xref: /freebsd/sys/net/iflib.c (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
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 	if (txq->ift_sds.ifsd_map != NULL) {
1700 		map = txq->ift_sds.ifsd_map[i];
1701 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1702 		bus_dmamap_unload(txq->ift_buf_tag, map);
1703 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1704 		txq->ift_sds.ifsd_map[i] = NULL;
1705 	}
1706 
1707 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1708 		map = txq->ift_sds.ifsd_tso_map[i];
1709 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1710 		    BUS_DMASYNC_POSTWRITE);
1711 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1712 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1713 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1714 	}
1715 }
1716 
1717 static void
1718 iflib_txq_destroy(iflib_txq_t txq)
1719 {
1720 	if_ctx_t ctx = txq->ift_ctx;
1721 
1722 	for (int i = 0; i < txq->ift_size; i++)
1723 		iflib_txsd_destroy(ctx, txq, i);
1724 
1725 	if (txq->ift_br != NULL) {
1726 		ifmp_ring_free(txq->ift_br);
1727 		txq->ift_br = NULL;
1728 	}
1729 
1730 	mtx_destroy(&txq->ift_mtx);
1731 
1732 	if (txq->ift_sds.ifsd_map != NULL) {
1733 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1734 		txq->ift_sds.ifsd_map = NULL;
1735 	}
1736 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1737 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1738 		txq->ift_sds.ifsd_tso_map = NULL;
1739 	}
1740 	if (txq->ift_sds.ifsd_m != NULL) {
1741 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1742 		txq->ift_sds.ifsd_m = NULL;
1743 	}
1744 	if (txq->ift_buf_tag != NULL) {
1745 		bus_dma_tag_destroy(txq->ift_buf_tag);
1746 		txq->ift_buf_tag = NULL;
1747 	}
1748 	if (txq->ift_tso_buf_tag != NULL) {
1749 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1750 		txq->ift_tso_buf_tag = NULL;
1751 	}
1752 	if (txq->ift_ifdi != NULL) {
1753 		free(txq->ift_ifdi, M_IFLIB);
1754 	}
1755 }
1756 
1757 static void
1758 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1759 {
1760 	struct mbuf **mp;
1761 
1762 	mp = &txq->ift_sds.ifsd_m[i];
1763 	if (*mp == NULL)
1764 		return;
1765 
1766 	if (txq->ift_sds.ifsd_map != NULL) {
1767 		bus_dmamap_sync(txq->ift_buf_tag,
1768 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1769 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1770 	}
1771 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1772 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1773 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1774 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1775 		    txq->ift_sds.ifsd_tso_map[i]);
1776 	}
1777 	m_free(*mp);
1778 	DBG_COUNTER_INC(tx_frees);
1779 	*mp = NULL;
1780 }
1781 
1782 static int
1783 iflib_txq_setup(iflib_txq_t txq)
1784 {
1785 	if_ctx_t ctx = txq->ift_ctx;
1786 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1787 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1788 	iflib_dma_info_t di;
1789 	int i;
1790 
1791 	/* Set number of descriptors available */
1792 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1793 	/* XXX make configurable */
1794 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1795 
1796 	/* Reset indices */
1797 	txq->ift_cidx_processed = 0;
1798 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1799 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1800 
1801 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1802 		bzero((void *)di->idi_vaddr, di->idi_size);
1803 
1804 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1805 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1806 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1807 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1808 	return (0);
1809 }
1810 
1811 /*********************************************************************
1812  *
1813  *  Allocate DMA resources for RX buffers as well as memory for the RX
1814  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1815  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1816  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1817  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1818  *  packet, the maximum number of entries we'll need is equal to the
1819  *  number of hardware receive descriptors that we've allocated.
1820  *
1821  **********************************************************************/
1822 static int
1823 iflib_rxsd_alloc(iflib_rxq_t rxq)
1824 {
1825 	if_ctx_t ctx = rxq->ifr_ctx;
1826 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1827 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1828 	device_t dev = ctx->ifc_dev;
1829 	iflib_fl_t fl;
1830 	int			err;
1831 
1832 	MPASS(scctx->isc_nrxd[0] > 0);
1833 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1834 
1835 	fl = rxq->ifr_fl;
1836 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1837 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1838 		/* Set up DMA tag for RX buffers. */
1839 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1840 					 1, 0,			/* alignment, bounds */
1841 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1842 					 BUS_SPACE_MAXADDR,	/* highaddr */
1843 					 NULL, NULL,		/* filter, filterarg */
1844 					 sctx->isc_rx_maxsize,	/* maxsize */
1845 					 sctx->isc_rx_nsegments,	/* nsegments */
1846 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1847 					 0,			/* flags */
1848 					 NULL,			/* lockfunc */
1849 					 NULL,			/* lockarg */
1850 					 &fl->ifl_buf_tag);
1851 		if (err) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX DMA tag: %d\n", err);
1854 			goto fail;
1855 		}
1856 
1857 		/* Allocate memory for the RX mbuf map. */
1858 		if (!(fl->ifl_sds.ifsd_m =
1859 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1860 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1861 			device_printf(dev,
1862 			    "Unable to allocate RX mbuf map memory\n");
1863 			err = ENOMEM;
1864 			goto fail;
1865 		}
1866 
1867 		/* Allocate memory for the direct RX cluster pointer map. */
1868 		if (!(fl->ifl_sds.ifsd_cl =
1869 		      (caddr_t *) malloc(sizeof(caddr_t) *
1870 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1871 			device_printf(dev,
1872 			    "Unable to allocate RX cluster map memory\n");
1873 			err = ENOMEM;
1874 			goto fail;
1875 		}
1876 
1877 		/* Allocate memory for the RX cluster bus address map. */
1878 		if (!(fl->ifl_sds.ifsd_ba =
1879 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1880 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1881 			device_printf(dev,
1882 			    "Unable to allocate RX bus address map memory\n");
1883 			err = ENOMEM;
1884 			goto fail;
1885 		}
1886 
1887 		/*
1888 		 * Create the DMA maps for RX buffers.
1889 		 */
1890 		if (!(fl->ifl_sds.ifsd_map =
1891 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1892 			device_printf(dev,
1893 			    "Unable to allocate RX buffer DMA map memory\n");
1894 			err = ENOMEM;
1895 			goto fail;
1896 		}
1897 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1898 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1899 			    &fl->ifl_sds.ifsd_map[i]);
1900 			if (err != 0) {
1901 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1902 				goto fail;
1903 			}
1904 		}
1905 	}
1906 	return (0);
1907 
1908 fail:
1909 	iflib_rx_structures_free(ctx);
1910 	return (err);
1911 }
1912 
1913 
1914 /*
1915  * Internal service routines
1916  */
1917 
1918 struct rxq_refill_cb_arg {
1919 	int               error;
1920 	bus_dma_segment_t seg;
1921 	int               nseg;
1922 };
1923 
1924 static void
1925 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1926 {
1927 	struct rxq_refill_cb_arg *cb_arg = arg;
1928 
1929 	cb_arg->error = error;
1930 	cb_arg->seg = segs[0];
1931 	cb_arg->nseg = nseg;
1932 }
1933 
1934 /**
1935  * _iflib_fl_refill - refill an rxq free-buffer list
1936  * @ctx: the iflib context
1937  * @fl: the free list to refill
1938  * @count: the number of new buffers to allocate
1939  *
1940  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1941  * The caller must assure that @count does not exceed the queue's capacity.
1942  */
1943 static void
1944 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1945 {
1946 	struct if_rxd_update iru;
1947 	struct rxq_refill_cb_arg cb_arg;
1948 	struct mbuf *m;
1949 	caddr_t cl, *sd_cl;
1950 	struct mbuf **sd_m;
1951 	bus_dmamap_t *sd_map;
1952 	bus_addr_t bus_addr, *sd_ba;
1953 	int err, frag_idx, i, idx, n, pidx;
1954 	qidx_t credits;
1955 
1956 	sd_m = fl->ifl_sds.ifsd_m;
1957 	sd_map = fl->ifl_sds.ifsd_map;
1958 	sd_cl = fl->ifl_sds.ifsd_cl;
1959 	sd_ba = fl->ifl_sds.ifsd_ba;
1960 	pidx = fl->ifl_pidx;
1961 	idx = pidx;
1962 	frag_idx = fl->ifl_fragidx;
1963 	credits = fl->ifl_credits;
1964 
1965 	i = 0;
1966 	n = count;
1967 	MPASS(n > 0);
1968 	MPASS(credits + n <= fl->ifl_size);
1969 
1970 	if (pidx < fl->ifl_cidx)
1971 		MPASS(pidx + n <= fl->ifl_cidx);
1972 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1973 		MPASS(fl->ifl_gen == 0);
1974 	if (pidx > fl->ifl_cidx)
1975 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1976 
1977 	DBG_COUNTER_INC(fl_refills);
1978 	if (n > 8)
1979 		DBG_COUNTER_INC(fl_refills_large);
1980 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1981 	while (n--) {
1982 		/*
1983 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1984 		 * initialized after rx.
1985 		 *
1986 		 * If the cluster is still set then we know a minimum sized packet was received
1987 		 */
1988 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1989 		    &frag_idx);
1990 		if (frag_idx < 0)
1991 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1992 		MPASS(frag_idx >= 0);
1993 		if ((cl = sd_cl[frag_idx]) == NULL) {
1994 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1995 				break;
1996 
1997 			cb_arg.error = 0;
1998 			MPASS(sd_map != NULL);
1999 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2000 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2001 			    BUS_DMA_NOWAIT);
2002 			if (err != 0 || cb_arg.error) {
2003 				/*
2004 				 * !zone_pack ?
2005 				 */
2006 				if (fl->ifl_zone == zone_pack)
2007 					uma_zfree(fl->ifl_zone, cl);
2008 				break;
2009 			}
2010 
2011 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2012 			sd_cl[frag_idx] = cl;
2013 #if MEMORY_LOGGING
2014 			fl->ifl_cl_enqueued++;
2015 #endif
2016 		} else {
2017 			bus_addr = sd_ba[frag_idx];
2018 		}
2019 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2020 		    BUS_DMASYNC_PREREAD);
2021 
2022 		if (sd_m[frag_idx] == NULL) {
2023 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2024 				break;
2025 			}
2026 			sd_m[frag_idx] = m;
2027 		}
2028 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2029 #if MEMORY_LOGGING
2030 		fl->ifl_m_enqueued++;
2031 #endif
2032 
2033 		DBG_COUNTER_INC(rx_allocs);
2034 		fl->ifl_rxd_idxs[i] = frag_idx;
2035 		fl->ifl_bus_addrs[i] = bus_addr;
2036 		fl->ifl_vm_addrs[i] = cl;
2037 		credits++;
2038 		i++;
2039 		MPASS(credits <= fl->ifl_size);
2040 		if (++idx == fl->ifl_size) {
2041 			fl->ifl_gen = 1;
2042 			idx = 0;
2043 		}
2044 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2045 			iru.iru_pidx = pidx;
2046 			iru.iru_count = i;
2047 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2048 			i = 0;
2049 			pidx = idx;
2050 			fl->ifl_pidx = idx;
2051 			fl->ifl_credits = credits;
2052 		}
2053 	}
2054 
2055 	if (i) {
2056 		iru.iru_pidx = pidx;
2057 		iru.iru_count = i;
2058 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2059 		fl->ifl_pidx = idx;
2060 		fl->ifl_credits = credits;
2061 	}
2062 	DBG_COUNTER_INC(rxd_flush);
2063 	if (fl->ifl_pidx == 0)
2064 		pidx = fl->ifl_size - 1;
2065 	else
2066 		pidx = fl->ifl_pidx - 1;
2067 
2068 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2069 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2070 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2071 	fl->ifl_fragidx = frag_idx;
2072 }
2073 
2074 static __inline void
2075 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2076 {
2077 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2078 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2079 #ifdef INVARIANTS
2080 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2081 #endif
2082 
2083 	MPASS(fl->ifl_credits <= fl->ifl_size);
2084 	MPASS(reclaimable == delta);
2085 
2086 	if (reclaimable > 0)
2087 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2088 }
2089 
2090 uint8_t
2091 iflib_in_detach(if_ctx_t ctx)
2092 {
2093 	bool in_detach;
2094 
2095 	STATE_LOCK(ctx);
2096 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2097 	STATE_UNLOCK(ctx);
2098 	return (in_detach);
2099 }
2100 
2101 static void
2102 iflib_fl_bufs_free(iflib_fl_t fl)
2103 {
2104 	iflib_dma_info_t idi = fl->ifl_ifdi;
2105 	bus_dmamap_t sd_map;
2106 	uint32_t i;
2107 
2108 	for (i = 0; i < fl->ifl_size; i++) {
2109 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2110 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2111 
2112 		if (*sd_cl != NULL) {
2113 			sd_map = fl->ifl_sds.ifsd_map[i];
2114 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2115 			    BUS_DMASYNC_POSTREAD);
2116 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2117 			if (*sd_cl != NULL)
2118 				uma_zfree(fl->ifl_zone, *sd_cl);
2119 			if (*sd_m != NULL) {
2120 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2121 				uma_zfree(zone_mbuf, *sd_m);
2122 			}
2123 		} else {
2124 			MPASS(*sd_cl == NULL);
2125 			MPASS(*sd_m == NULL);
2126 		}
2127 #if MEMORY_LOGGING
2128 		fl->ifl_m_dequeued++;
2129 		fl->ifl_cl_dequeued++;
2130 #endif
2131 		*sd_cl = NULL;
2132 		*sd_m = NULL;
2133 	}
2134 #ifdef INVARIANTS
2135 	for (i = 0; i < fl->ifl_size; i++) {
2136 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2137 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2138 	}
2139 #endif
2140 	/*
2141 	 * Reset free list values
2142 	 */
2143 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2144 	bzero(idi->idi_vaddr, idi->idi_size);
2145 }
2146 
2147 /*********************************************************************
2148  *
2149  *  Initialize a free list and its buffers.
2150  *
2151  **********************************************************************/
2152 static int
2153 iflib_fl_setup(iflib_fl_t fl)
2154 {
2155 	iflib_rxq_t rxq = fl->ifl_rxq;
2156 	if_ctx_t ctx = rxq->ifr_ctx;
2157 
2158 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2159 	/*
2160 	** Free current RX buffer structs and their mbufs
2161 	*/
2162 	iflib_fl_bufs_free(fl);
2163 	/* Now replenish the mbufs */
2164 	MPASS(fl->ifl_credits == 0);
2165 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2166 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2167 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2168 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2169 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2170 
2171 
2172 	/* avoid pre-allocating zillions of clusters to an idle card
2173 	 * potentially speeding up attach
2174 	 */
2175 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2176 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2177 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2178 		return (ENOBUFS);
2179 	/*
2180 	 * handle failure
2181 	 */
2182 	MPASS(rxq != NULL);
2183 	MPASS(fl->ifl_ifdi != NULL);
2184 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2185 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2186 	return (0);
2187 }
2188 
2189 /*********************************************************************
2190  *
2191  *  Free receive ring data structures
2192  *
2193  **********************************************************************/
2194 static void
2195 iflib_rx_sds_free(iflib_rxq_t rxq)
2196 {
2197 	iflib_fl_t fl;
2198 	int i, j;
2199 
2200 	if (rxq->ifr_fl != NULL) {
2201 		for (i = 0; i < rxq->ifr_nfl; i++) {
2202 			fl = &rxq->ifr_fl[i];
2203 			if (fl->ifl_buf_tag != NULL) {
2204 				if (fl->ifl_sds.ifsd_map != NULL) {
2205 					for (j = 0; j < fl->ifl_size; j++) {
2206 						bus_dmamap_sync(
2207 						    fl->ifl_buf_tag,
2208 						    fl->ifl_sds.ifsd_map[j],
2209 						    BUS_DMASYNC_POSTREAD);
2210 						bus_dmamap_unload(
2211 						    fl->ifl_buf_tag,
2212 						    fl->ifl_sds.ifsd_map[j]);
2213 						bus_dmamap_destroy(
2214 						    fl->ifl_buf_tag,
2215 						    fl->ifl_sds.ifsd_map[j]);
2216 					}
2217 				}
2218 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2219 				fl->ifl_buf_tag = NULL;
2220 			}
2221 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2222 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2223 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2224 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2225 			fl->ifl_sds.ifsd_m = NULL;
2226 			fl->ifl_sds.ifsd_cl = NULL;
2227 			fl->ifl_sds.ifsd_ba = NULL;
2228 			fl->ifl_sds.ifsd_map = NULL;
2229 		}
2230 		free(rxq->ifr_fl, M_IFLIB);
2231 		rxq->ifr_fl = NULL;
2232 		free(rxq->ifr_ifdi, M_IFLIB);
2233 		rxq->ifr_ifdi = NULL;
2234 		rxq->ifr_cq_cidx = 0;
2235 	}
2236 }
2237 
2238 /*
2239  * Timer routine
2240  */
2241 static void
2242 iflib_timer(void *arg)
2243 {
2244 	iflib_txq_t txq = arg;
2245 	if_ctx_t ctx = txq->ift_ctx;
2246 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2247 	uint64_t this_tick = ticks;
2248 	uint32_t reset_on = hz / 2;
2249 
2250 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2251 		return;
2252 
2253 	/*
2254 	** Check on the state of the TX queue(s), this
2255 	** can be done without the lock because its RO
2256 	** and the HUNG state will be static if set.
2257 	*/
2258 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2259 		txq->ift_last_timer_tick = this_tick;
2260 		IFDI_TIMER(ctx, txq->ift_id);
2261 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2262 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2263 		     (sctx->isc_pause_frames == 0)))
2264 			goto hung;
2265 
2266 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2267 		    ifmp_ring_is_stalled(txq->ift_br)) {
2268 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2269 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2270 		}
2271 		txq->ift_cleaned_prev = txq->ift_cleaned;
2272 	}
2273 #ifdef DEV_NETMAP
2274 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2275 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2276 #endif
2277 	/* handle any laggards */
2278 	if (txq->ift_db_pending)
2279 		GROUPTASK_ENQUEUE(&txq->ift_task);
2280 
2281 	sctx->isc_pause_frames = 0;
2282 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2283 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2284 	return;
2285 
2286  hung:
2287 	device_printf(ctx->ifc_dev,
2288 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2289 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2290 	STATE_LOCK(ctx);
2291 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2292 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2293 	iflib_admin_intr_deferred(ctx);
2294 	STATE_UNLOCK(ctx);
2295 }
2296 
2297 static void
2298 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2299 {
2300 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2301 
2302 	/*
2303 	 * XXX don't set the max_frame_size to larger
2304 	 * than the hardware can handle
2305 	 */
2306 	if (sctx->isc_max_frame_size <= MCLBYTES)
2307 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2308 	else
2309 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2310 }
2311 
2312 uint32_t
2313 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2314 {
2315 
2316 	return (ctx->ifc_rx_mbuf_sz);
2317 }
2318 
2319 static void
2320 iflib_init_locked(if_ctx_t ctx)
2321 {
2322 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2323 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2324 	if_t ifp = ctx->ifc_ifp;
2325 	iflib_fl_t fl;
2326 	iflib_txq_t txq;
2327 	iflib_rxq_t rxq;
2328 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2329 
2330 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2331 	IFDI_INTR_DISABLE(ctx);
2332 
2333 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2334 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2335 	/* Set hardware offload abilities */
2336 	if_clearhwassist(ifp);
2337 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2338 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2339 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2340 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2341 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2342 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2343 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2344 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2345 
2346 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2347 		CALLOUT_LOCK(txq);
2348 		callout_stop(&txq->ift_timer);
2349 		CALLOUT_UNLOCK(txq);
2350 		iflib_netmap_txq_init(ctx, txq);
2351 	}
2352 
2353 	/*
2354 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2355 	 * that drivers can use the value when setting up the hardware receive
2356 	 * buffers.
2357 	 */
2358 	iflib_calc_rx_mbuf_sz(ctx);
2359 
2360 #ifdef INVARIANTS
2361 	i = if_getdrvflags(ifp);
2362 #endif
2363 	IFDI_INIT(ctx);
2364 	MPASS(if_getdrvflags(ifp) == i);
2365 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2366 		/* XXX this should really be done on a per-queue basis */
2367 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2368 			MPASS(rxq->ifr_id == i);
2369 			iflib_netmap_rxq_init(ctx, rxq);
2370 			continue;
2371 		}
2372 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2373 			if (iflib_fl_setup(fl)) {
2374 				device_printf(ctx->ifc_dev,
2375 				    "setting up free list %d failed - "
2376 				    "check cluster settings\n", j);
2377 				goto done;
2378 			}
2379 		}
2380 	}
2381 done:
2382 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2383 	IFDI_INTR_ENABLE(ctx);
2384 	txq = ctx->ifc_txqs;
2385 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2386 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2387 			txq->ift_timer.c_cpu);
2388 }
2389 
2390 static int
2391 iflib_media_change(if_t ifp)
2392 {
2393 	if_ctx_t ctx = if_getsoftc(ifp);
2394 	int err;
2395 
2396 	CTX_LOCK(ctx);
2397 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2398 		iflib_init_locked(ctx);
2399 	CTX_UNLOCK(ctx);
2400 	return (err);
2401 }
2402 
2403 static void
2404 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2405 {
2406 	if_ctx_t ctx = if_getsoftc(ifp);
2407 
2408 	CTX_LOCK(ctx);
2409 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2410 	IFDI_MEDIA_STATUS(ctx, ifmr);
2411 	CTX_UNLOCK(ctx);
2412 }
2413 
2414 void
2415 iflib_stop(if_ctx_t ctx)
2416 {
2417 	iflib_txq_t txq = ctx->ifc_txqs;
2418 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2419 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2420 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2421 	iflib_dma_info_t di;
2422 	iflib_fl_t fl;
2423 	int i, j;
2424 
2425 	/* Tell the stack that the interface is no longer active */
2426 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2427 
2428 	IFDI_INTR_DISABLE(ctx);
2429 	DELAY(1000);
2430 	IFDI_STOP(ctx);
2431 	DELAY(1000);
2432 
2433 	iflib_debug_reset();
2434 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2435 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2436 		/* make sure all transmitters have completed before proceeding XXX */
2437 
2438 		CALLOUT_LOCK(txq);
2439 		callout_stop(&txq->ift_timer);
2440 		CALLOUT_UNLOCK(txq);
2441 
2442 		/* clean any enqueued buffers */
2443 		iflib_ifmp_purge(txq);
2444 		/* Free any existing tx buffers. */
2445 		for (j = 0; j < txq->ift_size; j++) {
2446 			iflib_txsd_free(ctx, txq, j);
2447 		}
2448 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2449 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2450 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2451 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2452 		txq->ift_pullups = 0;
2453 		ifmp_ring_reset_stats(txq->ift_br);
2454 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2455 			bzero((void *)di->idi_vaddr, di->idi_size);
2456 	}
2457 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2458 		/* make sure all transmitters have completed before proceeding XXX */
2459 
2460 		rxq->ifr_cq_cidx = 0;
2461 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2462 			bzero((void *)di->idi_vaddr, di->idi_size);
2463 		/* also resets the free lists pidx/cidx */
2464 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2465 			iflib_fl_bufs_free(fl);
2466 	}
2467 }
2468 
2469 static inline caddr_t
2470 calc_next_rxd(iflib_fl_t fl, int cidx)
2471 {
2472 	qidx_t size;
2473 	int nrxd;
2474 	caddr_t start, end, cur, next;
2475 
2476 	nrxd = fl->ifl_size;
2477 	size = fl->ifl_rxd_size;
2478 	start = fl->ifl_ifdi->idi_vaddr;
2479 
2480 	if (__predict_false(size == 0))
2481 		return (start);
2482 	cur = start + size*cidx;
2483 	end = start + size*nrxd;
2484 	next = CACHE_PTR_NEXT(cur);
2485 	return (next < end ? next : start);
2486 }
2487 
2488 static inline void
2489 prefetch_pkts(iflib_fl_t fl, int cidx)
2490 {
2491 	int nextptr;
2492 	int nrxd = fl->ifl_size;
2493 	caddr_t next_rxd;
2494 
2495 
2496 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2497 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2498 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2499 	next_rxd = calc_next_rxd(fl, cidx);
2500 	prefetch(next_rxd);
2501 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2502 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2503 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2504 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2505 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2506 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2507 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2508 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2509 }
2510 
2511 static struct mbuf *
2512 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2513     int *pf_rv, if_rxd_info_t ri)
2514 {
2515 	bus_dmamap_t map;
2516 	iflib_fl_t fl;
2517 	caddr_t payload;
2518 	struct mbuf *m;
2519 	int flid, cidx, len, next;
2520 
2521 	map = NULL;
2522 	flid = irf->irf_flid;
2523 	cidx = irf->irf_idx;
2524 	fl = &rxq->ifr_fl[flid];
2525 	sd->ifsd_fl = fl;
2526 	sd->ifsd_cidx = cidx;
2527 	m = fl->ifl_sds.ifsd_m[cidx];
2528 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2529 	fl->ifl_credits--;
2530 #if MEMORY_LOGGING
2531 	fl->ifl_m_dequeued++;
2532 #endif
2533 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2534 		prefetch_pkts(fl, cidx);
2535 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2536 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2537 	map = fl->ifl_sds.ifsd_map[cidx];
2538 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2539 
2540 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2541 	MPASS(fl->ifl_cidx == cidx);
2542 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2543 
2544 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2545 		payload  = *sd->ifsd_cl;
2546 		payload +=  ri->iri_pad;
2547 		len = ri->iri_len - ri->iri_pad;
2548 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2549 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2550 		switch (*pf_rv) {
2551 		case PFIL_DROPPED:
2552 		case PFIL_CONSUMED:
2553 			/*
2554 			 * The filter ate it.  Everything is recycled.
2555 			 */
2556 			m = NULL;
2557 			unload = 0;
2558 			break;
2559 		case PFIL_REALLOCED:
2560 			/*
2561 			 * The filter copied it.  Everything is recycled.
2562 			 */
2563 			m = pfil_mem2mbuf(payload);
2564 			unload = 0;
2565 			break;
2566 		case PFIL_PASS:
2567 			/*
2568 			 * Filter said it was OK, so receive like
2569 			 * normal
2570 			 */
2571 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2572 			break;
2573 		default:
2574 			MPASS(0);
2575 		}
2576 	} else {
2577 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2578 		*pf_rv = PFIL_PASS;
2579 	}
2580 
2581 	if (unload)
2582 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2583 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2584 	if (__predict_false(fl->ifl_cidx == 0))
2585 		fl->ifl_gen = 0;
2586 	bit_clear(fl->ifl_rx_bitmap, cidx);
2587 	return (m);
2588 }
2589 
2590 static struct mbuf *
2591 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2592 {
2593 	struct mbuf *m, *mh, *mt;
2594 	caddr_t cl;
2595 	int  *pf_rv_ptr, flags, i, padlen;
2596 	bool consumed;
2597 
2598 	i = 0;
2599 	mh = NULL;
2600 	consumed = false;
2601 	*pf_rv = PFIL_PASS;
2602 	pf_rv_ptr = pf_rv;
2603 	do {
2604 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2605 		    pf_rv_ptr, ri);
2606 
2607 		MPASS(*sd->ifsd_cl != NULL);
2608 
2609 		/*
2610 		 * Exclude zero-length frags & frags from
2611 		 * packets the filter has consumed or dropped
2612 		 */
2613 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2614 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2615 			if (mh == NULL) {
2616 				/* everything saved here */
2617 				consumed = true;
2618 				pf_rv_ptr = NULL;
2619 				continue;
2620 			}
2621 			/* XXX we can save the cluster here, but not the mbuf */
2622 			m_init(m, M_NOWAIT, MT_DATA, 0);
2623 			m_free(m);
2624 			continue;
2625 		}
2626 		if (mh == NULL) {
2627 			flags = M_PKTHDR|M_EXT;
2628 			mh = mt = m;
2629 			padlen = ri->iri_pad;
2630 		} else {
2631 			flags = M_EXT;
2632 			mt->m_next = m;
2633 			mt = m;
2634 			/* assuming padding is only on the first fragment */
2635 			padlen = 0;
2636 		}
2637 		cl = *sd->ifsd_cl;
2638 		*sd->ifsd_cl = NULL;
2639 
2640 		/* Can these two be made one ? */
2641 		m_init(m, M_NOWAIT, MT_DATA, flags);
2642 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2643 		/*
2644 		 * These must follow m_init and m_cljset
2645 		 */
2646 		m->m_data += padlen;
2647 		ri->iri_len -= padlen;
2648 		m->m_len = ri->iri_frags[i].irf_len;
2649 	} while (++i < ri->iri_nfrags);
2650 
2651 	return (mh);
2652 }
2653 
2654 /*
2655  * Process one software descriptor
2656  */
2657 static struct mbuf *
2658 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2659 {
2660 	struct if_rxsd sd;
2661 	struct mbuf *m;
2662 	int pf_rv;
2663 
2664 	/* should I merge this back in now that the two paths are basically duplicated? */
2665 	if (ri->iri_nfrags == 1 &&
2666 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2667 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2668 		    &pf_rv, ri);
2669 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2670 			return (m);
2671 		if (pf_rv == PFIL_PASS) {
2672 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2673 #ifndef __NO_STRICT_ALIGNMENT
2674 			if (!IP_ALIGNED(m))
2675 				m->m_data += 2;
2676 #endif
2677 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2678 			m->m_len = ri->iri_frags[0].irf_len;
2679 		}
2680 	} else {
2681 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2682 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2683 			return (m);
2684 	}
2685 	m->m_pkthdr.len = ri->iri_len;
2686 	m->m_pkthdr.rcvif = ri->iri_ifp;
2687 	m->m_flags |= ri->iri_flags;
2688 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2689 	m->m_pkthdr.flowid = ri->iri_flowid;
2690 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2691 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2692 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2693 	return (m);
2694 }
2695 
2696 #if defined(INET6) || defined(INET)
2697 static void
2698 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2699 {
2700 	CURVNET_SET(lc->ifp->if_vnet);
2701 #if defined(INET6)
2702 	*v6 = V_ip6_forwarding;
2703 #endif
2704 #if defined(INET)
2705 	*v4 = V_ipforwarding;
2706 #endif
2707 	CURVNET_RESTORE();
2708 }
2709 
2710 /*
2711  * Returns true if it's possible this packet could be LROed.
2712  * if it returns false, it is guaranteed that tcp_lro_rx()
2713  * would not return zero.
2714  */
2715 static bool
2716 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2717 {
2718 	struct ether_header *eh;
2719 
2720 	eh = mtod(m, struct ether_header *);
2721 	switch (eh->ether_type) {
2722 #if defined(INET6)
2723 		case htons(ETHERTYPE_IPV6):
2724 			return (!v6_forwarding);
2725 #endif
2726 #if defined (INET)
2727 		case htons(ETHERTYPE_IP):
2728 			return (!v4_forwarding);
2729 #endif
2730 	}
2731 
2732 	return false;
2733 }
2734 #else
2735 static void
2736 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2737 {
2738 }
2739 #endif
2740 
2741 static bool
2742 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2743 {
2744 	if_t ifp;
2745 	if_ctx_t ctx = rxq->ifr_ctx;
2746 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2747 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2748 	int avail, i;
2749 	qidx_t *cidxp;
2750 	struct if_rxd_info ri;
2751 	int err, budget_left, rx_bytes, rx_pkts;
2752 	iflib_fl_t fl;
2753 	int lro_enabled;
2754 	bool v4_forwarding, v6_forwarding, lro_possible;
2755 
2756 	/*
2757 	 * XXX early demux data packets so that if_input processing only handles
2758 	 * acks in interrupt context
2759 	 */
2760 	struct mbuf *m, *mh, *mt, *mf;
2761 
2762 	lro_possible = v4_forwarding = v6_forwarding = false;
2763 	ifp = ctx->ifc_ifp;
2764 	mh = mt = NULL;
2765 	MPASS(budget > 0);
2766 	rx_pkts	= rx_bytes = 0;
2767 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2768 		cidxp = &rxq->ifr_cq_cidx;
2769 	else
2770 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2771 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2772 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2773 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2774 		DBG_COUNTER_INC(rx_unavail);
2775 		return (false);
2776 	}
2777 
2778 	/* pfil needs the vnet to be set */
2779 	CURVNET_SET_QUIET(ifp->if_vnet);
2780 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2781 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2782 			DBG_COUNTER_INC(rx_ctx_inactive);
2783 			break;
2784 		}
2785 		/*
2786 		 * Reset client set fields to their default values
2787 		 */
2788 		rxd_info_zero(&ri);
2789 		ri.iri_qsidx = rxq->ifr_id;
2790 		ri.iri_cidx = *cidxp;
2791 		ri.iri_ifp = ifp;
2792 		ri.iri_frags = rxq->ifr_frags;
2793 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2794 
2795 		if (err)
2796 			goto err;
2797 		rx_pkts += 1;
2798 		rx_bytes += ri.iri_len;
2799 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2800 			*cidxp = ri.iri_cidx;
2801 			/* Update our consumer index */
2802 			/* XXX NB: shurd - check if this is still safe */
2803 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2804 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2805 			/* was this only a completion queue message? */
2806 			if (__predict_false(ri.iri_nfrags == 0))
2807 				continue;
2808 		}
2809 		MPASS(ri.iri_nfrags != 0);
2810 		MPASS(ri.iri_len != 0);
2811 
2812 		/* will advance the cidx on the corresponding free lists */
2813 		m = iflib_rxd_pkt_get(rxq, &ri);
2814 		avail--;
2815 		budget_left--;
2816 		if (avail == 0 && budget_left)
2817 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2818 
2819 		if (__predict_false(m == NULL))
2820 			continue;
2821 
2822 		/* imm_pkt: -- cxgb */
2823 		if (mh == NULL)
2824 			mh = mt = m;
2825 		else {
2826 			mt->m_nextpkt = m;
2827 			mt = m;
2828 		}
2829 	}
2830 	CURVNET_RESTORE();
2831 	/* make sure that we can refill faster than drain */
2832 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2833 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2834 
2835 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2836 	if (lro_enabled)
2837 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2838 	mt = mf = NULL;
2839 	while (mh != NULL) {
2840 		m = mh;
2841 		mh = mh->m_nextpkt;
2842 		m->m_nextpkt = NULL;
2843 #ifndef __NO_STRICT_ALIGNMENT
2844 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2845 			continue;
2846 #endif
2847 		rx_bytes += m->m_pkthdr.len;
2848 		rx_pkts++;
2849 #if defined(INET6) || defined(INET)
2850 		if (lro_enabled) {
2851 			if (!lro_possible) {
2852 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2853 				if (lro_possible && mf != NULL) {
2854 					ifp->if_input(ifp, mf);
2855 					DBG_COUNTER_INC(rx_if_input);
2856 					mt = mf = NULL;
2857 				}
2858 			}
2859 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2860 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2861 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2862 					continue;
2863 			}
2864 		}
2865 #endif
2866 		if (lro_possible) {
2867 			ifp->if_input(ifp, m);
2868 			DBG_COUNTER_INC(rx_if_input);
2869 			continue;
2870 		}
2871 
2872 		if (mf == NULL)
2873 			mf = m;
2874 		if (mt != NULL)
2875 			mt->m_nextpkt = m;
2876 		mt = m;
2877 	}
2878 	if (mf != NULL) {
2879 		ifp->if_input(ifp, mf);
2880 		DBG_COUNTER_INC(rx_if_input);
2881 	}
2882 
2883 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2884 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2885 
2886 	/*
2887 	 * Flush any outstanding LRO work
2888 	 */
2889 #if defined(INET6) || defined(INET)
2890 	tcp_lro_flush_all(&rxq->ifr_lc);
2891 #endif
2892 	if (avail)
2893 		return true;
2894 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2895 err:
2896 	STATE_LOCK(ctx);
2897 	ctx->ifc_flags |= IFC_DO_RESET;
2898 	iflib_admin_intr_deferred(ctx);
2899 	STATE_UNLOCK(ctx);
2900 	return (false);
2901 }
2902 
2903 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2904 static inline qidx_t
2905 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2906 {
2907 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2908 	qidx_t minthresh = txq->ift_size / 8;
2909 	if (in_use > 4*minthresh)
2910 		return (notify_count);
2911 	if (in_use > 2*minthresh)
2912 		return (notify_count >> 1);
2913 	if (in_use > minthresh)
2914 		return (notify_count >> 3);
2915 	return (0);
2916 }
2917 
2918 static inline qidx_t
2919 txq_max_rs_deferred(iflib_txq_t txq)
2920 {
2921 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2922 	qidx_t minthresh = txq->ift_size / 8;
2923 	if (txq->ift_in_use > 4*minthresh)
2924 		return (notify_count);
2925 	if (txq->ift_in_use > 2*minthresh)
2926 		return (notify_count >> 1);
2927 	if (txq->ift_in_use > minthresh)
2928 		return (notify_count >> 2);
2929 	return (2);
2930 }
2931 
2932 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2933 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2934 
2935 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2936 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2937 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2938 
2939 /* forward compatibility for cxgb */
2940 #define FIRST_QSET(ctx) 0
2941 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2942 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2943 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2944 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2945 
2946 /* XXX we should be setting this to something other than zero */
2947 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2948 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2949     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2950 
2951 static inline bool
2952 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2953 {
2954 	qidx_t dbval, max;
2955 	bool rang;
2956 
2957 	rang = false;
2958 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2959 	if (ring || txq->ift_db_pending >= max) {
2960 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2961 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2962 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2963 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2964 		txq->ift_db_pending = txq->ift_npending = 0;
2965 		rang = true;
2966 	}
2967 	return (rang);
2968 }
2969 
2970 #ifdef PKT_DEBUG
2971 static void
2972 print_pkt(if_pkt_info_t pi)
2973 {
2974 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2975 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2976 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2977 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2978 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2979 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2980 }
2981 #endif
2982 
2983 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2984 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2985 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2986 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2987 
2988 static int
2989 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2990 {
2991 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2992 	struct ether_vlan_header *eh;
2993 	struct mbuf *m;
2994 
2995 	m = *mp;
2996 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2997 	    M_WRITABLE(m) == 0) {
2998 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2999 			return (ENOMEM);
3000 		} else {
3001 			m_freem(*mp);
3002 			DBG_COUNTER_INC(tx_frees);
3003 			*mp = m;
3004 		}
3005 	}
3006 
3007 	/*
3008 	 * Determine where frame payload starts.
3009 	 * Jump over vlan headers if already present,
3010 	 * helpful for QinQ too.
3011 	 */
3012 	if (__predict_false(m->m_len < sizeof(*eh))) {
3013 		txq->ift_pullups++;
3014 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3015 			return (ENOMEM);
3016 	}
3017 	eh = mtod(m, struct ether_vlan_header *);
3018 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3019 		pi->ipi_etype = ntohs(eh->evl_proto);
3020 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3021 	} else {
3022 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3023 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3024 	}
3025 
3026 	switch (pi->ipi_etype) {
3027 #ifdef INET
3028 	case ETHERTYPE_IP:
3029 	{
3030 		struct mbuf *n;
3031 		struct ip *ip = NULL;
3032 		struct tcphdr *th = NULL;
3033 		int minthlen;
3034 
3035 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3036 		if (__predict_false(m->m_len < minthlen)) {
3037 			/*
3038 			 * if this code bloat is causing too much of a hit
3039 			 * move it to a separate function and mark it noinline
3040 			 */
3041 			if (m->m_len == pi->ipi_ehdrlen) {
3042 				n = m->m_next;
3043 				MPASS(n);
3044 				if (n->m_len >= sizeof(*ip))  {
3045 					ip = (struct ip *)n->m_data;
3046 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3047 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3048 				} else {
3049 					txq->ift_pullups++;
3050 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3051 						return (ENOMEM);
3052 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3053 				}
3054 			} else {
3055 				txq->ift_pullups++;
3056 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3057 					return (ENOMEM);
3058 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3059 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3060 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3061 			}
3062 		} else {
3063 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3064 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3065 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3066 		}
3067 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3068 		pi->ipi_ipproto = ip->ip_p;
3069 		pi->ipi_flags |= IPI_TX_IPV4;
3070 
3071 		/* TCP checksum offload may require TCP header length */
3072 		if (IS_TX_OFFLOAD4(pi)) {
3073 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3074 				if (__predict_false(th == NULL)) {
3075 					txq->ift_pullups++;
3076 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3077 						return (ENOMEM);
3078 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3079 				}
3080 				pi->ipi_tcp_hflags = th->th_flags;
3081 				pi->ipi_tcp_hlen = th->th_off << 2;
3082 				pi->ipi_tcp_seq = th->th_seq;
3083 			}
3084 			if (IS_TSO4(pi)) {
3085 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3086 					return (ENXIO);
3087 				/*
3088 				 * TSO always requires hardware checksum offload.
3089 				 */
3090 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3091 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3092 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3093 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3094 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3095 					ip->ip_sum = 0;
3096 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3097 				}
3098 			}
3099 		}
3100 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3101                        ip->ip_sum = 0;
3102 
3103 		break;
3104 	}
3105 #endif
3106 #ifdef INET6
3107 	case ETHERTYPE_IPV6:
3108 	{
3109 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3110 		struct tcphdr *th;
3111 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3112 
3113 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3114 			txq->ift_pullups++;
3115 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3116 				return (ENOMEM);
3117 		}
3118 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3119 
3120 		/* XXX-BZ this will go badly in case of ext hdrs. */
3121 		pi->ipi_ipproto = ip6->ip6_nxt;
3122 		pi->ipi_flags |= IPI_TX_IPV6;
3123 
3124 		/* TCP checksum offload may require TCP header length */
3125 		if (IS_TX_OFFLOAD6(pi)) {
3126 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3127 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3128 					txq->ift_pullups++;
3129 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3130 						return (ENOMEM);
3131 				}
3132 				pi->ipi_tcp_hflags = th->th_flags;
3133 				pi->ipi_tcp_hlen = th->th_off << 2;
3134 				pi->ipi_tcp_seq = th->th_seq;
3135 			}
3136 			if (IS_TSO6(pi)) {
3137 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3138 					return (ENXIO);
3139 				/*
3140 				 * TSO always requires hardware checksum offload.
3141 				 */
3142 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3143 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3144 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3145 			}
3146 		}
3147 		break;
3148 	}
3149 #endif
3150 	default:
3151 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3152 		pi->ipi_ip_hlen = 0;
3153 		break;
3154 	}
3155 	*mp = m;
3156 
3157 	return (0);
3158 }
3159 
3160 /*
3161  * If dodgy hardware rejects the scatter gather chain we've handed it
3162  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3163  * m_defrag'd mbufs
3164  */
3165 static __noinline struct mbuf *
3166 iflib_remove_mbuf(iflib_txq_t txq)
3167 {
3168 	int ntxd, pidx;
3169 	struct mbuf *m, **ifsd_m;
3170 
3171 	ifsd_m = txq->ift_sds.ifsd_m;
3172 	ntxd = txq->ift_size;
3173 	pidx = txq->ift_pidx & (ntxd - 1);
3174 	ifsd_m = txq->ift_sds.ifsd_m;
3175 	m = ifsd_m[pidx];
3176 	ifsd_m[pidx] = NULL;
3177 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3178 	if (txq->ift_sds.ifsd_tso_map != NULL)
3179 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3180 		    txq->ift_sds.ifsd_tso_map[pidx]);
3181 #if MEMORY_LOGGING
3182 	txq->ift_dequeued++;
3183 #endif
3184 	return (m);
3185 }
3186 
3187 static inline caddr_t
3188 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3189 {
3190 	qidx_t size;
3191 	int ntxd;
3192 	caddr_t start, end, cur, next;
3193 
3194 	ntxd = txq->ift_size;
3195 	size = txq->ift_txd_size[qid];
3196 	start = txq->ift_ifdi[qid].idi_vaddr;
3197 
3198 	if (__predict_false(size == 0))
3199 		return (start);
3200 	cur = start + size*cidx;
3201 	end = start + size*ntxd;
3202 	next = CACHE_PTR_NEXT(cur);
3203 	return (next < end ? next : start);
3204 }
3205 
3206 /*
3207  * Pad an mbuf to ensure a minimum ethernet frame size.
3208  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3209  */
3210 static __noinline int
3211 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3212 {
3213 	/*
3214 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3215 	 * and ARP message is the smallest common payload I can think of
3216 	 */
3217 	static char pad[18];	/* just zeros */
3218 	int n;
3219 	struct mbuf *new_head;
3220 
3221 	if (!M_WRITABLE(*m_head)) {
3222 		new_head = m_dup(*m_head, M_NOWAIT);
3223 		if (new_head == NULL) {
3224 			m_freem(*m_head);
3225 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3226 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3227 			DBG_COUNTER_INC(tx_frees);
3228 			return ENOMEM;
3229 		}
3230 		m_freem(*m_head);
3231 		*m_head = new_head;
3232 	}
3233 
3234 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3235 	     n > 0; n -= sizeof(pad))
3236 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3237 			break;
3238 
3239 	if (n > 0) {
3240 		m_freem(*m_head);
3241 		device_printf(dev, "cannot pad short frame\n");
3242 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3243 		DBG_COUNTER_INC(tx_frees);
3244 		return (ENOBUFS);
3245 	}
3246 
3247 	return 0;
3248 }
3249 
3250 static int
3251 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3252 {
3253 	if_ctx_t		ctx;
3254 	if_shared_ctx_t		sctx;
3255 	if_softc_ctx_t		scctx;
3256 	bus_dma_tag_t		buf_tag;
3257 	bus_dma_segment_t	*segs;
3258 	struct mbuf		*m_head, **ifsd_m;
3259 	void			*next_txd;
3260 	bus_dmamap_t		map;
3261 	struct if_pkt_info	pi;
3262 	int remap = 0;
3263 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3264 
3265 	ctx = txq->ift_ctx;
3266 	sctx = ctx->ifc_sctx;
3267 	scctx = &ctx->ifc_softc_ctx;
3268 	segs = txq->ift_segs;
3269 	ntxd = txq->ift_size;
3270 	m_head = *m_headp;
3271 	map = NULL;
3272 
3273 	/*
3274 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3275 	 */
3276 	cidx = txq->ift_cidx;
3277 	pidx = txq->ift_pidx;
3278 	if (ctx->ifc_flags & IFC_PREFETCH) {
3279 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3280 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3281 			next_txd = calc_next_txd(txq, cidx, 0);
3282 			prefetch(next_txd);
3283 		}
3284 
3285 		/* prefetch the next cache line of mbuf pointers and flags */
3286 		prefetch(&txq->ift_sds.ifsd_m[next]);
3287 		prefetch(&txq->ift_sds.ifsd_map[next]);
3288 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3289 	}
3290 	map = txq->ift_sds.ifsd_map[pidx];
3291 	ifsd_m = txq->ift_sds.ifsd_m;
3292 
3293 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3294 		buf_tag = txq->ift_tso_buf_tag;
3295 		max_segs = scctx->isc_tx_tso_segments_max;
3296 		map = txq->ift_sds.ifsd_tso_map[pidx];
3297 		MPASS(buf_tag != NULL);
3298 		MPASS(max_segs > 0);
3299 	} else {
3300 		buf_tag = txq->ift_buf_tag;
3301 		max_segs = scctx->isc_tx_nsegments;
3302 		map = txq->ift_sds.ifsd_map[pidx];
3303 	}
3304 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3305 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3306 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3307 		if (err) {
3308 			DBG_COUNTER_INC(encap_txd_encap_fail);
3309 			return err;
3310 		}
3311 	}
3312 	m_head = *m_headp;
3313 
3314 	pkt_info_zero(&pi);
3315 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3316 	pi.ipi_pidx = pidx;
3317 	pi.ipi_qsidx = txq->ift_id;
3318 	pi.ipi_len = m_head->m_pkthdr.len;
3319 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3320 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3321 
3322 	/* deliberate bitwise OR to make one condition */
3323 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3324 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3325 			DBG_COUNTER_INC(encap_txd_encap_fail);
3326 			return (err);
3327 		}
3328 		m_head = *m_headp;
3329 	}
3330 
3331 retry:
3332 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3333 	    BUS_DMA_NOWAIT);
3334 defrag:
3335 	if (__predict_false(err)) {
3336 		switch (err) {
3337 		case EFBIG:
3338 			/* try collapse once and defrag once */
3339 			if (remap == 0) {
3340 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3341 				/* try defrag if collapsing fails */
3342 				if (m_head == NULL)
3343 					remap++;
3344 			}
3345 			if (remap == 1) {
3346 				txq->ift_mbuf_defrag++;
3347 				m_head = m_defrag(*m_headp, M_NOWAIT);
3348 			}
3349 			/*
3350 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3351 			 * failed to map an mbuf that was run through m_defrag
3352 			 */
3353 			MPASS(remap <= 1);
3354 			if (__predict_false(m_head == NULL || remap > 1))
3355 				goto defrag_failed;
3356 			remap++;
3357 			*m_headp = m_head;
3358 			goto retry;
3359 			break;
3360 		case ENOMEM:
3361 			txq->ift_no_tx_dma_setup++;
3362 			break;
3363 		default:
3364 			txq->ift_no_tx_dma_setup++;
3365 			m_freem(*m_headp);
3366 			DBG_COUNTER_INC(tx_frees);
3367 			*m_headp = NULL;
3368 			break;
3369 		}
3370 		txq->ift_map_failed++;
3371 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3372 		DBG_COUNTER_INC(encap_txd_encap_fail);
3373 		return (err);
3374 	}
3375 	ifsd_m[pidx] = m_head;
3376 	/*
3377 	 * XXX assumes a 1 to 1 relationship between segments and
3378 	 *        descriptors - this does not hold true on all drivers, e.g.
3379 	 *        cxgb
3380 	 */
3381 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3382 		txq->ift_no_desc_avail++;
3383 		bus_dmamap_unload(buf_tag, map);
3384 		DBG_COUNTER_INC(encap_txq_avail_fail);
3385 		DBG_COUNTER_INC(encap_txd_encap_fail);
3386 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3387 			GROUPTASK_ENQUEUE(&txq->ift_task);
3388 		return (ENOBUFS);
3389 	}
3390 	/*
3391 	 * On Intel cards we can greatly reduce the number of TX interrupts
3392 	 * we see by only setting report status on every Nth descriptor.
3393 	 * However, this also means that the driver will need to keep track
3394 	 * of the descriptors that RS was set on to check them for the DD bit.
3395 	 */
3396 	txq->ift_rs_pending += nsegs + 1;
3397 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3398 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3399 		pi.ipi_flags |= IPI_TX_INTR;
3400 		txq->ift_rs_pending = 0;
3401 	}
3402 
3403 	pi.ipi_segs = segs;
3404 	pi.ipi_nsegs = nsegs;
3405 
3406 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3407 #ifdef PKT_DEBUG
3408 	print_pkt(&pi);
3409 #endif
3410 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3411 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3412 		DBG_COUNTER_INC(tx_encap);
3413 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3414 
3415 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3416 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3417 			ndesc += txq->ift_size;
3418 			txq->ift_gen = 1;
3419 		}
3420 		/*
3421 		 * drivers can need as many as
3422 		 * two sentinels
3423 		 */
3424 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3425 		MPASS(pi.ipi_new_pidx != pidx);
3426 		MPASS(ndesc > 0);
3427 		txq->ift_in_use += ndesc;
3428 
3429 		/*
3430 		 * We update the last software descriptor again here because there may
3431 		 * be a sentinel and/or there may be more mbufs than segments
3432 		 */
3433 		txq->ift_pidx = pi.ipi_new_pidx;
3434 		txq->ift_npending += pi.ipi_ndescs;
3435 	} else {
3436 		*m_headp = m_head = iflib_remove_mbuf(txq);
3437 		if (err == EFBIG) {
3438 			txq->ift_txd_encap_efbig++;
3439 			if (remap < 2) {
3440 				remap = 1;
3441 				goto defrag;
3442 			}
3443 		}
3444 		goto defrag_failed;
3445 	}
3446 	/*
3447 	 * err can't possibly be non-zero here, so we don't neet to test it
3448 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3449 	 */
3450 	return (err);
3451 
3452 defrag_failed:
3453 	txq->ift_mbuf_defrag_failed++;
3454 	txq->ift_map_failed++;
3455 	m_freem(*m_headp);
3456 	DBG_COUNTER_INC(tx_frees);
3457 	*m_headp = NULL;
3458 	DBG_COUNTER_INC(encap_txd_encap_fail);
3459 	return (ENOMEM);
3460 }
3461 
3462 static void
3463 iflib_tx_desc_free(iflib_txq_t txq, int n)
3464 {
3465 	uint32_t qsize, cidx, mask, gen;
3466 	struct mbuf *m, **ifsd_m;
3467 	bool do_prefetch;
3468 
3469 	cidx = txq->ift_cidx;
3470 	gen = txq->ift_gen;
3471 	qsize = txq->ift_size;
3472 	mask = qsize-1;
3473 	ifsd_m = txq->ift_sds.ifsd_m;
3474 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3475 
3476 	while (n-- > 0) {
3477 		if (do_prefetch) {
3478 			prefetch(ifsd_m[(cidx + 3) & mask]);
3479 			prefetch(ifsd_m[(cidx + 4) & mask]);
3480 		}
3481 		if ((m = ifsd_m[cidx]) != NULL) {
3482 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3483 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3484 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3485 				    txq->ift_sds.ifsd_tso_map[cidx],
3486 				    BUS_DMASYNC_POSTWRITE);
3487 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3488 				    txq->ift_sds.ifsd_tso_map[cidx]);
3489 			} else {
3490 				bus_dmamap_sync(txq->ift_buf_tag,
3491 				    txq->ift_sds.ifsd_map[cidx],
3492 				    BUS_DMASYNC_POSTWRITE);
3493 				bus_dmamap_unload(txq->ift_buf_tag,
3494 				    txq->ift_sds.ifsd_map[cidx]);
3495 			}
3496 			/* XXX we don't support any drivers that batch packets yet */
3497 			MPASS(m->m_nextpkt == NULL);
3498 			m_freem(m);
3499 			ifsd_m[cidx] = NULL;
3500 #if MEMORY_LOGGING
3501 			txq->ift_dequeued++;
3502 #endif
3503 			DBG_COUNTER_INC(tx_frees);
3504 		}
3505 		if (__predict_false(++cidx == qsize)) {
3506 			cidx = 0;
3507 			gen = 0;
3508 		}
3509 	}
3510 	txq->ift_cidx = cidx;
3511 	txq->ift_gen = gen;
3512 }
3513 
3514 static __inline int
3515 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3516 {
3517 	int reclaim;
3518 	if_ctx_t ctx = txq->ift_ctx;
3519 
3520 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3521 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3522 
3523 	/*
3524 	 * Need a rate-limiting check so that this isn't called every time
3525 	 */
3526 	iflib_tx_credits_update(ctx, txq);
3527 	reclaim = DESC_RECLAIMABLE(txq);
3528 
3529 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3530 #ifdef INVARIANTS
3531 		if (iflib_verbose_debug) {
3532 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3533 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3534 			       reclaim, thresh);
3535 
3536 		}
3537 #endif
3538 		return (0);
3539 	}
3540 	iflib_tx_desc_free(txq, reclaim);
3541 	txq->ift_cleaned += reclaim;
3542 	txq->ift_in_use -= reclaim;
3543 
3544 	return (reclaim);
3545 }
3546 
3547 static struct mbuf **
3548 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3549 {
3550 	int next, size;
3551 	struct mbuf **items;
3552 
3553 	size = r->size;
3554 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3555 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3556 
3557 	prefetch(items[(cidx + offset) & (size-1)]);
3558 	if (remaining > 1) {
3559 		prefetch2cachelines(&items[next]);
3560 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3561 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3562 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3563 	}
3564 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3565 }
3566 
3567 static void
3568 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3569 {
3570 
3571 	ifmp_ring_check_drainage(txq->ift_br, budget);
3572 }
3573 
3574 static uint32_t
3575 iflib_txq_can_drain(struct ifmp_ring *r)
3576 {
3577 	iflib_txq_t txq = r->cookie;
3578 	if_ctx_t ctx = txq->ift_ctx;
3579 
3580 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3581 		return (1);
3582 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3583 	    BUS_DMASYNC_POSTREAD);
3584 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3585 	    false));
3586 }
3587 
3588 static uint32_t
3589 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3590 {
3591 	iflib_txq_t txq = r->cookie;
3592 	if_ctx_t ctx = txq->ift_ctx;
3593 	if_t ifp = ctx->ifc_ifp;
3594 	struct mbuf *m, **mp;
3595 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3596 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3597 	bool do_prefetch, rang, ring;
3598 
3599 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3600 			    !LINK_ACTIVE(ctx))) {
3601 		DBG_COUNTER_INC(txq_drain_notready);
3602 		return (0);
3603 	}
3604 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3605 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3606 	avail = IDXDIFF(pidx, cidx, r->size);
3607 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3608 		DBG_COUNTER_INC(txq_drain_flushing);
3609 		for (i = 0; i < avail; i++) {
3610 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3611 				m_free(r->items[(cidx + i) & (r->size-1)]);
3612 			r->items[(cidx + i) & (r->size-1)] = NULL;
3613 		}
3614 		return (avail);
3615 	}
3616 
3617 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3618 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3619 		CALLOUT_LOCK(txq);
3620 		callout_stop(&txq->ift_timer);
3621 		CALLOUT_UNLOCK(txq);
3622 		DBG_COUNTER_INC(txq_drain_oactive);
3623 		return (0);
3624 	}
3625 	if (reclaimed)
3626 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3627 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3628 	count = MIN(avail, TX_BATCH_SIZE);
3629 #ifdef INVARIANTS
3630 	if (iflib_verbose_debug)
3631 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3632 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3633 #endif
3634 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3635 	txq_avail = TXQ_AVAIL(txq);
3636 	err = 0;
3637 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3638 		int rem = do_prefetch ? count - i : 0;
3639 
3640 		mp = _ring_peek_one(r, cidx, i, rem);
3641 		MPASS(mp != NULL && *mp != NULL);
3642 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3643 			consumed++;
3644 			continue;
3645 		}
3646 		in_use_prev = txq->ift_in_use;
3647 		err = iflib_encap(txq, mp);
3648 		if (__predict_false(err)) {
3649 			/* no room - bail out */
3650 			if (err == ENOBUFS)
3651 				break;
3652 			consumed++;
3653 			/* we can't send this packet - skip it */
3654 			continue;
3655 		}
3656 		consumed++;
3657 		pkt_sent++;
3658 		m = *mp;
3659 		DBG_COUNTER_INC(tx_sent);
3660 		bytes_sent += m->m_pkthdr.len;
3661 		mcast_sent += !!(m->m_flags & M_MCAST);
3662 		txq_avail = TXQ_AVAIL(txq);
3663 
3664 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3665 		ETHER_BPF_MTAP(ifp, m);
3666 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3667 			break;
3668 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3669 	}
3670 
3671 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3672 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3673 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3674 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3675 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3676 	if (mcast_sent)
3677 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3678 #ifdef INVARIANTS
3679 	if (iflib_verbose_debug)
3680 		printf("consumed=%d\n", consumed);
3681 #endif
3682 	return (consumed);
3683 }
3684 
3685 static uint32_t
3686 iflib_txq_drain_always(struct ifmp_ring *r)
3687 {
3688 	return (1);
3689 }
3690 
3691 static uint32_t
3692 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3693 {
3694 	int i, avail;
3695 	struct mbuf **mp;
3696 	iflib_txq_t txq;
3697 
3698 	txq = r->cookie;
3699 
3700 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3701 	CALLOUT_LOCK(txq);
3702 	callout_stop(&txq->ift_timer);
3703 	CALLOUT_UNLOCK(txq);
3704 
3705 	avail = IDXDIFF(pidx, cidx, r->size);
3706 	for (i = 0; i < avail; i++) {
3707 		mp = _ring_peek_one(r, cidx, i, avail - i);
3708 		if (__predict_false(*mp == (struct mbuf *)txq))
3709 			continue;
3710 		m_freem(*mp);
3711 		DBG_COUNTER_INC(tx_frees);
3712 	}
3713 	MPASS(ifmp_ring_is_stalled(r) == 0);
3714 	return (avail);
3715 }
3716 
3717 static void
3718 iflib_ifmp_purge(iflib_txq_t txq)
3719 {
3720 	struct ifmp_ring *r;
3721 
3722 	r = txq->ift_br;
3723 	r->drain = iflib_txq_drain_free;
3724 	r->can_drain = iflib_txq_drain_always;
3725 
3726 	ifmp_ring_check_drainage(r, r->size);
3727 
3728 	r->drain = iflib_txq_drain;
3729 	r->can_drain = iflib_txq_can_drain;
3730 }
3731 
3732 static void
3733 _task_fn_tx(void *context)
3734 {
3735 	iflib_txq_t txq = context;
3736 	if_ctx_t ctx = txq->ift_ctx;
3737 #if defined(ALTQ) || defined(DEV_NETMAP)
3738 	if_t ifp = ctx->ifc_ifp;
3739 #endif
3740 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3741 
3742 #ifdef IFLIB_DIAGNOSTICS
3743 	txq->ift_cpu_exec_count[curcpu]++;
3744 #endif
3745 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3746 		return;
3747 #ifdef DEV_NETMAP
3748 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3749 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3750 		    BUS_DMASYNC_POSTREAD);
3751 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3752 			netmap_tx_irq(ifp, txq->ift_id);
3753 		if (ctx->ifc_flags & IFC_LEGACY)
3754 			IFDI_INTR_ENABLE(ctx);
3755 		else
3756 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3757 		return;
3758 	}
3759 #endif
3760 #ifdef ALTQ
3761 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3762 		iflib_altq_if_start(ifp);
3763 #endif
3764 	if (txq->ift_db_pending)
3765 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3766 	else if (!abdicate)
3767 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3768 	/*
3769 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3770 	 */
3771 	if (abdicate)
3772 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3773 	if (ctx->ifc_flags & IFC_LEGACY)
3774 		IFDI_INTR_ENABLE(ctx);
3775 	else
3776 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3777 }
3778 
3779 static void
3780 _task_fn_rx(void *context)
3781 {
3782 	iflib_rxq_t rxq = context;
3783 	if_ctx_t ctx = rxq->ifr_ctx;
3784 	bool more;
3785 	uint16_t budget;
3786 
3787 #ifdef IFLIB_DIAGNOSTICS
3788 	rxq->ifr_cpu_exec_count[curcpu]++;
3789 #endif
3790 	DBG_COUNTER_INC(task_fn_rxs);
3791 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3792 		return;
3793 	more = true;
3794 #ifdef DEV_NETMAP
3795 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3796 		u_int work = 0;
3797 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3798 			more = false;
3799 		}
3800 	}
3801 #endif
3802 	budget = ctx->ifc_sysctl_rx_budget;
3803 	if (budget == 0)
3804 		budget = 16;	/* XXX */
3805 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3806 		if (ctx->ifc_flags & IFC_LEGACY)
3807 			IFDI_INTR_ENABLE(ctx);
3808 		else
3809 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3810 		DBG_COUNTER_INC(rx_intr_enables);
3811 	}
3812 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3813 		return;
3814 	if (more)
3815 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3816 }
3817 
3818 static void
3819 _task_fn_admin(void *context)
3820 {
3821 	if_ctx_t ctx = context;
3822 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3823 	iflib_txq_t txq;
3824 	int i;
3825 	bool oactive, running, do_reset, do_watchdog, in_detach;
3826 	uint32_t reset_on = hz / 2;
3827 
3828 	STATE_LOCK(ctx);
3829 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3830 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3831 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3832 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3833 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3834 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3835 	STATE_UNLOCK(ctx);
3836 
3837 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3838 		return;
3839 	if (in_detach)
3840 		return;
3841 
3842 	CTX_LOCK(ctx);
3843 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3844 		CALLOUT_LOCK(txq);
3845 		callout_stop(&txq->ift_timer);
3846 		CALLOUT_UNLOCK(txq);
3847 	}
3848 	if (do_watchdog) {
3849 		ctx->ifc_watchdog_events++;
3850 		IFDI_WATCHDOG_RESET(ctx);
3851 	}
3852 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3853 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3854 #ifdef DEV_NETMAP
3855 		reset_on = hz / 2;
3856 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3857 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3858 #endif
3859 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3860 	}
3861 	IFDI_LINK_INTR_ENABLE(ctx);
3862 	if (do_reset)
3863 		iflib_if_init_locked(ctx);
3864 	CTX_UNLOCK(ctx);
3865 
3866 	if (LINK_ACTIVE(ctx) == 0)
3867 		return;
3868 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3869 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3870 }
3871 
3872 
3873 static void
3874 _task_fn_iov(void *context)
3875 {
3876 	if_ctx_t ctx = context;
3877 
3878 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3879 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3880 		return;
3881 
3882 	CTX_LOCK(ctx);
3883 	IFDI_VFLR_HANDLE(ctx);
3884 	CTX_UNLOCK(ctx);
3885 }
3886 
3887 static int
3888 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3889 {
3890 	int err;
3891 	if_int_delay_info_t info;
3892 	if_ctx_t ctx;
3893 
3894 	info = (if_int_delay_info_t)arg1;
3895 	ctx = info->iidi_ctx;
3896 	info->iidi_req = req;
3897 	info->iidi_oidp = oidp;
3898 	CTX_LOCK(ctx);
3899 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3900 	CTX_UNLOCK(ctx);
3901 	return (err);
3902 }
3903 
3904 /*********************************************************************
3905  *
3906  *  IFNET FUNCTIONS
3907  *
3908  **********************************************************************/
3909 
3910 static void
3911 iflib_if_init_locked(if_ctx_t ctx)
3912 {
3913 	iflib_stop(ctx);
3914 	iflib_init_locked(ctx);
3915 }
3916 
3917 
3918 static void
3919 iflib_if_init(void *arg)
3920 {
3921 	if_ctx_t ctx = arg;
3922 
3923 	CTX_LOCK(ctx);
3924 	iflib_if_init_locked(ctx);
3925 	CTX_UNLOCK(ctx);
3926 }
3927 
3928 static int
3929 iflib_if_transmit(if_t ifp, struct mbuf *m)
3930 {
3931 	if_ctx_t	ctx = if_getsoftc(ifp);
3932 
3933 	iflib_txq_t txq;
3934 	int err, qidx;
3935 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3936 
3937 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3938 		DBG_COUNTER_INC(tx_frees);
3939 		m_freem(m);
3940 		return (ENETDOWN);
3941 	}
3942 
3943 	MPASS(m->m_nextpkt == NULL);
3944 	/* ALTQ-enabled interfaces always use queue 0. */
3945 	qidx = 0;
3946 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3947 		qidx = QIDX(ctx, m);
3948 	/*
3949 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3950 	 */
3951 	txq = &ctx->ifc_txqs[qidx];
3952 
3953 #ifdef DRIVER_BACKPRESSURE
3954 	if (txq->ift_closed) {
3955 		while (m != NULL) {
3956 			next = m->m_nextpkt;
3957 			m->m_nextpkt = NULL;
3958 			m_freem(m);
3959 			DBG_COUNTER_INC(tx_frees);
3960 			m = next;
3961 		}
3962 		return (ENOBUFS);
3963 	}
3964 #endif
3965 #ifdef notyet
3966 	qidx = count = 0;
3967 	mp = marr;
3968 	next = m;
3969 	do {
3970 		count++;
3971 		next = next->m_nextpkt;
3972 	} while (next != NULL);
3973 
3974 	if (count > nitems(marr))
3975 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3976 			/* XXX check nextpkt */
3977 			m_freem(m);
3978 			/* XXX simplify for now */
3979 			DBG_COUNTER_INC(tx_frees);
3980 			return (ENOBUFS);
3981 		}
3982 	for (next = m, i = 0; next != NULL; i++) {
3983 		mp[i] = next;
3984 		next = next->m_nextpkt;
3985 		mp[i]->m_nextpkt = NULL;
3986 	}
3987 #endif
3988 	DBG_COUNTER_INC(tx_seen);
3989 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3990 
3991 	if (abdicate)
3992 		GROUPTASK_ENQUEUE(&txq->ift_task);
3993  	if (err) {
3994 		if (!abdicate)
3995 			GROUPTASK_ENQUEUE(&txq->ift_task);
3996 		/* support forthcoming later */
3997 #ifdef DRIVER_BACKPRESSURE
3998 		txq->ift_closed = TRUE;
3999 #endif
4000 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4001 		m_freem(m);
4002 		DBG_COUNTER_INC(tx_frees);
4003 	}
4004 
4005 	return (err);
4006 }
4007 
4008 #ifdef ALTQ
4009 /*
4010  * The overall approach to integrating iflib with ALTQ is to continue to use
4011  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4012  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4013  * is redundant/unnecessary, but doing so minimizes the amount of
4014  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4015  * redundantly queueing to an intermediate mp_ring is swamped by the
4016  * performance limitations inherent in using ALTQ.
4017  *
4018  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4019  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4020  * given interface.  If ALTQ is enabled for an interface, then all
4021  * transmitted packets for that interface will be submitted to the ALTQ
4022  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4023  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4024  * update stats that the iflib machinery handles, and which is sensitve to
4025  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4026  * will be installed as the start routine for use by ALTQ facilities that
4027  * need to trigger queue drains on a scheduled basis.
4028  *
4029  */
4030 static void
4031 iflib_altq_if_start(if_t ifp)
4032 {
4033 	struct ifaltq *ifq = &ifp->if_snd;
4034 	struct mbuf *m;
4035 
4036 	IFQ_LOCK(ifq);
4037 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4038 	while (m != NULL) {
4039 		iflib_if_transmit(ifp, m);
4040 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4041 	}
4042 	IFQ_UNLOCK(ifq);
4043 }
4044 
4045 static int
4046 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4047 {
4048 	int err;
4049 
4050 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4051 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4052 		if (err == 0)
4053 			iflib_altq_if_start(ifp);
4054 	} else
4055 		err = iflib_if_transmit(ifp, m);
4056 
4057 	return (err);
4058 }
4059 #endif /* ALTQ */
4060 
4061 static void
4062 iflib_if_qflush(if_t ifp)
4063 {
4064 	if_ctx_t ctx = if_getsoftc(ifp);
4065 	iflib_txq_t txq = ctx->ifc_txqs;
4066 	int i;
4067 
4068 	STATE_LOCK(ctx);
4069 	ctx->ifc_flags |= IFC_QFLUSH;
4070 	STATE_UNLOCK(ctx);
4071 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4072 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4073 			iflib_txq_check_drain(txq, 0);
4074 	STATE_LOCK(ctx);
4075 	ctx->ifc_flags &= ~IFC_QFLUSH;
4076 	STATE_UNLOCK(ctx);
4077 
4078 	/*
4079 	 * When ALTQ is enabled, this will also take care of purging the
4080 	 * ALTQ queue(s).
4081 	 */
4082 	if_qflush(ifp);
4083 }
4084 
4085 
4086 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4087 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4088 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4089 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4090 
4091 static int
4092 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4093 {
4094 	if_ctx_t ctx = if_getsoftc(ifp);
4095 	struct ifreq	*ifr = (struct ifreq *)data;
4096 #if defined(INET) || defined(INET6)
4097 	struct ifaddr	*ifa = (struct ifaddr *)data;
4098 #endif
4099 	bool		avoid_reset = false;
4100 	int		err = 0, reinit = 0, bits;
4101 
4102 	switch (command) {
4103 	case SIOCSIFADDR:
4104 #ifdef INET
4105 		if (ifa->ifa_addr->sa_family == AF_INET)
4106 			avoid_reset = true;
4107 #endif
4108 #ifdef INET6
4109 		if (ifa->ifa_addr->sa_family == AF_INET6)
4110 			avoid_reset = true;
4111 #endif
4112 		/*
4113 		** Calling init results in link renegotiation,
4114 		** so we avoid doing it when possible.
4115 		*/
4116 		if (avoid_reset) {
4117 			if_setflagbits(ifp, IFF_UP,0);
4118 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4119 				reinit = 1;
4120 #ifdef INET
4121 			if (!(if_getflags(ifp) & IFF_NOARP))
4122 				arp_ifinit(ifp, ifa);
4123 #endif
4124 		} else
4125 			err = ether_ioctl(ifp, command, data);
4126 		break;
4127 	case SIOCSIFMTU:
4128 		CTX_LOCK(ctx);
4129 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4130 			CTX_UNLOCK(ctx);
4131 			break;
4132 		}
4133 		bits = if_getdrvflags(ifp);
4134 		/* stop the driver and free any clusters before proceeding */
4135 		iflib_stop(ctx);
4136 
4137 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4138 			STATE_LOCK(ctx);
4139 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4140 				ctx->ifc_flags |= IFC_MULTISEG;
4141 			else
4142 				ctx->ifc_flags &= ~IFC_MULTISEG;
4143 			STATE_UNLOCK(ctx);
4144 			err = if_setmtu(ifp, ifr->ifr_mtu);
4145 		}
4146 		iflib_init_locked(ctx);
4147 		STATE_LOCK(ctx);
4148 		if_setdrvflags(ifp, bits);
4149 		STATE_UNLOCK(ctx);
4150 		CTX_UNLOCK(ctx);
4151 		break;
4152 	case SIOCSIFFLAGS:
4153 		CTX_LOCK(ctx);
4154 		if (if_getflags(ifp) & IFF_UP) {
4155 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4156 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4157 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4158 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4159 				}
4160 			} else
4161 				reinit = 1;
4162 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4163 			iflib_stop(ctx);
4164 		}
4165 		ctx->ifc_if_flags = if_getflags(ifp);
4166 		CTX_UNLOCK(ctx);
4167 		break;
4168 	case SIOCADDMULTI:
4169 	case SIOCDELMULTI:
4170 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4171 			CTX_LOCK(ctx);
4172 			IFDI_INTR_DISABLE(ctx);
4173 			IFDI_MULTI_SET(ctx);
4174 			IFDI_INTR_ENABLE(ctx);
4175 			CTX_UNLOCK(ctx);
4176 		}
4177 		break;
4178 	case SIOCSIFMEDIA:
4179 		CTX_LOCK(ctx);
4180 		IFDI_MEDIA_SET(ctx);
4181 		CTX_UNLOCK(ctx);
4182 		/* FALLTHROUGH */
4183 	case SIOCGIFMEDIA:
4184 	case SIOCGIFXMEDIA:
4185 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4186 		break;
4187 	case SIOCGI2C:
4188 	{
4189 		struct ifi2creq i2c;
4190 
4191 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4192 		if (err != 0)
4193 			break;
4194 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4195 			err = EINVAL;
4196 			break;
4197 		}
4198 		if (i2c.len > sizeof(i2c.data)) {
4199 			err = EINVAL;
4200 			break;
4201 		}
4202 
4203 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4204 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4205 			    sizeof(i2c));
4206 		break;
4207 	}
4208 	case SIOCSIFCAP:
4209 	{
4210 		int mask, setmask, oldmask;
4211 
4212 		oldmask = if_getcapenable(ifp);
4213 		mask = ifr->ifr_reqcap ^ oldmask;
4214 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4215 		setmask = 0;
4216 #ifdef TCP_OFFLOAD
4217 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4218 #endif
4219 		setmask |= (mask & IFCAP_FLAGS);
4220 		setmask |= (mask & IFCAP_WOL);
4221 
4222 		/*
4223 		 * If any RX csum has changed, change all the ones that
4224 		 * are supported by the driver.
4225 		 */
4226 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4227 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4228 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4229 		}
4230 
4231 		/*
4232 		 * want to ensure that traffic has stopped before we change any of the flags
4233 		 */
4234 		if (setmask) {
4235 			CTX_LOCK(ctx);
4236 			bits = if_getdrvflags(ifp);
4237 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4238 				iflib_stop(ctx);
4239 			STATE_LOCK(ctx);
4240 			if_togglecapenable(ifp, setmask);
4241 			STATE_UNLOCK(ctx);
4242 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4243 				iflib_init_locked(ctx);
4244 			STATE_LOCK(ctx);
4245 			if_setdrvflags(ifp, bits);
4246 			STATE_UNLOCK(ctx);
4247 			CTX_UNLOCK(ctx);
4248 		}
4249 		if_vlancap(ifp);
4250 		break;
4251 	}
4252 	case SIOCGPRIVATE_0:
4253 	case SIOCSDRVSPEC:
4254 	case SIOCGDRVSPEC:
4255 		CTX_LOCK(ctx);
4256 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4257 		CTX_UNLOCK(ctx);
4258 		break;
4259 	default:
4260 		err = ether_ioctl(ifp, command, data);
4261 		break;
4262 	}
4263 	if (reinit)
4264 		iflib_if_init(ctx);
4265 	return (err);
4266 }
4267 
4268 static uint64_t
4269 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4270 {
4271 	if_ctx_t ctx = if_getsoftc(ifp);
4272 
4273 	return (IFDI_GET_COUNTER(ctx, cnt));
4274 }
4275 
4276 /*********************************************************************
4277  *
4278  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4279  *
4280  **********************************************************************/
4281 
4282 static void
4283 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4284 {
4285 	if_ctx_t ctx = if_getsoftc(ifp);
4286 
4287 	if ((void *)ctx != arg)
4288 		return;
4289 
4290 	if ((vtag == 0) || (vtag > 4095))
4291 		return;
4292 
4293 	if (iflib_in_detach(ctx))
4294 		return;
4295 
4296 	CTX_LOCK(ctx);
4297 	IFDI_VLAN_REGISTER(ctx, vtag);
4298 	/* Re-init to load the changes */
4299 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4300 		iflib_if_init_locked(ctx);
4301 	CTX_UNLOCK(ctx);
4302 }
4303 
4304 static void
4305 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4306 {
4307 	if_ctx_t ctx = if_getsoftc(ifp);
4308 
4309 	if ((void *)ctx != arg)
4310 		return;
4311 
4312 	if ((vtag == 0) || (vtag > 4095))
4313 		return;
4314 
4315 	CTX_LOCK(ctx);
4316 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4317 	/* Re-init to load the changes */
4318 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4319 		iflib_if_init_locked(ctx);
4320 	CTX_UNLOCK(ctx);
4321 }
4322 
4323 static void
4324 iflib_led_func(void *arg, int onoff)
4325 {
4326 	if_ctx_t ctx = arg;
4327 
4328 	CTX_LOCK(ctx);
4329 	IFDI_LED_FUNC(ctx, onoff);
4330 	CTX_UNLOCK(ctx);
4331 }
4332 
4333 /*********************************************************************
4334  *
4335  *  BUS FUNCTION DEFINITIONS
4336  *
4337  **********************************************************************/
4338 
4339 int
4340 iflib_device_probe(device_t dev)
4341 {
4342 	const pci_vendor_info_t *ent;
4343 	if_shared_ctx_t sctx;
4344 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4345 	uint16_t pci_vendor_id;
4346 
4347 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4348 		return (ENOTSUP);
4349 
4350 	pci_vendor_id = pci_get_vendor(dev);
4351 	pci_device_id = pci_get_device(dev);
4352 	pci_subvendor_id = pci_get_subvendor(dev);
4353 	pci_subdevice_id = pci_get_subdevice(dev);
4354 	pci_rev_id = pci_get_revid(dev);
4355 	if (sctx->isc_parse_devinfo != NULL)
4356 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4357 
4358 	ent = sctx->isc_vendor_info;
4359 	while (ent->pvi_vendor_id != 0) {
4360 		if (pci_vendor_id != ent->pvi_vendor_id) {
4361 			ent++;
4362 			continue;
4363 		}
4364 		if ((pci_device_id == ent->pvi_device_id) &&
4365 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4366 		     (ent->pvi_subvendor_id == 0)) &&
4367 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4368 		     (ent->pvi_subdevice_id == 0)) &&
4369 		    ((pci_rev_id == ent->pvi_rev_id) ||
4370 		     (ent->pvi_rev_id == 0))) {
4371 
4372 			device_set_desc_copy(dev, ent->pvi_name);
4373 			/* this needs to be changed to zero if the bus probing code
4374 			 * ever stops re-probing on best match because the sctx
4375 			 * may have its values over written by register calls
4376 			 * in subsequent probes
4377 			 */
4378 			return (BUS_PROBE_DEFAULT);
4379 		}
4380 		ent++;
4381 	}
4382 	return (ENXIO);
4383 }
4384 
4385 int
4386 iflib_device_probe_vendor(device_t dev)
4387 {
4388 	int probe;
4389 
4390 	probe = iflib_device_probe(dev);
4391 	if (probe == BUS_PROBE_DEFAULT)
4392 		return (BUS_PROBE_VENDOR);
4393 	else
4394 		return (probe);
4395 }
4396 
4397 static void
4398 iflib_reset_qvalues(if_ctx_t ctx)
4399 {
4400 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4401 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4402 	device_t dev = ctx->ifc_dev;
4403 	int i;
4404 
4405 	if (ctx->ifc_sysctl_ntxqs != 0)
4406 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4407 	if (ctx->ifc_sysctl_nrxqs != 0)
4408 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4409 
4410 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4411 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4412 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4413 		else
4414 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4415 	}
4416 
4417 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4418 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4419 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4420 		else
4421 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4422 	}
4423 
4424 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4425 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4426 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4427 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4428 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4429 		}
4430 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4431 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4432 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4433 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4434 		}
4435 		if (!powerof2(scctx->isc_nrxd[i])) {
4436 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4437 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4438 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4439 		}
4440 	}
4441 
4442 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4443 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4444 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4445 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4446 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4447 		}
4448 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4449 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4450 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4451 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4452 		}
4453 		if (!powerof2(scctx->isc_ntxd[i])) {
4454 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4455 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4456 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4457 		}
4458 	}
4459 }
4460 
4461 static void
4462 iflib_add_pfil(if_ctx_t ctx)
4463 {
4464 	struct pfil_head *pfil;
4465 	struct pfil_head_args pa;
4466 	iflib_rxq_t rxq;
4467 	int i;
4468 
4469 	pa.pa_version = PFIL_VERSION;
4470 	pa.pa_flags = PFIL_IN;
4471 	pa.pa_type = PFIL_TYPE_ETHERNET;
4472 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4473 	pfil = pfil_head_register(&pa);
4474 
4475 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4476 		rxq->pfil = pfil;
4477 	}
4478 }
4479 
4480 static void
4481 iflib_rem_pfil(if_ctx_t ctx)
4482 {
4483 	struct pfil_head *pfil;
4484 	iflib_rxq_t rxq;
4485 	int i;
4486 
4487 	rxq = ctx->ifc_rxqs;
4488 	pfil = rxq->pfil;
4489 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4490 		rxq->pfil = NULL;
4491 	}
4492 	pfil_head_unregister(pfil);
4493 }
4494 
4495 static uint16_t
4496 get_ctx_core_offset(if_ctx_t ctx)
4497 {
4498 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4499 	struct cpu_offset *op;
4500 	uint16_t qc;
4501 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4502 
4503 	if (ret != CORE_OFFSET_UNSPECIFIED)
4504 		return (ret);
4505 
4506 	if (ctx->ifc_sysctl_separate_txrx)
4507 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4508 	else
4509 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4510 
4511 	mtx_lock(&cpu_offset_mtx);
4512 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4513 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4514 			ret = op->offset;
4515 			op->offset += qc;
4516 			MPASS(op->refcount < UINT_MAX);
4517 			op->refcount++;
4518 			break;
4519 		}
4520 	}
4521 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4522 		ret = 0;
4523 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4524 		    M_NOWAIT | M_ZERO);
4525 		if (op == NULL) {
4526 			device_printf(ctx->ifc_dev,
4527 			    "allocation for cpu offset failed.\n");
4528 		} else {
4529 			op->offset = qc;
4530 			op->refcount = 1;
4531 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4532 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4533 		}
4534 	}
4535 	mtx_unlock(&cpu_offset_mtx);
4536 
4537 	return (ret);
4538 }
4539 
4540 static void
4541 unref_ctx_core_offset(if_ctx_t ctx)
4542 {
4543 	struct cpu_offset *op, *top;
4544 
4545 	mtx_lock(&cpu_offset_mtx);
4546 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4547 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4548 			MPASS(op->refcount > 0);
4549 			op->refcount--;
4550 			if (op->refcount == 0) {
4551 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4552 				free(op, M_IFLIB);
4553 			}
4554 			break;
4555 		}
4556 	}
4557 	mtx_unlock(&cpu_offset_mtx);
4558 }
4559 
4560 int
4561 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4562 {
4563 	if_ctx_t ctx;
4564 	if_t ifp;
4565 	if_softc_ctx_t scctx;
4566 	kobjop_desc_t kobj_desc;
4567 	kobj_method_t *kobj_method;
4568 	int err, msix, rid;
4569 	uint16_t main_rxq, main_txq;
4570 
4571 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4572 
4573 	if (sc == NULL) {
4574 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4575 		device_set_softc(dev, ctx);
4576 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4577 	}
4578 
4579 	ctx->ifc_sctx = sctx;
4580 	ctx->ifc_dev = dev;
4581 	ctx->ifc_softc = sc;
4582 
4583 	if ((err = iflib_register(ctx)) != 0) {
4584 		device_printf(dev, "iflib_register failed %d\n", err);
4585 		goto fail_ctx_free;
4586 	}
4587 	iflib_add_device_sysctl_pre(ctx);
4588 
4589 	scctx = &ctx->ifc_softc_ctx;
4590 	ifp = ctx->ifc_ifp;
4591 
4592 	iflib_reset_qvalues(ctx);
4593 	CTX_LOCK(ctx);
4594 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4595 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4596 		goto fail_unlock;
4597 	}
4598 	_iflib_pre_assert(scctx);
4599 	ctx->ifc_txrx = *scctx->isc_txrx;
4600 
4601 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4602 		ctx->ifc_mediap = scctx->isc_media;
4603 
4604 #ifdef INVARIANTS
4605 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4606 		MPASS(scctx->isc_tx_csum_flags);
4607 #endif
4608 
4609 	if_setcapabilities(ifp,
4610 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4611 	if_setcapenable(ifp,
4612 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4613 
4614 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4615 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4616 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4617 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4618 
4619 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4620 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4621 
4622 	/* XXX change for per-queue sizes */
4623 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4624 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4625 
4626 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4627 	    MAX_SINGLE_PACKET_FRACTION)
4628 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4629 		    MAX_SINGLE_PACKET_FRACTION);
4630 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4631 	    MAX_SINGLE_PACKET_FRACTION)
4632 		scctx->isc_tx_tso_segments_max = max(1,
4633 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4634 
4635 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4636 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4637 		/*
4638 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4639 		 * but some MACs do.
4640 		 */
4641 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4642 		    IP_MAXPACKET));
4643 		/*
4644 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4645 		 * into account.  In the worst case, each of these calls will
4646 		 * add another mbuf and, thus, the requirement for another DMA
4647 		 * segment.  So for best performance, it doesn't make sense to
4648 		 * advertize a maximum of TSO segments that typically will
4649 		 * require defragmentation in iflib_encap().
4650 		 */
4651 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4652 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4653 	}
4654 	if (scctx->isc_rss_table_size == 0)
4655 		scctx->isc_rss_table_size = 64;
4656 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4657 
4658 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4659 	/* XXX format name */
4660 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4661 	    NULL, NULL, "admin");
4662 
4663 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4664 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4665 		device_printf(dev, "Unable to fetch CPU list\n");
4666 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4667 	}
4668 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4669 
4670 	/*
4671 	** Now set up MSI or MSI-X, should return us the number of supported
4672 	** vectors (will be 1 for a legacy interrupt and MSI).
4673 	*/
4674 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4675 		msix = scctx->isc_vectors;
4676 	} else if (scctx->isc_msix_bar != 0)
4677 	       /*
4678 		* The simple fact that isc_msix_bar is not 0 does not mean we
4679 		* we have a good value there that is known to work.
4680 		*/
4681 		msix = iflib_msix_init(ctx);
4682 	else {
4683 		scctx->isc_vectors = 1;
4684 		scctx->isc_ntxqsets = 1;
4685 		scctx->isc_nrxqsets = 1;
4686 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4687 		msix = 0;
4688 	}
4689 	/* Get memory for the station queues */
4690 	if ((err = iflib_queues_alloc(ctx))) {
4691 		device_printf(dev, "Unable to allocate queue memory\n");
4692 		goto fail_intr_free;
4693 	}
4694 
4695 	if ((err = iflib_qset_structures_setup(ctx)))
4696 		goto fail_queues;
4697 
4698 	/*
4699 	 * Now that we know how many queues there are, get the core offset.
4700 	 */
4701 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4702 
4703 	/*
4704 	 * Group taskqueues aren't properly set up until SMP is started,
4705 	 * so we disable interrupts until we can handle them post
4706 	 * SI_SUB_SMP.
4707 	 *
4708 	 * XXX: disabling interrupts doesn't actually work, at least for
4709 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4710 	 * we do null handling and depend on this not causing too large an
4711 	 * interrupt storm.
4712 	 */
4713 	IFDI_INTR_DISABLE(ctx);
4714 
4715 	if (msix > 1) {
4716 		/*
4717 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4718 		 * aren't the default NULL implementation.
4719 		 */
4720 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4721 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4722 		    kobj_desc);
4723 		if (kobj_method == &kobj_desc->deflt) {
4724 			device_printf(dev,
4725 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4726 			err = EOPNOTSUPP;
4727 			goto fail_queues;
4728 		}
4729 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4730 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4731 		    kobj_desc);
4732 		if (kobj_method == &kobj_desc->deflt) {
4733 			device_printf(dev,
4734 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4735 			err = EOPNOTSUPP;
4736 			goto fail_queues;
4737 		}
4738 
4739 		/*
4740 		 * Assign the MSI-X vectors.
4741 		 * Note that the default NULL ifdi_msix_intr_assign method will
4742 		 * fail here, too.
4743 		 */
4744 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4745 		if (err != 0) {
4746 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4747 			    err);
4748 			goto fail_queues;
4749 		}
4750 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4751 		rid = 0;
4752 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4753 			MPASS(msix == 1);
4754 			rid = 1;
4755 		}
4756 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4757 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4758 			goto fail_queues;
4759 		}
4760 	} else {
4761 		device_printf(dev,
4762 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4763 		err = ENODEV;
4764 		goto fail_intr_free;
4765 	}
4766 
4767 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4768 
4769 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4770 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4771 		goto fail_detach;
4772 	}
4773 
4774 	/*
4775 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4776 	 * This must appear after the call to ether_ifattach() because
4777 	 * ether_ifattach() sets if_hdrlen to the default value.
4778 	 */
4779 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4780 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4781 
4782 	if ((err = iflib_netmap_attach(ctx))) {
4783 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4784 		goto fail_detach;
4785 	}
4786 	*ctxp = ctx;
4787 
4788 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4789 
4790 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4791 	iflib_add_device_sysctl_post(ctx);
4792 	iflib_add_pfil(ctx);
4793 	ctx->ifc_flags |= IFC_INIT_DONE;
4794 	CTX_UNLOCK(ctx);
4795 
4796 	return (0);
4797 
4798 fail_detach:
4799 	ether_ifdetach(ctx->ifc_ifp);
4800 fail_intr_free:
4801 	iflib_free_intr_mem(ctx);
4802 fail_queues:
4803 	iflib_tx_structures_free(ctx);
4804 	iflib_rx_structures_free(ctx);
4805 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4806 	IFDI_DETACH(ctx);
4807 fail_unlock:
4808 	CTX_UNLOCK(ctx);
4809 	iflib_deregister(ctx);
4810 fail_ctx_free:
4811 	device_set_softc(ctx->ifc_dev, NULL);
4812         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4813                 free(ctx->ifc_softc, M_IFLIB);
4814         free(ctx, M_IFLIB);
4815 	return (err);
4816 }
4817 
4818 int
4819 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4820 					  struct iflib_cloneattach_ctx *clctx)
4821 {
4822 	int err;
4823 	if_ctx_t ctx;
4824 	if_t ifp;
4825 	if_softc_ctx_t scctx;
4826 	int i;
4827 	void *sc;
4828 	uint16_t main_txq;
4829 	uint16_t main_rxq;
4830 
4831 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4832 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4833 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4834 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4835 		ctx->ifc_flags |= IFC_PSEUDO;
4836 
4837 	ctx->ifc_sctx = sctx;
4838 	ctx->ifc_softc = sc;
4839 	ctx->ifc_dev = dev;
4840 
4841 	if ((err = iflib_register(ctx)) != 0) {
4842 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4843 		goto fail_ctx_free;
4844 	}
4845 	iflib_add_device_sysctl_pre(ctx);
4846 
4847 	scctx = &ctx->ifc_softc_ctx;
4848 	ifp = ctx->ifc_ifp;
4849 
4850 	iflib_reset_qvalues(ctx);
4851 	CTX_LOCK(ctx);
4852 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4853 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4854 		goto fail_unlock;
4855 	}
4856 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4857 		ether_gen_addr(ifp, &ctx->ifc_mac);
4858 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4859 								clctx->cc_params)) != 0) {
4860 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4861 		goto fail_ctx_free;
4862 	}
4863 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4864 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4865 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4866 
4867 #ifdef INVARIANTS
4868 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4869 		MPASS(scctx->isc_tx_csum_flags);
4870 #endif
4871 
4872 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4873 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4874 
4875 	ifp->if_flags |= IFF_NOGROUP;
4876 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4877 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4878 
4879 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4880 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4881 			goto fail_detach;
4882 		}
4883 		*ctxp = ctx;
4884 
4885 		/*
4886 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4887 		 * This must appear after the call to ether_ifattach() because
4888 		 * ether_ifattach() sets if_hdrlen to the default value.
4889 		 */
4890 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4891 			if_setifheaderlen(ifp,
4892 			    sizeof(struct ether_vlan_header));
4893 
4894 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4895 		iflib_add_device_sysctl_post(ctx);
4896 		ctx->ifc_flags |= IFC_INIT_DONE;
4897 		return (0);
4898 	}
4899 	_iflib_pre_assert(scctx);
4900 	ctx->ifc_txrx = *scctx->isc_txrx;
4901 
4902 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4903 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4904 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4905 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4906 
4907 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4908 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4909 
4910 	/* XXX change for per-queue sizes */
4911 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4912 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4913 
4914 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4915 	    MAX_SINGLE_PACKET_FRACTION)
4916 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4917 		    MAX_SINGLE_PACKET_FRACTION);
4918 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4919 	    MAX_SINGLE_PACKET_FRACTION)
4920 		scctx->isc_tx_tso_segments_max = max(1,
4921 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4922 
4923 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4924 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4925 		/*
4926 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4927 		 * but some MACs do.
4928 		 */
4929 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4930 		    IP_MAXPACKET));
4931 		/*
4932 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4933 		 * into account.  In the worst case, each of these calls will
4934 		 * add another mbuf and, thus, the requirement for another DMA
4935 		 * segment.  So for best performance, it doesn't make sense to
4936 		 * advertize a maximum of TSO segments that typically will
4937 		 * require defragmentation in iflib_encap().
4938 		 */
4939 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4940 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4941 	}
4942 	if (scctx->isc_rss_table_size == 0)
4943 		scctx->isc_rss_table_size = 64;
4944 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4945 
4946 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4947 	/* XXX format name */
4948 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4949 	    NULL, NULL, "admin");
4950 
4951 	/* XXX --- can support > 1 -- but keep it simple for now */
4952 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4953 
4954 	/* Get memory for the station queues */
4955 	if ((err = iflib_queues_alloc(ctx))) {
4956 		device_printf(dev, "Unable to allocate queue memory\n");
4957 		goto fail_iflib_detach;
4958 	}
4959 
4960 	if ((err = iflib_qset_structures_setup(ctx))) {
4961 		device_printf(dev, "qset structure setup failed %d\n", err);
4962 		goto fail_queues;
4963 	}
4964 
4965 	/*
4966 	 * XXX What if anything do we want to do about interrupts?
4967 	 */
4968 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4969 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4970 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4971 		goto fail_detach;
4972 	}
4973 
4974 	/*
4975 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4976 	 * This must appear after the call to ether_ifattach() because
4977 	 * ether_ifattach() sets if_hdrlen to the default value.
4978 	 */
4979 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4980 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4981 
4982 	/* XXX handle more than one queue */
4983 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4984 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4985 
4986 	*ctxp = ctx;
4987 
4988 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4989 	iflib_add_device_sysctl_post(ctx);
4990 	ctx->ifc_flags |= IFC_INIT_DONE;
4991 	CTX_UNLOCK(ctx);
4992 
4993 	return (0);
4994 fail_detach:
4995 	ether_ifdetach(ctx->ifc_ifp);
4996 fail_queues:
4997 	iflib_tx_structures_free(ctx);
4998 	iflib_rx_structures_free(ctx);
4999 fail_iflib_detach:
5000 	IFDI_DETACH(ctx);
5001 fail_unlock:
5002 	CTX_UNLOCK(ctx);
5003 	iflib_deregister(ctx);
5004 fail_ctx_free:
5005 	free(ctx->ifc_softc, M_IFLIB);
5006 	free(ctx, M_IFLIB);
5007 	return (err);
5008 }
5009 
5010 int
5011 iflib_pseudo_deregister(if_ctx_t ctx)
5012 {
5013 	if_t ifp = ctx->ifc_ifp;
5014 	iflib_txq_t txq;
5015 	iflib_rxq_t rxq;
5016 	int i, j;
5017 	struct taskqgroup *tqg;
5018 	iflib_fl_t fl;
5019 
5020 	/* Unregister VLAN event handlers early */
5021 	iflib_unregister_vlan_handlers(ctx);
5022 
5023 	ether_ifdetach(ifp);
5024 	/* XXX drain any dependent tasks */
5025 	tqg = qgroup_if_io_tqg;
5026 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5027 		callout_drain(&txq->ift_timer);
5028 		if (txq->ift_task.gt_uniq != NULL)
5029 			taskqgroup_detach(tqg, &txq->ift_task);
5030 	}
5031 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5032 		if (rxq->ifr_task.gt_uniq != NULL)
5033 			taskqgroup_detach(tqg, &rxq->ifr_task);
5034 
5035 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5036 			free(fl->ifl_rx_bitmap, M_IFLIB);
5037 	}
5038 	tqg = qgroup_if_config_tqg;
5039 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5040 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5041 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5042 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5043 
5044 	iflib_tx_structures_free(ctx);
5045 	iflib_rx_structures_free(ctx);
5046 
5047 	iflib_deregister(ctx);
5048 
5049 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5050 		free(ctx->ifc_softc, M_IFLIB);
5051 	free(ctx, M_IFLIB);
5052 	return (0);
5053 }
5054 
5055 int
5056 iflib_device_attach(device_t dev)
5057 {
5058 	if_ctx_t ctx;
5059 	if_shared_ctx_t sctx;
5060 
5061 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5062 		return (ENOTSUP);
5063 
5064 	pci_enable_busmaster(dev);
5065 
5066 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5067 }
5068 
5069 int
5070 iflib_device_deregister(if_ctx_t ctx)
5071 {
5072 	if_t ifp = ctx->ifc_ifp;
5073 	iflib_txq_t txq;
5074 	iflib_rxq_t rxq;
5075 	device_t dev = ctx->ifc_dev;
5076 	int i, j;
5077 	struct taskqgroup *tqg;
5078 	iflib_fl_t fl;
5079 
5080 	/* Make sure VLANS are not using driver */
5081 	if (if_vlantrunkinuse(ifp)) {
5082 		device_printf(dev, "Vlan in use, detach first\n");
5083 		return (EBUSY);
5084 	}
5085 #ifdef PCI_IOV
5086 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5087 		device_printf(dev, "SR-IOV in use; detach first.\n");
5088 		return (EBUSY);
5089 	}
5090 #endif
5091 
5092 	STATE_LOCK(ctx);
5093 	ctx->ifc_flags |= IFC_IN_DETACH;
5094 	STATE_UNLOCK(ctx);
5095 
5096 	/* Unregister VLAN handlers before calling iflib_stop() */
5097 	iflib_unregister_vlan_handlers(ctx);
5098 
5099 	iflib_netmap_detach(ifp);
5100 	ether_ifdetach(ifp);
5101 
5102 	CTX_LOCK(ctx);
5103 	iflib_stop(ctx);
5104 	CTX_UNLOCK(ctx);
5105 
5106 	iflib_rem_pfil(ctx);
5107 	if (ctx->ifc_led_dev != NULL)
5108 		led_destroy(ctx->ifc_led_dev);
5109 	/* XXX drain any dependent tasks */
5110 	tqg = qgroup_if_io_tqg;
5111 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5112 		callout_drain(&txq->ift_timer);
5113 		if (txq->ift_task.gt_uniq != NULL)
5114 			taskqgroup_detach(tqg, &txq->ift_task);
5115 	}
5116 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5117 		if (rxq->ifr_task.gt_uniq != NULL)
5118 			taskqgroup_detach(tqg, &rxq->ifr_task);
5119 
5120 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5121 			free(fl->ifl_rx_bitmap, M_IFLIB);
5122 	}
5123 	tqg = qgroup_if_config_tqg;
5124 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5125 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5126 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5127 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5128 	CTX_LOCK(ctx);
5129 	IFDI_DETACH(ctx);
5130 	CTX_UNLOCK(ctx);
5131 
5132 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5133 	iflib_free_intr_mem(ctx);
5134 
5135 	bus_generic_detach(dev);
5136 
5137 	iflib_tx_structures_free(ctx);
5138 	iflib_rx_structures_free(ctx);
5139 
5140 	iflib_deregister(ctx);
5141 
5142 	device_set_softc(ctx->ifc_dev, NULL);
5143 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5144 		free(ctx->ifc_softc, M_IFLIB);
5145 	unref_ctx_core_offset(ctx);
5146 	free(ctx, M_IFLIB);
5147 	return (0);
5148 }
5149 
5150 static void
5151 iflib_free_intr_mem(if_ctx_t ctx)
5152 {
5153 
5154 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5155 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5156 	}
5157 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5158 		pci_release_msi(ctx->ifc_dev);
5159 	}
5160 	if (ctx->ifc_msix_mem != NULL) {
5161 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5162 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5163 		ctx->ifc_msix_mem = NULL;
5164 	}
5165 }
5166 
5167 int
5168 iflib_device_detach(device_t dev)
5169 {
5170 	if_ctx_t ctx = device_get_softc(dev);
5171 
5172 	return (iflib_device_deregister(ctx));
5173 }
5174 
5175 int
5176 iflib_device_suspend(device_t dev)
5177 {
5178 	if_ctx_t ctx = device_get_softc(dev);
5179 
5180 	CTX_LOCK(ctx);
5181 	IFDI_SUSPEND(ctx);
5182 	CTX_UNLOCK(ctx);
5183 
5184 	return bus_generic_suspend(dev);
5185 }
5186 int
5187 iflib_device_shutdown(device_t dev)
5188 {
5189 	if_ctx_t ctx = device_get_softc(dev);
5190 
5191 	CTX_LOCK(ctx);
5192 	IFDI_SHUTDOWN(ctx);
5193 	CTX_UNLOCK(ctx);
5194 
5195 	return bus_generic_suspend(dev);
5196 }
5197 
5198 
5199 int
5200 iflib_device_resume(device_t dev)
5201 {
5202 	if_ctx_t ctx = device_get_softc(dev);
5203 	iflib_txq_t txq = ctx->ifc_txqs;
5204 
5205 	CTX_LOCK(ctx);
5206 	IFDI_RESUME(ctx);
5207 	iflib_if_init_locked(ctx);
5208 	CTX_UNLOCK(ctx);
5209 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5210 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5211 
5212 	return (bus_generic_resume(dev));
5213 }
5214 
5215 int
5216 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5217 {
5218 	int error;
5219 	if_ctx_t ctx = device_get_softc(dev);
5220 
5221 	CTX_LOCK(ctx);
5222 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5223 	CTX_UNLOCK(ctx);
5224 
5225 	return (error);
5226 }
5227 
5228 void
5229 iflib_device_iov_uninit(device_t dev)
5230 {
5231 	if_ctx_t ctx = device_get_softc(dev);
5232 
5233 	CTX_LOCK(ctx);
5234 	IFDI_IOV_UNINIT(ctx);
5235 	CTX_UNLOCK(ctx);
5236 }
5237 
5238 int
5239 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5240 {
5241 	int error;
5242 	if_ctx_t ctx = device_get_softc(dev);
5243 
5244 	CTX_LOCK(ctx);
5245 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5246 	CTX_UNLOCK(ctx);
5247 
5248 	return (error);
5249 }
5250 
5251 /*********************************************************************
5252  *
5253  *  MODULE FUNCTION DEFINITIONS
5254  *
5255  **********************************************************************/
5256 
5257 /*
5258  * - Start a fast taskqueue thread for each core
5259  * - Start a taskqueue for control operations
5260  */
5261 static int
5262 iflib_module_init(void)
5263 {
5264 	return (0);
5265 }
5266 
5267 static int
5268 iflib_module_event_handler(module_t mod, int what, void *arg)
5269 {
5270 	int err;
5271 
5272 	switch (what) {
5273 	case MOD_LOAD:
5274 		if ((err = iflib_module_init()) != 0)
5275 			return (err);
5276 		break;
5277 	case MOD_UNLOAD:
5278 		return (EBUSY);
5279 	default:
5280 		return (EOPNOTSUPP);
5281 	}
5282 
5283 	return (0);
5284 }
5285 
5286 /*********************************************************************
5287  *
5288  *  PUBLIC FUNCTION DEFINITIONS
5289  *     ordered as in iflib.h
5290  *
5291  **********************************************************************/
5292 
5293 
5294 static void
5295 _iflib_assert(if_shared_ctx_t sctx)
5296 {
5297 	int i;
5298 
5299 	MPASS(sctx->isc_tx_maxsize);
5300 	MPASS(sctx->isc_tx_maxsegsize);
5301 
5302 	MPASS(sctx->isc_rx_maxsize);
5303 	MPASS(sctx->isc_rx_nsegments);
5304 	MPASS(sctx->isc_rx_maxsegsize);
5305 
5306 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5307 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5308 		MPASS(sctx->isc_nrxd_min[i]);
5309 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5310 		MPASS(sctx->isc_nrxd_max[i]);
5311 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5312 		MPASS(sctx->isc_nrxd_default[i]);
5313 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5314 	}
5315 
5316 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5317 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5318 		MPASS(sctx->isc_ntxd_min[i]);
5319 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5320 		MPASS(sctx->isc_ntxd_max[i]);
5321 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5322 		MPASS(sctx->isc_ntxd_default[i]);
5323 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5324 	}
5325 }
5326 
5327 static void
5328 _iflib_pre_assert(if_softc_ctx_t scctx)
5329 {
5330 
5331 	MPASS(scctx->isc_txrx->ift_txd_encap);
5332 	MPASS(scctx->isc_txrx->ift_txd_flush);
5333 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5334 	MPASS(scctx->isc_txrx->ift_rxd_available);
5335 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5336 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5337 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5338 }
5339 
5340 static int
5341 iflib_register(if_ctx_t ctx)
5342 {
5343 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5344 	driver_t *driver = sctx->isc_driver;
5345 	device_t dev = ctx->ifc_dev;
5346 	if_t ifp;
5347 
5348 	_iflib_assert(sctx);
5349 
5350 	CTX_LOCK_INIT(ctx);
5351 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5352 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5353 	if (ifp == NULL) {
5354 		device_printf(dev, "can not allocate ifnet structure\n");
5355 		return (ENOMEM);
5356 	}
5357 
5358 	/*
5359 	 * Initialize our context's device specific methods
5360 	 */
5361 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5362 	kobj_class_compile((kobj_class_t) driver);
5363 
5364 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5365 	if_setsoftc(ifp, ctx);
5366 	if_setdev(ifp, dev);
5367 	if_setinitfn(ifp, iflib_if_init);
5368 	if_setioctlfn(ifp, iflib_if_ioctl);
5369 #ifdef ALTQ
5370 	if_setstartfn(ifp, iflib_altq_if_start);
5371 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5372 	if_setsendqready(ifp);
5373 #else
5374 	if_settransmitfn(ifp, iflib_if_transmit);
5375 #endif
5376 	if_setqflushfn(ifp, iflib_if_qflush);
5377 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5378 
5379 	ctx->ifc_vlan_attach_event =
5380 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5381 							  EVENTHANDLER_PRI_FIRST);
5382 	ctx->ifc_vlan_detach_event =
5383 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5384 							  EVENTHANDLER_PRI_FIRST);
5385 
5386 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5387 		ctx->ifc_mediap = &ctx->ifc_media;
5388 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5389 		    iflib_media_change, iflib_media_status);
5390 	}
5391 	return (0);
5392 }
5393 
5394 static void
5395 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5396 {
5397 	/* Unregister VLAN events */
5398 	if (ctx->ifc_vlan_attach_event != NULL) {
5399 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5400 		ctx->ifc_vlan_attach_event = NULL;
5401 	}
5402 	if (ctx->ifc_vlan_detach_event != NULL) {
5403 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5404 		ctx->ifc_vlan_detach_event = NULL;
5405 	}
5406 
5407 }
5408 
5409 static void
5410 iflib_deregister(if_ctx_t ctx)
5411 {
5412 	if_t ifp = ctx->ifc_ifp;
5413 
5414 	/* Remove all media */
5415 	ifmedia_removeall(&ctx->ifc_media);
5416 
5417 	/* Ensure that VLAN event handlers are unregistered */
5418 	iflib_unregister_vlan_handlers(ctx);
5419 
5420 	/* Release kobject reference */
5421 	kobj_delete((kobj_t) ctx, NULL);
5422 
5423 	/* Free the ifnet structure */
5424 	if_free(ifp);
5425 
5426 	STATE_LOCK_DESTROY(ctx);
5427 
5428 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5429 	CTX_LOCK_DESTROY(ctx);
5430 }
5431 
5432 static int
5433 iflib_queues_alloc(if_ctx_t ctx)
5434 {
5435 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5436 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5437 	device_t dev = ctx->ifc_dev;
5438 	int nrxqsets = scctx->isc_nrxqsets;
5439 	int ntxqsets = scctx->isc_ntxqsets;
5440 	iflib_txq_t txq;
5441 	iflib_rxq_t rxq;
5442 	iflib_fl_t fl = NULL;
5443 	int i, j, cpu, err, txconf, rxconf;
5444 	iflib_dma_info_t ifdip;
5445 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5446 	uint32_t *txqsizes = scctx->isc_txqsizes;
5447 	uint8_t nrxqs = sctx->isc_nrxqs;
5448 	uint8_t ntxqs = sctx->isc_ntxqs;
5449 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5450 	caddr_t *vaddrs;
5451 	uint64_t *paddrs;
5452 
5453 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5454 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5455 
5456 	/* Allocate the TX ring struct memory */
5457 	if (!(ctx->ifc_txqs =
5458 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5459 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5460 		device_printf(dev, "Unable to allocate TX ring memory\n");
5461 		err = ENOMEM;
5462 		goto fail;
5463 	}
5464 
5465 	/* Now allocate the RX */
5466 	if (!(ctx->ifc_rxqs =
5467 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5468 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5469 		device_printf(dev, "Unable to allocate RX ring memory\n");
5470 		err = ENOMEM;
5471 		goto rx_fail;
5472 	}
5473 
5474 	txq = ctx->ifc_txqs;
5475 	rxq = ctx->ifc_rxqs;
5476 
5477 	/*
5478 	 * XXX handle allocation failure
5479 	 */
5480 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5481 		/* Set up some basics */
5482 
5483 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5484 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5485 			device_printf(dev,
5486 			    "Unable to allocate TX DMA info memory\n");
5487 			err = ENOMEM;
5488 			goto err_tx_desc;
5489 		}
5490 		txq->ift_ifdi = ifdip;
5491 		for (j = 0; j < ntxqs; j++, ifdip++) {
5492 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5493 				device_printf(dev,
5494 				    "Unable to allocate TX descriptors\n");
5495 				err = ENOMEM;
5496 				goto err_tx_desc;
5497 			}
5498 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5499 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5500 		}
5501 		txq->ift_ctx = ctx;
5502 		txq->ift_id = i;
5503 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5504 			txq->ift_br_offset = 1;
5505 		} else {
5506 			txq->ift_br_offset = 0;
5507 		}
5508 		/* XXX fix this */
5509 		txq->ift_timer.c_cpu = cpu;
5510 
5511 		if (iflib_txsd_alloc(txq)) {
5512 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5513 			err = ENOMEM;
5514 			goto err_tx_desc;
5515 		}
5516 
5517 		/* Initialize the TX lock */
5518 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5519 		    device_get_nameunit(dev), txq->ift_id);
5520 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5521 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5522 
5523 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5524 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5525 		if (err) {
5526 			/* XXX free any allocated rings */
5527 			device_printf(dev, "Unable to allocate buf_ring\n");
5528 			goto err_tx_desc;
5529 		}
5530 	}
5531 
5532 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5533 		/* Set up some basics */
5534 
5535 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5536 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5537 			device_printf(dev,
5538 			    "Unable to allocate RX DMA info memory\n");
5539 			err = ENOMEM;
5540 			goto err_tx_desc;
5541 		}
5542 
5543 		rxq->ifr_ifdi = ifdip;
5544 		/* XXX this needs to be changed if #rx queues != #tx queues */
5545 		rxq->ifr_ntxqirq = 1;
5546 		rxq->ifr_txqid[0] = i;
5547 		for (j = 0; j < nrxqs; j++, ifdip++) {
5548 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5549 				device_printf(dev,
5550 				    "Unable to allocate RX descriptors\n");
5551 				err = ENOMEM;
5552 				goto err_tx_desc;
5553 			}
5554 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5555 		}
5556 		rxq->ifr_ctx = ctx;
5557 		rxq->ifr_id = i;
5558 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5559 			rxq->ifr_fl_offset = 1;
5560 		} else {
5561 			rxq->ifr_fl_offset = 0;
5562 		}
5563 		rxq->ifr_nfl = nfree_lists;
5564 		if (!(fl =
5565 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5566 			device_printf(dev, "Unable to allocate free list memory\n");
5567 			err = ENOMEM;
5568 			goto err_tx_desc;
5569 		}
5570 		rxq->ifr_fl = fl;
5571 		for (j = 0; j < nfree_lists; j++) {
5572 			fl[j].ifl_rxq = rxq;
5573 			fl[j].ifl_id = j;
5574 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5575 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5576 		}
5577 		/* Allocate receive buffers for the ring */
5578 		if (iflib_rxsd_alloc(rxq)) {
5579 			device_printf(dev,
5580 			    "Critical Failure setting up receive buffers\n");
5581 			err = ENOMEM;
5582 			goto err_rx_desc;
5583 		}
5584 
5585 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5586 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5587 			    M_WAITOK);
5588 	}
5589 
5590 	/* TXQs */
5591 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5592 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5593 	for (i = 0; i < ntxqsets; i++) {
5594 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5595 
5596 		for (j = 0; j < ntxqs; j++, di++) {
5597 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5598 			paddrs[i*ntxqs + j] = di->idi_paddr;
5599 		}
5600 	}
5601 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5602 		device_printf(ctx->ifc_dev,
5603 		    "Unable to allocate device TX queue\n");
5604 		iflib_tx_structures_free(ctx);
5605 		free(vaddrs, M_IFLIB);
5606 		free(paddrs, M_IFLIB);
5607 		goto err_rx_desc;
5608 	}
5609 	free(vaddrs, M_IFLIB);
5610 	free(paddrs, M_IFLIB);
5611 
5612 	/* RXQs */
5613 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5614 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5615 	for (i = 0; i < nrxqsets; i++) {
5616 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5617 
5618 		for (j = 0; j < nrxqs; j++, di++) {
5619 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5620 			paddrs[i*nrxqs + j] = di->idi_paddr;
5621 		}
5622 	}
5623 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5624 		device_printf(ctx->ifc_dev,
5625 		    "Unable to allocate device RX queue\n");
5626 		iflib_tx_structures_free(ctx);
5627 		free(vaddrs, M_IFLIB);
5628 		free(paddrs, M_IFLIB);
5629 		goto err_rx_desc;
5630 	}
5631 	free(vaddrs, M_IFLIB);
5632 	free(paddrs, M_IFLIB);
5633 
5634 	return (0);
5635 
5636 /* XXX handle allocation failure changes */
5637 err_rx_desc:
5638 err_tx_desc:
5639 rx_fail:
5640 	if (ctx->ifc_rxqs != NULL)
5641 		free(ctx->ifc_rxqs, M_IFLIB);
5642 	ctx->ifc_rxqs = NULL;
5643 	if (ctx->ifc_txqs != NULL)
5644 		free(ctx->ifc_txqs, M_IFLIB);
5645 	ctx->ifc_txqs = NULL;
5646 fail:
5647 	return (err);
5648 }
5649 
5650 static int
5651 iflib_tx_structures_setup(if_ctx_t ctx)
5652 {
5653 	iflib_txq_t txq = ctx->ifc_txqs;
5654 	int i;
5655 
5656 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5657 		iflib_txq_setup(txq);
5658 
5659 	return (0);
5660 }
5661 
5662 static void
5663 iflib_tx_structures_free(if_ctx_t ctx)
5664 {
5665 	iflib_txq_t txq = ctx->ifc_txqs;
5666 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5667 	int i, j;
5668 
5669 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5670 		for (j = 0; j < sctx->isc_ntxqs; j++)
5671 			iflib_dma_free(&txq->ift_ifdi[j]);
5672 		iflib_txq_destroy(txq);
5673 	}
5674 	free(ctx->ifc_txqs, M_IFLIB);
5675 	ctx->ifc_txqs = NULL;
5676 	IFDI_QUEUES_FREE(ctx);
5677 }
5678 
5679 /*********************************************************************
5680  *
5681  *  Initialize all receive rings.
5682  *
5683  **********************************************************************/
5684 static int
5685 iflib_rx_structures_setup(if_ctx_t ctx)
5686 {
5687 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5688 	int q;
5689 #if defined(INET6) || defined(INET)
5690 	int err, i;
5691 #endif
5692 
5693 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5694 #if defined(INET6) || defined(INET)
5695 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5696 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5697 			    TCP_LRO_ENTRIES, min(1024,
5698 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5699 			if (err != 0) {
5700 				device_printf(ctx->ifc_dev,
5701 				    "LRO Initialization failed!\n");
5702 				goto fail;
5703 			}
5704 		}
5705 #endif
5706 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5707 	}
5708 	return (0);
5709 #if defined(INET6) || defined(INET)
5710 fail:
5711 	/*
5712 	 * Free LRO resources allocated so far, we will only handle
5713 	 * the rings that completed, the failing case will have
5714 	 * cleaned up for itself.  'q' failed, so its the terminus.
5715 	 */
5716 	rxq = ctx->ifc_rxqs;
5717 	for (i = 0; i < q; ++i, rxq++) {
5718 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5719 			tcp_lro_free(&rxq->ifr_lc);
5720 	}
5721 	return (err);
5722 #endif
5723 }
5724 
5725 /*********************************************************************
5726  *
5727  *  Free all receive rings.
5728  *
5729  **********************************************************************/
5730 static void
5731 iflib_rx_structures_free(if_ctx_t ctx)
5732 {
5733 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5734 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5735 	int i, j;
5736 
5737 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5738 		for (j = 0; j < sctx->isc_nrxqs; j++)
5739 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5740 		iflib_rx_sds_free(rxq);
5741 #if defined(INET6) || defined(INET)
5742 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5743 			tcp_lro_free(&rxq->ifr_lc);
5744 #endif
5745 	}
5746 	free(ctx->ifc_rxqs, M_IFLIB);
5747 	ctx->ifc_rxqs = NULL;
5748 }
5749 
5750 static int
5751 iflib_qset_structures_setup(if_ctx_t ctx)
5752 {
5753 	int err;
5754 
5755 	/*
5756 	 * It is expected that the caller takes care of freeing queues if this
5757 	 * fails.
5758 	 */
5759 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5760 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5761 		return (err);
5762 	}
5763 
5764 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5765 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5766 
5767 	return (err);
5768 }
5769 
5770 int
5771 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5772 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5773 {
5774 
5775 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5776 }
5777 
5778 #ifdef SMP
5779 static int
5780 find_nth(if_ctx_t ctx, int qid)
5781 {
5782 	cpuset_t cpus;
5783 	int i, cpuid, eqid, count;
5784 
5785 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5786 	count = CPU_COUNT(&cpus);
5787 	eqid = qid % count;
5788 	/* clear up to the qid'th bit */
5789 	for (i = 0; i < eqid; i++) {
5790 		cpuid = CPU_FFS(&cpus);
5791 		MPASS(cpuid != 0);
5792 		CPU_CLR(cpuid-1, &cpus);
5793 	}
5794 	cpuid = CPU_FFS(&cpus);
5795 	MPASS(cpuid != 0);
5796 	return (cpuid-1);
5797 }
5798 
5799 #ifdef SCHED_ULE
5800 extern struct cpu_group *cpu_top;              /* CPU topology */
5801 
5802 static int
5803 find_child_with_core(int cpu, struct cpu_group *grp)
5804 {
5805 	int i;
5806 
5807 	if (grp->cg_children == 0)
5808 		return -1;
5809 
5810 	MPASS(grp->cg_child);
5811 	for (i = 0; i < grp->cg_children; i++) {
5812 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5813 			return i;
5814 	}
5815 
5816 	return -1;
5817 }
5818 
5819 /*
5820  * Find the nth "close" core to the specified core
5821  * "close" is defined as the deepest level that shares
5822  * at least an L2 cache.  With threads, this will be
5823  * threads on the same core.  If the shared cache is L3
5824  * or higher, simply returns the same core.
5825  */
5826 static int
5827 find_close_core(int cpu, int core_offset)
5828 {
5829 	struct cpu_group *grp;
5830 	int i;
5831 	int fcpu;
5832 	cpuset_t cs;
5833 
5834 	grp = cpu_top;
5835 	if (grp == NULL)
5836 		return cpu;
5837 	i = 0;
5838 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5839 		/* If the child only has one cpu, don't descend */
5840 		if (grp->cg_child[i].cg_count <= 1)
5841 			break;
5842 		grp = &grp->cg_child[i];
5843 	}
5844 
5845 	/* If they don't share at least an L2 cache, use the same CPU */
5846 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5847 		return cpu;
5848 
5849 	/* Now pick one */
5850 	CPU_COPY(&grp->cg_mask, &cs);
5851 
5852 	/* Add the selected CPU offset to core offset. */
5853 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5854 		if (fcpu - 1 == cpu)
5855 			break;
5856 		CPU_CLR(fcpu - 1, &cs);
5857 	}
5858 	MPASS(fcpu);
5859 
5860 	core_offset += i;
5861 
5862 	CPU_COPY(&grp->cg_mask, &cs);
5863 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5864 		MPASS(CPU_FFS(&cs));
5865 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5866 	}
5867 	MPASS(CPU_FFS(&cs));
5868 	return CPU_FFS(&cs) - 1;
5869 }
5870 #else
5871 static int
5872 find_close_core(int cpu, int core_offset __unused)
5873 {
5874 	return cpu;
5875 }
5876 #endif
5877 
5878 static int
5879 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5880 {
5881 	switch (type) {
5882 	case IFLIB_INTR_TX:
5883 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5884 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5885 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5886 	case IFLIB_INTR_RX:
5887 	case IFLIB_INTR_RXTX:
5888 		/* RX queues get the specified core */
5889 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5890 	default:
5891 		return -1;
5892 	}
5893 }
5894 #else
5895 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5896 #define find_close_core(cpuid, tid)	CPU_FIRST()
5897 #define find_nth(ctx, gid)		CPU_FIRST()
5898 #endif
5899 
5900 /* Just to avoid copy/paste */
5901 static inline int
5902 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5903     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5904     const char *name)
5905 {
5906 	device_t dev;
5907 	int co, cpuid, err, tid;
5908 
5909 	dev = ctx->ifc_dev;
5910 	co = ctx->ifc_sysctl_core_offset;
5911 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5912 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5913 	cpuid = find_nth(ctx, qid + co);
5914 	tid = get_core_offset(ctx, type, qid);
5915 	if (tid < 0) {
5916 		device_printf(dev, "get_core_offset failed\n");
5917 		return (EOPNOTSUPP);
5918 	}
5919 	cpuid = find_close_core(cpuid, tid);
5920 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5921 	    name);
5922 	if (err) {
5923 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5924 		return (err);
5925 	}
5926 #ifdef notyet
5927 	if (cpuid > ctx->ifc_cpuid_highest)
5928 		ctx->ifc_cpuid_highest = cpuid;
5929 #endif
5930 	return (0);
5931 }
5932 
5933 int
5934 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5935 			iflib_intr_type_t type, driver_filter_t *filter,
5936 			void *filter_arg, int qid, const char *name)
5937 {
5938 	device_t dev;
5939 	struct grouptask *gtask;
5940 	struct taskqgroup *tqg;
5941 	iflib_filter_info_t info;
5942 	gtask_fn_t *fn;
5943 	int tqrid, err;
5944 	driver_filter_t *intr_fast;
5945 	void *q;
5946 
5947 	info = &ctx->ifc_filter_info;
5948 	tqrid = rid;
5949 
5950 	switch (type) {
5951 	/* XXX merge tx/rx for netmap? */
5952 	case IFLIB_INTR_TX:
5953 		q = &ctx->ifc_txqs[qid];
5954 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5955 		gtask = &ctx->ifc_txqs[qid].ift_task;
5956 		tqg = qgroup_if_io_tqg;
5957 		fn = _task_fn_tx;
5958 		intr_fast = iflib_fast_intr;
5959 		GROUPTASK_INIT(gtask, 0, fn, q);
5960 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5961 		break;
5962 	case IFLIB_INTR_RX:
5963 		q = &ctx->ifc_rxqs[qid];
5964 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5965 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5966 		tqg = qgroup_if_io_tqg;
5967 		fn = _task_fn_rx;
5968 		intr_fast = iflib_fast_intr;
5969 		GROUPTASK_INIT(gtask, 0, fn, q);
5970 		break;
5971 	case IFLIB_INTR_RXTX:
5972 		q = &ctx->ifc_rxqs[qid];
5973 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5974 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5975 		tqg = qgroup_if_io_tqg;
5976 		fn = _task_fn_rx;
5977 		intr_fast = iflib_fast_intr_rxtx;
5978 		GROUPTASK_INIT(gtask, 0, fn, q);
5979 		break;
5980 	case IFLIB_INTR_ADMIN:
5981 		q = ctx;
5982 		tqrid = -1;
5983 		info = &ctx->ifc_filter_info;
5984 		gtask = &ctx->ifc_admin_task;
5985 		tqg = qgroup_if_config_tqg;
5986 		fn = _task_fn_admin;
5987 		intr_fast = iflib_fast_intr_ctx;
5988 		break;
5989 	default:
5990 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5991 		    __func__);
5992 		return (EINVAL);
5993 	}
5994 
5995 	info->ifi_filter = filter;
5996 	info->ifi_filter_arg = filter_arg;
5997 	info->ifi_task = gtask;
5998 	info->ifi_ctx = q;
5999 
6000 	dev = ctx->ifc_dev;
6001 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6002 	if (err != 0) {
6003 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6004 		return (err);
6005 	}
6006 	if (type == IFLIB_INTR_ADMIN)
6007 		return (0);
6008 
6009 	if (tqrid != -1) {
6010 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6011 		    q, name);
6012 		if (err)
6013 			return (err);
6014 	} else {
6015 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6016 	}
6017 
6018 	return (0);
6019 }
6020 
6021 void
6022 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6023 {
6024 	struct grouptask *gtask;
6025 	struct taskqgroup *tqg;
6026 	gtask_fn_t *fn;
6027 	void *q;
6028 	int err;
6029 
6030 	switch (type) {
6031 	case IFLIB_INTR_TX:
6032 		q = &ctx->ifc_txqs[qid];
6033 		gtask = &ctx->ifc_txqs[qid].ift_task;
6034 		tqg = qgroup_if_io_tqg;
6035 		fn = _task_fn_tx;
6036 		break;
6037 	case IFLIB_INTR_RX:
6038 		q = &ctx->ifc_rxqs[qid];
6039 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6040 		tqg = qgroup_if_io_tqg;
6041 		fn = _task_fn_rx;
6042 		break;
6043 	case IFLIB_INTR_IOV:
6044 		q = ctx;
6045 		gtask = &ctx->ifc_vflr_task;
6046 		tqg = qgroup_if_config_tqg;
6047 		fn = _task_fn_iov;
6048 		break;
6049 	default:
6050 		panic("unknown net intr type");
6051 	}
6052 	GROUPTASK_INIT(gtask, 0, fn, q);
6053 	if (irq != NULL) {
6054 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6055 		    q, name);
6056 		if (err)
6057 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6058 			    irq->ii_res, name);
6059 	} else {
6060 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6061 	}
6062 }
6063 
6064 void
6065 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6066 {
6067 
6068 	if (irq->ii_tag)
6069 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6070 
6071 	if (irq->ii_res)
6072 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6073 		    rman_get_rid(irq->ii_res), irq->ii_res);
6074 }
6075 
6076 static int
6077 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6078 {
6079 	iflib_txq_t txq = ctx->ifc_txqs;
6080 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6081 	if_irq_t irq = &ctx->ifc_legacy_irq;
6082 	iflib_filter_info_t info;
6083 	device_t dev;
6084 	struct grouptask *gtask;
6085 	struct resource *res;
6086 	struct taskqgroup *tqg;
6087 	gtask_fn_t *fn;
6088 	void *q;
6089 	int err, tqrid;
6090 	bool rx_only;
6091 
6092 	q = &ctx->ifc_rxqs[0];
6093 	info = &rxq[0].ifr_filter_info;
6094 	gtask = &rxq[0].ifr_task;
6095 	tqg = qgroup_if_io_tqg;
6096 	tqrid = *rid;
6097 	fn = _task_fn_rx;
6098 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6099 
6100 	ctx->ifc_flags |= IFC_LEGACY;
6101 	info->ifi_filter = filter;
6102 	info->ifi_filter_arg = filter_arg;
6103 	info->ifi_task = gtask;
6104 	info->ifi_ctx = rx_only ? ctx : q;
6105 
6106 	dev = ctx->ifc_dev;
6107 	/* We allocate a single interrupt resource */
6108 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6109 	    iflib_fast_intr_rxtx, NULL, info, name);
6110 	if (err != 0)
6111 		return (err);
6112 	GROUPTASK_INIT(gtask, 0, fn, q);
6113 	res = irq->ii_res;
6114 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6115 
6116 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6117 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6118 	    "tx");
6119 	return (0);
6120 }
6121 
6122 void
6123 iflib_led_create(if_ctx_t ctx)
6124 {
6125 
6126 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6127 	    device_get_nameunit(ctx->ifc_dev));
6128 }
6129 
6130 void
6131 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6132 {
6133 
6134 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6135 }
6136 
6137 void
6138 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6139 {
6140 
6141 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6142 }
6143 
6144 void
6145 iflib_admin_intr_deferred(if_ctx_t ctx)
6146 {
6147 #ifdef INVARIANTS
6148 	struct grouptask *gtask;
6149 
6150 	gtask = &ctx->ifc_admin_task;
6151 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6152 #endif
6153 
6154 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6155 }
6156 
6157 void
6158 iflib_iov_intr_deferred(if_ctx_t ctx)
6159 {
6160 
6161 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6162 }
6163 
6164 void
6165 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6166 {
6167 
6168 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6169 	    name);
6170 }
6171 
6172 void
6173 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6174 	const char *name)
6175 {
6176 
6177 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6178 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6179 	    name);
6180 }
6181 
6182 void
6183 iflib_config_gtask_deinit(struct grouptask *gtask)
6184 {
6185 
6186 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6187 }
6188 
6189 void
6190 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6191 {
6192 	if_t ifp = ctx->ifc_ifp;
6193 	iflib_txq_t txq = ctx->ifc_txqs;
6194 
6195 	if_setbaudrate(ifp, baudrate);
6196 	if (baudrate >= IF_Gbps(10)) {
6197 		STATE_LOCK(ctx);
6198 		ctx->ifc_flags |= IFC_PREFETCH;
6199 		STATE_UNLOCK(ctx);
6200 	}
6201 	/* If link down, disable watchdog */
6202 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6203 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6204 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6205 	}
6206 	ctx->ifc_link_state = link_state;
6207 	if_link_state_change(ifp, link_state);
6208 }
6209 
6210 static int
6211 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6212 {
6213 	int credits;
6214 #ifdef INVARIANTS
6215 	int credits_pre = txq->ift_cidx_processed;
6216 #endif
6217 
6218 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6219 	    BUS_DMASYNC_POSTREAD);
6220 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6221 		return (0);
6222 
6223 	txq->ift_processed += credits;
6224 	txq->ift_cidx_processed += credits;
6225 
6226 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6227 	if (txq->ift_cidx_processed >= txq->ift_size)
6228 		txq->ift_cidx_processed -= txq->ift_size;
6229 	return (credits);
6230 }
6231 
6232 static int
6233 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6234 {
6235 	iflib_fl_t fl;
6236 	u_int i;
6237 
6238 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6239 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6240 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6241 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6242 	    budget));
6243 }
6244 
6245 void
6246 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6247 	const char *description, if_int_delay_info_t info,
6248 	int offset, int value)
6249 {
6250 	info->iidi_ctx = ctx;
6251 	info->iidi_offset = offset;
6252 	info->iidi_value = value;
6253 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6254 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6255 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6256 	    info, 0, iflib_sysctl_int_delay, "I", description);
6257 }
6258 
6259 struct sx *
6260 iflib_ctx_lock_get(if_ctx_t ctx)
6261 {
6262 
6263 	return (&ctx->ifc_ctx_sx);
6264 }
6265 
6266 static int
6267 iflib_msix_init(if_ctx_t ctx)
6268 {
6269 	device_t dev = ctx->ifc_dev;
6270 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6271 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6272 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6273 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6274 
6275 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6276 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6277 
6278 	if (bootverbose)
6279 		device_printf(dev, "msix_init qsets capped at %d\n",
6280 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6281 
6282 	/* Override by tuneable */
6283 	if (scctx->isc_disable_msix)
6284 		goto msi;
6285 
6286 	/* First try MSI-X */
6287 	if ((msgs = pci_msix_count(dev)) == 0) {
6288 		if (bootverbose)
6289 			device_printf(dev, "MSI-X not supported or disabled\n");
6290 		goto msi;
6291 	}
6292 
6293 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6294 	/*
6295 	 * bar == -1 => "trust me I know what I'm doing"
6296 	 * Some drivers are for hardware that is so shoddily
6297 	 * documented that no one knows which bars are which
6298 	 * so the developer has to map all bars. This hack
6299 	 * allows shoddy garbage to use MSI-X in this framework.
6300 	 */
6301 	if (bar != -1) {
6302 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6303 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6304 		if (ctx->ifc_msix_mem == NULL) {
6305 			device_printf(dev, "Unable to map MSI-X table\n");
6306 			goto msi;
6307 		}
6308 	}
6309 
6310 	admincnt = sctx->isc_admin_intrcnt;
6311 #if IFLIB_DEBUG
6312 	/* use only 1 qset in debug mode */
6313 	queuemsgs = min(msgs - admincnt, 1);
6314 #else
6315 	queuemsgs = msgs - admincnt;
6316 #endif
6317 #ifdef RSS
6318 	queues = imin(queuemsgs, rss_getnumbuckets());
6319 #else
6320 	queues = queuemsgs;
6321 #endif
6322 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6323 	if (bootverbose)
6324 		device_printf(dev,
6325 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6326 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6327 #ifdef  RSS
6328 	/* If we're doing RSS, clamp at the number of RSS buckets */
6329 	if (queues > rss_getnumbuckets())
6330 		queues = rss_getnumbuckets();
6331 #endif
6332 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6333 		rx_queues = iflib_num_rx_queues;
6334 	else
6335 		rx_queues = queues;
6336 
6337 	if (rx_queues > scctx->isc_nrxqsets)
6338 		rx_queues = scctx->isc_nrxqsets;
6339 
6340 	/*
6341 	 * We want this to be all logical CPUs by default
6342 	 */
6343 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6344 		tx_queues = iflib_num_tx_queues;
6345 	else
6346 		tx_queues = mp_ncpus;
6347 
6348 	if (tx_queues > scctx->isc_ntxqsets)
6349 		tx_queues = scctx->isc_ntxqsets;
6350 
6351 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6352 #ifdef INVARIANTS
6353 		if (tx_queues != rx_queues)
6354 			device_printf(dev,
6355 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6356 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6357 #endif
6358 		tx_queues = min(rx_queues, tx_queues);
6359 		rx_queues = min(rx_queues, tx_queues);
6360 	}
6361 
6362 	vectors = rx_queues + admincnt;
6363 	if (msgs < vectors) {
6364 		device_printf(dev,
6365 		    "insufficient number of MSI-X vectors "
6366 		    "(supported %d, need %d)\n", msgs, vectors);
6367 		goto msi;
6368 	}
6369 
6370 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6371 	    tx_queues);
6372 	msgs = vectors;
6373 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6374 		if (vectors != msgs) {
6375 			device_printf(dev,
6376 			    "Unable to allocate sufficient MSI-X vectors "
6377 			    "(got %d, need %d)\n", vectors, msgs);
6378 			pci_release_msi(dev);
6379 			if (bar != -1) {
6380 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6381 				    ctx->ifc_msix_mem);
6382 				ctx->ifc_msix_mem = NULL;
6383 			}
6384 			goto msi;
6385 		}
6386 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6387 		    vectors);
6388 		scctx->isc_vectors = vectors;
6389 		scctx->isc_nrxqsets = rx_queues;
6390 		scctx->isc_ntxqsets = tx_queues;
6391 		scctx->isc_intr = IFLIB_INTR_MSIX;
6392 
6393 		return (vectors);
6394 	} else {
6395 		device_printf(dev,
6396 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6397 		    err);
6398 		if (bar != -1) {
6399 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6400 			    ctx->ifc_msix_mem);
6401 			ctx->ifc_msix_mem = NULL;
6402 		}
6403 	}
6404 
6405 msi:
6406 	vectors = pci_msi_count(dev);
6407 	scctx->isc_nrxqsets = 1;
6408 	scctx->isc_ntxqsets = 1;
6409 	scctx->isc_vectors = vectors;
6410 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6411 		device_printf(dev,"Using an MSI interrupt\n");
6412 		scctx->isc_intr = IFLIB_INTR_MSI;
6413 	} else {
6414 		scctx->isc_vectors = 1;
6415 		device_printf(dev,"Using a Legacy interrupt\n");
6416 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6417 	}
6418 
6419 	return (vectors);
6420 }
6421 
6422 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6423 
6424 static int
6425 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6426 {
6427 	int rc;
6428 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6429 	struct sbuf *sb;
6430 	const char *ring_state = "UNKNOWN";
6431 
6432 	/* XXX needed ? */
6433 	rc = sysctl_wire_old_buffer(req, 0);
6434 	MPASS(rc == 0);
6435 	if (rc != 0)
6436 		return (rc);
6437 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6438 	MPASS(sb != NULL);
6439 	if (sb == NULL)
6440 		return (ENOMEM);
6441 	if (state[3] <= 3)
6442 		ring_state = ring_states[state[3]];
6443 
6444 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6445 		    state[0], state[1], state[2], ring_state);
6446 	rc = sbuf_finish(sb);
6447 	sbuf_delete(sb);
6448         return(rc);
6449 }
6450 
6451 enum iflib_ndesc_handler {
6452 	IFLIB_NTXD_HANDLER,
6453 	IFLIB_NRXD_HANDLER,
6454 };
6455 
6456 static int
6457 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6458 {
6459 	if_ctx_t ctx = (void *)arg1;
6460 	enum iflib_ndesc_handler type = arg2;
6461 	char buf[256] = {0};
6462 	qidx_t *ndesc;
6463 	char *p, *next;
6464 	int nqs, rc, i;
6465 
6466 	nqs = 8;
6467 	switch(type) {
6468 	case IFLIB_NTXD_HANDLER:
6469 		ndesc = ctx->ifc_sysctl_ntxds;
6470 		if (ctx->ifc_sctx)
6471 			nqs = ctx->ifc_sctx->isc_ntxqs;
6472 		break;
6473 	case IFLIB_NRXD_HANDLER:
6474 		ndesc = ctx->ifc_sysctl_nrxds;
6475 		if (ctx->ifc_sctx)
6476 			nqs = ctx->ifc_sctx->isc_nrxqs;
6477 		break;
6478 	default:
6479 		printf("%s: unhandled type\n", __func__);
6480 		return (EINVAL);
6481 	}
6482 	if (nqs == 0)
6483 		nqs = 8;
6484 
6485 	for (i=0; i<8; i++) {
6486 		if (i >= nqs)
6487 			break;
6488 		if (i)
6489 			strcat(buf, ",");
6490 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6491 	}
6492 
6493 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6494 	if (rc || req->newptr == NULL)
6495 		return rc;
6496 
6497 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6498 	    i++, p = strsep(&next, " ,")) {
6499 		ndesc[i] = strtoul(p, NULL, 10);
6500 	}
6501 
6502 	return(rc);
6503 }
6504 
6505 #define NAME_BUFLEN 32
6506 static void
6507 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6508 {
6509         device_t dev = iflib_get_dev(ctx);
6510 	struct sysctl_oid_list *child, *oid_list;
6511 	struct sysctl_ctx_list *ctx_list;
6512 	struct sysctl_oid *node;
6513 
6514 	ctx_list = device_get_sysctl_ctx(dev);
6515 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6516 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6517 						      CTLFLAG_RD, NULL, "IFLIB fields");
6518 	oid_list = SYSCTL_CHILDREN(node);
6519 
6520 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6521 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6522 		       "driver version");
6523 
6524 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6525 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6526 			"# of txqs to use, 0 => use default #");
6527 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6528 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6529 			"# of rxqs to use, 0 => use default #");
6530 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6531 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6532                        "permit #txq != #rxq");
6533 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6534                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6535                       "disable MSI-X (default 0)");
6536 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6537 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6538 		       "set the RX budget");
6539 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6540 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6541 		       "cause TX to abdicate instead of running to completion");
6542 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6543 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6544 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6545 		       "offset to start using cores at");
6546 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6547 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6548 		       "use separate cores for TX and RX");
6549 
6550 	/* XXX change for per-queue sizes */
6551 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6552 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6553                        mp_ndesc_handler, "A",
6554 		       "list of # of TX descriptors to use, 0 = use default #");
6555 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6556 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6557                        mp_ndesc_handler, "A",
6558 		       "list of # of RX descriptors to use, 0 = use default #");
6559 }
6560 
6561 static void
6562 iflib_add_device_sysctl_post(if_ctx_t ctx)
6563 {
6564 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6565 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6566         device_t dev = iflib_get_dev(ctx);
6567 	struct sysctl_oid_list *child;
6568 	struct sysctl_ctx_list *ctx_list;
6569 	iflib_fl_t fl;
6570 	iflib_txq_t txq;
6571 	iflib_rxq_t rxq;
6572 	int i, j;
6573 	char namebuf[NAME_BUFLEN];
6574 	char *qfmt;
6575 	struct sysctl_oid *queue_node, *fl_node, *node;
6576 	struct sysctl_oid_list *queue_list, *fl_list;
6577 	ctx_list = device_get_sysctl_ctx(dev);
6578 
6579 	node = ctx->ifc_sysctl_node;
6580 	child = SYSCTL_CHILDREN(node);
6581 
6582 	if (scctx->isc_ntxqsets > 100)
6583 		qfmt = "txq%03d";
6584 	else if (scctx->isc_ntxqsets > 10)
6585 		qfmt = "txq%02d";
6586 	else
6587 		qfmt = "txq%d";
6588 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6589 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6590 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6591 					     CTLFLAG_RD, NULL, "Queue Name");
6592 		queue_list = SYSCTL_CHILDREN(queue_node);
6593 #if MEMORY_LOGGING
6594 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6595 				CTLFLAG_RD,
6596 				&txq->ift_dequeued, "total mbufs freed");
6597 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6598 				CTLFLAG_RD,
6599 				&txq->ift_enqueued, "total mbufs enqueued");
6600 #endif
6601 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6602 				   CTLFLAG_RD,
6603 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6604 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6605 				   CTLFLAG_RD,
6606 				   &txq->ift_pullups, "# of times m_pullup was called");
6607 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6608 				   CTLFLAG_RD,
6609 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6610 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6611 				   CTLFLAG_RD,
6612 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6613 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6614 				   CTLFLAG_RD,
6615 				   &txq->ift_map_failed, "# of times DMA map failed");
6616 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6617 				   CTLFLAG_RD,
6618 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6619 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6620 				   CTLFLAG_RD,
6621 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6622 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6623 				   CTLFLAG_RD,
6624 				   &txq->ift_pidx, 1, "Producer Index");
6625 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6626 				   CTLFLAG_RD,
6627 				   &txq->ift_cidx, 1, "Consumer Index");
6628 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6629 				   CTLFLAG_RD,
6630 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6631 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6632 				   CTLFLAG_RD,
6633 				   &txq->ift_in_use, 1, "descriptors in use");
6634 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6635 				   CTLFLAG_RD,
6636 				   &txq->ift_processed, "descriptors procesed for clean");
6637 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6638 				   CTLFLAG_RD,
6639 				   &txq->ift_cleaned, "total cleaned");
6640 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6641 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6642 				0, mp_ring_state_handler, "A", "soft ring state");
6643 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6644 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6645 				       "# of enqueues to the mp_ring for this queue");
6646 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6647 				       CTLFLAG_RD, &txq->ift_br->drops,
6648 				       "# of drops in the mp_ring for this queue");
6649 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6650 				       CTLFLAG_RD, &txq->ift_br->starts,
6651 				       "# of normal consumer starts in the mp_ring for this queue");
6652 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6653 				       CTLFLAG_RD, &txq->ift_br->stalls,
6654 					       "# of consumer stalls in the mp_ring for this queue");
6655 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6656 			       CTLFLAG_RD, &txq->ift_br->restarts,
6657 				       "# of consumer restarts in the mp_ring for this queue");
6658 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6659 				       CTLFLAG_RD, &txq->ift_br->abdications,
6660 				       "# of consumer abdications in the mp_ring for this queue");
6661 	}
6662 
6663 	if (scctx->isc_nrxqsets > 100)
6664 		qfmt = "rxq%03d";
6665 	else if (scctx->isc_nrxqsets > 10)
6666 		qfmt = "rxq%02d";
6667 	else
6668 		qfmt = "rxq%d";
6669 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6670 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6671 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6672 					     CTLFLAG_RD, NULL, "Queue Name");
6673 		queue_list = SYSCTL_CHILDREN(queue_node);
6674 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6675 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6676 				       CTLFLAG_RD,
6677 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6678 		}
6679 
6680 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6681 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6682 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6683 						     CTLFLAG_RD, NULL, "freelist Name");
6684 			fl_list = SYSCTL_CHILDREN(fl_node);
6685 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6686 				       CTLFLAG_RD,
6687 				       &fl->ifl_pidx, 1, "Producer Index");
6688 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6689 				       CTLFLAG_RD,
6690 				       &fl->ifl_cidx, 1, "Consumer Index");
6691 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6692 				       CTLFLAG_RD,
6693 				       &fl->ifl_credits, 1, "credits available");
6694 #if MEMORY_LOGGING
6695 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6696 					CTLFLAG_RD,
6697 					&fl->ifl_m_enqueued, "mbufs allocated");
6698 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6699 					CTLFLAG_RD,
6700 					&fl->ifl_m_dequeued, "mbufs freed");
6701 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6702 					CTLFLAG_RD,
6703 					&fl->ifl_cl_enqueued, "clusters allocated");
6704 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6705 					CTLFLAG_RD,
6706 					&fl->ifl_cl_dequeued, "clusters freed");
6707 #endif
6708 
6709 		}
6710 	}
6711 
6712 }
6713 
6714 void
6715 iflib_request_reset(if_ctx_t ctx)
6716 {
6717 
6718 	STATE_LOCK(ctx);
6719 	ctx->ifc_flags |= IFC_DO_RESET;
6720 	STATE_UNLOCK(ctx);
6721 }
6722 
6723 #ifndef __NO_STRICT_ALIGNMENT
6724 static struct mbuf *
6725 iflib_fixup_rx(struct mbuf *m)
6726 {
6727 	struct mbuf *n;
6728 
6729 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6730 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6731 		m->m_data += ETHER_HDR_LEN;
6732 		n = m;
6733 	} else {
6734 		MGETHDR(n, M_NOWAIT, MT_DATA);
6735 		if (n == NULL) {
6736 			m_freem(m);
6737 			return (NULL);
6738 		}
6739 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6740 		m->m_data += ETHER_HDR_LEN;
6741 		m->m_len -= ETHER_HDR_LEN;
6742 		n->m_len = ETHER_HDR_LEN;
6743 		M_MOVE_PKTHDR(n, m);
6744 		n->m_next = m;
6745 	}
6746 	return (n);
6747 }
6748 #endif
6749 
6750 #ifdef DEBUGNET
6751 static void
6752 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6753 {
6754 	if_ctx_t ctx;
6755 
6756 	ctx = if_getsoftc(ifp);
6757 	CTX_LOCK(ctx);
6758 	*nrxr = NRXQSETS(ctx);
6759 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6760 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6761 	CTX_UNLOCK(ctx);
6762 }
6763 
6764 static void
6765 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6766 {
6767 	if_ctx_t ctx;
6768 	if_softc_ctx_t scctx;
6769 	iflib_fl_t fl;
6770 	iflib_rxq_t rxq;
6771 	int i, j;
6772 
6773 	ctx = if_getsoftc(ifp);
6774 	scctx = &ctx->ifc_softc_ctx;
6775 
6776 	switch (event) {
6777 	case DEBUGNET_START:
6778 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6779 			rxq = &ctx->ifc_rxqs[i];
6780 			for (j = 0; j < rxq->ifr_nfl; j++) {
6781 				fl = rxq->ifr_fl;
6782 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6783 			}
6784 		}
6785 		iflib_no_tx_batch = 1;
6786 		break;
6787 	default:
6788 		break;
6789 	}
6790 }
6791 
6792 static int
6793 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6794 {
6795 	if_ctx_t ctx;
6796 	iflib_txq_t txq;
6797 	int error;
6798 
6799 	ctx = if_getsoftc(ifp);
6800 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6801 	    IFF_DRV_RUNNING)
6802 		return (EBUSY);
6803 
6804 	txq = &ctx->ifc_txqs[0];
6805 	error = iflib_encap(txq, &m);
6806 	if (error == 0)
6807 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6808 	return (error);
6809 }
6810 
6811 static int
6812 iflib_debugnet_poll(if_t ifp, int count)
6813 {
6814 	if_ctx_t ctx;
6815 	if_softc_ctx_t scctx;
6816 	iflib_txq_t txq;
6817 	int i;
6818 
6819 	ctx = if_getsoftc(ifp);
6820 	scctx = &ctx->ifc_softc_ctx;
6821 
6822 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6823 	    IFF_DRV_RUNNING)
6824 		return (EBUSY);
6825 
6826 	txq = &ctx->ifc_txqs[0];
6827 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6828 
6829 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6830 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6831 	return (0);
6832 }
6833 #endif /* DEBUGNET */
6834