xref: /freebsd/sys/net/iflib.c (revision c1903b634c3db81f3283bc508e1feeb3fc8316db)
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/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.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_watchdog_events;
176 	struct cdev *ifc_led_dev;
177 	struct resource *ifc_msix_mem;
178 
179 	struct if_irq ifc_legacy_irq;
180 	struct grouptask ifc_admin_task;
181 	struct grouptask ifc_vflr_task;
182 	struct iflib_filter_info ifc_filter_info;
183 	struct ifmedia	ifc_media;
184 	struct ifmedia	*ifc_mediap;
185 
186 	struct sysctl_oid *ifc_sysctl_node;
187 	uint16_t ifc_sysctl_ntxqs;
188 	uint16_t ifc_sysctl_nrxqs;
189 	uint16_t ifc_sysctl_qs_eq_override;
190 	uint16_t ifc_sysctl_rx_budget;
191 	uint16_t ifc_sysctl_tx_abdicate;
192 	uint16_t ifc_sysctl_core_offset;
193 #define	CORE_OFFSET_UNSPECIFIED	0xffff
194 	uint8_t  ifc_sysctl_separate_txrx;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	struct ether_addr ifc_mac;
212 };
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (ctx->ifc_mediap);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 #define IFLIB_RX_COPY_THRESH		128
292 #define IFLIB_MAX_RX_REFRESH		32
293 /* The minimum descriptors per second before we start coalescing */
294 #define IFLIB_MIN_DESC_SEC		16384
295 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
296 #define IFLIB_QUEUE_IDLE		0
297 #define IFLIB_QUEUE_HUNG		1
298 #define IFLIB_QUEUE_WORKING		2
299 /* maximum number of txqs that can share an rx interrupt */
300 #define IFLIB_MAX_TX_SHARED_INTR	4
301 
302 /* this should really scale with ring size - this is a fairly arbitrary value */
303 #define TX_BATCH_SIZE			32
304 
305 #define IFLIB_RESTART_BUDGET		8
306 
307 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
308 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
309 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
310 
311 struct iflib_txq {
312 	qidx_t		ift_in_use;
313 	qidx_t		ift_cidx;
314 	qidx_t		ift_cidx_processed;
315 	qidx_t		ift_pidx;
316 	uint8_t		ift_gen;
317 	uint8_t		ift_br_offset;
318 	uint16_t	ift_npending;
319 	uint16_t	ift_db_pending;
320 	uint16_t	ift_rs_pending;
321 	/* implicit pad */
322 	uint8_t		ift_txd_size[8];
323 	uint64_t	ift_processed;
324 	uint64_t	ift_cleaned;
325 	uint64_t	ift_cleaned_prev;
326 #if MEMORY_LOGGING
327 	uint64_t	ift_enqueued;
328 	uint64_t	ift_dequeued;
329 #endif
330 	uint64_t	ift_no_tx_dma_setup;
331 	uint64_t	ift_no_desc_avail;
332 	uint64_t	ift_mbuf_defrag_failed;
333 	uint64_t	ift_mbuf_defrag;
334 	uint64_t	ift_map_failed;
335 	uint64_t	ift_txd_encap_efbig;
336 	uint64_t	ift_pullups;
337 	uint64_t	ift_last_timer_tick;
338 
339 	struct mtx	ift_mtx;
340 	struct mtx	ift_db_mtx;
341 
342 	/* constant values */
343 	if_ctx_t	ift_ctx;
344 	struct ifmp_ring        *ift_br;
345 	struct grouptask	ift_task;
346 	qidx_t		ift_size;
347 	uint16_t	ift_id;
348 	struct callout	ift_timer;
349 
350 	if_txsd_vec_t	ift_sds;
351 	uint8_t		ift_qstatus;
352 	uint8_t		ift_closed;
353 	uint8_t		ift_update_freq;
354 	struct iflib_filter_info ift_filter_info;
355 	bus_dma_tag_t	ift_buf_tag;
356 	bus_dma_tag_t	ift_tso_buf_tag;
357 	iflib_dma_info_t	ift_ifdi;
358 #define MTX_NAME_LEN 16
359 	char                    ift_mtx_name[MTX_NAME_LEN];
360 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
361 #ifdef IFLIB_DIAGNOSTICS
362 	uint64_t ift_cpu_exec_count[256];
363 #endif
364 } __aligned(CACHE_LINE_SIZE);
365 
366 struct iflib_fl {
367 	qidx_t		ifl_cidx;
368 	qidx_t		ifl_pidx;
369 	qidx_t		ifl_credits;
370 	uint8_t		ifl_gen;
371 	uint8_t		ifl_rxd_size;
372 #if MEMORY_LOGGING
373 	uint64_t	ifl_m_enqueued;
374 	uint64_t	ifl_m_dequeued;
375 	uint64_t	ifl_cl_enqueued;
376 	uint64_t	ifl_cl_dequeued;
377 #endif
378 	/* implicit pad */
379 	bitstr_t 	*ifl_rx_bitmap;
380 	qidx_t		ifl_fragidx;
381 	/* constant */
382 	qidx_t		ifl_size;
383 	uint16_t	ifl_buf_size;
384 	uint16_t	ifl_cltype;
385 	uma_zone_t	ifl_zone;
386 	iflib_rxsd_array_t	ifl_sds;
387 	iflib_rxq_t	ifl_rxq;
388 	uint8_t		ifl_id;
389 	bus_dma_tag_t	ifl_buf_tag;
390 	iflib_dma_info_t	ifl_ifdi;
391 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
392 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
393 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
394 }  __aligned(CACHE_LINE_SIZE);
395 
396 static inline qidx_t
397 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
398 {
399 	qidx_t used;
400 
401 	if (pidx > cidx)
402 		used = pidx - cidx;
403 	else if (pidx < cidx)
404 		used = size - cidx + pidx;
405 	else if (gen == 0 && pidx == cidx)
406 		used = 0;
407 	else if (gen == 1 && pidx == cidx)
408 		used = size;
409 	else
410 		panic("bad state");
411 
412 	return (used);
413 }
414 
415 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
416 
417 #define IDXDIFF(head, tail, wrap) \
418 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
419 
420 struct iflib_rxq {
421 	if_ctx_t	ifr_ctx;
422 	iflib_fl_t	ifr_fl;
423 	uint64_t	ifr_rx_irq;
424 	struct pfil_head	*pfil;
425 	/*
426 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
427 	 * the command queue consumer index.  Otherwise it's unused.
428 	 */
429 	qidx_t		ifr_cq_cidx;
430 	uint16_t	ifr_id;
431 	uint8_t		ifr_nfl;
432 	uint8_t		ifr_ntxqirq;
433 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
434 	uint8_t		ifr_fl_offset;
435 	struct lro_ctrl			ifr_lc;
436 	struct grouptask        ifr_task;
437 	struct iflib_filter_info ifr_filter_info;
438 	iflib_dma_info_t		ifr_ifdi;
439 
440 	/* dynamically allocate if any drivers need a value substantially larger than this */
441 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
442 #ifdef IFLIB_DIAGNOSTICS
443 	uint64_t ifr_cpu_exec_count[256];
444 #endif
445 }  __aligned(CACHE_LINE_SIZE);
446 
447 typedef struct if_rxsd {
448 	caddr_t *ifsd_cl;
449 	iflib_fl_t ifsd_fl;
450 	qidx_t ifsd_cidx;
451 } *if_rxsd_t;
452 
453 /* multiple of word size */
454 #ifdef __LP64__
455 #define PKT_INFO_SIZE	6
456 #define RXD_INFO_SIZE	5
457 #define PKT_TYPE uint64_t
458 #else
459 #define PKT_INFO_SIZE	11
460 #define RXD_INFO_SIZE	8
461 #define PKT_TYPE uint32_t
462 #endif
463 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
464 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
465 
466 typedef struct if_pkt_info_pad {
467 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
468 } *if_pkt_info_pad_t;
469 typedef struct if_rxd_info_pad {
470 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
471 } *if_rxd_info_pad_t;
472 
473 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
474 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
475 
476 
477 static inline void
478 pkt_info_zero(if_pkt_info_t pi)
479 {
480 	if_pkt_info_pad_t pi_pad;
481 
482 	pi_pad = (if_pkt_info_pad_t)pi;
483 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
484 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
485 #ifndef __LP64__
486 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
487 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
488 #endif
489 }
490 
491 static device_method_t iflib_pseudo_methods[] = {
492 	DEVMETHOD(device_attach, noop_attach),
493 	DEVMETHOD(device_detach, iflib_pseudo_detach),
494 	DEVMETHOD_END
495 };
496 
497 driver_t iflib_pseudodriver = {
498 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
499 };
500 
501 static inline void
502 rxd_info_zero(if_rxd_info_t ri)
503 {
504 	if_rxd_info_pad_t ri_pad;
505 	int i;
506 
507 	ri_pad = (if_rxd_info_pad_t)ri;
508 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
509 		ri_pad->rxd_val[i] = 0;
510 		ri_pad->rxd_val[i+1] = 0;
511 		ri_pad->rxd_val[i+2] = 0;
512 		ri_pad->rxd_val[i+3] = 0;
513 	}
514 #ifdef __LP64__
515 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
516 #endif
517 }
518 
519 /*
520  * Only allow a single packet to take up most 1/nth of the tx ring
521  */
522 #define MAX_SINGLE_PACKET_FRACTION 12
523 #define IF_BAD_DMA (bus_addr_t)-1
524 
525 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
526 
527 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
528 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
529 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
531 
532 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
533 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
534 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
535 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
539 
540 void
541 iflib_set_detach(if_ctx_t ctx)
542 {
543 	STATE_LOCK(ctx);
544 	ctx->ifc_flags |= IFC_IN_DETACH;
545 	STATE_UNLOCK(ctx);
546 }
547 
548 /* Our boot-time initialization hook */
549 static int	iflib_module_event_handler(module_t, int, void *);
550 
551 static moduledata_t iflib_moduledata = {
552 	"iflib",
553 	iflib_module_event_handler,
554 	NULL
555 };
556 
557 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
558 MODULE_VERSION(iflib, 1);
559 
560 MODULE_DEPEND(iflib, pci, 1, 1, 1);
561 MODULE_DEPEND(iflib, ether, 1, 1, 1);
562 
563 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
564 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
565 
566 #ifndef IFLIB_DEBUG_COUNTERS
567 #ifdef INVARIANTS
568 #define IFLIB_DEBUG_COUNTERS 1
569 #else
570 #define IFLIB_DEBUG_COUNTERS 0
571 #endif /* !INVARIANTS */
572 #endif
573 
574 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
575                    "iflib driver parameters");
576 
577 /*
578  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
579  */
580 static int iflib_min_tx_latency = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
582 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
583 static int iflib_no_tx_batch = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
585 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
586 
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# TX mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# TX mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# TX frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# RX allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
619 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
621 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
623 		   &iflib_txq_drain_notready, 0, "# drain notready");
624 
625 
626 static int iflib_encap_load_mbuf_fail;
627 static int iflib_encap_pad_mbuf_fail;
628 static int iflib_encap_txq_avail_fail;
629 static int iflib_encap_txd_encap_fail;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
632 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
634 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
636 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
638 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
639 
640 static int iflib_task_fn_rxs;
641 static int iflib_rx_intr_enables;
642 static int iflib_fast_intrs;
643 static int iflib_rx_unavail;
644 static int iflib_rx_ctx_inactive;
645 static int iflib_rx_if_input;
646 static int iflib_rxd_flush;
647 
648 static int iflib_verbose_debug;
649 
650 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
651 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
653 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
654 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
655 		   &iflib_fast_intrs, 0, "# fast_intr calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
657 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
659 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
661 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
663 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
664 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
665 		   &iflib_verbose_debug, 0, "enable verbose debugging");
666 
667 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
668 static void
669 iflib_debug_reset(void)
670 {
671 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
672 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
673 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
674 		iflib_txq_drain_notready =
675 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
676 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
677 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
678 		iflib_rx_unavail =
679 		iflib_rx_ctx_inactive = iflib_rx_if_input =
680 		iflib_rxd_flush = 0;
681 }
682 
683 #else
684 #define DBG_COUNTER_INC(name)
685 static void iflib_debug_reset(void) {}
686 #endif
687 
688 #define IFLIB_DEBUG 0
689 
690 static void iflib_tx_structures_free(if_ctx_t ctx);
691 static void iflib_rx_structures_free(if_ctx_t ctx);
692 static int iflib_queues_alloc(if_ctx_t ctx);
693 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
694 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
695 static int iflib_qset_structures_setup(if_ctx_t ctx);
696 static int iflib_msix_init(if_ctx_t ctx);
697 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
698 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
699 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
700 #ifdef ALTQ
701 static void iflib_altq_if_start(if_t ifp);
702 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
703 #endif
704 static int iflib_register(if_ctx_t);
705 static void iflib_deregister(if_ctx_t);
706 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
707 static void iflib_init_locked(if_ctx_t ctx);
708 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
709 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
710 static void iflib_ifmp_purge(iflib_txq_t txq);
711 static void _iflib_pre_assert(if_softc_ctx_t scctx);
712 static void iflib_if_init_locked(if_ctx_t ctx);
713 static void iflib_free_intr_mem(if_ctx_t ctx);
714 #ifndef __NO_STRICT_ALIGNMENT
715 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
716 #endif
717 
718 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
719     SLIST_HEAD_INITIALIZER(cpu_offsets);
720 struct cpu_offset {
721 	SLIST_ENTRY(cpu_offset) entries;
722 	cpuset_t	set;
723 	unsigned int	refcount;
724 	uint16_t	offset;
725 };
726 static struct mtx cpu_offset_mtx;
727 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
728     MTX_DEF);
729 
730 DEBUGNET_DEFINE(iflib);
731 
732 #ifdef DEV_NETMAP
733 #include <sys/selinfo.h>
734 #include <net/netmap.h>
735 #include <dev/netmap/netmap_kern.h>
736 
737 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
738 
739 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
740 
741 /*
742  * device-specific sysctl variables:
743  *
744  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
745  *	During regular operations the CRC is stripped, but on some
746  *	hardware reception of frames not multiple of 64 is slower,
747  *	so using crcstrip=0 helps in benchmarks.
748  *
749  * iflib_rx_miss, iflib_rx_miss_bufs:
750  *	count packets that might be missed due to lost interrupts.
751  */
752 SYSCTL_DECL(_dev_netmap);
753 /*
754  * The xl driver by default strips CRCs and we do not override it.
755  */
756 
757 int iflib_crcstrip = 1;
758 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
759     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
760 
761 int iflib_rx_miss, iflib_rx_miss_bufs;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
763     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
765     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
766 
767 /*
768  * Register/unregister. We are already under netmap lock.
769  * Only called on the first register or the last unregister.
770  */
771 static int
772 iflib_netmap_register(struct netmap_adapter *na, int onoff)
773 {
774 	if_t ifp = na->ifp;
775 	if_ctx_t ctx = ifp->if_softc;
776 	int status;
777 
778 	CTX_LOCK(ctx);
779 	IFDI_INTR_DISABLE(ctx);
780 
781 	/* Tell the stack that the interface is no longer active */
782 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
783 
784 	if (!CTX_IS_VF(ctx))
785 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
786 
787 	/* enable or disable flags and callbacks in na and ifp */
788 	if (onoff) {
789 		nm_set_native_flags(na);
790 	} else {
791 		nm_clear_native_flags(na);
792 	}
793 	iflib_stop(ctx);
794 	iflib_init_locked(ctx);
795 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
796 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
797 	if (status)
798 		nm_clear_native_flags(na);
799 	CTX_UNLOCK(ctx);
800 	return (status);
801 }
802 
803 static int
804 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
805 {
806 	struct netmap_adapter *na = kring->na;
807 	u_int const lim = kring->nkr_num_slots - 1;
808 	u_int head = kring->rhead;
809 	struct netmap_ring *ring = kring->ring;
810 	bus_dmamap_t *map;
811 	struct if_rxd_update iru;
812 	if_ctx_t ctx = rxq->ifr_ctx;
813 	iflib_fl_t fl = &rxq->ifr_fl[0];
814 	uint32_t refill_pidx, nic_i;
815 #if IFLIB_DEBUG_COUNTERS
816 	int rf_count = 0;
817 #endif
818 
819 	if (nm_i == head && __predict_true(!init))
820 		return 0;
821 	iru_init(&iru, rxq, 0 /* flid */);
822 	map = fl->ifl_sds.ifsd_map;
823 	refill_pidx = netmap_idx_k2n(kring, nm_i);
824 	/*
825 	 * IMPORTANT: we must leave one free slot in the ring,
826 	 * so move head back by one unit
827 	 */
828 	head = nm_prev(head, lim);
829 	nic_i = UINT_MAX;
830 	DBG_COUNTER_INC(fl_refills);
831 	while (nm_i != head) {
832 #if IFLIB_DEBUG_COUNTERS
833 		if (++rf_count == 9)
834 			DBG_COUNTER_INC(fl_refills_large);
835 #endif
836 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
837 			struct netmap_slot *slot = &ring->slot[nm_i];
838 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
839 			uint32_t nic_i_dma = refill_pidx;
840 			nic_i = netmap_idx_k2n(kring, nm_i);
841 
842 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
843 
844 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
845 			        return netmap_ring_reinit(kring);
846 
847 			fl->ifl_vm_addrs[tmp_pidx] = addr;
848 			if (__predict_false(init)) {
849 				netmap_load_map(na, fl->ifl_buf_tag,
850 				    map[nic_i], addr);
851 			} else if (slot->flags & NS_BUF_CHANGED) {
852 				/* buffer has changed, reload map */
853 				netmap_reload_map(na, fl->ifl_buf_tag,
854 				    map[nic_i], addr);
855 			}
856 			slot->flags &= ~NS_BUF_CHANGED;
857 
858 			nm_i = nm_next(nm_i, lim);
859 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
860 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
861 				continue;
862 
863 			iru.iru_pidx = refill_pidx;
864 			iru.iru_count = tmp_pidx+1;
865 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
866 			refill_pidx = nic_i;
867 			for (int n = 0; n < iru.iru_count; n++) {
868 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
869 						BUS_DMASYNC_PREREAD);
870 				/* XXX - change this to not use the netmap func*/
871 				nic_i_dma = nm_next(nic_i_dma, lim);
872 			}
873 		}
874 	}
875 	kring->nr_hwcur = head;
876 
877 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
878 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
879 	if (__predict_true(nic_i != UINT_MAX)) {
880 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
881 		DBG_COUNTER_INC(rxd_flush);
882 	}
883 	return (0);
884 }
885 
886 /*
887  * Reconcile kernel and user view of the transmit ring.
888  *
889  * All information is in the kring.
890  * Userspace wants to send packets up to the one before kring->rhead,
891  * kernel knows kring->nr_hwcur is the first unsent packet.
892  *
893  * Here we push packets out (as many as possible), and possibly
894  * reclaim buffers from previously completed transmission.
895  *
896  * The caller (netmap) guarantees that there is only one instance
897  * running at any time. Any interference with other driver
898  * methods should be handled by the individual drivers.
899  */
900 static int
901 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
902 {
903 	struct netmap_adapter *na = kring->na;
904 	if_t ifp = na->ifp;
905 	struct netmap_ring *ring = kring->ring;
906 	u_int nm_i;	/* index into the netmap kring */
907 	u_int nic_i;	/* index into the NIC ring */
908 	u_int n;
909 	u_int const lim = kring->nkr_num_slots - 1;
910 	u_int const head = kring->rhead;
911 	struct if_pkt_info pi;
912 
913 	/*
914 	 * interrupts on every tx packet are expensive so request
915 	 * them every half ring, or where NS_REPORT is set
916 	 */
917 	u_int report_frequency = kring->nkr_num_slots >> 1;
918 	/* device-specific */
919 	if_ctx_t ctx = ifp->if_softc;
920 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
921 
922 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
923 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
924 
925 	/*
926 	 * First part: process new packets to send.
927 	 * nm_i is the current index in the netmap kring,
928 	 * nic_i is the corresponding index in the NIC ring.
929 	 *
930 	 * If we have packets to send (nm_i != head)
931 	 * iterate over the netmap ring, fetch length and update
932 	 * the corresponding slot in the NIC ring. Some drivers also
933 	 * need to update the buffer's physical address in the NIC slot
934 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
935 	 *
936 	 * The netmap_reload_map() calls is especially expensive,
937 	 * even when (as in this case) the tag is 0, so do only
938 	 * when the buffer has actually changed.
939 	 *
940 	 * If possible do not set the report/intr bit on all slots,
941 	 * but only a few times per ring or when NS_REPORT is set.
942 	 *
943 	 * Finally, on 10G and faster drivers, it might be useful
944 	 * to prefetch the next slot and txr entry.
945 	 */
946 
947 	nm_i = kring->nr_hwcur;
948 	if (nm_i != head) {	/* we have new packets to send */
949 		pkt_info_zero(&pi);
950 		pi.ipi_segs = txq->ift_segs;
951 		pi.ipi_qsidx = kring->ring_id;
952 		nic_i = netmap_idx_k2n(kring, nm_i);
953 
954 		__builtin_prefetch(&ring->slot[nm_i]);
955 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
956 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
957 
958 		for (n = 0; nm_i != head; n++) {
959 			struct netmap_slot *slot = &ring->slot[nm_i];
960 			u_int len = slot->len;
961 			uint64_t paddr;
962 			void *addr = PNMB(na, slot, &paddr);
963 			int flags = (slot->flags & NS_REPORT ||
964 				nic_i == 0 || nic_i == report_frequency) ?
965 				IPI_TX_INTR : 0;
966 
967 			/* device-specific */
968 			pi.ipi_len = len;
969 			pi.ipi_segs[0].ds_addr = paddr;
970 			pi.ipi_segs[0].ds_len = len;
971 			pi.ipi_nsegs = 1;
972 			pi.ipi_ndescs = 0;
973 			pi.ipi_pidx = nic_i;
974 			pi.ipi_flags = flags;
975 
976 			/* Fill the slot in the NIC ring. */
977 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
978 			DBG_COUNTER_INC(tx_encap);
979 
980 			/* prefetch for next round */
981 			__builtin_prefetch(&ring->slot[nm_i + 1]);
982 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
983 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
984 
985 			NM_CHECK_ADDR_LEN(na, addr, len);
986 
987 			if (slot->flags & NS_BUF_CHANGED) {
988 				/* buffer has changed, reload map */
989 				netmap_reload_map(na, txq->ift_buf_tag,
990 				    txq->ift_sds.ifsd_map[nic_i], addr);
991 			}
992 			/* make sure changes to the buffer are synced */
993 			bus_dmamap_sync(txq->ift_buf_tag,
994 			    txq->ift_sds.ifsd_map[nic_i],
995 			    BUS_DMASYNC_PREWRITE);
996 
997 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
998 			nm_i = nm_next(nm_i, lim);
999 			nic_i = nm_next(nic_i, lim);
1000 		}
1001 		kring->nr_hwcur = nm_i;
1002 
1003 		/* synchronize the NIC ring */
1004 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1005 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1006 
1007 		/* (re)start the tx unit up to slot nic_i (excluded) */
1008 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1009 	}
1010 
1011 	/*
1012 	 * Second part: reclaim buffers for completed transmissions.
1013 	 *
1014 	 * If there are unclaimed buffers, attempt to reclaim them.
1015 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1016 	 * minimal delay, then trigger the tx handler which will spin in the
1017 	 * group task queue.
1018 	 */
1019 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1020 		if (iflib_tx_credits_update(ctx, txq)) {
1021 			/* some tx completed, increment avail */
1022 			nic_i = txq->ift_cidx_processed;
1023 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1024 		}
1025 	}
1026 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1027 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1028 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1029 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1030 	}
1031 	return (0);
1032 }
1033 
1034 /*
1035  * Reconcile kernel and user view of the receive ring.
1036  * Same as for the txsync, this routine must be efficient.
1037  * The caller guarantees a single invocations, but races against
1038  * the rest of the driver should be handled here.
1039  *
1040  * On call, kring->rhead is the first packet that userspace wants
1041  * to keep, and kring->rcur is the wakeup point.
1042  * The kernel has previously reported packets up to kring->rtail.
1043  *
1044  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1045  * of whether or not we received an interrupt.
1046  */
1047 static int
1048 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1049 {
1050 	struct netmap_adapter *na = kring->na;
1051 	struct netmap_ring *ring = kring->ring;
1052 	if_t ifp = na->ifp;
1053 	iflib_fl_t fl;
1054 	uint32_t nm_i;	/* index into the netmap ring */
1055 	uint32_t nic_i;	/* index into the NIC ring */
1056 	u_int i, n;
1057 	u_int const lim = kring->nkr_num_slots - 1;
1058 	u_int const head = kring->rhead;
1059 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1060 	struct if_rxd_info ri;
1061 
1062 	if_ctx_t ctx = ifp->if_softc;
1063 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1064 	if (head > lim)
1065 		return netmap_ring_reinit(kring);
1066 
1067 	/*
1068 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1069 	 */
1070 
1071 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1072 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1073 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1074 	}
1075 
1076 	/*
1077 	 * First part: import newly received packets.
1078 	 *
1079 	 * nm_i is the index of the next free slot in the netmap ring,
1080 	 * nic_i is the index of the next received packet in the NIC ring,
1081 	 * and they may differ in case if_init() has been called while
1082 	 * in netmap mode. For the receive ring we have
1083 	 *
1084 	 *	nic_i = rxr->next_check;
1085 	 *	nm_i = kring->nr_hwtail (previous)
1086 	 * and
1087 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1088 	 *
1089 	 * rxr->next_check is set to 0 on a ring reinit
1090 	 */
1091 	if (netmap_no_pendintr || force_update) {
1092 		int crclen = iflib_crcstrip ? 0 : 4;
1093 		int error, avail;
1094 
1095 		for (i = 0; i < rxq->ifr_nfl; i++) {
1096 			fl = &rxq->ifr_fl[i];
1097 			nic_i = fl->ifl_cidx;
1098 			nm_i = netmap_idx_n2k(kring, nic_i);
1099 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1100 			    rxq->ifr_id, nic_i, USHRT_MAX);
1101 			for (n = 0; avail > 0; n++, avail--) {
1102 				rxd_info_zero(&ri);
1103 				ri.iri_frags = rxq->ifr_frags;
1104 				ri.iri_qsidx = kring->ring_id;
1105 				ri.iri_ifp = ctx->ifc_ifp;
1106 				ri.iri_cidx = nic_i;
1107 
1108 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1109 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1110 				ring->slot[nm_i].flags = 0;
1111 				bus_dmamap_sync(fl->ifl_buf_tag,
1112 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1113 				nm_i = nm_next(nm_i, lim);
1114 				nic_i = nm_next(nic_i, lim);
1115 			}
1116 			if (n) { /* update the state variables */
1117 				if (netmap_no_pendintr && !force_update) {
1118 					/* diagnostics */
1119 					iflib_rx_miss ++;
1120 					iflib_rx_miss_bufs += n;
1121 				}
1122 				fl->ifl_cidx = nic_i;
1123 				kring->nr_hwtail = nm_i;
1124 			}
1125 			kring->nr_kflags &= ~NKR_PENDINTR;
1126 		}
1127 	}
1128 	/*
1129 	 * Second part: skip past packets that userspace has released.
1130 	 * (kring->nr_hwcur to head excluded),
1131 	 * and make the buffers available for reception.
1132 	 * As usual nm_i is the index in the netmap ring,
1133 	 * nic_i is the index in the NIC ring, and
1134 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1135 	 */
1136 	/* XXX not sure how this will work with multiple free lists */
1137 	nm_i = kring->nr_hwcur;
1138 
1139 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1140 }
1141 
1142 static void
1143 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1144 {
1145 	if_ctx_t ctx = na->ifp->if_softc;
1146 
1147 	CTX_LOCK(ctx);
1148 	if (onoff) {
1149 		IFDI_INTR_ENABLE(ctx);
1150 	} else {
1151 		IFDI_INTR_DISABLE(ctx);
1152 	}
1153 	CTX_UNLOCK(ctx);
1154 }
1155 
1156 
1157 static int
1158 iflib_netmap_attach(if_ctx_t ctx)
1159 {
1160 	struct netmap_adapter na;
1161 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1162 
1163 	bzero(&na, sizeof(na));
1164 
1165 	na.ifp = ctx->ifc_ifp;
1166 	na.na_flags = NAF_BDG_MAYSLEEP;
1167 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1168 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1169 
1170 	na.num_tx_desc = scctx->isc_ntxd[0];
1171 	na.num_rx_desc = scctx->isc_nrxd[0];
1172 	na.nm_txsync = iflib_netmap_txsync;
1173 	na.nm_rxsync = iflib_netmap_rxsync;
1174 	na.nm_register = iflib_netmap_register;
1175 	na.nm_intr = iflib_netmap_intr;
1176 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1177 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1178 	return (netmap_attach(&na));
1179 }
1180 
1181 static void
1182 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1183 {
1184 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1185 	struct netmap_slot *slot;
1186 
1187 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1188 	if (slot == NULL)
1189 		return;
1190 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1191 
1192 		/*
1193 		 * In netmap mode, set the map for the packet buffer.
1194 		 * NOTE: Some drivers (not this one) also need to set
1195 		 * the physical buffer address in the NIC ring.
1196 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1197 		 * netmap slot index, si
1198 		 */
1199 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1200 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1201 		    NMB(na, slot + si));
1202 	}
1203 }
1204 
1205 static void
1206 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1207 {
1208 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1209 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1210 	struct netmap_slot *slot;
1211 	uint32_t nm_i;
1212 
1213 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1214 	if (slot == NULL)
1215 		return;
1216 	nm_i = netmap_idx_n2k(kring, 0);
1217 	netmap_fl_refill(rxq, kring, nm_i, true);
1218 }
1219 
1220 static void
1221 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1222 {
1223 	struct netmap_kring *kring;
1224 	uint16_t txqid;
1225 
1226 	txqid = txq->ift_id;
1227 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1228 
1229 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1230 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1231 		    BUS_DMASYNC_POSTREAD);
1232 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1233 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1234 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1235 			if (hz < 2000)
1236 				*reset_on = 1;
1237 			else
1238 				*reset_on = hz / 1000;
1239 		}
1240 	}
1241 }
1242 
1243 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1244 
1245 #else
1246 #define iflib_netmap_txq_init(ctx, txq)
1247 #define iflib_netmap_rxq_init(ctx, rxq)
1248 #define iflib_netmap_detach(ifp)
1249 
1250 #define iflib_netmap_attach(ctx) (0)
1251 #define netmap_rx_irq(ifp, qid, budget) (0)
1252 #define netmap_tx_irq(ifp, qid) do {} while (0)
1253 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1254 #endif
1255 
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1258 prefetch(void *x)
1259 {
1260 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1261 }
1262 static __inline void
1263 prefetch2cachelines(void *x)
1264 {
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1268 #endif
1269 }
1270 #else
1271 #define prefetch(x)
1272 #define prefetch2cachelines(x)
1273 #endif
1274 
1275 static void
1276 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1277 {
1278 	iflib_fl_t fl;
1279 
1280 	fl = &rxq->ifr_fl[flid];
1281 	iru->iru_paddrs = fl->ifl_bus_addrs;
1282 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1283 	iru->iru_idxs = fl->ifl_rxd_idxs;
1284 	iru->iru_qsidx = rxq->ifr_id;
1285 	iru->iru_buf_size = fl->ifl_buf_size;
1286 	iru->iru_flidx = fl->ifl_id;
1287 }
1288 
1289 static void
1290 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1291 {
1292 	if (err)
1293 		return;
1294 	*(bus_addr_t *) arg = segs[0].ds_addr;
1295 }
1296 
1297 int
1298 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1299 {
1300 	int err;
1301 	device_t dev = ctx->ifc_dev;
1302 
1303 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1304 				align, 0,		/* alignment, bounds */
1305 				BUS_SPACE_MAXADDR,	/* lowaddr */
1306 				BUS_SPACE_MAXADDR,	/* highaddr */
1307 				NULL, NULL,		/* filter, filterarg */
1308 				size,			/* maxsize */
1309 				1,			/* nsegments */
1310 				size,			/* maxsegsize */
1311 				BUS_DMA_ALLOCNOW,	/* flags */
1312 				NULL,			/* lockfunc */
1313 				NULL,			/* lockarg */
1314 				&dma->idi_tag);
1315 	if (err) {
1316 		device_printf(dev,
1317 		    "%s: bus_dma_tag_create failed: %d\n",
1318 		    __func__, err);
1319 		goto fail_0;
1320 	}
1321 
1322 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1323 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1324 	if (err) {
1325 		device_printf(dev,
1326 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1327 		    __func__, (uintmax_t)size, err);
1328 		goto fail_1;
1329 	}
1330 
1331 	dma->idi_paddr = IF_BAD_DMA;
1332 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1333 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1334 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1335 		device_printf(dev,
1336 		    "%s: bus_dmamap_load failed: %d\n",
1337 		    __func__, err);
1338 		goto fail_2;
1339 	}
1340 
1341 	dma->idi_size = size;
1342 	return (0);
1343 
1344 fail_2:
1345 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1346 fail_1:
1347 	bus_dma_tag_destroy(dma->idi_tag);
1348 fail_0:
1349 	dma->idi_tag = NULL;
1350 
1351 	return (err);
1352 }
1353 
1354 int
1355 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1356 {
1357 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1358 
1359 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1360 
1361 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1362 }
1363 
1364 int
1365 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1366 {
1367 	int i, err;
1368 	iflib_dma_info_t *dmaiter;
1369 
1370 	dmaiter = dmalist;
1371 	for (i = 0; i < count; i++, dmaiter++) {
1372 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1373 			break;
1374 	}
1375 	if (err)
1376 		iflib_dma_free_multi(dmalist, i);
1377 	return (err);
1378 }
1379 
1380 void
1381 iflib_dma_free(iflib_dma_info_t dma)
1382 {
1383 	if (dma->idi_tag == NULL)
1384 		return;
1385 	if (dma->idi_paddr != IF_BAD_DMA) {
1386 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1387 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1388 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1389 		dma->idi_paddr = IF_BAD_DMA;
1390 	}
1391 	if (dma->idi_vaddr != NULL) {
1392 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1393 		dma->idi_vaddr = NULL;
1394 	}
1395 	bus_dma_tag_destroy(dma->idi_tag);
1396 	dma->idi_tag = NULL;
1397 }
1398 
1399 void
1400 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1401 {
1402 	int i;
1403 	iflib_dma_info_t *dmaiter = dmalist;
1404 
1405 	for (i = 0; i < count; i++, dmaiter++)
1406 		iflib_dma_free(*dmaiter);
1407 }
1408 
1409 #ifdef EARLY_AP_STARTUP
1410 static const int iflib_started = 1;
1411 #else
1412 /*
1413  * We used to abuse the smp_started flag to decide if the queues have been
1414  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1415  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1416  * is set.  Run a SYSINIT() strictly after that to just set a usable
1417  * completion flag.
1418  */
1419 
1420 static int iflib_started;
1421 
1422 static void
1423 iflib_record_started(void *arg)
1424 {
1425 	iflib_started = 1;
1426 }
1427 
1428 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1429 	iflib_record_started, NULL);
1430 #endif
1431 
1432 static int
1433 iflib_fast_intr(void *arg)
1434 {
1435 	iflib_filter_info_t info = arg;
1436 	struct grouptask *gtask = info->ifi_task;
1437 	int result;
1438 
1439 	if (!iflib_started)
1440 		return (FILTER_STRAY);
1441 
1442 	DBG_COUNTER_INC(fast_intrs);
1443 	if (info->ifi_filter != NULL) {
1444 		result = info->ifi_filter(info->ifi_filter_arg);
1445 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1446 			return (result);
1447 	}
1448 
1449 	GROUPTASK_ENQUEUE(gtask);
1450 	return (FILTER_HANDLED);
1451 }
1452 
1453 static int
1454 iflib_fast_intr_rxtx(void *arg)
1455 {
1456 	iflib_filter_info_t info = arg;
1457 	struct grouptask *gtask = info->ifi_task;
1458 	if_ctx_t ctx;
1459 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1460 	iflib_txq_t txq;
1461 	void *sc;
1462 	int i, cidx, result;
1463 	qidx_t txqid;
1464 	bool intr_enable, intr_legacy;
1465 
1466 	if (!iflib_started)
1467 		return (FILTER_STRAY);
1468 
1469 	DBG_COUNTER_INC(fast_intrs);
1470 	if (info->ifi_filter != NULL) {
1471 		result = info->ifi_filter(info->ifi_filter_arg);
1472 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1473 			return (result);
1474 	}
1475 
1476 	ctx = rxq->ifr_ctx;
1477 	sc = ctx->ifc_softc;
1478 	intr_enable = false;
1479 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1480 	MPASS(rxq->ifr_ntxqirq);
1481 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1482 		txqid = rxq->ifr_txqid[i];
1483 		txq = &ctx->ifc_txqs[txqid];
1484 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1485 		    BUS_DMASYNC_POSTREAD);
1486 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1487 			if (intr_legacy)
1488 				intr_enable = true;
1489 			else
1490 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1491 			continue;
1492 		}
1493 		GROUPTASK_ENQUEUE(&txq->ift_task);
1494 	}
1495 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1496 		cidx = rxq->ifr_cq_cidx;
1497 	else
1498 		cidx = rxq->ifr_fl[0].ifl_cidx;
1499 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1500 		GROUPTASK_ENQUEUE(gtask);
1501 	else {
1502 		if (intr_legacy)
1503 			intr_enable = true;
1504 		else
1505 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1506 		DBG_COUNTER_INC(rx_intr_enables);
1507 	}
1508 	if (intr_enable)
1509 		IFDI_INTR_ENABLE(ctx);
1510 	return (FILTER_HANDLED);
1511 }
1512 
1513 
1514 static int
1515 iflib_fast_intr_ctx(void *arg)
1516 {
1517 	iflib_filter_info_t info = arg;
1518 	struct grouptask *gtask = info->ifi_task;
1519 	int result;
1520 
1521 	if (!iflib_started)
1522 		return (FILTER_STRAY);
1523 
1524 	DBG_COUNTER_INC(fast_intrs);
1525 	if (info->ifi_filter != NULL) {
1526 		result = info->ifi_filter(info->ifi_filter_arg);
1527 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1528 			return (result);
1529 	}
1530 
1531 	GROUPTASK_ENQUEUE(gtask);
1532 	return (FILTER_HANDLED);
1533 }
1534 
1535 static int
1536 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1537 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1538 		 const char *name)
1539 {
1540 	struct resource *res;
1541 	void *tag = NULL;
1542 	device_t dev = ctx->ifc_dev;
1543 	int flags, i, rc;
1544 
1545 	flags = RF_ACTIVE;
1546 	if (ctx->ifc_flags & IFC_LEGACY)
1547 		flags |= RF_SHAREABLE;
1548 	MPASS(rid < 512);
1549 	i = rid;
1550 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1551 	if (res == NULL) {
1552 		device_printf(dev,
1553 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1554 		return (ENOMEM);
1555 	}
1556 	irq->ii_res = res;
1557 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1558 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1559 						filter, handler, arg, &tag);
1560 	if (rc != 0) {
1561 		device_printf(dev,
1562 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1563 					  rid, name ? name : "unknown", rc);
1564 		return (rc);
1565 	} else if (name)
1566 		bus_describe_intr(dev, res, tag, "%s", name);
1567 
1568 	irq->ii_tag = tag;
1569 	return (0);
1570 }
1571 
1572 /*********************************************************************
1573  *
1574  *  Allocate DMA resources for TX buffers as well as memory for the TX
1575  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1576  *  iflib_sw_tx_desc_array structure, storing all the information that
1577  *  is needed to transmit a packet on the wire.  This is called only
1578  *  once at attach, setup is done every reset.
1579  *
1580  **********************************************************************/
1581 static int
1582 iflib_txsd_alloc(iflib_txq_t txq)
1583 {
1584 	if_ctx_t ctx = txq->ift_ctx;
1585 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1586 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1587 	device_t dev = ctx->ifc_dev;
1588 	bus_size_t tsomaxsize;
1589 	int err, nsegments, ntsosegments;
1590 	bool tso;
1591 
1592 	nsegments = scctx->isc_tx_nsegments;
1593 	ntsosegments = scctx->isc_tx_tso_segments_max;
1594 	tsomaxsize = scctx->isc_tx_tso_size_max;
1595 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1596 		tsomaxsize += sizeof(struct ether_vlan_header);
1597 	MPASS(scctx->isc_ntxd[0] > 0);
1598 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1599 	MPASS(nsegments > 0);
1600 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1601 		MPASS(ntsosegments > 0);
1602 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1603 	}
1604 
1605 	/*
1606 	 * Set up DMA tags for TX buffers.
1607 	 */
1608 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1609 			       1, 0,			/* alignment, bounds */
1610 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1611 			       BUS_SPACE_MAXADDR,	/* highaddr */
1612 			       NULL, NULL,		/* filter, filterarg */
1613 			       sctx->isc_tx_maxsize,		/* maxsize */
1614 			       nsegments,	/* nsegments */
1615 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1616 			       0,			/* flags */
1617 			       NULL,			/* lockfunc */
1618 			       NULL,			/* lockfuncarg */
1619 			       &txq->ift_buf_tag))) {
1620 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1621 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1622 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1623 		goto fail;
1624 	}
1625 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1626 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1627 			       1, 0,			/* alignment, bounds */
1628 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1629 			       BUS_SPACE_MAXADDR,	/* highaddr */
1630 			       NULL, NULL,		/* filter, filterarg */
1631 			       tsomaxsize,		/* maxsize */
1632 			       ntsosegments,	/* nsegments */
1633 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1634 			       0,			/* flags */
1635 			       NULL,			/* lockfunc */
1636 			       NULL,			/* lockfuncarg */
1637 			       &txq->ift_tso_buf_tag))) {
1638 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1639 		    err);
1640 		goto fail;
1641 	}
1642 
1643 	/* Allocate memory for the TX mbuf map. */
1644 	if (!(txq->ift_sds.ifsd_m =
1645 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1646 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1647 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1648 		err = ENOMEM;
1649 		goto fail;
1650 	}
1651 
1652 	/*
1653 	 * Create the DMA maps for TX buffers.
1654 	 */
1655 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1656 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1657 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1658 		device_printf(dev,
1659 		    "Unable to allocate TX buffer DMA map memory\n");
1660 		err = ENOMEM;
1661 		goto fail;
1662 	}
1663 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1664 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1665 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1666 		device_printf(dev,
1667 		    "Unable to allocate TSO TX buffer map memory\n");
1668 		err = ENOMEM;
1669 		goto fail;
1670 	}
1671 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1672 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1673 		    &txq->ift_sds.ifsd_map[i]);
1674 		if (err != 0) {
1675 			device_printf(dev, "Unable to create TX DMA map\n");
1676 			goto fail;
1677 		}
1678 		if (!tso)
1679 			continue;
1680 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1681 		    &txq->ift_sds.ifsd_tso_map[i]);
1682 		if (err != 0) {
1683 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1684 			goto fail;
1685 		}
1686 	}
1687 	return (0);
1688 fail:
1689 	/* We free all, it handles case where we are in the middle */
1690 	iflib_tx_structures_free(ctx);
1691 	return (err);
1692 }
1693 
1694 static void
1695 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1696 {
1697 	bus_dmamap_t map;
1698 
1699 	map = NULL;
1700 	if (txq->ift_sds.ifsd_map != NULL)
1701 		map = txq->ift_sds.ifsd_map[i];
1702 	if (map != NULL) {
1703 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1704 		bus_dmamap_unload(txq->ift_buf_tag, map);
1705 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1706 		txq->ift_sds.ifsd_map[i] = NULL;
1707 	}
1708 
1709 	map = NULL;
1710 	if (txq->ift_sds.ifsd_tso_map != NULL)
1711 		map = txq->ift_sds.ifsd_tso_map[i];
1712 	if (map != NULL) {
1713 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1714 		    BUS_DMASYNC_POSTWRITE);
1715 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1716 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1717 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1718 	}
1719 }
1720 
1721 static void
1722 iflib_txq_destroy(iflib_txq_t txq)
1723 {
1724 	if_ctx_t ctx = txq->ift_ctx;
1725 
1726 	for (int i = 0; i < txq->ift_size; i++)
1727 		iflib_txsd_destroy(ctx, txq, i);
1728 
1729 	if (txq->ift_br != NULL) {
1730 		ifmp_ring_free(txq->ift_br);
1731 		txq->ift_br = NULL;
1732 	}
1733 
1734 	mtx_destroy(&txq->ift_mtx);
1735 
1736 	if (txq->ift_sds.ifsd_map != NULL) {
1737 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1738 		txq->ift_sds.ifsd_map = NULL;
1739 	}
1740 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1741 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1742 		txq->ift_sds.ifsd_tso_map = NULL;
1743 	}
1744 	if (txq->ift_sds.ifsd_m != NULL) {
1745 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1746 		txq->ift_sds.ifsd_m = NULL;
1747 	}
1748 	if (txq->ift_buf_tag != NULL) {
1749 		bus_dma_tag_destroy(txq->ift_buf_tag);
1750 		txq->ift_buf_tag = NULL;
1751 	}
1752 	if (txq->ift_tso_buf_tag != NULL) {
1753 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1754 		txq->ift_tso_buf_tag = NULL;
1755 	}
1756 	if (txq->ift_ifdi != NULL) {
1757 		free(txq->ift_ifdi, M_IFLIB);
1758 	}
1759 }
1760 
1761 static void
1762 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1763 {
1764 	struct mbuf **mp;
1765 
1766 	mp = &txq->ift_sds.ifsd_m[i];
1767 	if (*mp == NULL)
1768 		return;
1769 
1770 	if (txq->ift_sds.ifsd_map != NULL) {
1771 		bus_dmamap_sync(txq->ift_buf_tag,
1772 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1773 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1774 	}
1775 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1776 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1777 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1778 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1779 		    txq->ift_sds.ifsd_tso_map[i]);
1780 	}
1781 	m_free(*mp);
1782 	DBG_COUNTER_INC(tx_frees);
1783 	*mp = NULL;
1784 }
1785 
1786 static int
1787 iflib_txq_setup(iflib_txq_t txq)
1788 {
1789 	if_ctx_t ctx = txq->ift_ctx;
1790 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1791 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1792 	iflib_dma_info_t di;
1793 	int i;
1794 
1795 	/* Set number of descriptors available */
1796 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1797 	/* XXX make configurable */
1798 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1799 
1800 	/* Reset indices */
1801 	txq->ift_cidx_processed = 0;
1802 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1803 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1804 
1805 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1806 		bzero((void *)di->idi_vaddr, di->idi_size);
1807 
1808 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1809 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1810 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1811 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1812 	return (0);
1813 }
1814 
1815 /*********************************************************************
1816  *
1817  *  Allocate DMA resources for RX buffers as well as memory for the RX
1818  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1819  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1820  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1821  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1822  *  packet, the maximum number of entries we'll need is equal to the
1823  *  number of hardware receive descriptors that we've allocated.
1824  *
1825  **********************************************************************/
1826 static int
1827 iflib_rxsd_alloc(iflib_rxq_t rxq)
1828 {
1829 	if_ctx_t ctx = rxq->ifr_ctx;
1830 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1831 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1832 	device_t dev = ctx->ifc_dev;
1833 	iflib_fl_t fl;
1834 	int			err;
1835 
1836 	MPASS(scctx->isc_nrxd[0] > 0);
1837 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1838 
1839 	fl = rxq->ifr_fl;
1840 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1841 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1842 		/* Set up DMA tag for RX buffers. */
1843 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1844 					 1, 0,			/* alignment, bounds */
1845 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1846 					 BUS_SPACE_MAXADDR,	/* highaddr */
1847 					 NULL, NULL,		/* filter, filterarg */
1848 					 sctx->isc_rx_maxsize,	/* maxsize */
1849 					 sctx->isc_rx_nsegments,	/* nsegments */
1850 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1851 					 0,			/* flags */
1852 					 NULL,			/* lockfunc */
1853 					 NULL,			/* lockarg */
1854 					 &fl->ifl_buf_tag);
1855 		if (err) {
1856 			device_printf(dev,
1857 			    "Unable to allocate RX DMA tag: %d\n", err);
1858 			goto fail;
1859 		}
1860 
1861 		/* Allocate memory for the RX mbuf map. */
1862 		if (!(fl->ifl_sds.ifsd_m =
1863 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1864 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1865 			device_printf(dev,
1866 			    "Unable to allocate RX mbuf map memory\n");
1867 			err = ENOMEM;
1868 			goto fail;
1869 		}
1870 
1871 		/* Allocate memory for the direct RX cluster pointer map. */
1872 		if (!(fl->ifl_sds.ifsd_cl =
1873 		      (caddr_t *) malloc(sizeof(caddr_t) *
1874 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1875 			device_printf(dev,
1876 			    "Unable to allocate RX cluster map memory\n");
1877 			err = ENOMEM;
1878 			goto fail;
1879 		}
1880 
1881 		/* Allocate memory for the RX cluster bus address map. */
1882 		if (!(fl->ifl_sds.ifsd_ba =
1883 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1884 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1885 			device_printf(dev,
1886 			    "Unable to allocate RX bus address map memory\n");
1887 			err = ENOMEM;
1888 			goto fail;
1889 		}
1890 
1891 		/*
1892 		 * Create the DMA maps for RX buffers.
1893 		 */
1894 		if (!(fl->ifl_sds.ifsd_map =
1895 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1896 			device_printf(dev,
1897 			    "Unable to allocate RX buffer DMA map memory\n");
1898 			err = ENOMEM;
1899 			goto fail;
1900 		}
1901 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1902 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1903 			    &fl->ifl_sds.ifsd_map[i]);
1904 			if (err != 0) {
1905 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1906 				goto fail;
1907 			}
1908 		}
1909 	}
1910 	return (0);
1911 
1912 fail:
1913 	iflib_rx_structures_free(ctx);
1914 	return (err);
1915 }
1916 
1917 
1918 /*
1919  * Internal service routines
1920  */
1921 
1922 struct rxq_refill_cb_arg {
1923 	int               error;
1924 	bus_dma_segment_t seg;
1925 	int               nseg;
1926 };
1927 
1928 static void
1929 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1930 {
1931 	struct rxq_refill_cb_arg *cb_arg = arg;
1932 
1933 	cb_arg->error = error;
1934 	cb_arg->seg = segs[0];
1935 	cb_arg->nseg = nseg;
1936 }
1937 
1938 /**
1939  * _iflib_fl_refill - refill an rxq free-buffer list
1940  * @ctx: the iflib context
1941  * @fl: the free list to refill
1942  * @count: the number of new buffers to allocate
1943  *
1944  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1945  * The caller must assure that @count does not exceed the queue's capacity.
1946  */
1947 static void
1948 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1949 {
1950 	struct if_rxd_update iru;
1951 	struct rxq_refill_cb_arg cb_arg;
1952 	struct mbuf *m;
1953 	caddr_t cl, *sd_cl;
1954 	struct mbuf **sd_m;
1955 	bus_dmamap_t *sd_map;
1956 	bus_addr_t bus_addr, *sd_ba;
1957 	int err, frag_idx, i, idx, n, pidx;
1958 	qidx_t credits;
1959 
1960 	sd_m = fl->ifl_sds.ifsd_m;
1961 	sd_map = fl->ifl_sds.ifsd_map;
1962 	sd_cl = fl->ifl_sds.ifsd_cl;
1963 	sd_ba = fl->ifl_sds.ifsd_ba;
1964 	pidx = fl->ifl_pidx;
1965 	idx = pidx;
1966 	frag_idx = fl->ifl_fragidx;
1967 	credits = fl->ifl_credits;
1968 
1969 	i = 0;
1970 	n = count;
1971 	MPASS(n > 0);
1972 	MPASS(credits + n <= fl->ifl_size);
1973 
1974 	if (pidx < fl->ifl_cidx)
1975 		MPASS(pidx + n <= fl->ifl_cidx);
1976 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1977 		MPASS(fl->ifl_gen == 0);
1978 	if (pidx > fl->ifl_cidx)
1979 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1980 
1981 	DBG_COUNTER_INC(fl_refills);
1982 	if (n > 8)
1983 		DBG_COUNTER_INC(fl_refills_large);
1984 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1985 	while (n--) {
1986 		/*
1987 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1988 		 * initialized after rx.
1989 		 *
1990 		 * If the cluster is still set then we know a minimum sized packet was received
1991 		 */
1992 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1993 		    &frag_idx);
1994 		if (frag_idx < 0)
1995 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1996 		MPASS(frag_idx >= 0);
1997 		if ((cl = sd_cl[frag_idx]) == NULL) {
1998 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1999 				break;
2000 
2001 			cb_arg.error = 0;
2002 			MPASS(sd_map != NULL);
2003 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2004 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2005 			    BUS_DMA_NOWAIT);
2006 			if (err != 0 || cb_arg.error) {
2007 				/*
2008 				 * !zone_pack ?
2009 				 */
2010 				if (fl->ifl_zone == zone_pack)
2011 					uma_zfree(fl->ifl_zone, cl);
2012 				break;
2013 			}
2014 
2015 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2016 			sd_cl[frag_idx] = cl;
2017 #if MEMORY_LOGGING
2018 			fl->ifl_cl_enqueued++;
2019 #endif
2020 		} else {
2021 			bus_addr = sd_ba[frag_idx];
2022 		}
2023 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2024 		    BUS_DMASYNC_PREREAD);
2025 
2026 		if (sd_m[frag_idx] == NULL) {
2027 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2028 				break;
2029 			}
2030 			sd_m[frag_idx] = m;
2031 		}
2032 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2033 #if MEMORY_LOGGING
2034 		fl->ifl_m_enqueued++;
2035 #endif
2036 
2037 		DBG_COUNTER_INC(rx_allocs);
2038 		fl->ifl_rxd_idxs[i] = frag_idx;
2039 		fl->ifl_bus_addrs[i] = bus_addr;
2040 		fl->ifl_vm_addrs[i] = cl;
2041 		credits++;
2042 		i++;
2043 		MPASS(credits <= fl->ifl_size);
2044 		if (++idx == fl->ifl_size) {
2045 			fl->ifl_gen = 1;
2046 			idx = 0;
2047 		}
2048 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2049 			iru.iru_pidx = pidx;
2050 			iru.iru_count = i;
2051 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2052 			i = 0;
2053 			pidx = idx;
2054 			fl->ifl_pidx = idx;
2055 			fl->ifl_credits = credits;
2056 		}
2057 	}
2058 
2059 	if (i) {
2060 		iru.iru_pidx = pidx;
2061 		iru.iru_count = i;
2062 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2063 		fl->ifl_pidx = idx;
2064 		fl->ifl_credits = credits;
2065 	}
2066 	DBG_COUNTER_INC(rxd_flush);
2067 	if (fl->ifl_pidx == 0)
2068 		pidx = fl->ifl_size - 1;
2069 	else
2070 		pidx = fl->ifl_pidx - 1;
2071 
2072 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2073 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2074 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2075 	fl->ifl_fragidx = frag_idx;
2076 }
2077 
2078 static __inline void
2079 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2080 {
2081 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2082 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2083 #ifdef INVARIANTS
2084 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2085 #endif
2086 
2087 	MPASS(fl->ifl_credits <= fl->ifl_size);
2088 	MPASS(reclaimable == delta);
2089 
2090 	if (reclaimable > 0)
2091 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2092 }
2093 
2094 uint8_t
2095 iflib_in_detach(if_ctx_t ctx)
2096 {
2097 	bool in_detach;
2098 
2099 	STATE_LOCK(ctx);
2100 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2101 	STATE_UNLOCK(ctx);
2102 	return (in_detach);
2103 }
2104 
2105 static void
2106 iflib_fl_bufs_free(iflib_fl_t fl)
2107 {
2108 	iflib_dma_info_t idi = fl->ifl_ifdi;
2109 	bus_dmamap_t sd_map;
2110 	uint32_t i;
2111 
2112 	for (i = 0; i < fl->ifl_size; i++) {
2113 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2114 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2115 
2116 		if (*sd_cl != NULL) {
2117 			sd_map = fl->ifl_sds.ifsd_map[i];
2118 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2119 			    BUS_DMASYNC_POSTREAD);
2120 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2121 			if (*sd_cl != NULL)
2122 				uma_zfree(fl->ifl_zone, *sd_cl);
2123 			// XXX: Should this get moved out?
2124 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2125 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2126 			if (*sd_m != NULL) {
2127 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2128 				uma_zfree(zone_mbuf, *sd_m);
2129 			}
2130 		} else {
2131 			MPASS(*sd_cl == NULL);
2132 			MPASS(*sd_m == NULL);
2133 		}
2134 #if MEMORY_LOGGING
2135 		fl->ifl_m_dequeued++;
2136 		fl->ifl_cl_dequeued++;
2137 #endif
2138 		*sd_cl = NULL;
2139 		*sd_m = NULL;
2140 	}
2141 #ifdef INVARIANTS
2142 	for (i = 0; i < fl->ifl_size; i++) {
2143 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2144 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2145 	}
2146 #endif
2147 	/*
2148 	 * Reset free list values
2149 	 */
2150 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2151 	bzero(idi->idi_vaddr, idi->idi_size);
2152 }
2153 
2154 /*********************************************************************
2155  *
2156  *  Initialize a free list and its buffers.
2157  *
2158  **********************************************************************/
2159 static int
2160 iflib_fl_setup(iflib_fl_t fl)
2161 {
2162 	iflib_rxq_t rxq = fl->ifl_rxq;
2163 	if_ctx_t ctx = rxq->ifr_ctx;
2164 
2165 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2166 	/*
2167 	** Free current RX buffer structs and their mbufs
2168 	*/
2169 	iflib_fl_bufs_free(fl);
2170 	/* Now replenish the mbufs */
2171 	MPASS(fl->ifl_credits == 0);
2172 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2173 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2174 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2175 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2176 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2177 
2178 
2179 	/* avoid pre-allocating zillions of clusters to an idle card
2180 	 * potentially speeding up attach
2181 	 */
2182 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2183 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2184 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2185 		return (ENOBUFS);
2186 	/*
2187 	 * handle failure
2188 	 */
2189 	MPASS(rxq != NULL);
2190 	MPASS(fl->ifl_ifdi != NULL);
2191 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2192 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2193 	return (0);
2194 }
2195 
2196 /*********************************************************************
2197  *
2198  *  Free receive ring data structures
2199  *
2200  **********************************************************************/
2201 static void
2202 iflib_rx_sds_free(iflib_rxq_t rxq)
2203 {
2204 	iflib_fl_t fl;
2205 	int i, j;
2206 
2207 	if (rxq->ifr_fl != NULL) {
2208 		for (i = 0; i < rxq->ifr_nfl; i++) {
2209 			fl = &rxq->ifr_fl[i];
2210 			if (fl->ifl_buf_tag != NULL) {
2211 				if (fl->ifl_sds.ifsd_map != NULL) {
2212 					for (j = 0; j < fl->ifl_size; j++) {
2213 						if (fl->ifl_sds.ifsd_map[j] ==
2214 						    NULL)
2215 							continue;
2216 						bus_dmamap_sync(
2217 						    fl->ifl_buf_tag,
2218 						    fl->ifl_sds.ifsd_map[j],
2219 						    BUS_DMASYNC_POSTREAD);
2220 						bus_dmamap_unload(
2221 						    fl->ifl_buf_tag,
2222 						    fl->ifl_sds.ifsd_map[j]);
2223 					}
2224 				}
2225 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2226 				fl->ifl_buf_tag = NULL;
2227 			}
2228 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2229 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2230 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2231 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2232 			fl->ifl_sds.ifsd_m = NULL;
2233 			fl->ifl_sds.ifsd_cl = NULL;
2234 			fl->ifl_sds.ifsd_ba = NULL;
2235 			fl->ifl_sds.ifsd_map = NULL;
2236 		}
2237 		free(rxq->ifr_fl, M_IFLIB);
2238 		rxq->ifr_fl = NULL;
2239 		free(rxq->ifr_ifdi, M_IFLIB);
2240 		rxq->ifr_ifdi = NULL;
2241 		rxq->ifr_cq_cidx = 0;
2242 	}
2243 }
2244 
2245 /*
2246  * Timer routine
2247  */
2248 static void
2249 iflib_timer(void *arg)
2250 {
2251 	iflib_txq_t txq = arg;
2252 	if_ctx_t ctx = txq->ift_ctx;
2253 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2254 	uint64_t this_tick = ticks;
2255 	uint32_t reset_on = hz / 2;
2256 
2257 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2258 		return;
2259 
2260 	/*
2261 	** Check on the state of the TX queue(s), this
2262 	** can be done without the lock because its RO
2263 	** and the HUNG state will be static if set.
2264 	*/
2265 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2266 		txq->ift_last_timer_tick = this_tick;
2267 		IFDI_TIMER(ctx, txq->ift_id);
2268 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2269 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2270 		     (sctx->isc_pause_frames == 0)))
2271 			goto hung;
2272 
2273 		if (ifmp_ring_is_stalled(txq->ift_br))
2274 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2275 		txq->ift_cleaned_prev = txq->ift_cleaned;
2276 	}
2277 #ifdef DEV_NETMAP
2278 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2279 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2280 #endif
2281 	/* handle any laggards */
2282 	if (txq->ift_db_pending)
2283 		GROUPTASK_ENQUEUE(&txq->ift_task);
2284 
2285 	sctx->isc_pause_frames = 0;
2286 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2287 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2288 	return;
2289 
2290  hung:
2291 	device_printf(ctx->ifc_dev,
2292 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2293 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2294 	STATE_LOCK(ctx);
2295 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2296 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2297 	iflib_admin_intr_deferred(ctx);
2298 	STATE_UNLOCK(ctx);
2299 }
2300 
2301 static void
2302 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2303 {
2304 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2305 
2306 	/*
2307 	 * XXX don't set the max_frame_size to larger
2308 	 * than the hardware can handle
2309 	 */
2310 	if (sctx->isc_max_frame_size <= MCLBYTES)
2311 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2312 	else
2313 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2314 }
2315 
2316 uint32_t
2317 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2318 {
2319 
2320 	return (ctx->ifc_rx_mbuf_sz);
2321 }
2322 
2323 static void
2324 iflib_init_locked(if_ctx_t ctx)
2325 {
2326 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2327 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2328 	if_t ifp = ctx->ifc_ifp;
2329 	iflib_fl_t fl;
2330 	iflib_txq_t txq;
2331 	iflib_rxq_t rxq;
2332 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2333 
2334 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2335 	IFDI_INTR_DISABLE(ctx);
2336 
2337 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2338 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2339 	/* Set hardware offload abilities */
2340 	if_clearhwassist(ifp);
2341 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2342 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2343 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2344 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2345 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2346 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2347 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2348 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2349 
2350 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2351 		CALLOUT_LOCK(txq);
2352 		callout_stop(&txq->ift_timer);
2353 		CALLOUT_UNLOCK(txq);
2354 		iflib_netmap_txq_init(ctx, txq);
2355 	}
2356 
2357 	/*
2358 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2359 	 * that drivers can use the value when setting up the hardware receive
2360 	 * buffers.
2361 	 */
2362 	iflib_calc_rx_mbuf_sz(ctx);
2363 
2364 #ifdef INVARIANTS
2365 	i = if_getdrvflags(ifp);
2366 #endif
2367 	IFDI_INIT(ctx);
2368 	MPASS(if_getdrvflags(ifp) == i);
2369 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2370 		/* XXX this should really be done on a per-queue basis */
2371 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2372 			MPASS(rxq->ifr_id == i);
2373 			iflib_netmap_rxq_init(ctx, rxq);
2374 			continue;
2375 		}
2376 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2377 			if (iflib_fl_setup(fl)) {
2378 				device_printf(ctx->ifc_dev,
2379 				    "setting up free list %d failed - "
2380 				    "check cluster settings\n", j);
2381 				goto done;
2382 			}
2383 		}
2384 	}
2385 done:
2386 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2387 	IFDI_INTR_ENABLE(ctx);
2388 	txq = ctx->ifc_txqs;
2389 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2390 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2391 			txq->ift_timer.c_cpu);
2392 }
2393 
2394 static int
2395 iflib_media_change(if_t ifp)
2396 {
2397 	if_ctx_t ctx = if_getsoftc(ifp);
2398 	int err;
2399 
2400 	CTX_LOCK(ctx);
2401 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2402 		iflib_init_locked(ctx);
2403 	CTX_UNLOCK(ctx);
2404 	return (err);
2405 }
2406 
2407 static void
2408 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2409 {
2410 	if_ctx_t ctx = if_getsoftc(ifp);
2411 
2412 	CTX_LOCK(ctx);
2413 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2414 	IFDI_MEDIA_STATUS(ctx, ifmr);
2415 	CTX_UNLOCK(ctx);
2416 }
2417 
2418 void
2419 iflib_stop(if_ctx_t ctx)
2420 {
2421 	iflib_txq_t txq = ctx->ifc_txqs;
2422 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2423 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2424 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2425 	iflib_dma_info_t di;
2426 	iflib_fl_t fl;
2427 	int i, j;
2428 
2429 	/* Tell the stack that the interface is no longer active */
2430 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2431 
2432 	IFDI_INTR_DISABLE(ctx);
2433 	DELAY(1000);
2434 	IFDI_STOP(ctx);
2435 	DELAY(1000);
2436 
2437 	iflib_debug_reset();
2438 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2439 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2440 		/* make sure all transmitters have completed before proceeding XXX */
2441 
2442 		CALLOUT_LOCK(txq);
2443 		callout_stop(&txq->ift_timer);
2444 		CALLOUT_UNLOCK(txq);
2445 
2446 		/* clean any enqueued buffers */
2447 		iflib_ifmp_purge(txq);
2448 		/* Free any existing tx buffers. */
2449 		for (j = 0; j < txq->ift_size; j++) {
2450 			iflib_txsd_free(ctx, txq, j);
2451 		}
2452 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2453 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2454 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2455 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2456 		txq->ift_pullups = 0;
2457 		ifmp_ring_reset_stats(txq->ift_br);
2458 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2459 			bzero((void *)di->idi_vaddr, di->idi_size);
2460 	}
2461 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2462 		/* make sure all transmitters have completed before proceeding XXX */
2463 
2464 		rxq->ifr_cq_cidx = 0;
2465 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2466 			bzero((void *)di->idi_vaddr, di->idi_size);
2467 		/* also resets the free lists pidx/cidx */
2468 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2469 			iflib_fl_bufs_free(fl);
2470 	}
2471 }
2472 
2473 static inline caddr_t
2474 calc_next_rxd(iflib_fl_t fl, int cidx)
2475 {
2476 	qidx_t size;
2477 	int nrxd;
2478 	caddr_t start, end, cur, next;
2479 
2480 	nrxd = fl->ifl_size;
2481 	size = fl->ifl_rxd_size;
2482 	start = fl->ifl_ifdi->idi_vaddr;
2483 
2484 	if (__predict_false(size == 0))
2485 		return (start);
2486 	cur = start + size*cidx;
2487 	end = start + size*nrxd;
2488 	next = CACHE_PTR_NEXT(cur);
2489 	return (next < end ? next : start);
2490 }
2491 
2492 static inline void
2493 prefetch_pkts(iflib_fl_t fl, int cidx)
2494 {
2495 	int nextptr;
2496 	int nrxd = fl->ifl_size;
2497 	caddr_t next_rxd;
2498 
2499 
2500 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2501 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2502 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2503 	next_rxd = calc_next_rxd(fl, cidx);
2504 	prefetch(next_rxd);
2505 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2506 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2507 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2508 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2509 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2510 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2511 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2512 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2513 }
2514 
2515 static struct mbuf *
2516 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2517     int *pf_rv, if_rxd_info_t ri)
2518 {
2519 	bus_dmamap_t map;
2520 	iflib_fl_t fl;
2521 	caddr_t payload;
2522 	struct mbuf *m;
2523 	int flid, cidx, len, next;
2524 
2525 	map = NULL;
2526 	flid = irf->irf_flid;
2527 	cidx = irf->irf_idx;
2528 	fl = &rxq->ifr_fl[flid];
2529 	sd->ifsd_fl = fl;
2530 	sd->ifsd_cidx = cidx;
2531 	m = fl->ifl_sds.ifsd_m[cidx];
2532 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2533 	fl->ifl_credits--;
2534 #if MEMORY_LOGGING
2535 	fl->ifl_m_dequeued++;
2536 #endif
2537 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2538 		prefetch_pkts(fl, cidx);
2539 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2540 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2541 	map = fl->ifl_sds.ifsd_map[cidx];
2542 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2543 
2544 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2545 	MPASS(fl->ifl_cidx == cidx);
2546 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2547 
2548 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2549 		payload  = *sd->ifsd_cl;
2550 		payload +=  ri->iri_pad;
2551 		len = ri->iri_len - ri->iri_pad;
2552 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2553 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2554 		switch (*pf_rv) {
2555 		case PFIL_DROPPED:
2556 		case PFIL_CONSUMED:
2557 			/*
2558 			 * The filter ate it.  Everything is recycled.
2559 			 */
2560 			m = NULL;
2561 			unload = 0;
2562 			break;
2563 		case PFIL_REALLOCED:
2564 			/*
2565 			 * The filter copied it.  Everything is recycled.
2566 			 */
2567 			m = pfil_mem2mbuf(payload);
2568 			unload = 0;
2569 			break;
2570 		case PFIL_PASS:
2571 			/*
2572 			 * Filter said it was OK, so receive like
2573 			 * normal
2574 			 */
2575 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2576 			break;
2577 		default:
2578 			MPASS(0);
2579 		}
2580 	} else {
2581 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2582 		*pf_rv = PFIL_PASS;
2583 	}
2584 
2585 	if (unload)
2586 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2587 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2588 	if (__predict_false(fl->ifl_cidx == 0))
2589 		fl->ifl_gen = 0;
2590 	bit_clear(fl->ifl_rx_bitmap, cidx);
2591 	return (m);
2592 }
2593 
2594 static struct mbuf *
2595 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2596 {
2597 	struct mbuf *m, *mh, *mt;
2598 	caddr_t cl;
2599 	int  *pf_rv_ptr, flags, i, padlen;
2600 	bool consumed;
2601 
2602 	i = 0;
2603 	mh = NULL;
2604 	consumed = false;
2605 	*pf_rv = PFIL_PASS;
2606 	pf_rv_ptr = pf_rv;
2607 	do {
2608 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2609 		    pf_rv_ptr, ri);
2610 
2611 		MPASS(*sd->ifsd_cl != NULL);
2612 
2613 		/*
2614 		 * Exclude zero-length frags & frags from
2615 		 * packets the filter has consumed or dropped
2616 		 */
2617 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2618 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2619 			if (mh == NULL) {
2620 				/* everything saved here */
2621 				consumed = true;
2622 				pf_rv_ptr = NULL;
2623 				continue;
2624 			}
2625 			/* XXX we can save the cluster here, but not the mbuf */
2626 			m_init(m, M_NOWAIT, MT_DATA, 0);
2627 			m_free(m);
2628 			continue;
2629 		}
2630 		if (mh == NULL) {
2631 			flags = M_PKTHDR|M_EXT;
2632 			mh = mt = m;
2633 			padlen = ri->iri_pad;
2634 		} else {
2635 			flags = M_EXT;
2636 			mt->m_next = m;
2637 			mt = m;
2638 			/* assuming padding is only on the first fragment */
2639 			padlen = 0;
2640 		}
2641 		cl = *sd->ifsd_cl;
2642 		*sd->ifsd_cl = NULL;
2643 
2644 		/* Can these two be made one ? */
2645 		m_init(m, M_NOWAIT, MT_DATA, flags);
2646 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2647 		/*
2648 		 * These must follow m_init and m_cljset
2649 		 */
2650 		m->m_data += padlen;
2651 		ri->iri_len -= padlen;
2652 		m->m_len = ri->iri_frags[i].irf_len;
2653 	} while (++i < ri->iri_nfrags);
2654 
2655 	return (mh);
2656 }
2657 
2658 /*
2659  * Process one software descriptor
2660  */
2661 static struct mbuf *
2662 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2663 {
2664 	struct if_rxsd sd;
2665 	struct mbuf *m;
2666 	int pf_rv;
2667 
2668 	/* should I merge this back in now that the two paths are basically duplicated? */
2669 	if (ri->iri_nfrags == 1 &&
2670 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2671 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2672 		    &pf_rv, ri);
2673 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2674 			return (m);
2675 		if (pf_rv == PFIL_PASS) {
2676 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2677 #ifndef __NO_STRICT_ALIGNMENT
2678 			if (!IP_ALIGNED(m))
2679 				m->m_data += 2;
2680 #endif
2681 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2682 			m->m_len = ri->iri_frags[0].irf_len;
2683 		}
2684 	} else {
2685 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2686 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2687 			return (m);
2688 	}
2689 	m->m_pkthdr.len = ri->iri_len;
2690 	m->m_pkthdr.rcvif = ri->iri_ifp;
2691 	m->m_flags |= ri->iri_flags;
2692 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2693 	m->m_pkthdr.flowid = ri->iri_flowid;
2694 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2695 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2696 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2697 	return (m);
2698 }
2699 
2700 #if defined(INET6) || defined(INET)
2701 static void
2702 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2703 {
2704 	CURVNET_SET(lc->ifp->if_vnet);
2705 #if defined(INET6)
2706 	*v6 = V_ip6_forwarding;
2707 #endif
2708 #if defined(INET)
2709 	*v4 = V_ipforwarding;
2710 #endif
2711 	CURVNET_RESTORE();
2712 }
2713 
2714 /*
2715  * Returns true if it's possible this packet could be LROed.
2716  * if it returns false, it is guaranteed that tcp_lro_rx()
2717  * would not return zero.
2718  */
2719 static bool
2720 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2721 {
2722 	struct ether_header *eh;
2723 
2724 	eh = mtod(m, struct ether_header *);
2725 	switch (eh->ether_type) {
2726 #if defined(INET6)
2727 		case htons(ETHERTYPE_IPV6):
2728 			return (!v6_forwarding);
2729 #endif
2730 #if defined (INET)
2731 		case htons(ETHERTYPE_IP):
2732 			return (!v4_forwarding);
2733 #endif
2734 	}
2735 
2736 	return false;
2737 }
2738 #else
2739 static void
2740 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2741 {
2742 }
2743 #endif
2744 
2745 static bool
2746 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2747 {
2748 	if_t ifp;
2749 	if_ctx_t ctx = rxq->ifr_ctx;
2750 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2751 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2752 	int avail, i;
2753 	qidx_t *cidxp;
2754 	struct if_rxd_info ri;
2755 	int err, budget_left, rx_bytes, rx_pkts;
2756 	iflib_fl_t fl;
2757 	int lro_enabled;
2758 	bool v4_forwarding, v6_forwarding, lro_possible;
2759 
2760 	/*
2761 	 * XXX early demux data packets so that if_input processing only handles
2762 	 * acks in interrupt context
2763 	 */
2764 	struct mbuf *m, *mh, *mt, *mf;
2765 
2766 	lro_possible = v4_forwarding = v6_forwarding = false;
2767 	ifp = ctx->ifc_ifp;
2768 	mh = mt = NULL;
2769 	MPASS(budget > 0);
2770 	rx_pkts	= rx_bytes = 0;
2771 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2772 		cidxp = &rxq->ifr_cq_cidx;
2773 	else
2774 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2775 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2776 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2777 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2778 		DBG_COUNTER_INC(rx_unavail);
2779 		return (false);
2780 	}
2781 
2782 	/* pfil needs the vnet to be set */
2783 	CURVNET_SET_QUIET(ifp->if_vnet);
2784 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2785 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2786 			DBG_COUNTER_INC(rx_ctx_inactive);
2787 			break;
2788 		}
2789 		/*
2790 		 * Reset client set fields to their default values
2791 		 */
2792 		rxd_info_zero(&ri);
2793 		ri.iri_qsidx = rxq->ifr_id;
2794 		ri.iri_cidx = *cidxp;
2795 		ri.iri_ifp = ifp;
2796 		ri.iri_frags = rxq->ifr_frags;
2797 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2798 
2799 		if (err)
2800 			goto err;
2801 		rx_pkts += 1;
2802 		rx_bytes += ri.iri_len;
2803 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2804 			*cidxp = ri.iri_cidx;
2805 			/* Update our consumer index */
2806 			/* XXX NB: shurd - check if this is still safe */
2807 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2808 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2809 			/* was this only a completion queue message? */
2810 			if (__predict_false(ri.iri_nfrags == 0))
2811 				continue;
2812 		}
2813 		MPASS(ri.iri_nfrags != 0);
2814 		MPASS(ri.iri_len != 0);
2815 
2816 		/* will advance the cidx on the corresponding free lists */
2817 		m = iflib_rxd_pkt_get(rxq, &ri);
2818 		avail--;
2819 		budget_left--;
2820 		if (avail == 0 && budget_left)
2821 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2822 
2823 		if (__predict_false(m == NULL))
2824 			continue;
2825 
2826 		/* imm_pkt: -- cxgb */
2827 		if (mh == NULL)
2828 			mh = mt = m;
2829 		else {
2830 			mt->m_nextpkt = m;
2831 			mt = m;
2832 		}
2833 	}
2834 	CURVNET_RESTORE();
2835 	/* make sure that we can refill faster than drain */
2836 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2837 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2838 
2839 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2840 	if (lro_enabled)
2841 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2842 	mt = mf = NULL;
2843 	while (mh != NULL) {
2844 		m = mh;
2845 		mh = mh->m_nextpkt;
2846 		m->m_nextpkt = NULL;
2847 #ifndef __NO_STRICT_ALIGNMENT
2848 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2849 			continue;
2850 #endif
2851 		rx_bytes += m->m_pkthdr.len;
2852 		rx_pkts++;
2853 #if defined(INET6) || defined(INET)
2854 		if (lro_enabled) {
2855 			if (!lro_possible) {
2856 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2857 				if (lro_possible && mf != NULL) {
2858 					ifp->if_input(ifp, mf);
2859 					DBG_COUNTER_INC(rx_if_input);
2860 					mt = mf = NULL;
2861 				}
2862 			}
2863 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2864 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2865 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2866 					continue;
2867 			}
2868 		}
2869 #endif
2870 		if (lro_possible) {
2871 			ifp->if_input(ifp, m);
2872 			DBG_COUNTER_INC(rx_if_input);
2873 			continue;
2874 		}
2875 
2876 		if (mf == NULL)
2877 			mf = m;
2878 		if (mt != NULL)
2879 			mt->m_nextpkt = m;
2880 		mt = m;
2881 	}
2882 	if (mf != NULL) {
2883 		ifp->if_input(ifp, mf);
2884 		DBG_COUNTER_INC(rx_if_input);
2885 	}
2886 
2887 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2888 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2889 
2890 	/*
2891 	 * Flush any outstanding LRO work
2892 	 */
2893 #if defined(INET6) || defined(INET)
2894 	tcp_lro_flush_all(&rxq->ifr_lc);
2895 #endif
2896 	if (avail)
2897 		return true;
2898 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2899 err:
2900 	STATE_LOCK(ctx);
2901 	ctx->ifc_flags |= IFC_DO_RESET;
2902 	iflib_admin_intr_deferred(ctx);
2903 	STATE_UNLOCK(ctx);
2904 	return (false);
2905 }
2906 
2907 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2908 static inline qidx_t
2909 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2910 {
2911 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2912 	qidx_t minthresh = txq->ift_size / 8;
2913 	if (in_use > 4*minthresh)
2914 		return (notify_count);
2915 	if (in_use > 2*minthresh)
2916 		return (notify_count >> 1);
2917 	if (in_use > minthresh)
2918 		return (notify_count >> 3);
2919 	return (0);
2920 }
2921 
2922 static inline qidx_t
2923 txq_max_rs_deferred(iflib_txq_t txq)
2924 {
2925 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2926 	qidx_t minthresh = txq->ift_size / 8;
2927 	if (txq->ift_in_use > 4*minthresh)
2928 		return (notify_count);
2929 	if (txq->ift_in_use > 2*minthresh)
2930 		return (notify_count >> 1);
2931 	if (txq->ift_in_use > minthresh)
2932 		return (notify_count >> 2);
2933 	return (2);
2934 }
2935 
2936 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2937 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2938 
2939 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2940 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2941 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2942 
2943 /* forward compatibility for cxgb */
2944 #define FIRST_QSET(ctx) 0
2945 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2946 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2947 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2948 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2949 
2950 /* XXX we should be setting this to something other than zero */
2951 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2952 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2953     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2954 
2955 static inline bool
2956 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2957 {
2958 	qidx_t dbval, max;
2959 	bool rang;
2960 
2961 	rang = false;
2962 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2963 	if (ring || txq->ift_db_pending >= max) {
2964 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2965 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2966 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2967 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2968 		txq->ift_db_pending = txq->ift_npending = 0;
2969 		rang = true;
2970 	}
2971 	return (rang);
2972 }
2973 
2974 #ifdef PKT_DEBUG
2975 static void
2976 print_pkt(if_pkt_info_t pi)
2977 {
2978 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2979 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2980 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2981 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2982 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2983 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2984 }
2985 #endif
2986 
2987 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2988 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2989 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2990 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2991 
2992 static int
2993 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2994 {
2995 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2996 	struct ether_vlan_header *eh;
2997 	struct mbuf *m;
2998 
2999 	m = *mp;
3000 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3001 	    M_WRITABLE(m) == 0) {
3002 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3003 			return (ENOMEM);
3004 		} else {
3005 			m_freem(*mp);
3006 			DBG_COUNTER_INC(tx_frees);
3007 			*mp = m;
3008 		}
3009 	}
3010 
3011 	/*
3012 	 * Determine where frame payload starts.
3013 	 * Jump over vlan headers if already present,
3014 	 * helpful for QinQ too.
3015 	 */
3016 	if (__predict_false(m->m_len < sizeof(*eh))) {
3017 		txq->ift_pullups++;
3018 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3019 			return (ENOMEM);
3020 	}
3021 	eh = mtod(m, struct ether_vlan_header *);
3022 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3023 		pi->ipi_etype = ntohs(eh->evl_proto);
3024 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3025 	} else {
3026 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3027 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3028 	}
3029 
3030 	switch (pi->ipi_etype) {
3031 #ifdef INET
3032 	case ETHERTYPE_IP:
3033 	{
3034 		struct mbuf *n;
3035 		struct ip *ip = NULL;
3036 		struct tcphdr *th = NULL;
3037 		int minthlen;
3038 
3039 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3040 		if (__predict_false(m->m_len < minthlen)) {
3041 			/*
3042 			 * if this code bloat is causing too much of a hit
3043 			 * move it to a separate function and mark it noinline
3044 			 */
3045 			if (m->m_len == pi->ipi_ehdrlen) {
3046 				n = m->m_next;
3047 				MPASS(n);
3048 				if (n->m_len >= sizeof(*ip))  {
3049 					ip = (struct ip *)n->m_data;
3050 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3051 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3052 				} else {
3053 					txq->ift_pullups++;
3054 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3055 						return (ENOMEM);
3056 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3057 				}
3058 			} else {
3059 				txq->ift_pullups++;
3060 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3061 					return (ENOMEM);
3062 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3063 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3064 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3065 			}
3066 		} else {
3067 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3068 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3069 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3070 		}
3071 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3072 		pi->ipi_ipproto = ip->ip_p;
3073 		pi->ipi_flags |= IPI_TX_IPV4;
3074 
3075 		/* TCP checksum offload may require TCP header length */
3076 		if (IS_TX_OFFLOAD4(pi)) {
3077 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3078 				if (__predict_false(th == NULL)) {
3079 					txq->ift_pullups++;
3080 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3081 						return (ENOMEM);
3082 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3083 				}
3084 				pi->ipi_tcp_hflags = th->th_flags;
3085 				pi->ipi_tcp_hlen = th->th_off << 2;
3086 				pi->ipi_tcp_seq = th->th_seq;
3087 			}
3088 			if (IS_TSO4(pi)) {
3089 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3090 					return (ENXIO);
3091 				/*
3092 				 * TSO always requires hardware checksum offload.
3093 				 */
3094 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3095 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3096 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3097 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3098 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3099 					ip->ip_sum = 0;
3100 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3101 				}
3102 			}
3103 		}
3104 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3105                        ip->ip_sum = 0;
3106 
3107 		break;
3108 	}
3109 #endif
3110 #ifdef INET6
3111 	case ETHERTYPE_IPV6:
3112 	{
3113 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3114 		struct tcphdr *th;
3115 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3116 
3117 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3118 			txq->ift_pullups++;
3119 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3120 				return (ENOMEM);
3121 		}
3122 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3123 
3124 		/* XXX-BZ this will go badly in case of ext hdrs. */
3125 		pi->ipi_ipproto = ip6->ip6_nxt;
3126 		pi->ipi_flags |= IPI_TX_IPV6;
3127 
3128 		/* TCP checksum offload may require TCP header length */
3129 		if (IS_TX_OFFLOAD6(pi)) {
3130 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3131 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3132 					txq->ift_pullups++;
3133 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3134 						return (ENOMEM);
3135 				}
3136 				pi->ipi_tcp_hflags = th->th_flags;
3137 				pi->ipi_tcp_hlen = th->th_off << 2;
3138 				pi->ipi_tcp_seq = th->th_seq;
3139 			}
3140 			if (IS_TSO6(pi)) {
3141 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3142 					return (ENXIO);
3143 				/*
3144 				 * TSO always requires hardware checksum offload.
3145 				 */
3146 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3147 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3148 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3149 			}
3150 		}
3151 		break;
3152 	}
3153 #endif
3154 	default:
3155 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3156 		pi->ipi_ip_hlen = 0;
3157 		break;
3158 	}
3159 	*mp = m;
3160 
3161 	return (0);
3162 }
3163 
3164 /*
3165  * If dodgy hardware rejects the scatter gather chain we've handed it
3166  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3167  * m_defrag'd mbufs
3168  */
3169 static __noinline struct mbuf *
3170 iflib_remove_mbuf(iflib_txq_t txq)
3171 {
3172 	int ntxd, pidx;
3173 	struct mbuf *m, **ifsd_m;
3174 
3175 	ifsd_m = txq->ift_sds.ifsd_m;
3176 	ntxd = txq->ift_size;
3177 	pidx = txq->ift_pidx & (ntxd - 1);
3178 	ifsd_m = txq->ift_sds.ifsd_m;
3179 	m = ifsd_m[pidx];
3180 	ifsd_m[pidx] = NULL;
3181 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3182 	if (txq->ift_sds.ifsd_tso_map != NULL)
3183 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3184 		    txq->ift_sds.ifsd_tso_map[pidx]);
3185 #if MEMORY_LOGGING
3186 	txq->ift_dequeued++;
3187 #endif
3188 	return (m);
3189 }
3190 
3191 static inline caddr_t
3192 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3193 {
3194 	qidx_t size;
3195 	int ntxd;
3196 	caddr_t start, end, cur, next;
3197 
3198 	ntxd = txq->ift_size;
3199 	size = txq->ift_txd_size[qid];
3200 	start = txq->ift_ifdi[qid].idi_vaddr;
3201 
3202 	if (__predict_false(size == 0))
3203 		return (start);
3204 	cur = start + size*cidx;
3205 	end = start + size*ntxd;
3206 	next = CACHE_PTR_NEXT(cur);
3207 	return (next < end ? next : start);
3208 }
3209 
3210 /*
3211  * Pad an mbuf to ensure a minimum ethernet frame size.
3212  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3213  */
3214 static __noinline int
3215 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3216 {
3217 	/*
3218 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3219 	 * and ARP message is the smallest common payload I can think of
3220 	 */
3221 	static char pad[18];	/* just zeros */
3222 	int n;
3223 	struct mbuf *new_head;
3224 
3225 	if (!M_WRITABLE(*m_head)) {
3226 		new_head = m_dup(*m_head, M_NOWAIT);
3227 		if (new_head == NULL) {
3228 			m_freem(*m_head);
3229 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3230 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3231 			DBG_COUNTER_INC(tx_frees);
3232 			return ENOMEM;
3233 		}
3234 		m_freem(*m_head);
3235 		*m_head = new_head;
3236 	}
3237 
3238 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3239 	     n > 0; n -= sizeof(pad))
3240 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3241 			break;
3242 
3243 	if (n > 0) {
3244 		m_freem(*m_head);
3245 		device_printf(dev, "cannot pad short frame\n");
3246 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3247 		DBG_COUNTER_INC(tx_frees);
3248 		return (ENOBUFS);
3249 	}
3250 
3251 	return 0;
3252 }
3253 
3254 static int
3255 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3256 {
3257 	if_ctx_t		ctx;
3258 	if_shared_ctx_t		sctx;
3259 	if_softc_ctx_t		scctx;
3260 	bus_dma_tag_t		buf_tag;
3261 	bus_dma_segment_t	*segs;
3262 	struct mbuf		*m_head, **ifsd_m;
3263 	void			*next_txd;
3264 	bus_dmamap_t		map;
3265 	struct if_pkt_info	pi;
3266 	int remap = 0;
3267 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3268 
3269 	ctx = txq->ift_ctx;
3270 	sctx = ctx->ifc_sctx;
3271 	scctx = &ctx->ifc_softc_ctx;
3272 	segs = txq->ift_segs;
3273 	ntxd = txq->ift_size;
3274 	m_head = *m_headp;
3275 	map = NULL;
3276 
3277 	/*
3278 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3279 	 */
3280 	cidx = txq->ift_cidx;
3281 	pidx = txq->ift_pidx;
3282 	if (ctx->ifc_flags & IFC_PREFETCH) {
3283 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3284 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3285 			next_txd = calc_next_txd(txq, cidx, 0);
3286 			prefetch(next_txd);
3287 		}
3288 
3289 		/* prefetch the next cache line of mbuf pointers and flags */
3290 		prefetch(&txq->ift_sds.ifsd_m[next]);
3291 		prefetch(&txq->ift_sds.ifsd_map[next]);
3292 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3293 	}
3294 	map = txq->ift_sds.ifsd_map[pidx];
3295 	ifsd_m = txq->ift_sds.ifsd_m;
3296 
3297 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3298 		buf_tag = txq->ift_tso_buf_tag;
3299 		max_segs = scctx->isc_tx_tso_segments_max;
3300 		map = txq->ift_sds.ifsd_tso_map[pidx];
3301 		MPASS(buf_tag != NULL);
3302 		MPASS(max_segs > 0);
3303 	} else {
3304 		buf_tag = txq->ift_buf_tag;
3305 		max_segs = scctx->isc_tx_nsegments;
3306 		map = txq->ift_sds.ifsd_map[pidx];
3307 	}
3308 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3309 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3310 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3311 		if (err) {
3312 			DBG_COUNTER_INC(encap_txd_encap_fail);
3313 			return err;
3314 		}
3315 	}
3316 	m_head = *m_headp;
3317 
3318 	pkt_info_zero(&pi);
3319 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3320 	pi.ipi_pidx = pidx;
3321 	pi.ipi_qsidx = txq->ift_id;
3322 	pi.ipi_len = m_head->m_pkthdr.len;
3323 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3324 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3325 
3326 	/* deliberate bitwise OR to make one condition */
3327 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3328 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3329 			DBG_COUNTER_INC(encap_txd_encap_fail);
3330 			return (err);
3331 		}
3332 		m_head = *m_headp;
3333 	}
3334 
3335 retry:
3336 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3337 	    BUS_DMA_NOWAIT);
3338 defrag:
3339 	if (__predict_false(err)) {
3340 		switch (err) {
3341 		case EFBIG:
3342 			/* try collapse once and defrag once */
3343 			if (remap == 0) {
3344 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3345 				/* try defrag if collapsing fails */
3346 				if (m_head == NULL)
3347 					remap++;
3348 			}
3349 			if (remap == 1) {
3350 				txq->ift_mbuf_defrag++;
3351 				m_head = m_defrag(*m_headp, M_NOWAIT);
3352 			}
3353 			/*
3354 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3355 			 * failed to map an mbuf that was run through m_defrag
3356 			 */
3357 			MPASS(remap <= 1);
3358 			if (__predict_false(m_head == NULL || remap > 1))
3359 				goto defrag_failed;
3360 			remap++;
3361 			*m_headp = m_head;
3362 			goto retry;
3363 			break;
3364 		case ENOMEM:
3365 			txq->ift_no_tx_dma_setup++;
3366 			break;
3367 		default:
3368 			txq->ift_no_tx_dma_setup++;
3369 			m_freem(*m_headp);
3370 			DBG_COUNTER_INC(tx_frees);
3371 			*m_headp = NULL;
3372 			break;
3373 		}
3374 		txq->ift_map_failed++;
3375 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3376 		DBG_COUNTER_INC(encap_txd_encap_fail);
3377 		return (err);
3378 	}
3379 	ifsd_m[pidx] = m_head;
3380 	/*
3381 	 * XXX assumes a 1 to 1 relationship between segments and
3382 	 *        descriptors - this does not hold true on all drivers, e.g.
3383 	 *        cxgb
3384 	 */
3385 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3386 		txq->ift_no_desc_avail++;
3387 		bus_dmamap_unload(buf_tag, map);
3388 		DBG_COUNTER_INC(encap_txq_avail_fail);
3389 		DBG_COUNTER_INC(encap_txd_encap_fail);
3390 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3391 			GROUPTASK_ENQUEUE(&txq->ift_task);
3392 		return (ENOBUFS);
3393 	}
3394 	/*
3395 	 * On Intel cards we can greatly reduce the number of TX interrupts
3396 	 * we see by only setting report status on every Nth descriptor.
3397 	 * However, this also means that the driver will need to keep track
3398 	 * of the descriptors that RS was set on to check them for the DD bit.
3399 	 */
3400 	txq->ift_rs_pending += nsegs + 1;
3401 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3402 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3403 		pi.ipi_flags |= IPI_TX_INTR;
3404 		txq->ift_rs_pending = 0;
3405 	}
3406 
3407 	pi.ipi_segs = segs;
3408 	pi.ipi_nsegs = nsegs;
3409 
3410 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3411 #ifdef PKT_DEBUG
3412 	print_pkt(&pi);
3413 #endif
3414 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3415 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3416 		DBG_COUNTER_INC(tx_encap);
3417 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3418 
3419 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3420 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3421 			ndesc += txq->ift_size;
3422 			txq->ift_gen = 1;
3423 		}
3424 		/*
3425 		 * drivers can need as many as
3426 		 * two sentinels
3427 		 */
3428 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3429 		MPASS(pi.ipi_new_pidx != pidx);
3430 		MPASS(ndesc > 0);
3431 		txq->ift_in_use += ndesc;
3432 
3433 		/*
3434 		 * We update the last software descriptor again here because there may
3435 		 * be a sentinel and/or there may be more mbufs than segments
3436 		 */
3437 		txq->ift_pidx = pi.ipi_new_pidx;
3438 		txq->ift_npending += pi.ipi_ndescs;
3439 	} else {
3440 		*m_headp = m_head = iflib_remove_mbuf(txq);
3441 		if (err == EFBIG) {
3442 			txq->ift_txd_encap_efbig++;
3443 			if (remap < 2) {
3444 				remap = 1;
3445 				goto defrag;
3446 			}
3447 		}
3448 		goto defrag_failed;
3449 	}
3450 	/*
3451 	 * err can't possibly be non-zero here, so we don't neet to test it
3452 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3453 	 */
3454 	return (err);
3455 
3456 defrag_failed:
3457 	txq->ift_mbuf_defrag_failed++;
3458 	txq->ift_map_failed++;
3459 	m_freem(*m_headp);
3460 	DBG_COUNTER_INC(tx_frees);
3461 	*m_headp = NULL;
3462 	DBG_COUNTER_INC(encap_txd_encap_fail);
3463 	return (ENOMEM);
3464 }
3465 
3466 static void
3467 iflib_tx_desc_free(iflib_txq_t txq, int n)
3468 {
3469 	uint32_t qsize, cidx, mask, gen;
3470 	struct mbuf *m, **ifsd_m;
3471 	bool do_prefetch;
3472 
3473 	cidx = txq->ift_cidx;
3474 	gen = txq->ift_gen;
3475 	qsize = txq->ift_size;
3476 	mask = qsize-1;
3477 	ifsd_m = txq->ift_sds.ifsd_m;
3478 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3479 
3480 	while (n-- > 0) {
3481 		if (do_prefetch) {
3482 			prefetch(ifsd_m[(cidx + 3) & mask]);
3483 			prefetch(ifsd_m[(cidx + 4) & mask]);
3484 		}
3485 		if ((m = ifsd_m[cidx]) != NULL) {
3486 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3487 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3488 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3489 				    txq->ift_sds.ifsd_tso_map[cidx],
3490 				    BUS_DMASYNC_POSTWRITE);
3491 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3492 				    txq->ift_sds.ifsd_tso_map[cidx]);
3493 			} else {
3494 				bus_dmamap_sync(txq->ift_buf_tag,
3495 				    txq->ift_sds.ifsd_map[cidx],
3496 				    BUS_DMASYNC_POSTWRITE);
3497 				bus_dmamap_unload(txq->ift_buf_tag,
3498 				    txq->ift_sds.ifsd_map[cidx]);
3499 			}
3500 			/* XXX we don't support any drivers that batch packets yet */
3501 			MPASS(m->m_nextpkt == NULL);
3502 			m_freem(m);
3503 			ifsd_m[cidx] = NULL;
3504 #if MEMORY_LOGGING
3505 			txq->ift_dequeued++;
3506 #endif
3507 			DBG_COUNTER_INC(tx_frees);
3508 		}
3509 		if (__predict_false(++cidx == qsize)) {
3510 			cidx = 0;
3511 			gen = 0;
3512 		}
3513 	}
3514 	txq->ift_cidx = cidx;
3515 	txq->ift_gen = gen;
3516 }
3517 
3518 static __inline int
3519 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3520 {
3521 	int reclaim;
3522 	if_ctx_t ctx = txq->ift_ctx;
3523 
3524 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3525 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3526 
3527 	/*
3528 	 * Need a rate-limiting check so that this isn't called every time
3529 	 */
3530 	iflib_tx_credits_update(ctx, txq);
3531 	reclaim = DESC_RECLAIMABLE(txq);
3532 
3533 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3534 #ifdef INVARIANTS
3535 		if (iflib_verbose_debug) {
3536 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3537 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3538 			       reclaim, thresh);
3539 
3540 		}
3541 #endif
3542 		return (0);
3543 	}
3544 	iflib_tx_desc_free(txq, reclaim);
3545 	txq->ift_cleaned += reclaim;
3546 	txq->ift_in_use -= reclaim;
3547 
3548 	return (reclaim);
3549 }
3550 
3551 static struct mbuf **
3552 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3553 {
3554 	int next, size;
3555 	struct mbuf **items;
3556 
3557 	size = r->size;
3558 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3559 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3560 
3561 	prefetch(items[(cidx + offset) & (size-1)]);
3562 	if (remaining > 1) {
3563 		prefetch2cachelines(&items[next]);
3564 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3565 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3566 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3567 	}
3568 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3569 }
3570 
3571 static void
3572 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3573 {
3574 
3575 	ifmp_ring_check_drainage(txq->ift_br, budget);
3576 }
3577 
3578 static uint32_t
3579 iflib_txq_can_drain(struct ifmp_ring *r)
3580 {
3581 	iflib_txq_t txq = r->cookie;
3582 	if_ctx_t ctx = txq->ift_ctx;
3583 
3584 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3585 		return (1);
3586 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3587 	    BUS_DMASYNC_POSTREAD);
3588 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3589 	    false));
3590 }
3591 
3592 static uint32_t
3593 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3594 {
3595 	iflib_txq_t txq = r->cookie;
3596 	if_ctx_t ctx = txq->ift_ctx;
3597 	if_t ifp = ctx->ifc_ifp;
3598 	struct mbuf *m, **mp;
3599 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3600 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3601 	bool do_prefetch, rang, ring;
3602 
3603 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3604 			    !LINK_ACTIVE(ctx))) {
3605 		DBG_COUNTER_INC(txq_drain_notready);
3606 		return (0);
3607 	}
3608 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3609 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3610 	avail = IDXDIFF(pidx, cidx, r->size);
3611 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3612 		DBG_COUNTER_INC(txq_drain_flushing);
3613 		for (i = 0; i < avail; i++) {
3614 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3615 				m_free(r->items[(cidx + i) & (r->size-1)]);
3616 			r->items[(cidx + i) & (r->size-1)] = NULL;
3617 		}
3618 		return (avail);
3619 	}
3620 
3621 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3622 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3623 		CALLOUT_LOCK(txq);
3624 		callout_stop(&txq->ift_timer);
3625 		CALLOUT_UNLOCK(txq);
3626 		DBG_COUNTER_INC(txq_drain_oactive);
3627 		return (0);
3628 	}
3629 	if (reclaimed)
3630 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3631 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3632 	count = MIN(avail, TX_BATCH_SIZE);
3633 #ifdef INVARIANTS
3634 	if (iflib_verbose_debug)
3635 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3636 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3637 #endif
3638 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3639 	txq_avail = TXQ_AVAIL(txq);
3640 	err = 0;
3641 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3642 		int rem = do_prefetch ? count - i : 0;
3643 
3644 		mp = _ring_peek_one(r, cidx, i, rem);
3645 		MPASS(mp != NULL && *mp != NULL);
3646 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3647 			consumed++;
3648 			continue;
3649 		}
3650 		in_use_prev = txq->ift_in_use;
3651 		err = iflib_encap(txq, mp);
3652 		if (__predict_false(err)) {
3653 			/* no room - bail out */
3654 			if (err == ENOBUFS)
3655 				break;
3656 			consumed++;
3657 			/* we can't send this packet - skip it */
3658 			continue;
3659 		}
3660 		consumed++;
3661 		pkt_sent++;
3662 		m = *mp;
3663 		DBG_COUNTER_INC(tx_sent);
3664 		bytes_sent += m->m_pkthdr.len;
3665 		mcast_sent += !!(m->m_flags & M_MCAST);
3666 		txq_avail = TXQ_AVAIL(txq);
3667 
3668 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3669 		ETHER_BPF_MTAP(ifp, m);
3670 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3671 			break;
3672 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3673 	}
3674 
3675 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3676 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3677 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3678 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3679 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3680 	if (mcast_sent)
3681 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3682 #ifdef INVARIANTS
3683 	if (iflib_verbose_debug)
3684 		printf("consumed=%d\n", consumed);
3685 #endif
3686 	return (consumed);
3687 }
3688 
3689 static uint32_t
3690 iflib_txq_drain_always(struct ifmp_ring *r)
3691 {
3692 	return (1);
3693 }
3694 
3695 static uint32_t
3696 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3697 {
3698 	int i, avail;
3699 	struct mbuf **mp;
3700 	iflib_txq_t txq;
3701 
3702 	txq = r->cookie;
3703 
3704 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3705 	CALLOUT_LOCK(txq);
3706 	callout_stop(&txq->ift_timer);
3707 	CALLOUT_UNLOCK(txq);
3708 
3709 	avail = IDXDIFF(pidx, cidx, r->size);
3710 	for (i = 0; i < avail; i++) {
3711 		mp = _ring_peek_one(r, cidx, i, avail - i);
3712 		if (__predict_false(*mp == (struct mbuf *)txq))
3713 			continue;
3714 		m_freem(*mp);
3715 		DBG_COUNTER_INC(tx_frees);
3716 	}
3717 	MPASS(ifmp_ring_is_stalled(r) == 0);
3718 	return (avail);
3719 }
3720 
3721 static void
3722 iflib_ifmp_purge(iflib_txq_t txq)
3723 {
3724 	struct ifmp_ring *r;
3725 
3726 	r = txq->ift_br;
3727 	r->drain = iflib_txq_drain_free;
3728 	r->can_drain = iflib_txq_drain_always;
3729 
3730 	ifmp_ring_check_drainage(r, r->size);
3731 
3732 	r->drain = iflib_txq_drain;
3733 	r->can_drain = iflib_txq_can_drain;
3734 }
3735 
3736 static void
3737 _task_fn_tx(void *context)
3738 {
3739 	iflib_txq_t txq = context;
3740 	if_ctx_t ctx = txq->ift_ctx;
3741 #if defined(ALTQ) || defined(DEV_NETMAP)
3742 	if_t ifp = ctx->ifc_ifp;
3743 #endif
3744 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3745 
3746 #ifdef IFLIB_DIAGNOSTICS
3747 	txq->ift_cpu_exec_count[curcpu]++;
3748 #endif
3749 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3750 		return;
3751 #ifdef DEV_NETMAP
3752 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3753 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3754 		    BUS_DMASYNC_POSTREAD);
3755 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3756 			netmap_tx_irq(ifp, txq->ift_id);
3757 		if (ctx->ifc_flags & IFC_LEGACY)
3758 			IFDI_INTR_ENABLE(ctx);
3759 		else
3760 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3761 		return;
3762 	}
3763 #endif
3764 #ifdef ALTQ
3765 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3766 		iflib_altq_if_start(ifp);
3767 #endif
3768 	if (txq->ift_db_pending)
3769 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3770 	else if (!abdicate)
3771 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3772 	/*
3773 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3774 	 */
3775 	if (abdicate)
3776 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3777 	if (ctx->ifc_flags & IFC_LEGACY)
3778 		IFDI_INTR_ENABLE(ctx);
3779 	else
3780 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3781 }
3782 
3783 static void
3784 _task_fn_rx(void *context)
3785 {
3786 	iflib_rxq_t rxq = context;
3787 	if_ctx_t ctx = rxq->ifr_ctx;
3788 	bool more;
3789 	uint16_t budget;
3790 
3791 #ifdef IFLIB_DIAGNOSTICS
3792 	rxq->ifr_cpu_exec_count[curcpu]++;
3793 #endif
3794 	DBG_COUNTER_INC(task_fn_rxs);
3795 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3796 		return;
3797 	more = true;
3798 #ifdef DEV_NETMAP
3799 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3800 		u_int work = 0;
3801 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3802 			more = false;
3803 		}
3804 	}
3805 #endif
3806 	budget = ctx->ifc_sysctl_rx_budget;
3807 	if (budget == 0)
3808 		budget = 16;	/* XXX */
3809 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3810 		if (ctx->ifc_flags & IFC_LEGACY)
3811 			IFDI_INTR_ENABLE(ctx);
3812 		else
3813 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3814 		DBG_COUNTER_INC(rx_intr_enables);
3815 	}
3816 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3817 		return;
3818 	if (more)
3819 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3820 }
3821 
3822 static void
3823 _task_fn_admin(void *context)
3824 {
3825 	if_ctx_t ctx = context;
3826 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3827 	iflib_txq_t txq;
3828 	int i;
3829 	bool oactive, running, do_reset, do_watchdog, in_detach;
3830 	uint32_t reset_on = hz / 2;
3831 
3832 	STATE_LOCK(ctx);
3833 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3834 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3835 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3836 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3837 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3838 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3839 	STATE_UNLOCK(ctx);
3840 
3841 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3842 		return;
3843 	if (in_detach)
3844 		return;
3845 
3846 	CTX_LOCK(ctx);
3847 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3848 		CALLOUT_LOCK(txq);
3849 		callout_stop(&txq->ift_timer);
3850 		CALLOUT_UNLOCK(txq);
3851 	}
3852 	if (do_watchdog) {
3853 		ctx->ifc_watchdog_events++;
3854 		IFDI_WATCHDOG_RESET(ctx);
3855 	}
3856 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3857 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3858 #ifdef DEV_NETMAP
3859 		reset_on = hz / 2;
3860 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3861 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3862 #endif
3863 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3864 	}
3865 	IFDI_LINK_INTR_ENABLE(ctx);
3866 	if (do_reset)
3867 		iflib_if_init_locked(ctx);
3868 	CTX_UNLOCK(ctx);
3869 
3870 	if (LINK_ACTIVE(ctx) == 0)
3871 		return;
3872 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3873 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3874 }
3875 
3876 
3877 static void
3878 _task_fn_iov(void *context)
3879 {
3880 	if_ctx_t ctx = context;
3881 
3882 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3883 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3884 		return;
3885 
3886 	CTX_LOCK(ctx);
3887 	IFDI_VFLR_HANDLE(ctx);
3888 	CTX_UNLOCK(ctx);
3889 }
3890 
3891 static int
3892 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3893 {
3894 	int err;
3895 	if_int_delay_info_t info;
3896 	if_ctx_t ctx;
3897 
3898 	info = (if_int_delay_info_t)arg1;
3899 	ctx = info->iidi_ctx;
3900 	info->iidi_req = req;
3901 	info->iidi_oidp = oidp;
3902 	CTX_LOCK(ctx);
3903 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3904 	CTX_UNLOCK(ctx);
3905 	return (err);
3906 }
3907 
3908 /*********************************************************************
3909  *
3910  *  IFNET FUNCTIONS
3911  *
3912  **********************************************************************/
3913 
3914 static void
3915 iflib_if_init_locked(if_ctx_t ctx)
3916 {
3917 	iflib_stop(ctx);
3918 	iflib_init_locked(ctx);
3919 }
3920 
3921 
3922 static void
3923 iflib_if_init(void *arg)
3924 {
3925 	if_ctx_t ctx = arg;
3926 
3927 	CTX_LOCK(ctx);
3928 	iflib_if_init_locked(ctx);
3929 	CTX_UNLOCK(ctx);
3930 }
3931 
3932 static int
3933 iflib_if_transmit(if_t ifp, struct mbuf *m)
3934 {
3935 	if_ctx_t	ctx = if_getsoftc(ifp);
3936 
3937 	iflib_txq_t txq;
3938 	int err, qidx;
3939 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3940 
3941 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3942 		DBG_COUNTER_INC(tx_frees);
3943 		m_freem(m);
3944 		return (ENETDOWN);
3945 	}
3946 
3947 	MPASS(m->m_nextpkt == NULL);
3948 	/* ALTQ-enabled interfaces always use queue 0. */
3949 	qidx = 0;
3950 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3951 		qidx = QIDX(ctx, m);
3952 	/*
3953 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3954 	 */
3955 	txq = &ctx->ifc_txqs[qidx];
3956 
3957 #ifdef DRIVER_BACKPRESSURE
3958 	if (txq->ift_closed) {
3959 		while (m != NULL) {
3960 			next = m->m_nextpkt;
3961 			m->m_nextpkt = NULL;
3962 			m_freem(m);
3963 			DBG_COUNTER_INC(tx_frees);
3964 			m = next;
3965 		}
3966 		return (ENOBUFS);
3967 	}
3968 #endif
3969 #ifdef notyet
3970 	qidx = count = 0;
3971 	mp = marr;
3972 	next = m;
3973 	do {
3974 		count++;
3975 		next = next->m_nextpkt;
3976 	} while (next != NULL);
3977 
3978 	if (count > nitems(marr))
3979 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3980 			/* XXX check nextpkt */
3981 			m_freem(m);
3982 			/* XXX simplify for now */
3983 			DBG_COUNTER_INC(tx_frees);
3984 			return (ENOBUFS);
3985 		}
3986 	for (next = m, i = 0; next != NULL; i++) {
3987 		mp[i] = next;
3988 		next = next->m_nextpkt;
3989 		mp[i]->m_nextpkt = NULL;
3990 	}
3991 #endif
3992 	DBG_COUNTER_INC(tx_seen);
3993 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3994 
3995 	if (abdicate)
3996 		GROUPTASK_ENQUEUE(&txq->ift_task);
3997  	if (err) {
3998 		if (!abdicate)
3999 			GROUPTASK_ENQUEUE(&txq->ift_task);
4000 		/* support forthcoming later */
4001 #ifdef DRIVER_BACKPRESSURE
4002 		txq->ift_closed = TRUE;
4003 #endif
4004 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4005 		m_freem(m);
4006 		DBG_COUNTER_INC(tx_frees);
4007 	}
4008 
4009 	return (err);
4010 }
4011 
4012 #ifdef ALTQ
4013 /*
4014  * The overall approach to integrating iflib with ALTQ is to continue to use
4015  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4016  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4017  * is redundant/unnecessary, but doing so minimizes the amount of
4018  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4019  * redundantly queueing to an intermediate mp_ring is swamped by the
4020  * performance limitations inherent in using ALTQ.
4021  *
4022  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4023  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4024  * given interface.  If ALTQ is enabled for an interface, then all
4025  * transmitted packets for that interface will be submitted to the ALTQ
4026  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4027  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4028  * update stats that the iflib machinery handles, and which is sensitve to
4029  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4030  * will be installed as the start routine for use by ALTQ facilities that
4031  * need to trigger queue drains on a scheduled basis.
4032  *
4033  */
4034 static void
4035 iflib_altq_if_start(if_t ifp)
4036 {
4037 	struct ifaltq *ifq = &ifp->if_snd;
4038 	struct mbuf *m;
4039 
4040 	IFQ_LOCK(ifq);
4041 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4042 	while (m != NULL) {
4043 		iflib_if_transmit(ifp, m);
4044 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4045 	}
4046 	IFQ_UNLOCK(ifq);
4047 }
4048 
4049 static int
4050 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4051 {
4052 	int err;
4053 
4054 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4055 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4056 		if (err == 0)
4057 			iflib_altq_if_start(ifp);
4058 	} else
4059 		err = iflib_if_transmit(ifp, m);
4060 
4061 	return (err);
4062 }
4063 #endif /* ALTQ */
4064 
4065 static void
4066 iflib_if_qflush(if_t ifp)
4067 {
4068 	if_ctx_t ctx = if_getsoftc(ifp);
4069 	iflib_txq_t txq = ctx->ifc_txqs;
4070 	int i;
4071 
4072 	STATE_LOCK(ctx);
4073 	ctx->ifc_flags |= IFC_QFLUSH;
4074 	STATE_UNLOCK(ctx);
4075 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4076 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4077 			iflib_txq_check_drain(txq, 0);
4078 	STATE_LOCK(ctx);
4079 	ctx->ifc_flags &= ~IFC_QFLUSH;
4080 	STATE_UNLOCK(ctx);
4081 
4082 	/*
4083 	 * When ALTQ is enabled, this will also take care of purging the
4084 	 * ALTQ queue(s).
4085 	 */
4086 	if_qflush(ifp);
4087 }
4088 
4089 
4090 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4091 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4092 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4093 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4094 
4095 static int
4096 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4097 {
4098 	if_ctx_t ctx = if_getsoftc(ifp);
4099 	struct ifreq	*ifr = (struct ifreq *)data;
4100 #if defined(INET) || defined(INET6)
4101 	struct ifaddr	*ifa = (struct ifaddr *)data;
4102 #endif
4103 	bool		avoid_reset = false;
4104 	int		err = 0, reinit = 0, bits;
4105 
4106 	switch (command) {
4107 	case SIOCSIFADDR:
4108 #ifdef INET
4109 		if (ifa->ifa_addr->sa_family == AF_INET)
4110 			avoid_reset = true;
4111 #endif
4112 #ifdef INET6
4113 		if (ifa->ifa_addr->sa_family == AF_INET6)
4114 			avoid_reset = true;
4115 #endif
4116 		/*
4117 		** Calling init results in link renegotiation,
4118 		** so we avoid doing it when possible.
4119 		*/
4120 		if (avoid_reset) {
4121 			if_setflagbits(ifp, IFF_UP,0);
4122 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4123 				reinit = 1;
4124 #ifdef INET
4125 			if (!(if_getflags(ifp) & IFF_NOARP))
4126 				arp_ifinit(ifp, ifa);
4127 #endif
4128 		} else
4129 			err = ether_ioctl(ifp, command, data);
4130 		break;
4131 	case SIOCSIFMTU:
4132 		CTX_LOCK(ctx);
4133 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4134 			CTX_UNLOCK(ctx);
4135 			break;
4136 		}
4137 		bits = if_getdrvflags(ifp);
4138 		/* stop the driver and free any clusters before proceeding */
4139 		iflib_stop(ctx);
4140 
4141 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4142 			STATE_LOCK(ctx);
4143 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4144 				ctx->ifc_flags |= IFC_MULTISEG;
4145 			else
4146 				ctx->ifc_flags &= ~IFC_MULTISEG;
4147 			STATE_UNLOCK(ctx);
4148 			err = if_setmtu(ifp, ifr->ifr_mtu);
4149 		}
4150 		iflib_init_locked(ctx);
4151 		STATE_LOCK(ctx);
4152 		if_setdrvflags(ifp, bits);
4153 		STATE_UNLOCK(ctx);
4154 		CTX_UNLOCK(ctx);
4155 		break;
4156 	case SIOCSIFFLAGS:
4157 		CTX_LOCK(ctx);
4158 		if (if_getflags(ifp) & IFF_UP) {
4159 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4160 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4161 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4162 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4163 				}
4164 			} else
4165 				reinit = 1;
4166 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4167 			iflib_stop(ctx);
4168 		}
4169 		ctx->ifc_if_flags = if_getflags(ifp);
4170 		CTX_UNLOCK(ctx);
4171 		break;
4172 	case SIOCADDMULTI:
4173 	case SIOCDELMULTI:
4174 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4175 			CTX_LOCK(ctx);
4176 			IFDI_INTR_DISABLE(ctx);
4177 			IFDI_MULTI_SET(ctx);
4178 			IFDI_INTR_ENABLE(ctx);
4179 			CTX_UNLOCK(ctx);
4180 		}
4181 		break;
4182 	case SIOCSIFMEDIA:
4183 		CTX_LOCK(ctx);
4184 		IFDI_MEDIA_SET(ctx);
4185 		CTX_UNLOCK(ctx);
4186 		/* FALLTHROUGH */
4187 	case SIOCGIFMEDIA:
4188 	case SIOCGIFXMEDIA:
4189 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4190 		break;
4191 	case SIOCGI2C:
4192 	{
4193 		struct ifi2creq i2c;
4194 
4195 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4196 		if (err != 0)
4197 			break;
4198 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4199 			err = EINVAL;
4200 			break;
4201 		}
4202 		if (i2c.len > sizeof(i2c.data)) {
4203 			err = EINVAL;
4204 			break;
4205 		}
4206 
4207 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4208 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4209 			    sizeof(i2c));
4210 		break;
4211 	}
4212 	case SIOCSIFCAP:
4213 	{
4214 		int mask, setmask, oldmask;
4215 
4216 		oldmask = if_getcapenable(ifp);
4217 		mask = ifr->ifr_reqcap ^ oldmask;
4218 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4219 		setmask = 0;
4220 #ifdef TCP_OFFLOAD
4221 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4222 #endif
4223 		setmask |= (mask & IFCAP_FLAGS);
4224 		setmask |= (mask & IFCAP_WOL);
4225 
4226 		/*
4227 		 * If any RX csum has changed, change all the ones that
4228 		 * are supported by the driver.
4229 		 */
4230 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4231 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4232 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4233 		}
4234 
4235 		/*
4236 		 * want to ensure that traffic has stopped before we change any of the flags
4237 		 */
4238 		if (setmask) {
4239 			CTX_LOCK(ctx);
4240 			bits = if_getdrvflags(ifp);
4241 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4242 				iflib_stop(ctx);
4243 			STATE_LOCK(ctx);
4244 			if_togglecapenable(ifp, setmask);
4245 			STATE_UNLOCK(ctx);
4246 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4247 				iflib_init_locked(ctx);
4248 			STATE_LOCK(ctx);
4249 			if_setdrvflags(ifp, bits);
4250 			STATE_UNLOCK(ctx);
4251 			CTX_UNLOCK(ctx);
4252 		}
4253 		if_vlancap(ifp);
4254 		break;
4255 	}
4256 	case SIOCGPRIVATE_0:
4257 	case SIOCSDRVSPEC:
4258 	case SIOCGDRVSPEC:
4259 		CTX_LOCK(ctx);
4260 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4261 		CTX_UNLOCK(ctx);
4262 		break;
4263 	default:
4264 		err = ether_ioctl(ifp, command, data);
4265 		break;
4266 	}
4267 	if (reinit)
4268 		iflib_if_init(ctx);
4269 	return (err);
4270 }
4271 
4272 static uint64_t
4273 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4274 {
4275 	if_ctx_t ctx = if_getsoftc(ifp);
4276 
4277 	return (IFDI_GET_COUNTER(ctx, cnt));
4278 }
4279 
4280 /*********************************************************************
4281  *
4282  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4283  *
4284  **********************************************************************/
4285 
4286 static void
4287 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4288 {
4289 	if_ctx_t ctx = if_getsoftc(ifp);
4290 
4291 	if ((void *)ctx != arg)
4292 		return;
4293 
4294 	if ((vtag == 0) || (vtag > 4095))
4295 		return;
4296 
4297 	if (iflib_in_detach(ctx))
4298 		return;
4299 
4300 	CTX_LOCK(ctx);
4301 	IFDI_VLAN_REGISTER(ctx, vtag);
4302 	/* Re-init to load the changes */
4303 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4304 		iflib_if_init_locked(ctx);
4305 	CTX_UNLOCK(ctx);
4306 }
4307 
4308 static void
4309 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4310 {
4311 	if_ctx_t ctx = if_getsoftc(ifp);
4312 
4313 	if ((void *)ctx != arg)
4314 		return;
4315 
4316 	if ((vtag == 0) || (vtag > 4095))
4317 		return;
4318 
4319 	CTX_LOCK(ctx);
4320 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4321 	/* Re-init to load the changes */
4322 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4323 		iflib_if_init_locked(ctx);
4324 	CTX_UNLOCK(ctx);
4325 }
4326 
4327 static void
4328 iflib_led_func(void *arg, int onoff)
4329 {
4330 	if_ctx_t ctx = arg;
4331 
4332 	CTX_LOCK(ctx);
4333 	IFDI_LED_FUNC(ctx, onoff);
4334 	CTX_UNLOCK(ctx);
4335 }
4336 
4337 /*********************************************************************
4338  *
4339  *  BUS FUNCTION DEFINITIONS
4340  *
4341  **********************************************************************/
4342 
4343 int
4344 iflib_device_probe(device_t dev)
4345 {
4346 	const pci_vendor_info_t *ent;
4347 	if_shared_ctx_t sctx;
4348 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4349 	uint16_t pci_vendor_id;
4350 
4351 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4352 		return (ENOTSUP);
4353 
4354 	pci_vendor_id = pci_get_vendor(dev);
4355 	pci_device_id = pci_get_device(dev);
4356 	pci_subvendor_id = pci_get_subvendor(dev);
4357 	pci_subdevice_id = pci_get_subdevice(dev);
4358 	pci_rev_id = pci_get_revid(dev);
4359 	if (sctx->isc_parse_devinfo != NULL)
4360 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4361 
4362 	ent = sctx->isc_vendor_info;
4363 	while (ent->pvi_vendor_id != 0) {
4364 		if (pci_vendor_id != ent->pvi_vendor_id) {
4365 			ent++;
4366 			continue;
4367 		}
4368 		if ((pci_device_id == ent->pvi_device_id) &&
4369 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4370 		     (ent->pvi_subvendor_id == 0)) &&
4371 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4372 		     (ent->pvi_subdevice_id == 0)) &&
4373 		    ((pci_rev_id == ent->pvi_rev_id) ||
4374 		     (ent->pvi_rev_id == 0))) {
4375 
4376 			device_set_desc_copy(dev, ent->pvi_name);
4377 			/* this needs to be changed to zero if the bus probing code
4378 			 * ever stops re-probing on best match because the sctx
4379 			 * may have its values over written by register calls
4380 			 * in subsequent probes
4381 			 */
4382 			return (BUS_PROBE_DEFAULT);
4383 		}
4384 		ent++;
4385 	}
4386 	return (ENXIO);
4387 }
4388 
4389 int
4390 iflib_device_probe_vendor(device_t dev)
4391 {
4392 	int probe;
4393 
4394 	probe = iflib_device_probe(dev);
4395 	if (probe == BUS_PROBE_DEFAULT)
4396 		return (BUS_PROBE_VENDOR);
4397 	else
4398 		return (probe);
4399 }
4400 
4401 static void
4402 iflib_reset_qvalues(if_ctx_t ctx)
4403 {
4404 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4405 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4406 	device_t dev = ctx->ifc_dev;
4407 	int i;
4408 
4409 	if (ctx->ifc_sysctl_ntxqs != 0)
4410 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4411 	if (ctx->ifc_sysctl_nrxqs != 0)
4412 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4413 
4414 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4415 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4416 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4417 		else
4418 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4419 	}
4420 
4421 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4422 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4423 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4424 		else
4425 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4426 	}
4427 
4428 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4429 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4430 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4431 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4432 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4433 		}
4434 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4435 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4436 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4437 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4438 		}
4439 		if (!powerof2(scctx->isc_nrxd[i])) {
4440 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4441 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4442 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4443 		}
4444 	}
4445 
4446 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4447 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4448 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4449 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4450 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4451 		}
4452 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4453 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4454 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4455 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4456 		}
4457 		if (!powerof2(scctx->isc_ntxd[i])) {
4458 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4459 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4460 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4461 		}
4462 	}
4463 }
4464 
4465 static void
4466 iflib_add_pfil(if_ctx_t ctx)
4467 {
4468 	struct pfil_head *pfil;
4469 	struct pfil_head_args pa;
4470 	iflib_rxq_t rxq;
4471 	int i;
4472 
4473 	pa.pa_version = PFIL_VERSION;
4474 	pa.pa_flags = PFIL_IN;
4475 	pa.pa_type = PFIL_TYPE_ETHERNET;
4476 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4477 	pfil = pfil_head_register(&pa);
4478 
4479 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4480 		rxq->pfil = pfil;
4481 	}
4482 }
4483 
4484 static void
4485 iflib_rem_pfil(if_ctx_t ctx)
4486 {
4487 	struct pfil_head *pfil;
4488 	iflib_rxq_t rxq;
4489 	int i;
4490 
4491 	rxq = ctx->ifc_rxqs;
4492 	pfil = rxq->pfil;
4493 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4494 		rxq->pfil = NULL;
4495 	}
4496 	pfil_head_unregister(pfil);
4497 }
4498 
4499 static uint16_t
4500 get_ctx_core_offset(if_ctx_t ctx)
4501 {
4502 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4503 	struct cpu_offset *op;
4504 	uint16_t qc;
4505 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4506 
4507 	if (ret != CORE_OFFSET_UNSPECIFIED)
4508 		return (ret);
4509 
4510 	if (ctx->ifc_sysctl_separate_txrx)
4511 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4512 	else
4513 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4514 
4515 	mtx_lock(&cpu_offset_mtx);
4516 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4517 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4518 			ret = op->offset;
4519 			op->offset += qc;
4520 			MPASS(op->refcount < UINT_MAX);
4521 			op->refcount++;
4522 			break;
4523 		}
4524 	}
4525 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4526 		ret = 0;
4527 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4528 		    M_NOWAIT | M_ZERO);
4529 		if (op == NULL) {
4530 			device_printf(ctx->ifc_dev,
4531 			    "allocation for cpu offset failed.\n");
4532 		} else {
4533 			op->offset = qc;
4534 			op->refcount = 1;
4535 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4536 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4537 		}
4538 	}
4539 	mtx_unlock(&cpu_offset_mtx);
4540 
4541 	return (ret);
4542 }
4543 
4544 static void
4545 unref_ctx_core_offset(if_ctx_t ctx)
4546 {
4547 	struct cpu_offset *op, *top;
4548 
4549 	mtx_lock(&cpu_offset_mtx);
4550 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4551 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4552 			MPASS(op->refcount > 0);
4553 			op->refcount--;
4554 			if (op->refcount == 0) {
4555 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4556 				free(op, M_IFLIB);
4557 			}
4558 			break;
4559 		}
4560 	}
4561 	mtx_unlock(&cpu_offset_mtx);
4562 }
4563 
4564 int
4565 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4566 {
4567 	if_ctx_t ctx;
4568 	if_t ifp;
4569 	if_softc_ctx_t scctx;
4570 	kobjop_desc_t kobj_desc;
4571 	kobj_method_t *kobj_method;
4572 	int err, msix, rid;
4573 	uint16_t main_rxq, main_txq;
4574 
4575 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4576 
4577 	if (sc == NULL) {
4578 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4579 		device_set_softc(dev, ctx);
4580 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4581 	}
4582 
4583 	ctx->ifc_sctx = sctx;
4584 	ctx->ifc_dev = dev;
4585 	ctx->ifc_softc = sc;
4586 
4587 	if ((err = iflib_register(ctx)) != 0) {
4588 		device_printf(dev, "iflib_register failed %d\n", err);
4589 		goto fail_ctx_free;
4590 	}
4591 	iflib_add_device_sysctl_pre(ctx);
4592 
4593 	scctx = &ctx->ifc_softc_ctx;
4594 	ifp = ctx->ifc_ifp;
4595 
4596 	iflib_reset_qvalues(ctx);
4597 	CTX_LOCK(ctx);
4598 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4599 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4600 		goto fail_unlock;
4601 	}
4602 	_iflib_pre_assert(scctx);
4603 	ctx->ifc_txrx = *scctx->isc_txrx;
4604 
4605 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4606 		ctx->ifc_mediap = scctx->isc_media;
4607 
4608 #ifdef INVARIANTS
4609 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4610 		MPASS(scctx->isc_tx_csum_flags);
4611 #endif
4612 
4613 	if_setcapabilities(ifp,
4614 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4615 	if_setcapenable(ifp,
4616 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4617 
4618 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4619 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4620 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4621 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4622 
4623 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4624 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4625 
4626 	/* XXX change for per-queue sizes */
4627 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4628 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4629 
4630 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4631 	    MAX_SINGLE_PACKET_FRACTION)
4632 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4633 		    MAX_SINGLE_PACKET_FRACTION);
4634 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4635 	    MAX_SINGLE_PACKET_FRACTION)
4636 		scctx->isc_tx_tso_segments_max = max(1,
4637 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4638 
4639 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4640 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4641 		/*
4642 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4643 		 * but some MACs do.
4644 		 */
4645 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4646 		    IP_MAXPACKET));
4647 		/*
4648 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4649 		 * into account.  In the worst case, each of these calls will
4650 		 * add another mbuf and, thus, the requirement for another DMA
4651 		 * segment.  So for best performance, it doesn't make sense to
4652 		 * advertize a maximum of TSO segments that typically will
4653 		 * require defragmentation in iflib_encap().
4654 		 */
4655 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4656 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4657 	}
4658 	if (scctx->isc_rss_table_size == 0)
4659 		scctx->isc_rss_table_size = 64;
4660 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4661 
4662 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4663 	/* XXX format name */
4664 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4665 	    NULL, NULL, "admin");
4666 
4667 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4668 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4669 		device_printf(dev, "Unable to fetch CPU list\n");
4670 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4671 	}
4672 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4673 
4674 	/*
4675 	** Now set up MSI or MSI-X, should return us the number of supported
4676 	** vectors (will be 1 for a legacy interrupt and MSI).
4677 	*/
4678 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4679 		msix = scctx->isc_vectors;
4680 	} else if (scctx->isc_msix_bar != 0)
4681 	       /*
4682 		* The simple fact that isc_msix_bar is not 0 does not mean we
4683 		* we have a good value there that is known to work.
4684 		*/
4685 		msix = iflib_msix_init(ctx);
4686 	else {
4687 		scctx->isc_vectors = 1;
4688 		scctx->isc_ntxqsets = 1;
4689 		scctx->isc_nrxqsets = 1;
4690 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4691 		msix = 0;
4692 	}
4693 	/* Get memory for the station queues */
4694 	if ((err = iflib_queues_alloc(ctx))) {
4695 		device_printf(dev, "Unable to allocate queue memory\n");
4696 		goto fail_intr_free;
4697 	}
4698 
4699 	if ((err = iflib_qset_structures_setup(ctx)))
4700 		goto fail_queues;
4701 
4702 	/*
4703 	 * Now that we know how many queues there are, get the core offset.
4704 	 */
4705 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4706 
4707 	/*
4708 	 * Group taskqueues aren't properly set up until SMP is started,
4709 	 * so we disable interrupts until we can handle them post
4710 	 * SI_SUB_SMP.
4711 	 *
4712 	 * XXX: disabling interrupts doesn't actually work, at least for
4713 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4714 	 * we do null handling and depend on this not causing too large an
4715 	 * interrupt storm.
4716 	 */
4717 	IFDI_INTR_DISABLE(ctx);
4718 
4719 	if (msix > 1) {
4720 		/*
4721 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4722 		 * aren't the default NULL implementation.
4723 		 */
4724 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4725 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4726 		    kobj_desc);
4727 		if (kobj_method == &kobj_desc->deflt) {
4728 			device_printf(dev,
4729 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4730 			err = EOPNOTSUPP;
4731 			goto fail_queues;
4732 		}
4733 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4734 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4735 		    kobj_desc);
4736 		if (kobj_method == &kobj_desc->deflt) {
4737 			device_printf(dev,
4738 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4739 			err = EOPNOTSUPP;
4740 			goto fail_queues;
4741 		}
4742 
4743 		/*
4744 		 * Assign the MSI-X vectors.
4745 		 * Note that the default NULL ifdi_msix_intr_assign method will
4746 		 * fail here, too.
4747 		 */
4748 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4749 		if (err != 0) {
4750 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4751 			    err);
4752 			goto fail_queues;
4753 		}
4754 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4755 		rid = 0;
4756 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4757 			MPASS(msix == 1);
4758 			rid = 1;
4759 		}
4760 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4761 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4762 			goto fail_queues;
4763 		}
4764 	} else {
4765 		device_printf(dev,
4766 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4767 		err = ENODEV;
4768 		goto fail_intr_free;
4769 	}
4770 
4771 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4772 
4773 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4774 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4775 		goto fail_detach;
4776 	}
4777 
4778 	/*
4779 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4780 	 * This must appear after the call to ether_ifattach() because
4781 	 * ether_ifattach() sets if_hdrlen to the default value.
4782 	 */
4783 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4784 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4785 
4786 	if ((err = iflib_netmap_attach(ctx))) {
4787 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4788 		goto fail_detach;
4789 	}
4790 	*ctxp = ctx;
4791 
4792 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4793 
4794 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4795 	iflib_add_device_sysctl_post(ctx);
4796 	iflib_add_pfil(ctx);
4797 	ctx->ifc_flags |= IFC_INIT_DONE;
4798 	CTX_UNLOCK(ctx);
4799 
4800 	return (0);
4801 
4802 fail_detach:
4803 	ether_ifdetach(ctx->ifc_ifp);
4804 fail_intr_free:
4805 	iflib_free_intr_mem(ctx);
4806 fail_queues:
4807 	iflib_tx_structures_free(ctx);
4808 	iflib_rx_structures_free(ctx);
4809 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4810 	IFDI_DETACH(ctx);
4811 fail_unlock:
4812 	CTX_UNLOCK(ctx);
4813 	iflib_deregister(ctx);
4814 fail_ctx_free:
4815 	device_set_softc(ctx->ifc_dev, NULL);
4816         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4817                 free(ctx->ifc_softc, M_IFLIB);
4818         free(ctx, M_IFLIB);
4819 	return (err);
4820 }
4821 
4822 int
4823 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4824 					  struct iflib_cloneattach_ctx *clctx)
4825 {
4826 	int err;
4827 	if_ctx_t ctx;
4828 	if_t ifp;
4829 	if_softc_ctx_t scctx;
4830 	int i;
4831 	void *sc;
4832 	uint16_t main_txq;
4833 	uint16_t main_rxq;
4834 
4835 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4836 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4837 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4838 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4839 		ctx->ifc_flags |= IFC_PSEUDO;
4840 
4841 	ctx->ifc_sctx = sctx;
4842 	ctx->ifc_softc = sc;
4843 	ctx->ifc_dev = dev;
4844 
4845 	if ((err = iflib_register(ctx)) != 0) {
4846 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4847 		goto fail_ctx_free;
4848 	}
4849 	iflib_add_device_sysctl_pre(ctx);
4850 
4851 	scctx = &ctx->ifc_softc_ctx;
4852 	ifp = ctx->ifc_ifp;
4853 
4854 	iflib_reset_qvalues(ctx);
4855 	CTX_LOCK(ctx);
4856 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4857 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4858 		goto fail_unlock;
4859 	}
4860 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4861 		ether_gen_addr(ifp, &ctx->ifc_mac);
4862 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4863 								clctx->cc_params)) != 0) {
4864 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4865 		goto fail_ctx_free;
4866 	}
4867 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4868 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4869 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4870 
4871 #ifdef INVARIANTS
4872 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4873 		MPASS(scctx->isc_tx_csum_flags);
4874 #endif
4875 
4876 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4877 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4878 
4879 	ifp->if_flags |= IFF_NOGROUP;
4880 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4881 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4882 
4883 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4884 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4885 			goto fail_detach;
4886 		}
4887 		*ctxp = ctx;
4888 
4889 		/*
4890 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4891 		 * This must appear after the call to ether_ifattach() because
4892 		 * ether_ifattach() sets if_hdrlen to the default value.
4893 		 */
4894 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4895 			if_setifheaderlen(ifp,
4896 			    sizeof(struct ether_vlan_header));
4897 
4898 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4899 		iflib_add_device_sysctl_post(ctx);
4900 		ctx->ifc_flags |= IFC_INIT_DONE;
4901 		return (0);
4902 	}
4903 	_iflib_pre_assert(scctx);
4904 	ctx->ifc_txrx = *scctx->isc_txrx;
4905 
4906 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4907 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4908 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4909 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4910 
4911 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4912 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4913 
4914 	/* XXX change for per-queue sizes */
4915 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4916 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4917 
4918 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4919 	    MAX_SINGLE_PACKET_FRACTION)
4920 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4921 		    MAX_SINGLE_PACKET_FRACTION);
4922 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4923 	    MAX_SINGLE_PACKET_FRACTION)
4924 		scctx->isc_tx_tso_segments_max = max(1,
4925 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4926 
4927 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4928 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4929 		/*
4930 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4931 		 * but some MACs do.
4932 		 */
4933 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4934 		    IP_MAXPACKET));
4935 		/*
4936 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4937 		 * into account.  In the worst case, each of these calls will
4938 		 * add another mbuf and, thus, the requirement for another DMA
4939 		 * segment.  So for best performance, it doesn't make sense to
4940 		 * advertize a maximum of TSO segments that typically will
4941 		 * require defragmentation in iflib_encap().
4942 		 */
4943 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4944 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4945 	}
4946 	if (scctx->isc_rss_table_size == 0)
4947 		scctx->isc_rss_table_size = 64;
4948 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4949 
4950 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4951 	/* XXX format name */
4952 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4953 	    NULL, NULL, "admin");
4954 
4955 	/* XXX --- can support > 1 -- but keep it simple for now */
4956 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4957 
4958 	/* Get memory for the station queues */
4959 	if ((err = iflib_queues_alloc(ctx))) {
4960 		device_printf(dev, "Unable to allocate queue memory\n");
4961 		goto fail_iflib_detach;
4962 	}
4963 
4964 	if ((err = iflib_qset_structures_setup(ctx))) {
4965 		device_printf(dev, "qset structure setup failed %d\n", err);
4966 		goto fail_queues;
4967 	}
4968 
4969 	/*
4970 	 * XXX What if anything do we want to do about interrupts?
4971 	 */
4972 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4973 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4974 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4975 		goto fail_detach;
4976 	}
4977 
4978 	/*
4979 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4980 	 * This must appear after the call to ether_ifattach() because
4981 	 * ether_ifattach() sets if_hdrlen to the default value.
4982 	 */
4983 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4984 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4985 
4986 	/* XXX handle more than one queue */
4987 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4988 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4989 
4990 	*ctxp = ctx;
4991 
4992 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4993 	iflib_add_device_sysctl_post(ctx);
4994 	ctx->ifc_flags |= IFC_INIT_DONE;
4995 	CTX_UNLOCK(ctx);
4996 
4997 	return (0);
4998 fail_detach:
4999 	ether_ifdetach(ctx->ifc_ifp);
5000 fail_queues:
5001 	iflib_tx_structures_free(ctx);
5002 	iflib_rx_structures_free(ctx);
5003 fail_iflib_detach:
5004 	IFDI_DETACH(ctx);
5005 fail_unlock:
5006 	CTX_UNLOCK(ctx);
5007 	iflib_deregister(ctx);
5008 fail_ctx_free:
5009 	free(ctx->ifc_softc, M_IFLIB);
5010 	free(ctx, M_IFLIB);
5011 	return (err);
5012 }
5013 
5014 int
5015 iflib_pseudo_deregister(if_ctx_t ctx)
5016 {
5017 	if_t ifp = ctx->ifc_ifp;
5018 	iflib_txq_t txq;
5019 	iflib_rxq_t rxq;
5020 	int i, j;
5021 	struct taskqgroup *tqg;
5022 	iflib_fl_t fl;
5023 
5024 	/* Unregister VLAN event handlers early */
5025 	iflib_unregister_vlan_handlers(ctx);
5026 
5027 	ether_ifdetach(ifp);
5028 	/* XXX drain any dependent tasks */
5029 	tqg = qgroup_if_io_tqg;
5030 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5031 		callout_drain(&txq->ift_timer);
5032 		if (txq->ift_task.gt_uniq != NULL)
5033 			taskqgroup_detach(tqg, &txq->ift_task);
5034 	}
5035 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5036 		if (rxq->ifr_task.gt_uniq != NULL)
5037 			taskqgroup_detach(tqg, &rxq->ifr_task);
5038 
5039 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5040 			free(fl->ifl_rx_bitmap, M_IFLIB);
5041 	}
5042 	tqg = qgroup_if_config_tqg;
5043 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5044 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5045 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5046 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5047 
5048 	iflib_tx_structures_free(ctx);
5049 	iflib_rx_structures_free(ctx);
5050 
5051 	iflib_deregister(ctx);
5052 
5053 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5054 		free(ctx->ifc_softc, M_IFLIB);
5055 	free(ctx, M_IFLIB);
5056 	return (0);
5057 }
5058 
5059 int
5060 iflib_device_attach(device_t dev)
5061 {
5062 	if_ctx_t ctx;
5063 	if_shared_ctx_t sctx;
5064 
5065 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5066 		return (ENOTSUP);
5067 
5068 	pci_enable_busmaster(dev);
5069 
5070 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5071 }
5072 
5073 int
5074 iflib_device_deregister(if_ctx_t ctx)
5075 {
5076 	if_t ifp = ctx->ifc_ifp;
5077 	iflib_txq_t txq;
5078 	iflib_rxq_t rxq;
5079 	device_t dev = ctx->ifc_dev;
5080 	int i, j;
5081 	struct taskqgroup *tqg;
5082 	iflib_fl_t fl;
5083 
5084 	/* Make sure VLANS are not using driver */
5085 	if (if_vlantrunkinuse(ifp)) {
5086 		device_printf(dev, "Vlan in use, detach first\n");
5087 		return (EBUSY);
5088 	}
5089 #ifdef PCI_IOV
5090 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5091 		device_printf(dev, "SR-IOV in use; detach first.\n");
5092 		return (EBUSY);
5093 	}
5094 #endif
5095 
5096 	STATE_LOCK(ctx);
5097 	ctx->ifc_flags |= IFC_IN_DETACH;
5098 	STATE_UNLOCK(ctx);
5099 
5100 	/* Unregister VLAN handlers before calling iflib_stop() */
5101 	iflib_unregister_vlan_handlers(ctx);
5102 
5103 	iflib_netmap_detach(ifp);
5104 	ether_ifdetach(ifp);
5105 
5106 	CTX_LOCK(ctx);
5107 	iflib_stop(ctx);
5108 	CTX_UNLOCK(ctx);
5109 
5110 	iflib_rem_pfil(ctx);
5111 	if (ctx->ifc_led_dev != NULL)
5112 		led_destroy(ctx->ifc_led_dev);
5113 	/* XXX drain any dependent tasks */
5114 	tqg = qgroup_if_io_tqg;
5115 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5116 		callout_drain(&txq->ift_timer);
5117 		if (txq->ift_task.gt_uniq != NULL)
5118 			taskqgroup_detach(tqg, &txq->ift_task);
5119 	}
5120 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5121 		if (rxq->ifr_task.gt_uniq != NULL)
5122 			taskqgroup_detach(tqg, &rxq->ifr_task);
5123 
5124 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5125 			free(fl->ifl_rx_bitmap, M_IFLIB);
5126 	}
5127 	tqg = qgroup_if_config_tqg;
5128 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5129 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5130 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5131 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5132 	CTX_LOCK(ctx);
5133 	IFDI_DETACH(ctx);
5134 	CTX_UNLOCK(ctx);
5135 
5136 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5137 	iflib_free_intr_mem(ctx);
5138 
5139 	bus_generic_detach(dev);
5140 
5141 	iflib_tx_structures_free(ctx);
5142 	iflib_rx_structures_free(ctx);
5143 
5144 	iflib_deregister(ctx);
5145 
5146 	device_set_softc(ctx->ifc_dev, NULL);
5147 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5148 		free(ctx->ifc_softc, M_IFLIB);
5149 	unref_ctx_core_offset(ctx);
5150 	free(ctx, M_IFLIB);
5151 	return (0);
5152 }
5153 
5154 static void
5155 iflib_free_intr_mem(if_ctx_t ctx)
5156 {
5157 
5158 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5159 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5160 	}
5161 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5162 		pci_release_msi(ctx->ifc_dev);
5163 	}
5164 	if (ctx->ifc_msix_mem != NULL) {
5165 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5166 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5167 		ctx->ifc_msix_mem = NULL;
5168 	}
5169 }
5170 
5171 int
5172 iflib_device_detach(device_t dev)
5173 {
5174 	if_ctx_t ctx = device_get_softc(dev);
5175 
5176 	return (iflib_device_deregister(ctx));
5177 }
5178 
5179 int
5180 iflib_device_suspend(device_t dev)
5181 {
5182 	if_ctx_t ctx = device_get_softc(dev);
5183 
5184 	CTX_LOCK(ctx);
5185 	IFDI_SUSPEND(ctx);
5186 	CTX_UNLOCK(ctx);
5187 
5188 	return bus_generic_suspend(dev);
5189 }
5190 int
5191 iflib_device_shutdown(device_t dev)
5192 {
5193 	if_ctx_t ctx = device_get_softc(dev);
5194 
5195 	CTX_LOCK(ctx);
5196 	IFDI_SHUTDOWN(ctx);
5197 	CTX_UNLOCK(ctx);
5198 
5199 	return bus_generic_suspend(dev);
5200 }
5201 
5202 
5203 int
5204 iflib_device_resume(device_t dev)
5205 {
5206 	if_ctx_t ctx = device_get_softc(dev);
5207 	iflib_txq_t txq = ctx->ifc_txqs;
5208 
5209 	CTX_LOCK(ctx);
5210 	IFDI_RESUME(ctx);
5211 	iflib_if_init_locked(ctx);
5212 	CTX_UNLOCK(ctx);
5213 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5214 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5215 
5216 	return (bus_generic_resume(dev));
5217 }
5218 
5219 int
5220 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5221 {
5222 	int error;
5223 	if_ctx_t ctx = device_get_softc(dev);
5224 
5225 	CTX_LOCK(ctx);
5226 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5227 	CTX_UNLOCK(ctx);
5228 
5229 	return (error);
5230 }
5231 
5232 void
5233 iflib_device_iov_uninit(device_t dev)
5234 {
5235 	if_ctx_t ctx = device_get_softc(dev);
5236 
5237 	CTX_LOCK(ctx);
5238 	IFDI_IOV_UNINIT(ctx);
5239 	CTX_UNLOCK(ctx);
5240 }
5241 
5242 int
5243 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5244 {
5245 	int error;
5246 	if_ctx_t ctx = device_get_softc(dev);
5247 
5248 	CTX_LOCK(ctx);
5249 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5250 	CTX_UNLOCK(ctx);
5251 
5252 	return (error);
5253 }
5254 
5255 /*********************************************************************
5256  *
5257  *  MODULE FUNCTION DEFINITIONS
5258  *
5259  **********************************************************************/
5260 
5261 /*
5262  * - Start a fast taskqueue thread for each core
5263  * - Start a taskqueue for control operations
5264  */
5265 static int
5266 iflib_module_init(void)
5267 {
5268 	return (0);
5269 }
5270 
5271 static int
5272 iflib_module_event_handler(module_t mod, int what, void *arg)
5273 {
5274 	int err;
5275 
5276 	switch (what) {
5277 	case MOD_LOAD:
5278 		if ((err = iflib_module_init()) != 0)
5279 			return (err);
5280 		break;
5281 	case MOD_UNLOAD:
5282 		return (EBUSY);
5283 	default:
5284 		return (EOPNOTSUPP);
5285 	}
5286 
5287 	return (0);
5288 }
5289 
5290 /*********************************************************************
5291  *
5292  *  PUBLIC FUNCTION DEFINITIONS
5293  *     ordered as in iflib.h
5294  *
5295  **********************************************************************/
5296 
5297 
5298 static void
5299 _iflib_assert(if_shared_ctx_t sctx)
5300 {
5301 	int i;
5302 
5303 	MPASS(sctx->isc_tx_maxsize);
5304 	MPASS(sctx->isc_tx_maxsegsize);
5305 
5306 	MPASS(sctx->isc_rx_maxsize);
5307 	MPASS(sctx->isc_rx_nsegments);
5308 	MPASS(sctx->isc_rx_maxsegsize);
5309 
5310 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5311 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5312 		MPASS(sctx->isc_nrxd_min[i]);
5313 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5314 		MPASS(sctx->isc_nrxd_max[i]);
5315 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5316 		MPASS(sctx->isc_nrxd_default[i]);
5317 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5318 	}
5319 
5320 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5321 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5322 		MPASS(sctx->isc_ntxd_min[i]);
5323 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5324 		MPASS(sctx->isc_ntxd_max[i]);
5325 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5326 		MPASS(sctx->isc_ntxd_default[i]);
5327 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5328 	}
5329 }
5330 
5331 static void
5332 _iflib_pre_assert(if_softc_ctx_t scctx)
5333 {
5334 
5335 	MPASS(scctx->isc_txrx->ift_txd_encap);
5336 	MPASS(scctx->isc_txrx->ift_txd_flush);
5337 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5338 	MPASS(scctx->isc_txrx->ift_rxd_available);
5339 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5340 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5341 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5342 }
5343 
5344 static int
5345 iflib_register(if_ctx_t ctx)
5346 {
5347 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5348 	driver_t *driver = sctx->isc_driver;
5349 	device_t dev = ctx->ifc_dev;
5350 	if_t ifp;
5351 
5352 	_iflib_assert(sctx);
5353 
5354 	CTX_LOCK_INIT(ctx);
5355 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5356 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5357 	if (ifp == NULL) {
5358 		device_printf(dev, "can not allocate ifnet structure\n");
5359 		return (ENOMEM);
5360 	}
5361 
5362 	/*
5363 	 * Initialize our context's device specific methods
5364 	 */
5365 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5366 	kobj_class_compile((kobj_class_t) driver);
5367 
5368 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5369 	if_setsoftc(ifp, ctx);
5370 	if_setdev(ifp, dev);
5371 	if_setinitfn(ifp, iflib_if_init);
5372 	if_setioctlfn(ifp, iflib_if_ioctl);
5373 #ifdef ALTQ
5374 	if_setstartfn(ifp, iflib_altq_if_start);
5375 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5376 	if_setsendqready(ifp);
5377 #else
5378 	if_settransmitfn(ifp, iflib_if_transmit);
5379 #endif
5380 	if_setqflushfn(ifp, iflib_if_qflush);
5381 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5382 
5383 	ctx->ifc_vlan_attach_event =
5384 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5385 							  EVENTHANDLER_PRI_FIRST);
5386 	ctx->ifc_vlan_detach_event =
5387 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5388 							  EVENTHANDLER_PRI_FIRST);
5389 
5390 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5391 		ctx->ifc_mediap = &ctx->ifc_media;
5392 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5393 		    iflib_media_change, iflib_media_status);
5394 	}
5395 	return (0);
5396 }
5397 
5398 static void
5399 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5400 {
5401 	/* Unregister VLAN events */
5402 	if (ctx->ifc_vlan_attach_event != NULL) {
5403 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5404 		ctx->ifc_vlan_attach_event = NULL;
5405 	}
5406 	if (ctx->ifc_vlan_detach_event != NULL) {
5407 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5408 		ctx->ifc_vlan_detach_event = NULL;
5409 	}
5410 
5411 }
5412 
5413 static void
5414 iflib_deregister(if_ctx_t ctx)
5415 {
5416 	if_t ifp = ctx->ifc_ifp;
5417 
5418 	/* Remove all media */
5419 	ifmedia_removeall(&ctx->ifc_media);
5420 
5421 	/* Ensure that VLAN event handlers are unregistered */
5422 	iflib_unregister_vlan_handlers(ctx);
5423 
5424 	/* Release kobject reference */
5425 	kobj_delete((kobj_t) ctx, NULL);
5426 
5427 	/* Free the ifnet structure */
5428 	if_free(ifp);
5429 
5430 	STATE_LOCK_DESTROY(ctx);
5431 
5432 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5433 	CTX_LOCK_DESTROY(ctx);
5434 }
5435 
5436 static int
5437 iflib_queues_alloc(if_ctx_t ctx)
5438 {
5439 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5440 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5441 	device_t dev = ctx->ifc_dev;
5442 	int nrxqsets = scctx->isc_nrxqsets;
5443 	int ntxqsets = scctx->isc_ntxqsets;
5444 	iflib_txq_t txq;
5445 	iflib_rxq_t rxq;
5446 	iflib_fl_t fl = NULL;
5447 	int i, j, cpu, err, txconf, rxconf;
5448 	iflib_dma_info_t ifdip;
5449 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5450 	uint32_t *txqsizes = scctx->isc_txqsizes;
5451 	uint8_t nrxqs = sctx->isc_nrxqs;
5452 	uint8_t ntxqs = sctx->isc_ntxqs;
5453 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5454 	caddr_t *vaddrs;
5455 	uint64_t *paddrs;
5456 
5457 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5458 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5459 
5460 	/* Allocate the TX ring struct memory */
5461 	if (!(ctx->ifc_txqs =
5462 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5463 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5464 		device_printf(dev, "Unable to allocate TX ring memory\n");
5465 		err = ENOMEM;
5466 		goto fail;
5467 	}
5468 
5469 	/* Now allocate the RX */
5470 	if (!(ctx->ifc_rxqs =
5471 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5472 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5473 		device_printf(dev, "Unable to allocate RX ring memory\n");
5474 		err = ENOMEM;
5475 		goto rx_fail;
5476 	}
5477 
5478 	txq = ctx->ifc_txqs;
5479 	rxq = ctx->ifc_rxqs;
5480 
5481 	/*
5482 	 * XXX handle allocation failure
5483 	 */
5484 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5485 		/* Set up some basics */
5486 
5487 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5488 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5489 			device_printf(dev,
5490 			    "Unable to allocate TX DMA info memory\n");
5491 			err = ENOMEM;
5492 			goto err_tx_desc;
5493 		}
5494 		txq->ift_ifdi = ifdip;
5495 		for (j = 0; j < ntxqs; j++, ifdip++) {
5496 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5497 				device_printf(dev,
5498 				    "Unable to allocate TX descriptors\n");
5499 				err = ENOMEM;
5500 				goto err_tx_desc;
5501 			}
5502 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5503 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5504 		}
5505 		txq->ift_ctx = ctx;
5506 		txq->ift_id = i;
5507 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5508 			txq->ift_br_offset = 1;
5509 		} else {
5510 			txq->ift_br_offset = 0;
5511 		}
5512 		/* XXX fix this */
5513 		txq->ift_timer.c_cpu = cpu;
5514 
5515 		if (iflib_txsd_alloc(txq)) {
5516 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5517 			err = ENOMEM;
5518 			goto err_tx_desc;
5519 		}
5520 
5521 		/* Initialize the TX lock */
5522 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5523 		    device_get_nameunit(dev), txq->ift_id);
5524 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5525 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5526 
5527 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5528 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5529 		if (err) {
5530 			/* XXX free any allocated rings */
5531 			device_printf(dev, "Unable to allocate buf_ring\n");
5532 			goto err_tx_desc;
5533 		}
5534 	}
5535 
5536 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5537 		/* Set up some basics */
5538 
5539 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5540 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5541 			device_printf(dev,
5542 			    "Unable to allocate RX DMA info memory\n");
5543 			err = ENOMEM;
5544 			goto err_tx_desc;
5545 		}
5546 
5547 		rxq->ifr_ifdi = ifdip;
5548 		/* XXX this needs to be changed if #rx queues != #tx queues */
5549 		rxq->ifr_ntxqirq = 1;
5550 		rxq->ifr_txqid[0] = i;
5551 		for (j = 0; j < nrxqs; j++, ifdip++) {
5552 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5553 				device_printf(dev,
5554 				    "Unable to allocate RX descriptors\n");
5555 				err = ENOMEM;
5556 				goto err_tx_desc;
5557 			}
5558 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5559 		}
5560 		rxq->ifr_ctx = ctx;
5561 		rxq->ifr_id = i;
5562 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5563 			rxq->ifr_fl_offset = 1;
5564 		} else {
5565 			rxq->ifr_fl_offset = 0;
5566 		}
5567 		rxq->ifr_nfl = nfree_lists;
5568 		if (!(fl =
5569 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5570 			device_printf(dev, "Unable to allocate free list memory\n");
5571 			err = ENOMEM;
5572 			goto err_tx_desc;
5573 		}
5574 		rxq->ifr_fl = fl;
5575 		for (j = 0; j < nfree_lists; j++) {
5576 			fl[j].ifl_rxq = rxq;
5577 			fl[j].ifl_id = j;
5578 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5579 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5580 		}
5581 		/* Allocate receive buffers for the ring */
5582 		if (iflib_rxsd_alloc(rxq)) {
5583 			device_printf(dev,
5584 			    "Critical Failure setting up receive buffers\n");
5585 			err = ENOMEM;
5586 			goto err_rx_desc;
5587 		}
5588 
5589 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5590 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5591 			    M_WAITOK);
5592 	}
5593 
5594 	/* TXQs */
5595 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5596 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5597 	for (i = 0; i < ntxqsets; i++) {
5598 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5599 
5600 		for (j = 0; j < ntxqs; j++, di++) {
5601 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5602 			paddrs[i*ntxqs + j] = di->idi_paddr;
5603 		}
5604 	}
5605 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5606 		device_printf(ctx->ifc_dev,
5607 		    "Unable to allocate device TX queue\n");
5608 		iflib_tx_structures_free(ctx);
5609 		free(vaddrs, M_IFLIB);
5610 		free(paddrs, M_IFLIB);
5611 		goto err_rx_desc;
5612 	}
5613 	free(vaddrs, M_IFLIB);
5614 	free(paddrs, M_IFLIB);
5615 
5616 	/* RXQs */
5617 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5618 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5619 	for (i = 0; i < nrxqsets; i++) {
5620 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5621 
5622 		for (j = 0; j < nrxqs; j++, di++) {
5623 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5624 			paddrs[i*nrxqs + j] = di->idi_paddr;
5625 		}
5626 	}
5627 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5628 		device_printf(ctx->ifc_dev,
5629 		    "Unable to allocate device RX queue\n");
5630 		iflib_tx_structures_free(ctx);
5631 		free(vaddrs, M_IFLIB);
5632 		free(paddrs, M_IFLIB);
5633 		goto err_rx_desc;
5634 	}
5635 	free(vaddrs, M_IFLIB);
5636 	free(paddrs, M_IFLIB);
5637 
5638 	return (0);
5639 
5640 /* XXX handle allocation failure changes */
5641 err_rx_desc:
5642 err_tx_desc:
5643 rx_fail:
5644 	if (ctx->ifc_rxqs != NULL)
5645 		free(ctx->ifc_rxqs, M_IFLIB);
5646 	ctx->ifc_rxqs = NULL;
5647 	if (ctx->ifc_txqs != NULL)
5648 		free(ctx->ifc_txqs, M_IFLIB);
5649 	ctx->ifc_txqs = NULL;
5650 fail:
5651 	return (err);
5652 }
5653 
5654 static int
5655 iflib_tx_structures_setup(if_ctx_t ctx)
5656 {
5657 	iflib_txq_t txq = ctx->ifc_txqs;
5658 	int i;
5659 
5660 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5661 		iflib_txq_setup(txq);
5662 
5663 	return (0);
5664 }
5665 
5666 static void
5667 iflib_tx_structures_free(if_ctx_t ctx)
5668 {
5669 	iflib_txq_t txq = ctx->ifc_txqs;
5670 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5671 	int i, j;
5672 
5673 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5674 		for (j = 0; j < sctx->isc_ntxqs; j++)
5675 			iflib_dma_free(&txq->ift_ifdi[j]);
5676 		iflib_txq_destroy(txq);
5677 	}
5678 	free(ctx->ifc_txqs, M_IFLIB);
5679 	ctx->ifc_txqs = NULL;
5680 	IFDI_QUEUES_FREE(ctx);
5681 }
5682 
5683 /*********************************************************************
5684  *
5685  *  Initialize all receive rings.
5686  *
5687  **********************************************************************/
5688 static int
5689 iflib_rx_structures_setup(if_ctx_t ctx)
5690 {
5691 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5692 	int q;
5693 #if defined(INET6) || defined(INET)
5694 	int err, i;
5695 #endif
5696 
5697 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5698 #if defined(INET6) || defined(INET)
5699 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5700 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5701 			    TCP_LRO_ENTRIES, min(1024,
5702 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5703 			if (err != 0) {
5704 				device_printf(ctx->ifc_dev,
5705 				    "LRO Initialization failed!\n");
5706 				goto fail;
5707 			}
5708 		}
5709 #endif
5710 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5711 	}
5712 	return (0);
5713 #if defined(INET6) || defined(INET)
5714 fail:
5715 	/*
5716 	 * Free LRO resources allocated so far, we will only handle
5717 	 * the rings that completed, the failing case will have
5718 	 * cleaned up for itself.  'q' failed, so its the terminus.
5719 	 */
5720 	rxq = ctx->ifc_rxqs;
5721 	for (i = 0; i < q; ++i, rxq++) {
5722 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5723 			tcp_lro_free(&rxq->ifr_lc);
5724 	}
5725 	return (err);
5726 #endif
5727 }
5728 
5729 /*********************************************************************
5730  *
5731  *  Free all receive rings.
5732  *
5733  **********************************************************************/
5734 static void
5735 iflib_rx_structures_free(if_ctx_t ctx)
5736 {
5737 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5738 	int i;
5739 
5740 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5741 		iflib_rx_sds_free(rxq);
5742 #if defined(INET6) || defined(INET)
5743 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5744 			tcp_lro_free(&rxq->ifr_lc);
5745 #endif
5746 	}
5747 	free(ctx->ifc_rxqs, M_IFLIB);
5748 	ctx->ifc_rxqs = NULL;
5749 }
5750 
5751 static int
5752 iflib_qset_structures_setup(if_ctx_t ctx)
5753 {
5754 	int err;
5755 
5756 	/*
5757 	 * It is expected that the caller takes care of freeing queues if this
5758 	 * fails.
5759 	 */
5760 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5761 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5762 		return (err);
5763 	}
5764 
5765 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5766 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5767 
5768 	return (err);
5769 }
5770 
5771 int
5772 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5773 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5774 {
5775 
5776 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5777 }
5778 
5779 #ifdef SMP
5780 static int
5781 find_nth(if_ctx_t ctx, int qid)
5782 {
5783 	cpuset_t cpus;
5784 	int i, cpuid, eqid, count;
5785 
5786 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5787 	count = CPU_COUNT(&cpus);
5788 	eqid = qid % count;
5789 	/* clear up to the qid'th bit */
5790 	for (i = 0; i < eqid; i++) {
5791 		cpuid = CPU_FFS(&cpus);
5792 		MPASS(cpuid != 0);
5793 		CPU_CLR(cpuid-1, &cpus);
5794 	}
5795 	cpuid = CPU_FFS(&cpus);
5796 	MPASS(cpuid != 0);
5797 	return (cpuid-1);
5798 }
5799 
5800 #ifdef SCHED_ULE
5801 extern struct cpu_group *cpu_top;              /* CPU topology */
5802 
5803 static int
5804 find_child_with_core(int cpu, struct cpu_group *grp)
5805 {
5806 	int i;
5807 
5808 	if (grp->cg_children == 0)
5809 		return -1;
5810 
5811 	MPASS(grp->cg_child);
5812 	for (i = 0; i < grp->cg_children; i++) {
5813 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5814 			return i;
5815 	}
5816 
5817 	return -1;
5818 }
5819 
5820 /*
5821  * Find the nth "close" core to the specified core
5822  * "close" is defined as the deepest level that shares
5823  * at least an L2 cache.  With threads, this will be
5824  * threads on the same core.  If the shared cache is L3
5825  * or higher, simply returns the same core.
5826  */
5827 static int
5828 find_close_core(int cpu, int core_offset)
5829 {
5830 	struct cpu_group *grp;
5831 	int i;
5832 	int fcpu;
5833 	cpuset_t cs;
5834 
5835 	grp = cpu_top;
5836 	if (grp == NULL)
5837 		return cpu;
5838 	i = 0;
5839 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5840 		/* If the child only has one cpu, don't descend */
5841 		if (grp->cg_child[i].cg_count <= 1)
5842 			break;
5843 		grp = &grp->cg_child[i];
5844 	}
5845 
5846 	/* If they don't share at least an L2 cache, use the same CPU */
5847 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5848 		return cpu;
5849 
5850 	/* Now pick one */
5851 	CPU_COPY(&grp->cg_mask, &cs);
5852 
5853 	/* Add the selected CPU offset to core offset. */
5854 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5855 		if (fcpu - 1 == cpu)
5856 			break;
5857 		CPU_CLR(fcpu - 1, &cs);
5858 	}
5859 	MPASS(fcpu);
5860 
5861 	core_offset += i;
5862 
5863 	CPU_COPY(&grp->cg_mask, &cs);
5864 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5865 		MPASS(CPU_FFS(&cs));
5866 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5867 	}
5868 	MPASS(CPU_FFS(&cs));
5869 	return CPU_FFS(&cs) - 1;
5870 }
5871 #else
5872 static int
5873 find_close_core(int cpu, int core_offset __unused)
5874 {
5875 	return cpu;
5876 }
5877 #endif
5878 
5879 static int
5880 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5881 {
5882 	switch (type) {
5883 	case IFLIB_INTR_TX:
5884 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5885 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5886 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5887 	case IFLIB_INTR_RX:
5888 	case IFLIB_INTR_RXTX:
5889 		/* RX queues get the specified core */
5890 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5891 	default:
5892 		return -1;
5893 	}
5894 }
5895 #else
5896 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5897 #define find_close_core(cpuid, tid)	CPU_FIRST()
5898 #define find_nth(ctx, gid)		CPU_FIRST()
5899 #endif
5900 
5901 /* Just to avoid copy/paste */
5902 static inline int
5903 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5904     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5905     const char *name)
5906 {
5907 	device_t dev;
5908 	int co, cpuid, err, tid;
5909 
5910 	dev = ctx->ifc_dev;
5911 	co = ctx->ifc_sysctl_core_offset;
5912 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5913 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5914 	cpuid = find_nth(ctx, qid + co);
5915 	tid = get_core_offset(ctx, type, qid);
5916 	if (tid < 0) {
5917 		device_printf(dev, "get_core_offset failed\n");
5918 		return (EOPNOTSUPP);
5919 	}
5920 	cpuid = find_close_core(cpuid, tid);
5921 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5922 	    name);
5923 	if (err) {
5924 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5925 		return (err);
5926 	}
5927 #ifdef notyet
5928 	if (cpuid > ctx->ifc_cpuid_highest)
5929 		ctx->ifc_cpuid_highest = cpuid;
5930 #endif
5931 	return (0);
5932 }
5933 
5934 int
5935 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5936 			iflib_intr_type_t type, driver_filter_t *filter,
5937 			void *filter_arg, int qid, const char *name)
5938 {
5939 	device_t dev;
5940 	struct grouptask *gtask;
5941 	struct taskqgroup *tqg;
5942 	iflib_filter_info_t info;
5943 	gtask_fn_t *fn;
5944 	int tqrid, err;
5945 	driver_filter_t *intr_fast;
5946 	void *q;
5947 
5948 	info = &ctx->ifc_filter_info;
5949 	tqrid = rid;
5950 
5951 	switch (type) {
5952 	/* XXX merge tx/rx for netmap? */
5953 	case IFLIB_INTR_TX:
5954 		q = &ctx->ifc_txqs[qid];
5955 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5956 		gtask = &ctx->ifc_txqs[qid].ift_task;
5957 		tqg = qgroup_if_io_tqg;
5958 		fn = _task_fn_tx;
5959 		intr_fast = iflib_fast_intr;
5960 		GROUPTASK_INIT(gtask, 0, fn, q);
5961 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5962 		break;
5963 	case IFLIB_INTR_RX:
5964 		q = &ctx->ifc_rxqs[qid];
5965 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5966 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5967 		tqg = qgroup_if_io_tqg;
5968 		fn = _task_fn_rx;
5969 		intr_fast = iflib_fast_intr;
5970 		GROUPTASK_INIT(gtask, 0, fn, q);
5971 		break;
5972 	case IFLIB_INTR_RXTX:
5973 		q = &ctx->ifc_rxqs[qid];
5974 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5975 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5976 		tqg = qgroup_if_io_tqg;
5977 		fn = _task_fn_rx;
5978 		intr_fast = iflib_fast_intr_rxtx;
5979 		GROUPTASK_INIT(gtask, 0, fn, q);
5980 		break;
5981 	case IFLIB_INTR_ADMIN:
5982 		q = ctx;
5983 		tqrid = -1;
5984 		info = &ctx->ifc_filter_info;
5985 		gtask = &ctx->ifc_admin_task;
5986 		tqg = qgroup_if_config_tqg;
5987 		fn = _task_fn_admin;
5988 		intr_fast = iflib_fast_intr_ctx;
5989 		break;
5990 	default:
5991 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5992 		    __func__);
5993 		return (EINVAL);
5994 	}
5995 
5996 	info->ifi_filter = filter;
5997 	info->ifi_filter_arg = filter_arg;
5998 	info->ifi_task = gtask;
5999 	info->ifi_ctx = q;
6000 
6001 	dev = ctx->ifc_dev;
6002 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6003 	if (err != 0) {
6004 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6005 		return (err);
6006 	}
6007 	if (type == IFLIB_INTR_ADMIN)
6008 		return (0);
6009 
6010 	if (tqrid != -1) {
6011 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6012 		    q, name);
6013 		if (err)
6014 			return (err);
6015 	} else {
6016 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6017 	}
6018 
6019 	return (0);
6020 }
6021 
6022 void
6023 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6024 {
6025 	struct grouptask *gtask;
6026 	struct taskqgroup *tqg;
6027 	gtask_fn_t *fn;
6028 	void *q;
6029 	int err;
6030 
6031 	switch (type) {
6032 	case IFLIB_INTR_TX:
6033 		q = &ctx->ifc_txqs[qid];
6034 		gtask = &ctx->ifc_txqs[qid].ift_task;
6035 		tqg = qgroup_if_io_tqg;
6036 		fn = _task_fn_tx;
6037 		break;
6038 	case IFLIB_INTR_RX:
6039 		q = &ctx->ifc_rxqs[qid];
6040 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6041 		tqg = qgroup_if_io_tqg;
6042 		fn = _task_fn_rx;
6043 		break;
6044 	case IFLIB_INTR_IOV:
6045 		q = ctx;
6046 		gtask = &ctx->ifc_vflr_task;
6047 		tqg = qgroup_if_config_tqg;
6048 		fn = _task_fn_iov;
6049 		break;
6050 	default:
6051 		panic("unknown net intr type");
6052 	}
6053 	GROUPTASK_INIT(gtask, 0, fn, q);
6054 	if (irq != NULL) {
6055 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6056 		    q, name);
6057 		if (err)
6058 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6059 			    irq->ii_res, name);
6060 	} else {
6061 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6062 	}
6063 }
6064 
6065 void
6066 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6067 {
6068 
6069 	if (irq->ii_tag)
6070 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6071 
6072 	if (irq->ii_res)
6073 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6074 		    rman_get_rid(irq->ii_res), irq->ii_res);
6075 }
6076 
6077 static int
6078 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6079 {
6080 	iflib_txq_t txq = ctx->ifc_txqs;
6081 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6082 	if_irq_t irq = &ctx->ifc_legacy_irq;
6083 	iflib_filter_info_t info;
6084 	device_t dev;
6085 	struct grouptask *gtask;
6086 	struct resource *res;
6087 	struct taskqgroup *tqg;
6088 	gtask_fn_t *fn;
6089 	void *q;
6090 	int err, tqrid;
6091 	bool rx_only;
6092 
6093 	q = &ctx->ifc_rxqs[0];
6094 	info = &rxq[0].ifr_filter_info;
6095 	gtask = &rxq[0].ifr_task;
6096 	tqg = qgroup_if_io_tqg;
6097 	tqrid = *rid;
6098 	fn = _task_fn_rx;
6099 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6100 
6101 	ctx->ifc_flags |= IFC_LEGACY;
6102 	info->ifi_filter = filter;
6103 	info->ifi_filter_arg = filter_arg;
6104 	info->ifi_task = gtask;
6105 	info->ifi_ctx = rx_only ? ctx : q;
6106 
6107 	dev = ctx->ifc_dev;
6108 	/* We allocate a single interrupt resource */
6109 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6110 	    iflib_fast_intr_rxtx, NULL, info, name);
6111 	if (err != 0)
6112 		return (err);
6113 	GROUPTASK_INIT(gtask, 0, fn, q);
6114 	res = irq->ii_res;
6115 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6116 
6117 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6118 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6119 	    "tx");
6120 	return (0);
6121 }
6122 
6123 void
6124 iflib_led_create(if_ctx_t ctx)
6125 {
6126 
6127 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6128 	    device_get_nameunit(ctx->ifc_dev));
6129 }
6130 
6131 void
6132 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6133 {
6134 
6135 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6136 }
6137 
6138 void
6139 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6140 {
6141 
6142 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6143 }
6144 
6145 void
6146 iflib_admin_intr_deferred(if_ctx_t ctx)
6147 {
6148 #ifdef INVARIANTS
6149 	struct grouptask *gtask;
6150 
6151 	gtask = &ctx->ifc_admin_task;
6152 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6153 #endif
6154 
6155 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6156 }
6157 
6158 void
6159 iflib_iov_intr_deferred(if_ctx_t ctx)
6160 {
6161 
6162 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6163 }
6164 
6165 void
6166 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6167 {
6168 
6169 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6170 	    name);
6171 }
6172 
6173 void
6174 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6175 	const char *name)
6176 {
6177 
6178 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6179 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6180 	    name);
6181 }
6182 
6183 void
6184 iflib_config_gtask_deinit(struct grouptask *gtask)
6185 {
6186 
6187 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6188 }
6189 
6190 void
6191 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6192 {
6193 	if_t ifp = ctx->ifc_ifp;
6194 	iflib_txq_t txq = ctx->ifc_txqs;
6195 
6196 	if_setbaudrate(ifp, baudrate);
6197 	if (baudrate >= IF_Gbps(10)) {
6198 		STATE_LOCK(ctx);
6199 		ctx->ifc_flags |= IFC_PREFETCH;
6200 		STATE_UNLOCK(ctx);
6201 	}
6202 	/* If link down, disable watchdog */
6203 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6204 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6205 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6206 	}
6207 	ctx->ifc_link_state = link_state;
6208 	if_link_state_change(ifp, link_state);
6209 }
6210 
6211 static int
6212 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6213 {
6214 	int credits;
6215 #ifdef INVARIANTS
6216 	int credits_pre = txq->ift_cidx_processed;
6217 #endif
6218 
6219 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6220 	    BUS_DMASYNC_POSTREAD);
6221 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6222 		return (0);
6223 
6224 	txq->ift_processed += credits;
6225 	txq->ift_cidx_processed += credits;
6226 
6227 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6228 	if (txq->ift_cidx_processed >= txq->ift_size)
6229 		txq->ift_cidx_processed -= txq->ift_size;
6230 	return (credits);
6231 }
6232 
6233 static int
6234 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6235 {
6236 	iflib_fl_t fl;
6237 	u_int i;
6238 
6239 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6240 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6241 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6242 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6243 	    budget));
6244 }
6245 
6246 void
6247 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6248 	const char *description, if_int_delay_info_t info,
6249 	int offset, int value)
6250 {
6251 	info->iidi_ctx = ctx;
6252 	info->iidi_offset = offset;
6253 	info->iidi_value = value;
6254 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6255 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6256 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6257 	    info, 0, iflib_sysctl_int_delay, "I", description);
6258 }
6259 
6260 struct sx *
6261 iflib_ctx_lock_get(if_ctx_t ctx)
6262 {
6263 
6264 	return (&ctx->ifc_ctx_sx);
6265 }
6266 
6267 static int
6268 iflib_msix_init(if_ctx_t ctx)
6269 {
6270 	device_t dev = ctx->ifc_dev;
6271 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6272 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6273 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6274 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6275 
6276 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6277 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6278 
6279 	if (bootverbose)
6280 		device_printf(dev, "msix_init qsets capped at %d\n",
6281 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6282 
6283 	/* Override by tuneable */
6284 	if (scctx->isc_disable_msix)
6285 		goto msi;
6286 
6287 	/* First try MSI-X */
6288 	if ((msgs = pci_msix_count(dev)) == 0) {
6289 		if (bootverbose)
6290 			device_printf(dev, "MSI-X not supported or disabled\n");
6291 		goto msi;
6292 	}
6293 
6294 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6295 	/*
6296 	 * bar == -1 => "trust me I know what I'm doing"
6297 	 * Some drivers are for hardware that is so shoddily
6298 	 * documented that no one knows which bars are which
6299 	 * so the developer has to map all bars. This hack
6300 	 * allows shoddy garbage to use MSI-X in this framework.
6301 	 */
6302 	if (bar != -1) {
6303 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6304 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6305 		if (ctx->ifc_msix_mem == NULL) {
6306 			device_printf(dev, "Unable to map MSI-X table\n");
6307 			goto msi;
6308 		}
6309 	}
6310 
6311 	admincnt = sctx->isc_admin_intrcnt;
6312 #if IFLIB_DEBUG
6313 	/* use only 1 qset in debug mode */
6314 	queuemsgs = min(msgs - admincnt, 1);
6315 #else
6316 	queuemsgs = msgs - admincnt;
6317 #endif
6318 #ifdef RSS
6319 	queues = imin(queuemsgs, rss_getnumbuckets());
6320 #else
6321 	queues = queuemsgs;
6322 #endif
6323 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6324 	if (bootverbose)
6325 		device_printf(dev,
6326 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6327 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6328 #ifdef  RSS
6329 	/* If we're doing RSS, clamp at the number of RSS buckets */
6330 	if (queues > rss_getnumbuckets())
6331 		queues = rss_getnumbuckets();
6332 #endif
6333 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6334 		rx_queues = iflib_num_rx_queues;
6335 	else
6336 		rx_queues = queues;
6337 
6338 	if (rx_queues > scctx->isc_nrxqsets)
6339 		rx_queues = scctx->isc_nrxqsets;
6340 
6341 	/*
6342 	 * We want this to be all logical CPUs by default
6343 	 */
6344 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6345 		tx_queues = iflib_num_tx_queues;
6346 	else
6347 		tx_queues = mp_ncpus;
6348 
6349 	if (tx_queues > scctx->isc_ntxqsets)
6350 		tx_queues = scctx->isc_ntxqsets;
6351 
6352 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6353 #ifdef INVARIANTS
6354 		if (tx_queues != rx_queues)
6355 			device_printf(dev,
6356 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6357 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6358 #endif
6359 		tx_queues = min(rx_queues, tx_queues);
6360 		rx_queues = min(rx_queues, tx_queues);
6361 	}
6362 
6363 	vectors = rx_queues + admincnt;
6364 	if (msgs < vectors) {
6365 		device_printf(dev,
6366 		    "insufficient number of MSI-X vectors "
6367 		    "(supported %d, need %d)\n", msgs, vectors);
6368 		goto msi;
6369 	}
6370 
6371 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6372 	    tx_queues);
6373 	msgs = vectors;
6374 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6375 		if (vectors != msgs) {
6376 			device_printf(dev,
6377 			    "Unable to allocate sufficient MSI-X vectors "
6378 			    "(got %d, need %d)\n", vectors, msgs);
6379 			pci_release_msi(dev);
6380 			if (bar != -1) {
6381 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6382 				    ctx->ifc_msix_mem);
6383 				ctx->ifc_msix_mem = NULL;
6384 			}
6385 			goto msi;
6386 		}
6387 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6388 		    vectors);
6389 		scctx->isc_vectors = vectors;
6390 		scctx->isc_nrxqsets = rx_queues;
6391 		scctx->isc_ntxqsets = tx_queues;
6392 		scctx->isc_intr = IFLIB_INTR_MSIX;
6393 
6394 		return (vectors);
6395 	} else {
6396 		device_printf(dev,
6397 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6398 		    err);
6399 		if (bar != -1) {
6400 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6401 			    ctx->ifc_msix_mem);
6402 			ctx->ifc_msix_mem = NULL;
6403 		}
6404 	}
6405 
6406 msi:
6407 	vectors = pci_msi_count(dev);
6408 	scctx->isc_nrxqsets = 1;
6409 	scctx->isc_ntxqsets = 1;
6410 	scctx->isc_vectors = vectors;
6411 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6412 		device_printf(dev,"Using an MSI interrupt\n");
6413 		scctx->isc_intr = IFLIB_INTR_MSI;
6414 	} else {
6415 		scctx->isc_vectors = 1;
6416 		device_printf(dev,"Using a Legacy interrupt\n");
6417 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6418 	}
6419 
6420 	return (vectors);
6421 }
6422 
6423 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6424 
6425 static int
6426 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6427 {
6428 	int rc;
6429 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6430 	struct sbuf *sb;
6431 	const char *ring_state = "UNKNOWN";
6432 
6433 	/* XXX needed ? */
6434 	rc = sysctl_wire_old_buffer(req, 0);
6435 	MPASS(rc == 0);
6436 	if (rc != 0)
6437 		return (rc);
6438 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6439 	MPASS(sb != NULL);
6440 	if (sb == NULL)
6441 		return (ENOMEM);
6442 	if (state[3] <= 3)
6443 		ring_state = ring_states[state[3]];
6444 
6445 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6446 		    state[0], state[1], state[2], ring_state);
6447 	rc = sbuf_finish(sb);
6448 	sbuf_delete(sb);
6449         return(rc);
6450 }
6451 
6452 enum iflib_ndesc_handler {
6453 	IFLIB_NTXD_HANDLER,
6454 	IFLIB_NRXD_HANDLER,
6455 };
6456 
6457 static int
6458 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6459 {
6460 	if_ctx_t ctx = (void *)arg1;
6461 	enum iflib_ndesc_handler type = arg2;
6462 	char buf[256] = {0};
6463 	qidx_t *ndesc;
6464 	char *p, *next;
6465 	int nqs, rc, i;
6466 
6467 	nqs = 8;
6468 	switch(type) {
6469 	case IFLIB_NTXD_HANDLER:
6470 		ndesc = ctx->ifc_sysctl_ntxds;
6471 		if (ctx->ifc_sctx)
6472 			nqs = ctx->ifc_sctx->isc_ntxqs;
6473 		break;
6474 	case IFLIB_NRXD_HANDLER:
6475 		ndesc = ctx->ifc_sysctl_nrxds;
6476 		if (ctx->ifc_sctx)
6477 			nqs = ctx->ifc_sctx->isc_nrxqs;
6478 		break;
6479 	default:
6480 		printf("%s: unhandled type\n", __func__);
6481 		return (EINVAL);
6482 	}
6483 	if (nqs == 0)
6484 		nqs = 8;
6485 
6486 	for (i=0; i<8; i++) {
6487 		if (i >= nqs)
6488 			break;
6489 		if (i)
6490 			strcat(buf, ",");
6491 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6492 	}
6493 
6494 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6495 	if (rc || req->newptr == NULL)
6496 		return rc;
6497 
6498 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6499 	    i++, p = strsep(&next, " ,")) {
6500 		ndesc[i] = strtoul(p, NULL, 10);
6501 	}
6502 
6503 	return(rc);
6504 }
6505 
6506 #define NAME_BUFLEN 32
6507 static void
6508 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6509 {
6510         device_t dev = iflib_get_dev(ctx);
6511 	struct sysctl_oid_list *child, *oid_list;
6512 	struct sysctl_ctx_list *ctx_list;
6513 	struct sysctl_oid *node;
6514 
6515 	ctx_list = device_get_sysctl_ctx(dev);
6516 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6517 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6518 						      CTLFLAG_RD, NULL, "IFLIB fields");
6519 	oid_list = SYSCTL_CHILDREN(node);
6520 
6521 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6522 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6523 		       "driver version");
6524 
6525 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6526 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6527 			"# of txqs to use, 0 => use default #");
6528 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6529 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6530 			"# of rxqs to use, 0 => use default #");
6531 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6532 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6533                        "permit #txq != #rxq");
6534 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6535                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6536                       "disable MSI-X (default 0)");
6537 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6538 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6539 		       "set the RX budget");
6540 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6541 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6542 		       "cause TX to abdicate instead of running to completion");
6543 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6544 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6545 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6546 		       "offset to start using cores at");
6547 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6548 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6549 		       "use separate cores for TX and RX");
6550 
6551 	/* XXX change for per-queue sizes */
6552 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6553 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6554                        mp_ndesc_handler, "A",
6555 		       "list of # of TX descriptors to use, 0 = use default #");
6556 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6557 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6558                        mp_ndesc_handler, "A",
6559 		       "list of # of RX descriptors to use, 0 = use default #");
6560 }
6561 
6562 static void
6563 iflib_add_device_sysctl_post(if_ctx_t ctx)
6564 {
6565 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6566 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6567         device_t dev = iflib_get_dev(ctx);
6568 	struct sysctl_oid_list *child;
6569 	struct sysctl_ctx_list *ctx_list;
6570 	iflib_fl_t fl;
6571 	iflib_txq_t txq;
6572 	iflib_rxq_t rxq;
6573 	int i, j;
6574 	char namebuf[NAME_BUFLEN];
6575 	char *qfmt;
6576 	struct sysctl_oid *queue_node, *fl_node, *node;
6577 	struct sysctl_oid_list *queue_list, *fl_list;
6578 	ctx_list = device_get_sysctl_ctx(dev);
6579 
6580 	node = ctx->ifc_sysctl_node;
6581 	child = SYSCTL_CHILDREN(node);
6582 
6583 	if (scctx->isc_ntxqsets > 100)
6584 		qfmt = "txq%03d";
6585 	else if (scctx->isc_ntxqsets > 10)
6586 		qfmt = "txq%02d";
6587 	else
6588 		qfmt = "txq%d";
6589 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6590 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6591 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6592 					     CTLFLAG_RD, NULL, "Queue Name");
6593 		queue_list = SYSCTL_CHILDREN(queue_node);
6594 #if MEMORY_LOGGING
6595 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6596 				CTLFLAG_RD,
6597 				&txq->ift_dequeued, "total mbufs freed");
6598 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6599 				CTLFLAG_RD,
6600 				&txq->ift_enqueued, "total mbufs enqueued");
6601 #endif
6602 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6603 				   CTLFLAG_RD,
6604 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6605 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6606 				   CTLFLAG_RD,
6607 				   &txq->ift_pullups, "# of times m_pullup was called");
6608 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6609 				   CTLFLAG_RD,
6610 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6611 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6612 				   CTLFLAG_RD,
6613 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6614 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6615 				   CTLFLAG_RD,
6616 				   &txq->ift_map_failed, "# of times DMA map failed");
6617 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6618 				   CTLFLAG_RD,
6619 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6620 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6621 				   CTLFLAG_RD,
6622 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6623 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6624 				   CTLFLAG_RD,
6625 				   &txq->ift_pidx, 1, "Producer Index");
6626 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6627 				   CTLFLAG_RD,
6628 				   &txq->ift_cidx, 1, "Consumer Index");
6629 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6630 				   CTLFLAG_RD,
6631 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6632 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6633 				   CTLFLAG_RD,
6634 				   &txq->ift_in_use, 1, "descriptors in use");
6635 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6636 				   CTLFLAG_RD,
6637 				   &txq->ift_processed, "descriptors procesed for clean");
6638 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6639 				   CTLFLAG_RD,
6640 				   &txq->ift_cleaned, "total cleaned");
6641 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6642 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6643 				0, mp_ring_state_handler, "A", "soft ring state");
6644 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6645 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6646 				       "# of enqueues to the mp_ring for this queue");
6647 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6648 				       CTLFLAG_RD, &txq->ift_br->drops,
6649 				       "# of drops in the mp_ring for this queue");
6650 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6651 				       CTLFLAG_RD, &txq->ift_br->starts,
6652 				       "# of normal consumer starts in the mp_ring for this queue");
6653 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6654 				       CTLFLAG_RD, &txq->ift_br->stalls,
6655 					       "# of consumer stalls in the mp_ring for this queue");
6656 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6657 			       CTLFLAG_RD, &txq->ift_br->restarts,
6658 				       "# of consumer restarts in the mp_ring for this queue");
6659 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6660 				       CTLFLAG_RD, &txq->ift_br->abdications,
6661 				       "# of consumer abdications in the mp_ring for this queue");
6662 	}
6663 
6664 	if (scctx->isc_nrxqsets > 100)
6665 		qfmt = "rxq%03d";
6666 	else if (scctx->isc_nrxqsets > 10)
6667 		qfmt = "rxq%02d";
6668 	else
6669 		qfmt = "rxq%d";
6670 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6671 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6672 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6673 					     CTLFLAG_RD, NULL, "Queue Name");
6674 		queue_list = SYSCTL_CHILDREN(queue_node);
6675 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6676 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6677 				       CTLFLAG_RD,
6678 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6679 		}
6680 
6681 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6682 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6683 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6684 						     CTLFLAG_RD, NULL, "freelist Name");
6685 			fl_list = SYSCTL_CHILDREN(fl_node);
6686 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6687 				       CTLFLAG_RD,
6688 				       &fl->ifl_pidx, 1, "Producer Index");
6689 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6690 				       CTLFLAG_RD,
6691 				       &fl->ifl_cidx, 1, "Consumer Index");
6692 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6693 				       CTLFLAG_RD,
6694 				       &fl->ifl_credits, 1, "credits available");
6695 #if MEMORY_LOGGING
6696 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6697 					CTLFLAG_RD,
6698 					&fl->ifl_m_enqueued, "mbufs allocated");
6699 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6700 					CTLFLAG_RD,
6701 					&fl->ifl_m_dequeued, "mbufs freed");
6702 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6703 					CTLFLAG_RD,
6704 					&fl->ifl_cl_enqueued, "clusters allocated");
6705 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6706 					CTLFLAG_RD,
6707 					&fl->ifl_cl_dequeued, "clusters freed");
6708 #endif
6709 
6710 		}
6711 	}
6712 
6713 }
6714 
6715 void
6716 iflib_request_reset(if_ctx_t ctx)
6717 {
6718 
6719 	STATE_LOCK(ctx);
6720 	ctx->ifc_flags |= IFC_DO_RESET;
6721 	STATE_UNLOCK(ctx);
6722 }
6723 
6724 #ifndef __NO_STRICT_ALIGNMENT
6725 static struct mbuf *
6726 iflib_fixup_rx(struct mbuf *m)
6727 {
6728 	struct mbuf *n;
6729 
6730 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6731 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6732 		m->m_data += ETHER_HDR_LEN;
6733 		n = m;
6734 	} else {
6735 		MGETHDR(n, M_NOWAIT, MT_DATA);
6736 		if (n == NULL) {
6737 			m_freem(m);
6738 			return (NULL);
6739 		}
6740 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6741 		m->m_data += ETHER_HDR_LEN;
6742 		m->m_len -= ETHER_HDR_LEN;
6743 		n->m_len = ETHER_HDR_LEN;
6744 		M_MOVE_PKTHDR(n, m);
6745 		n->m_next = m;
6746 	}
6747 	return (n);
6748 }
6749 #endif
6750 
6751 #ifdef DEBUGNET
6752 static void
6753 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6754 {
6755 	if_ctx_t ctx;
6756 
6757 	ctx = if_getsoftc(ifp);
6758 	CTX_LOCK(ctx);
6759 	*nrxr = NRXQSETS(ctx);
6760 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6761 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6762 	CTX_UNLOCK(ctx);
6763 }
6764 
6765 static void
6766 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6767 {
6768 	if_ctx_t ctx;
6769 	if_softc_ctx_t scctx;
6770 	iflib_fl_t fl;
6771 	iflib_rxq_t rxq;
6772 	int i, j;
6773 
6774 	ctx = if_getsoftc(ifp);
6775 	scctx = &ctx->ifc_softc_ctx;
6776 
6777 	switch (event) {
6778 	case DEBUGNET_START:
6779 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6780 			rxq = &ctx->ifc_rxqs[i];
6781 			for (j = 0; j < rxq->ifr_nfl; j++) {
6782 				fl = rxq->ifr_fl;
6783 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6784 			}
6785 		}
6786 		iflib_no_tx_batch = 1;
6787 		break;
6788 	default:
6789 		break;
6790 	}
6791 }
6792 
6793 static int
6794 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6795 {
6796 	if_ctx_t ctx;
6797 	iflib_txq_t txq;
6798 	int error;
6799 
6800 	ctx = if_getsoftc(ifp);
6801 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6802 	    IFF_DRV_RUNNING)
6803 		return (EBUSY);
6804 
6805 	txq = &ctx->ifc_txqs[0];
6806 	error = iflib_encap(txq, &m);
6807 	if (error == 0)
6808 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6809 	return (error);
6810 }
6811 
6812 static int
6813 iflib_debugnet_poll(if_t ifp, int count)
6814 {
6815 	if_ctx_t ctx;
6816 	if_softc_ctx_t scctx;
6817 	iflib_txq_t txq;
6818 	int i;
6819 
6820 	ctx = if_getsoftc(ifp);
6821 	scctx = &ctx->ifc_softc_ctx;
6822 
6823 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6824 	    IFF_DRV_RUNNING)
6825 		return (EBUSY);
6826 
6827 	txq = &ctx->ifc_txqs[0];
6828 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6829 
6830 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6831 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6832 	return (0);
6833 }
6834 #endif /* DEBUGNET */
6835