xref: /freebsd/sys/net/iflib.c (revision 123af6ec70016f5556da5972d4d63c7d175c06d3)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/jail.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/md5.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/kobj.h>
47 #include <sys/rman.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
57 
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
62 #include <net/bpf.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
92 
93 #include "ifdi_if.h"
94 
95 #ifdef PCI_IOV
96 #include <dev/pci/pci_iov.h>
97 #endif
98 
99 #include <sys/bitstring.h>
100 /*
101  * enable accounting of every mbuf as it comes in to and goes out of
102  * iflib's software descriptor references
103  */
104 #define MEMORY_LOGGING 0
105 /*
106  * Enable mbuf vectors for compressing long mbuf chains
107  */
108 
109 /*
110  * NB:
111  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
112  *   we prefetch needs to be determined by the time spent in m_free vis a vis
113  *   the cost of a prefetch. This will of course vary based on the workload:
114  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
115  *        is quite expensive, thus suggesting very little prefetch.
116  *      - small packet forwarding which is just returning a single mbuf to
117  *        UMA will typically be very fast vis a vis the cost of a memory
118  *        access.
119  */
120 
121 
122 /*
123  * File organization:
124  *  - private structures
125  *  - iflib private utility functions
126  *  - ifnet functions
127  *  - vlan registry and other exported functions
128  *  - iflib public core functions
129  *
130  *
131  */
132 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155 	/*
156 	 * Pointer to hardware driver's softc
157 	 */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	iflib_txq_t ifc_txqs;
170 	iflib_rxq_t ifc_rxqs;
171 	uint32_t ifc_if_flags;
172 	uint32_t ifc_flags;
173 	uint32_t ifc_max_fl_buf_size;
174 	uint32_t ifc_rx_mbuf_sz;
175 
176 	int ifc_link_state;
177 	int ifc_link_irq;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 
188 	struct sysctl_oid *ifc_sysctl_node;
189 	uint16_t ifc_sysctl_ntxqs;
190 	uint16_t ifc_sysctl_nrxqs;
191 	uint16_t ifc_sysctl_qs_eq_override;
192 	uint16_t ifc_sysctl_rx_budget;
193 	uint16_t ifc_sysctl_tx_abdicate;
194 
195 	qidx_t ifc_sysctl_ntxds[8];
196 	qidx_t ifc_sysctl_nrxds[8];
197 	struct if_txrx ifc_txrx;
198 #define isc_txd_encap  ifc_txrx.ift_txd_encap
199 #define isc_txd_flush  ifc_txrx.ift_txd_flush
200 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
201 #define isc_rxd_available ifc_txrx.ift_rxd_available
202 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
203 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
204 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
208 	eventhandler_tag ifc_vlan_attach_event;
209 	eventhandler_tag ifc_vlan_detach_event;
210 	uint8_t ifc_mac[ETHER_ADDR_LEN];
211 	char ifc_mtx_name[16];
212 };
213 
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (&ctx->ifc_media);
241 }
242 
243 uint32_t
244 iflib_get_flags(if_ctx_t ctx)
245 {
246 	return (ctx->ifc_flags);
247 }
248 
249 void
250 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
251 {
252 
253 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
254 }
255 
256 if_softc_ctx_t
257 iflib_get_softc_ctx(if_ctx_t ctx)
258 {
259 
260 	return (&ctx->ifc_softc_ctx);
261 }
262 
263 if_shared_ctx_t
264 iflib_get_sctx(if_ctx_t ctx)
265 {
266 
267 	return (ctx->ifc_sctx);
268 }
269 
270 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
271 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
272 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 
274 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
275 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 
277 typedef struct iflib_sw_rx_desc_array {
278 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
280 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
281 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
282 } iflib_rxsd_array_t;
283 
284 typedef struct iflib_sw_tx_desc_array {
285 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
286 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 } if_txsd_vec_t;
289 
290 
291 /* magic number that should be high enough for any hardware */
292 #define IFLIB_MAX_TX_SEGS		128
293 #define IFLIB_RX_COPY_THRESH		128
294 #define IFLIB_MAX_RX_REFRESH		32
295 /* The minimum descriptors per second before we start coalescing */
296 #define IFLIB_MIN_DESC_SEC		16384
297 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
298 #define IFLIB_QUEUE_IDLE		0
299 #define IFLIB_QUEUE_HUNG		1
300 #define IFLIB_QUEUE_WORKING		2
301 /* maximum number of txqs that can share an rx interrupt */
302 #define IFLIB_MAX_TX_SHARED_INTR	4
303 
304 /* this should really scale with ring size - this is a fairly arbitrary value */
305 #define TX_BATCH_SIZE			32
306 
307 #define IFLIB_RESTART_BUDGET		8
308 
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 struct iflib_txq {
314 	qidx_t		ift_in_use;
315 	qidx_t		ift_cidx;
316 	qidx_t		ift_cidx_processed;
317 	qidx_t		ift_pidx;
318 	uint8_t		ift_gen;
319 	uint8_t		ift_br_offset;
320 	uint16_t	ift_npending;
321 	uint16_t	ift_db_pending;
322 	uint16_t	ift_rs_pending;
323 	/* implicit pad */
324 	uint8_t		ift_txd_size[8];
325 	uint64_t	ift_processed;
326 	uint64_t	ift_cleaned;
327 	uint64_t	ift_cleaned_prev;
328 #if MEMORY_LOGGING
329 	uint64_t	ift_enqueued;
330 	uint64_t	ift_dequeued;
331 #endif
332 	uint64_t	ift_no_tx_dma_setup;
333 	uint64_t	ift_no_desc_avail;
334 	uint64_t	ift_mbuf_defrag_failed;
335 	uint64_t	ift_mbuf_defrag;
336 	uint64_t	ift_map_failed;
337 	uint64_t	ift_txd_encap_efbig;
338 	uint64_t	ift_pullups;
339 	uint64_t	ift_last_timer_tick;
340 
341 	struct mtx	ift_mtx;
342 	struct mtx	ift_db_mtx;
343 
344 	/* constant values */
345 	if_ctx_t	ift_ctx;
346 	struct ifmp_ring        *ift_br;
347 	struct grouptask	ift_task;
348 	qidx_t		ift_size;
349 	uint16_t	ift_id;
350 	struct callout	ift_timer;
351 
352 	if_txsd_vec_t	ift_sds;
353 	uint8_t		ift_qstatus;
354 	uint8_t		ift_closed;
355 	uint8_t		ift_update_freq;
356 	struct iflib_filter_info ift_filter_info;
357 	bus_dma_tag_t	ift_buf_tag;
358 	bus_dma_tag_t	ift_tso_buf_tag;
359 	iflib_dma_info_t	ift_ifdi;
360 #define MTX_NAME_LEN 16
361 	char                    ift_mtx_name[MTX_NAME_LEN];
362 	char                    ift_db_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 
383 	bitstr_t 	*ifl_rx_bitmap;
384 	qidx_t		ifl_fragidx;
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t	ifl_buf_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
397 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
398 }  __aligned(CACHE_LINE_SIZE);
399 
400 static inline qidx_t
401 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
402 {
403 	qidx_t used;
404 
405 	if (pidx > cidx)
406 		used = pidx - cidx;
407 	else if (pidx < cidx)
408 		used = size - cidx + pidx;
409 	else if (gen == 0 && pidx == cidx)
410 		used = 0;
411 	else if (gen == 1 && pidx == cidx)
412 		used = size;
413 	else
414 		panic("bad state");
415 
416 	return (used);
417 }
418 
419 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
420 
421 #define IDXDIFF(head, tail, wrap) \
422 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
423 
424 struct iflib_rxq {
425 	/* If there is a separate completion queue -
426 	 * these are the cq cidx and pidx. Otherwise
427 	 * these are unused.
428 	 */
429 	qidx_t		ifr_size;
430 	qidx_t		ifr_cq_cidx;
431 	qidx_t		ifr_cq_pidx;
432 	uint8_t		ifr_cq_gen;
433 	uint8_t		ifr_fl_offset;
434 
435 	if_ctx_t	ifr_ctx;
436 	iflib_fl_t	ifr_fl;
437 	uint64_t	ifr_rx_irq;
438 	uint16_t	ifr_id;
439 	uint8_t		ifr_lro_enabled;
440 	uint8_t		ifr_nfl;
441 	uint8_t		ifr_ntxqirq;
442 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
443 	struct lro_ctrl			ifr_lc;
444 	struct grouptask        ifr_task;
445 	struct iflib_filter_info ifr_filter_info;
446 	iflib_dma_info_t		ifr_ifdi;
447 
448 	/* dynamically allocate if any drivers need a value substantially larger than this */
449 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
450 #ifdef IFLIB_DIAGNOSTICS
451 	uint64_t ifr_cpu_exec_count[256];
452 #endif
453 }  __aligned(CACHE_LINE_SIZE);
454 
455 typedef struct if_rxsd {
456 	caddr_t *ifsd_cl;
457 	struct mbuf **ifsd_m;
458 	iflib_fl_t ifsd_fl;
459 	qidx_t ifsd_cidx;
460 } *if_rxsd_t;
461 
462 /* multiple of word size */
463 #ifdef __LP64__
464 #define PKT_INFO_SIZE	6
465 #define RXD_INFO_SIZE	5
466 #define PKT_TYPE uint64_t
467 #else
468 #define PKT_INFO_SIZE	11
469 #define RXD_INFO_SIZE	8
470 #define PKT_TYPE uint32_t
471 #endif
472 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
473 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
474 
475 typedef struct if_pkt_info_pad {
476 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
477 } *if_pkt_info_pad_t;
478 typedef struct if_rxd_info_pad {
479 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
480 } *if_rxd_info_pad_t;
481 
482 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
483 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
484 
485 
486 static inline void
487 pkt_info_zero(if_pkt_info_t pi)
488 {
489 	if_pkt_info_pad_t pi_pad;
490 
491 	pi_pad = (if_pkt_info_pad_t)pi;
492 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
493 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
494 #ifndef __LP64__
495 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
496 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
497 #endif
498 }
499 
500 static device_method_t iflib_pseudo_methods[] = {
501 	DEVMETHOD(device_attach, noop_attach),
502 	DEVMETHOD(device_detach, iflib_pseudo_detach),
503 	DEVMETHOD_END
504 };
505 
506 driver_t iflib_pseudodriver = {
507 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
508 };
509 
510 static inline void
511 rxd_info_zero(if_rxd_info_t ri)
512 {
513 	if_rxd_info_pad_t ri_pad;
514 	int i;
515 
516 	ri_pad = (if_rxd_info_pad_t)ri;
517 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
518 		ri_pad->rxd_val[i] = 0;
519 		ri_pad->rxd_val[i+1] = 0;
520 		ri_pad->rxd_val[i+2] = 0;
521 		ri_pad->rxd_val[i+3] = 0;
522 	}
523 #ifdef __LP64__
524 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
525 #endif
526 }
527 
528 /*
529  * Only allow a single packet to take up most 1/nth of the tx ring
530  */
531 #define MAX_SINGLE_PACKET_FRACTION 12
532 #define IF_BAD_DMA (bus_addr_t)-1
533 
534 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
535 
536 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
537 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
538 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
539 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
540 
541 
542 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
543 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
544 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
545 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
546 
547 
548 
549 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
550 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
551 
552 void
553 iflib_set_detach(if_ctx_t ctx)
554 {
555 	STATE_LOCK(ctx);
556 	ctx->ifc_flags |= IFC_IN_DETACH;
557 	STATE_UNLOCK(ctx);
558 }
559 
560 /* Our boot-time initialization hook */
561 static int	iflib_module_event_handler(module_t, int, void *);
562 
563 static moduledata_t iflib_moduledata = {
564 	"iflib",
565 	iflib_module_event_handler,
566 	NULL
567 };
568 
569 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
570 MODULE_VERSION(iflib, 1);
571 
572 MODULE_DEPEND(iflib, pci, 1, 1, 1);
573 MODULE_DEPEND(iflib, ether, 1, 1, 1);
574 
575 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
576 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
577 
578 #ifndef IFLIB_DEBUG_COUNTERS
579 #ifdef INVARIANTS
580 #define IFLIB_DEBUG_COUNTERS 1
581 #else
582 #define IFLIB_DEBUG_COUNTERS 0
583 #endif /* !INVARIANTS */
584 #endif
585 
586 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
587                    "iflib driver parameters");
588 
589 /*
590  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
591  */
592 static int iflib_min_tx_latency = 0;
593 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
594 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
595 static int iflib_no_tx_batch = 0;
596 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
597 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
598 
599 
600 #if IFLIB_DEBUG_COUNTERS
601 
602 static int iflib_tx_seen;
603 static int iflib_tx_sent;
604 static int iflib_tx_encap;
605 static int iflib_rx_allocs;
606 static int iflib_fl_refills;
607 static int iflib_fl_refills_large;
608 static int iflib_tx_frees;
609 
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
611 		   &iflib_tx_seen, 0, "# tx mbufs seen");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
613 		   &iflib_tx_sent, 0, "# tx mbufs sent");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
615 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
616 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
617 		   &iflib_tx_frees, 0, "# tx frees");
618 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
619 		   &iflib_rx_allocs, 0, "# rx allocations");
620 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
621 		   &iflib_fl_refills, 0, "# refills");
622 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
623 		   &iflib_fl_refills_large, 0, "# large refills");
624 
625 
626 static int iflib_txq_drain_flushing;
627 static int iflib_txq_drain_oactive;
628 static int iflib_txq_drain_notready;
629 
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
631 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
633 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
634 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
635 		   &iflib_txq_drain_notready, 0, "# drain notready");
636 
637 
638 static int iflib_encap_load_mbuf_fail;
639 static int iflib_encap_pad_mbuf_fail;
640 static int iflib_encap_txq_avail_fail;
641 static int iflib_encap_txd_encap_fail;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
644 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
646 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
647 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
648 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
649 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
650 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
651 
652 static int iflib_task_fn_rxs;
653 static int iflib_rx_intr_enables;
654 static int iflib_fast_intrs;
655 static int iflib_rx_unavail;
656 static int iflib_rx_ctx_inactive;
657 static int iflib_rx_if_input;
658 static int iflib_rx_mbuf_null;
659 static int iflib_rxd_flush;
660 
661 static int iflib_verbose_debug;
662 
663 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
664 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
666 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
667 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
668 		   &iflib_fast_intrs, 0, "# fast_intr calls");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
670 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
671 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
672 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
673 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
674 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
675 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
676 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
677 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
678 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
679 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
680 		   &iflib_verbose_debug, 0, "enable verbose debugging");
681 
682 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
683 static void
684 iflib_debug_reset(void)
685 {
686 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
687 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
688 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
689 		iflib_txq_drain_notready =
690 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
691 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
692 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
693 		iflib_rx_unavail =
694 		iflib_rx_ctx_inactive = iflib_rx_if_input =
695 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
696 }
697 
698 #else
699 #define DBG_COUNTER_INC(name)
700 static void iflib_debug_reset(void) {}
701 #endif
702 
703 #define IFLIB_DEBUG 0
704 
705 static void iflib_tx_structures_free(if_ctx_t ctx);
706 static void iflib_rx_structures_free(if_ctx_t ctx);
707 static int iflib_queues_alloc(if_ctx_t ctx);
708 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
709 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
710 static int iflib_qset_structures_setup(if_ctx_t ctx);
711 static int iflib_msix_init(if_ctx_t ctx);
712 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
713 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
714 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
715 #ifdef ALTQ
716 static void iflib_altq_if_start(if_t ifp);
717 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
718 #endif
719 static int iflib_register(if_ctx_t);
720 static void iflib_init_locked(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
722 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
723 static void iflib_ifmp_purge(iflib_txq_t txq);
724 static void _iflib_pre_assert(if_softc_ctx_t scctx);
725 static void iflib_if_init_locked(if_ctx_t ctx);
726 static void iflib_free_intr_mem(if_ctx_t ctx);
727 #ifndef __NO_STRICT_ALIGNMENT
728 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
729 #endif
730 
731 NETDUMP_DEFINE(iflib);
732 
733 #ifdef DEV_NETMAP
734 #include <sys/selinfo.h>
735 #include <net/netmap.h>
736 #include <dev/netmap/netmap_kern.h>
737 
738 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
739 
740 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
741 
742 /*
743  * device-specific sysctl variables:
744  *
745  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
746  *	During regular operations the CRC is stripped, but on some
747  *	hardware reception of frames not multiple of 64 is slower,
748  *	so using crcstrip=0 helps in benchmarks.
749  *
750  * iflib_rx_miss, iflib_rx_miss_bufs:
751  *	count packets that might be missed due to lost interrupts.
752  */
753 SYSCTL_DECL(_dev_netmap);
754 /*
755  * The xl driver by default strips CRCs and we do not override it.
756  */
757 
758 int iflib_crcstrip = 1;
759 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
760     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
761 
762 int iflib_rx_miss, iflib_rx_miss_bufs;
763 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
764     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
765 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
766     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
767 
768 /*
769  * Register/unregister. We are already under netmap lock.
770  * Only called on the first register or the last unregister.
771  */
772 static int
773 iflib_netmap_register(struct netmap_adapter *na, int onoff)
774 {
775 	struct ifnet *ifp = na->ifp;
776 	if_ctx_t ctx = ifp->if_softc;
777 	int status;
778 
779 	CTX_LOCK(ctx);
780 	IFDI_INTR_DISABLE(ctx);
781 
782 	/* Tell the stack that the interface is no longer active */
783 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
784 
785 	if (!CTX_IS_VF(ctx))
786 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
787 
788 	/* enable or disable flags and callbacks in na and ifp */
789 	if (onoff) {
790 		nm_set_native_flags(na);
791 	} else {
792 		nm_clear_native_flags(na);
793 	}
794 	iflib_stop(ctx);
795 	iflib_init_locked(ctx);
796 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
797 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
798 	if (status)
799 		nm_clear_native_flags(na);
800 	CTX_UNLOCK(ctx);
801 	return (status);
802 }
803 
804 static int
805 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
806 {
807 	struct netmap_adapter *na = kring->na;
808 	u_int const lim = kring->nkr_num_slots - 1;
809 	u_int head = kring->rhead;
810 	struct netmap_ring *ring = kring->ring;
811 	bus_dmamap_t *map;
812 	struct if_rxd_update iru;
813 	if_ctx_t ctx = rxq->ifr_ctx;
814 	iflib_fl_t fl = &rxq->ifr_fl[0];
815 	uint32_t refill_pidx, nic_i;
816 #if IFLIB_DEBUG_COUNTERS
817 	int rf_count = 0;
818 #endif
819 
820 	if (nm_i == head && __predict_true(!init))
821 		return 0;
822 	iru_init(&iru, rxq, 0 /* flid */);
823 	map = fl->ifl_sds.ifsd_map;
824 	refill_pidx = netmap_idx_k2n(kring, nm_i);
825 	/*
826 	 * IMPORTANT: we must leave one free slot in the ring,
827 	 * so move head back by one unit
828 	 */
829 	head = nm_prev(head, lim);
830 	nic_i = UINT_MAX;
831 	DBG_COUNTER_INC(fl_refills);
832 	while (nm_i != head) {
833 #if IFLIB_DEBUG_COUNTERS
834 		if (++rf_count == 9)
835 			DBG_COUNTER_INC(fl_refills_large);
836 #endif
837 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
838 			struct netmap_slot *slot = &ring->slot[nm_i];
839 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
840 			uint32_t nic_i_dma = refill_pidx;
841 			nic_i = netmap_idx_k2n(kring, nm_i);
842 
843 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
844 
845 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
846 			        return netmap_ring_reinit(kring);
847 
848 			fl->ifl_vm_addrs[tmp_pidx] = addr;
849 			if (__predict_false(init)) {
850 				netmap_load_map(na, fl->ifl_buf_tag,
851 				    map[nic_i], addr);
852 			} else if (slot->flags & NS_BUF_CHANGED) {
853 				/* buffer has changed, reload map */
854 				netmap_reload_map(na, fl->ifl_buf_tag,
855 				    map[nic_i], addr);
856 			}
857 			slot->flags &= ~NS_BUF_CHANGED;
858 
859 			nm_i = nm_next(nm_i, lim);
860 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
861 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
862 				continue;
863 
864 			iru.iru_pidx = refill_pidx;
865 			iru.iru_count = tmp_pidx+1;
866 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
867 			refill_pidx = nic_i;
868 			for (int n = 0; n < iru.iru_count; n++) {
869 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
870 						BUS_DMASYNC_PREREAD);
871 				/* XXX - change this to not use the netmap func*/
872 				nic_i_dma = nm_next(nic_i_dma, lim);
873 			}
874 		}
875 	}
876 	kring->nr_hwcur = head;
877 
878 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
879 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
880 	if (__predict_true(nic_i != UINT_MAX)) {
881 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
882 		DBG_COUNTER_INC(rxd_flush);
883 	}
884 	return (0);
885 }
886 
887 /*
888  * Reconcile kernel and user view of the transmit ring.
889  *
890  * All information is in the kring.
891  * Userspace wants to send packets up to the one before kring->rhead,
892  * kernel knows kring->nr_hwcur is the first unsent packet.
893  *
894  * Here we push packets out (as many as possible), and possibly
895  * reclaim buffers from previously completed transmission.
896  *
897  * The caller (netmap) guarantees that there is only one instance
898  * running at any time. Any interference with other driver
899  * methods should be handled by the individual drivers.
900  */
901 static int
902 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
903 {
904 	struct netmap_adapter *na = kring->na;
905 	struct ifnet *ifp = na->ifp;
906 	struct netmap_ring *ring = kring->ring;
907 	u_int nm_i;	/* index into the netmap kring */
908 	u_int nic_i;	/* index into the NIC ring */
909 	u_int n;
910 	u_int const lim = kring->nkr_num_slots - 1;
911 	u_int const head = kring->rhead;
912 	struct if_pkt_info pi;
913 
914 	/*
915 	 * interrupts on every tx packet are expensive so request
916 	 * them every half ring, or where NS_REPORT is set
917 	 */
918 	u_int report_frequency = kring->nkr_num_slots >> 1;
919 	/* device-specific */
920 	if_ctx_t ctx = ifp->if_softc;
921 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
922 
923 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
924 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
925 
926 	/*
927 	 * First part: process new packets to send.
928 	 * nm_i is the current index in the netmap kring,
929 	 * nic_i is the corresponding index in the NIC ring.
930 	 *
931 	 * If we have packets to send (nm_i != head)
932 	 * iterate over the netmap ring, fetch length and update
933 	 * the corresponding slot in the NIC ring. Some drivers also
934 	 * need to update the buffer's physical address in the NIC slot
935 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
936 	 *
937 	 * The netmap_reload_map() calls is especially expensive,
938 	 * even when (as in this case) the tag is 0, so do only
939 	 * when the buffer has actually changed.
940 	 *
941 	 * If possible do not set the report/intr bit on all slots,
942 	 * but only a few times per ring or when NS_REPORT is set.
943 	 *
944 	 * Finally, on 10G and faster drivers, it might be useful
945 	 * to prefetch the next slot and txr entry.
946 	 */
947 
948 	nm_i = kring->nr_hwcur;
949 	if (nm_i != head) {	/* we have new packets to send */
950 		pkt_info_zero(&pi);
951 		pi.ipi_segs = txq->ift_segs;
952 		pi.ipi_qsidx = kring->ring_id;
953 		nic_i = netmap_idx_k2n(kring, nm_i);
954 
955 		__builtin_prefetch(&ring->slot[nm_i]);
956 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
957 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
958 
959 		for (n = 0; nm_i != head; n++) {
960 			struct netmap_slot *slot = &ring->slot[nm_i];
961 			u_int len = slot->len;
962 			uint64_t paddr;
963 			void *addr = PNMB(na, slot, &paddr);
964 			int flags = (slot->flags & NS_REPORT ||
965 				nic_i == 0 || nic_i == report_frequency) ?
966 				IPI_TX_INTR : 0;
967 
968 			/* device-specific */
969 			pi.ipi_len = len;
970 			pi.ipi_segs[0].ds_addr = paddr;
971 			pi.ipi_segs[0].ds_len = len;
972 			pi.ipi_nsegs = 1;
973 			pi.ipi_ndescs = 0;
974 			pi.ipi_pidx = nic_i;
975 			pi.ipi_flags = flags;
976 
977 			/* Fill the slot in the NIC ring. */
978 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
979 			DBG_COUNTER_INC(tx_encap);
980 
981 			/* prefetch for next round */
982 			__builtin_prefetch(&ring->slot[nm_i + 1]);
983 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
984 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
985 
986 			NM_CHECK_ADDR_LEN(na, addr, len);
987 
988 			if (slot->flags & NS_BUF_CHANGED) {
989 				/* buffer has changed, reload map */
990 				netmap_reload_map(na, txq->ift_buf_tag,
991 				    txq->ift_sds.ifsd_map[nic_i], addr);
992 			}
993 			/* make sure changes to the buffer are synced */
994 			bus_dmamap_sync(txq->ift_buf_tag,
995 			    txq->ift_sds.ifsd_map[nic_i],
996 			    BUS_DMASYNC_PREWRITE);
997 
998 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
999 			nm_i = nm_next(nm_i, lim);
1000 			nic_i = nm_next(nic_i, lim);
1001 		}
1002 		kring->nr_hwcur = nm_i;
1003 
1004 		/* synchronize the NIC ring */
1005 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1006 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1007 
1008 		/* (re)start the tx unit up to slot nic_i (excluded) */
1009 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1010 	}
1011 
1012 	/*
1013 	 * Second part: reclaim buffers for completed transmissions.
1014 	 *
1015 	 * If there are unclaimed buffers, attempt to reclaim them.
1016 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1017 	 * minimal delay, then trigger the tx handler which will spin in the
1018 	 * group task queue.
1019 	 */
1020 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1021 		if (iflib_tx_credits_update(ctx, txq)) {
1022 			/* some tx completed, increment avail */
1023 			nic_i = txq->ift_cidx_processed;
1024 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1025 		}
1026 	}
1027 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1028 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1029 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1030 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1031 	}
1032 	return (0);
1033 }
1034 
1035 /*
1036  * Reconcile kernel and user view of the receive ring.
1037  * Same as for the txsync, this routine must be efficient.
1038  * The caller guarantees a single invocations, but races against
1039  * the rest of the driver should be handled here.
1040  *
1041  * On call, kring->rhead is the first packet that userspace wants
1042  * to keep, and kring->rcur is the wakeup point.
1043  * The kernel has previously reported packets up to kring->rtail.
1044  *
1045  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1046  * of whether or not we received an interrupt.
1047  */
1048 static int
1049 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1050 {
1051 	struct netmap_adapter *na = kring->na;
1052 	struct netmap_ring *ring = kring->ring;
1053 	iflib_fl_t fl;
1054 	uint32_t nm_i;	/* index into the netmap ring */
1055 	uint32_t nic_i;	/* index into the NIC ring */
1056 	u_int i, n;
1057 	u_int const lim = kring->nkr_num_slots - 1;
1058 	u_int const head = kring->rhead;
1059 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1060 	struct if_rxd_info ri;
1061 
1062 	struct ifnet *ifp = na->ifp;
1063 	if_ctx_t ctx = ifp->if_softc;
1064 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1065 	if (head > lim)
1066 		return netmap_ring_reinit(kring);
1067 
1068 	/*
1069 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1070 	 */
1071 
1072 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1073 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1074 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1075 	}
1076 
1077 	/*
1078 	 * First part: import newly received packets.
1079 	 *
1080 	 * nm_i is the index of the next free slot in the netmap ring,
1081 	 * nic_i is the index of the next received packet in the NIC ring,
1082 	 * and they may differ in case if_init() has been called while
1083 	 * in netmap mode. For the receive ring we have
1084 	 *
1085 	 *	nic_i = rxr->next_check;
1086 	 *	nm_i = kring->nr_hwtail (previous)
1087 	 * and
1088 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1089 	 *
1090 	 * rxr->next_check is set to 0 on a ring reinit
1091 	 */
1092 	if (netmap_no_pendintr || force_update) {
1093 		int crclen = iflib_crcstrip ? 0 : 4;
1094 		int error, avail;
1095 
1096 		for (i = 0; i < rxq->ifr_nfl; i++) {
1097 			fl = &rxq->ifr_fl[i];
1098 			nic_i = fl->ifl_cidx;
1099 			nm_i = netmap_idx_n2k(kring, nic_i);
1100 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1101 			    rxq->ifr_id, nic_i, USHRT_MAX);
1102 			for (n = 0; avail > 0; n++, avail--) {
1103 				rxd_info_zero(&ri);
1104 				ri.iri_frags = rxq->ifr_frags;
1105 				ri.iri_qsidx = kring->ring_id;
1106 				ri.iri_ifp = ctx->ifc_ifp;
1107 				ri.iri_cidx = nic_i;
1108 
1109 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1110 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1111 				ring->slot[nm_i].flags = 0;
1112 				bus_dmamap_sync(fl->ifl_buf_tag,
1113 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1114 				nm_i = nm_next(nm_i, lim);
1115 				nic_i = nm_next(nic_i, lim);
1116 			}
1117 			if (n) { /* update the state variables */
1118 				if (netmap_no_pendintr && !force_update) {
1119 					/* diagnostics */
1120 					iflib_rx_miss ++;
1121 					iflib_rx_miss_bufs += n;
1122 				}
1123 				fl->ifl_cidx = nic_i;
1124 				kring->nr_hwtail = nm_i;
1125 			}
1126 			kring->nr_kflags &= ~NKR_PENDINTR;
1127 		}
1128 	}
1129 	/*
1130 	 * Second part: skip past packets that userspace has released.
1131 	 * (kring->nr_hwcur to head excluded),
1132 	 * and make the buffers available for reception.
1133 	 * As usual nm_i is the index in the netmap ring,
1134 	 * nic_i is the index in the NIC ring, and
1135 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1136 	 */
1137 	/* XXX not sure how this will work with multiple free lists */
1138 	nm_i = kring->nr_hwcur;
1139 
1140 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1141 }
1142 
1143 static void
1144 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1145 {
1146 	struct ifnet *ifp = na->ifp;
1147 	if_ctx_t ctx = ifp->if_softc;
1148 
1149 	CTX_LOCK(ctx);
1150 	if (onoff) {
1151 		IFDI_INTR_ENABLE(ctx);
1152 	} else {
1153 		IFDI_INTR_DISABLE(ctx);
1154 	}
1155 	CTX_UNLOCK(ctx);
1156 }
1157 
1158 
1159 static int
1160 iflib_netmap_attach(if_ctx_t ctx)
1161 {
1162 	struct netmap_adapter na;
1163 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1164 
1165 	bzero(&na, sizeof(na));
1166 
1167 	na.ifp = ctx->ifc_ifp;
1168 	na.na_flags = NAF_BDG_MAYSLEEP;
1169 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1170 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1171 
1172 	na.num_tx_desc = scctx->isc_ntxd[0];
1173 	na.num_rx_desc = scctx->isc_nrxd[0];
1174 	na.nm_txsync = iflib_netmap_txsync;
1175 	na.nm_rxsync = iflib_netmap_rxsync;
1176 	na.nm_register = iflib_netmap_register;
1177 	na.nm_intr = iflib_netmap_intr;
1178 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1179 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1180 	return (netmap_attach(&na));
1181 }
1182 
1183 static void
1184 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1185 {
1186 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1187 	struct netmap_slot *slot;
1188 
1189 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1190 	if (slot == NULL)
1191 		return;
1192 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1193 
1194 		/*
1195 		 * In netmap mode, set the map for the packet buffer.
1196 		 * NOTE: Some drivers (not this one) also need to set
1197 		 * the physical buffer address in the NIC ring.
1198 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1199 		 * netmap slot index, si
1200 		 */
1201 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1202 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1203 		    NMB(na, slot + si));
1204 	}
1205 }
1206 
1207 static void
1208 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1209 {
1210 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1211 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1212 	struct netmap_slot *slot;
1213 	uint32_t nm_i;
1214 
1215 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1216 	if (slot == NULL)
1217 		return;
1218 	nm_i = netmap_idx_n2k(kring, 0);
1219 	netmap_fl_refill(rxq, kring, nm_i, true);
1220 }
1221 
1222 static void
1223 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1224 {
1225 	struct netmap_kring *kring;
1226 	uint16_t txqid;
1227 
1228 	txqid = txq->ift_id;
1229 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1230 
1231 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1232 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1233 		    BUS_DMASYNC_POSTREAD);
1234 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1235 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1236 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1237 			if (hz < 2000)
1238 				*reset_on = 1;
1239 			else
1240 				*reset_on = hz / 1000;
1241 		}
1242 	}
1243 }
1244 
1245 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1246 
1247 #else
1248 #define iflib_netmap_txq_init(ctx, txq)
1249 #define iflib_netmap_rxq_init(ctx, rxq)
1250 #define iflib_netmap_detach(ifp)
1251 
1252 #define iflib_netmap_attach(ctx) (0)
1253 #define netmap_rx_irq(ifp, qid, budget) (0)
1254 #define netmap_tx_irq(ifp, qid) do {} while (0)
1255 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1256 
1257 #endif
1258 
1259 #if defined(__i386__) || defined(__amd64__)
1260 static __inline void
1261 prefetch(void *x)
1262 {
1263 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1264 }
1265 static __inline void
1266 prefetch2cachelines(void *x)
1267 {
1268 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1269 #if (CACHE_LINE_SIZE < 128)
1270 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1271 #endif
1272 }
1273 #else
1274 #define prefetch(x)
1275 #define prefetch2cachelines(x)
1276 #endif
1277 
1278 static void
1279 iflib_gen_mac(if_ctx_t ctx)
1280 {
1281 	struct thread *td;
1282 	MD5_CTX mdctx;
1283 	char uuid[HOSTUUIDLEN+1];
1284 	char buf[HOSTUUIDLEN+16];
1285 	uint8_t *mac;
1286 	unsigned char digest[16];
1287 
1288 	td = curthread;
1289 	mac = ctx->ifc_mac;
1290 	uuid[HOSTUUIDLEN] = 0;
1291 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1292 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1293 	/*
1294 	 * Generate a pseudo-random, deterministic MAC
1295 	 * address based on the UUID and unit number.
1296 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1297 	 */
1298 	MD5Init(&mdctx);
1299 	MD5Update(&mdctx, buf, strlen(buf));
1300 	MD5Final(digest, &mdctx);
1301 
1302 	mac[0] = 0x58;
1303 	mac[1] = 0x9C;
1304 	mac[2] = 0xFC;
1305 	mac[3] = digest[0];
1306 	mac[4] = digest[1];
1307 	mac[5] = digest[2];
1308 }
1309 
1310 static void
1311 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1312 {
1313 	iflib_fl_t fl;
1314 
1315 	fl = &rxq->ifr_fl[flid];
1316 	iru->iru_paddrs = fl->ifl_bus_addrs;
1317 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1318 	iru->iru_idxs = fl->ifl_rxd_idxs;
1319 	iru->iru_qsidx = rxq->ifr_id;
1320 	iru->iru_buf_size = fl->ifl_buf_size;
1321 	iru->iru_flidx = fl->ifl_id;
1322 }
1323 
1324 static void
1325 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1326 {
1327 	if (err)
1328 		return;
1329 	*(bus_addr_t *) arg = segs[0].ds_addr;
1330 }
1331 
1332 int
1333 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1334 {
1335 	int err;
1336 	device_t dev = ctx->ifc_dev;
1337 
1338 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1339 				align, 0,		/* alignment, bounds */
1340 				BUS_SPACE_MAXADDR,	/* lowaddr */
1341 				BUS_SPACE_MAXADDR,	/* highaddr */
1342 				NULL, NULL,		/* filter, filterarg */
1343 				size,			/* maxsize */
1344 				1,			/* nsegments */
1345 				size,			/* maxsegsize */
1346 				BUS_DMA_ALLOCNOW,	/* flags */
1347 				NULL,			/* lockfunc */
1348 				NULL,			/* lockarg */
1349 				&dma->idi_tag);
1350 	if (err) {
1351 		device_printf(dev,
1352 		    "%s: bus_dma_tag_create failed: %d\n",
1353 		    __func__, err);
1354 		goto fail_0;
1355 	}
1356 
1357 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1358 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1359 	if (err) {
1360 		device_printf(dev,
1361 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1362 		    __func__, (uintmax_t)size, err);
1363 		goto fail_1;
1364 	}
1365 
1366 	dma->idi_paddr = IF_BAD_DMA;
1367 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1368 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1369 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1370 		device_printf(dev,
1371 		    "%s: bus_dmamap_load failed: %d\n",
1372 		    __func__, err);
1373 		goto fail_2;
1374 	}
1375 
1376 	dma->idi_size = size;
1377 	return (0);
1378 
1379 fail_2:
1380 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1381 fail_1:
1382 	bus_dma_tag_destroy(dma->idi_tag);
1383 fail_0:
1384 	dma->idi_tag = NULL;
1385 
1386 	return (err);
1387 }
1388 
1389 int
1390 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1391 {
1392 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1393 
1394 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1395 
1396 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1397 }
1398 
1399 int
1400 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1401 {
1402 	int i, err;
1403 	iflib_dma_info_t *dmaiter;
1404 
1405 	dmaiter = dmalist;
1406 	for (i = 0; i < count; i++, dmaiter++) {
1407 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1408 			break;
1409 	}
1410 	if (err)
1411 		iflib_dma_free_multi(dmalist, i);
1412 	return (err);
1413 }
1414 
1415 void
1416 iflib_dma_free(iflib_dma_info_t dma)
1417 {
1418 	if (dma->idi_tag == NULL)
1419 		return;
1420 	if (dma->idi_paddr != IF_BAD_DMA) {
1421 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1422 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1423 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1424 		dma->idi_paddr = IF_BAD_DMA;
1425 	}
1426 	if (dma->idi_vaddr != NULL) {
1427 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1428 		dma->idi_vaddr = NULL;
1429 	}
1430 	bus_dma_tag_destroy(dma->idi_tag);
1431 	dma->idi_tag = NULL;
1432 }
1433 
1434 void
1435 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1436 {
1437 	int i;
1438 	iflib_dma_info_t *dmaiter = dmalist;
1439 
1440 	for (i = 0; i < count; i++, dmaiter++)
1441 		iflib_dma_free(*dmaiter);
1442 }
1443 
1444 #ifdef EARLY_AP_STARTUP
1445 static const int iflib_started = 1;
1446 #else
1447 /*
1448  * We used to abuse the smp_started flag to decide if the queues have been
1449  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1450  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1451  * is set.  Run a SYSINIT() strictly after that to just set a usable
1452  * completion flag.
1453  */
1454 
1455 static int iflib_started;
1456 
1457 static void
1458 iflib_record_started(void *arg)
1459 {
1460 	iflib_started = 1;
1461 }
1462 
1463 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1464 	iflib_record_started, NULL);
1465 #endif
1466 
1467 static int
1468 iflib_fast_intr(void *arg)
1469 {
1470 	iflib_filter_info_t info = arg;
1471 	struct grouptask *gtask = info->ifi_task;
1472 	int result;
1473 
1474 	if (!iflib_started)
1475 		return (FILTER_STRAY);
1476 
1477 	DBG_COUNTER_INC(fast_intrs);
1478 	if (info->ifi_filter != NULL) {
1479 		result = info->ifi_filter(info->ifi_filter_arg);
1480 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1481 			return (result);
1482 	}
1483 
1484 	GROUPTASK_ENQUEUE(gtask);
1485 	return (FILTER_HANDLED);
1486 }
1487 
1488 static int
1489 iflib_fast_intr_rxtx(void *arg)
1490 {
1491 	iflib_filter_info_t info = arg;
1492 	struct grouptask *gtask = info->ifi_task;
1493 	if_ctx_t ctx;
1494 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1495 	iflib_txq_t txq;
1496 	void *sc;
1497 	int i, cidx, result;
1498 	qidx_t txqid;
1499 
1500 	if (!iflib_started)
1501 		return (FILTER_STRAY);
1502 
1503 	DBG_COUNTER_INC(fast_intrs);
1504 	if (info->ifi_filter != NULL) {
1505 		result = info->ifi_filter(info->ifi_filter_arg);
1506 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1507 			return (result);
1508 	}
1509 
1510 	ctx = rxq->ifr_ctx;
1511 	sc = ctx->ifc_softc;
1512 	MPASS(rxq->ifr_ntxqirq);
1513 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1514 		txqid = rxq->ifr_txqid[i];
1515 		txq = &ctx->ifc_txqs[txqid];
1516 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1517 		    BUS_DMASYNC_POSTREAD);
1518 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1519 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1520 			continue;
1521 		}
1522 		GROUPTASK_ENQUEUE(&txq->ift_task);
1523 	}
1524 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1525 		cidx = rxq->ifr_cq_cidx;
1526 	else
1527 		cidx = rxq->ifr_fl[0].ifl_cidx;
1528 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1529 		GROUPTASK_ENQUEUE(gtask);
1530 	else {
1531 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1532 		DBG_COUNTER_INC(rx_intr_enables);
1533 	}
1534 	return (FILTER_HANDLED);
1535 }
1536 
1537 
1538 static int
1539 iflib_fast_intr_ctx(void *arg)
1540 {
1541 	iflib_filter_info_t info = arg;
1542 	struct grouptask *gtask = info->ifi_task;
1543 	int result;
1544 
1545 	if (!iflib_started)
1546 		return (FILTER_STRAY);
1547 
1548 	DBG_COUNTER_INC(fast_intrs);
1549 	if (info->ifi_filter != NULL) {
1550 		result = info->ifi_filter(info->ifi_filter_arg);
1551 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1552 			return (result);
1553 	}
1554 
1555 	GROUPTASK_ENQUEUE(gtask);
1556 	return (FILTER_HANDLED);
1557 }
1558 
1559 static int
1560 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1561 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1562 		 const char *name)
1563 {
1564 	int rc, flags;
1565 	struct resource *res;
1566 	void *tag = NULL;
1567 	device_t dev = ctx->ifc_dev;
1568 
1569 	flags = RF_ACTIVE;
1570 	if (ctx->ifc_flags & IFC_LEGACY)
1571 		flags |= RF_SHAREABLE;
1572 	MPASS(rid < 512);
1573 	irq->ii_rid = rid;
1574 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1575 	if (res == NULL) {
1576 		device_printf(dev,
1577 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1578 		return (ENOMEM);
1579 	}
1580 	irq->ii_res = res;
1581 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1582 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1583 						filter, handler, arg, &tag);
1584 	if (rc != 0) {
1585 		device_printf(dev,
1586 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1587 					  rid, name ? name : "unknown", rc);
1588 		return (rc);
1589 	} else if (name)
1590 		bus_describe_intr(dev, res, tag, "%s", name);
1591 
1592 	irq->ii_tag = tag;
1593 	return (0);
1594 }
1595 
1596 
1597 /*********************************************************************
1598  *
1599  *  Allocate DMA resources for TX buffers as well as memory for the TX
1600  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1601  *  iflib_sw_tx_desc_array structure, storing all the information that
1602  *  is needed to transmit a packet on the wire.  This is called only
1603  *  once at attach, setup is done every reset.
1604  *
1605  **********************************************************************/
1606 static int
1607 iflib_txsd_alloc(iflib_txq_t txq)
1608 {
1609 	if_ctx_t ctx = txq->ift_ctx;
1610 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1611 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1612 	device_t dev = ctx->ifc_dev;
1613 	bus_size_t tsomaxsize;
1614 	int err, nsegments, ntsosegments;
1615 	bool tso;
1616 
1617 	nsegments = scctx->isc_tx_nsegments;
1618 	ntsosegments = scctx->isc_tx_tso_segments_max;
1619 	tsomaxsize = scctx->isc_tx_tso_size_max;
1620 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1621 		tsomaxsize += sizeof(struct ether_vlan_header);
1622 	MPASS(scctx->isc_ntxd[0] > 0);
1623 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1624 	MPASS(nsegments > 0);
1625 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1626 		MPASS(ntsosegments > 0);
1627 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1628 	}
1629 
1630 	/*
1631 	 * Set up DMA tags for TX buffers.
1632 	 */
1633 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1634 			       1, 0,			/* alignment, bounds */
1635 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1636 			       BUS_SPACE_MAXADDR,	/* highaddr */
1637 			       NULL, NULL,		/* filter, filterarg */
1638 			       sctx->isc_tx_maxsize,		/* maxsize */
1639 			       nsegments,	/* nsegments */
1640 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1641 			       0,			/* flags */
1642 			       NULL,			/* lockfunc */
1643 			       NULL,			/* lockfuncarg */
1644 			       &txq->ift_buf_tag))) {
1645 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1646 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1647 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1648 		goto fail;
1649 	}
1650 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1651 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1652 			       1, 0,			/* alignment, bounds */
1653 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1654 			       BUS_SPACE_MAXADDR,	/* highaddr */
1655 			       NULL, NULL,		/* filter, filterarg */
1656 			       tsomaxsize,		/* maxsize */
1657 			       ntsosegments,	/* nsegments */
1658 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1659 			       0,			/* flags */
1660 			       NULL,			/* lockfunc */
1661 			       NULL,			/* lockfuncarg */
1662 			       &txq->ift_tso_buf_tag))) {
1663 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1664 		    err);
1665 		goto fail;
1666 	}
1667 
1668 	/* Allocate memory for the TX mbuf map. */
1669 	if (!(txq->ift_sds.ifsd_m =
1670 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1671 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1672 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1673 		err = ENOMEM;
1674 		goto fail;
1675 	}
1676 
1677 	/*
1678 	 * Create the DMA maps for TX buffers.
1679 	 */
1680 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1681 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1682 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1683 		device_printf(dev,
1684 		    "Unable to allocate TX buffer DMA map memory\n");
1685 		err = ENOMEM;
1686 		goto fail;
1687 	}
1688 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1689 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1690 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1691 		device_printf(dev,
1692 		    "Unable to allocate TSO TX buffer map memory\n");
1693 		err = ENOMEM;
1694 		goto fail;
1695 	}
1696 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1697 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1698 		    &txq->ift_sds.ifsd_map[i]);
1699 		if (err != 0) {
1700 			device_printf(dev, "Unable to create TX DMA map\n");
1701 			goto fail;
1702 		}
1703 		if (!tso)
1704 			continue;
1705 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1706 		    &txq->ift_sds.ifsd_tso_map[i]);
1707 		if (err != 0) {
1708 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1709 			goto fail;
1710 		}
1711 	}
1712 	return (0);
1713 fail:
1714 	/* We free all, it handles case where we are in the middle */
1715 	iflib_tx_structures_free(ctx);
1716 	return (err);
1717 }
1718 
1719 static void
1720 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1721 {
1722 	bus_dmamap_t map;
1723 
1724 	map = NULL;
1725 	if (txq->ift_sds.ifsd_map != NULL)
1726 		map = txq->ift_sds.ifsd_map[i];
1727 	if (map != NULL) {
1728 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1729 		bus_dmamap_unload(txq->ift_buf_tag, map);
1730 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1731 		txq->ift_sds.ifsd_map[i] = NULL;
1732 	}
1733 
1734 	map = NULL;
1735 	if (txq->ift_sds.ifsd_tso_map != NULL)
1736 		map = txq->ift_sds.ifsd_tso_map[i];
1737 	if (map != NULL) {
1738 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1739 		    BUS_DMASYNC_POSTWRITE);
1740 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1741 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1742 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1743 	}
1744 }
1745 
1746 static void
1747 iflib_txq_destroy(iflib_txq_t txq)
1748 {
1749 	if_ctx_t ctx = txq->ift_ctx;
1750 
1751 	for (int i = 0; i < txq->ift_size; i++)
1752 		iflib_txsd_destroy(ctx, txq, i);
1753 	if (txq->ift_sds.ifsd_map != NULL) {
1754 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1755 		txq->ift_sds.ifsd_map = NULL;
1756 	}
1757 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1758 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1759 		txq->ift_sds.ifsd_tso_map = NULL;
1760 	}
1761 	if (txq->ift_sds.ifsd_m != NULL) {
1762 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1763 		txq->ift_sds.ifsd_m = NULL;
1764 	}
1765 	if (txq->ift_buf_tag != NULL) {
1766 		bus_dma_tag_destroy(txq->ift_buf_tag);
1767 		txq->ift_buf_tag = NULL;
1768 	}
1769 	if (txq->ift_tso_buf_tag != NULL) {
1770 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1771 		txq->ift_tso_buf_tag = NULL;
1772 	}
1773 }
1774 
1775 static void
1776 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1777 {
1778 	struct mbuf **mp;
1779 
1780 	mp = &txq->ift_sds.ifsd_m[i];
1781 	if (*mp == NULL)
1782 		return;
1783 
1784 	if (txq->ift_sds.ifsd_map != NULL) {
1785 		bus_dmamap_sync(txq->ift_buf_tag,
1786 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1787 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1788 	}
1789 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1790 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1791 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1792 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1793 		    txq->ift_sds.ifsd_tso_map[i]);
1794 	}
1795 	m_free(*mp);
1796 	DBG_COUNTER_INC(tx_frees);
1797 	*mp = NULL;
1798 }
1799 
1800 static int
1801 iflib_txq_setup(iflib_txq_t txq)
1802 {
1803 	if_ctx_t ctx = txq->ift_ctx;
1804 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1805 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1806 	iflib_dma_info_t di;
1807 	int i;
1808 
1809 	/* Set number of descriptors available */
1810 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1811 	/* XXX make configurable */
1812 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1813 
1814 	/* Reset indices */
1815 	txq->ift_cidx_processed = 0;
1816 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1817 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1818 
1819 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1820 		bzero((void *)di->idi_vaddr, di->idi_size);
1821 
1822 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1823 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1824 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1825 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1826 	return (0);
1827 }
1828 
1829 /*********************************************************************
1830  *
1831  *  Allocate DMA resources for RX buffers as well as memory for the RX
1832  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1833  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1834  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1835  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1836  *  packet, the maximum number of entries we'll need is equal to the
1837  *  number of hardware receive descriptors that we've allocated.
1838  *
1839  **********************************************************************/
1840 static int
1841 iflib_rxsd_alloc(iflib_rxq_t rxq)
1842 {
1843 	if_ctx_t ctx = rxq->ifr_ctx;
1844 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1845 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1846 	device_t dev = ctx->ifc_dev;
1847 	iflib_fl_t fl;
1848 	int			err;
1849 
1850 	MPASS(scctx->isc_nrxd[0] > 0);
1851 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1852 
1853 	fl = rxq->ifr_fl;
1854 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1855 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1856 		/* Set up DMA tag for RX buffers. */
1857 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1858 					 1, 0,			/* alignment, bounds */
1859 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1860 					 BUS_SPACE_MAXADDR,	/* highaddr */
1861 					 NULL, NULL,		/* filter, filterarg */
1862 					 sctx->isc_rx_maxsize,	/* maxsize */
1863 					 sctx->isc_rx_nsegments,	/* nsegments */
1864 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1865 					 0,			/* flags */
1866 					 NULL,			/* lockfunc */
1867 					 NULL,			/* lockarg */
1868 					 &fl->ifl_buf_tag);
1869 		if (err) {
1870 			device_printf(dev,
1871 			    "Unable to allocate RX DMA tag: %d\n", err);
1872 			goto fail;
1873 		}
1874 
1875 		/* Allocate memory for the RX mbuf map. */
1876 		if (!(fl->ifl_sds.ifsd_m =
1877 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1878 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1879 			device_printf(dev,
1880 			    "Unable to allocate RX mbuf map memory\n");
1881 			err = ENOMEM;
1882 			goto fail;
1883 		}
1884 
1885 		/* Allocate memory for the direct RX cluster pointer map. */
1886 		if (!(fl->ifl_sds.ifsd_cl =
1887 		      (caddr_t *) malloc(sizeof(caddr_t) *
1888 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1889 			device_printf(dev,
1890 			    "Unable to allocate RX cluster map memory\n");
1891 			err = ENOMEM;
1892 			goto fail;
1893 		}
1894 
1895 		/* Allocate memory for the RX cluster bus address map. */
1896 		if (!(fl->ifl_sds.ifsd_ba =
1897 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1898 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1899 			device_printf(dev,
1900 			    "Unable to allocate RX bus address map memory\n");
1901 			err = ENOMEM;
1902 			goto fail;
1903 		}
1904 
1905 		/*
1906 		 * Create the DMA maps for RX buffers.
1907 		 */
1908 		if (!(fl->ifl_sds.ifsd_map =
1909 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1910 			device_printf(dev,
1911 			    "Unable to allocate RX buffer DMA map memory\n");
1912 			err = ENOMEM;
1913 			goto fail;
1914 		}
1915 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1916 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1917 			    &fl->ifl_sds.ifsd_map[i]);
1918 			if (err != 0) {
1919 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1920 				goto fail;
1921 			}
1922 		}
1923 	}
1924 	return (0);
1925 
1926 fail:
1927 	iflib_rx_structures_free(ctx);
1928 	return (err);
1929 }
1930 
1931 
1932 /*
1933  * Internal service routines
1934  */
1935 
1936 struct rxq_refill_cb_arg {
1937 	int               error;
1938 	bus_dma_segment_t seg;
1939 	int               nseg;
1940 };
1941 
1942 static void
1943 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1944 {
1945 	struct rxq_refill_cb_arg *cb_arg = arg;
1946 
1947 	cb_arg->error = error;
1948 	cb_arg->seg = segs[0];
1949 	cb_arg->nseg = nseg;
1950 }
1951 
1952 /**
1953  *	rxq_refill - refill an rxq  free-buffer list
1954  *	@ctx: the iflib context
1955  *	@rxq: the free-list to refill
1956  *	@n: the number of new buffers to allocate
1957  *
1958  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1959  *	The caller must assure that @n does not exceed the queue's capacity.
1960  */
1961 static void
1962 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1963 {
1964 	struct if_rxd_update iru;
1965 	struct rxq_refill_cb_arg cb_arg;
1966 	struct mbuf *m;
1967 	caddr_t cl, *sd_cl;
1968 	struct mbuf **sd_m;
1969 	bus_dmamap_t *sd_map;
1970 	bus_addr_t bus_addr, *sd_ba;
1971 	int err, frag_idx, i, idx, n, pidx;
1972 	qidx_t credits;
1973 
1974 	sd_m = fl->ifl_sds.ifsd_m;
1975 	sd_map = fl->ifl_sds.ifsd_map;
1976 	sd_cl = fl->ifl_sds.ifsd_cl;
1977 	sd_ba = fl->ifl_sds.ifsd_ba;
1978 	pidx = fl->ifl_pidx;
1979 	idx = pidx;
1980 	frag_idx = fl->ifl_fragidx;
1981 	credits = fl->ifl_credits;
1982 
1983 	i = 0;
1984 	n = count;
1985 	MPASS(n > 0);
1986 	MPASS(credits + n <= fl->ifl_size);
1987 
1988 	if (pidx < fl->ifl_cidx)
1989 		MPASS(pidx + n <= fl->ifl_cidx);
1990 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1991 		MPASS(fl->ifl_gen == 0);
1992 	if (pidx > fl->ifl_cidx)
1993 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1994 
1995 	DBG_COUNTER_INC(fl_refills);
1996 	if (n > 8)
1997 		DBG_COUNTER_INC(fl_refills_large);
1998 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1999 	while (n--) {
2000 		/*
2001 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2002 		 * initialized after rx.
2003 		 *
2004 		 * If the cluster is still set then we know a minimum sized packet was received
2005 		 */
2006 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2007 		    &frag_idx);
2008 		if (frag_idx < 0)
2009 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2010 		MPASS(frag_idx >= 0);
2011 		if ((cl = sd_cl[frag_idx]) == NULL) {
2012 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
2013 				break;
2014 
2015 			cb_arg.error = 0;
2016 			MPASS(sd_map != NULL);
2017 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2018 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2019 			    BUS_DMA_NOWAIT);
2020 			if (err != 0 || cb_arg.error) {
2021 				/*
2022 				 * !zone_pack ?
2023 				 */
2024 				if (fl->ifl_zone == zone_pack)
2025 					uma_zfree(fl->ifl_zone, cl);
2026 				break;
2027 			}
2028 
2029 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2030 			sd_cl[frag_idx] = cl;
2031 #if MEMORY_LOGGING
2032 			fl->ifl_cl_enqueued++;
2033 #endif
2034 		} else {
2035 			bus_addr = sd_ba[frag_idx];
2036 		}
2037 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2038 		    BUS_DMASYNC_PREREAD);
2039 
2040 		MPASS(sd_m[frag_idx] == NULL);
2041 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2042 			break;
2043 		}
2044 		sd_m[frag_idx] = m;
2045 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2046 #if MEMORY_LOGGING
2047 		fl->ifl_m_enqueued++;
2048 #endif
2049 
2050 		DBG_COUNTER_INC(rx_allocs);
2051 		fl->ifl_rxd_idxs[i] = frag_idx;
2052 		fl->ifl_bus_addrs[i] = bus_addr;
2053 		fl->ifl_vm_addrs[i] = cl;
2054 		credits++;
2055 		i++;
2056 		MPASS(credits <= fl->ifl_size);
2057 		if (++idx == fl->ifl_size) {
2058 			fl->ifl_gen = 1;
2059 			idx = 0;
2060 		}
2061 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2062 			iru.iru_pidx = pidx;
2063 			iru.iru_count = i;
2064 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2065 			i = 0;
2066 			pidx = idx;
2067 			fl->ifl_pidx = idx;
2068 			fl->ifl_credits = credits;
2069 		}
2070 	}
2071 
2072 	if (i) {
2073 		iru.iru_pidx = pidx;
2074 		iru.iru_count = i;
2075 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2076 		fl->ifl_pidx = idx;
2077 		fl->ifl_credits = credits;
2078 	}
2079 	DBG_COUNTER_INC(rxd_flush);
2080 	if (fl->ifl_pidx == 0)
2081 		pidx = fl->ifl_size - 1;
2082 	else
2083 		pidx = fl->ifl_pidx - 1;
2084 
2085 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2086 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2087 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2088 	fl->ifl_fragidx = frag_idx;
2089 }
2090 
2091 static __inline void
2092 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2093 {
2094 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2095 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2096 #ifdef INVARIANTS
2097 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2098 #endif
2099 
2100 	MPASS(fl->ifl_credits <= fl->ifl_size);
2101 	MPASS(reclaimable == delta);
2102 
2103 	if (reclaimable > 0)
2104 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2105 }
2106 
2107 uint8_t
2108 iflib_in_detach(if_ctx_t ctx)
2109 {
2110 	bool in_detach;
2111 	STATE_LOCK(ctx);
2112 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2113 	STATE_UNLOCK(ctx);
2114 	return (in_detach);
2115 }
2116 
2117 static void
2118 iflib_fl_bufs_free(iflib_fl_t fl)
2119 {
2120 	iflib_dma_info_t idi = fl->ifl_ifdi;
2121 	bus_dmamap_t sd_map;
2122 	uint32_t i;
2123 
2124 	for (i = 0; i < fl->ifl_size; i++) {
2125 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2126 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2127 
2128 		if (*sd_cl != NULL) {
2129 			sd_map = fl->ifl_sds.ifsd_map[i];
2130 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2131 			    BUS_DMASYNC_POSTREAD);
2132 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2133 			if (*sd_cl != NULL)
2134 				uma_zfree(fl->ifl_zone, *sd_cl);
2135 			// XXX: Should this get moved out?
2136 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2137 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2138 			if (*sd_m != NULL) {
2139 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2140 				uma_zfree(zone_mbuf, *sd_m);
2141 			}
2142 		} else {
2143 			MPASS(*sd_cl == NULL);
2144 			MPASS(*sd_m == NULL);
2145 		}
2146 #if MEMORY_LOGGING
2147 		fl->ifl_m_dequeued++;
2148 		fl->ifl_cl_dequeued++;
2149 #endif
2150 		*sd_cl = NULL;
2151 		*sd_m = NULL;
2152 	}
2153 #ifdef INVARIANTS
2154 	for (i = 0; i < fl->ifl_size; i++) {
2155 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2156 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2157 	}
2158 #endif
2159 	/*
2160 	 * Reset free list values
2161 	 */
2162 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2163 	bzero(idi->idi_vaddr, idi->idi_size);
2164 }
2165 
2166 /*********************************************************************
2167  *
2168  *  Initialize a receive ring and its buffers.
2169  *
2170  **********************************************************************/
2171 static int
2172 iflib_fl_setup(iflib_fl_t fl)
2173 {
2174 	iflib_rxq_t rxq = fl->ifl_rxq;
2175 	if_ctx_t ctx = rxq->ifr_ctx;
2176 
2177 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2178 	/*
2179 	** Free current RX buffer structs and their mbufs
2180 	*/
2181 	iflib_fl_bufs_free(fl);
2182 	/* Now replenish the mbufs */
2183 	MPASS(fl->ifl_credits == 0);
2184 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2185 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2186 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2187 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2188 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2189 
2190 
2191 	/* avoid pre-allocating zillions of clusters to an idle card
2192 	 * potentially speeding up attach
2193 	 */
2194 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2195 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2196 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2197 		return (ENOBUFS);
2198 	/*
2199 	 * handle failure
2200 	 */
2201 	MPASS(rxq != NULL);
2202 	MPASS(fl->ifl_ifdi != NULL);
2203 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2204 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2205 	return (0);
2206 }
2207 
2208 /*********************************************************************
2209  *
2210  *  Free receive ring data structures
2211  *
2212  **********************************************************************/
2213 static void
2214 iflib_rx_sds_free(iflib_rxq_t rxq)
2215 {
2216 	iflib_fl_t fl;
2217 	int i, j;
2218 
2219 	if (rxq->ifr_fl != NULL) {
2220 		for (i = 0; i < rxq->ifr_nfl; i++) {
2221 			fl = &rxq->ifr_fl[i];
2222 			if (fl->ifl_buf_tag != NULL) {
2223 				if (fl->ifl_sds.ifsd_map != NULL) {
2224 					for (j = 0; j < fl->ifl_size; j++) {
2225 						if (fl->ifl_sds.ifsd_map[j] ==
2226 						    NULL)
2227 							continue;
2228 						bus_dmamap_sync(
2229 						    fl->ifl_buf_tag,
2230 						    fl->ifl_sds.ifsd_map[j],
2231 						    BUS_DMASYNC_POSTREAD);
2232 						bus_dmamap_unload(
2233 						    fl->ifl_buf_tag,
2234 						    fl->ifl_sds.ifsd_map[j]);
2235 					}
2236 				}
2237 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2238 				fl->ifl_buf_tag = NULL;
2239 			}
2240 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2241 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2242 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2243 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2244 			fl->ifl_sds.ifsd_m = NULL;
2245 			fl->ifl_sds.ifsd_cl = NULL;
2246 			fl->ifl_sds.ifsd_ba = NULL;
2247 			fl->ifl_sds.ifsd_map = NULL;
2248 		}
2249 		free(rxq->ifr_fl, M_IFLIB);
2250 		rxq->ifr_fl = NULL;
2251 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2252 	}
2253 }
2254 
2255 /*
2256  * MI independent logic
2257  *
2258  */
2259 static void
2260 iflib_timer(void *arg)
2261 {
2262 	iflib_txq_t txq = arg;
2263 	if_ctx_t ctx = txq->ift_ctx;
2264 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2265 	uint64_t this_tick = ticks;
2266 	uint32_t reset_on = hz / 2;
2267 
2268 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2269 		return;
2270 	/*
2271 	** Check on the state of the TX queue(s), this
2272 	** can be done without the lock because its RO
2273 	** and the HUNG state will be static if set.
2274 	*/
2275 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2276 		txq->ift_last_timer_tick = this_tick;
2277 		IFDI_TIMER(ctx, txq->ift_id);
2278 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2279 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2280 		     (sctx->isc_pause_frames == 0)))
2281 			goto hung;
2282 
2283 		if (ifmp_ring_is_stalled(txq->ift_br))
2284 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2285 		txq->ift_cleaned_prev = txq->ift_cleaned;
2286 	}
2287 #ifdef DEV_NETMAP
2288 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2289 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2290 #endif
2291 	/* handle any laggards */
2292 	if (txq->ift_db_pending)
2293 		GROUPTASK_ENQUEUE(&txq->ift_task);
2294 
2295 	sctx->isc_pause_frames = 0;
2296 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2297 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2298 	return;
2299  hung:
2300 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2301 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2302 	STATE_LOCK(ctx);
2303 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2304 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2305 	iflib_admin_intr_deferred(ctx);
2306 	STATE_UNLOCK(ctx);
2307 }
2308 
2309 static void
2310 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2311 {
2312 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2313 
2314 	/*
2315 	 * XXX don't set the max_frame_size to larger
2316 	 * than the hardware can handle
2317 	 */
2318 	if (sctx->isc_max_frame_size <= MCLBYTES)
2319 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2320 	else
2321 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2322 }
2323 
2324 uint32_t
2325 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2326 {
2327 	return (ctx->ifc_rx_mbuf_sz);
2328 }
2329 
2330 static void
2331 iflib_init_locked(if_ctx_t ctx)
2332 {
2333 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2334 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2335 	if_t ifp = ctx->ifc_ifp;
2336 	iflib_fl_t fl;
2337 	iflib_txq_t txq;
2338 	iflib_rxq_t rxq;
2339 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2340 
2341 
2342 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2343 	IFDI_INTR_DISABLE(ctx);
2344 
2345 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2346 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2347 	/* Set hardware offload abilities */
2348 	if_clearhwassist(ifp);
2349 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2350 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2351 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2352 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2353 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2354 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2355 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2356 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2357 
2358 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2359 		CALLOUT_LOCK(txq);
2360 		callout_stop(&txq->ift_timer);
2361 		CALLOUT_UNLOCK(txq);
2362 		iflib_netmap_txq_init(ctx, txq);
2363 	}
2364 
2365 	/*
2366 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2367 	 * that drivers can use the value when setting up the hardware receive
2368 	 * buffers.
2369 	 */
2370 	iflib_calc_rx_mbuf_sz(ctx);
2371 
2372 #ifdef INVARIANTS
2373 	i = if_getdrvflags(ifp);
2374 #endif
2375 	IFDI_INIT(ctx);
2376 	MPASS(if_getdrvflags(ifp) == i);
2377 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2378 		/* XXX this should really be done on a per-queue basis */
2379 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2380 			MPASS(rxq->ifr_id == i);
2381 			iflib_netmap_rxq_init(ctx, rxq);
2382 			continue;
2383 		}
2384 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2385 			if (iflib_fl_setup(fl)) {
2386 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2387 				goto done;
2388 			}
2389 		}
2390 	}
2391 done:
2392 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2393 	IFDI_INTR_ENABLE(ctx);
2394 	txq = ctx->ifc_txqs;
2395 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2396 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2397 			txq->ift_timer.c_cpu);
2398 }
2399 
2400 static int
2401 iflib_media_change(if_t ifp)
2402 {
2403 	if_ctx_t ctx = if_getsoftc(ifp);
2404 	int err;
2405 
2406 	CTX_LOCK(ctx);
2407 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2408 		iflib_init_locked(ctx);
2409 	CTX_UNLOCK(ctx);
2410 	return (err);
2411 }
2412 
2413 static void
2414 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2415 {
2416 	if_ctx_t ctx = if_getsoftc(ifp);
2417 
2418 	CTX_LOCK(ctx);
2419 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2420 	IFDI_MEDIA_STATUS(ctx, ifmr);
2421 	CTX_UNLOCK(ctx);
2422 }
2423 
2424 void
2425 iflib_stop(if_ctx_t ctx)
2426 {
2427 	iflib_txq_t txq = ctx->ifc_txqs;
2428 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2429 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2430 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2431 	iflib_dma_info_t di;
2432 	iflib_fl_t fl;
2433 	int i, j;
2434 
2435 	/* Tell the stack that the interface is no longer active */
2436 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2437 
2438 	IFDI_INTR_DISABLE(ctx);
2439 	DELAY(1000);
2440 	IFDI_STOP(ctx);
2441 	DELAY(1000);
2442 
2443 	iflib_debug_reset();
2444 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2445 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2446 		/* make sure all transmitters have completed before proceeding XXX */
2447 
2448 		CALLOUT_LOCK(txq);
2449 		callout_stop(&txq->ift_timer);
2450 		CALLOUT_UNLOCK(txq);
2451 
2452 		/* clean any enqueued buffers */
2453 		iflib_ifmp_purge(txq);
2454 		/* Free any existing tx buffers. */
2455 		for (j = 0; j < txq->ift_size; j++) {
2456 			iflib_txsd_free(ctx, txq, j);
2457 		}
2458 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2459 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2460 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2461 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2462 		txq->ift_pullups = 0;
2463 		ifmp_ring_reset_stats(txq->ift_br);
2464 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2465 			bzero((void *)di->idi_vaddr, di->idi_size);
2466 	}
2467 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2468 		/* make sure all transmitters have completed before proceeding XXX */
2469 
2470 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2471 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2472 			bzero((void *)di->idi_vaddr, di->idi_size);
2473 		/* also resets the free lists pidx/cidx */
2474 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2475 			iflib_fl_bufs_free(fl);
2476 	}
2477 }
2478 
2479 static inline caddr_t
2480 calc_next_rxd(iflib_fl_t fl, int cidx)
2481 {
2482 	qidx_t size;
2483 	int nrxd;
2484 	caddr_t start, end, cur, next;
2485 
2486 	nrxd = fl->ifl_size;
2487 	size = fl->ifl_rxd_size;
2488 	start = fl->ifl_ifdi->idi_vaddr;
2489 
2490 	if (__predict_false(size == 0))
2491 		return (start);
2492 	cur = start + size*cidx;
2493 	end = start + size*nrxd;
2494 	next = CACHE_PTR_NEXT(cur);
2495 	return (next < end ? next : start);
2496 }
2497 
2498 static inline void
2499 prefetch_pkts(iflib_fl_t fl, int cidx)
2500 {
2501 	int nextptr;
2502 	int nrxd = fl->ifl_size;
2503 	caddr_t next_rxd;
2504 
2505 
2506 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2507 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2508 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2509 	next_rxd = calc_next_rxd(fl, cidx);
2510 	prefetch(next_rxd);
2511 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2512 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2513 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2514 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2515 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2516 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2517 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2518 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2519 }
2520 
2521 static void
2522 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2523 {
2524 	int flid, cidx;
2525 	bus_dmamap_t map;
2526 	iflib_fl_t fl;
2527 	int next;
2528 
2529 	map = NULL;
2530 	flid = irf->irf_flid;
2531 	cidx = irf->irf_idx;
2532 	fl = &rxq->ifr_fl[flid];
2533 	sd->ifsd_fl = fl;
2534 	sd->ifsd_cidx = cidx;
2535 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2536 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2537 	fl->ifl_credits--;
2538 #if MEMORY_LOGGING
2539 	fl->ifl_m_dequeued++;
2540 #endif
2541 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2542 		prefetch_pkts(fl, cidx);
2543 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2544 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2545 	map = fl->ifl_sds.ifsd_map[cidx];
2546 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2547 
2548 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2549 	MPASS(fl->ifl_cidx == cidx);
2550 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2551 	if (unload)
2552 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2553 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2554 	if (__predict_false(fl->ifl_cidx == 0))
2555 		fl->ifl_gen = 0;
2556 	bit_clear(fl->ifl_rx_bitmap, cidx);
2557 }
2558 
2559 static struct mbuf *
2560 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2561 {
2562 	int i, padlen , flags;
2563 	struct mbuf *m, *mh, *mt;
2564 	caddr_t cl;
2565 
2566 	i = 0;
2567 	mh = NULL;
2568 	do {
2569 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2570 
2571 		MPASS(*sd->ifsd_cl != NULL);
2572 		MPASS(*sd->ifsd_m != NULL);
2573 
2574 		/* Don't include zero-length frags */
2575 		if (ri->iri_frags[i].irf_len == 0) {
2576 			/* XXX we can save the cluster here, but not the mbuf */
2577 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2578 			m_free(*sd->ifsd_m);
2579 			*sd->ifsd_m = NULL;
2580 			continue;
2581 		}
2582 		m = *sd->ifsd_m;
2583 		*sd->ifsd_m = NULL;
2584 		if (mh == NULL) {
2585 			flags = M_PKTHDR|M_EXT;
2586 			mh = mt = m;
2587 			padlen = ri->iri_pad;
2588 		} else {
2589 			flags = M_EXT;
2590 			mt->m_next = m;
2591 			mt = m;
2592 			/* assuming padding is only on the first fragment */
2593 			padlen = 0;
2594 		}
2595 		cl = *sd->ifsd_cl;
2596 		*sd->ifsd_cl = NULL;
2597 
2598 		/* Can these two be made one ? */
2599 		m_init(m, M_NOWAIT, MT_DATA, flags);
2600 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2601 		/*
2602 		 * These must follow m_init and m_cljset
2603 		 */
2604 		m->m_data += padlen;
2605 		ri->iri_len -= padlen;
2606 		m->m_len = ri->iri_frags[i].irf_len;
2607 	} while (++i < ri->iri_nfrags);
2608 
2609 	return (mh);
2610 }
2611 
2612 /*
2613  * Process one software descriptor
2614  */
2615 static struct mbuf *
2616 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2617 {
2618 	struct if_rxsd sd;
2619 	struct mbuf *m;
2620 
2621 	/* should I merge this back in now that the two paths are basically duplicated? */
2622 	if (ri->iri_nfrags == 1 &&
2623 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2624 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2625 		m = *sd.ifsd_m;
2626 		*sd.ifsd_m = NULL;
2627 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2628 #ifndef __NO_STRICT_ALIGNMENT
2629 		if (!IP_ALIGNED(m))
2630 			m->m_data += 2;
2631 #endif
2632 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2633 		m->m_len = ri->iri_frags[0].irf_len;
2634        } else {
2635 		m = assemble_segments(rxq, ri, &sd);
2636 	}
2637 	m->m_pkthdr.len = ri->iri_len;
2638 	m->m_pkthdr.rcvif = ri->iri_ifp;
2639 	m->m_flags |= ri->iri_flags;
2640 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2641 	m->m_pkthdr.flowid = ri->iri_flowid;
2642 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2643 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2644 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2645 	return (m);
2646 }
2647 
2648 #if defined(INET6) || defined(INET)
2649 static void
2650 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2651 {
2652 	CURVNET_SET(lc->ifp->if_vnet);
2653 #if defined(INET6)
2654 	*v6 = VNET(ip6_forwarding);
2655 #endif
2656 #if defined(INET)
2657 	*v4 = VNET(ipforwarding);
2658 #endif
2659 	CURVNET_RESTORE();
2660 }
2661 
2662 /*
2663  * Returns true if it's possible this packet could be LROed.
2664  * if it returns false, it is guaranteed that tcp_lro_rx()
2665  * would not return zero.
2666  */
2667 static bool
2668 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2669 {
2670 	struct ether_header *eh;
2671 	uint16_t eh_type;
2672 
2673 	eh = mtod(m, struct ether_header *);
2674 	eh_type = ntohs(eh->ether_type);
2675 	switch (eh_type) {
2676 #if defined(INET6)
2677 		case ETHERTYPE_IPV6:
2678 			return !v6_forwarding;
2679 #endif
2680 #if defined (INET)
2681 		case ETHERTYPE_IP:
2682 			return !v4_forwarding;
2683 #endif
2684 	}
2685 
2686 	return false;
2687 }
2688 #else
2689 static void
2690 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2691 {
2692 }
2693 #endif
2694 
2695 static bool
2696 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2697 {
2698 	if_ctx_t ctx = rxq->ifr_ctx;
2699 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2700 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2701 	int avail, i;
2702 	qidx_t *cidxp;
2703 	struct if_rxd_info ri;
2704 	int err, budget_left, rx_bytes, rx_pkts;
2705 	iflib_fl_t fl;
2706 	struct ifnet *ifp;
2707 	int lro_enabled;
2708 	bool v4_forwarding, v6_forwarding, lro_possible;
2709 
2710 	/*
2711 	 * XXX early demux data packets so that if_input processing only handles
2712 	 * acks in interrupt context
2713 	 */
2714 	struct mbuf *m, *mh, *mt, *mf;
2715 
2716 	lro_possible = v4_forwarding = v6_forwarding = false;
2717 	ifp = ctx->ifc_ifp;
2718 	mh = mt = NULL;
2719 	MPASS(budget > 0);
2720 	rx_pkts	= rx_bytes = 0;
2721 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2722 		cidxp = &rxq->ifr_cq_cidx;
2723 	else
2724 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2725 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2726 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2727 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2728 		DBG_COUNTER_INC(rx_unavail);
2729 		return (false);
2730 	}
2731 
2732 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2733 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2734 			DBG_COUNTER_INC(rx_ctx_inactive);
2735 			break;
2736 		}
2737 		/*
2738 		 * Reset client set fields to their default values
2739 		 */
2740 		rxd_info_zero(&ri);
2741 		ri.iri_qsidx = rxq->ifr_id;
2742 		ri.iri_cidx = *cidxp;
2743 		ri.iri_ifp = ifp;
2744 		ri.iri_frags = rxq->ifr_frags;
2745 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2746 
2747 		if (err)
2748 			goto err;
2749 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2750 			*cidxp = ri.iri_cidx;
2751 			/* Update our consumer index */
2752 			/* XXX NB: shurd - check if this is still safe */
2753 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2754 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2755 				rxq->ifr_cq_gen = 0;
2756 			}
2757 			/* was this only a completion queue message? */
2758 			if (__predict_false(ri.iri_nfrags == 0))
2759 				continue;
2760 		}
2761 		MPASS(ri.iri_nfrags != 0);
2762 		MPASS(ri.iri_len != 0);
2763 
2764 		/* will advance the cidx on the corresponding free lists */
2765 		m = iflib_rxd_pkt_get(rxq, &ri);
2766 		avail--;
2767 		budget_left--;
2768 		if (avail == 0 && budget_left)
2769 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2770 
2771 		if (__predict_false(m == NULL)) {
2772 			DBG_COUNTER_INC(rx_mbuf_null);
2773 			continue;
2774 		}
2775 		/* imm_pkt: -- cxgb */
2776 		if (mh == NULL)
2777 			mh = mt = m;
2778 		else {
2779 			mt->m_nextpkt = m;
2780 			mt = m;
2781 		}
2782 	}
2783 	/* make sure that we can refill faster than drain */
2784 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2785 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2786 
2787 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2788 	if (lro_enabled)
2789 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2790 	mt = mf = NULL;
2791 	while (mh != NULL) {
2792 		m = mh;
2793 		mh = mh->m_nextpkt;
2794 		m->m_nextpkt = NULL;
2795 #ifndef __NO_STRICT_ALIGNMENT
2796 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2797 			continue;
2798 #endif
2799 		rx_bytes += m->m_pkthdr.len;
2800 		rx_pkts++;
2801 #if defined(INET6) || defined(INET)
2802 		if (lro_enabled) {
2803 			if (!lro_possible) {
2804 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2805 				if (lro_possible && mf != NULL) {
2806 					ifp->if_input(ifp, mf);
2807 					DBG_COUNTER_INC(rx_if_input);
2808 					mt = mf = NULL;
2809 				}
2810 			}
2811 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2812 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2813 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2814 					continue;
2815 			}
2816 		}
2817 #endif
2818 		if (lro_possible) {
2819 			ifp->if_input(ifp, m);
2820 			DBG_COUNTER_INC(rx_if_input);
2821 			continue;
2822 		}
2823 
2824 		if (mf == NULL)
2825 			mf = m;
2826 		if (mt != NULL)
2827 			mt->m_nextpkt = m;
2828 		mt = m;
2829 	}
2830 	if (mf != NULL) {
2831 		ifp->if_input(ifp, mf);
2832 		DBG_COUNTER_INC(rx_if_input);
2833 	}
2834 
2835 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2836 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2837 
2838 	/*
2839 	 * Flush any outstanding LRO work
2840 	 */
2841 #if defined(INET6) || defined(INET)
2842 	tcp_lro_flush_all(&rxq->ifr_lc);
2843 #endif
2844 	if (avail)
2845 		return true;
2846 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2847 err:
2848 	STATE_LOCK(ctx);
2849 	ctx->ifc_flags |= IFC_DO_RESET;
2850 	iflib_admin_intr_deferred(ctx);
2851 	STATE_UNLOCK(ctx);
2852 	return (false);
2853 }
2854 
2855 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2856 static inline qidx_t
2857 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2858 {
2859 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2860 	qidx_t minthresh = txq->ift_size / 8;
2861 	if (in_use > 4*minthresh)
2862 		return (notify_count);
2863 	if (in_use > 2*minthresh)
2864 		return (notify_count >> 1);
2865 	if (in_use > minthresh)
2866 		return (notify_count >> 3);
2867 	return (0);
2868 }
2869 
2870 static inline qidx_t
2871 txq_max_rs_deferred(iflib_txq_t txq)
2872 {
2873 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2874 	qidx_t minthresh = txq->ift_size / 8;
2875 	if (txq->ift_in_use > 4*minthresh)
2876 		return (notify_count);
2877 	if (txq->ift_in_use > 2*minthresh)
2878 		return (notify_count >> 1);
2879 	if (txq->ift_in_use > minthresh)
2880 		return (notify_count >> 2);
2881 	return (2);
2882 }
2883 
2884 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2885 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2886 
2887 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2888 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2889 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2890 
2891 /* forward compatibility for cxgb */
2892 #define FIRST_QSET(ctx) 0
2893 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2894 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2895 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2896 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2897 
2898 /* XXX we should be setting this to something other than zero */
2899 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2900 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2901     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2902 
2903 static inline bool
2904 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2905 {
2906 	qidx_t dbval, max;
2907 	bool rang;
2908 
2909 	rang = false;
2910 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2911 	if (ring || txq->ift_db_pending >= max) {
2912 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2913 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2914 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2915 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2916 		txq->ift_db_pending = txq->ift_npending = 0;
2917 		rang = true;
2918 	}
2919 	return (rang);
2920 }
2921 
2922 #ifdef PKT_DEBUG
2923 static void
2924 print_pkt(if_pkt_info_t pi)
2925 {
2926 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2927 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2928 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2929 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2930 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2931 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2932 }
2933 #endif
2934 
2935 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2936 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2937 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2938 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2939 
2940 static int
2941 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2942 {
2943 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2944 	struct ether_vlan_header *eh;
2945 	struct mbuf *m;
2946 
2947 	m = *mp;
2948 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2949 	    M_WRITABLE(m) == 0) {
2950 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2951 			return (ENOMEM);
2952 		} else {
2953 			m_freem(*mp);
2954 			DBG_COUNTER_INC(tx_frees);
2955 			*mp = m;
2956 		}
2957 	}
2958 
2959 	/*
2960 	 * Determine where frame payload starts.
2961 	 * Jump over vlan headers if already present,
2962 	 * helpful for QinQ too.
2963 	 */
2964 	if (__predict_false(m->m_len < sizeof(*eh))) {
2965 		txq->ift_pullups++;
2966 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2967 			return (ENOMEM);
2968 	}
2969 	eh = mtod(m, struct ether_vlan_header *);
2970 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2971 		pi->ipi_etype = ntohs(eh->evl_proto);
2972 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2973 	} else {
2974 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2975 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2976 	}
2977 
2978 	switch (pi->ipi_etype) {
2979 #ifdef INET
2980 	case ETHERTYPE_IP:
2981 	{
2982 		struct mbuf *n;
2983 		struct ip *ip = NULL;
2984 		struct tcphdr *th = NULL;
2985 		int minthlen;
2986 
2987 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2988 		if (__predict_false(m->m_len < minthlen)) {
2989 			/*
2990 			 * if this code bloat is causing too much of a hit
2991 			 * move it to a separate function and mark it noinline
2992 			 */
2993 			if (m->m_len == pi->ipi_ehdrlen) {
2994 				n = m->m_next;
2995 				MPASS(n);
2996 				if (n->m_len >= sizeof(*ip))  {
2997 					ip = (struct ip *)n->m_data;
2998 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2999 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3000 				} else {
3001 					txq->ift_pullups++;
3002 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3003 						return (ENOMEM);
3004 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3005 				}
3006 			} else {
3007 				txq->ift_pullups++;
3008 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3009 					return (ENOMEM);
3010 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3011 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3012 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3013 			}
3014 		} else {
3015 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3016 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3017 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3018 		}
3019 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3020 		pi->ipi_ipproto = ip->ip_p;
3021 		pi->ipi_flags |= IPI_TX_IPV4;
3022 
3023 		/* TCP checksum offload may require TCP header length */
3024 		if (IS_TX_OFFLOAD4(pi)) {
3025 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3026 				if (__predict_false(th == NULL)) {
3027 					txq->ift_pullups++;
3028 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3029 						return (ENOMEM);
3030 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3031 				}
3032 				pi->ipi_tcp_hflags = th->th_flags;
3033 				pi->ipi_tcp_hlen = th->th_off << 2;
3034 				pi->ipi_tcp_seq = th->th_seq;
3035 			}
3036 			if (IS_TSO4(pi)) {
3037 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3038 					return (ENXIO);
3039 				/*
3040 				 * TSO always requires hardware checksum offload.
3041 				 */
3042 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3043 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3044 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3045 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3046 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3047 					ip->ip_sum = 0;
3048 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3049 				}
3050 			}
3051 		}
3052 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3053                        ip->ip_sum = 0;
3054 
3055 		break;
3056 	}
3057 #endif
3058 #ifdef INET6
3059 	case ETHERTYPE_IPV6:
3060 	{
3061 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3062 		struct tcphdr *th;
3063 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3064 
3065 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3066 			txq->ift_pullups++;
3067 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3068 				return (ENOMEM);
3069 		}
3070 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3071 
3072 		/* XXX-BZ this will go badly in case of ext hdrs. */
3073 		pi->ipi_ipproto = ip6->ip6_nxt;
3074 		pi->ipi_flags |= IPI_TX_IPV6;
3075 
3076 		/* TCP checksum offload may require TCP header length */
3077 		if (IS_TX_OFFLOAD6(pi)) {
3078 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3079 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3080 					txq->ift_pullups++;
3081 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3082 						return (ENOMEM);
3083 				}
3084 				pi->ipi_tcp_hflags = th->th_flags;
3085 				pi->ipi_tcp_hlen = th->th_off << 2;
3086 				pi->ipi_tcp_seq = th->th_seq;
3087 			}
3088 			if (IS_TSO6(pi)) {
3089 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3090 					return (ENXIO);
3091 				/*
3092 				 * TSO always requires hardware checksum offload.
3093 				 */
3094 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3095 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3096 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3097 			}
3098 		}
3099 		break;
3100 	}
3101 #endif
3102 	default:
3103 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3104 		pi->ipi_ip_hlen = 0;
3105 		break;
3106 	}
3107 	*mp = m;
3108 
3109 	return (0);
3110 }
3111 
3112 /*
3113  * If dodgy hardware rejects the scatter gather chain we've handed it
3114  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3115  * m_defrag'd mbufs
3116  */
3117 static __noinline struct mbuf *
3118 iflib_remove_mbuf(iflib_txq_t txq)
3119 {
3120 	int ntxd, pidx;
3121 	struct mbuf *m, **ifsd_m;
3122 
3123 	ifsd_m = txq->ift_sds.ifsd_m;
3124 	ntxd = txq->ift_size;
3125 	pidx = txq->ift_pidx & (ntxd - 1);
3126 	ifsd_m = txq->ift_sds.ifsd_m;
3127 	m = ifsd_m[pidx];
3128 	ifsd_m[pidx] = NULL;
3129 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3130 	if (txq->ift_sds.ifsd_tso_map != NULL)
3131 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3132 		    txq->ift_sds.ifsd_tso_map[pidx]);
3133 #if MEMORY_LOGGING
3134 	txq->ift_dequeued++;
3135 #endif
3136 	return (m);
3137 }
3138 
3139 static inline caddr_t
3140 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3141 {
3142 	qidx_t size;
3143 	int ntxd;
3144 	caddr_t start, end, cur, next;
3145 
3146 	ntxd = txq->ift_size;
3147 	size = txq->ift_txd_size[qid];
3148 	start = txq->ift_ifdi[qid].idi_vaddr;
3149 
3150 	if (__predict_false(size == 0))
3151 		return (start);
3152 	cur = start + size*cidx;
3153 	end = start + size*ntxd;
3154 	next = CACHE_PTR_NEXT(cur);
3155 	return (next < end ? next : start);
3156 }
3157 
3158 /*
3159  * Pad an mbuf to ensure a minimum ethernet frame size.
3160  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3161  */
3162 static __noinline int
3163 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3164 {
3165 	/*
3166 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3167 	 * and ARP message is the smallest common payload I can think of
3168 	 */
3169 	static char pad[18];	/* just zeros */
3170 	int n;
3171 	struct mbuf *new_head;
3172 
3173 	if (!M_WRITABLE(*m_head)) {
3174 		new_head = m_dup(*m_head, M_NOWAIT);
3175 		if (new_head == NULL) {
3176 			m_freem(*m_head);
3177 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3178 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3179 			DBG_COUNTER_INC(tx_frees);
3180 			return ENOMEM;
3181 		}
3182 		m_freem(*m_head);
3183 		*m_head = new_head;
3184 	}
3185 
3186 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3187 	     n > 0; n -= sizeof(pad))
3188 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3189 			break;
3190 
3191 	if (n > 0) {
3192 		m_freem(*m_head);
3193 		device_printf(dev, "cannot pad short frame\n");
3194 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3195 		DBG_COUNTER_INC(tx_frees);
3196 		return (ENOBUFS);
3197 	}
3198 
3199 	return 0;
3200 }
3201 
3202 static int
3203 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3204 {
3205 	if_ctx_t		ctx;
3206 	if_shared_ctx_t		sctx;
3207 	if_softc_ctx_t		scctx;
3208 	bus_dma_tag_t		buf_tag;
3209 	bus_dma_segment_t	*segs;
3210 	struct mbuf		*m_head, **ifsd_m;
3211 	void			*next_txd;
3212 	bus_dmamap_t		map;
3213 	struct if_pkt_info	pi;
3214 	int remap = 0;
3215 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3216 
3217 	ctx = txq->ift_ctx;
3218 	sctx = ctx->ifc_sctx;
3219 	scctx = &ctx->ifc_softc_ctx;
3220 	segs = txq->ift_segs;
3221 	ntxd = txq->ift_size;
3222 	m_head = *m_headp;
3223 	map = NULL;
3224 
3225 	/*
3226 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3227 	 */
3228 	cidx = txq->ift_cidx;
3229 	pidx = txq->ift_pidx;
3230 	if (ctx->ifc_flags & IFC_PREFETCH) {
3231 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3232 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3233 			next_txd = calc_next_txd(txq, cidx, 0);
3234 			prefetch(next_txd);
3235 		}
3236 
3237 		/* prefetch the next cache line of mbuf pointers and flags */
3238 		prefetch(&txq->ift_sds.ifsd_m[next]);
3239 		prefetch(&txq->ift_sds.ifsd_map[next]);
3240 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3241 	}
3242 	map = txq->ift_sds.ifsd_map[pidx];
3243 	ifsd_m = txq->ift_sds.ifsd_m;
3244 
3245 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3246 		buf_tag = txq->ift_tso_buf_tag;
3247 		max_segs = scctx->isc_tx_tso_segments_max;
3248 		map = txq->ift_sds.ifsd_tso_map[pidx];
3249 		MPASS(buf_tag != NULL);
3250 		MPASS(max_segs > 0);
3251 	} else {
3252 		buf_tag = txq->ift_buf_tag;
3253 		max_segs = scctx->isc_tx_nsegments;
3254 		map = txq->ift_sds.ifsd_map[pidx];
3255 	}
3256 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3257 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3258 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3259 		if (err) {
3260 			DBG_COUNTER_INC(encap_txd_encap_fail);
3261 			return err;
3262 		}
3263 	}
3264 	m_head = *m_headp;
3265 
3266 	pkt_info_zero(&pi);
3267 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3268 	pi.ipi_pidx = pidx;
3269 	pi.ipi_qsidx = txq->ift_id;
3270 	pi.ipi_len = m_head->m_pkthdr.len;
3271 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3272 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3273 
3274 	/* deliberate bitwise OR to make one condition */
3275 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3276 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3277 			DBG_COUNTER_INC(encap_txd_encap_fail);
3278 			return (err);
3279 		}
3280 		m_head = *m_headp;
3281 	}
3282 
3283 retry:
3284 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3285 	    BUS_DMA_NOWAIT);
3286 defrag:
3287 	if (__predict_false(err)) {
3288 		switch (err) {
3289 		case EFBIG:
3290 			/* try collapse once and defrag once */
3291 			if (remap == 0) {
3292 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3293 				/* try defrag if collapsing fails */
3294 				if (m_head == NULL)
3295 					remap++;
3296 			}
3297 			if (remap == 1) {
3298 				txq->ift_mbuf_defrag++;
3299 				m_head = m_defrag(*m_headp, M_NOWAIT);
3300 			}
3301 			/*
3302 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3303 			 * failed to map an mbuf that was run through m_defrag
3304 			 */
3305 			MPASS(remap <= 1);
3306 			if (__predict_false(m_head == NULL || remap > 1))
3307 				goto defrag_failed;
3308 			remap++;
3309 			*m_headp = m_head;
3310 			goto retry;
3311 			break;
3312 		case ENOMEM:
3313 			txq->ift_no_tx_dma_setup++;
3314 			break;
3315 		default:
3316 			txq->ift_no_tx_dma_setup++;
3317 			m_freem(*m_headp);
3318 			DBG_COUNTER_INC(tx_frees);
3319 			*m_headp = NULL;
3320 			break;
3321 		}
3322 		txq->ift_map_failed++;
3323 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3324 		DBG_COUNTER_INC(encap_txd_encap_fail);
3325 		return (err);
3326 	}
3327 	ifsd_m[pidx] = m_head;
3328 	/*
3329 	 * XXX assumes a 1 to 1 relationship between segments and
3330 	 *        descriptors - this does not hold true on all drivers, e.g.
3331 	 *        cxgb
3332 	 */
3333 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3334 		txq->ift_no_desc_avail++;
3335 		bus_dmamap_unload(buf_tag, map);
3336 		DBG_COUNTER_INC(encap_txq_avail_fail);
3337 		DBG_COUNTER_INC(encap_txd_encap_fail);
3338 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3339 			GROUPTASK_ENQUEUE(&txq->ift_task);
3340 		return (ENOBUFS);
3341 	}
3342 	/*
3343 	 * On Intel cards we can greatly reduce the number of TX interrupts
3344 	 * we see by only setting report status on every Nth descriptor.
3345 	 * However, this also means that the driver will need to keep track
3346 	 * of the descriptors that RS was set on to check them for the DD bit.
3347 	 */
3348 	txq->ift_rs_pending += nsegs + 1;
3349 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3350 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3351 		pi.ipi_flags |= IPI_TX_INTR;
3352 		txq->ift_rs_pending = 0;
3353 	}
3354 
3355 	pi.ipi_segs = segs;
3356 	pi.ipi_nsegs = nsegs;
3357 
3358 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3359 #ifdef PKT_DEBUG
3360 	print_pkt(&pi);
3361 #endif
3362 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3363 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3364 		DBG_COUNTER_INC(tx_encap);
3365 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3366 
3367 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3368 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3369 			ndesc += txq->ift_size;
3370 			txq->ift_gen = 1;
3371 		}
3372 		/*
3373 		 * drivers can need as many as
3374 		 * two sentinels
3375 		 */
3376 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3377 		MPASS(pi.ipi_new_pidx != pidx);
3378 		MPASS(ndesc > 0);
3379 		txq->ift_in_use += ndesc;
3380 
3381 		/*
3382 		 * We update the last software descriptor again here because there may
3383 		 * be a sentinel and/or there may be more mbufs than segments
3384 		 */
3385 		txq->ift_pidx = pi.ipi_new_pidx;
3386 		txq->ift_npending += pi.ipi_ndescs;
3387 	} else {
3388 		*m_headp = m_head = iflib_remove_mbuf(txq);
3389 		if (err == EFBIG) {
3390 			txq->ift_txd_encap_efbig++;
3391 			if (remap < 2) {
3392 				remap = 1;
3393 				goto defrag;
3394 			}
3395 		}
3396 		goto defrag_failed;
3397 	}
3398 	/*
3399 	 * err can't possibly be non-zero here, so we don't neet to test it
3400 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3401 	 */
3402 	return (err);
3403 
3404 defrag_failed:
3405 	txq->ift_mbuf_defrag_failed++;
3406 	txq->ift_map_failed++;
3407 	m_freem(*m_headp);
3408 	DBG_COUNTER_INC(tx_frees);
3409 	*m_headp = NULL;
3410 	DBG_COUNTER_INC(encap_txd_encap_fail);
3411 	return (ENOMEM);
3412 }
3413 
3414 static void
3415 iflib_tx_desc_free(iflib_txq_t txq, int n)
3416 {
3417 	uint32_t qsize, cidx, mask, gen;
3418 	struct mbuf *m, **ifsd_m;
3419 	bool do_prefetch;
3420 
3421 	cidx = txq->ift_cidx;
3422 	gen = txq->ift_gen;
3423 	qsize = txq->ift_size;
3424 	mask = qsize-1;
3425 	ifsd_m = txq->ift_sds.ifsd_m;
3426 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3427 
3428 	while (n-- > 0) {
3429 		if (do_prefetch) {
3430 			prefetch(ifsd_m[(cidx + 3) & mask]);
3431 			prefetch(ifsd_m[(cidx + 4) & mask]);
3432 		}
3433 		if ((m = ifsd_m[cidx]) != NULL) {
3434 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3435 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3436 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3437 				    txq->ift_sds.ifsd_tso_map[cidx],
3438 				    BUS_DMASYNC_POSTWRITE);
3439 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3440 				    txq->ift_sds.ifsd_tso_map[cidx]);
3441 			} else {
3442 				bus_dmamap_sync(txq->ift_buf_tag,
3443 				    txq->ift_sds.ifsd_map[cidx],
3444 				    BUS_DMASYNC_POSTWRITE);
3445 				bus_dmamap_unload(txq->ift_buf_tag,
3446 				    txq->ift_sds.ifsd_map[cidx]);
3447 			}
3448 			/* XXX we don't support any drivers that batch packets yet */
3449 			MPASS(m->m_nextpkt == NULL);
3450 			m_freem(m);
3451 			ifsd_m[cidx] = NULL;
3452 #if MEMORY_LOGGING
3453 			txq->ift_dequeued++;
3454 #endif
3455 			DBG_COUNTER_INC(tx_frees);
3456 		}
3457 		if (__predict_false(++cidx == qsize)) {
3458 			cidx = 0;
3459 			gen = 0;
3460 		}
3461 	}
3462 	txq->ift_cidx = cidx;
3463 	txq->ift_gen = gen;
3464 }
3465 
3466 static __inline int
3467 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3468 {
3469 	int reclaim;
3470 	if_ctx_t ctx = txq->ift_ctx;
3471 
3472 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3473 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3474 
3475 	/*
3476 	 * Need a rate-limiting check so that this isn't called every time
3477 	 */
3478 	iflib_tx_credits_update(ctx, txq);
3479 	reclaim = DESC_RECLAIMABLE(txq);
3480 
3481 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3482 #ifdef INVARIANTS
3483 		if (iflib_verbose_debug) {
3484 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3485 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3486 			       reclaim, thresh);
3487 
3488 		}
3489 #endif
3490 		return (0);
3491 	}
3492 	iflib_tx_desc_free(txq, reclaim);
3493 	txq->ift_cleaned += reclaim;
3494 	txq->ift_in_use -= reclaim;
3495 
3496 	return (reclaim);
3497 }
3498 
3499 static struct mbuf **
3500 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3501 {
3502 	int next, size;
3503 	struct mbuf **items;
3504 
3505 	size = r->size;
3506 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3507 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3508 
3509 	prefetch(items[(cidx + offset) & (size-1)]);
3510 	if (remaining > 1) {
3511 		prefetch2cachelines(&items[next]);
3512 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3513 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3514 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3515 	}
3516 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3517 }
3518 
3519 static void
3520 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3521 {
3522 
3523 	ifmp_ring_check_drainage(txq->ift_br, budget);
3524 }
3525 
3526 static uint32_t
3527 iflib_txq_can_drain(struct ifmp_ring *r)
3528 {
3529 	iflib_txq_t txq = r->cookie;
3530 	if_ctx_t ctx = txq->ift_ctx;
3531 
3532 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3533 		return (1);
3534 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3535 	    BUS_DMASYNC_POSTREAD);
3536 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3537 	    false));
3538 }
3539 
3540 static uint32_t
3541 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3542 {
3543 	iflib_txq_t txq = r->cookie;
3544 	if_ctx_t ctx = txq->ift_ctx;
3545 	struct ifnet *ifp = ctx->ifc_ifp;
3546 	struct mbuf **mp, *m;
3547 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3548 	int reclaimed, err, in_use_prev, desc_used;
3549 	bool do_prefetch, ring, rang;
3550 
3551 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3552 			    !LINK_ACTIVE(ctx))) {
3553 		DBG_COUNTER_INC(txq_drain_notready);
3554 		return (0);
3555 	}
3556 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3557 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3558 	avail = IDXDIFF(pidx, cidx, r->size);
3559 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3560 		DBG_COUNTER_INC(txq_drain_flushing);
3561 		for (i = 0; i < avail; i++) {
3562 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3563 				m_free(r->items[(cidx + i) & (r->size-1)]);
3564 			r->items[(cidx + i) & (r->size-1)] = NULL;
3565 		}
3566 		return (avail);
3567 	}
3568 
3569 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3570 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3571 		CALLOUT_LOCK(txq);
3572 		callout_stop(&txq->ift_timer);
3573 		CALLOUT_UNLOCK(txq);
3574 		DBG_COUNTER_INC(txq_drain_oactive);
3575 		return (0);
3576 	}
3577 	if (reclaimed)
3578 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3579 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3580 	count = MIN(avail, TX_BATCH_SIZE);
3581 #ifdef INVARIANTS
3582 	if (iflib_verbose_debug)
3583 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3584 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3585 #endif
3586 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3587 	avail = TXQ_AVAIL(txq);
3588 	err = 0;
3589 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3590 		int rem = do_prefetch ? count - i : 0;
3591 
3592 		mp = _ring_peek_one(r, cidx, i, rem);
3593 		MPASS(mp != NULL && *mp != NULL);
3594 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3595 			consumed++;
3596 			reclaimed++;
3597 			continue;
3598 		}
3599 		in_use_prev = txq->ift_in_use;
3600 		err = iflib_encap(txq, mp);
3601 		if (__predict_false(err)) {
3602 			/* no room - bail out */
3603 			if (err == ENOBUFS)
3604 				break;
3605 			consumed++;
3606 			/* we can't send this packet - skip it */
3607 			continue;
3608 		}
3609 		consumed++;
3610 		pkt_sent++;
3611 		m = *mp;
3612 		DBG_COUNTER_INC(tx_sent);
3613 		bytes_sent += m->m_pkthdr.len;
3614 		mcast_sent += !!(m->m_flags & M_MCAST);
3615 		avail = TXQ_AVAIL(txq);
3616 
3617 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3618 		desc_used += (txq->ift_in_use - in_use_prev);
3619 		ETHER_BPF_MTAP(ifp, m);
3620 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3621 			break;
3622 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3623 	}
3624 
3625 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3626 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3627 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3628 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3629 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3630 	if (mcast_sent)
3631 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3632 #ifdef INVARIANTS
3633 	if (iflib_verbose_debug)
3634 		printf("consumed=%d\n", consumed);
3635 #endif
3636 	return (consumed);
3637 }
3638 
3639 static uint32_t
3640 iflib_txq_drain_always(struct ifmp_ring *r)
3641 {
3642 	return (1);
3643 }
3644 
3645 static uint32_t
3646 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3647 {
3648 	int i, avail;
3649 	struct mbuf **mp;
3650 	iflib_txq_t txq;
3651 
3652 	txq = r->cookie;
3653 
3654 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3655 	CALLOUT_LOCK(txq);
3656 	callout_stop(&txq->ift_timer);
3657 	CALLOUT_UNLOCK(txq);
3658 
3659 	avail = IDXDIFF(pidx, cidx, r->size);
3660 	for (i = 0; i < avail; i++) {
3661 		mp = _ring_peek_one(r, cidx, i, avail - i);
3662 		if (__predict_false(*mp == (struct mbuf *)txq))
3663 			continue;
3664 		m_freem(*mp);
3665 		DBG_COUNTER_INC(tx_frees);
3666 	}
3667 	MPASS(ifmp_ring_is_stalled(r) == 0);
3668 	return (avail);
3669 }
3670 
3671 static void
3672 iflib_ifmp_purge(iflib_txq_t txq)
3673 {
3674 	struct ifmp_ring *r;
3675 
3676 	r = txq->ift_br;
3677 	r->drain = iflib_txq_drain_free;
3678 	r->can_drain = iflib_txq_drain_always;
3679 
3680 	ifmp_ring_check_drainage(r, r->size);
3681 
3682 	r->drain = iflib_txq_drain;
3683 	r->can_drain = iflib_txq_can_drain;
3684 }
3685 
3686 static void
3687 _task_fn_tx(void *context)
3688 {
3689 	iflib_txq_t txq = context;
3690 	if_ctx_t ctx = txq->ift_ctx;
3691 #if defined(ALTQ) || defined(DEV_NETMAP)
3692 	if_t ifp = ctx->ifc_ifp;
3693 #endif
3694 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3695 
3696 #ifdef IFLIB_DIAGNOSTICS
3697 	txq->ift_cpu_exec_count[curcpu]++;
3698 #endif
3699 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3700 		return;
3701 #ifdef DEV_NETMAP
3702 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3703 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3704 		    BUS_DMASYNC_POSTREAD);
3705 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3706 			netmap_tx_irq(ifp, txq->ift_id);
3707 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3708 		return;
3709 	}
3710 #endif
3711 #ifdef ALTQ
3712 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3713 		iflib_altq_if_start(ifp);
3714 #endif
3715 	if (txq->ift_db_pending)
3716 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3717 	else if (!abdicate)
3718 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3719 	/*
3720 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3721 	 */
3722 	if (abdicate)
3723 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3724 	if (ctx->ifc_flags & IFC_LEGACY)
3725 		IFDI_INTR_ENABLE(ctx);
3726 	else {
3727 #ifdef INVARIANTS
3728 		int rc =
3729 #endif
3730 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3731 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3732 	}
3733 }
3734 
3735 static void
3736 _task_fn_rx(void *context)
3737 {
3738 	iflib_rxq_t rxq = context;
3739 	if_ctx_t ctx = rxq->ifr_ctx;
3740 	bool more;
3741 	uint16_t budget;
3742 
3743 #ifdef IFLIB_DIAGNOSTICS
3744 	rxq->ifr_cpu_exec_count[curcpu]++;
3745 #endif
3746 	DBG_COUNTER_INC(task_fn_rxs);
3747 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3748 		return;
3749 	more = true;
3750 #ifdef DEV_NETMAP
3751 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3752 		u_int work = 0;
3753 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3754 			more = false;
3755 		}
3756 	}
3757 #endif
3758 	budget = ctx->ifc_sysctl_rx_budget;
3759 	if (budget == 0)
3760 		budget = 16;	/* XXX */
3761 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3762 		if (ctx->ifc_flags & IFC_LEGACY)
3763 			IFDI_INTR_ENABLE(ctx);
3764 		else {
3765 #ifdef INVARIANTS
3766 			int rc =
3767 #endif
3768 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3769 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3770 			DBG_COUNTER_INC(rx_intr_enables);
3771 		}
3772 	}
3773 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3774 		return;
3775 	if (more)
3776 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3777 }
3778 
3779 static void
3780 _task_fn_admin(void *context)
3781 {
3782 	if_ctx_t ctx = context;
3783 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3784 	iflib_txq_t txq;
3785 	int i;
3786 	bool oactive, running, do_reset, do_watchdog, in_detach;
3787 	uint32_t reset_on = hz / 2;
3788 
3789 	STATE_LOCK(ctx);
3790 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3791 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3792 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3793 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3794 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3795 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3796 	STATE_UNLOCK(ctx);
3797 
3798 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3799 		return;
3800 	if (in_detach)
3801 		return;
3802 
3803 	CTX_LOCK(ctx);
3804 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3805 		CALLOUT_LOCK(txq);
3806 		callout_stop(&txq->ift_timer);
3807 		CALLOUT_UNLOCK(txq);
3808 	}
3809 	if (do_watchdog) {
3810 		ctx->ifc_watchdog_events++;
3811 		IFDI_WATCHDOG_RESET(ctx);
3812 	}
3813 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3814 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3815 #ifdef DEV_NETMAP
3816 		reset_on = hz / 2;
3817 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3818 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3819 #endif
3820 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3821 	}
3822 	IFDI_LINK_INTR_ENABLE(ctx);
3823 	if (do_reset)
3824 		iflib_if_init_locked(ctx);
3825 	CTX_UNLOCK(ctx);
3826 
3827 	if (LINK_ACTIVE(ctx) == 0)
3828 		return;
3829 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3830 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3831 }
3832 
3833 
3834 static void
3835 _task_fn_iov(void *context)
3836 {
3837 	if_ctx_t ctx = context;
3838 
3839 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3840 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3841 		return;
3842 
3843 	CTX_LOCK(ctx);
3844 	IFDI_VFLR_HANDLE(ctx);
3845 	CTX_UNLOCK(ctx);
3846 }
3847 
3848 static int
3849 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3850 {
3851 	int err;
3852 	if_int_delay_info_t info;
3853 	if_ctx_t ctx;
3854 
3855 	info = (if_int_delay_info_t)arg1;
3856 	ctx = info->iidi_ctx;
3857 	info->iidi_req = req;
3858 	info->iidi_oidp = oidp;
3859 	CTX_LOCK(ctx);
3860 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3861 	CTX_UNLOCK(ctx);
3862 	return (err);
3863 }
3864 
3865 /*********************************************************************
3866  *
3867  *  IFNET FUNCTIONS
3868  *
3869  **********************************************************************/
3870 
3871 static void
3872 iflib_if_init_locked(if_ctx_t ctx)
3873 {
3874 	iflib_stop(ctx);
3875 	iflib_init_locked(ctx);
3876 }
3877 
3878 
3879 static void
3880 iflib_if_init(void *arg)
3881 {
3882 	if_ctx_t ctx = arg;
3883 
3884 	CTX_LOCK(ctx);
3885 	iflib_if_init_locked(ctx);
3886 	CTX_UNLOCK(ctx);
3887 }
3888 
3889 static int
3890 iflib_if_transmit(if_t ifp, struct mbuf *m)
3891 {
3892 	if_ctx_t	ctx = if_getsoftc(ifp);
3893 
3894 	iflib_txq_t txq;
3895 	int err, qidx;
3896 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3897 
3898 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3899 		DBG_COUNTER_INC(tx_frees);
3900 		m_freem(m);
3901 		return (ENOBUFS);
3902 	}
3903 
3904 	MPASS(m->m_nextpkt == NULL);
3905 	/* ALTQ-enabled interfaces always use queue 0. */
3906 	qidx = 0;
3907 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3908 		qidx = QIDX(ctx, m);
3909 	/*
3910 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3911 	 */
3912 	txq = &ctx->ifc_txqs[qidx];
3913 
3914 #ifdef DRIVER_BACKPRESSURE
3915 	if (txq->ift_closed) {
3916 		while (m != NULL) {
3917 			next = m->m_nextpkt;
3918 			m->m_nextpkt = NULL;
3919 			m_freem(m);
3920 			DBG_COUNTER_INC(tx_frees);
3921 			m = next;
3922 		}
3923 		return (ENOBUFS);
3924 	}
3925 #endif
3926 #ifdef notyet
3927 	qidx = count = 0;
3928 	mp = marr;
3929 	next = m;
3930 	do {
3931 		count++;
3932 		next = next->m_nextpkt;
3933 	} while (next != NULL);
3934 
3935 	if (count > nitems(marr))
3936 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3937 			/* XXX check nextpkt */
3938 			m_freem(m);
3939 			/* XXX simplify for now */
3940 			DBG_COUNTER_INC(tx_frees);
3941 			return (ENOBUFS);
3942 		}
3943 	for (next = m, i = 0; next != NULL; i++) {
3944 		mp[i] = next;
3945 		next = next->m_nextpkt;
3946 		mp[i]->m_nextpkt = NULL;
3947 	}
3948 #endif
3949 	DBG_COUNTER_INC(tx_seen);
3950 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3951 
3952 	if (abdicate)
3953 		GROUPTASK_ENQUEUE(&txq->ift_task);
3954  	if (err) {
3955 		if (!abdicate)
3956 			GROUPTASK_ENQUEUE(&txq->ift_task);
3957 		/* support forthcoming later */
3958 #ifdef DRIVER_BACKPRESSURE
3959 		txq->ift_closed = TRUE;
3960 #endif
3961 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3962 		m_freem(m);
3963 		DBG_COUNTER_INC(tx_frees);
3964 	}
3965 
3966 	return (err);
3967 }
3968 
3969 #ifdef ALTQ
3970 /*
3971  * The overall approach to integrating iflib with ALTQ is to continue to use
3972  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3973  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3974  * is redundant/unnecessary, but doing so minimizes the amount of
3975  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3976  * redundantly queueing to an intermediate mp_ring is swamped by the
3977  * performance limitations inherent in using ALTQ.
3978  *
3979  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3980  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3981  * given interface.  If ALTQ is enabled for an interface, then all
3982  * transmitted packets for that interface will be submitted to the ALTQ
3983  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3984  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3985  * update stats that the iflib machinery handles, and which is sensitve to
3986  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3987  * will be installed as the start routine for use by ALTQ facilities that
3988  * need to trigger queue drains on a scheduled basis.
3989  *
3990  */
3991 static void
3992 iflib_altq_if_start(if_t ifp)
3993 {
3994 	struct ifaltq *ifq = &ifp->if_snd;
3995 	struct mbuf *m;
3996 
3997 	IFQ_LOCK(ifq);
3998 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3999 	while (m != NULL) {
4000 		iflib_if_transmit(ifp, m);
4001 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4002 	}
4003 	IFQ_UNLOCK(ifq);
4004 }
4005 
4006 static int
4007 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4008 {
4009 	int err;
4010 
4011 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4012 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4013 		if (err == 0)
4014 			iflib_altq_if_start(ifp);
4015 	} else
4016 		err = iflib_if_transmit(ifp, m);
4017 
4018 	return (err);
4019 }
4020 #endif /* ALTQ */
4021 
4022 static void
4023 iflib_if_qflush(if_t ifp)
4024 {
4025 	if_ctx_t ctx = if_getsoftc(ifp);
4026 	iflib_txq_t txq = ctx->ifc_txqs;
4027 	int i;
4028 
4029 	STATE_LOCK(ctx);
4030 	ctx->ifc_flags |= IFC_QFLUSH;
4031 	STATE_UNLOCK(ctx);
4032 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4033 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4034 			iflib_txq_check_drain(txq, 0);
4035 	STATE_LOCK(ctx);
4036 	ctx->ifc_flags &= ~IFC_QFLUSH;
4037 	STATE_UNLOCK(ctx);
4038 
4039 	/*
4040 	 * When ALTQ is enabled, this will also take care of purging the
4041 	 * ALTQ queue(s).
4042 	 */
4043 	if_qflush(ifp);
4044 }
4045 
4046 
4047 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4048 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4049 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4050 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4051 
4052 static int
4053 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4054 {
4055 	if_ctx_t ctx = if_getsoftc(ifp);
4056 	struct ifreq	*ifr = (struct ifreq *)data;
4057 #if defined(INET) || defined(INET6)
4058 	struct ifaddr	*ifa = (struct ifaddr *)data;
4059 #endif
4060 	bool		avoid_reset = FALSE;
4061 	int		err = 0, reinit = 0, bits;
4062 
4063 	switch (command) {
4064 	case SIOCSIFADDR:
4065 #ifdef INET
4066 		if (ifa->ifa_addr->sa_family == AF_INET)
4067 			avoid_reset = TRUE;
4068 #endif
4069 #ifdef INET6
4070 		if (ifa->ifa_addr->sa_family == AF_INET6)
4071 			avoid_reset = TRUE;
4072 #endif
4073 		/*
4074 		** Calling init results in link renegotiation,
4075 		** so we avoid doing it when possible.
4076 		*/
4077 		if (avoid_reset) {
4078 			if_setflagbits(ifp, IFF_UP,0);
4079 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4080 				reinit = 1;
4081 #ifdef INET
4082 			if (!(if_getflags(ifp) & IFF_NOARP))
4083 				arp_ifinit(ifp, ifa);
4084 #endif
4085 		} else
4086 			err = ether_ioctl(ifp, command, data);
4087 		break;
4088 	case SIOCSIFMTU:
4089 		CTX_LOCK(ctx);
4090 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4091 			CTX_UNLOCK(ctx);
4092 			break;
4093 		}
4094 		bits = if_getdrvflags(ifp);
4095 		/* stop the driver and free any clusters before proceeding */
4096 		iflib_stop(ctx);
4097 
4098 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4099 			STATE_LOCK(ctx);
4100 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4101 				ctx->ifc_flags |= IFC_MULTISEG;
4102 			else
4103 				ctx->ifc_flags &= ~IFC_MULTISEG;
4104 			STATE_UNLOCK(ctx);
4105 			err = if_setmtu(ifp, ifr->ifr_mtu);
4106 		}
4107 		iflib_init_locked(ctx);
4108 		STATE_LOCK(ctx);
4109 		if_setdrvflags(ifp, bits);
4110 		STATE_UNLOCK(ctx);
4111 		CTX_UNLOCK(ctx);
4112 		break;
4113 	case SIOCSIFFLAGS:
4114 		CTX_LOCK(ctx);
4115 		if (if_getflags(ifp) & IFF_UP) {
4116 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4117 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4118 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4119 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4120 				}
4121 			} else
4122 				reinit = 1;
4123 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4124 			iflib_stop(ctx);
4125 		}
4126 		ctx->ifc_if_flags = if_getflags(ifp);
4127 		CTX_UNLOCK(ctx);
4128 		break;
4129 	case SIOCADDMULTI:
4130 	case SIOCDELMULTI:
4131 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4132 			CTX_LOCK(ctx);
4133 			IFDI_INTR_DISABLE(ctx);
4134 			IFDI_MULTI_SET(ctx);
4135 			IFDI_INTR_ENABLE(ctx);
4136 			CTX_UNLOCK(ctx);
4137 		}
4138 		break;
4139 	case SIOCSIFMEDIA:
4140 		CTX_LOCK(ctx);
4141 		IFDI_MEDIA_SET(ctx);
4142 		CTX_UNLOCK(ctx);
4143 		/* falls thru */
4144 	case SIOCGIFMEDIA:
4145 	case SIOCGIFXMEDIA:
4146 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4147 		break;
4148 	case SIOCGI2C:
4149 	{
4150 		struct ifi2creq i2c;
4151 
4152 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4153 		if (err != 0)
4154 			break;
4155 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4156 			err = EINVAL;
4157 			break;
4158 		}
4159 		if (i2c.len > sizeof(i2c.data)) {
4160 			err = EINVAL;
4161 			break;
4162 		}
4163 
4164 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4165 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4166 			    sizeof(i2c));
4167 		break;
4168 	}
4169 	case SIOCSIFCAP:
4170 	{
4171 		int mask, setmask, oldmask;
4172 
4173 		oldmask = if_getcapenable(ifp);
4174 		mask = ifr->ifr_reqcap ^ oldmask;
4175 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4176 		setmask = 0;
4177 #ifdef TCP_OFFLOAD
4178 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4179 #endif
4180 		setmask |= (mask & IFCAP_FLAGS);
4181 		setmask |= (mask & IFCAP_WOL);
4182 
4183 		/*
4184 		 * If any RX csum has changed, change all the ones that
4185 		 * are supported by the driver.
4186 		 */
4187 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4188 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4189 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4190 		}
4191 
4192 		/*
4193 		 * want to ensure that traffic has stopped before we change any of the flags
4194 		 */
4195 		if (setmask) {
4196 			CTX_LOCK(ctx);
4197 			bits = if_getdrvflags(ifp);
4198 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4199 				iflib_stop(ctx);
4200 			STATE_LOCK(ctx);
4201 			if_togglecapenable(ifp, setmask);
4202 			STATE_UNLOCK(ctx);
4203 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4204 				iflib_init_locked(ctx);
4205 			STATE_LOCK(ctx);
4206 			if_setdrvflags(ifp, bits);
4207 			STATE_UNLOCK(ctx);
4208 			CTX_UNLOCK(ctx);
4209 		}
4210 		if_vlancap(ifp);
4211 		break;
4212 	}
4213 	case SIOCGPRIVATE_0:
4214 	case SIOCSDRVSPEC:
4215 	case SIOCGDRVSPEC:
4216 		CTX_LOCK(ctx);
4217 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4218 		CTX_UNLOCK(ctx);
4219 		break;
4220 	default:
4221 		err = ether_ioctl(ifp, command, data);
4222 		break;
4223 	}
4224 	if (reinit)
4225 		iflib_if_init(ctx);
4226 	return (err);
4227 }
4228 
4229 static uint64_t
4230 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4231 {
4232 	if_ctx_t ctx = if_getsoftc(ifp);
4233 
4234 	return (IFDI_GET_COUNTER(ctx, cnt));
4235 }
4236 
4237 /*********************************************************************
4238  *
4239  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4240  *
4241  **********************************************************************/
4242 
4243 static void
4244 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4245 {
4246 	if_ctx_t ctx = if_getsoftc(ifp);
4247 
4248 	if ((void *)ctx != arg)
4249 		return;
4250 
4251 	if ((vtag == 0) || (vtag > 4095))
4252 		return;
4253 
4254 	CTX_LOCK(ctx);
4255 	IFDI_VLAN_REGISTER(ctx, vtag);
4256 	/* Re-init to load the changes */
4257 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4258 		iflib_if_init_locked(ctx);
4259 	CTX_UNLOCK(ctx);
4260 }
4261 
4262 static void
4263 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4264 {
4265 	if_ctx_t ctx = if_getsoftc(ifp);
4266 
4267 	if ((void *)ctx != arg)
4268 		return;
4269 
4270 	if ((vtag == 0) || (vtag > 4095))
4271 		return;
4272 
4273 	CTX_LOCK(ctx);
4274 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4275 	/* Re-init to load the changes */
4276 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4277 		iflib_if_init_locked(ctx);
4278 	CTX_UNLOCK(ctx);
4279 }
4280 
4281 static void
4282 iflib_led_func(void *arg, int onoff)
4283 {
4284 	if_ctx_t ctx = arg;
4285 
4286 	CTX_LOCK(ctx);
4287 	IFDI_LED_FUNC(ctx, onoff);
4288 	CTX_UNLOCK(ctx);
4289 }
4290 
4291 /*********************************************************************
4292  *
4293  *  BUS FUNCTION DEFINITIONS
4294  *
4295  **********************************************************************/
4296 
4297 int
4298 iflib_device_probe(device_t dev)
4299 {
4300 	pci_vendor_info_t *ent;
4301 
4302 	uint16_t	pci_vendor_id, pci_device_id;
4303 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4304 	uint16_t	pci_rev_id;
4305 	if_shared_ctx_t sctx;
4306 
4307 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4308 		return (ENOTSUP);
4309 
4310 	pci_vendor_id = pci_get_vendor(dev);
4311 	pci_device_id = pci_get_device(dev);
4312 	pci_subvendor_id = pci_get_subvendor(dev);
4313 	pci_subdevice_id = pci_get_subdevice(dev);
4314 	pci_rev_id = pci_get_revid(dev);
4315 	if (sctx->isc_parse_devinfo != NULL)
4316 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4317 
4318 	ent = sctx->isc_vendor_info;
4319 	while (ent->pvi_vendor_id != 0) {
4320 		if (pci_vendor_id != ent->pvi_vendor_id) {
4321 			ent++;
4322 			continue;
4323 		}
4324 		if ((pci_device_id == ent->pvi_device_id) &&
4325 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4326 		     (ent->pvi_subvendor_id == 0)) &&
4327 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4328 		     (ent->pvi_subdevice_id == 0)) &&
4329 		    ((pci_rev_id == ent->pvi_rev_id) ||
4330 		     (ent->pvi_rev_id == 0))) {
4331 
4332 			device_set_desc_copy(dev, ent->pvi_name);
4333 			/* this needs to be changed to zero if the bus probing code
4334 			 * ever stops re-probing on best match because the sctx
4335 			 * may have its values over written by register calls
4336 			 * in subsequent probes
4337 			 */
4338 			return (BUS_PROBE_DEFAULT);
4339 		}
4340 		ent++;
4341 	}
4342 	return (ENXIO);
4343 }
4344 
4345 static void
4346 iflib_reset_qvalues(if_ctx_t ctx)
4347 {
4348 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4349 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4350 	device_t dev = ctx->ifc_dev;
4351 	int i;
4352 
4353 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4354 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4355 	/*
4356 	 * XXX sanity check that ntxd & nrxd are a power of 2
4357 	 */
4358 	if (ctx->ifc_sysctl_ntxqs != 0)
4359 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4360 	if (ctx->ifc_sysctl_nrxqs != 0)
4361 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4362 
4363 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4364 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4365 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4366 		else
4367 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4368 	}
4369 
4370 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4371 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4372 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4373 		else
4374 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4375 	}
4376 
4377 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4378 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4379 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4380 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4381 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4382 		}
4383 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4384 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4385 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4386 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4387 		}
4388 	}
4389 
4390 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4391 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4392 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4393 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4394 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4395 		}
4396 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4397 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4398 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4399 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4400 		}
4401 	}
4402 }
4403 
4404 int
4405 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4406 {
4407 	int err, rid, msix;
4408 	if_ctx_t ctx;
4409 	if_t ifp;
4410 	if_softc_ctx_t scctx;
4411 	int i;
4412 	uint16_t main_txq;
4413 	uint16_t main_rxq;
4414 
4415 
4416 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4417 
4418 	if (sc == NULL) {
4419 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4420 		device_set_softc(dev, ctx);
4421 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4422 	}
4423 
4424 	ctx->ifc_sctx = sctx;
4425 	ctx->ifc_dev = dev;
4426 	ctx->ifc_softc = sc;
4427 
4428 	if ((err = iflib_register(ctx)) != 0) {
4429 		device_printf(dev, "iflib_register failed %d\n", err);
4430 		goto fail_ctx_free;
4431 	}
4432 	iflib_add_device_sysctl_pre(ctx);
4433 
4434 	scctx = &ctx->ifc_softc_ctx;
4435 	ifp = ctx->ifc_ifp;
4436 
4437 	iflib_reset_qvalues(ctx);
4438 	CTX_LOCK(ctx);
4439 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4440 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4441 		goto fail_unlock;
4442 	}
4443 	_iflib_pre_assert(scctx);
4444 	ctx->ifc_txrx = *scctx->isc_txrx;
4445 
4446 #ifdef INVARIANTS
4447 	MPASS(scctx->isc_capabilities);
4448 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4449 		MPASS(scctx->isc_tx_csum_flags);
4450 #endif
4451 
4452 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4453 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4454 
4455 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4456 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4457 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4458 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4459 
4460 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4461 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4462 
4463 	/* XXX change for per-queue sizes */
4464 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4465 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4466 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4467 		if (!powerof2(scctx->isc_nrxd[i])) {
4468 			/* round down instead? */
4469 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4470 			err = EINVAL;
4471 			goto fail_iflib_detach;
4472 		}
4473 	}
4474 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4475 		if (!powerof2(scctx->isc_ntxd[i])) {
4476 			device_printf(dev,
4477 			    "# tx descriptors must be a power of 2");
4478 			err = EINVAL;
4479 			goto fail_iflib_detach;
4480 		}
4481 	}
4482 
4483 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4484 	    MAX_SINGLE_PACKET_FRACTION)
4485 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4486 		    MAX_SINGLE_PACKET_FRACTION);
4487 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4488 	    MAX_SINGLE_PACKET_FRACTION)
4489 		scctx->isc_tx_tso_segments_max = max(1,
4490 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4491 
4492 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4493 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4494 		/*
4495 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4496 		 * but some MACs do.
4497 		 */
4498 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4499 		    IP_MAXPACKET));
4500 		/*
4501 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4502 		 * into account.  In the worst case, each of these calls will
4503 		 * add another mbuf and, thus, the requirement for another DMA
4504 		 * segment.  So for best performance, it doesn't make sense to
4505 		 * advertize a maximum of TSO segments that typically will
4506 		 * require defragmentation in iflib_encap().
4507 		 */
4508 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4509 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4510 	}
4511 	if (scctx->isc_rss_table_size == 0)
4512 		scctx->isc_rss_table_size = 64;
4513 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4514 
4515 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4516 	/* XXX format name */
4517 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4518 	    NULL, NULL, "admin");
4519 
4520 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4521 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4522 		device_printf(dev, "Unable to fetch CPU list\n");
4523 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4524 	}
4525 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4526 
4527 	/*
4528 	** Now set up MSI or MSI-X, should return us the number of supported
4529 	** vectors (will be 1 for a legacy interrupt and MSI).
4530 	*/
4531 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4532 		msix = scctx->isc_vectors;
4533 	} else if (scctx->isc_msix_bar != 0)
4534 	       /*
4535 		* The simple fact that isc_msix_bar is not 0 does not mean we
4536 		* we have a good value there that is known to work.
4537 		*/
4538 		msix = iflib_msix_init(ctx);
4539 	else {
4540 		scctx->isc_vectors = 1;
4541 		scctx->isc_ntxqsets = 1;
4542 		scctx->isc_nrxqsets = 1;
4543 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4544 		msix = 0;
4545 	}
4546 	/* Get memory for the station queues */
4547 	if ((err = iflib_queues_alloc(ctx))) {
4548 		device_printf(dev, "Unable to allocate queue memory\n");
4549 		goto fail_intr_free;
4550 	}
4551 
4552 	if ((err = iflib_qset_structures_setup(ctx)))
4553 		goto fail_queues;
4554 
4555 	/*
4556 	 * Group taskqueues aren't properly set up until SMP is started,
4557 	 * so we disable interrupts until we can handle them post
4558 	 * SI_SUB_SMP.
4559 	 *
4560 	 * XXX: disabling interrupts doesn't actually work, at least for
4561 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4562 	 * we do null handling and depend on this not causing too large an
4563 	 * interrupt storm.
4564 	 */
4565 	IFDI_INTR_DISABLE(ctx);
4566 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4567 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4568 		goto fail_queues;
4569 	}
4570 	if (msix <= 1) {
4571 		rid = 0;
4572 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4573 			MPASS(msix == 1);
4574 			rid = 1;
4575 		}
4576 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4577 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4578 			goto fail_queues;
4579 		}
4580 	}
4581 
4582 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4583 
4584 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4585 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4586 		goto fail_detach;
4587 	}
4588 
4589 	/*
4590 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4591 	 * This must appear after the call to ether_ifattach() because
4592 	 * ether_ifattach() sets if_hdrlen to the default value.
4593 	 */
4594 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4595 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4596 
4597 	if ((err = iflib_netmap_attach(ctx))) {
4598 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4599 		goto fail_detach;
4600 	}
4601 	*ctxp = ctx;
4602 
4603 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4604 
4605 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4606 	iflib_add_device_sysctl_post(ctx);
4607 	ctx->ifc_flags |= IFC_INIT_DONE;
4608 	CTX_UNLOCK(ctx);
4609 	return (0);
4610 
4611 fail_detach:
4612 	ether_ifdetach(ctx->ifc_ifp);
4613 fail_intr_free:
4614 	iflib_free_intr_mem(ctx);
4615 fail_queues:
4616 	iflib_tx_structures_free(ctx);
4617 	iflib_rx_structures_free(ctx);
4618 fail_iflib_detach:
4619 	IFDI_DETACH(ctx);
4620 fail_unlock:
4621 	CTX_UNLOCK(ctx);
4622 fail_ctx_free:
4623         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4624                 free(ctx->ifc_softc, M_IFLIB);
4625         free(ctx, M_IFLIB);
4626 	return (err);
4627 }
4628 
4629 int
4630 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4631 					  struct iflib_cloneattach_ctx *clctx)
4632 {
4633 	int err;
4634 	if_ctx_t ctx;
4635 	if_t ifp;
4636 	if_softc_ctx_t scctx;
4637 	int i;
4638 	void *sc;
4639 	uint16_t main_txq;
4640 	uint16_t main_rxq;
4641 
4642 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4643 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4644 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4645 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4646 		ctx->ifc_flags |= IFC_PSEUDO;
4647 
4648 	ctx->ifc_sctx = sctx;
4649 	ctx->ifc_softc = sc;
4650 	ctx->ifc_dev = dev;
4651 
4652 	if ((err = iflib_register(ctx)) != 0) {
4653 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4654 		goto fail_ctx_free;
4655 	}
4656 	iflib_add_device_sysctl_pre(ctx);
4657 
4658 	scctx = &ctx->ifc_softc_ctx;
4659 	ifp = ctx->ifc_ifp;
4660 
4661 	/*
4662 	 * XXX sanity check that ntxd & nrxd are a power of 2
4663 	 */
4664 	iflib_reset_qvalues(ctx);
4665 
4666 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4667 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4668 		goto fail_ctx_free;
4669 	}
4670 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4671 		iflib_gen_mac(ctx);
4672 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4673 								clctx->cc_params)) != 0) {
4674 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4675 		goto fail_ctx_free;
4676 	}
4677 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4678 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4679 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4680 
4681 #ifdef INVARIANTS
4682 	MPASS(scctx->isc_capabilities);
4683 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4684 		MPASS(scctx->isc_tx_csum_flags);
4685 #endif
4686 
4687 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4688 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4689 
4690 	ifp->if_flags |= IFF_NOGROUP;
4691 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4692 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4693 
4694 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4695 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4696 			goto fail_detach;
4697 		}
4698 		*ctxp = ctx;
4699 
4700 		/*
4701 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4702 		 * This must appear after the call to ether_ifattach() because
4703 		 * ether_ifattach() sets if_hdrlen to the default value.
4704 		 */
4705 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4706 			if_setifheaderlen(ifp,
4707 			    sizeof(struct ether_vlan_header));
4708 
4709 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4710 		iflib_add_device_sysctl_post(ctx);
4711 		ctx->ifc_flags |= IFC_INIT_DONE;
4712 		return (0);
4713 	}
4714 	_iflib_pre_assert(scctx);
4715 	ctx->ifc_txrx = *scctx->isc_txrx;
4716 
4717 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4718 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4719 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4720 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4721 
4722 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4723 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4724 
4725 	/* XXX change for per-queue sizes */
4726 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4727 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4728 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4729 		if (!powerof2(scctx->isc_nrxd[i])) {
4730 			/* round down instead? */
4731 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4732 			err = EINVAL;
4733 			goto fail_iflib_detach;
4734 		}
4735 	}
4736 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4737 		if (!powerof2(scctx->isc_ntxd[i])) {
4738 			device_printf(dev,
4739 			    "# tx descriptors must be a power of 2");
4740 			err = EINVAL;
4741 			goto fail_iflib_detach;
4742 		}
4743 	}
4744 
4745 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4746 	    MAX_SINGLE_PACKET_FRACTION)
4747 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4748 		    MAX_SINGLE_PACKET_FRACTION);
4749 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4750 	    MAX_SINGLE_PACKET_FRACTION)
4751 		scctx->isc_tx_tso_segments_max = max(1,
4752 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4753 
4754 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4755 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4756 		/*
4757 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4758 		 * but some MACs do.
4759 		 */
4760 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4761 		    IP_MAXPACKET));
4762 		/*
4763 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4764 		 * into account.  In the worst case, each of these calls will
4765 		 * add another mbuf and, thus, the requirement for another DMA
4766 		 * segment.  So for best performance, it doesn't make sense to
4767 		 * advertize a maximum of TSO segments that typically will
4768 		 * require defragmentation in iflib_encap().
4769 		 */
4770 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4771 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4772 	}
4773 	if (scctx->isc_rss_table_size == 0)
4774 		scctx->isc_rss_table_size = 64;
4775 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4776 
4777 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4778 	/* XXX format name */
4779 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4780 	    NULL, NULL, "admin");
4781 
4782 	/* XXX --- can support > 1 -- but keep it simple for now */
4783 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4784 
4785 	/* Get memory for the station queues */
4786 	if ((err = iflib_queues_alloc(ctx))) {
4787 		device_printf(dev, "Unable to allocate queue memory\n");
4788 		goto fail_iflib_detach;
4789 	}
4790 
4791 	if ((err = iflib_qset_structures_setup(ctx))) {
4792 		device_printf(dev, "qset structure setup failed %d\n", err);
4793 		goto fail_queues;
4794 	}
4795 
4796 	/*
4797 	 * XXX What if anything do we want to do about interrupts?
4798 	 */
4799 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4800 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4801 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4802 		goto fail_detach;
4803 	}
4804 
4805 	/*
4806 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4807 	 * This must appear after the call to ether_ifattach() because
4808 	 * ether_ifattach() sets if_hdrlen to the default value.
4809 	 */
4810 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4811 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4812 
4813 	/* XXX handle more than one queue */
4814 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4815 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4816 
4817 	*ctxp = ctx;
4818 
4819 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4820 	iflib_add_device_sysctl_post(ctx);
4821 	ctx->ifc_flags |= IFC_INIT_DONE;
4822 	return (0);
4823 fail_detach:
4824 	ether_ifdetach(ctx->ifc_ifp);
4825 fail_queues:
4826 	iflib_tx_structures_free(ctx);
4827 	iflib_rx_structures_free(ctx);
4828 fail_iflib_detach:
4829 	IFDI_DETACH(ctx);
4830 fail_ctx_free:
4831 	free(ctx->ifc_softc, M_IFLIB);
4832 	free(ctx, M_IFLIB);
4833 	return (err);
4834 }
4835 
4836 int
4837 iflib_pseudo_deregister(if_ctx_t ctx)
4838 {
4839 	if_t ifp = ctx->ifc_ifp;
4840 	iflib_txq_t txq;
4841 	iflib_rxq_t rxq;
4842 	int i, j;
4843 	struct taskqgroup *tqg;
4844 	iflib_fl_t fl;
4845 
4846 	/* Unregister VLAN events */
4847 	if (ctx->ifc_vlan_attach_event != NULL)
4848 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4849 	if (ctx->ifc_vlan_detach_event != NULL)
4850 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4851 
4852 	ether_ifdetach(ifp);
4853 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4854 	CTX_LOCK_DESTROY(ctx);
4855 	/* XXX drain any dependent tasks */
4856 	tqg = qgroup_if_io_tqg;
4857 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4858 		callout_drain(&txq->ift_timer);
4859 		if (txq->ift_task.gt_uniq != NULL)
4860 			taskqgroup_detach(tqg, &txq->ift_task);
4861 	}
4862 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4863 		if (rxq->ifr_task.gt_uniq != NULL)
4864 			taskqgroup_detach(tqg, &rxq->ifr_task);
4865 
4866 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4867 			free(fl->ifl_rx_bitmap, M_IFLIB);
4868 	}
4869 	tqg = qgroup_if_config_tqg;
4870 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4871 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4872 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4873 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4874 
4875 	if_free(ifp);
4876 
4877 	iflib_tx_structures_free(ctx);
4878 	iflib_rx_structures_free(ctx);
4879 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4880 		free(ctx->ifc_softc, M_IFLIB);
4881 	free(ctx, M_IFLIB);
4882 	return (0);
4883 }
4884 
4885 int
4886 iflib_device_attach(device_t dev)
4887 {
4888 	if_ctx_t ctx;
4889 	if_shared_ctx_t sctx;
4890 
4891 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4892 		return (ENOTSUP);
4893 
4894 	pci_enable_busmaster(dev);
4895 
4896 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4897 }
4898 
4899 int
4900 iflib_device_deregister(if_ctx_t ctx)
4901 {
4902 	if_t ifp = ctx->ifc_ifp;
4903 	iflib_txq_t txq;
4904 	iflib_rxq_t rxq;
4905 	device_t dev = ctx->ifc_dev;
4906 	int i, j;
4907 	struct taskqgroup *tqg;
4908 	iflib_fl_t fl;
4909 
4910 	/* Make sure VLANS are not using driver */
4911 	if (if_vlantrunkinuse(ifp)) {
4912 		device_printf(dev, "Vlan in use, detach first\n");
4913 		return (EBUSY);
4914 	}
4915 #ifdef PCI_IOV
4916 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4917 		device_printf(dev, "SR-IOV in use; detach first.\n");
4918 		return (EBUSY);
4919 	}
4920 #endif
4921 
4922 	STATE_LOCK(ctx);
4923 	ctx->ifc_flags |= IFC_IN_DETACH;
4924 	STATE_UNLOCK(ctx);
4925 
4926 	CTX_LOCK(ctx);
4927 	iflib_stop(ctx);
4928 	CTX_UNLOCK(ctx);
4929 
4930 	/* Unregister VLAN events */
4931 	if (ctx->ifc_vlan_attach_event != NULL)
4932 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4933 	if (ctx->ifc_vlan_detach_event != NULL)
4934 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4935 
4936 	iflib_netmap_detach(ifp);
4937 	ether_ifdetach(ifp);
4938 	if (ctx->ifc_led_dev != NULL)
4939 		led_destroy(ctx->ifc_led_dev);
4940 	/* XXX drain any dependent tasks */
4941 	tqg = qgroup_if_io_tqg;
4942 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4943 		callout_drain(&txq->ift_timer);
4944 		if (txq->ift_task.gt_uniq != NULL)
4945 			taskqgroup_detach(tqg, &txq->ift_task);
4946 	}
4947 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4948 		if (rxq->ifr_task.gt_uniq != NULL)
4949 			taskqgroup_detach(tqg, &rxq->ifr_task);
4950 
4951 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4952 			free(fl->ifl_rx_bitmap, M_IFLIB);
4953 	}
4954 	tqg = qgroup_if_config_tqg;
4955 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4956 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4957 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4958 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4959 	CTX_LOCK(ctx);
4960 	IFDI_DETACH(ctx);
4961 	CTX_UNLOCK(ctx);
4962 
4963 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4964 	CTX_LOCK_DESTROY(ctx);
4965 	device_set_softc(ctx->ifc_dev, NULL);
4966 	iflib_free_intr_mem(ctx);
4967 
4968 	bus_generic_detach(dev);
4969 	if_free(ifp);
4970 
4971 	iflib_tx_structures_free(ctx);
4972 	iflib_rx_structures_free(ctx);
4973 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4974 		free(ctx->ifc_softc, M_IFLIB);
4975 	STATE_LOCK_DESTROY(ctx);
4976 	free(ctx, M_IFLIB);
4977 	return (0);
4978 }
4979 
4980 static void
4981 iflib_free_intr_mem(if_ctx_t ctx)
4982 {
4983 
4984 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4985 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4986 	}
4987 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4988 		pci_release_msi(ctx->ifc_dev);
4989 	}
4990 	if (ctx->ifc_msix_mem != NULL) {
4991 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4992 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
4993 		ctx->ifc_msix_mem = NULL;
4994 	}
4995 }
4996 
4997 int
4998 iflib_device_detach(device_t dev)
4999 {
5000 	if_ctx_t ctx = device_get_softc(dev);
5001 
5002 	return (iflib_device_deregister(ctx));
5003 }
5004 
5005 int
5006 iflib_device_suspend(device_t dev)
5007 {
5008 	if_ctx_t ctx = device_get_softc(dev);
5009 
5010 	CTX_LOCK(ctx);
5011 	IFDI_SUSPEND(ctx);
5012 	CTX_UNLOCK(ctx);
5013 
5014 	return bus_generic_suspend(dev);
5015 }
5016 int
5017 iflib_device_shutdown(device_t dev)
5018 {
5019 	if_ctx_t ctx = device_get_softc(dev);
5020 
5021 	CTX_LOCK(ctx);
5022 	IFDI_SHUTDOWN(ctx);
5023 	CTX_UNLOCK(ctx);
5024 
5025 	return bus_generic_suspend(dev);
5026 }
5027 
5028 
5029 int
5030 iflib_device_resume(device_t dev)
5031 {
5032 	if_ctx_t ctx = device_get_softc(dev);
5033 	iflib_txq_t txq = ctx->ifc_txqs;
5034 
5035 	CTX_LOCK(ctx);
5036 	IFDI_RESUME(ctx);
5037 	iflib_if_init_locked(ctx);
5038 	CTX_UNLOCK(ctx);
5039 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5040 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5041 
5042 	return (bus_generic_resume(dev));
5043 }
5044 
5045 int
5046 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5047 {
5048 	int error;
5049 	if_ctx_t ctx = device_get_softc(dev);
5050 
5051 	CTX_LOCK(ctx);
5052 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5053 	CTX_UNLOCK(ctx);
5054 
5055 	return (error);
5056 }
5057 
5058 void
5059 iflib_device_iov_uninit(device_t dev)
5060 {
5061 	if_ctx_t ctx = device_get_softc(dev);
5062 
5063 	CTX_LOCK(ctx);
5064 	IFDI_IOV_UNINIT(ctx);
5065 	CTX_UNLOCK(ctx);
5066 }
5067 
5068 int
5069 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5070 {
5071 	int error;
5072 	if_ctx_t ctx = device_get_softc(dev);
5073 
5074 	CTX_LOCK(ctx);
5075 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5076 	CTX_UNLOCK(ctx);
5077 
5078 	return (error);
5079 }
5080 
5081 /*********************************************************************
5082  *
5083  *  MODULE FUNCTION DEFINITIONS
5084  *
5085  **********************************************************************/
5086 
5087 /*
5088  * - Start a fast taskqueue thread for each core
5089  * - Start a taskqueue for control operations
5090  */
5091 static int
5092 iflib_module_init(void)
5093 {
5094 	return (0);
5095 }
5096 
5097 static int
5098 iflib_module_event_handler(module_t mod, int what, void *arg)
5099 {
5100 	int err;
5101 
5102 	switch (what) {
5103 	case MOD_LOAD:
5104 		if ((err = iflib_module_init()) != 0)
5105 			return (err);
5106 		break;
5107 	case MOD_UNLOAD:
5108 		return (EBUSY);
5109 	default:
5110 		return (EOPNOTSUPP);
5111 	}
5112 
5113 	return (0);
5114 }
5115 
5116 /*********************************************************************
5117  *
5118  *  PUBLIC FUNCTION DEFINITIONS
5119  *     ordered as in iflib.h
5120  *
5121  **********************************************************************/
5122 
5123 
5124 static void
5125 _iflib_assert(if_shared_ctx_t sctx)
5126 {
5127 	MPASS(sctx->isc_tx_maxsize);
5128 	MPASS(sctx->isc_tx_maxsegsize);
5129 
5130 	MPASS(sctx->isc_rx_maxsize);
5131 	MPASS(sctx->isc_rx_nsegments);
5132 	MPASS(sctx->isc_rx_maxsegsize);
5133 
5134 	MPASS(sctx->isc_nrxd_min[0]);
5135 	MPASS(sctx->isc_nrxd_max[0]);
5136 	MPASS(sctx->isc_nrxd_default[0]);
5137 	MPASS(sctx->isc_ntxd_min[0]);
5138 	MPASS(sctx->isc_ntxd_max[0]);
5139 	MPASS(sctx->isc_ntxd_default[0]);
5140 }
5141 
5142 static void
5143 _iflib_pre_assert(if_softc_ctx_t scctx)
5144 {
5145 
5146 	MPASS(scctx->isc_txrx->ift_txd_encap);
5147 	MPASS(scctx->isc_txrx->ift_txd_flush);
5148 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5149 	MPASS(scctx->isc_txrx->ift_rxd_available);
5150 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5151 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5152 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5153 }
5154 
5155 static int
5156 iflib_register(if_ctx_t ctx)
5157 {
5158 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5159 	driver_t *driver = sctx->isc_driver;
5160 	device_t dev = ctx->ifc_dev;
5161 	if_t ifp;
5162 
5163 	_iflib_assert(sctx);
5164 
5165 	CTX_LOCK_INIT(ctx);
5166 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5167 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5168 	if (ifp == NULL) {
5169 		device_printf(dev, "can not allocate ifnet structure\n");
5170 		return (ENOMEM);
5171 	}
5172 
5173 	/*
5174 	 * Initialize our context's device specific methods
5175 	 */
5176 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5177 	kobj_class_compile((kobj_class_t) driver);
5178 	driver->refs++;
5179 
5180 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5181 	if_setsoftc(ifp, ctx);
5182 	if_setdev(ifp, dev);
5183 	if_setinitfn(ifp, iflib_if_init);
5184 	if_setioctlfn(ifp, iflib_if_ioctl);
5185 #ifdef ALTQ
5186 	if_setstartfn(ifp, iflib_altq_if_start);
5187 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5188 	if_setsendqready(ifp);
5189 #else
5190 	if_settransmitfn(ifp, iflib_if_transmit);
5191 #endif
5192 	if_setqflushfn(ifp, iflib_if_qflush);
5193 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5194 
5195 	ctx->ifc_vlan_attach_event =
5196 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5197 							  EVENTHANDLER_PRI_FIRST);
5198 	ctx->ifc_vlan_detach_event =
5199 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5200 							  EVENTHANDLER_PRI_FIRST);
5201 
5202 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5203 					 iflib_media_change, iflib_media_status);
5204 
5205 	return (0);
5206 }
5207 
5208 
5209 static int
5210 iflib_queues_alloc(if_ctx_t ctx)
5211 {
5212 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5213 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5214 	device_t dev = ctx->ifc_dev;
5215 	int nrxqsets = scctx->isc_nrxqsets;
5216 	int ntxqsets = scctx->isc_ntxqsets;
5217 	iflib_txq_t txq;
5218 	iflib_rxq_t rxq;
5219 	iflib_fl_t fl = NULL;
5220 	int i, j, cpu, err, txconf, rxconf;
5221 	iflib_dma_info_t ifdip;
5222 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5223 	uint32_t *txqsizes = scctx->isc_txqsizes;
5224 	uint8_t nrxqs = sctx->isc_nrxqs;
5225 	uint8_t ntxqs = sctx->isc_ntxqs;
5226 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5227 	caddr_t *vaddrs;
5228 	uint64_t *paddrs;
5229 
5230 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5231 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5232 
5233 	/* Allocate the TX ring struct memory */
5234 	if (!(ctx->ifc_txqs =
5235 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5236 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5237 		device_printf(dev, "Unable to allocate TX ring memory\n");
5238 		err = ENOMEM;
5239 		goto fail;
5240 	}
5241 
5242 	/* Now allocate the RX */
5243 	if (!(ctx->ifc_rxqs =
5244 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5245 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5246 		device_printf(dev, "Unable to allocate RX ring memory\n");
5247 		err = ENOMEM;
5248 		goto rx_fail;
5249 	}
5250 
5251 	txq = ctx->ifc_txqs;
5252 	rxq = ctx->ifc_rxqs;
5253 
5254 	/*
5255 	 * XXX handle allocation failure
5256 	 */
5257 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5258 		/* Set up some basics */
5259 
5260 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5261 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5262 			device_printf(dev,
5263 			    "Unable to allocate TX DMA info memory\n");
5264 			err = ENOMEM;
5265 			goto err_tx_desc;
5266 		}
5267 		txq->ift_ifdi = ifdip;
5268 		for (j = 0; j < ntxqs; j++, ifdip++) {
5269 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5270 				device_printf(dev,
5271 				    "Unable to allocate TX descriptors\n");
5272 				err = ENOMEM;
5273 				goto err_tx_desc;
5274 			}
5275 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5276 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5277 		}
5278 		txq->ift_ctx = ctx;
5279 		txq->ift_id = i;
5280 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5281 			txq->ift_br_offset = 1;
5282 		} else {
5283 			txq->ift_br_offset = 0;
5284 		}
5285 		/* XXX fix this */
5286 		txq->ift_timer.c_cpu = cpu;
5287 
5288 		if (iflib_txsd_alloc(txq)) {
5289 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5290 			err = ENOMEM;
5291 			goto err_tx_desc;
5292 		}
5293 
5294 		/* Initialize the TX lock */
5295 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5296 		    device_get_nameunit(dev), txq->ift_id);
5297 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5298 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5299 
5300 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5301 			 device_get_nameunit(dev), txq->ift_id);
5302 
5303 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5304 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5305 		if (err) {
5306 			/* XXX free any allocated rings */
5307 			device_printf(dev, "Unable to allocate buf_ring\n");
5308 			goto err_tx_desc;
5309 		}
5310 	}
5311 
5312 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5313 		/* Set up some basics */
5314 
5315 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5316 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5317 			device_printf(dev,
5318 			    "Unable to allocate RX DMA info memory\n");
5319 			err = ENOMEM;
5320 			goto err_tx_desc;
5321 		}
5322 
5323 		rxq->ifr_ifdi = ifdip;
5324 		/* XXX this needs to be changed if #rx queues != #tx queues */
5325 		rxq->ifr_ntxqirq = 1;
5326 		rxq->ifr_txqid[0] = i;
5327 		for (j = 0; j < nrxqs; j++, ifdip++) {
5328 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5329 				device_printf(dev,
5330 				    "Unable to allocate RX descriptors\n");
5331 				err = ENOMEM;
5332 				goto err_tx_desc;
5333 			}
5334 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5335 		}
5336 		rxq->ifr_ctx = ctx;
5337 		rxq->ifr_id = i;
5338 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5339 			rxq->ifr_fl_offset = 1;
5340 		} else {
5341 			rxq->ifr_fl_offset = 0;
5342 		}
5343 		rxq->ifr_nfl = nfree_lists;
5344 		if (!(fl =
5345 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5346 			device_printf(dev, "Unable to allocate free list memory\n");
5347 			err = ENOMEM;
5348 			goto err_tx_desc;
5349 		}
5350 		rxq->ifr_fl = fl;
5351 		for (j = 0; j < nfree_lists; j++) {
5352 			fl[j].ifl_rxq = rxq;
5353 			fl[j].ifl_id = j;
5354 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5355 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5356 		}
5357 		/* Allocate receive buffers for the ring */
5358 		if (iflib_rxsd_alloc(rxq)) {
5359 			device_printf(dev,
5360 			    "Critical Failure setting up receive buffers\n");
5361 			err = ENOMEM;
5362 			goto err_rx_desc;
5363 		}
5364 
5365 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5366 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5367 			    M_WAITOK);
5368 	}
5369 
5370 	/* TXQs */
5371 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5372 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5373 	for (i = 0; i < ntxqsets; i++) {
5374 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5375 
5376 		for (j = 0; j < ntxqs; j++, di++) {
5377 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5378 			paddrs[i*ntxqs + j] = di->idi_paddr;
5379 		}
5380 	}
5381 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5382 		device_printf(ctx->ifc_dev,
5383 		    "Unable to allocate device TX queue\n");
5384 		iflib_tx_structures_free(ctx);
5385 		free(vaddrs, M_IFLIB);
5386 		free(paddrs, M_IFLIB);
5387 		goto err_rx_desc;
5388 	}
5389 	free(vaddrs, M_IFLIB);
5390 	free(paddrs, M_IFLIB);
5391 
5392 	/* RXQs */
5393 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5394 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5395 	for (i = 0; i < nrxqsets; i++) {
5396 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5397 
5398 		for (j = 0; j < nrxqs; j++, di++) {
5399 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5400 			paddrs[i*nrxqs + j] = di->idi_paddr;
5401 		}
5402 	}
5403 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5404 		device_printf(ctx->ifc_dev,
5405 		    "Unable to allocate device RX queue\n");
5406 		iflib_tx_structures_free(ctx);
5407 		free(vaddrs, M_IFLIB);
5408 		free(paddrs, M_IFLIB);
5409 		goto err_rx_desc;
5410 	}
5411 	free(vaddrs, M_IFLIB);
5412 	free(paddrs, M_IFLIB);
5413 
5414 	return (0);
5415 
5416 /* XXX handle allocation failure changes */
5417 err_rx_desc:
5418 err_tx_desc:
5419 rx_fail:
5420 	if (ctx->ifc_rxqs != NULL)
5421 		free(ctx->ifc_rxqs, M_IFLIB);
5422 	ctx->ifc_rxqs = NULL;
5423 	if (ctx->ifc_txqs != NULL)
5424 		free(ctx->ifc_txqs, M_IFLIB);
5425 	ctx->ifc_txqs = NULL;
5426 fail:
5427 	return (err);
5428 }
5429 
5430 static int
5431 iflib_tx_structures_setup(if_ctx_t ctx)
5432 {
5433 	iflib_txq_t txq = ctx->ifc_txqs;
5434 	int i;
5435 
5436 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5437 		iflib_txq_setup(txq);
5438 
5439 	return (0);
5440 }
5441 
5442 static void
5443 iflib_tx_structures_free(if_ctx_t ctx)
5444 {
5445 	iflib_txq_t txq = ctx->ifc_txqs;
5446 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5447 	int i, j;
5448 
5449 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5450 		iflib_txq_destroy(txq);
5451 		for (j = 0; j < sctx->isc_ntxqs; j++)
5452 			iflib_dma_free(&txq->ift_ifdi[j]);
5453 	}
5454 	free(ctx->ifc_txqs, M_IFLIB);
5455 	ctx->ifc_txqs = NULL;
5456 	IFDI_QUEUES_FREE(ctx);
5457 }
5458 
5459 /*********************************************************************
5460  *
5461  *  Initialize all receive rings.
5462  *
5463  **********************************************************************/
5464 static int
5465 iflib_rx_structures_setup(if_ctx_t ctx)
5466 {
5467 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5468 	int q;
5469 #if defined(INET6) || defined(INET)
5470 	int i, err;
5471 #endif
5472 
5473 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5474 #if defined(INET6) || defined(INET)
5475 		tcp_lro_free(&rxq->ifr_lc);
5476 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5477 		    TCP_LRO_ENTRIES, min(1024,
5478 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5479 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5480 			goto fail;
5481 		}
5482 		rxq->ifr_lro_enabled = TRUE;
5483 #endif
5484 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5485 	}
5486 	return (0);
5487 #if defined(INET6) || defined(INET)
5488 fail:
5489 	/*
5490 	 * Free RX software descriptors allocated so far, we will only handle
5491 	 * the rings that completed, the failing case will have
5492 	 * cleaned up for itself. 'q' failed, so its the terminus.
5493 	 */
5494 	rxq = ctx->ifc_rxqs;
5495 	for (i = 0; i < q; ++i, rxq++) {
5496 		iflib_rx_sds_free(rxq);
5497 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5498 	}
5499 	return (err);
5500 #endif
5501 }
5502 
5503 /*********************************************************************
5504  *
5505  *  Free all receive rings.
5506  *
5507  **********************************************************************/
5508 static void
5509 iflib_rx_structures_free(if_ctx_t ctx)
5510 {
5511 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5512 
5513 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5514 		iflib_rx_sds_free(rxq);
5515 	}
5516 	free(ctx->ifc_rxqs, M_IFLIB);
5517 	ctx->ifc_rxqs = NULL;
5518 }
5519 
5520 static int
5521 iflib_qset_structures_setup(if_ctx_t ctx)
5522 {
5523 	int err;
5524 
5525 	/*
5526 	 * It is expected that the caller takes care of freeing queues if this
5527 	 * fails.
5528 	 */
5529 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5530 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5531 		return (err);
5532 	}
5533 
5534 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5535 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5536 
5537 	return (err);
5538 }
5539 
5540 int
5541 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5542 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5543 {
5544 
5545 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5546 }
5547 
5548 #ifdef SMP
5549 static int
5550 find_nth(if_ctx_t ctx, int qid)
5551 {
5552 	cpuset_t cpus;
5553 	int i, cpuid, eqid, count;
5554 
5555 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5556 	count = CPU_COUNT(&cpus);
5557 	eqid = qid % count;
5558 	/* clear up to the qid'th bit */
5559 	for (i = 0; i < eqid; i++) {
5560 		cpuid = CPU_FFS(&cpus);
5561 		MPASS(cpuid != 0);
5562 		CPU_CLR(cpuid-1, &cpus);
5563 	}
5564 	cpuid = CPU_FFS(&cpus);
5565 	MPASS(cpuid != 0);
5566 	return (cpuid-1);
5567 }
5568 
5569 #ifdef SCHED_ULE
5570 extern struct cpu_group *cpu_top;              /* CPU topology */
5571 
5572 static int
5573 find_child_with_core(int cpu, struct cpu_group *grp)
5574 {
5575 	int i;
5576 
5577 	if (grp->cg_children == 0)
5578 		return -1;
5579 
5580 	MPASS(grp->cg_child);
5581 	for (i = 0; i < grp->cg_children; i++) {
5582 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5583 			return i;
5584 	}
5585 
5586 	return -1;
5587 }
5588 
5589 /*
5590  * Find the nth "close" core to the specified core
5591  * "close" is defined as the deepest level that shares
5592  * at least an L2 cache.  With threads, this will be
5593  * threads on the same core.  If the sahred cache is L3
5594  * or higher, simply returns the same core.
5595  */
5596 static int
5597 find_close_core(int cpu, int core_offset)
5598 {
5599 	struct cpu_group *grp;
5600 	int i;
5601 	int fcpu;
5602 	cpuset_t cs;
5603 
5604 	grp = cpu_top;
5605 	if (grp == NULL)
5606 		return cpu;
5607 	i = 0;
5608 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5609 		/* If the child only has one cpu, don't descend */
5610 		if (grp->cg_child[i].cg_count <= 1)
5611 			break;
5612 		grp = &grp->cg_child[i];
5613 	}
5614 
5615 	/* If they don't share at least an L2 cache, use the same CPU */
5616 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5617 		return cpu;
5618 
5619 	/* Now pick one */
5620 	CPU_COPY(&grp->cg_mask, &cs);
5621 
5622 	/* Add the selected CPU offset to core offset. */
5623 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5624 		if (fcpu - 1 == cpu)
5625 			break;
5626 		CPU_CLR(fcpu - 1, &cs);
5627 	}
5628 	MPASS(fcpu);
5629 
5630 	core_offset += i;
5631 
5632 	CPU_COPY(&grp->cg_mask, &cs);
5633 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5634 		MPASS(CPU_FFS(&cs));
5635 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5636 	}
5637 	MPASS(CPU_FFS(&cs));
5638 	return CPU_FFS(&cs) - 1;
5639 }
5640 #else
5641 static int
5642 find_close_core(int cpu, int core_offset __unused)
5643 {
5644 	return cpu;
5645 }
5646 #endif
5647 
5648 static int
5649 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5650 {
5651 	switch (type) {
5652 	case IFLIB_INTR_TX:
5653 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5654 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5655 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5656 	case IFLIB_INTR_RX:
5657 	case IFLIB_INTR_RXTX:
5658 		/* RX queues get the specified core */
5659 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5660 	default:
5661 		return -1;
5662 	}
5663 }
5664 #else
5665 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5666 #define find_close_core(cpuid, tid)	CPU_FIRST()
5667 #define find_nth(ctx, gid)		CPU_FIRST()
5668 #endif
5669 
5670 /* Just to avoid copy/paste */
5671 static inline int
5672 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5673     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5674     const char *name)
5675 {
5676 	device_t dev;
5677 	int err, cpuid, tid;
5678 
5679 	dev = ctx->ifc_dev;
5680 	cpuid = find_nth(ctx, qid);
5681 	tid = get_core_offset(ctx, type, qid);
5682 	MPASS(tid >= 0);
5683 	cpuid = find_close_core(cpuid, tid);
5684 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5685 	    name);
5686 	if (err) {
5687 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5688 		return (err);
5689 	}
5690 #ifdef notyet
5691 	if (cpuid > ctx->ifc_cpuid_highest)
5692 		ctx->ifc_cpuid_highest = cpuid;
5693 #endif
5694 	return 0;
5695 }
5696 
5697 int
5698 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5699 			iflib_intr_type_t type, driver_filter_t *filter,
5700 			void *filter_arg, int qid, const char *name)
5701 {
5702 	device_t dev;
5703 	struct grouptask *gtask;
5704 	struct taskqgroup *tqg;
5705 	iflib_filter_info_t info;
5706 	gtask_fn_t *fn;
5707 	int tqrid, err;
5708 	driver_filter_t *intr_fast;
5709 	void *q;
5710 
5711 	info = &ctx->ifc_filter_info;
5712 	tqrid = rid;
5713 
5714 	switch (type) {
5715 	/* XXX merge tx/rx for netmap? */
5716 	case IFLIB_INTR_TX:
5717 		q = &ctx->ifc_txqs[qid];
5718 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5719 		gtask = &ctx->ifc_txqs[qid].ift_task;
5720 		tqg = qgroup_if_io_tqg;
5721 		fn = _task_fn_tx;
5722 		intr_fast = iflib_fast_intr;
5723 		GROUPTASK_INIT(gtask, 0, fn, q);
5724 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5725 		break;
5726 	case IFLIB_INTR_RX:
5727 		q = &ctx->ifc_rxqs[qid];
5728 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5729 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5730 		tqg = qgroup_if_io_tqg;
5731 		fn = _task_fn_rx;
5732 		intr_fast = iflib_fast_intr;
5733 		GROUPTASK_INIT(gtask, 0, fn, q);
5734 		break;
5735 	case IFLIB_INTR_RXTX:
5736 		q = &ctx->ifc_rxqs[qid];
5737 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5738 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5739 		tqg = qgroup_if_io_tqg;
5740 		fn = _task_fn_rx;
5741 		intr_fast = iflib_fast_intr_rxtx;
5742 		GROUPTASK_INIT(gtask, 0, fn, q);
5743 		break;
5744 	case IFLIB_INTR_ADMIN:
5745 		q = ctx;
5746 		tqrid = -1;
5747 		info = &ctx->ifc_filter_info;
5748 		gtask = &ctx->ifc_admin_task;
5749 		tqg = qgroup_if_config_tqg;
5750 		fn = _task_fn_admin;
5751 		intr_fast = iflib_fast_intr_ctx;
5752 		break;
5753 	default:
5754 		panic("unknown net intr type");
5755 	}
5756 
5757 	info->ifi_filter = filter;
5758 	info->ifi_filter_arg = filter_arg;
5759 	info->ifi_task = gtask;
5760 	info->ifi_ctx = q;
5761 
5762 	dev = ctx->ifc_dev;
5763 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5764 	if (err != 0) {
5765 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5766 		return (err);
5767 	}
5768 	if (type == IFLIB_INTR_ADMIN)
5769 		return (0);
5770 
5771 	if (tqrid != -1) {
5772 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5773 		    q, name);
5774 		if (err)
5775 			return (err);
5776 	} else {
5777 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5778 	}
5779 
5780 	return (0);
5781 }
5782 
5783 void
5784 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5785 {
5786 	struct grouptask *gtask;
5787 	struct taskqgroup *tqg;
5788 	gtask_fn_t *fn;
5789 	void *q;
5790 	int err;
5791 
5792 	switch (type) {
5793 	case IFLIB_INTR_TX:
5794 		q = &ctx->ifc_txqs[qid];
5795 		gtask = &ctx->ifc_txqs[qid].ift_task;
5796 		tqg = qgroup_if_io_tqg;
5797 		fn = _task_fn_tx;
5798 		break;
5799 	case IFLIB_INTR_RX:
5800 		q = &ctx->ifc_rxqs[qid];
5801 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5802 		tqg = qgroup_if_io_tqg;
5803 		fn = _task_fn_rx;
5804 		break;
5805 	case IFLIB_INTR_IOV:
5806 		q = ctx;
5807 		gtask = &ctx->ifc_vflr_task;
5808 		tqg = qgroup_if_config_tqg;
5809 		fn = _task_fn_iov;
5810 		break;
5811 	default:
5812 		panic("unknown net intr type");
5813 	}
5814 	GROUPTASK_INIT(gtask, 0, fn, q);
5815 	if (irq != NULL) {
5816 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5817 		    q, name);
5818 		if (err)
5819 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5820 			    irq->ii_res, name);
5821 	} else {
5822 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5823 	}
5824 }
5825 
5826 void
5827 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5828 {
5829 
5830 	if (irq->ii_tag)
5831 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5832 
5833 	if (irq->ii_res)
5834 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5835 		    rman_get_rid(irq->ii_res), irq->ii_res);
5836 }
5837 
5838 static int
5839 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5840 {
5841 	iflib_txq_t txq = ctx->ifc_txqs;
5842 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5843 	if_irq_t irq = &ctx->ifc_legacy_irq;
5844 	iflib_filter_info_t info;
5845 	device_t dev;
5846 	struct grouptask *gtask;
5847 	struct resource *res;
5848 	struct taskqgroup *tqg;
5849 	gtask_fn_t *fn;
5850 	int tqrid;
5851 	void *q;
5852 	int err;
5853 
5854 	q = &ctx->ifc_rxqs[0];
5855 	info = &rxq[0].ifr_filter_info;
5856 	gtask = &rxq[0].ifr_task;
5857 	tqg = qgroup_if_io_tqg;
5858 	tqrid = irq->ii_rid = *rid;
5859 	fn = _task_fn_rx;
5860 
5861 	ctx->ifc_flags |= IFC_LEGACY;
5862 	info->ifi_filter = filter;
5863 	info->ifi_filter_arg = filter_arg;
5864 	info->ifi_task = gtask;
5865 	info->ifi_ctx = ctx;
5866 
5867 	dev = ctx->ifc_dev;
5868 	/* We allocate a single interrupt resource */
5869 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5870 		return (err);
5871 	GROUPTASK_INIT(gtask, 0, fn, q);
5872 	res = irq->ii_res;
5873 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
5874 
5875 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5876 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
5877 	    "tx");
5878 	return (0);
5879 }
5880 
5881 void
5882 iflib_led_create(if_ctx_t ctx)
5883 {
5884 
5885 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5886 	    device_get_nameunit(ctx->ifc_dev));
5887 }
5888 
5889 void
5890 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5891 {
5892 
5893 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5894 }
5895 
5896 void
5897 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5898 {
5899 
5900 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5901 }
5902 
5903 void
5904 iflib_admin_intr_deferred(if_ctx_t ctx)
5905 {
5906 #ifdef INVARIANTS
5907 	struct grouptask *gtask;
5908 
5909 	gtask = &ctx->ifc_admin_task;
5910 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5911 #endif
5912 
5913 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5914 }
5915 
5916 void
5917 iflib_iov_intr_deferred(if_ctx_t ctx)
5918 {
5919 
5920 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5921 }
5922 
5923 void
5924 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5925 {
5926 
5927 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
5928 	    name);
5929 }
5930 
5931 void
5932 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5933 	const char *name)
5934 {
5935 
5936 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5937 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
5938 	    name);
5939 }
5940 
5941 void
5942 iflib_config_gtask_deinit(struct grouptask *gtask)
5943 {
5944 
5945 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5946 }
5947 
5948 void
5949 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5950 {
5951 	if_t ifp = ctx->ifc_ifp;
5952 	iflib_txq_t txq = ctx->ifc_txqs;
5953 
5954 	if_setbaudrate(ifp, baudrate);
5955 	if (baudrate >= IF_Gbps(10)) {
5956 		STATE_LOCK(ctx);
5957 		ctx->ifc_flags |= IFC_PREFETCH;
5958 		STATE_UNLOCK(ctx);
5959 	}
5960 	/* If link down, disable watchdog */
5961 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5962 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5963 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5964 	}
5965 	ctx->ifc_link_state = link_state;
5966 	if_link_state_change(ifp, link_state);
5967 }
5968 
5969 static int
5970 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5971 {
5972 	int credits;
5973 #ifdef INVARIANTS
5974 	int credits_pre = txq->ift_cidx_processed;
5975 #endif
5976 
5977 	if (ctx->isc_txd_credits_update == NULL)
5978 		return (0);
5979 
5980 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5981 	    BUS_DMASYNC_POSTREAD);
5982 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5983 		return (0);
5984 
5985 	txq->ift_processed += credits;
5986 	txq->ift_cidx_processed += credits;
5987 
5988 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5989 	if (txq->ift_cidx_processed >= txq->ift_size)
5990 		txq->ift_cidx_processed -= txq->ift_size;
5991 	return (credits);
5992 }
5993 
5994 static int
5995 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5996 {
5997 	iflib_fl_t fl;
5998 	u_int i;
5999 
6000 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6001 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6002 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6003 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6004 	    budget));
6005 }
6006 
6007 void
6008 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6009 	const char *description, if_int_delay_info_t info,
6010 	int offset, int value)
6011 {
6012 	info->iidi_ctx = ctx;
6013 	info->iidi_offset = offset;
6014 	info->iidi_value = value;
6015 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6016 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6017 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6018 	    info, 0, iflib_sysctl_int_delay, "I", description);
6019 }
6020 
6021 struct sx *
6022 iflib_ctx_lock_get(if_ctx_t ctx)
6023 {
6024 
6025 	return (&ctx->ifc_ctx_sx);
6026 }
6027 
6028 static int
6029 iflib_msix_init(if_ctx_t ctx)
6030 {
6031 	device_t dev = ctx->ifc_dev;
6032 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6033 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6034 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6035 	int iflib_num_tx_queues, iflib_num_rx_queues;
6036 	int err, admincnt, bar;
6037 
6038 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6039 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6040 
6041 	if (bootverbose)
6042 		device_printf(dev, "msix_init qsets capped at %d\n",
6043 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6044 
6045 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6046 	admincnt = sctx->isc_admin_intrcnt;
6047 	/* Override by tuneable */
6048 	if (scctx->isc_disable_msix)
6049 		goto msi;
6050 
6051 	/* First try MSI-X */
6052 	if ((msgs = pci_msix_count(dev)) == 0) {
6053 		if (bootverbose)
6054 			device_printf(dev, "MSI-X not supported or disabled\n");
6055 		goto msi;
6056 	}
6057 	/*
6058 	 * bar == -1 => "trust me I know what I'm doing"
6059 	 * Some drivers are for hardware that is so shoddily
6060 	 * documented that no one knows which bars are which
6061 	 * so the developer has to map all bars. This hack
6062 	 * allows shoddy garbage to use MSI-X in this framework.
6063 	 */
6064 	if (bar != -1) {
6065 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6066 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6067 		if (ctx->ifc_msix_mem == NULL) {
6068 			device_printf(dev, "Unable to map MSI-X table\n");
6069 			goto msi;
6070 		}
6071 	}
6072 #if IFLIB_DEBUG
6073 	/* use only 1 qset in debug mode */
6074 	queuemsgs = min(msgs - admincnt, 1);
6075 #else
6076 	queuemsgs = msgs - admincnt;
6077 #endif
6078 #ifdef RSS
6079 	queues = imin(queuemsgs, rss_getnumbuckets());
6080 #else
6081 	queues = queuemsgs;
6082 #endif
6083 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6084 	if (bootverbose)
6085 		device_printf(dev,
6086 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6087 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6088 #ifdef  RSS
6089 	/* If we're doing RSS, clamp at the number of RSS buckets */
6090 	if (queues > rss_getnumbuckets())
6091 		queues = rss_getnumbuckets();
6092 #endif
6093 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6094 		rx_queues = iflib_num_rx_queues;
6095 	else
6096 		rx_queues = queues;
6097 
6098 	if (rx_queues > scctx->isc_nrxqsets)
6099 		rx_queues = scctx->isc_nrxqsets;
6100 
6101 	/*
6102 	 * We want this to be all logical CPUs by default
6103 	 */
6104 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6105 		tx_queues = iflib_num_tx_queues;
6106 	else
6107 		tx_queues = mp_ncpus;
6108 
6109 	if (tx_queues > scctx->isc_ntxqsets)
6110 		tx_queues = scctx->isc_ntxqsets;
6111 
6112 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6113 #ifdef INVARIANTS
6114 		if (tx_queues != rx_queues)
6115 			device_printf(dev,
6116 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6117 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6118 #endif
6119 		tx_queues = min(rx_queues, tx_queues);
6120 		rx_queues = min(rx_queues, tx_queues);
6121 	}
6122 
6123 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6124 	    rx_queues, tx_queues);
6125 
6126 	vectors = rx_queues + admincnt;
6127 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6128 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6129 		    vectors);
6130 		scctx->isc_vectors = vectors;
6131 		scctx->isc_nrxqsets = rx_queues;
6132 		scctx->isc_ntxqsets = tx_queues;
6133 		scctx->isc_intr = IFLIB_INTR_MSIX;
6134 
6135 		return (vectors);
6136 	} else {
6137 		device_printf(dev,
6138 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6139 		    vectors, err);
6140 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6141 		    ctx->ifc_msix_mem);
6142 		ctx->ifc_msix_mem = NULL;
6143 	}
6144 msi:
6145 	vectors = pci_msi_count(dev);
6146 	scctx->isc_nrxqsets = 1;
6147 	scctx->isc_ntxqsets = 1;
6148 	scctx->isc_vectors = vectors;
6149 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6150 		device_printf(dev,"Using an MSI interrupt\n");
6151 		scctx->isc_intr = IFLIB_INTR_MSI;
6152 	} else {
6153 		scctx->isc_vectors = 1;
6154 		device_printf(dev,"Using a Legacy interrupt\n");
6155 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6156 	}
6157 
6158 	return (vectors);
6159 }
6160 
6161 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6162 
6163 static int
6164 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6165 {
6166 	int rc;
6167 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6168 	struct sbuf *sb;
6169 	const char *ring_state = "UNKNOWN";
6170 
6171 	/* XXX needed ? */
6172 	rc = sysctl_wire_old_buffer(req, 0);
6173 	MPASS(rc == 0);
6174 	if (rc != 0)
6175 		return (rc);
6176 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6177 	MPASS(sb != NULL);
6178 	if (sb == NULL)
6179 		return (ENOMEM);
6180 	if (state[3] <= 3)
6181 		ring_state = ring_states[state[3]];
6182 
6183 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6184 		    state[0], state[1], state[2], ring_state);
6185 	rc = sbuf_finish(sb);
6186 	sbuf_delete(sb);
6187         return(rc);
6188 }
6189 
6190 enum iflib_ndesc_handler {
6191 	IFLIB_NTXD_HANDLER,
6192 	IFLIB_NRXD_HANDLER,
6193 };
6194 
6195 static int
6196 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6197 {
6198 	if_ctx_t ctx = (void *)arg1;
6199 	enum iflib_ndesc_handler type = arg2;
6200 	char buf[256] = {0};
6201 	qidx_t *ndesc;
6202 	char *p, *next;
6203 	int nqs, rc, i;
6204 
6205 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6206 
6207 	nqs = 8;
6208 	switch(type) {
6209 	case IFLIB_NTXD_HANDLER:
6210 		ndesc = ctx->ifc_sysctl_ntxds;
6211 		if (ctx->ifc_sctx)
6212 			nqs = ctx->ifc_sctx->isc_ntxqs;
6213 		break;
6214 	case IFLIB_NRXD_HANDLER:
6215 		ndesc = ctx->ifc_sysctl_nrxds;
6216 		if (ctx->ifc_sctx)
6217 			nqs = ctx->ifc_sctx->isc_nrxqs;
6218 		break;
6219 	default:
6220 			panic("unhandled type");
6221 	}
6222 	if (nqs == 0)
6223 		nqs = 8;
6224 
6225 	for (i=0; i<8; i++) {
6226 		if (i >= nqs)
6227 			break;
6228 		if (i)
6229 			strcat(buf, ",");
6230 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6231 	}
6232 
6233 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6234 	if (rc || req->newptr == NULL)
6235 		return rc;
6236 
6237 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6238 	    i++, p = strsep(&next, " ,")) {
6239 		ndesc[i] = strtoul(p, NULL, 10);
6240 	}
6241 
6242 	return(rc);
6243 }
6244 
6245 #define NAME_BUFLEN 32
6246 static void
6247 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6248 {
6249         device_t dev = iflib_get_dev(ctx);
6250 	struct sysctl_oid_list *child, *oid_list;
6251 	struct sysctl_ctx_list *ctx_list;
6252 	struct sysctl_oid *node;
6253 
6254 	ctx_list = device_get_sysctl_ctx(dev);
6255 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6256 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6257 						      CTLFLAG_RD, NULL, "IFLIB fields");
6258 	oid_list = SYSCTL_CHILDREN(node);
6259 
6260 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6261 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6262 		       "driver version");
6263 
6264 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6265 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6266 			"# of txqs to use, 0 => use default #");
6267 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6268 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6269 			"# of rxqs to use, 0 => use default #");
6270 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6271 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6272                        "permit #txq != #rxq");
6273 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6274                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6275                       "disable MSI-X (default 0)");
6276 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6277 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6278                        "set the rx budget");
6279 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6280 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6281 		       "cause tx to abdicate instead of running to completion");
6282 
6283 	/* XXX change for per-queue sizes */
6284 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6285 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6286                        mp_ndesc_handler, "A",
6287                        "list of # of tx descriptors to use, 0 = use default #");
6288 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6289 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6290                        mp_ndesc_handler, "A",
6291                        "list of # of rx descriptors to use, 0 = use default #");
6292 }
6293 
6294 static void
6295 iflib_add_device_sysctl_post(if_ctx_t ctx)
6296 {
6297 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6298 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6299         device_t dev = iflib_get_dev(ctx);
6300 	struct sysctl_oid_list *child;
6301 	struct sysctl_ctx_list *ctx_list;
6302 	iflib_fl_t fl;
6303 	iflib_txq_t txq;
6304 	iflib_rxq_t rxq;
6305 	int i, j;
6306 	char namebuf[NAME_BUFLEN];
6307 	char *qfmt;
6308 	struct sysctl_oid *queue_node, *fl_node, *node;
6309 	struct sysctl_oid_list *queue_list, *fl_list;
6310 	ctx_list = device_get_sysctl_ctx(dev);
6311 
6312 	node = ctx->ifc_sysctl_node;
6313 	child = SYSCTL_CHILDREN(node);
6314 
6315 	if (scctx->isc_ntxqsets > 100)
6316 		qfmt = "txq%03d";
6317 	else if (scctx->isc_ntxqsets > 10)
6318 		qfmt = "txq%02d";
6319 	else
6320 		qfmt = "txq%d";
6321 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6322 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6323 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6324 					     CTLFLAG_RD, NULL, "Queue Name");
6325 		queue_list = SYSCTL_CHILDREN(queue_node);
6326 #if MEMORY_LOGGING
6327 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6328 				CTLFLAG_RD,
6329 				&txq->ift_dequeued, "total mbufs freed");
6330 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6331 				CTLFLAG_RD,
6332 				&txq->ift_enqueued, "total mbufs enqueued");
6333 #endif
6334 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6335 				   CTLFLAG_RD,
6336 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6337 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6338 				   CTLFLAG_RD,
6339 				   &txq->ift_pullups, "# of times m_pullup was called");
6340 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6341 				   CTLFLAG_RD,
6342 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6343 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6344 				   CTLFLAG_RD,
6345 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6346 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6347 				   CTLFLAG_RD,
6348 				   &txq->ift_map_failed, "# of times dma map failed");
6349 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6350 				   CTLFLAG_RD,
6351 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6352 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6353 				   CTLFLAG_RD,
6354 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6355 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6356 				   CTLFLAG_RD,
6357 				   &txq->ift_pidx, 1, "Producer Index");
6358 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6359 				   CTLFLAG_RD,
6360 				   &txq->ift_cidx, 1, "Consumer Index");
6361 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6362 				   CTLFLAG_RD,
6363 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6364 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6365 				   CTLFLAG_RD,
6366 				   &txq->ift_in_use, 1, "descriptors in use");
6367 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6368 				   CTLFLAG_RD,
6369 				   &txq->ift_processed, "descriptors procesed for clean");
6370 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6371 				   CTLFLAG_RD,
6372 				   &txq->ift_cleaned, "total cleaned");
6373 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6374 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6375 				0, mp_ring_state_handler, "A", "soft ring state");
6376 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6377 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6378 				       "# of enqueues to the mp_ring for this queue");
6379 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6380 				       CTLFLAG_RD, &txq->ift_br->drops,
6381 				       "# of drops in the mp_ring for this queue");
6382 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6383 				       CTLFLAG_RD, &txq->ift_br->starts,
6384 				       "# of normal consumer starts in the mp_ring for this queue");
6385 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6386 				       CTLFLAG_RD, &txq->ift_br->stalls,
6387 					       "# of consumer stalls in the mp_ring for this queue");
6388 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6389 			       CTLFLAG_RD, &txq->ift_br->restarts,
6390 				       "# of consumer restarts in the mp_ring for this queue");
6391 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6392 				       CTLFLAG_RD, &txq->ift_br->abdications,
6393 				       "# of consumer abdications in the mp_ring for this queue");
6394 	}
6395 
6396 	if (scctx->isc_nrxqsets > 100)
6397 		qfmt = "rxq%03d";
6398 	else if (scctx->isc_nrxqsets > 10)
6399 		qfmt = "rxq%02d";
6400 	else
6401 		qfmt = "rxq%d";
6402 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6403 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6404 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6405 					     CTLFLAG_RD, NULL, "Queue Name");
6406 		queue_list = SYSCTL_CHILDREN(queue_node);
6407 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6408 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6409 				       CTLFLAG_RD,
6410 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6411 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6412 				       CTLFLAG_RD,
6413 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6414 		}
6415 
6416 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6417 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6418 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6419 						     CTLFLAG_RD, NULL, "freelist Name");
6420 			fl_list = SYSCTL_CHILDREN(fl_node);
6421 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6422 				       CTLFLAG_RD,
6423 				       &fl->ifl_pidx, 1, "Producer Index");
6424 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6425 				       CTLFLAG_RD,
6426 				       &fl->ifl_cidx, 1, "Consumer Index");
6427 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6428 				       CTLFLAG_RD,
6429 				       &fl->ifl_credits, 1, "credits available");
6430 #if MEMORY_LOGGING
6431 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6432 					CTLFLAG_RD,
6433 					&fl->ifl_m_enqueued, "mbufs allocated");
6434 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6435 					CTLFLAG_RD,
6436 					&fl->ifl_m_dequeued, "mbufs freed");
6437 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6438 					CTLFLAG_RD,
6439 					&fl->ifl_cl_enqueued, "clusters allocated");
6440 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6441 					CTLFLAG_RD,
6442 					&fl->ifl_cl_dequeued, "clusters freed");
6443 #endif
6444 
6445 		}
6446 	}
6447 
6448 }
6449 
6450 void
6451 iflib_request_reset(if_ctx_t ctx)
6452 {
6453 
6454 	STATE_LOCK(ctx);
6455 	ctx->ifc_flags |= IFC_DO_RESET;
6456 	STATE_UNLOCK(ctx);
6457 }
6458 
6459 #ifndef __NO_STRICT_ALIGNMENT
6460 static struct mbuf *
6461 iflib_fixup_rx(struct mbuf *m)
6462 {
6463 	struct mbuf *n;
6464 
6465 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6466 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6467 		m->m_data += ETHER_HDR_LEN;
6468 		n = m;
6469 	} else {
6470 		MGETHDR(n, M_NOWAIT, MT_DATA);
6471 		if (n == NULL) {
6472 			m_freem(m);
6473 			return (NULL);
6474 		}
6475 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6476 		m->m_data += ETHER_HDR_LEN;
6477 		m->m_len -= ETHER_HDR_LEN;
6478 		n->m_len = ETHER_HDR_LEN;
6479 		M_MOVE_PKTHDR(n, m);
6480 		n->m_next = m;
6481 	}
6482 	return (n);
6483 }
6484 #endif
6485 
6486 #ifdef NETDUMP
6487 static void
6488 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6489 {
6490 	if_ctx_t ctx;
6491 
6492 	ctx = if_getsoftc(ifp);
6493 	CTX_LOCK(ctx);
6494 	*nrxr = NRXQSETS(ctx);
6495 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6496 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6497 	CTX_UNLOCK(ctx);
6498 }
6499 
6500 static void
6501 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6502 {
6503 	if_ctx_t ctx;
6504 	if_softc_ctx_t scctx;
6505 	iflib_fl_t fl;
6506 	iflib_rxq_t rxq;
6507 	int i, j;
6508 
6509 	ctx = if_getsoftc(ifp);
6510 	scctx = &ctx->ifc_softc_ctx;
6511 
6512 	switch (event) {
6513 	case NETDUMP_START:
6514 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6515 			rxq = &ctx->ifc_rxqs[i];
6516 			for (j = 0; j < rxq->ifr_nfl; j++) {
6517 				fl = rxq->ifr_fl;
6518 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6519 			}
6520 		}
6521 		iflib_no_tx_batch = 1;
6522 		break;
6523 	default:
6524 		break;
6525 	}
6526 }
6527 
6528 static int
6529 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6530 {
6531 	if_ctx_t ctx;
6532 	iflib_txq_t txq;
6533 	int error;
6534 
6535 	ctx = if_getsoftc(ifp);
6536 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6537 	    IFF_DRV_RUNNING)
6538 		return (EBUSY);
6539 
6540 	txq = &ctx->ifc_txqs[0];
6541 	error = iflib_encap(txq, &m);
6542 	if (error == 0)
6543 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6544 	return (error);
6545 }
6546 
6547 static int
6548 iflib_netdump_poll(struct ifnet *ifp, int count)
6549 {
6550 	if_ctx_t ctx;
6551 	if_softc_ctx_t scctx;
6552 	iflib_txq_t txq;
6553 	int i;
6554 
6555 	ctx = if_getsoftc(ifp);
6556 	scctx = &ctx->ifc_softc_ctx;
6557 
6558 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6559 	    IFF_DRV_RUNNING)
6560 		return (EBUSY);
6561 
6562 	txq = &ctx->ifc_txqs[0];
6563 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6564 
6565 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6566 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6567 	return (0);
6568 }
6569 #endif /* NETDUMP */
6570