_INC(tx_frees);
3329 			*m_headp = NULL;
3330 			break;
3331 		}
3332 		txq->ift_map_failed++;
3333 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3334 		return (err);
3335 	}
3336 
3337 	/*
3338 	 * XXX assumes a 1 to 1 relationship between segments and
3339 	 *        descriptors - this does not hold true on all drivers, e.g.
3340 	 *        cxgb
3341 	 */
3342 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3343 		txq->ift_no_desc_avail++;
3344 		if (map != NULL)
3345 			bus_dmamap_unload(desc_tag, map);
3346 		DBG_COUNTER_INC(encap_txq_avail_fail);
3347 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3348 			GROUPTASK_ENQUEUE(&txq->ift_task);
3349 		return (ENOBUFS);
3350 	}
3351 	/*
3352 	 * On Intel cards we can greatly reduce the number of TX interrupts
3353 	 * we see by only setting report status on every Nth descriptor.
3354 	 * However, this also means that the driver will need to keep track
3355 	 * of the descriptors that RS was set on to check them for the DD bit.
3356 	 */
3357 	txq->ift_rs_pending += nsegs + 1;
3358 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3359 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3360 		pi.ipi_flags |= IPI_TX_INTR;
3361 		txq->ift_rs_pending = 0;
3362 	}
3363 
3364 	pi.ipi_segs = segs;
3365 	pi.ipi_nsegs = nsegs;
3366 
3367 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3368 #ifdef PKT_DEBUG
3369 	print_pkt(&pi);
3370 #endif
3371 	if (map != NULL)
3372 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3373 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3374 		if (map != NULL)
3375 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3376 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3377 		DBG_COUNTER_INC(tx_encap);
3378 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3379 
3380 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3381 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3382 			ndesc += txq->ift_size;
3383 			txq->ift_gen = 1;
3384 		}
3385 		/*
3386 		 * drivers can need as many as
3387 		 * two sentinels
3388 		 */
3389 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3390 		MPASS(pi.ipi_new_pidx != pidx);
3391 		MPASS(ndesc > 0);
3392 		txq->ift_in_use += ndesc;
3393 
3394 		/*
3395 		 * We update the last software descriptor again here because there may
3396 		 * be a sentinel and/or there may be more mbufs than segments
3397 		 */
3398 		txq->ift_pidx = pi.ipi_new_pidx;
3399 		txq->ift_npending += pi.ipi_ndescs;
3400 	} else {
3401 		*m_headp = m_head = iflib_remove_mbuf(txq);
3402 		if (err == EFBIG) {
3403 			txq->ift_txd_encap_efbig++;
3404 			if (remap < 2) {
3405 				remap = 1;
3406 				goto defrag;
3407 			}
3408 		}
3409 		DBG_COUNTER_INC(encap_txd_encap_fail);
3410 		goto defrag_failed;
3411 	}
3412 	return (err);
3413 
3414 defrag_failed:
3415 	txq->ift_mbuf_defrag_failed++;
3416 	txq->ift_map_failed++;
3417 	m_freem(*m_headp);
3418 	DBG_COUNTER_INC(tx_frees);
3419 	*m_headp = NULL;
3420 	return (ENOMEM);
3421 }
3422 
3423 static void
3424 iflib_tx_desc_free(iflib_txq_t txq, int n)
3425 {
3426 	int hasmap;
3427 	uint32_t qsize, cidx, mask, gen;
3428 	struct mbuf *m, **ifsd_m;
3429 	uint8_t *ifsd_flags;
3430 	bus_dmamap_t *ifsd_map;
3431 	bool do_prefetch;
3432 
3433 	cidx = txq->ift_cidx;
3434 	gen = txq->ift_gen;
3435 	qsize = txq->ift_size;
3436 	mask = qsize-1;
3437 	hasmap = txq->ift_sds.ifsd_map != NULL;
3438 	ifsd_flags = txq->ift_sds.ifsd_flags;
3439 	ifsd_m = txq->ift_sds.ifsd_m;
3440 	ifsd_map = txq->ift_sds.ifsd_map;
3441 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3442 
3443 	while (n-- > 0) {
3444 		if (do_prefetch) {
3445 			prefetch(ifsd_m[(cidx + 3) & mask]);
3446 			prefetch(ifsd_m[(cidx + 4) & mask]);
3447 		}
3448 		if (ifsd_m[cidx] != NULL) {
3449 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3450 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3451 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3452 				/*
3453 				 * does it matter if it's not the TSO tag? If so we'll
3454 				 * have to add the type to flags
3455 				 */
3456 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3457 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3458 			}
3459 			if ((m = ifsd_m[cidx]) != NULL) {
3460 				/* XXX we don't support any drivers that batch packets yet */
3461 				MPASS(m->m_nextpkt == NULL);
3462 				/* if the number of clusters exceeds the number of segments
3463 				 * there won't be space on the ring to save a pointer to each
3464 				 * cluster so we simply free the list here
3465 				 */
3466 				if (m->m_flags & M_TOOBIG) {
3467 					m_freem(m);
3468 				} else {
3469 					m_free(m);
3470 				}
3471 				ifsd_m[cidx] = NULL;
3472 #if MEMORY_LOGGING
3473 				txq->ift_dequeued++;
3474 #endif
3475 				DBG_COUNTER_INC(tx_frees);
3476 			}
3477 		}
3478 		if (__predict_false(++cidx == qsize)) {
3479 			cidx = 0;
3480 			gen = 0;
3481 		}
3482 	}
3483 	txq->ift_cidx = cidx;
3484 	txq->ift_gen = gen;
3485 }
3486 
3487 static __inline int
3488 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3489 {
3490 	int reclaim;
3491 	if_ctx_t ctx = txq->ift_ctx;
3492 
3493 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3494 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3495 
3496 	/*
3497 	 * Need a rate-limiting check so that this isn't called every time
3498 	 */
3499 	iflib_tx_credits_update(ctx, txq);
3500 	reclaim = DESC_RECLAIMABLE(txq);
3501 
3502 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3503 #ifdef INVARIANTS
3504 		if (iflib_verbose_debug) {
3505 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3506 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3507 			       reclaim, thresh);
3508 
3509 		}
3510 #endif
3511 		return (0);
3512 	}
3513 	iflib_tx_desc_free(txq, reclaim);
3514 	txq->ift_cleaned += reclaim;
3515 	txq->ift_in_use -= reclaim;
3516 
3517 	return (reclaim);
3518 }
3519 
3520 static struct mbuf **
3521 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3522 {
3523 	int next, size;
3524 	struct mbuf **items;
3525 
3526 	size = r->size;
3527 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3528 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3529 
3530 	prefetch(items[(cidx + offset) & (size-1)]);
3531 	if (remaining > 1) {
3532 		prefetch2cachelines(&items[next]);
3533 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3534 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3535 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3536 	}
3537 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3538 }
3539 
3540 static void
3541 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3542 {
3543 
3544 	ifmp_ring_check_drainage(txq->ift_br, budget);
3545 }
3546 
3547 static uint32_t
3548 iflib_txq_can_drain(struct ifmp_ring *r)
3549 {
3550 	iflib_txq_t txq = r->cookie;
3551 	if_ctx_t ctx = txq->ift_ctx;
3552 
3553 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3554 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3555 }
3556 
3557 static uint32_t
3558 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3559 {
3560 	iflib_txq_t txq = r->cookie;
3561 	if_ctx_t ctx = txq->ift_ctx;
3562 	struct ifnet *ifp = ctx->ifc_ifp;
3563 	struct mbuf **mp, *m;
3564 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3565 	int reclaimed, err, in_use_prev, desc_used;
3566 	bool do_prefetch, ring, rang;
3567 
3568 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3569 			    !LINK_ACTIVE(ctx))) {
3570 		DBG_COUNTER_INC(txq_drain_notready);
3571 		return (0);
3572 	}
3573 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3574 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3575 	avail = IDXDIFF(pidx, cidx, r->size);
3576 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3577 		DBG_COUNTER_INC(txq_drain_flushing);
3578 		for (i = 0; i < avail; i++) {
3579 			m_free(r->items[(cidx + i) & (r->size-1)]);
3580 			r->items[(cidx + i) & (r->size-1)] = NULL;
3581 		}
3582 		return (avail);
3583 	}
3584 
3585 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3586 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3587 		CALLOUT_LOCK(txq);
3588 		callout_stop(&txq->ift_timer);
3589 		CALLOUT_UNLOCK(txq);
3590 		DBG_COUNTER_INC(txq_drain_oactive);
3591 		return (0);
3592 	}
3593 	if (reclaimed)
3594 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3595 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3596 	count = MIN(avail, TX_BATCH_SIZE);
3597 #ifdef INVARIANTS
3598 	if (iflib_verbose_debug)
3599 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3600 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3601 #endif
3602 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3603 	avail = TXQ_AVAIL(txq);
3604 	err = 0;
3605 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3606 		int rem = do_prefetch ? count - i : 0;
3607 
3608 		mp = _ring_peek_one(r, cidx, i, rem);
3609 		MPASS(mp != NULL && *mp != NULL);
3610 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3611 			consumed++;
3612 			reclaimed++;
3613 			continue;
3614 		}
3615 		in_use_prev = txq->ift_in_use;
3616 		err = iflib_encap(txq, mp);
3617 		if (__predict_false(err)) {
3618 			DBG_COUNTER_INC(txq_drain_encapfail);
3619 			/* no room - bail out */
3620 			if (err == ENOBUFS)
3621 				break;
3622 			consumed++;
3623 			DBG_COUNTER_INC(txq_drain_encapfail);
3624 			/* we can't send this packet - skip it */
3625 			continue;
3626 		}
3627 		consumed++;
3628 		pkt_sent++;
3629 		m = *mp;
3630 		DBG_COUNTER_INC(tx_sent);
3631 		bytes_sent += m->m_pkthdr.len;
3632 		mcast_sent += !!(m->m_flags & M_MCAST);
3633 		avail = TXQ_AVAIL(txq);
3634 
3635 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3636 		desc_used += (txq->ift_in_use - in_use_prev);
3637 		ETHER_BPF_MTAP(ifp, m);
3638 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3639 			break;
3640 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3641 	}
3642 
3643 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3644 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3645 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3646 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3647 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3648 	if (mcast_sent)
3649 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3650 #ifdef INVARIANTS
3651 	if (iflib_verbose_debug)
3652 		printf("consumed=%d\n", consumed);
3653 #endif
3654 	return (consumed);
3655 }
3656 
3657 static uint32_t
3658 iflib_txq_drain_always(struct ifmp_ring *r)
3659 {
3660 	return (1);
3661 }
3662 
3663 static uint32_t
3664 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3665 {
3666 	int i, avail;
3667 	struct mbuf **mp;
3668 	iflib_txq_t txq;
3669 
3670 	txq = r->cookie;
3671 
3672 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3673 	CALLOUT_LOCK(txq);
3674 	callout_stop(&txq->ift_timer);
3675 	CALLOUT_UNLOCK(txq);
3676 
3677 	avail = IDXDIFF(pidx, cidx, r->size);
3678 	for (i = 0; i < avail; i++) {
3679 		mp = _ring_peek_one(r, cidx, i, avail - i);
3680 		if (__predict_false(*mp == (struct mbuf *)txq))
3681 			continue;
3682 		m_freem(*mp);
3683 	}
3684 	MPASS(ifmp_ring_is_stalled(r) == 0);
3685 	return (avail);
3686 }
3687 
3688 static void
3689 iflib_ifmp_purge(iflib_txq_t txq)
3690 {
3691 	struct ifmp_ring *r;
3692 
3693 	r = txq->ift_br;
3694 	r->drain = iflib_txq_drain_free;
3695 	r->can_drain = iflib_txq_drain_always;
3696 
3697 	ifmp_ring_check_drainage(r, r->size);
3698 
3699 	r->drain = iflib_txq_drain;
3700 	r->can_drain = iflib_txq_can_drain;
3701 }
3702 
3703 static void
3704 _task_fn_tx(void *context)
3705 {
3706 	iflib_txq_t txq = context;
3707 	if_ctx_t ctx = txq->ift_ctx;
3708 	struct ifnet *ifp = ctx->ifc_ifp;
3709 
3710 #ifdef IFLIB_DIAGNOSTICS
3711 	txq->ift_cpu_exec_count[curcpu]++;
3712 #endif
3713 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3714 		return;
3715 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3716 		/*
3717 		 * If there are no available credits, and TX IRQs are not in use,
3718 		 * re-schedule the task immediately.
3719 		 */
3720 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3721 			netmap_tx_irq(ifp, txq->ift_id);
3722 		else {
3723 #ifdef DEV_NETMAP
3724 			if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
3725 				struct netmap_kring *kring = NA(ctx->ifc_ifp)->tx_rings[txq->ift_id];
3726 
3727 				if (kring->nr_hwtail != nm_prev(kring->rhead, kring->nkr_num_slots - 1))
3728 					GROUPTASK_ENQUEUE(&txq->ift_task);
3729 			}
3730 #endif
3731 		}
3732 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3733 		return;
3734 	}
3735 	if (txq->ift_db_pending)
3736 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3737 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3738 	if (ctx->ifc_flags & IFC_LEGACY)
3739 		IFDI_INTR_ENABLE(ctx);
3740 	else {
3741 #ifdef INVARIANTS
3742 		int rc =
3743 #endif
3744 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3745 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3746 	}
3747 }
3748 
3749 static void
3750 _task_fn_rx(void *context)
3751 {
3752 	iflib_rxq_t rxq = context;
3753 	if_ctx_t ctx = rxq->ifr_ctx;
3754 	bool more;
3755 	uint16_t budget;
3756 
3757 #ifdef IFLIB_DIAGNOSTICS
3758 	rxq->ifr_cpu_exec_count[curcpu]++;
3759 #endif
3760 	DBG_COUNTER_INC(task_fn_rxs);
3761 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3762 		return;
3763 	more = true;
3764 #ifdef DEV_NETMAP
3765 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3766 		u_int work = 0;
3767 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3768 			more = false;
3769 		}
3770 	}
3771 #endif
3772 	budget = ctx->ifc_sysctl_rx_budget;
3773 	if (budget == 0)
3774 		budget = 16;	/* XXX */
3775 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3776 		if (ctx->ifc_flags & IFC_LEGACY)
3777 			IFDI_INTR_ENABLE(ctx);
3778 		else {
3779 #ifdef INVARIANTS
3780 			int rc =
3781 #endif
3782 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3783 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3784 			DBG_COUNTER_INC(rx_intr_enables);
3785 		}
3786 	}
3787 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3788 		return;
3789 	if (more)
3790 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3791 }
3792 
3793 static void
3794 _task_fn_admin(void *context)
3795 {
3796 	if_ctx_t ctx = context;
3797 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3798 	iflib_txq_t txq;
3799 	int i;
3800 	bool oactive, running, do_reset, do_watchdog;
3801 
3802 	STATE_LOCK(ctx);
3803 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3804 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3805 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3806 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3807 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3808 	STATE_UNLOCK(ctx);
3809 
3810 	if ((!running & !oactive) &&
3811 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3812 		return;
3813 
3814 	CTX_LOCK(ctx);
3815 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3816 		CALLOUT_LOCK(txq);
3817 		callout_stop(&txq->ift_timer);
3818 		CALLOUT_UNLOCK(txq);
3819 	}
3820 	if (do_watchdog) {
3821 		ctx->ifc_watchdog_events++;
3822 		IFDI_WATCHDOG_RESET(ctx);
3823 	}
3824 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3825 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3826 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3827 	IFDI_LINK_INTR_ENABLE(ctx);
3828 	if (do_reset)
3829 		iflib_if_init_locked(ctx);
3830 	CTX_UNLOCK(ctx);
3831 
3832 	if (LINK_ACTIVE(ctx) == 0)
3833 		return;
3834 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3835 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3836 }
3837 
3838 
3839 static void
3840 _task_fn_iov(void *context)
3841 {
3842 	if_ctx_t ctx = context;
3843 
3844 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3845 		return;
3846 
3847 	CTX_LOCK(ctx);
3848 	IFDI_VFLR_HANDLE(ctx);
3849 	CTX_UNLOCK(ctx);
3850 }
3851 
3852 static int
3853 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3854 {
3855 	int err;
3856 	if_int_delay_info_t info;
3857 	if_ctx_t ctx;
3858 
3859 	info = (if_int_delay_info_t)arg1;
3860 	ctx = info->iidi_ctx;
3861 	info->iidi_req = req;
3862 	info->iidi_oidp = oidp;
3863 	CTX_LOCK(ctx);
3864 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3865 	CTX_UNLOCK(ctx);
3866 	return (err);
3867 }
3868 
3869 /*********************************************************************
3870  *
3871  *  IFNET FUNCTIONS
3872  *
3873  **********************************************************************/
3874 
3875 static void
3876 iflib_if_init_locked(if_ctx_t ctx)
3877 {
3878 	iflib_stop(ctx);
3879 	iflib_init_locked(ctx);
3880 }
3881 
3882 
3883 static void
3884 iflib_if_init(void *arg)
3885 {
3886 	if_ctx_t ctx = arg;
3887 
3888 	CTX_LOCK(ctx);
3889 	iflib_if_init_locked(ctx);
3890 	CTX_UNLOCK(ctx);
3891 }
3892 
3893 static int
3894 iflib_if_transmit(if_t ifp, struct mbuf *m)
3895 {
3896 	if_ctx_t	ctx = if_getsoftc(ifp);
3897 
3898 	iflib_txq_t txq;
3899 	int err, qidx;
3900 
3901 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3902 		DBG_COUNTER_INC(tx_frees);
3903 		m_freem(m);
3904 		return (ENOBUFS);
3905 	}
3906 
3907 	MPASS(m->m_nextpkt == NULL);
3908 	qidx = 0;
3909 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3910 		qidx = QIDX(ctx, m);
3911 	/*
3912 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3913 	 */
3914 	txq = &ctx->ifc_txqs[qidx];
3915 
3916 #ifdef DRIVER_BACKPRESSURE
3917 	if (txq->ift_closed) {
3918 		while (m != NULL) {
3919 			next = m->m_nextpkt;
3920 			m->m_nextpkt = NULL;
3921 			m_freem(m);
3922 			m = next;
3923 		}
3924 		return (ENOBUFS);
3925 	}
3926 #endif
3927 #ifdef notyet
3928 	qidx = count = 0;
3929 	mp = marr;
3930 	next = m;
3931 	do {
3932 		count++;
3933 		next = next->m_nextpkt;
3934 	} while (next != NULL);
3935 
3936 	if (count > nitems(marr))
3937 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3938 			/* XXX check nextpkt */
3939 			m_freem(m);
3940 			/* XXX simplify for now */
3941 			DBG_COUNTER_INC(tx_frees);
3942 			return (ENOBUFS);
3943 		}
3944 	for (next = m, i = 0; next != NULL; i++) {
3945 		mp[i] = next;
3946 		next = next->m_nextpkt;
3947 		mp[i]->m_nextpkt = NULL;
3948 	}
3949 #endif
3950 	DBG_COUNTER_INC(tx_seen);
3951 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3952 
3953 	GROUPTASK_ENQUEUE(&txq->ift_task);
3954 	if (err) {
3955 		/* support forthcoming later */
3956 #ifdef DRIVER_BACKPRESSURE
3957 		txq->ift_closed = TRUE;
3958 #endif
3959 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3960 		m_freem(m);
3961 	}
3962 
3963 	return (err);
3964 }
3965 
3966 static void
3967 iflib_if_qflush(if_t ifp)
3968 {
3969 	if_ctx_t ctx = if_getsoftc(ifp);
3970 	iflib_txq_t txq = ctx->ifc_txqs;
3971 	int i;
3972 
3973 	STATE_LOCK(ctx);
3974 	ctx->ifc_flags |= IFC_QFLUSH;
3975 	STATE_UNLOCK(ctx);
3976 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3977 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3978 			iflib_txq_check_drain(txq, 0);
3979 	STATE_LOCK(ctx);
3980 	ctx->ifc_flags &= ~IFC_QFLUSH;
3981 	STATE_UNLOCK(ctx);
3982 
3983 	if_qflush(ifp);
3984 }
3985 
3986 
3987 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3988 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3989 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3990 
3991 static int
3992 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3993 {
3994 	if_ctx_t ctx = if_getsoftc(ifp);
3995 	struct ifreq	*ifr = (struct ifreq *)data;
3996 #if defined(INET) || defined(INET6)
3997 	struct ifaddr	*ifa = (struct ifaddr *)data;
3998 #endif
3999 	bool		avoid_reset = FALSE;
4000 	int		err = 0, reinit = 0, bits;
4001 
4002 	switch (command) {
4003 	case SIOCSIFADDR:
4004 #ifdef INET
4005 		if (ifa->ifa_addr->sa_family == AF_INET)
4006 			avoid_reset = TRUE;
4007 #endif
4008 #ifdef INET6
4009 		if (ifa->ifa_addr->sa_family == AF_INET6)
4010 			avoid_reset = TRUE;
4011 #endif
4012 		/*
4013 		** Calling init results in link renegotiation,
4014 		** so we avoid doing it when possible.
4015 		*/
4016 		if (avoid_reset) {
4017 			if_setflagbits(ifp, IFF_UP,0);
4018 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
4019 				reinit = 1;
4020 #ifdef INET
4021 			if (!(if_getflags(ifp) & IFF_NOARP))
4022 				arp_ifinit(ifp, ifa);
4023 #endif
4024 		} else
4025 			err = ether_ioctl(ifp, command, data);
4026 		break;
4027 	case SIOCSIFMTU:
4028 		CTX_LOCK(ctx);
4029 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4030 			CTX_UNLOCK(ctx);
4031 			break;
4032 		}
4033 		bits = if_getdrvflags(ifp);
4034 		/* stop the driver and free any clusters before proceeding */
4035 		iflib_stop(ctx);
4036 
4037 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4038 			STATE_LOCK(ctx);
4039 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4040 				ctx->ifc_flags |= IFC_MULTISEG;
4041 			else
4042 				ctx->ifc_flags &= ~IFC_MULTISEG;
4043 			STATE_UNLOCK(ctx);
4044 			err = if_setmtu(ifp, ifr->ifr_mtu);
4045 		}
4046 		iflib_init_locked(ctx);
4047 		STATE_LOCK(ctx);
4048 		if_setdrvflags(ifp, bits);
4049 		STATE_UNLOCK(ctx);
4050 		CTX_UNLOCK(ctx);
4051 		break;
4052 	case SIOCSIFFLAGS:
4053 		CTX_LOCK(ctx);
4054 		if (if_getflags(ifp) & IFF_UP) {
4055 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4056 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4057 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4058 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4059 				}
4060 			} else
4061 				reinit = 1;
4062 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4063 			iflib_stop(ctx);
4064 		}
4065 		ctx->ifc_if_flags = if_getflags(ifp);
4066 		CTX_UNLOCK(ctx);
4067 		break;
4068 	case SIOCADDMULTI:
4069 	case SIOCDELMULTI:
4070 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4071 			CTX_LOCK(ctx);
4072 			IFDI_INTR_DISABLE(ctx);
4073 			IFDI_MULTI_SET(ctx);
4074 			IFDI_INTR_ENABLE(ctx);
4075 			CTX_UNLOCK(ctx);
4076 		}
4077 		break;
4078 	case SIOCSIFMEDIA:
4079 		CTX_LOCK(ctx);
4080 		IFDI_MEDIA_SET(ctx);
4081 		CTX_UNLOCK(ctx);
4082 		/* falls thru */
4083 	case SIOCGIFMEDIA:
4084 	case SIOCGIFXMEDIA:
4085 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4086 		break;
4087 	case SIOCGI2C:
4088 	{
4089 		struct ifi2creq i2c;
4090 
4091 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4092 		if (err != 0)
4093 			break;
4094 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4095 			err = EINVAL;
4096 			break;
4097 		}
4098 		if (i2c.len > sizeof(i2c.data)) {
4099 			err = EINVAL;
4100 			break;
4101 		}
4102 
4103 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4104 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4105 			    sizeof(i2c));
4106 		break;
4107 	}
4108 	case SIOCSIFCAP:
4109 	{
4110 		int mask, setmask;
4111 
4112 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4113 		setmask = 0;
4114 #ifdef TCP_OFFLOAD
4115 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4116 #endif
4117 		setmask |= (mask & IFCAP_FLAGS);
4118 
4119 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4120 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4121 		if ((mask & IFCAP_WOL) &&
4122 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4123 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4124 		if_vlancap(ifp);
4125 		/*
4126 		 * want to ensure that traffic has stopped before we change any of the flags
4127 		 */
4128 		if (setmask) {
4129 			CTX_LOCK(ctx);
4130 			bits = if_getdrvflags(ifp);
4131 			if (bits & IFF_DRV_RUNNING)
4132 				iflib_stop(ctx);
4133 			STATE_LOCK(ctx);
4134 			if_togglecapenable(ifp, setmask);
4135 			STATE_UNLOCK(ctx);
4136 			if (bits & IFF_DRV_RUNNING)
4137 				iflib_init_locked(ctx);
4138 			STATE_LOCK(ctx);
4139 			if_setdrvflags(ifp, bits);
4140 			STATE_UNLOCK(ctx);
4141 			CTX_UNLOCK(ctx);
4142 		}
4143 		break;
4144 	    }
4145 	case SIOCGPRIVATE_0:
4146 	case SIOCSDRVSPEC:
4147 	case SIOCGDRVSPEC:
4148 		CTX_LOCK(ctx);
4149 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4150 		CTX_UNLOCK(ctx);
4151 		break;
4152 	default:
4153 		err = ether_ioctl(ifp, command, data);
4154 		break;
4155 	}
4156 	if (reinit)
4157 		iflib_if_init(ctx);
4158 	return (err);
4159 }
4160 
4161 static uint64_t
4162 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4163 {
4164 	if_ctx_t ctx = if_getsoftc(ifp);
4165 
4166 	return (IFDI_GET_COUNTER(ctx, cnt));
4167 }
4168 
4169 /*********************************************************************
4170  *
4171  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4172  *
4173  **********************************************************************/
4174 
4175 static void
4176 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4177 {
4178 	if_ctx_t ctx = if_getsoftc(ifp);
4179 
4180 	if ((void *)ctx != arg)
4181 		return;
4182 
4183 	if ((vtag == 0) || (vtag > 4095))
4184 		return;
4185 
4186 	CTX_LOCK(ctx);
4187 	IFDI_VLAN_REGISTER(ctx, vtag);
4188 	/* Re-init to load the changes */
4189 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4190 		iflib_if_init_locked(ctx);
4191 	CTX_UNLOCK(ctx);
4192 }
4193 
4194 static void
4195 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4196 {
4197 	if_ctx_t ctx = if_getsoftc(ifp);
4198 
4199 	if ((void *)ctx != arg)
4200 		return;
4201 
4202 	if ((vtag == 0) || (vtag > 4095))
4203 		return;
4204 
4205 	CTX_LOCK(ctx);
4206 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4207 	/* Re-init to load the changes */
4208 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4209 		iflib_if_init_locked(ctx);
4210 	CTX_UNLOCK(ctx);
4211 }
4212 
4213 static void
4214 iflib_led_func(void *arg, int onoff)
4215 {
4216 	if_ctx_t ctx = arg;
4217 
4218 	CTX_LOCK(ctx);
4219 	IFDI_LED_FUNC(ctx, onoff);
4220 	CTX_UNLOCK(ctx);
4221 }
4222 
4223 /*********************************************************************
4224  *
4225  *  BUS FUNCTION DEFINITIONS
4226  *
4227  **********************************************************************/
4228 
4229 int
4230 iflib_device_probe(device_t dev)
4231 {
4232 	pci_vendor_info_t *ent;
4233 
4234 	uint16_t	pci_vendor_id, pci_device_id;
4235 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4236 	uint16_t	pci_rev_id;
4237 	if_shared_ctx_t sctx;
4238 
4239 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4240 		return (ENOTSUP);
4241 
4242 	pci_vendor_id = pci_get_vendor(dev);
4243 	pci_device_id = pci_get_device(dev);
4244 	pci_subvendor_id = pci_get_subvendor(dev);
4245 	pci_subdevice_id = pci_get_subdevice(dev);
4246 	pci_rev_id = pci_get_revid(dev);
4247 	if (sctx->isc_parse_devinfo != NULL)
4248 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4249 
4250 	ent = sctx->isc_vendor_info;
4251 	while (ent->pvi_vendor_id != 0) {
4252 		if (pci_vendor_id != ent->pvi_vendor_id) {
4253 			ent++;
4254 			continue;
4255 		}
4256 		if ((pci_device_id == ent->pvi_device_id) &&
4257 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4258 		     (ent->pvi_subvendor_id == 0)) &&
4259 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4260 		     (ent->pvi_subdevice_id == 0)) &&
4261 		    ((pci_rev_id == ent->pvi_rev_id) ||
4262 		     (ent->pvi_rev_id == 0))) {
4263 
4264 			device_set_desc_copy(dev, ent->pvi_name);
4265 			/* this needs to be changed to zero if the bus probing code
4266 			 * ever stops re-probing on best match because the sctx
4267 			 * may have its values over written by register calls
4268 			 * in subsequent probes
4269 			 */
4270 			return (BUS_PROBE_DEFAULT);
4271 		}
4272 		ent++;
4273 	}
4274 	return (ENXIO);
4275 }
4276 
4277 static void
4278 iflib_reset_qvalues(if_ctx_t ctx)
4279 {
4280 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4281 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4282 	device_t dev = ctx->ifc_dev;
4283 	int i;
4284 
4285 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4286 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4287 	/*
4288 	 * XXX sanity check that ntxd & nrxd are a power of 2
4289 	 */
4290 	if (ctx->ifc_sysctl_ntxqs != 0)
4291 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4292 	if (ctx->ifc_sysctl_nrxqs != 0)
4293 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4294 
4295 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4296 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4297 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4298 		else
4299 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4300 	}
4301 
4302 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4303 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4304 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4305 		else
4306 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4307 	}
4308 
4309 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4310 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4311 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4312 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4313 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4314 		}
4315 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4316 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4317 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4318 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4319 		}
4320 	}
4321 
4322 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4323 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4324 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4325 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4326 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4327 		}
4328 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4329 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4330 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4331 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4332 		}
4333 	}
4334 }
4335 
4336 int
4337 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4338 {
4339 	int err, rid, msix;
4340 	if_ctx_t ctx;
4341 	if_t ifp;
4342 	if_softc_ctx_t scctx;
4343 	int i;
4344 	uint16_t main_txq;
4345 	uint16_t main_rxq;
4346 
4347 
4348 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4349 
4350 	if (sc == NULL) {
4351 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4352 		device_set_softc(dev, ctx);
4353 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4354 	}
4355 
4356 	ctx->ifc_sctx = sctx;
4357 	ctx->ifc_dev = dev;
4358 	ctx->ifc_softc = sc;
4359 
4360 	if ((err = iflib_register(ctx)) != 0) {
4361 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4362 			free(sc, M_IFLIB);
4363 		free(ctx, M_IFLIB);
4364 		device_printf(dev, "iflib_register failed %d\n", err);
4365 		return (err);
4366 	}
4367 	iflib_add_device_sysctl_pre(ctx);
4368 
4369 	scctx = &ctx->ifc_softc_ctx;
4370 	ifp = ctx->ifc_ifp;
4371 
4372 	iflib_reset_qvalues(ctx);
4373 	CTX_LOCK(ctx);
4374 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4375 		CTX_UNLOCK(ctx);
4376 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4377 		return (err);
4378 	}
4379 	_iflib_pre_assert(scctx);
4380 	ctx->ifc_txrx = *scctx->isc_txrx;
4381 
4382 #ifdef INVARIANTS
4383 	MPASS(scctx->isc_capenable);
4384 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4385 		MPASS(scctx->isc_tx_csum_flags);
4386 #endif
4387 
4388 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4389 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4390 
4391 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4392 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4393 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4394 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4395 
4396 #ifdef ACPI_DMAR
4397 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4398 		ctx->ifc_flags |= IFC_DMAR;
4399 #elif !(defined(__i386__) || defined(__amd64__))
4400 	/* set unconditionally for !x86 */
4401 	ctx->ifc_flags |= IFC_DMAR;
4402 #endif
4403 
4404 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4405 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4406 
4407 	/* XXX change for per-queue sizes */
4408 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4409 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4410 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4411 		if (!powerof2(scctx->isc_nrxd[i])) {
4412 			/* round down instead? */
4413 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4414 			err = EINVAL;
4415 			goto fail;
4416 		}
4417 	}
4418 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4419 		if (!powerof2(scctx->isc_ntxd[i])) {
4420 			device_printf(dev,
4421 			    "# tx descriptors must be a power of 2");
4422 			err = EINVAL;
4423 			goto fail;
4424 		}
4425 	}
4426 
4427 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4428 	    MAX_SINGLE_PACKET_FRACTION)
4429 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4430 		    MAX_SINGLE_PACKET_FRACTION);
4431 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4432 	    MAX_SINGLE_PACKET_FRACTION)
4433 		scctx->isc_tx_tso_segments_max = max(1,
4434 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4435 
4436 	/*
4437 	 * Protect the stack against modern hardware
4438 	 */
4439 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4440 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4441 
4442 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4443 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4444 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4445 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4446 	if (scctx->isc_rss_table_size == 0)
4447 		scctx->isc_rss_table_size = 64;
4448 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4449 
4450 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4451 	/* XXX format name */
4452 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4453 
4454 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4455 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4456 		device_printf(dev, "Unable to fetch CPU list\n");
4457 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4458 	}
4459 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4460 
4461 	/*
4462 	** Now setup MSI or MSI/X, should
4463 	** return us the number of supported
4464 	** vectors. (Will be 1 for MSI)
4465 	*/
4466 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4467 		msix = scctx->isc_vectors;
4468 	} else if (scctx->isc_msix_bar != 0)
4469 	       /*
4470 		* The simple fact that isc_msix_bar is not 0 does not mean we
4471 		* we have a good value there that is known to work.
4472 		*/
4473 		msix = iflib_msix_init(ctx);
4474 	else {
4475 		scctx->isc_vectors = 1;
4476 		scctx->isc_ntxqsets = 1;
4477 		scctx->isc_nrxqsets = 1;
4478 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4479 		msix = 0;
4480 	}
4481 	/* Get memory for the station queues */
4482 	if ((err = iflib_queues_alloc(ctx))) {
4483 		device_printf(dev, "Unable to allocate queue memory\n");
4484 		goto fail;
4485 	}
4486 
4487 	if ((err = iflib_qset_structures_setup(ctx)))
4488 		goto fail_queues;
4489 
4490 	/*
4491 	 * Group taskqueues aren't properly set up until SMP is started,
4492 	 * so we disable interrupts until we can handle them post
4493 	 * SI_SUB_SMP.
4494 	 *
4495 	 * XXX: disabling interrupts doesn't actually work, at least for
4496 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4497 	 * we do null handling and depend on this not causing too large an
4498 	 * interrupt storm.
4499 	 */
4500 	IFDI_INTR_DISABLE(ctx);
4501 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4502 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4503 		goto fail_intr_free;
4504 	}
4505 	if (msix <= 1) {
4506 		rid = 0;
4507 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4508 			MPASS(msix == 1);
4509 			rid = 1;
4510 		}
4511 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4512 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4513 			goto fail_intr_free;
4514 		}
4515 	}
4516 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4517 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4518 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4519 		goto fail_detach;
4520 	}
4521 	if ((err = iflib_netmap_attach(ctx))) {
4522 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4523 		goto fail_detach;
4524 	}
4525 	*ctxp = ctx;
4526 
4527 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4528 
4529 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4530 	iflib_add_device_sysctl_post(ctx);
4531 	ctx->ifc_flags |= IFC_INIT_DONE;
4532 	CTX_UNLOCK(ctx);
4533 	return (0);
4534 fail_detach:
4535 	ether_ifdetach(ctx->ifc_ifp);
4536 fail_intr_free:
4537 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4538 		pci_release_msi(ctx->ifc_dev);
4539 fail_queues:
4540 	iflib_tx_structures_free(ctx);
4541 	iflib_rx_structures_free(ctx);
4542 fail:
4543 	IFDI_DETACH(ctx);
4544 	CTX_UNLOCK(ctx);
4545 	return (err);
4546 }
4547 
4548 int
4549 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4550 					  struct iflib_cloneattach_ctx *clctx)
4551 {
4552 	int err;
4553 	if_ctx_t ctx;
4554 	if_t ifp;
4555 	if_softc_ctx_t scctx;
4556 	int i;
4557 	void *sc;
4558 	uint16_t main_txq;
4559 	uint16_t main_rxq;
4560 
4561 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4562 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4563 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4564 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4565 		ctx->ifc_flags |= IFC_PSEUDO;
4566 
4567 	ctx->ifc_sctx = sctx;
4568 	ctx->ifc_softc = sc;
4569 	ctx->ifc_dev = dev;
4570 
4571 	if ((err = iflib_register(ctx)) != 0) {
4572 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4573 		free(sc, M_IFLIB);
4574 		free(ctx, M_IFLIB);
4575 		return (err);
4576 	}
4577 	iflib_add_device_sysctl_pre(ctx);
4578 
4579 	scctx = &ctx->ifc_softc_ctx;
4580 	ifp = ctx->ifc_ifp;
4581 
4582 	/*
4583 	 * XXX sanity check that ntxd & nrxd are a power of 2
4584 	 */
4585 	iflib_reset_qvalues(ctx);
4586 
4587 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4588 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4589 		return (err);
4590 	}
4591 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4592 		iflib_gen_mac(ctx);
4593 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4594 								clctx->cc_params)) != 0) {
4595 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4596 		return (err);
4597 	}
4598 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4599 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4600 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4601 
4602 #ifdef INVARIANTS
4603 	MPASS(scctx->isc_capenable);
4604 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4605 		MPASS(scctx->isc_tx_csum_flags);
4606 #endif
4607 
4608 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4609 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4610 
4611 	ifp->if_flags |= IFF_NOGROUP;
4612 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4613 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4614 
4615 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4616 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4617 			goto fail_detach;
4618 		}
4619 		*ctxp = ctx;
4620 
4621 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4622 		iflib_add_device_sysctl_post(ctx);
4623 		ctx->ifc_flags |= IFC_INIT_DONE;
4624 		return (0);
4625 	}
4626 	_iflib_pre_assert(scctx);
4627 	ctx->ifc_txrx = *scctx->isc_txrx;
4628 
4629 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4630 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4631 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4632 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4633 
4634 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4635 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4636 
4637 	/* XXX change for per-queue sizes */
4638 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4639 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4640 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4641 		if (!powerof2(scctx->isc_nrxd[i])) {
4642 			/* round down instead? */
4643 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4644 			err = EINVAL;
4645 			goto fail;
4646 		}
4647 	}
4648 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4649 		if (!powerof2(scctx->isc_ntxd[i])) {
4650 			device_printf(dev,
4651 			    "# tx descriptors must be a power of 2");
4652 			err = EINVAL;
4653 			goto fail;
4654 		}
4655 	}
4656 
4657 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4658 	    MAX_SINGLE_PACKET_FRACTION)
4659 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4660 		    MAX_SINGLE_PACKET_FRACTION);
4661 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4662 	    MAX_SINGLE_PACKET_FRACTION)
4663 		scctx->isc_tx_tso_segments_max = max(1,
4664 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4665 
4666 	/*
4667 	 * Protect the stack against modern hardware
4668 	 */
4669 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4670 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4671 
4672 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4673 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4674 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4675 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4676 	if (scctx->isc_rss_table_size == 0)
4677 		scctx->isc_rss_table_size = 64;
4678 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4679 
4680 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4681 	/* XXX format name */
4682 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4683 
4684 	/* XXX --- can support > 1 -- but keep it simple for now */
4685 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4686 
4687 	/* Get memory for the station queues */
4688 	if ((err = iflib_queues_alloc(ctx))) {
4689 		device_printf(dev, "Unable to allocate queue memory\n");
4690 		goto fail;
4691 	}
4692 
4693 	if ((err = iflib_qset_structures_setup(ctx))) {
4694 		device_printf(dev, "qset structure setup failed %d\n", err);
4695 		goto fail_queues;
4696 	}
4697 	/*
4698 	 * XXX What if anything do we want to do about interrupts?
4699 	 */
4700 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4701 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4702 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4703 		goto fail_detach;
4704 	}
4705 	/* XXX handle more than one queue */
4706 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4707 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4708 
4709 	*ctxp = ctx;
4710 
4711 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4712 	iflib_add_device_sysctl_post(ctx);
4713 	ctx->ifc_flags |= IFC_INIT_DONE;
4714 	return (0);
4715 fail_detach:
4716 	ether_ifdetach(ctx->ifc_ifp);
4717 fail_queues:
4718 	iflib_tx_structures_free(ctx);
4719 	iflib_rx_structures_free(ctx);
4720 fail:
4721 	IFDI_DETACH(ctx);
4722 	return (err);
4723 }
4724 
4725 int
4726 iflib_pseudo_deregister(if_ctx_t ctx)
4727 {
4728 	if_t ifp = ctx->ifc_ifp;
4729 	iflib_txq_t txq;
4730 	iflib_rxq_t rxq;
4731 	int i, j;
4732 	struct taskqgroup *tqg;
4733 	iflib_fl_t fl;
4734 
4735 	/* Unregister VLAN events */
4736 	if (ctx->ifc_vlan_attach_event != NULL)
4737 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4738 	if (ctx->ifc_vlan_detach_event != NULL)
4739 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4740 
4741 	ether_ifdetach(ifp);
4742 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4743 	CTX_LOCK_DESTROY(ctx);
4744 	/* XXX drain any dependent tasks */
4745 	tqg = qgroup_if_io_tqg;
4746 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4747 		callout_drain(&txq->ift_timer);
4748 		if (txq->ift_task.gt_uniq != NULL)
4749 			taskqgroup_detach(tqg, &txq->ift_task);
4750 	}
4751 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4752 		if (rxq->ifr_task.gt_uniq != NULL)
4753 			taskqgroup_detach(tqg, &rxq->ifr_task);
4754 
4755 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4756 			free(fl->ifl_rx_bitmap, M_IFLIB);
4757 	}
4758 	tqg = qgroup_if_config_tqg;
4759 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4760 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4761 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4762 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4763 
4764 	if_free(ifp);
4765 
4766 	iflib_tx_structures_free(ctx);
4767 	iflib_rx_structures_free(ctx);
4768 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4769 		free(ctx->ifc_softc, M_IFLIB);
4770 	free(ctx, M_IFLIB);
4771 	return (0);
4772 }
4773 
4774 int
4775 iflib_device_attach(device_t dev)
4776 {
4777 	if_ctx_t ctx;
4778 	if_shared_ctx_t sctx;
4779 
4780 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4781 		return (ENOTSUP);
4782 
4783 	pci_enable_busmaster(dev);
4784 
4785 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4786 }
4787 
4788 int
4789 iflib_device_deregister(if_ctx_t ctx)
4790 {
4791 	if_t ifp = ctx->ifc_ifp;
4792 	iflib_txq_t txq;
4793 	iflib_rxq_t rxq;
4794 	device_t dev = ctx->ifc_dev;
4795 	int i, j;
4796 	struct taskqgroup *tqg;
4797 	iflib_fl_t fl;
4798 
4799 	/* Make sure VLANS are not using driver */
4800 	if (if_vlantrunkinuse(ifp)) {
4801 		device_printf(dev,"Vlan in use, detach first\n");
4802 		return (EBUSY);
4803 	}
4804 
4805 	CTX_LOCK(ctx);
4806 	ctx->ifc_in_detach = 1;
4807 	iflib_stop(ctx);
4808 	CTX_UNLOCK(ctx);
4809 
4810 	/* Unregister VLAN events */
4811 	if (ctx->ifc_vlan_attach_event != NULL)
4812 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4813 	if (ctx->ifc_vlan_detach_event != NULL)
4814 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4815 
4816 	iflib_netmap_detach(ifp);
4817 	ether_ifdetach(ifp);
4818 	if (ctx->ifc_led_dev != NULL)
4819 		led_destroy(ctx->ifc_led_dev);
4820 	/* XXX drain any dependent tasks */
4821 	tqg = qgroup_if_io_tqg;
4822 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4823 		callout_drain(&txq->ift_timer);
4824 		if (txq->ift_task.gt_uniq != NULL)
4825 			taskqgroup_detach(tqg, &txq->ift_task);
4826 	}
4827 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4828 		if (rxq->ifr_task.gt_uniq != NULL)
4829 			taskqgroup_detach(tqg, &rxq->ifr_task);
4830 
4831 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4832 			free(fl->ifl_rx_bitmap, M_IFLIB);
4833 
4834 	}
4835 	tqg = qgroup_if_config_tqg;
4836 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4837 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4838 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4839 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4840 	CTX_LOCK(ctx);
4841 	IFDI_DETACH(ctx);
4842 	CTX_UNLOCK(ctx);
4843 
4844 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4845 	CTX_LOCK_DESTROY(ctx);
4846 	device_set_softc(ctx->ifc_dev, NULL);
4847 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4848 		pci_release_msi(dev);
4849 	}
4850 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4851 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4852 	}
4853 	if (ctx->ifc_msix_mem != NULL) {
4854 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4855 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4856 		ctx->ifc_msix_mem = NULL;
4857 	}
4858 
4859 	bus_generic_detach(dev);
4860 	if_free(ifp);
4861 
4862 	iflib_tx_structures_free(ctx);
4863 	iflib_rx_structures_free(ctx);
4864 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4865 		free(ctx->ifc_softc, M_IFLIB);
4866 	free(ctx, M_IFLIB);
4867 	return (0);
4868 }
4869 
4870 
4871 int
4872 iflib_device_detach(device_t dev)
4873 {
4874 	if_ctx_t ctx = device_get_softc(dev);
4875 
4876 	return (iflib_device_deregister(ctx));
4877 }
4878 
4879 int
4880 iflib_device_suspend(device_t dev)
4881 {
4882 	if_ctx_t ctx = device_get_softc(dev);
4883 
4884 	CTX_LOCK(ctx);
4885 	IFDI_SUSPEND(ctx);
4886 	CTX_UNLOCK(ctx);
4887 
4888 	return bus_generic_suspend(dev);
4889 }
4890 int
4891 iflib_device_shutdown(device_t dev)
4892 {
4893 	if_ctx_t ctx = device_get_softc(dev);
4894 
4895 	CTX_LOCK(ctx);
4896 	IFDI_SHUTDOWN(ctx);
4897 	CTX_UNLOCK(ctx);
4898 
4899 	return bus_generic_suspend(dev);
4900 }
4901 
4902 
4903 int
4904 iflib_device_resume(device_t dev)
4905 {
4906 	if_ctx_t ctx = device_get_softc(dev);
4907 	iflib_txq_t txq = ctx->ifc_txqs;
4908 
4909 	CTX_LOCK(ctx);
4910 	IFDI_RESUME(ctx);
4911 	iflib_init_locked(ctx);
4912 	CTX_UNLOCK(ctx);
4913 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4914 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4915 
4916 	return (bus_generic_resume(dev));
4917 }
4918 
4919 int
4920 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4921 {
4922 	int error;
4923 	if_ctx_t ctx = device_get_softc(dev);
4924 
4925 	CTX_LOCK(ctx);
4926 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4927 	CTX_UNLOCK(ctx);
4928 
4929 	return (error);
4930 }
4931 
4932 void
4933 iflib_device_iov_uninit(device_t dev)
4934 {
4935 	if_ctx_t ctx = device_get_softc(dev);
4936 
4937 	CTX_LOCK(ctx);
4938 	IFDI_IOV_UNINIT(ctx);
4939 	CTX_UNLOCK(ctx);
4940 }
4941 
4942 int
4943 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4944 {
4945 	int error;
4946 	if_ctx_t ctx = device_get_softc(dev);
4947 
4948 	CTX_LOCK(ctx);
4949 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4950 	CTX_UNLOCK(ctx);
4951 
4952 	return (error);
4953 }
4954 
4955 /*********************************************************************
4956  *
4957  *  MODULE FUNCTION DEFINITIONS
4958  *
4959  **********************************************************************/
4960 
4961 /*
4962  * - Start a fast taskqueue thread for each core
4963  * - Start a taskqueue for control operations
4964  */
4965 static int
4966 iflib_module_init(void)
4967 {
4968 	return (0);
4969 }
4970 
4971 static int
4972 iflib_module_event_handler(module_t mod, int what, void *arg)
4973 {
4974 	int err;
4975 
4976 	switch (what) {
4977 	case MOD_LOAD:
4978 		if ((err = iflib_module_init()) != 0)
4979 			return (err);
4980 		break;
4981 	case MOD_UNLOAD:
4982 		return (EBUSY);
4983 	default:
4984 		return (EOPNOTSUPP);
4985 	}
4986 
4987 	return (0);
4988 }
4989 
4990 /*********************************************************************
4991  *
4992  *  PUBLIC FUNCTION DEFINITIONS
4993  *     ordered as in iflib.h
4994  *
4995  **********************************************************************/
4996 
4997 
4998 static void
4999 _iflib_assert(if_shared_ctx_t sctx)
5000 {
5001 	MPASS(sctx->isc_tx_maxsize);
5002 	MPASS(sctx->isc_tx_maxsegsize);
5003 
5004 	MPASS(sctx->isc_rx_maxsize);
5005 	MPASS(sctx->isc_rx_nsegments);
5006 	MPASS(sctx->isc_rx_maxsegsize);
5007 
5008 	MPASS(sctx->isc_nrxd_min[0]);
5009 	MPASS(sctx->isc_nrxd_max[0]);
5010 	MPASS(sctx->isc_nrxd_default[0]);
5011 	MPASS(sctx->isc_ntxd_min[0]);
5012 	MPASS(sctx->isc_ntxd_max[0]);
5013 	MPASS(sctx->isc_ntxd_default[0]);
5014 }
5015 
5016 static void
5017 _iflib_pre_assert(if_softc_ctx_t scctx)
5018 {
5019 
5020 	MPASS(scctx->isc_txrx->ift_txd_encap);
5021 	MPASS(scctx->isc_txrx->ift_txd_flush);
5022 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5023 	MPASS(scctx->isc_txrx->ift_rxd_available);
5024 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5025 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5026 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5027 }
5028 
5029 static int
5030 iflib_register(if_ctx_t ctx)
5031 {
5032 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5033 	driver_t *driver = sctx->isc_driver;
5034 	device_t dev = ctx->ifc_dev;
5035 	if_t ifp;
5036 
5037 	_iflib_assert(sctx);
5038 
5039 	CTX_LOCK_INIT(ctx);
5040 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5041 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5042 	if (ifp == NULL) {
5043 		device_printf(dev, "can not allocate ifnet structure\n");
5044 		return (ENOMEM);
5045 	}
5046 
5047 	/*
5048 	 * Initialize our context's device specific methods
5049 	 */
5050 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5051 	kobj_class_compile((kobj_class_t) driver);
5052 	driver->refs++;
5053 
5054 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5055 	if_setsoftc(ifp, ctx);
5056 	if_setdev(ifp, dev);
5057 	if_setinitfn(ifp, iflib_if_init);
5058 	if_setioctlfn(ifp, iflib_if_ioctl);
5059 	if_settransmitfn(ifp, iflib_if_transmit);
5060 	if_setqflushfn(ifp, iflib_if_qflush);
5061 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5062 
5063 	ctx->ifc_vlan_attach_event =
5064 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5065 							  EVENTHANDLER_PRI_FIRST);
5066 	ctx->ifc_vlan_detach_event =
5067 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5068 							  EVENTHANDLER_PRI_FIRST);
5069 
5070 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5071 					 iflib_media_change, iflib_media_status);
5072 
5073 	return (0);
5074 }
5075 
5076 
5077 static int
5078 iflib_queues_alloc(if_ctx_t ctx)
5079 {
5080 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5081 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5082 	device_t dev = ctx->ifc_dev;
5083 	int nrxqsets = scctx->isc_nrxqsets;
5084 	int ntxqsets = scctx->isc_ntxqsets;
5085 	iflib_txq_t txq;
5086 	iflib_rxq_t rxq;
5087 	iflib_fl_t fl = NULL;
5088 	int i, j, cpu, err, txconf, rxconf;
5089 	iflib_dma_info_t ifdip;
5090 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5091 	uint32_t *txqsizes = scctx->isc_txqsizes;
5092 	uint8_t nrxqs = sctx->isc_nrxqs;
5093 	uint8_t ntxqs = sctx->isc_ntxqs;
5094 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5095 	caddr_t *vaddrs;
5096 	uint64_t *paddrs;
5097 
5098 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5099 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5100 
5101 /* Allocate the TX ring struct memory */
5102 	if (!(ctx->ifc_txqs =
5103 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5104 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5105 		device_printf(dev, "Unable to allocate TX ring memory\n");
5106 		err = ENOMEM;
5107 		goto fail;
5108 	}
5109 
5110 	/* Now allocate the RX */
5111 	if (!(ctx->ifc_rxqs =
5112 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5113 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5114 		device_printf(dev, "Unable to allocate RX ring memory\n");
5115 		err = ENOMEM;
5116 		goto rx_fail;
5117 	}
5118 
5119 	txq = ctx->ifc_txqs;
5120 	rxq = ctx->ifc_rxqs;
5121 
5122 	/*
5123 	 * XXX handle allocation failure
5124 	 */
5125 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5126 		/* Set up some basics */
5127 
5128 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5129 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5130 			err = ENOMEM;
5131 			goto err_tx_desc;
5132 		}
5133 		txq->ift_ifdi = ifdip;
5134 		for (j = 0; j < ntxqs; j++, ifdip++) {
5135 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5136 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5137 				err = ENOMEM;
5138 				goto err_tx_desc;
5139 			}
5140 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5141 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5142 		}
5143 		txq->ift_ctx = ctx;
5144 		txq->ift_id = i;
5145 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5146 			txq->ift_br_offset = 1;
5147 		} else {
5148 			txq->ift_br_offset = 0;
5149 		}
5150 		/* XXX fix this */
5151 		txq->ift_timer.c_cpu = cpu;
5152 
5153 		if (iflib_txsd_alloc(txq)) {
5154 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5155 			err = ENOMEM;
5156 			goto err_tx_desc;
5157 		}
5158 
5159 		/* Initialize the TX lock */
5160 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5161 		    device_get_nameunit(dev), txq->ift_id);
5162 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5163 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5164 
5165 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5166 			 device_get_nameunit(dev), txq->ift_id);
5167 
5168 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5169 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5170 		if (err) {
5171 			/* XXX free any allocated rings */
5172 			device_printf(dev, "Unable to allocate buf_ring\n");
5173 			goto err_tx_desc;
5174 		}
5175 	}
5176 
5177 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5178 		/* Set up some basics */
5179 
5180 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5181 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5182 			err = ENOMEM;
5183 			goto err_tx_desc;
5184 		}
5185 
5186 		rxq->ifr_ifdi = ifdip;
5187 		/* XXX this needs to be changed if #rx queues != #tx queues */
5188 		rxq->ifr_ntxqirq = 1;
5189 		rxq->ifr_txqid[0] = i;
5190 		for (j = 0; j < nrxqs; j++, ifdip++) {
5191 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5192 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5193 				err = ENOMEM;
5194 				goto err_tx_desc;
5195 			}
5196 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5197 		}
5198 		rxq->ifr_ctx = ctx;
5199 		rxq->ifr_id = i;
5200 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5201 			rxq->ifr_fl_offset = 1;
5202 		} else {
5203 			rxq->ifr_fl_offset = 0;
5204 		}
5205 		rxq->ifr_nfl = nfree_lists;
5206 		if (!(fl =
5207 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5208 			device_printf(dev, "Unable to allocate free list memory\n");
5209 			err = ENOMEM;
5210 			goto err_tx_desc;
5211 		}
5212 		rxq->ifr_fl = fl;
5213 		for (j = 0; j < nfree_lists; j++) {
5214 			fl[j].ifl_rxq = rxq;
5215 			fl[j].ifl_id = j;
5216 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5217 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5218 		}
5219         /* Allocate receive buffers for the ring*/
5220 		if (iflib_rxsd_alloc(rxq)) {
5221 			device_printf(dev,
5222 			    "Critical Failure setting up receive buffers\n");
5223 			err = ENOMEM;
5224 			goto err_rx_desc;
5225 		}
5226 
5227 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5228 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5229 	}
5230 
5231 	/* TXQs */
5232 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5233 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5234 	for (i = 0; i < ntxqsets; i++) {
5235 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5236 
5237 		for (j = 0; j < ntxqs; j++, di++) {
5238 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5239 			paddrs[i*ntxqs + j] = di->idi_paddr;
5240 		}
5241 	}
5242 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5243 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5244 		iflib_tx_structures_free(ctx);
5245 		free(vaddrs, M_IFLIB);
5246 		free(paddrs, M_IFLIB);
5247 		goto err_rx_desc;
5248 	}
5249 	free(vaddrs, M_IFLIB);
5250 	free(paddrs, M_IFLIB);
5251 
5252 	/* RXQs */
5253 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5254 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5255 	for (i = 0; i < nrxqsets; i++) {
5256 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5257 
5258 		for (j = 0; j < nrxqs; j++, di++) {
5259 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5260 			paddrs[i*nrxqs + j] = di->idi_paddr;
5261 		}
5262 	}
5263 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5264 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5265 		iflib_tx_structures_free(ctx);
5266 		free(vaddrs, M_IFLIB);
5267 		free(paddrs, M_IFLIB);
5268 		goto err_rx_desc;
5269 	}
5270 	free(vaddrs, M_IFLIB);
5271 	free(paddrs, M_IFLIB);
5272 
5273 	return (0);
5274 
5275 /* XXX handle allocation failure changes */
5276 err_rx_desc:
5277 err_tx_desc:
5278 rx_fail:
5279 	if (ctx->ifc_rxqs != NULL)
5280 		free(ctx->ifc_rxqs, M_IFLIB);
5281 	ctx->ifc_rxqs = NULL;
5282 	if (ctx->ifc_txqs != NULL)
5283 		free(ctx->ifc_txqs, M_IFLIB);
5284 	ctx->ifc_txqs = NULL;
5285 fail:
5286 	return (err);
5287 }
5288 
5289 static int
5290 iflib_tx_structures_setup(if_ctx_t ctx)
5291 {
5292 	iflib_txq_t txq = ctx->ifc_txqs;
5293 	int i;
5294 
5295 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5296 		iflib_txq_setup(txq);
5297 
5298 	return (0);
5299 }
5300 
5301 static void
5302 iflib_tx_structures_free(if_ctx_t ctx)
5303 {
5304 	iflib_txq_t txq = ctx->ifc_txqs;
5305 	int i, j;
5306 
5307 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5308 		iflib_txq_destroy(txq);
5309 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5310 			iflib_dma_free(&txq->ift_ifdi[j]);
5311 	}
5312 	free(ctx->ifc_txqs, M_IFLIB);
5313 	ctx->ifc_txqs = NULL;
5314 	IFDI_QUEUES_FREE(ctx);
5315 }
5316 
5317 /*********************************************************************
5318  *
5319  *  Initialize all receive rings.
5320  *
5321  **********************************************************************/
5322 static int
5323 iflib_rx_structures_setup(if_ctx_t ctx)
5324 {
5325 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5326 	int q;
5327 #if defined(INET6) || defined(INET)
5328 	int i, err;
5329 #endif
5330 
5331 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5332 #if defined(INET6) || defined(INET)
5333 		tcp_lro_free(&rxq->ifr_lc);
5334 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5335 		    TCP_LRO_ENTRIES, min(1024,
5336 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5337 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5338 			goto fail;
5339 		}
5340 		rxq->ifr_lro_enabled = TRUE;
5341 #endif
5342 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5343 	}
5344 	return (0);
5345 #if defined(INET6) || defined(INET)
5346 fail:
5347 	/*
5348 	 * Free RX software descriptors allocated so far, we will only handle
5349 	 * the rings that completed, the failing case will have
5350 	 * cleaned up for itself. 'q' failed, so its the terminus.
5351 	 */
5352 	rxq = ctx->ifc_rxqs;
5353 	for (i = 0; i < q; ++i, rxq++) {
5354 		iflib_rx_sds_free(rxq);
5355 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5356 	}
5357 	return (err);
5358 #endif
5359 }
5360 
5361 /*********************************************************************
5362  *
5363  *  Free all receive rings.
5364  *
5365  **********************************************************************/
5366 static void
5367 iflib_rx_structures_free(if_ctx_t ctx)
5368 {
5369 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5370 
5371 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5372 		iflib_rx_sds_free(rxq);
5373 	}
5374 }
5375 
5376 static int
5377 iflib_qset_structures_setup(if_ctx_t ctx)
5378 {
5379 	int err;
5380 
5381 	/*
5382 	 * It is expected that the caller takes care of freeing queues if this
5383 	 * fails.
5384 	 */
5385 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5386 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5387 		return (err);
5388 	}
5389 
5390 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5391 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5392 
5393 	return (err);
5394 }
5395 
5396 int
5397 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5398 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5399 {
5400 
5401 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5402 }
5403 
5404 #ifdef SMP
5405 static int
5406 find_nth(if_ctx_t ctx, int qid)
5407 {
5408 	cpuset_t cpus;
5409 	int i, cpuid, eqid, count;
5410 
5411 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5412 	count = CPU_COUNT(&cpus);
5413 	eqid = qid % count;
5414 	/* clear up to the qid'th bit */
5415 	for (i = 0; i < eqid; i++) {
5416 		cpuid = CPU_FFS(&cpus);
5417 		MPASS(cpuid != 0);
5418 		CPU_CLR(cpuid-1, &cpus);
5419 	}
5420 	cpuid = CPU_FFS(&cpus);
5421 	MPASS(cpuid != 0);
5422 	return (cpuid-1);
5423 }
5424 
5425 #ifdef SCHED_ULE
5426 extern struct cpu_group *cpu_top;              /* CPU topology */
5427 
5428 static int
5429 find_child_with_core(int cpu, struct cpu_group *grp)
5430 {
5431 	int i;
5432 
5433 	if (grp->cg_children == 0)
5434 		return -1;
5435 
5436 	MPASS(grp->cg_child);
5437 	for (i = 0; i < grp->cg_children; i++) {
5438 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5439 			return i;
5440 	}
5441 
5442 	return -1;
5443 }
5444 
5445 /*
5446  * Find the nth "close" core to the specified core
5447  * "close" is defined as the deepest level that shares
5448  * at least an L2 cache.  With threads, this will be
5449  * threads on the same core.  If the sahred cache is L3
5450  * or higher, simply returns the same core.
5451  */
5452 static int
5453 find_close_core(int cpu, int core_offset)
5454 {
5455 	struct cpu_group *grp;
5456 	int i;
5457 	int fcpu;
5458 	cpuset_t cs;
5459 
5460 	grp = cpu_top;
5461 	if (grp == NULL)
5462 		return cpu;
5463 	i = 0;
5464 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5465 		/* If the child only has one cpu, don't descend */
5466 		if (grp->cg_child[i].cg_count <= 1)
5467 			break;
5468 		grp = &grp->cg_child[i];
5469 	}
5470 
5471 	/* If they don't share at least an L2 cache, use the same CPU */
5472 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5473 		return cpu;
5474 
5475 	/* Now pick one */
5476 	CPU_COPY(&grp->cg_mask, &cs);
5477 
5478 	/* Add the selected CPU offset to core offset. */
5479 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5480 		if (fcpu - 1 == cpu)
5481 			break;
5482 		CPU_CLR(fcpu - 1, &cs);
5483 	}
5484 	MPASS(fcpu);
5485 
5486 	core_offset += i;
5487 
5488 	CPU_COPY(&grp->cg_mask, &cs);
5489 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5490 		MPASS(CPU_FFS(&cs));
5491 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5492 	}
5493 	MPASS(CPU_FFS(&cs));
5494 	return CPU_FFS(&cs) - 1;
5495 }
5496 #else
5497 static int
5498 find_close_core(int cpu, int core_offset __unused)
5499 {
5500 	return cpu;
5501 }
5502 #endif
5503 
5504 static int
5505 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5506 {
5507 	switch (type) {
5508 	case IFLIB_INTR_TX:
5509 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5510 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5511 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5512 	case IFLIB_INTR_RX:
5513 	case IFLIB_INTR_RXTX:
5514 		/* RX queues get the specified core */
5515 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5516 	default:
5517 		return -1;
5518 	}
5519 }
5520 #else
5521 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5522 #define find_close_core(cpuid, tid)	CPU_FIRST()
5523 #define find_nth(ctx, gid)		CPU_FIRST()
5524 #endif
5525 
5526 /* Just to avoid copy/paste */
5527 static inline int
5528 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5529     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5530 {
5531 	int cpuid;
5532 	int err, tid;
5533 
5534 	cpuid = find_nth(ctx, qid);
5535 	tid = get_core_offset(ctx, type, qid);
5536 	MPASS(tid >= 0);
5537 	cpuid = find_close_core(cpuid, tid);
5538 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5539 	if (err) {
5540 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5541 		return (err);
5542 	}
5543 #ifdef notyet
5544 	if (cpuid > ctx->ifc_cpuid_highest)
5545 		ctx->ifc_cpuid_highest = cpuid;
5546 #endif
5547 	return 0;
5548 }
5549 
5550 int
5551 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5552 			iflib_intr_type_t type, driver_filter_t *filter,
5553 			void *filter_arg, int qid, const char *name)
5554 {
5555 	struct grouptask *gtask;
5556 	struct taskqgroup *tqg;
5557 	iflib_filter_info_t info;
5558 	gtask_fn_t *fn;
5559 	int tqrid, err;
5560 	driver_filter_t *intr_fast;
5561 	void *q;
5562 
5563 	info = &ctx->ifc_filter_info;
5564 	tqrid = rid;
5565 
5566 	switch (type) {
5567 	/* XXX merge tx/rx for netmap? */
5568 	case IFLIB_INTR_TX:
5569 		q = &ctx->ifc_txqs[qid];
5570 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5571 		gtask = &ctx->ifc_txqs[qid].ift_task;
5572 		tqg = qgroup_if_io_tqg;
5573 		fn = _task_fn_tx;
5574 		intr_fast = iflib_fast_intr;
5575 		GROUPTASK_INIT(gtask, 0, fn, q);
5576 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5577 		break;
5578 	case IFLIB_INTR_RX:
5579 		q = &ctx->ifc_rxqs[qid];
5580 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5581 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5582 		tqg = qgroup_if_io_tqg;
5583 		fn = _task_fn_rx;
5584 		intr_fast = iflib_fast_intr;
5585 		GROUPTASK_INIT(gtask, 0, fn, q);
5586 		break;
5587 	case IFLIB_INTR_RXTX:
5588 		q = &ctx->ifc_rxqs[qid];
5589 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5590 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5591 		tqg = qgroup_if_io_tqg;
5592 		fn = _task_fn_rx;
5593 		intr_fast = iflib_fast_intr_rxtx;
5594 		GROUPTASK_INIT(gtask, 0, fn, q);
5595 		break;
5596 	case IFLIB_INTR_ADMIN:
5597 		q = ctx;
5598 		tqrid = -1;
5599 		info = &ctx->ifc_filter_info;
5600 		gtask = &ctx->ifc_admin_task;
5601 		tqg = qgroup_if_config_tqg;
5602 		fn = _task_fn_admin;
5603 		intr_fast = iflib_fast_intr_ctx;
5604 		break;
5605 	default:
5606 		panic("unknown net intr type");
5607 	}
5608 
5609 	info->ifi_filter = filter;
5610 	info->ifi_filter_arg = filter_arg;
5611 	info->ifi_task = gtask;
5612 	info->ifi_ctx = q;
5613 
5614 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5615 	if (err != 0) {
5616 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5617 		return (err);
5618 	}
5619 	if (type == IFLIB_INTR_ADMIN)
5620 		return (0);
5621 
5622 	if (tqrid != -1) {
5623 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5624 		if (err)
5625 			return (err);
5626 	} else {
5627 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5628 	}
5629 
5630 	return (0);
5631 }
5632 
5633 void
5634 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, const char *name)
5635 {
5636 	struct grouptask *gtask;
5637 	struct taskqgroup *tqg;
5638 	gtask_fn_t *fn;
5639 	void *q;
5640 	int irq_num = -1;
5641 	int err;
5642 
5643 	switch (type) {
5644 	case IFLIB_INTR_TX:
5645 		q = &ctx->ifc_txqs[qid];
5646 		gtask = &ctx->ifc_txqs[qid].ift_task;
5647 		tqg = qgroup_if_io_tqg;
5648 		fn = _task_fn_tx;
5649 		if (irq != NULL)
5650 			irq_num = rman_get_start(irq->ii_res);
5651 		break;
5652 	case IFLIB_INTR_RX:
5653 		q = &ctx->ifc_rxqs[qid];
5654 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5655 		tqg = qgroup_if_io_tqg;
5656 		fn = _task_fn_rx;
5657 		if (irq != NULL)
5658 			irq_num = rman_get_start(irq->ii_res);
5659 		break;
5660 	case IFLIB_INTR_IOV:
5661 		q = ctx;
5662 		gtask = &ctx->ifc_vflr_task;
5663 		tqg = qgroup_if_config_tqg;
5664 		fn = _task_fn_iov;
5665 		break;
5666 	default:
5667 		panic("unknown net intr type");
5668 	}
5669 	GROUPTASK_INIT(gtask, 0, fn, q);
5670 	if (irq_num != -1) {
5671 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5672 		if (err)
5673 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5674 	}
5675 	else {
5676 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5677 	}
5678 }
5679 
5680 void
5681 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5682 {
5683 	if (irq->ii_tag)
5684 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5685 
5686 	if (irq->ii_res)
5687 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5688 }
5689 
5690 static int
5691 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5692 {
5693 	iflib_txq_t txq = ctx->ifc_txqs;
5694 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5695 	if_irq_t irq = &ctx->ifc_legacy_irq;
5696 	iflib_filter_info_t info;
5697 	struct grouptask *gtask;
5698 	struct taskqgroup *tqg;
5699 	gtask_fn_t *fn;
5700 	int tqrid;
5701 	void *q;
5702 	int err;
5703 
5704 	q = &ctx->ifc_rxqs[0];
5705 	info = &rxq[0].ifr_filter_info;
5706 	gtask = &rxq[0].ifr_task;
5707 	tqg = qgroup_if_io_tqg;
5708 	tqrid = irq->ii_rid = *rid;
5709 	fn = _task_fn_rx;
5710 
5711 	ctx->ifc_flags |= IFC_LEGACY;
5712 	info->ifi_filter = filter;
5713 	info->ifi_filter_arg = filter_arg;
5714 	info->ifi_task = gtask;
5715 	info->ifi_ctx = ctx;
5716 
5717 	/* We allocate a single interrupt resource */
5718 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5719 		return (err);
5720 	GROUPTASK_INIT(gtask, 0, fn, q);
5721 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5722 
5723 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5724 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5725 	return (0);
5726 }
5727 
5728 void
5729 iflib_led_create(if_ctx_t ctx)
5730 {
5731 
5732 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5733 	    device_get_nameunit(ctx->ifc_dev));
5734 }
5735 
5736 void
5737 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5738 {
5739 
5740 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5741 }
5742 
5743 void
5744 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5745 {
5746 
5747 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5748 }
5749 
5750 void
5751 iflib_admin_intr_deferred(if_ctx_t ctx)
5752 {
5753 #ifdef INVARIANTS
5754 	struct grouptask *gtask;
5755 
5756 	gtask = &ctx->ifc_admin_task;
5757 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5758 #endif
5759 
5760 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5761 }
5762 
5763 void
5764 iflib_iov_intr_deferred(if_ctx_t ctx)
5765 {
5766 
5767 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5768 }
5769 
5770 void
5771 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5772 {
5773 
5774 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5775 }
5776 
5777 void
5778 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5779 	const char *name)
5780 {
5781 
5782 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5783 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5784 }
5785 
5786 void
5787 iflib_config_gtask_deinit(struct grouptask *gtask)
5788 {
5789 
5790 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5791 }
5792 
5793 void
5794 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5795 {
5796 	if_t ifp = ctx->ifc_ifp;
5797 	iflib_txq_t txq = ctx->ifc_txqs;
5798 
5799 	if_setbaudrate(ifp, baudrate);
5800 	if (baudrate >= IF_Gbps(10)) {
5801 		STATE_LOCK(ctx);
5802 		ctx->ifc_flags |= IFC_PREFETCH;
5803 		STATE_UNLOCK(ctx);
5804 	}
5805 	/* If link down, disable watchdog */
5806 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5807 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5808 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5809 	}
5810 	ctx->ifc_link_state = link_state;
5811 	if_link_state_change(ifp, link_state);
5812 }
5813 
5814 static int
5815 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5816 {
5817 	int credits;
5818 #ifdef INVARIANTS
5819 	int credits_pre = txq->ift_cidx_processed;
5820 #endif
5821 
5822 	if (ctx->isc_txd_credits_update == NULL)
5823 		return (0);
5824 
5825 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5826 		return (0);
5827 
5828 	txq->ift_processed += credits;
5829 	txq->ift_cidx_processed += credits;
5830 
5831 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5832 	if (txq->ift_cidx_processed >= txq->ift_size)
5833 		txq->ift_cidx_processed -= txq->ift_size;
5834 	return (credits);
5835 }
5836 
5837 static int
5838 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5839 {
5840 
5841 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5842 	    budget));
5843 }
5844 
5845 void
5846 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5847 	const char *description, if_int_delay_info_t info,
5848 	int offset, int value)
5849 {
5850 	info->iidi_ctx = ctx;
5851 	info->iidi_offset = offset;
5852 	info->iidi_value = value;
5853 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5854 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5855 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5856 	    info, 0, iflib_sysctl_int_delay, "I", description);
5857 }
5858 
5859 struct sx *
5860 iflib_ctx_lock_get(if_ctx_t ctx)
5861 {
5862 
5863 	return (&ctx->ifc_ctx_sx);
5864 }
5865 
5866 static int
5867 iflib_msix_init(if_ctx_t ctx)
5868 {
5869 	device_t dev = ctx->ifc_dev;
5870 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5871 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5872 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5873 	int iflib_num_tx_queues, iflib_num_rx_queues;
5874 	int err, admincnt, bar;
5875 
5876 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5877 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5878 
5879 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5880 
5881 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5882 	admincnt = sctx->isc_admin_intrcnt;
5883 	/* Override by global tuneable */
5884 	{
5885 		int i;
5886 		size_t len = sizeof(i);
5887 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5888 		if (err == 0) {
5889 			if (i == 0)
5890 				goto msi;
5891 		}
5892 		else {
5893 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5894 		}
5895 	}
5896 	/* Override by tuneable */
5897 	if (scctx->isc_disable_msix)
5898 		goto msi;
5899 
5900 	/*
5901 	** When used in a virtualized environment
5902 	** PCI BUSMASTER capability may not be set
5903 	** so explicity set it here and rewrite
5904 	** the ENABLE in the MSIX control register
5905 	** at this point to cause the host to
5906 	** successfully initialize us.
5907 	*/
5908 	{
5909 		int msix_ctrl, rid;
5910 
5911  		pci_enable_busmaster(dev);
5912 		rid = 0;
5913 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5914 			rid += PCIR_MSIX_CTRL;
5915 			msix_ctrl = pci_read_config(dev, rid, 2);
5916 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5917 			pci_write_config(dev, rid, msix_ctrl, 2);
5918 		} else {
5919 			device_printf(dev, "PCIY_MSIX capability not found; "
5920 			                   "or rid %d == 0.\n", rid);
5921 			goto msi;
5922 		}
5923 	}
5924 
5925 	/*
5926 	 * bar == -1 => "trust me I know what I'm doing"
5927 	 * Some drivers are for hardware that is so shoddily
5928 	 * documented that no one knows which bars are which
5929 	 * so the developer has to map all bars. This hack
5930 	 * allows shoddy garbage to use msix in this framework.
5931 	 */
5932 	if (bar != -1) {
5933 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5934 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5935 		if (ctx->ifc_msix_mem == NULL) {
5936 			/* May not be enabled */
5937 			device_printf(dev, "Unable to map MSIX table \n");
5938 			goto msi;
5939 		}
5940 	}
5941 	/* First try MSI/X */
5942 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5943 		device_printf(dev, "System has MSIX disabled \n");
5944 		bus_release_resource(dev, SYS_RES_MEMORY,
5945 		    bar, ctx->ifc_msix_mem);
5946 		ctx->ifc_msix_mem = NULL;
5947 		goto msi;
5948 	}
5949 #if IFLIB_DEBUG
5950 	/* use only 1 qset in debug mode */
5951 	queuemsgs = min(msgs - admincnt, 1);
5952 #else
5953 	queuemsgs = msgs - admincnt;
5954 #endif
5955 #ifdef RSS
5956 	queues = imin(queuemsgs, rss_getnumbuckets());
5957 #else
5958 	queues = queuemsgs;
5959 #endif
5960 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5961 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5962 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5963 #ifdef  RSS
5964 	/* If we're doing RSS, clamp at the number of RSS buckets */
5965 	if (queues > rss_getnumbuckets())
5966 		queues = rss_getnumbuckets();
5967 #endif
5968 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5969 		rx_queues = iflib_num_rx_queues;
5970 	else
5971 		rx_queues = queues;
5972 
5973 	if (rx_queues > scctx->isc_nrxqsets)
5974 		rx_queues = scctx->isc_nrxqsets;
5975 
5976 	/*
5977 	 * We want this to be all logical CPUs by default
5978 	 */
5979 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5980 		tx_queues = iflib_num_tx_queues;
5981 	else
5982 		tx_queues = mp_ncpus;
5983 
5984 	if (tx_queues > scctx->isc_ntxqsets)
5985 		tx_queues = scctx->isc_ntxqsets;
5986 
5987 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5988 #ifdef INVARIANTS
5989 		if (tx_queues != rx_queues)
5990 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5991 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5992 #endif
5993 		tx_queues = min(rx_queues, tx_queues);
5994 		rx_queues = min(rx_queues, tx_queues);
5995 	}
5996 
5997 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5998 
5999 	vectors = rx_queues + admincnt;
6000 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6001 		device_printf(dev,
6002 					  "Using MSIX interrupts with %d vectors\n", vectors);
6003 		scctx->isc_vectors = vectors;
6004 		scctx->isc_nrxqsets = rx_queues;
6005 		scctx->isc_ntxqsets = tx_queues;
6006 		scctx->isc_intr = IFLIB_INTR_MSIX;
6007 
6008 		return (vectors);
6009 	} else {
6010 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6011 	}
6012 msi:
6013 	vectors = pci_msi_count(dev);
6014 	scctx->isc_nrxqsets = 1;
6015 	scctx->isc_ntxqsets = 1;
6016 	scctx->isc_vectors = vectors;
6017 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6018 		device_printf(dev,"Using an MSI interrupt\n");
6019 		scctx->isc_intr = IFLIB_INTR_MSI;
6020 	} else {
6021 		device_printf(dev,"Using a Legacy interrupt\n");
6022 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6023 	}
6024 
6025 	return (vectors);
6026 }
6027 
6028 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6029 
6030 static int
6031 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6032 {
6033 	int rc;
6034 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6035 	struct sbuf *sb;
6036 	char *ring_state = "UNKNOWN";
6037 
6038 	/* XXX needed ? */
6039 	rc = sysctl_wire_old_buffer(req, 0);
6040 	MPASS(rc == 0);
6041 	if (rc != 0)
6042 		return (rc);
6043 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6044 	MPASS(sb != NULL);
6045 	if (sb == NULL)
6046 		return (ENOMEM);
6047 	if (state[3] <= 3)
6048 		ring_state = ring_states[state[3]];
6049 
6050 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6051 		    state[0], state[1], state[2], ring_state);
6052 	rc = sbuf_finish(sb);
6053 	sbuf_delete(sb);
6054         return(rc);
6055 }
6056 
6057 enum iflib_ndesc_handler {
6058 	IFLIB_NTXD_HANDLER,
6059 	IFLIB_NRXD_HANDLER,
6060 };
6061 
6062 static int
6063 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6064 {
6065 	if_ctx_t ctx = (void *)arg1;
6066 	enum iflib_ndesc_handler type = arg2;
6067 	char buf[256] = {0};
6068 	qidx_t *ndesc;
6069 	char *p, *next;
6070 	int nqs, rc, i;
6071 
6072 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6073 
6074 	nqs = 8;
6075 	switch(type) {
6076 	case IFLIB_NTXD_HANDLER:
6077 		ndesc = ctx->ifc_sysctl_ntxds;
6078 		if (ctx->ifc_sctx)
6079 			nqs = ctx->ifc_sctx->isc_ntxqs;
6080 		break;
6081 	case IFLIB_NRXD_HANDLER:
6082 		ndesc = ctx->ifc_sysctl_nrxds;
6083 		if (ctx->ifc_sctx)
6084 			nqs = ctx->ifc_sctx->isc_nrxqs;
6085 		break;
6086 	default:
6087 			panic("unhandled type");
6088 	}
6089 	if (nqs == 0)
6090 		nqs = 8;
6091 
6092 	for (i=0; i<8; i++) {
6093 		if (i >= nqs)
6094 			break;
6095 		if (i)
6096 			strcat(buf, ",");
6097 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6098 	}
6099 
6100 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6101 	if (rc || req->newptr == NULL)
6102 		return rc;
6103 
6104 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6105 	    i++, p = strsep(&next, " ,")) {
6106 		ndesc[i] = strtoul(p, NULL, 10);
6107 	}
6108 
6109 	return(rc);
6110 }
6111 
6112 #define NAME_BUFLEN 32
6113 static void
6114 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6115 {
6116         device_t dev = iflib_get_dev(ctx);
6117 	struct sysctl_oid_list *child, *oid_list;
6118 	struct sysctl_ctx_list *ctx_list;
6119 	struct sysctl_oid *node;
6120 
6121 	ctx_list = device_get_sysctl_ctx(dev);
6122 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6123 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6124 						      CTLFLAG_RD, NULL, "IFLIB fields");
6125 	oid_list = SYSCTL_CHILDREN(node);
6126 
6127 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6128 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6129 		       "driver version");
6130 
6131 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6132 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6133 			"# of txqs to use, 0 => use default #");
6134 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6135 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6136 			"# of rxqs to use, 0 => use default #");
6137 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6138 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6139                        "permit #txq != #rxq");
6140 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6141                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6142                       "disable MSIX (default 0)");
6143 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6144 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6145                        "set the rx budget");
6146 
6147 	/* XXX change for per-queue sizes */
6148 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6149 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6150                        mp_ndesc_handler, "A",
6151                        "list of # of tx descriptors to use, 0 = use default #");
6152 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6153 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6154                        mp_ndesc_handler, "A",
6155                        "list of # of rx descriptors to use, 0 = use default #");
6156 }
6157 
6158 static void
6159 iflib_add_device_sysctl_post(if_ctx_t ctx)
6160 {
6161 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6162 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6163         device_t dev = iflib_get_dev(ctx);
6164 	struct sysctl_oid_list *child;
6165 	struct sysctl_ctx_list *ctx_list;
6166 	iflib_fl_t fl;
6167 	iflib_txq_t txq;
6168 	iflib_rxq_t rxq;
6169 	int i, j;
6170 	char namebuf[NAME_BUFLEN];
6171 	char *qfmt;
6172 	struct sysctl_oid *queue_node, *fl_node, *node;
6173 	struct sysctl_oid_list *queue_list, *fl_list;
6174 	ctx_list = device_get_sysctl_ctx(dev);
6175 
6176 	node = ctx->ifc_sysctl_node;
6177 	child = SYSCTL_CHILDREN(node);
6178 
6179 	if (scctx->isc_ntxqsets > 100)
6180 		qfmt = "txq%03d";
6181 	else if (scctx->isc_ntxqsets > 10)
6182 		qfmt = "txq%02d";
6183 	else
6184 		qfmt = "txq%d";
6185 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6186 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6187 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6188 					     CTLFLAG_RD, NULL, "Queue Name");
6189 		queue_list = SYSCTL_CHILDREN(queue_node);
6190 #if MEMORY_LOGGING
6191 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6192 				CTLFLAG_RD,
6193 				&txq->ift_dequeued, "total mbufs freed");
6194 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6195 				CTLFLAG_RD,
6196 				&txq->ift_enqueued, "total mbufs enqueued");
6197 #endif
6198 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6199 				   CTLFLAG_RD,
6200 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6201 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6202 				   CTLFLAG_RD,
6203 				   &txq->ift_pullups, "# of times m_pullup was called");
6204 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6205 				   CTLFLAG_RD,
6206 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6207 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6208 				   CTLFLAG_RD,
6209 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6210 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6211 				   CTLFLAG_RD,
6212 				   &txq->ift_map_failed, "# of times dma map failed");
6213 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6214 				   CTLFLAG_RD,
6215 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6216 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6217 				   CTLFLAG_RD,
6218 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6219 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6220 				   CTLFLAG_RD,
6221 				   &txq->ift_pidx, 1, "Producer Index");
6222 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6223 				   CTLFLAG_RD,
6224 				   &txq->ift_cidx, 1, "Consumer Index");
6225 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6226 				   CTLFLAG_RD,
6227 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6228 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6229 				   CTLFLAG_RD,
6230 				   &txq->ift_in_use, 1, "descriptors in use");
6231 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6232 				   CTLFLAG_RD,
6233 				   &txq->ift_processed, "descriptors procesed for clean");
6234 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6235 				   CTLFLAG_RD,
6236 				   &txq->ift_cleaned, "total cleaned");
6237 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6238 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6239 				0, mp_ring_state_handler, "A", "soft ring state");
6240 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6241 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6242 				       "# of enqueues to the mp_ring for this queue");
6243 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6244 				       CTLFLAG_RD, &txq->ift_br->drops,
6245 				       "# of drops in the mp_ring for this queue");
6246 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6247 				       CTLFLAG_RD, &txq->ift_br->starts,
6248 				       "# of normal consumer starts in the mp_ring for this queue");
6249 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6250 				       CTLFLAG_RD, &txq->ift_br->stalls,
6251 					       "# of consumer stalls in the mp_ring for this queue");
6252 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6253 			       CTLFLAG_RD, &txq->ift_br->restarts,
6254 				       "# of consumer restarts in the mp_ring for this queue");
6255 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6256 				       CTLFLAG_RD, &txq->ift_br->abdications,
6257 				       "# of consumer abdications in the mp_ring for this queue");
6258 	}
6259 
6260 	if (scctx->isc_nrxqsets > 100)
6261 		qfmt = "rxq%03d";
6262 	else if (scctx->isc_nrxqsets > 10)
6263 		qfmt = "rxq%02d";
6264 	else
6265 		qfmt = "rxq%d";
6266 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6267 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6268 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6269 					     CTLFLAG_RD, NULL, "Queue Name");
6270 		queue_list = SYSCTL_CHILDREN(queue_node);
6271 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6272 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6273 				       CTLFLAG_RD,
6274 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6275 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6276 				       CTLFLAG_RD,
6277 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6278 		}
6279 
6280 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6281 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6282 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6283 						     CTLFLAG_RD, NULL, "freelist Name");
6284 			fl_list = SYSCTL_CHILDREN(fl_node);
6285 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6286 				       CTLFLAG_RD,
6287 				       &fl->ifl_pidx, 1, "Producer Index");
6288 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6289 				       CTLFLAG_RD,
6290 				       &fl->ifl_cidx, 1, "Consumer Index");
6291 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6292 				       CTLFLAG_RD,
6293 				       &fl->ifl_credits, 1, "credits available");
6294 #if MEMORY_LOGGING
6295 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6296 					CTLFLAG_RD,
6297 					&fl->ifl_m_enqueued, "mbufs allocated");
6298 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6299 					CTLFLAG_RD,
6300 					&fl->ifl_m_dequeued, "mbufs freed");
6301 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6302 					CTLFLAG_RD,
6303 					&fl->ifl_cl_enqueued, "clusters allocated");
6304 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6305 					CTLFLAG_RD,
6306 					&fl->ifl_cl_dequeued, "clusters freed");
6307 #endif
6308 
6309 		}
6310 	}
6311 
6312 }
6313 
6314 #ifndef __NO_STRICT_ALIGNMENT
6315 static struct mbuf *
6316 iflib_fixup_rx(struct mbuf *m)
6317 {
6318 	struct mbuf *n;
6319 
6320 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6321 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6322 		m->m_data += ETHER_HDR_LEN;
6323 		n = m;
6324 	} else {
6325 		MGETHDR(n, M_NOWAIT, MT_DATA);
6326 		if (n == NULL) {
6327 			m_freem(m);
6328 			return (NULL);
6329 		}
6330 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6331 		m->m_data += ETHER_HDR_LEN;
6332 		m->m_len -= ETHER_HDR_LEN;
6333 		n->m_len = ETHER_HDR_LEN;
6334 		M_MOVE_PKTHDR(n, m);
6335 		n->m_next = m;
6336 	}
6337 	return (n);
6338 }
6339 #endif
6340 
6341 #ifdef NETDUMP
6342 static void
6343 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6344 {
6345 	if_ctx_t ctx;
6346 
6347 	ctx = if_getsoftc(ifp);
6348 	CTX_LOCK(ctx);
6349 	*nrxr = NRXQSETS(ctx);
6350 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6351 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6352 	CTX_UNLOCK(ctx);
6353 }
6354 
6355 static void
6356 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6357 {
6358 	if_ctx_t ctx;
6359 	if_softc_ctx_t scctx;
6360 	iflib_fl_t fl;
6361 	iflib_rxq_t rxq;
6362 	int i, j;
6363 
6364 	ctx = if_getsoftc(ifp);
6365 	scctx = &ctx->ifc_softc_ctx;
6366 
6367 	switch (event) {
6368 	case NETDUMP_START:
6369 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6370 			rxq = &ctx->ifc_rxqs[i];
6371 			for (j = 0; j < rxq->ifr_nfl; j++) {
6372 				fl = rxq->ifr_fl;
6373 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6374 			}
6375 		}
6376 		iflib_no_tx_batch = 1;
6377 		break;
6378 	default:
6379 		break;
6380 	}
6381 }
6382 
6383 static int
6384 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6385 {
6386 	if_ctx_t ctx;
6387 	iflib_txq_t txq;
6388 	int error;
6389 
6390 	ctx = if_getsoftc(ifp);
6391 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6392 	    IFF_DRV_RUNNING)
6393 		return (EBUSY);
6394 
6395 	txq = &ctx->ifc_txqs[0];
6396 	error = iflib_encap(txq, &m);
6397 	if (error == 0)
6398 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6399 	return (error);
6400 }
6401 
6402 static int
6403 iflib_netdump_poll(struct ifnet *ifp, int count)
6404 {
6405 	if_ctx_t ctx;
6406 	if_softc_ctx_t scctx;
6407 	iflib_txq_t txq;
6408 	int i;
6409 
6410 	ctx = if_getsoftc(ifp);
6411 	scctx = &ctx->ifc_softc_ctx;
6412 
6413 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6414 	    IFF_DRV_RUNNING)
6415 		return (EBUSY);
6416 
6417 	txq = &ctx->ifc_txqs[0];
6418 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6419 
6420 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6421 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6422 	return (0);
6423 }
6424 #endif /* NETDUMP */
6425