xref: /freebsd/sys/net/iflib.c (revision 66df505066f51e6d8411b966765d828817f88971)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 	uint8_t  ifc_sysctl_use_logical_cores;
199 	bool	 ifc_cpus_are_physical_cores;
200 
201 	qidx_t ifc_sysctl_ntxds[8];
202 	qidx_t ifc_sysctl_nrxds[8];
203 	struct if_txrx ifc_txrx;
204 #define isc_txd_encap  ifc_txrx.ift_txd_encap
205 #define isc_txd_flush  ifc_txrx.ift_txd_flush
206 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
207 #define isc_rxd_available ifc_txrx.ift_rxd_available
208 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 #ifdef DEV_NETMAP
353 	struct callout	ift_netmap_timer;
354 #endif /* DEV_NETMAP */
355 
356 	if_txsd_vec_t	ift_sds;
357 	uint8_t		ift_qstatus;
358 	uint8_t		ift_closed;
359 	uint8_t		ift_update_freq;
360 	struct iflib_filter_info ift_filter_info;
361 	bus_dma_tag_t	ift_buf_tag;
362 	bus_dma_tag_t	ift_tso_buf_tag;
363 	iflib_dma_info_t	ift_ifdi;
364 #define	MTX_NAME_LEN	32
365 	char                    ift_mtx_name[MTX_NAME_LEN];
366 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
367 #ifdef IFLIB_DIAGNOSTICS
368 	uint64_t ift_cpu_exec_count[256];
369 #endif
370 } __aligned(CACHE_LINE_SIZE);
371 
372 struct iflib_fl {
373 	qidx_t		ifl_cidx;
374 	qidx_t		ifl_pidx;
375 	qidx_t		ifl_credits;
376 	uint8_t		ifl_gen;
377 	uint8_t		ifl_rxd_size;
378 #if MEMORY_LOGGING
379 	uint64_t	ifl_m_enqueued;
380 	uint64_t	ifl_m_dequeued;
381 	uint64_t	ifl_cl_enqueued;
382 	uint64_t	ifl_cl_dequeued;
383 #endif
384 	/* implicit pad */
385 	bitstr_t 	*ifl_rx_bitmap;
386 	qidx_t		ifl_fragidx;
387 	/* constant */
388 	qidx_t		ifl_size;
389 	uint16_t	ifl_buf_size;
390 	uint16_t	ifl_cltype;
391 	uma_zone_t	ifl_zone;
392 	iflib_rxsd_array_t	ifl_sds;
393 	iflib_rxq_t	ifl_rxq;
394 	uint8_t		ifl_id;
395 	bus_dma_tag_t	ifl_buf_tag;
396 	iflib_dma_info_t	ifl_ifdi;
397 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
398 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
399 }  __aligned(CACHE_LINE_SIZE);
400 
401 static inline qidx_t
402 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
403 {
404 	qidx_t used;
405 
406 	if (pidx > cidx)
407 		used = pidx - cidx;
408 	else if (pidx < cidx)
409 		used = size - cidx + pidx;
410 	else if (gen == 0 && pidx == cidx)
411 		used = 0;
412 	else if (gen == 1 && pidx == cidx)
413 		used = size;
414 	else
415 		panic("bad state");
416 
417 	return (used);
418 }
419 
420 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
421 
422 #define IDXDIFF(head, tail, wrap) \
423 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
424 
425 struct iflib_rxq {
426 	if_ctx_t	ifr_ctx;
427 	iflib_fl_t	ifr_fl;
428 	uint64_t	ifr_rx_irq;
429 	struct pfil_head	*pfil;
430 	/*
431 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
432 	 * the completion queue consumer index.  Otherwise it's unused.
433 	 */
434 	qidx_t		ifr_cq_cidx;
435 	uint16_t	ifr_id;
436 	uint8_t		ifr_nfl;
437 	uint8_t		ifr_ntxqirq;
438 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
439 	uint8_t		ifr_fl_offset;
440 	struct lro_ctrl			ifr_lc;
441 	struct grouptask        ifr_task;
442 	struct callout		ifr_watchdog;
443 	struct iflib_filter_info ifr_filter_info;
444 	iflib_dma_info_t		ifr_ifdi;
445 
446 	/* dynamically allocate if any drivers need a value substantially larger than this */
447 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
448 #ifdef IFLIB_DIAGNOSTICS
449 	uint64_t ifr_cpu_exec_count[256];
450 #endif
451 }  __aligned(CACHE_LINE_SIZE);
452 
453 typedef struct if_rxsd {
454 	caddr_t *ifsd_cl;
455 	iflib_fl_t ifsd_fl;
456 } *if_rxsd_t;
457 
458 /* multiple of word size */
459 #ifdef __LP64__
460 #define PKT_INFO_SIZE	6
461 #define RXD_INFO_SIZE	5
462 #define PKT_TYPE uint64_t
463 #else
464 #define PKT_INFO_SIZE	11
465 #define RXD_INFO_SIZE	8
466 #define PKT_TYPE uint32_t
467 #endif
468 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
469 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
470 
471 typedef struct if_pkt_info_pad {
472 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
473 } *if_pkt_info_pad_t;
474 typedef struct if_rxd_info_pad {
475 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
476 } *if_rxd_info_pad_t;
477 
478 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
479 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
480 
481 static inline void
482 pkt_info_zero(if_pkt_info_t pi)
483 {
484 	if_pkt_info_pad_t pi_pad;
485 
486 	pi_pad = (if_pkt_info_pad_t)pi;
487 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
488 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
489 #ifndef __LP64__
490 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
491 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
492 #endif
493 }
494 
495 static device_method_t iflib_pseudo_methods[] = {
496 	DEVMETHOD(device_attach, noop_attach),
497 	DEVMETHOD(device_detach, iflib_pseudo_detach),
498 	DEVMETHOD_END
499 };
500 
501 driver_t iflib_pseudodriver = {
502 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
503 };
504 
505 static inline void
506 rxd_info_zero(if_rxd_info_t ri)
507 {
508 	if_rxd_info_pad_t ri_pad;
509 	int i;
510 
511 	ri_pad = (if_rxd_info_pad_t)ri;
512 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
513 		ri_pad->rxd_val[i] = 0;
514 		ri_pad->rxd_val[i+1] = 0;
515 		ri_pad->rxd_val[i+2] = 0;
516 		ri_pad->rxd_val[i+3] = 0;
517 	}
518 #ifdef __LP64__
519 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
520 #endif
521 }
522 
523 /*
524  * Only allow a single packet to take up most 1/nth of the tx ring
525  */
526 #define MAX_SINGLE_PACKET_FRACTION 12
527 #define IF_BAD_DMA (bus_addr_t)-1
528 
529 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
530 
531 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
532 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
535 
536 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
537 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
538 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
539 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
540 
541 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
542 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
543 
544 void
545 iflib_set_detach(if_ctx_t ctx)
546 {
547 	STATE_LOCK(ctx);
548 	ctx->ifc_flags |= IFC_IN_DETACH;
549 	STATE_UNLOCK(ctx);
550 }
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, "minimize transmit latency at the possible expense of throughput");
587 static int iflib_no_tx_batch = 0;
588 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
589 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
590 static int iflib_timer_default = 1000;
591 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
592 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
593 
594 
595 #if IFLIB_DEBUG_COUNTERS
596 
597 static int iflib_tx_seen;
598 static int iflib_tx_sent;
599 static int iflib_tx_encap;
600 static int iflib_rx_allocs;
601 static int iflib_fl_refills;
602 static int iflib_fl_refills_large;
603 static int iflib_tx_frees;
604 
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
606 		   &iflib_tx_seen, 0, "# TX mbufs seen");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
608 		   &iflib_tx_sent, 0, "# TX mbufs sent");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
610 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
612 		   &iflib_tx_frees, 0, "# TX frees");
613 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
614 		   &iflib_rx_allocs, 0, "# RX allocations");
615 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
616 		   &iflib_fl_refills, 0, "# refills");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
618 		   &iflib_fl_refills_large, 0, "# large refills");
619 
620 static int iflib_txq_drain_flushing;
621 static int iflib_txq_drain_oactive;
622 static int iflib_txq_drain_notready;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
625 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
627 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
629 		   &iflib_txq_drain_notready, 0, "# drain notready");
630 
631 static int iflib_encap_load_mbuf_fail;
632 static int iflib_encap_pad_mbuf_fail;
633 static int iflib_encap_txq_avail_fail;
634 static int iflib_encap_txd_encap_fail;
635 
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
639 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
641 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
643 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
644 
645 static int iflib_task_fn_rxs;
646 static int iflib_rx_intr_enables;
647 static int iflib_fast_intrs;
648 static int iflib_rx_unavail;
649 static int iflib_rx_ctx_inactive;
650 static int iflib_rx_if_input;
651 static int iflib_rxd_flush;
652 
653 static int iflib_verbose_debug;
654 
655 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
656 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
658 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
659 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
660 		   &iflib_fast_intrs, 0, "# fast_intr calls");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
662 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
664 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
666 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
668 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
669 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
670 		   &iflib_verbose_debug, 0, "enable verbose debugging");
671 
672 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
673 static void
674 iflib_debug_reset(void)
675 {
676 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
677 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
678 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
679 		iflib_txq_drain_notready =
680 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
681 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
682 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
683 		iflib_rx_unavail =
684 		iflib_rx_ctx_inactive = iflib_rx_if_input =
685 		iflib_rxd_flush = 0;
686 }
687 
688 #else
689 #define DBG_COUNTER_INC(name)
690 static void iflib_debug_reset(void) {}
691 #endif
692 
693 #define IFLIB_DEBUG 0
694 
695 static void iflib_tx_structures_free(if_ctx_t ctx);
696 static void iflib_rx_structures_free(if_ctx_t ctx);
697 static int iflib_queues_alloc(if_ctx_t ctx);
698 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
699 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
700 static int iflib_qset_structures_setup(if_ctx_t ctx);
701 static int iflib_msix_init(if_ctx_t ctx);
702 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
703 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
704 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
705 #ifdef ALTQ
706 static void iflib_altq_if_start(if_t ifp);
707 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
708 #endif
709 static int iflib_register(if_ctx_t);
710 static void iflib_deregister(if_ctx_t);
711 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
712 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
713 static void iflib_init_locked(if_ctx_t ctx);
714 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
715 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
716 static void iflib_ifmp_purge(iflib_txq_t txq);
717 static void _iflib_pre_assert(if_softc_ctx_t scctx);
718 static void iflib_if_init_locked(if_ctx_t ctx);
719 static void iflib_free_intr_mem(if_ctx_t ctx);
720 #ifndef __NO_STRICT_ALIGNMENT
721 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
722 #endif
723 
724 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
725     SLIST_HEAD_INITIALIZER(cpu_offsets);
726 struct cpu_offset {
727 	SLIST_ENTRY(cpu_offset) entries;
728 	cpuset_t	set;
729 	unsigned int	refcount;
730 	uint16_t	next_cpuid;
731 };
732 static struct mtx cpu_offset_mtx;
733 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
734     MTX_DEF);
735 
736 DEBUGNET_DEFINE(iflib);
737 
738 static int
739 iflib_num_rx_descs(if_ctx_t ctx)
740 {
741 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
742 	if_shared_ctx_t sctx = ctx->ifc_sctx;
743 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
744 
745 	return scctx->isc_nrxd[first_rxq];
746 }
747 
748 static int
749 iflib_num_tx_descs(if_ctx_t ctx)
750 {
751 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
752 	if_shared_ctx_t sctx = ctx->ifc_sctx;
753 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
754 
755 	return scctx->isc_ntxd[first_txq];
756 }
757 
758 #ifdef DEV_NETMAP
759 #include <sys/selinfo.h>
760 #include <net/netmap.h>
761 #include <dev/netmap/netmap_kern.h>
762 
763 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
764 
765 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
766 static void iflib_netmap_timer(void *arg);
767 
768 /*
769  * device-specific sysctl variables:
770  *
771  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
772  *	During regular operations the CRC is stripped, but on some
773  *	hardware reception of frames not multiple of 64 is slower,
774  *	so using crcstrip=0 helps in benchmarks.
775  *
776  * iflib_rx_miss, iflib_rx_miss_bufs:
777  *	count packets that might be missed due to lost interrupts.
778  */
779 SYSCTL_DECL(_dev_netmap);
780 /*
781  * The xl driver by default strips CRCs and we do not override it.
782  */
783 
784 int iflib_crcstrip = 1;
785 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
786     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
787 
788 int iflib_rx_miss, iflib_rx_miss_bufs;
789 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
790     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
791 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
792     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
793 
794 /*
795  * Register/unregister. We are already under netmap lock.
796  * Only called on the first register or the last unregister.
797  */
798 static int
799 iflib_netmap_register(struct netmap_adapter *na, int onoff)
800 {
801 	if_t ifp = na->ifp;
802 	if_ctx_t ctx = ifp->if_softc;
803 	int status;
804 
805 	CTX_LOCK(ctx);
806 	if (!CTX_IS_VF(ctx))
807 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
808 
809 	iflib_stop(ctx);
810 
811 	/*
812 	 * Enable (or disable) netmap flags, and intercept (or restore)
813 	 * ifp->if_transmit. This is done once the device has been stopped
814 	 * to prevent race conditions. Also, this must be done after
815 	 * calling netmap_disable_all_rings() and before calling
816 	 * netmap_enable_all_rings(), so that these two functions see the
817 	 * updated state of the NAF_NETMAP_ON bit.
818 	 */
819 	if (onoff) {
820 		nm_set_native_flags(na);
821 	} else {
822 		nm_clear_native_flags(na);
823 	}
824 
825 	iflib_init_locked(ctx);
826 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
827 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
828 	if (status)
829 		nm_clear_native_flags(na);
830 	CTX_UNLOCK(ctx);
831 	return (status);
832 }
833 
834 static int
835 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
836 {
837 	if_t ifp = na->ifp;
838 	if_ctx_t ctx = ifp->if_softc;
839 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
840 	iflib_fl_t fl = &rxq->ifr_fl[0];
841 
842 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
843 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
844 	info->num_tx_descs = iflib_num_tx_descs(ctx);
845 	info->num_rx_descs = iflib_num_rx_descs(ctx);
846 	info->rx_buf_maxsize = fl->ifl_buf_size;
847 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
848 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
849 		info->num_rx_descs, info->rx_buf_maxsize);
850 
851 	return 0;
852 }
853 
854 static int
855 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
856 {
857 	struct netmap_adapter *na = kring->na;
858 	u_int const lim = kring->nkr_num_slots - 1;
859 	struct netmap_ring *ring = kring->ring;
860 	bus_dmamap_t *map;
861 	struct if_rxd_update iru;
862 	if_ctx_t ctx = rxq->ifr_ctx;
863 	iflib_fl_t fl = &rxq->ifr_fl[0];
864 	u_int nic_i_first, nic_i;
865 	u_int nm_i;
866 	int i, n;
867 #if IFLIB_DEBUG_COUNTERS
868 	int rf_count = 0;
869 #endif
870 
871 	/*
872 	 * This function is used both at initialization and in rxsync.
873 	 * At initialization we need to prepare (with isc_rxd_refill())
874 	 * all the netmap buffers currently owned by the kernel, in
875 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
876 	 * (except for kring->nkr_hwofs). These may be less than
877 	 * kring->nkr_num_slots if netmap_reset() was called while
878 	 * an application using the kring that still owned some
879 	 * buffers.
880 	 * At rxsync time, both indexes point to the next buffer to be
881 	 * refilled.
882 	 * In any case we publish (with isc_rxd_flush()) up to
883 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
884 	 * pointer to overrun the head/cons pointer, although this is
885 	 * not necessary for some NICs (e.g. vmx).
886 	 */
887 	if (__predict_false(init)) {
888 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
889 	} else {
890 		n = kring->rhead - kring->nr_hwcur;
891 		if (n == 0)
892 			return (0); /* Nothing to do. */
893 		if (n < 0)
894 			n += kring->nkr_num_slots;
895 	}
896 
897 	iru_init(&iru, rxq, 0 /* flid */);
898 	map = fl->ifl_sds.ifsd_map;
899 	nic_i = fl->ifl_pidx;
900 	nm_i = netmap_idx_n2k(kring, nic_i);
901 	if (__predict_false(init)) {
902 		/*
903 		 * On init/reset, nic_i must be 0, and we must
904 		 * start to refill from hwtail (see netmap_reset()).
905 		 */
906 		MPASS(nic_i == 0);
907 		MPASS(nm_i == kring->nr_hwtail);
908 	} else
909 		MPASS(nm_i == kring->nr_hwcur);
910 	DBG_COUNTER_INC(fl_refills);
911 	while (n > 0) {
912 #if IFLIB_DEBUG_COUNTERS
913 		if (++rf_count == 9)
914 			DBG_COUNTER_INC(fl_refills_large);
915 #endif
916 		nic_i_first = nic_i;
917 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
918 			struct netmap_slot *slot = &ring->slot[nm_i];
919 			uint64_t paddr;
920 			void *addr = PNMB(na, slot, &paddr);
921 
922 			MPASS(i < IFLIB_MAX_RX_REFRESH);
923 
924 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
925 			        return netmap_ring_reinit(kring);
926 
927 			fl->ifl_bus_addrs[i] = paddr +
928 			    nm_get_offset(kring, slot);
929 			fl->ifl_rxd_idxs[i] = nic_i;
930 
931 			if (__predict_false(init)) {
932 				netmap_load_map(na, fl->ifl_buf_tag,
933 				    map[nic_i], addr);
934 			} else if (slot->flags & NS_BUF_CHANGED) {
935 				/* buffer has changed, reload map */
936 				netmap_reload_map(na, fl->ifl_buf_tag,
937 				    map[nic_i], addr);
938 			}
939 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
940 			    BUS_DMASYNC_PREREAD);
941 			slot->flags &= ~NS_BUF_CHANGED;
942 
943 			nm_i = nm_next(nm_i, lim);
944 			nic_i = nm_next(nic_i, lim);
945 		}
946 
947 		iru.iru_pidx = nic_i_first;
948 		iru.iru_count = i;
949 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
950 	}
951 	fl->ifl_pidx = nic_i;
952 	/*
953 	 * At the end of the loop we must have refilled everything
954 	 * we could possibly refill.
955 	 */
956 	MPASS(nm_i == kring->rhead);
957 	kring->nr_hwcur = nm_i;
958 
959 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
960 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
961 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
962 	    nm_prev(nic_i, lim));
963 	DBG_COUNTER_INC(rxd_flush);
964 
965 	return (0);
966 }
967 
968 #define NETMAP_TX_TIMER_US	90
969 
970 /*
971  * Reconcile kernel and user view of the transmit ring.
972  *
973  * All information is in the kring.
974  * Userspace wants to send packets up to the one before kring->rhead,
975  * kernel knows kring->nr_hwcur is the first unsent packet.
976  *
977  * Here we push packets out (as many as possible), and possibly
978  * reclaim buffers from previously completed transmission.
979  *
980  * The caller (netmap) guarantees that there is only one instance
981  * running at any time. Any interference with other driver
982  * methods should be handled by the individual drivers.
983  */
984 static int
985 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
986 {
987 	struct netmap_adapter *na = kring->na;
988 	if_t ifp = na->ifp;
989 	struct netmap_ring *ring = kring->ring;
990 	u_int nm_i;	/* index into the netmap kring */
991 	u_int nic_i;	/* index into the NIC ring */
992 	u_int n;
993 	u_int const lim = kring->nkr_num_slots - 1;
994 	u_int const head = kring->rhead;
995 	struct if_pkt_info pi;
996 	int tx_pkts = 0, tx_bytes = 0;
997 
998 	/*
999 	 * interrupts on every tx packet are expensive so request
1000 	 * them every half ring, or where NS_REPORT is set
1001 	 */
1002 	u_int report_frequency = kring->nkr_num_slots >> 1;
1003 	/* device-specific */
1004 	if_ctx_t ctx = ifp->if_softc;
1005 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1006 
1007 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1008 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1009 
1010 	/*
1011 	 * First part: process new packets to send.
1012 	 * nm_i is the current index in the netmap kring,
1013 	 * nic_i is the corresponding index in the NIC ring.
1014 	 *
1015 	 * If we have packets to send (nm_i != head)
1016 	 * iterate over the netmap ring, fetch length and update
1017 	 * the corresponding slot in the NIC ring. Some drivers also
1018 	 * need to update the buffer's physical address in the NIC slot
1019 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1020 	 *
1021 	 * The netmap_reload_map() calls is especially expensive,
1022 	 * even when (as in this case) the tag is 0, so do only
1023 	 * when the buffer has actually changed.
1024 	 *
1025 	 * If possible do not set the report/intr bit on all slots,
1026 	 * but only a few times per ring or when NS_REPORT is set.
1027 	 *
1028 	 * Finally, on 10G and faster drivers, it might be useful
1029 	 * to prefetch the next slot and txr entry.
1030 	 */
1031 
1032 	nm_i = kring->nr_hwcur;
1033 	if (nm_i != head) {	/* we have new packets to send */
1034 		uint32_t pkt_len = 0, seg_idx = 0;
1035 		int nic_i_start = -1, flags = 0;
1036 		pkt_info_zero(&pi);
1037 		pi.ipi_segs = txq->ift_segs;
1038 		pi.ipi_qsidx = kring->ring_id;
1039 		nic_i = netmap_idx_k2n(kring, nm_i);
1040 
1041 		__builtin_prefetch(&ring->slot[nm_i]);
1042 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1043 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1044 
1045 		for (n = 0; nm_i != head; n++) {
1046 			struct netmap_slot *slot = &ring->slot[nm_i];
1047 			uint64_t offset = nm_get_offset(kring, slot);
1048 			u_int len = slot->len;
1049 			uint64_t paddr;
1050 			void *addr = PNMB(na, slot, &paddr);
1051 
1052 			flags |= (slot->flags & NS_REPORT ||
1053 				nic_i == 0 || nic_i == report_frequency) ?
1054 				IPI_TX_INTR : 0;
1055 
1056 			/*
1057 			 * If this is the first packet fragment, save the
1058 			 * index of the first NIC slot for later.
1059 			 */
1060 			if (nic_i_start < 0)
1061 				nic_i_start = nic_i;
1062 
1063 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1064 			pi.ipi_segs[seg_idx].ds_len = len;
1065 			if (len) {
1066 				pkt_len += len;
1067 				seg_idx++;
1068 			}
1069 
1070 			if (!(slot->flags & NS_MOREFRAG)) {
1071 				pi.ipi_len = pkt_len;
1072 				pi.ipi_nsegs = seg_idx;
1073 				pi.ipi_pidx = nic_i_start;
1074 				pi.ipi_ndescs = 0;
1075 				pi.ipi_flags = flags;
1076 
1077 				/* Prepare the NIC TX ring. */
1078 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1079 				DBG_COUNTER_INC(tx_encap);
1080 
1081 				/* Update transmit counters */
1082 				tx_bytes += pi.ipi_len;
1083 				tx_pkts++;
1084 
1085 				/* Reinit per-packet info for the next one. */
1086 				flags = seg_idx = pkt_len = 0;
1087 				nic_i_start = -1;
1088 			}
1089 
1090 			/* prefetch for next round */
1091 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1092 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1093 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1094 
1095 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1096 
1097 			if (slot->flags & NS_BUF_CHANGED) {
1098 				/* buffer has changed, reload map */
1099 				netmap_reload_map(na, txq->ift_buf_tag,
1100 				    txq->ift_sds.ifsd_map[nic_i], addr);
1101 			}
1102 			/* make sure changes to the buffer are synced */
1103 			bus_dmamap_sync(txq->ift_buf_tag,
1104 			    txq->ift_sds.ifsd_map[nic_i],
1105 			    BUS_DMASYNC_PREWRITE);
1106 
1107 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1108 			nm_i = nm_next(nm_i, lim);
1109 			nic_i = nm_next(nic_i, lim);
1110 		}
1111 		kring->nr_hwcur = nm_i;
1112 
1113 		/* synchronize the NIC ring */
1114 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1115 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1116 
1117 		/* (re)start the tx unit up to slot nic_i (excluded) */
1118 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1119 	}
1120 
1121 	/*
1122 	 * Second part: reclaim buffers for completed transmissions.
1123 	 *
1124 	 * If there are unclaimed buffers, attempt to reclaim them.
1125 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1126 	 * trigger a per-tx-queue timer to try again later.
1127 	 */
1128 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1129 		if (iflib_tx_credits_update(ctx, txq)) {
1130 			/* some tx completed, increment avail */
1131 			nic_i = txq->ift_cidx_processed;
1132 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1133 		}
1134 	}
1135 
1136 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1137 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1138 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1139 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1140 			    iflib_netmap_timer, txq,
1141 			    txq->ift_netmap_timer.c_cpu, 0);
1142 		}
1143 
1144 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1145 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1146 
1147 	return (0);
1148 }
1149 
1150 /*
1151  * Reconcile kernel and user view of the receive ring.
1152  * Same as for the txsync, this routine must be efficient.
1153  * The caller guarantees a single invocations, but races against
1154  * the rest of the driver should be handled here.
1155  *
1156  * On call, kring->rhead is the first packet that userspace wants
1157  * to keep, and kring->rcur is the wakeup point.
1158  * The kernel has previously reported packets up to kring->rtail.
1159  *
1160  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1161  * of whether or not we received an interrupt.
1162  */
1163 static int
1164 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1165 {
1166 	struct netmap_adapter *na = kring->na;
1167 	struct netmap_ring *ring = kring->ring;
1168 	if_t ifp = na->ifp;
1169 	uint32_t nm_i;	/* index into the netmap ring */
1170 	uint32_t nic_i;	/* index into the NIC ring */
1171 	u_int n;
1172 	u_int const lim = kring->nkr_num_slots - 1;
1173 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1174 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1175 
1176 	if_ctx_t ctx = ifp->if_softc;
1177 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1178 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1179 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1180 	iflib_fl_t fl = &rxq->ifr_fl[0];
1181 	struct if_rxd_info ri;
1182 	qidx_t *cidxp;
1183 
1184 	/*
1185 	 * netmap only uses free list 0, to avoid out of order consumption
1186 	 * of receive buffers
1187 	 */
1188 
1189 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1190 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1191 
1192 	/*
1193 	 * First part: import newly received packets.
1194 	 *
1195 	 * nm_i is the index of the next free slot in the netmap ring,
1196 	 * nic_i is the index of the next received packet in the NIC ring
1197 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1198 	 * differ in case if_init() has been called while
1199 	 * in netmap mode. For the receive ring we have
1200 	 *
1201 	 *	nic_i = fl->ifl_cidx;
1202 	 *	nm_i = kring->nr_hwtail (previous)
1203 	 * and
1204 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1205 	 *
1206 	 * fl->ifl_cidx is set to 0 on a ring reinit
1207 	 */
1208 	if (netmap_no_pendintr || force_update) {
1209 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1210 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1211 		int crclen = iflib_crcstrip ? 0 : 4;
1212 		int error, avail;
1213 
1214 		/*
1215 		 * For the free list consumer index, we use the same
1216 		 * logic as in iflib_rxeof().
1217 		 */
1218 		if (have_rxcq)
1219 			cidxp = &rxq->ifr_cq_cidx;
1220 		else
1221 			cidxp = &fl->ifl_cidx;
1222 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1223 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1224 
1225 		nic_i = fl->ifl_cidx;
1226 		nm_i = netmap_idx_n2k(kring, nic_i);
1227 		MPASS(nm_i == kring->nr_hwtail);
1228 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1229 			rxd_info_zero(&ri);
1230 			ri.iri_frags = rxq->ifr_frags;
1231 			ri.iri_qsidx = kring->ring_id;
1232 			ri.iri_ifp = ctx->ifc_ifp;
1233 			ri.iri_cidx = *cidxp;
1234 
1235 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1236 			for (i = 0; i < ri.iri_nfrags; i++) {
1237 				if (error) {
1238 					ring->slot[nm_i].len = 0;
1239 					ring->slot[nm_i].flags = 0;
1240 				} else {
1241 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1242 					if (i == (ri.iri_nfrags - 1)) {
1243 						ring->slot[nm_i].len -= crclen;
1244 						ring->slot[nm_i].flags = 0;
1245 
1246 						/* Update receive counters */
1247 						rx_bytes += ri.iri_len;
1248 						rx_pkts++;
1249 					} else
1250 						ring->slot[nm_i].flags = NS_MOREFRAG;
1251 				}
1252 
1253 				bus_dmamap_sync(fl->ifl_buf_tag,
1254 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1255 				nm_i = nm_next(nm_i, lim);
1256 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1257 			}
1258 
1259 			if (have_rxcq) {
1260 				*cidxp = ri.iri_cidx;
1261 				while (*cidxp >= scctx->isc_nrxd[0])
1262 					*cidxp -= scctx->isc_nrxd[0];
1263 			}
1264 
1265 		}
1266 		if (n) { /* update the state variables */
1267 			if (netmap_no_pendintr && !force_update) {
1268 				/* diagnostics */
1269 				iflib_rx_miss ++;
1270 				iflib_rx_miss_bufs += n;
1271 			}
1272 			kring->nr_hwtail = nm_i;
1273 		}
1274 		kring->nr_kflags &= ~NKR_PENDINTR;
1275 	}
1276 	/*
1277 	 * Second part: skip past packets that userspace has released.
1278 	 * (kring->nr_hwcur to head excluded),
1279 	 * and make the buffers available for reception.
1280 	 * As usual nm_i is the index in the netmap ring,
1281 	 * nic_i is the index in the NIC ring, and
1282 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1283 	 */
1284 	netmap_fl_refill(rxq, kring, false);
1285 
1286 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1287 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1288 
1289 	return (0);
1290 }
1291 
1292 static void
1293 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1294 {
1295 	if_ctx_t ctx = na->ifp->if_softc;
1296 
1297 	CTX_LOCK(ctx);
1298 	if (onoff) {
1299 		IFDI_INTR_ENABLE(ctx);
1300 	} else {
1301 		IFDI_INTR_DISABLE(ctx);
1302 	}
1303 	CTX_UNLOCK(ctx);
1304 }
1305 
1306 static int
1307 iflib_netmap_attach(if_ctx_t ctx)
1308 {
1309 	struct netmap_adapter na;
1310 
1311 	bzero(&na, sizeof(na));
1312 
1313 	na.ifp = ctx->ifc_ifp;
1314 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1315 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1316 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1317 
1318 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1319 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1320 	na.nm_txsync = iflib_netmap_txsync;
1321 	na.nm_rxsync = iflib_netmap_rxsync;
1322 	na.nm_register = iflib_netmap_register;
1323 	na.nm_intr = iflib_netmap_intr;
1324 	na.nm_config = iflib_netmap_config;
1325 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1326 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1327 	return (netmap_attach(&na));
1328 }
1329 
1330 static int
1331 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1332 {
1333 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1334 	struct netmap_slot *slot;
1335 
1336 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1337 	if (slot == NULL)
1338 		return (0);
1339 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1340 		/*
1341 		 * In netmap mode, set the map for the packet buffer.
1342 		 * NOTE: Some drivers (not this one) also need to set
1343 		 * the physical buffer address in the NIC ring.
1344 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1345 		 * netmap slot index, si
1346 		 */
1347 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1348 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1349 		    NMB(na, slot + si));
1350 	}
1351 	return (1);
1352 }
1353 
1354 static int
1355 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1356 {
1357 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1358 	struct netmap_kring *kring;
1359 	struct netmap_slot *slot;
1360 
1361 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1362 	if (slot == NULL)
1363 		return (0);
1364 	kring = na->rx_rings[rxq->ifr_id];
1365 	netmap_fl_refill(rxq, kring, true);
1366 	return (1);
1367 }
1368 
1369 static void
1370 iflib_netmap_timer(void *arg)
1371 {
1372 	iflib_txq_t txq = arg;
1373 	if_ctx_t ctx = txq->ift_ctx;
1374 
1375 	/*
1376 	 * Wake up the netmap application, to give it a chance to
1377 	 * call txsync and reclaim more completed TX buffers.
1378 	 */
1379 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1380 }
1381 
1382 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1383 
1384 #else
1385 #define iflib_netmap_txq_init(ctx, txq) (0)
1386 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1387 #define iflib_netmap_detach(ifp)
1388 #define netmap_enable_all_rings(ifp)
1389 #define netmap_disable_all_rings(ifp)
1390 
1391 #define iflib_netmap_attach(ctx) (0)
1392 #define netmap_rx_irq(ifp, qid, budget) (0)
1393 #endif
1394 
1395 #if defined(__i386__) || defined(__amd64__)
1396 static __inline void
1397 prefetch(void *x)
1398 {
1399 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1400 }
1401 static __inline void
1402 prefetch2cachelines(void *x)
1403 {
1404 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1405 #if (CACHE_LINE_SIZE < 128)
1406 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1407 #endif
1408 }
1409 #else
1410 #define prefetch(x)
1411 #define prefetch2cachelines(x)
1412 #endif
1413 
1414 static void
1415 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1416 {
1417 	iflib_fl_t fl;
1418 
1419 	fl = &rxq->ifr_fl[flid];
1420 	iru->iru_paddrs = fl->ifl_bus_addrs;
1421 	iru->iru_idxs = fl->ifl_rxd_idxs;
1422 	iru->iru_qsidx = rxq->ifr_id;
1423 	iru->iru_buf_size = fl->ifl_buf_size;
1424 	iru->iru_flidx = fl->ifl_id;
1425 }
1426 
1427 static void
1428 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1429 {
1430 	if (err)
1431 		return;
1432 	*(bus_addr_t *) arg = segs[0].ds_addr;
1433 }
1434 
1435 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1436 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1437 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1438 
1439 int
1440 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1441 {
1442 	int err;
1443 	device_t dev = ctx->ifc_dev;
1444 	bus_addr_t lowaddr;
1445 
1446 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1447 
1448 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1449 				align, 0,		/* alignment, bounds */
1450 				lowaddr,		/* lowaddr */
1451 				BUS_SPACE_MAXADDR,	/* highaddr */
1452 				NULL, NULL,		/* filter, filterarg */
1453 				size,			/* maxsize */
1454 				1,			/* nsegments */
1455 				size,			/* maxsegsize */
1456 				BUS_DMA_ALLOCNOW,	/* flags */
1457 				NULL,			/* lockfunc */
1458 				NULL,			/* lockarg */
1459 				&dma->idi_tag);
1460 	if (err) {
1461 		device_printf(dev,
1462 		    "%s: bus_dma_tag_create failed: %d\n",
1463 		    __func__, err);
1464 		goto fail_0;
1465 	}
1466 
1467 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1468 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1469 	if (err) {
1470 		device_printf(dev,
1471 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1472 		    __func__, (uintmax_t)size, err);
1473 		goto fail_1;
1474 	}
1475 
1476 	dma->idi_paddr = IF_BAD_DMA;
1477 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1478 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1479 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1480 		device_printf(dev,
1481 		    "%s: bus_dmamap_load failed: %d\n",
1482 		    __func__, err);
1483 		goto fail_2;
1484 	}
1485 
1486 	dma->idi_size = size;
1487 	return (0);
1488 
1489 fail_2:
1490 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1491 fail_1:
1492 	bus_dma_tag_destroy(dma->idi_tag);
1493 fail_0:
1494 	dma->idi_tag = NULL;
1495 
1496 	return (err);
1497 }
1498 
1499 int
1500 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1501 {
1502 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1503 
1504 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1505 
1506 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1507 }
1508 
1509 int
1510 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1511 {
1512 	int i, err;
1513 	iflib_dma_info_t *dmaiter;
1514 
1515 	dmaiter = dmalist;
1516 	for (i = 0; i < count; i++, dmaiter++) {
1517 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1518 			break;
1519 	}
1520 	if (err)
1521 		iflib_dma_free_multi(dmalist, i);
1522 	return (err);
1523 }
1524 
1525 void
1526 iflib_dma_free(iflib_dma_info_t dma)
1527 {
1528 	if (dma->idi_tag == NULL)
1529 		return;
1530 	if (dma->idi_paddr != IF_BAD_DMA) {
1531 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1532 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1533 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1534 		dma->idi_paddr = IF_BAD_DMA;
1535 	}
1536 	if (dma->idi_vaddr != NULL) {
1537 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1538 		dma->idi_vaddr = NULL;
1539 	}
1540 	bus_dma_tag_destroy(dma->idi_tag);
1541 	dma->idi_tag = NULL;
1542 }
1543 
1544 void
1545 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1546 {
1547 	int i;
1548 	iflib_dma_info_t *dmaiter = dmalist;
1549 
1550 	for (i = 0; i < count; i++, dmaiter++)
1551 		iflib_dma_free(*dmaiter);
1552 }
1553 
1554 static int
1555 iflib_fast_intr(void *arg)
1556 {
1557 	iflib_filter_info_t info = arg;
1558 	struct grouptask *gtask = info->ifi_task;
1559 	int result;
1560 
1561 	DBG_COUNTER_INC(fast_intrs);
1562 	if (info->ifi_filter != NULL) {
1563 		result = info->ifi_filter(info->ifi_filter_arg);
1564 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1565 			return (result);
1566 	}
1567 
1568 	GROUPTASK_ENQUEUE(gtask);
1569 	return (FILTER_HANDLED);
1570 }
1571 
1572 static int
1573 iflib_fast_intr_rxtx(void *arg)
1574 {
1575 	iflib_filter_info_t info = arg;
1576 	struct grouptask *gtask = info->ifi_task;
1577 	if_ctx_t ctx;
1578 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1579 	iflib_txq_t txq;
1580 	void *sc;
1581 	int i, cidx, result;
1582 	qidx_t txqid;
1583 	bool intr_enable, intr_legacy;
1584 
1585 	DBG_COUNTER_INC(fast_intrs);
1586 	if (info->ifi_filter != NULL) {
1587 		result = info->ifi_filter(info->ifi_filter_arg);
1588 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1589 			return (result);
1590 	}
1591 
1592 	ctx = rxq->ifr_ctx;
1593 	sc = ctx->ifc_softc;
1594 	intr_enable = false;
1595 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1596 	MPASS(rxq->ifr_ntxqirq);
1597 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1598 		txqid = rxq->ifr_txqid[i];
1599 		txq = &ctx->ifc_txqs[txqid];
1600 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1601 		    BUS_DMASYNC_POSTREAD);
1602 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1603 			if (intr_legacy)
1604 				intr_enable = true;
1605 			else
1606 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1607 			continue;
1608 		}
1609 		GROUPTASK_ENQUEUE(&txq->ift_task);
1610 	}
1611 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1612 		cidx = rxq->ifr_cq_cidx;
1613 	else
1614 		cidx = rxq->ifr_fl[0].ifl_cidx;
1615 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1616 		GROUPTASK_ENQUEUE(gtask);
1617 	else {
1618 		if (intr_legacy)
1619 			intr_enable = true;
1620 		else
1621 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1622 		DBG_COUNTER_INC(rx_intr_enables);
1623 	}
1624 	if (intr_enable)
1625 		IFDI_INTR_ENABLE(ctx);
1626 	return (FILTER_HANDLED);
1627 }
1628 
1629 static int
1630 iflib_fast_intr_ctx(void *arg)
1631 {
1632 	iflib_filter_info_t info = arg;
1633 	struct grouptask *gtask = info->ifi_task;
1634 	int result;
1635 
1636 	DBG_COUNTER_INC(fast_intrs);
1637 	if (info->ifi_filter != NULL) {
1638 		result = info->ifi_filter(info->ifi_filter_arg);
1639 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1640 			return (result);
1641 	}
1642 
1643 	GROUPTASK_ENQUEUE(gtask);
1644 	return (FILTER_HANDLED);
1645 }
1646 
1647 static int
1648 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1649 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1650 		 const char *name)
1651 {
1652 	struct resource *res;
1653 	void *tag = NULL;
1654 	device_t dev = ctx->ifc_dev;
1655 	int flags, i, rc;
1656 
1657 	flags = RF_ACTIVE;
1658 	if (ctx->ifc_flags & IFC_LEGACY)
1659 		flags |= RF_SHAREABLE;
1660 	MPASS(rid < 512);
1661 	i = rid;
1662 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1663 	if (res == NULL) {
1664 		device_printf(dev,
1665 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1666 		return (ENOMEM);
1667 	}
1668 	irq->ii_res = res;
1669 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1670 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1671 						filter, handler, arg, &tag);
1672 	if (rc != 0) {
1673 		device_printf(dev,
1674 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1675 					  rid, name ? name : "unknown", rc);
1676 		return (rc);
1677 	} else if (name)
1678 		bus_describe_intr(dev, res, tag, "%s", name);
1679 
1680 	irq->ii_tag = tag;
1681 	return (0);
1682 }
1683 
1684 /*********************************************************************
1685  *
1686  *  Allocate DMA resources for TX buffers as well as memory for the TX
1687  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1688  *  iflib_sw_tx_desc_array structure, storing all the information that
1689  *  is needed to transmit a packet on the wire.  This is called only
1690  *  once at attach, setup is done every reset.
1691  *
1692  **********************************************************************/
1693 static int
1694 iflib_txsd_alloc(iflib_txq_t txq)
1695 {
1696 	if_ctx_t ctx = txq->ift_ctx;
1697 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1698 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1699 	device_t dev = ctx->ifc_dev;
1700 	bus_size_t tsomaxsize;
1701 	bus_addr_t lowaddr;
1702 	int err, nsegments, ntsosegments;
1703 	bool tso;
1704 
1705 	nsegments = scctx->isc_tx_nsegments;
1706 	ntsosegments = scctx->isc_tx_tso_segments_max;
1707 	tsomaxsize = scctx->isc_tx_tso_size_max;
1708 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1709 		tsomaxsize += sizeof(struct ether_vlan_header);
1710 	MPASS(scctx->isc_ntxd[0] > 0);
1711 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1712 	MPASS(nsegments > 0);
1713 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1714 		MPASS(ntsosegments > 0);
1715 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1716 	}
1717 
1718 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1719 
1720 	/*
1721 	 * Set up DMA tags for TX buffers.
1722 	 */
1723 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1724 			       1, 0,			/* alignment, bounds */
1725 			       lowaddr,			/* lowaddr */
1726 			       BUS_SPACE_MAXADDR,	/* highaddr */
1727 			       NULL, NULL,		/* filter, filterarg */
1728 			       sctx->isc_tx_maxsize,		/* maxsize */
1729 			       nsegments,	/* nsegments */
1730 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1731 			       0,			/* flags */
1732 			       NULL,			/* lockfunc */
1733 			       NULL,			/* lockfuncarg */
1734 			       &txq->ift_buf_tag))) {
1735 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1736 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1737 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1738 		goto fail;
1739 	}
1740 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1741 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1742 			       1, 0,			/* alignment, bounds */
1743 			       lowaddr,			/* lowaddr */
1744 			       BUS_SPACE_MAXADDR,	/* highaddr */
1745 			       NULL, NULL,		/* filter, filterarg */
1746 			       tsomaxsize,		/* maxsize */
1747 			       ntsosegments,	/* nsegments */
1748 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1749 			       0,			/* flags */
1750 			       NULL,			/* lockfunc */
1751 			       NULL,			/* lockfuncarg */
1752 			       &txq->ift_tso_buf_tag))) {
1753 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1754 		    err);
1755 		goto fail;
1756 	}
1757 
1758 	/* Allocate memory for the TX mbuf map. */
1759 	if (!(txq->ift_sds.ifsd_m =
1760 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1761 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1762 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1763 		err = ENOMEM;
1764 		goto fail;
1765 	}
1766 
1767 	/*
1768 	 * Create the DMA maps for TX buffers.
1769 	 */
1770 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1771 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1772 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1773 		device_printf(dev,
1774 		    "Unable to allocate TX buffer DMA map memory\n");
1775 		err = ENOMEM;
1776 		goto fail;
1777 	}
1778 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1779 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1780 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1781 		device_printf(dev,
1782 		    "Unable to allocate TSO TX buffer map memory\n");
1783 		err = ENOMEM;
1784 		goto fail;
1785 	}
1786 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1787 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1788 		    &txq->ift_sds.ifsd_map[i]);
1789 		if (err != 0) {
1790 			device_printf(dev, "Unable to create TX DMA map\n");
1791 			goto fail;
1792 		}
1793 		if (!tso)
1794 			continue;
1795 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1796 		    &txq->ift_sds.ifsd_tso_map[i]);
1797 		if (err != 0) {
1798 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1799 			goto fail;
1800 		}
1801 	}
1802 	return (0);
1803 fail:
1804 	/* We free all, it handles case where we are in the middle */
1805 	iflib_tx_structures_free(ctx);
1806 	return (err);
1807 }
1808 
1809 static void
1810 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1811 {
1812 	bus_dmamap_t map;
1813 
1814 	if (txq->ift_sds.ifsd_map != NULL) {
1815 		map = txq->ift_sds.ifsd_map[i];
1816 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1817 		bus_dmamap_unload(txq->ift_buf_tag, map);
1818 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1819 		txq->ift_sds.ifsd_map[i] = NULL;
1820 	}
1821 
1822 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1823 		map = txq->ift_sds.ifsd_tso_map[i];
1824 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1825 		    BUS_DMASYNC_POSTWRITE);
1826 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1827 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1828 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1829 	}
1830 }
1831 
1832 static void
1833 iflib_txq_destroy(iflib_txq_t txq)
1834 {
1835 	if_ctx_t ctx = txq->ift_ctx;
1836 
1837 	for (int i = 0; i < txq->ift_size; i++)
1838 		iflib_txsd_destroy(ctx, txq, i);
1839 
1840 	if (txq->ift_br != NULL) {
1841 		ifmp_ring_free(txq->ift_br);
1842 		txq->ift_br = NULL;
1843 	}
1844 
1845 	mtx_destroy(&txq->ift_mtx);
1846 
1847 	if (txq->ift_sds.ifsd_map != NULL) {
1848 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1849 		txq->ift_sds.ifsd_map = NULL;
1850 	}
1851 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1852 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1853 		txq->ift_sds.ifsd_tso_map = NULL;
1854 	}
1855 	if (txq->ift_sds.ifsd_m != NULL) {
1856 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1857 		txq->ift_sds.ifsd_m = NULL;
1858 	}
1859 	if (txq->ift_buf_tag != NULL) {
1860 		bus_dma_tag_destroy(txq->ift_buf_tag);
1861 		txq->ift_buf_tag = NULL;
1862 	}
1863 	if (txq->ift_tso_buf_tag != NULL) {
1864 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1865 		txq->ift_tso_buf_tag = NULL;
1866 	}
1867 	if (txq->ift_ifdi != NULL) {
1868 		free(txq->ift_ifdi, M_IFLIB);
1869 	}
1870 }
1871 
1872 static void
1873 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1874 {
1875 	struct mbuf **mp;
1876 
1877 	mp = &txq->ift_sds.ifsd_m[i];
1878 	if (*mp == NULL)
1879 		return;
1880 
1881 	if (txq->ift_sds.ifsd_map != NULL) {
1882 		bus_dmamap_sync(txq->ift_buf_tag,
1883 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1884 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1885 	}
1886 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1887 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1888 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1889 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1890 		    txq->ift_sds.ifsd_tso_map[i]);
1891 	}
1892 	m_freem(*mp);
1893 	DBG_COUNTER_INC(tx_frees);
1894 	*mp = NULL;
1895 }
1896 
1897 static int
1898 iflib_txq_setup(iflib_txq_t txq)
1899 {
1900 	if_ctx_t ctx = txq->ift_ctx;
1901 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1902 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1903 	iflib_dma_info_t di;
1904 	int i;
1905 
1906 	/* Set number of descriptors available */
1907 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1908 	/* XXX make configurable */
1909 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1910 
1911 	/* Reset indices */
1912 	txq->ift_cidx_processed = 0;
1913 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1914 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1915 
1916 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1917 		bzero((void *)di->idi_vaddr, di->idi_size);
1918 
1919 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1920 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1921 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1922 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1923 	return (0);
1924 }
1925 
1926 /*********************************************************************
1927  *
1928  *  Allocate DMA resources for RX buffers as well as memory for the RX
1929  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1930  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1931  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1932  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1933  *  packet, the maximum number of entries we'll need is equal to the
1934  *  number of hardware receive descriptors that we've allocated.
1935  *
1936  **********************************************************************/
1937 static int
1938 iflib_rxsd_alloc(iflib_rxq_t rxq)
1939 {
1940 	if_ctx_t ctx = rxq->ifr_ctx;
1941 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1942 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1943 	device_t dev = ctx->ifc_dev;
1944 	iflib_fl_t fl;
1945 	bus_addr_t lowaddr;
1946 	int			err;
1947 
1948 	MPASS(scctx->isc_nrxd[0] > 0);
1949 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1950 
1951 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1952 
1953 	fl = rxq->ifr_fl;
1954 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1955 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1956 		/* Set up DMA tag for RX buffers. */
1957 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1958 					 1, 0,			/* alignment, bounds */
1959 					 lowaddr,		/* lowaddr */
1960 					 BUS_SPACE_MAXADDR,	/* highaddr */
1961 					 NULL, NULL,		/* filter, filterarg */
1962 					 sctx->isc_rx_maxsize,	/* maxsize */
1963 					 sctx->isc_rx_nsegments,	/* nsegments */
1964 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1965 					 0,			/* flags */
1966 					 NULL,			/* lockfunc */
1967 					 NULL,			/* lockarg */
1968 					 &fl->ifl_buf_tag);
1969 		if (err) {
1970 			device_printf(dev,
1971 			    "Unable to allocate RX DMA tag: %d\n", err);
1972 			goto fail;
1973 		}
1974 
1975 		/* Allocate memory for the RX mbuf map. */
1976 		if (!(fl->ifl_sds.ifsd_m =
1977 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1978 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1979 			device_printf(dev,
1980 			    "Unable to allocate RX mbuf map memory\n");
1981 			err = ENOMEM;
1982 			goto fail;
1983 		}
1984 
1985 		/* Allocate memory for the direct RX cluster pointer map. */
1986 		if (!(fl->ifl_sds.ifsd_cl =
1987 		      (caddr_t *) malloc(sizeof(caddr_t) *
1988 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1989 			device_printf(dev,
1990 			    "Unable to allocate RX cluster map memory\n");
1991 			err = ENOMEM;
1992 			goto fail;
1993 		}
1994 
1995 		/* Allocate memory for the RX cluster bus address map. */
1996 		if (!(fl->ifl_sds.ifsd_ba =
1997 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1998 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1999 			device_printf(dev,
2000 			    "Unable to allocate RX bus address map memory\n");
2001 			err = ENOMEM;
2002 			goto fail;
2003 		}
2004 
2005 		/*
2006 		 * Create the DMA maps for RX buffers.
2007 		 */
2008 		if (!(fl->ifl_sds.ifsd_map =
2009 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2010 			device_printf(dev,
2011 			    "Unable to allocate RX buffer DMA map memory\n");
2012 			err = ENOMEM;
2013 			goto fail;
2014 		}
2015 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2016 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2017 			    &fl->ifl_sds.ifsd_map[i]);
2018 			if (err != 0) {
2019 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2020 				goto fail;
2021 			}
2022 		}
2023 	}
2024 	return (0);
2025 
2026 fail:
2027 	iflib_rx_structures_free(ctx);
2028 	return (err);
2029 }
2030 
2031 /*
2032  * Internal service routines
2033  */
2034 
2035 struct rxq_refill_cb_arg {
2036 	int               error;
2037 	bus_dma_segment_t seg;
2038 	int               nseg;
2039 };
2040 
2041 static void
2042 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2043 {
2044 	struct rxq_refill_cb_arg *cb_arg = arg;
2045 
2046 	cb_arg->error = error;
2047 	cb_arg->seg = segs[0];
2048 	cb_arg->nseg = nseg;
2049 }
2050 
2051 /**
2052  * iflib_fl_refill - refill an rxq free-buffer list
2053  * @ctx: the iflib context
2054  * @fl: the free list to refill
2055  * @count: the number of new buffers to allocate
2056  *
2057  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2058  * The caller must assure that @count does not exceed the queue's capacity
2059  * minus one (since we always leave a descriptor unavailable).
2060  */
2061 static uint8_t
2062 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2063 {
2064 	struct if_rxd_update iru;
2065 	struct rxq_refill_cb_arg cb_arg;
2066 	struct mbuf *m;
2067 	caddr_t cl, *sd_cl;
2068 	struct mbuf **sd_m;
2069 	bus_dmamap_t *sd_map;
2070 	bus_addr_t bus_addr, *sd_ba;
2071 	int err, frag_idx, i, idx, n, pidx;
2072 	qidx_t credits;
2073 
2074 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2075 
2076 	sd_m = fl->ifl_sds.ifsd_m;
2077 	sd_map = fl->ifl_sds.ifsd_map;
2078 	sd_cl = fl->ifl_sds.ifsd_cl;
2079 	sd_ba = fl->ifl_sds.ifsd_ba;
2080 	pidx = fl->ifl_pidx;
2081 	idx = pidx;
2082 	frag_idx = fl->ifl_fragidx;
2083 	credits = fl->ifl_credits;
2084 
2085 	i = 0;
2086 	n = count;
2087 	MPASS(n > 0);
2088 	MPASS(credits + n <= fl->ifl_size);
2089 
2090 	if (pidx < fl->ifl_cidx)
2091 		MPASS(pidx + n <= fl->ifl_cidx);
2092 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2093 		MPASS(fl->ifl_gen == 0);
2094 	if (pidx > fl->ifl_cidx)
2095 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2096 
2097 	DBG_COUNTER_INC(fl_refills);
2098 	if (n > 8)
2099 		DBG_COUNTER_INC(fl_refills_large);
2100 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2101 	while (n-- > 0) {
2102 		/*
2103 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2104 		 * initialized after rx.
2105 		 *
2106 		 * If the cluster is still set then we know a minimum sized
2107 		 * packet was received
2108 		 */
2109 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2110 		    &frag_idx);
2111 		if (frag_idx < 0)
2112 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2113 		MPASS(frag_idx >= 0);
2114 		if ((cl = sd_cl[frag_idx]) == NULL) {
2115 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2116 			if (__predict_false(cl == NULL))
2117 				break;
2118 
2119 			cb_arg.error = 0;
2120 			MPASS(sd_map != NULL);
2121 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2122 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2123 			    BUS_DMA_NOWAIT);
2124 			if (__predict_false(err != 0 || cb_arg.error)) {
2125 				uma_zfree(fl->ifl_zone, cl);
2126 				break;
2127 			}
2128 
2129 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2130 			sd_cl[frag_idx] = cl;
2131 #if MEMORY_LOGGING
2132 			fl->ifl_cl_enqueued++;
2133 #endif
2134 		} else {
2135 			bus_addr = sd_ba[frag_idx];
2136 		}
2137 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2138 		    BUS_DMASYNC_PREREAD);
2139 
2140 		if (sd_m[frag_idx] == NULL) {
2141 			m = m_gethdr_raw(M_NOWAIT, 0);
2142 			if (__predict_false(m == NULL))
2143 				break;
2144 			sd_m[frag_idx] = m;
2145 		}
2146 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2147 #if MEMORY_LOGGING
2148 		fl->ifl_m_enqueued++;
2149 #endif
2150 
2151 		DBG_COUNTER_INC(rx_allocs);
2152 		fl->ifl_rxd_idxs[i] = frag_idx;
2153 		fl->ifl_bus_addrs[i] = bus_addr;
2154 		credits++;
2155 		i++;
2156 		MPASS(credits <= fl->ifl_size);
2157 		if (++idx == fl->ifl_size) {
2158 #ifdef INVARIANTS
2159 			fl->ifl_gen = 1;
2160 #endif
2161 			idx = 0;
2162 		}
2163 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2164 			iru.iru_pidx = pidx;
2165 			iru.iru_count = i;
2166 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2167 			fl->ifl_pidx = idx;
2168 			fl->ifl_credits = credits;
2169 			pidx = idx;
2170 			i = 0;
2171 		}
2172 	}
2173 
2174 	if (n < count - 1) {
2175 		if (i != 0) {
2176 			iru.iru_pidx = pidx;
2177 			iru.iru_count = i;
2178 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2179 			fl->ifl_pidx = idx;
2180 			fl->ifl_credits = credits;
2181 		}
2182 		DBG_COUNTER_INC(rxd_flush);
2183 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2184 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2185 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2186 		    fl->ifl_id, fl->ifl_pidx);
2187 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2188 			fl->ifl_fragidx = frag_idx + 1;
2189 			if (fl->ifl_fragidx == fl->ifl_size)
2190 				fl->ifl_fragidx = 0;
2191 		} else {
2192 			fl->ifl_fragidx = frag_idx;
2193 		}
2194 	}
2195 
2196 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2197 }
2198 
2199 static inline uint8_t
2200 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2201 {
2202 	/*
2203 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2204 	 * This is important as it confuses most NICs. For instance,
2205 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2206 	 * RDH points to the next receive descriptor to be used by the NIC,
2207 	 * and RDT for the next receive descriptor to be published by the
2208 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2209 	 * The condition RDH == RDT means no descriptors are available to
2210 	 * the NIC, and thus it would be ambiguous if it also meant that
2211 	 * all the descriptors are available to the NIC.
2212 	 */
2213 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2214 #ifdef INVARIANTS
2215 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2216 #endif
2217 
2218 	MPASS(fl->ifl_credits <= fl->ifl_size);
2219 	MPASS(reclaimable == delta);
2220 
2221 	if (reclaimable > 0)
2222 		return (iflib_fl_refill(ctx, fl, reclaimable));
2223 	return (0);
2224 }
2225 
2226 uint8_t
2227 iflib_in_detach(if_ctx_t ctx)
2228 {
2229 	bool in_detach;
2230 
2231 	STATE_LOCK(ctx);
2232 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2233 	STATE_UNLOCK(ctx);
2234 	return (in_detach);
2235 }
2236 
2237 static void
2238 iflib_fl_bufs_free(iflib_fl_t fl)
2239 {
2240 	iflib_dma_info_t idi = fl->ifl_ifdi;
2241 	bus_dmamap_t sd_map;
2242 	uint32_t i;
2243 
2244 	for (i = 0; i < fl->ifl_size; i++) {
2245 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2246 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2247 
2248 		if (*sd_cl != NULL) {
2249 			sd_map = fl->ifl_sds.ifsd_map[i];
2250 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2251 			    BUS_DMASYNC_POSTREAD);
2252 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2253 			uma_zfree(fl->ifl_zone, *sd_cl);
2254 			*sd_cl = NULL;
2255 			if (*sd_m != NULL) {
2256 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2257 				m_free_raw(*sd_m);
2258 				*sd_m = NULL;
2259 			}
2260 		} else {
2261 			MPASS(*sd_m == NULL);
2262 		}
2263 #if MEMORY_LOGGING
2264 		fl->ifl_m_dequeued++;
2265 		fl->ifl_cl_dequeued++;
2266 #endif
2267 	}
2268 #ifdef INVARIANTS
2269 	for (i = 0; i < fl->ifl_size; i++) {
2270 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2271 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2272 	}
2273 #endif
2274 	/*
2275 	 * Reset free list values
2276 	 */
2277 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2278 	bzero(idi->idi_vaddr, idi->idi_size);
2279 }
2280 
2281 /*********************************************************************
2282  *
2283  *  Initialize a free list and its buffers.
2284  *
2285  **********************************************************************/
2286 static int
2287 iflib_fl_setup(iflib_fl_t fl)
2288 {
2289 	iflib_rxq_t rxq = fl->ifl_rxq;
2290 	if_ctx_t ctx = rxq->ifr_ctx;
2291 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2292 	int qidx;
2293 
2294 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2295 	/*
2296 	** Free current RX buffer structs and their mbufs
2297 	*/
2298 	iflib_fl_bufs_free(fl);
2299 	/* Now replenish the mbufs */
2300 	MPASS(fl->ifl_credits == 0);
2301 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2302 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2303 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2304 	else
2305 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2306 	/*
2307 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2308 	 * to the selected mbuf size.
2309 	 */
2310 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2311 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2312 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2313 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2314 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2315 
2316 	/*
2317 	 * Avoid pre-allocating zillions of clusters to an idle card
2318 	 * potentially speeding up attach. In any case make sure
2319 	 * to leave a descriptor unavailable. See the comment in
2320 	 * iflib_fl_refill_all().
2321 	 */
2322 	MPASS(fl->ifl_size > 0);
2323 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2324 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2325 		return (ENOBUFS);
2326 	/*
2327 	 * handle failure
2328 	 */
2329 	MPASS(rxq != NULL);
2330 	MPASS(fl->ifl_ifdi != NULL);
2331 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2332 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2333 	return (0);
2334 }
2335 
2336 /*********************************************************************
2337  *
2338  *  Free receive ring data structures
2339  *
2340  **********************************************************************/
2341 static void
2342 iflib_rx_sds_free(iflib_rxq_t rxq)
2343 {
2344 	iflib_fl_t fl;
2345 	int i, j;
2346 
2347 	if (rxq->ifr_fl != NULL) {
2348 		for (i = 0; i < rxq->ifr_nfl; i++) {
2349 			fl = &rxq->ifr_fl[i];
2350 			if (fl->ifl_buf_tag != NULL) {
2351 				if (fl->ifl_sds.ifsd_map != NULL) {
2352 					for (j = 0; j < fl->ifl_size; j++) {
2353 						bus_dmamap_sync(
2354 						    fl->ifl_buf_tag,
2355 						    fl->ifl_sds.ifsd_map[j],
2356 						    BUS_DMASYNC_POSTREAD);
2357 						bus_dmamap_unload(
2358 						    fl->ifl_buf_tag,
2359 						    fl->ifl_sds.ifsd_map[j]);
2360 						bus_dmamap_destroy(
2361 						    fl->ifl_buf_tag,
2362 						    fl->ifl_sds.ifsd_map[j]);
2363 					}
2364 				}
2365 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2366 				fl->ifl_buf_tag = NULL;
2367 			}
2368 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2369 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2370 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2371 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2372 			free(fl->ifl_rx_bitmap, M_IFLIB);
2373 			fl->ifl_sds.ifsd_m = NULL;
2374 			fl->ifl_sds.ifsd_cl = NULL;
2375 			fl->ifl_sds.ifsd_ba = NULL;
2376 			fl->ifl_sds.ifsd_map = NULL;
2377 			fl->ifl_rx_bitmap = NULL;
2378 		}
2379 		free(rxq->ifr_fl, M_IFLIB);
2380 		rxq->ifr_fl = NULL;
2381 		free(rxq->ifr_ifdi, M_IFLIB);
2382 		rxq->ifr_ifdi = NULL;
2383 		rxq->ifr_cq_cidx = 0;
2384 	}
2385 }
2386 
2387 /*
2388  * Timer routine
2389  */
2390 static void
2391 iflib_timer(void *arg)
2392 {
2393 	iflib_txq_t txq = arg;
2394 	if_ctx_t ctx = txq->ift_ctx;
2395 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2396 	uint64_t this_tick = ticks;
2397 
2398 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2399 		return;
2400 
2401 	/*
2402 	** Check on the state of the TX queue(s), this
2403 	** can be done without the lock because its RO
2404 	** and the HUNG state will be static if set.
2405 	*/
2406 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2407 		txq->ift_last_timer_tick = this_tick;
2408 		IFDI_TIMER(ctx, txq->ift_id);
2409 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2410 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2411 		     (sctx->isc_pause_frames == 0)))
2412 			goto hung;
2413 
2414 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2415 		    ifmp_ring_is_stalled(txq->ift_br)) {
2416 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2417 			    ("queue can't be marked as hung if interface is down"));
2418 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2419 		}
2420 		txq->ift_cleaned_prev = txq->ift_cleaned;
2421 	}
2422 	/* handle any laggards */
2423 	if (txq->ift_db_pending)
2424 		GROUPTASK_ENQUEUE(&txq->ift_task);
2425 
2426 	sctx->isc_pause_frames = 0;
2427 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2428 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2429 		    txq, txq->ift_timer.c_cpu);
2430 	return;
2431 
2432  hung:
2433 	device_printf(ctx->ifc_dev,
2434 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2435 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2436 	STATE_LOCK(ctx);
2437 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2438 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2439 	iflib_admin_intr_deferred(ctx);
2440 	STATE_UNLOCK(ctx);
2441 }
2442 
2443 static uint16_t
2444 iflib_get_mbuf_size_for(unsigned int size)
2445 {
2446 
2447 	if (size <= MCLBYTES)
2448 		return (MCLBYTES);
2449 	else
2450 		return (MJUMPAGESIZE);
2451 }
2452 
2453 static void
2454 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2455 {
2456 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2457 
2458 	/*
2459 	 * XXX don't set the max_frame_size to larger
2460 	 * than the hardware can handle
2461 	 */
2462 	ctx->ifc_rx_mbuf_sz =
2463 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2464 }
2465 
2466 uint32_t
2467 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2468 {
2469 
2470 	return (ctx->ifc_rx_mbuf_sz);
2471 }
2472 
2473 static void
2474 iflib_init_locked(if_ctx_t ctx)
2475 {
2476 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2477 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2478 	if_t ifp = ctx->ifc_ifp;
2479 	iflib_fl_t fl;
2480 	iflib_txq_t txq;
2481 	iflib_rxq_t rxq;
2482 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2483 
2484 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2485 	IFDI_INTR_DISABLE(ctx);
2486 
2487 	/*
2488 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2489 	 * called without first calling iflib_stop().
2490 	 */
2491 	netmap_disable_all_rings(ifp);
2492 
2493 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2494 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2495 	/* Set hardware offload abilities */
2496 	if_clearhwassist(ifp);
2497 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2498 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2499 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2500 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2501 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2502 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2503 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2504 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2505 
2506 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2507 		CALLOUT_LOCK(txq);
2508 		callout_stop(&txq->ift_timer);
2509 #ifdef DEV_NETMAP
2510 		callout_stop(&txq->ift_netmap_timer);
2511 #endif /* DEV_NETMAP */
2512 		CALLOUT_UNLOCK(txq);
2513 		(void)iflib_netmap_txq_init(ctx, txq);
2514 	}
2515 
2516 	/*
2517 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2518 	 * that drivers can use the value when setting up the hardware receive
2519 	 * buffers.
2520 	 */
2521 	iflib_calc_rx_mbuf_sz(ctx);
2522 
2523 #ifdef INVARIANTS
2524 	i = if_getdrvflags(ifp);
2525 #endif
2526 	IFDI_INIT(ctx);
2527 	MPASS(if_getdrvflags(ifp) == i);
2528 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2529 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2530 			/* This rxq is in netmap mode. Skip normal init. */
2531 			continue;
2532 		}
2533 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2534 			if (iflib_fl_setup(fl)) {
2535 				device_printf(ctx->ifc_dev,
2536 				    "setting up free list %d failed - "
2537 				    "check cluster settings\n", j);
2538 				goto done;
2539 			}
2540 		}
2541 	}
2542 done:
2543 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2544 	IFDI_INTR_ENABLE(ctx);
2545 	txq = ctx->ifc_txqs;
2546 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2547 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2548 			txq->ift_timer.c_cpu);
2549 
2550         /* Re-enable txsync/rxsync. */
2551 	netmap_enable_all_rings(ifp);
2552 }
2553 
2554 static int
2555 iflib_media_change(if_t ifp)
2556 {
2557 	if_ctx_t ctx = if_getsoftc(ifp);
2558 	int err;
2559 
2560 	CTX_LOCK(ctx);
2561 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2562 		iflib_if_init_locked(ctx);
2563 	CTX_UNLOCK(ctx);
2564 	return (err);
2565 }
2566 
2567 static void
2568 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2569 {
2570 	if_ctx_t ctx = if_getsoftc(ifp);
2571 
2572 	CTX_LOCK(ctx);
2573 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2574 	IFDI_MEDIA_STATUS(ctx, ifmr);
2575 	CTX_UNLOCK(ctx);
2576 }
2577 
2578 void
2579 iflib_stop(if_ctx_t ctx)
2580 {
2581 	iflib_txq_t txq = ctx->ifc_txqs;
2582 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2583 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2584 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2585 	iflib_dma_info_t di;
2586 	iflib_fl_t fl;
2587 	int i, j;
2588 
2589 	/* Tell the stack that the interface is no longer active */
2590 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2591 
2592 	IFDI_INTR_DISABLE(ctx);
2593 	DELAY(1000);
2594 	IFDI_STOP(ctx);
2595 	DELAY(1000);
2596 
2597 	/*
2598 	 * Stop any pending txsync/rxsync and prevent new ones
2599 	 * form starting. Processes blocked in poll() will get
2600 	 * POLLERR.
2601 	 */
2602 	netmap_disable_all_rings(ctx->ifc_ifp);
2603 
2604 	iflib_debug_reset();
2605 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2606 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2607 		/* make sure all transmitters have completed before proceeding XXX */
2608 
2609 		CALLOUT_LOCK(txq);
2610 		callout_stop(&txq->ift_timer);
2611 #ifdef DEV_NETMAP
2612 		callout_stop(&txq->ift_netmap_timer);
2613 #endif /* DEV_NETMAP */
2614 		CALLOUT_UNLOCK(txq);
2615 
2616 		/* clean any enqueued buffers */
2617 		iflib_ifmp_purge(txq);
2618 		/* Free any existing tx buffers. */
2619 		for (j = 0; j < txq->ift_size; j++) {
2620 			iflib_txsd_free(ctx, txq, j);
2621 		}
2622 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2623 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2624 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2625 			txq->ift_cidx = txq->ift_pidx;
2626 		else
2627 			txq->ift_cidx = txq->ift_pidx = 0;
2628 
2629 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2630 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2631 		txq->ift_pullups = 0;
2632 		ifmp_ring_reset_stats(txq->ift_br);
2633 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2634 			bzero((void *)di->idi_vaddr, di->idi_size);
2635 	}
2636 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2637 		gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2638 		    &rxq->ifr_task.gt_task);
2639 
2640 		rxq->ifr_cq_cidx = 0;
2641 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2642 			bzero((void *)di->idi_vaddr, di->idi_size);
2643 		/* also resets the free lists pidx/cidx */
2644 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2645 			iflib_fl_bufs_free(fl);
2646 	}
2647 }
2648 
2649 static inline caddr_t
2650 calc_next_rxd(iflib_fl_t fl, int cidx)
2651 {
2652 	qidx_t size;
2653 	int nrxd;
2654 	caddr_t start, end, cur, next;
2655 
2656 	nrxd = fl->ifl_size;
2657 	size = fl->ifl_rxd_size;
2658 	start = fl->ifl_ifdi->idi_vaddr;
2659 
2660 	if (__predict_false(size == 0))
2661 		return (start);
2662 	cur = start + size*cidx;
2663 	end = start + size*nrxd;
2664 	next = CACHE_PTR_NEXT(cur);
2665 	return (next < end ? next : start);
2666 }
2667 
2668 static inline void
2669 prefetch_pkts(iflib_fl_t fl, int cidx)
2670 {
2671 	int nextptr;
2672 	int nrxd = fl->ifl_size;
2673 	caddr_t next_rxd;
2674 
2675 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2676 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2677 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2678 	next_rxd = calc_next_rxd(fl, cidx);
2679 	prefetch(next_rxd);
2680 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2681 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2682 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2683 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2684 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2685 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2686 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2687 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2688 }
2689 
2690 static struct mbuf *
2691 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2692     int *pf_rv, if_rxd_info_t ri)
2693 {
2694 	bus_dmamap_t map;
2695 	iflib_fl_t fl;
2696 	caddr_t payload;
2697 	struct mbuf *m;
2698 	int flid, cidx, len, next;
2699 
2700 	map = NULL;
2701 	flid = irf->irf_flid;
2702 	cidx = irf->irf_idx;
2703 	fl = &rxq->ifr_fl[flid];
2704 	sd->ifsd_fl = fl;
2705 	m = fl->ifl_sds.ifsd_m[cidx];
2706 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2707 	fl->ifl_credits--;
2708 #if MEMORY_LOGGING
2709 	fl->ifl_m_dequeued++;
2710 #endif
2711 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2712 		prefetch_pkts(fl, cidx);
2713 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2714 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2715 	map = fl->ifl_sds.ifsd_map[cidx];
2716 
2717 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2718 
2719 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2720 	    irf->irf_len != 0) {
2721 		payload  = *sd->ifsd_cl;
2722 		payload +=  ri->iri_pad;
2723 		len = ri->iri_len - ri->iri_pad;
2724 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2725 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2726 		switch (*pf_rv) {
2727 		case PFIL_DROPPED:
2728 		case PFIL_CONSUMED:
2729 			/*
2730 			 * The filter ate it.  Everything is recycled.
2731 			 */
2732 			m = NULL;
2733 			unload = 0;
2734 			break;
2735 		case PFIL_REALLOCED:
2736 			/*
2737 			 * The filter copied it.  Everything is recycled.
2738 			 */
2739 			m = pfil_mem2mbuf(payload);
2740 			unload = 0;
2741 			break;
2742 		case PFIL_PASS:
2743 			/*
2744 			 * Filter said it was OK, so receive like
2745 			 * normal
2746 			 */
2747 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2748 			break;
2749 		default:
2750 			MPASS(0);
2751 		}
2752 	} else {
2753 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2754 		if (pf_rv != NULL)
2755 			*pf_rv = PFIL_PASS;
2756 	}
2757 
2758 	if (unload && irf->irf_len != 0)
2759 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2760 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2761 	if (__predict_false(fl->ifl_cidx == 0))
2762 		fl->ifl_gen = 0;
2763 	bit_clear(fl->ifl_rx_bitmap, cidx);
2764 	return (m);
2765 }
2766 
2767 static struct mbuf *
2768 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2769 {
2770 	struct mbuf *m, *mh, *mt;
2771 	caddr_t cl;
2772 	int  *pf_rv_ptr, flags, i, padlen;
2773 	bool consumed;
2774 
2775 	i = 0;
2776 	mh = NULL;
2777 	consumed = false;
2778 	*pf_rv = PFIL_PASS;
2779 	pf_rv_ptr = pf_rv;
2780 	do {
2781 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2782 		    pf_rv_ptr, ri);
2783 
2784 		MPASS(*sd->ifsd_cl != NULL);
2785 
2786 		/*
2787 		 * Exclude zero-length frags & frags from
2788 		 * packets the filter has consumed or dropped
2789 		 */
2790 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2791 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2792 			if (mh == NULL) {
2793 				/* everything saved here */
2794 				consumed = true;
2795 				pf_rv_ptr = NULL;
2796 				continue;
2797 			}
2798 			/* XXX we can save the cluster here, but not the mbuf */
2799 			m_init(m, M_NOWAIT, MT_DATA, 0);
2800 			m_free(m);
2801 			continue;
2802 		}
2803 		if (mh == NULL) {
2804 			flags = M_PKTHDR|M_EXT;
2805 			mh = mt = m;
2806 			padlen = ri->iri_pad;
2807 		} else {
2808 			flags = M_EXT;
2809 			mt->m_next = m;
2810 			mt = m;
2811 			/* assuming padding is only on the first fragment */
2812 			padlen = 0;
2813 		}
2814 		cl = *sd->ifsd_cl;
2815 		*sd->ifsd_cl = NULL;
2816 
2817 		/* Can these two be made one ? */
2818 		m_init(m, M_NOWAIT, MT_DATA, flags);
2819 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2820 		/*
2821 		 * These must follow m_init and m_cljset
2822 		 */
2823 		m->m_data += padlen;
2824 		ri->iri_len -= padlen;
2825 		m->m_len = ri->iri_frags[i].irf_len;
2826 	} while (++i < ri->iri_nfrags);
2827 
2828 	return (mh);
2829 }
2830 
2831 /*
2832  * Process one software descriptor
2833  */
2834 static struct mbuf *
2835 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2836 {
2837 	struct if_rxsd sd;
2838 	struct mbuf *m;
2839 	int pf_rv;
2840 
2841 	/* should I merge this back in now that the two paths are basically duplicated? */
2842 	if (ri->iri_nfrags == 1 &&
2843 	    ri->iri_frags[0].irf_len != 0 &&
2844 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2845 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2846 		    &pf_rv, ri);
2847 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2848 			return (m);
2849 		if (pf_rv == PFIL_PASS) {
2850 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2851 #ifndef __NO_STRICT_ALIGNMENT
2852 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2853 				m->m_data += 2;
2854 #endif
2855 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2856 			m->m_len = ri->iri_frags[0].irf_len;
2857 			m->m_data += ri->iri_pad;
2858 			ri->iri_len -= ri->iri_pad;
2859 		}
2860 	} else {
2861 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2862 		if (m == NULL)
2863 			return (NULL);
2864 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2865 			return (m);
2866 	}
2867 	m->m_pkthdr.len = ri->iri_len;
2868 	m->m_pkthdr.rcvif = ri->iri_ifp;
2869 	m->m_flags |= ri->iri_flags;
2870 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2871 	m->m_pkthdr.flowid = ri->iri_flowid;
2872 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2873 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2874 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2875 	return (m);
2876 }
2877 
2878 #if defined(INET6) || defined(INET)
2879 static void
2880 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2881 {
2882 	CURVNET_SET(lc->ifp->if_vnet);
2883 #if defined(INET6)
2884 	*v6 = V_ip6_forwarding;
2885 #endif
2886 #if defined(INET)
2887 	*v4 = V_ipforwarding;
2888 #endif
2889 	CURVNET_RESTORE();
2890 }
2891 
2892 /*
2893  * Returns true if it's possible this packet could be LROed.
2894  * if it returns false, it is guaranteed that tcp_lro_rx()
2895  * would not return zero.
2896  */
2897 static bool
2898 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2899 {
2900 	struct ether_header *eh;
2901 
2902 	eh = mtod(m, struct ether_header *);
2903 	switch (eh->ether_type) {
2904 #if defined(INET6)
2905 		case htons(ETHERTYPE_IPV6):
2906 			return (!v6_forwarding);
2907 #endif
2908 #if defined (INET)
2909 		case htons(ETHERTYPE_IP):
2910 			return (!v4_forwarding);
2911 #endif
2912 	}
2913 
2914 	return false;
2915 }
2916 #else
2917 static void
2918 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2919 {
2920 }
2921 #endif
2922 
2923 static void
2924 _task_fn_rx_watchdog(void *context)
2925 {
2926 	iflib_rxq_t rxq = context;
2927 
2928 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2929 }
2930 
2931 static uint8_t
2932 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2933 {
2934 	if_t ifp;
2935 	if_ctx_t ctx = rxq->ifr_ctx;
2936 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2937 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2938 	int avail, i;
2939 	qidx_t *cidxp;
2940 	struct if_rxd_info ri;
2941 	int err, budget_left, rx_bytes, rx_pkts;
2942 	iflib_fl_t fl;
2943 	int lro_enabled;
2944 	bool v4_forwarding, v6_forwarding, lro_possible;
2945 	uint8_t retval = 0;
2946 
2947 	/*
2948 	 * XXX early demux data packets so that if_input processing only handles
2949 	 * acks in interrupt context
2950 	 */
2951 	struct mbuf *m, *mh, *mt, *mf;
2952 
2953 	NET_EPOCH_ASSERT();
2954 
2955 	lro_possible = v4_forwarding = v6_forwarding = false;
2956 	ifp = ctx->ifc_ifp;
2957 	mh = mt = NULL;
2958 	MPASS(budget > 0);
2959 	rx_pkts	= rx_bytes = 0;
2960 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2961 		cidxp = &rxq->ifr_cq_cidx;
2962 	else
2963 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2964 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2965 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2966 			retval |= iflib_fl_refill_all(ctx, fl);
2967 		DBG_COUNTER_INC(rx_unavail);
2968 		return (retval);
2969 	}
2970 
2971 	/* pfil needs the vnet to be set */
2972 	CURVNET_SET_QUIET(ifp->if_vnet);
2973 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2974 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2975 			DBG_COUNTER_INC(rx_ctx_inactive);
2976 			break;
2977 		}
2978 		/*
2979 		 * Reset client set fields to their default values
2980 		 */
2981 		rxd_info_zero(&ri);
2982 		ri.iri_qsidx = rxq->ifr_id;
2983 		ri.iri_cidx = *cidxp;
2984 		ri.iri_ifp = ifp;
2985 		ri.iri_frags = rxq->ifr_frags;
2986 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2987 
2988 		if (err)
2989 			goto err;
2990 		rx_pkts += 1;
2991 		rx_bytes += ri.iri_len;
2992 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2993 			*cidxp = ri.iri_cidx;
2994 			/* Update our consumer index */
2995 			/* XXX NB: shurd - check if this is still safe */
2996 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2997 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2998 			/* was this only a completion queue message? */
2999 			if (__predict_false(ri.iri_nfrags == 0))
3000 				continue;
3001 		}
3002 		MPASS(ri.iri_nfrags != 0);
3003 		MPASS(ri.iri_len != 0);
3004 
3005 		/* will advance the cidx on the corresponding free lists */
3006 		m = iflib_rxd_pkt_get(rxq, &ri);
3007 		avail--;
3008 		budget_left--;
3009 		if (avail == 0 && budget_left)
3010 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3011 
3012 		if (__predict_false(m == NULL))
3013 			continue;
3014 
3015 		/* imm_pkt: -- cxgb */
3016 		if (mh == NULL)
3017 			mh = mt = m;
3018 		else {
3019 			mt->m_nextpkt = m;
3020 			mt = m;
3021 		}
3022 	}
3023 	CURVNET_RESTORE();
3024 	/* make sure that we can refill faster than drain */
3025 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3026 		retval |= iflib_fl_refill_all(ctx, fl);
3027 
3028 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3029 	if (lro_enabled)
3030 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3031 	mt = mf = NULL;
3032 	while (mh != NULL) {
3033 		m = mh;
3034 		mh = mh->m_nextpkt;
3035 		m->m_nextpkt = NULL;
3036 #ifndef __NO_STRICT_ALIGNMENT
3037 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3038 			continue;
3039 #endif
3040 #if defined(INET6) || defined(INET)
3041 		if (lro_enabled) {
3042 			if (!lro_possible) {
3043 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3044 				if (lro_possible && mf != NULL) {
3045 					ifp->if_input(ifp, mf);
3046 					DBG_COUNTER_INC(rx_if_input);
3047 					mt = mf = NULL;
3048 				}
3049 			}
3050 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3051 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3052 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3053 					continue;
3054 			}
3055 		}
3056 #endif
3057 		if (lro_possible) {
3058 			ifp->if_input(ifp, m);
3059 			DBG_COUNTER_INC(rx_if_input);
3060 			continue;
3061 		}
3062 
3063 		if (mf == NULL)
3064 			mf = m;
3065 		if (mt != NULL)
3066 			mt->m_nextpkt = m;
3067 		mt = m;
3068 	}
3069 	if (mf != NULL) {
3070 		ifp->if_input(ifp, mf);
3071 		DBG_COUNTER_INC(rx_if_input);
3072 	}
3073 
3074 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3075 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3076 
3077 	/*
3078 	 * Flush any outstanding LRO work
3079 	 */
3080 #if defined(INET6) || defined(INET)
3081 	tcp_lro_flush_all(&rxq->ifr_lc);
3082 #endif
3083 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3084 		retval |= IFLIB_RXEOF_MORE;
3085 	return (retval);
3086 err:
3087 	STATE_LOCK(ctx);
3088 	ctx->ifc_flags |= IFC_DO_RESET;
3089 	iflib_admin_intr_deferred(ctx);
3090 	STATE_UNLOCK(ctx);
3091 	return (0);
3092 }
3093 
3094 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3095 static inline qidx_t
3096 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3097 {
3098 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3099 	qidx_t minthresh = txq->ift_size / 8;
3100 	if (in_use > 4*minthresh)
3101 		return (notify_count);
3102 	if (in_use > 2*minthresh)
3103 		return (notify_count >> 1);
3104 	if (in_use > minthresh)
3105 		return (notify_count >> 3);
3106 	return (0);
3107 }
3108 
3109 static inline qidx_t
3110 txq_max_rs_deferred(iflib_txq_t txq)
3111 {
3112 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3113 	qidx_t minthresh = txq->ift_size / 8;
3114 	if (txq->ift_in_use > 4*minthresh)
3115 		return (notify_count);
3116 	if (txq->ift_in_use > 2*minthresh)
3117 		return (notify_count >> 1);
3118 	if (txq->ift_in_use > minthresh)
3119 		return (notify_count >> 2);
3120 	return (2);
3121 }
3122 
3123 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3124 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3125 
3126 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3127 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3128 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3129 
3130 /* forward compatibility for cxgb */
3131 #define FIRST_QSET(ctx) 0
3132 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3133 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3134 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3135 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3136 
3137 /* XXX we should be setting this to something other than zero */
3138 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3139 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3140     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3141 
3142 static inline bool
3143 iflib_txd_db_check(iflib_txq_t txq, int ring)
3144 {
3145 	if_ctx_t ctx = txq->ift_ctx;
3146 	qidx_t dbval, max;
3147 
3148 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3149 
3150 	/* force || threshold exceeded || at the edge of the ring */
3151 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3152 
3153 		/*
3154 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3155 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3156 		 * producer index explicitly (INTC).
3157 		 */
3158 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3159 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3160 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3161 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3162 
3163 		/*
3164 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3165 		 */
3166 		txq->ift_db_pending = txq->ift_npending = 0;
3167 		return (true);
3168 	}
3169 	return (false);
3170 }
3171 
3172 #ifdef PKT_DEBUG
3173 static void
3174 print_pkt(if_pkt_info_t pi)
3175 {
3176 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3177 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3178 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3179 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3180 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3181 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3182 }
3183 #endif
3184 
3185 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3186 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3187 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3188 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3189 
3190 static int
3191 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3192 {
3193 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3194 	struct ether_vlan_header *eh;
3195 	struct mbuf *m;
3196 
3197 	m = *mp;
3198 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3199 	    M_WRITABLE(m) == 0) {
3200 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3201 			return (ENOMEM);
3202 		} else {
3203 			m_freem(*mp);
3204 			DBG_COUNTER_INC(tx_frees);
3205 			*mp = m;
3206 		}
3207 	}
3208 
3209 	/*
3210 	 * Determine where frame payload starts.
3211 	 * Jump over vlan headers if already present,
3212 	 * helpful for QinQ too.
3213 	 */
3214 	if (__predict_false(m->m_len < sizeof(*eh))) {
3215 		txq->ift_pullups++;
3216 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3217 			return (ENOMEM);
3218 	}
3219 	eh = mtod(m, struct ether_vlan_header *);
3220 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3221 		pi->ipi_etype = ntohs(eh->evl_proto);
3222 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3223 	} else {
3224 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3225 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3226 	}
3227 
3228 	switch (pi->ipi_etype) {
3229 #ifdef INET
3230 	case ETHERTYPE_IP:
3231 	{
3232 		struct mbuf *n;
3233 		struct ip *ip = NULL;
3234 		struct tcphdr *th = NULL;
3235 		int minthlen;
3236 
3237 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3238 		if (__predict_false(m->m_len < minthlen)) {
3239 			/*
3240 			 * if this code bloat is causing too much of a hit
3241 			 * move it to a separate function and mark it noinline
3242 			 */
3243 			if (m->m_len == pi->ipi_ehdrlen) {
3244 				n = m->m_next;
3245 				MPASS(n);
3246 				if (n->m_len >= sizeof(*ip))  {
3247 					ip = (struct ip *)n->m_data;
3248 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3249 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3250 				} else {
3251 					txq->ift_pullups++;
3252 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3253 						return (ENOMEM);
3254 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3255 				}
3256 			} else {
3257 				txq->ift_pullups++;
3258 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3259 					return (ENOMEM);
3260 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3261 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3262 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3263 			}
3264 		} else {
3265 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3266 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3267 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3268 		}
3269 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3270 		pi->ipi_ipproto = ip->ip_p;
3271 		pi->ipi_flags |= IPI_TX_IPV4;
3272 
3273 		/* TCP checksum offload may require TCP header length */
3274 		if (IS_TX_OFFLOAD4(pi)) {
3275 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3276 				if (__predict_false(th == NULL)) {
3277 					txq->ift_pullups++;
3278 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3279 						return (ENOMEM);
3280 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3281 				}
3282 				pi->ipi_tcp_hflags = th->th_flags;
3283 				pi->ipi_tcp_hlen = th->th_off << 2;
3284 				pi->ipi_tcp_seq = th->th_seq;
3285 			}
3286 			if (IS_TSO4(pi)) {
3287 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3288 					return (ENXIO);
3289 				/*
3290 				 * TSO always requires hardware checksum offload.
3291 				 */
3292 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3293 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3294 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3295 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3296 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3297 					ip->ip_sum = 0;
3298 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3299 				}
3300 			}
3301 		}
3302 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3303                        ip->ip_sum = 0;
3304 
3305 		break;
3306 	}
3307 #endif
3308 #ifdef INET6
3309 	case ETHERTYPE_IPV6:
3310 	{
3311 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3312 		struct tcphdr *th;
3313 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3314 
3315 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3316 			txq->ift_pullups++;
3317 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3318 				return (ENOMEM);
3319 		}
3320 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3321 
3322 		/* XXX-BZ this will go badly in case of ext hdrs. */
3323 		pi->ipi_ipproto = ip6->ip6_nxt;
3324 		pi->ipi_flags |= IPI_TX_IPV6;
3325 
3326 		/* TCP checksum offload may require TCP header length */
3327 		if (IS_TX_OFFLOAD6(pi)) {
3328 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3329 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3330 					txq->ift_pullups++;
3331 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3332 						return (ENOMEM);
3333 				}
3334 				pi->ipi_tcp_hflags = th->th_flags;
3335 				pi->ipi_tcp_hlen = th->th_off << 2;
3336 				pi->ipi_tcp_seq = th->th_seq;
3337 			}
3338 			if (IS_TSO6(pi)) {
3339 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3340 					return (ENXIO);
3341 				/*
3342 				 * TSO always requires hardware checksum offload.
3343 				 */
3344 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3345 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3346 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3347 			}
3348 		}
3349 		break;
3350 	}
3351 #endif
3352 	default:
3353 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3354 		pi->ipi_ip_hlen = 0;
3355 		break;
3356 	}
3357 	*mp = m;
3358 
3359 	return (0);
3360 }
3361 
3362 /*
3363  * If dodgy hardware rejects the scatter gather chain we've handed it
3364  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3365  * m_defrag'd mbufs
3366  */
3367 static __noinline struct mbuf *
3368 iflib_remove_mbuf(iflib_txq_t txq)
3369 {
3370 	int ntxd, pidx;
3371 	struct mbuf *m, **ifsd_m;
3372 
3373 	ifsd_m = txq->ift_sds.ifsd_m;
3374 	ntxd = txq->ift_size;
3375 	pidx = txq->ift_pidx & (ntxd - 1);
3376 	ifsd_m = txq->ift_sds.ifsd_m;
3377 	m = ifsd_m[pidx];
3378 	ifsd_m[pidx] = NULL;
3379 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3380 	if (txq->ift_sds.ifsd_tso_map != NULL)
3381 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3382 		    txq->ift_sds.ifsd_tso_map[pidx]);
3383 #if MEMORY_LOGGING
3384 	txq->ift_dequeued++;
3385 #endif
3386 	return (m);
3387 }
3388 
3389 static inline caddr_t
3390 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3391 {
3392 	qidx_t size;
3393 	int ntxd;
3394 	caddr_t start, end, cur, next;
3395 
3396 	ntxd = txq->ift_size;
3397 	size = txq->ift_txd_size[qid];
3398 	start = txq->ift_ifdi[qid].idi_vaddr;
3399 
3400 	if (__predict_false(size == 0))
3401 		return (start);
3402 	cur = start + size*cidx;
3403 	end = start + size*ntxd;
3404 	next = CACHE_PTR_NEXT(cur);
3405 	return (next < end ? next : start);
3406 }
3407 
3408 /*
3409  * Pad an mbuf to ensure a minimum ethernet frame size.
3410  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3411  */
3412 static __noinline int
3413 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3414 {
3415 	/*
3416 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3417 	 * and ARP message is the smallest common payload I can think of
3418 	 */
3419 	static char pad[18];	/* just zeros */
3420 	int n;
3421 	struct mbuf *new_head;
3422 
3423 	if (!M_WRITABLE(*m_head)) {
3424 		new_head = m_dup(*m_head, M_NOWAIT);
3425 		if (new_head == NULL) {
3426 			m_freem(*m_head);
3427 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3428 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3429 			DBG_COUNTER_INC(tx_frees);
3430 			return ENOMEM;
3431 		}
3432 		m_freem(*m_head);
3433 		*m_head = new_head;
3434 	}
3435 
3436 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3437 	     n > 0; n -= sizeof(pad))
3438 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3439 			break;
3440 
3441 	if (n > 0) {
3442 		m_freem(*m_head);
3443 		device_printf(dev, "cannot pad short frame\n");
3444 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3445 		DBG_COUNTER_INC(tx_frees);
3446 		return (ENOBUFS);
3447 	}
3448 
3449 	return 0;
3450 }
3451 
3452 static int
3453 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3454 {
3455 	if_ctx_t		ctx;
3456 	if_shared_ctx_t		sctx;
3457 	if_softc_ctx_t		scctx;
3458 	bus_dma_tag_t		buf_tag;
3459 	bus_dma_segment_t	*segs;
3460 	struct mbuf		*m_head, **ifsd_m;
3461 	void			*next_txd;
3462 	bus_dmamap_t		map;
3463 	struct if_pkt_info	pi;
3464 	int remap = 0;
3465 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3466 
3467 	ctx = txq->ift_ctx;
3468 	sctx = ctx->ifc_sctx;
3469 	scctx = &ctx->ifc_softc_ctx;
3470 	segs = txq->ift_segs;
3471 	ntxd = txq->ift_size;
3472 	m_head = *m_headp;
3473 	map = NULL;
3474 
3475 	/*
3476 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3477 	 */
3478 	cidx = txq->ift_cidx;
3479 	pidx = txq->ift_pidx;
3480 	if (ctx->ifc_flags & IFC_PREFETCH) {
3481 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3482 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3483 			next_txd = calc_next_txd(txq, cidx, 0);
3484 			prefetch(next_txd);
3485 		}
3486 
3487 		/* prefetch the next cache line of mbuf pointers and flags */
3488 		prefetch(&txq->ift_sds.ifsd_m[next]);
3489 		prefetch(&txq->ift_sds.ifsd_map[next]);
3490 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3491 	}
3492 	map = txq->ift_sds.ifsd_map[pidx];
3493 	ifsd_m = txq->ift_sds.ifsd_m;
3494 
3495 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3496 		buf_tag = txq->ift_tso_buf_tag;
3497 		max_segs = scctx->isc_tx_tso_segments_max;
3498 		map = txq->ift_sds.ifsd_tso_map[pidx];
3499 		MPASS(buf_tag != NULL);
3500 		MPASS(max_segs > 0);
3501 	} else {
3502 		buf_tag = txq->ift_buf_tag;
3503 		max_segs = scctx->isc_tx_nsegments;
3504 		map = txq->ift_sds.ifsd_map[pidx];
3505 	}
3506 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3507 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3508 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3509 		if (err) {
3510 			DBG_COUNTER_INC(encap_txd_encap_fail);
3511 			return err;
3512 		}
3513 	}
3514 	m_head = *m_headp;
3515 
3516 	pkt_info_zero(&pi);
3517 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3518 	pi.ipi_pidx = pidx;
3519 	pi.ipi_qsidx = txq->ift_id;
3520 	pi.ipi_len = m_head->m_pkthdr.len;
3521 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3522 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3523 
3524 	/* deliberate bitwise OR to make one condition */
3525 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3526 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3527 			DBG_COUNTER_INC(encap_txd_encap_fail);
3528 			return (err);
3529 		}
3530 		m_head = *m_headp;
3531 	}
3532 
3533 retry:
3534 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3535 	    BUS_DMA_NOWAIT);
3536 defrag:
3537 	if (__predict_false(err)) {
3538 		switch (err) {
3539 		case EFBIG:
3540 			/* try collapse once and defrag once */
3541 			if (remap == 0) {
3542 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3543 				/* try defrag if collapsing fails */
3544 				if (m_head == NULL)
3545 					remap++;
3546 			}
3547 			if (remap == 1) {
3548 				txq->ift_mbuf_defrag++;
3549 				m_head = m_defrag(*m_headp, M_NOWAIT);
3550 			}
3551 			/*
3552 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3553 			 * failed to map an mbuf that was run through m_defrag
3554 			 */
3555 			MPASS(remap <= 1);
3556 			if (__predict_false(m_head == NULL || remap > 1))
3557 				goto defrag_failed;
3558 			remap++;
3559 			*m_headp = m_head;
3560 			goto retry;
3561 			break;
3562 		case ENOMEM:
3563 			txq->ift_no_tx_dma_setup++;
3564 			break;
3565 		default:
3566 			txq->ift_no_tx_dma_setup++;
3567 			m_freem(*m_headp);
3568 			DBG_COUNTER_INC(tx_frees);
3569 			*m_headp = NULL;
3570 			break;
3571 		}
3572 		txq->ift_map_failed++;
3573 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3574 		DBG_COUNTER_INC(encap_txd_encap_fail);
3575 		return (err);
3576 	}
3577 	ifsd_m[pidx] = m_head;
3578 	/*
3579 	 * XXX assumes a 1 to 1 relationship between segments and
3580 	 *        descriptors - this does not hold true on all drivers, e.g.
3581 	 *        cxgb
3582 	 */
3583 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3584 		txq->ift_no_desc_avail++;
3585 		bus_dmamap_unload(buf_tag, map);
3586 		DBG_COUNTER_INC(encap_txq_avail_fail);
3587 		DBG_COUNTER_INC(encap_txd_encap_fail);
3588 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3589 			GROUPTASK_ENQUEUE(&txq->ift_task);
3590 		return (ENOBUFS);
3591 	}
3592 	/*
3593 	 * On Intel cards we can greatly reduce the number of TX interrupts
3594 	 * we see by only setting report status on every Nth descriptor.
3595 	 * However, this also means that the driver will need to keep track
3596 	 * of the descriptors that RS was set on to check them for the DD bit.
3597 	 */
3598 	txq->ift_rs_pending += nsegs + 1;
3599 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3600 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3601 		pi.ipi_flags |= IPI_TX_INTR;
3602 		txq->ift_rs_pending = 0;
3603 	}
3604 
3605 	pi.ipi_segs = segs;
3606 	pi.ipi_nsegs = nsegs;
3607 
3608 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3609 #ifdef PKT_DEBUG
3610 	print_pkt(&pi);
3611 #endif
3612 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3613 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3614 		DBG_COUNTER_INC(tx_encap);
3615 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3616 
3617 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3618 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3619 			ndesc += txq->ift_size;
3620 			txq->ift_gen = 1;
3621 		}
3622 		/*
3623 		 * drivers can need as many as
3624 		 * two sentinels
3625 		 */
3626 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3627 		MPASS(pi.ipi_new_pidx != pidx);
3628 		MPASS(ndesc > 0);
3629 		txq->ift_in_use += ndesc;
3630 		txq->ift_db_pending += ndesc;
3631 
3632 		/*
3633 		 * We update the last software descriptor again here because there may
3634 		 * be a sentinel and/or there may be more mbufs than segments
3635 		 */
3636 		txq->ift_pidx = pi.ipi_new_pidx;
3637 		txq->ift_npending += pi.ipi_ndescs;
3638 	} else {
3639 		*m_headp = m_head = iflib_remove_mbuf(txq);
3640 		if (err == EFBIG) {
3641 			txq->ift_txd_encap_efbig++;
3642 			if (remap < 2) {
3643 				remap = 1;
3644 				goto defrag;
3645 			}
3646 		}
3647 		goto defrag_failed;
3648 	}
3649 	/*
3650 	 * err can't possibly be non-zero here, so we don't neet to test it
3651 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3652 	 */
3653 	return (err);
3654 
3655 defrag_failed:
3656 	txq->ift_mbuf_defrag_failed++;
3657 	txq->ift_map_failed++;
3658 	m_freem(*m_headp);
3659 	DBG_COUNTER_INC(tx_frees);
3660 	*m_headp = NULL;
3661 	DBG_COUNTER_INC(encap_txd_encap_fail);
3662 	return (ENOMEM);
3663 }
3664 
3665 static void
3666 iflib_tx_desc_free(iflib_txq_t txq, int n)
3667 {
3668 	uint32_t qsize, cidx, mask, gen;
3669 	struct mbuf *m, **ifsd_m;
3670 	bool do_prefetch;
3671 
3672 	cidx = txq->ift_cidx;
3673 	gen = txq->ift_gen;
3674 	qsize = txq->ift_size;
3675 	mask = qsize-1;
3676 	ifsd_m = txq->ift_sds.ifsd_m;
3677 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3678 
3679 	while (n-- > 0) {
3680 		if (do_prefetch) {
3681 			prefetch(ifsd_m[(cidx + 3) & mask]);
3682 			prefetch(ifsd_m[(cidx + 4) & mask]);
3683 		}
3684 		if ((m = ifsd_m[cidx]) != NULL) {
3685 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3686 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3687 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3688 				    txq->ift_sds.ifsd_tso_map[cidx],
3689 				    BUS_DMASYNC_POSTWRITE);
3690 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3691 				    txq->ift_sds.ifsd_tso_map[cidx]);
3692 			} else {
3693 				bus_dmamap_sync(txq->ift_buf_tag,
3694 				    txq->ift_sds.ifsd_map[cidx],
3695 				    BUS_DMASYNC_POSTWRITE);
3696 				bus_dmamap_unload(txq->ift_buf_tag,
3697 				    txq->ift_sds.ifsd_map[cidx]);
3698 			}
3699 			/* XXX we don't support any drivers that batch packets yet */
3700 			MPASS(m->m_nextpkt == NULL);
3701 			m_freem(m);
3702 			ifsd_m[cidx] = NULL;
3703 #if MEMORY_LOGGING
3704 			txq->ift_dequeued++;
3705 #endif
3706 			DBG_COUNTER_INC(tx_frees);
3707 		}
3708 		if (__predict_false(++cidx == qsize)) {
3709 			cidx = 0;
3710 			gen = 0;
3711 		}
3712 	}
3713 	txq->ift_cidx = cidx;
3714 	txq->ift_gen = gen;
3715 }
3716 
3717 static __inline int
3718 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3719 {
3720 	int reclaim;
3721 	if_ctx_t ctx = txq->ift_ctx;
3722 
3723 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3724 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3725 
3726 	/*
3727 	 * Need a rate-limiting check so that this isn't called every time
3728 	 */
3729 	iflib_tx_credits_update(ctx, txq);
3730 	reclaim = DESC_RECLAIMABLE(txq);
3731 
3732 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3733 #ifdef INVARIANTS
3734 		if (iflib_verbose_debug) {
3735 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3736 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3737 			       reclaim, thresh);
3738 		}
3739 #endif
3740 		return (0);
3741 	}
3742 	iflib_tx_desc_free(txq, reclaim);
3743 	txq->ift_cleaned += reclaim;
3744 	txq->ift_in_use -= reclaim;
3745 
3746 	return (reclaim);
3747 }
3748 
3749 static struct mbuf **
3750 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3751 {
3752 	int next, size;
3753 	struct mbuf **items;
3754 
3755 	size = r->size;
3756 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3757 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3758 
3759 	prefetch(items[(cidx + offset) & (size-1)]);
3760 	if (remaining > 1) {
3761 		prefetch2cachelines(&items[next]);
3762 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3763 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3764 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3765 	}
3766 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3767 }
3768 
3769 static void
3770 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3771 {
3772 
3773 	ifmp_ring_check_drainage(txq->ift_br, budget);
3774 }
3775 
3776 static uint32_t
3777 iflib_txq_can_drain(struct ifmp_ring *r)
3778 {
3779 	iflib_txq_t txq = r->cookie;
3780 	if_ctx_t ctx = txq->ift_ctx;
3781 
3782 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3783 		return (1);
3784 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3785 	    BUS_DMASYNC_POSTREAD);
3786 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3787 	    false));
3788 }
3789 
3790 static uint32_t
3791 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3792 {
3793 	iflib_txq_t txq = r->cookie;
3794 	if_ctx_t ctx = txq->ift_ctx;
3795 	if_t ifp = ctx->ifc_ifp;
3796 	struct mbuf *m, **mp;
3797 	int avail, bytes_sent, skipped, count, err, i;
3798 	int mcast_sent, pkt_sent, reclaimed;
3799 	bool do_prefetch, rang, ring;
3800 
3801 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3802 			    !LINK_ACTIVE(ctx))) {
3803 		DBG_COUNTER_INC(txq_drain_notready);
3804 		return (0);
3805 	}
3806 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3807 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3808 	avail = IDXDIFF(pidx, cidx, r->size);
3809 
3810 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3811 		/*
3812 		 * The driver is unloading so we need to free all pending packets.
3813 		 */
3814 		DBG_COUNTER_INC(txq_drain_flushing);
3815 		for (i = 0; i < avail; i++) {
3816 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3817 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3818 			r->items[(cidx + i) & (r->size-1)] = NULL;
3819 		}
3820 		return (avail);
3821 	}
3822 
3823 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3824 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3825 		CALLOUT_LOCK(txq);
3826 		callout_stop(&txq->ift_timer);
3827 		CALLOUT_UNLOCK(txq);
3828 		DBG_COUNTER_INC(txq_drain_oactive);
3829 		return (0);
3830 	}
3831 
3832 	/*
3833 	 * If we've reclaimed any packets this queue cannot be hung.
3834 	 */
3835 	if (reclaimed)
3836 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3837 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3838 	count = MIN(avail, TX_BATCH_SIZE);
3839 #ifdef INVARIANTS
3840 	if (iflib_verbose_debug)
3841 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3842 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3843 #endif
3844 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3845 	err = 0;
3846 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3847 		int rem = do_prefetch ? count - i : 0;
3848 
3849 		mp = _ring_peek_one(r, cidx, i, rem);
3850 		MPASS(mp != NULL && *mp != NULL);
3851 
3852 		/*
3853 		 * Completion interrupts will use the address of the txq
3854 		 * as a sentinel to enqueue _something_ in order to acquire
3855 		 * the lock on the mp_ring (there's no direct lock call).
3856 		 * We obviously whave to check for these sentinel cases
3857 		 * and skip them.
3858 		 */
3859 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3860 			skipped++;
3861 			continue;
3862 		}
3863 		err = iflib_encap(txq, mp);
3864 		if (__predict_false(err)) {
3865 			/* no room - bail out */
3866 			if (err == ENOBUFS)
3867 				break;
3868 			skipped++;
3869 			/* we can't send this packet - skip it */
3870 			continue;
3871 		}
3872 		pkt_sent++;
3873 		m = *mp;
3874 		DBG_COUNTER_INC(tx_sent);
3875 		bytes_sent += m->m_pkthdr.len;
3876 		mcast_sent += !!(m->m_flags & M_MCAST);
3877 
3878 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3879 			break;
3880 		ETHER_BPF_MTAP(ifp, m);
3881 		rang = iflib_txd_db_check(txq, false);
3882 	}
3883 
3884 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3885 	ring = rang ? false  : (iflib_min_tx_latency | err);
3886 	iflib_txd_db_check(txq, ring);
3887 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3888 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3889 	if (mcast_sent)
3890 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3891 #ifdef INVARIANTS
3892 	if (iflib_verbose_debug)
3893 		printf("consumed=%d\n", skipped + pkt_sent);
3894 #endif
3895 	return (skipped + pkt_sent);
3896 }
3897 
3898 static uint32_t
3899 iflib_txq_drain_always(struct ifmp_ring *r)
3900 {
3901 	return (1);
3902 }
3903 
3904 static uint32_t
3905 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3906 {
3907 	int i, avail;
3908 	struct mbuf **mp;
3909 	iflib_txq_t txq;
3910 
3911 	txq = r->cookie;
3912 
3913 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3914 	CALLOUT_LOCK(txq);
3915 	callout_stop(&txq->ift_timer);
3916 	CALLOUT_UNLOCK(txq);
3917 
3918 	avail = IDXDIFF(pidx, cidx, r->size);
3919 	for (i = 0; i < avail; i++) {
3920 		mp = _ring_peek_one(r, cidx, i, avail - i);
3921 		if (__predict_false(*mp == (struct mbuf *)txq))
3922 			continue;
3923 		m_freem(*mp);
3924 		DBG_COUNTER_INC(tx_frees);
3925 	}
3926 	MPASS(ifmp_ring_is_stalled(r) == 0);
3927 	return (avail);
3928 }
3929 
3930 static void
3931 iflib_ifmp_purge(iflib_txq_t txq)
3932 {
3933 	struct ifmp_ring *r;
3934 
3935 	r = txq->ift_br;
3936 	r->drain = iflib_txq_drain_free;
3937 	r->can_drain = iflib_txq_drain_always;
3938 
3939 	ifmp_ring_check_drainage(r, r->size);
3940 
3941 	r->drain = iflib_txq_drain;
3942 	r->can_drain = iflib_txq_can_drain;
3943 }
3944 
3945 static void
3946 _task_fn_tx(void *context)
3947 {
3948 	iflib_txq_t txq = context;
3949 	if_ctx_t ctx = txq->ift_ctx;
3950 	if_t ifp = ctx->ifc_ifp;
3951 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3952 
3953 #ifdef IFLIB_DIAGNOSTICS
3954 	txq->ift_cpu_exec_count[curcpu]++;
3955 #endif
3956 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3957 		return;
3958 #ifdef DEV_NETMAP
3959 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3960 	    netmap_tx_irq(ifp, txq->ift_id))
3961 		goto skip_ifmp;
3962 #endif
3963 #ifdef ALTQ
3964 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3965 		iflib_altq_if_start(ifp);
3966 #endif
3967 	if (txq->ift_db_pending)
3968 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3969 	else if (!abdicate)
3970 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3971 	/*
3972 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3973 	 */
3974 	if (abdicate)
3975 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3976 #ifdef DEV_NETMAP
3977 skip_ifmp:
3978 #endif
3979 	if (ctx->ifc_flags & IFC_LEGACY)
3980 		IFDI_INTR_ENABLE(ctx);
3981 	else
3982 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3983 }
3984 
3985 static void
3986 _task_fn_rx(void *context)
3987 {
3988 	iflib_rxq_t rxq = context;
3989 	if_ctx_t ctx = rxq->ifr_ctx;
3990 	uint8_t more;
3991 	uint16_t budget;
3992 #ifdef DEV_NETMAP
3993 	u_int work = 0;
3994 	int nmirq;
3995 #endif
3996 
3997 #ifdef IFLIB_DIAGNOSTICS
3998 	rxq->ifr_cpu_exec_count[curcpu]++;
3999 #endif
4000 	DBG_COUNTER_INC(task_fn_rxs);
4001 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4002 		return;
4003 #ifdef DEV_NETMAP
4004 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4005 	if (nmirq != NM_IRQ_PASS) {
4006 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4007 		goto skip_rxeof;
4008 	}
4009 #endif
4010 	budget = ctx->ifc_sysctl_rx_budget;
4011 	if (budget == 0)
4012 		budget = 16;	/* XXX */
4013 	more = iflib_rxeof(rxq, budget);
4014 #ifdef DEV_NETMAP
4015 skip_rxeof:
4016 #endif
4017 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4018 		if (ctx->ifc_flags & IFC_LEGACY)
4019 			IFDI_INTR_ENABLE(ctx);
4020 		else
4021 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4022 		DBG_COUNTER_INC(rx_intr_enables);
4023 	}
4024 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4025 		return;
4026 
4027 	if (more & IFLIB_RXEOF_MORE)
4028 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4029 	else if (more & IFLIB_RXEOF_EMPTY)
4030 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4031 }
4032 
4033 static void
4034 _task_fn_admin(void *context)
4035 {
4036 	if_ctx_t ctx = context;
4037 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4038 	iflib_txq_t txq;
4039 	int i;
4040 	bool oactive, running, do_reset, do_watchdog, in_detach;
4041 
4042 	STATE_LOCK(ctx);
4043 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4044 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4045 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4046 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4047 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4048 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4049 	STATE_UNLOCK(ctx);
4050 
4051 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4052 		return;
4053 	if (in_detach)
4054 		return;
4055 
4056 	CTX_LOCK(ctx);
4057 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4058 		CALLOUT_LOCK(txq);
4059 		callout_stop(&txq->ift_timer);
4060 		CALLOUT_UNLOCK(txq);
4061 	}
4062 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4063 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4064 	if (do_watchdog) {
4065 		ctx->ifc_watchdog_events++;
4066 		IFDI_WATCHDOG_RESET(ctx);
4067 	}
4068 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4069 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4070 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4071 		    txq->ift_timer.c_cpu);
4072 	}
4073 	IFDI_LINK_INTR_ENABLE(ctx);
4074 	if (do_reset)
4075 		iflib_if_init_locked(ctx);
4076 	CTX_UNLOCK(ctx);
4077 
4078 	if (LINK_ACTIVE(ctx) == 0)
4079 		return;
4080 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4081 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4082 }
4083 
4084 static void
4085 _task_fn_iov(void *context)
4086 {
4087 	if_ctx_t ctx = context;
4088 
4089 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4090 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4091 		return;
4092 
4093 	CTX_LOCK(ctx);
4094 	IFDI_VFLR_HANDLE(ctx);
4095 	CTX_UNLOCK(ctx);
4096 }
4097 
4098 static int
4099 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4100 {
4101 	int err;
4102 	if_int_delay_info_t info;
4103 	if_ctx_t ctx;
4104 
4105 	info = (if_int_delay_info_t)arg1;
4106 	ctx = info->iidi_ctx;
4107 	info->iidi_req = req;
4108 	info->iidi_oidp = oidp;
4109 	CTX_LOCK(ctx);
4110 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4111 	CTX_UNLOCK(ctx);
4112 	return (err);
4113 }
4114 
4115 /*********************************************************************
4116  *
4117  *  IFNET FUNCTIONS
4118  *
4119  **********************************************************************/
4120 
4121 static void
4122 iflib_if_init_locked(if_ctx_t ctx)
4123 {
4124 	iflib_stop(ctx);
4125 	iflib_init_locked(ctx);
4126 }
4127 
4128 static void
4129 iflib_if_init(void *arg)
4130 {
4131 	if_ctx_t ctx = arg;
4132 
4133 	CTX_LOCK(ctx);
4134 	iflib_if_init_locked(ctx);
4135 	CTX_UNLOCK(ctx);
4136 }
4137 
4138 static int
4139 iflib_if_transmit(if_t ifp, struct mbuf *m)
4140 {
4141 	if_ctx_t	ctx = if_getsoftc(ifp);
4142 
4143 	iflib_txq_t txq;
4144 	int err, qidx;
4145 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4146 
4147 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4148 		DBG_COUNTER_INC(tx_frees);
4149 		m_freem(m);
4150 		return (ENETDOWN);
4151 	}
4152 
4153 	MPASS(m->m_nextpkt == NULL);
4154 	/* ALTQ-enabled interfaces always use queue 0. */
4155 	qidx = 0;
4156 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4157 		qidx = QIDX(ctx, m);
4158 	/*
4159 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4160 	 */
4161 	txq = &ctx->ifc_txqs[qidx];
4162 
4163 #ifdef DRIVER_BACKPRESSURE
4164 	if (txq->ift_closed) {
4165 		while (m != NULL) {
4166 			next = m->m_nextpkt;
4167 			m->m_nextpkt = NULL;
4168 			m_freem(m);
4169 			DBG_COUNTER_INC(tx_frees);
4170 			m = next;
4171 		}
4172 		return (ENOBUFS);
4173 	}
4174 #endif
4175 #ifdef notyet
4176 	qidx = count = 0;
4177 	mp = marr;
4178 	next = m;
4179 	do {
4180 		count++;
4181 		next = next->m_nextpkt;
4182 	} while (next != NULL);
4183 
4184 	if (count > nitems(marr))
4185 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4186 			/* XXX check nextpkt */
4187 			m_freem(m);
4188 			/* XXX simplify for now */
4189 			DBG_COUNTER_INC(tx_frees);
4190 			return (ENOBUFS);
4191 		}
4192 	for (next = m, i = 0; next != NULL; i++) {
4193 		mp[i] = next;
4194 		next = next->m_nextpkt;
4195 		mp[i]->m_nextpkt = NULL;
4196 	}
4197 #endif
4198 	DBG_COUNTER_INC(tx_seen);
4199 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4200 
4201 	if (abdicate)
4202 		GROUPTASK_ENQUEUE(&txq->ift_task);
4203  	if (err) {
4204 		if (!abdicate)
4205 			GROUPTASK_ENQUEUE(&txq->ift_task);
4206 		/* support forthcoming later */
4207 #ifdef DRIVER_BACKPRESSURE
4208 		txq->ift_closed = TRUE;
4209 #endif
4210 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4211 		m_freem(m);
4212 		DBG_COUNTER_INC(tx_frees);
4213 	}
4214 
4215 	return (err);
4216 }
4217 
4218 #ifdef ALTQ
4219 /*
4220  * The overall approach to integrating iflib with ALTQ is to continue to use
4221  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4222  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4223  * is redundant/unnecessary, but doing so minimizes the amount of
4224  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4225  * redundantly queueing to an intermediate mp_ring is swamped by the
4226  * performance limitations inherent in using ALTQ.
4227  *
4228  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4229  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4230  * given interface.  If ALTQ is enabled for an interface, then all
4231  * transmitted packets for that interface will be submitted to the ALTQ
4232  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4233  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4234  * update stats that the iflib machinery handles, and which is sensitve to
4235  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4236  * will be installed as the start routine for use by ALTQ facilities that
4237  * need to trigger queue drains on a scheduled basis.
4238  *
4239  */
4240 static void
4241 iflib_altq_if_start(if_t ifp)
4242 {
4243 	struct ifaltq *ifq = &ifp->if_snd;
4244 	struct mbuf *m;
4245 
4246 	IFQ_LOCK(ifq);
4247 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4248 	while (m != NULL) {
4249 		iflib_if_transmit(ifp, m);
4250 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4251 	}
4252 	IFQ_UNLOCK(ifq);
4253 }
4254 
4255 static int
4256 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4257 {
4258 	int err;
4259 
4260 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4261 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4262 		if (err == 0)
4263 			iflib_altq_if_start(ifp);
4264 	} else
4265 		err = iflib_if_transmit(ifp, m);
4266 
4267 	return (err);
4268 }
4269 #endif /* ALTQ */
4270 
4271 static void
4272 iflib_if_qflush(if_t ifp)
4273 {
4274 	if_ctx_t ctx = if_getsoftc(ifp);
4275 	iflib_txq_t txq = ctx->ifc_txqs;
4276 	int i;
4277 
4278 	STATE_LOCK(ctx);
4279 	ctx->ifc_flags |= IFC_QFLUSH;
4280 	STATE_UNLOCK(ctx);
4281 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4282 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4283 			iflib_txq_check_drain(txq, 0);
4284 	STATE_LOCK(ctx);
4285 	ctx->ifc_flags &= ~IFC_QFLUSH;
4286 	STATE_UNLOCK(ctx);
4287 
4288 	/*
4289 	 * When ALTQ is enabled, this will also take care of purging the
4290 	 * ALTQ queue(s).
4291 	 */
4292 	if_qflush(ifp);
4293 }
4294 
4295 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4296 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4297 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4298 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4299 
4300 static int
4301 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4302 {
4303 	if_ctx_t ctx = if_getsoftc(ifp);
4304 	struct ifreq	*ifr = (struct ifreq *)data;
4305 #if defined(INET) || defined(INET6)
4306 	struct ifaddr	*ifa = (struct ifaddr *)data;
4307 #endif
4308 	bool		avoid_reset = false;
4309 	int		err = 0, reinit = 0, bits;
4310 
4311 	switch (command) {
4312 	case SIOCSIFADDR:
4313 #ifdef INET
4314 		if (ifa->ifa_addr->sa_family == AF_INET)
4315 			avoid_reset = true;
4316 #endif
4317 #ifdef INET6
4318 		if (ifa->ifa_addr->sa_family == AF_INET6)
4319 			avoid_reset = true;
4320 #endif
4321 		/*
4322 		** Calling init results in link renegotiation,
4323 		** so we avoid doing it when possible.
4324 		*/
4325 		if (avoid_reset) {
4326 			if_setflagbits(ifp, IFF_UP,0);
4327 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4328 				reinit = 1;
4329 #ifdef INET
4330 			if (!(if_getflags(ifp) & IFF_NOARP))
4331 				arp_ifinit(ifp, ifa);
4332 #endif
4333 		} else
4334 			err = ether_ioctl(ifp, command, data);
4335 		break;
4336 	case SIOCSIFMTU:
4337 		CTX_LOCK(ctx);
4338 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4339 			CTX_UNLOCK(ctx);
4340 			break;
4341 		}
4342 		bits = if_getdrvflags(ifp);
4343 		/* stop the driver and free any clusters before proceeding */
4344 		iflib_stop(ctx);
4345 
4346 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4347 			STATE_LOCK(ctx);
4348 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4349 				ctx->ifc_flags |= IFC_MULTISEG;
4350 			else
4351 				ctx->ifc_flags &= ~IFC_MULTISEG;
4352 			STATE_UNLOCK(ctx);
4353 			err = if_setmtu(ifp, ifr->ifr_mtu);
4354 		}
4355 		iflib_init_locked(ctx);
4356 		STATE_LOCK(ctx);
4357 		if_setdrvflags(ifp, bits);
4358 		STATE_UNLOCK(ctx);
4359 		CTX_UNLOCK(ctx);
4360 		break;
4361 	case SIOCSIFFLAGS:
4362 		CTX_LOCK(ctx);
4363 		if (if_getflags(ifp) & IFF_UP) {
4364 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4365 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4366 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4367 					CTX_UNLOCK(ctx);
4368 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4369 					CTX_LOCK(ctx);
4370 				}
4371 			} else
4372 				reinit = 1;
4373 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4374 			iflib_stop(ctx);
4375 		}
4376 		ctx->ifc_if_flags = if_getflags(ifp);
4377 		CTX_UNLOCK(ctx);
4378 		break;
4379 	case SIOCADDMULTI:
4380 	case SIOCDELMULTI:
4381 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4382 			CTX_LOCK(ctx);
4383 			IFDI_INTR_DISABLE(ctx);
4384 			IFDI_MULTI_SET(ctx);
4385 			IFDI_INTR_ENABLE(ctx);
4386 			CTX_UNLOCK(ctx);
4387 		}
4388 		break;
4389 	case SIOCSIFMEDIA:
4390 		CTX_LOCK(ctx);
4391 		IFDI_MEDIA_SET(ctx);
4392 		CTX_UNLOCK(ctx);
4393 		/* FALLTHROUGH */
4394 	case SIOCGIFMEDIA:
4395 	case SIOCGIFXMEDIA:
4396 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4397 		break;
4398 	case SIOCGI2C:
4399 	{
4400 		struct ifi2creq i2c;
4401 
4402 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4403 		if (err != 0)
4404 			break;
4405 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4406 			err = EINVAL;
4407 			break;
4408 		}
4409 		if (i2c.len > sizeof(i2c.data)) {
4410 			err = EINVAL;
4411 			break;
4412 		}
4413 
4414 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4415 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4416 			    sizeof(i2c));
4417 		break;
4418 	}
4419 	case SIOCSIFCAP:
4420 	{
4421 		int mask, setmask, oldmask;
4422 
4423 		oldmask = if_getcapenable(ifp);
4424 		mask = ifr->ifr_reqcap ^ oldmask;
4425 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4426 		setmask = 0;
4427 #ifdef TCP_OFFLOAD
4428 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4429 #endif
4430 		setmask |= (mask & IFCAP_FLAGS);
4431 		setmask |= (mask & IFCAP_WOL);
4432 
4433 		/*
4434 		 * If any RX csum has changed, change all the ones that
4435 		 * are supported by the driver.
4436 		 */
4437 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4438 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4439 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4440 		}
4441 
4442 		/*
4443 		 * want to ensure that traffic has stopped before we change any of the flags
4444 		 */
4445 		if (setmask) {
4446 			CTX_LOCK(ctx);
4447 			bits = if_getdrvflags(ifp);
4448 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4449 				iflib_stop(ctx);
4450 			STATE_LOCK(ctx);
4451 			if_togglecapenable(ifp, setmask);
4452 			STATE_UNLOCK(ctx);
4453 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4454 				iflib_init_locked(ctx);
4455 			STATE_LOCK(ctx);
4456 			if_setdrvflags(ifp, bits);
4457 			STATE_UNLOCK(ctx);
4458 			CTX_UNLOCK(ctx);
4459 		}
4460 		if_vlancap(ifp);
4461 		break;
4462 	}
4463 	case SIOCGPRIVATE_0:
4464 	case SIOCSDRVSPEC:
4465 	case SIOCGDRVSPEC:
4466 		CTX_LOCK(ctx);
4467 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4468 		CTX_UNLOCK(ctx);
4469 		break;
4470 	default:
4471 		err = ether_ioctl(ifp, command, data);
4472 		break;
4473 	}
4474 	if (reinit)
4475 		iflib_if_init(ctx);
4476 	return (err);
4477 }
4478 
4479 static uint64_t
4480 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4481 {
4482 	if_ctx_t ctx = if_getsoftc(ifp);
4483 
4484 	return (IFDI_GET_COUNTER(ctx, cnt));
4485 }
4486 
4487 /*********************************************************************
4488  *
4489  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4490  *
4491  **********************************************************************/
4492 
4493 static void
4494 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4495 {
4496 	if_ctx_t ctx = if_getsoftc(ifp);
4497 
4498 	if ((void *)ctx != arg)
4499 		return;
4500 
4501 	if ((vtag == 0) || (vtag > 4095))
4502 		return;
4503 
4504 	if (iflib_in_detach(ctx))
4505 		return;
4506 
4507 	CTX_LOCK(ctx);
4508 	/* Driver may need all untagged packets to be flushed */
4509 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4510 		iflib_stop(ctx);
4511 	IFDI_VLAN_REGISTER(ctx, vtag);
4512 	/* Re-init to load the changes, if required */
4513 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4514 		iflib_init_locked(ctx);
4515 	CTX_UNLOCK(ctx);
4516 }
4517 
4518 static void
4519 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4520 {
4521 	if_ctx_t ctx = if_getsoftc(ifp);
4522 
4523 	if ((void *)ctx != arg)
4524 		return;
4525 
4526 	if ((vtag == 0) || (vtag > 4095))
4527 		return;
4528 
4529 	CTX_LOCK(ctx);
4530 	/* Driver may need all tagged packets to be flushed */
4531 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4532 		iflib_stop(ctx);
4533 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4534 	/* Re-init to load the changes, if required */
4535 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4536 		iflib_init_locked(ctx);
4537 	CTX_UNLOCK(ctx);
4538 }
4539 
4540 static void
4541 iflib_led_func(void *arg, int onoff)
4542 {
4543 	if_ctx_t ctx = arg;
4544 
4545 	CTX_LOCK(ctx);
4546 	IFDI_LED_FUNC(ctx, onoff);
4547 	CTX_UNLOCK(ctx);
4548 }
4549 
4550 /*********************************************************************
4551  *
4552  *  BUS FUNCTION DEFINITIONS
4553  *
4554  **********************************************************************/
4555 
4556 int
4557 iflib_device_probe(device_t dev)
4558 {
4559 	const pci_vendor_info_t *ent;
4560 	if_shared_ctx_t sctx;
4561 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4562 	uint16_t pci_vendor_id;
4563 
4564 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4565 		return (ENOTSUP);
4566 
4567 	pci_vendor_id = pci_get_vendor(dev);
4568 	pci_device_id = pci_get_device(dev);
4569 	pci_subvendor_id = pci_get_subvendor(dev);
4570 	pci_subdevice_id = pci_get_subdevice(dev);
4571 	pci_rev_id = pci_get_revid(dev);
4572 	if (sctx->isc_parse_devinfo != NULL)
4573 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4574 
4575 	ent = sctx->isc_vendor_info;
4576 	while (ent->pvi_vendor_id != 0) {
4577 		if (pci_vendor_id != ent->pvi_vendor_id) {
4578 			ent++;
4579 			continue;
4580 		}
4581 		if ((pci_device_id == ent->pvi_device_id) &&
4582 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4583 		     (ent->pvi_subvendor_id == 0)) &&
4584 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4585 		     (ent->pvi_subdevice_id == 0)) &&
4586 		    ((pci_rev_id == ent->pvi_rev_id) ||
4587 		     (ent->pvi_rev_id == 0))) {
4588 			device_set_desc_copy(dev, ent->pvi_name);
4589 			/* this needs to be changed to zero if the bus probing code
4590 			 * ever stops re-probing on best match because the sctx
4591 			 * may have its values over written by register calls
4592 			 * in subsequent probes
4593 			 */
4594 			return (BUS_PROBE_DEFAULT);
4595 		}
4596 		ent++;
4597 	}
4598 	return (ENXIO);
4599 }
4600 
4601 int
4602 iflib_device_probe_vendor(device_t dev)
4603 {
4604 	int probe;
4605 
4606 	probe = iflib_device_probe(dev);
4607 	if (probe == BUS_PROBE_DEFAULT)
4608 		return (BUS_PROBE_VENDOR);
4609 	else
4610 		return (probe);
4611 }
4612 
4613 static void
4614 iflib_reset_qvalues(if_ctx_t ctx)
4615 {
4616 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4617 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4618 	device_t dev = ctx->ifc_dev;
4619 	int i;
4620 
4621 	if (ctx->ifc_sysctl_ntxqs != 0)
4622 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4623 	if (ctx->ifc_sysctl_nrxqs != 0)
4624 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4625 
4626 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4627 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4628 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4629 		else
4630 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4631 	}
4632 
4633 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4634 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4635 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4636 		else
4637 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4638 	}
4639 
4640 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4641 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4642 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4643 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4644 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4645 		}
4646 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4647 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4648 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4649 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4650 		}
4651 		if (!powerof2(scctx->isc_nrxd[i])) {
4652 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4653 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4654 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4655 		}
4656 	}
4657 
4658 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4659 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4660 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4661 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4662 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4663 		}
4664 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4665 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4666 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4667 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4668 		}
4669 		if (!powerof2(scctx->isc_ntxd[i])) {
4670 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4671 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4672 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4673 		}
4674 	}
4675 }
4676 
4677 static void
4678 iflib_add_pfil(if_ctx_t ctx)
4679 {
4680 	struct pfil_head *pfil;
4681 	struct pfil_head_args pa;
4682 	iflib_rxq_t rxq;
4683 	int i;
4684 
4685 	pa.pa_version = PFIL_VERSION;
4686 	pa.pa_flags = PFIL_IN;
4687 	pa.pa_type = PFIL_TYPE_ETHERNET;
4688 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4689 	pfil = pfil_head_register(&pa);
4690 
4691 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4692 		rxq->pfil = pfil;
4693 	}
4694 }
4695 
4696 static void
4697 iflib_rem_pfil(if_ctx_t ctx)
4698 {
4699 	struct pfil_head *pfil;
4700 	iflib_rxq_t rxq;
4701 	int i;
4702 
4703 	rxq = ctx->ifc_rxqs;
4704 	pfil = rxq->pfil;
4705 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4706 		rxq->pfil = NULL;
4707 	}
4708 	pfil_head_unregister(pfil);
4709 }
4710 
4711 
4712 /*
4713  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4714  * cpuid and wrapping as necessary.
4715  */
4716 static unsigned int
4717 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4718 {
4719 	unsigned int first_valid;
4720 	unsigned int last_valid;
4721 
4722 	/* cpuid should always be in the valid set */
4723 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4724 
4725 	/* valid set should never be empty */
4726 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4727 
4728 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4729 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4730 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4731 	while (n > 0) {
4732 		do {
4733 			cpuid++;
4734 			if (cpuid > last_valid)
4735 				cpuid = first_valid;
4736 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4737 		n--;
4738 	}
4739 
4740 	return (cpuid);
4741 }
4742 
4743 #if defined(SMP) && defined(SCHED_ULE)
4744 extern struct cpu_group *cpu_top;              /* CPU topology */
4745 
4746 static int
4747 find_child_with_core(int cpu, struct cpu_group *grp)
4748 {
4749 	int i;
4750 
4751 	if (grp->cg_children == 0)
4752 		return -1;
4753 
4754 	MPASS(grp->cg_child);
4755 	for (i = 0; i < grp->cg_children; i++) {
4756 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4757 			return i;
4758 	}
4759 
4760 	return -1;
4761 }
4762 
4763 
4764 /*
4765  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4766  * does not distinguish among multiple L2 neighbors if the given CPU has
4767  * more than one (it will always return the same result in that case).
4768  */
4769 static int
4770 find_l2_neighbor(int cpu)
4771 {
4772 	struct cpu_group *grp;
4773 	int i;
4774 
4775 	grp = cpu_top;
4776 	if (grp == NULL)
4777 		return -1;
4778 
4779 	/*
4780 	 * Find the smallest CPU group that contains the given core.
4781 	 */
4782 	i = 0;
4783 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4784 		/*
4785 		 * If the smallest group containing the given CPU has less
4786 		 * than two members, we conclude the given CPU has no
4787 		 * L2 neighbor.
4788 		 */
4789 		if (grp->cg_child[i].cg_count <= 1)
4790 			return (-1);
4791 		grp = &grp->cg_child[i];
4792 	}
4793 
4794 	/* Must share L2. */
4795 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4796 		return -1;
4797 
4798 	/*
4799 	 * Select the first member of the set that isn't the reference
4800 	 * CPU, which at this point is guaranteed to exist.
4801 	 */
4802 	for (i = 0; i < CPU_SETSIZE; i++) {
4803 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4804 			return (i);
4805 	}
4806 
4807 	/* Should never be reached */
4808 	return (-1);
4809 }
4810 
4811 #else
4812 static int
4813 find_l2_neighbor(int cpu)
4814 {
4815 
4816 	return (-1);
4817 }
4818 #endif
4819 
4820 /*
4821  * CPU mapping behaviors
4822  * ---------------------
4823  * 'separate txrx' refers to the separate_txrx sysctl
4824  * 'use logical' refers to the use_logical_cores sysctl
4825  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4826  *
4827  *  separate     use     INTR
4828  *    txrx     logical   CPUS   result
4829  * ---------- --------- ------ ------------------------------------------------
4830  *     -          -       X     RX and TX queues mapped to consecutive physical
4831  *                              cores with RX/TX pairs on same core and excess
4832  *                              of either following
4833  *     -          X       X     RX and TX queues mapped to consecutive cores
4834  *                              of any type with RX/TX pairs on same core and
4835  *                              excess of either following
4836  *     X          -       X     RX and TX queues mapped to consecutive physical
4837  *                              cores; all RX then all TX
4838  *     X          X       X     RX queues mapped to consecutive physical cores
4839  *                              first, then TX queues mapped to L2 neighbor of
4840  *                              the corresponding RX queue if one exists,
4841  *                              otherwise to consecutive physical cores
4842  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4843  *                              any type with RX/TX pairs on same core and excess
4844  *                              of either following
4845  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4846  *                              any type; all RX then all TX
4847  */
4848 static unsigned int
4849 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4850     bool is_tx)
4851 {
4852 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4853 	unsigned int core_index;
4854 
4855 	if (ctx->ifc_sysctl_separate_txrx) {
4856 		/*
4857 		 * When using separate CPUs for TX and RX, the assignment
4858 		 * will always be of a consecutive CPU out of the set of
4859 		 * context CPUs, except for the specific case where the
4860 		 * context CPUs are phsyical cores, the use of logical cores
4861 		 * has been enabled, the assignment is for TX, the TX qid
4862 		 * corresponds to an RX qid, and the CPU assigned to the
4863 		 * corresponding RX queue has an L2 neighbor.
4864 		 */
4865 		if (ctx->ifc_sysctl_use_logical_cores &&
4866 		    ctx->ifc_cpus_are_physical_cores &&
4867 		    is_tx && qid < scctx->isc_nrxqsets) {
4868 			int l2_neighbor;
4869 			unsigned int rx_cpuid;
4870 
4871 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4872 			l2_neighbor = find_l2_neighbor(rx_cpuid);
4873 			if (l2_neighbor != -1) {
4874 				return (l2_neighbor);
4875 			}
4876 			/*
4877 			 * ... else fall through to the normal
4878 			 * consecutive-after-RX assignment scheme.
4879 			 *
4880 			 * Note that we are assuming that all RX queue CPUs
4881 			 * have an L2 neighbor, or all do not.  If a mixed
4882 			 * scenario is possible, we will have to keep track
4883 			 * separately of how many queues prior to this one
4884 			 * were not able to be assigned to an L2 neighbor.
4885 			 */
4886 		}
4887 		if (is_tx)
4888 			core_index = scctx->isc_nrxqsets + qid;
4889 		else
4890 			core_index = qid;
4891 	} else {
4892 		core_index = qid;
4893 	}
4894 
4895 	return (cpuid_advance(ctx, base_cpuid, core_index));
4896 }
4897 
4898 static uint16_t
4899 get_ctx_core_offset(if_ctx_t ctx)
4900 {
4901 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4902 	struct cpu_offset *op;
4903 	cpuset_t assigned_cpus;
4904 	unsigned int cores_consumed;
4905 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4906 	unsigned int first_valid;
4907 	unsigned int last_valid;
4908 	unsigned int i;
4909 
4910 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4911 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4912 
4913 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4914 		/*
4915 		 * Align the user-chosen base CPU ID to the next valid CPU
4916 		 * for this device.  If the chosen base CPU ID is smaller
4917 		 * than the first valid CPU or larger than the last valid
4918 		 * CPU, we assume the user does not know what the valid
4919 		 * range is for this device and is thinking in terms of a
4920 		 * zero-based reference frame, and so we shift the given
4921 		 * value into the valid range (and wrap accordingly) so the
4922 		 * intent is translated to the proper frame of reference.
4923 		 * If the base CPU ID is within the valid first/last, but
4924 		 * does not correspond to a valid CPU, it is advanced to the
4925 		 * next valid CPU (wrapping if necessary).
4926 		 */
4927 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
4928 			/* shift from zero-based to first_valid-based */
4929 			base_cpuid += first_valid;
4930 			/* wrap to range [first_valid, last_valid] */
4931 			base_cpuid = (base_cpuid - first_valid) %
4932 			    (last_valid - first_valid + 1);
4933 		}
4934 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
4935 			/*
4936 			 * base_cpuid is in [first_valid, last_valid], but
4937 			 * not a member of the valid set.  In this case,
4938 			 * there will always be a member of the valid set
4939 			 * with a CPU ID that is greater than base_cpuid,
4940 			 * and we simply advance to it.
4941 			 */
4942 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
4943 				base_cpuid++;
4944 		}
4945 		return (base_cpuid);
4946 	}
4947 
4948 	/*
4949 	 * Determine how many cores will be consumed by performing the CPU
4950 	 * assignments and counting how many of the assigned CPUs correspond
4951 	 * to CPUs in the set of context CPUs.  This is done using the CPU
4952 	 * ID first_valid as the base CPU ID, as the base CPU must be within
4953 	 * the set of context CPUs.
4954 	 *
4955 	 * Note not all assigned CPUs will be in the set of context CPUs
4956 	 * when separate CPUs are being allocated to TX and RX queues,
4957 	 * assignment to logical cores has been enabled, the set of context
4958 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
4959 	 * neighbors of CPUs that RX queues have been mapped to - in this
4960 	 * case we do only want to count how many CPUs in the set of context
4961 	 * CPUs have been consumed, as that determines the next CPU in that
4962 	 * set to start allocating at for the next device for which
4963 	 * core_offset is not set.
4964 	 */
4965 	CPU_ZERO(&assigned_cpus);
4966 	for (i = 0; i < scctx->isc_ntxqsets; i++)
4967 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
4968 		    &assigned_cpus);
4969 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4970 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
4971 		    &assigned_cpus);
4972 	CPU_AND(&assigned_cpus, &ctx->ifc_cpus);
4973 	cores_consumed = CPU_COUNT(&assigned_cpus);
4974 
4975 	mtx_lock(&cpu_offset_mtx);
4976 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4977 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4978 			base_cpuid = op->next_cpuid;
4979 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
4980 			    cores_consumed);
4981 			MPASS(op->refcount < UINT_MAX);
4982 			op->refcount++;
4983 			break;
4984 		}
4985 	}
4986 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
4987 		base_cpuid = first_valid;
4988 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4989 		    M_NOWAIT | M_ZERO);
4990 		if (op == NULL) {
4991 			device_printf(ctx->ifc_dev,
4992 			    "allocation for cpu offset failed.\n");
4993 		} else {
4994 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
4995 			    cores_consumed);
4996 			op->refcount = 1;
4997 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4998 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4999 		}
5000 	}
5001 	mtx_unlock(&cpu_offset_mtx);
5002 
5003 	return (base_cpuid);
5004 }
5005 
5006 static void
5007 unref_ctx_core_offset(if_ctx_t ctx)
5008 {
5009 	struct cpu_offset *op, *top;
5010 
5011 	mtx_lock(&cpu_offset_mtx);
5012 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5013 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5014 			MPASS(op->refcount > 0);
5015 			op->refcount--;
5016 			if (op->refcount == 0) {
5017 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5018 				free(op, M_IFLIB);
5019 			}
5020 			break;
5021 		}
5022 	}
5023 	mtx_unlock(&cpu_offset_mtx);
5024 }
5025 
5026 int
5027 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5028 {
5029 	if_ctx_t ctx;
5030 	if_t ifp;
5031 	if_softc_ctx_t scctx;
5032 	kobjop_desc_t kobj_desc;
5033 	kobj_method_t *kobj_method;
5034 	int err, msix, rid;
5035 	int num_txd, num_rxd;
5036 
5037 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5038 
5039 	if (sc == NULL) {
5040 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5041 		device_set_softc(dev, ctx);
5042 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5043 	}
5044 
5045 	ctx->ifc_sctx = sctx;
5046 	ctx->ifc_dev = dev;
5047 	ctx->ifc_softc = sc;
5048 
5049 	if ((err = iflib_register(ctx)) != 0) {
5050 		device_printf(dev, "iflib_register failed %d\n", err);
5051 		goto fail_ctx_free;
5052 	}
5053 	iflib_add_device_sysctl_pre(ctx);
5054 
5055 	scctx = &ctx->ifc_softc_ctx;
5056 	ifp = ctx->ifc_ifp;
5057 
5058 	iflib_reset_qvalues(ctx);
5059 	CTX_LOCK(ctx);
5060 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5061 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5062 		goto fail_unlock;
5063 	}
5064 	_iflib_pre_assert(scctx);
5065 	ctx->ifc_txrx = *scctx->isc_txrx;
5066 
5067 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5068 
5069 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5070 		ctx->ifc_mediap = scctx->isc_media;
5071 
5072 #ifdef INVARIANTS
5073 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5074 		MPASS(scctx->isc_tx_csum_flags);
5075 #endif
5076 
5077 	if_setcapabilities(ifp,
5078 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5079 	if_setcapenable(ifp,
5080 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5081 
5082 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5083 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5084 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5085 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5086 
5087 	num_txd = iflib_num_tx_descs(ctx);
5088 	num_rxd = iflib_num_rx_descs(ctx);
5089 
5090 	/* XXX change for per-queue sizes */
5091 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5092 	    num_txd, num_rxd);
5093 
5094 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5095 		scctx->isc_tx_nsegments = max(1, num_txd /
5096 		    MAX_SINGLE_PACKET_FRACTION);
5097 	if (scctx->isc_tx_tso_segments_max > num_txd /
5098 	    MAX_SINGLE_PACKET_FRACTION)
5099 		scctx->isc_tx_tso_segments_max = max(1,
5100 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5101 
5102 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5103 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5104 		/*
5105 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5106 		 * but some MACs do.
5107 		 */
5108 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5109 		    IP_MAXPACKET));
5110 		/*
5111 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5112 		 * into account.  In the worst case, each of these calls will
5113 		 * add another mbuf and, thus, the requirement for another DMA
5114 		 * segment.  So for best performance, it doesn't make sense to
5115 		 * advertize a maximum of TSO segments that typically will
5116 		 * require defragmentation in iflib_encap().
5117 		 */
5118 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5119 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5120 	}
5121 	if (scctx->isc_rss_table_size == 0)
5122 		scctx->isc_rss_table_size = 64;
5123 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5124 
5125 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5126 	/* XXX format name */
5127 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5128 	    NULL, NULL, "admin");
5129 
5130 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5131 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5132 		device_printf(dev, "Unable to fetch CPU list\n");
5133 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5134 		ctx->ifc_cpus_are_physical_cores = false;
5135 	} else
5136 		ctx->ifc_cpus_are_physical_cores = true;
5137 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5138 
5139 	/*
5140 	** Now set up MSI or MSI-X, should return us the number of supported
5141 	** vectors (will be 1 for a legacy interrupt and MSI).
5142 	*/
5143 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5144 		msix = scctx->isc_vectors;
5145 	} else if (scctx->isc_msix_bar != 0)
5146 	       /*
5147 		* The simple fact that isc_msix_bar is not 0 does not mean we
5148 		* we have a good value there that is known to work.
5149 		*/
5150 		msix = iflib_msix_init(ctx);
5151 	else {
5152 		scctx->isc_vectors = 1;
5153 		scctx->isc_ntxqsets = 1;
5154 		scctx->isc_nrxqsets = 1;
5155 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5156 		msix = 0;
5157 	}
5158 	/* Get memory for the station queues */
5159 	if ((err = iflib_queues_alloc(ctx))) {
5160 		device_printf(dev, "Unable to allocate queue memory\n");
5161 		goto fail_intr_free;
5162 	}
5163 
5164 	if ((err = iflib_qset_structures_setup(ctx)))
5165 		goto fail_queues;
5166 
5167 	/*
5168 	 * Now that we know how many queues there are, get the core offset.
5169 	 */
5170 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5171 
5172 	if (msix > 1) {
5173 		/*
5174 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5175 		 * aren't the default NULL implementation.
5176 		 */
5177 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5178 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5179 		    kobj_desc);
5180 		if (kobj_method == &kobj_desc->deflt) {
5181 			device_printf(dev,
5182 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5183 			err = EOPNOTSUPP;
5184 			goto fail_queues;
5185 		}
5186 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5187 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5188 		    kobj_desc);
5189 		if (kobj_method == &kobj_desc->deflt) {
5190 			device_printf(dev,
5191 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5192 			err = EOPNOTSUPP;
5193 			goto fail_queues;
5194 		}
5195 
5196 		/*
5197 		 * Assign the MSI-X vectors.
5198 		 * Note that the default NULL ifdi_msix_intr_assign method will
5199 		 * fail here, too.
5200 		 */
5201 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5202 		if (err != 0) {
5203 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5204 			    err);
5205 			goto fail_queues;
5206 		}
5207 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5208 		rid = 0;
5209 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5210 			MPASS(msix == 1);
5211 			rid = 1;
5212 		}
5213 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5214 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5215 			goto fail_queues;
5216 		}
5217 	} else {
5218 		device_printf(dev,
5219 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5220 		err = ENODEV;
5221 		goto fail_queues;
5222 	}
5223 
5224 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5225 
5226 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5227 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5228 		goto fail_detach;
5229 	}
5230 
5231 	/*
5232 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5233 	 * This must appear after the call to ether_ifattach() because
5234 	 * ether_ifattach() sets if_hdrlen to the default value.
5235 	 */
5236 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5237 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5238 
5239 	if ((err = iflib_netmap_attach(ctx))) {
5240 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5241 		goto fail_detach;
5242 	}
5243 	*ctxp = ctx;
5244 
5245 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5246 
5247 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5248 	iflib_add_device_sysctl_post(ctx);
5249 	iflib_add_pfil(ctx);
5250 	ctx->ifc_flags |= IFC_INIT_DONE;
5251 	CTX_UNLOCK(ctx);
5252 
5253 	return (0);
5254 
5255 fail_detach:
5256 	ether_ifdetach(ctx->ifc_ifp);
5257 fail_queues:
5258 	iflib_tqg_detach(ctx);
5259 	iflib_tx_structures_free(ctx);
5260 	iflib_rx_structures_free(ctx);
5261 	IFDI_DETACH(ctx);
5262 	IFDI_QUEUES_FREE(ctx);
5263 fail_intr_free:
5264 	iflib_free_intr_mem(ctx);
5265 fail_unlock:
5266 	CTX_UNLOCK(ctx);
5267 	iflib_deregister(ctx);
5268 fail_ctx_free:
5269 	device_set_softc(ctx->ifc_dev, NULL);
5270         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5271                 free(ctx->ifc_softc, M_IFLIB);
5272         free(ctx, M_IFLIB);
5273 	return (err);
5274 }
5275 
5276 int
5277 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5278 					  struct iflib_cloneattach_ctx *clctx)
5279 {
5280 	int num_txd, num_rxd;
5281 	int err;
5282 	if_ctx_t ctx;
5283 	if_t ifp;
5284 	if_softc_ctx_t scctx;
5285 	int i;
5286 	void *sc;
5287 
5288 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5289 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5290 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5291 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5292 		ctx->ifc_flags |= IFC_PSEUDO;
5293 
5294 	ctx->ifc_sctx = sctx;
5295 	ctx->ifc_softc = sc;
5296 	ctx->ifc_dev = dev;
5297 
5298 	if ((err = iflib_register(ctx)) != 0) {
5299 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5300 		goto fail_ctx_free;
5301 	}
5302 	iflib_add_device_sysctl_pre(ctx);
5303 
5304 	scctx = &ctx->ifc_softc_ctx;
5305 	ifp = ctx->ifc_ifp;
5306 
5307 	iflib_reset_qvalues(ctx);
5308 	CTX_LOCK(ctx);
5309 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5310 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5311 		goto fail_unlock;
5312 	}
5313 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5314 		ether_gen_addr(ifp, &ctx->ifc_mac);
5315 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5316 								clctx->cc_params)) != 0) {
5317 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5318 		goto fail_unlock;
5319 	}
5320 #ifdef INVARIANTS
5321 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5322 		MPASS(scctx->isc_tx_csum_flags);
5323 #endif
5324 
5325 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5326 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5327 
5328 	ifp->if_flags |= IFF_NOGROUP;
5329 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5330 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5331 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5332 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5333 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5334 		} else {
5335 			if_attach(ctx->ifc_ifp);
5336 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5337 		}
5338 
5339 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5340 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5341 			goto fail_detach;
5342 		}
5343 		*ctxp = ctx;
5344 
5345 		/*
5346 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5347 		 * This must appear after the call to ether_ifattach() because
5348 		 * ether_ifattach() sets if_hdrlen to the default value.
5349 		 */
5350 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5351 			if_setifheaderlen(ifp,
5352 			    sizeof(struct ether_vlan_header));
5353 
5354 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5355 		iflib_add_device_sysctl_post(ctx);
5356 		ctx->ifc_flags |= IFC_INIT_DONE;
5357 		CTX_UNLOCK(ctx);
5358 		return (0);
5359 	}
5360 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5361 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5362 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5363 
5364 	_iflib_pre_assert(scctx);
5365 	ctx->ifc_txrx = *scctx->isc_txrx;
5366 
5367 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5368 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5369 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5370 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5371 
5372 	num_txd = iflib_num_tx_descs(ctx);
5373 	num_rxd = iflib_num_rx_descs(ctx);
5374 
5375 	/* XXX change for per-queue sizes */
5376 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5377 	    num_txd, num_rxd);
5378 
5379 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5380 		scctx->isc_tx_nsegments = max(1, num_txd /
5381 		    MAX_SINGLE_PACKET_FRACTION);
5382 	if (scctx->isc_tx_tso_segments_max > num_txd /
5383 	    MAX_SINGLE_PACKET_FRACTION)
5384 		scctx->isc_tx_tso_segments_max = max(1,
5385 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5386 
5387 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5388 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5389 		/*
5390 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5391 		 * but some MACs do.
5392 		 */
5393 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5394 		    IP_MAXPACKET));
5395 		/*
5396 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5397 		 * into account.  In the worst case, each of these calls will
5398 		 * add another mbuf and, thus, the requirement for another DMA
5399 		 * segment.  So for best performance, it doesn't make sense to
5400 		 * advertize a maximum of TSO segments that typically will
5401 		 * require defragmentation in iflib_encap().
5402 		 */
5403 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5404 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5405 	}
5406 	if (scctx->isc_rss_table_size == 0)
5407 		scctx->isc_rss_table_size = 64;
5408 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5409 
5410 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5411 	/* XXX format name */
5412 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5413 	    NULL, NULL, "admin");
5414 
5415 	/* XXX --- can support > 1 -- but keep it simple for now */
5416 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5417 
5418 	/* Get memory for the station queues */
5419 	if ((err = iflib_queues_alloc(ctx))) {
5420 		device_printf(dev, "Unable to allocate queue memory\n");
5421 		goto fail_iflib_detach;
5422 	}
5423 
5424 	if ((err = iflib_qset_structures_setup(ctx))) {
5425 		device_printf(dev, "qset structure setup failed %d\n", err);
5426 		goto fail_queues;
5427 	}
5428 
5429 	/*
5430 	 * XXX What if anything do we want to do about interrupts?
5431 	 */
5432 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5433 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5434 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5435 		goto fail_detach;
5436 	}
5437 
5438 	/*
5439 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5440 	 * This must appear after the call to ether_ifattach() because
5441 	 * ether_ifattach() sets if_hdrlen to the default value.
5442 	 */
5443 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5444 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5445 
5446 	/* XXX handle more than one queue */
5447 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5448 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5449 
5450 	*ctxp = ctx;
5451 
5452 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5453 	iflib_add_device_sysctl_post(ctx);
5454 	ctx->ifc_flags |= IFC_INIT_DONE;
5455 	CTX_UNLOCK(ctx);
5456 
5457 	return (0);
5458 fail_detach:
5459 	ether_ifdetach(ctx->ifc_ifp);
5460 fail_queues:
5461 	iflib_tqg_detach(ctx);
5462 	iflib_tx_structures_free(ctx);
5463 	iflib_rx_structures_free(ctx);
5464 fail_iflib_detach:
5465 	IFDI_DETACH(ctx);
5466 	IFDI_QUEUES_FREE(ctx);
5467 fail_unlock:
5468 	CTX_UNLOCK(ctx);
5469 	iflib_deregister(ctx);
5470 fail_ctx_free:
5471 	free(ctx->ifc_softc, M_IFLIB);
5472 	free(ctx, M_IFLIB);
5473 	return (err);
5474 }
5475 
5476 int
5477 iflib_pseudo_deregister(if_ctx_t ctx)
5478 {
5479 	if_t ifp = ctx->ifc_ifp;
5480 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5481 
5482 	/* Unregister VLAN event handlers early */
5483 	iflib_unregister_vlan_handlers(ctx);
5484 
5485 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5486 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5487 		bpfdetach(ifp);
5488 		if_detach(ifp);
5489 	} else {
5490 		ether_ifdetach(ifp);
5491 	}
5492 
5493 	iflib_tqg_detach(ctx);
5494 	iflib_tx_structures_free(ctx);
5495 	iflib_rx_structures_free(ctx);
5496 	IFDI_DETACH(ctx);
5497 	IFDI_QUEUES_FREE(ctx);
5498 
5499 	iflib_deregister(ctx);
5500 
5501 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5502 		free(ctx->ifc_softc, M_IFLIB);
5503 	free(ctx, M_IFLIB);
5504 	return (0);
5505 }
5506 
5507 int
5508 iflib_device_attach(device_t dev)
5509 {
5510 	if_ctx_t ctx;
5511 	if_shared_ctx_t sctx;
5512 
5513 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5514 		return (ENOTSUP);
5515 
5516 	pci_enable_busmaster(dev);
5517 
5518 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5519 }
5520 
5521 int
5522 iflib_device_deregister(if_ctx_t ctx)
5523 {
5524 	if_t ifp = ctx->ifc_ifp;
5525 	device_t dev = ctx->ifc_dev;
5526 
5527 	/* Make sure VLANS are not using driver */
5528 	if (if_vlantrunkinuse(ifp)) {
5529 		device_printf(dev, "Vlan in use, detach first\n");
5530 		return (EBUSY);
5531 	}
5532 #ifdef PCI_IOV
5533 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5534 		device_printf(dev, "SR-IOV in use; detach first.\n");
5535 		return (EBUSY);
5536 	}
5537 #endif
5538 
5539 	STATE_LOCK(ctx);
5540 	ctx->ifc_flags |= IFC_IN_DETACH;
5541 	STATE_UNLOCK(ctx);
5542 
5543 	/* Unregister VLAN handlers before calling iflib_stop() */
5544 	iflib_unregister_vlan_handlers(ctx);
5545 
5546 	iflib_netmap_detach(ifp);
5547 	ether_ifdetach(ifp);
5548 
5549 	CTX_LOCK(ctx);
5550 	iflib_stop(ctx);
5551 	CTX_UNLOCK(ctx);
5552 
5553 	iflib_rem_pfil(ctx);
5554 	if (ctx->ifc_led_dev != NULL)
5555 		led_destroy(ctx->ifc_led_dev);
5556 
5557 	iflib_tqg_detach(ctx);
5558 	iflib_tx_structures_free(ctx);
5559 	iflib_rx_structures_free(ctx);
5560 
5561 	CTX_LOCK(ctx);
5562 	IFDI_DETACH(ctx);
5563 	IFDI_QUEUES_FREE(ctx);
5564 	CTX_UNLOCK(ctx);
5565 
5566 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5567 	iflib_free_intr_mem(ctx);
5568 
5569 	bus_generic_detach(dev);
5570 
5571 	iflib_deregister(ctx);
5572 
5573 	device_set_softc(ctx->ifc_dev, NULL);
5574 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5575 		free(ctx->ifc_softc, M_IFLIB);
5576 	unref_ctx_core_offset(ctx);
5577 	free(ctx, M_IFLIB);
5578 	return (0);
5579 }
5580 
5581 static void
5582 iflib_tqg_detach(if_ctx_t ctx)
5583 {
5584 	iflib_txq_t txq;
5585 	iflib_rxq_t rxq;
5586 	int i;
5587 	struct taskqgroup *tqg;
5588 
5589 	/* XXX drain any dependent tasks */
5590 	tqg = qgroup_if_io_tqg;
5591 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5592 		callout_drain(&txq->ift_timer);
5593 #ifdef DEV_NETMAP
5594 		callout_drain(&txq->ift_netmap_timer);
5595 #endif /* DEV_NETMAP */
5596 		if (txq->ift_task.gt_uniq != NULL)
5597 			taskqgroup_detach(tqg, &txq->ift_task);
5598 	}
5599 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5600 		if (rxq->ifr_task.gt_uniq != NULL)
5601 			taskqgroup_detach(tqg, &rxq->ifr_task);
5602 	}
5603 	tqg = qgroup_if_config_tqg;
5604 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5605 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5606 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5607 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5608 }
5609 
5610 static void
5611 iflib_free_intr_mem(if_ctx_t ctx)
5612 {
5613 
5614 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5615 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5616 	}
5617 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5618 		pci_release_msi(ctx->ifc_dev);
5619 	}
5620 	if (ctx->ifc_msix_mem != NULL) {
5621 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5622 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5623 		ctx->ifc_msix_mem = NULL;
5624 	}
5625 }
5626 
5627 int
5628 iflib_device_detach(device_t dev)
5629 {
5630 	if_ctx_t ctx = device_get_softc(dev);
5631 
5632 	return (iflib_device_deregister(ctx));
5633 }
5634 
5635 int
5636 iflib_device_suspend(device_t dev)
5637 {
5638 	if_ctx_t ctx = device_get_softc(dev);
5639 
5640 	CTX_LOCK(ctx);
5641 	IFDI_SUSPEND(ctx);
5642 	CTX_UNLOCK(ctx);
5643 
5644 	return bus_generic_suspend(dev);
5645 }
5646 int
5647 iflib_device_shutdown(device_t dev)
5648 {
5649 	if_ctx_t ctx = device_get_softc(dev);
5650 
5651 	CTX_LOCK(ctx);
5652 	IFDI_SHUTDOWN(ctx);
5653 	CTX_UNLOCK(ctx);
5654 
5655 	return bus_generic_suspend(dev);
5656 }
5657 
5658 int
5659 iflib_device_resume(device_t dev)
5660 {
5661 	if_ctx_t ctx = device_get_softc(dev);
5662 	iflib_txq_t txq = ctx->ifc_txqs;
5663 
5664 	CTX_LOCK(ctx);
5665 	IFDI_RESUME(ctx);
5666 	iflib_if_init_locked(ctx);
5667 	CTX_UNLOCK(ctx);
5668 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5669 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5670 
5671 	return (bus_generic_resume(dev));
5672 }
5673 
5674 int
5675 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5676 {
5677 	int error;
5678 	if_ctx_t ctx = device_get_softc(dev);
5679 
5680 	CTX_LOCK(ctx);
5681 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5682 	CTX_UNLOCK(ctx);
5683 
5684 	return (error);
5685 }
5686 
5687 void
5688 iflib_device_iov_uninit(device_t dev)
5689 {
5690 	if_ctx_t ctx = device_get_softc(dev);
5691 
5692 	CTX_LOCK(ctx);
5693 	IFDI_IOV_UNINIT(ctx);
5694 	CTX_UNLOCK(ctx);
5695 }
5696 
5697 int
5698 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5699 {
5700 	int error;
5701 	if_ctx_t ctx = device_get_softc(dev);
5702 
5703 	CTX_LOCK(ctx);
5704 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5705 	CTX_UNLOCK(ctx);
5706 
5707 	return (error);
5708 }
5709 
5710 /*********************************************************************
5711  *
5712  *  MODULE FUNCTION DEFINITIONS
5713  *
5714  **********************************************************************/
5715 
5716 /*
5717  * - Start a fast taskqueue thread for each core
5718  * - Start a taskqueue for control operations
5719  */
5720 static int
5721 iflib_module_init(void)
5722 {
5723 	iflib_timer_default = hz / 2;
5724 	return (0);
5725 }
5726 
5727 static int
5728 iflib_module_event_handler(module_t mod, int what, void *arg)
5729 {
5730 	int err;
5731 
5732 	switch (what) {
5733 	case MOD_LOAD:
5734 		if ((err = iflib_module_init()) != 0)
5735 			return (err);
5736 		break;
5737 	case MOD_UNLOAD:
5738 		return (EBUSY);
5739 	default:
5740 		return (EOPNOTSUPP);
5741 	}
5742 
5743 	return (0);
5744 }
5745 
5746 /*********************************************************************
5747  *
5748  *  PUBLIC FUNCTION DEFINITIONS
5749  *     ordered as in iflib.h
5750  *
5751  **********************************************************************/
5752 
5753 static void
5754 _iflib_assert(if_shared_ctx_t sctx)
5755 {
5756 	int i;
5757 
5758 	MPASS(sctx->isc_tx_maxsize);
5759 	MPASS(sctx->isc_tx_maxsegsize);
5760 
5761 	MPASS(sctx->isc_rx_maxsize);
5762 	MPASS(sctx->isc_rx_nsegments);
5763 	MPASS(sctx->isc_rx_maxsegsize);
5764 
5765 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5766 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5767 		MPASS(sctx->isc_nrxd_min[i]);
5768 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5769 		MPASS(sctx->isc_nrxd_max[i]);
5770 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5771 		MPASS(sctx->isc_nrxd_default[i]);
5772 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5773 	}
5774 
5775 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5776 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5777 		MPASS(sctx->isc_ntxd_min[i]);
5778 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5779 		MPASS(sctx->isc_ntxd_max[i]);
5780 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5781 		MPASS(sctx->isc_ntxd_default[i]);
5782 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5783 	}
5784 }
5785 
5786 static void
5787 _iflib_pre_assert(if_softc_ctx_t scctx)
5788 {
5789 
5790 	MPASS(scctx->isc_txrx->ift_txd_encap);
5791 	MPASS(scctx->isc_txrx->ift_txd_flush);
5792 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5793 	MPASS(scctx->isc_txrx->ift_rxd_available);
5794 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5795 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5796 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5797 }
5798 
5799 static int
5800 iflib_register(if_ctx_t ctx)
5801 {
5802 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5803 	driver_t *driver = sctx->isc_driver;
5804 	device_t dev = ctx->ifc_dev;
5805 	if_t ifp;
5806 	u_char type;
5807 	int iflags;
5808 
5809 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5810 		_iflib_assert(sctx);
5811 
5812 	CTX_LOCK_INIT(ctx);
5813 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5814 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5815 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5816 			type = IFT_ETHER;
5817 		else
5818 			type = IFT_PPP;
5819 	} else
5820 		type = IFT_ETHER;
5821 	ifp = ctx->ifc_ifp = if_alloc(type);
5822 	if (ifp == NULL) {
5823 		device_printf(dev, "can not allocate ifnet structure\n");
5824 		return (ENOMEM);
5825 	}
5826 
5827 	/*
5828 	 * Initialize our context's device specific methods
5829 	 */
5830 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5831 	kobj_class_compile((kobj_class_t) driver);
5832 
5833 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5834 	if_setsoftc(ifp, ctx);
5835 	if_setdev(ifp, dev);
5836 	if_setinitfn(ifp, iflib_if_init);
5837 	if_setioctlfn(ifp, iflib_if_ioctl);
5838 #ifdef ALTQ
5839 	if_setstartfn(ifp, iflib_altq_if_start);
5840 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5841 	if_setsendqready(ifp);
5842 #else
5843 	if_settransmitfn(ifp, iflib_if_transmit);
5844 #endif
5845 	if_setqflushfn(ifp, iflib_if_qflush);
5846 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5847 
5848 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5849 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5850 		iflags |= IFF_POINTOPOINT;
5851 	else
5852 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5853 	if_setflags(ifp, iflags);
5854 	ctx->ifc_vlan_attach_event =
5855 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5856 							  EVENTHANDLER_PRI_FIRST);
5857 	ctx->ifc_vlan_detach_event =
5858 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5859 							  EVENTHANDLER_PRI_FIRST);
5860 
5861 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5862 		ctx->ifc_mediap = &ctx->ifc_media;
5863 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5864 		    iflib_media_change, iflib_media_status);
5865 	}
5866 	return (0);
5867 }
5868 
5869 static void
5870 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5871 {
5872 	/* Unregister VLAN events */
5873 	if (ctx->ifc_vlan_attach_event != NULL) {
5874 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5875 		ctx->ifc_vlan_attach_event = NULL;
5876 	}
5877 	if (ctx->ifc_vlan_detach_event != NULL) {
5878 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5879 		ctx->ifc_vlan_detach_event = NULL;
5880 	}
5881 
5882 }
5883 
5884 static void
5885 iflib_deregister(if_ctx_t ctx)
5886 {
5887 	if_t ifp = ctx->ifc_ifp;
5888 
5889 	/* Remove all media */
5890 	ifmedia_removeall(&ctx->ifc_media);
5891 
5892 	/* Ensure that VLAN event handlers are unregistered */
5893 	iflib_unregister_vlan_handlers(ctx);
5894 
5895 	/* Release kobject reference */
5896 	kobj_delete((kobj_t) ctx, NULL);
5897 
5898 	/* Free the ifnet structure */
5899 	if_free(ifp);
5900 
5901 	STATE_LOCK_DESTROY(ctx);
5902 
5903 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5904 	CTX_LOCK_DESTROY(ctx);
5905 }
5906 
5907 static int
5908 iflib_queues_alloc(if_ctx_t ctx)
5909 {
5910 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5911 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5912 	device_t dev = ctx->ifc_dev;
5913 	int nrxqsets = scctx->isc_nrxqsets;
5914 	int ntxqsets = scctx->isc_ntxqsets;
5915 	iflib_txq_t txq;
5916 	iflib_rxq_t rxq;
5917 	iflib_fl_t fl = NULL;
5918 	int i, j, cpu, err, txconf, rxconf;
5919 	iflib_dma_info_t ifdip;
5920 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5921 	uint32_t *txqsizes = scctx->isc_txqsizes;
5922 	uint8_t nrxqs = sctx->isc_nrxqs;
5923 	uint8_t ntxqs = sctx->isc_ntxqs;
5924 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5925 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5926 	caddr_t *vaddrs;
5927 	uint64_t *paddrs;
5928 
5929 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5930 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5931 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5932            ("there must be at least a rxq for each free list"));
5933 
5934 	/* Allocate the TX ring struct memory */
5935 	if (!(ctx->ifc_txqs =
5936 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5937 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5938 		device_printf(dev, "Unable to allocate TX ring memory\n");
5939 		err = ENOMEM;
5940 		goto fail;
5941 	}
5942 
5943 	/* Now allocate the RX */
5944 	if (!(ctx->ifc_rxqs =
5945 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5946 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5947 		device_printf(dev, "Unable to allocate RX ring memory\n");
5948 		err = ENOMEM;
5949 		goto rx_fail;
5950 	}
5951 
5952 	txq = ctx->ifc_txqs;
5953 	rxq = ctx->ifc_rxqs;
5954 
5955 	/*
5956 	 * XXX handle allocation failure
5957 	 */
5958 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5959 		/* Set up some basics */
5960 
5961 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5962 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5963 			device_printf(dev,
5964 			    "Unable to allocate TX DMA info memory\n");
5965 			err = ENOMEM;
5966 			goto err_tx_desc;
5967 		}
5968 		txq->ift_ifdi = ifdip;
5969 		for (j = 0; j < ntxqs; j++, ifdip++) {
5970 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5971 				device_printf(dev,
5972 				    "Unable to allocate TX descriptors\n");
5973 				err = ENOMEM;
5974 				goto err_tx_desc;
5975 			}
5976 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5977 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5978 		}
5979 		txq->ift_ctx = ctx;
5980 		txq->ift_id = i;
5981 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5982 			txq->ift_br_offset = 1;
5983 		} else {
5984 			txq->ift_br_offset = 0;
5985 		}
5986 
5987 		if (iflib_txsd_alloc(txq)) {
5988 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5989 			err = ENOMEM;
5990 			goto err_tx_desc;
5991 		}
5992 
5993 		/* Initialize the TX lock */
5994 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5995 		    device_get_nameunit(dev), txq->ift_id);
5996 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5997 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5998 		txq->ift_timer.c_cpu = cpu;
5999 #ifdef DEV_NETMAP
6000 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6001 		txq->ift_netmap_timer.c_cpu = cpu;
6002 #endif /* DEV_NETMAP */
6003 
6004 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6005 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6006 		if (err) {
6007 			/* XXX free any allocated rings */
6008 			device_printf(dev, "Unable to allocate buf_ring\n");
6009 			goto err_tx_desc;
6010 		}
6011 	}
6012 
6013 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6014 		/* Set up some basics */
6015 		callout_init(&rxq->ifr_watchdog, 1);
6016 
6017 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6018 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6019 			device_printf(dev,
6020 			    "Unable to allocate RX DMA info memory\n");
6021 			err = ENOMEM;
6022 			goto err_tx_desc;
6023 		}
6024 
6025 		rxq->ifr_ifdi = ifdip;
6026 		/* XXX this needs to be changed if #rx queues != #tx queues */
6027 		rxq->ifr_ntxqirq = 1;
6028 		rxq->ifr_txqid[0] = i;
6029 		for (j = 0; j < nrxqs; j++, ifdip++) {
6030 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6031 				device_printf(dev,
6032 				    "Unable to allocate RX descriptors\n");
6033 				err = ENOMEM;
6034 				goto err_tx_desc;
6035 			}
6036 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6037 		}
6038 		rxq->ifr_ctx = ctx;
6039 		rxq->ifr_id = i;
6040 		rxq->ifr_fl_offset = fl_offset;
6041 		rxq->ifr_nfl = nfree_lists;
6042 		if (!(fl =
6043 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6044 			device_printf(dev, "Unable to allocate free list memory\n");
6045 			err = ENOMEM;
6046 			goto err_tx_desc;
6047 		}
6048 		rxq->ifr_fl = fl;
6049 		for (j = 0; j < nfree_lists; j++) {
6050 			fl[j].ifl_rxq = rxq;
6051 			fl[j].ifl_id = j;
6052 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6053 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6054 		}
6055 		/* Allocate receive buffers for the ring */
6056 		if (iflib_rxsd_alloc(rxq)) {
6057 			device_printf(dev,
6058 			    "Critical Failure setting up receive buffers\n");
6059 			err = ENOMEM;
6060 			goto err_rx_desc;
6061 		}
6062 
6063 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6064 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6065 			    M_WAITOK);
6066 	}
6067 
6068 	/* TXQs */
6069 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6070 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6071 	for (i = 0; i < ntxqsets; i++) {
6072 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6073 
6074 		for (j = 0; j < ntxqs; j++, di++) {
6075 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6076 			paddrs[i*ntxqs + j] = di->idi_paddr;
6077 		}
6078 	}
6079 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6080 		device_printf(ctx->ifc_dev,
6081 		    "Unable to allocate device TX queue\n");
6082 		iflib_tx_structures_free(ctx);
6083 		free(vaddrs, M_IFLIB);
6084 		free(paddrs, M_IFLIB);
6085 		goto err_rx_desc;
6086 	}
6087 	free(vaddrs, M_IFLIB);
6088 	free(paddrs, M_IFLIB);
6089 
6090 	/* RXQs */
6091 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6092 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6093 	for (i = 0; i < nrxqsets; i++) {
6094 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6095 
6096 		for (j = 0; j < nrxqs; j++, di++) {
6097 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6098 			paddrs[i*nrxqs + j] = di->idi_paddr;
6099 		}
6100 	}
6101 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6102 		device_printf(ctx->ifc_dev,
6103 		    "Unable to allocate device RX queue\n");
6104 		iflib_tx_structures_free(ctx);
6105 		free(vaddrs, M_IFLIB);
6106 		free(paddrs, M_IFLIB);
6107 		goto err_rx_desc;
6108 	}
6109 	free(vaddrs, M_IFLIB);
6110 	free(paddrs, M_IFLIB);
6111 
6112 	return (0);
6113 
6114 /* XXX handle allocation failure changes */
6115 err_rx_desc:
6116 err_tx_desc:
6117 rx_fail:
6118 	if (ctx->ifc_rxqs != NULL)
6119 		free(ctx->ifc_rxqs, M_IFLIB);
6120 	ctx->ifc_rxqs = NULL;
6121 	if (ctx->ifc_txqs != NULL)
6122 		free(ctx->ifc_txqs, M_IFLIB);
6123 	ctx->ifc_txqs = NULL;
6124 fail:
6125 	return (err);
6126 }
6127 
6128 static int
6129 iflib_tx_structures_setup(if_ctx_t ctx)
6130 {
6131 	iflib_txq_t txq = ctx->ifc_txqs;
6132 	int i;
6133 
6134 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6135 		iflib_txq_setup(txq);
6136 
6137 	return (0);
6138 }
6139 
6140 static void
6141 iflib_tx_structures_free(if_ctx_t ctx)
6142 {
6143 	iflib_txq_t txq = ctx->ifc_txqs;
6144 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6145 	int i, j;
6146 
6147 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6148 		for (j = 0; j < sctx->isc_ntxqs; j++)
6149 			iflib_dma_free(&txq->ift_ifdi[j]);
6150 		iflib_txq_destroy(txq);
6151 	}
6152 	free(ctx->ifc_txqs, M_IFLIB);
6153 	ctx->ifc_txqs = NULL;
6154 }
6155 
6156 /*********************************************************************
6157  *
6158  *  Initialize all receive rings.
6159  *
6160  **********************************************************************/
6161 static int
6162 iflib_rx_structures_setup(if_ctx_t ctx)
6163 {
6164 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6165 	int q;
6166 #if defined(INET6) || defined(INET)
6167 	int err, i;
6168 #endif
6169 
6170 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6171 #if defined(INET6) || defined(INET)
6172 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6173 		    TCP_LRO_ENTRIES, min(1024,
6174 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6175 		if (err != 0) {
6176 			device_printf(ctx->ifc_dev,
6177 			    "LRO Initialization failed!\n");
6178 			goto fail;
6179 		}
6180 #endif
6181 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6182 	}
6183 	return (0);
6184 #if defined(INET6) || defined(INET)
6185 fail:
6186 	/*
6187 	 * Free LRO resources allocated so far, we will only handle
6188 	 * the rings that completed, the failing case will have
6189 	 * cleaned up for itself.  'q' failed, so its the terminus.
6190 	 */
6191 	rxq = ctx->ifc_rxqs;
6192 	for (i = 0; i < q; ++i, rxq++) {
6193 		tcp_lro_free(&rxq->ifr_lc);
6194 	}
6195 	return (err);
6196 #endif
6197 }
6198 
6199 /*********************************************************************
6200  *
6201  *  Free all receive rings.
6202  *
6203  **********************************************************************/
6204 static void
6205 iflib_rx_structures_free(if_ctx_t ctx)
6206 {
6207 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6208 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6209 	int i, j;
6210 
6211 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6212 		for (j = 0; j < sctx->isc_nrxqs; j++)
6213 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6214 		iflib_rx_sds_free(rxq);
6215 #if defined(INET6) || defined(INET)
6216 		tcp_lro_free(&rxq->ifr_lc);
6217 #endif
6218 	}
6219 	free(ctx->ifc_rxqs, M_IFLIB);
6220 	ctx->ifc_rxqs = NULL;
6221 }
6222 
6223 static int
6224 iflib_qset_structures_setup(if_ctx_t ctx)
6225 {
6226 	int err;
6227 
6228 	/*
6229 	 * It is expected that the caller takes care of freeing queues if this
6230 	 * fails.
6231 	 */
6232 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6233 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6234 		return (err);
6235 	}
6236 
6237 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6238 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6239 
6240 	return (err);
6241 }
6242 
6243 int
6244 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6245 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6246 {
6247 
6248 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6249 }
6250 
6251 /* Just to avoid copy/paste */
6252 static inline int
6253 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6254     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6255     const char *name)
6256 {
6257 	device_t dev;
6258 	unsigned int base_cpuid, cpuid;
6259 	int err;
6260 
6261 	dev = ctx->ifc_dev;
6262 	base_cpuid = ctx->ifc_sysctl_core_offset;
6263 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6264 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6265 	    irq ? irq->ii_res : NULL, name);
6266 	if (err) {
6267 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6268 		return (err);
6269 	}
6270 #ifdef notyet
6271 	if (cpuid > ctx->ifc_cpuid_highest)
6272 		ctx->ifc_cpuid_highest = cpuid;
6273 #endif
6274 	return (0);
6275 }
6276 
6277 int
6278 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6279 			iflib_intr_type_t type, driver_filter_t *filter,
6280 			void *filter_arg, int qid, const char *name)
6281 {
6282 	device_t dev;
6283 	struct grouptask *gtask;
6284 	struct taskqgroup *tqg;
6285 	iflib_filter_info_t info;
6286 	gtask_fn_t *fn;
6287 	int tqrid, err;
6288 	driver_filter_t *intr_fast;
6289 	void *q;
6290 
6291 	info = &ctx->ifc_filter_info;
6292 	tqrid = rid;
6293 
6294 	switch (type) {
6295 	/* XXX merge tx/rx for netmap? */
6296 	case IFLIB_INTR_TX:
6297 		q = &ctx->ifc_txqs[qid];
6298 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6299 		gtask = &ctx->ifc_txqs[qid].ift_task;
6300 		tqg = qgroup_if_io_tqg;
6301 		fn = _task_fn_tx;
6302 		intr_fast = iflib_fast_intr;
6303 		GROUPTASK_INIT(gtask, 0, fn, q);
6304 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6305 		break;
6306 	case IFLIB_INTR_RX:
6307 		q = &ctx->ifc_rxqs[qid];
6308 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6309 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6310 		tqg = qgroup_if_io_tqg;
6311 		fn = _task_fn_rx;
6312 		intr_fast = iflib_fast_intr;
6313 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6314 		break;
6315 	case IFLIB_INTR_RXTX:
6316 		q = &ctx->ifc_rxqs[qid];
6317 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6318 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6319 		tqg = qgroup_if_io_tqg;
6320 		fn = _task_fn_rx;
6321 		intr_fast = iflib_fast_intr_rxtx;
6322 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6323 		break;
6324 	case IFLIB_INTR_ADMIN:
6325 		q = ctx;
6326 		tqrid = -1;
6327 		info = &ctx->ifc_filter_info;
6328 		gtask = &ctx->ifc_admin_task;
6329 		tqg = qgroup_if_config_tqg;
6330 		fn = _task_fn_admin;
6331 		intr_fast = iflib_fast_intr_ctx;
6332 		break;
6333 	default:
6334 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6335 		    __func__);
6336 		return (EINVAL);
6337 	}
6338 
6339 	info->ifi_filter = filter;
6340 	info->ifi_filter_arg = filter_arg;
6341 	info->ifi_task = gtask;
6342 	info->ifi_ctx = q;
6343 
6344 	dev = ctx->ifc_dev;
6345 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6346 	if (err != 0) {
6347 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6348 		return (err);
6349 	}
6350 	if (type == IFLIB_INTR_ADMIN)
6351 		return (0);
6352 
6353 	if (tqrid != -1) {
6354 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6355 		    name);
6356 		if (err)
6357 			return (err);
6358 	} else {
6359 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6360 	}
6361 
6362 	return (0);
6363 }
6364 
6365 void
6366 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6367 {
6368 	device_t dev;
6369 	struct grouptask *gtask;
6370 	struct taskqgroup *tqg;
6371 	gtask_fn_t *fn;
6372 	void *q;
6373 	int err;
6374 
6375 	switch (type) {
6376 	case IFLIB_INTR_TX:
6377 		q = &ctx->ifc_txqs[qid];
6378 		gtask = &ctx->ifc_txqs[qid].ift_task;
6379 		tqg = qgroup_if_io_tqg;
6380 		fn = _task_fn_tx;
6381 		GROUPTASK_INIT(gtask, 0, fn, q);
6382 		break;
6383 	case IFLIB_INTR_RX:
6384 		q = &ctx->ifc_rxqs[qid];
6385 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6386 		tqg = qgroup_if_io_tqg;
6387 		fn = _task_fn_rx;
6388 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6389 		break;
6390 	case IFLIB_INTR_IOV:
6391 		q = ctx;
6392 		gtask = &ctx->ifc_vflr_task;
6393 		tqg = qgroup_if_config_tqg;
6394 		fn = _task_fn_iov;
6395 		GROUPTASK_INIT(gtask, 0, fn, q);
6396 		break;
6397 	default:
6398 		panic("unknown net intr type");
6399 	}
6400 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6401 	if (err) {
6402 		dev = ctx->ifc_dev;
6403 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6404 		    name);
6405 	}
6406 }
6407 
6408 void
6409 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6410 {
6411 
6412 	if (irq->ii_tag)
6413 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6414 
6415 	if (irq->ii_res)
6416 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6417 		    rman_get_rid(irq->ii_res), irq->ii_res);
6418 }
6419 
6420 static int
6421 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6422 {
6423 	iflib_txq_t txq = ctx->ifc_txqs;
6424 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6425 	if_irq_t irq = &ctx->ifc_legacy_irq;
6426 	iflib_filter_info_t info;
6427 	device_t dev;
6428 	struct grouptask *gtask;
6429 	struct resource *res;
6430 	struct taskqgroup *tqg;
6431 	void *q;
6432 	int err, tqrid;
6433 	bool rx_only;
6434 
6435 	q = &ctx->ifc_rxqs[0];
6436 	info = &rxq[0].ifr_filter_info;
6437 	gtask = &rxq[0].ifr_task;
6438 	tqg = qgroup_if_io_tqg;
6439 	tqrid = *rid;
6440 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6441 
6442 	ctx->ifc_flags |= IFC_LEGACY;
6443 	info->ifi_filter = filter;
6444 	info->ifi_filter_arg = filter_arg;
6445 	info->ifi_task = gtask;
6446 	info->ifi_ctx = rx_only ? ctx : q;
6447 
6448 	dev = ctx->ifc_dev;
6449 	/* We allocate a single interrupt resource */
6450 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6451 	    iflib_fast_intr_rxtx, NULL, info, name);
6452 	if (err != 0)
6453 		return (err);
6454 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6455 	res = irq->ii_res;
6456 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6457 
6458 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6459 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6460 	    "tx");
6461 	return (0);
6462 }
6463 
6464 void
6465 iflib_led_create(if_ctx_t ctx)
6466 {
6467 
6468 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6469 	    device_get_nameunit(ctx->ifc_dev));
6470 }
6471 
6472 void
6473 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6474 {
6475 
6476 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6477 }
6478 
6479 void
6480 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6481 {
6482 
6483 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6484 }
6485 
6486 void
6487 iflib_admin_intr_deferred(if_ctx_t ctx)
6488 {
6489 
6490 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6491 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6492 }
6493 
6494 void
6495 iflib_iov_intr_deferred(if_ctx_t ctx)
6496 {
6497 
6498 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6499 }
6500 
6501 void
6502 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6503 {
6504 
6505 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6506 	    name);
6507 }
6508 
6509 void
6510 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6511 	const char *name)
6512 {
6513 
6514 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6515 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6516 	    name);
6517 }
6518 
6519 void
6520 iflib_config_gtask_deinit(struct grouptask *gtask)
6521 {
6522 
6523 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6524 }
6525 
6526 void
6527 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6528 {
6529 	if_t ifp = ctx->ifc_ifp;
6530 	iflib_txq_t txq = ctx->ifc_txqs;
6531 
6532 	if_setbaudrate(ifp, baudrate);
6533 	if (baudrate >= IF_Gbps(10)) {
6534 		STATE_LOCK(ctx);
6535 		ctx->ifc_flags |= IFC_PREFETCH;
6536 		STATE_UNLOCK(ctx);
6537 	}
6538 	/* If link down, disable watchdog */
6539 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6540 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6541 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6542 	}
6543 	ctx->ifc_link_state = link_state;
6544 	if_link_state_change(ifp, link_state);
6545 }
6546 
6547 static int
6548 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6549 {
6550 	int credits;
6551 #ifdef INVARIANTS
6552 	int credits_pre = txq->ift_cidx_processed;
6553 #endif
6554 
6555 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6556 	    BUS_DMASYNC_POSTREAD);
6557 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6558 		return (0);
6559 
6560 	txq->ift_processed += credits;
6561 	txq->ift_cidx_processed += credits;
6562 
6563 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6564 	if (txq->ift_cidx_processed >= txq->ift_size)
6565 		txq->ift_cidx_processed -= txq->ift_size;
6566 	return (credits);
6567 }
6568 
6569 static int
6570 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6571 {
6572 	iflib_fl_t fl;
6573 	u_int i;
6574 
6575 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6576 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6577 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6578 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6579 	    budget));
6580 }
6581 
6582 void
6583 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6584 	const char *description, if_int_delay_info_t info,
6585 	int offset, int value)
6586 {
6587 	info->iidi_ctx = ctx;
6588 	info->iidi_offset = offset;
6589 	info->iidi_value = value;
6590 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6591 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6592 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6593 	    info, 0, iflib_sysctl_int_delay, "I", description);
6594 }
6595 
6596 struct sx *
6597 iflib_ctx_lock_get(if_ctx_t ctx)
6598 {
6599 
6600 	return (&ctx->ifc_ctx_sx);
6601 }
6602 
6603 static int
6604 iflib_msix_init(if_ctx_t ctx)
6605 {
6606 	device_t dev = ctx->ifc_dev;
6607 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6608 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6609 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6610 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6611 
6612 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6613 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6614 
6615 	if (bootverbose)
6616 		device_printf(dev, "msix_init qsets capped at %d\n",
6617 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6618 
6619 	/* Override by tuneable */
6620 	if (scctx->isc_disable_msix)
6621 		goto msi;
6622 
6623 	/* First try MSI-X */
6624 	if ((msgs = pci_msix_count(dev)) == 0) {
6625 		if (bootverbose)
6626 			device_printf(dev, "MSI-X not supported or disabled\n");
6627 		goto msi;
6628 	}
6629 
6630 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6631 	/*
6632 	 * bar == -1 => "trust me I know what I'm doing"
6633 	 * Some drivers are for hardware that is so shoddily
6634 	 * documented that no one knows which bars are which
6635 	 * so the developer has to map all bars. This hack
6636 	 * allows shoddy garbage to use MSI-X in this framework.
6637 	 */
6638 	if (bar != -1) {
6639 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6640 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6641 		if (ctx->ifc_msix_mem == NULL) {
6642 			device_printf(dev, "Unable to map MSI-X table\n");
6643 			goto msi;
6644 		}
6645 	}
6646 
6647 	admincnt = sctx->isc_admin_intrcnt;
6648 #if IFLIB_DEBUG
6649 	/* use only 1 qset in debug mode */
6650 	queuemsgs = min(msgs - admincnt, 1);
6651 #else
6652 	queuemsgs = msgs - admincnt;
6653 #endif
6654 #ifdef RSS
6655 	queues = imin(queuemsgs, rss_getnumbuckets());
6656 #else
6657 	queues = queuemsgs;
6658 #endif
6659 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6660 	if (bootverbose)
6661 		device_printf(dev,
6662 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6663 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6664 #ifdef  RSS
6665 	/* If we're doing RSS, clamp at the number of RSS buckets */
6666 	if (queues > rss_getnumbuckets())
6667 		queues = rss_getnumbuckets();
6668 #endif
6669 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6670 		rx_queues = iflib_num_rx_queues;
6671 	else
6672 		rx_queues = queues;
6673 
6674 	if (rx_queues > scctx->isc_nrxqsets)
6675 		rx_queues = scctx->isc_nrxqsets;
6676 
6677 	/*
6678 	 * We want this to be all logical CPUs by default
6679 	 */
6680 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6681 		tx_queues = iflib_num_tx_queues;
6682 	else
6683 		tx_queues = mp_ncpus;
6684 
6685 	if (tx_queues > scctx->isc_ntxqsets)
6686 		tx_queues = scctx->isc_ntxqsets;
6687 
6688 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6689 #ifdef INVARIANTS
6690 		if (tx_queues != rx_queues)
6691 			device_printf(dev,
6692 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6693 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6694 #endif
6695 		tx_queues = min(rx_queues, tx_queues);
6696 		rx_queues = min(rx_queues, tx_queues);
6697 	}
6698 
6699 	vectors = rx_queues + admincnt;
6700 	if (msgs < vectors) {
6701 		device_printf(dev,
6702 		    "insufficient number of MSI-X vectors "
6703 		    "(supported %d, need %d)\n", msgs, vectors);
6704 		goto msi;
6705 	}
6706 
6707 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6708 	    tx_queues);
6709 	msgs = vectors;
6710 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6711 		if (vectors != msgs) {
6712 			device_printf(dev,
6713 			    "Unable to allocate sufficient MSI-X vectors "
6714 			    "(got %d, need %d)\n", vectors, msgs);
6715 			pci_release_msi(dev);
6716 			if (bar != -1) {
6717 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6718 				    ctx->ifc_msix_mem);
6719 				ctx->ifc_msix_mem = NULL;
6720 			}
6721 			goto msi;
6722 		}
6723 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6724 		    vectors);
6725 		scctx->isc_vectors = vectors;
6726 		scctx->isc_nrxqsets = rx_queues;
6727 		scctx->isc_ntxqsets = tx_queues;
6728 		scctx->isc_intr = IFLIB_INTR_MSIX;
6729 
6730 		return (vectors);
6731 	} else {
6732 		device_printf(dev,
6733 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6734 		    err);
6735 		if (bar != -1) {
6736 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6737 			    ctx->ifc_msix_mem);
6738 			ctx->ifc_msix_mem = NULL;
6739 		}
6740 	}
6741 
6742 msi:
6743 	vectors = pci_msi_count(dev);
6744 	scctx->isc_nrxqsets = 1;
6745 	scctx->isc_ntxqsets = 1;
6746 	scctx->isc_vectors = vectors;
6747 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6748 		device_printf(dev,"Using an MSI interrupt\n");
6749 		scctx->isc_intr = IFLIB_INTR_MSI;
6750 	} else {
6751 		scctx->isc_vectors = 1;
6752 		device_printf(dev,"Using a Legacy interrupt\n");
6753 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6754 	}
6755 
6756 	return (vectors);
6757 }
6758 
6759 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6760 
6761 static int
6762 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6763 {
6764 	int rc;
6765 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6766 	struct sbuf *sb;
6767 	const char *ring_state = "UNKNOWN";
6768 
6769 	/* XXX needed ? */
6770 	rc = sysctl_wire_old_buffer(req, 0);
6771 	MPASS(rc == 0);
6772 	if (rc != 0)
6773 		return (rc);
6774 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6775 	MPASS(sb != NULL);
6776 	if (sb == NULL)
6777 		return (ENOMEM);
6778 	if (state[3] <= 3)
6779 		ring_state = ring_states[state[3]];
6780 
6781 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6782 		    state[0], state[1], state[2], ring_state);
6783 	rc = sbuf_finish(sb);
6784 	sbuf_delete(sb);
6785         return(rc);
6786 }
6787 
6788 enum iflib_ndesc_handler {
6789 	IFLIB_NTXD_HANDLER,
6790 	IFLIB_NRXD_HANDLER,
6791 };
6792 
6793 static int
6794 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6795 {
6796 	if_ctx_t ctx = (void *)arg1;
6797 	enum iflib_ndesc_handler type = arg2;
6798 	char buf[256] = {0};
6799 	qidx_t *ndesc;
6800 	char *p, *next;
6801 	int nqs, rc, i;
6802 
6803 	nqs = 8;
6804 	switch(type) {
6805 	case IFLIB_NTXD_HANDLER:
6806 		ndesc = ctx->ifc_sysctl_ntxds;
6807 		if (ctx->ifc_sctx)
6808 			nqs = ctx->ifc_sctx->isc_ntxqs;
6809 		break;
6810 	case IFLIB_NRXD_HANDLER:
6811 		ndesc = ctx->ifc_sysctl_nrxds;
6812 		if (ctx->ifc_sctx)
6813 			nqs = ctx->ifc_sctx->isc_nrxqs;
6814 		break;
6815 	default:
6816 		printf("%s: unhandled type\n", __func__);
6817 		return (EINVAL);
6818 	}
6819 	if (nqs == 0)
6820 		nqs = 8;
6821 
6822 	for (i=0; i<8; i++) {
6823 		if (i >= nqs)
6824 			break;
6825 		if (i)
6826 			strcat(buf, ",");
6827 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6828 	}
6829 
6830 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6831 	if (rc || req->newptr == NULL)
6832 		return rc;
6833 
6834 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6835 	    i++, p = strsep(&next, " ,")) {
6836 		ndesc[i] = strtoul(p, NULL, 10);
6837 	}
6838 
6839 	return(rc);
6840 }
6841 
6842 #define NAME_BUFLEN 32
6843 static void
6844 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6845 {
6846         device_t dev = iflib_get_dev(ctx);
6847 	struct sysctl_oid_list *child, *oid_list;
6848 	struct sysctl_ctx_list *ctx_list;
6849 	struct sysctl_oid *node;
6850 
6851 	ctx_list = device_get_sysctl_ctx(dev);
6852 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6853 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6854 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6855 	oid_list = SYSCTL_CHILDREN(node);
6856 
6857 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6858 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6859 		       "driver version");
6860 
6861 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6862 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6863 			"# of txqs to use, 0 => use default #");
6864 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6865 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6866 			"# of rxqs to use, 0 => use default #");
6867 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6868 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6869                        "permit #txq != #rxq");
6870 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6871                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6872                       "disable MSI-X (default 0)");
6873 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6874 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6875 		       "set the RX budget");
6876 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6877 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6878 		       "cause TX to abdicate instead of running to completion");
6879 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6880 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6881 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6882 		       "offset to start using cores at");
6883 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6884 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6885 		       "use separate cores for TX and RX");
6886 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6887 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6888 		      "try to make use of logical cores for TX and RX");
6889 
6890 	/* XXX change for per-queue sizes */
6891 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6892 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6893 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6894 	    "list of # of TX descriptors to use, 0 = use default #");
6895 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6896 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6897 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6898 	    "list of # of RX descriptors to use, 0 = use default #");
6899 }
6900 
6901 static void
6902 iflib_add_device_sysctl_post(if_ctx_t ctx)
6903 {
6904 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6905 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6906         device_t dev = iflib_get_dev(ctx);
6907 	struct sysctl_oid_list *child;
6908 	struct sysctl_ctx_list *ctx_list;
6909 	iflib_fl_t fl;
6910 	iflib_txq_t txq;
6911 	iflib_rxq_t rxq;
6912 	int i, j;
6913 	char namebuf[NAME_BUFLEN];
6914 	char *qfmt;
6915 	struct sysctl_oid *queue_node, *fl_node, *node;
6916 	struct sysctl_oid_list *queue_list, *fl_list;
6917 	ctx_list = device_get_sysctl_ctx(dev);
6918 
6919 	node = ctx->ifc_sysctl_node;
6920 	child = SYSCTL_CHILDREN(node);
6921 
6922 	if (scctx->isc_ntxqsets > 100)
6923 		qfmt = "txq%03d";
6924 	else if (scctx->isc_ntxqsets > 10)
6925 		qfmt = "txq%02d";
6926 	else
6927 		qfmt = "txq%d";
6928 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6929 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6930 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6931 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6932 		queue_list = SYSCTL_CHILDREN(queue_node);
6933 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6934 			       CTLFLAG_RD,
6935 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
6936 #if MEMORY_LOGGING
6937 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6938 				CTLFLAG_RD,
6939 				&txq->ift_dequeued, "total mbufs freed");
6940 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6941 				CTLFLAG_RD,
6942 				&txq->ift_enqueued, "total mbufs enqueued");
6943 #endif
6944 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6945 				   CTLFLAG_RD,
6946 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6947 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6948 				   CTLFLAG_RD,
6949 				   &txq->ift_pullups, "# of times m_pullup was called");
6950 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6951 				   CTLFLAG_RD,
6952 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6953 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6954 				   CTLFLAG_RD,
6955 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6956 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6957 				   CTLFLAG_RD,
6958 				   &txq->ift_map_failed, "# of times DMA map failed");
6959 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6960 				   CTLFLAG_RD,
6961 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6962 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6963 				   CTLFLAG_RD,
6964 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6965 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6966 				   CTLFLAG_RD,
6967 				   &txq->ift_pidx, 1, "Producer Index");
6968 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6969 				   CTLFLAG_RD,
6970 				   &txq->ift_cidx, 1, "Consumer Index");
6971 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6972 				   CTLFLAG_RD,
6973 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6974 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6975 				   CTLFLAG_RD,
6976 				   &txq->ift_in_use, 1, "descriptors in use");
6977 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6978 				   CTLFLAG_RD,
6979 				   &txq->ift_processed, "descriptors procesed for clean");
6980 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6981 				   CTLFLAG_RD,
6982 				   &txq->ift_cleaned, "total cleaned");
6983 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6984 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6985 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6986 		    mp_ring_state_handler, "A", "soft ring state");
6987 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6988 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6989 				       "# of enqueues to the mp_ring for this queue");
6990 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6991 				       CTLFLAG_RD, &txq->ift_br->drops,
6992 				       "# of drops in the mp_ring for this queue");
6993 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6994 				       CTLFLAG_RD, &txq->ift_br->starts,
6995 				       "# of normal consumer starts in the mp_ring for this queue");
6996 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6997 				       CTLFLAG_RD, &txq->ift_br->stalls,
6998 					       "# of consumer stalls in the mp_ring for this queue");
6999 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7000 			       CTLFLAG_RD, &txq->ift_br->restarts,
7001 				       "# of consumer restarts in the mp_ring for this queue");
7002 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7003 				       CTLFLAG_RD, &txq->ift_br->abdications,
7004 				       "# of consumer abdications in the mp_ring for this queue");
7005 	}
7006 
7007 	if (scctx->isc_nrxqsets > 100)
7008 		qfmt = "rxq%03d";
7009 	else if (scctx->isc_nrxqsets > 10)
7010 		qfmt = "rxq%02d";
7011 	else
7012 		qfmt = "rxq%d";
7013 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7014 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7015 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7016 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7017 		queue_list = SYSCTL_CHILDREN(queue_node);
7018 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7019 			       CTLFLAG_RD,
7020 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7021 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7022 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7023 				       CTLFLAG_RD,
7024 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7025 		}
7026 
7027 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7028 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7029 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7030 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7031 			fl_list = SYSCTL_CHILDREN(fl_node);
7032 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7033 				       CTLFLAG_RD,
7034 				       &fl->ifl_pidx, 1, "Producer Index");
7035 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7036 				       CTLFLAG_RD,
7037 				       &fl->ifl_cidx, 1, "Consumer Index");
7038 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7039 				       CTLFLAG_RD,
7040 				       &fl->ifl_credits, 1, "credits available");
7041 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7042 				       CTLFLAG_RD,
7043 				       &fl->ifl_buf_size, 1, "buffer size");
7044 #if MEMORY_LOGGING
7045 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7046 					CTLFLAG_RD,
7047 					&fl->ifl_m_enqueued, "mbufs allocated");
7048 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7049 					CTLFLAG_RD,
7050 					&fl->ifl_m_dequeued, "mbufs freed");
7051 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7052 					CTLFLAG_RD,
7053 					&fl->ifl_cl_enqueued, "clusters allocated");
7054 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7055 					CTLFLAG_RD,
7056 					&fl->ifl_cl_dequeued, "clusters freed");
7057 #endif
7058 		}
7059 	}
7060 
7061 }
7062 
7063 void
7064 iflib_request_reset(if_ctx_t ctx)
7065 {
7066 
7067 	STATE_LOCK(ctx);
7068 	ctx->ifc_flags |= IFC_DO_RESET;
7069 	STATE_UNLOCK(ctx);
7070 }
7071 
7072 #ifndef __NO_STRICT_ALIGNMENT
7073 static struct mbuf *
7074 iflib_fixup_rx(struct mbuf *m)
7075 {
7076 	struct mbuf *n;
7077 
7078 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7079 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7080 		m->m_data += ETHER_HDR_LEN;
7081 		n = m;
7082 	} else {
7083 		MGETHDR(n, M_NOWAIT, MT_DATA);
7084 		if (n == NULL) {
7085 			m_freem(m);
7086 			return (NULL);
7087 		}
7088 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7089 		m->m_data += ETHER_HDR_LEN;
7090 		m->m_len -= ETHER_HDR_LEN;
7091 		n->m_len = ETHER_HDR_LEN;
7092 		M_MOVE_PKTHDR(n, m);
7093 		n->m_next = m;
7094 	}
7095 	return (n);
7096 }
7097 #endif
7098 
7099 #ifdef DEBUGNET
7100 static void
7101 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7102 {
7103 	if_ctx_t ctx;
7104 
7105 	ctx = if_getsoftc(ifp);
7106 	CTX_LOCK(ctx);
7107 	*nrxr = NRXQSETS(ctx);
7108 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7109 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7110 	CTX_UNLOCK(ctx);
7111 }
7112 
7113 static void
7114 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7115 {
7116 	if_ctx_t ctx;
7117 	if_softc_ctx_t scctx;
7118 	iflib_fl_t fl;
7119 	iflib_rxq_t rxq;
7120 	int i, j;
7121 
7122 	ctx = if_getsoftc(ifp);
7123 	scctx = &ctx->ifc_softc_ctx;
7124 
7125 	switch (event) {
7126 	case DEBUGNET_START:
7127 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7128 			rxq = &ctx->ifc_rxqs[i];
7129 			for (j = 0; j < rxq->ifr_nfl; j++) {
7130 				fl = rxq->ifr_fl;
7131 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7132 			}
7133 		}
7134 		iflib_no_tx_batch = 1;
7135 		break;
7136 	default:
7137 		break;
7138 	}
7139 }
7140 
7141 static int
7142 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7143 {
7144 	if_ctx_t ctx;
7145 	iflib_txq_t txq;
7146 	int error;
7147 
7148 	ctx = if_getsoftc(ifp);
7149 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7150 	    IFF_DRV_RUNNING)
7151 		return (EBUSY);
7152 
7153 	txq = &ctx->ifc_txqs[0];
7154 	error = iflib_encap(txq, &m);
7155 	if (error == 0)
7156 		(void)iflib_txd_db_check(txq, true);
7157 	return (error);
7158 }
7159 
7160 static int
7161 iflib_debugnet_poll(if_t ifp, int count)
7162 {
7163 	struct epoch_tracker et;
7164 	if_ctx_t ctx;
7165 	if_softc_ctx_t scctx;
7166 	iflib_txq_t txq;
7167 	int i;
7168 
7169 	ctx = if_getsoftc(ifp);
7170 	scctx = &ctx->ifc_softc_ctx;
7171 
7172 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7173 	    IFF_DRV_RUNNING)
7174 		return (EBUSY);
7175 
7176 	txq = &ctx->ifc_txqs[0];
7177 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7178 
7179 	NET_EPOCH_ENTER(et);
7180 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7181 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7182 	NET_EPOCH_EXIT(et);
7183 	return (0);
7184 }
7185 #endif /* DEBUGNET */
7186