xref: /freebsd/sys/net/iflib.c (revision 8881d206f4e68b564c2c5f50fc717086fc3e827a)
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  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
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 	uint8_t  ifc_sysctl_use_logical_cores;
199 	bool	 ifc_cpus_are_physical_cores;
200 
201 	qidx_t ifc_sysctl_ntxds[8];
202 	qidx_t ifc_sysctl_nrxds[8];
203 	struct if_txrx ifc_txrx;
204 #define isc_txd_encap  ifc_txrx.ift_txd_encap
205 #define isc_txd_flush  ifc_txrx.ift_txd_flush
206 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
207 #define isc_rxd_available ifc_txrx.ift_rxd_available
208 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 #define isc_txq_select ifc_txrx.ift_txq_select
213 	eventhandler_tag ifc_vlan_attach_event;
214 	eventhandler_tag ifc_vlan_detach_event;
215 	struct ether_addr ifc_mac;
216 };
217 
218 void *
219 iflib_get_softc(if_ctx_t ctx)
220 {
221 
222 	return (ctx->ifc_softc);
223 }
224 
225 device_t
226 iflib_get_dev(if_ctx_t ctx)
227 {
228 
229 	return (ctx->ifc_dev);
230 }
231 
232 if_t
233 iflib_get_ifp(if_ctx_t ctx)
234 {
235 
236 	return (ctx->ifc_ifp);
237 }
238 
239 struct ifmedia *
240 iflib_get_media(if_ctx_t ctx)
241 {
242 
243 	return (ctx->ifc_mediap);
244 }
245 
246 uint32_t
247 iflib_get_flags(if_ctx_t ctx)
248 {
249 	return (ctx->ifc_flags);
250 }
251 
252 void
253 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
254 {
255 
256 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
257 }
258 
259 if_softc_ctx_t
260 iflib_get_softc_ctx(if_ctx_t ctx)
261 {
262 
263 	return (&ctx->ifc_softc_ctx);
264 }
265 
266 if_shared_ctx_t
267 iflib_get_sctx(if_ctx_t ctx)
268 {
269 
270 	return (ctx->ifc_sctx);
271 }
272 
273 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
274 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
275 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
276 
277 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
278 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
279 
280 typedef struct iflib_sw_rx_desc_array {
281 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
282 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
283 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
284 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
285 } iflib_rxsd_array_t;
286 
287 typedef struct iflib_sw_tx_desc_array {
288 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
289 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
290 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
291 } if_txsd_vec_t;
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 #define IFLIB_RX_COPY_THRESH		128
296 #define IFLIB_MAX_RX_REFRESH		32
297 /* The minimum descriptors per second before we start coalescing */
298 #define IFLIB_MIN_DESC_SEC		16384
299 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
300 #define IFLIB_QUEUE_IDLE		0
301 #define IFLIB_QUEUE_HUNG		1
302 #define IFLIB_QUEUE_WORKING		2
303 /* maximum number of txqs that can share an rx interrupt */
304 #define IFLIB_MAX_TX_SHARED_INTR	4
305 
306 /* this should really scale with ring size - this is a fairly arbitrary value */
307 #define TX_BATCH_SIZE			32
308 
309 #define IFLIB_RESTART_BUDGET		8
310 
311 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
312 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
313 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
314 
315 struct iflib_txq {
316 	qidx_t		ift_in_use;
317 	qidx_t		ift_cidx;
318 	qidx_t		ift_cidx_processed;
319 	qidx_t		ift_pidx;
320 	uint8_t		ift_gen;
321 	uint8_t		ift_br_offset;
322 	uint16_t	ift_npending;
323 	uint16_t	ift_db_pending;
324 	uint16_t	ift_rs_pending;
325 	/* implicit pad */
326 	uint8_t		ift_txd_size[8];
327 	uint64_t	ift_processed;
328 	uint64_t	ift_cleaned;
329 	uint64_t	ift_cleaned_prev;
330 #if MEMORY_LOGGING
331 	uint64_t	ift_enqueued;
332 	uint64_t	ift_dequeued;
333 #endif
334 	uint64_t	ift_no_tx_dma_setup;
335 	uint64_t	ift_no_desc_avail;
336 	uint64_t	ift_mbuf_defrag_failed;
337 	uint64_t	ift_mbuf_defrag;
338 	uint64_t	ift_map_failed;
339 	uint64_t	ift_txd_encap_efbig;
340 	uint64_t	ift_pullups;
341 	uint64_t	ift_last_timer_tick;
342 
343 	struct mtx	ift_mtx;
344 	struct mtx	ift_db_mtx;
345 
346 	/* constant values */
347 	if_ctx_t	ift_ctx;
348 	struct ifmp_ring        *ift_br;
349 	struct grouptask	ift_task;
350 	qidx_t		ift_size;
351 	uint16_t	ift_id;
352 	struct callout	ift_timer;
353 #ifdef DEV_NETMAP
354 	struct callout	ift_netmap_timer;
355 #endif /* DEV_NETMAP */
356 
357 	if_txsd_vec_t	ift_sds;
358 	uint8_t		ift_qstatus;
359 	uint8_t		ift_closed;
360 	uint8_t		ift_update_freq;
361 	struct iflib_filter_info ift_filter_info;
362 	bus_dma_tag_t	ift_buf_tag;
363 	bus_dma_tag_t	ift_tso_buf_tag;
364 	iflib_dma_info_t	ift_ifdi;
365 #define	MTX_NAME_LEN	32
366 	char                    ift_mtx_name[MTX_NAME_LEN];
367 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
368 #ifdef IFLIB_DIAGNOSTICS
369 	uint64_t ift_cpu_exec_count[256];
370 #endif
371 } __aligned(CACHE_LINE_SIZE);
372 
373 struct iflib_fl {
374 	qidx_t		ifl_cidx;
375 	qidx_t		ifl_pidx;
376 	qidx_t		ifl_credits;
377 	uint8_t		ifl_gen;
378 	uint8_t		ifl_rxd_size;
379 #if MEMORY_LOGGING
380 	uint64_t	ifl_m_enqueued;
381 	uint64_t	ifl_m_dequeued;
382 	uint64_t	ifl_cl_enqueued;
383 	uint64_t	ifl_cl_dequeued;
384 #endif
385 	/* implicit pad */
386 	bitstr_t 	*ifl_rx_bitmap;
387 	qidx_t		ifl_fragidx;
388 	/* constant */
389 	qidx_t		ifl_size;
390 	uint16_t	ifl_buf_size;
391 	uint16_t	ifl_cltype;
392 	uma_zone_t	ifl_zone;
393 	iflib_rxsd_array_t	ifl_sds;
394 	iflib_rxq_t	ifl_rxq;
395 	uint8_t		ifl_id;
396 	bus_dma_tag_t	ifl_buf_tag;
397 	iflib_dma_info_t	ifl_ifdi;
398 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
399 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
400 }  __aligned(CACHE_LINE_SIZE);
401 
402 static inline qidx_t
403 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
404 {
405 	qidx_t used;
406 
407 	if (pidx > cidx)
408 		used = pidx - cidx;
409 	else if (pidx < cidx)
410 		used = size - cidx + pidx;
411 	else if (gen == 0 && pidx == cidx)
412 		used = 0;
413 	else if (gen == 1 && pidx == cidx)
414 		used = size;
415 	else
416 		panic("bad state");
417 
418 	return (used);
419 }
420 
421 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 
423 #define IDXDIFF(head, tail, wrap) \
424 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
425 
426 struct iflib_rxq {
427 	if_ctx_t	ifr_ctx;
428 	iflib_fl_t	ifr_fl;
429 	uint64_t	ifr_rx_irq;
430 	struct pfil_head	*pfil;
431 	/*
432 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
433 	 * the completion queue consumer index.  Otherwise it's unused.
434 	 */
435 	qidx_t		ifr_cq_cidx;
436 	uint16_t	ifr_id;
437 	uint8_t		ifr_nfl;
438 	uint8_t		ifr_ntxqirq;
439 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
440 	uint8_t		ifr_fl_offset;
441 	struct lro_ctrl			ifr_lc;
442 	struct grouptask        ifr_task;
443 	struct callout		ifr_watchdog;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	iflib_fl_t ifsd_fl;
457 } *if_rxsd_t;
458 
459 /* multiple of word size */
460 #ifdef __LP64__
461 #define PKT_INFO_SIZE	6
462 #define RXD_INFO_SIZE	5
463 #define PKT_TYPE uint64_t
464 #else
465 #define PKT_INFO_SIZE	11
466 #define RXD_INFO_SIZE	8
467 #define PKT_TYPE uint32_t
468 #endif
469 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
470 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
471 
472 typedef struct if_pkt_info_pad {
473 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
474 } *if_pkt_info_pad_t;
475 typedef struct if_rxd_info_pad {
476 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
477 } *if_rxd_info_pad_t;
478 
479 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
480 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
481 
482 static inline void
483 pkt_info_zero(if_pkt_info_t pi)
484 {
485 	if_pkt_info_pad_t pi_pad;
486 
487 	pi_pad = (if_pkt_info_pad_t)pi;
488 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
489 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
490 #ifndef __LP64__
491 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
492 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
493 #endif
494 }
495 
496 static device_method_t iflib_pseudo_methods[] = {
497 	DEVMETHOD(device_attach, noop_attach),
498 	DEVMETHOD(device_detach, iflib_pseudo_detach),
499 	DEVMETHOD_END
500 };
501 
502 driver_t iflib_pseudodriver = {
503 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
504 };
505 
506 static inline void
507 rxd_info_zero(if_rxd_info_t ri)
508 {
509 	if_rxd_info_pad_t ri_pad;
510 	int i;
511 
512 	ri_pad = (if_rxd_info_pad_t)ri;
513 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
514 		ri_pad->rxd_val[i] = 0;
515 		ri_pad->rxd_val[i+1] = 0;
516 		ri_pad->rxd_val[i+2] = 0;
517 		ri_pad->rxd_val[i+3] = 0;
518 	}
519 #ifdef __LP64__
520 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
521 #endif
522 }
523 
524 /*
525  * Only allow a single packet to take up most 1/nth of the tx ring
526  */
527 #define MAX_SINGLE_PACKET_FRACTION 12
528 #define IF_BAD_DMA (bus_addr_t)-1
529 
530 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
531 
532 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
533 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
535 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
536 
537 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
538 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
539 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
540 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
541 
542 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
543 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
544 
545 void
546 iflib_set_detach(if_ctx_t ctx)
547 {
548 	STATE_LOCK(ctx);
549 	ctx->ifc_flags |= IFC_IN_DETACH;
550 	STATE_UNLOCK(ctx);
551 }
552 
553 /* Our boot-time initialization hook */
554 static int	iflib_module_event_handler(module_t, int, void *);
555 
556 static moduledata_t iflib_moduledata = {
557 	"iflib",
558 	iflib_module_event_handler,
559 	NULL
560 };
561 
562 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
563 MODULE_VERSION(iflib, 1);
564 
565 MODULE_DEPEND(iflib, pci, 1, 1, 1);
566 MODULE_DEPEND(iflib, ether, 1, 1, 1);
567 
568 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
569 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
570 
571 #ifndef IFLIB_DEBUG_COUNTERS
572 #ifdef INVARIANTS
573 #define IFLIB_DEBUG_COUNTERS 1
574 #else
575 #define IFLIB_DEBUG_COUNTERS 0
576 #endif /* !INVARIANTS */
577 #endif
578 
579 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
580     "iflib driver parameters");
581 
582 /*
583  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
584  */
585 static int iflib_min_tx_latency = 0;
586 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
587 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
588 static int iflib_no_tx_batch = 0;
589 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
590 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
591 static int iflib_timer_default = 1000;
592 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
593 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
594 
595 
596 #if IFLIB_DEBUG_COUNTERS
597 
598 static int iflib_tx_seen;
599 static int iflib_tx_sent;
600 static int iflib_tx_encap;
601 static int iflib_rx_allocs;
602 static int iflib_fl_refills;
603 static int iflib_fl_refills_large;
604 static int iflib_tx_frees;
605 
606 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
607 		   &iflib_tx_seen, 0, "# TX mbufs seen");
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
609 		   &iflib_tx_sent, 0, "# TX mbufs sent");
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
611 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
613 		   &iflib_tx_frees, 0, "# TX frees");
614 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
615 		   &iflib_rx_allocs, 0, "# RX allocations");
616 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
617 		   &iflib_fl_refills, 0, "# refills");
618 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
619 		   &iflib_fl_refills_large, 0, "# large refills");
620 
621 static int iflib_txq_drain_flushing;
622 static int iflib_txq_drain_oactive;
623 static int iflib_txq_drain_notready;
624 
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
626 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
627 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
628 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
630 		   &iflib_txq_drain_notready, 0, "# drain notready");
631 
632 static int iflib_encap_load_mbuf_fail;
633 static int iflib_encap_pad_mbuf_fail;
634 static int iflib_encap_txq_avail_fail;
635 static int iflib_encap_txd_encap_fail;
636 
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
638 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
640 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
642 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
644 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
645 
646 static int iflib_task_fn_rxs;
647 static int iflib_rx_intr_enables;
648 static int iflib_fast_intrs;
649 static int iflib_rx_unavail;
650 static int iflib_rx_ctx_inactive;
651 static int iflib_rx_if_input;
652 static int iflib_rxd_flush;
653 
654 static int iflib_verbose_debug;
655 
656 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
657 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
659 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
660 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
661 		   &iflib_fast_intrs, 0, "# fast_intr calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
663 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
665 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
667 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
669 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
670 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
671 		   &iflib_verbose_debug, 0, "enable verbose debugging");
672 
673 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
674 static void
675 iflib_debug_reset(void)
676 {
677 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
678 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
679 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
680 		iflib_txq_drain_notready =
681 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
682 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
683 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
684 		iflib_rx_unavail =
685 		iflib_rx_ctx_inactive = iflib_rx_if_input =
686 		iflib_rxd_flush = 0;
687 }
688 
689 #else
690 #define DBG_COUNTER_INC(name)
691 static void iflib_debug_reset(void) {}
692 #endif
693 
694 #define IFLIB_DEBUG 0
695 
696 static void iflib_tx_structures_free(if_ctx_t ctx);
697 static void iflib_rx_structures_free(if_ctx_t ctx);
698 static int iflib_queues_alloc(if_ctx_t ctx);
699 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
700 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
701 static int iflib_qset_structures_setup(if_ctx_t ctx);
702 static int iflib_msix_init(if_ctx_t ctx);
703 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
704 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
705 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
706 #ifdef ALTQ
707 static void iflib_altq_if_start(if_t ifp);
708 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
709 #endif
710 static int iflib_register(if_ctx_t);
711 static void iflib_deregister(if_ctx_t);
712 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
713 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
714 static void iflib_init_locked(if_ctx_t ctx);
715 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
716 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
717 static void iflib_ifmp_purge(iflib_txq_t txq);
718 static void _iflib_pre_assert(if_softc_ctx_t scctx);
719 static void iflib_if_init_locked(if_ctx_t ctx);
720 static void iflib_free_intr_mem(if_ctx_t ctx);
721 #ifndef __NO_STRICT_ALIGNMENT
722 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
723 #endif
724 
725 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
726     SLIST_HEAD_INITIALIZER(cpu_offsets);
727 struct cpu_offset {
728 	SLIST_ENTRY(cpu_offset) entries;
729 	cpuset_t	set;
730 	unsigned int	refcount;
731 	uint16_t	next_cpuid;
732 };
733 static struct mtx cpu_offset_mtx;
734 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
735     MTX_DEF);
736 
737 DEBUGNET_DEFINE(iflib);
738 
739 static int
740 iflib_num_rx_descs(if_ctx_t ctx)
741 {
742 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
743 	if_shared_ctx_t sctx = ctx->ifc_sctx;
744 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
745 
746 	return scctx->isc_nrxd[first_rxq];
747 }
748 
749 static int
750 iflib_num_tx_descs(if_ctx_t ctx)
751 {
752 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
753 	if_shared_ctx_t sctx = ctx->ifc_sctx;
754 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
755 
756 	return scctx->isc_ntxd[first_txq];
757 }
758 
759 #ifdef DEV_NETMAP
760 #include <sys/selinfo.h>
761 #include <net/netmap.h>
762 #include <dev/netmap/netmap_kern.h>
763 
764 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
765 
766 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
767 static void iflib_netmap_timer(void *arg);
768 
769 /*
770  * device-specific sysctl variables:
771  *
772  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
773  *	During regular operations the CRC is stripped, but on some
774  *	hardware reception of frames not multiple of 64 is slower,
775  *	so using crcstrip=0 helps in benchmarks.
776  *
777  * iflib_rx_miss, iflib_rx_miss_bufs:
778  *	count packets that might be missed due to lost interrupts.
779  */
780 SYSCTL_DECL(_dev_netmap);
781 /*
782  * The xl driver by default strips CRCs and we do not override it.
783  */
784 
785 int iflib_crcstrip = 1;
786 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
787     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
788 
789 int iflib_rx_miss, iflib_rx_miss_bufs;
790 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
791     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
792 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
793     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
794 
795 /*
796  * Register/unregister. We are already under netmap lock.
797  * Only called on the first register or the last unregister.
798  */
799 static int
800 iflib_netmap_register(struct netmap_adapter *na, int onoff)
801 {
802 	if_t ifp = na->ifp;
803 	if_ctx_t ctx = ifp->if_softc;
804 	int status;
805 
806 	CTX_LOCK(ctx);
807 	if (!CTX_IS_VF(ctx))
808 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
809 
810 	iflib_stop(ctx);
811 
812 	/*
813 	 * Enable (or disable) netmap flags, and intercept (or restore)
814 	 * ifp->if_transmit. This is done once the device has been stopped
815 	 * to prevent race conditions. Also, this must be done after
816 	 * calling netmap_disable_all_rings() and before calling
817 	 * netmap_enable_all_rings(), so that these two functions see the
818 	 * updated state of the NAF_NETMAP_ON bit.
819 	 */
820 	if (onoff) {
821 		nm_set_native_flags(na);
822 	} else {
823 		nm_clear_native_flags(na);
824 	}
825 
826 	iflib_init_locked(ctx);
827 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
828 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
829 	if (status)
830 		nm_clear_native_flags(na);
831 	CTX_UNLOCK(ctx);
832 	return (status);
833 }
834 
835 static int
836 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
837 {
838 	if_t ifp = na->ifp;
839 	if_ctx_t ctx = ifp->if_softc;
840 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
841 	iflib_fl_t fl = &rxq->ifr_fl[0];
842 
843 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
844 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
845 	info->num_tx_descs = iflib_num_tx_descs(ctx);
846 	info->num_rx_descs = iflib_num_rx_descs(ctx);
847 	info->rx_buf_maxsize = fl->ifl_buf_size;
848 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
849 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
850 		info->num_rx_descs, info->rx_buf_maxsize);
851 
852 	return 0;
853 }
854 
855 static int
856 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
857 {
858 	struct netmap_adapter *na = kring->na;
859 	u_int const lim = kring->nkr_num_slots - 1;
860 	struct netmap_ring *ring = kring->ring;
861 	bus_dmamap_t *map;
862 	struct if_rxd_update iru;
863 	if_ctx_t ctx = rxq->ifr_ctx;
864 	iflib_fl_t fl = &rxq->ifr_fl[0];
865 	u_int nic_i_first, nic_i;
866 	u_int nm_i;
867 	int i, n;
868 #if IFLIB_DEBUG_COUNTERS
869 	int rf_count = 0;
870 #endif
871 
872 	/*
873 	 * This function is used both at initialization and in rxsync.
874 	 * At initialization we need to prepare (with isc_rxd_refill())
875 	 * all the netmap buffers currently owned by the kernel, in
876 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
877 	 * (except for kring->nkr_hwofs). These may be less than
878 	 * kring->nkr_num_slots if netmap_reset() was called while
879 	 * an application using the kring that still owned some
880 	 * buffers.
881 	 * At rxsync time, both indexes point to the next buffer to be
882 	 * refilled.
883 	 * In any case we publish (with isc_rxd_flush()) up to
884 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
885 	 * pointer to overrun the head/cons pointer, although this is
886 	 * not necessary for some NICs (e.g. vmx).
887 	 */
888 	if (__predict_false(init)) {
889 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
890 	} else {
891 		n = kring->rhead - kring->nr_hwcur;
892 		if (n == 0)
893 			return (0); /* Nothing to do. */
894 		if (n < 0)
895 			n += kring->nkr_num_slots;
896 	}
897 
898 	iru_init(&iru, rxq, 0 /* flid */);
899 	map = fl->ifl_sds.ifsd_map;
900 	nic_i = fl->ifl_pidx;
901 	nm_i = netmap_idx_n2k(kring, nic_i);
902 	if (__predict_false(init)) {
903 		/*
904 		 * On init/reset, nic_i must be 0, and we must
905 		 * start to refill from hwtail (see netmap_reset()).
906 		 */
907 		MPASS(nic_i == 0);
908 		MPASS(nm_i == kring->nr_hwtail);
909 	} else
910 		MPASS(nm_i == kring->nr_hwcur);
911 	DBG_COUNTER_INC(fl_refills);
912 	while (n > 0) {
913 #if IFLIB_DEBUG_COUNTERS
914 		if (++rf_count == 9)
915 			DBG_COUNTER_INC(fl_refills_large);
916 #endif
917 		nic_i_first = nic_i;
918 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
919 			struct netmap_slot *slot = &ring->slot[nm_i];
920 			uint64_t paddr;
921 			void *addr = PNMB(na, slot, &paddr);
922 
923 			MPASS(i < IFLIB_MAX_RX_REFRESH);
924 
925 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
926 			        return netmap_ring_reinit(kring);
927 
928 			fl->ifl_bus_addrs[i] = paddr +
929 			    nm_get_offset(kring, slot);
930 			fl->ifl_rxd_idxs[i] = nic_i;
931 
932 			if (__predict_false(init)) {
933 				netmap_load_map(na, fl->ifl_buf_tag,
934 				    map[nic_i], addr);
935 			} else if (slot->flags & NS_BUF_CHANGED) {
936 				/* buffer has changed, reload map */
937 				netmap_reload_map(na, fl->ifl_buf_tag,
938 				    map[nic_i], addr);
939 			}
940 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
941 			    BUS_DMASYNC_PREREAD);
942 			slot->flags &= ~NS_BUF_CHANGED;
943 
944 			nm_i = nm_next(nm_i, lim);
945 			nic_i = nm_next(nic_i, lim);
946 		}
947 
948 		iru.iru_pidx = nic_i_first;
949 		iru.iru_count = i;
950 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
951 	}
952 	fl->ifl_pidx = nic_i;
953 	/*
954 	 * At the end of the loop we must have refilled everything
955 	 * we could possibly refill.
956 	 */
957 	MPASS(nm_i == kring->rhead);
958 	kring->nr_hwcur = nm_i;
959 
960 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
961 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
962 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
963 	    nm_prev(nic_i, lim));
964 	DBG_COUNTER_INC(rxd_flush);
965 
966 	return (0);
967 }
968 
969 #define NETMAP_TX_TIMER_US	90
970 
971 /*
972  * Reconcile kernel and user view of the transmit ring.
973  *
974  * All information is in the kring.
975  * Userspace wants to send packets up to the one before kring->rhead,
976  * kernel knows kring->nr_hwcur is the first unsent packet.
977  *
978  * Here we push packets out (as many as possible), and possibly
979  * reclaim buffers from previously completed transmission.
980  *
981  * The caller (netmap) guarantees that there is only one instance
982  * running at any time. Any interference with other driver
983  * methods should be handled by the individual drivers.
984  */
985 static int
986 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
987 {
988 	struct netmap_adapter *na = kring->na;
989 	if_t ifp = na->ifp;
990 	struct netmap_ring *ring = kring->ring;
991 	u_int nm_i;	/* index into the netmap kring */
992 	u_int nic_i;	/* index into the NIC ring */
993 	u_int n;
994 	u_int const lim = kring->nkr_num_slots - 1;
995 	u_int const head = kring->rhead;
996 	struct if_pkt_info pi;
997 	int tx_pkts = 0, tx_bytes = 0;
998 
999 	/*
1000 	 * interrupts on every tx packet are expensive so request
1001 	 * them every half ring, or where NS_REPORT is set
1002 	 */
1003 	u_int report_frequency = kring->nkr_num_slots >> 1;
1004 	/* device-specific */
1005 	if_ctx_t ctx = ifp->if_softc;
1006 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1007 
1008 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1009 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1010 
1011 	/*
1012 	 * First part: process new packets to send.
1013 	 * nm_i is the current index in the netmap kring,
1014 	 * nic_i is the corresponding index in the NIC ring.
1015 	 *
1016 	 * If we have packets to send (nm_i != head)
1017 	 * iterate over the netmap ring, fetch length and update
1018 	 * the corresponding slot in the NIC ring. Some drivers also
1019 	 * need to update the buffer's physical address in the NIC slot
1020 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1021 	 *
1022 	 * The netmap_reload_map() calls is especially expensive,
1023 	 * even when (as in this case) the tag is 0, so do only
1024 	 * when the buffer has actually changed.
1025 	 *
1026 	 * If possible do not set the report/intr bit on all slots,
1027 	 * but only a few times per ring or when NS_REPORT is set.
1028 	 *
1029 	 * Finally, on 10G and faster drivers, it might be useful
1030 	 * to prefetch the next slot and txr entry.
1031 	 */
1032 
1033 	nm_i = kring->nr_hwcur;
1034 	if (nm_i != head) {	/* we have new packets to send */
1035 		uint32_t pkt_len = 0, seg_idx = 0;
1036 		int nic_i_start = -1, flags = 0;
1037 		pkt_info_zero(&pi);
1038 		pi.ipi_segs = txq->ift_segs;
1039 		pi.ipi_qsidx = kring->ring_id;
1040 		nic_i = netmap_idx_k2n(kring, nm_i);
1041 
1042 		__builtin_prefetch(&ring->slot[nm_i]);
1043 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1044 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1045 
1046 		for (n = 0; nm_i != head; n++) {
1047 			struct netmap_slot *slot = &ring->slot[nm_i];
1048 			uint64_t offset = nm_get_offset(kring, slot);
1049 			u_int len = slot->len;
1050 			uint64_t paddr;
1051 			void *addr = PNMB(na, slot, &paddr);
1052 
1053 			flags |= (slot->flags & NS_REPORT ||
1054 				nic_i == 0 || nic_i == report_frequency) ?
1055 				IPI_TX_INTR : 0;
1056 
1057 			/*
1058 			 * If this is the first packet fragment, save the
1059 			 * index of the first NIC slot for later.
1060 			 */
1061 			if (nic_i_start < 0)
1062 				nic_i_start = nic_i;
1063 
1064 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1065 			pi.ipi_segs[seg_idx].ds_len = len;
1066 			if (len) {
1067 				pkt_len += len;
1068 				seg_idx++;
1069 			}
1070 
1071 			if (!(slot->flags & NS_MOREFRAG)) {
1072 				pi.ipi_len = pkt_len;
1073 				pi.ipi_nsegs = seg_idx;
1074 				pi.ipi_pidx = nic_i_start;
1075 				pi.ipi_ndescs = 0;
1076 				pi.ipi_flags = flags;
1077 
1078 				/* Prepare the NIC TX ring. */
1079 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1080 				DBG_COUNTER_INC(tx_encap);
1081 
1082 				/* Update transmit counters */
1083 				tx_bytes += pi.ipi_len;
1084 				tx_pkts++;
1085 
1086 				/* Reinit per-packet info for the next one. */
1087 				flags = seg_idx = pkt_len = 0;
1088 				nic_i_start = -1;
1089 			}
1090 
1091 			/* prefetch for next round */
1092 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1093 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1094 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1095 
1096 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1097 
1098 			if (slot->flags & NS_BUF_CHANGED) {
1099 				/* buffer has changed, reload map */
1100 				netmap_reload_map(na, txq->ift_buf_tag,
1101 				    txq->ift_sds.ifsd_map[nic_i], addr);
1102 			}
1103 			/* make sure changes to the buffer are synced */
1104 			bus_dmamap_sync(txq->ift_buf_tag,
1105 			    txq->ift_sds.ifsd_map[nic_i],
1106 			    BUS_DMASYNC_PREWRITE);
1107 
1108 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1109 			nm_i = nm_next(nm_i, lim);
1110 			nic_i = nm_next(nic_i, lim);
1111 		}
1112 		kring->nr_hwcur = nm_i;
1113 
1114 		/* synchronize the NIC ring */
1115 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1116 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1117 
1118 		/* (re)start the tx unit up to slot nic_i (excluded) */
1119 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1120 	}
1121 
1122 	/*
1123 	 * Second part: reclaim buffers for completed transmissions.
1124 	 *
1125 	 * If there are unclaimed buffers, attempt to reclaim them.
1126 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1127 	 * trigger a per-tx-queue timer to try again later.
1128 	 */
1129 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1130 		if (iflib_tx_credits_update(ctx, txq)) {
1131 			/* some tx completed, increment avail */
1132 			nic_i = txq->ift_cidx_processed;
1133 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1134 		}
1135 	}
1136 
1137 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1138 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1139 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1140 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1141 			    iflib_netmap_timer, txq,
1142 			    txq->ift_netmap_timer.c_cpu, 0);
1143 		}
1144 
1145 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1146 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1147 
1148 	return (0);
1149 }
1150 
1151 /*
1152  * Reconcile kernel and user view of the receive ring.
1153  * Same as for the txsync, this routine must be efficient.
1154  * The caller guarantees a single invocations, but races against
1155  * the rest of the driver should be handled here.
1156  *
1157  * On call, kring->rhead is the first packet that userspace wants
1158  * to keep, and kring->rcur is the wakeup point.
1159  * The kernel has previously reported packets up to kring->rtail.
1160  *
1161  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1162  * of whether or not we received an interrupt.
1163  */
1164 static int
1165 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1166 {
1167 	struct netmap_adapter *na = kring->na;
1168 	struct netmap_ring *ring = kring->ring;
1169 	if_t ifp = na->ifp;
1170 	uint32_t nm_i;	/* index into the netmap ring */
1171 	uint32_t nic_i;	/* index into the NIC ring */
1172 	u_int n;
1173 	u_int const lim = kring->nkr_num_slots - 1;
1174 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1175 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1176 
1177 	if_ctx_t ctx = ifp->if_softc;
1178 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1179 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1180 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1181 	iflib_fl_t fl = &rxq->ifr_fl[0];
1182 	struct if_rxd_info ri;
1183 	qidx_t *cidxp;
1184 
1185 	/*
1186 	 * netmap only uses free list 0, to avoid out of order consumption
1187 	 * of receive buffers
1188 	 */
1189 
1190 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1191 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1192 
1193 	/*
1194 	 * First part: import newly received packets.
1195 	 *
1196 	 * nm_i is the index of the next free slot in the netmap ring,
1197 	 * nic_i is the index of the next received packet in the NIC ring
1198 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1199 	 * differ in case if_init() has been called while
1200 	 * in netmap mode. For the receive ring we have
1201 	 *
1202 	 *	nic_i = fl->ifl_cidx;
1203 	 *	nm_i = kring->nr_hwtail (previous)
1204 	 * and
1205 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1206 	 *
1207 	 * fl->ifl_cidx is set to 0 on a ring reinit
1208 	 */
1209 	if (netmap_no_pendintr || force_update) {
1210 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1211 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1212 		int crclen = iflib_crcstrip ? 0 : 4;
1213 		int error, avail;
1214 
1215 		/*
1216 		 * For the free list consumer index, we use the same
1217 		 * logic as in iflib_rxeof().
1218 		 */
1219 		if (have_rxcq)
1220 			cidxp = &rxq->ifr_cq_cidx;
1221 		else
1222 			cidxp = &fl->ifl_cidx;
1223 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1224 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1225 
1226 		nic_i = fl->ifl_cidx;
1227 		nm_i = netmap_idx_n2k(kring, nic_i);
1228 		MPASS(nm_i == kring->nr_hwtail);
1229 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1230 			rxd_info_zero(&ri);
1231 			ri.iri_frags = rxq->ifr_frags;
1232 			ri.iri_qsidx = kring->ring_id;
1233 			ri.iri_ifp = ctx->ifc_ifp;
1234 			ri.iri_cidx = *cidxp;
1235 
1236 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1237 			for (i = 0; i < ri.iri_nfrags; i++) {
1238 				if (error) {
1239 					ring->slot[nm_i].len = 0;
1240 					ring->slot[nm_i].flags = 0;
1241 				} else {
1242 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1243 					if (i == (ri.iri_nfrags - 1)) {
1244 						ring->slot[nm_i].len -= crclen;
1245 						ring->slot[nm_i].flags = 0;
1246 
1247 						/* Update receive counters */
1248 						rx_bytes += ri.iri_len;
1249 						rx_pkts++;
1250 					} else
1251 						ring->slot[nm_i].flags = NS_MOREFRAG;
1252 				}
1253 
1254 				bus_dmamap_sync(fl->ifl_buf_tag,
1255 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1256 				nm_i = nm_next(nm_i, lim);
1257 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1258 			}
1259 
1260 			if (have_rxcq) {
1261 				*cidxp = ri.iri_cidx;
1262 				while (*cidxp >= scctx->isc_nrxd[0])
1263 					*cidxp -= scctx->isc_nrxd[0];
1264 			}
1265 
1266 		}
1267 		if (n) { /* update the state variables */
1268 			if (netmap_no_pendintr && !force_update) {
1269 				/* diagnostics */
1270 				iflib_rx_miss ++;
1271 				iflib_rx_miss_bufs += n;
1272 			}
1273 			kring->nr_hwtail = nm_i;
1274 		}
1275 		kring->nr_kflags &= ~NKR_PENDINTR;
1276 	}
1277 	/*
1278 	 * Second part: skip past packets that userspace has released.
1279 	 * (kring->nr_hwcur to head excluded),
1280 	 * and make the buffers available for reception.
1281 	 * As usual nm_i is the index in the netmap ring,
1282 	 * nic_i is the index in the NIC ring, and
1283 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1284 	 */
1285 	netmap_fl_refill(rxq, kring, false);
1286 
1287 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1288 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1289 
1290 	return (0);
1291 }
1292 
1293 static void
1294 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1295 {
1296 	if_ctx_t ctx = na->ifp->if_softc;
1297 
1298 	CTX_LOCK(ctx);
1299 	if (onoff) {
1300 		IFDI_INTR_ENABLE(ctx);
1301 	} else {
1302 		IFDI_INTR_DISABLE(ctx);
1303 	}
1304 	CTX_UNLOCK(ctx);
1305 }
1306 
1307 static int
1308 iflib_netmap_attach(if_ctx_t ctx)
1309 {
1310 	struct netmap_adapter na;
1311 
1312 	bzero(&na, sizeof(na));
1313 
1314 	na.ifp = ctx->ifc_ifp;
1315 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1316 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1317 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1318 
1319 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1320 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1321 	na.nm_txsync = iflib_netmap_txsync;
1322 	na.nm_rxsync = iflib_netmap_rxsync;
1323 	na.nm_register = iflib_netmap_register;
1324 	na.nm_intr = iflib_netmap_intr;
1325 	na.nm_config = iflib_netmap_config;
1326 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1327 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1328 	return (netmap_attach(&na));
1329 }
1330 
1331 static int
1332 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1333 {
1334 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1335 	struct netmap_slot *slot;
1336 
1337 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1338 	if (slot == NULL)
1339 		return (0);
1340 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1341 		/*
1342 		 * In netmap mode, set the map for the packet buffer.
1343 		 * NOTE: Some drivers (not this one) also need to set
1344 		 * the physical buffer address in the NIC ring.
1345 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1346 		 * netmap slot index, si
1347 		 */
1348 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1349 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1350 		    NMB(na, slot + si));
1351 	}
1352 	return (1);
1353 }
1354 
1355 static int
1356 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1357 {
1358 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1359 	struct netmap_kring *kring;
1360 	struct netmap_slot *slot;
1361 
1362 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1363 	if (slot == NULL)
1364 		return (0);
1365 	kring = na->rx_rings[rxq->ifr_id];
1366 	netmap_fl_refill(rxq, kring, true);
1367 	return (1);
1368 }
1369 
1370 static void
1371 iflib_netmap_timer(void *arg)
1372 {
1373 	iflib_txq_t txq = arg;
1374 	if_ctx_t ctx = txq->ift_ctx;
1375 
1376 	/*
1377 	 * Wake up the netmap application, to give it a chance to
1378 	 * call txsync and reclaim more completed TX buffers.
1379 	 */
1380 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1381 }
1382 
1383 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1384 
1385 #else
1386 #define iflib_netmap_txq_init(ctx, txq) (0)
1387 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1388 #define iflib_netmap_detach(ifp)
1389 #define netmap_enable_all_rings(ifp)
1390 #define netmap_disable_all_rings(ifp)
1391 
1392 #define iflib_netmap_attach(ctx) (0)
1393 #define netmap_rx_irq(ifp, qid, budget) (0)
1394 #endif
1395 
1396 #if defined(__i386__) || defined(__amd64__)
1397 static __inline void
1398 prefetch(void *x)
1399 {
1400 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1401 }
1402 static __inline void
1403 prefetch2cachelines(void *x)
1404 {
1405 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1406 #if (CACHE_LINE_SIZE < 128)
1407 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1408 #endif
1409 }
1410 #else
1411 #define prefetch(x)
1412 #define prefetch2cachelines(x)
1413 #endif
1414 
1415 static void
1416 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1417 {
1418 	iflib_fl_t fl;
1419 
1420 	fl = &rxq->ifr_fl[flid];
1421 	iru->iru_paddrs = fl->ifl_bus_addrs;
1422 	iru->iru_idxs = fl->ifl_rxd_idxs;
1423 	iru->iru_qsidx = rxq->ifr_id;
1424 	iru->iru_buf_size = fl->ifl_buf_size;
1425 	iru->iru_flidx = fl->ifl_id;
1426 }
1427 
1428 static void
1429 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1430 {
1431 	if (err)
1432 		return;
1433 	*(bus_addr_t *) arg = segs[0].ds_addr;
1434 }
1435 
1436 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1437 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1438 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1439 
1440 int
1441 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1442 {
1443 	int err;
1444 	device_t dev = ctx->ifc_dev;
1445 	bus_addr_t lowaddr;
1446 
1447 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1448 
1449 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1450 				align, 0,		/* alignment, bounds */
1451 				lowaddr,		/* lowaddr */
1452 				BUS_SPACE_MAXADDR,	/* highaddr */
1453 				NULL, NULL,		/* filter, filterarg */
1454 				size,			/* maxsize */
1455 				1,			/* nsegments */
1456 				size,			/* maxsegsize */
1457 				BUS_DMA_ALLOCNOW,	/* flags */
1458 				NULL,			/* lockfunc */
1459 				NULL,			/* lockarg */
1460 				&dma->idi_tag);
1461 	if (err) {
1462 		device_printf(dev,
1463 		    "%s: bus_dma_tag_create failed: %d\n",
1464 		    __func__, err);
1465 		goto fail_0;
1466 	}
1467 
1468 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1469 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1470 	if (err) {
1471 		device_printf(dev,
1472 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1473 		    __func__, (uintmax_t)size, err);
1474 		goto fail_1;
1475 	}
1476 
1477 	dma->idi_paddr = IF_BAD_DMA;
1478 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1479 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1480 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1481 		device_printf(dev,
1482 		    "%s: bus_dmamap_load failed: %d\n",
1483 		    __func__, err);
1484 		goto fail_2;
1485 	}
1486 
1487 	dma->idi_size = size;
1488 	return (0);
1489 
1490 fail_2:
1491 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1492 fail_1:
1493 	bus_dma_tag_destroy(dma->idi_tag);
1494 fail_0:
1495 	dma->idi_tag = NULL;
1496 
1497 	return (err);
1498 }
1499 
1500 int
1501 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1502 {
1503 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1504 
1505 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1506 
1507 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1508 }
1509 
1510 int
1511 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1512 {
1513 	int i, err;
1514 	iflib_dma_info_t *dmaiter;
1515 
1516 	dmaiter = dmalist;
1517 	for (i = 0; i < count; i++, dmaiter++) {
1518 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1519 			break;
1520 	}
1521 	if (err)
1522 		iflib_dma_free_multi(dmalist, i);
1523 	return (err);
1524 }
1525 
1526 void
1527 iflib_dma_free(iflib_dma_info_t dma)
1528 {
1529 	if (dma->idi_tag == NULL)
1530 		return;
1531 	if (dma->idi_paddr != IF_BAD_DMA) {
1532 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1533 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1534 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1535 		dma->idi_paddr = IF_BAD_DMA;
1536 	}
1537 	if (dma->idi_vaddr != NULL) {
1538 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1539 		dma->idi_vaddr = NULL;
1540 	}
1541 	bus_dma_tag_destroy(dma->idi_tag);
1542 	dma->idi_tag = NULL;
1543 }
1544 
1545 void
1546 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1547 {
1548 	int i;
1549 	iflib_dma_info_t *dmaiter = dmalist;
1550 
1551 	for (i = 0; i < count; i++, dmaiter++)
1552 		iflib_dma_free(*dmaiter);
1553 }
1554 
1555 static int
1556 iflib_fast_intr(void *arg)
1557 {
1558 	iflib_filter_info_t info = arg;
1559 	struct grouptask *gtask = info->ifi_task;
1560 	int result;
1561 
1562 	DBG_COUNTER_INC(fast_intrs);
1563 	if (info->ifi_filter != NULL) {
1564 		result = info->ifi_filter(info->ifi_filter_arg);
1565 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1566 			return (result);
1567 	}
1568 
1569 	GROUPTASK_ENQUEUE(gtask);
1570 	return (FILTER_HANDLED);
1571 }
1572 
1573 static int
1574 iflib_fast_intr_rxtx(void *arg)
1575 {
1576 	iflib_filter_info_t info = arg;
1577 	struct grouptask *gtask = info->ifi_task;
1578 	if_ctx_t ctx;
1579 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1580 	iflib_txq_t txq;
1581 	void *sc;
1582 	int i, cidx, result;
1583 	qidx_t txqid;
1584 	bool intr_enable, intr_legacy;
1585 
1586 	DBG_COUNTER_INC(fast_intrs);
1587 	if (info->ifi_filter != NULL) {
1588 		result = info->ifi_filter(info->ifi_filter_arg);
1589 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1590 			return (result);
1591 	}
1592 
1593 	ctx = rxq->ifr_ctx;
1594 	sc = ctx->ifc_softc;
1595 	intr_enable = false;
1596 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1597 	MPASS(rxq->ifr_ntxqirq);
1598 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1599 		txqid = rxq->ifr_txqid[i];
1600 		txq = &ctx->ifc_txqs[txqid];
1601 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1602 		    BUS_DMASYNC_POSTREAD);
1603 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1604 			if (intr_legacy)
1605 				intr_enable = true;
1606 			else
1607 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1608 			continue;
1609 		}
1610 		GROUPTASK_ENQUEUE(&txq->ift_task);
1611 	}
1612 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1613 		cidx = rxq->ifr_cq_cidx;
1614 	else
1615 		cidx = rxq->ifr_fl[0].ifl_cidx;
1616 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1617 		GROUPTASK_ENQUEUE(gtask);
1618 	else {
1619 		if (intr_legacy)
1620 			intr_enable = true;
1621 		else
1622 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1623 		DBG_COUNTER_INC(rx_intr_enables);
1624 	}
1625 	if (intr_enable)
1626 		IFDI_INTR_ENABLE(ctx);
1627 	return (FILTER_HANDLED);
1628 }
1629 
1630 static int
1631 iflib_fast_intr_ctx(void *arg)
1632 {
1633 	iflib_filter_info_t info = arg;
1634 	struct grouptask *gtask = info->ifi_task;
1635 	int result;
1636 
1637 	DBG_COUNTER_INC(fast_intrs);
1638 	if (info->ifi_filter != NULL) {
1639 		result = info->ifi_filter(info->ifi_filter_arg);
1640 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1641 			return (result);
1642 	}
1643 
1644 	GROUPTASK_ENQUEUE(gtask);
1645 	return (FILTER_HANDLED);
1646 }
1647 
1648 static int
1649 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1650 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1651 		 const char *name)
1652 {
1653 	struct resource *res;
1654 	void *tag = NULL;
1655 	device_t dev = ctx->ifc_dev;
1656 	int flags, i, rc;
1657 
1658 	flags = RF_ACTIVE;
1659 	if (ctx->ifc_flags & IFC_LEGACY)
1660 		flags |= RF_SHAREABLE;
1661 	MPASS(rid < 512);
1662 	i = rid;
1663 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1664 	if (res == NULL) {
1665 		device_printf(dev,
1666 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1667 		return (ENOMEM);
1668 	}
1669 	irq->ii_res = res;
1670 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1671 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1672 						filter, handler, arg, &tag);
1673 	if (rc != 0) {
1674 		device_printf(dev,
1675 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1676 					  rid, name ? name : "unknown", rc);
1677 		return (rc);
1678 	} else if (name)
1679 		bus_describe_intr(dev, res, tag, "%s", name);
1680 
1681 	irq->ii_tag = tag;
1682 	return (0);
1683 }
1684 
1685 /*********************************************************************
1686  *
1687  *  Allocate DMA resources for TX buffers as well as memory for the TX
1688  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1689  *  iflib_sw_tx_desc_array structure, storing all the information that
1690  *  is needed to transmit a packet on the wire.  This is called only
1691  *  once at attach, setup is done every reset.
1692  *
1693  **********************************************************************/
1694 static int
1695 iflib_txsd_alloc(iflib_txq_t txq)
1696 {
1697 	if_ctx_t ctx = txq->ift_ctx;
1698 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1699 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1700 	device_t dev = ctx->ifc_dev;
1701 	bus_size_t tsomaxsize;
1702 	bus_addr_t lowaddr;
1703 	int err, nsegments, ntsosegments;
1704 	bool tso;
1705 
1706 	nsegments = scctx->isc_tx_nsegments;
1707 	ntsosegments = scctx->isc_tx_tso_segments_max;
1708 	tsomaxsize = scctx->isc_tx_tso_size_max;
1709 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1710 		tsomaxsize += sizeof(struct ether_vlan_header);
1711 	MPASS(scctx->isc_ntxd[0] > 0);
1712 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1713 	MPASS(nsegments > 0);
1714 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1715 		MPASS(ntsosegments > 0);
1716 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1717 	}
1718 
1719 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1720 
1721 	/*
1722 	 * Set up DMA tags for TX buffers.
1723 	 */
1724 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1725 			       1, 0,			/* alignment, bounds */
1726 			       lowaddr,			/* lowaddr */
1727 			       BUS_SPACE_MAXADDR,	/* highaddr */
1728 			       NULL, NULL,		/* filter, filterarg */
1729 			       sctx->isc_tx_maxsize,		/* maxsize */
1730 			       nsegments,	/* nsegments */
1731 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1732 			       0,			/* flags */
1733 			       NULL,			/* lockfunc */
1734 			       NULL,			/* lockfuncarg */
1735 			       &txq->ift_buf_tag))) {
1736 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1737 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1738 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1739 		goto fail;
1740 	}
1741 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1742 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1743 			       1, 0,			/* alignment, bounds */
1744 			       lowaddr,			/* lowaddr */
1745 			       BUS_SPACE_MAXADDR,	/* highaddr */
1746 			       NULL, NULL,		/* filter, filterarg */
1747 			       tsomaxsize,		/* maxsize */
1748 			       ntsosegments,	/* nsegments */
1749 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1750 			       0,			/* flags */
1751 			       NULL,			/* lockfunc */
1752 			       NULL,			/* lockfuncarg */
1753 			       &txq->ift_tso_buf_tag))) {
1754 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1755 		    err);
1756 		goto fail;
1757 	}
1758 
1759 	/* Allocate memory for the TX mbuf map. */
1760 	if (!(txq->ift_sds.ifsd_m =
1761 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1762 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1763 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1764 		err = ENOMEM;
1765 		goto fail;
1766 	}
1767 
1768 	/*
1769 	 * Create the DMA maps for TX buffers.
1770 	 */
1771 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1772 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1773 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1774 		device_printf(dev,
1775 		    "Unable to allocate TX buffer DMA map memory\n");
1776 		err = ENOMEM;
1777 		goto fail;
1778 	}
1779 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1780 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1781 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1782 		device_printf(dev,
1783 		    "Unable to allocate TSO TX buffer map memory\n");
1784 		err = ENOMEM;
1785 		goto fail;
1786 	}
1787 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1788 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1789 		    &txq->ift_sds.ifsd_map[i]);
1790 		if (err != 0) {
1791 			device_printf(dev, "Unable to create TX DMA map\n");
1792 			goto fail;
1793 		}
1794 		if (!tso)
1795 			continue;
1796 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1797 		    &txq->ift_sds.ifsd_tso_map[i]);
1798 		if (err != 0) {
1799 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1800 			goto fail;
1801 		}
1802 	}
1803 	return (0);
1804 fail:
1805 	/* We free all, it handles case where we are in the middle */
1806 	iflib_tx_structures_free(ctx);
1807 	return (err);
1808 }
1809 
1810 static void
1811 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1812 {
1813 	bus_dmamap_t map;
1814 
1815 	if (txq->ift_sds.ifsd_map != NULL) {
1816 		map = txq->ift_sds.ifsd_map[i];
1817 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1818 		bus_dmamap_unload(txq->ift_buf_tag, map);
1819 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1820 		txq->ift_sds.ifsd_map[i] = NULL;
1821 	}
1822 
1823 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1824 		map = txq->ift_sds.ifsd_tso_map[i];
1825 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1826 		    BUS_DMASYNC_POSTWRITE);
1827 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1828 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1829 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1830 	}
1831 }
1832 
1833 static void
1834 iflib_txq_destroy(iflib_txq_t txq)
1835 {
1836 	if_ctx_t ctx = txq->ift_ctx;
1837 
1838 	for (int i = 0; i < txq->ift_size; i++)
1839 		iflib_txsd_destroy(ctx, txq, i);
1840 
1841 	if (txq->ift_br != NULL) {
1842 		ifmp_ring_free(txq->ift_br);
1843 		txq->ift_br = NULL;
1844 	}
1845 
1846 	mtx_destroy(&txq->ift_mtx);
1847 
1848 	if (txq->ift_sds.ifsd_map != NULL) {
1849 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1850 		txq->ift_sds.ifsd_map = NULL;
1851 	}
1852 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1853 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1854 		txq->ift_sds.ifsd_tso_map = NULL;
1855 	}
1856 	if (txq->ift_sds.ifsd_m != NULL) {
1857 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1858 		txq->ift_sds.ifsd_m = NULL;
1859 	}
1860 	if (txq->ift_buf_tag != NULL) {
1861 		bus_dma_tag_destroy(txq->ift_buf_tag);
1862 		txq->ift_buf_tag = NULL;
1863 	}
1864 	if (txq->ift_tso_buf_tag != NULL) {
1865 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1866 		txq->ift_tso_buf_tag = NULL;
1867 	}
1868 	if (txq->ift_ifdi != NULL) {
1869 		free(txq->ift_ifdi, M_IFLIB);
1870 	}
1871 }
1872 
1873 static void
1874 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1875 {
1876 	struct mbuf **mp;
1877 
1878 	mp = &txq->ift_sds.ifsd_m[i];
1879 	if (*mp == NULL)
1880 		return;
1881 
1882 	if (txq->ift_sds.ifsd_map != NULL) {
1883 		bus_dmamap_sync(txq->ift_buf_tag,
1884 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1885 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1886 	}
1887 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1888 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1889 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1890 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1891 		    txq->ift_sds.ifsd_tso_map[i]);
1892 	}
1893 	m_freem(*mp);
1894 	DBG_COUNTER_INC(tx_frees);
1895 	*mp = NULL;
1896 }
1897 
1898 static int
1899 iflib_txq_setup(iflib_txq_t txq)
1900 {
1901 	if_ctx_t ctx = txq->ift_ctx;
1902 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1903 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1904 	iflib_dma_info_t di;
1905 	int i;
1906 
1907 	/* Set number of descriptors available */
1908 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1909 	/* XXX make configurable */
1910 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1911 
1912 	/* Reset indices */
1913 	txq->ift_cidx_processed = 0;
1914 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1915 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1916 
1917 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1918 		bzero((void *)di->idi_vaddr, di->idi_size);
1919 
1920 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1921 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1922 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1923 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1924 	return (0);
1925 }
1926 
1927 /*********************************************************************
1928  *
1929  *  Allocate DMA resources for RX buffers as well as memory for the RX
1930  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1931  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1932  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1933  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1934  *  packet, the maximum number of entries we'll need is equal to the
1935  *  number of hardware receive descriptors that we've allocated.
1936  *
1937  **********************************************************************/
1938 static int
1939 iflib_rxsd_alloc(iflib_rxq_t rxq)
1940 {
1941 	if_ctx_t ctx = rxq->ifr_ctx;
1942 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1943 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1944 	device_t dev = ctx->ifc_dev;
1945 	iflib_fl_t fl;
1946 	bus_addr_t lowaddr;
1947 	int			err;
1948 
1949 	MPASS(scctx->isc_nrxd[0] > 0);
1950 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1951 
1952 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1953 
1954 	fl = rxq->ifr_fl;
1955 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1956 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1957 		/* Set up DMA tag for RX buffers. */
1958 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1959 					 1, 0,			/* alignment, bounds */
1960 					 lowaddr,		/* lowaddr */
1961 					 BUS_SPACE_MAXADDR,	/* highaddr */
1962 					 NULL, NULL,		/* filter, filterarg */
1963 					 sctx->isc_rx_maxsize,	/* maxsize */
1964 					 sctx->isc_rx_nsegments,	/* nsegments */
1965 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1966 					 0,			/* flags */
1967 					 NULL,			/* lockfunc */
1968 					 NULL,			/* lockarg */
1969 					 &fl->ifl_buf_tag);
1970 		if (err) {
1971 			device_printf(dev,
1972 			    "Unable to allocate RX DMA tag: %d\n", err);
1973 			goto fail;
1974 		}
1975 
1976 		/* Allocate memory for the RX mbuf map. */
1977 		if (!(fl->ifl_sds.ifsd_m =
1978 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1979 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1980 			device_printf(dev,
1981 			    "Unable to allocate RX mbuf map memory\n");
1982 			err = ENOMEM;
1983 			goto fail;
1984 		}
1985 
1986 		/* Allocate memory for the direct RX cluster pointer map. */
1987 		if (!(fl->ifl_sds.ifsd_cl =
1988 		      (caddr_t *) malloc(sizeof(caddr_t) *
1989 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1990 			device_printf(dev,
1991 			    "Unable to allocate RX cluster map memory\n");
1992 			err = ENOMEM;
1993 			goto fail;
1994 		}
1995 
1996 		/* Allocate memory for the RX cluster bus address map. */
1997 		if (!(fl->ifl_sds.ifsd_ba =
1998 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1999 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2000 			device_printf(dev,
2001 			    "Unable to allocate RX bus address map memory\n");
2002 			err = ENOMEM;
2003 			goto fail;
2004 		}
2005 
2006 		/*
2007 		 * Create the DMA maps for RX buffers.
2008 		 */
2009 		if (!(fl->ifl_sds.ifsd_map =
2010 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2011 			device_printf(dev,
2012 			    "Unable to allocate RX buffer DMA map memory\n");
2013 			err = ENOMEM;
2014 			goto fail;
2015 		}
2016 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2017 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2018 			    &fl->ifl_sds.ifsd_map[i]);
2019 			if (err != 0) {
2020 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2021 				goto fail;
2022 			}
2023 		}
2024 	}
2025 	return (0);
2026 
2027 fail:
2028 	iflib_rx_structures_free(ctx);
2029 	return (err);
2030 }
2031 
2032 /*
2033  * Internal service routines
2034  */
2035 
2036 struct rxq_refill_cb_arg {
2037 	int               error;
2038 	bus_dma_segment_t seg;
2039 	int               nseg;
2040 };
2041 
2042 static void
2043 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2044 {
2045 	struct rxq_refill_cb_arg *cb_arg = arg;
2046 
2047 	cb_arg->error = error;
2048 	cb_arg->seg = segs[0];
2049 	cb_arg->nseg = nseg;
2050 }
2051 
2052 /**
2053  * iflib_fl_refill - refill an rxq free-buffer list
2054  * @ctx: the iflib context
2055  * @fl: the free list to refill
2056  * @count: the number of new buffers to allocate
2057  *
2058  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2059  * The caller must assure that @count does not exceed the queue's capacity
2060  * minus one (since we always leave a descriptor unavailable).
2061  */
2062 static uint8_t
2063 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2064 {
2065 	struct if_rxd_update iru;
2066 	struct rxq_refill_cb_arg cb_arg;
2067 	struct mbuf *m;
2068 	caddr_t cl, *sd_cl;
2069 	struct mbuf **sd_m;
2070 	bus_dmamap_t *sd_map;
2071 	bus_addr_t bus_addr, *sd_ba;
2072 	int err, frag_idx, i, idx, n, pidx;
2073 	qidx_t credits;
2074 
2075 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2076 
2077 	sd_m = fl->ifl_sds.ifsd_m;
2078 	sd_map = fl->ifl_sds.ifsd_map;
2079 	sd_cl = fl->ifl_sds.ifsd_cl;
2080 	sd_ba = fl->ifl_sds.ifsd_ba;
2081 	pidx = fl->ifl_pidx;
2082 	idx = pidx;
2083 	frag_idx = fl->ifl_fragidx;
2084 	credits = fl->ifl_credits;
2085 
2086 	i = 0;
2087 	n = count;
2088 	MPASS(n > 0);
2089 	MPASS(credits + n <= fl->ifl_size);
2090 
2091 	if (pidx < fl->ifl_cidx)
2092 		MPASS(pidx + n <= fl->ifl_cidx);
2093 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2094 		MPASS(fl->ifl_gen == 0);
2095 	if (pidx > fl->ifl_cidx)
2096 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2097 
2098 	DBG_COUNTER_INC(fl_refills);
2099 	if (n > 8)
2100 		DBG_COUNTER_INC(fl_refills_large);
2101 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2102 	while (n-- > 0) {
2103 		/*
2104 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2105 		 * initialized after rx.
2106 		 *
2107 		 * If the cluster is still set then we know a minimum sized
2108 		 * packet was received
2109 		 */
2110 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2111 		    &frag_idx);
2112 		if (frag_idx < 0)
2113 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2114 		MPASS(frag_idx >= 0);
2115 		if ((cl = sd_cl[frag_idx]) == NULL) {
2116 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2117 			if (__predict_false(cl == NULL))
2118 				break;
2119 
2120 			cb_arg.error = 0;
2121 			MPASS(sd_map != NULL);
2122 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2123 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2124 			    BUS_DMA_NOWAIT);
2125 			if (__predict_false(err != 0 || cb_arg.error)) {
2126 				uma_zfree(fl->ifl_zone, cl);
2127 				break;
2128 			}
2129 
2130 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2131 			sd_cl[frag_idx] = cl;
2132 #if MEMORY_LOGGING
2133 			fl->ifl_cl_enqueued++;
2134 #endif
2135 		} else {
2136 			bus_addr = sd_ba[frag_idx];
2137 		}
2138 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2139 		    BUS_DMASYNC_PREREAD);
2140 
2141 		if (sd_m[frag_idx] == NULL) {
2142 			m = m_gethdr_raw(M_NOWAIT, 0);
2143 			if (__predict_false(m == NULL))
2144 				break;
2145 			sd_m[frag_idx] = m;
2146 		}
2147 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2148 #if MEMORY_LOGGING
2149 		fl->ifl_m_enqueued++;
2150 #endif
2151 
2152 		DBG_COUNTER_INC(rx_allocs);
2153 		fl->ifl_rxd_idxs[i] = frag_idx;
2154 		fl->ifl_bus_addrs[i] = bus_addr;
2155 		credits++;
2156 		i++;
2157 		MPASS(credits <= fl->ifl_size);
2158 		if (++idx == fl->ifl_size) {
2159 #ifdef INVARIANTS
2160 			fl->ifl_gen = 1;
2161 #endif
2162 			idx = 0;
2163 		}
2164 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2165 			iru.iru_pidx = pidx;
2166 			iru.iru_count = i;
2167 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2168 			fl->ifl_pidx = idx;
2169 			fl->ifl_credits = credits;
2170 			pidx = idx;
2171 			i = 0;
2172 		}
2173 	}
2174 
2175 	if (n < count - 1) {
2176 		if (i != 0) {
2177 			iru.iru_pidx = pidx;
2178 			iru.iru_count = i;
2179 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2180 			fl->ifl_pidx = idx;
2181 			fl->ifl_credits = credits;
2182 		}
2183 		DBG_COUNTER_INC(rxd_flush);
2184 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2185 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2186 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2187 		    fl->ifl_id, fl->ifl_pidx);
2188 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2189 			fl->ifl_fragidx = frag_idx + 1;
2190 			if (fl->ifl_fragidx == fl->ifl_size)
2191 				fl->ifl_fragidx = 0;
2192 		} else {
2193 			fl->ifl_fragidx = frag_idx;
2194 		}
2195 	}
2196 
2197 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2198 }
2199 
2200 static inline uint8_t
2201 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2202 {
2203 	/*
2204 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2205 	 * This is important as it confuses most NICs. For instance,
2206 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2207 	 * RDH points to the next receive descriptor to be used by the NIC,
2208 	 * and RDT for the next receive descriptor to be published by the
2209 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2210 	 * The condition RDH == RDT means no descriptors are available to
2211 	 * the NIC, and thus it would be ambiguous if it also meant that
2212 	 * all the descriptors are available to the NIC.
2213 	 */
2214 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2215 #ifdef INVARIANTS
2216 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2217 #endif
2218 
2219 	MPASS(fl->ifl_credits <= fl->ifl_size);
2220 	MPASS(reclaimable == delta);
2221 
2222 	if (reclaimable > 0)
2223 		return (iflib_fl_refill(ctx, fl, reclaimable));
2224 	return (0);
2225 }
2226 
2227 uint8_t
2228 iflib_in_detach(if_ctx_t ctx)
2229 {
2230 	bool in_detach;
2231 
2232 	STATE_LOCK(ctx);
2233 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2234 	STATE_UNLOCK(ctx);
2235 	return (in_detach);
2236 }
2237 
2238 static void
2239 iflib_fl_bufs_free(iflib_fl_t fl)
2240 {
2241 	iflib_dma_info_t idi = fl->ifl_ifdi;
2242 	bus_dmamap_t sd_map;
2243 	uint32_t i;
2244 
2245 	for (i = 0; i < fl->ifl_size; i++) {
2246 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2247 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2248 
2249 		if (*sd_cl != NULL) {
2250 			sd_map = fl->ifl_sds.ifsd_map[i];
2251 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2252 			    BUS_DMASYNC_POSTREAD);
2253 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2254 			uma_zfree(fl->ifl_zone, *sd_cl);
2255 			*sd_cl = NULL;
2256 			if (*sd_m != NULL) {
2257 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2258 				m_free_raw(*sd_m);
2259 				*sd_m = NULL;
2260 			}
2261 		} else {
2262 			MPASS(*sd_m == NULL);
2263 		}
2264 #if MEMORY_LOGGING
2265 		fl->ifl_m_dequeued++;
2266 		fl->ifl_cl_dequeued++;
2267 #endif
2268 	}
2269 #ifdef INVARIANTS
2270 	for (i = 0; i < fl->ifl_size; i++) {
2271 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2272 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2273 	}
2274 #endif
2275 	/*
2276 	 * Reset free list values
2277 	 */
2278 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2279 	bzero(idi->idi_vaddr, idi->idi_size);
2280 }
2281 
2282 /*********************************************************************
2283  *
2284  *  Initialize a free list and its buffers.
2285  *
2286  **********************************************************************/
2287 static int
2288 iflib_fl_setup(iflib_fl_t fl)
2289 {
2290 	iflib_rxq_t rxq = fl->ifl_rxq;
2291 	if_ctx_t ctx = rxq->ifr_ctx;
2292 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2293 	int qidx;
2294 
2295 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2296 	/*
2297 	** Free current RX buffer structs and their mbufs
2298 	*/
2299 	iflib_fl_bufs_free(fl);
2300 	/* Now replenish the mbufs */
2301 	MPASS(fl->ifl_credits == 0);
2302 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2303 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2304 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2305 	else
2306 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2307 	/*
2308 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2309 	 * to the selected mbuf size.
2310 	 */
2311 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2312 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2313 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2314 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2315 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2316 
2317 	/*
2318 	 * Avoid pre-allocating zillions of clusters to an idle card
2319 	 * potentially speeding up attach. In any case make sure
2320 	 * to leave a descriptor unavailable. See the comment in
2321 	 * iflib_fl_refill_all().
2322 	 */
2323 	MPASS(fl->ifl_size > 0);
2324 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2325 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2326 		return (ENOBUFS);
2327 	/*
2328 	 * handle failure
2329 	 */
2330 	MPASS(rxq != NULL);
2331 	MPASS(fl->ifl_ifdi != NULL);
2332 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2333 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2334 	return (0);
2335 }
2336 
2337 /*********************************************************************
2338  *
2339  *  Free receive ring data structures
2340  *
2341  **********************************************************************/
2342 static void
2343 iflib_rx_sds_free(iflib_rxq_t rxq)
2344 {
2345 	iflib_fl_t fl;
2346 	int i, j;
2347 
2348 	if (rxq->ifr_fl != NULL) {
2349 		for (i = 0; i < rxq->ifr_nfl; i++) {
2350 			fl = &rxq->ifr_fl[i];
2351 			if (fl->ifl_buf_tag != NULL) {
2352 				if (fl->ifl_sds.ifsd_map != NULL) {
2353 					for (j = 0; j < fl->ifl_size; j++) {
2354 						bus_dmamap_sync(
2355 						    fl->ifl_buf_tag,
2356 						    fl->ifl_sds.ifsd_map[j],
2357 						    BUS_DMASYNC_POSTREAD);
2358 						bus_dmamap_unload(
2359 						    fl->ifl_buf_tag,
2360 						    fl->ifl_sds.ifsd_map[j]);
2361 						bus_dmamap_destroy(
2362 						    fl->ifl_buf_tag,
2363 						    fl->ifl_sds.ifsd_map[j]);
2364 					}
2365 				}
2366 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2367 				fl->ifl_buf_tag = NULL;
2368 			}
2369 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2370 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2371 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2372 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2373 			free(fl->ifl_rx_bitmap, M_IFLIB);
2374 			fl->ifl_sds.ifsd_m = NULL;
2375 			fl->ifl_sds.ifsd_cl = NULL;
2376 			fl->ifl_sds.ifsd_ba = NULL;
2377 			fl->ifl_sds.ifsd_map = NULL;
2378 			fl->ifl_rx_bitmap = NULL;
2379 		}
2380 		free(rxq->ifr_fl, M_IFLIB);
2381 		rxq->ifr_fl = NULL;
2382 		free(rxq->ifr_ifdi, M_IFLIB);
2383 		rxq->ifr_ifdi = NULL;
2384 		rxq->ifr_cq_cidx = 0;
2385 	}
2386 }
2387 
2388 /*
2389  * Timer routine
2390  */
2391 static void
2392 iflib_timer(void *arg)
2393 {
2394 	iflib_txq_t txq = arg;
2395 	if_ctx_t ctx = txq->ift_ctx;
2396 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2397 	uint64_t this_tick = ticks;
2398 
2399 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2400 		return;
2401 
2402 	/*
2403 	** Check on the state of the TX queue(s), this
2404 	** can be done without the lock because its RO
2405 	** and the HUNG state will be static if set.
2406 	*/
2407 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2408 		txq->ift_last_timer_tick = this_tick;
2409 		IFDI_TIMER(ctx, txq->ift_id);
2410 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2411 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2412 		     (sctx->isc_pause_frames == 0)))
2413 			goto hung;
2414 
2415 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2416 		    ifmp_ring_is_stalled(txq->ift_br)) {
2417 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2418 			    ("queue can't be marked as hung if interface is down"));
2419 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2420 		}
2421 		txq->ift_cleaned_prev = txq->ift_cleaned;
2422 	}
2423 	/* handle any laggards */
2424 	if (txq->ift_db_pending)
2425 		GROUPTASK_ENQUEUE(&txq->ift_task);
2426 
2427 	sctx->isc_pause_frames = 0;
2428 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2429 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2430 		    txq, txq->ift_timer.c_cpu);
2431 	return;
2432 
2433  hung:
2434 	device_printf(ctx->ifc_dev,
2435 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2436 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2437 	STATE_LOCK(ctx);
2438 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2439 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2440 	iflib_admin_intr_deferred(ctx);
2441 	STATE_UNLOCK(ctx);
2442 }
2443 
2444 static uint16_t
2445 iflib_get_mbuf_size_for(unsigned int size)
2446 {
2447 
2448 	if (size <= MCLBYTES)
2449 		return (MCLBYTES);
2450 	else
2451 		return (MJUMPAGESIZE);
2452 }
2453 
2454 static void
2455 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2456 {
2457 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2458 
2459 	/*
2460 	 * XXX don't set the max_frame_size to larger
2461 	 * than the hardware can handle
2462 	 */
2463 	ctx->ifc_rx_mbuf_sz =
2464 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2465 }
2466 
2467 uint32_t
2468 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2469 {
2470 
2471 	return (ctx->ifc_rx_mbuf_sz);
2472 }
2473 
2474 static void
2475 iflib_init_locked(if_ctx_t ctx)
2476 {
2477 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2478 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2479 	if_t ifp = ctx->ifc_ifp;
2480 	iflib_fl_t fl;
2481 	iflib_txq_t txq;
2482 	iflib_rxq_t rxq;
2483 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2484 
2485 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2486 	IFDI_INTR_DISABLE(ctx);
2487 
2488 	/*
2489 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2490 	 * called without first calling iflib_stop().
2491 	 */
2492 	netmap_disable_all_rings(ifp);
2493 
2494 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2495 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2496 	/* Set hardware offload abilities */
2497 	if_clearhwassist(ifp);
2498 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2499 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2500 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2501 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2502 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2503 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2504 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2505 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2506 
2507 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2508 		CALLOUT_LOCK(txq);
2509 		callout_stop(&txq->ift_timer);
2510 #ifdef DEV_NETMAP
2511 		callout_stop(&txq->ift_netmap_timer);
2512 #endif /* DEV_NETMAP */
2513 		CALLOUT_UNLOCK(txq);
2514 		(void)iflib_netmap_txq_init(ctx, txq);
2515 	}
2516 
2517 	/*
2518 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2519 	 * that drivers can use the value when setting up the hardware receive
2520 	 * buffers.
2521 	 */
2522 	iflib_calc_rx_mbuf_sz(ctx);
2523 
2524 #ifdef INVARIANTS
2525 	i = if_getdrvflags(ifp);
2526 #endif
2527 	IFDI_INIT(ctx);
2528 	MPASS(if_getdrvflags(ifp) == i);
2529 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2530 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2531 			/* This rxq is in netmap mode. Skip normal init. */
2532 			continue;
2533 		}
2534 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2535 			if (iflib_fl_setup(fl)) {
2536 				device_printf(ctx->ifc_dev,
2537 				    "setting up free list %d failed - "
2538 				    "check cluster settings\n", j);
2539 				goto done;
2540 			}
2541 		}
2542 	}
2543 done:
2544 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2545 	IFDI_INTR_ENABLE(ctx);
2546 	txq = ctx->ifc_txqs;
2547 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2548 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2549 			txq->ift_timer.c_cpu);
2550 
2551         /* Re-enable txsync/rxsync. */
2552 	netmap_enable_all_rings(ifp);
2553 }
2554 
2555 static int
2556 iflib_media_change(if_t ifp)
2557 {
2558 	if_ctx_t ctx = if_getsoftc(ifp);
2559 	int err;
2560 
2561 	CTX_LOCK(ctx);
2562 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2563 		iflib_if_init_locked(ctx);
2564 	CTX_UNLOCK(ctx);
2565 	return (err);
2566 }
2567 
2568 static void
2569 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2570 {
2571 	if_ctx_t ctx = if_getsoftc(ifp);
2572 
2573 	CTX_LOCK(ctx);
2574 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2575 	IFDI_MEDIA_STATUS(ctx, ifmr);
2576 	CTX_UNLOCK(ctx);
2577 }
2578 
2579 void
2580 iflib_stop(if_ctx_t ctx)
2581 {
2582 	iflib_txq_t txq = ctx->ifc_txqs;
2583 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2584 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2585 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2586 	iflib_dma_info_t di;
2587 	iflib_fl_t fl;
2588 	int i, j;
2589 
2590 	/* Tell the stack that the interface is no longer active */
2591 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2592 
2593 	IFDI_INTR_DISABLE(ctx);
2594 	DELAY(1000);
2595 	IFDI_STOP(ctx);
2596 	DELAY(1000);
2597 
2598 	/*
2599 	 * Stop any pending txsync/rxsync and prevent new ones
2600 	 * form starting. Processes blocked in poll() will get
2601 	 * POLLERR.
2602 	 */
2603 	netmap_disable_all_rings(ctx->ifc_ifp);
2604 
2605 	iflib_debug_reset();
2606 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2607 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2608 		/* make sure all transmitters have completed before proceeding XXX */
2609 
2610 		CALLOUT_LOCK(txq);
2611 		callout_stop(&txq->ift_timer);
2612 #ifdef DEV_NETMAP
2613 		callout_stop(&txq->ift_netmap_timer);
2614 #endif /* DEV_NETMAP */
2615 		CALLOUT_UNLOCK(txq);
2616 
2617 		/* clean any enqueued buffers */
2618 		iflib_ifmp_purge(txq);
2619 		/* Free any existing tx buffers. */
2620 		for (j = 0; j < txq->ift_size; j++) {
2621 			iflib_txsd_free(ctx, txq, j);
2622 		}
2623 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2624 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2625 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2626 			txq->ift_cidx = txq->ift_pidx;
2627 		else
2628 			txq->ift_cidx = txq->ift_pidx = 0;
2629 
2630 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2631 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2632 		txq->ift_pullups = 0;
2633 		ifmp_ring_reset_stats(txq->ift_br);
2634 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2635 			bzero((void *)di->idi_vaddr, di->idi_size);
2636 	}
2637 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2638 		gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2639 		    &rxq->ifr_task.gt_task);
2640 
2641 		rxq->ifr_cq_cidx = 0;
2642 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2643 			bzero((void *)di->idi_vaddr, di->idi_size);
2644 		/* also resets the free lists pidx/cidx */
2645 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2646 			iflib_fl_bufs_free(fl);
2647 	}
2648 }
2649 
2650 static inline caddr_t
2651 calc_next_rxd(iflib_fl_t fl, int cidx)
2652 {
2653 	qidx_t size;
2654 	int nrxd;
2655 	caddr_t start, end, cur, next;
2656 
2657 	nrxd = fl->ifl_size;
2658 	size = fl->ifl_rxd_size;
2659 	start = fl->ifl_ifdi->idi_vaddr;
2660 
2661 	if (__predict_false(size == 0))
2662 		return (start);
2663 	cur = start + size*cidx;
2664 	end = start + size*nrxd;
2665 	next = CACHE_PTR_NEXT(cur);
2666 	return (next < end ? next : start);
2667 }
2668 
2669 static inline void
2670 prefetch_pkts(iflib_fl_t fl, int cidx)
2671 {
2672 	int nextptr;
2673 	int nrxd = fl->ifl_size;
2674 	caddr_t next_rxd;
2675 
2676 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2677 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2678 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2679 	next_rxd = calc_next_rxd(fl, cidx);
2680 	prefetch(next_rxd);
2681 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2682 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2683 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2684 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2685 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2686 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2687 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2688 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2689 }
2690 
2691 static struct mbuf *
2692 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2693     int *pf_rv, if_rxd_info_t ri)
2694 {
2695 	bus_dmamap_t map;
2696 	iflib_fl_t fl;
2697 	caddr_t payload;
2698 	struct mbuf *m;
2699 	int flid, cidx, len, next;
2700 
2701 	map = NULL;
2702 	flid = irf->irf_flid;
2703 	cidx = irf->irf_idx;
2704 	fl = &rxq->ifr_fl[flid];
2705 	sd->ifsd_fl = fl;
2706 	m = fl->ifl_sds.ifsd_m[cidx];
2707 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2708 	fl->ifl_credits--;
2709 #if MEMORY_LOGGING
2710 	fl->ifl_m_dequeued++;
2711 #endif
2712 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2713 		prefetch_pkts(fl, cidx);
2714 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2715 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2716 	map = fl->ifl_sds.ifsd_map[cidx];
2717 
2718 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2719 
2720 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2721 	    irf->irf_len != 0) {
2722 		payload  = *sd->ifsd_cl;
2723 		payload +=  ri->iri_pad;
2724 		len = ri->iri_len - ri->iri_pad;
2725 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2726 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2727 		switch (*pf_rv) {
2728 		case PFIL_DROPPED:
2729 		case PFIL_CONSUMED:
2730 			/*
2731 			 * The filter ate it.  Everything is recycled.
2732 			 */
2733 			m = NULL;
2734 			unload = 0;
2735 			break;
2736 		case PFIL_REALLOCED:
2737 			/*
2738 			 * The filter copied it.  Everything is recycled.
2739 			 */
2740 			m = pfil_mem2mbuf(payload);
2741 			unload = 0;
2742 			break;
2743 		case PFIL_PASS:
2744 			/*
2745 			 * Filter said it was OK, so receive like
2746 			 * normal
2747 			 */
2748 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2749 			break;
2750 		default:
2751 			MPASS(0);
2752 		}
2753 	} else {
2754 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2755 		if (pf_rv != NULL)
2756 			*pf_rv = PFIL_PASS;
2757 	}
2758 
2759 	if (unload && irf->irf_len != 0)
2760 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2761 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2762 	if (__predict_false(fl->ifl_cidx == 0))
2763 		fl->ifl_gen = 0;
2764 	bit_clear(fl->ifl_rx_bitmap, cidx);
2765 	return (m);
2766 }
2767 
2768 static struct mbuf *
2769 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2770 {
2771 	struct mbuf *m, *mh, *mt;
2772 	caddr_t cl;
2773 	int  *pf_rv_ptr, flags, i, padlen;
2774 	bool consumed;
2775 
2776 	i = 0;
2777 	mh = NULL;
2778 	consumed = false;
2779 	*pf_rv = PFIL_PASS;
2780 	pf_rv_ptr = pf_rv;
2781 	do {
2782 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2783 		    pf_rv_ptr, ri);
2784 
2785 		MPASS(*sd->ifsd_cl != NULL);
2786 
2787 		/*
2788 		 * Exclude zero-length frags & frags from
2789 		 * packets the filter has consumed or dropped
2790 		 */
2791 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2792 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2793 			if (mh == NULL) {
2794 				/* everything saved here */
2795 				consumed = true;
2796 				pf_rv_ptr = NULL;
2797 				continue;
2798 			}
2799 			/* XXX we can save the cluster here, but not the mbuf */
2800 			m_init(m, M_NOWAIT, MT_DATA, 0);
2801 			m_free(m);
2802 			continue;
2803 		}
2804 		if (mh == NULL) {
2805 			flags = M_PKTHDR|M_EXT;
2806 			mh = mt = m;
2807 			padlen = ri->iri_pad;
2808 		} else {
2809 			flags = M_EXT;
2810 			mt->m_next = m;
2811 			mt = m;
2812 			/* assuming padding is only on the first fragment */
2813 			padlen = 0;
2814 		}
2815 		cl = *sd->ifsd_cl;
2816 		*sd->ifsd_cl = NULL;
2817 
2818 		/* Can these two be made one ? */
2819 		m_init(m, M_NOWAIT, MT_DATA, flags);
2820 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2821 		/*
2822 		 * These must follow m_init and m_cljset
2823 		 */
2824 		m->m_data += padlen;
2825 		ri->iri_len -= padlen;
2826 		m->m_len = ri->iri_frags[i].irf_len;
2827 	} while (++i < ri->iri_nfrags);
2828 
2829 	return (mh);
2830 }
2831 
2832 /*
2833  * Process one software descriptor
2834  */
2835 static struct mbuf *
2836 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2837 {
2838 	struct if_rxsd sd;
2839 	struct mbuf *m;
2840 	int pf_rv;
2841 
2842 	/* should I merge this back in now that the two paths are basically duplicated? */
2843 	if (ri->iri_nfrags == 1 &&
2844 	    ri->iri_frags[0].irf_len != 0 &&
2845 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2846 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2847 		    &pf_rv, ri);
2848 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2849 			return (m);
2850 		if (pf_rv == PFIL_PASS) {
2851 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2852 #ifndef __NO_STRICT_ALIGNMENT
2853 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2854 				m->m_data += 2;
2855 #endif
2856 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2857 			m->m_len = ri->iri_frags[0].irf_len;
2858 			m->m_data += ri->iri_pad;
2859 			ri->iri_len -= ri->iri_pad;
2860 		}
2861 	} else {
2862 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2863 		if (m == NULL)
2864 			return (NULL);
2865 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2866 			return (m);
2867 	}
2868 	m->m_pkthdr.len = ri->iri_len;
2869 	m->m_pkthdr.rcvif = ri->iri_ifp;
2870 	m->m_flags |= ri->iri_flags;
2871 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2872 	m->m_pkthdr.flowid = ri->iri_flowid;
2873 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2874 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2875 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2876 	return (m);
2877 }
2878 
2879 #if defined(INET6) || defined(INET)
2880 static void
2881 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2882 {
2883 	CURVNET_SET(lc->ifp->if_vnet);
2884 #if defined(INET6)
2885 	*v6 = V_ip6_forwarding;
2886 #endif
2887 #if defined(INET)
2888 	*v4 = V_ipforwarding;
2889 #endif
2890 	CURVNET_RESTORE();
2891 }
2892 
2893 /*
2894  * Returns true if it's possible this packet could be LROed.
2895  * if it returns false, it is guaranteed that tcp_lro_rx()
2896  * would not return zero.
2897  */
2898 static bool
2899 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2900 {
2901 	struct ether_header *eh;
2902 
2903 	eh = mtod(m, struct ether_header *);
2904 	switch (eh->ether_type) {
2905 #if defined(INET6)
2906 		case htons(ETHERTYPE_IPV6):
2907 			return (!v6_forwarding);
2908 #endif
2909 #if defined (INET)
2910 		case htons(ETHERTYPE_IP):
2911 			return (!v4_forwarding);
2912 #endif
2913 	}
2914 
2915 	return false;
2916 }
2917 #else
2918 static void
2919 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2920 {
2921 }
2922 #endif
2923 
2924 static void
2925 _task_fn_rx_watchdog(void *context)
2926 {
2927 	iflib_rxq_t rxq = context;
2928 
2929 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2930 }
2931 
2932 static uint8_t
2933 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2934 {
2935 	if_t ifp;
2936 	if_ctx_t ctx = rxq->ifr_ctx;
2937 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2938 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2939 	int avail, i;
2940 	qidx_t *cidxp;
2941 	struct if_rxd_info ri;
2942 	int err, budget_left, rx_bytes, rx_pkts;
2943 	iflib_fl_t fl;
2944 	int lro_enabled;
2945 	bool v4_forwarding, v6_forwarding, lro_possible;
2946 	uint8_t retval = 0;
2947 
2948 	/*
2949 	 * XXX early demux data packets so that if_input processing only handles
2950 	 * acks in interrupt context
2951 	 */
2952 	struct mbuf *m, *mh, *mt, *mf;
2953 
2954 	NET_EPOCH_ASSERT();
2955 
2956 	lro_possible = v4_forwarding = v6_forwarding = false;
2957 	ifp = ctx->ifc_ifp;
2958 	mh = mt = NULL;
2959 	MPASS(budget > 0);
2960 	rx_pkts	= rx_bytes = 0;
2961 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2962 		cidxp = &rxq->ifr_cq_cidx;
2963 	else
2964 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2965 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2966 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2967 			retval |= iflib_fl_refill_all(ctx, fl);
2968 		DBG_COUNTER_INC(rx_unavail);
2969 		return (retval);
2970 	}
2971 
2972 	/* pfil needs the vnet to be set */
2973 	CURVNET_SET_QUIET(ifp->if_vnet);
2974 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2975 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2976 			DBG_COUNTER_INC(rx_ctx_inactive);
2977 			break;
2978 		}
2979 		/*
2980 		 * Reset client set fields to their default values
2981 		 */
2982 		rxd_info_zero(&ri);
2983 		ri.iri_qsidx = rxq->ifr_id;
2984 		ri.iri_cidx = *cidxp;
2985 		ri.iri_ifp = ifp;
2986 		ri.iri_frags = rxq->ifr_frags;
2987 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2988 
2989 		if (err)
2990 			goto err;
2991 		rx_pkts += 1;
2992 		rx_bytes += ri.iri_len;
2993 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2994 			*cidxp = ri.iri_cidx;
2995 			/* Update our consumer index */
2996 			/* XXX NB: shurd - check if this is still safe */
2997 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2998 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2999 			/* was this only a completion queue message? */
3000 			if (__predict_false(ri.iri_nfrags == 0))
3001 				continue;
3002 		}
3003 		MPASS(ri.iri_nfrags != 0);
3004 		MPASS(ri.iri_len != 0);
3005 
3006 		/* will advance the cidx on the corresponding free lists */
3007 		m = iflib_rxd_pkt_get(rxq, &ri);
3008 		avail--;
3009 		budget_left--;
3010 		if (avail == 0 && budget_left)
3011 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3012 
3013 		if (__predict_false(m == NULL))
3014 			continue;
3015 
3016 		/* imm_pkt: -- cxgb */
3017 		if (mh == NULL)
3018 			mh = mt = m;
3019 		else {
3020 			mt->m_nextpkt = m;
3021 			mt = m;
3022 		}
3023 	}
3024 	CURVNET_RESTORE();
3025 	/* make sure that we can refill faster than drain */
3026 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3027 		retval |= iflib_fl_refill_all(ctx, fl);
3028 
3029 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3030 	if (lro_enabled)
3031 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3032 	mt = mf = NULL;
3033 	while (mh != NULL) {
3034 		m = mh;
3035 		mh = mh->m_nextpkt;
3036 		m->m_nextpkt = NULL;
3037 #ifndef __NO_STRICT_ALIGNMENT
3038 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3039 			continue;
3040 #endif
3041 #if defined(INET6) || defined(INET)
3042 		if (lro_enabled) {
3043 			if (!lro_possible) {
3044 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3045 				if (lro_possible && mf != NULL) {
3046 					ifp->if_input(ifp, mf);
3047 					DBG_COUNTER_INC(rx_if_input);
3048 					mt = mf = NULL;
3049 				}
3050 			}
3051 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3052 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3053 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3054 					continue;
3055 			}
3056 		}
3057 #endif
3058 		if (lro_possible) {
3059 			ifp->if_input(ifp, m);
3060 			DBG_COUNTER_INC(rx_if_input);
3061 			continue;
3062 		}
3063 
3064 		if (mf == NULL)
3065 			mf = m;
3066 		if (mt != NULL)
3067 			mt->m_nextpkt = m;
3068 		mt = m;
3069 	}
3070 	if (mf != NULL) {
3071 		ifp->if_input(ifp, mf);
3072 		DBG_COUNTER_INC(rx_if_input);
3073 	}
3074 
3075 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3076 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3077 
3078 	/*
3079 	 * Flush any outstanding LRO work
3080 	 */
3081 #if defined(INET6) || defined(INET)
3082 	tcp_lro_flush_all(&rxq->ifr_lc);
3083 #endif
3084 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3085 		retval |= IFLIB_RXEOF_MORE;
3086 	return (retval);
3087 err:
3088 	STATE_LOCK(ctx);
3089 	ctx->ifc_flags |= IFC_DO_RESET;
3090 	iflib_admin_intr_deferred(ctx);
3091 	STATE_UNLOCK(ctx);
3092 	return (0);
3093 }
3094 
3095 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3096 static inline qidx_t
3097 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3098 {
3099 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3100 	qidx_t minthresh = txq->ift_size / 8;
3101 	if (in_use > 4*minthresh)
3102 		return (notify_count);
3103 	if (in_use > 2*minthresh)
3104 		return (notify_count >> 1);
3105 	if (in_use > minthresh)
3106 		return (notify_count >> 3);
3107 	return (0);
3108 }
3109 
3110 static inline qidx_t
3111 txq_max_rs_deferred(iflib_txq_t txq)
3112 {
3113 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3114 	qidx_t minthresh = txq->ift_size / 8;
3115 	if (txq->ift_in_use > 4*minthresh)
3116 		return (notify_count);
3117 	if (txq->ift_in_use > 2*minthresh)
3118 		return (notify_count >> 1);
3119 	if (txq->ift_in_use > minthresh)
3120 		return (notify_count >> 2);
3121 	return (2);
3122 }
3123 
3124 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3125 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3126 
3127 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3128 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3129 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3130 
3131 /* forward compatibility for cxgb */
3132 #define FIRST_QSET(ctx) 0
3133 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3134 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3135 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3136 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3137 
3138 /* XXX we should be setting this to something other than zero */
3139 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3140 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3141     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3142 
3143 static inline bool
3144 iflib_txd_db_check(iflib_txq_t txq, int ring)
3145 {
3146 	if_ctx_t ctx = txq->ift_ctx;
3147 	qidx_t dbval, max;
3148 
3149 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3150 
3151 	/* force || threshold exceeded || at the edge of the ring */
3152 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3153 
3154 		/*
3155 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3156 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3157 		 * producer index explicitly (INTC).
3158 		 */
3159 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3160 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3161 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3162 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3163 
3164 		/*
3165 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3166 		 */
3167 		txq->ift_db_pending = txq->ift_npending = 0;
3168 		return (true);
3169 	}
3170 	return (false);
3171 }
3172 
3173 #ifdef PKT_DEBUG
3174 static void
3175 print_pkt(if_pkt_info_t pi)
3176 {
3177 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3178 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3179 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3180 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3181 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3182 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3183 }
3184 #endif
3185 
3186 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3187 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3188 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3189 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3190 
3191 static int
3192 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3193 {
3194 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3195 	struct ether_vlan_header *eh;
3196 	struct mbuf *m;
3197 
3198 	m = *mp;
3199 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3200 	    M_WRITABLE(m) == 0) {
3201 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3202 			return (ENOMEM);
3203 		} else {
3204 			m_freem(*mp);
3205 			DBG_COUNTER_INC(tx_frees);
3206 			*mp = m;
3207 		}
3208 	}
3209 
3210 	/*
3211 	 * Determine where frame payload starts.
3212 	 * Jump over vlan headers if already present,
3213 	 * helpful for QinQ too.
3214 	 */
3215 	if (__predict_false(m->m_len < sizeof(*eh))) {
3216 		txq->ift_pullups++;
3217 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3218 			return (ENOMEM);
3219 	}
3220 	eh = mtod(m, struct ether_vlan_header *);
3221 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3222 		pi->ipi_etype = ntohs(eh->evl_proto);
3223 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3224 	} else {
3225 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3226 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3227 	}
3228 
3229 	switch (pi->ipi_etype) {
3230 #ifdef INET
3231 	case ETHERTYPE_IP:
3232 	{
3233 		struct mbuf *n;
3234 		struct ip *ip = NULL;
3235 		struct tcphdr *th = NULL;
3236 		int minthlen;
3237 
3238 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3239 		if (__predict_false(m->m_len < minthlen)) {
3240 			/*
3241 			 * if this code bloat is causing too much of a hit
3242 			 * move it to a separate function and mark it noinline
3243 			 */
3244 			if (m->m_len == pi->ipi_ehdrlen) {
3245 				n = m->m_next;
3246 				MPASS(n);
3247 				if (n->m_len >= sizeof(*ip))  {
3248 					ip = (struct ip *)n->m_data;
3249 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3250 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3251 				} else {
3252 					txq->ift_pullups++;
3253 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3254 						return (ENOMEM);
3255 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3256 				}
3257 			} else {
3258 				txq->ift_pullups++;
3259 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3260 					return (ENOMEM);
3261 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3262 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3263 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3264 			}
3265 		} else {
3266 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3267 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3268 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3269 		}
3270 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3271 		pi->ipi_ipproto = ip->ip_p;
3272 		pi->ipi_flags |= IPI_TX_IPV4;
3273 
3274 		/* TCP checksum offload may require TCP header length */
3275 		if (IS_TX_OFFLOAD4(pi)) {
3276 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3277 				if (__predict_false(th == NULL)) {
3278 					txq->ift_pullups++;
3279 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3280 						return (ENOMEM);
3281 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3282 				}
3283 				pi->ipi_tcp_hflags = th->th_flags;
3284 				pi->ipi_tcp_hlen = th->th_off << 2;
3285 				pi->ipi_tcp_seq = th->th_seq;
3286 			}
3287 			if (IS_TSO4(pi)) {
3288 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3289 					return (ENXIO);
3290 				/*
3291 				 * TSO always requires hardware checksum offload.
3292 				 */
3293 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3294 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3295 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3296 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3297 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3298 					ip->ip_sum = 0;
3299 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3300 				}
3301 			}
3302 		}
3303 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3304                        ip->ip_sum = 0;
3305 
3306 		break;
3307 	}
3308 #endif
3309 #ifdef INET6
3310 	case ETHERTYPE_IPV6:
3311 	{
3312 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3313 		struct tcphdr *th;
3314 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3315 
3316 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3317 			txq->ift_pullups++;
3318 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3319 				return (ENOMEM);
3320 		}
3321 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3322 
3323 		/* XXX-BZ this will go badly in case of ext hdrs. */
3324 		pi->ipi_ipproto = ip6->ip6_nxt;
3325 		pi->ipi_flags |= IPI_TX_IPV6;
3326 
3327 		/* TCP checksum offload may require TCP header length */
3328 		if (IS_TX_OFFLOAD6(pi)) {
3329 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3330 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3331 					txq->ift_pullups++;
3332 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3333 						return (ENOMEM);
3334 				}
3335 				pi->ipi_tcp_hflags = th->th_flags;
3336 				pi->ipi_tcp_hlen = th->th_off << 2;
3337 				pi->ipi_tcp_seq = th->th_seq;
3338 			}
3339 			if (IS_TSO6(pi)) {
3340 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3341 					return (ENXIO);
3342 				/*
3343 				 * TSO always requires hardware checksum offload.
3344 				 */
3345 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3346 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3347 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3348 			}
3349 		}
3350 		break;
3351 	}
3352 #endif
3353 	default:
3354 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3355 		pi->ipi_ip_hlen = 0;
3356 		break;
3357 	}
3358 	*mp = m;
3359 
3360 	return (0);
3361 }
3362 
3363 /*
3364  * If dodgy hardware rejects the scatter gather chain we've handed it
3365  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3366  * m_defrag'd mbufs
3367  */
3368 static __noinline struct mbuf *
3369 iflib_remove_mbuf(iflib_txq_t txq)
3370 {
3371 	int ntxd, pidx;
3372 	struct mbuf *m, **ifsd_m;
3373 
3374 	ifsd_m = txq->ift_sds.ifsd_m;
3375 	ntxd = txq->ift_size;
3376 	pidx = txq->ift_pidx & (ntxd - 1);
3377 	ifsd_m = txq->ift_sds.ifsd_m;
3378 	m = ifsd_m[pidx];
3379 	ifsd_m[pidx] = NULL;
3380 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3381 	if (txq->ift_sds.ifsd_tso_map != NULL)
3382 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3383 		    txq->ift_sds.ifsd_tso_map[pidx]);
3384 #if MEMORY_LOGGING
3385 	txq->ift_dequeued++;
3386 #endif
3387 	return (m);
3388 }
3389 
3390 static inline caddr_t
3391 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3392 {
3393 	qidx_t size;
3394 	int ntxd;
3395 	caddr_t start, end, cur, next;
3396 
3397 	ntxd = txq->ift_size;
3398 	size = txq->ift_txd_size[qid];
3399 	start = txq->ift_ifdi[qid].idi_vaddr;
3400 
3401 	if (__predict_false(size == 0))
3402 		return (start);
3403 	cur = start + size*cidx;
3404 	end = start + size*ntxd;
3405 	next = CACHE_PTR_NEXT(cur);
3406 	return (next < end ? next : start);
3407 }
3408 
3409 /*
3410  * Pad an mbuf to ensure a minimum ethernet frame size.
3411  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3412  */
3413 static __noinline int
3414 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3415 {
3416 	/*
3417 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3418 	 * and ARP message is the smallest common payload I can think of
3419 	 */
3420 	static char pad[18];	/* just zeros */
3421 	int n;
3422 	struct mbuf *new_head;
3423 
3424 	if (!M_WRITABLE(*m_head)) {
3425 		new_head = m_dup(*m_head, M_NOWAIT);
3426 		if (new_head == NULL) {
3427 			m_freem(*m_head);
3428 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3429 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3430 			DBG_COUNTER_INC(tx_frees);
3431 			return ENOMEM;
3432 		}
3433 		m_freem(*m_head);
3434 		*m_head = new_head;
3435 	}
3436 
3437 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3438 	     n > 0; n -= sizeof(pad))
3439 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3440 			break;
3441 
3442 	if (n > 0) {
3443 		m_freem(*m_head);
3444 		device_printf(dev, "cannot pad short frame\n");
3445 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3446 		DBG_COUNTER_INC(tx_frees);
3447 		return (ENOBUFS);
3448 	}
3449 
3450 	return 0;
3451 }
3452 
3453 static int
3454 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3455 {
3456 	if_ctx_t		ctx;
3457 	if_shared_ctx_t		sctx;
3458 	if_softc_ctx_t		scctx;
3459 	bus_dma_tag_t		buf_tag;
3460 	bus_dma_segment_t	*segs;
3461 	struct mbuf		*m_head, **ifsd_m;
3462 	void			*next_txd;
3463 	bus_dmamap_t		map;
3464 	struct if_pkt_info	pi;
3465 	int remap = 0;
3466 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3467 
3468 	ctx = txq->ift_ctx;
3469 	sctx = ctx->ifc_sctx;
3470 	scctx = &ctx->ifc_softc_ctx;
3471 	segs = txq->ift_segs;
3472 	ntxd = txq->ift_size;
3473 	m_head = *m_headp;
3474 	map = NULL;
3475 
3476 	/*
3477 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3478 	 */
3479 	cidx = txq->ift_cidx;
3480 	pidx = txq->ift_pidx;
3481 	if (ctx->ifc_flags & IFC_PREFETCH) {
3482 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3483 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3484 			next_txd = calc_next_txd(txq, cidx, 0);
3485 			prefetch(next_txd);
3486 		}
3487 
3488 		/* prefetch the next cache line of mbuf pointers and flags */
3489 		prefetch(&txq->ift_sds.ifsd_m[next]);
3490 		prefetch(&txq->ift_sds.ifsd_map[next]);
3491 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3492 	}
3493 	map = txq->ift_sds.ifsd_map[pidx];
3494 	ifsd_m = txq->ift_sds.ifsd_m;
3495 
3496 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3497 		buf_tag = txq->ift_tso_buf_tag;
3498 		max_segs = scctx->isc_tx_tso_segments_max;
3499 		map = txq->ift_sds.ifsd_tso_map[pidx];
3500 		MPASS(buf_tag != NULL);
3501 		MPASS(max_segs > 0);
3502 	} else {
3503 		buf_tag = txq->ift_buf_tag;
3504 		max_segs = scctx->isc_tx_nsegments;
3505 		map = txq->ift_sds.ifsd_map[pidx];
3506 	}
3507 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3508 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3509 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3510 		if (err) {
3511 			DBG_COUNTER_INC(encap_txd_encap_fail);
3512 			return err;
3513 		}
3514 	}
3515 	m_head = *m_headp;
3516 
3517 	pkt_info_zero(&pi);
3518 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3519 	pi.ipi_pidx = pidx;
3520 	pi.ipi_qsidx = txq->ift_id;
3521 	pi.ipi_len = m_head->m_pkthdr.len;
3522 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3523 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3524 
3525 	/* deliberate bitwise OR to make one condition */
3526 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3527 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3528 			DBG_COUNTER_INC(encap_txd_encap_fail);
3529 			return (err);
3530 		}
3531 		m_head = *m_headp;
3532 	}
3533 
3534 retry:
3535 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3536 	    BUS_DMA_NOWAIT);
3537 defrag:
3538 	if (__predict_false(err)) {
3539 		switch (err) {
3540 		case EFBIG:
3541 			/* try collapse once and defrag once */
3542 			if (remap == 0) {
3543 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3544 				/* try defrag if collapsing fails */
3545 				if (m_head == NULL)
3546 					remap++;
3547 			}
3548 			if (remap == 1) {
3549 				txq->ift_mbuf_defrag++;
3550 				m_head = m_defrag(*m_headp, M_NOWAIT);
3551 			}
3552 			/*
3553 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3554 			 * failed to map an mbuf that was run through m_defrag
3555 			 */
3556 			MPASS(remap <= 1);
3557 			if (__predict_false(m_head == NULL || remap > 1))
3558 				goto defrag_failed;
3559 			remap++;
3560 			*m_headp = m_head;
3561 			goto retry;
3562 			break;
3563 		case ENOMEM:
3564 			txq->ift_no_tx_dma_setup++;
3565 			break;
3566 		default:
3567 			txq->ift_no_tx_dma_setup++;
3568 			m_freem(*m_headp);
3569 			DBG_COUNTER_INC(tx_frees);
3570 			*m_headp = NULL;
3571 			break;
3572 		}
3573 		txq->ift_map_failed++;
3574 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3575 		DBG_COUNTER_INC(encap_txd_encap_fail);
3576 		return (err);
3577 	}
3578 	ifsd_m[pidx] = m_head;
3579 	/*
3580 	 * XXX assumes a 1 to 1 relationship between segments and
3581 	 *        descriptors - this does not hold true on all drivers, e.g.
3582 	 *        cxgb
3583 	 */
3584 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3585 		txq->ift_no_desc_avail++;
3586 		bus_dmamap_unload(buf_tag, map);
3587 		DBG_COUNTER_INC(encap_txq_avail_fail);
3588 		DBG_COUNTER_INC(encap_txd_encap_fail);
3589 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3590 			GROUPTASK_ENQUEUE(&txq->ift_task);
3591 		return (ENOBUFS);
3592 	}
3593 	/*
3594 	 * On Intel cards we can greatly reduce the number of TX interrupts
3595 	 * we see by only setting report status on every Nth descriptor.
3596 	 * However, this also means that the driver will need to keep track
3597 	 * of the descriptors that RS was set on to check them for the DD bit.
3598 	 */
3599 	txq->ift_rs_pending += nsegs + 1;
3600 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3601 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3602 		pi.ipi_flags |= IPI_TX_INTR;
3603 		txq->ift_rs_pending = 0;
3604 	}
3605 
3606 	pi.ipi_segs = segs;
3607 	pi.ipi_nsegs = nsegs;
3608 
3609 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3610 #ifdef PKT_DEBUG
3611 	print_pkt(&pi);
3612 #endif
3613 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3614 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3615 		DBG_COUNTER_INC(tx_encap);
3616 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3617 
3618 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3619 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3620 			ndesc += txq->ift_size;
3621 			txq->ift_gen = 1;
3622 		}
3623 		/*
3624 		 * drivers can need as many as
3625 		 * two sentinels
3626 		 */
3627 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3628 		MPASS(pi.ipi_new_pidx != pidx);
3629 		MPASS(ndesc > 0);
3630 		txq->ift_in_use += ndesc;
3631 		txq->ift_db_pending += ndesc;
3632 
3633 		/*
3634 		 * We update the last software descriptor again here because there may
3635 		 * be a sentinel and/or there may be more mbufs than segments
3636 		 */
3637 		txq->ift_pidx = pi.ipi_new_pidx;
3638 		txq->ift_npending += pi.ipi_ndescs;
3639 	} else {
3640 		*m_headp = m_head = iflib_remove_mbuf(txq);
3641 		if (err == EFBIG) {
3642 			txq->ift_txd_encap_efbig++;
3643 			if (remap < 2) {
3644 				remap = 1;
3645 				goto defrag;
3646 			}
3647 		}
3648 		goto defrag_failed;
3649 	}
3650 	/*
3651 	 * err can't possibly be non-zero here, so we don't neet to test it
3652 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3653 	 */
3654 	return (err);
3655 
3656 defrag_failed:
3657 	txq->ift_mbuf_defrag_failed++;
3658 	txq->ift_map_failed++;
3659 	m_freem(*m_headp);
3660 	DBG_COUNTER_INC(tx_frees);
3661 	*m_headp = NULL;
3662 	DBG_COUNTER_INC(encap_txd_encap_fail);
3663 	return (ENOMEM);
3664 }
3665 
3666 static void
3667 iflib_tx_desc_free(iflib_txq_t txq, int n)
3668 {
3669 	uint32_t qsize, cidx, mask, gen;
3670 	struct mbuf *m, **ifsd_m;
3671 	bool do_prefetch;
3672 
3673 	cidx = txq->ift_cidx;
3674 	gen = txq->ift_gen;
3675 	qsize = txq->ift_size;
3676 	mask = qsize-1;
3677 	ifsd_m = txq->ift_sds.ifsd_m;
3678 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3679 
3680 	while (n-- > 0) {
3681 		if (do_prefetch) {
3682 			prefetch(ifsd_m[(cidx + 3) & mask]);
3683 			prefetch(ifsd_m[(cidx + 4) & mask]);
3684 		}
3685 		if ((m = ifsd_m[cidx]) != NULL) {
3686 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3687 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3688 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3689 				    txq->ift_sds.ifsd_tso_map[cidx],
3690 				    BUS_DMASYNC_POSTWRITE);
3691 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3692 				    txq->ift_sds.ifsd_tso_map[cidx]);
3693 			} else {
3694 				bus_dmamap_sync(txq->ift_buf_tag,
3695 				    txq->ift_sds.ifsd_map[cidx],
3696 				    BUS_DMASYNC_POSTWRITE);
3697 				bus_dmamap_unload(txq->ift_buf_tag,
3698 				    txq->ift_sds.ifsd_map[cidx]);
3699 			}
3700 			/* XXX we don't support any drivers that batch packets yet */
3701 			MPASS(m->m_nextpkt == NULL);
3702 			m_freem(m);
3703 			ifsd_m[cidx] = NULL;
3704 #if MEMORY_LOGGING
3705 			txq->ift_dequeued++;
3706 #endif
3707 			DBG_COUNTER_INC(tx_frees);
3708 		}
3709 		if (__predict_false(++cidx == qsize)) {
3710 			cidx = 0;
3711 			gen = 0;
3712 		}
3713 	}
3714 	txq->ift_cidx = cidx;
3715 	txq->ift_gen = gen;
3716 }
3717 
3718 static __inline int
3719 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3720 {
3721 	int reclaim;
3722 	if_ctx_t ctx = txq->ift_ctx;
3723 
3724 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3725 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3726 
3727 	/*
3728 	 * Need a rate-limiting check so that this isn't called every time
3729 	 */
3730 	iflib_tx_credits_update(ctx, txq);
3731 	reclaim = DESC_RECLAIMABLE(txq);
3732 
3733 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3734 #ifdef INVARIANTS
3735 		if (iflib_verbose_debug) {
3736 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3737 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3738 			       reclaim, thresh);
3739 		}
3740 #endif
3741 		return (0);
3742 	}
3743 	iflib_tx_desc_free(txq, reclaim);
3744 	txq->ift_cleaned += reclaim;
3745 	txq->ift_in_use -= reclaim;
3746 
3747 	return (reclaim);
3748 }
3749 
3750 static struct mbuf **
3751 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3752 {
3753 	int next, size;
3754 	struct mbuf **items;
3755 
3756 	size = r->size;
3757 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3758 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3759 
3760 	prefetch(items[(cidx + offset) & (size-1)]);
3761 	if (remaining > 1) {
3762 		prefetch2cachelines(&items[next]);
3763 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3764 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3765 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3766 	}
3767 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3768 }
3769 
3770 static void
3771 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3772 {
3773 
3774 	ifmp_ring_check_drainage(txq->ift_br, budget);
3775 }
3776 
3777 static uint32_t
3778 iflib_txq_can_drain(struct ifmp_ring *r)
3779 {
3780 	iflib_txq_t txq = r->cookie;
3781 	if_ctx_t ctx = txq->ift_ctx;
3782 
3783 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3784 		return (1);
3785 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3786 	    BUS_DMASYNC_POSTREAD);
3787 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3788 	    false));
3789 }
3790 
3791 static uint32_t
3792 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3793 {
3794 	iflib_txq_t txq = r->cookie;
3795 	if_ctx_t ctx = txq->ift_ctx;
3796 	if_t ifp = ctx->ifc_ifp;
3797 	struct mbuf *m, **mp;
3798 	int avail, bytes_sent, skipped, count, err, i;
3799 	int mcast_sent, pkt_sent, reclaimed;
3800 	bool do_prefetch, rang, ring;
3801 
3802 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3803 			    !LINK_ACTIVE(ctx))) {
3804 		DBG_COUNTER_INC(txq_drain_notready);
3805 		return (0);
3806 	}
3807 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3808 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3809 	avail = IDXDIFF(pidx, cidx, r->size);
3810 
3811 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3812 		/*
3813 		 * The driver is unloading so we need to free all pending packets.
3814 		 */
3815 		DBG_COUNTER_INC(txq_drain_flushing);
3816 		for (i = 0; i < avail; i++) {
3817 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3818 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3819 			r->items[(cidx + i) & (r->size-1)] = NULL;
3820 		}
3821 		return (avail);
3822 	}
3823 
3824 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3825 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3826 		CALLOUT_LOCK(txq);
3827 		callout_stop(&txq->ift_timer);
3828 		CALLOUT_UNLOCK(txq);
3829 		DBG_COUNTER_INC(txq_drain_oactive);
3830 		return (0);
3831 	}
3832 
3833 	/*
3834 	 * If we've reclaimed any packets this queue cannot be hung.
3835 	 */
3836 	if (reclaimed)
3837 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3838 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3839 	count = MIN(avail, TX_BATCH_SIZE);
3840 #ifdef INVARIANTS
3841 	if (iflib_verbose_debug)
3842 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3843 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3844 #endif
3845 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3846 	err = 0;
3847 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3848 		int rem = do_prefetch ? count - i : 0;
3849 
3850 		mp = _ring_peek_one(r, cidx, i, rem);
3851 		MPASS(mp != NULL && *mp != NULL);
3852 
3853 		/*
3854 		 * Completion interrupts will use the address of the txq
3855 		 * as a sentinel to enqueue _something_ in order to acquire
3856 		 * the lock on the mp_ring (there's no direct lock call).
3857 		 * We obviously whave to check for these sentinel cases
3858 		 * and skip them.
3859 		 */
3860 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3861 			skipped++;
3862 			continue;
3863 		}
3864 		err = iflib_encap(txq, mp);
3865 		if (__predict_false(err)) {
3866 			/* no room - bail out */
3867 			if (err == ENOBUFS)
3868 				break;
3869 			skipped++;
3870 			/* we can't send this packet - skip it */
3871 			continue;
3872 		}
3873 		pkt_sent++;
3874 		m = *mp;
3875 		DBG_COUNTER_INC(tx_sent);
3876 		bytes_sent += m->m_pkthdr.len;
3877 		mcast_sent += !!(m->m_flags & M_MCAST);
3878 
3879 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3880 			break;
3881 		ETHER_BPF_MTAP(ifp, m);
3882 		rang = iflib_txd_db_check(txq, false);
3883 	}
3884 
3885 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3886 	ring = rang ? false  : (iflib_min_tx_latency | err);
3887 	iflib_txd_db_check(txq, ring);
3888 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3889 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3890 	if (mcast_sent)
3891 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3892 #ifdef INVARIANTS
3893 	if (iflib_verbose_debug)
3894 		printf("consumed=%d\n", skipped + pkt_sent);
3895 #endif
3896 	return (skipped + pkt_sent);
3897 }
3898 
3899 static uint32_t
3900 iflib_txq_drain_always(struct ifmp_ring *r)
3901 {
3902 	return (1);
3903 }
3904 
3905 static uint32_t
3906 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3907 {
3908 	int i, avail;
3909 	struct mbuf **mp;
3910 	iflib_txq_t txq;
3911 
3912 	txq = r->cookie;
3913 
3914 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3915 	CALLOUT_LOCK(txq);
3916 	callout_stop(&txq->ift_timer);
3917 	CALLOUT_UNLOCK(txq);
3918 
3919 	avail = IDXDIFF(pidx, cidx, r->size);
3920 	for (i = 0; i < avail; i++) {
3921 		mp = _ring_peek_one(r, cidx, i, avail - i);
3922 		if (__predict_false(*mp == (struct mbuf *)txq))
3923 			continue;
3924 		m_freem(*mp);
3925 		DBG_COUNTER_INC(tx_frees);
3926 	}
3927 	MPASS(ifmp_ring_is_stalled(r) == 0);
3928 	return (avail);
3929 }
3930 
3931 static void
3932 iflib_ifmp_purge(iflib_txq_t txq)
3933 {
3934 	struct ifmp_ring *r;
3935 
3936 	r = txq->ift_br;
3937 	r->drain = iflib_txq_drain_free;
3938 	r->can_drain = iflib_txq_drain_always;
3939 
3940 	ifmp_ring_check_drainage(r, r->size);
3941 
3942 	r->drain = iflib_txq_drain;
3943 	r->can_drain = iflib_txq_can_drain;
3944 }
3945 
3946 static void
3947 _task_fn_tx(void *context)
3948 {
3949 	iflib_txq_t txq = context;
3950 	if_ctx_t ctx = txq->ift_ctx;
3951 	if_t ifp = ctx->ifc_ifp;
3952 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3953 
3954 #ifdef IFLIB_DIAGNOSTICS
3955 	txq->ift_cpu_exec_count[curcpu]++;
3956 #endif
3957 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3958 		return;
3959 #ifdef DEV_NETMAP
3960 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3961 	    netmap_tx_irq(ifp, txq->ift_id))
3962 		goto skip_ifmp;
3963 #endif
3964 #ifdef ALTQ
3965 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3966 		iflib_altq_if_start(ifp);
3967 #endif
3968 	if (txq->ift_db_pending)
3969 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3970 	else if (!abdicate)
3971 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3972 	/*
3973 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3974 	 */
3975 	if (abdicate)
3976 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3977 #ifdef DEV_NETMAP
3978 skip_ifmp:
3979 #endif
3980 	if (ctx->ifc_flags & IFC_LEGACY)
3981 		IFDI_INTR_ENABLE(ctx);
3982 	else
3983 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3984 }
3985 
3986 static void
3987 _task_fn_rx(void *context)
3988 {
3989 	iflib_rxq_t rxq = context;
3990 	if_ctx_t ctx = rxq->ifr_ctx;
3991 	uint8_t more;
3992 	uint16_t budget;
3993 #ifdef DEV_NETMAP
3994 	u_int work = 0;
3995 	int nmirq;
3996 #endif
3997 
3998 #ifdef IFLIB_DIAGNOSTICS
3999 	rxq->ifr_cpu_exec_count[curcpu]++;
4000 #endif
4001 	DBG_COUNTER_INC(task_fn_rxs);
4002 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4003 		return;
4004 #ifdef DEV_NETMAP
4005 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4006 	if (nmirq != NM_IRQ_PASS) {
4007 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4008 		goto skip_rxeof;
4009 	}
4010 #endif
4011 	budget = ctx->ifc_sysctl_rx_budget;
4012 	if (budget == 0)
4013 		budget = 16;	/* XXX */
4014 	more = iflib_rxeof(rxq, budget);
4015 #ifdef DEV_NETMAP
4016 skip_rxeof:
4017 #endif
4018 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4019 		if (ctx->ifc_flags & IFC_LEGACY)
4020 			IFDI_INTR_ENABLE(ctx);
4021 		else
4022 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4023 		DBG_COUNTER_INC(rx_intr_enables);
4024 	}
4025 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4026 		return;
4027 
4028 	if (more & IFLIB_RXEOF_MORE)
4029 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4030 	else if (more & IFLIB_RXEOF_EMPTY)
4031 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4032 }
4033 
4034 static void
4035 _task_fn_admin(void *context)
4036 {
4037 	if_ctx_t ctx = context;
4038 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4039 	iflib_txq_t txq;
4040 	int i;
4041 	bool oactive, running, do_reset, do_watchdog, in_detach;
4042 
4043 	STATE_LOCK(ctx);
4044 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4045 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4046 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4047 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4048 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4049 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4050 	STATE_UNLOCK(ctx);
4051 
4052 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4053 		return;
4054 	if (in_detach)
4055 		return;
4056 
4057 	CTX_LOCK(ctx);
4058 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4059 		CALLOUT_LOCK(txq);
4060 		callout_stop(&txq->ift_timer);
4061 		CALLOUT_UNLOCK(txq);
4062 	}
4063 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4064 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4065 	if (do_watchdog) {
4066 		ctx->ifc_watchdog_events++;
4067 		IFDI_WATCHDOG_RESET(ctx);
4068 	}
4069 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4070 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4071 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4072 		    txq->ift_timer.c_cpu);
4073 	}
4074 	IFDI_LINK_INTR_ENABLE(ctx);
4075 	if (do_reset)
4076 		iflib_if_init_locked(ctx);
4077 	CTX_UNLOCK(ctx);
4078 
4079 	if (LINK_ACTIVE(ctx) == 0)
4080 		return;
4081 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4082 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4083 }
4084 
4085 static void
4086 _task_fn_iov(void *context)
4087 {
4088 	if_ctx_t ctx = context;
4089 
4090 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4091 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4092 		return;
4093 
4094 	CTX_LOCK(ctx);
4095 	IFDI_VFLR_HANDLE(ctx);
4096 	CTX_UNLOCK(ctx);
4097 }
4098 
4099 static int
4100 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4101 {
4102 	int err;
4103 	if_int_delay_info_t info;
4104 	if_ctx_t ctx;
4105 
4106 	info = (if_int_delay_info_t)arg1;
4107 	ctx = info->iidi_ctx;
4108 	info->iidi_req = req;
4109 	info->iidi_oidp = oidp;
4110 	CTX_LOCK(ctx);
4111 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4112 	CTX_UNLOCK(ctx);
4113 	return (err);
4114 }
4115 
4116 /*********************************************************************
4117  *
4118  *  IFNET FUNCTIONS
4119  *
4120  **********************************************************************/
4121 
4122 static void
4123 iflib_if_init_locked(if_ctx_t ctx)
4124 {
4125 	iflib_stop(ctx);
4126 	iflib_init_locked(ctx);
4127 }
4128 
4129 static void
4130 iflib_if_init(void *arg)
4131 {
4132 	if_ctx_t ctx = arg;
4133 
4134 	CTX_LOCK(ctx);
4135 	iflib_if_init_locked(ctx);
4136 	CTX_UNLOCK(ctx);
4137 }
4138 
4139 static int
4140 iflib_if_transmit(if_t ifp, struct mbuf *m)
4141 {
4142 	if_ctx_t	ctx = if_getsoftc(ifp);
4143 
4144 	iflib_txq_t txq;
4145 	int err, qidx;
4146 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4147 
4148 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4149 		DBG_COUNTER_INC(tx_frees);
4150 		m_freem(m);
4151 		return (ENETDOWN);
4152 	}
4153 
4154 	MPASS(m->m_nextpkt == NULL);
4155 	/* ALTQ-enabled interfaces always use queue 0. */
4156 	qidx = 0;
4157 	/* Use driver-supplied queue selection method if it exists */
4158 	if (ctx->isc_txq_select)
4159 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4160 	/* If not, use iflib's standard method */
4161 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4162 		qidx = QIDX(ctx, m);
4163 
4164 	/* Set TX queue */
4165 	txq = &ctx->ifc_txqs[qidx];
4166 
4167 #ifdef DRIVER_BACKPRESSURE
4168 	if (txq->ift_closed) {
4169 		while (m != NULL) {
4170 			next = m->m_nextpkt;
4171 			m->m_nextpkt = NULL;
4172 			m_freem(m);
4173 			DBG_COUNTER_INC(tx_frees);
4174 			m = next;
4175 		}
4176 		return (ENOBUFS);
4177 	}
4178 #endif
4179 #ifdef notyet
4180 	qidx = count = 0;
4181 	mp = marr;
4182 	next = m;
4183 	do {
4184 		count++;
4185 		next = next->m_nextpkt;
4186 	} while (next != NULL);
4187 
4188 	if (count > nitems(marr))
4189 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4190 			/* XXX check nextpkt */
4191 			m_freem(m);
4192 			/* XXX simplify for now */
4193 			DBG_COUNTER_INC(tx_frees);
4194 			return (ENOBUFS);
4195 		}
4196 	for (next = m, i = 0; next != NULL; i++) {
4197 		mp[i] = next;
4198 		next = next->m_nextpkt;
4199 		mp[i]->m_nextpkt = NULL;
4200 	}
4201 #endif
4202 	DBG_COUNTER_INC(tx_seen);
4203 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4204 
4205 	if (abdicate)
4206 		GROUPTASK_ENQUEUE(&txq->ift_task);
4207  	if (err) {
4208 		if (!abdicate)
4209 			GROUPTASK_ENQUEUE(&txq->ift_task);
4210 		/* support forthcoming later */
4211 #ifdef DRIVER_BACKPRESSURE
4212 		txq->ift_closed = TRUE;
4213 #endif
4214 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4215 		m_freem(m);
4216 		DBG_COUNTER_INC(tx_frees);
4217 	}
4218 
4219 	return (err);
4220 }
4221 
4222 #ifdef ALTQ
4223 /*
4224  * The overall approach to integrating iflib with ALTQ is to continue to use
4225  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4226  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4227  * is redundant/unnecessary, but doing so minimizes the amount of
4228  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4229  * redundantly queueing to an intermediate mp_ring is swamped by the
4230  * performance limitations inherent in using ALTQ.
4231  *
4232  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4233  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4234  * given interface.  If ALTQ is enabled for an interface, then all
4235  * transmitted packets for that interface will be submitted to the ALTQ
4236  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4237  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4238  * update stats that the iflib machinery handles, and which is sensitve to
4239  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4240  * will be installed as the start routine for use by ALTQ facilities that
4241  * need to trigger queue drains on a scheduled basis.
4242  *
4243  */
4244 static void
4245 iflib_altq_if_start(if_t ifp)
4246 {
4247 	struct ifaltq *ifq = &ifp->if_snd;
4248 	struct mbuf *m;
4249 
4250 	IFQ_LOCK(ifq);
4251 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4252 	while (m != NULL) {
4253 		iflib_if_transmit(ifp, m);
4254 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4255 	}
4256 	IFQ_UNLOCK(ifq);
4257 }
4258 
4259 static int
4260 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4261 {
4262 	int err;
4263 
4264 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4265 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4266 		if (err == 0)
4267 			iflib_altq_if_start(ifp);
4268 	} else
4269 		err = iflib_if_transmit(ifp, m);
4270 
4271 	return (err);
4272 }
4273 #endif /* ALTQ */
4274 
4275 static void
4276 iflib_if_qflush(if_t ifp)
4277 {
4278 	if_ctx_t ctx = if_getsoftc(ifp);
4279 	iflib_txq_t txq = ctx->ifc_txqs;
4280 	int i;
4281 
4282 	STATE_LOCK(ctx);
4283 	ctx->ifc_flags |= IFC_QFLUSH;
4284 	STATE_UNLOCK(ctx);
4285 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4286 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4287 			iflib_txq_check_drain(txq, 0);
4288 	STATE_LOCK(ctx);
4289 	ctx->ifc_flags &= ~IFC_QFLUSH;
4290 	STATE_UNLOCK(ctx);
4291 
4292 	/*
4293 	 * When ALTQ is enabled, this will also take care of purging the
4294 	 * ALTQ queue(s).
4295 	 */
4296 	if_qflush(ifp);
4297 }
4298 
4299 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4300 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4301 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4302 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4303 
4304 static int
4305 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4306 {
4307 	if_ctx_t ctx = if_getsoftc(ifp);
4308 	struct ifreq	*ifr = (struct ifreq *)data;
4309 #if defined(INET) || defined(INET6)
4310 	struct ifaddr	*ifa = (struct ifaddr *)data;
4311 #endif
4312 	bool		avoid_reset = false;
4313 	int		err = 0, reinit = 0, bits;
4314 
4315 	switch (command) {
4316 	case SIOCSIFADDR:
4317 #ifdef INET
4318 		if (ifa->ifa_addr->sa_family == AF_INET)
4319 			avoid_reset = true;
4320 #endif
4321 #ifdef INET6
4322 		if (ifa->ifa_addr->sa_family == AF_INET6)
4323 			avoid_reset = true;
4324 #endif
4325 		/*
4326 		** Calling init results in link renegotiation,
4327 		** so we avoid doing it when possible.
4328 		*/
4329 		if (avoid_reset) {
4330 			if_setflagbits(ifp, IFF_UP,0);
4331 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4332 				reinit = 1;
4333 #ifdef INET
4334 			if (!(if_getflags(ifp) & IFF_NOARP))
4335 				arp_ifinit(ifp, ifa);
4336 #endif
4337 		} else
4338 			err = ether_ioctl(ifp, command, data);
4339 		break;
4340 	case SIOCSIFMTU:
4341 		CTX_LOCK(ctx);
4342 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4343 			CTX_UNLOCK(ctx);
4344 			break;
4345 		}
4346 		bits = if_getdrvflags(ifp);
4347 		/* stop the driver and free any clusters before proceeding */
4348 		iflib_stop(ctx);
4349 
4350 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4351 			STATE_LOCK(ctx);
4352 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4353 				ctx->ifc_flags |= IFC_MULTISEG;
4354 			else
4355 				ctx->ifc_flags &= ~IFC_MULTISEG;
4356 			STATE_UNLOCK(ctx);
4357 			err = if_setmtu(ifp, ifr->ifr_mtu);
4358 		}
4359 		iflib_init_locked(ctx);
4360 		STATE_LOCK(ctx);
4361 		if_setdrvflags(ifp, bits);
4362 		STATE_UNLOCK(ctx);
4363 		CTX_UNLOCK(ctx);
4364 		break;
4365 	case SIOCSIFFLAGS:
4366 		CTX_LOCK(ctx);
4367 		if (if_getflags(ifp) & IFF_UP) {
4368 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4369 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4370 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4371 					CTX_UNLOCK(ctx);
4372 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4373 					CTX_LOCK(ctx);
4374 				}
4375 			} else
4376 				reinit = 1;
4377 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4378 			iflib_stop(ctx);
4379 		}
4380 		ctx->ifc_if_flags = if_getflags(ifp);
4381 		CTX_UNLOCK(ctx);
4382 		break;
4383 	case SIOCADDMULTI:
4384 	case SIOCDELMULTI:
4385 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4386 			CTX_LOCK(ctx);
4387 			IFDI_INTR_DISABLE(ctx);
4388 			IFDI_MULTI_SET(ctx);
4389 			IFDI_INTR_ENABLE(ctx);
4390 			CTX_UNLOCK(ctx);
4391 		}
4392 		break;
4393 	case SIOCSIFMEDIA:
4394 		CTX_LOCK(ctx);
4395 		IFDI_MEDIA_SET(ctx);
4396 		CTX_UNLOCK(ctx);
4397 		/* FALLTHROUGH */
4398 	case SIOCGIFMEDIA:
4399 	case SIOCGIFXMEDIA:
4400 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4401 		break;
4402 	case SIOCGI2C:
4403 	{
4404 		struct ifi2creq i2c;
4405 
4406 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4407 		if (err != 0)
4408 			break;
4409 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4410 			err = EINVAL;
4411 			break;
4412 		}
4413 		if (i2c.len > sizeof(i2c.data)) {
4414 			err = EINVAL;
4415 			break;
4416 		}
4417 
4418 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4419 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4420 			    sizeof(i2c));
4421 		break;
4422 	}
4423 	case SIOCSIFCAP:
4424 	{
4425 		int mask, setmask, oldmask;
4426 
4427 		oldmask = if_getcapenable(ifp);
4428 		mask = ifr->ifr_reqcap ^ oldmask;
4429 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4430 		setmask = 0;
4431 #ifdef TCP_OFFLOAD
4432 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4433 #endif
4434 		setmask |= (mask & IFCAP_FLAGS);
4435 		setmask |= (mask & IFCAP_WOL);
4436 
4437 		/*
4438 		 * If any RX csum has changed, change all the ones that
4439 		 * are supported by the driver.
4440 		 */
4441 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4442 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4443 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4444 		}
4445 
4446 		/*
4447 		 * want to ensure that traffic has stopped before we change any of the flags
4448 		 */
4449 		if (setmask) {
4450 			CTX_LOCK(ctx);
4451 			bits = if_getdrvflags(ifp);
4452 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4453 				iflib_stop(ctx);
4454 			STATE_LOCK(ctx);
4455 			if_togglecapenable(ifp, setmask);
4456 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4457 			STATE_UNLOCK(ctx);
4458 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4459 				iflib_init_locked(ctx);
4460 			STATE_LOCK(ctx);
4461 			if_setdrvflags(ifp, bits);
4462 			STATE_UNLOCK(ctx);
4463 			CTX_UNLOCK(ctx);
4464 		}
4465 		if_vlancap(ifp);
4466 		break;
4467 	}
4468 	case SIOCGPRIVATE_0:
4469 	case SIOCSDRVSPEC:
4470 	case SIOCGDRVSPEC:
4471 		CTX_LOCK(ctx);
4472 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4473 		CTX_UNLOCK(ctx);
4474 		break;
4475 	default:
4476 		err = ether_ioctl(ifp, command, data);
4477 		break;
4478 	}
4479 	if (reinit)
4480 		iflib_if_init(ctx);
4481 	return (err);
4482 }
4483 
4484 static uint64_t
4485 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4486 {
4487 	if_ctx_t ctx = if_getsoftc(ifp);
4488 
4489 	return (IFDI_GET_COUNTER(ctx, cnt));
4490 }
4491 
4492 /*********************************************************************
4493  *
4494  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4495  *
4496  **********************************************************************/
4497 
4498 static void
4499 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4500 {
4501 	if_ctx_t ctx = if_getsoftc(ifp);
4502 
4503 	if ((void *)ctx != arg)
4504 		return;
4505 
4506 	if ((vtag == 0) || (vtag > 4095))
4507 		return;
4508 
4509 	if (iflib_in_detach(ctx))
4510 		return;
4511 
4512 	CTX_LOCK(ctx);
4513 	/* Driver may need all untagged packets to be flushed */
4514 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4515 		iflib_stop(ctx);
4516 	IFDI_VLAN_REGISTER(ctx, vtag);
4517 	/* Re-init to load the changes, if required */
4518 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4519 		iflib_init_locked(ctx);
4520 	CTX_UNLOCK(ctx);
4521 }
4522 
4523 static void
4524 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4525 {
4526 	if_ctx_t ctx = if_getsoftc(ifp);
4527 
4528 	if ((void *)ctx != arg)
4529 		return;
4530 
4531 	if ((vtag == 0) || (vtag > 4095))
4532 		return;
4533 
4534 	CTX_LOCK(ctx);
4535 	/* Driver may need all tagged packets to be flushed */
4536 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4537 		iflib_stop(ctx);
4538 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4539 	/* Re-init to load the changes, if required */
4540 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4541 		iflib_init_locked(ctx);
4542 	CTX_UNLOCK(ctx);
4543 }
4544 
4545 static void
4546 iflib_led_func(void *arg, int onoff)
4547 {
4548 	if_ctx_t ctx = arg;
4549 
4550 	CTX_LOCK(ctx);
4551 	IFDI_LED_FUNC(ctx, onoff);
4552 	CTX_UNLOCK(ctx);
4553 }
4554 
4555 /*********************************************************************
4556  *
4557  *  BUS FUNCTION DEFINITIONS
4558  *
4559  **********************************************************************/
4560 
4561 int
4562 iflib_device_probe(device_t dev)
4563 {
4564 	const pci_vendor_info_t *ent;
4565 	if_shared_ctx_t sctx;
4566 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4567 	uint16_t pci_vendor_id;
4568 
4569 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4570 		return (ENOTSUP);
4571 
4572 	pci_vendor_id = pci_get_vendor(dev);
4573 	pci_device_id = pci_get_device(dev);
4574 	pci_subvendor_id = pci_get_subvendor(dev);
4575 	pci_subdevice_id = pci_get_subdevice(dev);
4576 	pci_rev_id = pci_get_revid(dev);
4577 	if (sctx->isc_parse_devinfo != NULL)
4578 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4579 
4580 	ent = sctx->isc_vendor_info;
4581 	while (ent->pvi_vendor_id != 0) {
4582 		if (pci_vendor_id != ent->pvi_vendor_id) {
4583 			ent++;
4584 			continue;
4585 		}
4586 		if ((pci_device_id == ent->pvi_device_id) &&
4587 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4588 		     (ent->pvi_subvendor_id == 0)) &&
4589 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4590 		     (ent->pvi_subdevice_id == 0)) &&
4591 		    ((pci_rev_id == ent->pvi_rev_id) ||
4592 		     (ent->pvi_rev_id == 0))) {
4593 			device_set_desc_copy(dev, ent->pvi_name);
4594 			/* this needs to be changed to zero if the bus probing code
4595 			 * ever stops re-probing on best match because the sctx
4596 			 * may have its values over written by register calls
4597 			 * in subsequent probes
4598 			 */
4599 			return (BUS_PROBE_DEFAULT);
4600 		}
4601 		ent++;
4602 	}
4603 	return (ENXIO);
4604 }
4605 
4606 int
4607 iflib_device_probe_vendor(device_t dev)
4608 {
4609 	int probe;
4610 
4611 	probe = iflib_device_probe(dev);
4612 	if (probe == BUS_PROBE_DEFAULT)
4613 		return (BUS_PROBE_VENDOR);
4614 	else
4615 		return (probe);
4616 }
4617 
4618 static void
4619 iflib_reset_qvalues(if_ctx_t ctx)
4620 {
4621 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4622 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4623 	device_t dev = ctx->ifc_dev;
4624 	int i;
4625 
4626 	if (ctx->ifc_sysctl_ntxqs != 0)
4627 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4628 	if (ctx->ifc_sysctl_nrxqs != 0)
4629 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4630 
4631 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4632 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4633 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4634 		else
4635 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4636 	}
4637 
4638 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4639 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4640 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4641 		else
4642 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4643 	}
4644 
4645 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4646 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4647 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4648 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4649 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4650 		}
4651 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4652 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4653 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4654 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4655 		}
4656 		if (!powerof2(scctx->isc_nrxd[i])) {
4657 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4658 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4659 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4660 		}
4661 	}
4662 
4663 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4664 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4665 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4666 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4667 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4668 		}
4669 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4670 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4671 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4672 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4673 		}
4674 		if (!powerof2(scctx->isc_ntxd[i])) {
4675 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4676 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4677 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4678 		}
4679 	}
4680 }
4681 
4682 static void
4683 iflib_add_pfil(if_ctx_t ctx)
4684 {
4685 	struct pfil_head *pfil;
4686 	struct pfil_head_args pa;
4687 	iflib_rxq_t rxq;
4688 	int i;
4689 
4690 	pa.pa_version = PFIL_VERSION;
4691 	pa.pa_flags = PFIL_IN;
4692 	pa.pa_type = PFIL_TYPE_ETHERNET;
4693 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4694 	pfil = pfil_head_register(&pa);
4695 
4696 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4697 		rxq->pfil = pfil;
4698 	}
4699 }
4700 
4701 static void
4702 iflib_rem_pfil(if_ctx_t ctx)
4703 {
4704 	struct pfil_head *pfil;
4705 	iflib_rxq_t rxq;
4706 	int i;
4707 
4708 	rxq = ctx->ifc_rxqs;
4709 	pfil = rxq->pfil;
4710 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4711 		rxq->pfil = NULL;
4712 	}
4713 	pfil_head_unregister(pfil);
4714 }
4715 
4716 
4717 /*
4718  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4719  * cpuid and wrapping as necessary.
4720  */
4721 static unsigned int
4722 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4723 {
4724 	unsigned int first_valid;
4725 	unsigned int last_valid;
4726 
4727 	/* cpuid should always be in the valid set */
4728 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4729 
4730 	/* valid set should never be empty */
4731 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4732 
4733 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4734 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4735 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4736 	while (n > 0) {
4737 		do {
4738 			cpuid++;
4739 			if (cpuid > last_valid)
4740 				cpuid = first_valid;
4741 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4742 		n--;
4743 	}
4744 
4745 	return (cpuid);
4746 }
4747 
4748 #if defined(SMP) && defined(SCHED_ULE)
4749 extern struct cpu_group *cpu_top;              /* CPU topology */
4750 
4751 static int
4752 find_child_with_core(int cpu, struct cpu_group *grp)
4753 {
4754 	int i;
4755 
4756 	if (grp->cg_children == 0)
4757 		return -1;
4758 
4759 	MPASS(grp->cg_child);
4760 	for (i = 0; i < grp->cg_children; i++) {
4761 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4762 			return i;
4763 	}
4764 
4765 	return -1;
4766 }
4767 
4768 
4769 /*
4770  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4771  * does not distinguish among multiple L2 neighbors if the given CPU has
4772  * more than one (it will always return the same result in that case).
4773  */
4774 static int
4775 find_l2_neighbor(int cpu)
4776 {
4777 	struct cpu_group *grp;
4778 	int i;
4779 
4780 	grp = cpu_top;
4781 	if (grp == NULL)
4782 		return -1;
4783 
4784 	/*
4785 	 * Find the smallest CPU group that contains the given core.
4786 	 */
4787 	i = 0;
4788 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4789 		/*
4790 		 * If the smallest group containing the given CPU has less
4791 		 * than two members, we conclude the given CPU has no
4792 		 * L2 neighbor.
4793 		 */
4794 		if (grp->cg_child[i].cg_count <= 1)
4795 			return (-1);
4796 		grp = &grp->cg_child[i];
4797 	}
4798 
4799 	/* Must share L2. */
4800 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4801 		return -1;
4802 
4803 	/*
4804 	 * Select the first member of the set that isn't the reference
4805 	 * CPU, which at this point is guaranteed to exist.
4806 	 */
4807 	for (i = 0; i < CPU_SETSIZE; i++) {
4808 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4809 			return (i);
4810 	}
4811 
4812 	/* Should never be reached */
4813 	return (-1);
4814 }
4815 
4816 #else
4817 static int
4818 find_l2_neighbor(int cpu)
4819 {
4820 
4821 	return (-1);
4822 }
4823 #endif
4824 
4825 /*
4826  * CPU mapping behaviors
4827  * ---------------------
4828  * 'separate txrx' refers to the separate_txrx sysctl
4829  * 'use logical' refers to the use_logical_cores sysctl
4830  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4831  *
4832  *  separate     use     INTR
4833  *    txrx     logical   CPUS   result
4834  * ---------- --------- ------ ------------------------------------------------
4835  *     -          -       X     RX and TX queues mapped to consecutive physical
4836  *                              cores with RX/TX pairs on same core and excess
4837  *                              of either following
4838  *     -          X       X     RX and TX queues mapped to consecutive cores
4839  *                              of any type with RX/TX pairs on same core and
4840  *                              excess of either following
4841  *     X          -       X     RX and TX queues mapped to consecutive physical
4842  *                              cores; all RX then all TX
4843  *     X          X       X     RX queues mapped to consecutive physical cores
4844  *                              first, then TX queues mapped to L2 neighbor of
4845  *                              the corresponding RX queue if one exists,
4846  *                              otherwise to consecutive physical cores
4847  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4848  *                              any type with RX/TX pairs on same core and excess
4849  *                              of either following
4850  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4851  *                              any type; all RX then all TX
4852  */
4853 static unsigned int
4854 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4855     bool is_tx)
4856 {
4857 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4858 	unsigned int core_index;
4859 
4860 	if (ctx->ifc_sysctl_separate_txrx) {
4861 		/*
4862 		 * When using separate CPUs for TX and RX, the assignment
4863 		 * will always be of a consecutive CPU out of the set of
4864 		 * context CPUs, except for the specific case where the
4865 		 * context CPUs are phsyical cores, the use of logical cores
4866 		 * has been enabled, the assignment is for TX, the TX qid
4867 		 * corresponds to an RX qid, and the CPU assigned to the
4868 		 * corresponding RX queue has an L2 neighbor.
4869 		 */
4870 		if (ctx->ifc_sysctl_use_logical_cores &&
4871 		    ctx->ifc_cpus_are_physical_cores &&
4872 		    is_tx && qid < scctx->isc_nrxqsets) {
4873 			int l2_neighbor;
4874 			unsigned int rx_cpuid;
4875 
4876 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4877 			l2_neighbor = find_l2_neighbor(rx_cpuid);
4878 			if (l2_neighbor != -1) {
4879 				return (l2_neighbor);
4880 			}
4881 			/*
4882 			 * ... else fall through to the normal
4883 			 * consecutive-after-RX assignment scheme.
4884 			 *
4885 			 * Note that we are assuming that all RX queue CPUs
4886 			 * have an L2 neighbor, or all do not.  If a mixed
4887 			 * scenario is possible, we will have to keep track
4888 			 * separately of how many queues prior to this one
4889 			 * were not able to be assigned to an L2 neighbor.
4890 			 */
4891 		}
4892 		if (is_tx)
4893 			core_index = scctx->isc_nrxqsets + qid;
4894 		else
4895 			core_index = qid;
4896 	} else {
4897 		core_index = qid;
4898 	}
4899 
4900 	return (cpuid_advance(ctx, base_cpuid, core_index));
4901 }
4902 
4903 static uint16_t
4904 get_ctx_core_offset(if_ctx_t ctx)
4905 {
4906 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4907 	struct cpu_offset *op;
4908 	cpuset_t assigned_cpus;
4909 	unsigned int cores_consumed;
4910 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4911 	unsigned int first_valid;
4912 	unsigned int last_valid;
4913 	unsigned int i;
4914 
4915 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4916 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4917 
4918 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4919 		/*
4920 		 * Align the user-chosen base CPU ID to the next valid CPU
4921 		 * for this device.  If the chosen base CPU ID is smaller
4922 		 * than the first valid CPU or larger than the last valid
4923 		 * CPU, we assume the user does not know what the valid
4924 		 * range is for this device and is thinking in terms of a
4925 		 * zero-based reference frame, and so we shift the given
4926 		 * value into the valid range (and wrap accordingly) so the
4927 		 * intent is translated to the proper frame of reference.
4928 		 * If the base CPU ID is within the valid first/last, but
4929 		 * does not correspond to a valid CPU, it is advanced to the
4930 		 * next valid CPU (wrapping if necessary).
4931 		 */
4932 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
4933 			/* shift from zero-based to first_valid-based */
4934 			base_cpuid += first_valid;
4935 			/* wrap to range [first_valid, last_valid] */
4936 			base_cpuid = (base_cpuid - first_valid) %
4937 			    (last_valid - first_valid + 1);
4938 		}
4939 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
4940 			/*
4941 			 * base_cpuid is in [first_valid, last_valid], but
4942 			 * not a member of the valid set.  In this case,
4943 			 * there will always be a member of the valid set
4944 			 * with a CPU ID that is greater than base_cpuid,
4945 			 * and we simply advance to it.
4946 			 */
4947 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
4948 				base_cpuid++;
4949 		}
4950 		return (base_cpuid);
4951 	}
4952 
4953 	/*
4954 	 * Determine how many cores will be consumed by performing the CPU
4955 	 * assignments and counting how many of the assigned CPUs correspond
4956 	 * to CPUs in the set of context CPUs.  This is done using the CPU
4957 	 * ID first_valid as the base CPU ID, as the base CPU must be within
4958 	 * the set of context CPUs.
4959 	 *
4960 	 * Note not all assigned CPUs will be in the set of context CPUs
4961 	 * when separate CPUs are being allocated to TX and RX queues,
4962 	 * assignment to logical cores has been enabled, the set of context
4963 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
4964 	 * neighbors of CPUs that RX queues have been mapped to - in this
4965 	 * case we do only want to count how many CPUs in the set of context
4966 	 * CPUs have been consumed, as that determines the next CPU in that
4967 	 * set to start allocating at for the next device for which
4968 	 * core_offset is not set.
4969 	 */
4970 	CPU_ZERO(&assigned_cpus);
4971 	for (i = 0; i < scctx->isc_ntxqsets; i++)
4972 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
4973 		    &assigned_cpus);
4974 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4975 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
4976 		    &assigned_cpus);
4977 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
4978 	cores_consumed = CPU_COUNT(&assigned_cpus);
4979 
4980 	mtx_lock(&cpu_offset_mtx);
4981 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4982 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4983 			base_cpuid = op->next_cpuid;
4984 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
4985 			    cores_consumed);
4986 			MPASS(op->refcount < UINT_MAX);
4987 			op->refcount++;
4988 			break;
4989 		}
4990 	}
4991 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
4992 		base_cpuid = first_valid;
4993 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4994 		    M_NOWAIT | M_ZERO);
4995 		if (op == NULL) {
4996 			device_printf(ctx->ifc_dev,
4997 			    "allocation for cpu offset failed.\n");
4998 		} else {
4999 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5000 			    cores_consumed);
5001 			op->refcount = 1;
5002 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5003 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5004 		}
5005 	}
5006 	mtx_unlock(&cpu_offset_mtx);
5007 
5008 	return (base_cpuid);
5009 }
5010 
5011 static void
5012 unref_ctx_core_offset(if_ctx_t ctx)
5013 {
5014 	struct cpu_offset *op, *top;
5015 
5016 	mtx_lock(&cpu_offset_mtx);
5017 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5018 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5019 			MPASS(op->refcount > 0);
5020 			op->refcount--;
5021 			if (op->refcount == 0) {
5022 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5023 				free(op, M_IFLIB);
5024 			}
5025 			break;
5026 		}
5027 	}
5028 	mtx_unlock(&cpu_offset_mtx);
5029 }
5030 
5031 int
5032 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5033 {
5034 	if_ctx_t ctx;
5035 	if_t ifp;
5036 	if_softc_ctx_t scctx;
5037 	kobjop_desc_t kobj_desc;
5038 	kobj_method_t *kobj_method;
5039 	int err, msix, rid;
5040 	int num_txd, num_rxd;
5041 
5042 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5043 
5044 	if (sc == NULL) {
5045 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5046 		device_set_softc(dev, ctx);
5047 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5048 	}
5049 
5050 	ctx->ifc_sctx = sctx;
5051 	ctx->ifc_dev = dev;
5052 	ctx->ifc_softc = sc;
5053 
5054 	if ((err = iflib_register(ctx)) != 0) {
5055 		device_printf(dev, "iflib_register failed %d\n", err);
5056 		goto fail_ctx_free;
5057 	}
5058 	iflib_add_device_sysctl_pre(ctx);
5059 
5060 	scctx = &ctx->ifc_softc_ctx;
5061 	ifp = ctx->ifc_ifp;
5062 
5063 	iflib_reset_qvalues(ctx);
5064 	IFNET_WLOCK();
5065 	CTX_LOCK(ctx);
5066 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5067 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5068 		goto fail_unlock;
5069 	}
5070 	_iflib_pre_assert(scctx);
5071 	ctx->ifc_txrx = *scctx->isc_txrx;
5072 
5073 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5074 
5075 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5076 		ctx->ifc_mediap = scctx->isc_media;
5077 
5078 #ifdef INVARIANTS
5079 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5080 		MPASS(scctx->isc_tx_csum_flags);
5081 #endif
5082 
5083 	if_setcapabilities(ifp,
5084 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5085 	if_setcapenable(ifp,
5086 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5087 
5088 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5089 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5090 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5091 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5092 
5093 	num_txd = iflib_num_tx_descs(ctx);
5094 	num_rxd = iflib_num_rx_descs(ctx);
5095 
5096 	/* XXX change for per-queue sizes */
5097 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5098 	    num_txd, num_rxd);
5099 
5100 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5101 		scctx->isc_tx_nsegments = max(1, num_txd /
5102 		    MAX_SINGLE_PACKET_FRACTION);
5103 	if (scctx->isc_tx_tso_segments_max > num_txd /
5104 	    MAX_SINGLE_PACKET_FRACTION)
5105 		scctx->isc_tx_tso_segments_max = max(1,
5106 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5107 
5108 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5109 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5110 		/*
5111 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5112 		 * but some MACs do.
5113 		 */
5114 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5115 		    IP_MAXPACKET));
5116 		/*
5117 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5118 		 * into account.  In the worst case, each of these calls will
5119 		 * add another mbuf and, thus, the requirement for another DMA
5120 		 * segment.  So for best performance, it doesn't make sense to
5121 		 * advertize a maximum of TSO segments that typically will
5122 		 * require defragmentation in iflib_encap().
5123 		 */
5124 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5125 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5126 	}
5127 	if (scctx->isc_rss_table_size == 0)
5128 		scctx->isc_rss_table_size = 64;
5129 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5130 
5131 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5132 	/* XXX format name */
5133 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5134 	    NULL, NULL, "admin");
5135 
5136 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5137 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5138 		device_printf(dev, "Unable to fetch CPU list\n");
5139 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5140 		ctx->ifc_cpus_are_physical_cores = false;
5141 	} else
5142 		ctx->ifc_cpus_are_physical_cores = true;
5143 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5144 
5145 	/*
5146 	** Now set up MSI or MSI-X, should return us the number of supported
5147 	** vectors (will be 1 for a legacy interrupt and MSI).
5148 	*/
5149 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5150 		msix = scctx->isc_vectors;
5151 	} else if (scctx->isc_msix_bar != 0)
5152 	       /*
5153 		* The simple fact that isc_msix_bar is not 0 does not mean we
5154 		* we have a good value there that is known to work.
5155 		*/
5156 		msix = iflib_msix_init(ctx);
5157 	else {
5158 		scctx->isc_vectors = 1;
5159 		scctx->isc_ntxqsets = 1;
5160 		scctx->isc_nrxqsets = 1;
5161 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5162 		msix = 0;
5163 	}
5164 	/* Get memory for the station queues */
5165 	if ((err = iflib_queues_alloc(ctx))) {
5166 		device_printf(dev, "Unable to allocate queue memory\n");
5167 		goto fail_intr_free;
5168 	}
5169 
5170 	if ((err = iflib_qset_structures_setup(ctx)))
5171 		goto fail_queues;
5172 
5173 	/*
5174 	 * Now that we know how many queues there are, get the core offset.
5175 	 */
5176 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5177 
5178 	if (msix > 1) {
5179 		/*
5180 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5181 		 * aren't the default NULL implementation.
5182 		 */
5183 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5184 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5185 		    kobj_desc);
5186 		if (kobj_method == &kobj_desc->deflt) {
5187 			device_printf(dev,
5188 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5189 			err = EOPNOTSUPP;
5190 			goto fail_queues;
5191 		}
5192 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5193 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5194 		    kobj_desc);
5195 		if (kobj_method == &kobj_desc->deflt) {
5196 			device_printf(dev,
5197 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5198 			err = EOPNOTSUPP;
5199 			goto fail_queues;
5200 		}
5201 
5202 		/*
5203 		 * Assign the MSI-X vectors.
5204 		 * Note that the default NULL ifdi_msix_intr_assign method will
5205 		 * fail here, too.
5206 		 */
5207 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5208 		if (err != 0) {
5209 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5210 			    err);
5211 			goto fail_queues;
5212 		}
5213 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5214 		rid = 0;
5215 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5216 			MPASS(msix == 1);
5217 			rid = 1;
5218 		}
5219 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5220 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5221 			goto fail_queues;
5222 		}
5223 	} else {
5224 		device_printf(dev,
5225 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5226 		err = ENODEV;
5227 		goto fail_queues;
5228 	}
5229 
5230 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5231 
5232 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5233 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5234 		goto fail_detach;
5235 	}
5236 
5237 	/*
5238 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5239 	 * This must appear after the call to ether_ifattach() because
5240 	 * ether_ifattach() sets if_hdrlen to the default value.
5241 	 */
5242 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5243 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5244 
5245 	if ((err = iflib_netmap_attach(ctx))) {
5246 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5247 		goto fail_detach;
5248 	}
5249 	*ctxp = ctx;
5250 
5251 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5252 
5253 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5254 	iflib_add_device_sysctl_post(ctx);
5255 	iflib_add_pfil(ctx);
5256 	ctx->ifc_flags |= IFC_INIT_DONE;
5257 	CTX_UNLOCK(ctx);
5258 	IFNET_WUNLOCK();
5259 
5260 	return (0);
5261 
5262 fail_detach:
5263 	ether_ifdetach(ctx->ifc_ifp);
5264 fail_queues:
5265 	iflib_tqg_detach(ctx);
5266 	iflib_tx_structures_free(ctx);
5267 	iflib_rx_structures_free(ctx);
5268 	IFDI_DETACH(ctx);
5269 	IFDI_QUEUES_FREE(ctx);
5270 fail_intr_free:
5271 	iflib_free_intr_mem(ctx);
5272 fail_unlock:
5273 	CTX_UNLOCK(ctx);
5274 	IFNET_WUNLOCK();
5275 	iflib_deregister(ctx);
5276 fail_ctx_free:
5277 	device_set_softc(ctx->ifc_dev, NULL);
5278         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5279                 free(ctx->ifc_softc, M_IFLIB);
5280         free(ctx, M_IFLIB);
5281 	return (err);
5282 }
5283 
5284 int
5285 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5286 					  struct iflib_cloneattach_ctx *clctx)
5287 {
5288 	int num_txd, num_rxd;
5289 	int err;
5290 	if_ctx_t ctx;
5291 	if_t ifp;
5292 	if_softc_ctx_t scctx;
5293 	int i;
5294 	void *sc;
5295 
5296 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5297 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5298 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5299 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5300 		ctx->ifc_flags |= IFC_PSEUDO;
5301 
5302 	ctx->ifc_sctx = sctx;
5303 	ctx->ifc_softc = sc;
5304 	ctx->ifc_dev = dev;
5305 
5306 	if ((err = iflib_register(ctx)) != 0) {
5307 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5308 		goto fail_ctx_free;
5309 	}
5310 	iflib_add_device_sysctl_pre(ctx);
5311 
5312 	scctx = &ctx->ifc_softc_ctx;
5313 	ifp = ctx->ifc_ifp;
5314 
5315 	iflib_reset_qvalues(ctx);
5316 	CTX_LOCK(ctx);
5317 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5318 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5319 		goto fail_unlock;
5320 	}
5321 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5322 		ether_gen_addr(ifp, &ctx->ifc_mac);
5323 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5324 								clctx->cc_params)) != 0) {
5325 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5326 		goto fail_unlock;
5327 	}
5328 #ifdef INVARIANTS
5329 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5330 		MPASS(scctx->isc_tx_csum_flags);
5331 #endif
5332 
5333 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5334 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5335 
5336 	ifp->if_flags |= IFF_NOGROUP;
5337 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5338 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5339 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5340 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5341 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5342 		} else {
5343 			if_attach(ctx->ifc_ifp);
5344 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5345 		}
5346 
5347 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5348 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5349 			goto fail_detach;
5350 		}
5351 		*ctxp = ctx;
5352 
5353 		/*
5354 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5355 		 * This must appear after the call to ether_ifattach() because
5356 		 * ether_ifattach() sets if_hdrlen to the default value.
5357 		 */
5358 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5359 			if_setifheaderlen(ifp,
5360 			    sizeof(struct ether_vlan_header));
5361 
5362 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5363 		iflib_add_device_sysctl_post(ctx);
5364 		ctx->ifc_flags |= IFC_INIT_DONE;
5365 		CTX_UNLOCK(ctx);
5366 		return (0);
5367 	}
5368 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5369 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5370 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5371 
5372 	_iflib_pre_assert(scctx);
5373 	ctx->ifc_txrx = *scctx->isc_txrx;
5374 
5375 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5376 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5377 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5378 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5379 
5380 	num_txd = iflib_num_tx_descs(ctx);
5381 	num_rxd = iflib_num_rx_descs(ctx);
5382 
5383 	/* XXX change for per-queue sizes */
5384 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5385 	    num_txd, num_rxd);
5386 
5387 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5388 		scctx->isc_tx_nsegments = max(1, num_txd /
5389 		    MAX_SINGLE_PACKET_FRACTION);
5390 	if (scctx->isc_tx_tso_segments_max > num_txd /
5391 	    MAX_SINGLE_PACKET_FRACTION)
5392 		scctx->isc_tx_tso_segments_max = max(1,
5393 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5394 
5395 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5396 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5397 		/*
5398 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5399 		 * but some MACs do.
5400 		 */
5401 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5402 		    IP_MAXPACKET));
5403 		/*
5404 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5405 		 * into account.  In the worst case, each of these calls will
5406 		 * add another mbuf and, thus, the requirement for another DMA
5407 		 * segment.  So for best performance, it doesn't make sense to
5408 		 * advertize a maximum of TSO segments that typically will
5409 		 * require defragmentation in iflib_encap().
5410 		 */
5411 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5412 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5413 	}
5414 	if (scctx->isc_rss_table_size == 0)
5415 		scctx->isc_rss_table_size = 64;
5416 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5417 
5418 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5419 	/* XXX format name */
5420 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5421 	    NULL, NULL, "admin");
5422 
5423 	/* XXX --- can support > 1 -- but keep it simple for now */
5424 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5425 
5426 	/* Get memory for the station queues */
5427 	if ((err = iflib_queues_alloc(ctx))) {
5428 		device_printf(dev, "Unable to allocate queue memory\n");
5429 		goto fail_iflib_detach;
5430 	}
5431 
5432 	if ((err = iflib_qset_structures_setup(ctx))) {
5433 		device_printf(dev, "qset structure setup failed %d\n", err);
5434 		goto fail_queues;
5435 	}
5436 
5437 	/*
5438 	 * XXX What if anything do we want to do about interrupts?
5439 	 */
5440 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5441 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5442 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5443 		goto fail_detach;
5444 	}
5445 
5446 	/*
5447 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5448 	 * This must appear after the call to ether_ifattach() because
5449 	 * ether_ifattach() sets if_hdrlen to the default value.
5450 	 */
5451 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5452 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5453 
5454 	/* XXX handle more than one queue */
5455 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5456 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5457 
5458 	*ctxp = ctx;
5459 
5460 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5461 	iflib_add_device_sysctl_post(ctx);
5462 	ctx->ifc_flags |= IFC_INIT_DONE;
5463 	CTX_UNLOCK(ctx);
5464 
5465 	return (0);
5466 fail_detach:
5467 	ether_ifdetach(ctx->ifc_ifp);
5468 fail_queues:
5469 	iflib_tqg_detach(ctx);
5470 	iflib_tx_structures_free(ctx);
5471 	iflib_rx_structures_free(ctx);
5472 fail_iflib_detach:
5473 	IFDI_DETACH(ctx);
5474 	IFDI_QUEUES_FREE(ctx);
5475 fail_unlock:
5476 	CTX_UNLOCK(ctx);
5477 	iflib_deregister(ctx);
5478 fail_ctx_free:
5479 	free(ctx->ifc_softc, M_IFLIB);
5480 	free(ctx, M_IFLIB);
5481 	return (err);
5482 }
5483 
5484 int
5485 iflib_pseudo_deregister(if_ctx_t ctx)
5486 {
5487 	if_t ifp = ctx->ifc_ifp;
5488 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5489 
5490 	/* Unregister VLAN event handlers early */
5491 	iflib_unregister_vlan_handlers(ctx);
5492 
5493 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5494 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5495 		bpfdetach(ifp);
5496 		if_detach(ifp);
5497 	} else {
5498 		ether_ifdetach(ifp);
5499 	}
5500 
5501 	iflib_tqg_detach(ctx);
5502 	iflib_tx_structures_free(ctx);
5503 	iflib_rx_structures_free(ctx);
5504 	IFDI_DETACH(ctx);
5505 	IFDI_QUEUES_FREE(ctx);
5506 
5507 	iflib_deregister(ctx);
5508 
5509 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5510 		free(ctx->ifc_softc, M_IFLIB);
5511 	free(ctx, M_IFLIB);
5512 	return (0);
5513 }
5514 
5515 int
5516 iflib_device_attach(device_t dev)
5517 {
5518 	if_ctx_t ctx;
5519 	if_shared_ctx_t sctx;
5520 
5521 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5522 		return (ENOTSUP);
5523 
5524 	pci_enable_busmaster(dev);
5525 
5526 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5527 }
5528 
5529 int
5530 iflib_device_deregister(if_ctx_t ctx)
5531 {
5532 	if_t ifp = ctx->ifc_ifp;
5533 	device_t dev = ctx->ifc_dev;
5534 
5535 	/* Make sure VLANS are not using driver */
5536 	if (if_vlantrunkinuse(ifp)) {
5537 		device_printf(dev, "Vlan in use, detach first\n");
5538 		return (EBUSY);
5539 	}
5540 #ifdef PCI_IOV
5541 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5542 		device_printf(dev, "SR-IOV in use; detach first.\n");
5543 		return (EBUSY);
5544 	}
5545 #endif
5546 
5547 	STATE_LOCK(ctx);
5548 	ctx->ifc_flags |= IFC_IN_DETACH;
5549 	STATE_UNLOCK(ctx);
5550 
5551 	/* Unregister VLAN handlers before calling iflib_stop() */
5552 	iflib_unregister_vlan_handlers(ctx);
5553 
5554 	iflib_netmap_detach(ifp);
5555 	ether_ifdetach(ifp);
5556 
5557 	CTX_LOCK(ctx);
5558 	iflib_stop(ctx);
5559 	CTX_UNLOCK(ctx);
5560 
5561 	iflib_rem_pfil(ctx);
5562 	if (ctx->ifc_led_dev != NULL)
5563 		led_destroy(ctx->ifc_led_dev);
5564 
5565 	iflib_tqg_detach(ctx);
5566 	iflib_tx_structures_free(ctx);
5567 	iflib_rx_structures_free(ctx);
5568 
5569 	CTX_LOCK(ctx);
5570 	IFDI_DETACH(ctx);
5571 	IFDI_QUEUES_FREE(ctx);
5572 	CTX_UNLOCK(ctx);
5573 
5574 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5575 	iflib_free_intr_mem(ctx);
5576 
5577 	bus_generic_detach(dev);
5578 
5579 	iflib_deregister(ctx);
5580 
5581 	device_set_softc(ctx->ifc_dev, NULL);
5582 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5583 		free(ctx->ifc_softc, M_IFLIB);
5584 	unref_ctx_core_offset(ctx);
5585 	free(ctx, M_IFLIB);
5586 	return (0);
5587 }
5588 
5589 static void
5590 iflib_tqg_detach(if_ctx_t ctx)
5591 {
5592 	iflib_txq_t txq;
5593 	iflib_rxq_t rxq;
5594 	int i;
5595 	struct taskqgroup *tqg;
5596 
5597 	/* XXX drain any dependent tasks */
5598 	tqg = qgroup_if_io_tqg;
5599 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5600 		callout_drain(&txq->ift_timer);
5601 #ifdef DEV_NETMAP
5602 		callout_drain(&txq->ift_netmap_timer);
5603 #endif /* DEV_NETMAP */
5604 		if (txq->ift_task.gt_uniq != NULL)
5605 			taskqgroup_detach(tqg, &txq->ift_task);
5606 	}
5607 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5608 		if (rxq->ifr_task.gt_uniq != NULL)
5609 			taskqgroup_detach(tqg, &rxq->ifr_task);
5610 	}
5611 	tqg = qgroup_if_config_tqg;
5612 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5613 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5614 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5615 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5616 }
5617 
5618 static void
5619 iflib_free_intr_mem(if_ctx_t ctx)
5620 {
5621 
5622 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5623 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5624 	}
5625 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5626 		pci_release_msi(ctx->ifc_dev);
5627 	}
5628 	if (ctx->ifc_msix_mem != NULL) {
5629 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5630 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5631 		ctx->ifc_msix_mem = NULL;
5632 	}
5633 }
5634 
5635 int
5636 iflib_device_detach(device_t dev)
5637 {
5638 	if_ctx_t ctx = device_get_softc(dev);
5639 
5640 	return (iflib_device_deregister(ctx));
5641 }
5642 
5643 int
5644 iflib_device_suspend(device_t dev)
5645 {
5646 	if_ctx_t ctx = device_get_softc(dev);
5647 
5648 	CTX_LOCK(ctx);
5649 	IFDI_SUSPEND(ctx);
5650 	CTX_UNLOCK(ctx);
5651 
5652 	return bus_generic_suspend(dev);
5653 }
5654 int
5655 iflib_device_shutdown(device_t dev)
5656 {
5657 	if_ctx_t ctx = device_get_softc(dev);
5658 
5659 	CTX_LOCK(ctx);
5660 	IFDI_SHUTDOWN(ctx);
5661 	CTX_UNLOCK(ctx);
5662 
5663 	return bus_generic_suspend(dev);
5664 }
5665 
5666 int
5667 iflib_device_resume(device_t dev)
5668 {
5669 	if_ctx_t ctx = device_get_softc(dev);
5670 	iflib_txq_t txq = ctx->ifc_txqs;
5671 
5672 	CTX_LOCK(ctx);
5673 	IFDI_RESUME(ctx);
5674 	iflib_if_init_locked(ctx);
5675 	CTX_UNLOCK(ctx);
5676 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5677 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5678 
5679 	return (bus_generic_resume(dev));
5680 }
5681 
5682 int
5683 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5684 {
5685 	int error;
5686 	if_ctx_t ctx = device_get_softc(dev);
5687 
5688 	CTX_LOCK(ctx);
5689 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5690 	CTX_UNLOCK(ctx);
5691 
5692 	return (error);
5693 }
5694 
5695 void
5696 iflib_device_iov_uninit(device_t dev)
5697 {
5698 	if_ctx_t ctx = device_get_softc(dev);
5699 
5700 	CTX_LOCK(ctx);
5701 	IFDI_IOV_UNINIT(ctx);
5702 	CTX_UNLOCK(ctx);
5703 }
5704 
5705 int
5706 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5707 {
5708 	int error;
5709 	if_ctx_t ctx = device_get_softc(dev);
5710 
5711 	CTX_LOCK(ctx);
5712 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5713 	CTX_UNLOCK(ctx);
5714 
5715 	return (error);
5716 }
5717 
5718 /*********************************************************************
5719  *
5720  *  MODULE FUNCTION DEFINITIONS
5721  *
5722  **********************************************************************/
5723 
5724 /*
5725  * - Start a fast taskqueue thread for each core
5726  * - Start a taskqueue for control operations
5727  */
5728 static int
5729 iflib_module_init(void)
5730 {
5731 	iflib_timer_default = hz / 2;
5732 	return (0);
5733 }
5734 
5735 static int
5736 iflib_module_event_handler(module_t mod, int what, void *arg)
5737 {
5738 	int err;
5739 
5740 	switch (what) {
5741 	case MOD_LOAD:
5742 		if ((err = iflib_module_init()) != 0)
5743 			return (err);
5744 		break;
5745 	case MOD_UNLOAD:
5746 		return (EBUSY);
5747 	default:
5748 		return (EOPNOTSUPP);
5749 	}
5750 
5751 	return (0);
5752 }
5753 
5754 /*********************************************************************
5755  *
5756  *  PUBLIC FUNCTION DEFINITIONS
5757  *     ordered as in iflib.h
5758  *
5759  **********************************************************************/
5760 
5761 static void
5762 _iflib_assert(if_shared_ctx_t sctx)
5763 {
5764 	int i;
5765 
5766 	MPASS(sctx->isc_tx_maxsize);
5767 	MPASS(sctx->isc_tx_maxsegsize);
5768 
5769 	MPASS(sctx->isc_rx_maxsize);
5770 	MPASS(sctx->isc_rx_nsegments);
5771 	MPASS(sctx->isc_rx_maxsegsize);
5772 
5773 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5774 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5775 		MPASS(sctx->isc_nrxd_min[i]);
5776 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5777 		MPASS(sctx->isc_nrxd_max[i]);
5778 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5779 		MPASS(sctx->isc_nrxd_default[i]);
5780 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5781 	}
5782 
5783 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5784 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5785 		MPASS(sctx->isc_ntxd_min[i]);
5786 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5787 		MPASS(sctx->isc_ntxd_max[i]);
5788 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5789 		MPASS(sctx->isc_ntxd_default[i]);
5790 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5791 	}
5792 }
5793 
5794 static void
5795 _iflib_pre_assert(if_softc_ctx_t scctx)
5796 {
5797 
5798 	MPASS(scctx->isc_txrx->ift_txd_encap);
5799 	MPASS(scctx->isc_txrx->ift_txd_flush);
5800 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5801 	MPASS(scctx->isc_txrx->ift_rxd_available);
5802 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5803 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5804 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5805 }
5806 
5807 static int
5808 iflib_register(if_ctx_t ctx)
5809 {
5810 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5811 	driver_t *driver = sctx->isc_driver;
5812 	device_t dev = ctx->ifc_dev;
5813 	if_t ifp;
5814 	u_char type;
5815 	int iflags;
5816 
5817 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5818 		_iflib_assert(sctx);
5819 
5820 	CTX_LOCK_INIT(ctx);
5821 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5822 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5823 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5824 			type = IFT_ETHER;
5825 		else
5826 			type = IFT_PPP;
5827 	} else
5828 		type = IFT_ETHER;
5829 	ifp = ctx->ifc_ifp = if_alloc(type);
5830 	if (ifp == NULL) {
5831 		device_printf(dev, "can not allocate ifnet structure\n");
5832 		return (ENOMEM);
5833 	}
5834 
5835 	/*
5836 	 * Initialize our context's device specific methods
5837 	 */
5838 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5839 	kobj_class_compile((kobj_class_t) driver);
5840 
5841 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5842 	if_setsoftc(ifp, ctx);
5843 	if_setdev(ifp, dev);
5844 	if_setinitfn(ifp, iflib_if_init);
5845 	if_setioctlfn(ifp, iflib_if_ioctl);
5846 #ifdef ALTQ
5847 	if_setstartfn(ifp, iflib_altq_if_start);
5848 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5849 	if_setsendqready(ifp);
5850 #else
5851 	if_settransmitfn(ifp, iflib_if_transmit);
5852 #endif
5853 	if_setqflushfn(ifp, iflib_if_qflush);
5854 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5855 
5856 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5857 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5858 		iflags |= IFF_POINTOPOINT;
5859 	else
5860 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5861 	if_setflags(ifp, iflags);
5862 	ctx->ifc_vlan_attach_event =
5863 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5864 							  EVENTHANDLER_PRI_FIRST);
5865 	ctx->ifc_vlan_detach_event =
5866 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5867 							  EVENTHANDLER_PRI_FIRST);
5868 
5869 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5870 		ctx->ifc_mediap = &ctx->ifc_media;
5871 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5872 		    iflib_media_change, iflib_media_status);
5873 	}
5874 	return (0);
5875 }
5876 
5877 static void
5878 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5879 {
5880 	/* Unregister VLAN events */
5881 	if (ctx->ifc_vlan_attach_event != NULL) {
5882 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5883 		ctx->ifc_vlan_attach_event = NULL;
5884 	}
5885 	if (ctx->ifc_vlan_detach_event != NULL) {
5886 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5887 		ctx->ifc_vlan_detach_event = NULL;
5888 	}
5889 
5890 }
5891 
5892 static void
5893 iflib_deregister(if_ctx_t ctx)
5894 {
5895 	if_t ifp = ctx->ifc_ifp;
5896 
5897 	/* Remove all media */
5898 	ifmedia_removeall(&ctx->ifc_media);
5899 
5900 	/* Ensure that VLAN event handlers are unregistered */
5901 	iflib_unregister_vlan_handlers(ctx);
5902 
5903 	/* Release kobject reference */
5904 	kobj_delete((kobj_t) ctx, NULL);
5905 
5906 	/* Free the ifnet structure */
5907 	if_free(ifp);
5908 
5909 	STATE_LOCK_DESTROY(ctx);
5910 
5911 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5912 	CTX_LOCK_DESTROY(ctx);
5913 }
5914 
5915 static int
5916 iflib_queues_alloc(if_ctx_t ctx)
5917 {
5918 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5919 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5920 	device_t dev = ctx->ifc_dev;
5921 	int nrxqsets = scctx->isc_nrxqsets;
5922 	int ntxqsets = scctx->isc_ntxqsets;
5923 	iflib_txq_t txq;
5924 	iflib_rxq_t rxq;
5925 	iflib_fl_t fl = NULL;
5926 	int i, j, cpu, err, txconf, rxconf;
5927 	iflib_dma_info_t ifdip;
5928 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5929 	uint32_t *txqsizes = scctx->isc_txqsizes;
5930 	uint8_t nrxqs = sctx->isc_nrxqs;
5931 	uint8_t ntxqs = sctx->isc_ntxqs;
5932 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5933 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5934 	caddr_t *vaddrs;
5935 	uint64_t *paddrs;
5936 
5937 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5938 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5939 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5940            ("there must be at least a rxq for each free list"));
5941 
5942 	/* Allocate the TX ring struct memory */
5943 	if (!(ctx->ifc_txqs =
5944 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5945 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5946 		device_printf(dev, "Unable to allocate TX ring memory\n");
5947 		err = ENOMEM;
5948 		goto fail;
5949 	}
5950 
5951 	/* Now allocate the RX */
5952 	if (!(ctx->ifc_rxqs =
5953 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5954 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5955 		device_printf(dev, "Unable to allocate RX ring memory\n");
5956 		err = ENOMEM;
5957 		goto rx_fail;
5958 	}
5959 
5960 	txq = ctx->ifc_txqs;
5961 	rxq = ctx->ifc_rxqs;
5962 
5963 	/*
5964 	 * XXX handle allocation failure
5965 	 */
5966 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5967 		/* Set up some basics */
5968 
5969 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5970 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5971 			device_printf(dev,
5972 			    "Unable to allocate TX DMA info memory\n");
5973 			err = ENOMEM;
5974 			goto err_tx_desc;
5975 		}
5976 		txq->ift_ifdi = ifdip;
5977 		for (j = 0; j < ntxqs; j++, ifdip++) {
5978 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5979 				device_printf(dev,
5980 				    "Unable to allocate TX descriptors\n");
5981 				err = ENOMEM;
5982 				goto err_tx_desc;
5983 			}
5984 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5985 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5986 		}
5987 		txq->ift_ctx = ctx;
5988 		txq->ift_id = i;
5989 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5990 			txq->ift_br_offset = 1;
5991 		} else {
5992 			txq->ift_br_offset = 0;
5993 		}
5994 
5995 		if (iflib_txsd_alloc(txq)) {
5996 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5997 			err = ENOMEM;
5998 			goto err_tx_desc;
5999 		}
6000 
6001 		/* Initialize the TX lock */
6002 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6003 		    device_get_nameunit(dev), txq->ift_id);
6004 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6005 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6006 		txq->ift_timer.c_cpu = cpu;
6007 #ifdef DEV_NETMAP
6008 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6009 		txq->ift_netmap_timer.c_cpu = cpu;
6010 #endif /* DEV_NETMAP */
6011 
6012 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6013 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6014 		if (err) {
6015 			/* XXX free any allocated rings */
6016 			device_printf(dev, "Unable to allocate buf_ring\n");
6017 			goto err_tx_desc;
6018 		}
6019 	}
6020 
6021 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6022 		/* Set up some basics */
6023 		callout_init(&rxq->ifr_watchdog, 1);
6024 
6025 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6026 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6027 			device_printf(dev,
6028 			    "Unable to allocate RX DMA info memory\n");
6029 			err = ENOMEM;
6030 			goto err_tx_desc;
6031 		}
6032 
6033 		rxq->ifr_ifdi = ifdip;
6034 		/* XXX this needs to be changed if #rx queues != #tx queues */
6035 		rxq->ifr_ntxqirq = 1;
6036 		rxq->ifr_txqid[0] = i;
6037 		for (j = 0; j < nrxqs; j++, ifdip++) {
6038 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6039 				device_printf(dev,
6040 				    "Unable to allocate RX descriptors\n");
6041 				err = ENOMEM;
6042 				goto err_tx_desc;
6043 			}
6044 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6045 		}
6046 		rxq->ifr_ctx = ctx;
6047 		rxq->ifr_id = i;
6048 		rxq->ifr_fl_offset = fl_offset;
6049 		rxq->ifr_nfl = nfree_lists;
6050 		if (!(fl =
6051 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6052 			device_printf(dev, "Unable to allocate free list memory\n");
6053 			err = ENOMEM;
6054 			goto err_tx_desc;
6055 		}
6056 		rxq->ifr_fl = fl;
6057 		for (j = 0; j < nfree_lists; j++) {
6058 			fl[j].ifl_rxq = rxq;
6059 			fl[j].ifl_id = j;
6060 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6061 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6062 		}
6063 		/* Allocate receive buffers for the ring */
6064 		if (iflib_rxsd_alloc(rxq)) {
6065 			device_printf(dev,
6066 			    "Critical Failure setting up receive buffers\n");
6067 			err = ENOMEM;
6068 			goto err_rx_desc;
6069 		}
6070 
6071 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6072 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6073 			    M_WAITOK);
6074 	}
6075 
6076 	/* TXQs */
6077 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6078 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6079 	for (i = 0; i < ntxqsets; i++) {
6080 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6081 
6082 		for (j = 0; j < ntxqs; j++, di++) {
6083 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6084 			paddrs[i*ntxqs + j] = di->idi_paddr;
6085 		}
6086 	}
6087 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6088 		device_printf(ctx->ifc_dev,
6089 		    "Unable to allocate device TX queue\n");
6090 		iflib_tx_structures_free(ctx);
6091 		free(vaddrs, M_IFLIB);
6092 		free(paddrs, M_IFLIB);
6093 		goto err_rx_desc;
6094 	}
6095 	free(vaddrs, M_IFLIB);
6096 	free(paddrs, M_IFLIB);
6097 
6098 	/* RXQs */
6099 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6100 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6101 	for (i = 0; i < nrxqsets; i++) {
6102 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6103 
6104 		for (j = 0; j < nrxqs; j++, di++) {
6105 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6106 			paddrs[i*nrxqs + j] = di->idi_paddr;
6107 		}
6108 	}
6109 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6110 		device_printf(ctx->ifc_dev,
6111 		    "Unable to allocate device RX queue\n");
6112 		iflib_tx_structures_free(ctx);
6113 		free(vaddrs, M_IFLIB);
6114 		free(paddrs, M_IFLIB);
6115 		goto err_rx_desc;
6116 	}
6117 	free(vaddrs, M_IFLIB);
6118 	free(paddrs, M_IFLIB);
6119 
6120 	return (0);
6121 
6122 /* XXX handle allocation failure changes */
6123 err_rx_desc:
6124 err_tx_desc:
6125 rx_fail:
6126 	if (ctx->ifc_rxqs != NULL)
6127 		free(ctx->ifc_rxqs, M_IFLIB);
6128 	ctx->ifc_rxqs = NULL;
6129 	if (ctx->ifc_txqs != NULL)
6130 		free(ctx->ifc_txqs, M_IFLIB);
6131 	ctx->ifc_txqs = NULL;
6132 fail:
6133 	return (err);
6134 }
6135 
6136 static int
6137 iflib_tx_structures_setup(if_ctx_t ctx)
6138 {
6139 	iflib_txq_t txq = ctx->ifc_txqs;
6140 	int i;
6141 
6142 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6143 		iflib_txq_setup(txq);
6144 
6145 	return (0);
6146 }
6147 
6148 static void
6149 iflib_tx_structures_free(if_ctx_t ctx)
6150 {
6151 	iflib_txq_t txq = ctx->ifc_txqs;
6152 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6153 	int i, j;
6154 
6155 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6156 		for (j = 0; j < sctx->isc_ntxqs; j++)
6157 			iflib_dma_free(&txq->ift_ifdi[j]);
6158 		iflib_txq_destroy(txq);
6159 	}
6160 	free(ctx->ifc_txqs, M_IFLIB);
6161 	ctx->ifc_txqs = NULL;
6162 }
6163 
6164 /*********************************************************************
6165  *
6166  *  Initialize all receive rings.
6167  *
6168  **********************************************************************/
6169 static int
6170 iflib_rx_structures_setup(if_ctx_t ctx)
6171 {
6172 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6173 	int q;
6174 #if defined(INET6) || defined(INET)
6175 	int err, i;
6176 #endif
6177 
6178 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6179 #if defined(INET6) || defined(INET)
6180 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6181 		    TCP_LRO_ENTRIES, min(1024,
6182 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6183 		if (err != 0) {
6184 			device_printf(ctx->ifc_dev,
6185 			    "LRO Initialization failed!\n");
6186 			goto fail;
6187 		}
6188 #endif
6189 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6190 	}
6191 	return (0);
6192 #if defined(INET6) || defined(INET)
6193 fail:
6194 	/*
6195 	 * Free LRO resources allocated so far, we will only handle
6196 	 * the rings that completed, the failing case will have
6197 	 * cleaned up for itself.  'q' failed, so its the terminus.
6198 	 */
6199 	rxq = ctx->ifc_rxqs;
6200 	for (i = 0; i < q; ++i, rxq++) {
6201 		tcp_lro_free(&rxq->ifr_lc);
6202 	}
6203 	return (err);
6204 #endif
6205 }
6206 
6207 /*********************************************************************
6208  *
6209  *  Free all receive rings.
6210  *
6211  **********************************************************************/
6212 static void
6213 iflib_rx_structures_free(if_ctx_t ctx)
6214 {
6215 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6216 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6217 	int i, j;
6218 
6219 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6220 		for (j = 0; j < sctx->isc_nrxqs; j++)
6221 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6222 		iflib_rx_sds_free(rxq);
6223 #if defined(INET6) || defined(INET)
6224 		tcp_lro_free(&rxq->ifr_lc);
6225 #endif
6226 	}
6227 	free(ctx->ifc_rxqs, M_IFLIB);
6228 	ctx->ifc_rxqs = NULL;
6229 }
6230 
6231 static int
6232 iflib_qset_structures_setup(if_ctx_t ctx)
6233 {
6234 	int err;
6235 
6236 	/*
6237 	 * It is expected that the caller takes care of freeing queues if this
6238 	 * fails.
6239 	 */
6240 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6241 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6242 		return (err);
6243 	}
6244 
6245 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6246 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6247 
6248 	return (err);
6249 }
6250 
6251 int
6252 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6253 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6254 {
6255 
6256 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6257 }
6258 
6259 /* Just to avoid copy/paste */
6260 static inline int
6261 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6262     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6263     const char *name)
6264 {
6265 	device_t dev;
6266 	unsigned int base_cpuid, cpuid;
6267 	int err;
6268 
6269 	dev = ctx->ifc_dev;
6270 	base_cpuid = ctx->ifc_sysctl_core_offset;
6271 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6272 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6273 	    irq ? irq->ii_res : NULL, name);
6274 	if (err) {
6275 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6276 		return (err);
6277 	}
6278 #ifdef notyet
6279 	if (cpuid > ctx->ifc_cpuid_highest)
6280 		ctx->ifc_cpuid_highest = cpuid;
6281 #endif
6282 	return (0);
6283 }
6284 
6285 int
6286 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6287 			iflib_intr_type_t type, driver_filter_t *filter,
6288 			void *filter_arg, int qid, const char *name)
6289 {
6290 	device_t dev;
6291 	struct grouptask *gtask;
6292 	struct taskqgroup *tqg;
6293 	iflib_filter_info_t info;
6294 	gtask_fn_t *fn;
6295 	int tqrid, err;
6296 	driver_filter_t *intr_fast;
6297 	void *q;
6298 
6299 	info = &ctx->ifc_filter_info;
6300 	tqrid = rid;
6301 
6302 	switch (type) {
6303 	/* XXX merge tx/rx for netmap? */
6304 	case IFLIB_INTR_TX:
6305 		q = &ctx->ifc_txqs[qid];
6306 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6307 		gtask = &ctx->ifc_txqs[qid].ift_task;
6308 		tqg = qgroup_if_io_tqg;
6309 		fn = _task_fn_tx;
6310 		intr_fast = iflib_fast_intr;
6311 		GROUPTASK_INIT(gtask, 0, fn, q);
6312 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6313 		break;
6314 	case IFLIB_INTR_RX:
6315 		q = &ctx->ifc_rxqs[qid];
6316 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6317 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6318 		tqg = qgroup_if_io_tqg;
6319 		fn = _task_fn_rx;
6320 		intr_fast = iflib_fast_intr;
6321 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6322 		break;
6323 	case IFLIB_INTR_RXTX:
6324 		q = &ctx->ifc_rxqs[qid];
6325 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6326 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6327 		tqg = qgroup_if_io_tqg;
6328 		fn = _task_fn_rx;
6329 		intr_fast = iflib_fast_intr_rxtx;
6330 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6331 		break;
6332 	case IFLIB_INTR_ADMIN:
6333 		q = ctx;
6334 		tqrid = -1;
6335 		info = &ctx->ifc_filter_info;
6336 		gtask = &ctx->ifc_admin_task;
6337 		tqg = qgroup_if_config_tqg;
6338 		fn = _task_fn_admin;
6339 		intr_fast = iflib_fast_intr_ctx;
6340 		break;
6341 	default:
6342 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6343 		    __func__);
6344 		return (EINVAL);
6345 	}
6346 
6347 	info->ifi_filter = filter;
6348 	info->ifi_filter_arg = filter_arg;
6349 	info->ifi_task = gtask;
6350 	info->ifi_ctx = q;
6351 
6352 	dev = ctx->ifc_dev;
6353 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6354 	if (err != 0) {
6355 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6356 		return (err);
6357 	}
6358 	if (type == IFLIB_INTR_ADMIN)
6359 		return (0);
6360 
6361 	if (tqrid != -1) {
6362 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6363 		    name);
6364 		if (err)
6365 			return (err);
6366 	} else {
6367 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6368 	}
6369 
6370 	return (0);
6371 }
6372 
6373 void
6374 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6375 {
6376 	device_t dev;
6377 	struct grouptask *gtask;
6378 	struct taskqgroup *tqg;
6379 	gtask_fn_t *fn;
6380 	void *q;
6381 	int err;
6382 
6383 	switch (type) {
6384 	case IFLIB_INTR_TX:
6385 		q = &ctx->ifc_txqs[qid];
6386 		gtask = &ctx->ifc_txqs[qid].ift_task;
6387 		tqg = qgroup_if_io_tqg;
6388 		fn = _task_fn_tx;
6389 		GROUPTASK_INIT(gtask, 0, fn, q);
6390 		break;
6391 	case IFLIB_INTR_RX:
6392 		q = &ctx->ifc_rxqs[qid];
6393 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6394 		tqg = qgroup_if_io_tqg;
6395 		fn = _task_fn_rx;
6396 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6397 		break;
6398 	case IFLIB_INTR_IOV:
6399 		q = ctx;
6400 		gtask = &ctx->ifc_vflr_task;
6401 		tqg = qgroup_if_config_tqg;
6402 		fn = _task_fn_iov;
6403 		GROUPTASK_INIT(gtask, 0, fn, q);
6404 		break;
6405 	default:
6406 		panic("unknown net intr type");
6407 	}
6408 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6409 	if (err) {
6410 		dev = ctx->ifc_dev;
6411 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6412 		    name);
6413 	}
6414 }
6415 
6416 void
6417 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6418 {
6419 
6420 	if (irq->ii_tag)
6421 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6422 
6423 	if (irq->ii_res)
6424 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6425 		    rman_get_rid(irq->ii_res), irq->ii_res);
6426 }
6427 
6428 static int
6429 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6430 {
6431 	iflib_txq_t txq = ctx->ifc_txqs;
6432 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6433 	if_irq_t irq = &ctx->ifc_legacy_irq;
6434 	iflib_filter_info_t info;
6435 	device_t dev;
6436 	struct grouptask *gtask;
6437 	struct resource *res;
6438 	struct taskqgroup *tqg;
6439 	void *q;
6440 	int err, tqrid;
6441 	bool rx_only;
6442 
6443 	q = &ctx->ifc_rxqs[0];
6444 	info = &rxq[0].ifr_filter_info;
6445 	gtask = &rxq[0].ifr_task;
6446 	tqg = qgroup_if_io_tqg;
6447 	tqrid = *rid;
6448 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6449 
6450 	ctx->ifc_flags |= IFC_LEGACY;
6451 	info->ifi_filter = filter;
6452 	info->ifi_filter_arg = filter_arg;
6453 	info->ifi_task = gtask;
6454 	info->ifi_ctx = rx_only ? ctx : q;
6455 
6456 	dev = ctx->ifc_dev;
6457 	/* We allocate a single interrupt resource */
6458 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6459 	    iflib_fast_intr_rxtx, NULL, info, name);
6460 	if (err != 0)
6461 		return (err);
6462 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6463 	res = irq->ii_res;
6464 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6465 
6466 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6467 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6468 	    "tx");
6469 	return (0);
6470 }
6471 
6472 void
6473 iflib_led_create(if_ctx_t ctx)
6474 {
6475 
6476 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6477 	    device_get_nameunit(ctx->ifc_dev));
6478 }
6479 
6480 void
6481 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6482 {
6483 
6484 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6485 }
6486 
6487 void
6488 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6489 {
6490 
6491 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6492 }
6493 
6494 void
6495 iflib_admin_intr_deferred(if_ctx_t ctx)
6496 {
6497 
6498 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6499 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6500 }
6501 
6502 void
6503 iflib_iov_intr_deferred(if_ctx_t ctx)
6504 {
6505 
6506 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6507 }
6508 
6509 void
6510 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6511 {
6512 
6513 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6514 	    name);
6515 }
6516 
6517 void
6518 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6519 	const char *name)
6520 {
6521 
6522 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6523 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6524 	    name);
6525 }
6526 
6527 void
6528 iflib_config_gtask_deinit(struct grouptask *gtask)
6529 {
6530 
6531 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6532 }
6533 
6534 void
6535 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6536 {
6537 	if_t ifp = ctx->ifc_ifp;
6538 	iflib_txq_t txq = ctx->ifc_txqs;
6539 
6540 	if_setbaudrate(ifp, baudrate);
6541 	if (baudrate >= IF_Gbps(10)) {
6542 		STATE_LOCK(ctx);
6543 		ctx->ifc_flags |= IFC_PREFETCH;
6544 		STATE_UNLOCK(ctx);
6545 	}
6546 	/* If link down, disable watchdog */
6547 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6548 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6549 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6550 	}
6551 	ctx->ifc_link_state = link_state;
6552 	if_link_state_change(ifp, link_state);
6553 }
6554 
6555 static int
6556 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6557 {
6558 	int credits;
6559 #ifdef INVARIANTS
6560 	int credits_pre = txq->ift_cidx_processed;
6561 #endif
6562 
6563 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6564 	    BUS_DMASYNC_POSTREAD);
6565 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6566 		return (0);
6567 
6568 	txq->ift_processed += credits;
6569 	txq->ift_cidx_processed += credits;
6570 
6571 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6572 	if (txq->ift_cidx_processed >= txq->ift_size)
6573 		txq->ift_cidx_processed -= txq->ift_size;
6574 	return (credits);
6575 }
6576 
6577 static int
6578 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6579 {
6580 	iflib_fl_t fl;
6581 	u_int i;
6582 
6583 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6584 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6585 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6586 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6587 	    budget));
6588 }
6589 
6590 void
6591 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6592 	const char *description, if_int_delay_info_t info,
6593 	int offset, int value)
6594 {
6595 	info->iidi_ctx = ctx;
6596 	info->iidi_offset = offset;
6597 	info->iidi_value = value;
6598 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6599 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6600 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6601 	    info, 0, iflib_sysctl_int_delay, "I", description);
6602 }
6603 
6604 struct sx *
6605 iflib_ctx_lock_get(if_ctx_t ctx)
6606 {
6607 
6608 	return (&ctx->ifc_ctx_sx);
6609 }
6610 
6611 static int
6612 iflib_msix_init(if_ctx_t ctx)
6613 {
6614 	device_t dev = ctx->ifc_dev;
6615 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6616 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6617 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6618 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6619 
6620 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6621 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6622 
6623 	if (bootverbose)
6624 		device_printf(dev, "msix_init qsets capped at %d\n",
6625 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6626 
6627 	/* Override by tuneable */
6628 	if (scctx->isc_disable_msix)
6629 		goto msi;
6630 
6631 	/* First try MSI-X */
6632 	if ((msgs = pci_msix_count(dev)) == 0) {
6633 		if (bootverbose)
6634 			device_printf(dev, "MSI-X not supported or disabled\n");
6635 		goto msi;
6636 	}
6637 
6638 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6639 	/*
6640 	 * bar == -1 => "trust me I know what I'm doing"
6641 	 * Some drivers are for hardware that is so shoddily
6642 	 * documented that no one knows which bars are which
6643 	 * so the developer has to map all bars. This hack
6644 	 * allows shoddy garbage to use MSI-X in this framework.
6645 	 */
6646 	if (bar != -1) {
6647 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6648 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6649 		if (ctx->ifc_msix_mem == NULL) {
6650 			device_printf(dev, "Unable to map MSI-X table\n");
6651 			goto msi;
6652 		}
6653 	}
6654 
6655 	admincnt = sctx->isc_admin_intrcnt;
6656 #if IFLIB_DEBUG
6657 	/* use only 1 qset in debug mode */
6658 	queuemsgs = min(msgs - admincnt, 1);
6659 #else
6660 	queuemsgs = msgs - admincnt;
6661 #endif
6662 #ifdef RSS
6663 	queues = imin(queuemsgs, rss_getnumbuckets());
6664 #else
6665 	queues = queuemsgs;
6666 #endif
6667 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6668 	if (bootverbose)
6669 		device_printf(dev,
6670 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6671 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6672 #ifdef  RSS
6673 	/* If we're doing RSS, clamp at the number of RSS buckets */
6674 	if (queues > rss_getnumbuckets())
6675 		queues = rss_getnumbuckets();
6676 #endif
6677 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6678 		rx_queues = iflib_num_rx_queues;
6679 	else
6680 		rx_queues = queues;
6681 
6682 	if (rx_queues > scctx->isc_nrxqsets)
6683 		rx_queues = scctx->isc_nrxqsets;
6684 
6685 	/*
6686 	 * We want this to be all logical CPUs by default
6687 	 */
6688 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6689 		tx_queues = iflib_num_tx_queues;
6690 	else
6691 		tx_queues = mp_ncpus;
6692 
6693 	if (tx_queues > scctx->isc_ntxqsets)
6694 		tx_queues = scctx->isc_ntxqsets;
6695 
6696 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6697 #ifdef INVARIANTS
6698 		if (tx_queues != rx_queues)
6699 			device_printf(dev,
6700 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6701 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6702 #endif
6703 		tx_queues = min(rx_queues, tx_queues);
6704 		rx_queues = min(rx_queues, tx_queues);
6705 	}
6706 
6707 	vectors = rx_queues + admincnt;
6708 	if (msgs < vectors) {
6709 		device_printf(dev,
6710 		    "insufficient number of MSI-X vectors "
6711 		    "(supported %d, need %d)\n", msgs, vectors);
6712 		goto msi;
6713 	}
6714 
6715 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6716 	    tx_queues);
6717 	msgs = vectors;
6718 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6719 		if (vectors != msgs) {
6720 			device_printf(dev,
6721 			    "Unable to allocate sufficient MSI-X vectors "
6722 			    "(got %d, need %d)\n", vectors, msgs);
6723 			pci_release_msi(dev);
6724 			if (bar != -1) {
6725 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6726 				    ctx->ifc_msix_mem);
6727 				ctx->ifc_msix_mem = NULL;
6728 			}
6729 			goto msi;
6730 		}
6731 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6732 		    vectors);
6733 		scctx->isc_vectors = vectors;
6734 		scctx->isc_nrxqsets = rx_queues;
6735 		scctx->isc_ntxqsets = tx_queues;
6736 		scctx->isc_intr = IFLIB_INTR_MSIX;
6737 
6738 		return (vectors);
6739 	} else {
6740 		device_printf(dev,
6741 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6742 		    err);
6743 		if (bar != -1) {
6744 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6745 			    ctx->ifc_msix_mem);
6746 			ctx->ifc_msix_mem = NULL;
6747 		}
6748 	}
6749 
6750 msi:
6751 	vectors = pci_msi_count(dev);
6752 	scctx->isc_nrxqsets = 1;
6753 	scctx->isc_ntxqsets = 1;
6754 	scctx->isc_vectors = vectors;
6755 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6756 		device_printf(dev,"Using an MSI interrupt\n");
6757 		scctx->isc_intr = IFLIB_INTR_MSI;
6758 	} else {
6759 		scctx->isc_vectors = 1;
6760 		device_printf(dev,"Using a Legacy interrupt\n");
6761 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6762 	}
6763 
6764 	return (vectors);
6765 }
6766 
6767 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6768 
6769 static int
6770 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6771 {
6772 	int rc;
6773 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6774 	struct sbuf *sb;
6775 	const char *ring_state = "UNKNOWN";
6776 
6777 	/* XXX needed ? */
6778 	rc = sysctl_wire_old_buffer(req, 0);
6779 	MPASS(rc == 0);
6780 	if (rc != 0)
6781 		return (rc);
6782 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6783 	MPASS(sb != NULL);
6784 	if (sb == NULL)
6785 		return (ENOMEM);
6786 	if (state[3] <= 3)
6787 		ring_state = ring_states[state[3]];
6788 
6789 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6790 		    state[0], state[1], state[2], ring_state);
6791 	rc = sbuf_finish(sb);
6792 	sbuf_delete(sb);
6793         return(rc);
6794 }
6795 
6796 enum iflib_ndesc_handler {
6797 	IFLIB_NTXD_HANDLER,
6798 	IFLIB_NRXD_HANDLER,
6799 };
6800 
6801 static int
6802 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6803 {
6804 	if_ctx_t ctx = (void *)arg1;
6805 	enum iflib_ndesc_handler type = arg2;
6806 	char buf[256] = {0};
6807 	qidx_t *ndesc;
6808 	char *p, *next;
6809 	int nqs, rc, i;
6810 
6811 	nqs = 8;
6812 	switch(type) {
6813 	case IFLIB_NTXD_HANDLER:
6814 		ndesc = ctx->ifc_sysctl_ntxds;
6815 		if (ctx->ifc_sctx)
6816 			nqs = ctx->ifc_sctx->isc_ntxqs;
6817 		break;
6818 	case IFLIB_NRXD_HANDLER:
6819 		ndesc = ctx->ifc_sysctl_nrxds;
6820 		if (ctx->ifc_sctx)
6821 			nqs = ctx->ifc_sctx->isc_nrxqs;
6822 		break;
6823 	default:
6824 		printf("%s: unhandled type\n", __func__);
6825 		return (EINVAL);
6826 	}
6827 	if (nqs == 0)
6828 		nqs = 8;
6829 
6830 	for (i=0; i<8; i++) {
6831 		if (i >= nqs)
6832 			break;
6833 		if (i)
6834 			strcat(buf, ",");
6835 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6836 	}
6837 
6838 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6839 	if (rc || req->newptr == NULL)
6840 		return rc;
6841 
6842 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6843 	    i++, p = strsep(&next, " ,")) {
6844 		ndesc[i] = strtoul(p, NULL, 10);
6845 	}
6846 
6847 	return(rc);
6848 }
6849 
6850 #define NAME_BUFLEN 32
6851 static void
6852 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6853 {
6854         device_t dev = iflib_get_dev(ctx);
6855 	struct sysctl_oid_list *child, *oid_list;
6856 	struct sysctl_ctx_list *ctx_list;
6857 	struct sysctl_oid *node;
6858 
6859 	ctx_list = device_get_sysctl_ctx(dev);
6860 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6861 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6862 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6863 	oid_list = SYSCTL_CHILDREN(node);
6864 
6865 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6866 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6867 		       "driver version");
6868 
6869 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6870 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6871 			"# of txqs to use, 0 => use default #");
6872 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6873 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6874 			"# of rxqs to use, 0 => use default #");
6875 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6876 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6877                        "permit #txq != #rxq");
6878 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6879                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6880                       "disable MSI-X (default 0)");
6881 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6882 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6883 		       "set the RX budget");
6884 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6885 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6886 		       "cause TX to abdicate instead of running to completion");
6887 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6888 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6889 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6890 		       "offset to start using cores at");
6891 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6892 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6893 		       "use separate cores for TX and RX");
6894 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6895 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6896 		      "try to make use of logical cores for TX and RX");
6897 
6898 	/* XXX change for per-queue sizes */
6899 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6900 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6901 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6902 	    "list of # of TX descriptors to use, 0 = use default #");
6903 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6904 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6905 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6906 	    "list of # of RX descriptors to use, 0 = use default #");
6907 }
6908 
6909 static void
6910 iflib_add_device_sysctl_post(if_ctx_t ctx)
6911 {
6912 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6913 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6914         device_t dev = iflib_get_dev(ctx);
6915 	struct sysctl_oid_list *child;
6916 	struct sysctl_ctx_list *ctx_list;
6917 	iflib_fl_t fl;
6918 	iflib_txq_t txq;
6919 	iflib_rxq_t rxq;
6920 	int i, j;
6921 	char namebuf[NAME_BUFLEN];
6922 	char *qfmt;
6923 	struct sysctl_oid *queue_node, *fl_node, *node;
6924 	struct sysctl_oid_list *queue_list, *fl_list;
6925 	ctx_list = device_get_sysctl_ctx(dev);
6926 
6927 	node = ctx->ifc_sysctl_node;
6928 	child = SYSCTL_CHILDREN(node);
6929 
6930 	if (scctx->isc_ntxqsets > 100)
6931 		qfmt = "txq%03d";
6932 	else if (scctx->isc_ntxqsets > 10)
6933 		qfmt = "txq%02d";
6934 	else
6935 		qfmt = "txq%d";
6936 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6937 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6938 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6939 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6940 		queue_list = SYSCTL_CHILDREN(queue_node);
6941 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6942 			       CTLFLAG_RD,
6943 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
6944 #if MEMORY_LOGGING
6945 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6946 				CTLFLAG_RD,
6947 				&txq->ift_dequeued, "total mbufs freed");
6948 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6949 				CTLFLAG_RD,
6950 				&txq->ift_enqueued, "total mbufs enqueued");
6951 #endif
6952 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6953 				   CTLFLAG_RD,
6954 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6955 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6956 				   CTLFLAG_RD,
6957 				   &txq->ift_pullups, "# of times m_pullup was called");
6958 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6959 				   CTLFLAG_RD,
6960 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6961 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6962 				   CTLFLAG_RD,
6963 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6964 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6965 				   CTLFLAG_RD,
6966 				   &txq->ift_map_failed, "# of times DMA map failed");
6967 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6968 				   CTLFLAG_RD,
6969 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6970 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6971 				   CTLFLAG_RD,
6972 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6973 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6974 				   CTLFLAG_RD,
6975 				   &txq->ift_pidx, 1, "Producer Index");
6976 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6977 				   CTLFLAG_RD,
6978 				   &txq->ift_cidx, 1, "Consumer Index");
6979 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6980 				   CTLFLAG_RD,
6981 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6982 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6983 				   CTLFLAG_RD,
6984 				   &txq->ift_in_use, 1, "descriptors in use");
6985 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6986 				   CTLFLAG_RD,
6987 				   &txq->ift_processed, "descriptors procesed for clean");
6988 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6989 				   CTLFLAG_RD,
6990 				   &txq->ift_cleaned, "total cleaned");
6991 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6992 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6993 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6994 		    mp_ring_state_handler, "A", "soft ring state");
6995 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6996 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6997 				       "# of enqueues to the mp_ring for this queue");
6998 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6999 				       CTLFLAG_RD, &txq->ift_br->drops,
7000 				       "# of drops in the mp_ring for this queue");
7001 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
7002 				       CTLFLAG_RD, &txq->ift_br->starts,
7003 				       "# of normal consumer starts in the mp_ring for this queue");
7004 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
7005 				       CTLFLAG_RD, &txq->ift_br->stalls,
7006 					       "# of consumer stalls in the mp_ring for this queue");
7007 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7008 			       CTLFLAG_RD, &txq->ift_br->restarts,
7009 				       "# of consumer restarts in the mp_ring for this queue");
7010 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7011 				       CTLFLAG_RD, &txq->ift_br->abdications,
7012 				       "# of consumer abdications in the mp_ring for this queue");
7013 	}
7014 
7015 	if (scctx->isc_nrxqsets > 100)
7016 		qfmt = "rxq%03d";
7017 	else if (scctx->isc_nrxqsets > 10)
7018 		qfmt = "rxq%02d";
7019 	else
7020 		qfmt = "rxq%d";
7021 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7022 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7023 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7024 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7025 		queue_list = SYSCTL_CHILDREN(queue_node);
7026 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7027 			       CTLFLAG_RD,
7028 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7029 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7030 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7031 				       CTLFLAG_RD,
7032 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7033 		}
7034 
7035 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7036 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7037 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7038 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7039 			fl_list = SYSCTL_CHILDREN(fl_node);
7040 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7041 				       CTLFLAG_RD,
7042 				       &fl->ifl_pidx, 1, "Producer Index");
7043 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7044 				       CTLFLAG_RD,
7045 				       &fl->ifl_cidx, 1, "Consumer Index");
7046 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7047 				       CTLFLAG_RD,
7048 				       &fl->ifl_credits, 1, "credits available");
7049 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7050 				       CTLFLAG_RD,
7051 				       &fl->ifl_buf_size, 1, "buffer size");
7052 #if MEMORY_LOGGING
7053 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7054 					CTLFLAG_RD,
7055 					&fl->ifl_m_enqueued, "mbufs allocated");
7056 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7057 					CTLFLAG_RD,
7058 					&fl->ifl_m_dequeued, "mbufs freed");
7059 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7060 					CTLFLAG_RD,
7061 					&fl->ifl_cl_enqueued, "clusters allocated");
7062 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7063 					CTLFLAG_RD,
7064 					&fl->ifl_cl_dequeued, "clusters freed");
7065 #endif
7066 		}
7067 	}
7068 
7069 }
7070 
7071 void
7072 iflib_request_reset(if_ctx_t ctx)
7073 {
7074 
7075 	STATE_LOCK(ctx);
7076 	ctx->ifc_flags |= IFC_DO_RESET;
7077 	STATE_UNLOCK(ctx);
7078 }
7079 
7080 #ifndef __NO_STRICT_ALIGNMENT
7081 static struct mbuf *
7082 iflib_fixup_rx(struct mbuf *m)
7083 {
7084 	struct mbuf *n;
7085 
7086 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7087 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7088 		m->m_data += ETHER_HDR_LEN;
7089 		n = m;
7090 	} else {
7091 		MGETHDR(n, M_NOWAIT, MT_DATA);
7092 		if (n == NULL) {
7093 			m_freem(m);
7094 			return (NULL);
7095 		}
7096 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7097 		m->m_data += ETHER_HDR_LEN;
7098 		m->m_len -= ETHER_HDR_LEN;
7099 		n->m_len = ETHER_HDR_LEN;
7100 		M_MOVE_PKTHDR(n, m);
7101 		n->m_next = m;
7102 	}
7103 	return (n);
7104 }
7105 #endif
7106 
7107 #ifdef DEBUGNET
7108 static void
7109 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7110 {
7111 	if_ctx_t ctx;
7112 
7113 	ctx = if_getsoftc(ifp);
7114 	CTX_LOCK(ctx);
7115 	*nrxr = NRXQSETS(ctx);
7116 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7117 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7118 	CTX_UNLOCK(ctx);
7119 }
7120 
7121 static void
7122 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7123 {
7124 	if_ctx_t ctx;
7125 	if_softc_ctx_t scctx;
7126 	iflib_fl_t fl;
7127 	iflib_rxq_t rxq;
7128 	int i, j;
7129 
7130 	ctx = if_getsoftc(ifp);
7131 	scctx = &ctx->ifc_softc_ctx;
7132 
7133 	switch (event) {
7134 	case DEBUGNET_START:
7135 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7136 			rxq = &ctx->ifc_rxqs[i];
7137 			for (j = 0; j < rxq->ifr_nfl; j++) {
7138 				fl = rxq->ifr_fl;
7139 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7140 			}
7141 		}
7142 		iflib_no_tx_batch = 1;
7143 		break;
7144 	default:
7145 		break;
7146 	}
7147 }
7148 
7149 static int
7150 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7151 {
7152 	if_ctx_t ctx;
7153 	iflib_txq_t txq;
7154 	int error;
7155 
7156 	ctx = if_getsoftc(ifp);
7157 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7158 	    IFF_DRV_RUNNING)
7159 		return (EBUSY);
7160 
7161 	txq = &ctx->ifc_txqs[0];
7162 	error = iflib_encap(txq, &m);
7163 	if (error == 0)
7164 		(void)iflib_txd_db_check(txq, true);
7165 	return (error);
7166 }
7167 
7168 static int
7169 iflib_debugnet_poll(if_t ifp, int count)
7170 {
7171 	struct epoch_tracker et;
7172 	if_ctx_t ctx;
7173 	if_softc_ctx_t scctx;
7174 	iflib_txq_t txq;
7175 	int i;
7176 
7177 	ctx = if_getsoftc(ifp);
7178 	scctx = &ctx->ifc_softc_ctx;
7179 
7180 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7181 	    IFF_DRV_RUNNING)
7182 		return (EBUSY);
7183 
7184 	txq = &ctx->ifc_txqs[0];
7185 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7186 
7187 	NET_EPOCH_ENTER(et);
7188 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7189 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7190 	NET_EPOCH_EXIT(et);
7191 	return (0);
7192 }
7193 #endif /* DEBUGNET */
7194