xref: /freebsd/sys/net/iflib.c (revision 52c81be11a107cdedb865a274b5567b0c95c0308)
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 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t	ift_buf_tag;
359 	bus_dma_tag_t	ift_tso_buf_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define	MTX_NAME_LEN	32
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the command queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 
480 static inline void
481 pkt_info_zero(if_pkt_info_t pi)
482 {
483 	if_pkt_info_pad_t pi_pad;
484 
485 	pi_pad = (if_pkt_info_pad_t)pi;
486 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
487 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
488 #ifndef __LP64__
489 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
490 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
491 #endif
492 }
493 
494 static device_method_t iflib_pseudo_methods[] = {
495 	DEVMETHOD(device_attach, noop_attach),
496 	DEVMETHOD(device_detach, iflib_pseudo_detach),
497 	DEVMETHOD_END
498 };
499 
500 driver_t iflib_pseudodriver = {
501 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
502 };
503 
504 static inline void
505 rxd_info_zero(if_rxd_info_t ri)
506 {
507 	if_rxd_info_pad_t ri_pad;
508 	int i;
509 
510 	ri_pad = (if_rxd_info_pad_t)ri;
511 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
512 		ri_pad->rxd_val[i] = 0;
513 		ri_pad->rxd_val[i+1] = 0;
514 		ri_pad->rxd_val[i+2] = 0;
515 		ri_pad->rxd_val[i+3] = 0;
516 	}
517 #ifdef __LP64__
518 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
519 #endif
520 }
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA (bus_addr_t)-1
527 
528 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
531 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
534 
535 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
536 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
537 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
538 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
539 
540 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
541 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
542 
543 void
544 iflib_set_detach(if_ctx_t ctx)
545 {
546 	STATE_LOCK(ctx);
547 	ctx->ifc_flags |= IFC_IN_DETACH;
548 	STATE_UNLOCK(ctx);
549 }
550 
551 /* Our boot-time initialization hook */
552 static int	iflib_module_event_handler(module_t, int, void *);
553 
554 static moduledata_t iflib_moduledata = {
555 	"iflib",
556 	iflib_module_event_handler,
557 	NULL
558 };
559 
560 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
561 MODULE_VERSION(iflib, 1);
562 
563 MODULE_DEPEND(iflib, pci, 1, 1, 1);
564 MODULE_DEPEND(iflib, ether, 1, 1, 1);
565 
566 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
567 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
568 
569 #ifndef IFLIB_DEBUG_COUNTERS
570 #ifdef INVARIANTS
571 #define IFLIB_DEBUG_COUNTERS 1
572 #else
573 #define IFLIB_DEBUG_COUNTERS 0
574 #endif /* !INVARIANTS */
575 #endif
576 
577 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
578     "iflib driver parameters");
579 
580 /*
581  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
582  */
583 static int iflib_min_tx_latency = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
585 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
586 static int iflib_no_tx_batch = 0;
587 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
588 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
589 
590 
591 #if IFLIB_DEBUG_COUNTERS
592 
593 static int iflib_tx_seen;
594 static int iflib_tx_sent;
595 static int iflib_tx_encap;
596 static int iflib_rx_allocs;
597 static int iflib_fl_refills;
598 static int iflib_fl_refills_large;
599 static int iflib_tx_frees;
600 
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
602 		   &iflib_tx_seen, 0, "# TX mbufs seen");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
604 		   &iflib_tx_sent, 0, "# TX mbufs sent");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
606 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
608 		   &iflib_tx_frees, 0, "# TX frees");
609 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
610 		   &iflib_rx_allocs, 0, "# RX allocations");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
612 		   &iflib_fl_refills, 0, "# refills");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
614 		   &iflib_fl_refills_large, 0, "# large refills");
615 
616 
617 static int iflib_txq_drain_flushing;
618 static int iflib_txq_drain_oactive;
619 static int iflib_txq_drain_notready;
620 
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
622 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
623 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
624 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
626 		   &iflib_txq_drain_notready, 0, "# drain notready");
627 
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 #ifdef DEV_NETMAP
737 #include <sys/selinfo.h>
738 #include <net/netmap.h>
739 #include <dev/netmap/netmap_kern.h>
740 
741 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
742 
743 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
744 
745 /*
746  * device-specific sysctl variables:
747  *
748  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
749  *	During regular operations the CRC is stripped, but on some
750  *	hardware reception of frames not multiple of 64 is slower,
751  *	so using crcstrip=0 helps in benchmarks.
752  *
753  * iflib_rx_miss, iflib_rx_miss_bufs:
754  *	count packets that might be missed due to lost interrupts.
755  */
756 SYSCTL_DECL(_dev_netmap);
757 /*
758  * The xl driver by default strips CRCs and we do not override it.
759  */
760 
761 int iflib_crcstrip = 1;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
763     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
764 
765 int iflib_rx_miss, iflib_rx_miss_bufs;
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
767     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
768 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
769     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
770 
771 /*
772  * Register/unregister. We are already under netmap lock.
773  * Only called on the first register or the last unregister.
774  */
775 static int
776 iflib_netmap_register(struct netmap_adapter *na, int onoff)
777 {
778 	if_t ifp = na->ifp;
779 	if_ctx_t ctx = ifp->if_softc;
780 	int status;
781 
782 	CTX_LOCK(ctx);
783 	IFDI_INTR_DISABLE(ctx);
784 
785 	/* Tell the stack that the interface is no longer active */
786 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
787 
788 	if (!CTX_IS_VF(ctx))
789 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
790 
791 	iflib_stop(ctx);
792 
793 	/*
794 	 * Enable (or disable) netmap flags, and intercept (or restore)
795 	 * ifp->if_transmit. This is done once the device has been stopped
796 	 * to prevent race conditions.
797 	 */
798 	if (onoff) {
799 		nm_set_native_flags(na);
800 	} else {
801 		nm_clear_native_flags(na);
802 	}
803 
804 	iflib_init_locked(ctx);
805 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
806 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
807 	if (status)
808 		nm_clear_native_flags(na);
809 	CTX_UNLOCK(ctx);
810 	return (status);
811 }
812 
813 static int
814 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
815 {
816 	struct netmap_adapter *na = kring->na;
817 	u_int const lim = kring->nkr_num_slots - 1;
818 	u_int head = kring->rhead;
819 	struct netmap_ring *ring = kring->ring;
820 	bus_dmamap_t *map;
821 	struct if_rxd_update iru;
822 	if_ctx_t ctx = rxq->ifr_ctx;
823 	iflib_fl_t fl = &rxq->ifr_fl[0];
824 	uint32_t refill_pidx, nic_i;
825 #if IFLIB_DEBUG_COUNTERS
826 	int rf_count = 0;
827 #endif
828 
829 	if (nm_i == head && __predict_true(!init))
830 		return 0;
831 	iru_init(&iru, rxq, 0 /* flid */);
832 	map = fl->ifl_sds.ifsd_map;
833 	refill_pidx = netmap_idx_k2n(kring, nm_i);
834 	/*
835 	 * IMPORTANT: we must leave one free slot in the ring,
836 	 * so move head back by one unit
837 	 */
838 	head = nm_prev(head, lim);
839 	nic_i = UINT_MAX;
840 	DBG_COUNTER_INC(fl_refills);
841 	while (nm_i != head) {
842 #if IFLIB_DEBUG_COUNTERS
843 		if (++rf_count == 9)
844 			DBG_COUNTER_INC(fl_refills_large);
845 #endif
846 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
847 			struct netmap_slot *slot = &ring->slot[nm_i];
848 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
849 			uint32_t nic_i_dma = refill_pidx;
850 			nic_i = netmap_idx_k2n(kring, nm_i);
851 
852 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
853 
854 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
855 			        return netmap_ring_reinit(kring);
856 
857 			fl->ifl_vm_addrs[tmp_pidx] = addr;
858 			if (__predict_false(init)) {
859 				netmap_load_map(na, fl->ifl_buf_tag,
860 				    map[nic_i], addr);
861 			} else if (slot->flags & NS_BUF_CHANGED) {
862 				/* buffer has changed, reload map */
863 				netmap_reload_map(na, fl->ifl_buf_tag,
864 				    map[nic_i], addr);
865 			}
866 			slot->flags &= ~NS_BUF_CHANGED;
867 
868 			nm_i = nm_next(nm_i, lim);
869 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
870 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
871 				continue;
872 
873 			iru.iru_pidx = refill_pidx;
874 			iru.iru_count = tmp_pidx+1;
875 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
876 			refill_pidx = nic_i;
877 			for (int n = 0; n < iru.iru_count; n++) {
878 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
879 						BUS_DMASYNC_PREREAD);
880 				/* XXX - change this to not use the netmap func*/
881 				nic_i_dma = nm_next(nic_i_dma, lim);
882 			}
883 		}
884 	}
885 	kring->nr_hwcur = head;
886 
887 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
888 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
889 	if (__predict_true(nic_i != UINT_MAX)) {
890 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
891 		DBG_COUNTER_INC(rxd_flush);
892 	}
893 	return (0);
894 }
895 
896 /*
897  * Reconcile kernel and user view of the transmit ring.
898  *
899  * All information is in the kring.
900  * Userspace wants to send packets up to the one before kring->rhead,
901  * kernel knows kring->nr_hwcur is the first unsent packet.
902  *
903  * Here we push packets out (as many as possible), and possibly
904  * reclaim buffers from previously completed transmission.
905  *
906  * The caller (netmap) guarantees that there is only one instance
907  * running at any time. Any interference with other driver
908  * methods should be handled by the individual drivers.
909  */
910 static int
911 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
912 {
913 	struct netmap_adapter *na = kring->na;
914 	if_t ifp = na->ifp;
915 	struct netmap_ring *ring = kring->ring;
916 	u_int nm_i;	/* index into the netmap kring */
917 	u_int nic_i;	/* index into the NIC ring */
918 	u_int n;
919 	u_int const lim = kring->nkr_num_slots - 1;
920 	u_int const head = kring->rhead;
921 	struct if_pkt_info pi;
922 
923 	/*
924 	 * interrupts on every tx packet are expensive so request
925 	 * them every half ring, or where NS_REPORT is set
926 	 */
927 	u_int report_frequency = kring->nkr_num_slots >> 1;
928 	/* device-specific */
929 	if_ctx_t ctx = ifp->if_softc;
930 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
931 
932 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
933 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
934 
935 	/*
936 	 * First part: process new packets to send.
937 	 * nm_i is the current index in the netmap kring,
938 	 * nic_i is the corresponding index in the NIC ring.
939 	 *
940 	 * If we have packets to send (nm_i != head)
941 	 * iterate over the netmap ring, fetch length and update
942 	 * the corresponding slot in the NIC ring. Some drivers also
943 	 * need to update the buffer's physical address in the NIC slot
944 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
945 	 *
946 	 * The netmap_reload_map() calls is especially expensive,
947 	 * even when (as in this case) the tag is 0, so do only
948 	 * when the buffer has actually changed.
949 	 *
950 	 * If possible do not set the report/intr bit on all slots,
951 	 * but only a few times per ring or when NS_REPORT is set.
952 	 *
953 	 * Finally, on 10G and faster drivers, it might be useful
954 	 * to prefetch the next slot and txr entry.
955 	 */
956 
957 	nm_i = kring->nr_hwcur;
958 	if (nm_i != head) {	/* we have new packets to send */
959 		pkt_info_zero(&pi);
960 		pi.ipi_segs = txq->ift_segs;
961 		pi.ipi_qsidx = kring->ring_id;
962 		nic_i = netmap_idx_k2n(kring, nm_i);
963 
964 		__builtin_prefetch(&ring->slot[nm_i]);
965 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
966 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
967 
968 		for (n = 0; nm_i != head; n++) {
969 			struct netmap_slot *slot = &ring->slot[nm_i];
970 			u_int len = slot->len;
971 			uint64_t paddr;
972 			void *addr = PNMB(na, slot, &paddr);
973 			int flags = (slot->flags & NS_REPORT ||
974 				nic_i == 0 || nic_i == report_frequency) ?
975 				IPI_TX_INTR : 0;
976 
977 			/* device-specific */
978 			pi.ipi_len = len;
979 			pi.ipi_segs[0].ds_addr = paddr;
980 			pi.ipi_segs[0].ds_len = len;
981 			pi.ipi_nsegs = 1;
982 			pi.ipi_ndescs = 0;
983 			pi.ipi_pidx = nic_i;
984 			pi.ipi_flags = flags;
985 
986 			/* Fill the slot in the NIC ring. */
987 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
988 			DBG_COUNTER_INC(tx_encap);
989 
990 			/* prefetch for next round */
991 			__builtin_prefetch(&ring->slot[nm_i + 1]);
992 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
993 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
994 
995 			NM_CHECK_ADDR_LEN(na, addr, len);
996 
997 			if (slot->flags & NS_BUF_CHANGED) {
998 				/* buffer has changed, reload map */
999 				netmap_reload_map(na, txq->ift_buf_tag,
1000 				    txq->ift_sds.ifsd_map[nic_i], addr);
1001 			}
1002 			/* make sure changes to the buffer are synced */
1003 			bus_dmamap_sync(txq->ift_buf_tag,
1004 			    txq->ift_sds.ifsd_map[nic_i],
1005 			    BUS_DMASYNC_PREWRITE);
1006 
1007 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1008 			nm_i = nm_next(nm_i, lim);
1009 			nic_i = nm_next(nic_i, lim);
1010 		}
1011 		kring->nr_hwcur = nm_i;
1012 
1013 		/* synchronize the NIC ring */
1014 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1015 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1016 
1017 		/* (re)start the tx unit up to slot nic_i (excluded) */
1018 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1019 	}
1020 
1021 	/*
1022 	 * Second part: reclaim buffers for completed transmissions.
1023 	 *
1024 	 * If there are unclaimed buffers, attempt to reclaim them.
1025 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1026 	 * minimal delay, then trigger the tx handler which will spin in the
1027 	 * group task queue.
1028 	 */
1029 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1030 		if (iflib_tx_credits_update(ctx, txq)) {
1031 			/* some tx completed, increment avail */
1032 			nic_i = txq->ift_cidx_processed;
1033 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1034 		}
1035 	}
1036 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1037 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1038 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1039 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1040 	}
1041 	return (0);
1042 }
1043 
1044 /*
1045  * Reconcile kernel and user view of the receive ring.
1046  * Same as for the txsync, this routine must be efficient.
1047  * The caller guarantees a single invocations, but races against
1048  * the rest of the driver should be handled here.
1049  *
1050  * On call, kring->rhead is the first packet that userspace wants
1051  * to keep, and kring->rcur is the wakeup point.
1052  * The kernel has previously reported packets up to kring->rtail.
1053  *
1054  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1055  * of whether or not we received an interrupt.
1056  */
1057 static int
1058 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1059 {
1060 	struct netmap_adapter *na = kring->na;
1061 	struct netmap_ring *ring = kring->ring;
1062 	if_t ifp = na->ifp;
1063 	iflib_fl_t fl;
1064 	uint32_t nm_i;	/* index into the netmap ring */
1065 	uint32_t nic_i;	/* index into the NIC ring */
1066 	u_int i, n;
1067 	u_int const lim = kring->nkr_num_slots - 1;
1068 	u_int const head = kring->rhead;
1069 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1070 	struct if_rxd_info ri;
1071 
1072 	if_ctx_t ctx = ifp->if_softc;
1073 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1074 	if (head > lim)
1075 		return netmap_ring_reinit(kring);
1076 
1077 	/*
1078 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1079 	 */
1080 
1081 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1082 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1083 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1084 	}
1085 
1086 	/*
1087 	 * First part: import newly received packets.
1088 	 *
1089 	 * nm_i is the index of the next free slot in the netmap ring,
1090 	 * nic_i is the index of the next received packet in the NIC ring,
1091 	 * and they may differ in case if_init() has been called while
1092 	 * in netmap mode. For the receive ring we have
1093 	 *
1094 	 *	nic_i = rxr->next_check;
1095 	 *	nm_i = kring->nr_hwtail (previous)
1096 	 * and
1097 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1098 	 *
1099 	 * rxr->next_check is set to 0 on a ring reinit
1100 	 */
1101 	if (netmap_no_pendintr || force_update) {
1102 		int crclen = iflib_crcstrip ? 0 : 4;
1103 		int error, avail;
1104 
1105 		for (i = 0; i < rxq->ifr_nfl; i++) {
1106 			fl = &rxq->ifr_fl[i];
1107 			nic_i = fl->ifl_cidx;
1108 			nm_i = netmap_idx_n2k(kring, nic_i);
1109 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1110 			    rxq->ifr_id, nic_i, USHRT_MAX);
1111 			for (n = 0; avail > 0; n++, avail--) {
1112 				rxd_info_zero(&ri);
1113 				ri.iri_frags = rxq->ifr_frags;
1114 				ri.iri_qsidx = kring->ring_id;
1115 				ri.iri_ifp = ctx->ifc_ifp;
1116 				ri.iri_cidx = nic_i;
1117 
1118 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1119 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1120 				ring->slot[nm_i].flags = 0;
1121 				bus_dmamap_sync(fl->ifl_buf_tag,
1122 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1123 				nm_i = nm_next(nm_i, lim);
1124 				nic_i = nm_next(nic_i, lim);
1125 			}
1126 			if (n) { /* update the state variables */
1127 				if (netmap_no_pendintr && !force_update) {
1128 					/* diagnostics */
1129 					iflib_rx_miss ++;
1130 					iflib_rx_miss_bufs += n;
1131 				}
1132 				fl->ifl_cidx = nic_i;
1133 				kring->nr_hwtail = nm_i;
1134 			}
1135 			kring->nr_kflags &= ~NKR_PENDINTR;
1136 		}
1137 	}
1138 	/*
1139 	 * Second part: skip past packets that userspace has released.
1140 	 * (kring->nr_hwcur to head excluded),
1141 	 * and make the buffers available for reception.
1142 	 * As usual nm_i is the index in the netmap ring,
1143 	 * nic_i is the index in the NIC ring, and
1144 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1145 	 */
1146 	/* XXX not sure how this will work with multiple free lists */
1147 	nm_i = kring->nr_hwcur;
1148 
1149 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1150 }
1151 
1152 static void
1153 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1154 {
1155 	if_ctx_t ctx = na->ifp->if_softc;
1156 
1157 	CTX_LOCK(ctx);
1158 	if (onoff) {
1159 		IFDI_INTR_ENABLE(ctx);
1160 	} else {
1161 		IFDI_INTR_DISABLE(ctx);
1162 	}
1163 	CTX_UNLOCK(ctx);
1164 }
1165 
1166 
1167 static int
1168 iflib_netmap_attach(if_ctx_t ctx)
1169 {
1170 	struct netmap_adapter na;
1171 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1172 
1173 	bzero(&na, sizeof(na));
1174 
1175 	na.ifp = ctx->ifc_ifp;
1176 	na.na_flags = NAF_BDG_MAYSLEEP;
1177 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1178 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1179 
1180 	na.num_tx_desc = scctx->isc_ntxd[0];
1181 	na.num_rx_desc = scctx->isc_nrxd[0];
1182 	na.nm_txsync = iflib_netmap_txsync;
1183 	na.nm_rxsync = iflib_netmap_rxsync;
1184 	na.nm_register = iflib_netmap_register;
1185 	na.nm_intr = iflib_netmap_intr;
1186 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1187 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1188 	return (netmap_attach(&na));
1189 }
1190 
1191 static void
1192 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1193 {
1194 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1195 	struct netmap_slot *slot;
1196 
1197 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1198 	if (slot == NULL)
1199 		return;
1200 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1201 
1202 		/*
1203 		 * In netmap mode, set the map for the packet buffer.
1204 		 * NOTE: Some drivers (not this one) also need to set
1205 		 * the physical buffer address in the NIC ring.
1206 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1207 		 * netmap slot index, si
1208 		 */
1209 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1210 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1211 		    NMB(na, slot + si));
1212 	}
1213 }
1214 
1215 static void
1216 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1217 {
1218 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1219 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1220 	struct netmap_slot *slot;
1221 	uint32_t nm_i;
1222 
1223 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1224 	if (slot == NULL)
1225 		return;
1226 	nm_i = netmap_idx_n2k(kring, 0);
1227 	netmap_fl_refill(rxq, kring, nm_i, true);
1228 }
1229 
1230 static void
1231 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1232 {
1233 	struct netmap_kring *kring;
1234 	uint16_t txqid;
1235 
1236 	txqid = txq->ift_id;
1237 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1238 
1239 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1240 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1241 		    BUS_DMASYNC_POSTREAD);
1242 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1243 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1244 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1245 			if (hz < 2000)
1246 				*reset_on = 1;
1247 			else
1248 				*reset_on = hz / 1000;
1249 		}
1250 	}
1251 }
1252 
1253 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1254 
1255 #else
1256 #define iflib_netmap_txq_init(ctx, txq)
1257 #define iflib_netmap_rxq_init(ctx, rxq)
1258 #define iflib_netmap_detach(ifp)
1259 
1260 #define iflib_netmap_attach(ctx) (0)
1261 #define netmap_rx_irq(ifp, qid, budget) (0)
1262 #define netmap_tx_irq(ifp, qid) do {} while (0)
1263 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1264 #endif
1265 
1266 #if defined(__i386__) || defined(__amd64__)
1267 static __inline void
1268 prefetch(void *x)
1269 {
1270 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1271 }
1272 static __inline void
1273 prefetch2cachelines(void *x)
1274 {
1275 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1276 #if (CACHE_LINE_SIZE < 128)
1277 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1278 #endif
1279 }
1280 #else
1281 #define prefetch(x)
1282 #define prefetch2cachelines(x)
1283 #endif
1284 
1285 static void
1286 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1287 {
1288 	iflib_fl_t fl;
1289 
1290 	fl = &rxq->ifr_fl[flid];
1291 	iru->iru_paddrs = fl->ifl_bus_addrs;
1292 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1293 	iru->iru_idxs = fl->ifl_rxd_idxs;
1294 	iru->iru_qsidx = rxq->ifr_id;
1295 	iru->iru_buf_size = fl->ifl_buf_size;
1296 	iru->iru_flidx = fl->ifl_id;
1297 }
1298 
1299 static void
1300 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1301 {
1302 	if (err)
1303 		return;
1304 	*(bus_addr_t *) arg = segs[0].ds_addr;
1305 }
1306 
1307 int
1308 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1309 {
1310 	int err;
1311 	device_t dev = ctx->ifc_dev;
1312 
1313 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1314 				align, 0,		/* alignment, bounds */
1315 				BUS_SPACE_MAXADDR,	/* lowaddr */
1316 				BUS_SPACE_MAXADDR,	/* highaddr */
1317 				NULL, NULL,		/* filter, filterarg */
1318 				size,			/* maxsize */
1319 				1,			/* nsegments */
1320 				size,			/* maxsegsize */
1321 				BUS_DMA_ALLOCNOW,	/* flags */
1322 				NULL,			/* lockfunc */
1323 				NULL,			/* lockarg */
1324 				&dma->idi_tag);
1325 	if (err) {
1326 		device_printf(dev,
1327 		    "%s: bus_dma_tag_create failed: %d\n",
1328 		    __func__, err);
1329 		goto fail_0;
1330 	}
1331 
1332 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1333 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1334 	if (err) {
1335 		device_printf(dev,
1336 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1337 		    __func__, (uintmax_t)size, err);
1338 		goto fail_1;
1339 	}
1340 
1341 	dma->idi_paddr = IF_BAD_DMA;
1342 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1343 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1344 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1345 		device_printf(dev,
1346 		    "%s: bus_dmamap_load failed: %d\n",
1347 		    __func__, err);
1348 		goto fail_2;
1349 	}
1350 
1351 	dma->idi_size = size;
1352 	return (0);
1353 
1354 fail_2:
1355 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1356 fail_1:
1357 	bus_dma_tag_destroy(dma->idi_tag);
1358 fail_0:
1359 	dma->idi_tag = NULL;
1360 
1361 	return (err);
1362 }
1363 
1364 int
1365 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1366 {
1367 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1368 
1369 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1370 
1371 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1372 }
1373 
1374 int
1375 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1376 {
1377 	int i, err;
1378 	iflib_dma_info_t *dmaiter;
1379 
1380 	dmaiter = dmalist;
1381 	for (i = 0; i < count; i++, dmaiter++) {
1382 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1383 			break;
1384 	}
1385 	if (err)
1386 		iflib_dma_free_multi(dmalist, i);
1387 	return (err);
1388 }
1389 
1390 void
1391 iflib_dma_free(iflib_dma_info_t dma)
1392 {
1393 	if (dma->idi_tag == NULL)
1394 		return;
1395 	if (dma->idi_paddr != IF_BAD_DMA) {
1396 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1397 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1398 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1399 		dma->idi_paddr = IF_BAD_DMA;
1400 	}
1401 	if (dma->idi_vaddr != NULL) {
1402 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1403 		dma->idi_vaddr = NULL;
1404 	}
1405 	bus_dma_tag_destroy(dma->idi_tag);
1406 	dma->idi_tag = NULL;
1407 }
1408 
1409 void
1410 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1411 {
1412 	int i;
1413 	iflib_dma_info_t *dmaiter = dmalist;
1414 
1415 	for (i = 0; i < count; i++, dmaiter++)
1416 		iflib_dma_free(*dmaiter);
1417 }
1418 
1419 static int
1420 iflib_fast_intr(void *arg)
1421 {
1422 	iflib_filter_info_t info = arg;
1423 	struct grouptask *gtask = info->ifi_task;
1424 	int result;
1425 
1426 	DBG_COUNTER_INC(fast_intrs);
1427 	if (info->ifi_filter != NULL) {
1428 		result = info->ifi_filter(info->ifi_filter_arg);
1429 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1430 			return (result);
1431 	}
1432 
1433 	GROUPTASK_ENQUEUE(gtask);
1434 	return (FILTER_HANDLED);
1435 }
1436 
1437 static int
1438 iflib_fast_intr_rxtx(void *arg)
1439 {
1440 	iflib_filter_info_t info = arg;
1441 	struct grouptask *gtask = info->ifi_task;
1442 	if_ctx_t ctx;
1443 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1444 	iflib_txq_t txq;
1445 	void *sc;
1446 	int i, cidx, result;
1447 	qidx_t txqid;
1448 	bool intr_enable, intr_legacy;
1449 
1450 	DBG_COUNTER_INC(fast_intrs);
1451 	if (info->ifi_filter != NULL) {
1452 		result = info->ifi_filter(info->ifi_filter_arg);
1453 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1454 			return (result);
1455 	}
1456 
1457 	ctx = rxq->ifr_ctx;
1458 	sc = ctx->ifc_softc;
1459 	intr_enable = false;
1460 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1461 	MPASS(rxq->ifr_ntxqirq);
1462 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1463 		txqid = rxq->ifr_txqid[i];
1464 		txq = &ctx->ifc_txqs[txqid];
1465 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1466 		    BUS_DMASYNC_POSTREAD);
1467 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1468 			if (intr_legacy)
1469 				intr_enable = true;
1470 			else
1471 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1472 			continue;
1473 		}
1474 		GROUPTASK_ENQUEUE(&txq->ift_task);
1475 	}
1476 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1477 		cidx = rxq->ifr_cq_cidx;
1478 	else
1479 		cidx = rxq->ifr_fl[0].ifl_cidx;
1480 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1481 		GROUPTASK_ENQUEUE(gtask);
1482 	else {
1483 		if (intr_legacy)
1484 			intr_enable = true;
1485 		else
1486 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1487 		DBG_COUNTER_INC(rx_intr_enables);
1488 	}
1489 	if (intr_enable)
1490 		IFDI_INTR_ENABLE(ctx);
1491 	return (FILTER_HANDLED);
1492 }
1493 
1494 
1495 static int
1496 iflib_fast_intr_ctx(void *arg)
1497 {
1498 	iflib_filter_info_t info = arg;
1499 	struct grouptask *gtask = info->ifi_task;
1500 	int result;
1501 
1502 	DBG_COUNTER_INC(fast_intrs);
1503 	if (info->ifi_filter != NULL) {
1504 		result = info->ifi_filter(info->ifi_filter_arg);
1505 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1506 			return (result);
1507 	}
1508 
1509 	GROUPTASK_ENQUEUE(gtask);
1510 	return (FILTER_HANDLED);
1511 }
1512 
1513 static int
1514 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1515 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1516 		 const char *name)
1517 {
1518 	struct resource *res;
1519 	void *tag = NULL;
1520 	device_t dev = ctx->ifc_dev;
1521 	int flags, i, rc;
1522 
1523 	flags = RF_ACTIVE;
1524 	if (ctx->ifc_flags & IFC_LEGACY)
1525 		flags |= RF_SHAREABLE;
1526 	MPASS(rid < 512);
1527 	i = rid;
1528 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1529 	if (res == NULL) {
1530 		device_printf(dev,
1531 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1532 		return (ENOMEM);
1533 	}
1534 	irq->ii_res = res;
1535 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1536 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1537 						filter, handler, arg, &tag);
1538 	if (rc != 0) {
1539 		device_printf(dev,
1540 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1541 					  rid, name ? name : "unknown", rc);
1542 		return (rc);
1543 	} else if (name)
1544 		bus_describe_intr(dev, res, tag, "%s", name);
1545 
1546 	irq->ii_tag = tag;
1547 	return (0);
1548 }
1549 
1550 /*********************************************************************
1551  *
1552  *  Allocate DMA resources for TX buffers as well as memory for the TX
1553  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1554  *  iflib_sw_tx_desc_array structure, storing all the information that
1555  *  is needed to transmit a packet on the wire.  This is called only
1556  *  once at attach, setup is done every reset.
1557  *
1558  **********************************************************************/
1559 static int
1560 iflib_txsd_alloc(iflib_txq_t txq)
1561 {
1562 	if_ctx_t ctx = txq->ift_ctx;
1563 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1564 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1565 	device_t dev = ctx->ifc_dev;
1566 	bus_size_t tsomaxsize;
1567 	int err, nsegments, ntsosegments;
1568 	bool tso;
1569 
1570 	nsegments = scctx->isc_tx_nsegments;
1571 	ntsosegments = scctx->isc_tx_tso_segments_max;
1572 	tsomaxsize = scctx->isc_tx_tso_size_max;
1573 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1574 		tsomaxsize += sizeof(struct ether_vlan_header);
1575 	MPASS(scctx->isc_ntxd[0] > 0);
1576 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1577 	MPASS(nsegments > 0);
1578 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1579 		MPASS(ntsosegments > 0);
1580 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1581 	}
1582 
1583 	/*
1584 	 * Set up DMA tags for TX buffers.
1585 	 */
1586 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1587 			       1, 0,			/* alignment, bounds */
1588 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1589 			       BUS_SPACE_MAXADDR,	/* highaddr */
1590 			       NULL, NULL,		/* filter, filterarg */
1591 			       sctx->isc_tx_maxsize,		/* maxsize */
1592 			       nsegments,	/* nsegments */
1593 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1594 			       0,			/* flags */
1595 			       NULL,			/* lockfunc */
1596 			       NULL,			/* lockfuncarg */
1597 			       &txq->ift_buf_tag))) {
1598 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1599 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1600 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1601 		goto fail;
1602 	}
1603 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1604 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1605 			       1, 0,			/* alignment, bounds */
1606 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1607 			       BUS_SPACE_MAXADDR,	/* highaddr */
1608 			       NULL, NULL,		/* filter, filterarg */
1609 			       tsomaxsize,		/* maxsize */
1610 			       ntsosegments,	/* nsegments */
1611 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1612 			       0,			/* flags */
1613 			       NULL,			/* lockfunc */
1614 			       NULL,			/* lockfuncarg */
1615 			       &txq->ift_tso_buf_tag))) {
1616 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1617 		    err);
1618 		goto fail;
1619 	}
1620 
1621 	/* Allocate memory for the TX mbuf map. */
1622 	if (!(txq->ift_sds.ifsd_m =
1623 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1624 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1625 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1626 		err = ENOMEM;
1627 		goto fail;
1628 	}
1629 
1630 	/*
1631 	 * Create the DMA maps for TX buffers.
1632 	 */
1633 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1634 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1635 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1636 		device_printf(dev,
1637 		    "Unable to allocate TX buffer DMA map memory\n");
1638 		err = ENOMEM;
1639 		goto fail;
1640 	}
1641 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1642 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1643 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1644 		device_printf(dev,
1645 		    "Unable to allocate TSO TX buffer map memory\n");
1646 		err = ENOMEM;
1647 		goto fail;
1648 	}
1649 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1650 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1651 		    &txq->ift_sds.ifsd_map[i]);
1652 		if (err != 0) {
1653 			device_printf(dev, "Unable to create TX DMA map\n");
1654 			goto fail;
1655 		}
1656 		if (!tso)
1657 			continue;
1658 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1659 		    &txq->ift_sds.ifsd_tso_map[i]);
1660 		if (err != 0) {
1661 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1662 			goto fail;
1663 		}
1664 	}
1665 	return (0);
1666 fail:
1667 	/* We free all, it handles case where we are in the middle */
1668 	iflib_tx_structures_free(ctx);
1669 	return (err);
1670 }
1671 
1672 static void
1673 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1674 {
1675 	bus_dmamap_t map;
1676 
1677 	if (txq->ift_sds.ifsd_map != NULL) {
1678 		map = txq->ift_sds.ifsd_map[i];
1679 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1680 		bus_dmamap_unload(txq->ift_buf_tag, map);
1681 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1682 		txq->ift_sds.ifsd_map[i] = NULL;
1683 	}
1684 
1685 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1686 		map = txq->ift_sds.ifsd_tso_map[i];
1687 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1688 		    BUS_DMASYNC_POSTWRITE);
1689 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1690 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1691 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1692 	}
1693 }
1694 
1695 static void
1696 iflib_txq_destroy(iflib_txq_t txq)
1697 {
1698 	if_ctx_t ctx = txq->ift_ctx;
1699 
1700 	for (int i = 0; i < txq->ift_size; i++)
1701 		iflib_txsd_destroy(ctx, txq, i);
1702 
1703 	if (txq->ift_br != NULL) {
1704 		ifmp_ring_free(txq->ift_br);
1705 		txq->ift_br = NULL;
1706 	}
1707 
1708 	mtx_destroy(&txq->ift_mtx);
1709 
1710 	if (txq->ift_sds.ifsd_map != NULL) {
1711 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1712 		txq->ift_sds.ifsd_map = NULL;
1713 	}
1714 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1715 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1716 		txq->ift_sds.ifsd_tso_map = NULL;
1717 	}
1718 	if (txq->ift_sds.ifsd_m != NULL) {
1719 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1720 		txq->ift_sds.ifsd_m = NULL;
1721 	}
1722 	if (txq->ift_buf_tag != NULL) {
1723 		bus_dma_tag_destroy(txq->ift_buf_tag);
1724 		txq->ift_buf_tag = NULL;
1725 	}
1726 	if (txq->ift_tso_buf_tag != NULL) {
1727 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1728 		txq->ift_tso_buf_tag = NULL;
1729 	}
1730 	if (txq->ift_ifdi != NULL) {
1731 		free(txq->ift_ifdi, M_IFLIB);
1732 	}
1733 }
1734 
1735 static void
1736 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1737 {
1738 	struct mbuf **mp;
1739 
1740 	mp = &txq->ift_sds.ifsd_m[i];
1741 	if (*mp == NULL)
1742 		return;
1743 
1744 	if (txq->ift_sds.ifsd_map != NULL) {
1745 		bus_dmamap_sync(txq->ift_buf_tag,
1746 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1747 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1748 	}
1749 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1750 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1751 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1752 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1753 		    txq->ift_sds.ifsd_tso_map[i]);
1754 	}
1755 	m_free(*mp);
1756 	DBG_COUNTER_INC(tx_frees);
1757 	*mp = NULL;
1758 }
1759 
1760 static int
1761 iflib_txq_setup(iflib_txq_t txq)
1762 {
1763 	if_ctx_t ctx = txq->ift_ctx;
1764 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1765 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1766 	iflib_dma_info_t di;
1767 	int i;
1768 
1769 	/* Set number of descriptors available */
1770 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1771 	/* XXX make configurable */
1772 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1773 
1774 	/* Reset indices */
1775 	txq->ift_cidx_processed = 0;
1776 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1777 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1778 
1779 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1780 		bzero((void *)di->idi_vaddr, di->idi_size);
1781 
1782 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1783 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1784 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1785 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1786 	return (0);
1787 }
1788 
1789 /*********************************************************************
1790  *
1791  *  Allocate DMA resources for RX buffers as well as memory for the RX
1792  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1793  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1794  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1795  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1796  *  packet, the maximum number of entries we'll need is equal to the
1797  *  number of hardware receive descriptors that we've allocated.
1798  *
1799  **********************************************************************/
1800 static int
1801 iflib_rxsd_alloc(iflib_rxq_t rxq)
1802 {
1803 	if_ctx_t ctx = rxq->ifr_ctx;
1804 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1805 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1806 	device_t dev = ctx->ifc_dev;
1807 	iflib_fl_t fl;
1808 	int			err;
1809 
1810 	MPASS(scctx->isc_nrxd[0] > 0);
1811 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1812 
1813 	fl = rxq->ifr_fl;
1814 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1815 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1816 		/* Set up DMA tag for RX buffers. */
1817 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1818 					 1, 0,			/* alignment, bounds */
1819 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1820 					 BUS_SPACE_MAXADDR,	/* highaddr */
1821 					 NULL, NULL,		/* filter, filterarg */
1822 					 sctx->isc_rx_maxsize,	/* maxsize */
1823 					 sctx->isc_rx_nsegments,	/* nsegments */
1824 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1825 					 0,			/* flags */
1826 					 NULL,			/* lockfunc */
1827 					 NULL,			/* lockarg */
1828 					 &fl->ifl_buf_tag);
1829 		if (err) {
1830 			device_printf(dev,
1831 			    "Unable to allocate RX DMA tag: %d\n", err);
1832 			goto fail;
1833 		}
1834 
1835 		/* Allocate memory for the RX mbuf map. */
1836 		if (!(fl->ifl_sds.ifsd_m =
1837 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1838 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1839 			device_printf(dev,
1840 			    "Unable to allocate RX mbuf map memory\n");
1841 			err = ENOMEM;
1842 			goto fail;
1843 		}
1844 
1845 		/* Allocate memory for the direct RX cluster pointer map. */
1846 		if (!(fl->ifl_sds.ifsd_cl =
1847 		      (caddr_t *) malloc(sizeof(caddr_t) *
1848 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1849 			device_printf(dev,
1850 			    "Unable to allocate RX cluster map memory\n");
1851 			err = ENOMEM;
1852 			goto fail;
1853 		}
1854 
1855 		/* Allocate memory for the RX cluster bus address map. */
1856 		if (!(fl->ifl_sds.ifsd_ba =
1857 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1858 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1859 			device_printf(dev,
1860 			    "Unable to allocate RX bus address map memory\n");
1861 			err = ENOMEM;
1862 			goto fail;
1863 		}
1864 
1865 		/*
1866 		 * Create the DMA maps for RX buffers.
1867 		 */
1868 		if (!(fl->ifl_sds.ifsd_map =
1869 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1870 			device_printf(dev,
1871 			    "Unable to allocate RX buffer DMA map memory\n");
1872 			err = ENOMEM;
1873 			goto fail;
1874 		}
1875 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1876 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1877 			    &fl->ifl_sds.ifsd_map[i]);
1878 			if (err != 0) {
1879 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1880 				goto fail;
1881 			}
1882 		}
1883 	}
1884 	return (0);
1885 
1886 fail:
1887 	iflib_rx_structures_free(ctx);
1888 	return (err);
1889 }
1890 
1891 
1892 /*
1893  * Internal service routines
1894  */
1895 
1896 struct rxq_refill_cb_arg {
1897 	int               error;
1898 	bus_dma_segment_t seg;
1899 	int               nseg;
1900 };
1901 
1902 static void
1903 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1904 {
1905 	struct rxq_refill_cb_arg *cb_arg = arg;
1906 
1907 	cb_arg->error = error;
1908 	cb_arg->seg = segs[0];
1909 	cb_arg->nseg = nseg;
1910 }
1911 
1912 /**
1913  * _iflib_fl_refill - refill an rxq free-buffer list
1914  * @ctx: the iflib context
1915  * @fl: the free list to refill
1916  * @count: the number of new buffers to allocate
1917  *
1918  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1919  * The caller must assure that @count does not exceed the queue's capacity.
1920  */
1921 static uint8_t
1922 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1923 {
1924 	struct if_rxd_update iru;
1925 	struct rxq_refill_cb_arg cb_arg;
1926 	struct mbuf *m;
1927 	caddr_t cl, *sd_cl;
1928 	struct mbuf **sd_m;
1929 	bus_dmamap_t *sd_map;
1930 	bus_addr_t bus_addr, *sd_ba;
1931 	int err, frag_idx, i, idx, n, pidx;
1932 	qidx_t credits;
1933 
1934 	sd_m = fl->ifl_sds.ifsd_m;
1935 	sd_map = fl->ifl_sds.ifsd_map;
1936 	sd_cl = fl->ifl_sds.ifsd_cl;
1937 	sd_ba = fl->ifl_sds.ifsd_ba;
1938 	pidx = fl->ifl_pidx;
1939 	idx = pidx;
1940 	frag_idx = fl->ifl_fragidx;
1941 	credits = fl->ifl_credits;
1942 
1943 	i = 0;
1944 	n = count;
1945 	MPASS(n > 0);
1946 	MPASS(credits + n <= fl->ifl_size);
1947 
1948 	if (pidx < fl->ifl_cidx)
1949 		MPASS(pidx + n <= fl->ifl_cidx);
1950 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1951 		MPASS(fl->ifl_gen == 0);
1952 	if (pidx > fl->ifl_cidx)
1953 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1954 
1955 	DBG_COUNTER_INC(fl_refills);
1956 	if (n > 8)
1957 		DBG_COUNTER_INC(fl_refills_large);
1958 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1959 	while (n--) {
1960 		/*
1961 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1962 		 * initialized after rx.
1963 		 *
1964 		 * If the cluster is still set then we know a minimum sized packet was received
1965 		 */
1966 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1967 		    &frag_idx);
1968 		if (frag_idx < 0)
1969 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1970 		MPASS(frag_idx >= 0);
1971 		if ((cl = sd_cl[frag_idx]) == NULL) {
1972 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1973 				break;
1974 
1975 			cb_arg.error = 0;
1976 			MPASS(sd_map != NULL);
1977 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1978 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1979 			    BUS_DMA_NOWAIT);
1980 			if (err != 0 || cb_arg.error) {
1981 				/*
1982 				 * !zone_pack ?
1983 				 */
1984 				if (fl->ifl_zone == zone_pack)
1985 					uma_zfree(fl->ifl_zone, cl);
1986 				break;
1987 			}
1988 
1989 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1990 			sd_cl[frag_idx] = cl;
1991 #if MEMORY_LOGGING
1992 			fl->ifl_cl_enqueued++;
1993 #endif
1994 		} else {
1995 			bus_addr = sd_ba[frag_idx];
1996 		}
1997 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
1998 		    BUS_DMASYNC_PREREAD);
1999 
2000 		if (sd_m[frag_idx] == NULL) {
2001 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2002 				break;
2003 			}
2004 			sd_m[frag_idx] = m;
2005 		}
2006 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2007 #if MEMORY_LOGGING
2008 		fl->ifl_m_enqueued++;
2009 #endif
2010 
2011 		DBG_COUNTER_INC(rx_allocs);
2012 		fl->ifl_rxd_idxs[i] = frag_idx;
2013 		fl->ifl_bus_addrs[i] = bus_addr;
2014 		fl->ifl_vm_addrs[i] = cl;
2015 		credits++;
2016 		i++;
2017 		MPASS(credits <= fl->ifl_size);
2018 		if (++idx == fl->ifl_size) {
2019 			fl->ifl_gen = 1;
2020 			idx = 0;
2021 		}
2022 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2023 			iru.iru_pidx = pidx;
2024 			iru.iru_count = i;
2025 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2026 			i = 0;
2027 			pidx = idx;
2028 			fl->ifl_pidx = idx;
2029 			fl->ifl_credits = credits;
2030 		}
2031 	}
2032 
2033 	if (i) {
2034 		iru.iru_pidx = pidx;
2035 		iru.iru_count = i;
2036 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2037 		fl->ifl_pidx = idx;
2038 		fl->ifl_credits = credits;
2039 	}
2040 	DBG_COUNTER_INC(rxd_flush);
2041 	if (fl->ifl_pidx == 0)
2042 		pidx = fl->ifl_size - 1;
2043 	else
2044 		pidx = fl->ifl_pidx - 1;
2045 
2046 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2047 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2048 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2049 	fl->ifl_fragidx = frag_idx + 1;
2050 	if (fl->ifl_fragidx == fl->ifl_size)
2051 		fl->ifl_fragidx = 0;
2052 
2053 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2054 }
2055 
2056 static __inline uint8_t
2057 __iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2058 {
2059 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2060 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2061 #ifdef INVARIANTS
2062 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2063 #endif
2064 
2065 	MPASS(fl->ifl_credits <= fl->ifl_size);
2066 	MPASS(reclaimable == delta);
2067 
2068 	if (reclaimable > 0)
2069 		return (_iflib_fl_refill(ctx, fl, reclaimable));
2070 	return (0);
2071 }
2072 
2073 uint8_t
2074 iflib_in_detach(if_ctx_t ctx)
2075 {
2076 	bool in_detach;
2077 
2078 	STATE_LOCK(ctx);
2079 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2080 	STATE_UNLOCK(ctx);
2081 	return (in_detach);
2082 }
2083 
2084 static void
2085 iflib_fl_bufs_free(iflib_fl_t fl)
2086 {
2087 	iflib_dma_info_t idi = fl->ifl_ifdi;
2088 	bus_dmamap_t sd_map;
2089 	uint32_t i;
2090 
2091 	for (i = 0; i < fl->ifl_size; i++) {
2092 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2093 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2094 
2095 		if (*sd_cl != NULL) {
2096 			sd_map = fl->ifl_sds.ifsd_map[i];
2097 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2098 			    BUS_DMASYNC_POSTREAD);
2099 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2100 			if (*sd_cl != NULL)
2101 				uma_zfree(fl->ifl_zone, *sd_cl);
2102 			if (*sd_m != NULL) {
2103 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2104 				uma_zfree(zone_mbuf, *sd_m);
2105 			}
2106 		} else {
2107 			MPASS(*sd_cl == NULL);
2108 			MPASS(*sd_m == NULL);
2109 		}
2110 #if MEMORY_LOGGING
2111 		fl->ifl_m_dequeued++;
2112 		fl->ifl_cl_dequeued++;
2113 #endif
2114 		*sd_cl = NULL;
2115 		*sd_m = NULL;
2116 	}
2117 #ifdef INVARIANTS
2118 	for (i = 0; i < fl->ifl_size; i++) {
2119 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2120 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2121 	}
2122 #endif
2123 	/*
2124 	 * Reset free list values
2125 	 */
2126 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2127 	bzero(idi->idi_vaddr, idi->idi_size);
2128 }
2129 
2130 /*********************************************************************
2131  *
2132  *  Initialize a free list and its buffers.
2133  *
2134  **********************************************************************/
2135 static int
2136 iflib_fl_setup(iflib_fl_t fl)
2137 {
2138 	iflib_rxq_t rxq = fl->ifl_rxq;
2139 	if_ctx_t ctx = rxq->ifr_ctx;
2140 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2141 	int qidx;
2142 
2143 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2144 	/*
2145 	** Free current RX buffer structs and their mbufs
2146 	*/
2147 	iflib_fl_bufs_free(fl);
2148 	/* Now replenish the mbufs */
2149 	MPASS(fl->ifl_credits == 0);
2150 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2151 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2152 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2153 	else
2154 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2155 	/*
2156 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2157 	 * to the selected mbuf size.
2158 	 */
2159 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2160 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2161 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2162 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2163 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2164 
2165 
2166 	/* avoid pre-allocating zillions of clusters to an idle card
2167 	 * potentially speeding up attach
2168 	 */
2169 	(void) _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2170 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2171 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2172 		return (ENOBUFS);
2173 	/*
2174 	 * handle failure
2175 	 */
2176 	MPASS(rxq != NULL);
2177 	MPASS(fl->ifl_ifdi != NULL);
2178 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2179 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2180 	return (0);
2181 }
2182 
2183 /*********************************************************************
2184  *
2185  *  Free receive ring data structures
2186  *
2187  **********************************************************************/
2188 static void
2189 iflib_rx_sds_free(iflib_rxq_t rxq)
2190 {
2191 	iflib_fl_t fl;
2192 	int i, j;
2193 
2194 	if (rxq->ifr_fl != NULL) {
2195 		for (i = 0; i < rxq->ifr_nfl; i++) {
2196 			fl = &rxq->ifr_fl[i];
2197 			if (fl->ifl_buf_tag != NULL) {
2198 				if (fl->ifl_sds.ifsd_map != NULL) {
2199 					for (j = 0; j < fl->ifl_size; j++) {
2200 						bus_dmamap_sync(
2201 						    fl->ifl_buf_tag,
2202 						    fl->ifl_sds.ifsd_map[j],
2203 						    BUS_DMASYNC_POSTREAD);
2204 						bus_dmamap_unload(
2205 						    fl->ifl_buf_tag,
2206 						    fl->ifl_sds.ifsd_map[j]);
2207 						bus_dmamap_destroy(
2208 						    fl->ifl_buf_tag,
2209 						    fl->ifl_sds.ifsd_map[j]);
2210 					}
2211 				}
2212 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2213 				fl->ifl_buf_tag = NULL;
2214 			}
2215 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2216 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2217 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2219 			fl->ifl_sds.ifsd_m = NULL;
2220 			fl->ifl_sds.ifsd_cl = NULL;
2221 			fl->ifl_sds.ifsd_ba = NULL;
2222 			fl->ifl_sds.ifsd_map = NULL;
2223 		}
2224 		free(rxq->ifr_fl, M_IFLIB);
2225 		rxq->ifr_fl = NULL;
2226 		free(rxq->ifr_ifdi, M_IFLIB);
2227 		rxq->ifr_ifdi = NULL;
2228 		rxq->ifr_cq_cidx = 0;
2229 	}
2230 }
2231 
2232 /*
2233  * Timer routine
2234  */
2235 static void
2236 iflib_timer(void *arg)
2237 {
2238 	iflib_txq_t txq = arg;
2239 	if_ctx_t ctx = txq->ift_ctx;
2240 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2241 	uint64_t this_tick = ticks;
2242 	uint32_t reset_on = hz / 2;
2243 
2244 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2245 		return;
2246 
2247 	/*
2248 	** Check on the state of the TX queue(s), this
2249 	** can be done without the lock because its RO
2250 	** and the HUNG state will be static if set.
2251 	*/
2252 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2253 		txq->ift_last_timer_tick = this_tick;
2254 		IFDI_TIMER(ctx, txq->ift_id);
2255 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2256 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2257 		     (sctx->isc_pause_frames == 0)))
2258 			goto hung;
2259 
2260 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2261 		    ifmp_ring_is_stalled(txq->ift_br)) {
2262 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2263 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2264 		}
2265 		txq->ift_cleaned_prev = txq->ift_cleaned;
2266 	}
2267 #ifdef DEV_NETMAP
2268 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2269 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2270 #endif
2271 	/* handle any laggards */
2272 	if (txq->ift_db_pending)
2273 		GROUPTASK_ENQUEUE(&txq->ift_task);
2274 
2275 	sctx->isc_pause_frames = 0;
2276 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2277 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2278 	return;
2279 
2280  hung:
2281 	device_printf(ctx->ifc_dev,
2282 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2283 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2284 	STATE_LOCK(ctx);
2285 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2286 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2287 	iflib_admin_intr_deferred(ctx);
2288 	STATE_UNLOCK(ctx);
2289 }
2290 
2291 static uint16_t
2292 iflib_get_mbuf_size_for(unsigned int size)
2293 {
2294 
2295 	if (size <= MCLBYTES)
2296 		return (MCLBYTES);
2297 	else
2298 		return (MJUMPAGESIZE);
2299 }
2300 
2301 static void
2302 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2303 {
2304 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2305 
2306 	/*
2307 	 * XXX don't set the max_frame_size to larger
2308 	 * than the hardware can handle
2309 	 */
2310 	ctx->ifc_rx_mbuf_sz =
2311 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2312 }
2313 
2314 uint32_t
2315 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2316 {
2317 
2318 	return (ctx->ifc_rx_mbuf_sz);
2319 }
2320 
2321 static void
2322 iflib_init_locked(if_ctx_t ctx)
2323 {
2324 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2325 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2326 	if_t ifp = ctx->ifc_ifp;
2327 	iflib_fl_t fl;
2328 	iflib_txq_t txq;
2329 	iflib_rxq_t rxq;
2330 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2331 
2332 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2333 	IFDI_INTR_DISABLE(ctx);
2334 
2335 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2336 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2337 	/* Set hardware offload abilities */
2338 	if_clearhwassist(ifp);
2339 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2340 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2341 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2342 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2343 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2344 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2345 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2346 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2347 
2348 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2349 		CALLOUT_LOCK(txq);
2350 		callout_stop(&txq->ift_timer);
2351 		CALLOUT_UNLOCK(txq);
2352 		iflib_netmap_txq_init(ctx, txq);
2353 	}
2354 
2355 	/*
2356 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2357 	 * that drivers can use the value when setting up the hardware receive
2358 	 * buffers.
2359 	 */
2360 	iflib_calc_rx_mbuf_sz(ctx);
2361 
2362 #ifdef INVARIANTS
2363 	i = if_getdrvflags(ifp);
2364 #endif
2365 	IFDI_INIT(ctx);
2366 	MPASS(if_getdrvflags(ifp) == i);
2367 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2368 		/* XXX this should really be done on a per-queue basis */
2369 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2370 			MPASS(rxq->ifr_id == i);
2371 			iflib_netmap_rxq_init(ctx, rxq);
2372 			continue;
2373 		}
2374 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2375 			if (iflib_fl_setup(fl)) {
2376 				device_printf(ctx->ifc_dev,
2377 				    "setting up free list %d failed - "
2378 				    "check cluster settings\n", j);
2379 				goto done;
2380 			}
2381 		}
2382 	}
2383 done:
2384 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2385 	IFDI_INTR_ENABLE(ctx);
2386 	txq = ctx->ifc_txqs;
2387 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2388 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2389 			txq->ift_timer.c_cpu);
2390 }
2391 
2392 static int
2393 iflib_media_change(if_t ifp)
2394 {
2395 	if_ctx_t ctx = if_getsoftc(ifp);
2396 	int err;
2397 
2398 	CTX_LOCK(ctx);
2399 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2400 		iflib_init_locked(ctx);
2401 	CTX_UNLOCK(ctx);
2402 	return (err);
2403 }
2404 
2405 static void
2406 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2407 {
2408 	if_ctx_t ctx = if_getsoftc(ifp);
2409 
2410 	CTX_LOCK(ctx);
2411 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2412 	IFDI_MEDIA_STATUS(ctx, ifmr);
2413 	CTX_UNLOCK(ctx);
2414 }
2415 
2416 void
2417 iflib_stop(if_ctx_t ctx)
2418 {
2419 	iflib_txq_t txq = ctx->ifc_txqs;
2420 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2421 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2422 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2423 	iflib_dma_info_t di;
2424 	iflib_fl_t fl;
2425 	int i, j;
2426 
2427 	/* Tell the stack that the interface is no longer active */
2428 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2429 
2430 	IFDI_INTR_DISABLE(ctx);
2431 	DELAY(1000);
2432 	IFDI_STOP(ctx);
2433 	DELAY(1000);
2434 
2435 	iflib_debug_reset();
2436 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2437 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2438 		/* make sure all transmitters have completed before proceeding XXX */
2439 
2440 		CALLOUT_LOCK(txq);
2441 		callout_stop(&txq->ift_timer);
2442 		CALLOUT_UNLOCK(txq);
2443 
2444 		/* clean any enqueued buffers */
2445 		iflib_ifmp_purge(txq);
2446 		/* Free any existing tx buffers. */
2447 		for (j = 0; j < txq->ift_size; j++) {
2448 			iflib_txsd_free(ctx, txq, j);
2449 		}
2450 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2451 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2452 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2453 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2454 		txq->ift_pullups = 0;
2455 		ifmp_ring_reset_stats(txq->ift_br);
2456 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2457 			bzero((void *)di->idi_vaddr, di->idi_size);
2458 	}
2459 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2460 		/* make sure all transmitters have completed before proceeding XXX */
2461 
2462 		rxq->ifr_cq_cidx = 0;
2463 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2464 			bzero((void *)di->idi_vaddr, di->idi_size);
2465 		/* also resets the free lists pidx/cidx */
2466 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2467 			iflib_fl_bufs_free(fl);
2468 	}
2469 }
2470 
2471 static inline caddr_t
2472 calc_next_rxd(iflib_fl_t fl, int cidx)
2473 {
2474 	qidx_t size;
2475 	int nrxd;
2476 	caddr_t start, end, cur, next;
2477 
2478 	nrxd = fl->ifl_size;
2479 	size = fl->ifl_rxd_size;
2480 	start = fl->ifl_ifdi->idi_vaddr;
2481 
2482 	if (__predict_false(size == 0))
2483 		return (start);
2484 	cur = start + size*cidx;
2485 	end = start + size*nrxd;
2486 	next = CACHE_PTR_NEXT(cur);
2487 	return (next < end ? next : start);
2488 }
2489 
2490 static inline void
2491 prefetch_pkts(iflib_fl_t fl, int cidx)
2492 {
2493 	int nextptr;
2494 	int nrxd = fl->ifl_size;
2495 	caddr_t next_rxd;
2496 
2497 
2498 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2499 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2500 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2501 	next_rxd = calc_next_rxd(fl, cidx);
2502 	prefetch(next_rxd);
2503 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2504 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2505 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2506 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2507 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2508 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2509 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2510 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2511 }
2512 
2513 static struct mbuf *
2514 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2515     int *pf_rv, if_rxd_info_t ri)
2516 {
2517 	bus_dmamap_t map;
2518 	iflib_fl_t fl;
2519 	caddr_t payload;
2520 	struct mbuf *m;
2521 	int flid, cidx, len, next;
2522 
2523 	map = NULL;
2524 	flid = irf->irf_flid;
2525 	cidx = irf->irf_idx;
2526 	fl = &rxq->ifr_fl[flid];
2527 	sd->ifsd_fl = fl;
2528 	m = fl->ifl_sds.ifsd_m[cidx];
2529 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2530 	fl->ifl_credits--;
2531 #if MEMORY_LOGGING
2532 	fl->ifl_m_dequeued++;
2533 #endif
2534 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2535 		prefetch_pkts(fl, cidx);
2536 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2537 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2538 	map = fl->ifl_sds.ifsd_map[cidx];
2539 
2540 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2541 
2542 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2543 	    irf->irf_len != 0) {
2544 		payload  = *sd->ifsd_cl;
2545 		payload +=  ri->iri_pad;
2546 		len = ri->iri_len - ri->iri_pad;
2547 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2548 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2549 		switch (*pf_rv) {
2550 		case PFIL_DROPPED:
2551 		case PFIL_CONSUMED:
2552 			/*
2553 			 * The filter ate it.  Everything is recycled.
2554 			 */
2555 			m = NULL;
2556 			unload = 0;
2557 			break;
2558 		case PFIL_REALLOCED:
2559 			/*
2560 			 * The filter copied it.  Everything is recycled.
2561 			 */
2562 			m = pfil_mem2mbuf(payload);
2563 			unload = 0;
2564 			break;
2565 		case PFIL_PASS:
2566 			/*
2567 			 * Filter said it was OK, so receive like
2568 			 * normal
2569 			 */
2570 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2571 			break;
2572 		default:
2573 			MPASS(0);
2574 		}
2575 	} else {
2576 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2577 		*pf_rv = PFIL_PASS;
2578 	}
2579 
2580 	if (unload && irf->irf_len != 0)
2581 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2582 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2583 	if (__predict_false(fl->ifl_cidx == 0))
2584 		fl->ifl_gen = 0;
2585 	bit_clear(fl->ifl_rx_bitmap, cidx);
2586 	return (m);
2587 }
2588 
2589 static struct mbuf *
2590 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2591 {
2592 	struct mbuf *m, *mh, *mt;
2593 	caddr_t cl;
2594 	int  *pf_rv_ptr, flags, i, padlen;
2595 	bool consumed;
2596 
2597 	i = 0;
2598 	mh = NULL;
2599 	consumed = false;
2600 	*pf_rv = PFIL_PASS;
2601 	pf_rv_ptr = pf_rv;
2602 	do {
2603 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2604 		    pf_rv_ptr, ri);
2605 
2606 		MPASS(*sd->ifsd_cl != NULL);
2607 
2608 		/*
2609 		 * Exclude zero-length frags & frags from
2610 		 * packets the filter has consumed or dropped
2611 		 */
2612 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2613 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2614 			if (mh == NULL) {
2615 				/* everything saved here */
2616 				consumed = true;
2617 				pf_rv_ptr = NULL;
2618 				continue;
2619 			}
2620 			/* XXX we can save the cluster here, but not the mbuf */
2621 			m_init(m, M_NOWAIT, MT_DATA, 0);
2622 			m_free(m);
2623 			continue;
2624 		}
2625 		if (mh == NULL) {
2626 			flags = M_PKTHDR|M_EXT;
2627 			mh = mt = m;
2628 			padlen = ri->iri_pad;
2629 		} else {
2630 			flags = M_EXT;
2631 			mt->m_next = m;
2632 			mt = m;
2633 			/* assuming padding is only on the first fragment */
2634 			padlen = 0;
2635 		}
2636 		cl = *sd->ifsd_cl;
2637 		*sd->ifsd_cl = NULL;
2638 
2639 		/* Can these two be made one ? */
2640 		m_init(m, M_NOWAIT, MT_DATA, flags);
2641 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2642 		/*
2643 		 * These must follow m_init and m_cljset
2644 		 */
2645 		m->m_data += padlen;
2646 		ri->iri_len -= padlen;
2647 		m->m_len = ri->iri_frags[i].irf_len;
2648 	} while (++i < ri->iri_nfrags);
2649 
2650 	return (mh);
2651 }
2652 
2653 /*
2654  * Process one software descriptor
2655  */
2656 static struct mbuf *
2657 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2658 {
2659 	struct if_rxsd sd;
2660 	struct mbuf *m;
2661 	int pf_rv;
2662 
2663 	/* should I merge this back in now that the two paths are basically duplicated? */
2664 	if (ri->iri_nfrags == 1 &&
2665 	    ri->iri_frags[0].irf_len != 0 &&
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 (m == NULL)
2683 			return (NULL);
2684 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2685 			return (m);
2686 	}
2687 	m->m_pkthdr.len = ri->iri_len;
2688 	m->m_pkthdr.rcvif = ri->iri_ifp;
2689 	m->m_flags |= ri->iri_flags;
2690 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2691 	m->m_pkthdr.flowid = ri->iri_flowid;
2692 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2693 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2694 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2695 	return (m);
2696 }
2697 
2698 #if defined(INET6) || defined(INET)
2699 static void
2700 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2701 {
2702 	CURVNET_SET(lc->ifp->if_vnet);
2703 #if defined(INET6)
2704 	*v6 = V_ip6_forwarding;
2705 #endif
2706 #if defined(INET)
2707 	*v4 = V_ipforwarding;
2708 #endif
2709 	CURVNET_RESTORE();
2710 }
2711 
2712 /*
2713  * Returns true if it's possible this packet could be LROed.
2714  * if it returns false, it is guaranteed that tcp_lro_rx()
2715  * would not return zero.
2716  */
2717 static bool
2718 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2719 {
2720 	struct ether_header *eh;
2721 
2722 	eh = mtod(m, struct ether_header *);
2723 	switch (eh->ether_type) {
2724 #if defined(INET6)
2725 		case htons(ETHERTYPE_IPV6):
2726 			return (!v6_forwarding);
2727 #endif
2728 #if defined (INET)
2729 		case htons(ETHERTYPE_IP):
2730 			return (!v4_forwarding);
2731 #endif
2732 	}
2733 
2734 	return false;
2735 }
2736 #else
2737 static void
2738 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2739 {
2740 }
2741 #endif
2742 
2743 static void
2744 _task_fn_rx_watchdog(void *context)
2745 {
2746 	iflib_rxq_t rxq = context;
2747 
2748 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2749 }
2750 
2751 static uint8_t
2752 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2753 {
2754 	if_t ifp;
2755 	if_ctx_t ctx = rxq->ifr_ctx;
2756 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2757 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2758 	int avail, i;
2759 	qidx_t *cidxp;
2760 	struct if_rxd_info ri;
2761 	int err, budget_left, rx_bytes, rx_pkts;
2762 	iflib_fl_t fl;
2763 	int lro_enabled;
2764 	bool v4_forwarding, v6_forwarding, lro_possible;
2765 	uint8_t retval = 0;
2766 
2767 	/*
2768 	 * XXX early demux data packets so that if_input processing only handles
2769 	 * acks in interrupt context
2770 	 */
2771 	struct mbuf *m, *mh, *mt, *mf;
2772 
2773 	NET_EPOCH_ASSERT();
2774 
2775 	lro_possible = v4_forwarding = v6_forwarding = false;
2776 	ifp = ctx->ifc_ifp;
2777 	mh = mt = NULL;
2778 	MPASS(budget > 0);
2779 	rx_pkts	= rx_bytes = 0;
2780 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2781 		cidxp = &rxq->ifr_cq_cidx;
2782 	else
2783 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2784 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2785 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2786 			retval |= __iflib_fl_refill_all(ctx, fl);
2787 		DBG_COUNTER_INC(rx_unavail);
2788 		return (retval);
2789 	}
2790 
2791 	/* pfil needs the vnet to be set */
2792 	CURVNET_SET_QUIET(ifp->if_vnet);
2793 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2794 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2795 			DBG_COUNTER_INC(rx_ctx_inactive);
2796 			break;
2797 		}
2798 		/*
2799 		 * Reset client set fields to their default values
2800 		 */
2801 		rxd_info_zero(&ri);
2802 		ri.iri_qsidx = rxq->ifr_id;
2803 		ri.iri_cidx = *cidxp;
2804 		ri.iri_ifp = ifp;
2805 		ri.iri_frags = rxq->ifr_frags;
2806 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2807 
2808 		if (err)
2809 			goto err;
2810 		rx_pkts += 1;
2811 		rx_bytes += ri.iri_len;
2812 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2813 			*cidxp = ri.iri_cidx;
2814 			/* Update our consumer index */
2815 			/* XXX NB: shurd - check if this is still safe */
2816 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2817 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2818 			/* was this only a completion queue message? */
2819 			if (__predict_false(ri.iri_nfrags == 0))
2820 				continue;
2821 		}
2822 		MPASS(ri.iri_nfrags != 0);
2823 		MPASS(ri.iri_len != 0);
2824 
2825 		/* will advance the cidx on the corresponding free lists */
2826 		m = iflib_rxd_pkt_get(rxq, &ri);
2827 		avail--;
2828 		budget_left--;
2829 		if (avail == 0 && budget_left)
2830 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2831 
2832 		if (__predict_false(m == NULL))
2833 			continue;
2834 
2835 		/* imm_pkt: -- cxgb */
2836 		if (mh == NULL)
2837 			mh = mt = m;
2838 		else {
2839 			mt->m_nextpkt = m;
2840 			mt = m;
2841 		}
2842 	}
2843 	CURVNET_RESTORE();
2844 	/* make sure that we can refill faster than drain */
2845 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2846 		retval |= __iflib_fl_refill_all(ctx, fl);
2847 
2848 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2849 	if (lro_enabled)
2850 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2851 	mt = mf = NULL;
2852 	while (mh != NULL) {
2853 		m = mh;
2854 		mh = mh->m_nextpkt;
2855 		m->m_nextpkt = NULL;
2856 #ifndef __NO_STRICT_ALIGNMENT
2857 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2858 			continue;
2859 #endif
2860 		rx_bytes += m->m_pkthdr.len;
2861 		rx_pkts++;
2862 #if defined(INET6) || defined(INET)
2863 		if (lro_enabled) {
2864 			if (!lro_possible) {
2865 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2866 				if (lro_possible && mf != NULL) {
2867 					ifp->if_input(ifp, mf);
2868 					DBG_COUNTER_INC(rx_if_input);
2869 					mt = mf = NULL;
2870 				}
2871 			}
2872 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2873 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2874 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2875 					continue;
2876 			}
2877 		}
2878 #endif
2879 		if (lro_possible) {
2880 			ifp->if_input(ifp, m);
2881 			DBG_COUNTER_INC(rx_if_input);
2882 			continue;
2883 		}
2884 
2885 		if (mf == NULL)
2886 			mf = m;
2887 		if (mt != NULL)
2888 			mt->m_nextpkt = m;
2889 		mt = m;
2890 	}
2891 	if (mf != NULL) {
2892 		ifp->if_input(ifp, mf);
2893 		DBG_COUNTER_INC(rx_if_input);
2894 	}
2895 
2896 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2897 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2898 
2899 	/*
2900 	 * Flush any outstanding LRO work
2901 	 */
2902 #if defined(INET6) || defined(INET)
2903 	tcp_lro_flush_all(&rxq->ifr_lc);
2904 #endif
2905 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2906 		retval |= IFLIB_RXEOF_MORE;
2907 	return (retval);
2908 err:
2909 	STATE_LOCK(ctx);
2910 	ctx->ifc_flags |= IFC_DO_RESET;
2911 	iflib_admin_intr_deferred(ctx);
2912 	STATE_UNLOCK(ctx);
2913 	return (0);
2914 }
2915 
2916 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2917 static inline qidx_t
2918 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2919 {
2920 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2921 	qidx_t minthresh = txq->ift_size / 8;
2922 	if (in_use > 4*minthresh)
2923 		return (notify_count);
2924 	if (in_use > 2*minthresh)
2925 		return (notify_count >> 1);
2926 	if (in_use > minthresh)
2927 		return (notify_count >> 3);
2928 	return (0);
2929 }
2930 
2931 static inline qidx_t
2932 txq_max_rs_deferred(iflib_txq_t txq)
2933 {
2934 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2935 	qidx_t minthresh = txq->ift_size / 8;
2936 	if (txq->ift_in_use > 4*minthresh)
2937 		return (notify_count);
2938 	if (txq->ift_in_use > 2*minthresh)
2939 		return (notify_count >> 1);
2940 	if (txq->ift_in_use > minthresh)
2941 		return (notify_count >> 2);
2942 	return (2);
2943 }
2944 
2945 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2946 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2947 
2948 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2949 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2950 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2951 
2952 /* forward compatibility for cxgb */
2953 #define FIRST_QSET(ctx) 0
2954 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2955 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2956 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2957 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2958 
2959 /* XXX we should be setting this to something other than zero */
2960 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2961 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2962     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2963 
2964 static inline bool
2965 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2966 {
2967 	qidx_t dbval, max;
2968 	bool rang;
2969 
2970 	rang = false;
2971 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2972 	if (ring || txq->ift_db_pending >= max) {
2973 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2974 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2975 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2976 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2977 		txq->ift_db_pending = txq->ift_npending = 0;
2978 		rang = true;
2979 	}
2980 	return (rang);
2981 }
2982 
2983 #ifdef PKT_DEBUG
2984 static void
2985 print_pkt(if_pkt_info_t pi)
2986 {
2987 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2988 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2989 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2990 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2991 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2992 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2993 }
2994 #endif
2995 
2996 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2997 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2998 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2999 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3000 
3001 static int
3002 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3003 {
3004 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3005 	struct ether_vlan_header *eh;
3006 	struct mbuf *m;
3007 
3008 	m = *mp;
3009 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3010 	    M_WRITABLE(m) == 0) {
3011 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3012 			return (ENOMEM);
3013 		} else {
3014 			m_freem(*mp);
3015 			DBG_COUNTER_INC(tx_frees);
3016 			*mp = m;
3017 		}
3018 	}
3019 
3020 	/*
3021 	 * Determine where frame payload starts.
3022 	 * Jump over vlan headers if already present,
3023 	 * helpful for QinQ too.
3024 	 */
3025 	if (__predict_false(m->m_len < sizeof(*eh))) {
3026 		txq->ift_pullups++;
3027 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3028 			return (ENOMEM);
3029 	}
3030 	eh = mtod(m, struct ether_vlan_header *);
3031 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3032 		pi->ipi_etype = ntohs(eh->evl_proto);
3033 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3034 	} else {
3035 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3036 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3037 	}
3038 
3039 	switch (pi->ipi_etype) {
3040 #ifdef INET
3041 	case ETHERTYPE_IP:
3042 	{
3043 		struct mbuf *n;
3044 		struct ip *ip = NULL;
3045 		struct tcphdr *th = NULL;
3046 		int minthlen;
3047 
3048 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3049 		if (__predict_false(m->m_len < minthlen)) {
3050 			/*
3051 			 * if this code bloat is causing too much of a hit
3052 			 * move it to a separate function and mark it noinline
3053 			 */
3054 			if (m->m_len == pi->ipi_ehdrlen) {
3055 				n = m->m_next;
3056 				MPASS(n);
3057 				if (n->m_len >= sizeof(*ip))  {
3058 					ip = (struct ip *)n->m_data;
3059 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3060 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3061 				} else {
3062 					txq->ift_pullups++;
3063 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3064 						return (ENOMEM);
3065 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3066 				}
3067 			} else {
3068 				txq->ift_pullups++;
3069 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3070 					return (ENOMEM);
3071 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3072 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3073 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3074 			}
3075 		} else {
3076 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3077 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3078 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3079 		}
3080 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3081 		pi->ipi_ipproto = ip->ip_p;
3082 		pi->ipi_flags |= IPI_TX_IPV4;
3083 
3084 		/* TCP checksum offload may require TCP header length */
3085 		if (IS_TX_OFFLOAD4(pi)) {
3086 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3087 				if (__predict_false(th == NULL)) {
3088 					txq->ift_pullups++;
3089 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3090 						return (ENOMEM);
3091 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3092 				}
3093 				pi->ipi_tcp_hflags = th->th_flags;
3094 				pi->ipi_tcp_hlen = th->th_off << 2;
3095 				pi->ipi_tcp_seq = th->th_seq;
3096 			}
3097 			if (IS_TSO4(pi)) {
3098 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3099 					return (ENXIO);
3100 				/*
3101 				 * TSO always requires hardware checksum offload.
3102 				 */
3103 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3104 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3105 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3106 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3107 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3108 					ip->ip_sum = 0;
3109 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3110 				}
3111 			}
3112 		}
3113 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3114                        ip->ip_sum = 0;
3115 
3116 		break;
3117 	}
3118 #endif
3119 #ifdef INET6
3120 	case ETHERTYPE_IPV6:
3121 	{
3122 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3123 		struct tcphdr *th;
3124 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3125 
3126 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3127 			txq->ift_pullups++;
3128 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3129 				return (ENOMEM);
3130 		}
3131 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3132 
3133 		/* XXX-BZ this will go badly in case of ext hdrs. */
3134 		pi->ipi_ipproto = ip6->ip6_nxt;
3135 		pi->ipi_flags |= IPI_TX_IPV6;
3136 
3137 		/* TCP checksum offload may require TCP header length */
3138 		if (IS_TX_OFFLOAD6(pi)) {
3139 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3140 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3141 					txq->ift_pullups++;
3142 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3143 						return (ENOMEM);
3144 				}
3145 				pi->ipi_tcp_hflags = th->th_flags;
3146 				pi->ipi_tcp_hlen = th->th_off << 2;
3147 				pi->ipi_tcp_seq = th->th_seq;
3148 			}
3149 			if (IS_TSO6(pi)) {
3150 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3151 					return (ENXIO);
3152 				/*
3153 				 * TSO always requires hardware checksum offload.
3154 				 */
3155 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3156 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3157 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3158 			}
3159 		}
3160 		break;
3161 	}
3162 #endif
3163 	default:
3164 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3165 		pi->ipi_ip_hlen = 0;
3166 		break;
3167 	}
3168 	*mp = m;
3169 
3170 	return (0);
3171 }
3172 
3173 /*
3174  * If dodgy hardware rejects the scatter gather chain we've handed it
3175  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3176  * m_defrag'd mbufs
3177  */
3178 static __noinline struct mbuf *
3179 iflib_remove_mbuf(iflib_txq_t txq)
3180 {
3181 	int ntxd, pidx;
3182 	struct mbuf *m, **ifsd_m;
3183 
3184 	ifsd_m = txq->ift_sds.ifsd_m;
3185 	ntxd = txq->ift_size;
3186 	pidx = txq->ift_pidx & (ntxd - 1);
3187 	ifsd_m = txq->ift_sds.ifsd_m;
3188 	m = ifsd_m[pidx];
3189 	ifsd_m[pidx] = NULL;
3190 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3191 	if (txq->ift_sds.ifsd_tso_map != NULL)
3192 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3193 		    txq->ift_sds.ifsd_tso_map[pidx]);
3194 #if MEMORY_LOGGING
3195 	txq->ift_dequeued++;
3196 #endif
3197 	return (m);
3198 }
3199 
3200 static inline caddr_t
3201 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3202 {
3203 	qidx_t size;
3204 	int ntxd;
3205 	caddr_t start, end, cur, next;
3206 
3207 	ntxd = txq->ift_size;
3208 	size = txq->ift_txd_size[qid];
3209 	start = txq->ift_ifdi[qid].idi_vaddr;
3210 
3211 	if (__predict_false(size == 0))
3212 		return (start);
3213 	cur = start + size*cidx;
3214 	end = start + size*ntxd;
3215 	next = CACHE_PTR_NEXT(cur);
3216 	return (next < end ? next : start);
3217 }
3218 
3219 /*
3220  * Pad an mbuf to ensure a minimum ethernet frame size.
3221  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3222  */
3223 static __noinline int
3224 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3225 {
3226 	/*
3227 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3228 	 * and ARP message is the smallest common payload I can think of
3229 	 */
3230 	static char pad[18];	/* just zeros */
3231 	int n;
3232 	struct mbuf *new_head;
3233 
3234 	if (!M_WRITABLE(*m_head)) {
3235 		new_head = m_dup(*m_head, M_NOWAIT);
3236 		if (new_head == NULL) {
3237 			m_freem(*m_head);
3238 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3239 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3240 			DBG_COUNTER_INC(tx_frees);
3241 			return ENOMEM;
3242 		}
3243 		m_freem(*m_head);
3244 		*m_head = new_head;
3245 	}
3246 
3247 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3248 	     n > 0; n -= sizeof(pad))
3249 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3250 			break;
3251 
3252 	if (n > 0) {
3253 		m_freem(*m_head);
3254 		device_printf(dev, "cannot pad short frame\n");
3255 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3256 		DBG_COUNTER_INC(tx_frees);
3257 		return (ENOBUFS);
3258 	}
3259 
3260 	return 0;
3261 }
3262 
3263 static int
3264 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3265 {
3266 	if_ctx_t		ctx;
3267 	if_shared_ctx_t		sctx;
3268 	if_softc_ctx_t		scctx;
3269 	bus_dma_tag_t		buf_tag;
3270 	bus_dma_segment_t	*segs;
3271 	struct mbuf		*m_head, **ifsd_m;
3272 	void			*next_txd;
3273 	bus_dmamap_t		map;
3274 	struct if_pkt_info	pi;
3275 	int remap = 0;
3276 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3277 
3278 	ctx = txq->ift_ctx;
3279 	sctx = ctx->ifc_sctx;
3280 	scctx = &ctx->ifc_softc_ctx;
3281 	segs = txq->ift_segs;
3282 	ntxd = txq->ift_size;
3283 	m_head = *m_headp;
3284 	map = NULL;
3285 
3286 	/*
3287 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3288 	 */
3289 	cidx = txq->ift_cidx;
3290 	pidx = txq->ift_pidx;
3291 	if (ctx->ifc_flags & IFC_PREFETCH) {
3292 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3293 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3294 			next_txd = calc_next_txd(txq, cidx, 0);
3295 			prefetch(next_txd);
3296 		}
3297 
3298 		/* prefetch the next cache line of mbuf pointers and flags */
3299 		prefetch(&txq->ift_sds.ifsd_m[next]);
3300 		prefetch(&txq->ift_sds.ifsd_map[next]);
3301 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3302 	}
3303 	map = txq->ift_sds.ifsd_map[pidx];
3304 	ifsd_m = txq->ift_sds.ifsd_m;
3305 
3306 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3307 		buf_tag = txq->ift_tso_buf_tag;
3308 		max_segs = scctx->isc_tx_tso_segments_max;
3309 		map = txq->ift_sds.ifsd_tso_map[pidx];
3310 		MPASS(buf_tag != NULL);
3311 		MPASS(max_segs > 0);
3312 	} else {
3313 		buf_tag = txq->ift_buf_tag;
3314 		max_segs = scctx->isc_tx_nsegments;
3315 		map = txq->ift_sds.ifsd_map[pidx];
3316 	}
3317 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3318 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3319 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3320 		if (err) {
3321 			DBG_COUNTER_INC(encap_txd_encap_fail);
3322 			return err;
3323 		}
3324 	}
3325 	m_head = *m_headp;
3326 
3327 	pkt_info_zero(&pi);
3328 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3329 	pi.ipi_pidx = pidx;
3330 	pi.ipi_qsidx = txq->ift_id;
3331 	pi.ipi_len = m_head->m_pkthdr.len;
3332 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3333 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3334 
3335 	/* deliberate bitwise OR to make one condition */
3336 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3337 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3338 			DBG_COUNTER_INC(encap_txd_encap_fail);
3339 			return (err);
3340 		}
3341 		m_head = *m_headp;
3342 	}
3343 
3344 retry:
3345 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3346 	    BUS_DMA_NOWAIT);
3347 defrag:
3348 	if (__predict_false(err)) {
3349 		switch (err) {
3350 		case EFBIG:
3351 			/* try collapse once and defrag once */
3352 			if (remap == 0) {
3353 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3354 				/* try defrag if collapsing fails */
3355 				if (m_head == NULL)
3356 					remap++;
3357 			}
3358 			if (remap == 1) {
3359 				txq->ift_mbuf_defrag++;
3360 				m_head = m_defrag(*m_headp, M_NOWAIT);
3361 			}
3362 			/*
3363 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3364 			 * failed to map an mbuf that was run through m_defrag
3365 			 */
3366 			MPASS(remap <= 1);
3367 			if (__predict_false(m_head == NULL || remap > 1))
3368 				goto defrag_failed;
3369 			remap++;
3370 			*m_headp = m_head;
3371 			goto retry;
3372 			break;
3373 		case ENOMEM:
3374 			txq->ift_no_tx_dma_setup++;
3375 			break;
3376 		default:
3377 			txq->ift_no_tx_dma_setup++;
3378 			m_freem(*m_headp);
3379 			DBG_COUNTER_INC(tx_frees);
3380 			*m_headp = NULL;
3381 			break;
3382 		}
3383 		txq->ift_map_failed++;
3384 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3385 		DBG_COUNTER_INC(encap_txd_encap_fail);
3386 		return (err);
3387 	}
3388 	ifsd_m[pidx] = m_head;
3389 	/*
3390 	 * XXX assumes a 1 to 1 relationship between segments and
3391 	 *        descriptors - this does not hold true on all drivers, e.g.
3392 	 *        cxgb
3393 	 */
3394 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3395 		txq->ift_no_desc_avail++;
3396 		bus_dmamap_unload(buf_tag, map);
3397 		DBG_COUNTER_INC(encap_txq_avail_fail);
3398 		DBG_COUNTER_INC(encap_txd_encap_fail);
3399 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3400 			GROUPTASK_ENQUEUE(&txq->ift_task);
3401 		return (ENOBUFS);
3402 	}
3403 	/*
3404 	 * On Intel cards we can greatly reduce the number of TX interrupts
3405 	 * we see by only setting report status on every Nth descriptor.
3406 	 * However, this also means that the driver will need to keep track
3407 	 * of the descriptors that RS was set on to check them for the DD bit.
3408 	 */
3409 	txq->ift_rs_pending += nsegs + 1;
3410 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3411 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3412 		pi.ipi_flags |= IPI_TX_INTR;
3413 		txq->ift_rs_pending = 0;
3414 	}
3415 
3416 	pi.ipi_segs = segs;
3417 	pi.ipi_nsegs = nsegs;
3418 
3419 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3420 #ifdef PKT_DEBUG
3421 	print_pkt(&pi);
3422 #endif
3423 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3424 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3425 		DBG_COUNTER_INC(tx_encap);
3426 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3427 
3428 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3429 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3430 			ndesc += txq->ift_size;
3431 			txq->ift_gen = 1;
3432 		}
3433 		/*
3434 		 * drivers can need as many as
3435 		 * two sentinels
3436 		 */
3437 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3438 		MPASS(pi.ipi_new_pidx != pidx);
3439 		MPASS(ndesc > 0);
3440 		txq->ift_in_use += ndesc;
3441 
3442 		/*
3443 		 * We update the last software descriptor again here because there may
3444 		 * be a sentinel and/or there may be more mbufs than segments
3445 		 */
3446 		txq->ift_pidx = pi.ipi_new_pidx;
3447 		txq->ift_npending += pi.ipi_ndescs;
3448 	} else {
3449 		*m_headp = m_head = iflib_remove_mbuf(txq);
3450 		if (err == EFBIG) {
3451 			txq->ift_txd_encap_efbig++;
3452 			if (remap < 2) {
3453 				remap = 1;
3454 				goto defrag;
3455 			}
3456 		}
3457 		goto defrag_failed;
3458 	}
3459 	/*
3460 	 * err can't possibly be non-zero here, so we don't neet to test it
3461 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3462 	 */
3463 	return (err);
3464 
3465 defrag_failed:
3466 	txq->ift_mbuf_defrag_failed++;
3467 	txq->ift_map_failed++;
3468 	m_freem(*m_headp);
3469 	DBG_COUNTER_INC(tx_frees);
3470 	*m_headp = NULL;
3471 	DBG_COUNTER_INC(encap_txd_encap_fail);
3472 	return (ENOMEM);
3473 }
3474 
3475 static void
3476 iflib_tx_desc_free(iflib_txq_t txq, int n)
3477 {
3478 	uint32_t qsize, cidx, mask, gen;
3479 	struct mbuf *m, **ifsd_m;
3480 	bool do_prefetch;
3481 
3482 	cidx = txq->ift_cidx;
3483 	gen = txq->ift_gen;
3484 	qsize = txq->ift_size;
3485 	mask = qsize-1;
3486 	ifsd_m = txq->ift_sds.ifsd_m;
3487 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3488 
3489 	while (n-- > 0) {
3490 		if (do_prefetch) {
3491 			prefetch(ifsd_m[(cidx + 3) & mask]);
3492 			prefetch(ifsd_m[(cidx + 4) & mask]);
3493 		}
3494 		if ((m = ifsd_m[cidx]) != NULL) {
3495 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3496 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3497 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3498 				    txq->ift_sds.ifsd_tso_map[cidx],
3499 				    BUS_DMASYNC_POSTWRITE);
3500 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3501 				    txq->ift_sds.ifsd_tso_map[cidx]);
3502 			} else {
3503 				bus_dmamap_sync(txq->ift_buf_tag,
3504 				    txq->ift_sds.ifsd_map[cidx],
3505 				    BUS_DMASYNC_POSTWRITE);
3506 				bus_dmamap_unload(txq->ift_buf_tag,
3507 				    txq->ift_sds.ifsd_map[cidx]);
3508 			}
3509 			/* XXX we don't support any drivers that batch packets yet */
3510 			MPASS(m->m_nextpkt == NULL);
3511 			m_freem(m);
3512 			ifsd_m[cidx] = NULL;
3513 #if MEMORY_LOGGING
3514 			txq->ift_dequeued++;
3515 #endif
3516 			DBG_COUNTER_INC(tx_frees);
3517 		}
3518 		if (__predict_false(++cidx == qsize)) {
3519 			cidx = 0;
3520 			gen = 0;
3521 		}
3522 	}
3523 	txq->ift_cidx = cidx;
3524 	txq->ift_gen = gen;
3525 }
3526 
3527 static __inline int
3528 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3529 {
3530 	int reclaim;
3531 	if_ctx_t ctx = txq->ift_ctx;
3532 
3533 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3534 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3535 
3536 	/*
3537 	 * Need a rate-limiting check so that this isn't called every time
3538 	 */
3539 	iflib_tx_credits_update(ctx, txq);
3540 	reclaim = DESC_RECLAIMABLE(txq);
3541 
3542 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3543 #ifdef INVARIANTS
3544 		if (iflib_verbose_debug) {
3545 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3546 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3547 			       reclaim, thresh);
3548 
3549 		}
3550 #endif
3551 		return (0);
3552 	}
3553 	iflib_tx_desc_free(txq, reclaim);
3554 	txq->ift_cleaned += reclaim;
3555 	txq->ift_in_use -= reclaim;
3556 
3557 	return (reclaim);
3558 }
3559 
3560 static struct mbuf **
3561 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3562 {
3563 	int next, size;
3564 	struct mbuf **items;
3565 
3566 	size = r->size;
3567 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3568 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3569 
3570 	prefetch(items[(cidx + offset) & (size-1)]);
3571 	if (remaining > 1) {
3572 		prefetch2cachelines(&items[next]);
3573 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3574 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3575 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3576 	}
3577 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3578 }
3579 
3580 static void
3581 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3582 {
3583 
3584 	ifmp_ring_check_drainage(txq->ift_br, budget);
3585 }
3586 
3587 static uint32_t
3588 iflib_txq_can_drain(struct ifmp_ring *r)
3589 {
3590 	iflib_txq_t txq = r->cookie;
3591 	if_ctx_t ctx = txq->ift_ctx;
3592 
3593 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3594 		return (1);
3595 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3596 	    BUS_DMASYNC_POSTREAD);
3597 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3598 	    false));
3599 }
3600 
3601 static uint32_t
3602 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3603 {
3604 	iflib_txq_t txq = r->cookie;
3605 	if_ctx_t ctx = txq->ift_ctx;
3606 	if_t ifp = ctx->ifc_ifp;
3607 	struct mbuf *m, **mp;
3608 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3609 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3610 	bool do_prefetch, rang, ring;
3611 
3612 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3613 			    !LINK_ACTIVE(ctx))) {
3614 		DBG_COUNTER_INC(txq_drain_notready);
3615 		return (0);
3616 	}
3617 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3618 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3619 	avail = IDXDIFF(pidx, cidx, r->size);
3620 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3621 		DBG_COUNTER_INC(txq_drain_flushing);
3622 		for (i = 0; i < avail; i++) {
3623 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3624 				m_free(r->items[(cidx + i) & (r->size-1)]);
3625 			r->items[(cidx + i) & (r->size-1)] = NULL;
3626 		}
3627 		return (avail);
3628 	}
3629 
3630 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3631 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3632 		CALLOUT_LOCK(txq);
3633 		callout_stop(&txq->ift_timer);
3634 		CALLOUT_UNLOCK(txq);
3635 		DBG_COUNTER_INC(txq_drain_oactive);
3636 		return (0);
3637 	}
3638 	if (reclaimed)
3639 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3640 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3641 	count = MIN(avail, TX_BATCH_SIZE);
3642 #ifdef INVARIANTS
3643 	if (iflib_verbose_debug)
3644 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3645 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3646 #endif
3647 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3648 	txq_avail = TXQ_AVAIL(txq);
3649 	err = 0;
3650 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3651 		int rem = do_prefetch ? count - i : 0;
3652 
3653 		mp = _ring_peek_one(r, cidx, i, rem);
3654 		MPASS(mp != NULL && *mp != NULL);
3655 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3656 			consumed++;
3657 			continue;
3658 		}
3659 		in_use_prev = txq->ift_in_use;
3660 		err = iflib_encap(txq, mp);
3661 		if (__predict_false(err)) {
3662 			/* no room - bail out */
3663 			if (err == ENOBUFS)
3664 				break;
3665 			consumed++;
3666 			/* we can't send this packet - skip it */
3667 			continue;
3668 		}
3669 		consumed++;
3670 		pkt_sent++;
3671 		m = *mp;
3672 		DBG_COUNTER_INC(tx_sent);
3673 		bytes_sent += m->m_pkthdr.len;
3674 		mcast_sent += !!(m->m_flags & M_MCAST);
3675 		txq_avail = TXQ_AVAIL(txq);
3676 
3677 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3678 		ETHER_BPF_MTAP(ifp, m);
3679 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3680 			break;
3681 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3682 	}
3683 
3684 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3685 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3686 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3687 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3688 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3689 	if (mcast_sent)
3690 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3691 #ifdef INVARIANTS
3692 	if (iflib_verbose_debug)
3693 		printf("consumed=%d\n", consumed);
3694 #endif
3695 	return (consumed);
3696 }
3697 
3698 static uint32_t
3699 iflib_txq_drain_always(struct ifmp_ring *r)
3700 {
3701 	return (1);
3702 }
3703 
3704 static uint32_t
3705 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3706 {
3707 	int i, avail;
3708 	struct mbuf **mp;
3709 	iflib_txq_t txq;
3710 
3711 	txq = r->cookie;
3712 
3713 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3714 	CALLOUT_LOCK(txq);
3715 	callout_stop(&txq->ift_timer);
3716 	CALLOUT_UNLOCK(txq);
3717 
3718 	avail = IDXDIFF(pidx, cidx, r->size);
3719 	for (i = 0; i < avail; i++) {
3720 		mp = _ring_peek_one(r, cidx, i, avail - i);
3721 		if (__predict_false(*mp == (struct mbuf *)txq))
3722 			continue;
3723 		m_freem(*mp);
3724 		DBG_COUNTER_INC(tx_frees);
3725 	}
3726 	MPASS(ifmp_ring_is_stalled(r) == 0);
3727 	return (avail);
3728 }
3729 
3730 static void
3731 iflib_ifmp_purge(iflib_txq_t txq)
3732 {
3733 	struct ifmp_ring *r;
3734 
3735 	r = txq->ift_br;
3736 	r->drain = iflib_txq_drain_free;
3737 	r->can_drain = iflib_txq_drain_always;
3738 
3739 	ifmp_ring_check_drainage(r, r->size);
3740 
3741 	r->drain = iflib_txq_drain;
3742 	r->can_drain = iflib_txq_can_drain;
3743 }
3744 
3745 static void
3746 _task_fn_tx(void *context)
3747 {
3748 	iflib_txq_t txq = context;
3749 	if_ctx_t ctx = txq->ift_ctx;
3750 #if defined(ALTQ) || defined(DEV_NETMAP)
3751 	if_t ifp = ctx->ifc_ifp;
3752 #endif
3753 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3754 
3755 #ifdef IFLIB_DIAGNOSTICS
3756 	txq->ift_cpu_exec_count[curcpu]++;
3757 #endif
3758 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3759 		return;
3760 #ifdef DEV_NETMAP
3761 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3762 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3763 		    BUS_DMASYNC_POSTREAD);
3764 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3765 			netmap_tx_irq(ifp, txq->ift_id);
3766 		if (ctx->ifc_flags & IFC_LEGACY)
3767 			IFDI_INTR_ENABLE(ctx);
3768 		else
3769 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3770 		return;
3771 	}
3772 #endif
3773 #ifdef ALTQ
3774 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3775 		iflib_altq_if_start(ifp);
3776 #endif
3777 	if (txq->ift_db_pending)
3778 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3779 	else if (!abdicate)
3780 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3781 	/*
3782 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3783 	 */
3784 	if (abdicate)
3785 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3786 	if (ctx->ifc_flags & IFC_LEGACY)
3787 		IFDI_INTR_ENABLE(ctx);
3788 	else
3789 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3790 }
3791 
3792 static void
3793 _task_fn_rx(void *context)
3794 {
3795 	iflib_rxq_t rxq = context;
3796 	if_ctx_t ctx = rxq->ifr_ctx;
3797 	uint8_t more;
3798 	uint16_t budget;
3799 #ifdef DEV_NETMAP
3800 	u_int work = 0;
3801 	int nmirq;
3802 #endif
3803 
3804 #ifdef IFLIB_DIAGNOSTICS
3805 	rxq->ifr_cpu_exec_count[curcpu]++;
3806 #endif
3807 	DBG_COUNTER_INC(task_fn_rxs);
3808 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3809 		return;
3810 #ifdef DEV_NETMAP
3811 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3812 	if (nmirq != NM_IRQ_PASS) {
3813 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3814 		goto skip_rxeof;
3815 	}
3816 #endif
3817 	budget = ctx->ifc_sysctl_rx_budget;
3818 	if (budget == 0)
3819 		budget = 16;	/* XXX */
3820 	more = iflib_rxeof(rxq, budget);
3821 #ifdef DEV_NETMAP
3822 skip_rxeof:
3823 #endif
3824 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3825 		if (ctx->ifc_flags & IFC_LEGACY)
3826 			IFDI_INTR_ENABLE(ctx);
3827 		else
3828 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3829 		DBG_COUNTER_INC(rx_intr_enables);
3830 	}
3831 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3832 		return;
3833 
3834 	if (more & IFLIB_RXEOF_MORE)
3835 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3836 	else if (more & IFLIB_RXEOF_EMPTY)
3837 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3838 }
3839 
3840 static void
3841 _task_fn_admin(void *context)
3842 {
3843 	if_ctx_t ctx = context;
3844 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3845 	iflib_txq_t txq;
3846 	int i;
3847 	bool oactive, running, do_reset, do_watchdog, in_detach;
3848 	uint32_t reset_on = hz / 2;
3849 
3850 	STATE_LOCK(ctx);
3851 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3852 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3853 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3854 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3855 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3856 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3857 	STATE_UNLOCK(ctx);
3858 
3859 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3860 		return;
3861 	if (in_detach)
3862 		return;
3863 
3864 	CTX_LOCK(ctx);
3865 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3866 		CALLOUT_LOCK(txq);
3867 		callout_stop(&txq->ift_timer);
3868 		CALLOUT_UNLOCK(txq);
3869 	}
3870 	if (do_watchdog) {
3871 		ctx->ifc_watchdog_events++;
3872 		IFDI_WATCHDOG_RESET(ctx);
3873 	}
3874 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3875 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3876 #ifdef DEV_NETMAP
3877 		reset_on = hz / 2;
3878 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3879 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3880 #endif
3881 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3882 	}
3883 	IFDI_LINK_INTR_ENABLE(ctx);
3884 	if (do_reset)
3885 		iflib_if_init_locked(ctx);
3886 	CTX_UNLOCK(ctx);
3887 
3888 	if (LINK_ACTIVE(ctx) == 0)
3889 		return;
3890 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3891 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3892 }
3893 
3894 
3895 static void
3896 _task_fn_iov(void *context)
3897 {
3898 	if_ctx_t ctx = context;
3899 
3900 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3901 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3902 		return;
3903 
3904 	CTX_LOCK(ctx);
3905 	IFDI_VFLR_HANDLE(ctx);
3906 	CTX_UNLOCK(ctx);
3907 }
3908 
3909 static int
3910 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3911 {
3912 	int err;
3913 	if_int_delay_info_t info;
3914 	if_ctx_t ctx;
3915 
3916 	info = (if_int_delay_info_t)arg1;
3917 	ctx = info->iidi_ctx;
3918 	info->iidi_req = req;
3919 	info->iidi_oidp = oidp;
3920 	CTX_LOCK(ctx);
3921 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3922 	CTX_UNLOCK(ctx);
3923 	return (err);
3924 }
3925 
3926 /*********************************************************************
3927  *
3928  *  IFNET FUNCTIONS
3929  *
3930  **********************************************************************/
3931 
3932 static void
3933 iflib_if_init_locked(if_ctx_t ctx)
3934 {
3935 	iflib_stop(ctx);
3936 	iflib_init_locked(ctx);
3937 }
3938 
3939 
3940 static void
3941 iflib_if_init(void *arg)
3942 {
3943 	if_ctx_t ctx = arg;
3944 
3945 	CTX_LOCK(ctx);
3946 	iflib_if_init_locked(ctx);
3947 	CTX_UNLOCK(ctx);
3948 }
3949 
3950 static int
3951 iflib_if_transmit(if_t ifp, struct mbuf *m)
3952 {
3953 	if_ctx_t	ctx = if_getsoftc(ifp);
3954 
3955 	iflib_txq_t txq;
3956 	int err, qidx;
3957 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3958 
3959 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3960 		DBG_COUNTER_INC(tx_frees);
3961 		m_freem(m);
3962 		return (ENETDOWN);
3963 	}
3964 
3965 	MPASS(m->m_nextpkt == NULL);
3966 	/* ALTQ-enabled interfaces always use queue 0. */
3967 	qidx = 0;
3968 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3969 		qidx = QIDX(ctx, m);
3970 	/*
3971 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3972 	 */
3973 	txq = &ctx->ifc_txqs[qidx];
3974 
3975 #ifdef DRIVER_BACKPRESSURE
3976 	if (txq->ift_closed) {
3977 		while (m != NULL) {
3978 			next = m->m_nextpkt;
3979 			m->m_nextpkt = NULL;
3980 			m_freem(m);
3981 			DBG_COUNTER_INC(tx_frees);
3982 			m = next;
3983 		}
3984 		return (ENOBUFS);
3985 	}
3986 #endif
3987 #ifdef notyet
3988 	qidx = count = 0;
3989 	mp = marr;
3990 	next = m;
3991 	do {
3992 		count++;
3993 		next = next->m_nextpkt;
3994 	} while (next != NULL);
3995 
3996 	if (count > nitems(marr))
3997 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3998 			/* XXX check nextpkt */
3999 			m_freem(m);
4000 			/* XXX simplify for now */
4001 			DBG_COUNTER_INC(tx_frees);
4002 			return (ENOBUFS);
4003 		}
4004 	for (next = m, i = 0; next != NULL; i++) {
4005 		mp[i] = next;
4006 		next = next->m_nextpkt;
4007 		mp[i]->m_nextpkt = NULL;
4008 	}
4009 #endif
4010 	DBG_COUNTER_INC(tx_seen);
4011 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4012 
4013 	if (abdicate)
4014 		GROUPTASK_ENQUEUE(&txq->ift_task);
4015  	if (err) {
4016 		if (!abdicate)
4017 			GROUPTASK_ENQUEUE(&txq->ift_task);
4018 		/* support forthcoming later */
4019 #ifdef DRIVER_BACKPRESSURE
4020 		txq->ift_closed = TRUE;
4021 #endif
4022 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4023 		m_freem(m);
4024 		DBG_COUNTER_INC(tx_frees);
4025 	}
4026 
4027 	return (err);
4028 }
4029 
4030 #ifdef ALTQ
4031 /*
4032  * The overall approach to integrating iflib with ALTQ is to continue to use
4033  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4034  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4035  * is redundant/unnecessary, but doing so minimizes the amount of
4036  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4037  * redundantly queueing to an intermediate mp_ring is swamped by the
4038  * performance limitations inherent in using ALTQ.
4039  *
4040  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4041  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4042  * given interface.  If ALTQ is enabled for an interface, then all
4043  * transmitted packets for that interface will be submitted to the ALTQ
4044  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4045  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4046  * update stats that the iflib machinery handles, and which is sensitve to
4047  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4048  * will be installed as the start routine for use by ALTQ facilities that
4049  * need to trigger queue drains on a scheduled basis.
4050  *
4051  */
4052 static void
4053 iflib_altq_if_start(if_t ifp)
4054 {
4055 	struct ifaltq *ifq = &ifp->if_snd;
4056 	struct mbuf *m;
4057 
4058 	IFQ_LOCK(ifq);
4059 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4060 	while (m != NULL) {
4061 		iflib_if_transmit(ifp, m);
4062 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4063 	}
4064 	IFQ_UNLOCK(ifq);
4065 }
4066 
4067 static int
4068 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4069 {
4070 	int err;
4071 
4072 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4073 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4074 		if (err == 0)
4075 			iflib_altq_if_start(ifp);
4076 	} else
4077 		err = iflib_if_transmit(ifp, m);
4078 
4079 	return (err);
4080 }
4081 #endif /* ALTQ */
4082 
4083 static void
4084 iflib_if_qflush(if_t ifp)
4085 {
4086 	if_ctx_t ctx = if_getsoftc(ifp);
4087 	iflib_txq_t txq = ctx->ifc_txqs;
4088 	int i;
4089 
4090 	STATE_LOCK(ctx);
4091 	ctx->ifc_flags |= IFC_QFLUSH;
4092 	STATE_UNLOCK(ctx);
4093 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4094 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4095 			iflib_txq_check_drain(txq, 0);
4096 	STATE_LOCK(ctx);
4097 	ctx->ifc_flags &= ~IFC_QFLUSH;
4098 	STATE_UNLOCK(ctx);
4099 
4100 	/*
4101 	 * When ALTQ is enabled, this will also take care of purging the
4102 	 * ALTQ queue(s).
4103 	 */
4104 	if_qflush(ifp);
4105 }
4106 
4107 
4108 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4109 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4110 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4111 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4112 
4113 static int
4114 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4115 {
4116 	if_ctx_t ctx = if_getsoftc(ifp);
4117 	struct ifreq	*ifr = (struct ifreq *)data;
4118 #if defined(INET) || defined(INET6)
4119 	struct ifaddr	*ifa = (struct ifaddr *)data;
4120 #endif
4121 	bool		avoid_reset = false;
4122 	int		err = 0, reinit = 0, bits;
4123 
4124 	switch (command) {
4125 	case SIOCSIFADDR:
4126 #ifdef INET
4127 		if (ifa->ifa_addr->sa_family == AF_INET)
4128 			avoid_reset = true;
4129 #endif
4130 #ifdef INET6
4131 		if (ifa->ifa_addr->sa_family == AF_INET6)
4132 			avoid_reset = true;
4133 #endif
4134 		/*
4135 		** Calling init results in link renegotiation,
4136 		** so we avoid doing it when possible.
4137 		*/
4138 		if (avoid_reset) {
4139 			if_setflagbits(ifp, IFF_UP,0);
4140 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4141 				reinit = 1;
4142 #ifdef INET
4143 			if (!(if_getflags(ifp) & IFF_NOARP))
4144 				arp_ifinit(ifp, ifa);
4145 #endif
4146 		} else
4147 			err = ether_ioctl(ifp, command, data);
4148 		break;
4149 	case SIOCSIFMTU:
4150 		CTX_LOCK(ctx);
4151 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4152 			CTX_UNLOCK(ctx);
4153 			break;
4154 		}
4155 		bits = if_getdrvflags(ifp);
4156 		/* stop the driver and free any clusters before proceeding */
4157 		iflib_stop(ctx);
4158 
4159 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4160 			STATE_LOCK(ctx);
4161 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4162 				ctx->ifc_flags |= IFC_MULTISEG;
4163 			else
4164 				ctx->ifc_flags &= ~IFC_MULTISEG;
4165 			STATE_UNLOCK(ctx);
4166 			err = if_setmtu(ifp, ifr->ifr_mtu);
4167 		}
4168 		iflib_init_locked(ctx);
4169 		STATE_LOCK(ctx);
4170 		if_setdrvflags(ifp, bits);
4171 		STATE_UNLOCK(ctx);
4172 		CTX_UNLOCK(ctx);
4173 		break;
4174 	case SIOCSIFFLAGS:
4175 		CTX_LOCK(ctx);
4176 		if (if_getflags(ifp) & IFF_UP) {
4177 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4178 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4179 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4180 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4181 				}
4182 			} else
4183 				reinit = 1;
4184 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4185 			iflib_stop(ctx);
4186 		}
4187 		ctx->ifc_if_flags = if_getflags(ifp);
4188 		CTX_UNLOCK(ctx);
4189 		break;
4190 	case SIOCADDMULTI:
4191 	case SIOCDELMULTI:
4192 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4193 			CTX_LOCK(ctx);
4194 			IFDI_INTR_DISABLE(ctx);
4195 			IFDI_MULTI_SET(ctx);
4196 			IFDI_INTR_ENABLE(ctx);
4197 			CTX_UNLOCK(ctx);
4198 		}
4199 		break;
4200 	case SIOCSIFMEDIA:
4201 		CTX_LOCK(ctx);
4202 		IFDI_MEDIA_SET(ctx);
4203 		CTX_UNLOCK(ctx);
4204 		/* FALLTHROUGH */
4205 	case SIOCGIFMEDIA:
4206 	case SIOCGIFXMEDIA:
4207 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4208 		break;
4209 	case SIOCGI2C:
4210 	{
4211 		struct ifi2creq i2c;
4212 
4213 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4214 		if (err != 0)
4215 			break;
4216 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4217 			err = EINVAL;
4218 			break;
4219 		}
4220 		if (i2c.len > sizeof(i2c.data)) {
4221 			err = EINVAL;
4222 			break;
4223 		}
4224 
4225 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4226 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4227 			    sizeof(i2c));
4228 		break;
4229 	}
4230 	case SIOCSIFCAP:
4231 	{
4232 		int mask, setmask, oldmask;
4233 
4234 		oldmask = if_getcapenable(ifp);
4235 		mask = ifr->ifr_reqcap ^ oldmask;
4236 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4237 		setmask = 0;
4238 #ifdef TCP_OFFLOAD
4239 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4240 #endif
4241 		setmask |= (mask & IFCAP_FLAGS);
4242 		setmask |= (mask & IFCAP_WOL);
4243 
4244 		/*
4245 		 * If any RX csum has changed, change all the ones that
4246 		 * are supported by the driver.
4247 		 */
4248 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4249 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4250 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4251 		}
4252 
4253 		/*
4254 		 * want to ensure that traffic has stopped before we change any of the flags
4255 		 */
4256 		if (setmask) {
4257 			CTX_LOCK(ctx);
4258 			bits = if_getdrvflags(ifp);
4259 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4260 				iflib_stop(ctx);
4261 			STATE_LOCK(ctx);
4262 			if_togglecapenable(ifp, setmask);
4263 			STATE_UNLOCK(ctx);
4264 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4265 				iflib_init_locked(ctx);
4266 			STATE_LOCK(ctx);
4267 			if_setdrvflags(ifp, bits);
4268 			STATE_UNLOCK(ctx);
4269 			CTX_UNLOCK(ctx);
4270 		}
4271 		if_vlancap(ifp);
4272 		break;
4273 	}
4274 	case SIOCGPRIVATE_0:
4275 	case SIOCSDRVSPEC:
4276 	case SIOCGDRVSPEC:
4277 		CTX_LOCK(ctx);
4278 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4279 		CTX_UNLOCK(ctx);
4280 		break;
4281 	default:
4282 		err = ether_ioctl(ifp, command, data);
4283 		break;
4284 	}
4285 	if (reinit)
4286 		iflib_if_init(ctx);
4287 	return (err);
4288 }
4289 
4290 static uint64_t
4291 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4292 {
4293 	if_ctx_t ctx = if_getsoftc(ifp);
4294 
4295 	return (IFDI_GET_COUNTER(ctx, cnt));
4296 }
4297 
4298 /*********************************************************************
4299  *
4300  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4301  *
4302  **********************************************************************/
4303 
4304 static void
4305 iflib_vlan_register(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 	if (iflib_in_detach(ctx))
4316 		return;
4317 
4318 	CTX_LOCK(ctx);
4319 	/* Driver may need all untagged packets to be flushed */
4320 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4321 		iflib_stop(ctx);
4322 	IFDI_VLAN_REGISTER(ctx, vtag);
4323 	/* Re-init to load the changes, if required */
4324 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4325 		iflib_init_locked(ctx);
4326 	CTX_UNLOCK(ctx);
4327 }
4328 
4329 static void
4330 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4331 {
4332 	if_ctx_t ctx = if_getsoftc(ifp);
4333 
4334 	if ((void *)ctx != arg)
4335 		return;
4336 
4337 	if ((vtag == 0) || (vtag > 4095))
4338 		return;
4339 
4340 	CTX_LOCK(ctx);
4341 	/* Driver may need all tagged packets to be flushed */
4342 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4343 		iflib_stop(ctx);
4344 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4345 	/* Re-init to load the changes, if required */
4346 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4347 		iflib_init_locked(ctx);
4348 	CTX_UNLOCK(ctx);
4349 }
4350 
4351 static void
4352 iflib_led_func(void *arg, int onoff)
4353 {
4354 	if_ctx_t ctx = arg;
4355 
4356 	CTX_LOCK(ctx);
4357 	IFDI_LED_FUNC(ctx, onoff);
4358 	CTX_UNLOCK(ctx);
4359 }
4360 
4361 /*********************************************************************
4362  *
4363  *  BUS FUNCTION DEFINITIONS
4364  *
4365  **********************************************************************/
4366 
4367 int
4368 iflib_device_probe(device_t dev)
4369 {
4370 	const pci_vendor_info_t *ent;
4371 	if_shared_ctx_t sctx;
4372 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4373 	uint16_t pci_vendor_id;
4374 
4375 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4376 		return (ENOTSUP);
4377 
4378 	pci_vendor_id = pci_get_vendor(dev);
4379 	pci_device_id = pci_get_device(dev);
4380 	pci_subvendor_id = pci_get_subvendor(dev);
4381 	pci_subdevice_id = pci_get_subdevice(dev);
4382 	pci_rev_id = pci_get_revid(dev);
4383 	if (sctx->isc_parse_devinfo != NULL)
4384 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4385 
4386 	ent = sctx->isc_vendor_info;
4387 	while (ent->pvi_vendor_id != 0) {
4388 		if (pci_vendor_id != ent->pvi_vendor_id) {
4389 			ent++;
4390 			continue;
4391 		}
4392 		if ((pci_device_id == ent->pvi_device_id) &&
4393 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4394 		     (ent->pvi_subvendor_id == 0)) &&
4395 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4396 		     (ent->pvi_subdevice_id == 0)) &&
4397 		    ((pci_rev_id == ent->pvi_rev_id) ||
4398 		     (ent->pvi_rev_id == 0))) {
4399 
4400 			device_set_desc_copy(dev, ent->pvi_name);
4401 			/* this needs to be changed to zero if the bus probing code
4402 			 * ever stops re-probing on best match because the sctx
4403 			 * may have its values over written by register calls
4404 			 * in subsequent probes
4405 			 */
4406 			return (BUS_PROBE_DEFAULT);
4407 		}
4408 		ent++;
4409 	}
4410 	return (ENXIO);
4411 }
4412 
4413 int
4414 iflib_device_probe_vendor(device_t dev)
4415 {
4416 	int probe;
4417 
4418 	probe = iflib_device_probe(dev);
4419 	if (probe == BUS_PROBE_DEFAULT)
4420 		return (BUS_PROBE_VENDOR);
4421 	else
4422 		return (probe);
4423 }
4424 
4425 static void
4426 iflib_reset_qvalues(if_ctx_t ctx)
4427 {
4428 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4429 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4430 	device_t dev = ctx->ifc_dev;
4431 	int i;
4432 
4433 	if (ctx->ifc_sysctl_ntxqs != 0)
4434 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4435 	if (ctx->ifc_sysctl_nrxqs != 0)
4436 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4437 
4438 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4439 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4440 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4441 		else
4442 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4443 	}
4444 
4445 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4446 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4447 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4448 		else
4449 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4450 	}
4451 
4452 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4453 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4454 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4455 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4456 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4457 		}
4458 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4459 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4460 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4461 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4462 		}
4463 		if (!powerof2(scctx->isc_nrxd[i])) {
4464 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4465 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4466 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4467 		}
4468 	}
4469 
4470 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4471 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4472 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4473 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4474 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4475 		}
4476 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4477 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4478 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4479 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4480 		}
4481 		if (!powerof2(scctx->isc_ntxd[i])) {
4482 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4483 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4484 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4485 		}
4486 	}
4487 }
4488 
4489 static void
4490 iflib_add_pfil(if_ctx_t ctx)
4491 {
4492 	struct pfil_head *pfil;
4493 	struct pfil_head_args pa;
4494 	iflib_rxq_t rxq;
4495 	int i;
4496 
4497 	pa.pa_version = PFIL_VERSION;
4498 	pa.pa_flags = PFIL_IN;
4499 	pa.pa_type = PFIL_TYPE_ETHERNET;
4500 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4501 	pfil = pfil_head_register(&pa);
4502 
4503 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4504 		rxq->pfil = pfil;
4505 	}
4506 }
4507 
4508 static void
4509 iflib_rem_pfil(if_ctx_t ctx)
4510 {
4511 	struct pfil_head *pfil;
4512 	iflib_rxq_t rxq;
4513 	int i;
4514 
4515 	rxq = ctx->ifc_rxqs;
4516 	pfil = rxq->pfil;
4517 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4518 		rxq->pfil = NULL;
4519 	}
4520 	pfil_head_unregister(pfil);
4521 }
4522 
4523 static uint16_t
4524 get_ctx_core_offset(if_ctx_t ctx)
4525 {
4526 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4527 	struct cpu_offset *op;
4528 	uint16_t qc;
4529 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4530 
4531 	if (ret != CORE_OFFSET_UNSPECIFIED)
4532 		return (ret);
4533 
4534 	if (ctx->ifc_sysctl_separate_txrx)
4535 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4536 	else
4537 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4538 
4539 	mtx_lock(&cpu_offset_mtx);
4540 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4541 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4542 			ret = op->offset;
4543 			op->offset += qc;
4544 			MPASS(op->refcount < UINT_MAX);
4545 			op->refcount++;
4546 			break;
4547 		}
4548 	}
4549 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4550 		ret = 0;
4551 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4552 		    M_NOWAIT | M_ZERO);
4553 		if (op == NULL) {
4554 			device_printf(ctx->ifc_dev,
4555 			    "allocation for cpu offset failed.\n");
4556 		} else {
4557 			op->offset = qc;
4558 			op->refcount = 1;
4559 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4560 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4561 		}
4562 	}
4563 	mtx_unlock(&cpu_offset_mtx);
4564 
4565 	return (ret);
4566 }
4567 
4568 static void
4569 unref_ctx_core_offset(if_ctx_t ctx)
4570 {
4571 	struct cpu_offset *op, *top;
4572 
4573 	mtx_lock(&cpu_offset_mtx);
4574 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4575 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4576 			MPASS(op->refcount > 0);
4577 			op->refcount--;
4578 			if (op->refcount == 0) {
4579 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4580 				free(op, M_IFLIB);
4581 			}
4582 			break;
4583 		}
4584 	}
4585 	mtx_unlock(&cpu_offset_mtx);
4586 }
4587 
4588 int
4589 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4590 {
4591 	if_ctx_t ctx;
4592 	if_t ifp;
4593 	if_softc_ctx_t scctx;
4594 	kobjop_desc_t kobj_desc;
4595 	kobj_method_t *kobj_method;
4596 	int err, msix, rid;
4597 	uint16_t main_rxq, main_txq;
4598 
4599 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4600 
4601 	if (sc == NULL) {
4602 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4603 		device_set_softc(dev, ctx);
4604 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4605 	}
4606 
4607 	ctx->ifc_sctx = sctx;
4608 	ctx->ifc_dev = dev;
4609 	ctx->ifc_softc = sc;
4610 
4611 	if ((err = iflib_register(ctx)) != 0) {
4612 		device_printf(dev, "iflib_register failed %d\n", err);
4613 		goto fail_ctx_free;
4614 	}
4615 	iflib_add_device_sysctl_pre(ctx);
4616 
4617 	scctx = &ctx->ifc_softc_ctx;
4618 	ifp = ctx->ifc_ifp;
4619 
4620 	iflib_reset_qvalues(ctx);
4621 	CTX_LOCK(ctx);
4622 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4623 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4624 		goto fail_unlock;
4625 	}
4626 	_iflib_pre_assert(scctx);
4627 	ctx->ifc_txrx = *scctx->isc_txrx;
4628 
4629 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4630 		ctx->ifc_mediap = scctx->isc_media;
4631 
4632 #ifdef INVARIANTS
4633 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4634 		MPASS(scctx->isc_tx_csum_flags);
4635 #endif
4636 
4637 	if_setcapabilities(ifp,
4638 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4639 	if_setcapenable(ifp,
4640 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4641 
4642 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4643 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4644 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4645 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4646 
4647 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4648 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4649 
4650 	/* XXX change for per-queue sizes */
4651 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4652 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4653 
4654 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4655 	    MAX_SINGLE_PACKET_FRACTION)
4656 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4657 		    MAX_SINGLE_PACKET_FRACTION);
4658 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4659 	    MAX_SINGLE_PACKET_FRACTION)
4660 		scctx->isc_tx_tso_segments_max = max(1,
4661 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4662 
4663 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4664 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4665 		/*
4666 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4667 		 * but some MACs do.
4668 		 */
4669 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4670 		    IP_MAXPACKET));
4671 		/*
4672 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4673 		 * into account.  In the worst case, each of these calls will
4674 		 * add another mbuf and, thus, the requirement for another DMA
4675 		 * segment.  So for best performance, it doesn't make sense to
4676 		 * advertize a maximum of TSO segments that typically will
4677 		 * require defragmentation in iflib_encap().
4678 		 */
4679 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4680 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4681 	}
4682 	if (scctx->isc_rss_table_size == 0)
4683 		scctx->isc_rss_table_size = 64;
4684 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4685 
4686 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4687 	/* XXX format name */
4688 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4689 	    NULL, NULL, "admin");
4690 
4691 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4692 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4693 		device_printf(dev, "Unable to fetch CPU list\n");
4694 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4695 	}
4696 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4697 
4698 	/*
4699 	** Now set up MSI or MSI-X, should return us the number of supported
4700 	** vectors (will be 1 for a legacy interrupt and MSI).
4701 	*/
4702 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4703 		msix = scctx->isc_vectors;
4704 	} else if (scctx->isc_msix_bar != 0)
4705 	       /*
4706 		* The simple fact that isc_msix_bar is not 0 does not mean we
4707 		* we have a good value there that is known to work.
4708 		*/
4709 		msix = iflib_msix_init(ctx);
4710 	else {
4711 		scctx->isc_vectors = 1;
4712 		scctx->isc_ntxqsets = 1;
4713 		scctx->isc_nrxqsets = 1;
4714 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4715 		msix = 0;
4716 	}
4717 	/* Get memory for the station queues */
4718 	if ((err = iflib_queues_alloc(ctx))) {
4719 		device_printf(dev, "Unable to allocate queue memory\n");
4720 		goto fail_intr_free;
4721 	}
4722 
4723 	if ((err = iflib_qset_structures_setup(ctx)))
4724 		goto fail_queues;
4725 
4726 	/*
4727 	 * Now that we know how many queues there are, get the core offset.
4728 	 */
4729 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4730 
4731 	if (msix > 1) {
4732 		/*
4733 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4734 		 * aren't the default NULL implementation.
4735 		 */
4736 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4737 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4738 		    kobj_desc);
4739 		if (kobj_method == &kobj_desc->deflt) {
4740 			device_printf(dev,
4741 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4742 			err = EOPNOTSUPP;
4743 			goto fail_queues;
4744 		}
4745 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4746 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4747 		    kobj_desc);
4748 		if (kobj_method == &kobj_desc->deflt) {
4749 			device_printf(dev,
4750 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4751 			err = EOPNOTSUPP;
4752 			goto fail_queues;
4753 		}
4754 
4755 		/*
4756 		 * Assign the MSI-X vectors.
4757 		 * Note that the default NULL ifdi_msix_intr_assign method will
4758 		 * fail here, too.
4759 		 */
4760 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4761 		if (err != 0) {
4762 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4763 			    err);
4764 			goto fail_queues;
4765 		}
4766 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4767 		rid = 0;
4768 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4769 			MPASS(msix == 1);
4770 			rid = 1;
4771 		}
4772 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4773 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4774 			goto fail_queues;
4775 		}
4776 	} else {
4777 		device_printf(dev,
4778 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4779 		err = ENODEV;
4780 		goto fail_intr_free;
4781 	}
4782 
4783 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4784 
4785 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4786 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4787 		goto fail_detach;
4788 	}
4789 
4790 	/*
4791 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4792 	 * This must appear after the call to ether_ifattach() because
4793 	 * ether_ifattach() sets if_hdrlen to the default value.
4794 	 */
4795 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4796 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4797 
4798 	if ((err = iflib_netmap_attach(ctx))) {
4799 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4800 		goto fail_detach;
4801 	}
4802 	*ctxp = ctx;
4803 
4804 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4805 
4806 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4807 	iflib_add_device_sysctl_post(ctx);
4808 	iflib_add_pfil(ctx);
4809 	ctx->ifc_flags |= IFC_INIT_DONE;
4810 	CTX_UNLOCK(ctx);
4811 
4812 	return (0);
4813 
4814 fail_detach:
4815 	ether_ifdetach(ctx->ifc_ifp);
4816 fail_intr_free:
4817 	iflib_free_intr_mem(ctx);
4818 fail_queues:
4819 	iflib_tx_structures_free(ctx);
4820 	iflib_rx_structures_free(ctx);
4821 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4822 	IFDI_DETACH(ctx);
4823 fail_unlock:
4824 	CTX_UNLOCK(ctx);
4825 	iflib_deregister(ctx);
4826 fail_ctx_free:
4827 	device_set_softc(ctx->ifc_dev, NULL);
4828         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4829                 free(ctx->ifc_softc, M_IFLIB);
4830         free(ctx, M_IFLIB);
4831 	return (err);
4832 }
4833 
4834 int
4835 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4836 					  struct iflib_cloneattach_ctx *clctx)
4837 {
4838 	int err;
4839 	if_ctx_t ctx;
4840 	if_t ifp;
4841 	if_softc_ctx_t scctx;
4842 	int i;
4843 	void *sc;
4844 	uint16_t main_txq;
4845 	uint16_t main_rxq;
4846 
4847 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4848 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4849 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4850 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4851 		ctx->ifc_flags |= IFC_PSEUDO;
4852 
4853 	ctx->ifc_sctx = sctx;
4854 	ctx->ifc_softc = sc;
4855 	ctx->ifc_dev = dev;
4856 
4857 	if ((err = iflib_register(ctx)) != 0) {
4858 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4859 		goto fail_ctx_free;
4860 	}
4861 	iflib_add_device_sysctl_pre(ctx);
4862 
4863 	scctx = &ctx->ifc_softc_ctx;
4864 	ifp = ctx->ifc_ifp;
4865 
4866 	iflib_reset_qvalues(ctx);
4867 	CTX_LOCK(ctx);
4868 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4869 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4870 		goto fail_unlock;
4871 	}
4872 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4873 		ether_gen_addr(ifp, &ctx->ifc_mac);
4874 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4875 								clctx->cc_params)) != 0) {
4876 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4877 		goto fail_ctx_free;
4878 	}
4879 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4880 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4881 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4882 
4883 #ifdef INVARIANTS
4884 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4885 		MPASS(scctx->isc_tx_csum_flags);
4886 #endif
4887 
4888 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4889 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4890 
4891 	ifp->if_flags |= IFF_NOGROUP;
4892 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4893 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4894 
4895 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4896 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4897 			goto fail_detach;
4898 		}
4899 		*ctxp = ctx;
4900 
4901 		/*
4902 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4903 		 * This must appear after the call to ether_ifattach() because
4904 		 * ether_ifattach() sets if_hdrlen to the default value.
4905 		 */
4906 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4907 			if_setifheaderlen(ifp,
4908 			    sizeof(struct ether_vlan_header));
4909 
4910 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4911 		iflib_add_device_sysctl_post(ctx);
4912 		ctx->ifc_flags |= IFC_INIT_DONE;
4913 		CTX_UNLOCK(ctx);
4914 		return (0);
4915 	}
4916 	_iflib_pre_assert(scctx);
4917 	ctx->ifc_txrx = *scctx->isc_txrx;
4918 
4919 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4920 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4921 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4922 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4923 
4924 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4925 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4926 
4927 	/* XXX change for per-queue sizes */
4928 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4929 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4930 
4931 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4932 	    MAX_SINGLE_PACKET_FRACTION)
4933 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4934 		    MAX_SINGLE_PACKET_FRACTION);
4935 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4936 	    MAX_SINGLE_PACKET_FRACTION)
4937 		scctx->isc_tx_tso_segments_max = max(1,
4938 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4939 
4940 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4941 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4942 		/*
4943 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4944 		 * but some MACs do.
4945 		 */
4946 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4947 		    IP_MAXPACKET));
4948 		/*
4949 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4950 		 * into account.  In the worst case, each of these calls will
4951 		 * add another mbuf and, thus, the requirement for another DMA
4952 		 * segment.  So for best performance, it doesn't make sense to
4953 		 * advertize a maximum of TSO segments that typically will
4954 		 * require defragmentation in iflib_encap().
4955 		 */
4956 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4957 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4958 	}
4959 	if (scctx->isc_rss_table_size == 0)
4960 		scctx->isc_rss_table_size = 64;
4961 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4962 
4963 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4964 	/* XXX format name */
4965 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4966 	    NULL, NULL, "admin");
4967 
4968 	/* XXX --- can support > 1 -- but keep it simple for now */
4969 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4970 
4971 	/* Get memory for the station queues */
4972 	if ((err = iflib_queues_alloc(ctx))) {
4973 		device_printf(dev, "Unable to allocate queue memory\n");
4974 		goto fail_iflib_detach;
4975 	}
4976 
4977 	if ((err = iflib_qset_structures_setup(ctx))) {
4978 		device_printf(dev, "qset structure setup failed %d\n", err);
4979 		goto fail_queues;
4980 	}
4981 
4982 	/*
4983 	 * XXX What if anything do we want to do about interrupts?
4984 	 */
4985 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4986 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4987 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4988 		goto fail_detach;
4989 	}
4990 
4991 	/*
4992 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4993 	 * This must appear after the call to ether_ifattach() because
4994 	 * ether_ifattach() sets if_hdrlen to the default value.
4995 	 */
4996 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4997 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4998 
4999 	/* XXX handle more than one queue */
5000 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5001 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5002 
5003 	*ctxp = ctx;
5004 
5005 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5006 	iflib_add_device_sysctl_post(ctx);
5007 	ctx->ifc_flags |= IFC_INIT_DONE;
5008 	CTX_UNLOCK(ctx);
5009 
5010 	return (0);
5011 fail_detach:
5012 	ether_ifdetach(ctx->ifc_ifp);
5013 fail_queues:
5014 	iflib_tx_structures_free(ctx);
5015 	iflib_rx_structures_free(ctx);
5016 fail_iflib_detach:
5017 	IFDI_DETACH(ctx);
5018 fail_unlock:
5019 	CTX_UNLOCK(ctx);
5020 	iflib_deregister(ctx);
5021 fail_ctx_free:
5022 	free(ctx->ifc_softc, M_IFLIB);
5023 	free(ctx, M_IFLIB);
5024 	return (err);
5025 }
5026 
5027 int
5028 iflib_pseudo_deregister(if_ctx_t ctx)
5029 {
5030 	if_t ifp = ctx->ifc_ifp;
5031 	iflib_txq_t txq;
5032 	iflib_rxq_t rxq;
5033 	int i, j;
5034 	struct taskqgroup *tqg;
5035 	iflib_fl_t fl;
5036 
5037 	/* Unregister VLAN event handlers early */
5038 	iflib_unregister_vlan_handlers(ctx);
5039 
5040 	ether_ifdetach(ifp);
5041 	/* XXX drain any dependent tasks */
5042 	tqg = qgroup_if_io_tqg;
5043 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5044 		callout_drain(&txq->ift_timer);
5045 		if (txq->ift_task.gt_uniq != NULL)
5046 			taskqgroup_detach(tqg, &txq->ift_task);
5047 	}
5048 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5049 		callout_drain(&rxq->ifr_watchdog);
5050 		if (rxq->ifr_task.gt_uniq != NULL)
5051 			taskqgroup_detach(tqg, &rxq->ifr_task);
5052 
5053 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5054 			free(fl->ifl_rx_bitmap, M_IFLIB);
5055 	}
5056 	tqg = qgroup_if_config_tqg;
5057 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5058 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5059 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5060 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5061 
5062 	iflib_tx_structures_free(ctx);
5063 	iflib_rx_structures_free(ctx);
5064 
5065 	iflib_deregister(ctx);
5066 
5067 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5068 		free(ctx->ifc_softc, M_IFLIB);
5069 	free(ctx, M_IFLIB);
5070 	return (0);
5071 }
5072 
5073 int
5074 iflib_device_attach(device_t dev)
5075 {
5076 	if_ctx_t ctx;
5077 	if_shared_ctx_t sctx;
5078 
5079 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5080 		return (ENOTSUP);
5081 
5082 	pci_enable_busmaster(dev);
5083 
5084 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5085 }
5086 
5087 int
5088 iflib_device_deregister(if_ctx_t ctx)
5089 {
5090 	if_t ifp = ctx->ifc_ifp;
5091 	iflib_txq_t txq;
5092 	iflib_rxq_t rxq;
5093 	device_t dev = ctx->ifc_dev;
5094 	int i, j;
5095 	struct taskqgroup *tqg;
5096 	iflib_fl_t fl;
5097 
5098 	/* Make sure VLANS are not using driver */
5099 	if (if_vlantrunkinuse(ifp)) {
5100 		device_printf(dev, "Vlan in use, detach first\n");
5101 		return (EBUSY);
5102 	}
5103 #ifdef PCI_IOV
5104 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5105 		device_printf(dev, "SR-IOV in use; detach first.\n");
5106 		return (EBUSY);
5107 	}
5108 #endif
5109 
5110 	STATE_LOCK(ctx);
5111 	ctx->ifc_flags |= IFC_IN_DETACH;
5112 	STATE_UNLOCK(ctx);
5113 
5114 	/* Unregister VLAN handlers before calling iflib_stop() */
5115 	iflib_unregister_vlan_handlers(ctx);
5116 
5117 	iflib_netmap_detach(ifp);
5118 	ether_ifdetach(ifp);
5119 
5120 	CTX_LOCK(ctx);
5121 	iflib_stop(ctx);
5122 	CTX_UNLOCK(ctx);
5123 
5124 	iflib_rem_pfil(ctx);
5125 	if (ctx->ifc_led_dev != NULL)
5126 		led_destroy(ctx->ifc_led_dev);
5127 	/* XXX drain any dependent tasks */
5128 	tqg = qgroup_if_io_tqg;
5129 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5130 		callout_drain(&txq->ift_timer);
5131 		if (txq->ift_task.gt_uniq != NULL)
5132 			taskqgroup_detach(tqg, &txq->ift_task);
5133 	}
5134 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5135 		if (rxq->ifr_task.gt_uniq != NULL)
5136 			taskqgroup_detach(tqg, &rxq->ifr_task);
5137 
5138 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5139 			free(fl->ifl_rx_bitmap, M_IFLIB);
5140 	}
5141 	tqg = qgroup_if_config_tqg;
5142 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5143 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5144 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5145 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5146 	CTX_LOCK(ctx);
5147 	IFDI_DETACH(ctx);
5148 	CTX_UNLOCK(ctx);
5149 
5150 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5151 	iflib_free_intr_mem(ctx);
5152 
5153 	bus_generic_detach(dev);
5154 
5155 	iflib_tx_structures_free(ctx);
5156 	iflib_rx_structures_free(ctx);
5157 
5158 	iflib_deregister(ctx);
5159 
5160 	device_set_softc(ctx->ifc_dev, NULL);
5161 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5162 		free(ctx->ifc_softc, M_IFLIB);
5163 	unref_ctx_core_offset(ctx);
5164 	free(ctx, M_IFLIB);
5165 	return (0);
5166 }
5167 
5168 static void
5169 iflib_free_intr_mem(if_ctx_t ctx)
5170 {
5171 
5172 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5173 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5174 	}
5175 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5176 		pci_release_msi(ctx->ifc_dev);
5177 	}
5178 	if (ctx->ifc_msix_mem != NULL) {
5179 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5180 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5181 		ctx->ifc_msix_mem = NULL;
5182 	}
5183 }
5184 
5185 int
5186 iflib_device_detach(device_t dev)
5187 {
5188 	if_ctx_t ctx = device_get_softc(dev);
5189 
5190 	return (iflib_device_deregister(ctx));
5191 }
5192 
5193 int
5194 iflib_device_suspend(device_t dev)
5195 {
5196 	if_ctx_t ctx = device_get_softc(dev);
5197 
5198 	CTX_LOCK(ctx);
5199 	IFDI_SUSPEND(ctx);
5200 	CTX_UNLOCK(ctx);
5201 
5202 	return bus_generic_suspend(dev);
5203 }
5204 int
5205 iflib_device_shutdown(device_t dev)
5206 {
5207 	if_ctx_t ctx = device_get_softc(dev);
5208 
5209 	CTX_LOCK(ctx);
5210 	IFDI_SHUTDOWN(ctx);
5211 	CTX_UNLOCK(ctx);
5212 
5213 	return bus_generic_suspend(dev);
5214 }
5215 
5216 
5217 int
5218 iflib_device_resume(device_t dev)
5219 {
5220 	if_ctx_t ctx = device_get_softc(dev);
5221 	iflib_txq_t txq = ctx->ifc_txqs;
5222 
5223 	CTX_LOCK(ctx);
5224 	IFDI_RESUME(ctx);
5225 	iflib_if_init_locked(ctx);
5226 	CTX_UNLOCK(ctx);
5227 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5228 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5229 
5230 	return (bus_generic_resume(dev));
5231 }
5232 
5233 int
5234 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5235 {
5236 	int error;
5237 	if_ctx_t ctx = device_get_softc(dev);
5238 
5239 	CTX_LOCK(ctx);
5240 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5241 	CTX_UNLOCK(ctx);
5242 
5243 	return (error);
5244 }
5245 
5246 void
5247 iflib_device_iov_uninit(device_t dev)
5248 {
5249 	if_ctx_t ctx = device_get_softc(dev);
5250 
5251 	CTX_LOCK(ctx);
5252 	IFDI_IOV_UNINIT(ctx);
5253 	CTX_UNLOCK(ctx);
5254 }
5255 
5256 int
5257 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5258 {
5259 	int error;
5260 	if_ctx_t ctx = device_get_softc(dev);
5261 
5262 	CTX_LOCK(ctx);
5263 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5264 	CTX_UNLOCK(ctx);
5265 
5266 	return (error);
5267 }
5268 
5269 /*********************************************************************
5270  *
5271  *  MODULE FUNCTION DEFINITIONS
5272  *
5273  **********************************************************************/
5274 
5275 /*
5276  * - Start a fast taskqueue thread for each core
5277  * - Start a taskqueue for control operations
5278  */
5279 static int
5280 iflib_module_init(void)
5281 {
5282 	return (0);
5283 }
5284 
5285 static int
5286 iflib_module_event_handler(module_t mod, int what, void *arg)
5287 {
5288 	int err;
5289 
5290 	switch (what) {
5291 	case MOD_LOAD:
5292 		if ((err = iflib_module_init()) != 0)
5293 			return (err);
5294 		break;
5295 	case MOD_UNLOAD:
5296 		return (EBUSY);
5297 	default:
5298 		return (EOPNOTSUPP);
5299 	}
5300 
5301 	return (0);
5302 }
5303 
5304 /*********************************************************************
5305  *
5306  *  PUBLIC FUNCTION DEFINITIONS
5307  *     ordered as in iflib.h
5308  *
5309  **********************************************************************/
5310 
5311 
5312 static void
5313 _iflib_assert(if_shared_ctx_t sctx)
5314 {
5315 	int i;
5316 
5317 	MPASS(sctx->isc_tx_maxsize);
5318 	MPASS(sctx->isc_tx_maxsegsize);
5319 
5320 	MPASS(sctx->isc_rx_maxsize);
5321 	MPASS(sctx->isc_rx_nsegments);
5322 	MPASS(sctx->isc_rx_maxsegsize);
5323 
5324 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5325 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5326 		MPASS(sctx->isc_nrxd_min[i]);
5327 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5328 		MPASS(sctx->isc_nrxd_max[i]);
5329 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5330 		MPASS(sctx->isc_nrxd_default[i]);
5331 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5332 	}
5333 
5334 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5335 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5336 		MPASS(sctx->isc_ntxd_min[i]);
5337 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5338 		MPASS(sctx->isc_ntxd_max[i]);
5339 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5340 		MPASS(sctx->isc_ntxd_default[i]);
5341 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5342 	}
5343 }
5344 
5345 static void
5346 _iflib_pre_assert(if_softc_ctx_t scctx)
5347 {
5348 
5349 	MPASS(scctx->isc_txrx->ift_txd_encap);
5350 	MPASS(scctx->isc_txrx->ift_txd_flush);
5351 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5352 	MPASS(scctx->isc_txrx->ift_rxd_available);
5353 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5354 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5355 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5356 }
5357 
5358 static int
5359 iflib_register(if_ctx_t ctx)
5360 {
5361 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5362 	driver_t *driver = sctx->isc_driver;
5363 	device_t dev = ctx->ifc_dev;
5364 	if_t ifp;
5365 
5366 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5367 		_iflib_assert(sctx);
5368 
5369 	CTX_LOCK_INIT(ctx);
5370 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5371 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5372 	if (ifp == NULL) {
5373 		device_printf(dev, "can not allocate ifnet structure\n");
5374 		return (ENOMEM);
5375 	}
5376 
5377 	/*
5378 	 * Initialize our context's device specific methods
5379 	 */
5380 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5381 	kobj_class_compile((kobj_class_t) driver);
5382 
5383 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5384 	if_setsoftc(ifp, ctx);
5385 	if_setdev(ifp, dev);
5386 	if_setinitfn(ifp, iflib_if_init);
5387 	if_setioctlfn(ifp, iflib_if_ioctl);
5388 #ifdef ALTQ
5389 	if_setstartfn(ifp, iflib_altq_if_start);
5390 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5391 	if_setsendqready(ifp);
5392 #else
5393 	if_settransmitfn(ifp, iflib_if_transmit);
5394 #endif
5395 	if_setqflushfn(ifp, iflib_if_qflush);
5396 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST |
5397 	    IFF_KNOWSEPOCH);
5398 
5399 	ctx->ifc_vlan_attach_event =
5400 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5401 							  EVENTHANDLER_PRI_FIRST);
5402 	ctx->ifc_vlan_detach_event =
5403 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5404 							  EVENTHANDLER_PRI_FIRST);
5405 
5406 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5407 		ctx->ifc_mediap = &ctx->ifc_media;
5408 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5409 		    iflib_media_change, iflib_media_status);
5410 	}
5411 	return (0);
5412 }
5413 
5414 static void
5415 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5416 {
5417 	/* Unregister VLAN events */
5418 	if (ctx->ifc_vlan_attach_event != NULL) {
5419 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5420 		ctx->ifc_vlan_attach_event = NULL;
5421 	}
5422 	if (ctx->ifc_vlan_detach_event != NULL) {
5423 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5424 		ctx->ifc_vlan_detach_event = NULL;
5425 	}
5426 
5427 }
5428 
5429 static void
5430 iflib_deregister(if_ctx_t ctx)
5431 {
5432 	if_t ifp = ctx->ifc_ifp;
5433 
5434 	/* Remove all media */
5435 	ifmedia_removeall(&ctx->ifc_media);
5436 
5437 	/* Ensure that VLAN event handlers are unregistered */
5438 	iflib_unregister_vlan_handlers(ctx);
5439 
5440 	/* Release kobject reference */
5441 	kobj_delete((kobj_t) ctx, NULL);
5442 
5443 	/* Free the ifnet structure */
5444 	if_free(ifp);
5445 
5446 	STATE_LOCK_DESTROY(ctx);
5447 
5448 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5449 	CTX_LOCK_DESTROY(ctx);
5450 }
5451 
5452 static int
5453 iflib_queues_alloc(if_ctx_t ctx)
5454 {
5455 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5456 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5457 	device_t dev = ctx->ifc_dev;
5458 	int nrxqsets = scctx->isc_nrxqsets;
5459 	int ntxqsets = scctx->isc_ntxqsets;
5460 	iflib_txq_t txq;
5461 	iflib_rxq_t rxq;
5462 	iflib_fl_t fl = NULL;
5463 	int i, j, cpu, err, txconf, rxconf;
5464 	iflib_dma_info_t ifdip;
5465 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5466 	uint32_t *txqsizes = scctx->isc_txqsizes;
5467 	uint8_t nrxqs = sctx->isc_nrxqs;
5468 	uint8_t ntxqs = sctx->isc_ntxqs;
5469 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5470 	caddr_t *vaddrs;
5471 	uint64_t *paddrs;
5472 
5473 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5474 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5475 
5476 	/* Allocate the TX ring struct memory */
5477 	if (!(ctx->ifc_txqs =
5478 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5479 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5480 		device_printf(dev, "Unable to allocate TX ring memory\n");
5481 		err = ENOMEM;
5482 		goto fail;
5483 	}
5484 
5485 	/* Now allocate the RX */
5486 	if (!(ctx->ifc_rxqs =
5487 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5488 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5489 		device_printf(dev, "Unable to allocate RX ring memory\n");
5490 		err = ENOMEM;
5491 		goto rx_fail;
5492 	}
5493 
5494 	txq = ctx->ifc_txqs;
5495 	rxq = ctx->ifc_rxqs;
5496 
5497 	/*
5498 	 * XXX handle allocation failure
5499 	 */
5500 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5501 		/* Set up some basics */
5502 
5503 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5504 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5505 			device_printf(dev,
5506 			    "Unable to allocate TX DMA info memory\n");
5507 			err = ENOMEM;
5508 			goto err_tx_desc;
5509 		}
5510 		txq->ift_ifdi = ifdip;
5511 		for (j = 0; j < ntxqs; j++, ifdip++) {
5512 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5513 				device_printf(dev,
5514 				    "Unable to allocate TX descriptors\n");
5515 				err = ENOMEM;
5516 				goto err_tx_desc;
5517 			}
5518 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5519 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5520 		}
5521 		txq->ift_ctx = ctx;
5522 		txq->ift_id = i;
5523 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5524 			txq->ift_br_offset = 1;
5525 		} else {
5526 			txq->ift_br_offset = 0;
5527 		}
5528 		/* XXX fix this */
5529 		txq->ift_timer.c_cpu = cpu;
5530 
5531 		if (iflib_txsd_alloc(txq)) {
5532 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5533 			err = ENOMEM;
5534 			goto err_tx_desc;
5535 		}
5536 
5537 		/* Initialize the TX lock */
5538 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5539 		    device_get_nameunit(dev), txq->ift_id);
5540 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5541 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5542 
5543 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5544 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5545 		if (err) {
5546 			/* XXX free any allocated rings */
5547 			device_printf(dev, "Unable to allocate buf_ring\n");
5548 			goto err_tx_desc;
5549 		}
5550 	}
5551 
5552 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5553 		/* Set up some basics */
5554 		callout_init(&rxq->ifr_watchdog, 1);
5555 
5556 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5557 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5558 			device_printf(dev,
5559 			    "Unable to allocate RX DMA info memory\n");
5560 			err = ENOMEM;
5561 			goto err_tx_desc;
5562 		}
5563 
5564 		rxq->ifr_ifdi = ifdip;
5565 		/* XXX this needs to be changed if #rx queues != #tx queues */
5566 		rxq->ifr_ntxqirq = 1;
5567 		rxq->ifr_txqid[0] = i;
5568 		for (j = 0; j < nrxqs; j++, ifdip++) {
5569 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5570 				device_printf(dev,
5571 				    "Unable to allocate RX descriptors\n");
5572 				err = ENOMEM;
5573 				goto err_tx_desc;
5574 			}
5575 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5576 		}
5577 		rxq->ifr_ctx = ctx;
5578 		rxq->ifr_id = i;
5579 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5580 			rxq->ifr_fl_offset = 1;
5581 		} else {
5582 			rxq->ifr_fl_offset = 0;
5583 		}
5584 		rxq->ifr_nfl = nfree_lists;
5585 		if (!(fl =
5586 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5587 			device_printf(dev, "Unable to allocate free list memory\n");
5588 			err = ENOMEM;
5589 			goto err_tx_desc;
5590 		}
5591 		rxq->ifr_fl = fl;
5592 		for (j = 0; j < nfree_lists; j++) {
5593 			fl[j].ifl_rxq = rxq;
5594 			fl[j].ifl_id = j;
5595 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5596 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5597 		}
5598 		/* Allocate receive buffers for the ring */
5599 		if (iflib_rxsd_alloc(rxq)) {
5600 			device_printf(dev,
5601 			    "Critical Failure setting up receive buffers\n");
5602 			err = ENOMEM;
5603 			goto err_rx_desc;
5604 		}
5605 
5606 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5607 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5608 			    M_WAITOK);
5609 	}
5610 
5611 	/* TXQs */
5612 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5613 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5614 	for (i = 0; i < ntxqsets; i++) {
5615 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5616 
5617 		for (j = 0; j < ntxqs; j++, di++) {
5618 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5619 			paddrs[i*ntxqs + j] = di->idi_paddr;
5620 		}
5621 	}
5622 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5623 		device_printf(ctx->ifc_dev,
5624 		    "Unable to allocate device TX queue\n");
5625 		iflib_tx_structures_free(ctx);
5626 		free(vaddrs, M_IFLIB);
5627 		free(paddrs, M_IFLIB);
5628 		goto err_rx_desc;
5629 	}
5630 	free(vaddrs, M_IFLIB);
5631 	free(paddrs, M_IFLIB);
5632 
5633 	/* RXQs */
5634 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5635 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5636 	for (i = 0; i < nrxqsets; i++) {
5637 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5638 
5639 		for (j = 0; j < nrxqs; j++, di++) {
5640 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5641 			paddrs[i*nrxqs + j] = di->idi_paddr;
5642 		}
5643 	}
5644 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5645 		device_printf(ctx->ifc_dev,
5646 		    "Unable to allocate device RX queue\n");
5647 		iflib_tx_structures_free(ctx);
5648 		free(vaddrs, M_IFLIB);
5649 		free(paddrs, M_IFLIB);
5650 		goto err_rx_desc;
5651 	}
5652 	free(vaddrs, M_IFLIB);
5653 	free(paddrs, M_IFLIB);
5654 
5655 	return (0);
5656 
5657 /* XXX handle allocation failure changes */
5658 err_rx_desc:
5659 err_tx_desc:
5660 rx_fail:
5661 	if (ctx->ifc_rxqs != NULL)
5662 		free(ctx->ifc_rxqs, M_IFLIB);
5663 	ctx->ifc_rxqs = NULL;
5664 	if (ctx->ifc_txqs != NULL)
5665 		free(ctx->ifc_txqs, M_IFLIB);
5666 	ctx->ifc_txqs = NULL;
5667 fail:
5668 	return (err);
5669 }
5670 
5671 static int
5672 iflib_tx_structures_setup(if_ctx_t ctx)
5673 {
5674 	iflib_txq_t txq = ctx->ifc_txqs;
5675 	int i;
5676 
5677 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5678 		iflib_txq_setup(txq);
5679 
5680 	return (0);
5681 }
5682 
5683 static void
5684 iflib_tx_structures_free(if_ctx_t ctx)
5685 {
5686 	iflib_txq_t txq = ctx->ifc_txqs;
5687 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5688 	int i, j;
5689 
5690 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5691 		for (j = 0; j < sctx->isc_ntxqs; j++)
5692 			iflib_dma_free(&txq->ift_ifdi[j]);
5693 		iflib_txq_destroy(txq);
5694 	}
5695 	free(ctx->ifc_txqs, M_IFLIB);
5696 	ctx->ifc_txqs = NULL;
5697 	IFDI_QUEUES_FREE(ctx);
5698 }
5699 
5700 /*********************************************************************
5701  *
5702  *  Initialize all receive rings.
5703  *
5704  **********************************************************************/
5705 static int
5706 iflib_rx_structures_setup(if_ctx_t ctx)
5707 {
5708 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5709 	int q;
5710 #if defined(INET6) || defined(INET)
5711 	int err, i;
5712 #endif
5713 
5714 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5715 #if defined(INET6) || defined(INET)
5716 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5717 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5718 			    TCP_LRO_ENTRIES, min(1024,
5719 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5720 			if (err != 0) {
5721 				device_printf(ctx->ifc_dev,
5722 				    "LRO Initialization failed!\n");
5723 				goto fail;
5724 			}
5725 		}
5726 #endif
5727 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5728 	}
5729 	return (0);
5730 #if defined(INET6) || defined(INET)
5731 fail:
5732 	/*
5733 	 * Free LRO resources allocated so far, we will only handle
5734 	 * the rings that completed, the failing case will have
5735 	 * cleaned up for itself.  'q' failed, so its the terminus.
5736 	 */
5737 	rxq = ctx->ifc_rxqs;
5738 	for (i = 0; i < q; ++i, rxq++) {
5739 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5740 			tcp_lro_free(&rxq->ifr_lc);
5741 	}
5742 	return (err);
5743 #endif
5744 }
5745 
5746 /*********************************************************************
5747  *
5748  *  Free all receive rings.
5749  *
5750  **********************************************************************/
5751 static void
5752 iflib_rx_structures_free(if_ctx_t ctx)
5753 {
5754 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5755 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5756 	int i, j;
5757 
5758 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5759 		for (j = 0; j < sctx->isc_nrxqs; j++)
5760 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5761 		iflib_rx_sds_free(rxq);
5762 #if defined(INET6) || defined(INET)
5763 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5764 			tcp_lro_free(&rxq->ifr_lc);
5765 #endif
5766 	}
5767 	free(ctx->ifc_rxqs, M_IFLIB);
5768 	ctx->ifc_rxqs = NULL;
5769 }
5770 
5771 static int
5772 iflib_qset_structures_setup(if_ctx_t ctx)
5773 {
5774 	int err;
5775 
5776 	/*
5777 	 * It is expected that the caller takes care of freeing queues if this
5778 	 * fails.
5779 	 */
5780 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5781 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5782 		return (err);
5783 	}
5784 
5785 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5786 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5787 
5788 	return (err);
5789 }
5790 
5791 int
5792 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5793 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5794 {
5795 
5796 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5797 }
5798 
5799 #ifdef SMP
5800 static int
5801 find_nth(if_ctx_t ctx, int qid)
5802 {
5803 	cpuset_t cpus;
5804 	int i, cpuid, eqid, count;
5805 
5806 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5807 	count = CPU_COUNT(&cpus);
5808 	eqid = qid % count;
5809 	/* clear up to the qid'th bit */
5810 	for (i = 0; i < eqid; i++) {
5811 		cpuid = CPU_FFS(&cpus);
5812 		MPASS(cpuid != 0);
5813 		CPU_CLR(cpuid-1, &cpus);
5814 	}
5815 	cpuid = CPU_FFS(&cpus);
5816 	MPASS(cpuid != 0);
5817 	return (cpuid-1);
5818 }
5819 
5820 #ifdef SCHED_ULE
5821 extern struct cpu_group *cpu_top;              /* CPU topology */
5822 
5823 static int
5824 find_child_with_core(int cpu, struct cpu_group *grp)
5825 {
5826 	int i;
5827 
5828 	if (grp->cg_children == 0)
5829 		return -1;
5830 
5831 	MPASS(grp->cg_child);
5832 	for (i = 0; i < grp->cg_children; i++) {
5833 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5834 			return i;
5835 	}
5836 
5837 	return -1;
5838 }
5839 
5840 /*
5841  * Find the nth "close" core to the specified core
5842  * "close" is defined as the deepest level that shares
5843  * at least an L2 cache.  With threads, this will be
5844  * threads on the same core.  If the shared cache is L3
5845  * or higher, simply returns the same core.
5846  */
5847 static int
5848 find_close_core(int cpu, int core_offset)
5849 {
5850 	struct cpu_group *grp;
5851 	int i;
5852 	int fcpu;
5853 	cpuset_t cs;
5854 
5855 	grp = cpu_top;
5856 	if (grp == NULL)
5857 		return cpu;
5858 	i = 0;
5859 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5860 		/* If the child only has one cpu, don't descend */
5861 		if (grp->cg_child[i].cg_count <= 1)
5862 			break;
5863 		grp = &grp->cg_child[i];
5864 	}
5865 
5866 	/* If they don't share at least an L2 cache, use the same CPU */
5867 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5868 		return cpu;
5869 
5870 	/* Now pick one */
5871 	CPU_COPY(&grp->cg_mask, &cs);
5872 
5873 	/* Add the selected CPU offset to core offset. */
5874 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5875 		if (fcpu - 1 == cpu)
5876 			break;
5877 		CPU_CLR(fcpu - 1, &cs);
5878 	}
5879 	MPASS(fcpu);
5880 
5881 	core_offset += i;
5882 
5883 	CPU_COPY(&grp->cg_mask, &cs);
5884 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5885 		MPASS(CPU_FFS(&cs));
5886 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5887 	}
5888 	MPASS(CPU_FFS(&cs));
5889 	return CPU_FFS(&cs) - 1;
5890 }
5891 #else
5892 static int
5893 find_close_core(int cpu, int core_offset __unused)
5894 {
5895 	return cpu;
5896 }
5897 #endif
5898 
5899 static int
5900 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5901 {
5902 	switch (type) {
5903 	case IFLIB_INTR_TX:
5904 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5905 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5906 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5907 	case IFLIB_INTR_RX:
5908 	case IFLIB_INTR_RXTX:
5909 		/* RX queues get the specified core */
5910 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5911 	default:
5912 		return -1;
5913 	}
5914 }
5915 #else
5916 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5917 #define find_close_core(cpuid, tid)	CPU_FIRST()
5918 #define find_nth(ctx, gid)		CPU_FIRST()
5919 #endif
5920 
5921 /* Just to avoid copy/paste */
5922 static inline int
5923 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5924     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5925     const char *name)
5926 {
5927 	device_t dev;
5928 	int co, cpuid, err, tid;
5929 
5930 	dev = ctx->ifc_dev;
5931 	co = ctx->ifc_sysctl_core_offset;
5932 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5933 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5934 	cpuid = find_nth(ctx, qid + co);
5935 	tid = get_core_offset(ctx, type, qid);
5936 	if (tid < 0) {
5937 		device_printf(dev, "get_core_offset failed\n");
5938 		return (EOPNOTSUPP);
5939 	}
5940 	cpuid = find_close_core(cpuid, tid);
5941 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5942 	    name);
5943 	if (err) {
5944 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5945 		return (err);
5946 	}
5947 #ifdef notyet
5948 	if (cpuid > ctx->ifc_cpuid_highest)
5949 		ctx->ifc_cpuid_highest = cpuid;
5950 #endif
5951 	return (0);
5952 }
5953 
5954 int
5955 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5956 			iflib_intr_type_t type, driver_filter_t *filter,
5957 			void *filter_arg, int qid, const char *name)
5958 {
5959 	device_t dev;
5960 	struct grouptask *gtask;
5961 	struct taskqgroup *tqg;
5962 	iflib_filter_info_t info;
5963 	gtask_fn_t *fn;
5964 	int tqrid, err;
5965 	driver_filter_t *intr_fast;
5966 	void *q;
5967 
5968 	info = &ctx->ifc_filter_info;
5969 	tqrid = rid;
5970 
5971 	switch (type) {
5972 	/* XXX merge tx/rx for netmap? */
5973 	case IFLIB_INTR_TX:
5974 		q = &ctx->ifc_txqs[qid];
5975 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5976 		gtask = &ctx->ifc_txqs[qid].ift_task;
5977 		tqg = qgroup_if_io_tqg;
5978 		fn = _task_fn_tx;
5979 		intr_fast = iflib_fast_intr;
5980 		GROUPTASK_INIT(gtask, 0, fn, q);
5981 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5982 		break;
5983 	case IFLIB_INTR_RX:
5984 		q = &ctx->ifc_rxqs[qid];
5985 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5986 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5987 		tqg = qgroup_if_io_tqg;
5988 		fn = _task_fn_rx;
5989 		intr_fast = iflib_fast_intr;
5990 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
5991 		break;
5992 	case IFLIB_INTR_RXTX:
5993 		q = &ctx->ifc_rxqs[qid];
5994 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5995 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5996 		tqg = qgroup_if_io_tqg;
5997 		fn = _task_fn_rx;
5998 		intr_fast = iflib_fast_intr_rxtx;
5999 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6000 		break;
6001 	case IFLIB_INTR_ADMIN:
6002 		q = ctx;
6003 		tqrid = -1;
6004 		info = &ctx->ifc_filter_info;
6005 		gtask = &ctx->ifc_admin_task;
6006 		tqg = qgroup_if_config_tqg;
6007 		fn = _task_fn_admin;
6008 		intr_fast = iflib_fast_intr_ctx;
6009 		break;
6010 	default:
6011 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6012 		    __func__);
6013 		return (EINVAL);
6014 	}
6015 
6016 	info->ifi_filter = filter;
6017 	info->ifi_filter_arg = filter_arg;
6018 	info->ifi_task = gtask;
6019 	info->ifi_ctx = q;
6020 
6021 	dev = ctx->ifc_dev;
6022 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6023 	if (err != 0) {
6024 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6025 		return (err);
6026 	}
6027 	if (type == IFLIB_INTR_ADMIN)
6028 		return (0);
6029 
6030 	if (tqrid != -1) {
6031 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6032 		    q, name);
6033 		if (err)
6034 			return (err);
6035 	} else {
6036 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6037 	}
6038 
6039 	return (0);
6040 }
6041 
6042 void
6043 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6044 {
6045 	struct grouptask *gtask;
6046 	struct taskqgroup *tqg;
6047 	gtask_fn_t *fn;
6048 	void *q;
6049 	int err;
6050 
6051 	switch (type) {
6052 	case IFLIB_INTR_TX:
6053 		q = &ctx->ifc_txqs[qid];
6054 		gtask = &ctx->ifc_txqs[qid].ift_task;
6055 		tqg = qgroup_if_io_tqg;
6056 		fn = _task_fn_tx;
6057 		GROUPTASK_INIT(gtask, 0, fn, q);
6058 		break;
6059 	case IFLIB_INTR_RX:
6060 		q = &ctx->ifc_rxqs[qid];
6061 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6062 		tqg = qgroup_if_io_tqg;
6063 		fn = _task_fn_rx;
6064 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6065 		break;
6066 	case IFLIB_INTR_IOV:
6067 		q = ctx;
6068 		gtask = &ctx->ifc_vflr_task;
6069 		tqg = qgroup_if_config_tqg;
6070 		fn = _task_fn_iov;
6071 		GROUPTASK_INIT(gtask, 0, fn, q);
6072 		break;
6073 	default:
6074 		panic("unknown net intr type");
6075 	}
6076 	if (irq != NULL) {
6077 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6078 		    q, name);
6079 		if (err)
6080 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6081 			    irq->ii_res, name);
6082 	} else {
6083 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6084 	}
6085 }
6086 
6087 void
6088 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6089 {
6090 
6091 	if (irq->ii_tag)
6092 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6093 
6094 	if (irq->ii_res)
6095 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6096 		    rman_get_rid(irq->ii_res), irq->ii_res);
6097 }
6098 
6099 static int
6100 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6101 {
6102 	iflib_txq_t txq = ctx->ifc_txqs;
6103 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6104 	if_irq_t irq = &ctx->ifc_legacy_irq;
6105 	iflib_filter_info_t info;
6106 	device_t dev;
6107 	struct grouptask *gtask;
6108 	struct resource *res;
6109 	struct taskqgroup *tqg;
6110 	void *q;
6111 	int err, tqrid;
6112 	bool rx_only;
6113 
6114 	q = &ctx->ifc_rxqs[0];
6115 	info = &rxq[0].ifr_filter_info;
6116 	gtask = &rxq[0].ifr_task;
6117 	tqg = qgroup_if_io_tqg;
6118 	tqrid = *rid;
6119 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6120 
6121 	ctx->ifc_flags |= IFC_LEGACY;
6122 	info->ifi_filter = filter;
6123 	info->ifi_filter_arg = filter_arg;
6124 	info->ifi_task = gtask;
6125 	info->ifi_ctx = rx_only ? ctx : q;
6126 
6127 	dev = ctx->ifc_dev;
6128 	/* We allocate a single interrupt resource */
6129 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6130 	    iflib_fast_intr_rxtx, NULL, info, name);
6131 	if (err != 0)
6132 		return (err);
6133 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6134 	res = irq->ii_res;
6135 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6136 
6137 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6138 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6139 	    "tx");
6140 	return (0);
6141 }
6142 
6143 void
6144 iflib_led_create(if_ctx_t ctx)
6145 {
6146 
6147 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6148 	    device_get_nameunit(ctx->ifc_dev));
6149 }
6150 
6151 void
6152 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6153 {
6154 
6155 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6156 }
6157 
6158 void
6159 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6160 {
6161 
6162 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6163 }
6164 
6165 void
6166 iflib_admin_intr_deferred(if_ctx_t ctx)
6167 {
6168 
6169 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6170 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6171 }
6172 
6173 void
6174 iflib_iov_intr_deferred(if_ctx_t ctx)
6175 {
6176 
6177 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6178 }
6179 
6180 void
6181 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6182 {
6183 
6184 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6185 	    name);
6186 }
6187 
6188 void
6189 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6190 	const char *name)
6191 {
6192 
6193 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6194 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6195 	    name);
6196 }
6197 
6198 void
6199 iflib_config_gtask_deinit(struct grouptask *gtask)
6200 {
6201 
6202 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6203 }
6204 
6205 void
6206 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6207 {
6208 	if_t ifp = ctx->ifc_ifp;
6209 	iflib_txq_t txq = ctx->ifc_txqs;
6210 
6211 	if_setbaudrate(ifp, baudrate);
6212 	if (baudrate >= IF_Gbps(10)) {
6213 		STATE_LOCK(ctx);
6214 		ctx->ifc_flags |= IFC_PREFETCH;
6215 		STATE_UNLOCK(ctx);
6216 	}
6217 	/* If link down, disable watchdog */
6218 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6219 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6220 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6221 	}
6222 	ctx->ifc_link_state = link_state;
6223 	if_link_state_change(ifp, link_state);
6224 }
6225 
6226 static int
6227 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6228 {
6229 	int credits;
6230 #ifdef INVARIANTS
6231 	int credits_pre = txq->ift_cidx_processed;
6232 #endif
6233 
6234 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6235 	    BUS_DMASYNC_POSTREAD);
6236 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6237 		return (0);
6238 
6239 	txq->ift_processed += credits;
6240 	txq->ift_cidx_processed += credits;
6241 
6242 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6243 	if (txq->ift_cidx_processed >= txq->ift_size)
6244 		txq->ift_cidx_processed -= txq->ift_size;
6245 	return (credits);
6246 }
6247 
6248 static int
6249 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6250 {
6251 	iflib_fl_t fl;
6252 	u_int i;
6253 
6254 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6255 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6256 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6257 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6258 	    budget));
6259 }
6260 
6261 void
6262 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6263 	const char *description, if_int_delay_info_t info,
6264 	int offset, int value)
6265 {
6266 	info->iidi_ctx = ctx;
6267 	info->iidi_offset = offset;
6268 	info->iidi_value = value;
6269 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6270 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6271 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6272 	    info, 0, iflib_sysctl_int_delay, "I", description);
6273 }
6274 
6275 struct sx *
6276 iflib_ctx_lock_get(if_ctx_t ctx)
6277 {
6278 
6279 	return (&ctx->ifc_ctx_sx);
6280 }
6281 
6282 static int
6283 iflib_msix_init(if_ctx_t ctx)
6284 {
6285 	device_t dev = ctx->ifc_dev;
6286 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6287 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6288 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6289 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6290 
6291 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6292 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6293 
6294 	if (bootverbose)
6295 		device_printf(dev, "msix_init qsets capped at %d\n",
6296 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6297 
6298 	/* Override by tuneable */
6299 	if (scctx->isc_disable_msix)
6300 		goto msi;
6301 
6302 	/* First try MSI-X */
6303 	if ((msgs = pci_msix_count(dev)) == 0) {
6304 		if (bootverbose)
6305 			device_printf(dev, "MSI-X not supported or disabled\n");
6306 		goto msi;
6307 	}
6308 
6309 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6310 	/*
6311 	 * bar == -1 => "trust me I know what I'm doing"
6312 	 * Some drivers are for hardware that is so shoddily
6313 	 * documented that no one knows which bars are which
6314 	 * so the developer has to map all bars. This hack
6315 	 * allows shoddy garbage to use MSI-X in this framework.
6316 	 */
6317 	if (bar != -1) {
6318 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6319 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6320 		if (ctx->ifc_msix_mem == NULL) {
6321 			device_printf(dev, "Unable to map MSI-X table\n");
6322 			goto msi;
6323 		}
6324 	}
6325 
6326 	admincnt = sctx->isc_admin_intrcnt;
6327 #if IFLIB_DEBUG
6328 	/* use only 1 qset in debug mode */
6329 	queuemsgs = min(msgs - admincnt, 1);
6330 #else
6331 	queuemsgs = msgs - admincnt;
6332 #endif
6333 #ifdef RSS
6334 	queues = imin(queuemsgs, rss_getnumbuckets());
6335 #else
6336 	queues = queuemsgs;
6337 #endif
6338 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6339 	if (bootverbose)
6340 		device_printf(dev,
6341 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6342 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6343 #ifdef  RSS
6344 	/* If we're doing RSS, clamp at the number of RSS buckets */
6345 	if (queues > rss_getnumbuckets())
6346 		queues = rss_getnumbuckets();
6347 #endif
6348 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6349 		rx_queues = iflib_num_rx_queues;
6350 	else
6351 		rx_queues = queues;
6352 
6353 	if (rx_queues > scctx->isc_nrxqsets)
6354 		rx_queues = scctx->isc_nrxqsets;
6355 
6356 	/*
6357 	 * We want this to be all logical CPUs by default
6358 	 */
6359 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6360 		tx_queues = iflib_num_tx_queues;
6361 	else
6362 		tx_queues = mp_ncpus;
6363 
6364 	if (tx_queues > scctx->isc_ntxqsets)
6365 		tx_queues = scctx->isc_ntxqsets;
6366 
6367 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6368 #ifdef INVARIANTS
6369 		if (tx_queues != rx_queues)
6370 			device_printf(dev,
6371 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6372 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6373 #endif
6374 		tx_queues = min(rx_queues, tx_queues);
6375 		rx_queues = min(rx_queues, tx_queues);
6376 	}
6377 
6378 	vectors = rx_queues + admincnt;
6379 	if (msgs < vectors) {
6380 		device_printf(dev,
6381 		    "insufficient number of MSI-X vectors "
6382 		    "(supported %d, need %d)\n", msgs, vectors);
6383 		goto msi;
6384 	}
6385 
6386 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6387 	    tx_queues);
6388 	msgs = vectors;
6389 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6390 		if (vectors != msgs) {
6391 			device_printf(dev,
6392 			    "Unable to allocate sufficient MSI-X vectors "
6393 			    "(got %d, need %d)\n", vectors, msgs);
6394 			pci_release_msi(dev);
6395 			if (bar != -1) {
6396 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6397 				    ctx->ifc_msix_mem);
6398 				ctx->ifc_msix_mem = NULL;
6399 			}
6400 			goto msi;
6401 		}
6402 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6403 		    vectors);
6404 		scctx->isc_vectors = vectors;
6405 		scctx->isc_nrxqsets = rx_queues;
6406 		scctx->isc_ntxqsets = tx_queues;
6407 		scctx->isc_intr = IFLIB_INTR_MSIX;
6408 
6409 		return (vectors);
6410 	} else {
6411 		device_printf(dev,
6412 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6413 		    err);
6414 		if (bar != -1) {
6415 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6416 			    ctx->ifc_msix_mem);
6417 			ctx->ifc_msix_mem = NULL;
6418 		}
6419 	}
6420 
6421 msi:
6422 	vectors = pci_msi_count(dev);
6423 	scctx->isc_nrxqsets = 1;
6424 	scctx->isc_ntxqsets = 1;
6425 	scctx->isc_vectors = vectors;
6426 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6427 		device_printf(dev,"Using an MSI interrupt\n");
6428 		scctx->isc_intr = IFLIB_INTR_MSI;
6429 	} else {
6430 		scctx->isc_vectors = 1;
6431 		device_printf(dev,"Using a Legacy interrupt\n");
6432 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6433 	}
6434 
6435 	return (vectors);
6436 }
6437 
6438 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6439 
6440 static int
6441 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6442 {
6443 	int rc;
6444 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6445 	struct sbuf *sb;
6446 	const char *ring_state = "UNKNOWN";
6447 
6448 	/* XXX needed ? */
6449 	rc = sysctl_wire_old_buffer(req, 0);
6450 	MPASS(rc == 0);
6451 	if (rc != 0)
6452 		return (rc);
6453 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6454 	MPASS(sb != NULL);
6455 	if (sb == NULL)
6456 		return (ENOMEM);
6457 	if (state[3] <= 3)
6458 		ring_state = ring_states[state[3]];
6459 
6460 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6461 		    state[0], state[1], state[2], ring_state);
6462 	rc = sbuf_finish(sb);
6463 	sbuf_delete(sb);
6464         return(rc);
6465 }
6466 
6467 enum iflib_ndesc_handler {
6468 	IFLIB_NTXD_HANDLER,
6469 	IFLIB_NRXD_HANDLER,
6470 };
6471 
6472 static int
6473 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6474 {
6475 	if_ctx_t ctx = (void *)arg1;
6476 	enum iflib_ndesc_handler type = arg2;
6477 	char buf[256] = {0};
6478 	qidx_t *ndesc;
6479 	char *p, *next;
6480 	int nqs, rc, i;
6481 
6482 	nqs = 8;
6483 	switch(type) {
6484 	case IFLIB_NTXD_HANDLER:
6485 		ndesc = ctx->ifc_sysctl_ntxds;
6486 		if (ctx->ifc_sctx)
6487 			nqs = ctx->ifc_sctx->isc_ntxqs;
6488 		break;
6489 	case IFLIB_NRXD_HANDLER:
6490 		ndesc = ctx->ifc_sysctl_nrxds;
6491 		if (ctx->ifc_sctx)
6492 			nqs = ctx->ifc_sctx->isc_nrxqs;
6493 		break;
6494 	default:
6495 		printf("%s: unhandled type\n", __func__);
6496 		return (EINVAL);
6497 	}
6498 	if (nqs == 0)
6499 		nqs = 8;
6500 
6501 	for (i=0; i<8; i++) {
6502 		if (i >= nqs)
6503 			break;
6504 		if (i)
6505 			strcat(buf, ",");
6506 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6507 	}
6508 
6509 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6510 	if (rc || req->newptr == NULL)
6511 		return rc;
6512 
6513 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6514 	    i++, p = strsep(&next, " ,")) {
6515 		ndesc[i] = strtoul(p, NULL, 10);
6516 	}
6517 
6518 	return(rc);
6519 }
6520 
6521 #define NAME_BUFLEN 32
6522 static void
6523 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6524 {
6525         device_t dev = iflib_get_dev(ctx);
6526 	struct sysctl_oid_list *child, *oid_list;
6527 	struct sysctl_ctx_list *ctx_list;
6528 	struct sysctl_oid *node;
6529 
6530 	ctx_list = device_get_sysctl_ctx(dev);
6531 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6532 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6533 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6534 	oid_list = SYSCTL_CHILDREN(node);
6535 
6536 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6537 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6538 		       "driver version");
6539 
6540 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6541 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6542 			"# of txqs to use, 0 => use default #");
6543 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6544 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6545 			"# of rxqs to use, 0 => use default #");
6546 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6547 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6548                        "permit #txq != #rxq");
6549 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6550                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6551                       "disable MSI-X (default 0)");
6552 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6553 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6554 		       "set the RX budget");
6555 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6556 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6557 		       "cause TX to abdicate instead of running to completion");
6558 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6559 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6560 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6561 		       "offset to start using cores at");
6562 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6563 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6564 		       "use separate cores for TX and RX");
6565 
6566 	/* XXX change for per-queue sizes */
6567 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6568 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6569 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6570 	    "list of # of TX descriptors to use, 0 = use default #");
6571 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6572 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6573 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6574 	    "list of # of RX descriptors to use, 0 = use default #");
6575 }
6576 
6577 static void
6578 iflib_add_device_sysctl_post(if_ctx_t ctx)
6579 {
6580 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6581 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6582         device_t dev = iflib_get_dev(ctx);
6583 	struct sysctl_oid_list *child;
6584 	struct sysctl_ctx_list *ctx_list;
6585 	iflib_fl_t fl;
6586 	iflib_txq_t txq;
6587 	iflib_rxq_t rxq;
6588 	int i, j;
6589 	char namebuf[NAME_BUFLEN];
6590 	char *qfmt;
6591 	struct sysctl_oid *queue_node, *fl_node, *node;
6592 	struct sysctl_oid_list *queue_list, *fl_list;
6593 	ctx_list = device_get_sysctl_ctx(dev);
6594 
6595 	node = ctx->ifc_sysctl_node;
6596 	child = SYSCTL_CHILDREN(node);
6597 
6598 	if (scctx->isc_ntxqsets > 100)
6599 		qfmt = "txq%03d";
6600 	else if (scctx->isc_ntxqsets > 10)
6601 		qfmt = "txq%02d";
6602 	else
6603 		qfmt = "txq%d";
6604 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6605 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6606 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6607 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6608 		queue_list = SYSCTL_CHILDREN(queue_node);
6609 #if MEMORY_LOGGING
6610 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6611 				CTLFLAG_RD,
6612 				&txq->ift_dequeued, "total mbufs freed");
6613 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6614 				CTLFLAG_RD,
6615 				&txq->ift_enqueued, "total mbufs enqueued");
6616 #endif
6617 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6618 				   CTLFLAG_RD,
6619 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6620 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6621 				   CTLFLAG_RD,
6622 				   &txq->ift_pullups, "# of times m_pullup was called");
6623 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6624 				   CTLFLAG_RD,
6625 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6626 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6627 				   CTLFLAG_RD,
6628 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6629 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6630 				   CTLFLAG_RD,
6631 				   &txq->ift_map_failed, "# of times DMA map failed");
6632 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6633 				   CTLFLAG_RD,
6634 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6635 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6636 				   CTLFLAG_RD,
6637 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6638 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6639 				   CTLFLAG_RD,
6640 				   &txq->ift_pidx, 1, "Producer Index");
6641 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6642 				   CTLFLAG_RD,
6643 				   &txq->ift_cidx, 1, "Consumer Index");
6644 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6645 				   CTLFLAG_RD,
6646 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6647 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6648 				   CTLFLAG_RD,
6649 				   &txq->ift_in_use, 1, "descriptors in use");
6650 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6651 				   CTLFLAG_RD,
6652 				   &txq->ift_processed, "descriptors procesed for clean");
6653 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6654 				   CTLFLAG_RD,
6655 				   &txq->ift_cleaned, "total cleaned");
6656 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6657 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6658 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6659 		    mp_ring_state_handler, "A", "soft ring state");
6660 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6661 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6662 				       "# of enqueues to the mp_ring for this queue");
6663 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6664 				       CTLFLAG_RD, &txq->ift_br->drops,
6665 				       "# of drops in the mp_ring for this queue");
6666 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6667 				       CTLFLAG_RD, &txq->ift_br->starts,
6668 				       "# of normal consumer starts in the mp_ring for this queue");
6669 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6670 				       CTLFLAG_RD, &txq->ift_br->stalls,
6671 					       "# of consumer stalls in the mp_ring for this queue");
6672 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6673 			       CTLFLAG_RD, &txq->ift_br->restarts,
6674 				       "# of consumer restarts in the mp_ring for this queue");
6675 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6676 				       CTLFLAG_RD, &txq->ift_br->abdications,
6677 				       "# of consumer abdications in the mp_ring for this queue");
6678 	}
6679 
6680 	if (scctx->isc_nrxqsets > 100)
6681 		qfmt = "rxq%03d";
6682 	else if (scctx->isc_nrxqsets > 10)
6683 		qfmt = "rxq%02d";
6684 	else
6685 		qfmt = "rxq%d";
6686 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6687 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6688 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6689 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6690 		queue_list = SYSCTL_CHILDREN(queue_node);
6691 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6692 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6693 				       CTLFLAG_RD,
6694 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6695 		}
6696 
6697 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6698 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6699 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6700 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6701 			fl_list = SYSCTL_CHILDREN(fl_node);
6702 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6703 				       CTLFLAG_RD,
6704 				       &fl->ifl_pidx, 1, "Producer Index");
6705 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6706 				       CTLFLAG_RD,
6707 				       &fl->ifl_cidx, 1, "Consumer Index");
6708 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6709 				       CTLFLAG_RD,
6710 				       &fl->ifl_credits, 1, "credits available");
6711 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6712 				       CTLFLAG_RD,
6713 				       &fl->ifl_buf_size, 1, "buffer size");
6714 #if MEMORY_LOGGING
6715 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6716 					CTLFLAG_RD,
6717 					&fl->ifl_m_enqueued, "mbufs allocated");
6718 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6719 					CTLFLAG_RD,
6720 					&fl->ifl_m_dequeued, "mbufs freed");
6721 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6722 					CTLFLAG_RD,
6723 					&fl->ifl_cl_enqueued, "clusters allocated");
6724 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6725 					CTLFLAG_RD,
6726 					&fl->ifl_cl_dequeued, "clusters freed");
6727 #endif
6728 
6729 		}
6730 	}
6731 
6732 }
6733 
6734 void
6735 iflib_request_reset(if_ctx_t ctx)
6736 {
6737 
6738 	STATE_LOCK(ctx);
6739 	ctx->ifc_flags |= IFC_DO_RESET;
6740 	STATE_UNLOCK(ctx);
6741 }
6742 
6743 #ifndef __NO_STRICT_ALIGNMENT
6744 static struct mbuf *
6745 iflib_fixup_rx(struct mbuf *m)
6746 {
6747 	struct mbuf *n;
6748 
6749 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6750 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6751 		m->m_data += ETHER_HDR_LEN;
6752 		n = m;
6753 	} else {
6754 		MGETHDR(n, M_NOWAIT, MT_DATA);
6755 		if (n == NULL) {
6756 			m_freem(m);
6757 			return (NULL);
6758 		}
6759 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6760 		m->m_data += ETHER_HDR_LEN;
6761 		m->m_len -= ETHER_HDR_LEN;
6762 		n->m_len = ETHER_HDR_LEN;
6763 		M_MOVE_PKTHDR(n, m);
6764 		n->m_next = m;
6765 	}
6766 	return (n);
6767 }
6768 #endif
6769 
6770 #ifdef DEBUGNET
6771 static void
6772 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6773 {
6774 	if_ctx_t ctx;
6775 
6776 	ctx = if_getsoftc(ifp);
6777 	CTX_LOCK(ctx);
6778 	*nrxr = NRXQSETS(ctx);
6779 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6780 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6781 	CTX_UNLOCK(ctx);
6782 }
6783 
6784 static void
6785 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6786 {
6787 	if_ctx_t ctx;
6788 	if_softc_ctx_t scctx;
6789 	iflib_fl_t fl;
6790 	iflib_rxq_t rxq;
6791 	int i, j;
6792 
6793 	ctx = if_getsoftc(ifp);
6794 	scctx = &ctx->ifc_softc_ctx;
6795 
6796 	switch (event) {
6797 	case DEBUGNET_START:
6798 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6799 			rxq = &ctx->ifc_rxqs[i];
6800 			for (j = 0; j < rxq->ifr_nfl; j++) {
6801 				fl = rxq->ifr_fl;
6802 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6803 			}
6804 		}
6805 		iflib_no_tx_batch = 1;
6806 		break;
6807 	default:
6808 		break;
6809 	}
6810 }
6811 
6812 static int
6813 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6814 {
6815 	if_ctx_t ctx;
6816 	iflib_txq_t txq;
6817 	int error;
6818 
6819 	ctx = if_getsoftc(ifp);
6820 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6821 	    IFF_DRV_RUNNING)
6822 		return (EBUSY);
6823 
6824 	txq = &ctx->ifc_txqs[0];
6825 	error = iflib_encap(txq, &m);
6826 	if (error == 0)
6827 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6828 	return (error);
6829 }
6830 
6831 static int
6832 iflib_debugnet_poll(if_t ifp, int count)
6833 {
6834 	struct epoch_tracker et;
6835 	if_ctx_t ctx;
6836 	if_softc_ctx_t scctx;
6837 	iflib_txq_t txq;
6838 	int i;
6839 
6840 	ctx = if_getsoftc(ifp);
6841 	scctx = &ctx->ifc_softc_ctx;
6842 
6843 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6844 	    IFF_DRV_RUNNING)
6845 		return (EBUSY);
6846 
6847 	txq = &ctx->ifc_txqs[0];
6848 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6849 
6850 	NET_EPOCH_ENTER(et);
6851 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6852 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6853 	NET_EPOCH_EXIT(et);
6854 	return (0);
6855 }
6856 #endif /* DEBUGNET */
6857