xref: /freebsd/sys/net/iflib.c (revision 7a7741af18d6c8a804cc643cb7ecda9d730c6aa6)
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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_acpi.h"
32 #include "opt_sched.h"
33 
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/bus.h>
37 #include <sys/eventhandler.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/module.h>
42 #include <sys/kobj.h>
43 #include <sys/rman.h>
44 #include <sys/sbuf.h>
45 #include <sys/smp.h>
46 #include <sys/socket.h>
47 #include <sys/sockio.h>
48 #include <sys/sysctl.h>
49 #include <sys/syslog.h>
50 #include <sys/taskqueue.h>
51 #include <sys/limits.h>
52 
53 #include <net/if.h>
54 #include <net/if_var.h>
55 #include <net/if_private.h>
56 #include <net/if_types.h>
57 #include <net/if_media.h>
58 #include <net/bpf.h>
59 #include <net/ethernet.h>
60 #include <net/mp_ring.h>
61 #include <net/debugnet.h>
62 #include <net/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #ifdef PCI_IOV
92 #include <dev/pci/pci_iov.h>
93 #endif
94 
95 #include <sys/bitstring.h>
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of
98  * iflib's software descriptor references
99  */
100 #define MEMORY_LOGGING 0
101 /*
102  * Enable mbuf vectors for compressing long mbuf chains
103  */
104 
105 /*
106  * NB:
107  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108  *   we prefetch needs to be determined by the time spent in m_free vis a vis
109  *   the cost of a prefetch. This will of course vary based on the workload:
110  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111  *        is quite expensive, thus suggesting very little prefetch.
112  *      - small packet forwarding which is just returning a single mbuf to
113  *        UMA will typically be very fast vis a vis the cost of a memory
114  *        access.
115  */
116 
117 /*
118  * File organization:
119  *  - private structures
120  *  - iflib private utility functions
121  *  - ifnet functions
122  *  - vlan registry and other exported functions
123  *  - iflib public core functions
124  *
125  *
126  */
127 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
128 
129 #define	IFLIB_RXEOF_MORE	(1U << 0)
130 #define	IFLIB_RXEOF_EMPTY	(2U << 0)
131 
132 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 static void iflib_tqg_detach(if_ctx_t ctx);
144 
145 typedef struct iflib_filter_info {
146 	driver_filter_t *ifi_filter;
147 	void *ifi_filter_arg;
148 	struct grouptask *ifi_task;
149 	void *ifi_ctx;
150 } *iflib_filter_info_t;
151 
152 struct iflib_ctx {
153 	KOBJ_FIELDS;
154 	/*
155 	 * Pointer to hardware driver's softc
156 	 */
157 	void *ifc_softc;
158 	device_t ifc_dev;
159 	if_t ifc_ifp;
160 
161 	cpuset_t ifc_cpus;
162 	if_shared_ctx_t ifc_sctx;
163 	struct if_softc_ctx ifc_softc_ctx;
164 
165 	struct sx ifc_ctx_sx;
166 	struct mtx ifc_state_mtx;
167 
168 	iflib_txq_t ifc_txqs;
169 	iflib_rxq_t ifc_rxqs;
170 	uint32_t ifc_if_flags;
171 	uint32_t ifc_flags;
172 	uint32_t ifc_max_fl_buf_size;
173 	uint32_t ifc_rx_mbuf_sz;
174 
175 	int ifc_link_state;
176 	int ifc_watchdog_events;
177 	struct cdev *ifc_led_dev;
178 	struct resource *ifc_msix_mem;
179 
180 	struct if_irq ifc_legacy_irq;
181 	struct grouptask ifc_admin_task;
182 	struct grouptask ifc_vflr_task;
183 	struct iflib_filter_info ifc_filter_info;
184 	struct ifmedia	ifc_media;
185 	struct ifmedia	*ifc_mediap;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 	uint16_t ifc_sysctl_core_offset;
194 #define	CORE_OFFSET_UNSPECIFIED	0xffff
195 	uint8_t  ifc_sysctl_separate_txrx;
196 	uint8_t  ifc_sysctl_use_logical_cores;
197 	uint16_t ifc_sysctl_extra_msix_vectors;
198 	bool	 ifc_cpus_are_physical_cores;
199 
200 	qidx_t ifc_sysctl_ntxds[8];
201 	qidx_t ifc_sysctl_nrxds[8];
202 	struct if_txrx ifc_txrx;
203 #define isc_txd_encap		ifc_txrx.ift_txd_encap
204 #define isc_txd_flush		ifc_txrx.ift_txd_flush
205 #define isc_txd_credits_update	ifc_txrx.ift_txd_credits_update
206 #define isc_rxd_available	ifc_txrx.ift_rxd_available
207 #define isc_rxd_pkt_get		ifc_txrx.ift_rxd_pkt_get
208 #define isc_rxd_refill		ifc_txrx.ift_rxd_refill
209 #define isc_rxd_flush		ifc_txrx.ift_rxd_flush
210 #define isc_legacy_intr		ifc_txrx.ift_legacy_intr
211 #define isc_txq_select		ifc_txrx.ift_txq_select
212 #define isc_txq_select_v2	ifc_txrx.ift_txq_select_v2
213 
214 	eventhandler_tag ifc_vlan_attach_event;
215 	eventhandler_tag ifc_vlan_detach_event;
216 	struct ether_addr ifc_mac;
217 };
218 
219 void *
220 iflib_get_softc(if_ctx_t ctx)
221 {
222 
223 	return (ctx->ifc_softc);
224 }
225 
226 device_t
227 iflib_get_dev(if_ctx_t ctx)
228 {
229 
230 	return (ctx->ifc_dev);
231 }
232 
233 if_t
234 iflib_get_ifp(if_ctx_t ctx)
235 {
236 
237 	return (ctx->ifc_ifp);
238 }
239 
240 struct ifmedia *
241 iflib_get_media(if_ctx_t ctx)
242 {
243 
244 	return (ctx->ifc_mediap);
245 }
246 
247 void
248 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
249 {
250 
251 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
252 }
253 
254 if_softc_ctx_t
255 iflib_get_softc_ctx(if_ctx_t ctx)
256 {
257 
258 	return (&ctx->ifc_softc_ctx);
259 }
260 
261 if_shared_ctx_t
262 iflib_get_sctx(if_ctx_t ctx)
263 {
264 
265 	return (ctx->ifc_sctx);
266 }
267 
268 uint16_t
269 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
270 {
271 
272 	return (ctx->ifc_sysctl_extra_msix_vectors);
273 }
274 
275 #define IP_ALIGNED(m)		((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
276 #define CACHE_PTR_INCREMENT	(CACHE_LINE_SIZE / sizeof(void *))
277 #define CACHE_PTR_NEXT(ptr)	((void *)(((uintptr_t)(ptr) + CACHE_LINE_SIZE - 1) & (CACHE_LINE_SIZE - 1)))
278 
279 #define LINK_ACTIVE(ctx)	((ctx)->ifc_link_state == LINK_STATE_UP)
280 #define CTX_IS_VF(ctx)		((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
281 
282 typedef struct iflib_sw_rx_desc_array {
283 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
284 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
285 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
286 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
287 } iflib_rxsd_array_t;
288 
289 typedef struct iflib_sw_tx_desc_array {
290 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
291 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
292 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
293 } if_txsd_vec_t;
294 
295 /* magic number that should be high enough for any hardware */
296 #define IFLIB_MAX_TX_SEGS		128
297 #define IFLIB_RX_COPY_THRESH		128
298 #define IFLIB_MAX_RX_REFRESH		32
299 /* The minimum descriptors per second before we start coalescing */
300 #define IFLIB_MIN_DESC_SEC		16384
301 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
302 #define IFLIB_QUEUE_IDLE		0
303 #define IFLIB_QUEUE_HUNG		1
304 #define IFLIB_QUEUE_WORKING		2
305 /* maximum number of txqs that can share an rx interrupt */
306 #define IFLIB_MAX_TX_SHARED_INTR	4
307 
308 /* this should really scale with ring size - this is a fairly arbitrary value */
309 #define TX_BATCH_SIZE			32
310 
311 #define IFLIB_RESTART_BUDGET		8
312 
313 #define	IFC_LEGACY		0x001
314 #define	IFC_QFLUSH		0x002
315 #define	IFC_MULTISEG		0x004
316 #define	IFC_SPARE1		0x008
317 #define	IFC_SC_ALLOCATED	0x010
318 #define	IFC_INIT_DONE		0x020
319 #define	IFC_PREFETCH		0x040
320 #define	IFC_DO_RESET		0x080
321 #define	IFC_DO_WATCHDOG		0x100
322 #define	IFC_SPARE0		0x200
323 #define	IFC_SPARE2		0x400
324 #define	IFC_IN_DETACH		0x800
325 
326 #define	IFC_NETMAP_TX_IRQ	0x80000000
327 
328 #define CSUM_OFFLOAD		(CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \
329 				 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \
330 				 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP)
331 
332 struct iflib_txq {
333 	qidx_t		ift_in_use;
334 	qidx_t		ift_cidx;
335 	qidx_t		ift_cidx_processed;
336 	qidx_t		ift_pidx;
337 	uint8_t		ift_gen;
338 	uint8_t		ift_br_offset;
339 	uint16_t	ift_npending;
340 	uint16_t	ift_db_pending;
341 	uint16_t	ift_rs_pending;
342 	/* implicit pad */
343 	uint8_t		ift_txd_size[8];
344 	uint64_t	ift_processed;
345 	uint64_t	ift_cleaned;
346 	uint64_t	ift_cleaned_prev;
347 #if MEMORY_LOGGING
348 	uint64_t	ift_enqueued;
349 	uint64_t	ift_dequeued;
350 #endif
351 	uint64_t	ift_no_tx_dma_setup;
352 	uint64_t	ift_no_desc_avail;
353 	uint64_t	ift_mbuf_defrag_failed;
354 	uint64_t	ift_mbuf_defrag;
355 	uint64_t	ift_map_failed;
356 	uint64_t	ift_txd_encap_efbig;
357 	uint64_t	ift_pullups;
358 	uint64_t	ift_last_timer_tick;
359 
360 	struct mtx	ift_mtx;
361 	struct mtx	ift_db_mtx;
362 
363 	/* constant values */
364 	if_ctx_t	ift_ctx;
365 	struct ifmp_ring        *ift_br;
366 	struct grouptask	ift_task;
367 	qidx_t		ift_size;
368 	uint16_t	ift_id;
369 	struct callout	ift_timer;
370 #ifdef DEV_NETMAP
371 	struct callout	ift_netmap_timer;
372 #endif /* DEV_NETMAP */
373 
374 	if_txsd_vec_t	ift_sds;
375 	uint8_t		ift_qstatus;
376 	uint8_t		ift_closed;
377 	uint8_t		ift_update_freq;
378 	struct iflib_filter_info ift_filter_info;
379 	bus_dma_tag_t	ift_buf_tag;
380 	bus_dma_tag_t	ift_tso_buf_tag;
381 	iflib_dma_info_t	ift_ifdi;
382 #define	MTX_NAME_LEN	32
383 	char                    ift_mtx_name[MTX_NAME_LEN];
384 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
385 #ifdef IFLIB_DIAGNOSTICS
386 	uint64_t ift_cpu_exec_count[256];
387 #endif
388 } __aligned(CACHE_LINE_SIZE);
389 
390 struct iflib_fl {
391 	qidx_t		ifl_cidx;
392 	qidx_t		ifl_pidx;
393 	qidx_t		ifl_credits;
394 	uint8_t		ifl_gen;
395 	uint8_t		ifl_rxd_size;
396 #if MEMORY_LOGGING
397 	uint64_t	ifl_m_enqueued;
398 	uint64_t	ifl_m_dequeued;
399 	uint64_t	ifl_cl_enqueued;
400 	uint64_t	ifl_cl_dequeued;
401 #endif
402 	/* implicit pad */
403 	bitstr_t 	*ifl_rx_bitmap;
404 	qidx_t		ifl_fragidx;
405 	/* constant */
406 	qidx_t		ifl_size;
407 	uint16_t	ifl_buf_size;
408 	uint16_t	ifl_cltype;
409 	uma_zone_t	ifl_zone;
410 	iflib_rxsd_array_t	ifl_sds;
411 	iflib_rxq_t	ifl_rxq;
412 	uint8_t		ifl_id;
413 	bus_dma_tag_t	ifl_buf_tag;
414 	iflib_dma_info_t	ifl_ifdi;
415 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
416 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
417 }  __aligned(CACHE_LINE_SIZE);
418 
419 static inline qidx_t
420 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
421 {
422 	qidx_t used;
423 
424 	if (pidx > cidx)
425 		used = pidx - cidx;
426 	else if (pidx < cidx)
427 		used = size - cidx + pidx;
428 	else if (gen == 0 && pidx == cidx)
429 		used = 0;
430 	else if (gen == 1 && pidx == cidx)
431 		used = size;
432 	else
433 		panic("bad state");
434 
435 	return (used);
436 }
437 
438 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
439 
440 #define IDXDIFF(head, tail, wrap) \
441 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
442 
443 struct iflib_rxq {
444 	if_ctx_t	ifr_ctx;
445 	iflib_fl_t	ifr_fl;
446 	uint64_t	ifr_rx_irq;
447 	struct pfil_head	*pfil;
448 	/*
449 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
450 	 * the completion queue consumer index.  Otherwise it's unused.
451 	 */
452 	qidx_t		ifr_cq_cidx;
453 	uint16_t	ifr_id;
454 	uint8_t		ifr_nfl;
455 	uint8_t		ifr_ntxqirq;
456 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
457 	uint8_t		ifr_fl_offset;
458 	struct lro_ctrl		ifr_lc;
459 	struct grouptask        ifr_task;
460 	struct callout		ifr_watchdog;
461 	struct iflib_filter_info ifr_filter_info;
462 	iflib_dma_info_t	ifr_ifdi;
463 
464 	/* dynamically allocate if any drivers need a value substantially larger than this */
465 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
466 #ifdef IFLIB_DIAGNOSTICS
467 	uint64_t ifr_cpu_exec_count[256];
468 #endif
469 }  __aligned(CACHE_LINE_SIZE);
470 
471 typedef struct if_rxsd {
472 	caddr_t *ifsd_cl;
473 	iflib_fl_t ifsd_fl;
474 } *if_rxsd_t;
475 
476 /* multiple of word size */
477 #ifdef __LP64__
478 #define PKT_INFO_SIZE	6
479 #define RXD_INFO_SIZE	5
480 #define PKT_TYPE uint64_t
481 #else
482 #define PKT_INFO_SIZE	11
483 #define RXD_INFO_SIZE	8
484 #define PKT_TYPE uint32_t
485 #endif
486 #define PKT_LOOP_BOUND	((PKT_INFO_SIZE / 3) * 3)
487 #define RXD_LOOP_BOUND	((RXD_INFO_SIZE / 4) * 4)
488 
489 typedef struct if_pkt_info_pad {
490 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
491 } *if_pkt_info_pad_t;
492 typedef struct if_rxd_info_pad {
493 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
494 } *if_rxd_info_pad_t;
495 
496 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
497 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
498 
499 static inline void
500 pkt_info_zero(if_pkt_info_t pi)
501 {
502 	if_pkt_info_pad_t pi_pad;
503 
504 	pi_pad = (if_pkt_info_pad_t)pi;
505 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
506 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
507 #ifndef __LP64__
508 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
509 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
510 #endif
511 }
512 
513 static inline void
514 rxd_info_zero(if_rxd_info_t ri)
515 {
516 	if_rxd_info_pad_t ri_pad;
517 	int i;
518 
519 	ri_pad = (if_rxd_info_pad_t)ri;
520 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
521 		ri_pad->rxd_val[i] = 0;
522 		ri_pad->rxd_val[i + 1] = 0;
523 		ri_pad->rxd_val[i + 2] = 0;
524 		ri_pad->rxd_val[i + 3] = 0;
525 	}
526 #ifdef __LP64__
527 	ri_pad->rxd_val[RXD_INFO_SIZE - 1] = 0;
528 #endif
529 }
530 
531 /*
532  * Only allow a single packet to take up most 1/nth of the tx ring
533  */
534 #define MAX_SINGLE_PACKET_FRACTION 12
535 #define IF_BAD_DMA	((bus_addr_t)-1)
536 
537 #define CTX_ACTIVE(ctx)	((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
538 
539 #define CTX_LOCK_INIT(_sc)	sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
540 #define CTX_LOCK(ctx)		sx_xlock(&(ctx)->ifc_ctx_sx)
541 #define CTX_UNLOCK(ctx)		sx_xunlock(&(ctx)->ifc_ctx_sx)
542 #define CTX_LOCK_DESTROY(ctx)	sx_destroy(&(ctx)->ifc_ctx_sx)
543 
544 #define STATE_LOCK_INIT(_sc, _name)	mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
545 #define STATE_LOCK(ctx)		mtx_lock(&(ctx)->ifc_state_mtx)
546 #define STATE_UNLOCK(ctx)	mtx_unlock(&(ctx)->ifc_state_mtx)
547 #define STATE_LOCK_DESTROY(ctx)	mtx_destroy(&(ctx)->ifc_state_mtx)
548 
549 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
550 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
551 
552 /* Our boot-time initialization hook */
553 static int	iflib_module_event_handler(module_t, int, void *);
554 
555 static moduledata_t iflib_moduledata = {
556 	"iflib",
557 	iflib_module_event_handler,
558 	NULL
559 };
560 
561 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
562 MODULE_VERSION(iflib, 1);
563 
564 MODULE_DEPEND(iflib, pci, 1, 1, 1);
565 MODULE_DEPEND(iflib, ether, 1, 1, 1);
566 
567 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
568 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
569 
570 #ifndef IFLIB_DEBUG_COUNTERS
571 #ifdef INVARIANTS
572 #define IFLIB_DEBUG_COUNTERS 1
573 #else
574 #define IFLIB_DEBUG_COUNTERS 0
575 #endif /* !INVARIANTS */
576 #endif
577 
578 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
579     "iflib driver parameters");
580 
581 /*
582  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
583  */
584 static int iflib_min_tx_latency = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
586     &iflib_min_tx_latency, 0,
587     "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,
591     "minimize transmit latency at the possible expense of throughput");
592 static int iflib_timer_default = 1000;
593 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
594     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
595 
596 
597 #if IFLIB_DEBUG_COUNTERS
598 
599 static int iflib_tx_seen;
600 static int iflib_tx_sent;
601 static int iflib_tx_encap;
602 static int iflib_rx_allocs;
603 static int iflib_fl_refills;
604 static int iflib_fl_refills_large;
605 static int iflib_tx_frees;
606 
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
608     "# TX mbufs seen");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
610     "# TX mbufs sent");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
612     "# TX mbufs encapped");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
614     "# TX frees");
615 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
616     "# RX allocations");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
618     "# refills");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
620     &iflib_fl_refills_large, 0, "# large refills");
621 
622 static int iflib_txq_drain_flushing;
623 static int iflib_txq_drain_oactive;
624 static int iflib_txq_drain_notready;
625 
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
627     &iflib_txq_drain_flushing, 0, "# drain flushes");
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
629     &iflib_txq_drain_oactive, 0, "# drain oactives");
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
631     &iflib_txq_drain_notready, 0, "# drain notready");
632 
633 static int iflib_encap_load_mbuf_fail;
634 static int iflib_encap_pad_mbuf_fail;
635 static int iflib_encap_txq_avail_fail;
636 static int iflib_encap_txd_encap_fail;
637 
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
639     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
641     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
643     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
645     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
646 
647 static int iflib_task_fn_rxs;
648 static int iflib_rx_intr_enables;
649 static int iflib_fast_intrs;
650 static int iflib_rx_unavail;
651 static int iflib_rx_ctx_inactive;
652 static int iflib_rx_if_input;
653 static int iflib_rxd_flush;
654 
655 static int iflib_verbose_debug;
656 
657 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
658     "# task_fn_rx calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
660     &iflib_rx_intr_enables, 0, "# RX intr enables");
661 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
662     "# fast_intr calls");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
664     "# times rxeof called with no available data");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
666     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
668     0, "# times rxeof called if_input");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
670     "# times rxd_flush called");
671 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
672     &iflib_verbose_debug, 0, "enable verbose debugging");
673 
674 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
675 static void
676 iflib_debug_reset(void)
677 {
678 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
679 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
680 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
681 		iflib_txq_drain_notready =
682 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
683 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
684 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
685 		iflib_rx_unavail =
686 		iflib_rx_ctx_inactive = iflib_rx_if_input =
687 		iflib_rxd_flush = 0;
688 }
689 
690 #else
691 #define DBG_COUNTER_INC(name)
692 static void iflib_debug_reset(void) {}
693 #endif
694 
695 #define IFLIB_DEBUG 0
696 
697 static void iflib_tx_structures_free(if_ctx_t ctx);
698 static void iflib_rx_structures_free(if_ctx_t ctx);
699 static int iflib_queues_alloc(if_ctx_t ctx);
700 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
701 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
702 static int iflib_qset_structures_setup(if_ctx_t ctx);
703 static int iflib_msix_init(if_ctx_t ctx);
704 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
705 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
706 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
707 #ifdef ALTQ
708 static void iflib_altq_if_start(if_t ifp);
709 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
710 #endif
711 static int iflib_register(if_ctx_t);
712 static void iflib_deregister(if_ctx_t);
713 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
714 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
715 static void iflib_init_locked(if_ctx_t ctx);
716 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
717 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
718 static void iflib_ifmp_purge(iflib_txq_t txq);
719 static void _iflib_pre_assert(if_softc_ctx_t scctx);
720 static void iflib_stop(if_ctx_t ctx);
721 static void iflib_if_init_locked(if_ctx_t ctx);
722 static void iflib_free_intr_mem(if_ctx_t ctx);
723 #ifndef __NO_STRICT_ALIGNMENT
724 static struct mbuf *iflib_fixup_rx(struct mbuf *m);
725 #endif
726 
727 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
728     SLIST_HEAD_INITIALIZER(cpu_offsets);
729 struct cpu_offset {
730 	SLIST_ENTRY(cpu_offset) entries;
731 	cpuset_t	set;
732 	unsigned int	refcount;
733 	uint16_t	next_cpuid;
734 };
735 static struct mtx cpu_offset_mtx;
736 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
737     MTX_DEF);
738 
739 DEBUGNET_DEFINE(iflib);
740 
741 static int
742 iflib_num_rx_descs(if_ctx_t ctx)
743 {
744 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
745 	if_shared_ctx_t sctx = ctx->ifc_sctx;
746 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
747 
748 	return (scctx->isc_nrxd[first_rxq]);
749 }
750 
751 static int
752 iflib_num_tx_descs(if_ctx_t ctx)
753 {
754 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
755 	if_shared_ctx_t sctx = ctx->ifc_sctx;
756 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
757 
758 	return (scctx->isc_ntxd[first_txq]);
759 }
760 
761 #ifdef DEV_NETMAP
762 #include <sys/selinfo.h>
763 #include <net/netmap.h>
764 #include <dev/netmap/netmap_kern.h>
765 
766 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
767 
768 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
769 static void iflib_netmap_timer(void *arg);
770 
771 /*
772  * device-specific sysctl variables:
773  *
774  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
775  *	During regular operations the CRC is stripped, but on some
776  *	hardware reception of frames not multiple of 64 is slower,
777  *	so using crcstrip=0 helps in benchmarks.
778  *
779  * iflib_rx_miss, iflib_rx_miss_bufs:
780  *	count packets that might be missed due to lost interrupts.
781  */
782 SYSCTL_DECL(_dev_netmap);
783 /*
784  * The xl driver by default strips CRCs and we do not override it.
785  */
786 
787 int iflib_crcstrip = 1;
788 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
789     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
790 
791 int iflib_rx_miss, iflib_rx_miss_bufs;
792 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
793     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
794 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
795     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
796 
797 /*
798  * Register/unregister. We are already under netmap lock.
799  * Only called on the first register or the last unregister.
800  */
801 static int
802 iflib_netmap_register(struct netmap_adapter *na, int onoff)
803 {
804 	if_t ifp = na->ifp;
805 	if_ctx_t ctx = if_getsoftc(ifp);
806 	int status;
807 
808 	CTX_LOCK(ctx);
809 	if (!CTX_IS_VF(ctx))
810 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
811 
812 	iflib_stop(ctx);
813 
814 	/*
815 	 * Enable (or disable) netmap flags, and intercept (or restore)
816 	 * ifp->if_transmit. This is done once the device has been stopped
817 	 * to prevent race conditions. Also, this must be done after
818 	 * calling netmap_disable_all_rings() and before calling
819 	 * netmap_enable_all_rings(), so that these two functions see the
820 	 * updated state of the NAF_NETMAP_ON bit.
821 	 */
822 	if (onoff) {
823 		nm_set_native_flags(na);
824 	} else {
825 		nm_clear_native_flags(na);
826 	}
827 
828 	iflib_init_locked(ctx);
829 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
830 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
831 	if (status)
832 		nm_clear_native_flags(na);
833 	CTX_UNLOCK(ctx);
834 	return (status);
835 }
836 
837 static int
838 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
839 {
840 	if_t ifp = na->ifp;
841 	if_ctx_t ctx = if_getsoftc(ifp);
842 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
843 	iflib_fl_t fl = &rxq->ifr_fl[0];
844 
845 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
846 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
847 	info->num_tx_descs = iflib_num_tx_descs(ctx);
848 	info->num_rx_descs = iflib_num_rx_descs(ctx);
849 	info->rx_buf_maxsize = fl->ifl_buf_size;
850 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
851 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
852 		info->num_rx_descs, info->rx_buf_maxsize);
853 
854 	return (0);
855 }
856 
857 static int
858 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
859 {
860 	struct netmap_adapter *na = kring->na;
861 	u_int const lim = kring->nkr_num_slots - 1;
862 	struct netmap_ring *ring = kring->ring;
863 	bus_dmamap_t *map;
864 	struct if_rxd_update iru;
865 	if_ctx_t ctx = rxq->ifr_ctx;
866 	iflib_fl_t fl = &rxq->ifr_fl[0];
867 	u_int nic_i_first, nic_i;
868 	u_int nm_i;
869 	int i, n;
870 #if IFLIB_DEBUG_COUNTERS
871 	int rf_count = 0;
872 #endif
873 
874 	/*
875 	 * This function is used both at initialization and in rxsync.
876 	 * At initialization we need to prepare (with isc_rxd_refill())
877 	 * all the netmap buffers currently owned by the kernel, in
878 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
879 	 * (except for kring->nkr_hwofs). These may be less than
880 	 * kring->nkr_num_slots if netmap_reset() was called while
881 	 * an application using the kring that still owned some
882 	 * buffers.
883 	 * At rxsync time, both indexes point to the next buffer to be
884 	 * refilled.
885 	 * In any case we publish (with isc_rxd_flush()) up to
886 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
887 	 * pointer to overrun the head/cons pointer, although this is
888 	 * not necessary for some NICs (e.g. vmx).
889 	 */
890 	if (__predict_false(init)) {
891 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
892 	} else {
893 		n = kring->rhead - kring->nr_hwcur;
894 		if (n == 0)
895 			return (0); /* Nothing to do. */
896 		if (n < 0)
897 			n += kring->nkr_num_slots;
898 	}
899 
900 	iru_init(&iru, rxq, 0 /* flid */);
901 	map = fl->ifl_sds.ifsd_map;
902 	nic_i = fl->ifl_pidx;
903 	nm_i = netmap_idx_n2k(kring, nic_i);
904 	if (__predict_false(init)) {
905 		/*
906 		 * On init/reset, nic_i must be 0, and we must
907 		 * start to refill from hwtail (see netmap_reset()).
908 		 */
909 		MPASS(nic_i == 0);
910 		MPASS(nm_i == kring->nr_hwtail);
911 	} else
912 		MPASS(nm_i == kring->nr_hwcur);
913 	DBG_COUNTER_INC(fl_refills);
914 	while (n > 0) {
915 #if IFLIB_DEBUG_COUNTERS
916 		if (++rf_count == 9)
917 			DBG_COUNTER_INC(fl_refills_large);
918 #endif
919 		nic_i_first = nic_i;
920 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
921 			struct netmap_slot *slot = &ring->slot[nm_i];
922 			uint64_t paddr;
923 			void *addr = PNMB(na, slot, &paddr);
924 
925 			MPASS(i < IFLIB_MAX_RX_REFRESH);
926 
927 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
928 			        return (netmap_ring_reinit(kring));
929 
930 			fl->ifl_bus_addrs[i] = paddr +
931 			    nm_get_offset(kring, slot);
932 			fl->ifl_rxd_idxs[i] = nic_i;
933 
934 			if (__predict_false(init)) {
935 				netmap_load_map(na, fl->ifl_buf_tag,
936 				    map[nic_i], addr);
937 			} else if (slot->flags & NS_BUF_CHANGED) {
938 				/* buffer has changed, reload map */
939 				netmap_reload_map(na, fl->ifl_buf_tag,
940 				    map[nic_i], addr);
941 			}
942 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
943 			    BUS_DMASYNC_PREREAD);
944 			slot->flags &= ~NS_BUF_CHANGED;
945 
946 			nm_i = nm_next(nm_i, lim);
947 			nic_i = nm_next(nic_i, lim);
948 		}
949 
950 		iru.iru_pidx = nic_i_first;
951 		iru.iru_count = i;
952 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
953 	}
954 	fl->ifl_pidx = nic_i;
955 	/*
956 	 * At the end of the loop we must have refilled everything
957 	 * we could possibly refill.
958 	 */
959 	MPASS(nm_i == kring->rhead);
960 	kring->nr_hwcur = nm_i;
961 
962 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
963 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
964 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
965 	    nm_prev(nic_i, lim));
966 	DBG_COUNTER_INC(rxd_flush);
967 
968 	return (0);
969 }
970 
971 #define NETMAP_TX_TIMER_US	90
972 
973 /*
974  * Reconcile kernel and user view of the transmit ring.
975  *
976  * All information is in the kring.
977  * Userspace wants to send packets up to the one before kring->rhead,
978  * kernel knows kring->nr_hwcur is the first unsent packet.
979  *
980  * Here we push packets out (as many as possible), and possibly
981  * reclaim buffers from previously completed transmission.
982  *
983  * The caller (netmap) guarantees that there is only one instance
984  * running at any time. Any interference with other driver
985  * methods should be handled by the individual drivers.
986  */
987 static int
988 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
989 {
990 	struct netmap_adapter *na = kring->na;
991 	if_t ifp = na->ifp;
992 	struct netmap_ring *ring = kring->ring;
993 	u_int nm_i;	/* index into the netmap kring */
994 	u_int nic_i;	/* index into the NIC ring */
995 	u_int const lim = kring->nkr_num_slots - 1;
996 	u_int const head = kring->rhead;
997 	struct if_pkt_info pi;
998 	int tx_pkts = 0, tx_bytes = 0;
999 
1000 	/*
1001 	 * interrupts on every tx packet are expensive so request
1002 	 * them every half ring, or where NS_REPORT is set
1003 	 */
1004 	u_int report_frequency = kring->nkr_num_slots >> 1;
1005 	/* device-specific */
1006 	if_ctx_t ctx = if_getsoftc(ifp);
1007 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1008 
1009 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1010 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1011 
1012 	/*
1013 	 * First part: process new packets to send.
1014 	 * nm_i is the current index in the netmap kring,
1015 	 * nic_i is the corresponding index in the NIC ring.
1016 	 *
1017 	 * If we have packets to send (nm_i != head)
1018 	 * iterate over the netmap ring, fetch length and update
1019 	 * the corresponding slot in the NIC ring. Some drivers also
1020 	 * need to update the buffer's physical address in the NIC slot
1021 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1022 	 *
1023 	 * The netmap_reload_map() calls is especially expensive,
1024 	 * even when (as in this case) the tag is 0, so do only
1025 	 * when the buffer has actually changed.
1026 	 *
1027 	 * If possible do not set the report/intr bit on all slots,
1028 	 * but only a few times per ring or when NS_REPORT is set.
1029 	 *
1030 	 * Finally, on 10G and faster drivers, it might be useful
1031 	 * to prefetch the next slot and txr entry.
1032 	 */
1033 
1034 	nm_i = kring->nr_hwcur;
1035 	if (nm_i != head) {	/* we have new packets to send */
1036 		uint32_t pkt_len = 0, seg_idx = 0;
1037 		int nic_i_start = -1, flags = 0;
1038 		pkt_info_zero(&pi);
1039 		pi.ipi_segs = txq->ift_segs;
1040 		pi.ipi_qsidx = kring->ring_id;
1041 		nic_i = netmap_idx_k2n(kring, nm_i);
1042 
1043 		__builtin_prefetch(&ring->slot[nm_i]);
1044 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1045 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1046 
1047 		while (nm_i != head) {
1048 			struct netmap_slot *slot = &ring->slot[nm_i];
1049 			uint64_t offset = nm_get_offset(kring, slot);
1050 			u_int len = slot->len;
1051 			uint64_t paddr;
1052 			void *addr = PNMB(na, slot, &paddr);
1053 
1054 			flags |= (slot->flags & NS_REPORT ||
1055 				nic_i == 0 || nic_i == report_frequency) ?
1056 				IPI_TX_INTR : 0;
1057 
1058 			/*
1059 			 * If this is the first packet fragment, save the
1060 			 * index of the first NIC slot for later.
1061 			 */
1062 			if (nic_i_start < 0)
1063 				nic_i_start = nic_i;
1064 
1065 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1066 			pi.ipi_segs[seg_idx].ds_len = len;
1067 			if (len) {
1068 				pkt_len += len;
1069 				seg_idx++;
1070 			}
1071 
1072 			if (!(slot->flags & NS_MOREFRAG)) {
1073 				pi.ipi_len = pkt_len;
1074 				pi.ipi_nsegs = seg_idx;
1075 				pi.ipi_pidx = nic_i_start;
1076 				pi.ipi_ndescs = 0;
1077 				pi.ipi_flags = flags;
1078 
1079 				/* Prepare the NIC TX ring. */
1080 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1081 				DBG_COUNTER_INC(tx_encap);
1082 
1083 				/* Update transmit counters */
1084 				tx_bytes += pi.ipi_len;
1085 				tx_pkts++;
1086 
1087 				/* Reinit per-packet info for the next one. */
1088 				flags = seg_idx = pkt_len = 0;
1089 				nic_i_start = -1;
1090 			}
1091 
1092 			/* prefetch for next round */
1093 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1094 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1095 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1096 
1097 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1098 
1099 			if (slot->flags & NS_BUF_CHANGED) {
1100 				/* buffer has changed, reload map */
1101 				netmap_reload_map(na, txq->ift_buf_tag,
1102 				    txq->ift_sds.ifsd_map[nic_i], addr);
1103 			}
1104 			/* make sure changes to the buffer are synced */
1105 			bus_dmamap_sync(txq->ift_buf_tag,
1106 			    txq->ift_sds.ifsd_map[nic_i],
1107 			    BUS_DMASYNC_PREWRITE);
1108 
1109 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1110 			nm_i = nm_next(nm_i, lim);
1111 			nic_i = nm_next(nic_i, lim);
1112 		}
1113 		kring->nr_hwcur = nm_i;
1114 
1115 		/* synchronize the NIC ring */
1116 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1117 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1118 
1119 		/* (re)start the tx unit up to slot nic_i (excluded) */
1120 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1121 	}
1122 
1123 	/*
1124 	 * Second part: reclaim buffers for completed transmissions.
1125 	 *
1126 	 * If there are unclaimed buffers, attempt to reclaim them.
1127 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1128 	 * trigger a per-tx-queue timer to try again later.
1129 	 */
1130 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1131 		if (iflib_tx_credits_update(ctx, txq)) {
1132 			/* some tx completed, increment avail */
1133 			nic_i = txq->ift_cidx_processed;
1134 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1135 		}
1136 	}
1137 
1138 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1139 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1140 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1141 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1142 			    iflib_netmap_timer, txq,
1143 			    txq->ift_netmap_timer.c_cpu, 0);
1144 		}
1145 
1146 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1147 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1148 
1149 	return (0);
1150 }
1151 
1152 /*
1153  * Reconcile kernel and user view of the receive ring.
1154  * Same as for the txsync, this routine must be efficient.
1155  * The caller guarantees a single invocations, but races against
1156  * the rest of the driver should be handled here.
1157  *
1158  * On call, kring->rhead is the first packet that userspace wants
1159  * to keep, and kring->rcur is the wakeup point.
1160  * The kernel has previously reported packets up to kring->rtail.
1161  *
1162  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1163  * of whether or not we received an interrupt.
1164  */
1165 static int
1166 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1167 {
1168 	struct netmap_adapter *na = kring->na;
1169 	struct netmap_ring *ring = kring->ring;
1170 	if_t ifp = na->ifp;
1171 	uint32_t nm_i;	/* index into the netmap ring */
1172 	uint32_t nic_i;	/* index into the NIC ring */
1173 	u_int n;
1174 	u_int const lim = kring->nkr_num_slots - 1;
1175 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1176 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1177 
1178 	if_ctx_t ctx = if_getsoftc(ifp);
1179 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1180 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1181 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1182 	iflib_fl_t fl = &rxq->ifr_fl[0];
1183 	struct if_rxd_info ri;
1184 	qidx_t *cidxp;
1185 
1186 	/*
1187 	 * netmap only uses free list 0, to avoid out of order consumption
1188 	 * of receive buffers
1189 	 */
1190 
1191 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1192 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1193 
1194 	/*
1195 	 * First part: import newly received packets.
1196 	 *
1197 	 * nm_i is the index of the next free slot in the netmap ring,
1198 	 * nic_i is the index of the next received packet in the NIC ring
1199 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1200 	 * differ in case if_init() has been called while
1201 	 * in netmap mode. For the receive ring we have
1202 	 *
1203 	 *	nic_i = fl->ifl_cidx;
1204 	 *	nm_i = kring->nr_hwtail (previous)
1205 	 * and
1206 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1207 	 *
1208 	 * fl->ifl_cidx is set to 0 on a ring reinit
1209 	 */
1210 	if (netmap_no_pendintr || force_update) {
1211 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1212 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1213 		int crclen = iflib_crcstrip ? 0 : 4;
1214 		int error, avail;
1215 
1216 		/*
1217 		 * For the free list consumer index, we use the same
1218 		 * logic as in iflib_rxeof().
1219 		 */
1220 		if (have_rxcq)
1221 			cidxp = &rxq->ifr_cq_cidx;
1222 		else
1223 			cidxp = &fl->ifl_cidx;
1224 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1225 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1226 
1227 		nic_i = fl->ifl_cidx;
1228 		nm_i = netmap_idx_n2k(kring, nic_i);
1229 		MPASS(nm_i == kring->nr_hwtail);
1230 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1231 			rxd_info_zero(&ri);
1232 			ri.iri_frags = rxq->ifr_frags;
1233 			ri.iri_qsidx = kring->ring_id;
1234 			ri.iri_ifp = ctx->ifc_ifp;
1235 			ri.iri_cidx = *cidxp;
1236 
1237 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1238 			for (i = 0; i < ri.iri_nfrags; i++) {
1239 				if (error) {
1240 					ring->slot[nm_i].len = 0;
1241 					ring->slot[nm_i].flags = 0;
1242 				} else {
1243 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1244 					if (i == (ri.iri_nfrags - 1)) {
1245 						ring->slot[nm_i].len -= crclen;
1246 						ring->slot[nm_i].flags = 0;
1247 
1248 						/* Update receive counters */
1249 						rx_bytes += ri.iri_len;
1250 						rx_pkts++;
1251 					} else
1252 						ring->slot[nm_i].flags = NS_MOREFRAG;
1253 				}
1254 
1255 				bus_dmamap_sync(fl->ifl_buf_tag,
1256 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1257 				nm_i = nm_next(nm_i, lim);
1258 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1259 			}
1260 
1261 			if (have_rxcq) {
1262 				*cidxp = ri.iri_cidx;
1263 				while (*cidxp >= scctx->isc_nrxd[0])
1264 					*cidxp -= scctx->isc_nrxd[0];
1265 			}
1266 
1267 		}
1268 		if (n) { /* update the state variables */
1269 			if (netmap_no_pendintr && !force_update) {
1270 				/* diagnostics */
1271 				iflib_rx_miss++;
1272 				iflib_rx_miss_bufs += n;
1273 			}
1274 			kring->nr_hwtail = nm_i;
1275 		}
1276 		kring->nr_kflags &= ~NKR_PENDINTR;
1277 	}
1278 	/*
1279 	 * Second part: skip past packets that userspace has released.
1280 	 * (kring->nr_hwcur to head excluded),
1281 	 * and make the buffers available for reception.
1282 	 * As usual nm_i is the index in the netmap ring,
1283 	 * nic_i is the index in the NIC ring, and
1284 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1285 	 */
1286 	netmap_fl_refill(rxq, kring, false);
1287 
1288 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1289 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1290 
1291 	return (0);
1292 }
1293 
1294 static void
1295 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1296 {
1297 	if_ctx_t ctx = if_getsoftc(na->ifp);
1298 
1299 	CTX_LOCK(ctx);
1300 	if (onoff) {
1301 		IFDI_INTR_ENABLE(ctx);
1302 	} else {
1303 		IFDI_INTR_DISABLE(ctx);
1304 	}
1305 	CTX_UNLOCK(ctx);
1306 }
1307 
1308 static int
1309 iflib_netmap_attach(if_ctx_t ctx)
1310 {
1311 	struct netmap_adapter na;
1312 
1313 	bzero(&na, sizeof(na));
1314 
1315 	na.ifp = ctx->ifc_ifp;
1316 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1317 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1318 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1319 
1320 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1321 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1322 	na.nm_txsync = iflib_netmap_txsync;
1323 	na.nm_rxsync = iflib_netmap_rxsync;
1324 	na.nm_register = iflib_netmap_register;
1325 	na.nm_intr = iflib_netmap_intr;
1326 	na.nm_config = iflib_netmap_config;
1327 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1328 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1329 	return (netmap_attach(&na));
1330 }
1331 
1332 static int
1333 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1334 {
1335 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1336 	struct netmap_slot *slot;
1337 
1338 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1339 	if (slot == NULL)
1340 		return (0);
1341 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1342 		/*
1343 		 * In netmap mode, set the map for the packet buffer.
1344 		 * NOTE: Some drivers (not this one) also need to set
1345 		 * the physical buffer address in the NIC ring.
1346 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1347 		 * netmap slot index, si
1348 		 */
1349 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1350 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1351 		    NMB(na, slot + si));
1352 	}
1353 	return (1);
1354 }
1355 
1356 static int
1357 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1358 {
1359 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1360 	struct netmap_kring *kring;
1361 	struct netmap_slot *slot;
1362 
1363 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1364 	if (slot == NULL)
1365 		return (0);
1366 	kring = na->rx_rings[rxq->ifr_id];
1367 	netmap_fl_refill(rxq, kring, true);
1368 	return (1);
1369 }
1370 
1371 static void
1372 iflib_netmap_timer(void *arg)
1373 {
1374 	iflib_txq_t txq = arg;
1375 	if_ctx_t ctx = txq->ift_ctx;
1376 
1377 	/*
1378 	 * Wake up the netmap application, to give it a chance to
1379 	 * call txsync and reclaim more completed TX buffers.
1380 	 */
1381 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1382 }
1383 
1384 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1385 
1386 #else
1387 #define iflib_netmap_txq_init(ctx, txq) (0)
1388 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1389 #define iflib_netmap_detach(ifp)
1390 #define netmap_enable_all_rings(ifp)
1391 #define netmap_disable_all_rings(ifp)
1392 
1393 #define iflib_netmap_attach(ctx) (0)
1394 #define netmap_rx_irq(ifp, qid, budget) (0)
1395 #endif
1396 
1397 #if defined(__i386__) || defined(__amd64__)
1398 static __inline void
1399 prefetch(void *x)
1400 {
1401 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1402 }
1403 
1404 static __inline void
1405 prefetch2cachelines(void *x)
1406 {
1407 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1408 #if (CACHE_LINE_SIZE < 128)
1409 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long)))));
1410 #endif
1411 }
1412 #else
1413 static __inline void
1414 prefetch(void *x)
1415 {
1416 }
1417 
1418 static __inline void
1419 prefetch2cachelines(void *x)
1420 {
1421 }
1422 #endif
1423 
1424 static void
1425 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1426 {
1427 	iflib_fl_t fl;
1428 
1429 	fl = &rxq->ifr_fl[flid];
1430 	iru->iru_paddrs = fl->ifl_bus_addrs;
1431 	iru->iru_idxs = fl->ifl_rxd_idxs;
1432 	iru->iru_qsidx = rxq->ifr_id;
1433 	iru->iru_buf_size = fl->ifl_buf_size;
1434 	iru->iru_flidx = fl->ifl_id;
1435 }
1436 
1437 static void
1438 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1439 {
1440 	if (err)
1441 		return;
1442 	*(bus_addr_t *) arg = segs[0].ds_addr;
1443 }
1444 
1445 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1446 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1447 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1448 
1449 int
1450 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1451 {
1452 	int err;
1453 	device_t dev = ctx->ifc_dev;
1454 	bus_addr_t lowaddr;
1455 
1456 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1457 
1458 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1459 				align, 0,		/* alignment, bounds */
1460 				lowaddr,		/* lowaddr */
1461 				BUS_SPACE_MAXADDR,	/* highaddr */
1462 				NULL, NULL,		/* filter, filterarg */
1463 				size,			/* maxsize */
1464 				1,			/* nsegments */
1465 				size,			/* maxsegsize */
1466 				BUS_DMA_ALLOCNOW,	/* flags */
1467 				NULL,			/* lockfunc */
1468 				NULL,			/* lockarg */
1469 				&dma->idi_tag);
1470 	if (err) {
1471 		device_printf(dev,
1472 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1473 		    __func__, err, size, align);
1474 		goto fail_0;
1475 	}
1476 
1477 	err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr,
1478 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1479 	if (err) {
1480 		device_printf(dev,
1481 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1482 		    __func__, (uintmax_t)size, err);
1483 		goto fail_1;
1484 	}
1485 
1486 	dma->idi_paddr = IF_BAD_DMA;
1487 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1488 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1489 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1490 		device_printf(dev,
1491 		    "%s: bus_dmamap_load failed: %d\n",
1492 		    __func__, err);
1493 		goto fail_2;
1494 	}
1495 
1496 	dma->idi_size = size;
1497 	return (0);
1498 
1499 fail_2:
1500 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1501 fail_1:
1502 	bus_dma_tag_destroy(dma->idi_tag);
1503 fail_0:
1504 	dma->idi_tag = NULL;
1505 
1506 	return (err);
1507 }
1508 
1509 int
1510 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1511 {
1512 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1513 
1514 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1515 
1516 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1517 }
1518 
1519 int
1520 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1521 {
1522 	int i, err;
1523 	iflib_dma_info_t *dmaiter;
1524 
1525 	dmaiter = dmalist;
1526 	for (i = 0; i < count; i++, dmaiter++) {
1527 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1528 			break;
1529 	}
1530 	if (err)
1531 		iflib_dma_free_multi(dmalist, i);
1532 	return (err);
1533 }
1534 
1535 void
1536 iflib_dma_free(iflib_dma_info_t dma)
1537 {
1538 	if (dma->idi_tag == NULL)
1539 		return;
1540 	if (dma->idi_paddr != IF_BAD_DMA) {
1541 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1542 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1543 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1544 		dma->idi_paddr = IF_BAD_DMA;
1545 	}
1546 	if (dma->idi_vaddr != NULL) {
1547 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1548 		dma->idi_vaddr = NULL;
1549 	}
1550 	bus_dma_tag_destroy(dma->idi_tag);
1551 	dma->idi_tag = NULL;
1552 }
1553 
1554 void
1555 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1556 {
1557 	int i;
1558 	iflib_dma_info_t *dmaiter = dmalist;
1559 
1560 	for (i = 0; i < count; i++, dmaiter++)
1561 		iflib_dma_free(*dmaiter);
1562 }
1563 
1564 static int
1565 iflib_fast_intr(void *arg)
1566 {
1567 	iflib_filter_info_t info = arg;
1568 	struct grouptask *gtask = info->ifi_task;
1569 	int result;
1570 
1571 	DBG_COUNTER_INC(fast_intrs);
1572 	if (info->ifi_filter != NULL) {
1573 		result = info->ifi_filter(info->ifi_filter_arg);
1574 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1575 			return (result);
1576 	}
1577 
1578 	GROUPTASK_ENQUEUE(gtask);
1579 	return (FILTER_HANDLED);
1580 }
1581 
1582 static int
1583 iflib_fast_intr_rxtx(void *arg)
1584 {
1585 	iflib_filter_info_t info = arg;
1586 	struct grouptask *gtask = info->ifi_task;
1587 	if_ctx_t ctx;
1588 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1589 	iflib_txq_t txq;
1590 	void *sc;
1591 	int i, cidx, result;
1592 	qidx_t txqid;
1593 	bool intr_enable, intr_legacy;
1594 
1595 	DBG_COUNTER_INC(fast_intrs);
1596 	if (info->ifi_filter != NULL) {
1597 		result = info->ifi_filter(info->ifi_filter_arg);
1598 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1599 			return (result);
1600 	}
1601 
1602 	ctx = rxq->ifr_ctx;
1603 	sc = ctx->ifc_softc;
1604 	intr_enable = false;
1605 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1606 	MPASS(rxq->ifr_ntxqirq);
1607 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1608 		txqid = rxq->ifr_txqid[i];
1609 		txq = &ctx->ifc_txqs[txqid];
1610 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1611 		    BUS_DMASYNC_POSTREAD);
1612 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1613 			if (intr_legacy)
1614 				intr_enable = true;
1615 			else
1616 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1617 			continue;
1618 		}
1619 		GROUPTASK_ENQUEUE(&txq->ift_task);
1620 	}
1621 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1622 		cidx = rxq->ifr_cq_cidx;
1623 	else
1624 		cidx = rxq->ifr_fl[0].ifl_cidx;
1625 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1626 		GROUPTASK_ENQUEUE(gtask);
1627 	else {
1628 		if (intr_legacy)
1629 			intr_enable = true;
1630 		else
1631 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1632 		DBG_COUNTER_INC(rx_intr_enables);
1633 	}
1634 	if (intr_enable)
1635 		IFDI_INTR_ENABLE(ctx);
1636 	return (FILTER_HANDLED);
1637 }
1638 
1639 static int
1640 iflib_fast_intr_ctx(void *arg)
1641 {
1642 	iflib_filter_info_t info = arg;
1643 	struct grouptask *gtask = info->ifi_task;
1644 	int result;
1645 
1646 	DBG_COUNTER_INC(fast_intrs);
1647 	if (info->ifi_filter != NULL) {
1648 		result = info->ifi_filter(info->ifi_filter_arg);
1649 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1650 			return (result);
1651 	}
1652 
1653 	if (gtask->gt_taskqueue != NULL)
1654 		GROUPTASK_ENQUEUE(gtask);
1655 	return (FILTER_HANDLED);
1656 }
1657 
1658 static int
1659 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1660 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1661 		 const char *name)
1662 {
1663 	struct resource *res;
1664 	void *tag = NULL;
1665 	device_t dev = ctx->ifc_dev;
1666 	int flags, i, rc;
1667 
1668 	flags = RF_ACTIVE;
1669 	if (ctx->ifc_flags & IFC_LEGACY)
1670 		flags |= RF_SHAREABLE;
1671 	MPASS(rid < 512);
1672 	i = rid;
1673 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1674 	if (res == NULL) {
1675 		device_printf(dev,
1676 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1677 		return (ENOMEM);
1678 	}
1679 	irq->ii_res = res;
1680 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1681 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1682 						filter, handler, arg, &tag);
1683 	if (rc != 0) {
1684 		device_printf(dev,
1685 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1686 					  rid, name ? name : "unknown", rc);
1687 		return (rc);
1688 	} else if (name)
1689 		bus_describe_intr(dev, res, tag, "%s", name);
1690 
1691 	irq->ii_tag = tag;
1692 	return (0);
1693 }
1694 
1695 /*********************************************************************
1696  *
1697  *  Allocate DMA resources for TX buffers as well as memory for the TX
1698  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1699  *  iflib_sw_tx_desc_array structure, storing all the information that
1700  *  is needed to transmit a packet on the wire.  This is called only
1701  *  once at attach, setup is done every reset.
1702  *
1703  **********************************************************************/
1704 static int
1705 iflib_txsd_alloc(iflib_txq_t txq)
1706 {
1707 	if_ctx_t ctx = txq->ift_ctx;
1708 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1709 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1710 	device_t dev = ctx->ifc_dev;
1711 	bus_size_t tsomaxsize;
1712 	bus_addr_t lowaddr;
1713 	int err, nsegments, ntsosegments;
1714 	bool tso;
1715 
1716 	nsegments = scctx->isc_tx_nsegments;
1717 	ntsosegments = scctx->isc_tx_tso_segments_max;
1718 	tsomaxsize = scctx->isc_tx_tso_size_max;
1719 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1720 		tsomaxsize += sizeof(struct ether_vlan_header);
1721 	MPASS(scctx->isc_ntxd[0] > 0);
1722 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1723 	MPASS(nsegments > 0);
1724 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1725 		MPASS(ntsosegments > 0);
1726 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1727 	}
1728 
1729 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1730 
1731 	/*
1732 	 * Set up DMA tags for TX buffers.
1733 	 */
1734 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1735 			       1, 0,			/* alignment, bounds */
1736 			       lowaddr,			/* lowaddr */
1737 			       BUS_SPACE_MAXADDR,	/* highaddr */
1738 			       NULL, NULL,		/* filter, filterarg */
1739 			       sctx->isc_tx_maxsize,		/* maxsize */
1740 			       nsegments,	/* nsegments */
1741 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1742 			       0,			/* flags */
1743 			       NULL,			/* lockfunc */
1744 			       NULL,			/* lockfuncarg */
1745 			       &txq->ift_buf_tag))) {
1746 		device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err);
1747 		device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1748 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1749 		goto fail;
1750 	}
1751 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1752 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1753 			       1, 0,			/* alignment, bounds */
1754 			       lowaddr,			/* lowaddr */
1755 			       BUS_SPACE_MAXADDR,	/* highaddr */
1756 			       NULL, NULL,		/* filter, filterarg */
1757 			       tsomaxsize,		/* maxsize */
1758 			       ntsosegments,	/* nsegments */
1759 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1760 			       0,			/* flags */
1761 			       NULL,			/* lockfunc */
1762 			       NULL,			/* lockfuncarg */
1763 			       &txq->ift_tso_buf_tag))) {
1764 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1765 		    err);
1766 		goto fail;
1767 	}
1768 
1769 	/* Allocate memory for the TX mbuf map. */
1770 	if (!(txq->ift_sds.ifsd_m =
1771 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1772 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1773 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1774 		err = ENOMEM;
1775 		goto fail;
1776 	}
1777 
1778 	/*
1779 	 * Create the DMA maps for TX buffers.
1780 	 */
1781 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1782 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1783 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1784 		device_printf(dev,
1785 		    "Unable to allocate TX buffer DMA map memory\n");
1786 		err = ENOMEM;
1787 		goto fail;
1788 	}
1789 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1790 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1791 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1792 		device_printf(dev,
1793 		    "Unable to allocate TSO TX buffer map memory\n");
1794 		err = ENOMEM;
1795 		goto fail;
1796 	}
1797 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1798 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1799 		    &txq->ift_sds.ifsd_map[i]);
1800 		if (err != 0) {
1801 			device_printf(dev, "Unable to create TX DMA map\n");
1802 			goto fail;
1803 		}
1804 		if (!tso)
1805 			continue;
1806 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1807 		    &txq->ift_sds.ifsd_tso_map[i]);
1808 		if (err != 0) {
1809 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1810 			goto fail;
1811 		}
1812 	}
1813 	return (0);
1814 fail:
1815 	/* We free all, it handles case where we are in the middle */
1816 	iflib_tx_structures_free(ctx);
1817 	return (err);
1818 }
1819 
1820 static void
1821 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1822 {
1823 	bus_dmamap_t map;
1824 
1825 	if (txq->ift_sds.ifsd_map != NULL) {
1826 		map = txq->ift_sds.ifsd_map[i];
1827 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1828 		bus_dmamap_unload(txq->ift_buf_tag, map);
1829 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1830 		txq->ift_sds.ifsd_map[i] = NULL;
1831 	}
1832 
1833 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1834 		map = txq->ift_sds.ifsd_tso_map[i];
1835 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1836 		    BUS_DMASYNC_POSTWRITE);
1837 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1838 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1839 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1840 	}
1841 }
1842 
1843 static void
1844 iflib_txq_destroy(iflib_txq_t txq)
1845 {
1846 	if_ctx_t ctx = txq->ift_ctx;
1847 
1848 	for (int i = 0; i < txq->ift_size; i++)
1849 		iflib_txsd_destroy(ctx, txq, i);
1850 
1851 	if (txq->ift_br != NULL) {
1852 		ifmp_ring_free(txq->ift_br);
1853 		txq->ift_br = NULL;
1854 	}
1855 
1856 	mtx_destroy(&txq->ift_mtx);
1857 
1858 	if (txq->ift_sds.ifsd_map != NULL) {
1859 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1860 		txq->ift_sds.ifsd_map = NULL;
1861 	}
1862 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1863 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1864 		txq->ift_sds.ifsd_tso_map = NULL;
1865 	}
1866 	if (txq->ift_sds.ifsd_m != NULL) {
1867 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1868 		txq->ift_sds.ifsd_m = NULL;
1869 	}
1870 	if (txq->ift_buf_tag != NULL) {
1871 		bus_dma_tag_destroy(txq->ift_buf_tag);
1872 		txq->ift_buf_tag = NULL;
1873 	}
1874 	if (txq->ift_tso_buf_tag != NULL) {
1875 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1876 		txq->ift_tso_buf_tag = NULL;
1877 	}
1878 	if (txq->ift_ifdi != NULL) {
1879 		free(txq->ift_ifdi, M_IFLIB);
1880 	}
1881 }
1882 
1883 static void
1884 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1885 {
1886 	struct mbuf **mp;
1887 
1888 	mp = &txq->ift_sds.ifsd_m[i];
1889 	if (*mp == NULL)
1890 		return;
1891 
1892 	if (txq->ift_sds.ifsd_map != NULL) {
1893 		bus_dmamap_sync(txq->ift_buf_tag,
1894 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1895 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1896 	}
1897 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1898 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1899 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1900 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1901 		    txq->ift_sds.ifsd_tso_map[i]);
1902 	}
1903 	m_freem(*mp);
1904 	DBG_COUNTER_INC(tx_frees);
1905 	*mp = NULL;
1906 }
1907 
1908 static int
1909 iflib_txq_setup(iflib_txq_t txq)
1910 {
1911 	if_ctx_t ctx = txq->ift_ctx;
1912 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1913 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1914 	iflib_dma_info_t di;
1915 	int i;
1916 
1917 	/* Set number of descriptors available */
1918 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1919 	/* XXX make configurable */
1920 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1921 
1922 	/* Reset indices */
1923 	txq->ift_cidx_processed = 0;
1924 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1925 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1926 
1927 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1928 		bzero((void *)di->idi_vaddr, di->idi_size);
1929 
1930 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1931 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1932 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1933 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1934 	return (0);
1935 }
1936 
1937 /*********************************************************************
1938  *
1939  *  Allocate DMA resources for RX buffers as well as memory for the RX
1940  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1941  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1942  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1943  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1944  *  packet, the maximum number of entries we'll need is equal to the
1945  *  number of hardware receive descriptors that we've allocated.
1946  *
1947  **********************************************************************/
1948 static int
1949 iflib_rxsd_alloc(iflib_rxq_t rxq)
1950 {
1951 	if_ctx_t ctx = rxq->ifr_ctx;
1952 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1953 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1954 	device_t dev = ctx->ifc_dev;
1955 	iflib_fl_t fl;
1956 	bus_addr_t lowaddr;
1957 	int err;
1958 
1959 	MPASS(scctx->isc_nrxd[0] > 0);
1960 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1961 
1962 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1963 
1964 	fl = rxq->ifr_fl;
1965 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1966 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1967 		/* Set up DMA tag for RX buffers. */
1968 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1969 					 1, 0,			/* alignment, bounds */
1970 					 lowaddr,		/* lowaddr */
1971 					 BUS_SPACE_MAXADDR,	/* highaddr */
1972 					 NULL, NULL,		/* filter, filterarg */
1973 					 sctx->isc_rx_maxsize,	/* maxsize */
1974 					 sctx->isc_rx_nsegments,	/* nsegments */
1975 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1976 					 0,			/* flags */
1977 					 NULL,			/* lockfunc */
1978 					 NULL,			/* lockarg */
1979 					 &fl->ifl_buf_tag);
1980 		if (err) {
1981 			device_printf(dev,
1982 			    "Unable to allocate RX DMA tag: %d\n", err);
1983 			goto fail;
1984 		}
1985 
1986 		/* Allocate memory for the RX mbuf map. */
1987 		if (!(fl->ifl_sds.ifsd_m =
1988 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1989 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1990 			device_printf(dev,
1991 			    "Unable to allocate RX mbuf map memory\n");
1992 			err = ENOMEM;
1993 			goto fail;
1994 		}
1995 
1996 		/* Allocate memory for the direct RX cluster pointer map. */
1997 		if (!(fl->ifl_sds.ifsd_cl =
1998 		      (caddr_t *) malloc(sizeof(caddr_t) *
1999 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2000 			device_printf(dev,
2001 			    "Unable to allocate RX cluster map memory\n");
2002 			err = ENOMEM;
2003 			goto fail;
2004 		}
2005 
2006 		/* Allocate memory for the RX cluster bus address map. */
2007 		if (!(fl->ifl_sds.ifsd_ba =
2008 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2009 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2010 			device_printf(dev,
2011 			    "Unable to allocate RX bus address map memory\n");
2012 			err = ENOMEM;
2013 			goto fail;
2014 		}
2015 
2016 		/*
2017 		 * Create the DMA maps for RX buffers.
2018 		 */
2019 		if (!(fl->ifl_sds.ifsd_map =
2020 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2021 			device_printf(dev,
2022 			    "Unable to allocate RX buffer DMA map memory\n");
2023 			err = ENOMEM;
2024 			goto fail;
2025 		}
2026 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2027 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2028 			    &fl->ifl_sds.ifsd_map[i]);
2029 			if (err != 0) {
2030 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2031 				goto fail;
2032 			}
2033 		}
2034 	}
2035 	return (0);
2036 
2037 fail:
2038 	iflib_rx_structures_free(ctx);
2039 	return (err);
2040 }
2041 
2042 /*
2043  * Internal service routines
2044  */
2045 
2046 struct rxq_refill_cb_arg {
2047 	int               error;
2048 	bus_dma_segment_t seg;
2049 	int               nseg;
2050 };
2051 
2052 static void
2053 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2054 {
2055 	struct rxq_refill_cb_arg *cb_arg = arg;
2056 
2057 	cb_arg->error = error;
2058 	cb_arg->seg = segs[0];
2059 	cb_arg->nseg = nseg;
2060 }
2061 
2062 /**
2063  * iflib_fl_refill - refill an rxq free-buffer list
2064  * @ctx: the iflib context
2065  * @fl: the free list to refill
2066  * @count: the number of new buffers to allocate
2067  *
2068  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2069  * The caller must assure that @count does not exceed the queue's capacity
2070  * minus one (since we always leave a descriptor unavailable).
2071  */
2072 static uint8_t
2073 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2074 {
2075 	struct if_rxd_update iru;
2076 	struct rxq_refill_cb_arg cb_arg;
2077 	struct mbuf *m;
2078 	caddr_t cl, *sd_cl;
2079 	struct mbuf **sd_m;
2080 	bus_dmamap_t *sd_map;
2081 	bus_addr_t bus_addr, *sd_ba;
2082 	int err, frag_idx, i, idx, n, pidx;
2083 	qidx_t credits;
2084 
2085 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2086 
2087 	sd_m = fl->ifl_sds.ifsd_m;
2088 	sd_map = fl->ifl_sds.ifsd_map;
2089 	sd_cl = fl->ifl_sds.ifsd_cl;
2090 	sd_ba = fl->ifl_sds.ifsd_ba;
2091 	pidx = fl->ifl_pidx;
2092 	idx = pidx;
2093 	frag_idx = fl->ifl_fragidx;
2094 	credits = fl->ifl_credits;
2095 
2096 	i = 0;
2097 	n = count;
2098 	MPASS(n > 0);
2099 	MPASS(credits + n <= fl->ifl_size);
2100 
2101 	if (pidx < fl->ifl_cidx)
2102 		MPASS(pidx + n <= fl->ifl_cidx);
2103 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2104 		MPASS(fl->ifl_gen == 0);
2105 	if (pidx > fl->ifl_cidx)
2106 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2107 
2108 	DBG_COUNTER_INC(fl_refills);
2109 	if (n > 8)
2110 		DBG_COUNTER_INC(fl_refills_large);
2111 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2112 	while (n-- > 0) {
2113 		/*
2114 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2115 		 * initialized after rx.
2116 		 *
2117 		 * If the cluster is still set then we know a minimum sized
2118 		 * packet was received
2119 		 */
2120 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2121 		    &frag_idx);
2122 		if (frag_idx < 0)
2123 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2124 		MPASS(frag_idx >= 0);
2125 		if ((cl = sd_cl[frag_idx]) == NULL) {
2126 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2127 			if (__predict_false(cl == NULL))
2128 				break;
2129 
2130 			cb_arg.error = 0;
2131 			MPASS(sd_map != NULL);
2132 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2133 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2134 			    BUS_DMA_NOWAIT);
2135 			if (__predict_false(err != 0 || cb_arg.error)) {
2136 				uma_zfree(fl->ifl_zone, cl);
2137 				break;
2138 			}
2139 
2140 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2141 			sd_cl[frag_idx] = cl;
2142 #if MEMORY_LOGGING
2143 			fl->ifl_cl_enqueued++;
2144 #endif
2145 		} else {
2146 			bus_addr = sd_ba[frag_idx];
2147 		}
2148 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2149 		    BUS_DMASYNC_PREREAD);
2150 
2151 		if (sd_m[frag_idx] == NULL) {
2152 			m = m_gethdr_raw(M_NOWAIT, 0);
2153 			if (__predict_false(m == NULL))
2154 				break;
2155 			sd_m[frag_idx] = m;
2156 		}
2157 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2158 #if MEMORY_LOGGING
2159 		fl->ifl_m_enqueued++;
2160 #endif
2161 
2162 		DBG_COUNTER_INC(rx_allocs);
2163 		fl->ifl_rxd_idxs[i] = frag_idx;
2164 		fl->ifl_bus_addrs[i] = bus_addr;
2165 		credits++;
2166 		i++;
2167 		MPASS(credits <= fl->ifl_size);
2168 		if (++idx == fl->ifl_size) {
2169 #ifdef INVARIANTS
2170 			fl->ifl_gen = 1;
2171 #endif
2172 			idx = 0;
2173 		}
2174 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2175 			iru.iru_pidx = pidx;
2176 			iru.iru_count = i;
2177 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2178 			fl->ifl_pidx = idx;
2179 			fl->ifl_credits = credits;
2180 			pidx = idx;
2181 			i = 0;
2182 		}
2183 	}
2184 
2185 	if (n < count - 1) {
2186 		if (i != 0) {
2187 			iru.iru_pidx = pidx;
2188 			iru.iru_count = i;
2189 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2190 			fl->ifl_pidx = idx;
2191 			fl->ifl_credits = credits;
2192 		}
2193 		DBG_COUNTER_INC(rxd_flush);
2194 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2195 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2196 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2197 		    fl->ifl_id, fl->ifl_pidx);
2198 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2199 			fl->ifl_fragidx = frag_idx + 1;
2200 			if (fl->ifl_fragidx == fl->ifl_size)
2201 				fl->ifl_fragidx = 0;
2202 		} else {
2203 			fl->ifl_fragidx = frag_idx;
2204 		}
2205 	}
2206 
2207 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2208 }
2209 
2210 static inline uint8_t
2211 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2212 {
2213 	/*
2214 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2215 	 * This is important as it confuses most NICs. For instance,
2216 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2217 	 * RDH points to the next receive descriptor to be used by the NIC,
2218 	 * and RDT for the next receive descriptor to be published by the
2219 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2220 	 * The condition RDH == RDT means no descriptors are available to
2221 	 * the NIC, and thus it would be ambiguous if it also meant that
2222 	 * all the descriptors are available to the NIC.
2223 	 */
2224 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2225 #ifdef INVARIANTS
2226 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2227 #endif
2228 
2229 	MPASS(fl->ifl_credits <= fl->ifl_size);
2230 	MPASS(reclaimable == delta);
2231 
2232 	if (reclaimable > 0)
2233 		return (iflib_fl_refill(ctx, fl, reclaimable));
2234 	return (0);
2235 }
2236 
2237 uint8_t
2238 iflib_in_detach(if_ctx_t ctx)
2239 {
2240 	bool in_detach;
2241 
2242 	STATE_LOCK(ctx);
2243 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2244 	STATE_UNLOCK(ctx);
2245 	return (in_detach);
2246 }
2247 
2248 static void
2249 iflib_fl_bufs_free(iflib_fl_t fl)
2250 {
2251 	iflib_dma_info_t idi = fl->ifl_ifdi;
2252 	bus_dmamap_t sd_map;
2253 	uint32_t i;
2254 
2255 	for (i = 0; i < fl->ifl_size; i++) {
2256 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2257 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2258 
2259 		if (*sd_cl != NULL) {
2260 			sd_map = fl->ifl_sds.ifsd_map[i];
2261 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2262 			    BUS_DMASYNC_POSTREAD);
2263 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2264 			uma_zfree(fl->ifl_zone, *sd_cl);
2265 			*sd_cl = NULL;
2266 			if (*sd_m != NULL) {
2267 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2268 				m_free_raw(*sd_m);
2269 				*sd_m = NULL;
2270 			}
2271 		} else {
2272 			MPASS(*sd_m == NULL);
2273 		}
2274 #if MEMORY_LOGGING
2275 		fl->ifl_m_dequeued++;
2276 		fl->ifl_cl_dequeued++;
2277 #endif
2278 	}
2279 #ifdef INVARIANTS
2280 	for (i = 0; i < fl->ifl_size; i++) {
2281 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2282 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2283 	}
2284 #endif
2285 	/*
2286 	 * Reset free list values
2287 	 */
2288 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2289 	bzero(idi->idi_vaddr, idi->idi_size);
2290 }
2291 
2292 /*********************************************************************
2293  *
2294  *  Initialize a free list and its buffers.
2295  *
2296  **********************************************************************/
2297 static int
2298 iflib_fl_setup(iflib_fl_t fl)
2299 {
2300 	iflib_rxq_t rxq = fl->ifl_rxq;
2301 	if_ctx_t ctx = rxq->ifr_ctx;
2302 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2303 	int qidx;
2304 
2305 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2306 	/*
2307 	 * Free current RX buffer structs and their mbufs
2308 	 */
2309 	iflib_fl_bufs_free(fl);
2310 	/* Now replenish the mbufs */
2311 	MPASS(fl->ifl_credits == 0);
2312 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2313 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2314 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2315 	else
2316 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2317 	/*
2318 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2319 	 * to the selected mbuf size.
2320 	 */
2321 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2322 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2323 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2324 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2325 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2326 
2327 	/*
2328 	 * Avoid pre-allocating zillions of clusters to an idle card
2329 	 * potentially speeding up attach. In any case make sure
2330 	 * to leave a descriptor unavailable. See the comment in
2331 	 * iflib_fl_refill_all().
2332 	 */
2333 	MPASS(fl->ifl_size > 0);
2334 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2335 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2336 		return (ENOBUFS);
2337 	/*
2338 	 * handle failure
2339 	 */
2340 	MPASS(rxq != NULL);
2341 	MPASS(fl->ifl_ifdi != NULL);
2342 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2343 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2344 	return (0);
2345 }
2346 
2347 /*********************************************************************
2348  *
2349  *  Free receive ring data structures
2350  *
2351  **********************************************************************/
2352 static void
2353 iflib_rx_sds_free(iflib_rxq_t rxq)
2354 {
2355 	iflib_fl_t fl;
2356 	int i, j;
2357 
2358 	if (rxq->ifr_fl != NULL) {
2359 		for (i = 0; i < rxq->ifr_nfl; i++) {
2360 			fl = &rxq->ifr_fl[i];
2361 			if (fl->ifl_buf_tag != NULL) {
2362 				if (fl->ifl_sds.ifsd_map != NULL) {
2363 					for (j = 0; j < fl->ifl_size; j++) {
2364 						bus_dmamap_sync(
2365 						    fl->ifl_buf_tag,
2366 						    fl->ifl_sds.ifsd_map[j],
2367 						    BUS_DMASYNC_POSTREAD);
2368 						bus_dmamap_unload(
2369 						    fl->ifl_buf_tag,
2370 						    fl->ifl_sds.ifsd_map[j]);
2371 						bus_dmamap_destroy(
2372 						    fl->ifl_buf_tag,
2373 						    fl->ifl_sds.ifsd_map[j]);
2374 					}
2375 				}
2376 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2377 				fl->ifl_buf_tag = NULL;
2378 			}
2379 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2380 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2381 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2382 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2383 			free(fl->ifl_rx_bitmap, M_IFLIB);
2384 			fl->ifl_sds.ifsd_m = NULL;
2385 			fl->ifl_sds.ifsd_cl = NULL;
2386 			fl->ifl_sds.ifsd_ba = NULL;
2387 			fl->ifl_sds.ifsd_map = NULL;
2388 			fl->ifl_rx_bitmap = NULL;
2389 		}
2390 		free(rxq->ifr_fl, M_IFLIB);
2391 		rxq->ifr_fl = NULL;
2392 		free(rxq->ifr_ifdi, M_IFLIB);
2393 		rxq->ifr_ifdi = NULL;
2394 		rxq->ifr_cq_cidx = 0;
2395 	}
2396 }
2397 
2398 /*
2399  * Timer routine
2400  */
2401 static void
2402 iflib_timer(void *arg)
2403 {
2404 	iflib_txq_t txq = arg;
2405 	if_ctx_t ctx = txq->ift_ctx;
2406 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2407 	uint64_t this_tick = ticks;
2408 
2409 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2410 		return;
2411 
2412 	/*
2413 	** Check on the state of the TX queue(s), this
2414 	** can be done without the lock because its RO
2415 	** and the HUNG state will be static if set.
2416 	*/
2417 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2418 		txq->ift_last_timer_tick = this_tick;
2419 		IFDI_TIMER(ctx, txq->ift_id);
2420 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2421 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2422 		     (sctx->isc_pause_frames == 0)))
2423 			goto hung;
2424 
2425 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2426 		    ifmp_ring_is_stalled(txq->ift_br)) {
2427 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2428 			    ("queue can't be marked as hung if interface is down"));
2429 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2430 		}
2431 		txq->ift_cleaned_prev = txq->ift_cleaned;
2432 	}
2433 	/* handle any laggards */
2434 	if (txq->ift_db_pending)
2435 		GROUPTASK_ENQUEUE(&txq->ift_task);
2436 
2437 	sctx->isc_pause_frames = 0;
2438 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2439 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2440 		    txq, txq->ift_timer.c_cpu);
2441 	return;
2442 
2443  hung:
2444 	device_printf(ctx->ifc_dev,
2445 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2446 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2447 	STATE_LOCK(ctx);
2448 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2449 	ctx->ifc_flags |= (IFC_DO_WATCHDOG | IFC_DO_RESET);
2450 	iflib_admin_intr_deferred(ctx);
2451 	STATE_UNLOCK(ctx);
2452 }
2453 
2454 static uint16_t
2455 iflib_get_mbuf_size_for(unsigned int size)
2456 {
2457 
2458 	if (size <= MCLBYTES)
2459 		return (MCLBYTES);
2460 	else
2461 		return (MJUMPAGESIZE);
2462 }
2463 
2464 static void
2465 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2466 {
2467 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2468 
2469 	/*
2470 	 * XXX don't set the max_frame_size to larger
2471 	 * than the hardware can handle
2472 	 */
2473 	ctx->ifc_rx_mbuf_sz =
2474 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2475 }
2476 
2477 uint32_t
2478 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2479 {
2480 
2481 	return (ctx->ifc_rx_mbuf_sz);
2482 }
2483 
2484 static void
2485 iflib_init_locked(if_ctx_t ctx)
2486 {
2487 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2488 	if_t ifp = ctx->ifc_ifp;
2489 	iflib_fl_t fl;
2490 	iflib_txq_t txq;
2491 	iflib_rxq_t rxq;
2492 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2493 
2494 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2495 	IFDI_INTR_DISABLE(ctx);
2496 
2497 	/*
2498 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2499 	 * called without first calling iflib_stop().
2500 	 */
2501 	netmap_disable_all_rings(ifp);
2502 
2503 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2504 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2505 	/* Set hardware offload abilities */
2506 	if_clearhwassist(ifp);
2507 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2508 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2509 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2510 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2511 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2512 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2513 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2514 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2515 
2516 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2517 		CALLOUT_LOCK(txq);
2518 		callout_stop(&txq->ift_timer);
2519 #ifdef DEV_NETMAP
2520 		callout_stop(&txq->ift_netmap_timer);
2521 #endif /* DEV_NETMAP */
2522 		CALLOUT_UNLOCK(txq);
2523 		(void)iflib_netmap_txq_init(ctx, txq);
2524 	}
2525 
2526 	/*
2527 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2528 	 * that drivers can use the value when setting up the hardware receive
2529 	 * buffers.
2530 	 */
2531 	iflib_calc_rx_mbuf_sz(ctx);
2532 
2533 #ifdef INVARIANTS
2534 	i = if_getdrvflags(ifp);
2535 #endif
2536 	IFDI_INIT(ctx);
2537 	MPASS(if_getdrvflags(ifp) == i);
2538 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2539 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2540 			/* This rxq is in netmap mode. Skip normal init. */
2541 			continue;
2542 		}
2543 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2544 			if (iflib_fl_setup(fl)) {
2545 				device_printf(ctx->ifc_dev,
2546 				    "setting up free list %d failed - "
2547 				    "check cluster settings\n", j);
2548 				goto done;
2549 			}
2550 		}
2551 	}
2552 done:
2553 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2554 	IFDI_INTR_ENABLE(ctx);
2555 	txq = ctx->ifc_txqs;
2556 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2557 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2558 			txq->ift_timer.c_cpu);
2559 
2560         /* Re-enable txsync/rxsync. */
2561 	netmap_enable_all_rings(ifp);
2562 }
2563 
2564 static int
2565 iflib_media_change(if_t ifp)
2566 {
2567 	if_ctx_t ctx = if_getsoftc(ifp);
2568 	int err;
2569 
2570 	CTX_LOCK(ctx);
2571 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2572 		iflib_if_init_locked(ctx);
2573 	CTX_UNLOCK(ctx);
2574 	return (err);
2575 }
2576 
2577 static void
2578 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2579 {
2580 	if_ctx_t ctx = if_getsoftc(ifp);
2581 
2582 	CTX_LOCK(ctx);
2583 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2584 	IFDI_MEDIA_STATUS(ctx, ifmr);
2585 	CTX_UNLOCK(ctx);
2586 }
2587 
2588 static void
2589 iflib_stop(if_ctx_t ctx)
2590 {
2591 	iflib_txq_t txq = ctx->ifc_txqs;
2592 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2593 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2594 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2595 	iflib_dma_info_t di;
2596 	iflib_fl_t fl;
2597 	int i, j;
2598 
2599 	/* Tell the stack that the interface is no longer active */
2600 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2601 
2602 	IFDI_INTR_DISABLE(ctx);
2603 	DELAY(1000);
2604 	IFDI_STOP(ctx);
2605 	DELAY(1000);
2606 
2607 	/*
2608 	 * Stop any pending txsync/rxsync and prevent new ones
2609 	 * form starting. Processes blocked in poll() will get
2610 	 * POLLERR.
2611 	 */
2612 	netmap_disable_all_rings(ctx->ifc_ifp);
2613 
2614 	iflib_debug_reset();
2615 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2616 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2617 		/* make sure all transmitters have completed before proceeding XXX */
2618 
2619 		CALLOUT_LOCK(txq);
2620 		callout_stop(&txq->ift_timer);
2621 #ifdef DEV_NETMAP
2622 		callout_stop(&txq->ift_netmap_timer);
2623 #endif /* DEV_NETMAP */
2624 		CALLOUT_UNLOCK(txq);
2625 
2626 		/* clean any enqueued buffers */
2627 		iflib_ifmp_purge(txq);
2628 		/* Free any existing tx buffers. */
2629 		for (j = 0; j < txq->ift_size; j++) {
2630 			iflib_txsd_free(ctx, txq, j);
2631 		}
2632 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2633 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2634 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2635 			txq->ift_cidx = txq->ift_pidx;
2636 		else
2637 			txq->ift_cidx = txq->ift_pidx = 0;
2638 
2639 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2640 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2641 		txq->ift_pullups = 0;
2642 		ifmp_ring_reset_stats(txq->ift_br);
2643 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2644 			bzero((void *)di->idi_vaddr, di->idi_size);
2645 	}
2646 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2647 		if (rxq->ifr_task.gt_taskqueue != NULL)
2648 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2649 				 &rxq->ifr_task.gt_task);
2650 
2651 		rxq->ifr_cq_cidx = 0;
2652 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2653 			bzero((void *)di->idi_vaddr, di->idi_size);
2654 		/* also resets the free lists pidx/cidx */
2655 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2656 			iflib_fl_bufs_free(fl);
2657 	}
2658 }
2659 
2660 static inline caddr_t
2661 calc_next_rxd(iflib_fl_t fl, int cidx)
2662 {
2663 	qidx_t size;
2664 	int nrxd;
2665 	caddr_t start, end, cur, next;
2666 
2667 	nrxd = fl->ifl_size;
2668 	size = fl->ifl_rxd_size;
2669 	start = fl->ifl_ifdi->idi_vaddr;
2670 
2671 	if (__predict_false(size == 0))
2672 		return (start);
2673 	cur = start + size * cidx;
2674 	end = start + size * nrxd;
2675 	next = CACHE_PTR_NEXT(cur);
2676 	return (next < end ? next : start);
2677 }
2678 
2679 static inline void
2680 prefetch_pkts(iflib_fl_t fl, int cidx)
2681 {
2682 	int nextptr;
2683 	int nrxd = fl->ifl_size;
2684 	caddr_t next_rxd;
2685 
2686 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1);
2687 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2688 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2689 	next_rxd = calc_next_rxd(fl, cidx);
2690 	prefetch(next_rxd);
2691 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]);
2692 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]);
2693 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]);
2694 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]);
2695 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]);
2696 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]);
2697 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]);
2698 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]);
2699 }
2700 
2701 static struct mbuf *
2702 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2703     int *pf_rv, if_rxd_info_t ri)
2704 {
2705 	bus_dmamap_t map;
2706 	iflib_fl_t fl;
2707 	caddr_t payload;
2708 	struct mbuf *m;
2709 	int flid, cidx, len, next;
2710 
2711 	map = NULL;
2712 	flid = irf->irf_flid;
2713 	cidx = irf->irf_idx;
2714 	fl = &rxq->ifr_fl[flid];
2715 	sd->ifsd_fl = fl;
2716 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2717 	fl->ifl_credits--;
2718 #if MEMORY_LOGGING
2719 	fl->ifl_m_dequeued++;
2720 #endif
2721 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2722 		prefetch_pkts(fl, cidx);
2723 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1);
2724 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2725 	map = fl->ifl_sds.ifsd_map[cidx];
2726 
2727 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2728 
2729 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2730 	    irf->irf_len != 0) {
2731 		payload  = *sd->ifsd_cl;
2732 		payload +=  ri->iri_pad;
2733 		len = ri->iri_len - ri->iri_pad;
2734 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2735 		switch (*pf_rv) {
2736 		case PFIL_DROPPED:
2737 		case PFIL_CONSUMED:
2738 			/*
2739 			 * The filter ate it.  Everything is recycled.
2740 			 */
2741 			m = NULL;
2742 			unload = 0;
2743 			break;
2744 		case PFIL_REALLOCED:
2745 			/*
2746 			 * The filter copied it.  Everything is recycled.
2747 			 * 'm' points at new mbuf.
2748 			 */
2749 			unload = 0;
2750 			break;
2751 		case PFIL_PASS:
2752 			/*
2753 			 * Filter said it was OK, so receive like
2754 			 * normal
2755 			 */
2756 			m = fl->ifl_sds.ifsd_m[cidx];
2757 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2758 			break;
2759 		default:
2760 			MPASS(0);
2761 		}
2762 	} else {
2763 		m = fl->ifl_sds.ifsd_m[cidx];
2764 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2765 		if (pf_rv != NULL)
2766 			*pf_rv = PFIL_PASS;
2767 	}
2768 
2769 	if (unload && irf->irf_len != 0)
2770 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2771 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1);
2772 	if (__predict_false(fl->ifl_cidx == 0))
2773 		fl->ifl_gen = 0;
2774 	bit_clear(fl->ifl_rx_bitmap, cidx);
2775 	return (m);
2776 }
2777 
2778 static struct mbuf *
2779 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2780 {
2781 	struct mbuf *m, *mh, *mt;
2782 	caddr_t cl;
2783 	int  *pf_rv_ptr, flags, i, padlen;
2784 	bool consumed;
2785 
2786 	i = 0;
2787 	mh = NULL;
2788 	consumed = false;
2789 	*pf_rv = PFIL_PASS;
2790 	pf_rv_ptr = pf_rv;
2791 	do {
2792 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2793 		    pf_rv_ptr, ri);
2794 
2795 		MPASS(*sd->ifsd_cl != NULL);
2796 
2797 		/*
2798 		 * Exclude zero-length frags & frags from
2799 		 * packets the filter has consumed or dropped
2800 		 */
2801 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2802 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2803 			if (mh == NULL) {
2804 				/* everything saved here */
2805 				consumed = true;
2806 				pf_rv_ptr = NULL;
2807 				continue;
2808 			}
2809 			/* XXX we can save the cluster here, but not the mbuf */
2810 			m_init(m, M_NOWAIT, MT_DATA, 0);
2811 			m_free(m);
2812 			continue;
2813 		}
2814 		if (mh == NULL) {
2815 			flags = M_PKTHDR | M_EXT;
2816 			mh = mt = m;
2817 			padlen = ri->iri_pad;
2818 		} else {
2819 			flags = M_EXT;
2820 			mt->m_next = m;
2821 			mt = m;
2822 			/* assuming padding is only on the first fragment */
2823 			padlen = 0;
2824 		}
2825 		cl = *sd->ifsd_cl;
2826 		*sd->ifsd_cl = NULL;
2827 
2828 		/* Can these two be made one ? */
2829 		m_init(m, M_NOWAIT, MT_DATA, flags);
2830 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2831 		/*
2832 		 * These must follow m_init and m_cljset
2833 		 */
2834 		m->m_data += padlen;
2835 		ri->iri_len -= padlen;
2836 		m->m_len = ri->iri_frags[i].irf_len;
2837 	} while (++i < ri->iri_nfrags);
2838 
2839 	return (mh);
2840 }
2841 
2842 /*
2843  * Process one software descriptor
2844  */
2845 static struct mbuf *
2846 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2847 {
2848 	struct if_rxsd sd;
2849 	struct mbuf *m;
2850 	int pf_rv;
2851 
2852 	/* should I merge this back in now that the two paths are basically duplicated? */
2853 	if (ri->iri_nfrags == 1 &&
2854 	    ri->iri_frags[0].irf_len != 0 &&
2855 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2856 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2857 		    &pf_rv, ri);
2858 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2859 			return (m);
2860 		if (pf_rv == PFIL_PASS) {
2861 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2862 #ifndef __NO_STRICT_ALIGNMENT
2863 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2864 				m->m_data += 2;
2865 #endif
2866 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2867 			m->m_len = ri->iri_frags[0].irf_len;
2868 			m->m_data += ri->iri_pad;
2869 			ri->iri_len -= ri->iri_pad;
2870 		}
2871 	} else {
2872 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2873 		if (m == NULL)
2874 			return (NULL);
2875 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2876 			return (m);
2877 	}
2878 	m->m_pkthdr.len = ri->iri_len;
2879 	m->m_pkthdr.rcvif = ri->iri_ifp;
2880 	m->m_flags |= ri->iri_flags;
2881 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2882 	m->m_pkthdr.flowid = ri->iri_flowid;
2883 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2884 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2885 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2886 	return (m);
2887 }
2888 
2889 #if defined(INET6) || defined(INET)
2890 static void
2891 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2892 {
2893 	CURVNET_SET(if_getvnet(lc->ifp));
2894 #if defined(INET6)
2895 	*v6 = V_ip6_forwarding;
2896 #endif
2897 #if defined(INET)
2898 	*v4 = V_ipforwarding;
2899 #endif
2900 	CURVNET_RESTORE();
2901 }
2902 
2903 /*
2904  * Returns true if it's possible this packet could be LROed.
2905  * if it returns false, it is guaranteed that tcp_lro_rx()
2906  * would not return zero.
2907  */
2908 static bool
2909 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2910 {
2911 	struct ether_header *eh;
2912 
2913 	eh = mtod(m, struct ether_header *);
2914 	switch (eh->ether_type) {
2915 #if defined(INET6)
2916 	case htons(ETHERTYPE_IPV6):
2917 		return (!v6_forwarding);
2918 #endif
2919 #if defined(INET)
2920 	case htons(ETHERTYPE_IP):
2921 		return (!v4_forwarding);
2922 #endif
2923 	}
2924 
2925 	return (false);
2926 }
2927 #else
2928 static void
2929 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2930 {
2931 }
2932 #endif
2933 
2934 static void
2935 _task_fn_rx_watchdog(void *context)
2936 {
2937 	iflib_rxq_t rxq = context;
2938 
2939 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2940 }
2941 
2942 static uint8_t
2943 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2944 {
2945 	if_t ifp;
2946 	if_ctx_t ctx = rxq->ifr_ctx;
2947 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2948 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2949 	int avail, i;
2950 	qidx_t *cidxp;
2951 	struct if_rxd_info ri;
2952 	int err, budget_left, rx_bytes, rx_pkts;
2953 	iflib_fl_t fl;
2954 	int lro_enabled;
2955 	bool v4_forwarding, v6_forwarding, lro_possible;
2956 	uint8_t retval = 0;
2957 
2958 	/*
2959 	 * XXX early demux data packets so that if_input processing only handles
2960 	 * acks in interrupt context
2961 	 */
2962 	struct mbuf *m, *mh, *mt, *mf;
2963 
2964 	NET_EPOCH_ASSERT();
2965 
2966 	lro_possible = v4_forwarding = v6_forwarding = false;
2967 	ifp = ctx->ifc_ifp;
2968 	mh = mt = NULL;
2969 	MPASS(budget > 0);
2970 	rx_pkts	= rx_bytes = 0;
2971 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2972 		cidxp = &rxq->ifr_cq_cidx;
2973 	else
2974 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2975 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2976 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2977 			retval |= iflib_fl_refill_all(ctx, fl);
2978 		DBG_COUNTER_INC(rx_unavail);
2979 		return (retval);
2980 	}
2981 
2982 	/* pfil needs the vnet to be set */
2983 	CURVNET_SET_QUIET(if_getvnet(ifp));
2984 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2985 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2986 			DBG_COUNTER_INC(rx_ctx_inactive);
2987 			break;
2988 		}
2989 		/*
2990 		 * Reset client set fields to their default values
2991 		 */
2992 		rxd_info_zero(&ri);
2993 		ri.iri_qsidx = rxq->ifr_id;
2994 		ri.iri_cidx = *cidxp;
2995 		ri.iri_ifp = ifp;
2996 		ri.iri_frags = rxq->ifr_frags;
2997 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2998 
2999 		if (err)
3000 			goto err;
3001 		rx_pkts += 1;
3002 		rx_bytes += ri.iri_len;
3003 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3004 			*cidxp = ri.iri_cidx;
3005 			/* Update our consumer index */
3006 			/* XXX NB: shurd - check if this is still safe */
3007 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3008 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3009 			/* was this only a completion queue message? */
3010 			if (__predict_false(ri.iri_nfrags == 0))
3011 				continue;
3012 		}
3013 		MPASS(ri.iri_nfrags != 0);
3014 		MPASS(ri.iri_len != 0);
3015 
3016 		/* will advance the cidx on the corresponding free lists */
3017 		m = iflib_rxd_pkt_get(rxq, &ri);
3018 		avail--;
3019 		budget_left--;
3020 		if (avail == 0 && budget_left)
3021 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3022 
3023 		if (__predict_false(m == NULL))
3024 			continue;
3025 
3026 		/* imm_pkt: -- cxgb */
3027 		if (mh == NULL)
3028 			mh = mt = m;
3029 		else {
3030 			mt->m_nextpkt = m;
3031 			mt = m;
3032 		}
3033 	}
3034 	CURVNET_RESTORE();
3035 	/* make sure that we can refill faster than drain */
3036 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3037 		retval |= iflib_fl_refill_all(ctx, fl);
3038 
3039 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3040 	if (lro_enabled)
3041 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3042 	mt = mf = NULL;
3043 	while (mh != NULL) {
3044 		m = mh;
3045 		mh = mh->m_nextpkt;
3046 		m->m_nextpkt = NULL;
3047 #ifndef __NO_STRICT_ALIGNMENT
3048 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3049 			continue;
3050 #endif
3051 #if defined(INET6) || defined(INET)
3052 		if (lro_enabled) {
3053 			if (!lro_possible) {
3054 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3055 				if (lro_possible && mf != NULL) {
3056 					if_input(ifp, mf);
3057 					DBG_COUNTER_INC(rx_if_input);
3058 					mt = mf = NULL;
3059 				}
3060 			}
3061 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC | CSUM_L4_VALID)) ==
3062 			    (CSUM_L4_CALC | CSUM_L4_VALID)) {
3063 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3064 					continue;
3065 			}
3066 		}
3067 #endif
3068 		if (lro_possible) {
3069 			if_input(ifp, m);
3070 			DBG_COUNTER_INC(rx_if_input);
3071 			continue;
3072 		}
3073 
3074 		if (mf == NULL)
3075 			mf = m;
3076 		if (mt != NULL)
3077 			mt->m_nextpkt = m;
3078 		mt = m;
3079 	}
3080 	if (mf != NULL) {
3081 		if_input(ifp, mf);
3082 		DBG_COUNTER_INC(rx_if_input);
3083 	}
3084 
3085 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3086 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3087 
3088 	/*
3089 	 * Flush any outstanding LRO work
3090 	 */
3091 #if defined(INET6) || defined(INET)
3092 	tcp_lro_flush_all(&rxq->ifr_lc);
3093 #endif
3094 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3095 		retval |= IFLIB_RXEOF_MORE;
3096 	return (retval);
3097 err:
3098 	STATE_LOCK(ctx);
3099 	ctx->ifc_flags |= IFC_DO_RESET;
3100 	iflib_admin_intr_deferred(ctx);
3101 	STATE_UNLOCK(ctx);
3102 	return (0);
3103 }
3104 
3105 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1)
3106 static inline qidx_t
3107 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3108 {
3109 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3110 	qidx_t minthresh = txq->ift_size / 8;
3111 	if (in_use > 4 * minthresh)
3112 		return (notify_count);
3113 	if (in_use > 2 * minthresh)
3114 		return (notify_count >> 1);
3115 	if (in_use > minthresh)
3116 		return (notify_count >> 3);
3117 	return (0);
3118 }
3119 
3120 static inline qidx_t
3121 txq_max_rs_deferred(iflib_txq_t txq)
3122 {
3123 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3124 	qidx_t minthresh = txq->ift_size / 8;
3125 	if (txq->ift_in_use > 4 * minthresh)
3126 		return (notify_count);
3127 	if (txq->ift_in_use > 2 * minthresh)
3128 		return (notify_count >> 1);
3129 	if (txq->ift_in_use > minthresh)
3130 		return (notify_count >> 2);
3131 	return (2);
3132 }
3133 
3134 #define M_CSUM_FLAGS(m)		((m)->m_pkthdr.csum_flags)
3135 #define M_HAS_VLANTAG(m)	(m->m_flags & M_VLANTAG)
3136 
3137 #define TXQ_MAX_DB_DEFERRED(txq, in_use)	txq_max_db_deferred((txq), (in_use))
3138 #define TXQ_MAX_RS_DEFERRED(txq)	txq_max_rs_deferred(txq)
3139 #define TXQ_MAX_DB_CONSUMED(size)	(size >> 4)
3140 
3141 /* forward compatibility for cxgb */
3142 #define FIRST_QSET(ctx) 0
3143 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3144 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3145 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3146 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3147 
3148 /* XXX we should be setting this to something other than zero */
3149 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3150 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3151     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3152 
3153 static inline bool
3154 iflib_txd_db_check(iflib_txq_t txq, int ring)
3155 {
3156 	if_ctx_t ctx = txq->ift_ctx;
3157 	qidx_t dbval, max;
3158 
3159 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3160 
3161 	/* force || threshold exceeded || at the edge of the ring */
3162 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3163 
3164 		/*
3165 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3166 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3167 		 * producer index explicitly (INTC).
3168 		 */
3169 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3170 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3171 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3172 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3173 
3174 		/*
3175 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3176 		 */
3177 		txq->ift_db_pending = txq->ift_npending = 0;
3178 		return (true);
3179 	}
3180 	return (false);
3181 }
3182 
3183 #ifdef PKT_DEBUG
3184 static void
3185 print_pkt(if_pkt_info_t pi)
3186 {
3187 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3188 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3189 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3190 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3191 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3192 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3193 }
3194 #endif
3195 
3196 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3197 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3198 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3199 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3200 
3201 /**
3202  * Parses out ethernet header information in the given mbuf.
3203  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3204  *
3205  * This will account for the VLAN header if present.
3206  *
3207  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3208  * types of packets.
3209  */
3210 static int
3211 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3212 {
3213 	struct ether_vlan_header *eh;
3214 	struct mbuf *m;
3215 
3216 	m = *mp;
3217 	if (__predict_false(m->m_len < sizeof(*eh))) {
3218 		(*pullups)++;
3219 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3220 			return (ENOMEM);
3221 	}
3222 	eh = mtod(m, struct ether_vlan_header *);
3223 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3224 		pi->ipi_etype = ntohs(eh->evl_proto);
3225 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3226 	} else {
3227 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3228 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3229 	}
3230 	*mp = m;
3231 
3232 	return (0);
3233 }
3234 
3235 /**
3236  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3237  * Currently this information includes: IP ToS value, IP header version/presence
3238  *
3239  * This is missing some checks and doesn't edit the packet content as it goes,
3240  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3241  */
3242 static int
3243 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3244 {
3245 	struct mbuf *m;
3246 	int err;
3247 
3248 	*pullups = 0;
3249 	m = *mp;
3250 	if (!M_WRITABLE(m)) {
3251 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3252 			return (ENOMEM);
3253 		} else {
3254 			m_freem(*mp);
3255 			DBG_COUNTER_INC(tx_frees);
3256 			*mp = m;
3257 		}
3258 	}
3259 
3260 	/* Fills out pi->ipi_etype */
3261 	err = iflib_parse_ether_header(pi, mp, pullups);
3262 	if (err)
3263 		return (err);
3264 	m = *mp;
3265 
3266 	switch (pi->ipi_etype) {
3267 #ifdef INET
3268 	case ETHERTYPE_IP:
3269 	{
3270 		struct mbuf *n;
3271 		struct ip *ip = NULL;
3272 		int miniplen;
3273 
3274 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3275 		if (__predict_false(m->m_len < miniplen)) {
3276 			/*
3277 			 * Check for common case where the first mbuf only contains
3278 			 * the Ethernet header
3279 			 */
3280 			if (m->m_len == pi->ipi_ehdrlen) {
3281 				n = m->m_next;
3282 				MPASS(n);
3283 				/* If next mbuf contains at least the minimal IP header, then stop */
3284 				if (n->m_len >= sizeof(*ip)) {
3285 					ip = (struct ip *)n->m_data;
3286 				} else {
3287 					(*pullups)++;
3288 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3289 						return (ENOMEM);
3290 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3291 				}
3292 			} else {
3293 				(*pullups)++;
3294 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3295 					return (ENOMEM);
3296 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3297 			}
3298 		} else {
3299 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3300 		}
3301 
3302 		/* Have the IPv4 header w/ no options here */
3303 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3304 		pi->ipi_ipproto = ip->ip_p;
3305 		pi->ipi_ip_tos = ip->ip_tos;
3306 		pi->ipi_flags |= IPI_TX_IPV4;
3307 
3308 		break;
3309 	}
3310 #endif
3311 #ifdef INET6
3312 	case ETHERTYPE_IPV6:
3313 	{
3314 		struct ip6_hdr *ip6;
3315 
3316 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3317 			(*pullups)++;
3318 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3319 				return (ENOMEM);
3320 		}
3321 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3322 
3323 		/* Have the IPv6 fixed header here */
3324 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3325 		pi->ipi_ipproto = ip6->ip6_nxt;
3326 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3327 		pi->ipi_flags |= IPI_TX_IPV6;
3328 
3329 		break;
3330 	}
3331 #endif
3332 	default:
3333 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3334 		pi->ipi_ip_hlen = 0;
3335 		break;
3336 	}
3337 	*mp = m;
3338 
3339 	return (0);
3340 
3341 }
3342 
3343 static int
3344 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3345 {
3346 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3347 	struct mbuf *m;
3348 	int err;
3349 
3350 	m = *mp;
3351 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3352 	    M_WRITABLE(m) == 0) {
3353 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3354 			return (ENOMEM);
3355 		} else {
3356 			m_freem(*mp);
3357 			DBG_COUNTER_INC(tx_frees);
3358 			*mp = m;
3359 		}
3360 	}
3361 
3362 	/* Fills out pi->ipi_etype */
3363 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3364 	if (__predict_false(err))
3365 		return (err);
3366 	m = *mp;
3367 
3368 	switch (pi->ipi_etype) {
3369 #ifdef INET
3370 	case ETHERTYPE_IP:
3371 	{
3372 		struct mbuf *n;
3373 		struct ip *ip = NULL;
3374 		struct tcphdr *th = NULL;
3375 		int minthlen;
3376 
3377 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3378 		if (__predict_false(m->m_len < minthlen)) {
3379 			/*
3380 			 * if this code bloat is causing too much of a hit
3381 			 * move it to a separate function and mark it noinline
3382 			 */
3383 			if (m->m_len == pi->ipi_ehdrlen) {
3384 				n = m->m_next;
3385 				MPASS(n);
3386 				if (n->m_len >= sizeof(*ip))  {
3387 					ip = (struct ip *)n->m_data;
3388 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3389 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3390 				} else {
3391 					txq->ift_pullups++;
3392 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3393 						return (ENOMEM);
3394 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3395 				}
3396 			} else {
3397 				txq->ift_pullups++;
3398 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3399 					return (ENOMEM);
3400 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3401 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3402 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3403 			}
3404 		} else {
3405 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3406 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3407 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3408 		}
3409 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3410 		pi->ipi_ipproto = ip->ip_p;
3411 		pi->ipi_ip_tos = ip->ip_tos;
3412 		pi->ipi_flags |= IPI_TX_IPV4;
3413 
3414 		/* TCP checksum offload may require TCP header length */
3415 		if (IS_TX_OFFLOAD4(pi)) {
3416 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3417 				if (__predict_false(th == NULL)) {
3418 					txq->ift_pullups++;
3419 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3420 						return (ENOMEM);
3421 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3422 				}
3423 				pi->ipi_tcp_hflags = th->th_flags;
3424 				pi->ipi_tcp_hlen = th->th_off << 2;
3425 				pi->ipi_tcp_seq = th->th_seq;
3426 			}
3427 			if (IS_TSO4(pi)) {
3428 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3429 					return (ENXIO);
3430 				/*
3431 				 * TSO always requires hardware checksum offload.
3432 				 */
3433 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3434 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3435 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3436 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3437 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3438 					ip->ip_sum = 0;
3439 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3440 				}
3441 			}
3442 		}
3443 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3444 			ip->ip_sum = 0;
3445 
3446 		break;
3447 	}
3448 #endif
3449 #ifdef INET6
3450 	case ETHERTYPE_IPV6:
3451 	{
3452 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3453 		struct tcphdr *th;
3454 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3455 
3456 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3457 			txq->ift_pullups++;
3458 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3459 				return (ENOMEM);
3460 		}
3461 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3462 
3463 		/* XXX-BZ this will go badly in case of ext hdrs. */
3464 		pi->ipi_ipproto = ip6->ip6_nxt;
3465 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3466 		pi->ipi_flags |= IPI_TX_IPV6;
3467 
3468 		/* TCP checksum offload may require TCP header length */
3469 		if (IS_TX_OFFLOAD6(pi)) {
3470 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3471 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3472 					txq->ift_pullups++;
3473 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3474 						return (ENOMEM);
3475 				}
3476 				pi->ipi_tcp_hflags = th->th_flags;
3477 				pi->ipi_tcp_hlen = th->th_off << 2;
3478 				pi->ipi_tcp_seq = th->th_seq;
3479 			}
3480 			if (IS_TSO6(pi)) {
3481 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3482 					return (ENXIO);
3483 				/*
3484 				 * TSO always requires hardware checksum offload.
3485 				 */
3486 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3487 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3488 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3489 			}
3490 		}
3491 		break;
3492 	}
3493 #endif
3494 	default:
3495 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3496 		pi->ipi_ip_hlen = 0;
3497 		break;
3498 	}
3499 	*mp = m;
3500 
3501 	return (0);
3502 }
3503 
3504 /*
3505  * If dodgy hardware rejects the scatter gather chain we've handed it
3506  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3507  * m_defrag'd mbufs
3508  */
3509 static __noinline struct mbuf *
3510 iflib_remove_mbuf(iflib_txq_t txq)
3511 {
3512 	int ntxd, pidx;
3513 	struct mbuf *m, **ifsd_m;
3514 
3515 	ifsd_m = txq->ift_sds.ifsd_m;
3516 	ntxd = txq->ift_size;
3517 	pidx = txq->ift_pidx & (ntxd - 1);
3518 	ifsd_m = txq->ift_sds.ifsd_m;
3519 	m = ifsd_m[pidx];
3520 	ifsd_m[pidx] = NULL;
3521 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3522 	if (txq->ift_sds.ifsd_tso_map != NULL)
3523 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3524 		    txq->ift_sds.ifsd_tso_map[pidx]);
3525 #if MEMORY_LOGGING
3526 	txq->ift_dequeued++;
3527 #endif
3528 	return (m);
3529 }
3530 
3531 static inline caddr_t
3532 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3533 {
3534 	qidx_t size;
3535 	int ntxd;
3536 	caddr_t start, end, cur, next;
3537 
3538 	ntxd = txq->ift_size;
3539 	size = txq->ift_txd_size[qid];
3540 	start = txq->ift_ifdi[qid].idi_vaddr;
3541 
3542 	if (__predict_false(size == 0))
3543 		return (start);
3544 	cur = start + size * cidx;
3545 	end = start + size * ntxd;
3546 	next = CACHE_PTR_NEXT(cur);
3547 	return (next < end ? next : start);
3548 }
3549 
3550 /*
3551  * Pad an mbuf to ensure a minimum ethernet frame size.
3552  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3553  */
3554 static __noinline int
3555 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3556 {
3557 	/*
3558 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3559 	 * and ARP message is the smallest common payload I can think of
3560 	 */
3561 	static char pad[18];	/* just zeros */
3562 	int n;
3563 	struct mbuf *new_head;
3564 
3565 	if (!M_WRITABLE(*m_head)) {
3566 		new_head = m_dup(*m_head, M_NOWAIT);
3567 		if (new_head == NULL) {
3568 			m_freem(*m_head);
3569 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3570 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3571 			DBG_COUNTER_INC(tx_frees);
3572 			return (ENOMEM);
3573 		}
3574 		m_freem(*m_head);
3575 		*m_head = new_head;
3576 	}
3577 
3578 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3579 	     n > 0; n -= sizeof(pad))
3580 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3581 			break;
3582 
3583 	if (n > 0) {
3584 		m_freem(*m_head);
3585 		device_printf(dev, "cannot pad short frame\n");
3586 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3587 		DBG_COUNTER_INC(tx_frees);
3588 		return (ENOBUFS);
3589 	}
3590 
3591 	return (0);
3592 }
3593 
3594 static int
3595 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3596 {
3597 	if_ctx_t		ctx;
3598 	if_shared_ctx_t		sctx;
3599 	if_softc_ctx_t		scctx;
3600 	bus_dma_tag_t		buf_tag;
3601 	bus_dma_segment_t	*segs;
3602 	struct mbuf		*m_head, **ifsd_m;
3603 	void			*next_txd;
3604 	bus_dmamap_t		map;
3605 	struct if_pkt_info	pi;
3606 	int remap = 0;
3607 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3608 
3609 	ctx = txq->ift_ctx;
3610 	sctx = ctx->ifc_sctx;
3611 	scctx = &ctx->ifc_softc_ctx;
3612 	segs = txq->ift_segs;
3613 	ntxd = txq->ift_size;
3614 	m_head = *m_headp;
3615 	map = NULL;
3616 
3617 	/*
3618 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3619 	 */
3620 	cidx = txq->ift_cidx;
3621 	pidx = txq->ift_pidx;
3622 	if (ctx->ifc_flags & IFC_PREFETCH) {
3623 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd - 1);
3624 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3625 			next_txd = calc_next_txd(txq, cidx, 0);
3626 			prefetch(next_txd);
3627 		}
3628 
3629 		/* prefetch the next cache line of mbuf pointers and flags */
3630 		prefetch(&txq->ift_sds.ifsd_m[next]);
3631 		prefetch(&txq->ift_sds.ifsd_map[next]);
3632 		next = (cidx + CACHE_LINE_SIZE) & (ntxd - 1);
3633 	}
3634 	map = txq->ift_sds.ifsd_map[pidx];
3635 	ifsd_m = txq->ift_sds.ifsd_m;
3636 
3637 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3638 		buf_tag = txq->ift_tso_buf_tag;
3639 		max_segs = scctx->isc_tx_tso_segments_max;
3640 		map = txq->ift_sds.ifsd_tso_map[pidx];
3641 		MPASS(buf_tag != NULL);
3642 		MPASS(max_segs > 0);
3643 	} else {
3644 		buf_tag = txq->ift_buf_tag;
3645 		max_segs = scctx->isc_tx_nsegments;
3646 		map = txq->ift_sds.ifsd_map[pidx];
3647 	}
3648 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3649 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3650 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3651 		if (err) {
3652 			DBG_COUNTER_INC(encap_txd_encap_fail);
3653 			return (err);
3654 		}
3655 	}
3656 	m_head = *m_headp;
3657 
3658 	pkt_info_zero(&pi);
3659 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST));
3660 	pi.ipi_pidx = pidx;
3661 	pi.ipi_qsidx = txq->ift_id;
3662 	pi.ipi_len = m_head->m_pkthdr.len;
3663 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3664 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3665 
3666 	/* deliberate bitwise OR to make one condition */
3667 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3668 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3669 			DBG_COUNTER_INC(encap_txd_encap_fail);
3670 			return (err);
3671 		}
3672 		m_head = *m_headp;
3673 	}
3674 
3675 retry:
3676 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3677 	    BUS_DMA_NOWAIT);
3678 defrag:
3679 	if (__predict_false(err)) {
3680 		switch (err) {
3681 		case EFBIG:
3682 			/* try collapse once and defrag once */
3683 			if (remap == 0) {
3684 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3685 				/* try defrag if collapsing fails */
3686 				if (m_head == NULL)
3687 					remap++;
3688 			}
3689 			if (remap == 1) {
3690 				txq->ift_mbuf_defrag++;
3691 				m_head = m_defrag(*m_headp, M_NOWAIT);
3692 			}
3693 			/*
3694 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3695 			 * failed to map an mbuf that was run through m_defrag
3696 			 */
3697 			MPASS(remap <= 1);
3698 			if (__predict_false(m_head == NULL || remap > 1))
3699 				goto defrag_failed;
3700 			remap++;
3701 			*m_headp = m_head;
3702 			goto retry;
3703 			break;
3704 		case ENOMEM:
3705 			txq->ift_no_tx_dma_setup++;
3706 			break;
3707 		default:
3708 			txq->ift_no_tx_dma_setup++;
3709 			m_freem(*m_headp);
3710 			DBG_COUNTER_INC(tx_frees);
3711 			*m_headp = NULL;
3712 			break;
3713 		}
3714 		txq->ift_map_failed++;
3715 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3716 		DBG_COUNTER_INC(encap_txd_encap_fail);
3717 		return (err);
3718 	}
3719 	ifsd_m[pidx] = m_head;
3720 	/*
3721 	 * XXX assumes a 1 to 1 relationship between segments and
3722 	 *        descriptors - this does not hold true on all drivers, e.g.
3723 	 *        cxgb
3724 	 */
3725 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3726 		txq->ift_no_desc_avail++;
3727 		bus_dmamap_unload(buf_tag, map);
3728 		DBG_COUNTER_INC(encap_txq_avail_fail);
3729 		DBG_COUNTER_INC(encap_txd_encap_fail);
3730 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3731 			GROUPTASK_ENQUEUE(&txq->ift_task);
3732 		return (ENOBUFS);
3733 	}
3734 	/*
3735 	 * On Intel cards we can greatly reduce the number of TX interrupts
3736 	 * we see by only setting report status on every Nth descriptor.
3737 	 * However, this also means that the driver will need to keep track
3738 	 * of the descriptors that RS was set on to check them for the DD bit.
3739 	 */
3740 	txq->ift_rs_pending += nsegs + 1;
3741 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3742 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3743 		pi.ipi_flags |= IPI_TX_INTR;
3744 		txq->ift_rs_pending = 0;
3745 	}
3746 
3747 	pi.ipi_segs = segs;
3748 	pi.ipi_nsegs = nsegs;
3749 
3750 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3751 #ifdef PKT_DEBUG
3752 	print_pkt(&pi);
3753 #endif
3754 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3755 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3756 		DBG_COUNTER_INC(tx_encap);
3757 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3758 
3759 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3760 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3761 			ndesc += txq->ift_size;
3762 			txq->ift_gen = 1;
3763 		}
3764 		/*
3765 		 * drivers can need as many as
3766 		 * two sentinels
3767 		 */
3768 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3769 		MPASS(pi.ipi_new_pidx != pidx);
3770 		MPASS(ndesc > 0);
3771 		txq->ift_in_use += ndesc;
3772 		txq->ift_db_pending += ndesc;
3773 
3774 		/*
3775 		 * We update the last software descriptor again here because there may
3776 		 * be a sentinel and/or there may be more mbufs than segments
3777 		 */
3778 		txq->ift_pidx = pi.ipi_new_pidx;
3779 		txq->ift_npending += pi.ipi_ndescs;
3780 	} else {
3781 		*m_headp = m_head = iflib_remove_mbuf(txq);
3782 		if (err == EFBIG) {
3783 			txq->ift_txd_encap_efbig++;
3784 			if (remap < 2) {
3785 				remap = 1;
3786 				goto defrag;
3787 			}
3788 		}
3789 		goto defrag_failed;
3790 	}
3791 	/*
3792 	 * err can't possibly be non-zero here, so we don't neet to test it
3793 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3794 	 */
3795 	return (err);
3796 
3797 defrag_failed:
3798 	txq->ift_mbuf_defrag_failed++;
3799 	txq->ift_map_failed++;
3800 	m_freem(*m_headp);
3801 	DBG_COUNTER_INC(tx_frees);
3802 	*m_headp = NULL;
3803 	DBG_COUNTER_INC(encap_txd_encap_fail);
3804 	return (ENOMEM);
3805 }
3806 
3807 static void
3808 iflib_tx_desc_free(iflib_txq_t txq, int n)
3809 {
3810 	uint32_t qsize, cidx, mask, gen;
3811 	struct mbuf *m, **ifsd_m;
3812 	bool do_prefetch;
3813 
3814 	cidx = txq->ift_cidx;
3815 	gen = txq->ift_gen;
3816 	qsize = txq->ift_size;
3817 	mask = qsize - 1;
3818 	ifsd_m = txq->ift_sds.ifsd_m;
3819 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3820 
3821 	while (n-- > 0) {
3822 		if (do_prefetch) {
3823 			prefetch(ifsd_m[(cidx + 3) & mask]);
3824 			prefetch(ifsd_m[(cidx + 4) & mask]);
3825 		}
3826 		if ((m = ifsd_m[cidx]) != NULL) {
3827 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3828 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3829 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3830 				    txq->ift_sds.ifsd_tso_map[cidx],
3831 				    BUS_DMASYNC_POSTWRITE);
3832 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3833 				    txq->ift_sds.ifsd_tso_map[cidx]);
3834 			} else {
3835 				bus_dmamap_sync(txq->ift_buf_tag,
3836 				    txq->ift_sds.ifsd_map[cidx],
3837 				    BUS_DMASYNC_POSTWRITE);
3838 				bus_dmamap_unload(txq->ift_buf_tag,
3839 				    txq->ift_sds.ifsd_map[cidx]);
3840 			}
3841 			/* XXX we don't support any drivers that batch packets yet */
3842 			MPASS(m->m_nextpkt == NULL);
3843 			m_freem(m);
3844 			ifsd_m[cidx] = NULL;
3845 #if MEMORY_LOGGING
3846 			txq->ift_dequeued++;
3847 #endif
3848 			DBG_COUNTER_INC(tx_frees);
3849 		}
3850 		if (__predict_false(++cidx == qsize)) {
3851 			cidx = 0;
3852 			gen = 0;
3853 		}
3854 	}
3855 	txq->ift_cidx = cidx;
3856 	txq->ift_gen = gen;
3857 }
3858 
3859 static __inline int
3860 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3861 {
3862 	int reclaim;
3863 	if_ctx_t ctx = txq->ift_ctx;
3864 
3865 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3866 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3867 
3868 	/*
3869 	 * Need a rate-limiting check so that this isn't called every time
3870 	 */
3871 	iflib_tx_credits_update(ctx, txq);
3872 	reclaim = DESC_RECLAIMABLE(txq);
3873 
3874 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3875 #ifdef INVARIANTS
3876 		if (iflib_verbose_debug) {
3877 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__,
3878 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3879 			       reclaim, thresh);
3880 		}
3881 #endif
3882 		return (0);
3883 	}
3884 	iflib_tx_desc_free(txq, reclaim);
3885 	txq->ift_cleaned += reclaim;
3886 	txq->ift_in_use -= reclaim;
3887 
3888 	return (reclaim);
3889 }
3890 
3891 static struct mbuf **
3892 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3893 {
3894 	int next, size;
3895 	struct mbuf **items;
3896 
3897 	size = r->size;
3898 	next = (cidx + CACHE_PTR_INCREMENT) & (size - 1);
3899 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3900 
3901 	prefetch(items[(cidx + offset) & (size - 1)]);
3902 	if (remaining > 1) {
3903 		prefetch2cachelines(&items[next]);
3904 		prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]);
3905 		prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]);
3906 		prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]);
3907 	}
3908 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)]));
3909 }
3910 
3911 static void
3912 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3913 {
3914 
3915 	ifmp_ring_check_drainage(txq->ift_br, budget);
3916 }
3917 
3918 static uint32_t
3919 iflib_txq_can_drain(struct ifmp_ring *r)
3920 {
3921 	iflib_txq_t txq = r->cookie;
3922 	if_ctx_t ctx = txq->ift_ctx;
3923 
3924 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3925 		return (1);
3926 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3927 	    BUS_DMASYNC_POSTREAD);
3928 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3929 	    false));
3930 }
3931 
3932 static uint32_t
3933 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3934 {
3935 	iflib_txq_t txq = r->cookie;
3936 	if_ctx_t ctx = txq->ift_ctx;
3937 	if_t ifp = ctx->ifc_ifp;
3938 	struct mbuf *m, **mp;
3939 	int avail, bytes_sent, skipped, count, err, i;
3940 	int mcast_sent, pkt_sent, reclaimed;
3941 	bool do_prefetch, rang, ring;
3942 
3943 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3944 			    !LINK_ACTIVE(ctx))) {
3945 		DBG_COUNTER_INC(txq_drain_notready);
3946 		return (0);
3947 	}
3948 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3949 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3950 	avail = IDXDIFF(pidx, cidx, r->size);
3951 
3952 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3953 		/*
3954 		 * The driver is unloading so we need to free all pending packets.
3955 		 */
3956 		DBG_COUNTER_INC(txq_drain_flushing);
3957 		for (i = 0; i < avail; i++) {
3958 			if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq))
3959 				m_freem(r->items[(cidx + i) & (r->size - 1)]);
3960 			r->items[(cidx + i) & (r->size - 1)] = NULL;
3961 		}
3962 		return (avail);
3963 	}
3964 
3965 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3966 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3967 		CALLOUT_LOCK(txq);
3968 		callout_stop(&txq->ift_timer);
3969 		CALLOUT_UNLOCK(txq);
3970 		DBG_COUNTER_INC(txq_drain_oactive);
3971 		return (0);
3972 	}
3973 
3974 	/*
3975 	 * If we've reclaimed any packets this queue cannot be hung.
3976 	 */
3977 	if (reclaimed)
3978 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3979 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3980 	count = MIN(avail, TX_BATCH_SIZE);
3981 #ifdef INVARIANTS
3982 	if (iflib_verbose_debug)
3983 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__,
3984 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3985 #endif
3986 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3987 	err = 0;
3988 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3989 		int rem = do_prefetch ? count - i : 0;
3990 
3991 		mp = _ring_peek_one(r, cidx, i, rem);
3992 		MPASS(mp != NULL && *mp != NULL);
3993 
3994 		/*
3995 		 * Completion interrupts will use the address of the txq
3996 		 * as a sentinel to enqueue _something_ in order to acquire
3997 		 * the lock on the mp_ring (there's no direct lock call).
3998 		 * We obviously whave to check for these sentinel cases
3999 		 * and skip them.
4000 		 */
4001 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4002 			skipped++;
4003 			continue;
4004 		}
4005 		err = iflib_encap(txq, mp);
4006 		if (__predict_false(err)) {
4007 			/* no room - bail out */
4008 			if (err == ENOBUFS)
4009 				break;
4010 			skipped++;
4011 			/* we can't send this packet - skip it */
4012 			continue;
4013 		}
4014 		pkt_sent++;
4015 		m = *mp;
4016 		DBG_COUNTER_INC(tx_sent);
4017 		bytes_sent += m->m_pkthdr.len;
4018 		mcast_sent += !!(m->m_flags & M_MCAST);
4019 
4020 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4021 			break;
4022 		ETHER_BPF_MTAP(ifp, m);
4023 		rang = iflib_txd_db_check(txq, false);
4024 	}
4025 
4026 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4027 	ring = rang ? false  : (iflib_min_tx_latency | err);
4028 	iflib_txd_db_check(txq, ring);
4029 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4030 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4031 	if (mcast_sent)
4032 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4033 #ifdef INVARIANTS
4034 	if (iflib_verbose_debug)
4035 		printf("consumed=%d\n", skipped + pkt_sent);
4036 #endif
4037 	return (skipped + pkt_sent);
4038 }
4039 
4040 static uint32_t
4041 iflib_txq_drain_always(struct ifmp_ring *r)
4042 {
4043 	return (1);
4044 }
4045 
4046 static uint32_t
4047 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4048 {
4049 	int i, avail;
4050 	struct mbuf **mp;
4051 	iflib_txq_t txq;
4052 
4053 	txq = r->cookie;
4054 
4055 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4056 	CALLOUT_LOCK(txq);
4057 	callout_stop(&txq->ift_timer);
4058 	CALLOUT_UNLOCK(txq);
4059 
4060 	avail = IDXDIFF(pidx, cidx, r->size);
4061 	for (i = 0; i < avail; i++) {
4062 		mp = _ring_peek_one(r, cidx, i, avail - i);
4063 		if (__predict_false(*mp == (struct mbuf *)txq))
4064 			continue;
4065 		m_freem(*mp);
4066 		DBG_COUNTER_INC(tx_frees);
4067 	}
4068 	MPASS(ifmp_ring_is_stalled(r) == 0);
4069 	return (avail);
4070 }
4071 
4072 static void
4073 iflib_ifmp_purge(iflib_txq_t txq)
4074 {
4075 	struct ifmp_ring *r;
4076 
4077 	r = txq->ift_br;
4078 	r->drain = iflib_txq_drain_free;
4079 	r->can_drain = iflib_txq_drain_always;
4080 
4081 	ifmp_ring_check_drainage(r, r->size);
4082 
4083 	r->drain = iflib_txq_drain;
4084 	r->can_drain = iflib_txq_can_drain;
4085 }
4086 
4087 static void
4088 _task_fn_tx(void *context)
4089 {
4090 	iflib_txq_t txq = context;
4091 	if_ctx_t ctx = txq->ift_ctx;
4092 	if_t ifp = ctx->ifc_ifp;
4093 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4094 
4095 #ifdef IFLIB_DIAGNOSTICS
4096 	txq->ift_cpu_exec_count[curcpu]++;
4097 #endif
4098 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4099 		return;
4100 #ifdef DEV_NETMAP
4101 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4102 	    netmap_tx_irq(ifp, txq->ift_id))
4103 		goto skip_ifmp;
4104 #endif
4105 #ifdef ALTQ
4106 	if (if_altq_is_enabled(ifp))
4107 		iflib_altq_if_start(ifp);
4108 #endif
4109 	if (txq->ift_db_pending)
4110 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4111 	else if (!abdicate)
4112 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4113 	/*
4114 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4115 	 */
4116 	if (abdicate)
4117 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4118 #ifdef DEV_NETMAP
4119 skip_ifmp:
4120 #endif
4121 	if (ctx->ifc_flags & IFC_LEGACY)
4122 		IFDI_INTR_ENABLE(ctx);
4123 	else
4124 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4125 }
4126 
4127 static void
4128 _task_fn_rx(void *context)
4129 {
4130 	iflib_rxq_t rxq = context;
4131 	if_ctx_t ctx = rxq->ifr_ctx;
4132 	uint8_t more;
4133 	uint16_t budget;
4134 #ifdef DEV_NETMAP
4135 	u_int work = 0;
4136 	int nmirq;
4137 #endif
4138 
4139 #ifdef IFLIB_DIAGNOSTICS
4140 	rxq->ifr_cpu_exec_count[curcpu]++;
4141 #endif
4142 	DBG_COUNTER_INC(task_fn_rxs);
4143 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4144 		return;
4145 #ifdef DEV_NETMAP
4146 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4147 	if (nmirq != NM_IRQ_PASS) {
4148 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4149 		goto skip_rxeof;
4150 	}
4151 #endif
4152 	budget = ctx->ifc_sysctl_rx_budget;
4153 	if (budget == 0)
4154 		budget = 16;	/* XXX */
4155 	more = iflib_rxeof(rxq, budget);
4156 #ifdef DEV_NETMAP
4157 skip_rxeof:
4158 #endif
4159 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4160 		if (ctx->ifc_flags & IFC_LEGACY)
4161 			IFDI_INTR_ENABLE(ctx);
4162 		else
4163 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4164 		DBG_COUNTER_INC(rx_intr_enables);
4165 	}
4166 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4167 		return;
4168 
4169 	if (more & IFLIB_RXEOF_MORE)
4170 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4171 	else if (more & IFLIB_RXEOF_EMPTY)
4172 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4173 }
4174 
4175 static void
4176 _task_fn_admin(void *context)
4177 {
4178 	if_ctx_t ctx = context;
4179 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4180 	iflib_txq_t txq;
4181 	int i;
4182 	bool oactive, running, do_reset, do_watchdog, in_detach;
4183 
4184 	STATE_LOCK(ctx);
4185 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4186 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4187 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4188 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4189 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4190 	ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_WATCHDOG);
4191 	STATE_UNLOCK(ctx);
4192 
4193 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4194 		return;
4195 	if (in_detach)
4196 		return;
4197 
4198 	CTX_LOCK(ctx);
4199 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4200 		CALLOUT_LOCK(txq);
4201 		callout_stop(&txq->ift_timer);
4202 		CALLOUT_UNLOCK(txq);
4203 	}
4204 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4205 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4206 	if (do_watchdog) {
4207 		ctx->ifc_watchdog_events++;
4208 		IFDI_WATCHDOG_RESET(ctx);
4209 	}
4210 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4211 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4212 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4213 		    txq->ift_timer.c_cpu);
4214 	}
4215 	IFDI_LINK_INTR_ENABLE(ctx);
4216 	if (do_reset)
4217 		iflib_if_init_locked(ctx);
4218 	CTX_UNLOCK(ctx);
4219 
4220 	if (LINK_ACTIVE(ctx) == 0)
4221 		return;
4222 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4223 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4224 }
4225 
4226 static void
4227 _task_fn_iov(void *context)
4228 {
4229 	if_ctx_t ctx = context;
4230 
4231 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4232 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4233 		return;
4234 
4235 	CTX_LOCK(ctx);
4236 	IFDI_VFLR_HANDLE(ctx);
4237 	CTX_UNLOCK(ctx);
4238 }
4239 
4240 static int
4241 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4242 {
4243 	int err;
4244 	if_int_delay_info_t info;
4245 	if_ctx_t ctx;
4246 
4247 	info = (if_int_delay_info_t)arg1;
4248 	ctx = info->iidi_ctx;
4249 	info->iidi_req = req;
4250 	info->iidi_oidp = oidp;
4251 	CTX_LOCK(ctx);
4252 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4253 	CTX_UNLOCK(ctx);
4254 	return (err);
4255 }
4256 
4257 /*********************************************************************
4258  *
4259  *  IFNET FUNCTIONS
4260  *
4261  **********************************************************************/
4262 
4263 static void
4264 iflib_if_init_locked(if_ctx_t ctx)
4265 {
4266 	iflib_stop(ctx);
4267 	iflib_init_locked(ctx);
4268 }
4269 
4270 static void
4271 iflib_if_init(void *arg)
4272 {
4273 	if_ctx_t ctx = arg;
4274 
4275 	CTX_LOCK(ctx);
4276 	iflib_if_init_locked(ctx);
4277 	CTX_UNLOCK(ctx);
4278 }
4279 
4280 static int
4281 iflib_if_transmit(if_t ifp, struct mbuf *m)
4282 {
4283 	if_ctx_t ctx = if_getsoftc(ifp);
4284 	iflib_txq_t txq;
4285 	int err, qidx;
4286 	int abdicate;
4287 
4288 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4289 		DBG_COUNTER_INC(tx_frees);
4290 		m_freem(m);
4291 		return (ENETDOWN);
4292 	}
4293 
4294 	MPASS(m->m_nextpkt == NULL);
4295 	/* ALTQ-enabled interfaces always use queue 0. */
4296 	qidx = 0;
4297 	/* Use driver-supplied queue selection method if it exists */
4298 	if (ctx->isc_txq_select_v2) {
4299 		struct if_pkt_info pi;
4300 		uint64_t early_pullups = 0;
4301 		pkt_info_zero(&pi);
4302 
4303 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4304 		if (__predict_false(err != 0)) {
4305 			/* Assign pullups for bad pkts to default queue */
4306 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4307 			DBG_COUNTER_INC(encap_txd_encap_fail);
4308 			return (err);
4309 		}
4310 		/* Let driver make queueing decision */
4311 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4312 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4313 	}
4314 	/* Backwards compatibility w/ simpler queue select */
4315 	else if (ctx->isc_txq_select)
4316 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4317 	/* If not, use iflib's standard method */
4318 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4319 		qidx = QIDX(ctx, m);
4320 
4321 	/* Set TX queue */
4322 	txq = &ctx->ifc_txqs[qidx];
4323 
4324 #ifdef DRIVER_BACKPRESSURE
4325 	if (txq->ift_closed) {
4326 		while (m != NULL) {
4327 			next = m->m_nextpkt;
4328 			m->m_nextpkt = NULL;
4329 			m_freem(m);
4330 			DBG_COUNTER_INC(tx_frees);
4331 			m = next;
4332 		}
4333 		return (ENOBUFS);
4334 	}
4335 #endif
4336 #ifdef notyet
4337 	qidx = count = 0;
4338 	mp = marr;
4339 	next = m;
4340 	do {
4341 		count++;
4342 		next = next->m_nextpkt;
4343 	} while (next != NULL);
4344 
4345 	if (count > nitems(marr))
4346 		if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4347 			/* XXX check nextpkt */
4348 			m_freem(m);
4349 			/* XXX simplify for now */
4350 			DBG_COUNTER_INC(tx_frees);
4351 			return (ENOBUFS);
4352 		}
4353 	for (next = m, i = 0; next != NULL; i++) {
4354 		mp[i] = next;
4355 		next = next->m_nextpkt;
4356 		mp[i]->m_nextpkt = NULL;
4357 	}
4358 #endif
4359 	DBG_COUNTER_INC(tx_seen);
4360 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4361 
4362 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4363 
4364 	if (abdicate)
4365 		GROUPTASK_ENQUEUE(&txq->ift_task);
4366  	if (err) {
4367 		if (!abdicate)
4368 			GROUPTASK_ENQUEUE(&txq->ift_task);
4369 		/* support forthcoming later */
4370 #ifdef DRIVER_BACKPRESSURE
4371 		txq->ift_closed = TRUE;
4372 #endif
4373 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4374 		m_freem(m);
4375 		DBG_COUNTER_INC(tx_frees);
4376 	}
4377 
4378 	return (err);
4379 }
4380 
4381 #ifdef ALTQ
4382 /*
4383  * The overall approach to integrating iflib with ALTQ is to continue to use
4384  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4385  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4386  * is redundant/unnecessary, but doing so minimizes the amount of
4387  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4388  * redundantly queueing to an intermediate mp_ring is swamped by the
4389  * performance limitations inherent in using ALTQ.
4390  *
4391  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4392  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4393  * given interface.  If ALTQ is enabled for an interface, then all
4394  * transmitted packets for that interface will be submitted to the ALTQ
4395  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4396  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4397  * update stats that the iflib machinery handles, and which is sensitve to
4398  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4399  * will be installed as the start routine for use by ALTQ facilities that
4400  * need to trigger queue drains on a scheduled basis.
4401  *
4402  */
4403 static void
4404 iflib_altq_if_start(if_t ifp)
4405 {
4406 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4407 	struct mbuf *m;
4408 
4409 	IFQ_LOCK(ifq);
4410 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4411 	while (m != NULL) {
4412 		iflib_if_transmit(ifp, m);
4413 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4414 	}
4415 	IFQ_UNLOCK(ifq);
4416 }
4417 
4418 static int
4419 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4420 {
4421 	int err;
4422 
4423 	if (if_altq_is_enabled(ifp)) {
4424 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4425 		if (err == 0)
4426 			iflib_altq_if_start(ifp);
4427 	} else
4428 		err = iflib_if_transmit(ifp, m);
4429 
4430 	return (err);
4431 }
4432 #endif /* ALTQ */
4433 
4434 static void
4435 iflib_if_qflush(if_t ifp)
4436 {
4437 	if_ctx_t ctx = if_getsoftc(ifp);
4438 	iflib_txq_t txq = ctx->ifc_txqs;
4439 	int i;
4440 
4441 	STATE_LOCK(ctx);
4442 	ctx->ifc_flags |= IFC_QFLUSH;
4443 	STATE_UNLOCK(ctx);
4444 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4445 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4446 			iflib_txq_check_drain(txq, 0);
4447 	STATE_LOCK(ctx);
4448 	ctx->ifc_flags &= ~IFC_QFLUSH;
4449 	STATE_UNLOCK(ctx);
4450 
4451 	/*
4452 	 * When ALTQ is enabled, this will also take care of purging the
4453 	 * ALTQ queue(s).
4454 	 */
4455 	if_qflush(ifp);
4456 }
4457 
4458 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4459 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4460 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4461 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4462 
4463 static int
4464 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4465 {
4466 	if_ctx_t ctx = if_getsoftc(ifp);
4467 	struct ifreq	*ifr = (struct ifreq *)data;
4468 #if defined(INET) || defined(INET6)
4469 	struct ifaddr	*ifa = (struct ifaddr *)data;
4470 #endif
4471 	bool		avoid_reset = false;
4472 	int		err = 0, reinit = 0, bits;
4473 
4474 	switch (command) {
4475 	case SIOCSIFADDR:
4476 #ifdef INET
4477 		if (ifa->ifa_addr->sa_family == AF_INET)
4478 			avoid_reset = true;
4479 #endif
4480 #ifdef INET6
4481 		if (ifa->ifa_addr->sa_family == AF_INET6)
4482 			avoid_reset = true;
4483 #endif
4484 		/*
4485 		 * Calling init results in link renegotiation,
4486 		 * so we avoid doing it when possible.
4487 		 */
4488 		if (avoid_reset) {
4489 			if_setflagbits(ifp, IFF_UP, 0);
4490 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4491 				reinit = 1;
4492 #ifdef INET
4493 			if (!(if_getflags(ifp) & IFF_NOARP))
4494 				arp_ifinit(ifp, ifa);
4495 #endif
4496 		} else
4497 			err = ether_ioctl(ifp, command, data);
4498 		break;
4499 	case SIOCSIFMTU:
4500 		CTX_LOCK(ctx);
4501 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4502 			CTX_UNLOCK(ctx);
4503 			break;
4504 		}
4505 		bits = if_getdrvflags(ifp);
4506 		/* stop the driver and free any clusters before proceeding */
4507 		iflib_stop(ctx);
4508 
4509 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4510 			STATE_LOCK(ctx);
4511 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4512 				ctx->ifc_flags |= IFC_MULTISEG;
4513 			else
4514 				ctx->ifc_flags &= ~IFC_MULTISEG;
4515 			STATE_UNLOCK(ctx);
4516 			err = if_setmtu(ifp, ifr->ifr_mtu);
4517 		}
4518 		iflib_init_locked(ctx);
4519 		STATE_LOCK(ctx);
4520 		if_setdrvflags(ifp, bits);
4521 		STATE_UNLOCK(ctx);
4522 		CTX_UNLOCK(ctx);
4523 		break;
4524 	case SIOCSIFFLAGS:
4525 		CTX_LOCK(ctx);
4526 		if (if_getflags(ifp) & IFF_UP) {
4527 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4528 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4529 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4530 					CTX_UNLOCK(ctx);
4531 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4532 					CTX_LOCK(ctx);
4533 				}
4534 			} else
4535 				reinit = 1;
4536 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4537 			iflib_stop(ctx);
4538 		}
4539 		ctx->ifc_if_flags = if_getflags(ifp);
4540 		CTX_UNLOCK(ctx);
4541 		break;
4542 	case SIOCADDMULTI:
4543 	case SIOCDELMULTI:
4544 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4545 			CTX_LOCK(ctx);
4546 			IFDI_INTR_DISABLE(ctx);
4547 			IFDI_MULTI_SET(ctx);
4548 			IFDI_INTR_ENABLE(ctx);
4549 			CTX_UNLOCK(ctx);
4550 		}
4551 		break;
4552 	case SIOCSIFMEDIA:
4553 		CTX_LOCK(ctx);
4554 		IFDI_MEDIA_SET(ctx);
4555 		CTX_UNLOCK(ctx);
4556 		/* FALLTHROUGH */
4557 	case SIOCGIFMEDIA:
4558 	case SIOCGIFXMEDIA:
4559 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4560 		break;
4561 	case SIOCGI2C:
4562 	{
4563 		struct ifi2creq i2c;
4564 
4565 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4566 		if (err != 0)
4567 			break;
4568 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4569 			err = EINVAL;
4570 			break;
4571 		}
4572 		if (i2c.len > sizeof(i2c.data)) {
4573 			err = EINVAL;
4574 			break;
4575 		}
4576 
4577 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4578 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4579 			    sizeof(i2c));
4580 		break;
4581 	}
4582 	case SIOCSIFCAP:
4583 	{
4584 		int mask, setmask, oldmask;
4585 
4586 		oldmask = if_getcapenable(ifp);
4587 		mask = ifr->ifr_reqcap ^ oldmask;
4588 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4589 		setmask = 0;
4590 #ifdef TCP_OFFLOAD
4591 		setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6);
4592 #endif
4593 		setmask |= (mask & IFCAP_FLAGS);
4594 		setmask |= (mask & IFCAP_WOL);
4595 
4596 		/*
4597 		 * If any RX csum has changed, change all the ones that
4598 		 * are supported by the driver.
4599 		 */
4600 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4601 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4602 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4603 		}
4604 
4605 		/*
4606 		 * want to ensure that traffic has stopped before we change any of the flags
4607 		 */
4608 		if (setmask) {
4609 			CTX_LOCK(ctx);
4610 			bits = if_getdrvflags(ifp);
4611 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4612 				iflib_stop(ctx);
4613 			STATE_LOCK(ctx);
4614 			if_togglecapenable(ifp, setmask);
4615 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4616 			STATE_UNLOCK(ctx);
4617 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4618 				iflib_init_locked(ctx);
4619 			STATE_LOCK(ctx);
4620 			if_setdrvflags(ifp, bits);
4621 			STATE_UNLOCK(ctx);
4622 			CTX_UNLOCK(ctx);
4623 		}
4624 		if_vlancap(ifp);
4625 		break;
4626 	}
4627 	case SIOCGPRIVATE_0:
4628 	case SIOCSDRVSPEC:
4629 	case SIOCGDRVSPEC:
4630 		CTX_LOCK(ctx);
4631 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4632 		CTX_UNLOCK(ctx);
4633 		break;
4634 	default:
4635 		err = ether_ioctl(ifp, command, data);
4636 		break;
4637 	}
4638 	if (reinit)
4639 		iflib_if_init(ctx);
4640 	return (err);
4641 }
4642 
4643 static uint64_t
4644 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4645 {
4646 	if_ctx_t ctx = if_getsoftc(ifp);
4647 
4648 	return (IFDI_GET_COUNTER(ctx, cnt));
4649 }
4650 
4651 /*********************************************************************
4652  *
4653  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4654  *
4655  **********************************************************************/
4656 
4657 static void
4658 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4659 {
4660 	if_ctx_t ctx = if_getsoftc(ifp);
4661 
4662 	if ((void *)ctx != arg)
4663 		return;
4664 
4665 	if ((vtag == 0) || (vtag > 4095))
4666 		return;
4667 
4668 	if (iflib_in_detach(ctx))
4669 		return;
4670 
4671 	CTX_LOCK(ctx);
4672 	/* Driver may need all untagged packets to be flushed */
4673 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4674 		iflib_stop(ctx);
4675 	IFDI_VLAN_REGISTER(ctx, vtag);
4676 	/* Re-init to load the changes, if required */
4677 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4678 		iflib_init_locked(ctx);
4679 	CTX_UNLOCK(ctx);
4680 }
4681 
4682 static void
4683 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4684 {
4685 	if_ctx_t ctx = if_getsoftc(ifp);
4686 
4687 	if ((void *)ctx != arg)
4688 		return;
4689 
4690 	if ((vtag == 0) || (vtag > 4095))
4691 		return;
4692 
4693 	CTX_LOCK(ctx);
4694 	/* Driver may need all tagged packets to be flushed */
4695 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4696 		iflib_stop(ctx);
4697 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4698 	/* Re-init to load the changes, if required */
4699 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4700 		iflib_init_locked(ctx);
4701 	CTX_UNLOCK(ctx);
4702 }
4703 
4704 static void
4705 iflib_led_func(void *arg, int onoff)
4706 {
4707 	if_ctx_t ctx = arg;
4708 
4709 	CTX_LOCK(ctx);
4710 	IFDI_LED_FUNC(ctx, onoff);
4711 	CTX_UNLOCK(ctx);
4712 }
4713 
4714 /*********************************************************************
4715  *
4716  *  BUS FUNCTION DEFINITIONS
4717  *
4718  **********************************************************************/
4719 
4720 int
4721 iflib_device_probe(device_t dev)
4722 {
4723 	const pci_vendor_info_t *ent;
4724 	if_shared_ctx_t sctx;
4725 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4726 	uint16_t pci_vendor_id;
4727 
4728 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4729 		return (ENOTSUP);
4730 
4731 	pci_vendor_id = pci_get_vendor(dev);
4732 	pci_device_id = pci_get_device(dev);
4733 	pci_subvendor_id = pci_get_subvendor(dev);
4734 	pci_subdevice_id = pci_get_subdevice(dev);
4735 	pci_rev_id = pci_get_revid(dev);
4736 	if (sctx->isc_parse_devinfo != NULL)
4737 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4738 
4739 	ent = sctx->isc_vendor_info;
4740 	while (ent->pvi_vendor_id != 0) {
4741 		if (pci_vendor_id != ent->pvi_vendor_id) {
4742 			ent++;
4743 			continue;
4744 		}
4745 		if ((pci_device_id == ent->pvi_device_id) &&
4746 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4747 		     (ent->pvi_subvendor_id == 0)) &&
4748 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4749 		     (ent->pvi_subdevice_id == 0)) &&
4750 		    ((pci_rev_id == ent->pvi_rev_id) ||
4751 		     (ent->pvi_rev_id == 0))) {
4752 			device_set_desc_copy(dev, ent->pvi_name);
4753 			/* this needs to be changed to zero if the bus probing code
4754 			 * ever stops re-probing on best match because the sctx
4755 			 * may have its values over written by register calls
4756 			 * in subsequent probes
4757 			 */
4758 			return (BUS_PROBE_DEFAULT);
4759 		}
4760 		ent++;
4761 	}
4762 	return (ENXIO);
4763 }
4764 
4765 int
4766 iflib_device_probe_vendor(device_t dev)
4767 {
4768 	int probe;
4769 
4770 	probe = iflib_device_probe(dev);
4771 	if (probe == BUS_PROBE_DEFAULT)
4772 		return (BUS_PROBE_VENDOR);
4773 	else
4774 		return (probe);
4775 }
4776 
4777 static void
4778 iflib_reset_qvalues(if_ctx_t ctx)
4779 {
4780 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4781 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4782 	device_t dev = ctx->ifc_dev;
4783 	int i;
4784 
4785 	if (ctx->ifc_sysctl_ntxqs != 0)
4786 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4787 	if (ctx->ifc_sysctl_nrxqs != 0)
4788 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4789 
4790 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4791 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4792 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4793 		else
4794 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4795 	}
4796 
4797 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4798 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4799 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4800 		else
4801 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4802 	}
4803 
4804 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4805 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4806 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4807 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4808 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4809 		}
4810 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4811 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4812 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4813 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4814 		}
4815 		if (!powerof2(scctx->isc_nrxd[i])) {
4816 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4817 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4818 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4819 		}
4820 	}
4821 
4822 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4823 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4824 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4825 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4826 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4827 		}
4828 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4829 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4830 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4831 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4832 		}
4833 		if (!powerof2(scctx->isc_ntxd[i])) {
4834 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4835 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4836 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4837 		}
4838 	}
4839 }
4840 
4841 static void
4842 iflib_add_pfil(if_ctx_t ctx)
4843 {
4844 	struct pfil_head *pfil;
4845 	struct pfil_head_args pa;
4846 	iflib_rxq_t rxq;
4847 	int i;
4848 
4849 	pa.pa_version = PFIL_VERSION;
4850 	pa.pa_flags = PFIL_IN;
4851 	pa.pa_type = PFIL_TYPE_ETHERNET;
4852 	pa.pa_headname = if_name(ctx->ifc_ifp);
4853 	pfil = pfil_head_register(&pa);
4854 
4855 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4856 		rxq->pfil = pfil;
4857 	}
4858 }
4859 
4860 static void
4861 iflib_rem_pfil(if_ctx_t ctx)
4862 {
4863 	struct pfil_head *pfil;
4864 	iflib_rxq_t rxq;
4865 	int i;
4866 
4867 	rxq = ctx->ifc_rxqs;
4868 	pfil = rxq->pfil;
4869 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4870 		rxq->pfil = NULL;
4871 	}
4872 	pfil_head_unregister(pfil);
4873 }
4874 
4875 
4876 /*
4877  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4878  * cpuid and wrapping as necessary.
4879  */
4880 static unsigned int
4881 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4882 {
4883 	unsigned int first_valid;
4884 	unsigned int last_valid;
4885 
4886 	/* cpuid should always be in the valid set */
4887 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4888 
4889 	/* valid set should never be empty */
4890 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4891 
4892 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4893 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4894 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4895 	while (n > 0) {
4896 		do {
4897 			cpuid++;
4898 			if (cpuid > last_valid)
4899 				cpuid = first_valid;
4900 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4901 		n--;
4902 	}
4903 
4904 	return (cpuid);
4905 }
4906 
4907 #if defined(SMP) && defined(SCHED_ULE)
4908 extern struct cpu_group *cpu_top;              /* CPU topology */
4909 
4910 static int
4911 find_child_with_core(int cpu, struct cpu_group *grp)
4912 {
4913 	int i;
4914 
4915 	if (grp->cg_children == 0)
4916 		return (-1);
4917 
4918 	MPASS(grp->cg_child);
4919 	for (i = 0; i < grp->cg_children; i++) {
4920 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4921 			return (i);
4922 	}
4923 
4924 	return (-1);
4925 }
4926 
4927 
4928 /*
4929  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4930  * does not distinguish among multiple L2 neighbors if the given CPU has
4931  * more than one (it will always return the same result in that case).
4932  */
4933 static int
4934 find_l2_neighbor(int cpu)
4935 {
4936 	struct cpu_group *grp;
4937 	int i;
4938 
4939 	grp = cpu_top;
4940 	if (grp == NULL)
4941 		return (-1);
4942 
4943 	/*
4944 	 * Find the smallest CPU group that contains the given core.
4945 	 */
4946 	i = 0;
4947 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4948 		/*
4949 		 * If the smallest group containing the given CPU has less
4950 		 * than two members, we conclude the given CPU has no
4951 		 * L2 neighbor.
4952 		 */
4953 		if (grp->cg_child[i].cg_count <= 1)
4954 			return (-1);
4955 		grp = &grp->cg_child[i];
4956 	}
4957 
4958 	/* Must share L2. */
4959 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4960 		return (-1);
4961 
4962 	/*
4963 	 * Select the first member of the set that isn't the reference
4964 	 * CPU, which at this point is guaranteed to exist.
4965 	 */
4966 	for (i = 0; i < CPU_SETSIZE; i++) {
4967 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4968 			return (i);
4969 	}
4970 
4971 	/* Should never be reached */
4972 	return (-1);
4973 }
4974 
4975 #else
4976 static int
4977 find_l2_neighbor(int cpu)
4978 {
4979 
4980 	return (-1);
4981 }
4982 #endif
4983 
4984 /*
4985  * CPU mapping behaviors
4986  * ---------------------
4987  * 'separate txrx' refers to the separate_txrx sysctl
4988  * 'use logical' refers to the use_logical_cores sysctl
4989  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4990  *
4991  *  separate     use     INTR
4992  *    txrx     logical   CPUS   result
4993  * ---------- --------- ------ ------------------------------------------------
4994  *     -          -       X     RX and TX queues mapped to consecutive physical
4995  *                              cores with RX/TX pairs on same core and excess
4996  *                              of either following
4997  *     -          X       X     RX and TX queues mapped to consecutive cores
4998  *                              of any type with RX/TX pairs on same core and
4999  *                              excess of either following
5000  *     X          -       X     RX and TX queues mapped to consecutive physical
5001  *                              cores; all RX then all TX
5002  *     X          X       X     RX queues mapped to consecutive physical cores
5003  *                              first, then TX queues mapped to L2 neighbor of
5004  *                              the corresponding RX queue if one exists,
5005  *                              otherwise to consecutive physical cores
5006  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5007  *                              any type with RX/TX pairs on same core and excess
5008  *                              of either following
5009  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5010  *                              any type; all RX then all TX
5011  */
5012 static unsigned int
5013 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5014     bool is_tx)
5015 {
5016 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5017 	unsigned int core_index;
5018 
5019 	if (ctx->ifc_sysctl_separate_txrx) {
5020 		/*
5021 		 * When using separate CPUs for TX and RX, the assignment
5022 		 * will always be of a consecutive CPU out of the set of
5023 		 * context CPUs, except for the specific case where the
5024 		 * context CPUs are phsyical cores, the use of logical cores
5025 		 * has been enabled, the assignment is for TX, the TX qid
5026 		 * corresponds to an RX qid, and the CPU assigned to the
5027 		 * corresponding RX queue has an L2 neighbor.
5028 		 */
5029 		if (ctx->ifc_sysctl_use_logical_cores &&
5030 		    ctx->ifc_cpus_are_physical_cores &&
5031 		    is_tx && qid < scctx->isc_nrxqsets) {
5032 			int l2_neighbor;
5033 			unsigned int rx_cpuid;
5034 
5035 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5036 			l2_neighbor = find_l2_neighbor(rx_cpuid);
5037 			if (l2_neighbor != -1) {
5038 				return (l2_neighbor);
5039 			}
5040 			/*
5041 			 * ... else fall through to the normal
5042 			 * consecutive-after-RX assignment scheme.
5043 			 *
5044 			 * Note that we are assuming that all RX queue CPUs
5045 			 * have an L2 neighbor, or all do not.  If a mixed
5046 			 * scenario is possible, we will have to keep track
5047 			 * separately of how many queues prior to this one
5048 			 * were not able to be assigned to an L2 neighbor.
5049 			 */
5050 		}
5051 		if (is_tx)
5052 			core_index = scctx->isc_nrxqsets + qid;
5053 		else
5054 			core_index = qid;
5055 	} else {
5056 		core_index = qid;
5057 	}
5058 
5059 	return (cpuid_advance(ctx, base_cpuid, core_index));
5060 }
5061 
5062 static uint16_t
5063 get_ctx_core_offset(if_ctx_t ctx)
5064 {
5065 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5066 	struct cpu_offset *op;
5067 	cpuset_t assigned_cpus;
5068 	unsigned int cores_consumed;
5069 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5070 	unsigned int first_valid;
5071 	unsigned int last_valid;
5072 	unsigned int i;
5073 
5074 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5075 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5076 
5077 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5078 		/*
5079 		 * Align the user-chosen base CPU ID to the next valid CPU
5080 		 * for this device.  If the chosen base CPU ID is smaller
5081 		 * than the first valid CPU or larger than the last valid
5082 		 * CPU, we assume the user does not know what the valid
5083 		 * range is for this device and is thinking in terms of a
5084 		 * zero-based reference frame, and so we shift the given
5085 		 * value into the valid range (and wrap accordingly) so the
5086 		 * intent is translated to the proper frame of reference.
5087 		 * If the base CPU ID is within the valid first/last, but
5088 		 * does not correspond to a valid CPU, it is advanced to the
5089 		 * next valid CPU (wrapping if necessary).
5090 		 */
5091 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5092 			/* shift from zero-based to first_valid-based */
5093 			base_cpuid += first_valid;
5094 			/* wrap to range [first_valid, last_valid] */
5095 			base_cpuid = (base_cpuid - first_valid) %
5096 			    (last_valid - first_valid + 1);
5097 		}
5098 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5099 			/*
5100 			 * base_cpuid is in [first_valid, last_valid], but
5101 			 * not a member of the valid set.  In this case,
5102 			 * there will always be a member of the valid set
5103 			 * with a CPU ID that is greater than base_cpuid,
5104 			 * and we simply advance to it.
5105 			 */
5106 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5107 				base_cpuid++;
5108 		}
5109 		return (base_cpuid);
5110 	}
5111 
5112 	/*
5113 	 * Determine how many cores will be consumed by performing the CPU
5114 	 * assignments and counting how many of the assigned CPUs correspond
5115 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5116 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5117 	 * the set of context CPUs.
5118 	 *
5119 	 * Note not all assigned CPUs will be in the set of context CPUs
5120 	 * when separate CPUs are being allocated to TX and RX queues,
5121 	 * assignment to logical cores has been enabled, the set of context
5122 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5123 	 * neighbors of CPUs that RX queues have been mapped to - in this
5124 	 * case we do only want to count how many CPUs in the set of context
5125 	 * CPUs have been consumed, as that determines the next CPU in that
5126 	 * set to start allocating at for the next device for which
5127 	 * core_offset is not set.
5128 	 */
5129 	CPU_ZERO(&assigned_cpus);
5130 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5131 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5132 		    &assigned_cpus);
5133 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5134 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5135 		    &assigned_cpus);
5136 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5137 	cores_consumed = CPU_COUNT(&assigned_cpus);
5138 
5139 	mtx_lock(&cpu_offset_mtx);
5140 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5141 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5142 			base_cpuid = op->next_cpuid;
5143 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5144 			    cores_consumed);
5145 			MPASS(op->refcount < UINT_MAX);
5146 			op->refcount++;
5147 			break;
5148 		}
5149 	}
5150 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5151 		base_cpuid = first_valid;
5152 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5153 		    M_NOWAIT | M_ZERO);
5154 		if (op == NULL) {
5155 			device_printf(ctx->ifc_dev,
5156 			    "allocation for cpu offset failed.\n");
5157 		} else {
5158 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5159 			    cores_consumed);
5160 			op->refcount = 1;
5161 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5162 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5163 		}
5164 	}
5165 	mtx_unlock(&cpu_offset_mtx);
5166 
5167 	return (base_cpuid);
5168 }
5169 
5170 static void
5171 unref_ctx_core_offset(if_ctx_t ctx)
5172 {
5173 	struct cpu_offset *op, *top;
5174 
5175 	mtx_lock(&cpu_offset_mtx);
5176 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5177 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5178 			MPASS(op->refcount > 0);
5179 			op->refcount--;
5180 			if (op->refcount == 0) {
5181 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5182 				free(op, M_IFLIB);
5183 			}
5184 			break;
5185 		}
5186 	}
5187 	mtx_unlock(&cpu_offset_mtx);
5188 }
5189 
5190 int
5191 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5192 {
5193 	if_ctx_t ctx;
5194 	if_t ifp;
5195 	if_softc_ctx_t scctx;
5196 	kobjop_desc_t kobj_desc;
5197 	kobj_method_t *kobj_method;
5198 	int err, msix, rid;
5199 	int num_txd, num_rxd;
5200 
5201 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO);
5202 
5203 	if (sc == NULL) {
5204 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO);
5205 		device_set_softc(dev, ctx);
5206 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5207 	}
5208 
5209 	ctx->ifc_sctx = sctx;
5210 	ctx->ifc_dev = dev;
5211 	ctx->ifc_softc = sc;
5212 
5213 	if ((err = iflib_register(ctx)) != 0) {
5214 		device_printf(dev, "iflib_register failed %d\n", err);
5215 		goto fail_ctx_free;
5216 	}
5217 	iflib_add_device_sysctl_pre(ctx);
5218 
5219 	scctx = &ctx->ifc_softc_ctx;
5220 	ifp = ctx->ifc_ifp;
5221 
5222 	iflib_reset_qvalues(ctx);
5223 	IFNET_WLOCK();
5224 	CTX_LOCK(ctx);
5225 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5226 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5227 		goto fail_unlock;
5228 	}
5229 	_iflib_pre_assert(scctx);
5230 	ctx->ifc_txrx = *scctx->isc_txrx;
5231 
5232 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5233 
5234 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5235 		ctx->ifc_mediap = scctx->isc_media;
5236 
5237 #ifdef INVARIANTS
5238 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5239 		MPASS(scctx->isc_tx_csum_flags);
5240 #endif
5241 
5242 	if_setcapabilities(ifp,
5243 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5244 	if_setcapenable(ifp,
5245 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5246 
5247 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5248 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5249 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5250 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5251 
5252 	num_txd = iflib_num_tx_descs(ctx);
5253 	num_rxd = iflib_num_rx_descs(ctx);
5254 
5255 	/* XXX change for per-queue sizes */
5256 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5257 	    num_txd, num_rxd);
5258 
5259 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5260 		scctx->isc_tx_nsegments = max(1, num_txd /
5261 		    MAX_SINGLE_PACKET_FRACTION);
5262 	if (scctx->isc_tx_tso_segments_max > num_txd /
5263 	    MAX_SINGLE_PACKET_FRACTION)
5264 		scctx->isc_tx_tso_segments_max = max(1,
5265 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5266 
5267 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5268 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5269 		/*
5270 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5271 		 * but some MACs do.
5272 		 */
5273 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5274 		    IP_MAXPACKET));
5275 		/*
5276 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5277 		 * into account.  In the worst case, each of these calls will
5278 		 * add another mbuf and, thus, the requirement for another DMA
5279 		 * segment.  So for best performance, it doesn't make sense to
5280 		 * advertize a maximum of TSO segments that typically will
5281 		 * require defragmentation in iflib_encap().
5282 		 */
5283 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5284 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5285 	}
5286 	if (scctx->isc_rss_table_size == 0)
5287 		scctx->isc_rss_table_size = 64;
5288 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1;
5289 
5290 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5291 	/* XXX format name */
5292 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5293 	    NULL, NULL, "admin");
5294 
5295 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5296 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5297 		device_printf(dev, "Unable to fetch CPU list\n");
5298 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5299 		ctx->ifc_cpus_are_physical_cores = false;
5300 	} else
5301 		ctx->ifc_cpus_are_physical_cores = true;
5302 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5303 
5304 	/*
5305 	 * Now set up MSI or MSI-X, should return us the number of supported
5306 	 * vectors (will be 1 for a legacy interrupt and MSI).
5307 	 */
5308 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5309 		msix = scctx->isc_vectors;
5310 	} else if (scctx->isc_msix_bar != 0)
5311 		/*
5312 		 * The simple fact that isc_msix_bar is not 0 does not mean we
5313 		 * we have a good value there that is known to work.
5314 		 */
5315 		msix = iflib_msix_init(ctx);
5316 	else {
5317 		scctx->isc_vectors = 1;
5318 		scctx->isc_ntxqsets = 1;
5319 		scctx->isc_nrxqsets = 1;
5320 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5321 		msix = 0;
5322 	}
5323 	/* Get memory for the station queues */
5324 	if ((err = iflib_queues_alloc(ctx))) {
5325 		device_printf(dev, "Unable to allocate queue memory\n");
5326 		goto fail_intr_free;
5327 	}
5328 
5329 	if ((err = iflib_qset_structures_setup(ctx)))
5330 		goto fail_queues;
5331 
5332 	/*
5333 	 * Now that we know how many queues there are, get the core offset.
5334 	 */
5335 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5336 
5337 	if (msix > 1) {
5338 		/*
5339 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5340 		 * aren't the default NULL implementation.
5341 		 */
5342 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5343 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5344 		    kobj_desc);
5345 		if (kobj_method == &kobj_desc->deflt) {
5346 			device_printf(dev,
5347 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5348 			err = EOPNOTSUPP;
5349 			goto fail_queues;
5350 		}
5351 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5352 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5353 		    kobj_desc);
5354 		if (kobj_method == &kobj_desc->deflt) {
5355 			device_printf(dev,
5356 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5357 			err = EOPNOTSUPP;
5358 			goto fail_queues;
5359 		}
5360 
5361 		/*
5362 		 * Assign the MSI-X vectors.
5363 		 * Note that the default NULL ifdi_msix_intr_assign method will
5364 		 * fail here, too.
5365 		 */
5366 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5367 		if (err != 0) {
5368 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5369 			    err);
5370 			goto fail_queues;
5371 		}
5372 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5373 		rid = 0;
5374 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5375 			MPASS(msix == 1);
5376 			rid = 1;
5377 		}
5378 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5379 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5380 			goto fail_queues;
5381 		}
5382 	} else {
5383 		device_printf(dev,
5384 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5385 		err = ENODEV;
5386 		goto fail_queues;
5387 	}
5388 
5389 	/*
5390 	 * It prevents a double-locking panic with iflib_media_status when
5391 	 * the driver loads.
5392 	 */
5393 	CTX_UNLOCK(ctx);
5394 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5395 	CTX_LOCK(ctx);
5396 
5397 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5398 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5399 		goto fail_detach;
5400 	}
5401 
5402 	/*
5403 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5404 	 * This must appear after the call to ether_ifattach() because
5405 	 * ether_ifattach() sets if_hdrlen to the default value.
5406 	 */
5407 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5408 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5409 
5410 	if ((err = iflib_netmap_attach(ctx))) {
5411 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5412 		goto fail_detach;
5413 	}
5414 	*ctxp = ctx;
5415 
5416 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5417 
5418 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5419 	iflib_add_device_sysctl_post(ctx);
5420 	iflib_add_pfil(ctx);
5421 	ctx->ifc_flags |= IFC_INIT_DONE;
5422 	CTX_UNLOCK(ctx);
5423 	IFNET_WUNLOCK();
5424 
5425 	return (0);
5426 
5427 fail_detach:
5428 	ether_ifdetach(ctx->ifc_ifp);
5429 fail_queues:
5430 	iflib_tqg_detach(ctx);
5431 	iflib_tx_structures_free(ctx);
5432 	iflib_rx_structures_free(ctx);
5433 	IFDI_DETACH(ctx);
5434 	IFDI_QUEUES_FREE(ctx);
5435 fail_intr_free:
5436 	iflib_free_intr_mem(ctx);
5437 fail_unlock:
5438 	CTX_UNLOCK(ctx);
5439 	IFNET_WUNLOCK();
5440 	iflib_deregister(ctx);
5441 fail_ctx_free:
5442 	device_set_softc(ctx->ifc_dev, NULL);
5443         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5444                 free(ctx->ifc_softc, M_IFLIB);
5445         free(ctx, M_IFLIB);
5446 	return (err);
5447 }
5448 
5449 int
5450 iflib_device_attach(device_t dev)
5451 {
5452 	if_ctx_t ctx;
5453 	if_shared_ctx_t sctx;
5454 
5455 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5456 		return (ENOTSUP);
5457 
5458 	pci_enable_busmaster(dev);
5459 
5460 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5461 }
5462 
5463 int
5464 iflib_device_deregister(if_ctx_t ctx)
5465 {
5466 	if_t ifp = ctx->ifc_ifp;
5467 	device_t dev = ctx->ifc_dev;
5468 
5469 	/* Make sure VLANS are not using driver */
5470 	if (if_vlantrunkinuse(ifp)) {
5471 		device_printf(dev, "Vlan in use, detach first\n");
5472 		return (EBUSY);
5473 	}
5474 #ifdef PCI_IOV
5475 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5476 		device_printf(dev, "SR-IOV in use; detach first.\n");
5477 		return (EBUSY);
5478 	}
5479 #endif
5480 
5481 	STATE_LOCK(ctx);
5482 	ctx->ifc_flags |= IFC_IN_DETACH;
5483 	STATE_UNLOCK(ctx);
5484 
5485 	/* Unregister VLAN handlers before calling iflib_stop() */
5486 	iflib_unregister_vlan_handlers(ctx);
5487 
5488 	iflib_netmap_detach(ifp);
5489 	ether_ifdetach(ifp);
5490 
5491 	CTX_LOCK(ctx);
5492 	iflib_stop(ctx);
5493 	CTX_UNLOCK(ctx);
5494 
5495 	iflib_rem_pfil(ctx);
5496 	if (ctx->ifc_led_dev != NULL)
5497 		led_destroy(ctx->ifc_led_dev);
5498 
5499 	iflib_tqg_detach(ctx);
5500 	iflib_tx_structures_free(ctx);
5501 	iflib_rx_structures_free(ctx);
5502 
5503 	CTX_LOCK(ctx);
5504 	IFDI_DETACH(ctx);
5505 	IFDI_QUEUES_FREE(ctx);
5506 	CTX_UNLOCK(ctx);
5507 
5508 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5509 	iflib_free_intr_mem(ctx);
5510 
5511 	bus_generic_detach(dev);
5512 
5513 	iflib_deregister(ctx);
5514 
5515 	device_set_softc(ctx->ifc_dev, NULL);
5516 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5517 		free(ctx->ifc_softc, M_IFLIB);
5518 	unref_ctx_core_offset(ctx);
5519 	free(ctx, M_IFLIB);
5520 	return (0);
5521 }
5522 
5523 static void
5524 iflib_tqg_detach(if_ctx_t ctx)
5525 {
5526 	iflib_txq_t txq;
5527 	iflib_rxq_t rxq;
5528 	int i;
5529 	struct taskqgroup *tqg;
5530 
5531 	/* XXX drain any dependent tasks */
5532 	tqg = qgroup_if_io_tqg;
5533 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5534 		callout_drain(&txq->ift_timer);
5535 #ifdef DEV_NETMAP
5536 		callout_drain(&txq->ift_netmap_timer);
5537 #endif /* DEV_NETMAP */
5538 		if (txq->ift_task.gt_uniq != NULL)
5539 			taskqgroup_detach(tqg, &txq->ift_task);
5540 	}
5541 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5542 		if (rxq->ifr_task.gt_uniq != NULL)
5543 			taskqgroup_detach(tqg, &rxq->ifr_task);
5544 	}
5545 	tqg = qgroup_if_config_tqg;
5546 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5547 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5548 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5549 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5550 }
5551 
5552 static void
5553 iflib_free_intr_mem(if_ctx_t ctx)
5554 {
5555 
5556 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5557 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5558 	}
5559 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5560 		pci_release_msi(ctx->ifc_dev);
5561 	}
5562 	if (ctx->ifc_msix_mem != NULL) {
5563 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5564 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5565 		ctx->ifc_msix_mem = NULL;
5566 	}
5567 }
5568 
5569 int
5570 iflib_device_detach(device_t dev)
5571 {
5572 	if_ctx_t ctx = device_get_softc(dev);
5573 
5574 	return (iflib_device_deregister(ctx));
5575 }
5576 
5577 int
5578 iflib_device_suspend(device_t dev)
5579 {
5580 	if_ctx_t ctx = device_get_softc(dev);
5581 
5582 	CTX_LOCK(ctx);
5583 	IFDI_SUSPEND(ctx);
5584 	CTX_UNLOCK(ctx);
5585 
5586 	return (bus_generic_suspend(dev));
5587 }
5588 int
5589 iflib_device_shutdown(device_t dev)
5590 {
5591 	if_ctx_t ctx = device_get_softc(dev);
5592 
5593 	CTX_LOCK(ctx);
5594 	IFDI_SHUTDOWN(ctx);
5595 	CTX_UNLOCK(ctx);
5596 
5597 	return (bus_generic_suspend(dev));
5598 }
5599 
5600 int
5601 iflib_device_resume(device_t dev)
5602 {
5603 	if_ctx_t ctx = device_get_softc(dev);
5604 	iflib_txq_t txq = ctx->ifc_txqs;
5605 
5606 	CTX_LOCK(ctx);
5607 	IFDI_RESUME(ctx);
5608 	iflib_if_init_locked(ctx);
5609 	CTX_UNLOCK(ctx);
5610 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5611 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5612 
5613 	return (bus_generic_resume(dev));
5614 }
5615 
5616 int
5617 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5618 {
5619 	int error;
5620 	if_ctx_t ctx = device_get_softc(dev);
5621 
5622 	CTX_LOCK(ctx);
5623 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5624 	CTX_UNLOCK(ctx);
5625 
5626 	return (error);
5627 }
5628 
5629 void
5630 iflib_device_iov_uninit(device_t dev)
5631 {
5632 	if_ctx_t ctx = device_get_softc(dev);
5633 
5634 	CTX_LOCK(ctx);
5635 	IFDI_IOV_UNINIT(ctx);
5636 	CTX_UNLOCK(ctx);
5637 }
5638 
5639 int
5640 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5641 {
5642 	int error;
5643 	if_ctx_t ctx = device_get_softc(dev);
5644 
5645 	CTX_LOCK(ctx);
5646 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5647 	CTX_UNLOCK(ctx);
5648 
5649 	return (error);
5650 }
5651 
5652 /*********************************************************************
5653  *
5654  *  MODULE FUNCTION DEFINITIONS
5655  *
5656  **********************************************************************/
5657 
5658 /*
5659  * - Start a fast taskqueue thread for each core
5660  * - Start a taskqueue for control operations
5661  */
5662 static int
5663 iflib_module_init(void)
5664 {
5665 	iflib_timer_default = hz / 2;
5666 	return (0);
5667 }
5668 
5669 static int
5670 iflib_module_event_handler(module_t mod, int what, void *arg)
5671 {
5672 	int err;
5673 
5674 	switch (what) {
5675 	case MOD_LOAD:
5676 		if ((err = iflib_module_init()) != 0)
5677 			return (err);
5678 		break;
5679 	case MOD_UNLOAD:
5680 		return (EBUSY);
5681 	default:
5682 		return (EOPNOTSUPP);
5683 	}
5684 
5685 	return (0);
5686 }
5687 
5688 /*********************************************************************
5689  *
5690  *  PUBLIC FUNCTION DEFINITIONS
5691  *     ordered as in iflib.h
5692  *
5693  **********************************************************************/
5694 
5695 static void
5696 _iflib_assert(if_shared_ctx_t sctx)
5697 {
5698 	int i;
5699 
5700 	MPASS(sctx->isc_tx_maxsize);
5701 	MPASS(sctx->isc_tx_maxsegsize);
5702 
5703 	MPASS(sctx->isc_rx_maxsize);
5704 	MPASS(sctx->isc_rx_nsegments);
5705 	MPASS(sctx->isc_rx_maxsegsize);
5706 
5707 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5708 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5709 		MPASS(sctx->isc_nrxd_min[i]);
5710 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5711 		MPASS(sctx->isc_nrxd_max[i]);
5712 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5713 		MPASS(sctx->isc_nrxd_default[i]);
5714 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5715 	}
5716 
5717 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5718 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5719 		MPASS(sctx->isc_ntxd_min[i]);
5720 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5721 		MPASS(sctx->isc_ntxd_max[i]);
5722 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5723 		MPASS(sctx->isc_ntxd_default[i]);
5724 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5725 	}
5726 }
5727 
5728 static void
5729 _iflib_pre_assert(if_softc_ctx_t scctx)
5730 {
5731 
5732 	MPASS(scctx->isc_txrx->ift_txd_encap);
5733 	MPASS(scctx->isc_txrx->ift_txd_flush);
5734 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5735 	MPASS(scctx->isc_txrx->ift_rxd_available);
5736 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5737 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5738 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5739 }
5740 
5741 static int
5742 iflib_register(if_ctx_t ctx)
5743 {
5744 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5745 	driver_t *driver = sctx->isc_driver;
5746 	device_t dev = ctx->ifc_dev;
5747 	if_t ifp;
5748 
5749 	_iflib_assert(sctx);
5750 
5751 	CTX_LOCK_INIT(ctx);
5752 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5753 	ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev);
5754 
5755 	/*
5756 	 * Initialize our context's device specific methods
5757 	 */
5758 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5759 	kobj_class_compile((kobj_class_t) driver);
5760 
5761 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5762 	if_setsoftc(ifp, ctx);
5763 	if_setdev(ifp, dev);
5764 	if_setinitfn(ifp, iflib_if_init);
5765 	if_setioctlfn(ifp, iflib_if_ioctl);
5766 #ifdef ALTQ
5767 	if_setstartfn(ifp, iflib_altq_if_start);
5768 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5769 	if_setsendqready(ifp);
5770 #else
5771 	if_settransmitfn(ifp, iflib_if_transmit);
5772 #endif
5773 	if_setqflushfn(ifp, iflib_if_qflush);
5774 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5775 	ctx->ifc_vlan_attach_event =
5776 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5777 							  EVENTHANDLER_PRI_FIRST);
5778 	ctx->ifc_vlan_detach_event =
5779 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5780 							  EVENTHANDLER_PRI_FIRST);
5781 
5782 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5783 		ctx->ifc_mediap = &ctx->ifc_media;
5784 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5785 		    iflib_media_change, iflib_media_status);
5786 	}
5787 	return (0);
5788 }
5789 
5790 static void
5791 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5792 {
5793 	/* Unregister VLAN events */
5794 	if (ctx->ifc_vlan_attach_event != NULL) {
5795 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5796 		ctx->ifc_vlan_attach_event = NULL;
5797 	}
5798 	if (ctx->ifc_vlan_detach_event != NULL) {
5799 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5800 		ctx->ifc_vlan_detach_event = NULL;
5801 	}
5802 
5803 }
5804 
5805 static void
5806 iflib_deregister(if_ctx_t ctx)
5807 {
5808 	if_t ifp = ctx->ifc_ifp;
5809 
5810 	/* Remove all media */
5811 	ifmedia_removeall(&ctx->ifc_media);
5812 
5813 	/* Ensure that VLAN event handlers are unregistered */
5814 	iflib_unregister_vlan_handlers(ctx);
5815 
5816 	/* Release kobject reference */
5817 	kobj_delete((kobj_t) ctx, NULL);
5818 
5819 	/* Free the ifnet structure */
5820 	if_free(ifp);
5821 
5822 	STATE_LOCK_DESTROY(ctx);
5823 
5824 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5825 	CTX_LOCK_DESTROY(ctx);
5826 }
5827 
5828 static int
5829 iflib_queues_alloc(if_ctx_t ctx)
5830 {
5831 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5832 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5833 	device_t dev = ctx->ifc_dev;
5834 	int nrxqsets = scctx->isc_nrxqsets;
5835 	int ntxqsets = scctx->isc_ntxqsets;
5836 	iflib_txq_t txq;
5837 	iflib_rxq_t rxq;
5838 	iflib_fl_t fl = NULL;
5839 	int i, j, cpu, err, txconf, rxconf;
5840 	iflib_dma_info_t ifdip;
5841 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5842 	uint32_t *txqsizes = scctx->isc_txqsizes;
5843 	uint8_t nrxqs = sctx->isc_nrxqs;
5844 	uint8_t ntxqs = sctx->isc_ntxqs;
5845 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5846 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5847 	caddr_t *vaddrs;
5848 	uint64_t *paddrs;
5849 
5850 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5851 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5852 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5853            ("there must be at least a rxq for each free list"));
5854 
5855 	/* Allocate the TX ring struct memory */
5856 	if (!(ctx->ifc_txqs =
5857 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5858 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5859 		device_printf(dev, "Unable to allocate TX ring memory\n");
5860 		err = ENOMEM;
5861 		goto fail;
5862 	}
5863 
5864 	/* Now allocate the RX */
5865 	if (!(ctx->ifc_rxqs =
5866 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5867 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5868 		device_printf(dev, "Unable to allocate RX ring memory\n");
5869 		err = ENOMEM;
5870 		goto rx_fail;
5871 	}
5872 
5873 	txq = ctx->ifc_txqs;
5874 	rxq = ctx->ifc_rxqs;
5875 
5876 	/*
5877 	 * XXX handle allocation failure
5878 	 */
5879 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5880 		/* Set up some basics */
5881 
5882 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5883 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5884 			device_printf(dev,
5885 			    "Unable to allocate TX DMA info memory\n");
5886 			err = ENOMEM;
5887 			goto err_tx_desc;
5888 		}
5889 		txq->ift_ifdi = ifdip;
5890 		for (j = 0; j < ntxqs; j++, ifdip++) {
5891 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5892 				device_printf(dev,
5893 				    "Unable to allocate TX descriptors\n");
5894 				err = ENOMEM;
5895 				goto err_tx_desc;
5896 			}
5897 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5898 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5899 		}
5900 		txq->ift_ctx = ctx;
5901 		txq->ift_id = i;
5902 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5903 			txq->ift_br_offset = 1;
5904 		} else {
5905 			txq->ift_br_offset = 0;
5906 		}
5907 
5908 		if (iflib_txsd_alloc(txq)) {
5909 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5910 			err = ENOMEM;
5911 			goto err_tx_desc;
5912 		}
5913 
5914 		/* Initialize the TX lock */
5915 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5916 		    device_get_nameunit(dev), txq->ift_id);
5917 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5918 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5919 		txq->ift_timer.c_cpu = cpu;
5920 #ifdef DEV_NETMAP
5921 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5922 		txq->ift_netmap_timer.c_cpu = cpu;
5923 #endif /* DEV_NETMAP */
5924 
5925 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5926 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5927 		if (err) {
5928 			/* XXX free any allocated rings */
5929 			device_printf(dev, "Unable to allocate buf_ring\n");
5930 			goto err_tx_desc;
5931 		}
5932 	}
5933 
5934 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5935 		/* Set up some basics */
5936 		callout_init(&rxq->ifr_watchdog, 1);
5937 
5938 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5939 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5940 			device_printf(dev,
5941 			    "Unable to allocate RX DMA info memory\n");
5942 			err = ENOMEM;
5943 			goto err_tx_desc;
5944 		}
5945 
5946 		rxq->ifr_ifdi = ifdip;
5947 		/* XXX this needs to be changed if #rx queues != #tx queues */
5948 		rxq->ifr_ntxqirq = 1;
5949 		rxq->ifr_txqid[0] = i;
5950 		for (j = 0; j < nrxqs; j++, ifdip++) {
5951 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5952 				device_printf(dev,
5953 				    "Unable to allocate RX descriptors\n");
5954 				err = ENOMEM;
5955 				goto err_tx_desc;
5956 			}
5957 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5958 		}
5959 		rxq->ifr_ctx = ctx;
5960 		rxq->ifr_id = i;
5961 		rxq->ifr_fl_offset = fl_offset;
5962 		rxq->ifr_nfl = nfree_lists;
5963 		if (!(fl =
5964 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5965 			device_printf(dev, "Unable to allocate free list memory\n");
5966 			err = ENOMEM;
5967 			goto err_tx_desc;
5968 		}
5969 		rxq->ifr_fl = fl;
5970 		for (j = 0; j < nfree_lists; j++) {
5971 			fl[j].ifl_rxq = rxq;
5972 			fl[j].ifl_id = j;
5973 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5974 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5975 		}
5976 		/* Allocate receive buffers for the ring */
5977 		if (iflib_rxsd_alloc(rxq)) {
5978 			device_printf(dev,
5979 			    "Critical Failure setting up receive buffers\n");
5980 			err = ENOMEM;
5981 			goto err_rx_desc;
5982 		}
5983 
5984 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5985 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5986 			    M_WAITOK);
5987 	}
5988 
5989 	/* TXQs */
5990 	vaddrs = malloc(sizeof(caddr_t)  * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
5991 	paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
5992 	for (i = 0; i < ntxqsets; i++) {
5993 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5994 
5995 		for (j = 0; j < ntxqs; j++, di++) {
5996 			vaddrs[i * ntxqs + j] = di->idi_vaddr;
5997 			paddrs[i * ntxqs + j] = di->idi_paddr;
5998 		}
5999 	}
6000 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6001 		device_printf(ctx->ifc_dev,
6002 		    "Unable to allocate device TX queue\n");
6003 		iflib_tx_structures_free(ctx);
6004 		free(vaddrs, M_IFLIB);
6005 		free(paddrs, M_IFLIB);
6006 		goto err_rx_desc;
6007 	}
6008 	free(vaddrs, M_IFLIB);
6009 	free(paddrs, M_IFLIB);
6010 
6011 	/* RXQs */
6012 	vaddrs = malloc(sizeof(caddr_t)  * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6013 	paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6014 	for (i = 0; i < nrxqsets; i++) {
6015 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6016 
6017 		for (j = 0; j < nrxqs; j++, di++) {
6018 			vaddrs[i * nrxqs + j] = di->idi_vaddr;
6019 			paddrs[i * nrxqs + j] = di->idi_paddr;
6020 		}
6021 	}
6022 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6023 		device_printf(ctx->ifc_dev,
6024 		    "Unable to allocate device RX queue\n");
6025 		iflib_tx_structures_free(ctx);
6026 		free(vaddrs, M_IFLIB);
6027 		free(paddrs, M_IFLIB);
6028 		goto err_rx_desc;
6029 	}
6030 	free(vaddrs, M_IFLIB);
6031 	free(paddrs, M_IFLIB);
6032 
6033 	return (0);
6034 
6035 /* XXX handle allocation failure changes */
6036 err_rx_desc:
6037 err_tx_desc:
6038 rx_fail:
6039 	if (ctx->ifc_rxqs != NULL)
6040 		free(ctx->ifc_rxqs, M_IFLIB);
6041 	ctx->ifc_rxqs = NULL;
6042 	if (ctx->ifc_txqs != NULL)
6043 		free(ctx->ifc_txqs, M_IFLIB);
6044 	ctx->ifc_txqs = NULL;
6045 fail:
6046 	return (err);
6047 }
6048 
6049 static int
6050 iflib_tx_structures_setup(if_ctx_t ctx)
6051 {
6052 	iflib_txq_t txq = ctx->ifc_txqs;
6053 	int i;
6054 
6055 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6056 		iflib_txq_setup(txq);
6057 
6058 	return (0);
6059 }
6060 
6061 static void
6062 iflib_tx_structures_free(if_ctx_t ctx)
6063 {
6064 	iflib_txq_t txq = ctx->ifc_txqs;
6065 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6066 	int i, j;
6067 
6068 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6069 		for (j = 0; j < sctx->isc_ntxqs; j++)
6070 			iflib_dma_free(&txq->ift_ifdi[j]);
6071 		iflib_txq_destroy(txq);
6072 	}
6073 	free(ctx->ifc_txqs, M_IFLIB);
6074 	ctx->ifc_txqs = NULL;
6075 }
6076 
6077 /*********************************************************************
6078  *
6079  *  Initialize all receive rings.
6080  *
6081  **********************************************************************/
6082 static int
6083 iflib_rx_structures_setup(if_ctx_t ctx)
6084 {
6085 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6086 	int q;
6087 #if defined(INET6) || defined(INET)
6088 	int err, i;
6089 #endif
6090 
6091 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6092 #if defined(INET6) || defined(INET)
6093 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6094 		    TCP_LRO_ENTRIES, min(1024,
6095 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6096 		if (err != 0) {
6097 			device_printf(ctx->ifc_dev,
6098 			    "LRO Initialization failed!\n");
6099 			goto fail;
6100 		}
6101 #endif
6102 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6103 	}
6104 	return (0);
6105 #if defined(INET6) || defined(INET)
6106 fail:
6107 	/*
6108 	 * Free LRO resources allocated so far, we will only handle
6109 	 * the rings that completed, the failing case will have
6110 	 * cleaned up for itself.  'q' failed, so its the terminus.
6111 	 */
6112 	rxq = ctx->ifc_rxqs;
6113 	for (i = 0; i < q; ++i, rxq++) {
6114 		tcp_lro_free(&rxq->ifr_lc);
6115 	}
6116 	return (err);
6117 #endif
6118 }
6119 
6120 /*********************************************************************
6121  *
6122  *  Free all receive rings.
6123  *
6124  **********************************************************************/
6125 static void
6126 iflib_rx_structures_free(if_ctx_t ctx)
6127 {
6128 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6129 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6130 	int i, j;
6131 
6132 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6133 		for (j = 0; j < sctx->isc_nrxqs; j++)
6134 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6135 		iflib_rx_sds_free(rxq);
6136 #if defined(INET6) || defined(INET)
6137 		tcp_lro_free(&rxq->ifr_lc);
6138 #endif
6139 	}
6140 	free(ctx->ifc_rxqs, M_IFLIB);
6141 	ctx->ifc_rxqs = NULL;
6142 }
6143 
6144 static int
6145 iflib_qset_structures_setup(if_ctx_t ctx)
6146 {
6147 	int err;
6148 
6149 	/*
6150 	 * It is expected that the caller takes care of freeing queues if this
6151 	 * fails.
6152 	 */
6153 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6154 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6155 		return (err);
6156 	}
6157 
6158 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6159 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6160 
6161 	return (err);
6162 }
6163 
6164 int
6165 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6166 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6167 {
6168 
6169 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6170 }
6171 
6172 /* Just to avoid copy/paste */
6173 static inline int
6174 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6175     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6176     const char *name)
6177 {
6178 	device_t dev;
6179 	unsigned int base_cpuid, cpuid;
6180 	int err;
6181 
6182 	dev = ctx->ifc_dev;
6183 	base_cpuid = ctx->ifc_sysctl_core_offset;
6184 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6185 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6186 	    irq ? irq->ii_res : NULL, name);
6187 	if (err) {
6188 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6189 		return (err);
6190 	}
6191 #ifdef notyet
6192 	if (cpuid > ctx->ifc_cpuid_highest)
6193 		ctx->ifc_cpuid_highest = cpuid;
6194 #endif
6195 	return (0);
6196 }
6197 
6198 /*
6199  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6200  * resources.
6201  *
6202  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6203  * only.
6204  *
6205  * XXX: Could be removed if subctx's dev has its intr resource allocation
6206  * methods replaced with custom ones?
6207  */
6208 int
6209 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6210 			       int rid, iflib_intr_type_t type,
6211 			       driver_filter_t *filter, void *filter_arg,
6212 			       int qid, const char *name)
6213 {
6214 	device_t dev, subdev;
6215 	struct grouptask *gtask;
6216 	struct taskqgroup *tqg;
6217 	iflib_filter_info_t info;
6218 	gtask_fn_t *fn;
6219 	int tqrid, err;
6220 	driver_filter_t *intr_fast;
6221 	void *q;
6222 
6223 	MPASS(ctx != NULL);
6224 	MPASS(subctx != NULL);
6225 
6226 	tqrid = rid;
6227 	dev = ctx->ifc_dev;
6228 	subdev = subctx->ifc_dev;
6229 
6230 	switch (type) {
6231 	case IFLIB_INTR_RXTX:
6232 		q = &subctx->ifc_rxqs[qid];
6233 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6234 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6235 		tqg = qgroup_if_io_tqg;
6236 		fn = _task_fn_rx;
6237 		intr_fast = iflib_fast_intr_rxtx;
6238 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6239 		break;
6240 	default:
6241 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6242 		    __func__, device_get_nameunit(subdev), type);
6243 		return (EINVAL);
6244 	}
6245 
6246 	info->ifi_filter = filter;
6247 	info->ifi_filter_arg = filter_arg;
6248 	info->ifi_task = gtask;
6249 	info->ifi_ctx = q;
6250 
6251 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6252 
6253 	/* Allocate interrupts from hardware using parent context */
6254 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6255 	if (err != 0) {
6256 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6257 		    device_get_nameunit(subdev), err);
6258 		return (err);
6259 	}
6260 
6261 	if (tqrid != -1) {
6262 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6263 		    name);
6264 		if (err)
6265 			return (err);
6266 	} else {
6267 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6268 	}
6269 
6270 	return (0);
6271 }
6272 
6273 int
6274 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6275 			iflib_intr_type_t type, driver_filter_t *filter,
6276 			void *filter_arg, int qid, const char *name)
6277 {
6278 	device_t dev;
6279 	struct grouptask *gtask;
6280 	struct taskqgroup *tqg;
6281 	iflib_filter_info_t info;
6282 	gtask_fn_t *fn;
6283 	int tqrid, err;
6284 	driver_filter_t *intr_fast;
6285 	void *q;
6286 
6287 	info = &ctx->ifc_filter_info;
6288 	tqrid = rid;
6289 
6290 	switch (type) {
6291 	/* XXX merge tx/rx for netmap? */
6292 	case IFLIB_INTR_TX:
6293 		q = &ctx->ifc_txqs[qid];
6294 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6295 		gtask = &ctx->ifc_txqs[qid].ift_task;
6296 		tqg = qgroup_if_io_tqg;
6297 		fn = _task_fn_tx;
6298 		intr_fast = iflib_fast_intr;
6299 		GROUPTASK_INIT(gtask, 0, fn, q);
6300 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6301 		break;
6302 	case IFLIB_INTR_RX:
6303 		q = &ctx->ifc_rxqs[qid];
6304 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6305 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6306 		tqg = qgroup_if_io_tqg;
6307 		fn = _task_fn_rx;
6308 		intr_fast = iflib_fast_intr;
6309 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6310 		break;
6311 	case IFLIB_INTR_RXTX:
6312 		q = &ctx->ifc_rxqs[qid];
6313 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6314 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6315 		tqg = qgroup_if_io_tqg;
6316 		fn = _task_fn_rx;
6317 		intr_fast = iflib_fast_intr_rxtx;
6318 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6319 		break;
6320 	case IFLIB_INTR_ADMIN:
6321 		q = ctx;
6322 		tqrid = -1;
6323 		info = &ctx->ifc_filter_info;
6324 		gtask = &ctx->ifc_admin_task;
6325 		tqg = qgroup_if_config_tqg;
6326 		fn = _task_fn_admin;
6327 		intr_fast = iflib_fast_intr_ctx;
6328 		break;
6329 	default:
6330 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6331 		    __func__);
6332 		return (EINVAL);
6333 	}
6334 
6335 	info->ifi_filter = filter;
6336 	info->ifi_filter_arg = filter_arg;
6337 	info->ifi_task = gtask;
6338 	info->ifi_ctx = q;
6339 
6340 	dev = ctx->ifc_dev;
6341 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6342 	if (err != 0) {
6343 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6344 		return (err);
6345 	}
6346 	if (type == IFLIB_INTR_ADMIN)
6347 		return (0);
6348 
6349 	if (tqrid != -1) {
6350 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6351 		    name);
6352 		if (err)
6353 			return (err);
6354 	} else {
6355 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6356 	}
6357 
6358 	return (0);
6359 }
6360 
6361 void
6362 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6363 			    void *arg, int qid, const char *name)
6364 {
6365 	device_t dev;
6366 	struct grouptask *gtask;
6367 	struct taskqgroup *tqg;
6368 	gtask_fn_t *fn;
6369 	void *q;
6370 	int err;
6371 
6372 	switch (type) {
6373 	case IFLIB_INTR_TX:
6374 		q = &ctx->ifc_txqs[qid];
6375 		gtask = &ctx->ifc_txqs[qid].ift_task;
6376 		tqg = qgroup_if_io_tqg;
6377 		fn = _task_fn_tx;
6378 		GROUPTASK_INIT(gtask, 0, fn, q);
6379 		break;
6380 	case IFLIB_INTR_RX:
6381 		q = &ctx->ifc_rxqs[qid];
6382 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6383 		tqg = qgroup_if_io_tqg;
6384 		fn = _task_fn_rx;
6385 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6386 		break;
6387 	case IFLIB_INTR_IOV:
6388 		q = ctx;
6389 		gtask = &ctx->ifc_vflr_task;
6390 		tqg = qgroup_if_config_tqg;
6391 		fn = _task_fn_iov;
6392 		GROUPTASK_INIT(gtask, 0, fn, q);
6393 		break;
6394 	default:
6395 		panic("unknown net intr type");
6396 	}
6397 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6398 	if (err) {
6399 		dev = ctx->ifc_dev;
6400 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6401 		    name);
6402 	}
6403 }
6404 
6405 void
6406 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6407 {
6408 
6409 	if (irq->ii_tag)
6410 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6411 
6412 	if (irq->ii_res)
6413 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6414 		    rman_get_rid(irq->ii_res), irq->ii_res);
6415 }
6416 
6417 static int
6418 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6419 {
6420 	iflib_txq_t txq = ctx->ifc_txqs;
6421 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6422 	if_irq_t irq = &ctx->ifc_legacy_irq;
6423 	iflib_filter_info_t info;
6424 	device_t dev;
6425 	struct grouptask *gtask;
6426 	struct resource *res;
6427 	int err, tqrid;
6428 	bool rx_only;
6429 
6430 	info = &rxq->ifr_filter_info;
6431 	gtask = &rxq->ifr_task;
6432 	tqrid = *rid;
6433 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6434 
6435 	ctx->ifc_flags |= IFC_LEGACY;
6436 	info->ifi_filter = filter;
6437 	info->ifi_filter_arg = filter_arg;
6438 	info->ifi_task = gtask;
6439 	info->ifi_ctx = rxq;
6440 
6441 	dev = ctx->ifc_dev;
6442 	/* We allocate a single interrupt resource */
6443 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr :
6444 	    iflib_fast_intr_rxtx, NULL, info, name);
6445 	if (err != 0)
6446 		return (err);
6447 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq);
6448 	res = irq->ii_res;
6449 	taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name);
6450 
6451 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6452 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6453 	    "tx");
6454 	return (0);
6455 }
6456 
6457 void
6458 iflib_led_create(if_ctx_t ctx)
6459 {
6460 
6461 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6462 	    device_get_nameunit(ctx->ifc_dev));
6463 }
6464 
6465 void
6466 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6467 {
6468 
6469 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6470 }
6471 
6472 void
6473 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6474 {
6475 
6476 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6477 }
6478 
6479 void
6480 iflib_admin_intr_deferred(if_ctx_t ctx)
6481 {
6482 
6483 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6484 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6485 }
6486 
6487 void
6488 iflib_iov_intr_deferred(if_ctx_t ctx)
6489 {
6490 
6491 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6492 }
6493 
6494 void
6495 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6496 {
6497 
6498 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6499 	    name);
6500 }
6501 
6502 void
6503 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6504 	const char *name)
6505 {
6506 
6507 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6508 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6509 	    name);
6510 }
6511 
6512 void
6513 iflib_config_gtask_deinit(struct grouptask *gtask)
6514 {
6515 
6516 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6517 }
6518 
6519 void
6520 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6521 {
6522 	if_t ifp = ctx->ifc_ifp;
6523 	iflib_txq_t txq = ctx->ifc_txqs;
6524 
6525 	if_setbaudrate(ifp, baudrate);
6526 	if (baudrate >= IF_Gbps(10)) {
6527 		STATE_LOCK(ctx);
6528 		ctx->ifc_flags |= IFC_PREFETCH;
6529 		STATE_UNLOCK(ctx);
6530 	}
6531 	/* If link down, disable watchdog */
6532 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6533 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6534 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6535 	}
6536 	ctx->ifc_link_state = link_state;
6537 	if_link_state_change(ifp, link_state);
6538 }
6539 
6540 static int
6541 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6542 {
6543 	int credits;
6544 #ifdef INVARIANTS
6545 	int credits_pre = txq->ift_cidx_processed;
6546 #endif
6547 
6548 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6549 	    BUS_DMASYNC_POSTREAD);
6550 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6551 		return (0);
6552 
6553 	txq->ift_processed += credits;
6554 	txq->ift_cidx_processed += credits;
6555 
6556 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6557 	if (txq->ift_cidx_processed >= txq->ift_size)
6558 		txq->ift_cidx_processed -= txq->ift_size;
6559 	return (credits);
6560 }
6561 
6562 static int
6563 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6564 {
6565 	iflib_fl_t fl;
6566 	u_int i;
6567 
6568 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6569 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6570 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6571 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6572 	    budget));
6573 }
6574 
6575 void
6576 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6577 	const char *description, if_int_delay_info_t info,
6578 	int offset, int value)
6579 {
6580 	info->iidi_ctx = ctx;
6581 	info->iidi_offset = offset;
6582 	info->iidi_value = value;
6583 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6584 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6585 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6586 	    info, 0, iflib_sysctl_int_delay, "I", description);
6587 }
6588 
6589 struct sx *
6590 iflib_ctx_lock_get(if_ctx_t ctx)
6591 {
6592 
6593 	return (&ctx->ifc_ctx_sx);
6594 }
6595 
6596 static int
6597 iflib_msix_init(if_ctx_t ctx)
6598 {
6599 	device_t dev = ctx->ifc_dev;
6600 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6601 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6602 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6603 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6604 
6605 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6606 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6607 
6608 	if (bootverbose)
6609 		device_printf(dev, "msix_init qsets capped at %d\n",
6610 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6611 
6612 	/* Override by tuneable */
6613 	if (scctx->isc_disable_msix)
6614 		goto msi;
6615 
6616 	/* First try MSI-X */
6617 	if ((msgs = pci_msix_count(dev)) == 0) {
6618 		if (bootverbose)
6619 			device_printf(dev, "MSI-X not supported or disabled\n");
6620 		goto msi;
6621 	}
6622 
6623 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6624 	/*
6625 	 * bar == -1 => "trust me I know what I'm doing"
6626 	 * Some drivers are for hardware that is so shoddily
6627 	 * documented that no one knows which bars are which
6628 	 * so the developer has to map all bars. This hack
6629 	 * allows shoddy garbage to use MSI-X in this framework.
6630 	 */
6631 	if (bar != -1) {
6632 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6633 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6634 		if (ctx->ifc_msix_mem == NULL) {
6635 			device_printf(dev, "Unable to map MSI-X table\n");
6636 			goto msi;
6637 		}
6638 	}
6639 
6640 	admincnt = sctx->isc_admin_intrcnt;
6641 #if IFLIB_DEBUG
6642 	/* use only 1 qset in debug mode */
6643 	queuemsgs = min(msgs - admincnt, 1);
6644 #else
6645 	queuemsgs = msgs - admincnt;
6646 #endif
6647 #ifdef RSS
6648 	queues = imin(queuemsgs, rss_getnumbuckets());
6649 #else
6650 	queues = queuemsgs;
6651 #endif
6652 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6653 	if (bootverbose)
6654 		device_printf(dev,
6655 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6656 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6657 #ifdef  RSS
6658 	/* If we're doing RSS, clamp at the number of RSS buckets */
6659 	if (queues > rss_getnumbuckets())
6660 		queues = rss_getnumbuckets();
6661 #endif
6662 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6663 		rx_queues = iflib_num_rx_queues;
6664 	else
6665 		rx_queues = queues;
6666 
6667 	if (rx_queues > scctx->isc_nrxqsets)
6668 		rx_queues = scctx->isc_nrxqsets;
6669 
6670 	/*
6671 	 * We want this to be all logical CPUs by default
6672 	 */
6673 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6674 		tx_queues = iflib_num_tx_queues;
6675 	else
6676 		tx_queues = mp_ncpus;
6677 
6678 	if (tx_queues > scctx->isc_ntxqsets)
6679 		tx_queues = scctx->isc_ntxqsets;
6680 
6681 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6682 #ifdef INVARIANTS
6683 		if (tx_queues != rx_queues)
6684 			device_printf(dev,
6685 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6686 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6687 #endif
6688 		tx_queues = min(rx_queues, tx_queues);
6689 		rx_queues = min(rx_queues, tx_queues);
6690 	}
6691 
6692 	vectors = rx_queues + admincnt;
6693 	if (msgs < vectors) {
6694 		device_printf(dev,
6695 		    "insufficient number of MSI-X vectors "
6696 		    "(supported %d, need %d)\n", msgs, vectors);
6697 		goto msi;
6698 	}
6699 
6700 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6701 	    tx_queues);
6702 	msgs = vectors;
6703 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6704 		if (vectors != msgs) {
6705 			device_printf(dev,
6706 			    "Unable to allocate sufficient MSI-X vectors "
6707 			    "(got %d, need %d)\n", vectors, msgs);
6708 			pci_release_msi(dev);
6709 			if (bar != -1) {
6710 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6711 				    ctx->ifc_msix_mem);
6712 				ctx->ifc_msix_mem = NULL;
6713 			}
6714 			goto msi;
6715 		}
6716 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6717 		    vectors);
6718 		scctx->isc_vectors = vectors;
6719 		scctx->isc_nrxqsets = rx_queues;
6720 		scctx->isc_ntxqsets = tx_queues;
6721 		scctx->isc_intr = IFLIB_INTR_MSIX;
6722 
6723 		return (vectors);
6724 	} else {
6725 		device_printf(dev,
6726 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6727 		    err);
6728 		if (bar != -1) {
6729 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6730 			    ctx->ifc_msix_mem);
6731 			ctx->ifc_msix_mem = NULL;
6732 		}
6733 	}
6734 
6735 msi:
6736 	vectors = pci_msi_count(dev);
6737 	scctx->isc_nrxqsets = 1;
6738 	scctx->isc_ntxqsets = 1;
6739 	scctx->isc_vectors = vectors;
6740 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6741 		device_printf(dev, "Using an MSI interrupt\n");
6742 		scctx->isc_intr = IFLIB_INTR_MSI;
6743 	} else {
6744 		scctx->isc_vectors = 1;
6745 		device_printf(dev, "Using a Legacy interrupt\n");
6746 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6747 	}
6748 
6749 	return (vectors);
6750 }
6751 
6752 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6753 
6754 static int
6755 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6756 {
6757 	int rc;
6758 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6759 	struct sbuf *sb;
6760 	const char *ring_state = "UNKNOWN";
6761 
6762 	/* XXX needed ? */
6763 	rc = sysctl_wire_old_buffer(req, 0);
6764 	MPASS(rc == 0);
6765 	if (rc != 0)
6766 		return (rc);
6767 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6768 	MPASS(sb != NULL);
6769 	if (sb == NULL)
6770 		return (ENOMEM);
6771 	if (state[3] <= 3)
6772 		ring_state = ring_states[state[3]];
6773 
6774 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6775 		    state[0], state[1], state[2], ring_state);
6776 	rc = sbuf_finish(sb);
6777 	sbuf_delete(sb);
6778         return (rc);
6779 }
6780 
6781 enum iflib_ndesc_handler {
6782 	IFLIB_NTXD_HANDLER,
6783 	IFLIB_NRXD_HANDLER,
6784 };
6785 
6786 static int
6787 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6788 {
6789 	if_ctx_t ctx = (void *)arg1;
6790 	enum iflib_ndesc_handler type = arg2;
6791 	char buf[256] = {0};
6792 	qidx_t *ndesc;
6793 	char *p, *next;
6794 	int nqs, rc, i;
6795 
6796 	nqs = 8;
6797 	switch (type) {
6798 	case IFLIB_NTXD_HANDLER:
6799 		ndesc = ctx->ifc_sysctl_ntxds;
6800 		if (ctx->ifc_sctx)
6801 			nqs = ctx->ifc_sctx->isc_ntxqs;
6802 		break;
6803 	case IFLIB_NRXD_HANDLER:
6804 		ndesc = ctx->ifc_sysctl_nrxds;
6805 		if (ctx->ifc_sctx)
6806 			nqs = ctx->ifc_sctx->isc_nrxqs;
6807 		break;
6808 	default:
6809 		printf("%s: unhandled type\n", __func__);
6810 		return (EINVAL);
6811 	}
6812 	if (nqs == 0)
6813 		nqs = 8;
6814 
6815 	for (i = 0; i < 8; i++) {
6816 		if (i >= nqs)
6817 			break;
6818 		if (i)
6819 			strcat(buf, ",");
6820 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6821 	}
6822 
6823 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6824 	if (rc || req->newptr == NULL)
6825 		return (rc);
6826 
6827 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6828 	    i++, p = strsep(&next, " ,")) {
6829 		ndesc[i] = strtoul(p, NULL, 10);
6830 	}
6831 
6832 	return (rc);
6833 }
6834 
6835 #define NAME_BUFLEN 32
6836 static void
6837 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6838 {
6839         device_t dev = iflib_get_dev(ctx);
6840 	struct sysctl_oid_list *child, *oid_list;
6841 	struct sysctl_ctx_list *ctx_list;
6842 	struct sysctl_oid *node;
6843 
6844 	ctx_list = device_get_sysctl_ctx(dev);
6845 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6846 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child,
6847 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
6848 	    "IFLIB fields");
6849 	oid_list = SYSCTL_CHILDREN(node);
6850 
6851 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6852 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
6853 
6854 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6855 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6856 	    "# of txqs to use, 0 => use default #");
6857 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6858 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6859 	    "# of rxqs to use, 0 => use default #");
6860 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6861 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6862 	    "permit #txq != #rxq");
6863 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6864 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6865 	    "disable MSI-X (default 0)");
6866 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6867 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
6868 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6869 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6870 	    "cause TX to abdicate instead of running to completion");
6871 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6872 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6873 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6874 	    "offset to start using cores at");
6875 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6876 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6877 	    "use separate cores for TX and RX");
6878 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6879 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6880 	    "try to make use of logical cores for TX and RX");
6881 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "use_extra_msix_vectors",
6882 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
6883 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
6884 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "allocated_msix_vectors",
6885        	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
6886 	    "total # of MSI-X vectors allocated by driver");
6887 
6888 	/* XXX change for per-queue sizes */
6889 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6890 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6891 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6892 	    "list of # of TX descriptors to use, 0 = use default #");
6893 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6894 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6895 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6896 	    "list of # of RX descriptors to use, 0 = use default #");
6897 }
6898 
6899 static void
6900 iflib_add_device_sysctl_post(if_ctx_t ctx)
6901 {
6902 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6903 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6904         device_t dev = iflib_get_dev(ctx);
6905 	struct sysctl_oid_list *child;
6906 	struct sysctl_ctx_list *ctx_list;
6907 	iflib_fl_t fl;
6908 	iflib_txq_t txq;
6909 	iflib_rxq_t rxq;
6910 	int i, j;
6911 	char namebuf[NAME_BUFLEN];
6912 	char *qfmt;
6913 	struct sysctl_oid *queue_node, *fl_node, *node;
6914 	struct sysctl_oid_list *queue_list, *fl_list;
6915 	ctx_list = device_get_sysctl_ctx(dev);
6916 
6917 	node = ctx->ifc_sysctl_node;
6918 	child = SYSCTL_CHILDREN(node);
6919 
6920 	if (scctx->isc_ntxqsets > 100)
6921 		qfmt = "txq%03d";
6922 	else if (scctx->isc_ntxqsets > 10)
6923 		qfmt = "txq%02d";
6924 	else
6925 		qfmt = "txq%d";
6926 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6927 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6928 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6929 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6930 		queue_list = SYSCTL_CHILDREN(queue_node);
6931 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6932 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
6933 		    "cpu this queue is bound to");
6934 #if MEMORY_LOGGING
6935 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6936 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
6937 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6938 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
6939 #endif
6940 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6941 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
6942 		    "# of times m_defrag was called");
6943 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6944 		    CTLFLAG_RD, &txq->ift_pullups,
6945 		    "# of times m_pullup was called");
6946 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6947 		    "mbuf_defrag_failed", CTLFLAG_RD,
6948 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6949 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6950 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
6951 		    "# of times no descriptors were available");
6952 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6953 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
6954 		    "# of times DMA map failed");
6955 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6956 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
6957 		    "# of times txd_encap returned EFBIG");
6958 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6959 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
6960 		    "# of times map failed for other than EFBIG");
6961 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6962 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
6963 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6964 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
6965 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
6966 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
6967 		    1, "Consumer Index seen by credit update");
6968 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6969 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
6970 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6971 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
6972 		    "descriptors procesed for clean");
6973 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6974 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
6975 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6976 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6977 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6978 		    mp_ring_state_handler, "A", "soft ring state");
6979 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6980 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
6981 		    "# of enqueues to the mp_ring for this queue");
6982 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6983 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
6984 		    "# of drops in the mp_ring for this queue");
6985 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6986 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
6987 		    "# of normal consumer starts in mp_ring for this queue");
6988 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6989 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
6990 		    "# of consumer stalls in the mp_ring for this queue");
6991 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6992 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
6993 		    "# of consumer restarts in the mp_ring for this queue");
6994 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6995 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
6996 		    "# of consumer abdications in the mp_ring for this queue");
6997 	}
6998 
6999 	if (scctx->isc_nrxqsets > 100)
7000 		qfmt = "rxq%03d";
7001 	else if (scctx->isc_nrxqsets > 10)
7002 		qfmt = "rxq%02d";
7003 	else
7004 		qfmt = "rxq%d";
7005 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7006 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7007 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7008 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7009 		queue_list = SYSCTL_CHILDREN(queue_node);
7010 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7011 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
7012 		    "cpu this queue is bound to");
7013 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7014 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7015 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
7016 			    "Consumer Index");
7017 		}
7018 
7019 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7020 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7021 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
7022 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
7023 			    NULL, "freelist Name");
7024 			fl_list = SYSCTL_CHILDREN(fl_node);
7025 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7026 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
7027 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7028 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
7029 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7030 			    CTLFLAG_RD, &fl->ifl_credits, 1,
7031 			    "credits available");
7032 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7033 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
7034 #if MEMORY_LOGGING
7035 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7036 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
7037 			    "mbufs allocated");
7038 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7039 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
7040 			    "mbufs freed");
7041 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7042 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
7043 			    "clusters allocated");
7044 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7045 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
7046 			    "clusters freed");
7047 #endif
7048 		}
7049 	}
7050 
7051 }
7052 
7053 void
7054 iflib_request_reset(if_ctx_t ctx)
7055 {
7056 
7057 	STATE_LOCK(ctx);
7058 	ctx->ifc_flags |= IFC_DO_RESET;
7059 	STATE_UNLOCK(ctx);
7060 }
7061 
7062 #ifndef __NO_STRICT_ALIGNMENT
7063 static struct mbuf *
7064 iflib_fixup_rx(struct mbuf *m)
7065 {
7066 	struct mbuf *n;
7067 
7068 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7069 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7070 		m->m_data += ETHER_HDR_LEN;
7071 		n = m;
7072 	} else {
7073 		MGETHDR(n, M_NOWAIT, MT_DATA);
7074 		if (n == NULL) {
7075 			m_freem(m);
7076 			return (NULL);
7077 		}
7078 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7079 		m->m_data += ETHER_HDR_LEN;
7080 		m->m_len -= ETHER_HDR_LEN;
7081 		n->m_len = ETHER_HDR_LEN;
7082 		M_MOVE_PKTHDR(n, m);
7083 		n->m_next = m;
7084 	}
7085 	return (n);
7086 }
7087 #endif
7088 
7089 #ifdef DEBUGNET
7090 static void
7091 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7092 {
7093 	if_ctx_t ctx;
7094 
7095 	ctx = if_getsoftc(ifp);
7096 	CTX_LOCK(ctx);
7097 	*nrxr = NRXQSETS(ctx);
7098 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7099 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7100 	CTX_UNLOCK(ctx);
7101 }
7102 
7103 static void
7104 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7105 {
7106 	if_ctx_t ctx;
7107 	if_softc_ctx_t scctx;
7108 	iflib_fl_t fl;
7109 	iflib_rxq_t rxq;
7110 	int i, j;
7111 
7112 	ctx = if_getsoftc(ifp);
7113 	scctx = &ctx->ifc_softc_ctx;
7114 
7115 	switch (event) {
7116 	case DEBUGNET_START:
7117 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7118 			rxq = &ctx->ifc_rxqs[i];
7119 			for (j = 0; j < rxq->ifr_nfl; j++) {
7120 				fl = rxq->ifr_fl;
7121 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7122 			}
7123 		}
7124 		iflib_no_tx_batch = 1;
7125 		break;
7126 	default:
7127 		break;
7128 	}
7129 }
7130 
7131 static int
7132 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7133 {
7134 	if_ctx_t ctx;
7135 	iflib_txq_t txq;
7136 	int error;
7137 
7138 	ctx = if_getsoftc(ifp);
7139 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7140 	    IFF_DRV_RUNNING)
7141 		return (EBUSY);
7142 
7143 	txq = &ctx->ifc_txqs[0];
7144 	error = iflib_encap(txq, &m);
7145 	if (error == 0)
7146 		(void)iflib_txd_db_check(txq, true);
7147 	return (error);
7148 }
7149 
7150 static int
7151 iflib_debugnet_poll(if_t ifp, int count)
7152 {
7153 	struct epoch_tracker et;
7154 	if_ctx_t ctx;
7155 	if_softc_ctx_t scctx;
7156 	iflib_txq_t txq;
7157 	int i;
7158 
7159 	ctx = if_getsoftc(ifp);
7160 	scctx = &ctx->ifc_softc_ctx;
7161 
7162 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7163 	    IFF_DRV_RUNNING)
7164 		return (EBUSY);
7165 
7166 	txq = &ctx->ifc_txqs[0];
7167 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7168 
7169 	NET_EPOCH_ENTER(et);
7170 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7171 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7172 	NET_EPOCH_EXIT(et);
7173 	return (0);
7174 }
7175 #endif /* DEBUGNET */
7176