xref: /freebsd/sys/net/iflib.c (revision 71ceff9f0b4a3add236a6458da0ea3671fedd0a1)
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/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/netdump/netdump.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 
144 typedef struct iflib_filter_info {
145 	driver_filter_t *ifi_filter;
146 	void *ifi_filter_arg;
147 	struct grouptask *ifi_task;
148 	void *ifi_ctx;
149 } *iflib_filter_info_t;
150 
151 struct iflib_ctx {
152 	KOBJ_FIELDS;
153 	/*
154 	 * Pointer to hardware driver's softc
155 	 */
156 	void *ifc_softc;
157 	device_t ifc_dev;
158 	if_t ifc_ifp;
159 
160 	cpuset_t ifc_cpus;
161 	if_shared_ctx_t ifc_sctx;
162 	struct if_softc_ctx ifc_softc_ctx;
163 
164 	struct sx ifc_ctx_sx;
165 	struct mtx ifc_state_mtx;
166 
167 	iflib_txq_t ifc_txqs;
168 	iflib_rxq_t ifc_rxqs;
169 	uint32_t ifc_if_flags;
170 	uint32_t ifc_flags;
171 	uint32_t ifc_max_fl_buf_size;
172 	uint32_t ifc_rx_mbuf_sz;
173 
174 	int ifc_link_state;
175 	int ifc_link_irq;
176 	int ifc_watchdog_events;
177 	struct cdev *ifc_led_dev;
178 	struct resource *ifc_msix_mem;
179 
180 	struct if_irq ifc_legacy_irq;
181 	struct grouptask ifc_admin_task;
182 	struct grouptask ifc_vflr_task;
183 	struct iflib_filter_info ifc_filter_info;
184 	struct ifmedia	ifc_media;
185 	struct ifmedia	*ifc_mediap;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 	uint16_t ifc_sysctl_core_offset;
194 #define	CORE_OFFSET_UNSPECIFIED	0xffff
195 	uint8_t  ifc_sysctl_separate_txrx;
196 
197 	qidx_t ifc_sysctl_ntxds[8];
198 	qidx_t ifc_sysctl_nrxds[8];
199 	struct if_txrx ifc_txrx;
200 #define isc_txd_encap  ifc_txrx.ift_txd_encap
201 #define isc_txd_flush  ifc_txrx.ift_txd_flush
202 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
203 #define isc_rxd_available ifc_txrx.ift_rxd_available
204 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	struct ether_addr ifc_mac;
213 	char ifc_mtx_name[16];
214 };
215 
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 #define IFLIB_RX_COPY_THRESH		128
296 #define IFLIB_MAX_RX_REFRESH		32
297 /* The minimum descriptors per second before we start coalescing */
298 #define IFLIB_MIN_DESC_SEC		16384
299 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
300 #define IFLIB_QUEUE_IDLE		0
301 #define IFLIB_QUEUE_HUNG		1
302 #define IFLIB_QUEUE_WORKING		2
303 /* maximum number of txqs that can share an rx interrupt */
304 #define IFLIB_MAX_TX_SHARED_INTR	4
305 
306 /* this should really scale with ring size - this is a fairly arbitrary value */
307 #define TX_BATCH_SIZE			32
308 
309 #define IFLIB_RESTART_BUDGET		8
310 
311 
312 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
313 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
314 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
315 struct iflib_txq {
316 	qidx_t		ift_in_use;
317 	qidx_t		ift_cidx;
318 	qidx_t		ift_cidx_processed;
319 	qidx_t		ift_pidx;
320 	uint8_t		ift_gen;
321 	uint8_t		ift_br_offset;
322 	uint16_t	ift_npending;
323 	uint16_t	ift_db_pending;
324 	uint16_t	ift_rs_pending;
325 	/* implicit pad */
326 	uint8_t		ift_txd_size[8];
327 	uint64_t	ift_processed;
328 	uint64_t	ift_cleaned;
329 	uint64_t	ift_cleaned_prev;
330 #if MEMORY_LOGGING
331 	uint64_t	ift_enqueued;
332 	uint64_t	ift_dequeued;
333 #endif
334 	uint64_t	ift_no_tx_dma_setup;
335 	uint64_t	ift_no_desc_avail;
336 	uint64_t	ift_mbuf_defrag_failed;
337 	uint64_t	ift_mbuf_defrag;
338 	uint64_t	ift_map_failed;
339 	uint64_t	ift_txd_encap_efbig;
340 	uint64_t	ift_pullups;
341 	uint64_t	ift_last_timer_tick;
342 
343 	struct mtx	ift_mtx;
344 	struct mtx	ift_db_mtx;
345 
346 	/* constant values */
347 	if_ctx_t	ift_ctx;
348 	struct ifmp_ring        *ift_br;
349 	struct grouptask	ift_task;
350 	qidx_t		ift_size;
351 	uint16_t	ift_id;
352 	struct callout	ift_timer;
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t	ift_buf_tag;
360 	bus_dma_tag_t	ift_tso_buf_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define MTX_NAME_LEN 16
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	char                    ift_db_mtx_name[MTX_NAME_LEN];
365 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
366 #ifdef IFLIB_DIAGNOSTICS
367 	uint64_t ift_cpu_exec_count[256];
368 #endif
369 } __aligned(CACHE_LINE_SIZE);
370 
371 struct iflib_fl {
372 	qidx_t		ifl_cidx;
373 	qidx_t		ifl_pidx;
374 	qidx_t		ifl_credits;
375 	uint8_t		ifl_gen;
376 	uint8_t		ifl_rxd_size;
377 #if MEMORY_LOGGING
378 	uint64_t	ifl_m_enqueued;
379 	uint64_t	ifl_m_dequeued;
380 	uint64_t	ifl_cl_enqueued;
381 	uint64_t	ifl_cl_dequeued;
382 #endif
383 	/* implicit pad */
384 
385 	bitstr_t 	*ifl_rx_bitmap;
386 	qidx_t		ifl_fragidx;
387 	/* constant */
388 	qidx_t		ifl_size;
389 	uint16_t	ifl_buf_size;
390 	uint16_t	ifl_cltype;
391 	uma_zone_t	ifl_zone;
392 	iflib_rxsd_array_t	ifl_sds;
393 	iflib_rxq_t	ifl_rxq;
394 	uint8_t		ifl_id;
395 	bus_dma_tag_t	ifl_buf_tag;
396 	iflib_dma_info_t	ifl_ifdi;
397 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
398 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
399 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
400 }  __aligned(CACHE_LINE_SIZE);
401 
402 static inline qidx_t
403 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
404 {
405 	qidx_t used;
406 
407 	if (pidx > cidx)
408 		used = pidx - cidx;
409 	else if (pidx < cidx)
410 		used = size - cidx + pidx;
411 	else if (gen == 0 && pidx == cidx)
412 		used = 0;
413 	else if (gen == 1 && pidx == cidx)
414 		used = size;
415 	else
416 		panic("bad state");
417 
418 	return (used);
419 }
420 
421 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 
423 #define IDXDIFF(head, tail, wrap) \
424 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
425 
426 struct iflib_rxq {
427 	/* If there is a separate completion queue -
428 	 * these are the cq cidx and pidx. Otherwise
429 	 * these are unused.
430 	 */
431 	qidx_t		ifr_size;
432 	qidx_t		ifr_cq_cidx;
433 	qidx_t		ifr_cq_pidx;
434 	uint8_t		ifr_cq_gen;
435 	uint8_t		ifr_fl_offset;
436 
437 	if_ctx_t	ifr_ctx;
438 	iflib_fl_t	ifr_fl;
439 	uint64_t	ifr_rx_irq;
440 	struct pfil_head	*pfil;
441 	uint16_t	ifr_id;
442 	uint8_t		ifr_lro_enabled;
443 	uint8_t		ifr_nfl;
444 	uint8_t		ifr_ntxqirq;
445 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
446 	struct lro_ctrl			ifr_lc;
447 	struct grouptask        ifr_task;
448 	struct iflib_filter_info ifr_filter_info;
449 	iflib_dma_info_t		ifr_ifdi;
450 
451 	/* dynamically allocate if any drivers need a value substantially larger than this */
452 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
453 #ifdef IFLIB_DIAGNOSTICS
454 	uint64_t ifr_cpu_exec_count[256];
455 #endif
456 }  __aligned(CACHE_LINE_SIZE);
457 
458 typedef struct if_rxsd {
459 	caddr_t *ifsd_cl;
460 	iflib_fl_t ifsd_fl;
461 	qidx_t ifsd_cidx;
462 } *if_rxsd_t;
463 
464 /* multiple of word size */
465 #ifdef __LP64__
466 #define PKT_INFO_SIZE	6
467 #define RXD_INFO_SIZE	5
468 #define PKT_TYPE uint64_t
469 #else
470 #define PKT_INFO_SIZE	11
471 #define RXD_INFO_SIZE	8
472 #define PKT_TYPE uint32_t
473 #endif
474 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
475 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
476 
477 typedef struct if_pkt_info_pad {
478 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
479 } *if_pkt_info_pad_t;
480 typedef struct if_rxd_info_pad {
481 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
482 } *if_rxd_info_pad_t;
483 
484 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
485 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
486 
487 
488 static inline void
489 pkt_info_zero(if_pkt_info_t pi)
490 {
491 	if_pkt_info_pad_t pi_pad;
492 
493 	pi_pad = (if_pkt_info_pad_t)pi;
494 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
495 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
496 #ifndef __LP64__
497 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
498 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
499 #endif
500 }
501 
502 static device_method_t iflib_pseudo_methods[] = {
503 	DEVMETHOD(device_attach, noop_attach),
504 	DEVMETHOD(device_detach, iflib_pseudo_detach),
505 	DEVMETHOD_END
506 };
507 
508 driver_t iflib_pseudodriver = {
509 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
510 };
511 
512 static inline void
513 rxd_info_zero(if_rxd_info_t ri)
514 {
515 	if_rxd_info_pad_t ri_pad;
516 	int i;
517 
518 	ri_pad = (if_rxd_info_pad_t)ri;
519 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
520 		ri_pad->rxd_val[i] = 0;
521 		ri_pad->rxd_val[i+1] = 0;
522 		ri_pad->rxd_val[i+2] = 0;
523 		ri_pad->rxd_val[i+3] = 0;
524 	}
525 #ifdef __LP64__
526 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
527 #endif
528 }
529 
530 /*
531  * Only allow a single packet to take up most 1/nth of the tx ring
532  */
533 #define MAX_SINGLE_PACKET_FRACTION 12
534 #define IF_BAD_DMA (bus_addr_t)-1
535 
536 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
537 
538 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
539 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
540 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
541 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
542 
543 
544 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
545 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
546 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
547 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
548 
549 
550 
551 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
552 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
553 
554 void
555 iflib_set_detach(if_ctx_t ctx)
556 {
557 	STATE_LOCK(ctx);
558 	ctx->ifc_flags |= IFC_IN_DETACH;
559 	STATE_UNLOCK(ctx);
560 }
561 
562 /* Our boot-time initialization hook */
563 static int	iflib_module_event_handler(module_t, int, void *);
564 
565 static moduledata_t iflib_moduledata = {
566 	"iflib",
567 	iflib_module_event_handler,
568 	NULL
569 };
570 
571 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
572 MODULE_VERSION(iflib, 1);
573 
574 MODULE_DEPEND(iflib, pci, 1, 1, 1);
575 MODULE_DEPEND(iflib, ether, 1, 1, 1);
576 
577 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
578 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
579 
580 #ifndef IFLIB_DEBUG_COUNTERS
581 #ifdef INVARIANTS
582 #define IFLIB_DEBUG_COUNTERS 1
583 #else
584 #define IFLIB_DEBUG_COUNTERS 0
585 #endif /* !INVARIANTS */
586 #endif
587 
588 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
589                    "iflib driver parameters");
590 
591 /*
592  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
593  */
594 static int iflib_min_tx_latency = 0;
595 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
596 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
597 static int iflib_no_tx_batch = 0;
598 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
599 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
600 
601 
602 #if IFLIB_DEBUG_COUNTERS
603 
604 static int iflib_tx_seen;
605 static int iflib_tx_sent;
606 static int iflib_tx_encap;
607 static int iflib_rx_allocs;
608 static int iflib_fl_refills;
609 static int iflib_fl_refills_large;
610 static int iflib_tx_frees;
611 
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
613 		   &iflib_tx_seen, 0, "# tx mbufs seen");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
615 		   &iflib_tx_sent, 0, "# tx mbufs sent");
616 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
617 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
618 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
619 		   &iflib_tx_frees, 0, "# tx frees");
620 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
621 		   &iflib_rx_allocs, 0, "# rx allocations");
622 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
623 		   &iflib_fl_refills, 0, "# refills");
624 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
625 		   &iflib_fl_refills_large, 0, "# large refills");
626 
627 
628 static int iflib_txq_drain_flushing;
629 static int iflib_txq_drain_oactive;
630 static int iflib_txq_drain_notready;
631 
632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
633 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
634 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
635 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
636 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
637 		   &iflib_txq_drain_notready, 0, "# drain notready");
638 
639 
640 static int iflib_encap_load_mbuf_fail;
641 static int iflib_encap_pad_mbuf_fail;
642 static int iflib_encap_txq_avail_fail;
643 static int iflib_encap_txd_encap_fail;
644 
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
646 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
647 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
648 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
649 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
650 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
651 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
652 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
653 
654 static int iflib_task_fn_rxs;
655 static int iflib_rx_intr_enables;
656 static int iflib_fast_intrs;
657 static int iflib_rx_unavail;
658 static int iflib_rx_ctx_inactive;
659 static int iflib_rx_if_input;
660 static int iflib_rxd_flush;
661 
662 static int iflib_verbose_debug;
663 
664 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
665 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
667 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
668 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
669 		   &iflib_fast_intrs, 0, "# fast_intr calls");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
671 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
673 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
675 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
677 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
678 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
679 		   &iflib_verbose_debug, 0, "enable verbose debugging");
680 
681 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
682 static void
683 iflib_debug_reset(void)
684 {
685 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
686 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
687 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
688 		iflib_txq_drain_notready =
689 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
690 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
691 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
692 		iflib_rx_unavail =
693 		iflib_rx_ctx_inactive = iflib_rx_if_input =
694 		iflib_rxd_flush = 0;
695 }
696 
697 #else
698 #define DBG_COUNTER_INC(name)
699 static void iflib_debug_reset(void) {}
700 #endif
701 
702 #define IFLIB_DEBUG 0
703 
704 static void iflib_tx_structures_free(if_ctx_t ctx);
705 static void iflib_rx_structures_free(if_ctx_t ctx);
706 static int iflib_queues_alloc(if_ctx_t ctx);
707 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
708 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
709 static int iflib_qset_structures_setup(if_ctx_t ctx);
710 static int iflib_msix_init(if_ctx_t ctx);
711 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
712 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
713 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
714 #ifdef ALTQ
715 static void iflib_altq_if_start(if_t ifp);
716 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
717 #endif
718 static int iflib_register(if_ctx_t);
719 static void iflib_init_locked(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
722 static void iflib_ifmp_purge(iflib_txq_t txq);
723 static void _iflib_pre_assert(if_softc_ctx_t scctx);
724 static void iflib_if_init_locked(if_ctx_t ctx);
725 static void iflib_free_intr_mem(if_ctx_t ctx);
726 #ifndef __NO_STRICT_ALIGNMENT
727 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
728 #endif
729 
730 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
731     SLIST_HEAD_INITIALIZER(cpu_offsets);
732 struct cpu_offset {
733 	SLIST_ENTRY(cpu_offset) entries;
734 	cpuset_t	set;
735 	unsigned int	refcount;
736 	uint16_t	offset;
737 };
738 static struct mtx cpu_offset_mtx;
739 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
740     MTX_DEF);
741 
742 NETDUMP_DEFINE(iflib);
743 
744 #ifdef DEV_NETMAP
745 #include <sys/selinfo.h>
746 #include <net/netmap.h>
747 #include <dev/netmap/netmap_kern.h>
748 
749 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
750 
751 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
752 
753 /*
754  * device-specific sysctl variables:
755  *
756  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
757  *	During regular operations the CRC is stripped, but on some
758  *	hardware reception of frames not multiple of 64 is slower,
759  *	so using crcstrip=0 helps in benchmarks.
760  *
761  * iflib_rx_miss, iflib_rx_miss_bufs:
762  *	count packets that might be missed due to lost interrupts.
763  */
764 SYSCTL_DECL(_dev_netmap);
765 /*
766  * The xl driver by default strips CRCs and we do not override it.
767  */
768 
769 int iflib_crcstrip = 1;
770 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
771     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
772 
773 int iflib_rx_miss, iflib_rx_miss_bufs;
774 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
775     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
776 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
777     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
778 
779 /*
780  * Register/unregister. We are already under netmap lock.
781  * Only called on the first register or the last unregister.
782  */
783 static int
784 iflib_netmap_register(struct netmap_adapter *na, int onoff)
785 {
786 	struct ifnet *ifp = na->ifp;
787 	if_ctx_t ctx = ifp->if_softc;
788 	int status;
789 
790 	CTX_LOCK(ctx);
791 	IFDI_INTR_DISABLE(ctx);
792 
793 	/* Tell the stack that the interface is no longer active */
794 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
795 
796 	if (!CTX_IS_VF(ctx))
797 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
798 
799 	/* enable or disable flags and callbacks in na and ifp */
800 	if (onoff) {
801 		nm_set_native_flags(na);
802 	} else {
803 		nm_clear_native_flags(na);
804 	}
805 	iflib_stop(ctx);
806 	iflib_init_locked(ctx);
807 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
808 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
809 	if (status)
810 		nm_clear_native_flags(na);
811 	CTX_UNLOCK(ctx);
812 	return (status);
813 }
814 
815 static int
816 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
817 {
818 	struct netmap_adapter *na = kring->na;
819 	u_int const lim = kring->nkr_num_slots - 1;
820 	u_int head = kring->rhead;
821 	struct netmap_ring *ring = kring->ring;
822 	bus_dmamap_t *map;
823 	struct if_rxd_update iru;
824 	if_ctx_t ctx = rxq->ifr_ctx;
825 	iflib_fl_t fl = &rxq->ifr_fl[0];
826 	uint32_t refill_pidx, nic_i;
827 #if IFLIB_DEBUG_COUNTERS
828 	int rf_count = 0;
829 #endif
830 
831 	if (nm_i == head && __predict_true(!init))
832 		return 0;
833 	iru_init(&iru, rxq, 0 /* flid */);
834 	map = fl->ifl_sds.ifsd_map;
835 	refill_pidx = netmap_idx_k2n(kring, nm_i);
836 	/*
837 	 * IMPORTANT: we must leave one free slot in the ring,
838 	 * so move head back by one unit
839 	 */
840 	head = nm_prev(head, lim);
841 	nic_i = UINT_MAX;
842 	DBG_COUNTER_INC(fl_refills);
843 	while (nm_i != head) {
844 #if IFLIB_DEBUG_COUNTERS
845 		if (++rf_count == 9)
846 			DBG_COUNTER_INC(fl_refills_large);
847 #endif
848 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
849 			struct netmap_slot *slot = &ring->slot[nm_i];
850 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
851 			uint32_t nic_i_dma = refill_pidx;
852 			nic_i = netmap_idx_k2n(kring, nm_i);
853 
854 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
855 
856 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
857 			        return netmap_ring_reinit(kring);
858 
859 			fl->ifl_vm_addrs[tmp_pidx] = addr;
860 			if (__predict_false(init)) {
861 				netmap_load_map(na, fl->ifl_buf_tag,
862 				    map[nic_i], addr);
863 			} else if (slot->flags & NS_BUF_CHANGED) {
864 				/* buffer has changed, reload map */
865 				netmap_reload_map(na, fl->ifl_buf_tag,
866 				    map[nic_i], addr);
867 			}
868 			slot->flags &= ~NS_BUF_CHANGED;
869 
870 			nm_i = nm_next(nm_i, lim);
871 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
872 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
873 				continue;
874 
875 			iru.iru_pidx = refill_pidx;
876 			iru.iru_count = tmp_pidx+1;
877 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
878 			refill_pidx = nic_i;
879 			for (int n = 0; n < iru.iru_count; n++) {
880 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
881 						BUS_DMASYNC_PREREAD);
882 				/* XXX - change this to not use the netmap func*/
883 				nic_i_dma = nm_next(nic_i_dma, lim);
884 			}
885 		}
886 	}
887 	kring->nr_hwcur = head;
888 
889 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
890 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
891 	if (__predict_true(nic_i != UINT_MAX)) {
892 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
893 		DBG_COUNTER_INC(rxd_flush);
894 	}
895 	return (0);
896 }
897 
898 /*
899  * Reconcile kernel and user view of the transmit ring.
900  *
901  * All information is in the kring.
902  * Userspace wants to send packets up to the one before kring->rhead,
903  * kernel knows kring->nr_hwcur is the first unsent packet.
904  *
905  * Here we push packets out (as many as possible), and possibly
906  * reclaim buffers from previously completed transmission.
907  *
908  * The caller (netmap) guarantees that there is only one instance
909  * running at any time. Any interference with other driver
910  * methods should be handled by the individual drivers.
911  */
912 static int
913 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
914 {
915 	struct netmap_adapter *na = kring->na;
916 	struct ifnet *ifp = na->ifp;
917 	struct netmap_ring *ring = kring->ring;
918 	u_int nm_i;	/* index into the netmap kring */
919 	u_int nic_i;	/* index into the NIC ring */
920 	u_int n;
921 	u_int const lim = kring->nkr_num_slots - 1;
922 	u_int const head = kring->rhead;
923 	struct if_pkt_info pi;
924 
925 	/*
926 	 * interrupts on every tx packet are expensive so request
927 	 * them every half ring, or where NS_REPORT is set
928 	 */
929 	u_int report_frequency = kring->nkr_num_slots >> 1;
930 	/* device-specific */
931 	if_ctx_t ctx = ifp->if_softc;
932 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
933 
934 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
935 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
936 
937 	/*
938 	 * First part: process new packets to send.
939 	 * nm_i is the current index in the netmap kring,
940 	 * nic_i is the corresponding index in the NIC ring.
941 	 *
942 	 * If we have packets to send (nm_i != head)
943 	 * iterate over the netmap ring, fetch length and update
944 	 * the corresponding slot in the NIC ring. Some drivers also
945 	 * need to update the buffer's physical address in the NIC slot
946 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
947 	 *
948 	 * The netmap_reload_map() calls is especially expensive,
949 	 * even when (as in this case) the tag is 0, so do only
950 	 * when the buffer has actually changed.
951 	 *
952 	 * If possible do not set the report/intr bit on all slots,
953 	 * but only a few times per ring or when NS_REPORT is set.
954 	 *
955 	 * Finally, on 10G and faster drivers, it might be useful
956 	 * to prefetch the next slot and txr entry.
957 	 */
958 
959 	nm_i = kring->nr_hwcur;
960 	if (nm_i != head) {	/* we have new packets to send */
961 		pkt_info_zero(&pi);
962 		pi.ipi_segs = txq->ift_segs;
963 		pi.ipi_qsidx = kring->ring_id;
964 		nic_i = netmap_idx_k2n(kring, nm_i);
965 
966 		__builtin_prefetch(&ring->slot[nm_i]);
967 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
968 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
969 
970 		for (n = 0; nm_i != head; n++) {
971 			struct netmap_slot *slot = &ring->slot[nm_i];
972 			u_int len = slot->len;
973 			uint64_t paddr;
974 			void *addr = PNMB(na, slot, &paddr);
975 			int flags = (slot->flags & NS_REPORT ||
976 				nic_i == 0 || nic_i == report_frequency) ?
977 				IPI_TX_INTR : 0;
978 
979 			/* device-specific */
980 			pi.ipi_len = len;
981 			pi.ipi_segs[0].ds_addr = paddr;
982 			pi.ipi_segs[0].ds_len = len;
983 			pi.ipi_nsegs = 1;
984 			pi.ipi_ndescs = 0;
985 			pi.ipi_pidx = nic_i;
986 			pi.ipi_flags = flags;
987 
988 			/* Fill the slot in the NIC ring. */
989 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
990 			DBG_COUNTER_INC(tx_encap);
991 
992 			/* prefetch for next round */
993 			__builtin_prefetch(&ring->slot[nm_i + 1]);
994 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
995 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
996 
997 			NM_CHECK_ADDR_LEN(na, addr, len);
998 
999 			if (slot->flags & NS_BUF_CHANGED) {
1000 				/* buffer has changed, reload map */
1001 				netmap_reload_map(na, txq->ift_buf_tag,
1002 				    txq->ift_sds.ifsd_map[nic_i], addr);
1003 			}
1004 			/* make sure changes to the buffer are synced */
1005 			bus_dmamap_sync(txq->ift_buf_tag,
1006 			    txq->ift_sds.ifsd_map[nic_i],
1007 			    BUS_DMASYNC_PREWRITE);
1008 
1009 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1010 			nm_i = nm_next(nm_i, lim);
1011 			nic_i = nm_next(nic_i, lim);
1012 		}
1013 		kring->nr_hwcur = nm_i;
1014 
1015 		/* synchronize the NIC ring */
1016 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1017 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1018 
1019 		/* (re)start the tx unit up to slot nic_i (excluded) */
1020 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1021 	}
1022 
1023 	/*
1024 	 * Second part: reclaim buffers for completed transmissions.
1025 	 *
1026 	 * If there are unclaimed buffers, attempt to reclaim them.
1027 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1028 	 * minimal delay, then trigger the tx handler which will spin in the
1029 	 * group task queue.
1030 	 */
1031 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1032 		if (iflib_tx_credits_update(ctx, txq)) {
1033 			/* some tx completed, increment avail */
1034 			nic_i = txq->ift_cidx_processed;
1035 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1036 		}
1037 	}
1038 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1039 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1040 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1041 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1042 	}
1043 	return (0);
1044 }
1045 
1046 /*
1047  * Reconcile kernel and user view of the receive ring.
1048  * Same as for the txsync, this routine must be efficient.
1049  * The caller guarantees a single invocations, but races against
1050  * the rest of the driver should be handled here.
1051  *
1052  * On call, kring->rhead is the first packet that userspace wants
1053  * to keep, and kring->rcur is the wakeup point.
1054  * The kernel has previously reported packets up to kring->rtail.
1055  *
1056  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1057  * of whether or not we received an interrupt.
1058  */
1059 static int
1060 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1061 {
1062 	struct netmap_adapter *na = kring->na;
1063 	struct netmap_ring *ring = kring->ring;
1064 	iflib_fl_t fl;
1065 	uint32_t nm_i;	/* index into the netmap ring */
1066 	uint32_t nic_i;	/* index into the NIC ring */
1067 	u_int i, n;
1068 	u_int const lim = kring->nkr_num_slots - 1;
1069 	u_int const head = kring->rhead;
1070 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1071 	struct if_rxd_info ri;
1072 
1073 	struct ifnet *ifp = na->ifp;
1074 	if_ctx_t ctx = ifp->if_softc;
1075 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1076 	if (head > lim)
1077 		return netmap_ring_reinit(kring);
1078 
1079 	/*
1080 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1081 	 */
1082 
1083 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1084 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1085 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1086 	}
1087 
1088 	/*
1089 	 * First part: import newly received packets.
1090 	 *
1091 	 * nm_i is the index of the next free slot in the netmap ring,
1092 	 * nic_i is the index of the next received packet in the NIC ring,
1093 	 * and they may differ in case if_init() has been called while
1094 	 * in netmap mode. For the receive ring we have
1095 	 *
1096 	 *	nic_i = rxr->next_check;
1097 	 *	nm_i = kring->nr_hwtail (previous)
1098 	 * and
1099 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1100 	 *
1101 	 * rxr->next_check is set to 0 on a ring reinit
1102 	 */
1103 	if (netmap_no_pendintr || force_update) {
1104 		int crclen = iflib_crcstrip ? 0 : 4;
1105 		int error, avail;
1106 
1107 		for (i = 0; i < rxq->ifr_nfl; i++) {
1108 			fl = &rxq->ifr_fl[i];
1109 			nic_i = fl->ifl_cidx;
1110 			nm_i = netmap_idx_n2k(kring, nic_i);
1111 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1112 			    rxq->ifr_id, nic_i, USHRT_MAX);
1113 			for (n = 0; avail > 0; n++, avail--) {
1114 				rxd_info_zero(&ri);
1115 				ri.iri_frags = rxq->ifr_frags;
1116 				ri.iri_qsidx = kring->ring_id;
1117 				ri.iri_ifp = ctx->ifc_ifp;
1118 				ri.iri_cidx = nic_i;
1119 
1120 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1121 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1122 				ring->slot[nm_i].flags = 0;
1123 				bus_dmamap_sync(fl->ifl_buf_tag,
1124 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1125 				nm_i = nm_next(nm_i, lim);
1126 				nic_i = nm_next(nic_i, lim);
1127 			}
1128 			if (n) { /* update the state variables */
1129 				if (netmap_no_pendintr && !force_update) {
1130 					/* diagnostics */
1131 					iflib_rx_miss ++;
1132 					iflib_rx_miss_bufs += n;
1133 				}
1134 				fl->ifl_cidx = nic_i;
1135 				kring->nr_hwtail = nm_i;
1136 			}
1137 			kring->nr_kflags &= ~NKR_PENDINTR;
1138 		}
1139 	}
1140 	/*
1141 	 * Second part: skip past packets that userspace has released.
1142 	 * (kring->nr_hwcur to head excluded),
1143 	 * and make the buffers available for reception.
1144 	 * As usual nm_i is the index in the netmap ring,
1145 	 * nic_i is the index in the NIC ring, and
1146 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1147 	 */
1148 	/* XXX not sure how this will work with multiple free lists */
1149 	nm_i = kring->nr_hwcur;
1150 
1151 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1152 }
1153 
1154 static void
1155 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1156 {
1157 	struct ifnet *ifp = na->ifp;
1158 	if_ctx_t ctx = ifp->if_softc;
1159 
1160 	CTX_LOCK(ctx);
1161 	if (onoff) {
1162 		IFDI_INTR_ENABLE(ctx);
1163 	} else {
1164 		IFDI_INTR_DISABLE(ctx);
1165 	}
1166 	CTX_UNLOCK(ctx);
1167 }
1168 
1169 
1170 static int
1171 iflib_netmap_attach(if_ctx_t ctx)
1172 {
1173 	struct netmap_adapter na;
1174 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1175 
1176 	bzero(&na, sizeof(na));
1177 
1178 	na.ifp = ctx->ifc_ifp;
1179 	na.na_flags = NAF_BDG_MAYSLEEP;
1180 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1181 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1182 
1183 	na.num_tx_desc = scctx->isc_ntxd[0];
1184 	na.num_rx_desc = scctx->isc_nrxd[0];
1185 	na.nm_txsync = iflib_netmap_txsync;
1186 	na.nm_rxsync = iflib_netmap_rxsync;
1187 	na.nm_register = iflib_netmap_register;
1188 	na.nm_intr = iflib_netmap_intr;
1189 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1190 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1191 	return (netmap_attach(&na));
1192 }
1193 
1194 static void
1195 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1196 {
1197 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1198 	struct netmap_slot *slot;
1199 
1200 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1201 	if (slot == NULL)
1202 		return;
1203 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1204 
1205 		/*
1206 		 * In netmap mode, set the map for the packet buffer.
1207 		 * NOTE: Some drivers (not this one) also need to set
1208 		 * the physical buffer address in the NIC ring.
1209 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1210 		 * netmap slot index, si
1211 		 */
1212 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1213 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1214 		    NMB(na, slot + si));
1215 	}
1216 }
1217 
1218 static void
1219 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1220 {
1221 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1222 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1223 	struct netmap_slot *slot;
1224 	uint32_t nm_i;
1225 
1226 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1227 	if (slot == NULL)
1228 		return;
1229 	nm_i = netmap_idx_n2k(kring, 0);
1230 	netmap_fl_refill(rxq, kring, nm_i, true);
1231 }
1232 
1233 static void
1234 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1235 {
1236 	struct netmap_kring *kring;
1237 	uint16_t txqid;
1238 
1239 	txqid = txq->ift_id;
1240 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1241 
1242 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1243 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1244 		    BUS_DMASYNC_POSTREAD);
1245 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1246 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1247 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1248 			if (hz < 2000)
1249 				*reset_on = 1;
1250 			else
1251 				*reset_on = hz / 1000;
1252 		}
1253 	}
1254 }
1255 
1256 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1257 
1258 #else
1259 #define iflib_netmap_txq_init(ctx, txq)
1260 #define iflib_netmap_rxq_init(ctx, rxq)
1261 #define iflib_netmap_detach(ifp)
1262 
1263 #define iflib_netmap_attach(ctx) (0)
1264 #define netmap_rx_irq(ifp, qid, budget) (0)
1265 #define netmap_tx_irq(ifp, qid) do {} while (0)
1266 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1267 
1268 #endif
1269 
1270 #if defined(__i386__) || defined(__amd64__)
1271 static __inline void
1272 prefetch(void *x)
1273 {
1274 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1275 }
1276 static __inline void
1277 prefetch2cachelines(void *x)
1278 {
1279 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1280 #if (CACHE_LINE_SIZE < 128)
1281 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1282 #endif
1283 }
1284 #else
1285 #define prefetch(x)
1286 #define prefetch2cachelines(x)
1287 #endif
1288 
1289 static void
1290 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1291 {
1292 	iflib_fl_t fl;
1293 
1294 	fl = &rxq->ifr_fl[flid];
1295 	iru->iru_paddrs = fl->ifl_bus_addrs;
1296 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1297 	iru->iru_idxs = fl->ifl_rxd_idxs;
1298 	iru->iru_qsidx = rxq->ifr_id;
1299 	iru->iru_buf_size = fl->ifl_buf_size;
1300 	iru->iru_flidx = fl->ifl_id;
1301 }
1302 
1303 static void
1304 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1305 {
1306 	if (err)
1307 		return;
1308 	*(bus_addr_t *) arg = segs[0].ds_addr;
1309 }
1310 
1311 int
1312 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1313 {
1314 	int err;
1315 	device_t dev = ctx->ifc_dev;
1316 
1317 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1318 				align, 0,		/* alignment, bounds */
1319 				BUS_SPACE_MAXADDR,	/* lowaddr */
1320 				BUS_SPACE_MAXADDR,	/* highaddr */
1321 				NULL, NULL,		/* filter, filterarg */
1322 				size,			/* maxsize */
1323 				1,			/* nsegments */
1324 				size,			/* maxsegsize */
1325 				BUS_DMA_ALLOCNOW,	/* flags */
1326 				NULL,			/* lockfunc */
1327 				NULL,			/* lockarg */
1328 				&dma->idi_tag);
1329 	if (err) {
1330 		device_printf(dev,
1331 		    "%s: bus_dma_tag_create failed: %d\n",
1332 		    __func__, err);
1333 		goto fail_0;
1334 	}
1335 
1336 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1337 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1338 	if (err) {
1339 		device_printf(dev,
1340 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1341 		    __func__, (uintmax_t)size, err);
1342 		goto fail_1;
1343 	}
1344 
1345 	dma->idi_paddr = IF_BAD_DMA;
1346 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1347 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1348 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1349 		device_printf(dev,
1350 		    "%s: bus_dmamap_load failed: %d\n",
1351 		    __func__, err);
1352 		goto fail_2;
1353 	}
1354 
1355 	dma->idi_size = size;
1356 	return (0);
1357 
1358 fail_2:
1359 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1360 fail_1:
1361 	bus_dma_tag_destroy(dma->idi_tag);
1362 fail_0:
1363 	dma->idi_tag = NULL;
1364 
1365 	return (err);
1366 }
1367 
1368 int
1369 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1370 {
1371 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1372 
1373 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1374 
1375 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1376 }
1377 
1378 int
1379 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1380 {
1381 	int i, err;
1382 	iflib_dma_info_t *dmaiter;
1383 
1384 	dmaiter = dmalist;
1385 	for (i = 0; i < count; i++, dmaiter++) {
1386 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1387 			break;
1388 	}
1389 	if (err)
1390 		iflib_dma_free_multi(dmalist, i);
1391 	return (err);
1392 }
1393 
1394 void
1395 iflib_dma_free(iflib_dma_info_t dma)
1396 {
1397 	if (dma->idi_tag == NULL)
1398 		return;
1399 	if (dma->idi_paddr != IF_BAD_DMA) {
1400 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1401 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1402 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1403 		dma->idi_paddr = IF_BAD_DMA;
1404 	}
1405 	if (dma->idi_vaddr != NULL) {
1406 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1407 		dma->idi_vaddr = NULL;
1408 	}
1409 	bus_dma_tag_destroy(dma->idi_tag);
1410 	dma->idi_tag = NULL;
1411 }
1412 
1413 void
1414 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1415 {
1416 	int i;
1417 	iflib_dma_info_t *dmaiter = dmalist;
1418 
1419 	for (i = 0; i < count; i++, dmaiter++)
1420 		iflib_dma_free(*dmaiter);
1421 }
1422 
1423 #ifdef EARLY_AP_STARTUP
1424 static const int iflib_started = 1;
1425 #else
1426 /*
1427  * We used to abuse the smp_started flag to decide if the queues have been
1428  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1429  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1430  * is set.  Run a SYSINIT() strictly after that to just set a usable
1431  * completion flag.
1432  */
1433 
1434 static int iflib_started;
1435 
1436 static void
1437 iflib_record_started(void *arg)
1438 {
1439 	iflib_started = 1;
1440 }
1441 
1442 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1443 	iflib_record_started, NULL);
1444 #endif
1445 
1446 static int
1447 iflib_fast_intr(void *arg)
1448 {
1449 	iflib_filter_info_t info = arg;
1450 	struct grouptask *gtask = info->ifi_task;
1451 	int result;
1452 
1453 	if (!iflib_started)
1454 		return (FILTER_STRAY);
1455 
1456 	DBG_COUNTER_INC(fast_intrs);
1457 	if (info->ifi_filter != NULL) {
1458 		result = info->ifi_filter(info->ifi_filter_arg);
1459 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1460 			return (result);
1461 	}
1462 
1463 	GROUPTASK_ENQUEUE(gtask);
1464 	return (FILTER_HANDLED);
1465 }
1466 
1467 static int
1468 iflib_fast_intr_rxtx(void *arg)
1469 {
1470 	iflib_filter_info_t info = arg;
1471 	struct grouptask *gtask = info->ifi_task;
1472 	if_ctx_t ctx;
1473 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1474 	iflib_txq_t txq;
1475 	void *sc;
1476 	int i, cidx, result;
1477 	qidx_t txqid;
1478 
1479 	if (!iflib_started)
1480 		return (FILTER_STRAY);
1481 
1482 	DBG_COUNTER_INC(fast_intrs);
1483 	if (info->ifi_filter != NULL) {
1484 		result = info->ifi_filter(info->ifi_filter_arg);
1485 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1486 			return (result);
1487 	}
1488 
1489 	ctx = rxq->ifr_ctx;
1490 	sc = ctx->ifc_softc;
1491 	MPASS(rxq->ifr_ntxqirq);
1492 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1493 		txqid = rxq->ifr_txqid[i];
1494 		txq = &ctx->ifc_txqs[txqid];
1495 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1496 		    BUS_DMASYNC_POSTREAD);
1497 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1498 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1499 			continue;
1500 		}
1501 		GROUPTASK_ENQUEUE(&txq->ift_task);
1502 	}
1503 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1504 		cidx = rxq->ifr_cq_cidx;
1505 	else
1506 		cidx = rxq->ifr_fl[0].ifl_cidx;
1507 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1508 		GROUPTASK_ENQUEUE(gtask);
1509 	else {
1510 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1511 		DBG_COUNTER_INC(rx_intr_enables);
1512 	}
1513 	return (FILTER_HANDLED);
1514 }
1515 
1516 
1517 static int
1518 iflib_fast_intr_ctx(void *arg)
1519 {
1520 	iflib_filter_info_t info = arg;
1521 	struct grouptask *gtask = info->ifi_task;
1522 	int result;
1523 
1524 	if (!iflib_started)
1525 		return (FILTER_STRAY);
1526 
1527 	DBG_COUNTER_INC(fast_intrs);
1528 	if (info->ifi_filter != NULL) {
1529 		result = info->ifi_filter(info->ifi_filter_arg);
1530 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1531 			return (result);
1532 	}
1533 
1534 	GROUPTASK_ENQUEUE(gtask);
1535 	return (FILTER_HANDLED);
1536 }
1537 
1538 static int
1539 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1540 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1541 		 const char *name)
1542 {
1543 	int rc, flags;
1544 	struct resource *res;
1545 	void *tag = NULL;
1546 	device_t dev = ctx->ifc_dev;
1547 
1548 	flags = RF_ACTIVE;
1549 	if (ctx->ifc_flags & IFC_LEGACY)
1550 		flags |= RF_SHAREABLE;
1551 	MPASS(rid < 512);
1552 	irq->ii_rid = rid;
1553 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1554 	if (res == NULL) {
1555 		device_printf(dev,
1556 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1557 		return (ENOMEM);
1558 	}
1559 	irq->ii_res = res;
1560 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1561 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1562 						filter, handler, arg, &tag);
1563 	if (rc != 0) {
1564 		device_printf(dev,
1565 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1566 					  rid, name ? name : "unknown", rc);
1567 		return (rc);
1568 	} else if (name)
1569 		bus_describe_intr(dev, res, tag, "%s", name);
1570 
1571 	irq->ii_tag = tag;
1572 	return (0);
1573 }
1574 
1575 
1576 /*********************************************************************
1577  *
1578  *  Allocate DMA resources for TX buffers as well as memory for the TX
1579  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1580  *  iflib_sw_tx_desc_array structure, storing all the information that
1581  *  is needed to transmit a packet on the wire.  This is called only
1582  *  once at attach, setup is done every reset.
1583  *
1584  **********************************************************************/
1585 static int
1586 iflib_txsd_alloc(iflib_txq_t txq)
1587 {
1588 	if_ctx_t ctx = txq->ift_ctx;
1589 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1590 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1591 	device_t dev = ctx->ifc_dev;
1592 	bus_size_t tsomaxsize;
1593 	int err, nsegments, ntsosegments;
1594 	bool tso;
1595 
1596 	nsegments = scctx->isc_tx_nsegments;
1597 	ntsosegments = scctx->isc_tx_tso_segments_max;
1598 	tsomaxsize = scctx->isc_tx_tso_size_max;
1599 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1600 		tsomaxsize += sizeof(struct ether_vlan_header);
1601 	MPASS(scctx->isc_ntxd[0] > 0);
1602 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1603 	MPASS(nsegments > 0);
1604 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1605 		MPASS(ntsosegments > 0);
1606 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1607 	}
1608 
1609 	/*
1610 	 * Set up DMA tags for TX buffers.
1611 	 */
1612 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1613 			       1, 0,			/* alignment, bounds */
1614 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1615 			       BUS_SPACE_MAXADDR,	/* highaddr */
1616 			       NULL, NULL,		/* filter, filterarg */
1617 			       sctx->isc_tx_maxsize,		/* maxsize */
1618 			       nsegments,	/* nsegments */
1619 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1620 			       0,			/* flags */
1621 			       NULL,			/* lockfunc */
1622 			       NULL,			/* lockfuncarg */
1623 			       &txq->ift_buf_tag))) {
1624 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1625 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1626 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1627 		goto fail;
1628 	}
1629 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1630 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1631 			       1, 0,			/* alignment, bounds */
1632 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1633 			       BUS_SPACE_MAXADDR,	/* highaddr */
1634 			       NULL, NULL,		/* filter, filterarg */
1635 			       tsomaxsize,		/* maxsize */
1636 			       ntsosegments,	/* nsegments */
1637 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1638 			       0,			/* flags */
1639 			       NULL,			/* lockfunc */
1640 			       NULL,			/* lockfuncarg */
1641 			       &txq->ift_tso_buf_tag))) {
1642 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1643 		    err);
1644 		goto fail;
1645 	}
1646 
1647 	/* Allocate memory for the TX mbuf map. */
1648 	if (!(txq->ift_sds.ifsd_m =
1649 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1650 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1651 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1652 		err = ENOMEM;
1653 		goto fail;
1654 	}
1655 
1656 	/*
1657 	 * Create the DMA maps for TX buffers.
1658 	 */
1659 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1660 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1661 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1662 		device_printf(dev,
1663 		    "Unable to allocate TX buffer DMA map memory\n");
1664 		err = ENOMEM;
1665 		goto fail;
1666 	}
1667 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1668 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1669 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1670 		device_printf(dev,
1671 		    "Unable to allocate TSO TX buffer map memory\n");
1672 		err = ENOMEM;
1673 		goto fail;
1674 	}
1675 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1676 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1677 		    &txq->ift_sds.ifsd_map[i]);
1678 		if (err != 0) {
1679 			device_printf(dev, "Unable to create TX DMA map\n");
1680 			goto fail;
1681 		}
1682 		if (!tso)
1683 			continue;
1684 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1685 		    &txq->ift_sds.ifsd_tso_map[i]);
1686 		if (err != 0) {
1687 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1688 			goto fail;
1689 		}
1690 	}
1691 	return (0);
1692 fail:
1693 	/* We free all, it handles case where we are in the middle */
1694 	iflib_tx_structures_free(ctx);
1695 	return (err);
1696 }
1697 
1698 static void
1699 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1700 {
1701 	bus_dmamap_t map;
1702 
1703 	map = NULL;
1704 	if (txq->ift_sds.ifsd_map != NULL)
1705 		map = txq->ift_sds.ifsd_map[i];
1706 	if (map != NULL) {
1707 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1708 		bus_dmamap_unload(txq->ift_buf_tag, map);
1709 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1710 		txq->ift_sds.ifsd_map[i] = NULL;
1711 	}
1712 
1713 	map = NULL;
1714 	if (txq->ift_sds.ifsd_tso_map != NULL)
1715 		map = txq->ift_sds.ifsd_tso_map[i];
1716 	if (map != NULL) {
1717 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1718 		    BUS_DMASYNC_POSTWRITE);
1719 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1720 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1721 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1722 	}
1723 }
1724 
1725 static void
1726 iflib_txq_destroy(iflib_txq_t txq)
1727 {
1728 	if_ctx_t ctx = txq->ift_ctx;
1729 
1730 	for (int i = 0; i < txq->ift_size; i++)
1731 		iflib_txsd_destroy(ctx, txq, i);
1732 	if (txq->ift_sds.ifsd_map != NULL) {
1733 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1734 		txq->ift_sds.ifsd_map = NULL;
1735 	}
1736 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1737 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1738 		txq->ift_sds.ifsd_tso_map = NULL;
1739 	}
1740 	if (txq->ift_sds.ifsd_m != NULL) {
1741 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1742 		txq->ift_sds.ifsd_m = NULL;
1743 	}
1744 	if (txq->ift_buf_tag != NULL) {
1745 		bus_dma_tag_destroy(txq->ift_buf_tag);
1746 		txq->ift_buf_tag = NULL;
1747 	}
1748 	if (txq->ift_tso_buf_tag != NULL) {
1749 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1750 		txq->ift_tso_buf_tag = NULL;
1751 	}
1752 }
1753 
1754 static void
1755 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1756 {
1757 	struct mbuf **mp;
1758 
1759 	mp = &txq->ift_sds.ifsd_m[i];
1760 	if (*mp == NULL)
1761 		return;
1762 
1763 	if (txq->ift_sds.ifsd_map != NULL) {
1764 		bus_dmamap_sync(txq->ift_buf_tag,
1765 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1766 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1767 	}
1768 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1769 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1770 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1771 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1772 		    txq->ift_sds.ifsd_tso_map[i]);
1773 	}
1774 	m_free(*mp);
1775 	DBG_COUNTER_INC(tx_frees);
1776 	*mp = NULL;
1777 }
1778 
1779 static int
1780 iflib_txq_setup(iflib_txq_t txq)
1781 {
1782 	if_ctx_t ctx = txq->ift_ctx;
1783 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1784 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1785 	iflib_dma_info_t di;
1786 	int i;
1787 
1788 	/* Set number of descriptors available */
1789 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1790 	/* XXX make configurable */
1791 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1792 
1793 	/* Reset indices */
1794 	txq->ift_cidx_processed = 0;
1795 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1796 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1797 
1798 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1799 		bzero((void *)di->idi_vaddr, di->idi_size);
1800 
1801 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1802 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1803 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1804 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1805 	return (0);
1806 }
1807 
1808 /*********************************************************************
1809  *
1810  *  Allocate DMA resources for RX buffers as well as memory for the RX
1811  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1812  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1813  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1814  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1815  *  packet, the maximum number of entries we'll need is equal to the
1816  *  number of hardware receive descriptors that we've allocated.
1817  *
1818  **********************************************************************/
1819 static int
1820 iflib_rxsd_alloc(iflib_rxq_t rxq)
1821 {
1822 	if_ctx_t ctx = rxq->ifr_ctx;
1823 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1824 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1825 	device_t dev = ctx->ifc_dev;
1826 	iflib_fl_t fl;
1827 	int			err;
1828 
1829 	MPASS(scctx->isc_nrxd[0] > 0);
1830 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1831 
1832 	fl = rxq->ifr_fl;
1833 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1834 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1835 		/* Set up DMA tag for RX buffers. */
1836 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1837 					 1, 0,			/* alignment, bounds */
1838 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1839 					 BUS_SPACE_MAXADDR,	/* highaddr */
1840 					 NULL, NULL,		/* filter, filterarg */
1841 					 sctx->isc_rx_maxsize,	/* maxsize */
1842 					 sctx->isc_rx_nsegments,	/* nsegments */
1843 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1844 					 0,			/* flags */
1845 					 NULL,			/* lockfunc */
1846 					 NULL,			/* lockarg */
1847 					 &fl->ifl_buf_tag);
1848 		if (err) {
1849 			device_printf(dev,
1850 			    "Unable to allocate RX DMA tag: %d\n", err);
1851 			goto fail;
1852 		}
1853 
1854 		/* Allocate memory for the RX mbuf map. */
1855 		if (!(fl->ifl_sds.ifsd_m =
1856 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1857 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1858 			device_printf(dev,
1859 			    "Unable to allocate RX mbuf map memory\n");
1860 			err = ENOMEM;
1861 			goto fail;
1862 		}
1863 
1864 		/* Allocate memory for the direct RX cluster pointer map. */
1865 		if (!(fl->ifl_sds.ifsd_cl =
1866 		      (caddr_t *) malloc(sizeof(caddr_t) *
1867 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1868 			device_printf(dev,
1869 			    "Unable to allocate RX cluster map memory\n");
1870 			err = ENOMEM;
1871 			goto fail;
1872 		}
1873 
1874 		/* Allocate memory for the RX cluster bus address map. */
1875 		if (!(fl->ifl_sds.ifsd_ba =
1876 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1877 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1878 			device_printf(dev,
1879 			    "Unable to allocate RX bus address map memory\n");
1880 			err = ENOMEM;
1881 			goto fail;
1882 		}
1883 
1884 		/*
1885 		 * Create the DMA maps for RX buffers.
1886 		 */
1887 		if (!(fl->ifl_sds.ifsd_map =
1888 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1889 			device_printf(dev,
1890 			    "Unable to allocate RX buffer DMA map memory\n");
1891 			err = ENOMEM;
1892 			goto fail;
1893 		}
1894 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1895 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1896 			    &fl->ifl_sds.ifsd_map[i]);
1897 			if (err != 0) {
1898 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1899 				goto fail;
1900 			}
1901 		}
1902 	}
1903 	return (0);
1904 
1905 fail:
1906 	iflib_rx_structures_free(ctx);
1907 	return (err);
1908 }
1909 
1910 
1911 /*
1912  * Internal service routines
1913  */
1914 
1915 struct rxq_refill_cb_arg {
1916 	int               error;
1917 	bus_dma_segment_t seg;
1918 	int               nseg;
1919 };
1920 
1921 static void
1922 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1923 {
1924 	struct rxq_refill_cb_arg *cb_arg = arg;
1925 
1926 	cb_arg->error = error;
1927 	cb_arg->seg = segs[0];
1928 	cb_arg->nseg = nseg;
1929 }
1930 
1931 /**
1932  *	rxq_refill - refill an rxq  free-buffer list
1933  *	@ctx: the iflib context
1934  *	@rxq: the free-list to refill
1935  *	@n: the number of new buffers to allocate
1936  *
1937  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1938  *	The caller must assure that @n does not exceed the queue's capacity.
1939  */
1940 static void
1941 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1942 {
1943 	struct if_rxd_update iru;
1944 	struct rxq_refill_cb_arg cb_arg;
1945 	struct mbuf *m;
1946 	caddr_t cl, *sd_cl;
1947 	struct mbuf **sd_m;
1948 	bus_dmamap_t *sd_map;
1949 	bus_addr_t bus_addr, *sd_ba;
1950 	int err, frag_idx, i, idx, n, pidx;
1951 	qidx_t credits;
1952 
1953 	sd_m = fl->ifl_sds.ifsd_m;
1954 	sd_map = fl->ifl_sds.ifsd_map;
1955 	sd_cl = fl->ifl_sds.ifsd_cl;
1956 	sd_ba = fl->ifl_sds.ifsd_ba;
1957 	pidx = fl->ifl_pidx;
1958 	idx = pidx;
1959 	frag_idx = fl->ifl_fragidx;
1960 	credits = fl->ifl_credits;
1961 
1962 	i = 0;
1963 	n = count;
1964 	MPASS(n > 0);
1965 	MPASS(credits + n <= fl->ifl_size);
1966 
1967 	if (pidx < fl->ifl_cidx)
1968 		MPASS(pidx + n <= fl->ifl_cidx);
1969 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1970 		MPASS(fl->ifl_gen == 0);
1971 	if (pidx > fl->ifl_cidx)
1972 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1973 
1974 	DBG_COUNTER_INC(fl_refills);
1975 	if (n > 8)
1976 		DBG_COUNTER_INC(fl_refills_large);
1977 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1978 	while (n--) {
1979 		/*
1980 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1981 		 * initialized after rx.
1982 		 *
1983 		 * If the cluster is still set then we know a minimum sized packet was received
1984 		 */
1985 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1986 		    &frag_idx);
1987 		if (frag_idx < 0)
1988 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1989 		MPASS(frag_idx >= 0);
1990 		if ((cl = sd_cl[frag_idx]) == NULL) {
1991 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1992 				break;
1993 
1994 			cb_arg.error = 0;
1995 			MPASS(sd_map != NULL);
1996 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1997 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1998 			    BUS_DMA_NOWAIT);
1999 			if (err != 0 || cb_arg.error) {
2000 				/*
2001 				 * !zone_pack ?
2002 				 */
2003 				if (fl->ifl_zone == zone_pack)
2004 					uma_zfree(fl->ifl_zone, cl);
2005 				break;
2006 			}
2007 
2008 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2009 			sd_cl[frag_idx] = cl;
2010 #if MEMORY_LOGGING
2011 			fl->ifl_cl_enqueued++;
2012 #endif
2013 		} else {
2014 			bus_addr = sd_ba[frag_idx];
2015 		}
2016 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2017 		    BUS_DMASYNC_PREREAD);
2018 
2019 		if (sd_m[frag_idx] == NULL) {
2020 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2021 				break;
2022 			}
2023 			sd_m[frag_idx] = m;
2024 		}
2025 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2026 #if MEMORY_LOGGING
2027 		fl->ifl_m_enqueued++;
2028 #endif
2029 
2030 		DBG_COUNTER_INC(rx_allocs);
2031 		fl->ifl_rxd_idxs[i] = frag_idx;
2032 		fl->ifl_bus_addrs[i] = bus_addr;
2033 		fl->ifl_vm_addrs[i] = cl;
2034 		credits++;
2035 		i++;
2036 		MPASS(credits <= fl->ifl_size);
2037 		if (++idx == fl->ifl_size) {
2038 			fl->ifl_gen = 1;
2039 			idx = 0;
2040 		}
2041 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2042 			iru.iru_pidx = pidx;
2043 			iru.iru_count = i;
2044 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2045 			i = 0;
2046 			pidx = idx;
2047 			fl->ifl_pidx = idx;
2048 			fl->ifl_credits = credits;
2049 		}
2050 	}
2051 
2052 	if (i) {
2053 		iru.iru_pidx = pidx;
2054 		iru.iru_count = i;
2055 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2056 		fl->ifl_pidx = idx;
2057 		fl->ifl_credits = credits;
2058 	}
2059 	DBG_COUNTER_INC(rxd_flush);
2060 	if (fl->ifl_pidx == 0)
2061 		pidx = fl->ifl_size - 1;
2062 	else
2063 		pidx = fl->ifl_pidx - 1;
2064 
2065 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2066 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2067 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2068 	fl->ifl_fragidx = frag_idx;
2069 }
2070 
2071 static __inline void
2072 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2073 {
2074 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2075 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2076 #ifdef INVARIANTS
2077 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2078 #endif
2079 
2080 	MPASS(fl->ifl_credits <= fl->ifl_size);
2081 	MPASS(reclaimable == delta);
2082 
2083 	if (reclaimable > 0)
2084 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2085 }
2086 
2087 uint8_t
2088 iflib_in_detach(if_ctx_t ctx)
2089 {
2090 	bool in_detach;
2091 	STATE_LOCK(ctx);
2092 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2093 	STATE_UNLOCK(ctx);
2094 	return (in_detach);
2095 }
2096 
2097 static void
2098 iflib_fl_bufs_free(iflib_fl_t fl)
2099 {
2100 	iflib_dma_info_t idi = fl->ifl_ifdi;
2101 	bus_dmamap_t sd_map;
2102 	uint32_t i;
2103 
2104 	for (i = 0; i < fl->ifl_size; i++) {
2105 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2106 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2107 
2108 		if (*sd_cl != NULL) {
2109 			sd_map = fl->ifl_sds.ifsd_map[i];
2110 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2111 			    BUS_DMASYNC_POSTREAD);
2112 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2113 			if (*sd_cl != NULL)
2114 				uma_zfree(fl->ifl_zone, *sd_cl);
2115 			// XXX: Should this get moved out?
2116 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2117 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2118 			if (*sd_m != NULL) {
2119 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2120 				uma_zfree(zone_mbuf, *sd_m);
2121 			}
2122 		} else {
2123 			MPASS(*sd_cl == NULL);
2124 			MPASS(*sd_m == NULL);
2125 		}
2126 #if MEMORY_LOGGING
2127 		fl->ifl_m_dequeued++;
2128 		fl->ifl_cl_dequeued++;
2129 #endif
2130 		*sd_cl = NULL;
2131 		*sd_m = NULL;
2132 	}
2133 #ifdef INVARIANTS
2134 	for (i = 0; i < fl->ifl_size; i++) {
2135 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2136 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2137 	}
2138 #endif
2139 	/*
2140 	 * Reset free list values
2141 	 */
2142 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2143 	bzero(idi->idi_vaddr, idi->idi_size);
2144 }
2145 
2146 /*********************************************************************
2147  *
2148  *  Initialize a receive ring and its buffers.
2149  *
2150  **********************************************************************/
2151 static int
2152 iflib_fl_setup(iflib_fl_t fl)
2153 {
2154 	iflib_rxq_t rxq = fl->ifl_rxq;
2155 	if_ctx_t ctx = rxq->ifr_ctx;
2156 
2157 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2158 	/*
2159 	** Free current RX buffer structs and their mbufs
2160 	*/
2161 	iflib_fl_bufs_free(fl);
2162 	/* Now replenish the mbufs */
2163 	MPASS(fl->ifl_credits == 0);
2164 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2165 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2166 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2167 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2168 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2169 
2170 
2171 	/* avoid pre-allocating zillions of clusters to an idle card
2172 	 * potentially speeding up attach
2173 	 */
2174 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2175 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2176 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2177 		return (ENOBUFS);
2178 	/*
2179 	 * handle failure
2180 	 */
2181 	MPASS(rxq != NULL);
2182 	MPASS(fl->ifl_ifdi != NULL);
2183 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2184 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2185 	return (0);
2186 }
2187 
2188 /*********************************************************************
2189  *
2190  *  Free receive ring data structures
2191  *
2192  **********************************************************************/
2193 static void
2194 iflib_rx_sds_free(iflib_rxq_t rxq)
2195 {
2196 	iflib_fl_t fl;
2197 	int i, j;
2198 
2199 	if (rxq->ifr_fl != NULL) {
2200 		for (i = 0; i < rxq->ifr_nfl; i++) {
2201 			fl = &rxq->ifr_fl[i];
2202 			if (fl->ifl_buf_tag != NULL) {
2203 				if (fl->ifl_sds.ifsd_map != NULL) {
2204 					for (j = 0; j < fl->ifl_size; j++) {
2205 						if (fl->ifl_sds.ifsd_map[j] ==
2206 						    NULL)
2207 							continue;
2208 						bus_dmamap_sync(
2209 						    fl->ifl_buf_tag,
2210 						    fl->ifl_sds.ifsd_map[j],
2211 						    BUS_DMASYNC_POSTREAD);
2212 						bus_dmamap_unload(
2213 						    fl->ifl_buf_tag,
2214 						    fl->ifl_sds.ifsd_map[j]);
2215 					}
2216 				}
2217 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2218 				fl->ifl_buf_tag = NULL;
2219 			}
2220 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2221 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2222 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2223 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2224 			fl->ifl_sds.ifsd_m = NULL;
2225 			fl->ifl_sds.ifsd_cl = NULL;
2226 			fl->ifl_sds.ifsd_ba = NULL;
2227 			fl->ifl_sds.ifsd_map = NULL;
2228 		}
2229 		free(rxq->ifr_fl, M_IFLIB);
2230 		rxq->ifr_fl = NULL;
2231 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2232 	}
2233 }
2234 
2235 /*
2236  * MI independent logic
2237  *
2238  */
2239 static void
2240 iflib_timer(void *arg)
2241 {
2242 	iflib_txq_t txq = arg;
2243 	if_ctx_t ctx = txq->ift_ctx;
2244 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2245 	uint64_t this_tick = ticks;
2246 	uint32_t reset_on = hz / 2;
2247 
2248 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2249 		return;
2250 	/*
2251 	** Check on the state of the TX queue(s), this
2252 	** can be done without the lock because its RO
2253 	** and the HUNG state will be static if set.
2254 	*/
2255 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2256 		txq->ift_last_timer_tick = this_tick;
2257 		IFDI_TIMER(ctx, txq->ift_id);
2258 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2259 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2260 		     (sctx->isc_pause_frames == 0)))
2261 			goto hung;
2262 
2263 		if (ifmp_ring_is_stalled(txq->ift_br))
2264 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2265 		txq->ift_cleaned_prev = txq->ift_cleaned;
2266 	}
2267 #ifdef DEV_NETMAP
2268 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2269 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2270 #endif
2271 	/* handle any laggards */
2272 	if (txq->ift_db_pending)
2273 		GROUPTASK_ENQUEUE(&txq->ift_task);
2274 
2275 	sctx->isc_pause_frames = 0;
2276 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2277 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2278 	return;
2279  hung:
2280 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2281 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2282 	STATE_LOCK(ctx);
2283 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2284 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2285 	iflib_admin_intr_deferred(ctx);
2286 	STATE_UNLOCK(ctx);
2287 }
2288 
2289 static void
2290 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2291 {
2292 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2293 
2294 	/*
2295 	 * XXX don't set the max_frame_size to larger
2296 	 * than the hardware can handle
2297 	 */
2298 	if (sctx->isc_max_frame_size <= MCLBYTES)
2299 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2300 	else
2301 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2302 }
2303 
2304 uint32_t
2305 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2306 {
2307 	return (ctx->ifc_rx_mbuf_sz);
2308 }
2309 
2310 static void
2311 iflib_init_locked(if_ctx_t ctx)
2312 {
2313 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2314 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2315 	if_t ifp = ctx->ifc_ifp;
2316 	iflib_fl_t fl;
2317 	iflib_txq_t txq;
2318 	iflib_rxq_t rxq;
2319 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2320 
2321 
2322 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2323 	IFDI_INTR_DISABLE(ctx);
2324 
2325 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2326 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2327 	/* Set hardware offload abilities */
2328 	if_clearhwassist(ifp);
2329 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2330 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2331 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2332 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2333 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2334 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2335 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2336 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2337 
2338 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2339 		CALLOUT_LOCK(txq);
2340 		callout_stop(&txq->ift_timer);
2341 		CALLOUT_UNLOCK(txq);
2342 		iflib_netmap_txq_init(ctx, txq);
2343 	}
2344 
2345 	/*
2346 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2347 	 * that drivers can use the value when setting up the hardware receive
2348 	 * buffers.
2349 	 */
2350 	iflib_calc_rx_mbuf_sz(ctx);
2351 
2352 #ifdef INVARIANTS
2353 	i = if_getdrvflags(ifp);
2354 #endif
2355 	IFDI_INIT(ctx);
2356 	MPASS(if_getdrvflags(ifp) == i);
2357 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2358 		/* XXX this should really be done on a per-queue basis */
2359 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2360 			MPASS(rxq->ifr_id == i);
2361 			iflib_netmap_rxq_init(ctx, rxq);
2362 			continue;
2363 		}
2364 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2365 			if (iflib_fl_setup(fl)) {
2366 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2367 				goto done;
2368 			}
2369 		}
2370 	}
2371 done:
2372 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2373 	IFDI_INTR_ENABLE(ctx);
2374 	txq = ctx->ifc_txqs;
2375 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2376 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2377 			txq->ift_timer.c_cpu);
2378 }
2379 
2380 static int
2381 iflib_media_change(if_t ifp)
2382 {
2383 	if_ctx_t ctx = if_getsoftc(ifp);
2384 	int err;
2385 
2386 	CTX_LOCK(ctx);
2387 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2388 		iflib_init_locked(ctx);
2389 	CTX_UNLOCK(ctx);
2390 	return (err);
2391 }
2392 
2393 static void
2394 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2395 {
2396 	if_ctx_t ctx = if_getsoftc(ifp);
2397 
2398 	CTX_LOCK(ctx);
2399 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2400 	IFDI_MEDIA_STATUS(ctx, ifmr);
2401 	CTX_UNLOCK(ctx);
2402 }
2403 
2404 void
2405 iflib_stop(if_ctx_t ctx)
2406 {
2407 	iflib_txq_t txq = ctx->ifc_txqs;
2408 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2409 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2410 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2411 	iflib_dma_info_t di;
2412 	iflib_fl_t fl;
2413 	int i, j;
2414 
2415 	/* Tell the stack that the interface is no longer active */
2416 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2417 
2418 	IFDI_INTR_DISABLE(ctx);
2419 	DELAY(1000);
2420 	IFDI_STOP(ctx);
2421 	DELAY(1000);
2422 
2423 	iflib_debug_reset();
2424 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2425 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2426 		/* make sure all transmitters have completed before proceeding XXX */
2427 
2428 		CALLOUT_LOCK(txq);
2429 		callout_stop(&txq->ift_timer);
2430 		CALLOUT_UNLOCK(txq);
2431 
2432 		/* clean any enqueued buffers */
2433 		iflib_ifmp_purge(txq);
2434 		/* Free any existing tx buffers. */
2435 		for (j = 0; j < txq->ift_size; j++) {
2436 			iflib_txsd_free(ctx, txq, j);
2437 		}
2438 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2439 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2440 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2441 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2442 		txq->ift_pullups = 0;
2443 		ifmp_ring_reset_stats(txq->ift_br);
2444 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2445 			bzero((void *)di->idi_vaddr, di->idi_size);
2446 	}
2447 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2448 		/* make sure all transmitters have completed before proceeding XXX */
2449 
2450 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2451 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2452 			bzero((void *)di->idi_vaddr, di->idi_size);
2453 		/* also resets the free lists pidx/cidx */
2454 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2455 			iflib_fl_bufs_free(fl);
2456 	}
2457 }
2458 
2459 static inline caddr_t
2460 calc_next_rxd(iflib_fl_t fl, int cidx)
2461 {
2462 	qidx_t size;
2463 	int nrxd;
2464 	caddr_t start, end, cur, next;
2465 
2466 	nrxd = fl->ifl_size;
2467 	size = fl->ifl_rxd_size;
2468 	start = fl->ifl_ifdi->idi_vaddr;
2469 
2470 	if (__predict_false(size == 0))
2471 		return (start);
2472 	cur = start + size*cidx;
2473 	end = start + size*nrxd;
2474 	next = CACHE_PTR_NEXT(cur);
2475 	return (next < end ? next : start);
2476 }
2477 
2478 static inline void
2479 prefetch_pkts(iflib_fl_t fl, int cidx)
2480 {
2481 	int nextptr;
2482 	int nrxd = fl->ifl_size;
2483 	caddr_t next_rxd;
2484 
2485 
2486 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2487 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2488 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2489 	next_rxd = calc_next_rxd(fl, cidx);
2490 	prefetch(next_rxd);
2491 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2492 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2493 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2494 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2495 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2496 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2497 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2498 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2499 }
2500 
2501 static struct mbuf *
2502 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2503     int *pf_rv, if_rxd_info_t ri)
2504 {
2505 	bus_dmamap_t map;
2506 	iflib_fl_t fl;
2507 	caddr_t payload;
2508 	struct mbuf *m;
2509 	int flid, cidx, len, next;
2510 
2511 	map = NULL;
2512 	flid = irf->irf_flid;
2513 	cidx = irf->irf_idx;
2514 	fl = &rxq->ifr_fl[flid];
2515 	sd->ifsd_fl = fl;
2516 	sd->ifsd_cidx = cidx;
2517 	m = fl->ifl_sds.ifsd_m[cidx];
2518 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2519 	fl->ifl_credits--;
2520 #if MEMORY_LOGGING
2521 	fl->ifl_m_dequeued++;
2522 #endif
2523 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2524 		prefetch_pkts(fl, cidx);
2525 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2526 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2527 	map = fl->ifl_sds.ifsd_map[cidx];
2528 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2529 
2530 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2531 	MPASS(fl->ifl_cidx == cidx);
2532 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2533 
2534 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2535 		payload  = *sd->ifsd_cl;
2536 		payload +=  ri->iri_pad;
2537 		len = ri->iri_len - ri->iri_pad;
2538 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2539 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2540 		switch (*pf_rv) {
2541 		case PFIL_DROPPED:
2542 		case PFIL_CONSUMED:
2543 			/*
2544 			 * The filter ate it.  Everything is recycled.
2545 			 */
2546 			m = NULL;
2547 			unload = 0;
2548 			break;
2549 		case PFIL_REALLOCED:
2550 			/*
2551 			 * The filter copied it.  Everything is recycled.
2552 			 */
2553 			m = pfil_mem2mbuf(payload);
2554 			unload = 0;
2555 			break;
2556 		case PFIL_PASS:
2557 			/*
2558 			 * Filter said it was OK, so receive like
2559 			 * normal
2560 			 */
2561 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2562 			break;
2563 		default:
2564 			MPASS(0);
2565 		}
2566 	} else {
2567 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2568 		*pf_rv = PFIL_PASS;
2569 	}
2570 
2571 	if (unload)
2572 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2573 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2574 	if (__predict_false(fl->ifl_cidx == 0))
2575 		fl->ifl_gen = 0;
2576 	bit_clear(fl->ifl_rx_bitmap, cidx);
2577 	return (m);
2578 }
2579 
2580 static struct mbuf *
2581 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2582 {
2583 	struct mbuf *m, *mh, *mt;
2584 	caddr_t cl;
2585 	int  *pf_rv_ptr, flags, i, padlen;
2586 	bool consumed;
2587 
2588 	i = 0;
2589 	mh = NULL;
2590 	consumed = false;
2591 	*pf_rv = PFIL_PASS;
2592 	pf_rv_ptr = pf_rv;
2593 	do {
2594 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2595 		    pf_rv_ptr, ri);
2596 
2597 		MPASS(*sd->ifsd_cl != NULL);
2598 
2599 		/*
2600 		 * Exclude zero-length frags & frags from
2601 		 * packets the filter has consumed or dropped
2602 		 */
2603 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2604 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2605 			if (mh == NULL) {
2606 				/* everything saved here */
2607 				consumed = true;
2608 				pf_rv_ptr = NULL;
2609 				continue;
2610 			}
2611 			/* XXX we can save the cluster here, but not the mbuf */
2612 			m_init(m, M_NOWAIT, MT_DATA, 0);
2613 			m_free(m);
2614 			continue;
2615 		}
2616 		if (mh == NULL) {
2617 			flags = M_PKTHDR|M_EXT;
2618 			mh = mt = m;
2619 			padlen = ri->iri_pad;
2620 		} else {
2621 			flags = M_EXT;
2622 			mt->m_next = m;
2623 			mt = m;
2624 			/* assuming padding is only on the first fragment */
2625 			padlen = 0;
2626 		}
2627 		cl = *sd->ifsd_cl;
2628 		*sd->ifsd_cl = NULL;
2629 
2630 		/* Can these two be made one ? */
2631 		m_init(m, M_NOWAIT, MT_DATA, flags);
2632 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2633 		/*
2634 		 * These must follow m_init and m_cljset
2635 		 */
2636 		m->m_data += padlen;
2637 		ri->iri_len -= padlen;
2638 		m->m_len = ri->iri_frags[i].irf_len;
2639 	} while (++i < ri->iri_nfrags);
2640 
2641 	return (mh);
2642 }
2643 
2644 /*
2645  * Process one software descriptor
2646  */
2647 static struct mbuf *
2648 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2649 {
2650 	struct if_rxsd sd;
2651 	struct mbuf *m;
2652 	int pf_rv;
2653 
2654 	/* should I merge this back in now that the two paths are basically duplicated? */
2655 	if (ri->iri_nfrags == 1 &&
2656 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2657 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2658 		    &pf_rv, ri);
2659 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2660 			return (m);
2661 		if (pf_rv == PFIL_PASS) {
2662 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2663 #ifndef __NO_STRICT_ALIGNMENT
2664 			if (!IP_ALIGNED(m))
2665 				m->m_data += 2;
2666 #endif
2667 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2668 			m->m_len = ri->iri_frags[0].irf_len;
2669 		}
2670 	} else {
2671 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2672 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2673 			return (m);
2674 	}
2675 	m->m_pkthdr.len = ri->iri_len;
2676 	m->m_pkthdr.rcvif = ri->iri_ifp;
2677 	m->m_flags |= ri->iri_flags;
2678 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2679 	m->m_pkthdr.flowid = ri->iri_flowid;
2680 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2681 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2682 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2683 	return (m);
2684 }
2685 
2686 #if defined(INET6) || defined(INET)
2687 static void
2688 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2689 {
2690 	CURVNET_SET(lc->ifp->if_vnet);
2691 #if defined(INET6)
2692 	*v6 = VNET(ip6_forwarding);
2693 #endif
2694 #if defined(INET)
2695 	*v4 = VNET(ipforwarding);
2696 #endif
2697 	CURVNET_RESTORE();
2698 }
2699 
2700 /*
2701  * Returns true if it's possible this packet could be LROed.
2702  * if it returns false, it is guaranteed that tcp_lro_rx()
2703  * would not return zero.
2704  */
2705 static bool
2706 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2707 {
2708 	struct ether_header *eh;
2709 	uint16_t eh_type;
2710 
2711 	eh = mtod(m, struct ether_header *);
2712 	eh_type = ntohs(eh->ether_type);
2713 	switch (eh_type) {
2714 #if defined(INET6)
2715 		case ETHERTYPE_IPV6:
2716 			return !v6_forwarding;
2717 #endif
2718 #if defined (INET)
2719 		case ETHERTYPE_IP:
2720 			return !v4_forwarding;
2721 #endif
2722 	}
2723 
2724 	return false;
2725 }
2726 #else
2727 static void
2728 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2729 {
2730 }
2731 #endif
2732 
2733 static bool
2734 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2735 {
2736 	if_ctx_t ctx = rxq->ifr_ctx;
2737 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2738 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2739 	int avail, i;
2740 	qidx_t *cidxp;
2741 	struct if_rxd_info ri;
2742 	int err, budget_left, rx_bytes, rx_pkts;
2743 	iflib_fl_t fl;
2744 	struct ifnet *ifp;
2745 	int lro_enabled;
2746 	bool v4_forwarding, v6_forwarding, lro_possible;
2747 
2748 	/*
2749 	 * XXX early demux data packets so that if_input processing only handles
2750 	 * acks in interrupt context
2751 	 */
2752 	struct mbuf *m, *mh, *mt, *mf;
2753 
2754 	lro_possible = v4_forwarding = v6_forwarding = false;
2755 	ifp = ctx->ifc_ifp;
2756 	mh = mt = NULL;
2757 	MPASS(budget > 0);
2758 	rx_pkts	= rx_bytes = 0;
2759 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2760 		cidxp = &rxq->ifr_cq_cidx;
2761 	else
2762 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2763 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2764 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2765 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2766 		DBG_COUNTER_INC(rx_unavail);
2767 		return (false);
2768 	}
2769 
2770 	/* pfil needs the vnet to be set */
2771 	CURVNET_SET_QUIET(ifp->if_vnet);
2772 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2773 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2774 			DBG_COUNTER_INC(rx_ctx_inactive);
2775 			break;
2776 		}
2777 		/*
2778 		 * Reset client set fields to their default values
2779 		 */
2780 		rxd_info_zero(&ri);
2781 		ri.iri_qsidx = rxq->ifr_id;
2782 		ri.iri_cidx = *cidxp;
2783 		ri.iri_ifp = ifp;
2784 		ri.iri_frags = rxq->ifr_frags;
2785 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2786 
2787 		if (err)
2788 			goto err;
2789 		rx_pkts += 1;
2790 		rx_bytes += ri.iri_len;
2791 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2792 			*cidxp = ri.iri_cidx;
2793 			/* Update our consumer index */
2794 			/* XXX NB: shurd - check if this is still safe */
2795 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2796 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2797 				rxq->ifr_cq_gen = 0;
2798 			}
2799 			/* was this only a completion queue message? */
2800 			if (__predict_false(ri.iri_nfrags == 0))
2801 				continue;
2802 		}
2803 		MPASS(ri.iri_nfrags != 0);
2804 		MPASS(ri.iri_len != 0);
2805 
2806 		/* will advance the cidx on the corresponding free lists */
2807 		m = iflib_rxd_pkt_get(rxq, &ri);
2808 		avail--;
2809 		budget_left--;
2810 		if (avail == 0 && budget_left)
2811 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2812 
2813 		if (__predict_false(m == NULL))
2814 			continue;
2815 
2816 		/* imm_pkt: -- cxgb */
2817 		if (mh == NULL)
2818 			mh = mt = m;
2819 		else {
2820 			mt->m_nextpkt = m;
2821 			mt = m;
2822 		}
2823 	}
2824 	CURVNET_RESTORE();
2825 	/* make sure that we can refill faster than drain */
2826 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2827 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2828 
2829 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2830 	if (lro_enabled)
2831 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2832 	mt = mf = NULL;
2833 	while (mh != NULL) {
2834 		m = mh;
2835 		mh = mh->m_nextpkt;
2836 		m->m_nextpkt = NULL;
2837 #ifndef __NO_STRICT_ALIGNMENT
2838 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2839 			continue;
2840 #endif
2841 		rx_bytes += m->m_pkthdr.len;
2842 		rx_pkts++;
2843 #if defined(INET6) || defined(INET)
2844 		if (lro_enabled) {
2845 			if (!lro_possible) {
2846 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2847 				if (lro_possible && mf != NULL) {
2848 					ifp->if_input(ifp, mf);
2849 					DBG_COUNTER_INC(rx_if_input);
2850 					mt = mf = NULL;
2851 				}
2852 			}
2853 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2854 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2855 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2856 					continue;
2857 			}
2858 		}
2859 #endif
2860 		if (lro_possible) {
2861 			ifp->if_input(ifp, m);
2862 			DBG_COUNTER_INC(rx_if_input);
2863 			continue;
2864 		}
2865 
2866 		if (mf == NULL)
2867 			mf = m;
2868 		if (mt != NULL)
2869 			mt->m_nextpkt = m;
2870 		mt = m;
2871 	}
2872 	if (mf != NULL) {
2873 		ifp->if_input(ifp, mf);
2874 		DBG_COUNTER_INC(rx_if_input);
2875 	}
2876 
2877 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2878 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2879 
2880 	/*
2881 	 * Flush any outstanding LRO work
2882 	 */
2883 #if defined(INET6) || defined(INET)
2884 	tcp_lro_flush_all(&rxq->ifr_lc);
2885 #endif
2886 	if (avail)
2887 		return true;
2888 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2889 err:
2890 	STATE_LOCK(ctx);
2891 	ctx->ifc_flags |= IFC_DO_RESET;
2892 	iflib_admin_intr_deferred(ctx);
2893 	STATE_UNLOCK(ctx);
2894 	return (false);
2895 }
2896 
2897 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2898 static inline qidx_t
2899 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2900 {
2901 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2902 	qidx_t minthresh = txq->ift_size / 8;
2903 	if (in_use > 4*minthresh)
2904 		return (notify_count);
2905 	if (in_use > 2*minthresh)
2906 		return (notify_count >> 1);
2907 	if (in_use > minthresh)
2908 		return (notify_count >> 3);
2909 	return (0);
2910 }
2911 
2912 static inline qidx_t
2913 txq_max_rs_deferred(iflib_txq_t txq)
2914 {
2915 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2916 	qidx_t minthresh = txq->ift_size / 8;
2917 	if (txq->ift_in_use > 4*minthresh)
2918 		return (notify_count);
2919 	if (txq->ift_in_use > 2*minthresh)
2920 		return (notify_count >> 1);
2921 	if (txq->ift_in_use > minthresh)
2922 		return (notify_count >> 2);
2923 	return (2);
2924 }
2925 
2926 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2927 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2928 
2929 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2930 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2931 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2932 
2933 /* forward compatibility for cxgb */
2934 #define FIRST_QSET(ctx) 0
2935 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2936 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2937 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2938 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2939 
2940 /* XXX we should be setting this to something other than zero */
2941 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2942 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2943     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2944 
2945 static inline bool
2946 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2947 {
2948 	qidx_t dbval, max;
2949 	bool rang;
2950 
2951 	rang = false;
2952 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2953 	if (ring || txq->ift_db_pending >= max) {
2954 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2955 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2956 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2957 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2958 		txq->ift_db_pending = txq->ift_npending = 0;
2959 		rang = true;
2960 	}
2961 	return (rang);
2962 }
2963 
2964 #ifdef PKT_DEBUG
2965 static void
2966 print_pkt(if_pkt_info_t pi)
2967 {
2968 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2969 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2970 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2971 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2972 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2973 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2974 }
2975 #endif
2976 
2977 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2978 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2979 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2980 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2981 
2982 static int
2983 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2984 {
2985 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2986 	struct ether_vlan_header *eh;
2987 	struct mbuf *m;
2988 
2989 	m = *mp;
2990 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2991 	    M_WRITABLE(m) == 0) {
2992 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2993 			return (ENOMEM);
2994 		} else {
2995 			m_freem(*mp);
2996 			DBG_COUNTER_INC(tx_frees);
2997 			*mp = m;
2998 		}
2999 	}
3000 
3001 	/*
3002 	 * Determine where frame payload starts.
3003 	 * Jump over vlan headers if already present,
3004 	 * helpful for QinQ too.
3005 	 */
3006 	if (__predict_false(m->m_len < sizeof(*eh))) {
3007 		txq->ift_pullups++;
3008 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3009 			return (ENOMEM);
3010 	}
3011 	eh = mtod(m, struct ether_vlan_header *);
3012 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3013 		pi->ipi_etype = ntohs(eh->evl_proto);
3014 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3015 	} else {
3016 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3017 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3018 	}
3019 
3020 	switch (pi->ipi_etype) {
3021 #ifdef INET
3022 	case ETHERTYPE_IP:
3023 	{
3024 		struct mbuf *n;
3025 		struct ip *ip = NULL;
3026 		struct tcphdr *th = NULL;
3027 		int minthlen;
3028 
3029 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3030 		if (__predict_false(m->m_len < minthlen)) {
3031 			/*
3032 			 * if this code bloat is causing too much of a hit
3033 			 * move it to a separate function and mark it noinline
3034 			 */
3035 			if (m->m_len == pi->ipi_ehdrlen) {
3036 				n = m->m_next;
3037 				MPASS(n);
3038 				if (n->m_len >= sizeof(*ip))  {
3039 					ip = (struct ip *)n->m_data;
3040 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3041 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3042 				} else {
3043 					txq->ift_pullups++;
3044 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3045 						return (ENOMEM);
3046 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3047 				}
3048 			} else {
3049 				txq->ift_pullups++;
3050 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3051 					return (ENOMEM);
3052 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3053 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3054 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3055 			}
3056 		} else {
3057 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3058 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3059 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3060 		}
3061 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3062 		pi->ipi_ipproto = ip->ip_p;
3063 		pi->ipi_flags |= IPI_TX_IPV4;
3064 
3065 		/* TCP checksum offload may require TCP header length */
3066 		if (IS_TX_OFFLOAD4(pi)) {
3067 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3068 				if (__predict_false(th == NULL)) {
3069 					txq->ift_pullups++;
3070 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3071 						return (ENOMEM);
3072 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3073 				}
3074 				pi->ipi_tcp_hflags = th->th_flags;
3075 				pi->ipi_tcp_hlen = th->th_off << 2;
3076 				pi->ipi_tcp_seq = th->th_seq;
3077 			}
3078 			if (IS_TSO4(pi)) {
3079 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3080 					return (ENXIO);
3081 				/*
3082 				 * TSO always requires hardware checksum offload.
3083 				 */
3084 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3085 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3086 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3087 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3088 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3089 					ip->ip_sum = 0;
3090 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3091 				}
3092 			}
3093 		}
3094 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3095                        ip->ip_sum = 0;
3096 
3097 		break;
3098 	}
3099 #endif
3100 #ifdef INET6
3101 	case ETHERTYPE_IPV6:
3102 	{
3103 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3104 		struct tcphdr *th;
3105 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3106 
3107 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3108 			txq->ift_pullups++;
3109 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3110 				return (ENOMEM);
3111 		}
3112 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3113 
3114 		/* XXX-BZ this will go badly in case of ext hdrs. */
3115 		pi->ipi_ipproto = ip6->ip6_nxt;
3116 		pi->ipi_flags |= IPI_TX_IPV6;
3117 
3118 		/* TCP checksum offload may require TCP header length */
3119 		if (IS_TX_OFFLOAD6(pi)) {
3120 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3121 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3122 					txq->ift_pullups++;
3123 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3124 						return (ENOMEM);
3125 				}
3126 				pi->ipi_tcp_hflags = th->th_flags;
3127 				pi->ipi_tcp_hlen = th->th_off << 2;
3128 				pi->ipi_tcp_seq = th->th_seq;
3129 			}
3130 			if (IS_TSO6(pi)) {
3131 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3132 					return (ENXIO);
3133 				/*
3134 				 * TSO always requires hardware checksum offload.
3135 				 */
3136 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3137 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3138 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3139 			}
3140 		}
3141 		break;
3142 	}
3143 #endif
3144 	default:
3145 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3146 		pi->ipi_ip_hlen = 0;
3147 		break;
3148 	}
3149 	*mp = m;
3150 
3151 	return (0);
3152 }
3153 
3154 /*
3155  * If dodgy hardware rejects the scatter gather chain we've handed it
3156  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3157  * m_defrag'd mbufs
3158  */
3159 static __noinline struct mbuf *
3160 iflib_remove_mbuf(iflib_txq_t txq)
3161 {
3162 	int ntxd, pidx;
3163 	struct mbuf *m, **ifsd_m;
3164 
3165 	ifsd_m = txq->ift_sds.ifsd_m;
3166 	ntxd = txq->ift_size;
3167 	pidx = txq->ift_pidx & (ntxd - 1);
3168 	ifsd_m = txq->ift_sds.ifsd_m;
3169 	m = ifsd_m[pidx];
3170 	ifsd_m[pidx] = NULL;
3171 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3172 	if (txq->ift_sds.ifsd_tso_map != NULL)
3173 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3174 		    txq->ift_sds.ifsd_tso_map[pidx]);
3175 #if MEMORY_LOGGING
3176 	txq->ift_dequeued++;
3177 #endif
3178 	return (m);
3179 }
3180 
3181 static inline caddr_t
3182 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3183 {
3184 	qidx_t size;
3185 	int ntxd;
3186 	caddr_t start, end, cur, next;
3187 
3188 	ntxd = txq->ift_size;
3189 	size = txq->ift_txd_size[qid];
3190 	start = txq->ift_ifdi[qid].idi_vaddr;
3191 
3192 	if (__predict_false(size == 0))
3193 		return (start);
3194 	cur = start + size*cidx;
3195 	end = start + size*ntxd;
3196 	next = CACHE_PTR_NEXT(cur);
3197 	return (next < end ? next : start);
3198 }
3199 
3200 /*
3201  * Pad an mbuf to ensure a minimum ethernet frame size.
3202  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3203  */
3204 static __noinline int
3205 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3206 {
3207 	/*
3208 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3209 	 * and ARP message is the smallest common payload I can think of
3210 	 */
3211 	static char pad[18];	/* just zeros */
3212 	int n;
3213 	struct mbuf *new_head;
3214 
3215 	if (!M_WRITABLE(*m_head)) {
3216 		new_head = m_dup(*m_head, M_NOWAIT);
3217 		if (new_head == NULL) {
3218 			m_freem(*m_head);
3219 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3220 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3221 			DBG_COUNTER_INC(tx_frees);
3222 			return ENOMEM;
3223 		}
3224 		m_freem(*m_head);
3225 		*m_head = new_head;
3226 	}
3227 
3228 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3229 	     n > 0; n -= sizeof(pad))
3230 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3231 			break;
3232 
3233 	if (n > 0) {
3234 		m_freem(*m_head);
3235 		device_printf(dev, "cannot pad short frame\n");
3236 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3237 		DBG_COUNTER_INC(tx_frees);
3238 		return (ENOBUFS);
3239 	}
3240 
3241 	return 0;
3242 }
3243 
3244 static int
3245 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3246 {
3247 	if_ctx_t		ctx;
3248 	if_shared_ctx_t		sctx;
3249 	if_softc_ctx_t		scctx;
3250 	bus_dma_tag_t		buf_tag;
3251 	bus_dma_segment_t	*segs;
3252 	struct mbuf		*m_head, **ifsd_m;
3253 	void			*next_txd;
3254 	bus_dmamap_t		map;
3255 	struct if_pkt_info	pi;
3256 	int remap = 0;
3257 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3258 
3259 	ctx = txq->ift_ctx;
3260 	sctx = ctx->ifc_sctx;
3261 	scctx = &ctx->ifc_softc_ctx;
3262 	segs = txq->ift_segs;
3263 	ntxd = txq->ift_size;
3264 	m_head = *m_headp;
3265 	map = NULL;
3266 
3267 	/*
3268 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3269 	 */
3270 	cidx = txq->ift_cidx;
3271 	pidx = txq->ift_pidx;
3272 	if (ctx->ifc_flags & IFC_PREFETCH) {
3273 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3274 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3275 			next_txd = calc_next_txd(txq, cidx, 0);
3276 			prefetch(next_txd);
3277 		}
3278 
3279 		/* prefetch the next cache line of mbuf pointers and flags */
3280 		prefetch(&txq->ift_sds.ifsd_m[next]);
3281 		prefetch(&txq->ift_sds.ifsd_map[next]);
3282 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3283 	}
3284 	map = txq->ift_sds.ifsd_map[pidx];
3285 	ifsd_m = txq->ift_sds.ifsd_m;
3286 
3287 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3288 		buf_tag = txq->ift_tso_buf_tag;
3289 		max_segs = scctx->isc_tx_tso_segments_max;
3290 		map = txq->ift_sds.ifsd_tso_map[pidx];
3291 		MPASS(buf_tag != NULL);
3292 		MPASS(max_segs > 0);
3293 	} else {
3294 		buf_tag = txq->ift_buf_tag;
3295 		max_segs = scctx->isc_tx_nsegments;
3296 		map = txq->ift_sds.ifsd_map[pidx];
3297 	}
3298 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3299 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3300 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3301 		if (err) {
3302 			DBG_COUNTER_INC(encap_txd_encap_fail);
3303 			return err;
3304 		}
3305 	}
3306 	m_head = *m_headp;
3307 
3308 	pkt_info_zero(&pi);
3309 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3310 	pi.ipi_pidx = pidx;
3311 	pi.ipi_qsidx = txq->ift_id;
3312 	pi.ipi_len = m_head->m_pkthdr.len;
3313 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3314 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3315 
3316 	/* deliberate bitwise OR to make one condition */
3317 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3318 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3319 			DBG_COUNTER_INC(encap_txd_encap_fail);
3320 			return (err);
3321 		}
3322 		m_head = *m_headp;
3323 	}
3324 
3325 retry:
3326 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3327 	    BUS_DMA_NOWAIT);
3328 defrag:
3329 	if (__predict_false(err)) {
3330 		switch (err) {
3331 		case EFBIG:
3332 			/* try collapse once and defrag once */
3333 			if (remap == 0) {
3334 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3335 				/* try defrag if collapsing fails */
3336 				if (m_head == NULL)
3337 					remap++;
3338 			}
3339 			if (remap == 1) {
3340 				txq->ift_mbuf_defrag++;
3341 				m_head = m_defrag(*m_headp, M_NOWAIT);
3342 			}
3343 			/*
3344 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3345 			 * failed to map an mbuf that was run through m_defrag
3346 			 */
3347 			MPASS(remap <= 1);
3348 			if (__predict_false(m_head == NULL || remap > 1))
3349 				goto defrag_failed;
3350 			remap++;
3351 			*m_headp = m_head;
3352 			goto retry;
3353 			break;
3354 		case ENOMEM:
3355 			txq->ift_no_tx_dma_setup++;
3356 			break;
3357 		default:
3358 			txq->ift_no_tx_dma_setup++;
3359 			m_freem(*m_headp);
3360 			DBG_COUNTER_INC(tx_frees);
3361 			*m_headp = NULL;
3362 			break;
3363 		}
3364 		txq->ift_map_failed++;
3365 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3366 		DBG_COUNTER_INC(encap_txd_encap_fail);
3367 		return (err);
3368 	}
3369 	ifsd_m[pidx] = m_head;
3370 	/*
3371 	 * XXX assumes a 1 to 1 relationship between segments and
3372 	 *        descriptors - this does not hold true on all drivers, e.g.
3373 	 *        cxgb
3374 	 */
3375 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3376 		txq->ift_no_desc_avail++;
3377 		bus_dmamap_unload(buf_tag, map);
3378 		DBG_COUNTER_INC(encap_txq_avail_fail);
3379 		DBG_COUNTER_INC(encap_txd_encap_fail);
3380 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3381 			GROUPTASK_ENQUEUE(&txq->ift_task);
3382 		return (ENOBUFS);
3383 	}
3384 	/*
3385 	 * On Intel cards we can greatly reduce the number of TX interrupts
3386 	 * we see by only setting report status on every Nth descriptor.
3387 	 * However, this also means that the driver will need to keep track
3388 	 * of the descriptors that RS was set on to check them for the DD bit.
3389 	 */
3390 	txq->ift_rs_pending += nsegs + 1;
3391 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3392 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3393 		pi.ipi_flags |= IPI_TX_INTR;
3394 		txq->ift_rs_pending = 0;
3395 	}
3396 
3397 	pi.ipi_segs = segs;
3398 	pi.ipi_nsegs = nsegs;
3399 
3400 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3401 #ifdef PKT_DEBUG
3402 	print_pkt(&pi);
3403 #endif
3404 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3405 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3406 		DBG_COUNTER_INC(tx_encap);
3407 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3408 
3409 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3410 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3411 			ndesc += txq->ift_size;
3412 			txq->ift_gen = 1;
3413 		}
3414 		/*
3415 		 * drivers can need as many as
3416 		 * two sentinels
3417 		 */
3418 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3419 		MPASS(pi.ipi_new_pidx != pidx);
3420 		MPASS(ndesc > 0);
3421 		txq->ift_in_use += ndesc;
3422 
3423 		/*
3424 		 * We update the last software descriptor again here because there may
3425 		 * be a sentinel and/or there may be more mbufs than segments
3426 		 */
3427 		txq->ift_pidx = pi.ipi_new_pidx;
3428 		txq->ift_npending += pi.ipi_ndescs;
3429 	} else {
3430 		*m_headp = m_head = iflib_remove_mbuf(txq);
3431 		if (err == EFBIG) {
3432 			txq->ift_txd_encap_efbig++;
3433 			if (remap < 2) {
3434 				remap = 1;
3435 				goto defrag;
3436 			}
3437 		}
3438 		goto defrag_failed;
3439 	}
3440 	/*
3441 	 * err can't possibly be non-zero here, so we don't neet to test it
3442 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3443 	 */
3444 	return (err);
3445 
3446 defrag_failed:
3447 	txq->ift_mbuf_defrag_failed++;
3448 	txq->ift_map_failed++;
3449 	m_freem(*m_headp);
3450 	DBG_COUNTER_INC(tx_frees);
3451 	*m_headp = NULL;
3452 	DBG_COUNTER_INC(encap_txd_encap_fail);
3453 	return (ENOMEM);
3454 }
3455 
3456 static void
3457 iflib_tx_desc_free(iflib_txq_t txq, int n)
3458 {
3459 	uint32_t qsize, cidx, mask, gen;
3460 	struct mbuf *m, **ifsd_m;
3461 	bool do_prefetch;
3462 
3463 	cidx = txq->ift_cidx;
3464 	gen = txq->ift_gen;
3465 	qsize = txq->ift_size;
3466 	mask = qsize-1;
3467 	ifsd_m = txq->ift_sds.ifsd_m;
3468 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3469 
3470 	while (n-- > 0) {
3471 		if (do_prefetch) {
3472 			prefetch(ifsd_m[(cidx + 3) & mask]);
3473 			prefetch(ifsd_m[(cidx + 4) & mask]);
3474 		}
3475 		if ((m = ifsd_m[cidx]) != NULL) {
3476 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3477 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3478 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3479 				    txq->ift_sds.ifsd_tso_map[cidx],
3480 				    BUS_DMASYNC_POSTWRITE);
3481 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3482 				    txq->ift_sds.ifsd_tso_map[cidx]);
3483 			} else {
3484 				bus_dmamap_sync(txq->ift_buf_tag,
3485 				    txq->ift_sds.ifsd_map[cidx],
3486 				    BUS_DMASYNC_POSTWRITE);
3487 				bus_dmamap_unload(txq->ift_buf_tag,
3488 				    txq->ift_sds.ifsd_map[cidx]);
3489 			}
3490 			/* XXX we don't support any drivers that batch packets yet */
3491 			MPASS(m->m_nextpkt == NULL);
3492 			m_freem(m);
3493 			ifsd_m[cidx] = NULL;
3494 #if MEMORY_LOGGING
3495 			txq->ift_dequeued++;
3496 #endif
3497 			DBG_COUNTER_INC(tx_frees);
3498 		}
3499 		if (__predict_false(++cidx == qsize)) {
3500 			cidx = 0;
3501 			gen = 0;
3502 		}
3503 	}
3504 	txq->ift_cidx = cidx;
3505 	txq->ift_gen = gen;
3506 }
3507 
3508 static __inline int
3509 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3510 {
3511 	int reclaim;
3512 	if_ctx_t ctx = txq->ift_ctx;
3513 
3514 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3515 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3516 
3517 	/*
3518 	 * Need a rate-limiting check so that this isn't called every time
3519 	 */
3520 	iflib_tx_credits_update(ctx, txq);
3521 	reclaim = DESC_RECLAIMABLE(txq);
3522 
3523 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3524 #ifdef INVARIANTS
3525 		if (iflib_verbose_debug) {
3526 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3527 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3528 			       reclaim, thresh);
3529 
3530 		}
3531 #endif
3532 		return (0);
3533 	}
3534 	iflib_tx_desc_free(txq, reclaim);
3535 	txq->ift_cleaned += reclaim;
3536 	txq->ift_in_use -= reclaim;
3537 
3538 	return (reclaim);
3539 }
3540 
3541 static struct mbuf **
3542 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3543 {
3544 	int next, size;
3545 	struct mbuf **items;
3546 
3547 	size = r->size;
3548 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3549 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3550 
3551 	prefetch(items[(cidx + offset) & (size-1)]);
3552 	if (remaining > 1) {
3553 		prefetch2cachelines(&items[next]);
3554 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3555 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3556 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3557 	}
3558 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3559 }
3560 
3561 static void
3562 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3563 {
3564 
3565 	ifmp_ring_check_drainage(txq->ift_br, budget);
3566 }
3567 
3568 static uint32_t
3569 iflib_txq_can_drain(struct ifmp_ring *r)
3570 {
3571 	iflib_txq_t txq = r->cookie;
3572 	if_ctx_t ctx = txq->ift_ctx;
3573 
3574 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3575 		return (1);
3576 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3577 	    BUS_DMASYNC_POSTREAD);
3578 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3579 	    false));
3580 }
3581 
3582 static uint32_t
3583 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3584 {
3585 	iflib_txq_t txq = r->cookie;
3586 	if_ctx_t ctx = txq->ift_ctx;
3587 	struct ifnet *ifp = ctx->ifc_ifp;
3588 	struct mbuf **mp, *m;
3589 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3590 	int reclaimed, err, in_use_prev, desc_used;
3591 	bool do_prefetch, ring, rang;
3592 
3593 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3594 			    !LINK_ACTIVE(ctx))) {
3595 		DBG_COUNTER_INC(txq_drain_notready);
3596 		return (0);
3597 	}
3598 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3599 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3600 	avail = IDXDIFF(pidx, cidx, r->size);
3601 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3602 		DBG_COUNTER_INC(txq_drain_flushing);
3603 		for (i = 0; i < avail; i++) {
3604 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3605 				m_free(r->items[(cidx + i) & (r->size-1)]);
3606 			r->items[(cidx + i) & (r->size-1)] = NULL;
3607 		}
3608 		return (avail);
3609 	}
3610 
3611 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3612 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3613 		CALLOUT_LOCK(txq);
3614 		callout_stop(&txq->ift_timer);
3615 		CALLOUT_UNLOCK(txq);
3616 		DBG_COUNTER_INC(txq_drain_oactive);
3617 		return (0);
3618 	}
3619 	if (reclaimed)
3620 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3621 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3622 	count = MIN(avail, TX_BATCH_SIZE);
3623 #ifdef INVARIANTS
3624 	if (iflib_verbose_debug)
3625 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3626 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3627 #endif
3628 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3629 	avail = TXQ_AVAIL(txq);
3630 	err = 0;
3631 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3632 		int rem = do_prefetch ? count - i : 0;
3633 
3634 		mp = _ring_peek_one(r, cidx, i, rem);
3635 		MPASS(mp != NULL && *mp != NULL);
3636 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3637 			consumed++;
3638 			reclaimed++;
3639 			continue;
3640 		}
3641 		in_use_prev = txq->ift_in_use;
3642 		err = iflib_encap(txq, mp);
3643 		if (__predict_false(err)) {
3644 			/* no room - bail out */
3645 			if (err == ENOBUFS)
3646 				break;
3647 			consumed++;
3648 			/* we can't send this packet - skip it */
3649 			continue;
3650 		}
3651 		consumed++;
3652 		pkt_sent++;
3653 		m = *mp;
3654 		DBG_COUNTER_INC(tx_sent);
3655 		bytes_sent += m->m_pkthdr.len;
3656 		mcast_sent += !!(m->m_flags & M_MCAST);
3657 		avail = TXQ_AVAIL(txq);
3658 
3659 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3660 		desc_used += (txq->ift_in_use - in_use_prev);
3661 		ETHER_BPF_MTAP(ifp, m);
3662 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3663 			break;
3664 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3665 	}
3666 
3667 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3668 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3669 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3670 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3671 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3672 	if (mcast_sent)
3673 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3674 #ifdef INVARIANTS
3675 	if (iflib_verbose_debug)
3676 		printf("consumed=%d\n", consumed);
3677 #endif
3678 	return (consumed);
3679 }
3680 
3681 static uint32_t
3682 iflib_txq_drain_always(struct ifmp_ring *r)
3683 {
3684 	return (1);
3685 }
3686 
3687 static uint32_t
3688 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3689 {
3690 	int i, avail;
3691 	struct mbuf **mp;
3692 	iflib_txq_t txq;
3693 
3694 	txq = r->cookie;
3695 
3696 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3697 	CALLOUT_LOCK(txq);
3698 	callout_stop(&txq->ift_timer);
3699 	CALLOUT_UNLOCK(txq);
3700 
3701 	avail = IDXDIFF(pidx, cidx, r->size);
3702 	for (i = 0; i < avail; i++) {
3703 		mp = _ring_peek_one(r, cidx, i, avail - i);
3704 		if (__predict_false(*mp == (struct mbuf *)txq))
3705 			continue;
3706 		m_freem(*mp);
3707 		DBG_COUNTER_INC(tx_frees);
3708 	}
3709 	MPASS(ifmp_ring_is_stalled(r) == 0);
3710 	return (avail);
3711 }
3712 
3713 static void
3714 iflib_ifmp_purge(iflib_txq_t txq)
3715 {
3716 	struct ifmp_ring *r;
3717 
3718 	r = txq->ift_br;
3719 	r->drain = iflib_txq_drain_free;
3720 	r->can_drain = iflib_txq_drain_always;
3721 
3722 	ifmp_ring_check_drainage(r, r->size);
3723 
3724 	r->drain = iflib_txq_drain;
3725 	r->can_drain = iflib_txq_can_drain;
3726 }
3727 
3728 static void
3729 _task_fn_tx(void *context)
3730 {
3731 	iflib_txq_t txq = context;
3732 	if_ctx_t ctx = txq->ift_ctx;
3733 #if defined(ALTQ) || defined(DEV_NETMAP)
3734 	if_t ifp = ctx->ifc_ifp;
3735 #endif
3736 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3737 
3738 #ifdef IFLIB_DIAGNOSTICS
3739 	txq->ift_cpu_exec_count[curcpu]++;
3740 #endif
3741 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3742 		return;
3743 #ifdef DEV_NETMAP
3744 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3745 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3746 		    BUS_DMASYNC_POSTREAD);
3747 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3748 			netmap_tx_irq(ifp, txq->ift_id);
3749 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3750 		return;
3751 	}
3752 #endif
3753 #ifdef ALTQ
3754 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3755 		iflib_altq_if_start(ifp);
3756 #endif
3757 	if (txq->ift_db_pending)
3758 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3759 	else if (!abdicate)
3760 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3761 	/*
3762 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3763 	 */
3764 	if (abdicate)
3765 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3766 	if (ctx->ifc_flags & IFC_LEGACY)
3767 		IFDI_INTR_ENABLE(ctx);
3768 	else {
3769 #ifdef INVARIANTS
3770 		int rc =
3771 #endif
3772 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3773 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3774 	}
3775 }
3776 
3777 static void
3778 _task_fn_rx(void *context)
3779 {
3780 	iflib_rxq_t rxq = context;
3781 	if_ctx_t ctx = rxq->ifr_ctx;
3782 	bool more;
3783 	uint16_t budget;
3784 
3785 #ifdef IFLIB_DIAGNOSTICS
3786 	rxq->ifr_cpu_exec_count[curcpu]++;
3787 #endif
3788 	DBG_COUNTER_INC(task_fn_rxs);
3789 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3790 		return;
3791 	more = true;
3792 #ifdef DEV_NETMAP
3793 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3794 		u_int work = 0;
3795 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3796 			more = false;
3797 		}
3798 	}
3799 #endif
3800 	budget = ctx->ifc_sysctl_rx_budget;
3801 	if (budget == 0)
3802 		budget = 16;	/* XXX */
3803 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3804 		if (ctx->ifc_flags & IFC_LEGACY)
3805 			IFDI_INTR_ENABLE(ctx);
3806 		else {
3807 #ifdef INVARIANTS
3808 			int rc =
3809 #endif
3810 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3811 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3812 			DBG_COUNTER_INC(rx_intr_enables);
3813 		}
3814 	}
3815 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3816 		return;
3817 	if (more)
3818 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3819 }
3820 
3821 static void
3822 _task_fn_admin(void *context)
3823 {
3824 	if_ctx_t ctx = context;
3825 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3826 	iflib_txq_t txq;
3827 	int i;
3828 	bool oactive, running, do_reset, do_watchdog, in_detach;
3829 	uint32_t reset_on = hz / 2;
3830 
3831 	STATE_LOCK(ctx);
3832 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3833 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3834 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3835 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3836 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3837 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3838 	STATE_UNLOCK(ctx);
3839 
3840 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3841 		return;
3842 	if (in_detach)
3843 		return;
3844 
3845 	CTX_LOCK(ctx);
3846 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3847 		CALLOUT_LOCK(txq);
3848 		callout_stop(&txq->ift_timer);
3849 		CALLOUT_UNLOCK(txq);
3850 	}
3851 	if (do_watchdog) {
3852 		ctx->ifc_watchdog_events++;
3853 		IFDI_WATCHDOG_RESET(ctx);
3854 	}
3855 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3856 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3857 #ifdef DEV_NETMAP
3858 		reset_on = hz / 2;
3859 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3860 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3861 #endif
3862 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3863 	}
3864 	IFDI_LINK_INTR_ENABLE(ctx);
3865 	if (do_reset)
3866 		iflib_if_init_locked(ctx);
3867 	CTX_UNLOCK(ctx);
3868 
3869 	if (LINK_ACTIVE(ctx) == 0)
3870 		return;
3871 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3872 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3873 }
3874 
3875 
3876 static void
3877 _task_fn_iov(void *context)
3878 {
3879 	if_ctx_t ctx = context;
3880 
3881 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3882 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3883 		return;
3884 
3885 	CTX_LOCK(ctx);
3886 	IFDI_VFLR_HANDLE(ctx);
3887 	CTX_UNLOCK(ctx);
3888 }
3889 
3890 static int
3891 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3892 {
3893 	int err;
3894 	if_int_delay_info_t info;
3895 	if_ctx_t ctx;
3896 
3897 	info = (if_int_delay_info_t)arg1;
3898 	ctx = info->iidi_ctx;
3899 	info->iidi_req = req;
3900 	info->iidi_oidp = oidp;
3901 	CTX_LOCK(ctx);
3902 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3903 	CTX_UNLOCK(ctx);
3904 	return (err);
3905 }
3906 
3907 /*********************************************************************
3908  *
3909  *  IFNET FUNCTIONS
3910  *
3911  **********************************************************************/
3912 
3913 static void
3914 iflib_if_init_locked(if_ctx_t ctx)
3915 {
3916 	iflib_stop(ctx);
3917 	iflib_init_locked(ctx);
3918 }
3919 
3920 
3921 static void
3922 iflib_if_init(void *arg)
3923 {
3924 	if_ctx_t ctx = arg;
3925 
3926 	CTX_LOCK(ctx);
3927 	iflib_if_init_locked(ctx);
3928 	CTX_UNLOCK(ctx);
3929 }
3930 
3931 static int
3932 iflib_if_transmit(if_t ifp, struct mbuf *m)
3933 {
3934 	if_ctx_t	ctx = if_getsoftc(ifp);
3935 
3936 	iflib_txq_t txq;
3937 	int err, qidx;
3938 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3939 
3940 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3941 		DBG_COUNTER_INC(tx_frees);
3942 		m_freem(m);
3943 		return (ENETDOWN);
3944 	}
3945 
3946 	MPASS(m->m_nextpkt == NULL);
3947 	/* ALTQ-enabled interfaces always use queue 0. */
3948 	qidx = 0;
3949 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3950 		qidx = QIDX(ctx, m);
3951 	/*
3952 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3953 	 */
3954 	txq = &ctx->ifc_txqs[qidx];
3955 
3956 #ifdef DRIVER_BACKPRESSURE
3957 	if (txq->ift_closed) {
3958 		while (m != NULL) {
3959 			next = m->m_nextpkt;
3960 			m->m_nextpkt = NULL;
3961 			m_freem(m);
3962 			DBG_COUNTER_INC(tx_frees);
3963 			m = next;
3964 		}
3965 		return (ENOBUFS);
3966 	}
3967 #endif
3968 #ifdef notyet
3969 	qidx = count = 0;
3970 	mp = marr;
3971 	next = m;
3972 	do {
3973 		count++;
3974 		next = next->m_nextpkt;
3975 	} while (next != NULL);
3976 
3977 	if (count > nitems(marr))
3978 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3979 			/* XXX check nextpkt */
3980 			m_freem(m);
3981 			/* XXX simplify for now */
3982 			DBG_COUNTER_INC(tx_frees);
3983 			return (ENOBUFS);
3984 		}
3985 	for (next = m, i = 0; next != NULL; i++) {
3986 		mp[i] = next;
3987 		next = next->m_nextpkt;
3988 		mp[i]->m_nextpkt = NULL;
3989 	}
3990 #endif
3991 	DBG_COUNTER_INC(tx_seen);
3992 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3993 
3994 	if (abdicate)
3995 		GROUPTASK_ENQUEUE(&txq->ift_task);
3996  	if (err) {
3997 		if (!abdicate)
3998 			GROUPTASK_ENQUEUE(&txq->ift_task);
3999 		/* support forthcoming later */
4000 #ifdef DRIVER_BACKPRESSURE
4001 		txq->ift_closed = TRUE;
4002 #endif
4003 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4004 		m_freem(m);
4005 		DBG_COUNTER_INC(tx_frees);
4006 	}
4007 
4008 	return (err);
4009 }
4010 
4011 #ifdef ALTQ
4012 /*
4013  * The overall approach to integrating iflib with ALTQ is to continue to use
4014  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4015  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4016  * is redundant/unnecessary, but doing so minimizes the amount of
4017  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4018  * redundantly queueing to an intermediate mp_ring is swamped by the
4019  * performance limitations inherent in using ALTQ.
4020  *
4021  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4022  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4023  * given interface.  If ALTQ is enabled for an interface, then all
4024  * transmitted packets for that interface will be submitted to the ALTQ
4025  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4026  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4027  * update stats that the iflib machinery handles, and which is sensitve to
4028  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4029  * will be installed as the start routine for use by ALTQ facilities that
4030  * need to trigger queue drains on a scheduled basis.
4031  *
4032  */
4033 static void
4034 iflib_altq_if_start(if_t ifp)
4035 {
4036 	struct ifaltq *ifq = &ifp->if_snd;
4037 	struct mbuf *m;
4038 
4039 	IFQ_LOCK(ifq);
4040 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4041 	while (m != NULL) {
4042 		iflib_if_transmit(ifp, m);
4043 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4044 	}
4045 	IFQ_UNLOCK(ifq);
4046 }
4047 
4048 static int
4049 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4050 {
4051 	int err;
4052 
4053 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4054 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4055 		if (err == 0)
4056 			iflib_altq_if_start(ifp);
4057 	} else
4058 		err = iflib_if_transmit(ifp, m);
4059 
4060 	return (err);
4061 }
4062 #endif /* ALTQ */
4063 
4064 static void
4065 iflib_if_qflush(if_t ifp)
4066 {
4067 	if_ctx_t ctx = if_getsoftc(ifp);
4068 	iflib_txq_t txq = ctx->ifc_txqs;
4069 	int i;
4070 
4071 	STATE_LOCK(ctx);
4072 	ctx->ifc_flags |= IFC_QFLUSH;
4073 	STATE_UNLOCK(ctx);
4074 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4075 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4076 			iflib_txq_check_drain(txq, 0);
4077 	STATE_LOCK(ctx);
4078 	ctx->ifc_flags &= ~IFC_QFLUSH;
4079 	STATE_UNLOCK(ctx);
4080 
4081 	/*
4082 	 * When ALTQ is enabled, this will also take care of purging the
4083 	 * ALTQ queue(s).
4084 	 */
4085 	if_qflush(ifp);
4086 }
4087 
4088 
4089 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4090 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4091 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4092 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4093 
4094 static int
4095 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4096 {
4097 	if_ctx_t ctx = if_getsoftc(ifp);
4098 	struct ifreq	*ifr = (struct ifreq *)data;
4099 #if defined(INET) || defined(INET6)
4100 	struct ifaddr	*ifa = (struct ifaddr *)data;
4101 #endif
4102 	bool		avoid_reset = FALSE;
4103 	int		err = 0, reinit = 0, bits;
4104 
4105 	switch (command) {
4106 	case SIOCSIFADDR:
4107 #ifdef INET
4108 		if (ifa->ifa_addr->sa_family == AF_INET)
4109 			avoid_reset = TRUE;
4110 #endif
4111 #ifdef INET6
4112 		if (ifa->ifa_addr->sa_family == AF_INET6)
4113 			avoid_reset = TRUE;
4114 #endif
4115 		/*
4116 		** Calling init results in link renegotiation,
4117 		** so we avoid doing it when possible.
4118 		*/
4119 		if (avoid_reset) {
4120 			if_setflagbits(ifp, IFF_UP,0);
4121 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4122 				reinit = 1;
4123 #ifdef INET
4124 			if (!(if_getflags(ifp) & IFF_NOARP))
4125 				arp_ifinit(ifp, ifa);
4126 #endif
4127 		} else
4128 			err = ether_ioctl(ifp, command, data);
4129 		break;
4130 	case SIOCSIFMTU:
4131 		CTX_LOCK(ctx);
4132 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4133 			CTX_UNLOCK(ctx);
4134 			break;
4135 		}
4136 		bits = if_getdrvflags(ifp);
4137 		/* stop the driver and free any clusters before proceeding */
4138 		iflib_stop(ctx);
4139 
4140 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4141 			STATE_LOCK(ctx);
4142 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4143 				ctx->ifc_flags |= IFC_MULTISEG;
4144 			else
4145 				ctx->ifc_flags &= ~IFC_MULTISEG;
4146 			STATE_UNLOCK(ctx);
4147 			err = if_setmtu(ifp, ifr->ifr_mtu);
4148 		}
4149 		iflib_init_locked(ctx);
4150 		STATE_LOCK(ctx);
4151 		if_setdrvflags(ifp, bits);
4152 		STATE_UNLOCK(ctx);
4153 		CTX_UNLOCK(ctx);
4154 		break;
4155 	case SIOCSIFFLAGS:
4156 		CTX_LOCK(ctx);
4157 		if (if_getflags(ifp) & IFF_UP) {
4158 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4159 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4160 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4161 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4162 				}
4163 			} else
4164 				reinit = 1;
4165 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4166 			iflib_stop(ctx);
4167 		}
4168 		ctx->ifc_if_flags = if_getflags(ifp);
4169 		CTX_UNLOCK(ctx);
4170 		break;
4171 	case SIOCADDMULTI:
4172 	case SIOCDELMULTI:
4173 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4174 			CTX_LOCK(ctx);
4175 			IFDI_INTR_DISABLE(ctx);
4176 			IFDI_MULTI_SET(ctx);
4177 			IFDI_INTR_ENABLE(ctx);
4178 			CTX_UNLOCK(ctx);
4179 		}
4180 		break;
4181 	case SIOCSIFMEDIA:
4182 		CTX_LOCK(ctx);
4183 		IFDI_MEDIA_SET(ctx);
4184 		CTX_UNLOCK(ctx);
4185 		/* falls thru */
4186 	case SIOCGIFMEDIA:
4187 	case SIOCGIFXMEDIA:
4188 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4189 		break;
4190 	case SIOCGI2C:
4191 	{
4192 		struct ifi2creq i2c;
4193 
4194 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4195 		if (err != 0)
4196 			break;
4197 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4198 			err = EINVAL;
4199 			break;
4200 		}
4201 		if (i2c.len > sizeof(i2c.data)) {
4202 			err = EINVAL;
4203 			break;
4204 		}
4205 
4206 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4207 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4208 			    sizeof(i2c));
4209 		break;
4210 	}
4211 	case SIOCSIFCAP:
4212 	{
4213 		int mask, setmask, oldmask;
4214 
4215 		oldmask = if_getcapenable(ifp);
4216 		mask = ifr->ifr_reqcap ^ oldmask;
4217 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4218 		setmask = 0;
4219 #ifdef TCP_OFFLOAD
4220 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4221 #endif
4222 		setmask |= (mask & IFCAP_FLAGS);
4223 		setmask |= (mask & IFCAP_WOL);
4224 
4225 		/*
4226 		 * If any RX csum has changed, change all the ones that
4227 		 * are supported by the driver.
4228 		 */
4229 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4230 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4231 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4232 		}
4233 
4234 		/*
4235 		 * want to ensure that traffic has stopped before we change any of the flags
4236 		 */
4237 		if (setmask) {
4238 			CTX_LOCK(ctx);
4239 			bits = if_getdrvflags(ifp);
4240 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4241 				iflib_stop(ctx);
4242 			STATE_LOCK(ctx);
4243 			if_togglecapenable(ifp, setmask);
4244 			STATE_UNLOCK(ctx);
4245 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4246 				iflib_init_locked(ctx);
4247 			STATE_LOCK(ctx);
4248 			if_setdrvflags(ifp, bits);
4249 			STATE_UNLOCK(ctx);
4250 			CTX_UNLOCK(ctx);
4251 		}
4252 		if_vlancap(ifp);
4253 		break;
4254 	}
4255 	case SIOCGPRIVATE_0:
4256 	case SIOCSDRVSPEC:
4257 	case SIOCGDRVSPEC:
4258 		CTX_LOCK(ctx);
4259 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4260 		CTX_UNLOCK(ctx);
4261 		break;
4262 	default:
4263 		err = ether_ioctl(ifp, command, data);
4264 		break;
4265 	}
4266 	if (reinit)
4267 		iflib_if_init(ctx);
4268 	return (err);
4269 }
4270 
4271 static uint64_t
4272 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4273 {
4274 	if_ctx_t ctx = if_getsoftc(ifp);
4275 
4276 	return (IFDI_GET_COUNTER(ctx, cnt));
4277 }
4278 
4279 /*********************************************************************
4280  *
4281  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4282  *
4283  **********************************************************************/
4284 
4285 static void
4286 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4287 {
4288 	if_ctx_t ctx = if_getsoftc(ifp);
4289 
4290 	if ((void *)ctx != arg)
4291 		return;
4292 
4293 	if ((vtag == 0) || (vtag > 4095))
4294 		return;
4295 
4296 	CTX_LOCK(ctx);
4297 	IFDI_VLAN_REGISTER(ctx, vtag);
4298 	/* Re-init to load the changes */
4299 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4300 		iflib_if_init_locked(ctx);
4301 	CTX_UNLOCK(ctx);
4302 }
4303 
4304 static void
4305 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4306 {
4307 	if_ctx_t ctx = if_getsoftc(ifp);
4308 
4309 	if ((void *)ctx != arg)
4310 		return;
4311 
4312 	if ((vtag == 0) || (vtag > 4095))
4313 		return;
4314 
4315 	CTX_LOCK(ctx);
4316 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4317 	/* Re-init to load the changes */
4318 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4319 		iflib_if_init_locked(ctx);
4320 	CTX_UNLOCK(ctx);
4321 }
4322 
4323 static void
4324 iflib_led_func(void *arg, int onoff)
4325 {
4326 	if_ctx_t ctx = arg;
4327 
4328 	CTX_LOCK(ctx);
4329 	IFDI_LED_FUNC(ctx, onoff);
4330 	CTX_UNLOCK(ctx);
4331 }
4332 
4333 /*********************************************************************
4334  *
4335  *  BUS FUNCTION DEFINITIONS
4336  *
4337  **********************************************************************/
4338 
4339 int
4340 iflib_device_probe(device_t dev)
4341 {
4342 	pci_vendor_info_t *ent;
4343 
4344 	uint16_t	pci_vendor_id, pci_device_id;
4345 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4346 	uint16_t	pci_rev_id;
4347 	if_shared_ctx_t sctx;
4348 
4349 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4350 		return (ENOTSUP);
4351 
4352 	pci_vendor_id = pci_get_vendor(dev);
4353 	pci_device_id = pci_get_device(dev);
4354 	pci_subvendor_id = pci_get_subvendor(dev);
4355 	pci_subdevice_id = pci_get_subdevice(dev);
4356 	pci_rev_id = pci_get_revid(dev);
4357 	if (sctx->isc_parse_devinfo != NULL)
4358 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4359 
4360 	ent = sctx->isc_vendor_info;
4361 	while (ent->pvi_vendor_id != 0) {
4362 		if (pci_vendor_id != ent->pvi_vendor_id) {
4363 			ent++;
4364 			continue;
4365 		}
4366 		if ((pci_device_id == ent->pvi_device_id) &&
4367 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4368 		     (ent->pvi_subvendor_id == 0)) &&
4369 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4370 		     (ent->pvi_subdevice_id == 0)) &&
4371 		    ((pci_rev_id == ent->pvi_rev_id) ||
4372 		     (ent->pvi_rev_id == 0))) {
4373 
4374 			device_set_desc_copy(dev, ent->pvi_name);
4375 			/* this needs to be changed to zero if the bus probing code
4376 			 * ever stops re-probing on best match because the sctx
4377 			 * may have its values over written by register calls
4378 			 * in subsequent probes
4379 			 */
4380 			return (BUS_PROBE_DEFAULT);
4381 		}
4382 		ent++;
4383 	}
4384 	return (ENXIO);
4385 }
4386 
4387 static void
4388 iflib_reset_qvalues(if_ctx_t ctx)
4389 {
4390 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4391 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4392 	device_t dev = ctx->ifc_dev;
4393 	int i;
4394 
4395 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4396 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4397 	/*
4398 	 * XXX sanity check that ntxd & nrxd are a power of 2
4399 	 */
4400 	if (ctx->ifc_sysctl_ntxqs != 0)
4401 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4402 	if (ctx->ifc_sysctl_nrxqs != 0)
4403 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4404 
4405 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4406 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4407 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4408 		else
4409 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4410 	}
4411 
4412 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4413 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4414 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4415 		else
4416 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4417 	}
4418 
4419 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4420 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4421 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4422 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4423 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4424 		}
4425 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4426 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4427 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4428 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4429 		}
4430 	}
4431 
4432 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4433 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4434 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4435 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4436 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4437 		}
4438 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4439 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4440 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4441 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4442 		}
4443 	}
4444 }
4445 
4446 static void
4447 iflib_add_pfil(if_ctx_t ctx)
4448 {
4449 	struct pfil_head *pfil;
4450 	struct pfil_head_args pa;
4451 	iflib_rxq_t rxq;
4452 	int i;
4453 
4454 	pa.pa_version = PFIL_VERSION;
4455 	pa.pa_flags = PFIL_IN;
4456 	pa.pa_type = PFIL_TYPE_ETHERNET;
4457 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4458 	pfil = pfil_head_register(&pa);
4459 
4460 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4461 		rxq->pfil = pfil;
4462 	}
4463 }
4464 
4465 static void
4466 iflib_rem_pfil(if_ctx_t ctx)
4467 {
4468 	struct pfil_head *pfil;
4469 	iflib_rxq_t rxq;
4470 	int i;
4471 
4472 	rxq = ctx->ifc_rxqs;
4473 	pfil = rxq->pfil;
4474 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4475 		rxq->pfil = NULL;
4476 	}
4477 	pfil_head_unregister(pfil);
4478 }
4479 
4480 static uint16_t
4481 get_ctx_core_offset(if_ctx_t ctx)
4482 {
4483 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4484 	struct cpu_offset *op;
4485 	uint16_t qc;
4486 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4487 
4488 	if (ret != CORE_OFFSET_UNSPECIFIED)
4489 		return (ret);
4490 
4491 	if (ctx->ifc_sysctl_separate_txrx)
4492 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4493 	else
4494 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4495 
4496 	mtx_lock(&cpu_offset_mtx);
4497 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4498 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4499 			ret = op->offset;
4500 			op->offset += qc;
4501 			MPASS(op->refcount < UINT_MAX);
4502 			op->refcount++;
4503 			break;
4504 		}
4505 	}
4506 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4507 		ret = 0;
4508 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4509 		    M_NOWAIT | M_ZERO);
4510 		if (op == NULL) {
4511 			device_printf(ctx->ifc_dev,
4512 			    "allocation for cpu offset failed.\n");
4513 		} else {
4514 			op->offset = qc;
4515 			op->refcount = 1;
4516 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4517 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4518 		}
4519 	}
4520 	mtx_unlock(&cpu_offset_mtx);
4521 
4522 	return (ret);
4523 }
4524 
4525 static void
4526 unref_ctx_core_offset(if_ctx_t ctx)
4527 {
4528 	struct cpu_offset *op, *top;
4529 
4530 	mtx_lock(&cpu_offset_mtx);
4531 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4532 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4533 			MPASS(op->refcount > 0);
4534 			op->refcount--;
4535 			if (op->refcount == 0) {
4536 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4537 				free(op, M_IFLIB);
4538 			}
4539 			break;
4540 		}
4541 	}
4542 	mtx_unlock(&cpu_offset_mtx);
4543 }
4544 
4545 int
4546 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4547 {
4548 	int err, rid, msix;
4549 	if_ctx_t ctx;
4550 	if_t ifp;
4551 	if_softc_ctx_t scctx;
4552 	int i;
4553 	uint16_t main_txq;
4554 	uint16_t main_rxq;
4555 
4556 
4557 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4558 
4559 	if (sc == NULL) {
4560 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4561 		device_set_softc(dev, ctx);
4562 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4563 	}
4564 
4565 	ctx->ifc_sctx = sctx;
4566 	ctx->ifc_dev = dev;
4567 	ctx->ifc_softc = sc;
4568 
4569 	if ((err = iflib_register(ctx)) != 0) {
4570 		device_printf(dev, "iflib_register failed %d\n", err);
4571 		goto fail_ctx_free;
4572 	}
4573 	iflib_add_device_sysctl_pre(ctx);
4574 
4575 	scctx = &ctx->ifc_softc_ctx;
4576 	ifp = ctx->ifc_ifp;
4577 
4578 	iflib_reset_qvalues(ctx);
4579 	CTX_LOCK(ctx);
4580 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4581 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4582 		goto fail_unlock;
4583 	}
4584 	_iflib_pre_assert(scctx);
4585 	ctx->ifc_txrx = *scctx->isc_txrx;
4586 
4587 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4588 		ctx->ifc_mediap = scctx->isc_media;
4589 
4590 #ifdef INVARIANTS
4591 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4592 		MPASS(scctx->isc_tx_csum_flags);
4593 #endif
4594 
4595 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4596 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4597 
4598 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4599 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4600 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4601 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4602 
4603 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4604 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4605 
4606 	/* XXX change for per-queue sizes */
4607 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4608 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4609 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4610 		if (!powerof2(scctx->isc_nrxd[i])) {
4611 			/* round down instead? */
4612 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4613 			err = EINVAL;
4614 			goto fail_iflib_detach;
4615 		}
4616 	}
4617 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4618 		if (!powerof2(scctx->isc_ntxd[i])) {
4619 			device_printf(dev,
4620 			    "# tx descriptors must be a power of 2");
4621 			err = EINVAL;
4622 			goto fail_iflib_detach;
4623 		}
4624 	}
4625 
4626 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4627 	    MAX_SINGLE_PACKET_FRACTION)
4628 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4629 		    MAX_SINGLE_PACKET_FRACTION);
4630 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4631 	    MAX_SINGLE_PACKET_FRACTION)
4632 		scctx->isc_tx_tso_segments_max = max(1,
4633 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4634 
4635 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4636 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4637 		/*
4638 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4639 		 * but some MACs do.
4640 		 */
4641 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4642 		    IP_MAXPACKET));
4643 		/*
4644 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4645 		 * into account.  In the worst case, each of these calls will
4646 		 * add another mbuf and, thus, the requirement for another DMA
4647 		 * segment.  So for best performance, it doesn't make sense to
4648 		 * advertize a maximum of TSO segments that typically will
4649 		 * require defragmentation in iflib_encap().
4650 		 */
4651 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4652 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4653 	}
4654 	if (scctx->isc_rss_table_size == 0)
4655 		scctx->isc_rss_table_size = 64;
4656 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4657 
4658 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4659 	/* XXX format name */
4660 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4661 	    NULL, NULL, "admin");
4662 
4663 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4664 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4665 		device_printf(dev, "Unable to fetch CPU list\n");
4666 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4667 	}
4668 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4669 
4670 	/*
4671 	** Now set up MSI or MSI-X, should return us the number of supported
4672 	** vectors (will be 1 for a legacy interrupt and MSI).
4673 	*/
4674 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4675 		msix = scctx->isc_vectors;
4676 	} else if (scctx->isc_msix_bar != 0)
4677 	       /*
4678 		* The simple fact that isc_msix_bar is not 0 does not mean we
4679 		* we have a good value there that is known to work.
4680 		*/
4681 		msix = iflib_msix_init(ctx);
4682 	else {
4683 		scctx->isc_vectors = 1;
4684 		scctx->isc_ntxqsets = 1;
4685 		scctx->isc_nrxqsets = 1;
4686 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4687 		msix = 0;
4688 	}
4689 	/* Get memory for the station queues */
4690 	if ((err = iflib_queues_alloc(ctx))) {
4691 		device_printf(dev, "Unable to allocate queue memory\n");
4692 		goto fail_intr_free;
4693 	}
4694 
4695 	if ((err = iflib_qset_structures_setup(ctx)))
4696 		goto fail_queues;
4697 
4698 	/*
4699 	 * Now that we know how many queues there are, get the core offset.
4700 	 */
4701 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4702 
4703 	/*
4704 	 * Group taskqueues aren't properly set up until SMP is started,
4705 	 * so we disable interrupts until we can handle them post
4706 	 * SI_SUB_SMP.
4707 	 *
4708 	 * XXX: disabling interrupts doesn't actually work, at least for
4709 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4710 	 * we do null handling and depend on this not causing too large an
4711 	 * interrupt storm.
4712 	 */
4713 	IFDI_INTR_DISABLE(ctx);
4714 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4715 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4716 		goto fail_queues;
4717 	}
4718 	if (msix <= 1) {
4719 		rid = 0;
4720 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4721 			MPASS(msix == 1);
4722 			rid = 1;
4723 		}
4724 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4725 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4726 			goto fail_queues;
4727 		}
4728 	}
4729 
4730 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4731 
4732 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4733 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4734 		goto fail_detach;
4735 	}
4736 
4737 	/*
4738 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4739 	 * This must appear after the call to ether_ifattach() because
4740 	 * ether_ifattach() sets if_hdrlen to the default value.
4741 	 */
4742 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4743 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4744 
4745 	if ((err = iflib_netmap_attach(ctx))) {
4746 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4747 		goto fail_detach;
4748 	}
4749 	*ctxp = ctx;
4750 
4751 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4752 
4753 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4754 	iflib_add_device_sysctl_post(ctx);
4755 	iflib_add_pfil(ctx);
4756 	ctx->ifc_flags |= IFC_INIT_DONE;
4757 	CTX_UNLOCK(ctx);
4758 	return (0);
4759 
4760 fail_detach:
4761 	ether_ifdetach(ctx->ifc_ifp);
4762 fail_intr_free:
4763 	iflib_free_intr_mem(ctx);
4764 fail_queues:
4765 	iflib_tx_structures_free(ctx);
4766 	iflib_rx_structures_free(ctx);
4767 fail_iflib_detach:
4768 	IFDI_DETACH(ctx);
4769 fail_unlock:
4770 	CTX_UNLOCK(ctx);
4771 fail_ctx_free:
4772         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4773                 free(ctx->ifc_softc, M_IFLIB);
4774         free(ctx, M_IFLIB);
4775 	return (err);
4776 }
4777 
4778 int
4779 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4780 					  struct iflib_cloneattach_ctx *clctx)
4781 {
4782 	int err;
4783 	if_ctx_t ctx;
4784 	if_t ifp;
4785 	if_softc_ctx_t scctx;
4786 	int i;
4787 	void *sc;
4788 	uint16_t main_txq;
4789 	uint16_t main_rxq;
4790 
4791 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4792 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4793 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4794 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4795 		ctx->ifc_flags |= IFC_PSEUDO;
4796 
4797 	ctx->ifc_sctx = sctx;
4798 	ctx->ifc_softc = sc;
4799 	ctx->ifc_dev = dev;
4800 
4801 	if ((err = iflib_register(ctx)) != 0) {
4802 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4803 		goto fail_ctx_free;
4804 	}
4805 	iflib_add_device_sysctl_pre(ctx);
4806 
4807 	scctx = &ctx->ifc_softc_ctx;
4808 	ifp = ctx->ifc_ifp;
4809 
4810 	/*
4811 	 * XXX sanity check that ntxd & nrxd are a power of 2
4812 	 */
4813 	iflib_reset_qvalues(ctx);
4814 	CTX_LOCK(ctx);
4815 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4816 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4817 		goto fail_unlock;
4818 	}
4819 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4820 		ether_gen_addr(ifp, &ctx->ifc_mac);
4821 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4822 								clctx->cc_params)) != 0) {
4823 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4824 		goto fail_ctx_free;
4825 	}
4826 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4827 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4828 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4829 
4830 #ifdef INVARIANTS
4831 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4832 		MPASS(scctx->isc_tx_csum_flags);
4833 #endif
4834 
4835 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4836 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4837 
4838 	ifp->if_flags |= IFF_NOGROUP;
4839 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4840 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4841 
4842 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4843 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4844 			goto fail_detach;
4845 		}
4846 		*ctxp = ctx;
4847 
4848 		/*
4849 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4850 		 * This must appear after the call to ether_ifattach() because
4851 		 * ether_ifattach() sets if_hdrlen to the default value.
4852 		 */
4853 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4854 			if_setifheaderlen(ifp,
4855 			    sizeof(struct ether_vlan_header));
4856 
4857 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4858 		iflib_add_device_sysctl_post(ctx);
4859 		ctx->ifc_flags |= IFC_INIT_DONE;
4860 		return (0);
4861 	}
4862 	_iflib_pre_assert(scctx);
4863 	ctx->ifc_txrx = *scctx->isc_txrx;
4864 
4865 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4866 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4867 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4868 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4869 
4870 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4871 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4872 
4873 	/* XXX change for per-queue sizes */
4874 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4875 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4876 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4877 		if (!powerof2(scctx->isc_nrxd[i])) {
4878 			/* round down instead? */
4879 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4880 			err = EINVAL;
4881 			goto fail_iflib_detach;
4882 		}
4883 	}
4884 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4885 		if (!powerof2(scctx->isc_ntxd[i])) {
4886 			device_printf(dev,
4887 			    "# tx descriptors must be a power of 2");
4888 			err = EINVAL;
4889 			goto fail_iflib_detach;
4890 		}
4891 	}
4892 
4893 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4894 	    MAX_SINGLE_PACKET_FRACTION)
4895 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4896 		    MAX_SINGLE_PACKET_FRACTION);
4897 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4898 	    MAX_SINGLE_PACKET_FRACTION)
4899 		scctx->isc_tx_tso_segments_max = max(1,
4900 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4901 
4902 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4903 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4904 		/*
4905 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4906 		 * but some MACs do.
4907 		 */
4908 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4909 		    IP_MAXPACKET));
4910 		/*
4911 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4912 		 * into account.  In the worst case, each of these calls will
4913 		 * add another mbuf and, thus, the requirement for another DMA
4914 		 * segment.  So for best performance, it doesn't make sense to
4915 		 * advertize a maximum of TSO segments that typically will
4916 		 * require defragmentation in iflib_encap().
4917 		 */
4918 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4919 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4920 	}
4921 	if (scctx->isc_rss_table_size == 0)
4922 		scctx->isc_rss_table_size = 64;
4923 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4924 
4925 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4926 	/* XXX format name */
4927 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4928 	    NULL, NULL, "admin");
4929 
4930 	/* XXX --- can support > 1 -- but keep it simple for now */
4931 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4932 
4933 	/* Get memory for the station queues */
4934 	if ((err = iflib_queues_alloc(ctx))) {
4935 		device_printf(dev, "Unable to allocate queue memory\n");
4936 		goto fail_iflib_detach;
4937 	}
4938 
4939 	if ((err = iflib_qset_structures_setup(ctx))) {
4940 		device_printf(dev, "qset structure setup failed %d\n", err);
4941 		goto fail_queues;
4942 	}
4943 
4944 	/*
4945 	 * XXX What if anything do we want to do about interrupts?
4946 	 */
4947 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4948 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4949 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4950 		goto fail_detach;
4951 	}
4952 
4953 	/*
4954 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4955 	 * This must appear after the call to ether_ifattach() because
4956 	 * ether_ifattach() sets if_hdrlen to the default value.
4957 	 */
4958 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4959 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4960 
4961 	/* XXX handle more than one queue */
4962 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4963 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4964 
4965 	*ctxp = ctx;
4966 
4967 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4968 	iflib_add_device_sysctl_post(ctx);
4969 	ctx->ifc_flags |= IFC_INIT_DONE;
4970 	CTX_UNLOCK(ctx);
4971 	return (0);
4972 fail_detach:
4973 	ether_ifdetach(ctx->ifc_ifp);
4974 fail_queues:
4975 	iflib_tx_structures_free(ctx);
4976 	iflib_rx_structures_free(ctx);
4977 fail_iflib_detach:
4978 	IFDI_DETACH(ctx);
4979 fail_unlock:
4980 	CTX_UNLOCK(ctx);
4981 fail_ctx_free:
4982 	free(ctx->ifc_softc, M_IFLIB);
4983 	free(ctx, M_IFLIB);
4984 	return (err);
4985 }
4986 
4987 int
4988 iflib_pseudo_deregister(if_ctx_t ctx)
4989 {
4990 	if_t ifp = ctx->ifc_ifp;
4991 	iflib_txq_t txq;
4992 	iflib_rxq_t rxq;
4993 	int i, j;
4994 	struct taskqgroup *tqg;
4995 	iflib_fl_t fl;
4996 
4997 	/* Unregister VLAN events */
4998 	if (ctx->ifc_vlan_attach_event != NULL)
4999 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5000 	if (ctx->ifc_vlan_detach_event != NULL)
5001 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5002 
5003 	ether_ifdetach(ifp);
5004 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5005 	CTX_LOCK_DESTROY(ctx);
5006 	/* XXX drain any dependent tasks */
5007 	tqg = qgroup_if_io_tqg;
5008 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5009 		callout_drain(&txq->ift_timer);
5010 		if (txq->ift_task.gt_uniq != NULL)
5011 			taskqgroup_detach(tqg, &txq->ift_task);
5012 	}
5013 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5014 		if (rxq->ifr_task.gt_uniq != NULL)
5015 			taskqgroup_detach(tqg, &rxq->ifr_task);
5016 
5017 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5018 			free(fl->ifl_rx_bitmap, M_IFLIB);
5019 	}
5020 	tqg = qgroup_if_config_tqg;
5021 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5022 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5023 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5024 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5025 
5026 	if_free(ifp);
5027 
5028 	iflib_tx_structures_free(ctx);
5029 	iflib_rx_structures_free(ctx);
5030 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5031 		free(ctx->ifc_softc, M_IFLIB);
5032 	free(ctx, M_IFLIB);
5033 	return (0);
5034 }
5035 
5036 int
5037 iflib_device_attach(device_t dev)
5038 {
5039 	if_ctx_t ctx;
5040 	if_shared_ctx_t sctx;
5041 
5042 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5043 		return (ENOTSUP);
5044 
5045 	pci_enable_busmaster(dev);
5046 
5047 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5048 }
5049 
5050 int
5051 iflib_device_deregister(if_ctx_t ctx)
5052 {
5053 	if_t ifp = ctx->ifc_ifp;
5054 	iflib_txq_t txq;
5055 	iflib_rxq_t rxq;
5056 	device_t dev = ctx->ifc_dev;
5057 	int i, j;
5058 	struct taskqgroup *tqg;
5059 	iflib_fl_t fl;
5060 
5061 	/* Make sure VLANS are not using driver */
5062 	if (if_vlantrunkinuse(ifp)) {
5063 		device_printf(dev, "Vlan in use, detach first\n");
5064 		return (EBUSY);
5065 	}
5066 #ifdef PCI_IOV
5067 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5068 		device_printf(dev, "SR-IOV in use; detach first.\n");
5069 		return (EBUSY);
5070 	}
5071 #endif
5072 
5073 	STATE_LOCK(ctx);
5074 	ctx->ifc_flags |= IFC_IN_DETACH;
5075 	STATE_UNLOCK(ctx);
5076 
5077 	CTX_LOCK(ctx);
5078 	iflib_stop(ctx);
5079 	CTX_UNLOCK(ctx);
5080 
5081 	/* Unregister VLAN events */
5082 	if (ctx->ifc_vlan_attach_event != NULL)
5083 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5084 	if (ctx->ifc_vlan_detach_event != NULL)
5085 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5086 
5087 	iflib_netmap_detach(ifp);
5088 	ether_ifdetach(ifp);
5089 	iflib_rem_pfil(ctx);
5090 	if (ctx->ifc_led_dev != NULL)
5091 		led_destroy(ctx->ifc_led_dev);
5092 	/* XXX drain any dependent tasks */
5093 	tqg = qgroup_if_io_tqg;
5094 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5095 		callout_drain(&txq->ift_timer);
5096 		if (txq->ift_task.gt_uniq != NULL)
5097 			taskqgroup_detach(tqg, &txq->ift_task);
5098 	}
5099 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5100 		if (rxq->ifr_task.gt_uniq != NULL)
5101 			taskqgroup_detach(tqg, &rxq->ifr_task);
5102 
5103 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5104 			free(fl->ifl_rx_bitmap, M_IFLIB);
5105 	}
5106 	tqg = qgroup_if_config_tqg;
5107 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5108 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5109 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5110 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5111 	CTX_LOCK(ctx);
5112 	IFDI_DETACH(ctx);
5113 	CTX_UNLOCK(ctx);
5114 
5115 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5116 	CTX_LOCK_DESTROY(ctx);
5117 	device_set_softc(ctx->ifc_dev, NULL);
5118 	iflib_free_intr_mem(ctx);
5119 
5120 	bus_generic_detach(dev);
5121 	if_free(ifp);
5122 
5123 	iflib_tx_structures_free(ctx);
5124 	iflib_rx_structures_free(ctx);
5125 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5126 		free(ctx->ifc_softc, M_IFLIB);
5127 	unref_ctx_core_offset(ctx);
5128 	STATE_LOCK_DESTROY(ctx);
5129 	free(ctx, M_IFLIB);
5130 	return (0);
5131 }
5132 
5133 static void
5134 iflib_free_intr_mem(if_ctx_t ctx)
5135 {
5136 
5137 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5138 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5139 	}
5140 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5141 		pci_release_msi(ctx->ifc_dev);
5142 	}
5143 	if (ctx->ifc_msix_mem != NULL) {
5144 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5145 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5146 		ctx->ifc_msix_mem = NULL;
5147 	}
5148 }
5149 
5150 int
5151 iflib_device_detach(device_t dev)
5152 {
5153 	if_ctx_t ctx = device_get_softc(dev);
5154 
5155 	return (iflib_device_deregister(ctx));
5156 }
5157 
5158 int
5159 iflib_device_suspend(device_t dev)
5160 {
5161 	if_ctx_t ctx = device_get_softc(dev);
5162 
5163 	CTX_LOCK(ctx);
5164 	IFDI_SUSPEND(ctx);
5165 	CTX_UNLOCK(ctx);
5166 
5167 	return bus_generic_suspend(dev);
5168 }
5169 int
5170 iflib_device_shutdown(device_t dev)
5171 {
5172 	if_ctx_t ctx = device_get_softc(dev);
5173 
5174 	CTX_LOCK(ctx);
5175 	IFDI_SHUTDOWN(ctx);
5176 	CTX_UNLOCK(ctx);
5177 
5178 	return bus_generic_suspend(dev);
5179 }
5180 
5181 
5182 int
5183 iflib_device_resume(device_t dev)
5184 {
5185 	if_ctx_t ctx = device_get_softc(dev);
5186 	iflib_txq_t txq = ctx->ifc_txqs;
5187 
5188 	CTX_LOCK(ctx);
5189 	IFDI_RESUME(ctx);
5190 	iflib_if_init_locked(ctx);
5191 	CTX_UNLOCK(ctx);
5192 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5193 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5194 
5195 	return (bus_generic_resume(dev));
5196 }
5197 
5198 int
5199 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5200 {
5201 	int error;
5202 	if_ctx_t ctx = device_get_softc(dev);
5203 
5204 	CTX_LOCK(ctx);
5205 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5206 	CTX_UNLOCK(ctx);
5207 
5208 	return (error);
5209 }
5210 
5211 void
5212 iflib_device_iov_uninit(device_t dev)
5213 {
5214 	if_ctx_t ctx = device_get_softc(dev);
5215 
5216 	CTX_LOCK(ctx);
5217 	IFDI_IOV_UNINIT(ctx);
5218 	CTX_UNLOCK(ctx);
5219 }
5220 
5221 int
5222 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5223 {
5224 	int error;
5225 	if_ctx_t ctx = device_get_softc(dev);
5226 
5227 	CTX_LOCK(ctx);
5228 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5229 	CTX_UNLOCK(ctx);
5230 
5231 	return (error);
5232 }
5233 
5234 /*********************************************************************
5235  *
5236  *  MODULE FUNCTION DEFINITIONS
5237  *
5238  **********************************************************************/
5239 
5240 /*
5241  * - Start a fast taskqueue thread for each core
5242  * - Start a taskqueue for control operations
5243  */
5244 static int
5245 iflib_module_init(void)
5246 {
5247 	return (0);
5248 }
5249 
5250 static int
5251 iflib_module_event_handler(module_t mod, int what, void *arg)
5252 {
5253 	int err;
5254 
5255 	switch (what) {
5256 	case MOD_LOAD:
5257 		if ((err = iflib_module_init()) != 0)
5258 			return (err);
5259 		break;
5260 	case MOD_UNLOAD:
5261 		return (EBUSY);
5262 	default:
5263 		return (EOPNOTSUPP);
5264 	}
5265 
5266 	return (0);
5267 }
5268 
5269 /*********************************************************************
5270  *
5271  *  PUBLIC FUNCTION DEFINITIONS
5272  *     ordered as in iflib.h
5273  *
5274  **********************************************************************/
5275 
5276 
5277 static void
5278 _iflib_assert(if_shared_ctx_t sctx)
5279 {
5280 	MPASS(sctx->isc_tx_maxsize);
5281 	MPASS(sctx->isc_tx_maxsegsize);
5282 
5283 	MPASS(sctx->isc_rx_maxsize);
5284 	MPASS(sctx->isc_rx_nsegments);
5285 	MPASS(sctx->isc_rx_maxsegsize);
5286 
5287 	MPASS(sctx->isc_nrxd_min[0]);
5288 	MPASS(sctx->isc_nrxd_max[0]);
5289 	MPASS(sctx->isc_nrxd_default[0]);
5290 	MPASS(sctx->isc_ntxd_min[0]);
5291 	MPASS(sctx->isc_ntxd_max[0]);
5292 	MPASS(sctx->isc_ntxd_default[0]);
5293 }
5294 
5295 static void
5296 _iflib_pre_assert(if_softc_ctx_t scctx)
5297 {
5298 
5299 	MPASS(scctx->isc_txrx->ift_txd_encap);
5300 	MPASS(scctx->isc_txrx->ift_txd_flush);
5301 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5302 	MPASS(scctx->isc_txrx->ift_rxd_available);
5303 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5304 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5305 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5306 }
5307 
5308 static int
5309 iflib_register(if_ctx_t ctx)
5310 {
5311 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5312 	driver_t *driver = sctx->isc_driver;
5313 	device_t dev = ctx->ifc_dev;
5314 	if_t ifp;
5315 
5316 	_iflib_assert(sctx);
5317 
5318 	CTX_LOCK_INIT(ctx);
5319 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5320 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5321 	if (ifp == NULL) {
5322 		device_printf(dev, "can not allocate ifnet structure\n");
5323 		return (ENOMEM);
5324 	}
5325 
5326 	/*
5327 	 * Initialize our context's device specific methods
5328 	 */
5329 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5330 	kobj_class_compile((kobj_class_t) driver);
5331 	driver->refs++;
5332 
5333 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5334 	if_setsoftc(ifp, ctx);
5335 	if_setdev(ifp, dev);
5336 	if_setinitfn(ifp, iflib_if_init);
5337 	if_setioctlfn(ifp, iflib_if_ioctl);
5338 #ifdef ALTQ
5339 	if_setstartfn(ifp, iflib_altq_if_start);
5340 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5341 	if_setsendqready(ifp);
5342 #else
5343 	if_settransmitfn(ifp, iflib_if_transmit);
5344 #endif
5345 	if_setqflushfn(ifp, iflib_if_qflush);
5346 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5347 
5348 	ctx->ifc_vlan_attach_event =
5349 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5350 							  EVENTHANDLER_PRI_FIRST);
5351 	ctx->ifc_vlan_detach_event =
5352 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5353 							  EVENTHANDLER_PRI_FIRST);
5354 
5355 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5356 		ctx->ifc_mediap = &ctx->ifc_media;
5357 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5358 		    iflib_media_change, iflib_media_status);
5359 	}
5360 	return (0);
5361 }
5362 
5363 
5364 static int
5365 iflib_queues_alloc(if_ctx_t ctx)
5366 {
5367 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5368 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5369 	device_t dev = ctx->ifc_dev;
5370 	int nrxqsets = scctx->isc_nrxqsets;
5371 	int ntxqsets = scctx->isc_ntxqsets;
5372 	iflib_txq_t txq;
5373 	iflib_rxq_t rxq;
5374 	iflib_fl_t fl = NULL;
5375 	int i, j, cpu, err, txconf, rxconf;
5376 	iflib_dma_info_t ifdip;
5377 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5378 	uint32_t *txqsizes = scctx->isc_txqsizes;
5379 	uint8_t nrxqs = sctx->isc_nrxqs;
5380 	uint8_t ntxqs = sctx->isc_ntxqs;
5381 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5382 	caddr_t *vaddrs;
5383 	uint64_t *paddrs;
5384 
5385 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5386 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5387 
5388 	/* Allocate the TX ring struct memory */
5389 	if (!(ctx->ifc_txqs =
5390 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5391 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5392 		device_printf(dev, "Unable to allocate TX ring memory\n");
5393 		err = ENOMEM;
5394 		goto fail;
5395 	}
5396 
5397 	/* Now allocate the RX */
5398 	if (!(ctx->ifc_rxqs =
5399 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5400 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5401 		device_printf(dev, "Unable to allocate RX ring memory\n");
5402 		err = ENOMEM;
5403 		goto rx_fail;
5404 	}
5405 
5406 	txq = ctx->ifc_txqs;
5407 	rxq = ctx->ifc_rxqs;
5408 
5409 	/*
5410 	 * XXX handle allocation failure
5411 	 */
5412 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5413 		/* Set up some basics */
5414 
5415 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5416 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5417 			device_printf(dev,
5418 			    "Unable to allocate TX DMA info memory\n");
5419 			err = ENOMEM;
5420 			goto err_tx_desc;
5421 		}
5422 		txq->ift_ifdi = ifdip;
5423 		for (j = 0; j < ntxqs; j++, ifdip++) {
5424 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5425 				device_printf(dev,
5426 				    "Unable to allocate TX descriptors\n");
5427 				err = ENOMEM;
5428 				goto err_tx_desc;
5429 			}
5430 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5431 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5432 		}
5433 		txq->ift_ctx = ctx;
5434 		txq->ift_id = i;
5435 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5436 			txq->ift_br_offset = 1;
5437 		} else {
5438 			txq->ift_br_offset = 0;
5439 		}
5440 		/* XXX fix this */
5441 		txq->ift_timer.c_cpu = cpu;
5442 
5443 		if (iflib_txsd_alloc(txq)) {
5444 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5445 			err = ENOMEM;
5446 			goto err_tx_desc;
5447 		}
5448 
5449 		/* Initialize the TX lock */
5450 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5451 		    device_get_nameunit(dev), txq->ift_id);
5452 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5453 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5454 
5455 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5456 			 device_get_nameunit(dev), txq->ift_id);
5457 
5458 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5459 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5460 		if (err) {
5461 			/* XXX free any allocated rings */
5462 			device_printf(dev, "Unable to allocate buf_ring\n");
5463 			goto err_tx_desc;
5464 		}
5465 	}
5466 
5467 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5468 		/* Set up some basics */
5469 
5470 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5471 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5472 			device_printf(dev,
5473 			    "Unable to allocate RX DMA info memory\n");
5474 			err = ENOMEM;
5475 			goto err_tx_desc;
5476 		}
5477 
5478 		rxq->ifr_ifdi = ifdip;
5479 		/* XXX this needs to be changed if #rx queues != #tx queues */
5480 		rxq->ifr_ntxqirq = 1;
5481 		rxq->ifr_txqid[0] = i;
5482 		for (j = 0; j < nrxqs; j++, ifdip++) {
5483 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5484 				device_printf(dev,
5485 				    "Unable to allocate RX descriptors\n");
5486 				err = ENOMEM;
5487 				goto err_tx_desc;
5488 			}
5489 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5490 		}
5491 		rxq->ifr_ctx = ctx;
5492 		rxq->ifr_id = i;
5493 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5494 			rxq->ifr_fl_offset = 1;
5495 		} else {
5496 			rxq->ifr_fl_offset = 0;
5497 		}
5498 		rxq->ifr_nfl = nfree_lists;
5499 		if (!(fl =
5500 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5501 			device_printf(dev, "Unable to allocate free list memory\n");
5502 			err = ENOMEM;
5503 			goto err_tx_desc;
5504 		}
5505 		rxq->ifr_fl = fl;
5506 		for (j = 0; j < nfree_lists; j++) {
5507 			fl[j].ifl_rxq = rxq;
5508 			fl[j].ifl_id = j;
5509 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5510 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5511 		}
5512 		/* Allocate receive buffers for the ring */
5513 		if (iflib_rxsd_alloc(rxq)) {
5514 			device_printf(dev,
5515 			    "Critical Failure setting up receive buffers\n");
5516 			err = ENOMEM;
5517 			goto err_rx_desc;
5518 		}
5519 
5520 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5521 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5522 			    M_WAITOK);
5523 	}
5524 
5525 	/* TXQs */
5526 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5527 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5528 	for (i = 0; i < ntxqsets; i++) {
5529 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5530 
5531 		for (j = 0; j < ntxqs; j++, di++) {
5532 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5533 			paddrs[i*ntxqs + j] = di->idi_paddr;
5534 		}
5535 	}
5536 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5537 		device_printf(ctx->ifc_dev,
5538 		    "Unable to allocate device TX queue\n");
5539 		iflib_tx_structures_free(ctx);
5540 		free(vaddrs, M_IFLIB);
5541 		free(paddrs, M_IFLIB);
5542 		goto err_rx_desc;
5543 	}
5544 	free(vaddrs, M_IFLIB);
5545 	free(paddrs, M_IFLIB);
5546 
5547 	/* RXQs */
5548 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5549 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5550 	for (i = 0; i < nrxqsets; i++) {
5551 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5552 
5553 		for (j = 0; j < nrxqs; j++, di++) {
5554 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5555 			paddrs[i*nrxqs + j] = di->idi_paddr;
5556 		}
5557 	}
5558 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5559 		device_printf(ctx->ifc_dev,
5560 		    "Unable to allocate device RX queue\n");
5561 		iflib_tx_structures_free(ctx);
5562 		free(vaddrs, M_IFLIB);
5563 		free(paddrs, M_IFLIB);
5564 		goto err_rx_desc;
5565 	}
5566 	free(vaddrs, M_IFLIB);
5567 	free(paddrs, M_IFLIB);
5568 
5569 	return (0);
5570 
5571 /* XXX handle allocation failure changes */
5572 err_rx_desc:
5573 err_tx_desc:
5574 rx_fail:
5575 	if (ctx->ifc_rxqs != NULL)
5576 		free(ctx->ifc_rxqs, M_IFLIB);
5577 	ctx->ifc_rxqs = NULL;
5578 	if (ctx->ifc_txqs != NULL)
5579 		free(ctx->ifc_txqs, M_IFLIB);
5580 	ctx->ifc_txqs = NULL;
5581 fail:
5582 	return (err);
5583 }
5584 
5585 static int
5586 iflib_tx_structures_setup(if_ctx_t ctx)
5587 {
5588 	iflib_txq_t txq = ctx->ifc_txqs;
5589 	int i;
5590 
5591 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5592 		iflib_txq_setup(txq);
5593 
5594 	return (0);
5595 }
5596 
5597 static void
5598 iflib_tx_structures_free(if_ctx_t ctx)
5599 {
5600 	iflib_txq_t txq = ctx->ifc_txqs;
5601 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5602 	int i, j;
5603 
5604 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5605 		iflib_txq_destroy(txq);
5606 		for (j = 0; j < sctx->isc_ntxqs; j++)
5607 			iflib_dma_free(&txq->ift_ifdi[j]);
5608 	}
5609 	free(ctx->ifc_txqs, M_IFLIB);
5610 	ctx->ifc_txqs = NULL;
5611 	IFDI_QUEUES_FREE(ctx);
5612 }
5613 
5614 /*********************************************************************
5615  *
5616  *  Initialize all receive rings.
5617  *
5618  **********************************************************************/
5619 static int
5620 iflib_rx_structures_setup(if_ctx_t ctx)
5621 {
5622 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5623 	int q;
5624 #if defined(INET6) || defined(INET)
5625 	int i, err;
5626 #endif
5627 
5628 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5629 #if defined(INET6) || defined(INET)
5630 		tcp_lro_free(&rxq->ifr_lc);
5631 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5632 		    TCP_LRO_ENTRIES, min(1024,
5633 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5634 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5635 			goto fail;
5636 		}
5637 		rxq->ifr_lro_enabled = TRUE;
5638 #endif
5639 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5640 	}
5641 	return (0);
5642 #if defined(INET6) || defined(INET)
5643 fail:
5644 	/*
5645 	 * Free RX software descriptors allocated so far, we will only handle
5646 	 * the rings that completed, the failing case will have
5647 	 * cleaned up for itself. 'q' failed, so its the terminus.
5648 	 */
5649 	rxq = ctx->ifc_rxqs;
5650 	for (i = 0; i < q; ++i, rxq++) {
5651 		iflib_rx_sds_free(rxq);
5652 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5653 	}
5654 	return (err);
5655 #endif
5656 }
5657 
5658 /*********************************************************************
5659  *
5660  *  Free all receive rings.
5661  *
5662  **********************************************************************/
5663 static void
5664 iflib_rx_structures_free(if_ctx_t ctx)
5665 {
5666 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5667 
5668 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5669 		iflib_rx_sds_free(rxq);
5670 	}
5671 	free(ctx->ifc_rxqs, M_IFLIB);
5672 	ctx->ifc_rxqs = NULL;
5673 }
5674 
5675 static int
5676 iflib_qset_structures_setup(if_ctx_t ctx)
5677 {
5678 	int err;
5679 
5680 	/*
5681 	 * It is expected that the caller takes care of freeing queues if this
5682 	 * fails.
5683 	 */
5684 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5685 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5686 		return (err);
5687 	}
5688 
5689 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5690 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5691 
5692 	return (err);
5693 }
5694 
5695 int
5696 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5697 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5698 {
5699 
5700 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5701 }
5702 
5703 #ifdef SMP
5704 static int
5705 find_nth(if_ctx_t ctx, int qid)
5706 {
5707 	cpuset_t cpus;
5708 	int i, cpuid, eqid, count;
5709 
5710 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5711 	count = CPU_COUNT(&cpus);
5712 	eqid = qid % count;
5713 	/* clear up to the qid'th bit */
5714 	for (i = 0; i < eqid; i++) {
5715 		cpuid = CPU_FFS(&cpus);
5716 		MPASS(cpuid != 0);
5717 		CPU_CLR(cpuid-1, &cpus);
5718 	}
5719 	cpuid = CPU_FFS(&cpus);
5720 	MPASS(cpuid != 0);
5721 	return (cpuid-1);
5722 }
5723 
5724 #ifdef SCHED_ULE
5725 extern struct cpu_group *cpu_top;              /* CPU topology */
5726 
5727 static int
5728 find_child_with_core(int cpu, struct cpu_group *grp)
5729 {
5730 	int i;
5731 
5732 	if (grp->cg_children == 0)
5733 		return -1;
5734 
5735 	MPASS(grp->cg_child);
5736 	for (i = 0; i < grp->cg_children; i++) {
5737 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5738 			return i;
5739 	}
5740 
5741 	return -1;
5742 }
5743 
5744 /*
5745  * Find the nth "close" core to the specified core
5746  * "close" is defined as the deepest level that shares
5747  * at least an L2 cache.  With threads, this will be
5748  * threads on the same core.  If the shared cache is L3
5749  * or higher, simply returns the same core.
5750  */
5751 static int
5752 find_close_core(int cpu, int core_offset)
5753 {
5754 	struct cpu_group *grp;
5755 	int i;
5756 	int fcpu;
5757 	cpuset_t cs;
5758 
5759 	grp = cpu_top;
5760 	if (grp == NULL)
5761 		return cpu;
5762 	i = 0;
5763 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5764 		/* If the child only has one cpu, don't descend */
5765 		if (grp->cg_child[i].cg_count <= 1)
5766 			break;
5767 		grp = &grp->cg_child[i];
5768 	}
5769 
5770 	/* If they don't share at least an L2 cache, use the same CPU */
5771 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5772 		return cpu;
5773 
5774 	/* Now pick one */
5775 	CPU_COPY(&grp->cg_mask, &cs);
5776 
5777 	/* Add the selected CPU offset to core offset. */
5778 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5779 		if (fcpu - 1 == cpu)
5780 			break;
5781 		CPU_CLR(fcpu - 1, &cs);
5782 	}
5783 	MPASS(fcpu);
5784 
5785 	core_offset += i;
5786 
5787 	CPU_COPY(&grp->cg_mask, &cs);
5788 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5789 		MPASS(CPU_FFS(&cs));
5790 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5791 	}
5792 	MPASS(CPU_FFS(&cs));
5793 	return CPU_FFS(&cs) - 1;
5794 }
5795 #else
5796 static int
5797 find_close_core(int cpu, int core_offset __unused)
5798 {
5799 	return cpu;
5800 }
5801 #endif
5802 
5803 static int
5804 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5805 {
5806 	switch (type) {
5807 	case IFLIB_INTR_TX:
5808 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5809 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5810 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5811 	case IFLIB_INTR_RX:
5812 	case IFLIB_INTR_RXTX:
5813 		/* RX queues get the specified core */
5814 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5815 	default:
5816 		return -1;
5817 	}
5818 }
5819 #else
5820 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5821 #define find_close_core(cpuid, tid)	CPU_FIRST()
5822 #define find_nth(ctx, gid)		CPU_FIRST()
5823 #endif
5824 
5825 /* Just to avoid copy/paste */
5826 static inline int
5827 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5828     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5829     const char *name)
5830 {
5831 	device_t dev;
5832 	int co, cpuid, err, tid;
5833 
5834 	dev = ctx->ifc_dev;
5835 	co = ctx->ifc_sysctl_core_offset;
5836 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5837 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5838 	cpuid = find_nth(ctx, qid + co);
5839 	tid = get_core_offset(ctx, type, qid);
5840 	MPASS(tid >= 0);
5841 	cpuid = find_close_core(cpuid, tid);
5842 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5843 	    name);
5844 	if (err) {
5845 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5846 		return (err);
5847 	}
5848 #ifdef notyet
5849 	if (cpuid > ctx->ifc_cpuid_highest)
5850 		ctx->ifc_cpuid_highest = cpuid;
5851 #endif
5852 	return 0;
5853 }
5854 
5855 int
5856 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5857 			iflib_intr_type_t type, driver_filter_t *filter,
5858 			void *filter_arg, int qid, const char *name)
5859 {
5860 	device_t dev;
5861 	struct grouptask *gtask;
5862 	struct taskqgroup *tqg;
5863 	iflib_filter_info_t info;
5864 	gtask_fn_t *fn;
5865 	int tqrid, err;
5866 	driver_filter_t *intr_fast;
5867 	void *q;
5868 
5869 	info = &ctx->ifc_filter_info;
5870 	tqrid = rid;
5871 
5872 	switch (type) {
5873 	/* XXX merge tx/rx for netmap? */
5874 	case IFLIB_INTR_TX:
5875 		q = &ctx->ifc_txqs[qid];
5876 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5877 		gtask = &ctx->ifc_txqs[qid].ift_task;
5878 		tqg = qgroup_if_io_tqg;
5879 		fn = _task_fn_tx;
5880 		intr_fast = iflib_fast_intr;
5881 		GROUPTASK_INIT(gtask, 0, fn, q);
5882 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5883 		break;
5884 	case IFLIB_INTR_RX:
5885 		q = &ctx->ifc_rxqs[qid];
5886 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5887 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5888 		tqg = qgroup_if_io_tqg;
5889 		fn = _task_fn_rx;
5890 		intr_fast = iflib_fast_intr;
5891 		GROUPTASK_INIT(gtask, 0, fn, q);
5892 		break;
5893 	case IFLIB_INTR_RXTX:
5894 		q = &ctx->ifc_rxqs[qid];
5895 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5896 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5897 		tqg = qgroup_if_io_tqg;
5898 		fn = _task_fn_rx;
5899 		intr_fast = iflib_fast_intr_rxtx;
5900 		GROUPTASK_INIT(gtask, 0, fn, q);
5901 		break;
5902 	case IFLIB_INTR_ADMIN:
5903 		q = ctx;
5904 		tqrid = -1;
5905 		info = &ctx->ifc_filter_info;
5906 		gtask = &ctx->ifc_admin_task;
5907 		tqg = qgroup_if_config_tqg;
5908 		fn = _task_fn_admin;
5909 		intr_fast = iflib_fast_intr_ctx;
5910 		break;
5911 	default:
5912 		panic("unknown net intr type");
5913 	}
5914 
5915 	info->ifi_filter = filter;
5916 	info->ifi_filter_arg = filter_arg;
5917 	info->ifi_task = gtask;
5918 	info->ifi_ctx = q;
5919 
5920 	dev = ctx->ifc_dev;
5921 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5922 	if (err != 0) {
5923 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5924 		return (err);
5925 	}
5926 	if (type == IFLIB_INTR_ADMIN)
5927 		return (0);
5928 
5929 	if (tqrid != -1) {
5930 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5931 		    q, name);
5932 		if (err)
5933 			return (err);
5934 	} else {
5935 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5936 	}
5937 
5938 	return (0);
5939 }
5940 
5941 void
5942 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5943 {
5944 	struct grouptask *gtask;
5945 	struct taskqgroup *tqg;
5946 	gtask_fn_t *fn;
5947 	void *q;
5948 	int err;
5949 
5950 	switch (type) {
5951 	case IFLIB_INTR_TX:
5952 		q = &ctx->ifc_txqs[qid];
5953 		gtask = &ctx->ifc_txqs[qid].ift_task;
5954 		tqg = qgroup_if_io_tqg;
5955 		fn = _task_fn_tx;
5956 		break;
5957 	case IFLIB_INTR_RX:
5958 		q = &ctx->ifc_rxqs[qid];
5959 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5960 		tqg = qgroup_if_io_tqg;
5961 		fn = _task_fn_rx;
5962 		break;
5963 	case IFLIB_INTR_IOV:
5964 		q = ctx;
5965 		gtask = &ctx->ifc_vflr_task;
5966 		tqg = qgroup_if_config_tqg;
5967 		fn = _task_fn_iov;
5968 		break;
5969 	default:
5970 		panic("unknown net intr type");
5971 	}
5972 	GROUPTASK_INIT(gtask, 0, fn, q);
5973 	if (irq != NULL) {
5974 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5975 		    q, name);
5976 		if (err)
5977 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5978 			    irq->ii_res, name);
5979 	} else {
5980 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5981 	}
5982 }
5983 
5984 void
5985 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5986 {
5987 
5988 	if (irq->ii_tag)
5989 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5990 
5991 	if (irq->ii_res)
5992 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5993 		    rman_get_rid(irq->ii_res), irq->ii_res);
5994 }
5995 
5996 static int
5997 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5998 {
5999 	iflib_txq_t txq = ctx->ifc_txqs;
6000 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6001 	if_irq_t irq = &ctx->ifc_legacy_irq;
6002 	iflib_filter_info_t info;
6003 	device_t dev;
6004 	struct grouptask *gtask;
6005 	struct resource *res;
6006 	struct taskqgroup *tqg;
6007 	gtask_fn_t *fn;
6008 	int tqrid;
6009 	void *q;
6010 	int err;
6011 
6012 	q = &ctx->ifc_rxqs[0];
6013 	info = &rxq[0].ifr_filter_info;
6014 	gtask = &rxq[0].ifr_task;
6015 	tqg = qgroup_if_io_tqg;
6016 	tqrid = irq->ii_rid = *rid;
6017 	fn = _task_fn_rx;
6018 
6019 	ctx->ifc_flags |= IFC_LEGACY;
6020 	info->ifi_filter = filter;
6021 	info->ifi_filter_arg = filter_arg;
6022 	info->ifi_task = gtask;
6023 	info->ifi_ctx = ctx;
6024 
6025 	dev = ctx->ifc_dev;
6026 	/* We allocate a single interrupt resource */
6027 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
6028 		return (err);
6029 	GROUPTASK_INIT(gtask, 0, fn, q);
6030 	res = irq->ii_res;
6031 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6032 
6033 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6034 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6035 	    "tx");
6036 	return (0);
6037 }
6038 
6039 void
6040 iflib_led_create(if_ctx_t ctx)
6041 {
6042 
6043 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6044 	    device_get_nameunit(ctx->ifc_dev));
6045 }
6046 
6047 void
6048 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6049 {
6050 
6051 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6052 }
6053 
6054 void
6055 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6056 {
6057 
6058 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6059 }
6060 
6061 void
6062 iflib_admin_intr_deferred(if_ctx_t ctx)
6063 {
6064 #ifdef INVARIANTS
6065 	struct grouptask *gtask;
6066 
6067 	gtask = &ctx->ifc_admin_task;
6068 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6069 #endif
6070 
6071 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6072 }
6073 
6074 void
6075 iflib_iov_intr_deferred(if_ctx_t ctx)
6076 {
6077 
6078 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6079 }
6080 
6081 void
6082 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
6083 {
6084 
6085 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6086 	    name);
6087 }
6088 
6089 void
6090 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6091 	const char *name)
6092 {
6093 
6094 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6095 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6096 	    name);
6097 }
6098 
6099 void
6100 iflib_config_gtask_deinit(struct grouptask *gtask)
6101 {
6102 
6103 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6104 }
6105 
6106 void
6107 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6108 {
6109 	if_t ifp = ctx->ifc_ifp;
6110 	iflib_txq_t txq = ctx->ifc_txqs;
6111 
6112 	if_setbaudrate(ifp, baudrate);
6113 	if (baudrate >= IF_Gbps(10)) {
6114 		STATE_LOCK(ctx);
6115 		ctx->ifc_flags |= IFC_PREFETCH;
6116 		STATE_UNLOCK(ctx);
6117 	}
6118 	/* If link down, disable watchdog */
6119 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6120 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6121 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6122 	}
6123 	ctx->ifc_link_state = link_state;
6124 	if_link_state_change(ifp, link_state);
6125 }
6126 
6127 static int
6128 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6129 {
6130 	int credits;
6131 #ifdef INVARIANTS
6132 	int credits_pre = txq->ift_cidx_processed;
6133 #endif
6134 
6135 	if (ctx->isc_txd_credits_update == NULL)
6136 		return (0);
6137 
6138 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6139 	    BUS_DMASYNC_POSTREAD);
6140 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6141 		return (0);
6142 
6143 	txq->ift_processed += credits;
6144 	txq->ift_cidx_processed += credits;
6145 
6146 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6147 	if (txq->ift_cidx_processed >= txq->ift_size)
6148 		txq->ift_cidx_processed -= txq->ift_size;
6149 	return (credits);
6150 }
6151 
6152 static int
6153 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6154 {
6155 	iflib_fl_t fl;
6156 	u_int i;
6157 
6158 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6159 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6160 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6161 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6162 	    budget));
6163 }
6164 
6165 void
6166 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6167 	const char *description, if_int_delay_info_t info,
6168 	int offset, int value)
6169 {
6170 	info->iidi_ctx = ctx;
6171 	info->iidi_offset = offset;
6172 	info->iidi_value = value;
6173 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6174 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6175 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6176 	    info, 0, iflib_sysctl_int_delay, "I", description);
6177 }
6178 
6179 struct sx *
6180 iflib_ctx_lock_get(if_ctx_t ctx)
6181 {
6182 
6183 	return (&ctx->ifc_ctx_sx);
6184 }
6185 
6186 static int
6187 iflib_msix_init(if_ctx_t ctx)
6188 {
6189 	device_t dev = ctx->ifc_dev;
6190 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6191 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6192 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6193 	int iflib_num_tx_queues, iflib_num_rx_queues;
6194 	int err, admincnt, bar;
6195 
6196 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6197 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6198 
6199 	if (bootverbose)
6200 		device_printf(dev, "msix_init qsets capped at %d\n",
6201 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6202 
6203 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6204 	admincnt = sctx->isc_admin_intrcnt;
6205 	/* Override by tuneable */
6206 	if (scctx->isc_disable_msix)
6207 		goto msi;
6208 
6209 	/* First try MSI-X */
6210 	if ((msgs = pci_msix_count(dev)) == 0) {
6211 		if (bootverbose)
6212 			device_printf(dev, "MSI-X not supported or disabled\n");
6213 		goto msi;
6214 	}
6215 	/*
6216 	 * bar == -1 => "trust me I know what I'm doing"
6217 	 * Some drivers are for hardware that is so shoddily
6218 	 * documented that no one knows which bars are which
6219 	 * so the developer has to map all bars. This hack
6220 	 * allows shoddy garbage to use MSI-X in this framework.
6221 	 */
6222 	if (bar != -1) {
6223 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6224 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6225 		if (ctx->ifc_msix_mem == NULL) {
6226 			device_printf(dev, "Unable to map MSI-X table\n");
6227 			goto msi;
6228 		}
6229 	}
6230 #if IFLIB_DEBUG
6231 	/* use only 1 qset in debug mode */
6232 	queuemsgs = min(msgs - admincnt, 1);
6233 #else
6234 	queuemsgs = msgs - admincnt;
6235 #endif
6236 #ifdef RSS
6237 	queues = imin(queuemsgs, rss_getnumbuckets());
6238 #else
6239 	queues = queuemsgs;
6240 #endif
6241 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6242 	if (bootverbose)
6243 		device_printf(dev,
6244 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6245 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6246 #ifdef  RSS
6247 	/* If we're doing RSS, clamp at the number of RSS buckets */
6248 	if (queues > rss_getnumbuckets())
6249 		queues = rss_getnumbuckets();
6250 #endif
6251 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6252 		rx_queues = iflib_num_rx_queues;
6253 	else
6254 		rx_queues = queues;
6255 
6256 	if (rx_queues > scctx->isc_nrxqsets)
6257 		rx_queues = scctx->isc_nrxqsets;
6258 
6259 	/*
6260 	 * We want this to be all logical CPUs by default
6261 	 */
6262 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6263 		tx_queues = iflib_num_tx_queues;
6264 	else
6265 		tx_queues = mp_ncpus;
6266 
6267 	if (tx_queues > scctx->isc_ntxqsets)
6268 		tx_queues = scctx->isc_ntxqsets;
6269 
6270 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6271 #ifdef INVARIANTS
6272 		if (tx_queues != rx_queues)
6273 			device_printf(dev,
6274 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6275 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6276 #endif
6277 		tx_queues = min(rx_queues, tx_queues);
6278 		rx_queues = min(rx_queues, tx_queues);
6279 	}
6280 
6281 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6282 	    rx_queues, tx_queues);
6283 
6284 	vectors = rx_queues + admincnt;
6285 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6286 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6287 		    vectors);
6288 		scctx->isc_vectors = vectors;
6289 		scctx->isc_nrxqsets = rx_queues;
6290 		scctx->isc_ntxqsets = tx_queues;
6291 		scctx->isc_intr = IFLIB_INTR_MSIX;
6292 
6293 		return (vectors);
6294 	} else {
6295 		device_printf(dev,
6296 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6297 		    vectors, err);
6298 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6299 		    ctx->ifc_msix_mem);
6300 		ctx->ifc_msix_mem = NULL;
6301 	}
6302 msi:
6303 	vectors = pci_msi_count(dev);
6304 	scctx->isc_nrxqsets = 1;
6305 	scctx->isc_ntxqsets = 1;
6306 	scctx->isc_vectors = vectors;
6307 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6308 		device_printf(dev,"Using an MSI interrupt\n");
6309 		scctx->isc_intr = IFLIB_INTR_MSI;
6310 	} else {
6311 		scctx->isc_vectors = 1;
6312 		device_printf(dev,"Using a Legacy interrupt\n");
6313 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6314 	}
6315 
6316 	return (vectors);
6317 }
6318 
6319 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6320 
6321 static int
6322 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6323 {
6324 	int rc;
6325 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6326 	struct sbuf *sb;
6327 	const char *ring_state = "UNKNOWN";
6328 
6329 	/* XXX needed ? */
6330 	rc = sysctl_wire_old_buffer(req, 0);
6331 	MPASS(rc == 0);
6332 	if (rc != 0)
6333 		return (rc);
6334 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6335 	MPASS(sb != NULL);
6336 	if (sb == NULL)
6337 		return (ENOMEM);
6338 	if (state[3] <= 3)
6339 		ring_state = ring_states[state[3]];
6340 
6341 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6342 		    state[0], state[1], state[2], ring_state);
6343 	rc = sbuf_finish(sb);
6344 	sbuf_delete(sb);
6345         return(rc);
6346 }
6347 
6348 enum iflib_ndesc_handler {
6349 	IFLIB_NTXD_HANDLER,
6350 	IFLIB_NRXD_HANDLER,
6351 };
6352 
6353 static int
6354 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6355 {
6356 	if_ctx_t ctx = (void *)arg1;
6357 	enum iflib_ndesc_handler type = arg2;
6358 	char buf[256] = {0};
6359 	qidx_t *ndesc;
6360 	char *p, *next;
6361 	int nqs, rc, i;
6362 
6363 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6364 
6365 	nqs = 8;
6366 	switch(type) {
6367 	case IFLIB_NTXD_HANDLER:
6368 		ndesc = ctx->ifc_sysctl_ntxds;
6369 		if (ctx->ifc_sctx)
6370 			nqs = ctx->ifc_sctx->isc_ntxqs;
6371 		break;
6372 	case IFLIB_NRXD_HANDLER:
6373 		ndesc = ctx->ifc_sysctl_nrxds;
6374 		if (ctx->ifc_sctx)
6375 			nqs = ctx->ifc_sctx->isc_nrxqs;
6376 		break;
6377 	default:
6378 			panic("unhandled type");
6379 	}
6380 	if (nqs == 0)
6381 		nqs = 8;
6382 
6383 	for (i=0; i<8; i++) {
6384 		if (i >= nqs)
6385 			break;
6386 		if (i)
6387 			strcat(buf, ",");
6388 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6389 	}
6390 
6391 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6392 	if (rc || req->newptr == NULL)
6393 		return rc;
6394 
6395 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6396 	    i++, p = strsep(&next, " ,")) {
6397 		ndesc[i] = strtoul(p, NULL, 10);
6398 	}
6399 
6400 	return(rc);
6401 }
6402 
6403 #define NAME_BUFLEN 32
6404 static void
6405 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6406 {
6407         device_t dev = iflib_get_dev(ctx);
6408 	struct sysctl_oid_list *child, *oid_list;
6409 	struct sysctl_ctx_list *ctx_list;
6410 	struct sysctl_oid *node;
6411 
6412 	ctx_list = device_get_sysctl_ctx(dev);
6413 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6414 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6415 						      CTLFLAG_RD, NULL, "IFLIB fields");
6416 	oid_list = SYSCTL_CHILDREN(node);
6417 
6418 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6419 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6420 		       "driver version");
6421 
6422 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6423 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6424 			"# of txqs to use, 0 => use default #");
6425 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6426 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6427 			"# of rxqs to use, 0 => use default #");
6428 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6429 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6430                        "permit #txq != #rxq");
6431 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6432                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6433                       "disable MSI-X (default 0)");
6434 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6435 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6436                        "set the rx budget");
6437 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6438 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6439 		       "cause tx to abdicate instead of running to completion");
6440 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6441 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6442 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6443 		       "offset to start using cores at");
6444 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6445 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6446 		       "use separate cores for TX and RX");
6447 
6448 	/* XXX change for per-queue sizes */
6449 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6450 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6451                        mp_ndesc_handler, "A",
6452                        "list of # of tx descriptors to use, 0 = use default #");
6453 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6454 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6455                        mp_ndesc_handler, "A",
6456                        "list of # of rx descriptors to use, 0 = use default #");
6457 }
6458 
6459 static void
6460 iflib_add_device_sysctl_post(if_ctx_t ctx)
6461 {
6462 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6463 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6464         device_t dev = iflib_get_dev(ctx);
6465 	struct sysctl_oid_list *child;
6466 	struct sysctl_ctx_list *ctx_list;
6467 	iflib_fl_t fl;
6468 	iflib_txq_t txq;
6469 	iflib_rxq_t rxq;
6470 	int i, j;
6471 	char namebuf[NAME_BUFLEN];
6472 	char *qfmt;
6473 	struct sysctl_oid *queue_node, *fl_node, *node;
6474 	struct sysctl_oid_list *queue_list, *fl_list;
6475 	ctx_list = device_get_sysctl_ctx(dev);
6476 
6477 	node = ctx->ifc_sysctl_node;
6478 	child = SYSCTL_CHILDREN(node);
6479 
6480 	if (scctx->isc_ntxqsets > 100)
6481 		qfmt = "txq%03d";
6482 	else if (scctx->isc_ntxqsets > 10)
6483 		qfmt = "txq%02d";
6484 	else
6485 		qfmt = "txq%d";
6486 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6487 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6488 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6489 					     CTLFLAG_RD, NULL, "Queue Name");
6490 		queue_list = SYSCTL_CHILDREN(queue_node);
6491 #if MEMORY_LOGGING
6492 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6493 				CTLFLAG_RD,
6494 				&txq->ift_dequeued, "total mbufs freed");
6495 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6496 				CTLFLAG_RD,
6497 				&txq->ift_enqueued, "total mbufs enqueued");
6498 #endif
6499 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6500 				   CTLFLAG_RD,
6501 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6502 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6503 				   CTLFLAG_RD,
6504 				   &txq->ift_pullups, "# of times m_pullup was called");
6505 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6506 				   CTLFLAG_RD,
6507 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6508 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6509 				   CTLFLAG_RD,
6510 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6511 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6512 				   CTLFLAG_RD,
6513 				   &txq->ift_map_failed, "# of times dma map failed");
6514 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6515 				   CTLFLAG_RD,
6516 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6517 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6518 				   CTLFLAG_RD,
6519 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6520 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6521 				   CTLFLAG_RD,
6522 				   &txq->ift_pidx, 1, "Producer Index");
6523 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6524 				   CTLFLAG_RD,
6525 				   &txq->ift_cidx, 1, "Consumer Index");
6526 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6527 				   CTLFLAG_RD,
6528 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6529 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6530 				   CTLFLAG_RD,
6531 				   &txq->ift_in_use, 1, "descriptors in use");
6532 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6533 				   CTLFLAG_RD,
6534 				   &txq->ift_processed, "descriptors procesed for clean");
6535 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6536 				   CTLFLAG_RD,
6537 				   &txq->ift_cleaned, "total cleaned");
6538 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6539 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6540 				0, mp_ring_state_handler, "A", "soft ring state");
6541 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6542 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6543 				       "# of enqueues to the mp_ring for this queue");
6544 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6545 				       CTLFLAG_RD, &txq->ift_br->drops,
6546 				       "# of drops in the mp_ring for this queue");
6547 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6548 				       CTLFLAG_RD, &txq->ift_br->starts,
6549 				       "# of normal consumer starts in the mp_ring for this queue");
6550 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6551 				       CTLFLAG_RD, &txq->ift_br->stalls,
6552 					       "# of consumer stalls in the mp_ring for this queue");
6553 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6554 			       CTLFLAG_RD, &txq->ift_br->restarts,
6555 				       "# of consumer restarts in the mp_ring for this queue");
6556 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6557 				       CTLFLAG_RD, &txq->ift_br->abdications,
6558 				       "# of consumer abdications in the mp_ring for this queue");
6559 	}
6560 
6561 	if (scctx->isc_nrxqsets > 100)
6562 		qfmt = "rxq%03d";
6563 	else if (scctx->isc_nrxqsets > 10)
6564 		qfmt = "rxq%02d";
6565 	else
6566 		qfmt = "rxq%d";
6567 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6568 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6569 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6570 					     CTLFLAG_RD, NULL, "Queue Name");
6571 		queue_list = SYSCTL_CHILDREN(queue_node);
6572 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6573 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6574 				       CTLFLAG_RD,
6575 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6576 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6577 				       CTLFLAG_RD,
6578 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6579 		}
6580 
6581 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6582 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6583 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6584 						     CTLFLAG_RD, NULL, "freelist Name");
6585 			fl_list = SYSCTL_CHILDREN(fl_node);
6586 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6587 				       CTLFLAG_RD,
6588 				       &fl->ifl_pidx, 1, "Producer Index");
6589 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6590 				       CTLFLAG_RD,
6591 				       &fl->ifl_cidx, 1, "Consumer Index");
6592 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6593 				       CTLFLAG_RD,
6594 				       &fl->ifl_credits, 1, "credits available");
6595 #if MEMORY_LOGGING
6596 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6597 					CTLFLAG_RD,
6598 					&fl->ifl_m_enqueued, "mbufs allocated");
6599 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6600 					CTLFLAG_RD,
6601 					&fl->ifl_m_dequeued, "mbufs freed");
6602 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6603 					CTLFLAG_RD,
6604 					&fl->ifl_cl_enqueued, "clusters allocated");
6605 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6606 					CTLFLAG_RD,
6607 					&fl->ifl_cl_dequeued, "clusters freed");
6608 #endif
6609 
6610 		}
6611 	}
6612 
6613 }
6614 
6615 void
6616 iflib_request_reset(if_ctx_t ctx)
6617 {
6618 
6619 	STATE_LOCK(ctx);
6620 	ctx->ifc_flags |= IFC_DO_RESET;
6621 	STATE_UNLOCK(ctx);
6622 }
6623 
6624 #ifndef __NO_STRICT_ALIGNMENT
6625 static struct mbuf *
6626 iflib_fixup_rx(struct mbuf *m)
6627 {
6628 	struct mbuf *n;
6629 
6630 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6631 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6632 		m->m_data += ETHER_HDR_LEN;
6633 		n = m;
6634 	} else {
6635 		MGETHDR(n, M_NOWAIT, MT_DATA);
6636 		if (n == NULL) {
6637 			m_freem(m);
6638 			return (NULL);
6639 		}
6640 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6641 		m->m_data += ETHER_HDR_LEN;
6642 		m->m_len -= ETHER_HDR_LEN;
6643 		n->m_len = ETHER_HDR_LEN;
6644 		M_MOVE_PKTHDR(n, m);
6645 		n->m_next = m;
6646 	}
6647 	return (n);
6648 }
6649 #endif
6650 
6651 #ifdef NETDUMP
6652 static void
6653 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6654 {
6655 	if_ctx_t ctx;
6656 
6657 	ctx = if_getsoftc(ifp);
6658 	CTX_LOCK(ctx);
6659 	*nrxr = NRXQSETS(ctx);
6660 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6661 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6662 	CTX_UNLOCK(ctx);
6663 }
6664 
6665 static void
6666 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6667 {
6668 	if_ctx_t ctx;
6669 	if_softc_ctx_t scctx;
6670 	iflib_fl_t fl;
6671 	iflib_rxq_t rxq;
6672 	int i, j;
6673 
6674 	ctx = if_getsoftc(ifp);
6675 	scctx = &ctx->ifc_softc_ctx;
6676 
6677 	switch (event) {
6678 	case NETDUMP_START:
6679 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6680 			rxq = &ctx->ifc_rxqs[i];
6681 			for (j = 0; j < rxq->ifr_nfl; j++) {
6682 				fl = rxq->ifr_fl;
6683 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6684 			}
6685 		}
6686 		iflib_no_tx_batch = 1;
6687 		break;
6688 	default:
6689 		break;
6690 	}
6691 }
6692 
6693 static int
6694 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6695 {
6696 	if_ctx_t ctx;
6697 	iflib_txq_t txq;
6698 	int error;
6699 
6700 	ctx = if_getsoftc(ifp);
6701 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6702 	    IFF_DRV_RUNNING)
6703 		return (EBUSY);
6704 
6705 	txq = &ctx->ifc_txqs[0];
6706 	error = iflib_encap(txq, &m);
6707 	if (error == 0)
6708 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6709 	return (error);
6710 }
6711 
6712 static int
6713 iflib_netdump_poll(struct ifnet *ifp, int count)
6714 {
6715 	if_ctx_t ctx;
6716 	if_softc_ctx_t scctx;
6717 	iflib_txq_t txq;
6718 	int i;
6719 
6720 	ctx = if_getsoftc(ifp);
6721 	scctx = &ctx->ifc_softc_ctx;
6722 
6723 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6724 	    IFF_DRV_RUNNING)
6725 		return (EBUSY);
6726 
6727 	txq = &ctx->ifc_txqs[0];
6728 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6729 
6730 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6731 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6732 	return (0);
6733 }
6734 #endif /* NETDUMP */
6735