xref: /freebsd/sys/net/iflib.c (revision 3dd5524264095ed8612c28908e13f80668eff2f9)
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_private.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/debugnet.h>
64 #include <net/pfil.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet6/ip6_var.h>
77 
78 #include <machine/bus.h>
79 #include <machine/in_cksum.h>
80 
81 #include <vm/vm.h>
82 #include <vm/pmap.h>
83 
84 #include <dev/led/led.h>
85 #include <dev/pci/pcireg.h>
86 #include <dev/pci/pcivar.h>
87 #include <dev/pci/pci_private.h>
88 
89 #include <net/iflib.h>
90 #include <net/iflib_private.h>
91 
92 #include "ifdi_if.h"
93 
94 #ifdef PCI_IOV
95 #include <dev/pci/pci_iov.h>
96 #endif
97 
98 #include <sys/bitstring.h>
99 /*
100  * enable accounting of every mbuf as it comes in to and goes out of
101  * iflib's software descriptor references
102  */
103 #define MEMORY_LOGGING 0
104 /*
105  * Enable mbuf vectors for compressing long mbuf chains
106  */
107 
108 /*
109  * NB:
110  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
111  *   we prefetch needs to be determined by the time spent in m_free vis a vis
112  *   the cost of a prefetch. This will of course vary based on the workload:
113  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
114  *        is quite expensive, thus suggesting very little prefetch.
115  *      - small packet forwarding which is just returning a single mbuf to
116  *        UMA will typically be very fast vis a vis the cost of a memory
117  *        access.
118  */
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 static void iflib_tqg_detach(if_ctx_t ctx);
147 
148 typedef struct iflib_filter_info {
149 	driver_filter_t *ifi_filter;
150 	void *ifi_filter_arg;
151 	struct grouptask *ifi_task;
152 	void *ifi_ctx;
153 } *iflib_filter_info_t;
154 
155 struct iflib_ctx {
156 	KOBJ_FIELDS;
157 	/*
158 	 * Pointer to hardware driver's softc
159 	 */
160 	void *ifc_softc;
161 	device_t ifc_dev;
162 	if_t ifc_ifp;
163 
164 	cpuset_t ifc_cpus;
165 	if_shared_ctx_t ifc_sctx;
166 	struct if_softc_ctx ifc_softc_ctx;
167 
168 	struct sx ifc_ctx_sx;
169 	struct mtx ifc_state_mtx;
170 
171 	iflib_txq_t ifc_txqs;
172 	iflib_rxq_t ifc_rxqs;
173 	uint32_t ifc_if_flags;
174 	uint32_t ifc_flags;
175 	uint32_t ifc_max_fl_buf_size;
176 	uint32_t ifc_rx_mbuf_sz;
177 
178 	int ifc_link_state;
179 	int ifc_watchdog_events;
180 	struct cdev *ifc_led_dev;
181 	struct resource *ifc_msix_mem;
182 
183 	struct if_irq ifc_legacy_irq;
184 	struct grouptask ifc_admin_task;
185 	struct grouptask ifc_vflr_task;
186 	struct iflib_filter_info ifc_filter_info;
187 	struct ifmedia	ifc_media;
188 	struct ifmedia	*ifc_mediap;
189 
190 	struct sysctl_oid *ifc_sysctl_node;
191 	uint16_t ifc_sysctl_ntxqs;
192 	uint16_t ifc_sysctl_nrxqs;
193 	uint16_t ifc_sysctl_qs_eq_override;
194 	uint16_t ifc_sysctl_rx_budget;
195 	uint16_t ifc_sysctl_tx_abdicate;
196 	uint16_t ifc_sysctl_core_offset;
197 #define	CORE_OFFSET_UNSPECIFIED	0xffff
198 	uint8_t  ifc_sysctl_separate_txrx;
199 	uint8_t  ifc_sysctl_use_logical_cores;
200 	bool	 ifc_cpus_are_physical_cores;
201 
202 	qidx_t ifc_sysctl_ntxds[8];
203 	qidx_t ifc_sysctl_nrxds[8];
204 	struct if_txrx ifc_txrx;
205 #define isc_txd_encap  ifc_txrx.ift_txd_encap
206 #define isc_txd_flush  ifc_txrx.ift_txd_flush
207 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
208 #define isc_rxd_available ifc_txrx.ift_rxd_available
209 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
212 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
213 #define isc_txq_select ifc_txrx.ift_txq_select
214 #define isc_txq_select_v2 ifc_txrx.ift_txq_select_v2
215 	eventhandler_tag ifc_vlan_attach_event;
216 	eventhandler_tag ifc_vlan_detach_event;
217 	struct ether_addr ifc_mac;
218 };
219 
220 void *
221 iflib_get_softc(if_ctx_t ctx)
222 {
223 
224 	return (ctx->ifc_softc);
225 }
226 
227 device_t
228 iflib_get_dev(if_ctx_t ctx)
229 {
230 
231 	return (ctx->ifc_dev);
232 }
233 
234 if_t
235 iflib_get_ifp(if_ctx_t ctx)
236 {
237 
238 	return (ctx->ifc_ifp);
239 }
240 
241 struct ifmedia *
242 iflib_get_media(if_ctx_t ctx)
243 {
244 
245 	return (ctx->ifc_mediap);
246 }
247 
248 uint32_t
249 iflib_get_flags(if_ctx_t ctx)
250 {
251 	return (ctx->ifc_flags);
252 }
253 
254 void
255 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
256 {
257 
258 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
259 }
260 
261 if_softc_ctx_t
262 iflib_get_softc_ctx(if_ctx_t ctx)
263 {
264 
265 	return (&ctx->ifc_softc_ctx);
266 }
267 
268 if_shared_ctx_t
269 iflib_get_sctx(if_ctx_t ctx)
270 {
271 
272 	return (ctx->ifc_sctx);
273 }
274 
275 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
276 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
277 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
278 
279 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
280 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
281 
282 typedef struct iflib_sw_rx_desc_array {
283 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
284 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
285 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
286 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
287 } iflib_rxsd_array_t;
288 
289 typedef struct iflib_sw_tx_desc_array {
290 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
291 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
292 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
293 } if_txsd_vec_t;
294 
295 /* magic number that should be high enough for any hardware */
296 #define IFLIB_MAX_TX_SEGS		128
297 #define IFLIB_RX_COPY_THRESH		128
298 #define IFLIB_MAX_RX_REFRESH		32
299 /* The minimum descriptors per second before we start coalescing */
300 #define IFLIB_MIN_DESC_SEC		16384
301 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
302 #define IFLIB_QUEUE_IDLE		0
303 #define IFLIB_QUEUE_HUNG		1
304 #define IFLIB_QUEUE_WORKING		2
305 /* maximum number of txqs that can share an rx interrupt */
306 #define IFLIB_MAX_TX_SHARED_INTR	4
307 
308 /* this should really scale with ring size - this is a fairly arbitrary value */
309 #define TX_BATCH_SIZE			32
310 
311 #define IFLIB_RESTART_BUDGET		8
312 
313 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
314 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
315 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
316 
317 struct iflib_txq {
318 	qidx_t		ift_in_use;
319 	qidx_t		ift_cidx;
320 	qidx_t		ift_cidx_processed;
321 	qidx_t		ift_pidx;
322 	uint8_t		ift_gen;
323 	uint8_t		ift_br_offset;
324 	uint16_t	ift_npending;
325 	uint16_t	ift_db_pending;
326 	uint16_t	ift_rs_pending;
327 	/* implicit pad */
328 	uint8_t		ift_txd_size[8];
329 	uint64_t	ift_processed;
330 	uint64_t	ift_cleaned;
331 	uint64_t	ift_cleaned_prev;
332 #if MEMORY_LOGGING
333 	uint64_t	ift_enqueued;
334 	uint64_t	ift_dequeued;
335 #endif
336 	uint64_t	ift_no_tx_dma_setup;
337 	uint64_t	ift_no_desc_avail;
338 	uint64_t	ift_mbuf_defrag_failed;
339 	uint64_t	ift_mbuf_defrag;
340 	uint64_t	ift_map_failed;
341 	uint64_t	ift_txd_encap_efbig;
342 	uint64_t	ift_pullups;
343 	uint64_t	ift_last_timer_tick;
344 
345 	struct mtx	ift_mtx;
346 	struct mtx	ift_db_mtx;
347 
348 	/* constant values */
349 	if_ctx_t	ift_ctx;
350 	struct ifmp_ring        *ift_br;
351 	struct grouptask	ift_task;
352 	qidx_t		ift_size;
353 	uint16_t	ift_id;
354 	struct callout	ift_timer;
355 #ifdef DEV_NETMAP
356 	struct callout	ift_netmap_timer;
357 #endif /* DEV_NETMAP */
358 
359 	if_txsd_vec_t	ift_sds;
360 	uint8_t		ift_qstatus;
361 	uint8_t		ift_closed;
362 	uint8_t		ift_update_freq;
363 	struct iflib_filter_info ift_filter_info;
364 	bus_dma_tag_t	ift_buf_tag;
365 	bus_dma_tag_t	ift_tso_buf_tag;
366 	iflib_dma_info_t	ift_ifdi;
367 #define	MTX_NAME_LEN	32
368 	char                    ift_mtx_name[MTX_NAME_LEN];
369 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
370 #ifdef IFLIB_DIAGNOSTICS
371 	uint64_t ift_cpu_exec_count[256];
372 #endif
373 } __aligned(CACHE_LINE_SIZE);
374 
375 struct iflib_fl {
376 	qidx_t		ifl_cidx;
377 	qidx_t		ifl_pidx;
378 	qidx_t		ifl_credits;
379 	uint8_t		ifl_gen;
380 	uint8_t		ifl_rxd_size;
381 #if MEMORY_LOGGING
382 	uint64_t	ifl_m_enqueued;
383 	uint64_t	ifl_m_dequeued;
384 	uint64_t	ifl_cl_enqueued;
385 	uint64_t	ifl_cl_dequeued;
386 #endif
387 	/* implicit pad */
388 	bitstr_t 	*ifl_rx_bitmap;
389 	qidx_t		ifl_fragidx;
390 	/* constant */
391 	qidx_t		ifl_size;
392 	uint16_t	ifl_buf_size;
393 	uint16_t	ifl_cltype;
394 	uma_zone_t	ifl_zone;
395 	iflib_rxsd_array_t	ifl_sds;
396 	iflib_rxq_t	ifl_rxq;
397 	uint8_t		ifl_id;
398 	bus_dma_tag_t	ifl_buf_tag;
399 	iflib_dma_info_t	ifl_ifdi;
400 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
401 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
402 }  __aligned(CACHE_LINE_SIZE);
403 
404 static inline qidx_t
405 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
406 {
407 	qidx_t used;
408 
409 	if (pidx > cidx)
410 		used = pidx - cidx;
411 	else if (pidx < cidx)
412 		used = size - cidx + pidx;
413 	else if (gen == 0 && pidx == cidx)
414 		used = 0;
415 	else if (gen == 1 && pidx == cidx)
416 		used = size;
417 	else
418 		panic("bad state");
419 
420 	return (used);
421 }
422 
423 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
424 
425 #define IDXDIFF(head, tail, wrap) \
426 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
427 
428 struct iflib_rxq {
429 	if_ctx_t	ifr_ctx;
430 	iflib_fl_t	ifr_fl;
431 	uint64_t	ifr_rx_irq;
432 	struct pfil_head	*pfil;
433 	/*
434 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
435 	 * the completion queue consumer index.  Otherwise it's unused.
436 	 */
437 	qidx_t		ifr_cq_cidx;
438 	uint16_t	ifr_id;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	uint8_t		ifr_fl_offset;
443 	struct lro_ctrl			ifr_lc;
444 	struct grouptask        ifr_task;
445 	struct callout		ifr_watchdog;
446 	struct iflib_filter_info ifr_filter_info;
447 	iflib_dma_info_t		ifr_ifdi;
448 
449 	/* dynamically allocate if any drivers need a value substantially larger than this */
450 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
451 #ifdef IFLIB_DIAGNOSTICS
452 	uint64_t ifr_cpu_exec_count[256];
453 #endif
454 }  __aligned(CACHE_LINE_SIZE);
455 
456 typedef struct if_rxsd {
457 	caddr_t *ifsd_cl;
458 	iflib_fl_t ifsd_fl;
459 } *if_rxsd_t;
460 
461 /* multiple of word size */
462 #ifdef __LP64__
463 #define PKT_INFO_SIZE	6
464 #define RXD_INFO_SIZE	5
465 #define PKT_TYPE uint64_t
466 #else
467 #define PKT_INFO_SIZE	11
468 #define RXD_INFO_SIZE	8
469 #define PKT_TYPE uint32_t
470 #endif
471 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
472 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
473 
474 typedef struct if_pkt_info_pad {
475 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
476 } *if_pkt_info_pad_t;
477 typedef struct if_rxd_info_pad {
478 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
479 } *if_rxd_info_pad_t;
480 
481 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
482 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
483 
484 static inline void
485 pkt_info_zero(if_pkt_info_t pi)
486 {
487 	if_pkt_info_pad_t pi_pad;
488 
489 	pi_pad = (if_pkt_info_pad_t)pi;
490 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
491 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
492 #ifndef __LP64__
493 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
494 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
495 #endif
496 }
497 
498 static device_method_t iflib_pseudo_methods[] = {
499 	DEVMETHOD(device_attach, noop_attach),
500 	DEVMETHOD(device_detach, iflib_pseudo_detach),
501 	DEVMETHOD_END
502 };
503 
504 driver_t iflib_pseudodriver = {
505 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
506 };
507 
508 static inline void
509 rxd_info_zero(if_rxd_info_t ri)
510 {
511 	if_rxd_info_pad_t ri_pad;
512 	int i;
513 
514 	ri_pad = (if_rxd_info_pad_t)ri;
515 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
516 		ri_pad->rxd_val[i] = 0;
517 		ri_pad->rxd_val[i+1] = 0;
518 		ri_pad->rxd_val[i+2] = 0;
519 		ri_pad->rxd_val[i+3] = 0;
520 	}
521 #ifdef __LP64__
522 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
523 #endif
524 }
525 
526 /*
527  * Only allow a single packet to take up most 1/nth of the tx ring
528  */
529 #define MAX_SINGLE_PACKET_FRACTION 12
530 #define IF_BAD_DMA (bus_addr_t)-1
531 
532 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
533 
534 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
535 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
536 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
537 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
538 
539 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
540 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
541 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
542 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
543 
544 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
545 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
546 
547 void
548 iflib_set_detach(if_ctx_t ctx)
549 {
550 	STATE_LOCK(ctx);
551 	ctx->ifc_flags |= IFC_IN_DETACH;
552 	STATE_UNLOCK(ctx);
553 }
554 
555 /* Our boot-time initialization hook */
556 static int	iflib_module_event_handler(module_t, int, void *);
557 
558 static moduledata_t iflib_moduledata = {
559 	"iflib",
560 	iflib_module_event_handler,
561 	NULL
562 };
563 
564 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
565 MODULE_VERSION(iflib, 1);
566 
567 MODULE_DEPEND(iflib, pci, 1, 1, 1);
568 MODULE_DEPEND(iflib, ether, 1, 1, 1);
569 
570 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
571 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
572 
573 #ifndef IFLIB_DEBUG_COUNTERS
574 #ifdef INVARIANTS
575 #define IFLIB_DEBUG_COUNTERS 1
576 #else
577 #define IFLIB_DEBUG_COUNTERS 0
578 #endif /* !INVARIANTS */
579 #endif
580 
581 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
582     "iflib driver parameters");
583 
584 /*
585  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
586  */
587 static int iflib_min_tx_latency = 0;
588 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
589 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
590 static int iflib_no_tx_batch = 0;
591 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
592 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
593 static int iflib_timer_default = 1000;
594 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
595 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
596 
597 
598 #if IFLIB_DEBUG_COUNTERS
599 
600 static int iflib_tx_seen;
601 static int iflib_tx_sent;
602 static int iflib_tx_encap;
603 static int iflib_rx_allocs;
604 static int iflib_fl_refills;
605 static int iflib_fl_refills_large;
606 static int iflib_tx_frees;
607 
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
609 		   &iflib_tx_seen, 0, "# TX mbufs seen");
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
611 		   &iflib_tx_sent, 0, "# TX mbufs sent");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
613 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
615 		   &iflib_tx_frees, 0, "# TX frees");
616 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
617 		   &iflib_rx_allocs, 0, "# RX allocations");
618 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
619 		   &iflib_fl_refills, 0, "# refills");
620 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
621 		   &iflib_fl_refills_large, 0, "# large refills");
622 
623 static int iflib_txq_drain_flushing;
624 static int iflib_txq_drain_oactive;
625 static int iflib_txq_drain_notready;
626 
627 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
628 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
630 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
632 		   &iflib_txq_drain_notready, 0, "# drain notready");
633 
634 static int iflib_encap_load_mbuf_fail;
635 static int iflib_encap_pad_mbuf_fail;
636 static int iflib_encap_txq_avail_fail;
637 static int iflib_encap_txd_encap_fail;
638 
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
640 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
642 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
644 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
646 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
647 
648 static int iflib_task_fn_rxs;
649 static int iflib_rx_intr_enables;
650 static int iflib_fast_intrs;
651 static int iflib_rx_unavail;
652 static int iflib_rx_ctx_inactive;
653 static int iflib_rx_if_input;
654 static int iflib_rxd_flush;
655 
656 static int iflib_verbose_debug;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
659 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
661 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
662 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
663 		   &iflib_fast_intrs, 0, "# fast_intr calls");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
665 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
667 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
669 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
671 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
672 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
673 		   &iflib_verbose_debug, 0, "enable verbose debugging");
674 
675 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
676 static void
677 iflib_debug_reset(void)
678 {
679 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
680 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
681 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
682 		iflib_txq_drain_notready =
683 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
684 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
685 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
686 		iflib_rx_unavail =
687 		iflib_rx_ctx_inactive = iflib_rx_if_input =
688 		iflib_rxd_flush = 0;
689 }
690 
691 #else
692 #define DBG_COUNTER_INC(name)
693 static void iflib_debug_reset(void) {}
694 #endif
695 
696 #define IFLIB_DEBUG 0
697 
698 static void iflib_tx_structures_free(if_ctx_t ctx);
699 static void iflib_rx_structures_free(if_ctx_t ctx);
700 static int iflib_queues_alloc(if_ctx_t ctx);
701 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
702 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
703 static int iflib_qset_structures_setup(if_ctx_t ctx);
704 static int iflib_msix_init(if_ctx_t ctx);
705 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
706 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
707 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
708 #ifdef ALTQ
709 static void iflib_altq_if_start(if_t ifp);
710 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
711 #endif
712 static int iflib_register(if_ctx_t);
713 static void iflib_deregister(if_ctx_t);
714 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
715 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
716 static void iflib_init_locked(if_ctx_t ctx);
717 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
718 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
719 static void iflib_ifmp_purge(iflib_txq_t txq);
720 static void _iflib_pre_assert(if_softc_ctx_t scctx);
721 static void iflib_if_init_locked(if_ctx_t ctx);
722 static void iflib_free_intr_mem(if_ctx_t ctx);
723 #ifndef __NO_STRICT_ALIGNMENT
724 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
725 #endif
726 
727 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
728     SLIST_HEAD_INITIALIZER(cpu_offsets);
729 struct cpu_offset {
730 	SLIST_ENTRY(cpu_offset) entries;
731 	cpuset_t	set;
732 	unsigned int	refcount;
733 	uint16_t	next_cpuid;
734 };
735 static struct mtx cpu_offset_mtx;
736 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
737     MTX_DEF);
738 
739 DEBUGNET_DEFINE(iflib);
740 
741 static int
742 iflib_num_rx_descs(if_ctx_t ctx)
743 {
744 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
745 	if_shared_ctx_t sctx = ctx->ifc_sctx;
746 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
747 
748 	return scctx->isc_nrxd[first_rxq];
749 }
750 
751 static int
752 iflib_num_tx_descs(if_ctx_t ctx)
753 {
754 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
755 	if_shared_ctx_t sctx = ctx->ifc_sctx;
756 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
757 
758 	return scctx->isc_ntxd[first_txq];
759 }
760 
761 #ifdef DEV_NETMAP
762 #include <sys/selinfo.h>
763 #include <net/netmap.h>
764 #include <dev/netmap/netmap_kern.h>
765 
766 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
767 
768 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
769 static void iflib_netmap_timer(void *arg);
770 
771 /*
772  * device-specific sysctl variables:
773  *
774  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
775  *	During regular operations the CRC is stripped, but on some
776  *	hardware reception of frames not multiple of 64 is slower,
777  *	so using crcstrip=0 helps in benchmarks.
778  *
779  * iflib_rx_miss, iflib_rx_miss_bufs:
780  *	count packets that might be missed due to lost interrupts.
781  */
782 SYSCTL_DECL(_dev_netmap);
783 /*
784  * The xl driver by default strips CRCs and we do not override it.
785  */
786 
787 int iflib_crcstrip = 1;
788 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
789     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
790 
791 int iflib_rx_miss, iflib_rx_miss_bufs;
792 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
793     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
794 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
795     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
796 
797 /*
798  * Register/unregister. We are already under netmap lock.
799  * Only called on the first register or the last unregister.
800  */
801 static int
802 iflib_netmap_register(struct netmap_adapter *na, int onoff)
803 {
804 	if_t ifp = na->ifp;
805 	if_ctx_t ctx = if_getsoftc(ifp);
806 	int status;
807 
808 	CTX_LOCK(ctx);
809 	if (!CTX_IS_VF(ctx))
810 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
811 
812 	iflib_stop(ctx);
813 
814 	/*
815 	 * Enable (or disable) netmap flags, and intercept (or restore)
816 	 * ifp->if_transmit. This is done once the device has been stopped
817 	 * to prevent race conditions. Also, this must be done after
818 	 * calling netmap_disable_all_rings() and before calling
819 	 * netmap_enable_all_rings(), so that these two functions see the
820 	 * updated state of the NAF_NETMAP_ON bit.
821 	 */
822 	if (onoff) {
823 		nm_set_native_flags(na);
824 	} else {
825 		nm_clear_native_flags(na);
826 	}
827 
828 	iflib_init_locked(ctx);
829 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
830 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
831 	if (status)
832 		nm_clear_native_flags(na);
833 	CTX_UNLOCK(ctx);
834 	return (status);
835 }
836 
837 static int
838 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
839 {
840 	if_t ifp = na->ifp;
841 	if_ctx_t ctx = if_getsoftc(ifp);
842 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
843 	iflib_fl_t fl = &rxq->ifr_fl[0];
844 
845 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
846 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
847 	info->num_tx_descs = iflib_num_tx_descs(ctx);
848 	info->num_rx_descs = iflib_num_rx_descs(ctx);
849 	info->rx_buf_maxsize = fl->ifl_buf_size;
850 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
851 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
852 		info->num_rx_descs, info->rx_buf_maxsize);
853 
854 	return 0;
855 }
856 
857 static int
858 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
859 {
860 	struct netmap_adapter *na = kring->na;
861 	u_int const lim = kring->nkr_num_slots - 1;
862 	struct netmap_ring *ring = kring->ring;
863 	bus_dmamap_t *map;
864 	struct if_rxd_update iru;
865 	if_ctx_t ctx = rxq->ifr_ctx;
866 	iflib_fl_t fl = &rxq->ifr_fl[0];
867 	u_int nic_i_first, nic_i;
868 	u_int nm_i;
869 	int i, n;
870 #if IFLIB_DEBUG_COUNTERS
871 	int rf_count = 0;
872 #endif
873 
874 	/*
875 	 * This function is used both at initialization and in rxsync.
876 	 * At initialization we need to prepare (with isc_rxd_refill())
877 	 * all the netmap buffers currently owned by the kernel, in
878 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
879 	 * (except for kring->nkr_hwofs). These may be less than
880 	 * kring->nkr_num_slots if netmap_reset() was called while
881 	 * an application using the kring that still owned some
882 	 * buffers.
883 	 * At rxsync time, both indexes point to the next buffer to be
884 	 * refilled.
885 	 * In any case we publish (with isc_rxd_flush()) up to
886 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
887 	 * pointer to overrun the head/cons pointer, although this is
888 	 * not necessary for some NICs (e.g. vmx).
889 	 */
890 	if (__predict_false(init)) {
891 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
892 	} else {
893 		n = kring->rhead - kring->nr_hwcur;
894 		if (n == 0)
895 			return (0); /* Nothing to do. */
896 		if (n < 0)
897 			n += kring->nkr_num_slots;
898 	}
899 
900 	iru_init(&iru, rxq, 0 /* flid */);
901 	map = fl->ifl_sds.ifsd_map;
902 	nic_i = fl->ifl_pidx;
903 	nm_i = netmap_idx_n2k(kring, nic_i);
904 	if (__predict_false(init)) {
905 		/*
906 		 * On init/reset, nic_i must be 0, and we must
907 		 * start to refill from hwtail (see netmap_reset()).
908 		 */
909 		MPASS(nic_i == 0);
910 		MPASS(nm_i == kring->nr_hwtail);
911 	} else
912 		MPASS(nm_i == kring->nr_hwcur);
913 	DBG_COUNTER_INC(fl_refills);
914 	while (n > 0) {
915 #if IFLIB_DEBUG_COUNTERS
916 		if (++rf_count == 9)
917 			DBG_COUNTER_INC(fl_refills_large);
918 #endif
919 		nic_i_first = nic_i;
920 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
921 			struct netmap_slot *slot = &ring->slot[nm_i];
922 			uint64_t paddr;
923 			void *addr = PNMB(na, slot, &paddr);
924 
925 			MPASS(i < IFLIB_MAX_RX_REFRESH);
926 
927 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
928 			        return netmap_ring_reinit(kring);
929 
930 			fl->ifl_bus_addrs[i] = paddr +
931 			    nm_get_offset(kring, slot);
932 			fl->ifl_rxd_idxs[i] = nic_i;
933 
934 			if (__predict_false(init)) {
935 				netmap_load_map(na, fl->ifl_buf_tag,
936 				    map[nic_i], addr);
937 			} else if (slot->flags & NS_BUF_CHANGED) {
938 				/* buffer has changed, reload map */
939 				netmap_reload_map(na, fl->ifl_buf_tag,
940 				    map[nic_i], addr);
941 			}
942 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
943 			    BUS_DMASYNC_PREREAD);
944 			slot->flags &= ~NS_BUF_CHANGED;
945 
946 			nm_i = nm_next(nm_i, lim);
947 			nic_i = nm_next(nic_i, lim);
948 		}
949 
950 		iru.iru_pidx = nic_i_first;
951 		iru.iru_count = i;
952 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
953 	}
954 	fl->ifl_pidx = nic_i;
955 	/*
956 	 * At the end of the loop we must have refilled everything
957 	 * we could possibly refill.
958 	 */
959 	MPASS(nm_i == kring->rhead);
960 	kring->nr_hwcur = nm_i;
961 
962 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
963 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
964 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
965 	    nm_prev(nic_i, lim));
966 	DBG_COUNTER_INC(rxd_flush);
967 
968 	return (0);
969 }
970 
971 #define NETMAP_TX_TIMER_US	90
972 
973 /*
974  * Reconcile kernel and user view of the transmit ring.
975  *
976  * All information is in the kring.
977  * Userspace wants to send packets up to the one before kring->rhead,
978  * kernel knows kring->nr_hwcur is the first unsent packet.
979  *
980  * Here we push packets out (as many as possible), and possibly
981  * reclaim buffers from previously completed transmission.
982  *
983  * The caller (netmap) guarantees that there is only one instance
984  * running at any time. Any interference with other driver
985  * methods should be handled by the individual drivers.
986  */
987 static int
988 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
989 {
990 	struct netmap_adapter *na = kring->na;
991 	if_t ifp = na->ifp;
992 	struct netmap_ring *ring = kring->ring;
993 	u_int nm_i;	/* index into the netmap kring */
994 	u_int nic_i;	/* index into the NIC ring */
995 	u_int const lim = kring->nkr_num_slots - 1;
996 	u_int const head = kring->rhead;
997 	struct if_pkt_info pi;
998 	int tx_pkts = 0, tx_bytes = 0;
999 
1000 	/*
1001 	 * interrupts on every tx packet are expensive so request
1002 	 * them every half ring, or where NS_REPORT is set
1003 	 */
1004 	u_int report_frequency = kring->nkr_num_slots >> 1;
1005 	/* device-specific */
1006 	if_ctx_t ctx = if_getsoftc(ifp);
1007 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1008 
1009 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1010 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1011 
1012 	/*
1013 	 * First part: process new packets to send.
1014 	 * nm_i is the current index in the netmap kring,
1015 	 * nic_i is the corresponding index in the NIC ring.
1016 	 *
1017 	 * If we have packets to send (nm_i != head)
1018 	 * iterate over the netmap ring, fetch length and update
1019 	 * the corresponding slot in the NIC ring. Some drivers also
1020 	 * need to update the buffer's physical address in the NIC slot
1021 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1022 	 *
1023 	 * The netmap_reload_map() calls is especially expensive,
1024 	 * even when (as in this case) the tag is 0, so do only
1025 	 * when the buffer has actually changed.
1026 	 *
1027 	 * If possible do not set the report/intr bit on all slots,
1028 	 * but only a few times per ring or when NS_REPORT is set.
1029 	 *
1030 	 * Finally, on 10G and faster drivers, it might be useful
1031 	 * to prefetch the next slot and txr entry.
1032 	 */
1033 
1034 	nm_i = kring->nr_hwcur;
1035 	if (nm_i != head) {	/* we have new packets to send */
1036 		uint32_t pkt_len = 0, seg_idx = 0;
1037 		int nic_i_start = -1, flags = 0;
1038 		pkt_info_zero(&pi);
1039 		pi.ipi_segs = txq->ift_segs;
1040 		pi.ipi_qsidx = kring->ring_id;
1041 		nic_i = netmap_idx_k2n(kring, nm_i);
1042 
1043 		__builtin_prefetch(&ring->slot[nm_i]);
1044 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1045 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1046 
1047 		while (nm_i != head) {
1048 			struct netmap_slot *slot = &ring->slot[nm_i];
1049 			uint64_t offset = nm_get_offset(kring, slot);
1050 			u_int len = slot->len;
1051 			uint64_t paddr;
1052 			void *addr = PNMB(na, slot, &paddr);
1053 
1054 			flags |= (slot->flags & NS_REPORT ||
1055 				nic_i == 0 || nic_i == report_frequency) ?
1056 				IPI_TX_INTR : 0;
1057 
1058 			/*
1059 			 * If this is the first packet fragment, save the
1060 			 * index of the first NIC slot for later.
1061 			 */
1062 			if (nic_i_start < 0)
1063 				nic_i_start = nic_i;
1064 
1065 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1066 			pi.ipi_segs[seg_idx].ds_len = len;
1067 			if (len) {
1068 				pkt_len += len;
1069 				seg_idx++;
1070 			}
1071 
1072 			if (!(slot->flags & NS_MOREFRAG)) {
1073 				pi.ipi_len = pkt_len;
1074 				pi.ipi_nsegs = seg_idx;
1075 				pi.ipi_pidx = nic_i_start;
1076 				pi.ipi_ndescs = 0;
1077 				pi.ipi_flags = flags;
1078 
1079 				/* Prepare the NIC TX ring. */
1080 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1081 				DBG_COUNTER_INC(tx_encap);
1082 
1083 				/* Update transmit counters */
1084 				tx_bytes += pi.ipi_len;
1085 				tx_pkts++;
1086 
1087 				/* Reinit per-packet info for the next one. */
1088 				flags = seg_idx = pkt_len = 0;
1089 				nic_i_start = -1;
1090 			}
1091 
1092 			/* prefetch for next round */
1093 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1094 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1095 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1096 
1097 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1098 
1099 			if (slot->flags & NS_BUF_CHANGED) {
1100 				/* buffer has changed, reload map */
1101 				netmap_reload_map(na, txq->ift_buf_tag,
1102 				    txq->ift_sds.ifsd_map[nic_i], addr);
1103 			}
1104 			/* make sure changes to the buffer are synced */
1105 			bus_dmamap_sync(txq->ift_buf_tag,
1106 			    txq->ift_sds.ifsd_map[nic_i],
1107 			    BUS_DMASYNC_PREWRITE);
1108 
1109 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1110 			nm_i = nm_next(nm_i, lim);
1111 			nic_i = nm_next(nic_i, lim);
1112 		}
1113 		kring->nr_hwcur = nm_i;
1114 
1115 		/* synchronize the NIC ring */
1116 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1117 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1118 
1119 		/* (re)start the tx unit up to slot nic_i (excluded) */
1120 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1121 	}
1122 
1123 	/*
1124 	 * Second part: reclaim buffers for completed transmissions.
1125 	 *
1126 	 * If there are unclaimed buffers, attempt to reclaim them.
1127 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1128 	 * trigger a per-tx-queue timer to try again later.
1129 	 */
1130 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1131 		if (iflib_tx_credits_update(ctx, txq)) {
1132 			/* some tx completed, increment avail */
1133 			nic_i = txq->ift_cidx_processed;
1134 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1135 		}
1136 	}
1137 
1138 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1139 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1140 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1141 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1142 			    iflib_netmap_timer, txq,
1143 			    txq->ift_netmap_timer.c_cpu, 0);
1144 		}
1145 
1146 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1147 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1148 
1149 	return (0);
1150 }
1151 
1152 /*
1153  * Reconcile kernel and user view of the receive ring.
1154  * Same as for the txsync, this routine must be efficient.
1155  * The caller guarantees a single invocations, but races against
1156  * the rest of the driver should be handled here.
1157  *
1158  * On call, kring->rhead is the first packet that userspace wants
1159  * to keep, and kring->rcur is the wakeup point.
1160  * The kernel has previously reported packets up to kring->rtail.
1161  *
1162  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1163  * of whether or not we received an interrupt.
1164  */
1165 static int
1166 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1167 {
1168 	struct netmap_adapter *na = kring->na;
1169 	struct netmap_ring *ring = kring->ring;
1170 	if_t ifp = na->ifp;
1171 	uint32_t nm_i;	/* index into the netmap ring */
1172 	uint32_t nic_i;	/* index into the NIC ring */
1173 	u_int n;
1174 	u_int const lim = kring->nkr_num_slots - 1;
1175 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1176 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1177 
1178 	if_ctx_t ctx = if_getsoftc(ifp);
1179 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1180 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1181 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1182 	iflib_fl_t fl = &rxq->ifr_fl[0];
1183 	struct if_rxd_info ri;
1184 	qidx_t *cidxp;
1185 
1186 	/*
1187 	 * netmap only uses free list 0, to avoid out of order consumption
1188 	 * of receive buffers
1189 	 */
1190 
1191 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1192 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1193 
1194 	/*
1195 	 * First part: import newly received packets.
1196 	 *
1197 	 * nm_i is the index of the next free slot in the netmap ring,
1198 	 * nic_i is the index of the next received packet in the NIC ring
1199 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1200 	 * differ in case if_init() has been called while
1201 	 * in netmap mode. For the receive ring we have
1202 	 *
1203 	 *	nic_i = fl->ifl_cidx;
1204 	 *	nm_i = kring->nr_hwtail (previous)
1205 	 * and
1206 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1207 	 *
1208 	 * fl->ifl_cidx is set to 0 on a ring reinit
1209 	 */
1210 	if (netmap_no_pendintr || force_update) {
1211 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1212 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1213 		int crclen = iflib_crcstrip ? 0 : 4;
1214 		int error, avail;
1215 
1216 		/*
1217 		 * For the free list consumer index, we use the same
1218 		 * logic as in iflib_rxeof().
1219 		 */
1220 		if (have_rxcq)
1221 			cidxp = &rxq->ifr_cq_cidx;
1222 		else
1223 			cidxp = &fl->ifl_cidx;
1224 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1225 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1226 
1227 		nic_i = fl->ifl_cidx;
1228 		nm_i = netmap_idx_n2k(kring, nic_i);
1229 		MPASS(nm_i == kring->nr_hwtail);
1230 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1231 			rxd_info_zero(&ri);
1232 			ri.iri_frags = rxq->ifr_frags;
1233 			ri.iri_qsidx = kring->ring_id;
1234 			ri.iri_ifp = ctx->ifc_ifp;
1235 			ri.iri_cidx = *cidxp;
1236 
1237 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1238 			for (i = 0; i < ri.iri_nfrags; i++) {
1239 				if (error) {
1240 					ring->slot[nm_i].len = 0;
1241 					ring->slot[nm_i].flags = 0;
1242 				} else {
1243 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1244 					if (i == (ri.iri_nfrags - 1)) {
1245 						ring->slot[nm_i].len -= crclen;
1246 						ring->slot[nm_i].flags = 0;
1247 
1248 						/* Update receive counters */
1249 						rx_bytes += ri.iri_len;
1250 						rx_pkts++;
1251 					} else
1252 						ring->slot[nm_i].flags = NS_MOREFRAG;
1253 				}
1254 
1255 				bus_dmamap_sync(fl->ifl_buf_tag,
1256 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1257 				nm_i = nm_next(nm_i, lim);
1258 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1259 			}
1260 
1261 			if (have_rxcq) {
1262 				*cidxp = ri.iri_cidx;
1263 				while (*cidxp >= scctx->isc_nrxd[0])
1264 					*cidxp -= scctx->isc_nrxd[0];
1265 			}
1266 
1267 		}
1268 		if (n) { /* update the state variables */
1269 			if (netmap_no_pendintr && !force_update) {
1270 				/* diagnostics */
1271 				iflib_rx_miss ++;
1272 				iflib_rx_miss_bufs += n;
1273 			}
1274 			kring->nr_hwtail = nm_i;
1275 		}
1276 		kring->nr_kflags &= ~NKR_PENDINTR;
1277 	}
1278 	/*
1279 	 * Second part: skip past packets that userspace has released.
1280 	 * (kring->nr_hwcur to head excluded),
1281 	 * and make the buffers available for reception.
1282 	 * As usual nm_i is the index in the netmap ring,
1283 	 * nic_i is the index in the NIC ring, and
1284 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1285 	 */
1286 	netmap_fl_refill(rxq, kring, false);
1287 
1288 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1289 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1290 
1291 	return (0);
1292 }
1293 
1294 static void
1295 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1296 {
1297 	if_ctx_t ctx = if_getsoftc(na->ifp);
1298 
1299 	CTX_LOCK(ctx);
1300 	if (onoff) {
1301 		IFDI_INTR_ENABLE(ctx);
1302 	} else {
1303 		IFDI_INTR_DISABLE(ctx);
1304 	}
1305 	CTX_UNLOCK(ctx);
1306 }
1307 
1308 static int
1309 iflib_netmap_attach(if_ctx_t ctx)
1310 {
1311 	struct netmap_adapter na;
1312 
1313 	bzero(&na, sizeof(na));
1314 
1315 	na.ifp = ctx->ifc_ifp;
1316 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1317 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1318 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1319 
1320 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1321 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1322 	na.nm_txsync = iflib_netmap_txsync;
1323 	na.nm_rxsync = iflib_netmap_rxsync;
1324 	na.nm_register = iflib_netmap_register;
1325 	na.nm_intr = iflib_netmap_intr;
1326 	na.nm_config = iflib_netmap_config;
1327 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1328 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1329 	return (netmap_attach(&na));
1330 }
1331 
1332 static int
1333 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1334 {
1335 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1336 	struct netmap_slot *slot;
1337 
1338 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1339 	if (slot == NULL)
1340 		return (0);
1341 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1342 		/*
1343 		 * In netmap mode, set the map for the packet buffer.
1344 		 * NOTE: Some drivers (not this one) also need to set
1345 		 * the physical buffer address in the NIC ring.
1346 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1347 		 * netmap slot index, si
1348 		 */
1349 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1350 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1351 		    NMB(na, slot + si));
1352 	}
1353 	return (1);
1354 }
1355 
1356 static int
1357 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1358 {
1359 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1360 	struct netmap_kring *kring;
1361 	struct netmap_slot *slot;
1362 
1363 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1364 	if (slot == NULL)
1365 		return (0);
1366 	kring = na->rx_rings[rxq->ifr_id];
1367 	netmap_fl_refill(rxq, kring, true);
1368 	return (1);
1369 }
1370 
1371 static void
1372 iflib_netmap_timer(void *arg)
1373 {
1374 	iflib_txq_t txq = arg;
1375 	if_ctx_t ctx = txq->ift_ctx;
1376 
1377 	/*
1378 	 * Wake up the netmap application, to give it a chance to
1379 	 * call txsync and reclaim more completed TX buffers.
1380 	 */
1381 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1382 }
1383 
1384 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1385 
1386 #else
1387 #define iflib_netmap_txq_init(ctx, txq) (0)
1388 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1389 #define iflib_netmap_detach(ifp)
1390 #define netmap_enable_all_rings(ifp)
1391 #define netmap_disable_all_rings(ifp)
1392 
1393 #define iflib_netmap_attach(ctx) (0)
1394 #define netmap_rx_irq(ifp, qid, budget) (0)
1395 #endif
1396 
1397 #if defined(__i386__) || defined(__amd64__)
1398 static __inline void
1399 prefetch(void *x)
1400 {
1401 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1402 }
1403 
1404 static __inline void
1405 prefetch2cachelines(void *x)
1406 {
1407 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1408 #if (CACHE_LINE_SIZE < 128)
1409 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1410 #endif
1411 }
1412 #else
1413 static __inline void
1414 prefetch(void *x)
1415 {
1416 }
1417 
1418 static __inline void
1419 prefetch2cachelines(void *x)
1420 {
1421 }
1422 #endif
1423 
1424 static void
1425 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1426 {
1427 	iflib_fl_t fl;
1428 
1429 	fl = &rxq->ifr_fl[flid];
1430 	iru->iru_paddrs = fl->ifl_bus_addrs;
1431 	iru->iru_idxs = fl->ifl_rxd_idxs;
1432 	iru->iru_qsidx = rxq->ifr_id;
1433 	iru->iru_buf_size = fl->ifl_buf_size;
1434 	iru->iru_flidx = fl->ifl_id;
1435 }
1436 
1437 static void
1438 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1439 {
1440 	if (err)
1441 		return;
1442 	*(bus_addr_t *) arg = segs[0].ds_addr;
1443 }
1444 
1445 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1446 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1447 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1448 
1449 int
1450 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1451 {
1452 	int err;
1453 	device_t dev = ctx->ifc_dev;
1454 	bus_addr_t lowaddr;
1455 
1456 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1457 
1458 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1459 				align, 0,		/* alignment, bounds */
1460 				lowaddr,		/* lowaddr */
1461 				BUS_SPACE_MAXADDR,	/* highaddr */
1462 				NULL, NULL,		/* filter, filterarg */
1463 				size,			/* maxsize */
1464 				1,			/* nsegments */
1465 				size,			/* maxsegsize */
1466 				BUS_DMA_ALLOCNOW,	/* flags */
1467 				NULL,			/* lockfunc */
1468 				NULL,			/* lockarg */
1469 				&dma->idi_tag);
1470 	if (err) {
1471 		device_printf(dev,
1472 		    "%s: bus_dma_tag_create failed: %d\n",
1473 		    __func__, err);
1474 		goto fail_0;
1475 	}
1476 
1477 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1478 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1479 	if (err) {
1480 		device_printf(dev,
1481 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1482 		    __func__, (uintmax_t)size, err);
1483 		goto fail_1;
1484 	}
1485 
1486 	dma->idi_paddr = IF_BAD_DMA;
1487 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1488 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1489 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1490 		device_printf(dev,
1491 		    "%s: bus_dmamap_load failed: %d\n",
1492 		    __func__, err);
1493 		goto fail_2;
1494 	}
1495 
1496 	dma->idi_size = size;
1497 	return (0);
1498 
1499 fail_2:
1500 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1501 fail_1:
1502 	bus_dma_tag_destroy(dma->idi_tag);
1503 fail_0:
1504 	dma->idi_tag = NULL;
1505 
1506 	return (err);
1507 }
1508 
1509 int
1510 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1511 {
1512 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1513 
1514 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1515 
1516 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1517 }
1518 
1519 int
1520 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1521 {
1522 	int i, err;
1523 	iflib_dma_info_t *dmaiter;
1524 
1525 	dmaiter = dmalist;
1526 	for (i = 0; i < count; i++, dmaiter++) {
1527 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1528 			break;
1529 	}
1530 	if (err)
1531 		iflib_dma_free_multi(dmalist, i);
1532 	return (err);
1533 }
1534 
1535 void
1536 iflib_dma_free(iflib_dma_info_t dma)
1537 {
1538 	if (dma->idi_tag == NULL)
1539 		return;
1540 	if (dma->idi_paddr != IF_BAD_DMA) {
1541 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1542 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1543 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1544 		dma->idi_paddr = IF_BAD_DMA;
1545 	}
1546 	if (dma->idi_vaddr != NULL) {
1547 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1548 		dma->idi_vaddr = NULL;
1549 	}
1550 	bus_dma_tag_destroy(dma->idi_tag);
1551 	dma->idi_tag = NULL;
1552 }
1553 
1554 void
1555 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1556 {
1557 	int i;
1558 	iflib_dma_info_t *dmaiter = dmalist;
1559 
1560 	for (i = 0; i < count; i++, dmaiter++)
1561 		iflib_dma_free(*dmaiter);
1562 }
1563 
1564 static int
1565 iflib_fast_intr(void *arg)
1566 {
1567 	iflib_filter_info_t info = arg;
1568 	struct grouptask *gtask = info->ifi_task;
1569 	int result;
1570 
1571 	DBG_COUNTER_INC(fast_intrs);
1572 	if (info->ifi_filter != NULL) {
1573 		result = info->ifi_filter(info->ifi_filter_arg);
1574 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1575 			return (result);
1576 	}
1577 
1578 	GROUPTASK_ENQUEUE(gtask);
1579 	return (FILTER_HANDLED);
1580 }
1581 
1582 static int
1583 iflib_fast_intr_rxtx(void *arg)
1584 {
1585 	iflib_filter_info_t info = arg;
1586 	struct grouptask *gtask = info->ifi_task;
1587 	if_ctx_t ctx;
1588 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1589 	iflib_txq_t txq;
1590 	void *sc;
1591 	int i, cidx, result;
1592 	qidx_t txqid;
1593 	bool intr_enable, intr_legacy;
1594 
1595 	DBG_COUNTER_INC(fast_intrs);
1596 	if (info->ifi_filter != NULL) {
1597 		result = info->ifi_filter(info->ifi_filter_arg);
1598 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1599 			return (result);
1600 	}
1601 
1602 	ctx = rxq->ifr_ctx;
1603 	sc = ctx->ifc_softc;
1604 	intr_enable = false;
1605 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1606 	MPASS(rxq->ifr_ntxqirq);
1607 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1608 		txqid = rxq->ifr_txqid[i];
1609 		txq = &ctx->ifc_txqs[txqid];
1610 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1611 		    BUS_DMASYNC_POSTREAD);
1612 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1613 			if (intr_legacy)
1614 				intr_enable = true;
1615 			else
1616 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1617 			continue;
1618 		}
1619 		GROUPTASK_ENQUEUE(&txq->ift_task);
1620 	}
1621 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1622 		cidx = rxq->ifr_cq_cidx;
1623 	else
1624 		cidx = rxq->ifr_fl[0].ifl_cidx;
1625 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1626 		GROUPTASK_ENQUEUE(gtask);
1627 	else {
1628 		if (intr_legacy)
1629 			intr_enable = true;
1630 		else
1631 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1632 		DBG_COUNTER_INC(rx_intr_enables);
1633 	}
1634 	if (intr_enable)
1635 		IFDI_INTR_ENABLE(ctx);
1636 	return (FILTER_HANDLED);
1637 }
1638 
1639 static int
1640 iflib_fast_intr_ctx(void *arg)
1641 {
1642 	iflib_filter_info_t info = arg;
1643 	struct grouptask *gtask = info->ifi_task;
1644 	int result;
1645 
1646 	DBG_COUNTER_INC(fast_intrs);
1647 	if (info->ifi_filter != NULL) {
1648 		result = info->ifi_filter(info->ifi_filter_arg);
1649 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1650 			return (result);
1651 	}
1652 
1653 	GROUPTASK_ENQUEUE(gtask);
1654 	return (FILTER_HANDLED);
1655 }
1656 
1657 static int
1658 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1659 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1660 		 const char *name)
1661 {
1662 	struct resource *res;
1663 	void *tag = NULL;
1664 	device_t dev = ctx->ifc_dev;
1665 	int flags, i, rc;
1666 
1667 	flags = RF_ACTIVE;
1668 	if (ctx->ifc_flags & IFC_LEGACY)
1669 		flags |= RF_SHAREABLE;
1670 	MPASS(rid < 512);
1671 	i = rid;
1672 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1673 	if (res == NULL) {
1674 		device_printf(dev,
1675 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1676 		return (ENOMEM);
1677 	}
1678 	irq->ii_res = res;
1679 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1680 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1681 						filter, handler, arg, &tag);
1682 	if (rc != 0) {
1683 		device_printf(dev,
1684 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1685 					  rid, name ? name : "unknown", rc);
1686 		return (rc);
1687 	} else if (name)
1688 		bus_describe_intr(dev, res, tag, "%s", name);
1689 
1690 	irq->ii_tag = tag;
1691 	return (0);
1692 }
1693 
1694 /*********************************************************************
1695  *
1696  *  Allocate DMA resources for TX buffers as well as memory for the TX
1697  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1698  *  iflib_sw_tx_desc_array structure, storing all the information that
1699  *  is needed to transmit a packet on the wire.  This is called only
1700  *  once at attach, setup is done every reset.
1701  *
1702  **********************************************************************/
1703 static int
1704 iflib_txsd_alloc(iflib_txq_t txq)
1705 {
1706 	if_ctx_t ctx = txq->ift_ctx;
1707 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1708 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1709 	device_t dev = ctx->ifc_dev;
1710 	bus_size_t tsomaxsize;
1711 	bus_addr_t lowaddr;
1712 	int err, nsegments, ntsosegments;
1713 	bool tso;
1714 
1715 	nsegments = scctx->isc_tx_nsegments;
1716 	ntsosegments = scctx->isc_tx_tso_segments_max;
1717 	tsomaxsize = scctx->isc_tx_tso_size_max;
1718 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1719 		tsomaxsize += sizeof(struct ether_vlan_header);
1720 	MPASS(scctx->isc_ntxd[0] > 0);
1721 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1722 	MPASS(nsegments > 0);
1723 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1724 		MPASS(ntsosegments > 0);
1725 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1726 	}
1727 
1728 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1729 
1730 	/*
1731 	 * Set up DMA tags for TX buffers.
1732 	 */
1733 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1734 			       1, 0,			/* alignment, bounds */
1735 			       lowaddr,			/* lowaddr */
1736 			       BUS_SPACE_MAXADDR,	/* highaddr */
1737 			       NULL, NULL,		/* filter, filterarg */
1738 			       sctx->isc_tx_maxsize,		/* maxsize */
1739 			       nsegments,	/* nsegments */
1740 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1741 			       0,			/* flags */
1742 			       NULL,			/* lockfunc */
1743 			       NULL,			/* lockfuncarg */
1744 			       &txq->ift_buf_tag))) {
1745 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1746 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1747 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1748 		goto fail;
1749 	}
1750 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1751 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1752 			       1, 0,			/* alignment, bounds */
1753 			       lowaddr,			/* lowaddr */
1754 			       BUS_SPACE_MAXADDR,	/* highaddr */
1755 			       NULL, NULL,		/* filter, filterarg */
1756 			       tsomaxsize,		/* maxsize */
1757 			       ntsosegments,	/* nsegments */
1758 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1759 			       0,			/* flags */
1760 			       NULL,			/* lockfunc */
1761 			       NULL,			/* lockfuncarg */
1762 			       &txq->ift_tso_buf_tag))) {
1763 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1764 		    err);
1765 		goto fail;
1766 	}
1767 
1768 	/* Allocate memory for the TX mbuf map. */
1769 	if (!(txq->ift_sds.ifsd_m =
1770 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1771 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1772 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1773 		err = ENOMEM;
1774 		goto fail;
1775 	}
1776 
1777 	/*
1778 	 * Create the DMA maps for TX buffers.
1779 	 */
1780 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1781 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1782 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1783 		device_printf(dev,
1784 		    "Unable to allocate TX buffer DMA map memory\n");
1785 		err = ENOMEM;
1786 		goto fail;
1787 	}
1788 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1789 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1790 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1791 		device_printf(dev,
1792 		    "Unable to allocate TSO TX buffer map memory\n");
1793 		err = ENOMEM;
1794 		goto fail;
1795 	}
1796 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1797 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1798 		    &txq->ift_sds.ifsd_map[i]);
1799 		if (err != 0) {
1800 			device_printf(dev, "Unable to create TX DMA map\n");
1801 			goto fail;
1802 		}
1803 		if (!tso)
1804 			continue;
1805 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1806 		    &txq->ift_sds.ifsd_tso_map[i]);
1807 		if (err != 0) {
1808 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1809 			goto fail;
1810 		}
1811 	}
1812 	return (0);
1813 fail:
1814 	/* We free all, it handles case where we are in the middle */
1815 	iflib_tx_structures_free(ctx);
1816 	return (err);
1817 }
1818 
1819 static void
1820 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1821 {
1822 	bus_dmamap_t map;
1823 
1824 	if (txq->ift_sds.ifsd_map != NULL) {
1825 		map = txq->ift_sds.ifsd_map[i];
1826 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1827 		bus_dmamap_unload(txq->ift_buf_tag, map);
1828 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1829 		txq->ift_sds.ifsd_map[i] = NULL;
1830 	}
1831 
1832 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1833 		map = txq->ift_sds.ifsd_tso_map[i];
1834 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1835 		    BUS_DMASYNC_POSTWRITE);
1836 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1837 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1838 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1839 	}
1840 }
1841 
1842 static void
1843 iflib_txq_destroy(iflib_txq_t txq)
1844 {
1845 	if_ctx_t ctx = txq->ift_ctx;
1846 
1847 	for (int i = 0; i < txq->ift_size; i++)
1848 		iflib_txsd_destroy(ctx, txq, i);
1849 
1850 	if (txq->ift_br != NULL) {
1851 		ifmp_ring_free(txq->ift_br);
1852 		txq->ift_br = NULL;
1853 	}
1854 
1855 	mtx_destroy(&txq->ift_mtx);
1856 
1857 	if (txq->ift_sds.ifsd_map != NULL) {
1858 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1859 		txq->ift_sds.ifsd_map = NULL;
1860 	}
1861 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1862 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1863 		txq->ift_sds.ifsd_tso_map = NULL;
1864 	}
1865 	if (txq->ift_sds.ifsd_m != NULL) {
1866 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1867 		txq->ift_sds.ifsd_m = NULL;
1868 	}
1869 	if (txq->ift_buf_tag != NULL) {
1870 		bus_dma_tag_destroy(txq->ift_buf_tag);
1871 		txq->ift_buf_tag = NULL;
1872 	}
1873 	if (txq->ift_tso_buf_tag != NULL) {
1874 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1875 		txq->ift_tso_buf_tag = NULL;
1876 	}
1877 	if (txq->ift_ifdi != NULL) {
1878 		free(txq->ift_ifdi, M_IFLIB);
1879 	}
1880 }
1881 
1882 static void
1883 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1884 {
1885 	struct mbuf **mp;
1886 
1887 	mp = &txq->ift_sds.ifsd_m[i];
1888 	if (*mp == NULL)
1889 		return;
1890 
1891 	if (txq->ift_sds.ifsd_map != NULL) {
1892 		bus_dmamap_sync(txq->ift_buf_tag,
1893 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1894 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1895 	}
1896 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1897 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1898 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1899 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1900 		    txq->ift_sds.ifsd_tso_map[i]);
1901 	}
1902 	m_freem(*mp);
1903 	DBG_COUNTER_INC(tx_frees);
1904 	*mp = NULL;
1905 }
1906 
1907 static int
1908 iflib_txq_setup(iflib_txq_t txq)
1909 {
1910 	if_ctx_t ctx = txq->ift_ctx;
1911 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1912 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1913 	iflib_dma_info_t di;
1914 	int i;
1915 
1916 	/* Set number of descriptors available */
1917 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1918 	/* XXX make configurable */
1919 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1920 
1921 	/* Reset indices */
1922 	txq->ift_cidx_processed = 0;
1923 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1924 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1925 
1926 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1927 		bzero((void *)di->idi_vaddr, di->idi_size);
1928 
1929 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1930 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1931 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1932 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1933 	return (0);
1934 }
1935 
1936 /*********************************************************************
1937  *
1938  *  Allocate DMA resources for RX buffers as well as memory for the RX
1939  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1940  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1941  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1942  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1943  *  packet, the maximum number of entries we'll need is equal to the
1944  *  number of hardware receive descriptors that we've allocated.
1945  *
1946  **********************************************************************/
1947 static int
1948 iflib_rxsd_alloc(iflib_rxq_t rxq)
1949 {
1950 	if_ctx_t ctx = rxq->ifr_ctx;
1951 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1952 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1953 	device_t dev = ctx->ifc_dev;
1954 	iflib_fl_t fl;
1955 	bus_addr_t lowaddr;
1956 	int			err;
1957 
1958 	MPASS(scctx->isc_nrxd[0] > 0);
1959 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1960 
1961 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1962 
1963 	fl = rxq->ifr_fl;
1964 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1965 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1966 		/* Set up DMA tag for RX buffers. */
1967 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1968 					 1, 0,			/* alignment, bounds */
1969 					 lowaddr,		/* lowaddr */
1970 					 BUS_SPACE_MAXADDR,	/* highaddr */
1971 					 NULL, NULL,		/* filter, filterarg */
1972 					 sctx->isc_rx_maxsize,	/* maxsize */
1973 					 sctx->isc_rx_nsegments,	/* nsegments */
1974 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1975 					 0,			/* flags */
1976 					 NULL,			/* lockfunc */
1977 					 NULL,			/* lockarg */
1978 					 &fl->ifl_buf_tag);
1979 		if (err) {
1980 			device_printf(dev,
1981 			    "Unable to allocate RX DMA tag: %d\n", err);
1982 			goto fail;
1983 		}
1984 
1985 		/* Allocate memory for the RX mbuf map. */
1986 		if (!(fl->ifl_sds.ifsd_m =
1987 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1988 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1989 			device_printf(dev,
1990 			    "Unable to allocate RX mbuf map memory\n");
1991 			err = ENOMEM;
1992 			goto fail;
1993 		}
1994 
1995 		/* Allocate memory for the direct RX cluster pointer map. */
1996 		if (!(fl->ifl_sds.ifsd_cl =
1997 		      (caddr_t *) malloc(sizeof(caddr_t) *
1998 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1999 			device_printf(dev,
2000 			    "Unable to allocate RX cluster map memory\n");
2001 			err = ENOMEM;
2002 			goto fail;
2003 		}
2004 
2005 		/* Allocate memory for the RX cluster bus address map. */
2006 		if (!(fl->ifl_sds.ifsd_ba =
2007 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2008 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2009 			device_printf(dev,
2010 			    "Unable to allocate RX bus address map memory\n");
2011 			err = ENOMEM;
2012 			goto fail;
2013 		}
2014 
2015 		/*
2016 		 * Create the DMA maps for RX buffers.
2017 		 */
2018 		if (!(fl->ifl_sds.ifsd_map =
2019 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2020 			device_printf(dev,
2021 			    "Unable to allocate RX buffer DMA map memory\n");
2022 			err = ENOMEM;
2023 			goto fail;
2024 		}
2025 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2026 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2027 			    &fl->ifl_sds.ifsd_map[i]);
2028 			if (err != 0) {
2029 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2030 				goto fail;
2031 			}
2032 		}
2033 	}
2034 	return (0);
2035 
2036 fail:
2037 	iflib_rx_structures_free(ctx);
2038 	return (err);
2039 }
2040 
2041 /*
2042  * Internal service routines
2043  */
2044 
2045 struct rxq_refill_cb_arg {
2046 	int               error;
2047 	bus_dma_segment_t seg;
2048 	int               nseg;
2049 };
2050 
2051 static void
2052 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2053 {
2054 	struct rxq_refill_cb_arg *cb_arg = arg;
2055 
2056 	cb_arg->error = error;
2057 	cb_arg->seg = segs[0];
2058 	cb_arg->nseg = nseg;
2059 }
2060 
2061 /**
2062  * iflib_fl_refill - refill an rxq free-buffer list
2063  * @ctx: the iflib context
2064  * @fl: the free list to refill
2065  * @count: the number of new buffers to allocate
2066  *
2067  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2068  * The caller must assure that @count does not exceed the queue's capacity
2069  * minus one (since we always leave a descriptor unavailable).
2070  */
2071 static uint8_t
2072 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2073 {
2074 	struct if_rxd_update iru;
2075 	struct rxq_refill_cb_arg cb_arg;
2076 	struct mbuf *m;
2077 	caddr_t cl, *sd_cl;
2078 	struct mbuf **sd_m;
2079 	bus_dmamap_t *sd_map;
2080 	bus_addr_t bus_addr, *sd_ba;
2081 	int err, frag_idx, i, idx, n, pidx;
2082 	qidx_t credits;
2083 
2084 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2085 
2086 	sd_m = fl->ifl_sds.ifsd_m;
2087 	sd_map = fl->ifl_sds.ifsd_map;
2088 	sd_cl = fl->ifl_sds.ifsd_cl;
2089 	sd_ba = fl->ifl_sds.ifsd_ba;
2090 	pidx = fl->ifl_pidx;
2091 	idx = pidx;
2092 	frag_idx = fl->ifl_fragidx;
2093 	credits = fl->ifl_credits;
2094 
2095 	i = 0;
2096 	n = count;
2097 	MPASS(n > 0);
2098 	MPASS(credits + n <= fl->ifl_size);
2099 
2100 	if (pidx < fl->ifl_cidx)
2101 		MPASS(pidx + n <= fl->ifl_cidx);
2102 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2103 		MPASS(fl->ifl_gen == 0);
2104 	if (pidx > fl->ifl_cidx)
2105 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2106 
2107 	DBG_COUNTER_INC(fl_refills);
2108 	if (n > 8)
2109 		DBG_COUNTER_INC(fl_refills_large);
2110 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2111 	while (n-- > 0) {
2112 		/*
2113 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2114 		 * initialized after rx.
2115 		 *
2116 		 * If the cluster is still set then we know a minimum sized
2117 		 * packet was received
2118 		 */
2119 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2120 		    &frag_idx);
2121 		if (frag_idx < 0)
2122 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2123 		MPASS(frag_idx >= 0);
2124 		if ((cl = sd_cl[frag_idx]) == NULL) {
2125 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2126 			if (__predict_false(cl == NULL))
2127 				break;
2128 
2129 			cb_arg.error = 0;
2130 			MPASS(sd_map != NULL);
2131 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2132 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2133 			    BUS_DMA_NOWAIT);
2134 			if (__predict_false(err != 0 || cb_arg.error)) {
2135 				uma_zfree(fl->ifl_zone, cl);
2136 				break;
2137 			}
2138 
2139 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2140 			sd_cl[frag_idx] = cl;
2141 #if MEMORY_LOGGING
2142 			fl->ifl_cl_enqueued++;
2143 #endif
2144 		} else {
2145 			bus_addr = sd_ba[frag_idx];
2146 		}
2147 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2148 		    BUS_DMASYNC_PREREAD);
2149 
2150 		if (sd_m[frag_idx] == NULL) {
2151 			m = m_gethdr_raw(M_NOWAIT, 0);
2152 			if (__predict_false(m == NULL))
2153 				break;
2154 			sd_m[frag_idx] = m;
2155 		}
2156 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2157 #if MEMORY_LOGGING
2158 		fl->ifl_m_enqueued++;
2159 #endif
2160 
2161 		DBG_COUNTER_INC(rx_allocs);
2162 		fl->ifl_rxd_idxs[i] = frag_idx;
2163 		fl->ifl_bus_addrs[i] = bus_addr;
2164 		credits++;
2165 		i++;
2166 		MPASS(credits <= fl->ifl_size);
2167 		if (++idx == fl->ifl_size) {
2168 #ifdef INVARIANTS
2169 			fl->ifl_gen = 1;
2170 #endif
2171 			idx = 0;
2172 		}
2173 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2174 			iru.iru_pidx = pidx;
2175 			iru.iru_count = i;
2176 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2177 			fl->ifl_pidx = idx;
2178 			fl->ifl_credits = credits;
2179 			pidx = idx;
2180 			i = 0;
2181 		}
2182 	}
2183 
2184 	if (n < count - 1) {
2185 		if (i != 0) {
2186 			iru.iru_pidx = pidx;
2187 			iru.iru_count = i;
2188 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2189 			fl->ifl_pidx = idx;
2190 			fl->ifl_credits = credits;
2191 		}
2192 		DBG_COUNTER_INC(rxd_flush);
2193 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2194 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2195 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2196 		    fl->ifl_id, fl->ifl_pidx);
2197 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2198 			fl->ifl_fragidx = frag_idx + 1;
2199 			if (fl->ifl_fragidx == fl->ifl_size)
2200 				fl->ifl_fragidx = 0;
2201 		} else {
2202 			fl->ifl_fragidx = frag_idx;
2203 		}
2204 	}
2205 
2206 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2207 }
2208 
2209 static inline uint8_t
2210 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2211 {
2212 	/*
2213 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2214 	 * This is important as it confuses most NICs. For instance,
2215 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2216 	 * RDH points to the next receive descriptor to be used by the NIC,
2217 	 * and RDT for the next receive descriptor to be published by the
2218 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2219 	 * The condition RDH == RDT means no descriptors are available to
2220 	 * the NIC, and thus it would be ambiguous if it also meant that
2221 	 * all the descriptors are available to the NIC.
2222 	 */
2223 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2224 #ifdef INVARIANTS
2225 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2226 #endif
2227 
2228 	MPASS(fl->ifl_credits <= fl->ifl_size);
2229 	MPASS(reclaimable == delta);
2230 
2231 	if (reclaimable > 0)
2232 		return (iflib_fl_refill(ctx, fl, reclaimable));
2233 	return (0);
2234 }
2235 
2236 uint8_t
2237 iflib_in_detach(if_ctx_t ctx)
2238 {
2239 	bool in_detach;
2240 
2241 	STATE_LOCK(ctx);
2242 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2243 	STATE_UNLOCK(ctx);
2244 	return (in_detach);
2245 }
2246 
2247 static void
2248 iflib_fl_bufs_free(iflib_fl_t fl)
2249 {
2250 	iflib_dma_info_t idi = fl->ifl_ifdi;
2251 	bus_dmamap_t sd_map;
2252 	uint32_t i;
2253 
2254 	for (i = 0; i < fl->ifl_size; i++) {
2255 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2256 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2257 
2258 		if (*sd_cl != NULL) {
2259 			sd_map = fl->ifl_sds.ifsd_map[i];
2260 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2261 			    BUS_DMASYNC_POSTREAD);
2262 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2263 			uma_zfree(fl->ifl_zone, *sd_cl);
2264 			*sd_cl = NULL;
2265 			if (*sd_m != NULL) {
2266 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2267 				m_free_raw(*sd_m);
2268 				*sd_m = NULL;
2269 			}
2270 		} else {
2271 			MPASS(*sd_m == NULL);
2272 		}
2273 #if MEMORY_LOGGING
2274 		fl->ifl_m_dequeued++;
2275 		fl->ifl_cl_dequeued++;
2276 #endif
2277 	}
2278 #ifdef INVARIANTS
2279 	for (i = 0; i < fl->ifl_size; i++) {
2280 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2281 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2282 	}
2283 #endif
2284 	/*
2285 	 * Reset free list values
2286 	 */
2287 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2288 	bzero(idi->idi_vaddr, idi->idi_size);
2289 }
2290 
2291 /*********************************************************************
2292  *
2293  *  Initialize a free list and its buffers.
2294  *
2295  **********************************************************************/
2296 static int
2297 iflib_fl_setup(iflib_fl_t fl)
2298 {
2299 	iflib_rxq_t rxq = fl->ifl_rxq;
2300 	if_ctx_t ctx = rxq->ifr_ctx;
2301 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2302 	int qidx;
2303 
2304 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2305 	/*
2306 	** Free current RX buffer structs and their mbufs
2307 	*/
2308 	iflib_fl_bufs_free(fl);
2309 	/* Now replenish the mbufs */
2310 	MPASS(fl->ifl_credits == 0);
2311 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2312 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2313 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2314 	else
2315 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2316 	/*
2317 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2318 	 * to the selected mbuf size.
2319 	 */
2320 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2321 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2322 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2323 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2324 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2325 
2326 	/*
2327 	 * Avoid pre-allocating zillions of clusters to an idle card
2328 	 * potentially speeding up attach. In any case make sure
2329 	 * to leave a descriptor unavailable. See the comment in
2330 	 * iflib_fl_refill_all().
2331 	 */
2332 	MPASS(fl->ifl_size > 0);
2333 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2334 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2335 		return (ENOBUFS);
2336 	/*
2337 	 * handle failure
2338 	 */
2339 	MPASS(rxq != NULL);
2340 	MPASS(fl->ifl_ifdi != NULL);
2341 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2342 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2343 	return (0);
2344 }
2345 
2346 /*********************************************************************
2347  *
2348  *  Free receive ring data structures
2349  *
2350  **********************************************************************/
2351 static void
2352 iflib_rx_sds_free(iflib_rxq_t rxq)
2353 {
2354 	iflib_fl_t fl;
2355 	int i, j;
2356 
2357 	if (rxq->ifr_fl != NULL) {
2358 		for (i = 0; i < rxq->ifr_nfl; i++) {
2359 			fl = &rxq->ifr_fl[i];
2360 			if (fl->ifl_buf_tag != NULL) {
2361 				if (fl->ifl_sds.ifsd_map != NULL) {
2362 					for (j = 0; j < fl->ifl_size; j++) {
2363 						bus_dmamap_sync(
2364 						    fl->ifl_buf_tag,
2365 						    fl->ifl_sds.ifsd_map[j],
2366 						    BUS_DMASYNC_POSTREAD);
2367 						bus_dmamap_unload(
2368 						    fl->ifl_buf_tag,
2369 						    fl->ifl_sds.ifsd_map[j]);
2370 						bus_dmamap_destroy(
2371 						    fl->ifl_buf_tag,
2372 						    fl->ifl_sds.ifsd_map[j]);
2373 					}
2374 				}
2375 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2376 				fl->ifl_buf_tag = NULL;
2377 			}
2378 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2379 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2380 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2381 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2382 			free(fl->ifl_rx_bitmap, M_IFLIB);
2383 			fl->ifl_sds.ifsd_m = NULL;
2384 			fl->ifl_sds.ifsd_cl = NULL;
2385 			fl->ifl_sds.ifsd_ba = NULL;
2386 			fl->ifl_sds.ifsd_map = NULL;
2387 			fl->ifl_rx_bitmap = NULL;
2388 		}
2389 		free(rxq->ifr_fl, M_IFLIB);
2390 		rxq->ifr_fl = NULL;
2391 		free(rxq->ifr_ifdi, M_IFLIB);
2392 		rxq->ifr_ifdi = NULL;
2393 		rxq->ifr_cq_cidx = 0;
2394 	}
2395 }
2396 
2397 /*
2398  * Timer routine
2399  */
2400 static void
2401 iflib_timer(void *arg)
2402 {
2403 	iflib_txq_t txq = arg;
2404 	if_ctx_t ctx = txq->ift_ctx;
2405 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2406 	uint64_t this_tick = ticks;
2407 
2408 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2409 		return;
2410 
2411 	/*
2412 	** Check on the state of the TX queue(s), this
2413 	** can be done without the lock because its RO
2414 	** and the HUNG state will be static if set.
2415 	*/
2416 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2417 		txq->ift_last_timer_tick = this_tick;
2418 		IFDI_TIMER(ctx, txq->ift_id);
2419 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2420 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2421 		     (sctx->isc_pause_frames == 0)))
2422 			goto hung;
2423 
2424 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2425 		    ifmp_ring_is_stalled(txq->ift_br)) {
2426 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2427 			    ("queue can't be marked as hung if interface is down"));
2428 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2429 		}
2430 		txq->ift_cleaned_prev = txq->ift_cleaned;
2431 	}
2432 	/* handle any laggards */
2433 	if (txq->ift_db_pending)
2434 		GROUPTASK_ENQUEUE(&txq->ift_task);
2435 
2436 	sctx->isc_pause_frames = 0;
2437 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2438 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2439 		    txq, txq->ift_timer.c_cpu);
2440 	return;
2441 
2442  hung:
2443 	device_printf(ctx->ifc_dev,
2444 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2445 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2446 	STATE_LOCK(ctx);
2447 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2448 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2449 	iflib_admin_intr_deferred(ctx);
2450 	STATE_UNLOCK(ctx);
2451 }
2452 
2453 static uint16_t
2454 iflib_get_mbuf_size_for(unsigned int size)
2455 {
2456 
2457 	if (size <= MCLBYTES)
2458 		return (MCLBYTES);
2459 	else
2460 		return (MJUMPAGESIZE);
2461 }
2462 
2463 static void
2464 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2465 {
2466 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2467 
2468 	/*
2469 	 * XXX don't set the max_frame_size to larger
2470 	 * than the hardware can handle
2471 	 */
2472 	ctx->ifc_rx_mbuf_sz =
2473 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2474 }
2475 
2476 uint32_t
2477 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2478 {
2479 
2480 	return (ctx->ifc_rx_mbuf_sz);
2481 }
2482 
2483 static void
2484 iflib_init_locked(if_ctx_t ctx)
2485 {
2486 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2487 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2488 	if_t ifp = ctx->ifc_ifp;
2489 	iflib_fl_t fl;
2490 	iflib_txq_t txq;
2491 	iflib_rxq_t rxq;
2492 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2493 
2494 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2495 	IFDI_INTR_DISABLE(ctx);
2496 
2497 	/*
2498 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2499 	 * called without first calling iflib_stop().
2500 	 */
2501 	netmap_disable_all_rings(ifp);
2502 
2503 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2504 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2505 	/* Set hardware offload abilities */
2506 	if_clearhwassist(ifp);
2507 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2508 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2509 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2510 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2511 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2512 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2513 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2514 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2515 
2516 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2517 		CALLOUT_LOCK(txq);
2518 		callout_stop(&txq->ift_timer);
2519 #ifdef DEV_NETMAP
2520 		callout_stop(&txq->ift_netmap_timer);
2521 #endif /* DEV_NETMAP */
2522 		CALLOUT_UNLOCK(txq);
2523 		(void)iflib_netmap_txq_init(ctx, txq);
2524 	}
2525 
2526 	/*
2527 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2528 	 * that drivers can use the value when setting up the hardware receive
2529 	 * buffers.
2530 	 */
2531 	iflib_calc_rx_mbuf_sz(ctx);
2532 
2533 #ifdef INVARIANTS
2534 	i = if_getdrvflags(ifp);
2535 #endif
2536 	IFDI_INIT(ctx);
2537 	MPASS(if_getdrvflags(ifp) == i);
2538 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2539 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2540 			/* This rxq is in netmap mode. Skip normal init. */
2541 			continue;
2542 		}
2543 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2544 			if (iflib_fl_setup(fl)) {
2545 				device_printf(ctx->ifc_dev,
2546 				    "setting up free list %d failed - "
2547 				    "check cluster settings\n", j);
2548 				goto done;
2549 			}
2550 		}
2551 	}
2552 done:
2553 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2554 	IFDI_INTR_ENABLE(ctx);
2555 	txq = ctx->ifc_txqs;
2556 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2557 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2558 			txq->ift_timer.c_cpu);
2559 
2560         /* Re-enable txsync/rxsync. */
2561 	netmap_enable_all_rings(ifp);
2562 }
2563 
2564 static int
2565 iflib_media_change(if_t ifp)
2566 {
2567 	if_ctx_t ctx = if_getsoftc(ifp);
2568 	int err;
2569 
2570 	CTX_LOCK(ctx);
2571 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2572 		iflib_if_init_locked(ctx);
2573 	CTX_UNLOCK(ctx);
2574 	return (err);
2575 }
2576 
2577 static void
2578 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2579 {
2580 	if_ctx_t ctx = if_getsoftc(ifp);
2581 
2582 	CTX_LOCK(ctx);
2583 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2584 	IFDI_MEDIA_STATUS(ctx, ifmr);
2585 	CTX_UNLOCK(ctx);
2586 }
2587 
2588 void
2589 iflib_stop(if_ctx_t ctx)
2590 {
2591 	iflib_txq_t txq = ctx->ifc_txqs;
2592 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2593 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2594 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2595 	iflib_dma_info_t di;
2596 	iflib_fl_t fl;
2597 	int i, j;
2598 
2599 	/* Tell the stack that the interface is no longer active */
2600 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2601 
2602 	IFDI_INTR_DISABLE(ctx);
2603 	DELAY(1000);
2604 	IFDI_STOP(ctx);
2605 	DELAY(1000);
2606 
2607 	/*
2608 	 * Stop any pending txsync/rxsync and prevent new ones
2609 	 * form starting. Processes blocked in poll() will get
2610 	 * POLLERR.
2611 	 */
2612 	netmap_disable_all_rings(ctx->ifc_ifp);
2613 
2614 	iflib_debug_reset();
2615 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2616 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2617 		/* make sure all transmitters have completed before proceeding XXX */
2618 
2619 		CALLOUT_LOCK(txq);
2620 		callout_stop(&txq->ift_timer);
2621 #ifdef DEV_NETMAP
2622 		callout_stop(&txq->ift_netmap_timer);
2623 #endif /* DEV_NETMAP */
2624 		CALLOUT_UNLOCK(txq);
2625 
2626 		/* clean any enqueued buffers */
2627 		iflib_ifmp_purge(txq);
2628 		/* Free any existing tx buffers. */
2629 		for (j = 0; j < txq->ift_size; j++) {
2630 			iflib_txsd_free(ctx, txq, j);
2631 		}
2632 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2633 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2634 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2635 			txq->ift_cidx = txq->ift_pidx;
2636 		else
2637 			txq->ift_cidx = txq->ift_pidx = 0;
2638 
2639 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2640 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2641 		txq->ift_pullups = 0;
2642 		ifmp_ring_reset_stats(txq->ift_br);
2643 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2644 			bzero((void *)di->idi_vaddr, di->idi_size);
2645 	}
2646 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2647 		if (rxq->ifr_task.gt_taskqueue != NULL)
2648 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2649 				 &rxq->ifr_task.gt_task);
2650 
2651 		rxq->ifr_cq_cidx = 0;
2652 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2653 			bzero((void *)di->idi_vaddr, di->idi_size);
2654 		/* also resets the free lists pidx/cidx */
2655 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2656 			iflib_fl_bufs_free(fl);
2657 	}
2658 }
2659 
2660 static inline caddr_t
2661 calc_next_rxd(iflib_fl_t fl, int cidx)
2662 {
2663 	qidx_t size;
2664 	int nrxd;
2665 	caddr_t start, end, cur, next;
2666 
2667 	nrxd = fl->ifl_size;
2668 	size = fl->ifl_rxd_size;
2669 	start = fl->ifl_ifdi->idi_vaddr;
2670 
2671 	if (__predict_false(size == 0))
2672 		return (start);
2673 	cur = start + size*cidx;
2674 	end = start + size*nrxd;
2675 	next = CACHE_PTR_NEXT(cur);
2676 	return (next < end ? next : start);
2677 }
2678 
2679 static inline void
2680 prefetch_pkts(iflib_fl_t fl, int cidx)
2681 {
2682 	int nextptr;
2683 	int nrxd = fl->ifl_size;
2684 	caddr_t next_rxd;
2685 
2686 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2687 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2688 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2689 	next_rxd = calc_next_rxd(fl, cidx);
2690 	prefetch(next_rxd);
2691 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2692 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2693 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2694 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2695 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2696 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2697 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2698 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2699 }
2700 
2701 static struct mbuf *
2702 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2703     int *pf_rv, if_rxd_info_t ri)
2704 {
2705 	bus_dmamap_t map;
2706 	iflib_fl_t fl;
2707 	caddr_t payload;
2708 	struct mbuf *m;
2709 	int flid, cidx, len, next;
2710 
2711 	map = NULL;
2712 	flid = irf->irf_flid;
2713 	cidx = irf->irf_idx;
2714 	fl = &rxq->ifr_fl[flid];
2715 	sd->ifsd_fl = fl;
2716 	m = fl->ifl_sds.ifsd_m[cidx];
2717 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2718 	fl->ifl_credits--;
2719 #if MEMORY_LOGGING
2720 	fl->ifl_m_dequeued++;
2721 #endif
2722 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2723 		prefetch_pkts(fl, cidx);
2724 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2725 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2726 	map = fl->ifl_sds.ifsd_map[cidx];
2727 
2728 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2729 
2730 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2731 	    irf->irf_len != 0) {
2732 		payload  = *sd->ifsd_cl;
2733 		payload +=  ri->iri_pad;
2734 		len = ri->iri_len - ri->iri_pad;
2735 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2736 		switch (*pf_rv) {
2737 		case PFIL_DROPPED:
2738 		case PFIL_CONSUMED:
2739 			/*
2740 			 * The filter ate it.  Everything is recycled.
2741 			 */
2742 			m = NULL;
2743 			unload = 0;
2744 			break;
2745 		case PFIL_REALLOCED:
2746 			/*
2747 			 * The filter copied it.  Everything is recycled.
2748 			 * 'm' points at new mbuf.
2749 			 */
2750 			unload = 0;
2751 			break;
2752 		case PFIL_PASS:
2753 			/*
2754 			 * Filter said it was OK, so receive like
2755 			 * normal
2756 			 */
2757 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2758 			break;
2759 		default:
2760 			MPASS(0);
2761 		}
2762 	} else {
2763 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2764 		if (pf_rv != NULL)
2765 			*pf_rv = PFIL_PASS;
2766 	}
2767 
2768 	if (unload && irf->irf_len != 0)
2769 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2770 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2771 	if (__predict_false(fl->ifl_cidx == 0))
2772 		fl->ifl_gen = 0;
2773 	bit_clear(fl->ifl_rx_bitmap, cidx);
2774 	return (m);
2775 }
2776 
2777 static struct mbuf *
2778 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2779 {
2780 	struct mbuf *m, *mh, *mt;
2781 	caddr_t cl;
2782 	int  *pf_rv_ptr, flags, i, padlen;
2783 	bool consumed;
2784 
2785 	i = 0;
2786 	mh = NULL;
2787 	consumed = false;
2788 	*pf_rv = PFIL_PASS;
2789 	pf_rv_ptr = pf_rv;
2790 	do {
2791 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2792 		    pf_rv_ptr, ri);
2793 
2794 		MPASS(*sd->ifsd_cl != NULL);
2795 
2796 		/*
2797 		 * Exclude zero-length frags & frags from
2798 		 * packets the filter has consumed or dropped
2799 		 */
2800 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2801 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2802 			if (mh == NULL) {
2803 				/* everything saved here */
2804 				consumed = true;
2805 				pf_rv_ptr = NULL;
2806 				continue;
2807 			}
2808 			/* XXX we can save the cluster here, but not the mbuf */
2809 			m_init(m, M_NOWAIT, MT_DATA, 0);
2810 			m_free(m);
2811 			continue;
2812 		}
2813 		if (mh == NULL) {
2814 			flags = M_PKTHDR|M_EXT;
2815 			mh = mt = m;
2816 			padlen = ri->iri_pad;
2817 		} else {
2818 			flags = M_EXT;
2819 			mt->m_next = m;
2820 			mt = m;
2821 			/* assuming padding is only on the first fragment */
2822 			padlen = 0;
2823 		}
2824 		cl = *sd->ifsd_cl;
2825 		*sd->ifsd_cl = NULL;
2826 
2827 		/* Can these two be made one ? */
2828 		m_init(m, M_NOWAIT, MT_DATA, flags);
2829 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2830 		/*
2831 		 * These must follow m_init and m_cljset
2832 		 */
2833 		m->m_data += padlen;
2834 		ri->iri_len -= padlen;
2835 		m->m_len = ri->iri_frags[i].irf_len;
2836 	} while (++i < ri->iri_nfrags);
2837 
2838 	return (mh);
2839 }
2840 
2841 /*
2842  * Process one software descriptor
2843  */
2844 static struct mbuf *
2845 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2846 {
2847 	struct if_rxsd sd;
2848 	struct mbuf *m;
2849 	int pf_rv;
2850 
2851 	/* should I merge this back in now that the two paths are basically duplicated? */
2852 	if (ri->iri_nfrags == 1 &&
2853 	    ri->iri_frags[0].irf_len != 0 &&
2854 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2855 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2856 		    &pf_rv, ri);
2857 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2858 			return (m);
2859 		if (pf_rv == PFIL_PASS) {
2860 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2861 #ifndef __NO_STRICT_ALIGNMENT
2862 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2863 				m->m_data += 2;
2864 #endif
2865 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2866 			m->m_len = ri->iri_frags[0].irf_len;
2867 			m->m_data += ri->iri_pad;
2868 			ri->iri_len -= ri->iri_pad;
2869 		}
2870 	} else {
2871 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2872 		if (m == NULL)
2873 			return (NULL);
2874 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2875 			return (m);
2876 	}
2877 	m->m_pkthdr.len = ri->iri_len;
2878 	m->m_pkthdr.rcvif = ri->iri_ifp;
2879 	m->m_flags |= ri->iri_flags;
2880 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2881 	m->m_pkthdr.flowid = ri->iri_flowid;
2882 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2883 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2884 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2885 	return (m);
2886 }
2887 
2888 #if defined(INET6) || defined(INET)
2889 static void
2890 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2891 {
2892 	CURVNET_SET(lc->ifp->if_vnet); /* XXX - DRVAPI */
2893 #if defined(INET6)
2894 	*v6 = V_ip6_forwarding;
2895 #endif
2896 #if defined(INET)
2897 	*v4 = V_ipforwarding;
2898 #endif
2899 	CURVNET_RESTORE();
2900 }
2901 
2902 /*
2903  * Returns true if it's possible this packet could be LROed.
2904  * if it returns false, it is guaranteed that tcp_lro_rx()
2905  * would not return zero.
2906  */
2907 static bool
2908 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2909 {
2910 	struct ether_header *eh;
2911 
2912 	eh = mtod(m, struct ether_header *);
2913 	switch (eh->ether_type) {
2914 #if defined(INET6)
2915 		case htons(ETHERTYPE_IPV6):
2916 			return (!v6_forwarding);
2917 #endif
2918 #if defined (INET)
2919 		case htons(ETHERTYPE_IP):
2920 			return (!v4_forwarding);
2921 #endif
2922 	}
2923 
2924 	return false;
2925 }
2926 #else
2927 static void
2928 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2929 {
2930 }
2931 #endif
2932 
2933 static void
2934 _task_fn_rx_watchdog(void *context)
2935 {
2936 	iflib_rxq_t rxq = context;
2937 
2938 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2939 }
2940 
2941 static uint8_t
2942 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2943 {
2944 	if_t ifp;
2945 	if_ctx_t ctx = rxq->ifr_ctx;
2946 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2947 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2948 	int avail, i;
2949 	qidx_t *cidxp;
2950 	struct if_rxd_info ri;
2951 	int err, budget_left, rx_bytes, rx_pkts;
2952 	iflib_fl_t fl;
2953 	int lro_enabled;
2954 	bool v4_forwarding, v6_forwarding, lro_possible;
2955 	uint8_t retval = 0;
2956 
2957 	/*
2958 	 * XXX early demux data packets so that if_input processing only handles
2959 	 * acks in interrupt context
2960 	 */
2961 	struct mbuf *m, *mh, *mt, *mf;
2962 
2963 	NET_EPOCH_ASSERT();
2964 
2965 	lro_possible = v4_forwarding = v6_forwarding = false;
2966 	ifp = ctx->ifc_ifp;
2967 	mh = mt = NULL;
2968 	MPASS(budget > 0);
2969 	rx_pkts	= rx_bytes = 0;
2970 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2971 		cidxp = &rxq->ifr_cq_cidx;
2972 	else
2973 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2974 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2975 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2976 			retval |= iflib_fl_refill_all(ctx, fl);
2977 		DBG_COUNTER_INC(rx_unavail);
2978 		return (retval);
2979 	}
2980 
2981 	/* pfil needs the vnet to be set */
2982 	CURVNET_SET_QUIET(ifp->if_vnet); /* XXX - DRVAPI */
2983 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2984 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2985 			DBG_COUNTER_INC(rx_ctx_inactive);
2986 			break;
2987 		}
2988 		/*
2989 		 * Reset client set fields to their default values
2990 		 */
2991 		rxd_info_zero(&ri);
2992 		ri.iri_qsidx = rxq->ifr_id;
2993 		ri.iri_cidx = *cidxp;
2994 		ri.iri_ifp = ifp;
2995 		ri.iri_frags = rxq->ifr_frags;
2996 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2997 
2998 		if (err)
2999 			goto err;
3000 		rx_pkts += 1;
3001 		rx_bytes += ri.iri_len;
3002 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3003 			*cidxp = ri.iri_cidx;
3004 			/* Update our consumer index */
3005 			/* XXX NB: shurd - check if this is still safe */
3006 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3007 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3008 			/* was this only a completion queue message? */
3009 			if (__predict_false(ri.iri_nfrags == 0))
3010 				continue;
3011 		}
3012 		MPASS(ri.iri_nfrags != 0);
3013 		MPASS(ri.iri_len != 0);
3014 
3015 		/* will advance the cidx on the corresponding free lists */
3016 		m = iflib_rxd_pkt_get(rxq, &ri);
3017 		avail--;
3018 		budget_left--;
3019 		if (avail == 0 && budget_left)
3020 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3021 
3022 		if (__predict_false(m == NULL))
3023 			continue;
3024 
3025 		/* imm_pkt: -- cxgb */
3026 		if (mh == NULL)
3027 			mh = mt = m;
3028 		else {
3029 			mt->m_nextpkt = m;
3030 			mt = m;
3031 		}
3032 	}
3033 	CURVNET_RESTORE();
3034 	/* make sure that we can refill faster than drain */
3035 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3036 		retval |= iflib_fl_refill_all(ctx, fl);
3037 
3038 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3039 	if (lro_enabled)
3040 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3041 	mt = mf = NULL;
3042 	while (mh != NULL) {
3043 		m = mh;
3044 		mh = mh->m_nextpkt;
3045 		m->m_nextpkt = NULL;
3046 #ifndef __NO_STRICT_ALIGNMENT
3047 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3048 			continue;
3049 #endif
3050 #if defined(INET6) || defined(INET)
3051 		if (lro_enabled) {
3052 			if (!lro_possible) {
3053 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3054 				if (lro_possible && mf != NULL) {
3055 					if_input(ifp, mf);
3056 					DBG_COUNTER_INC(rx_if_input);
3057 					mt = mf = NULL;
3058 				}
3059 			}
3060 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3061 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3062 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3063 					continue;
3064 			}
3065 		}
3066 #endif
3067 		if (lro_possible) {
3068 			if_input(ifp, m);
3069 			DBG_COUNTER_INC(rx_if_input);
3070 			continue;
3071 		}
3072 
3073 		if (mf == NULL)
3074 			mf = m;
3075 		if (mt != NULL)
3076 			mt->m_nextpkt = m;
3077 		mt = m;
3078 	}
3079 	if (mf != NULL) {
3080 		if_input(ifp, mf);
3081 		DBG_COUNTER_INC(rx_if_input);
3082 	}
3083 
3084 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3085 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3086 
3087 	/*
3088 	 * Flush any outstanding LRO work
3089 	 */
3090 #if defined(INET6) || defined(INET)
3091 	tcp_lro_flush_all(&rxq->ifr_lc);
3092 #endif
3093 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3094 		retval |= IFLIB_RXEOF_MORE;
3095 	return (retval);
3096 err:
3097 	STATE_LOCK(ctx);
3098 	ctx->ifc_flags |= IFC_DO_RESET;
3099 	iflib_admin_intr_deferred(ctx);
3100 	STATE_UNLOCK(ctx);
3101 	return (0);
3102 }
3103 
3104 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3105 static inline qidx_t
3106 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3107 {
3108 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3109 	qidx_t minthresh = txq->ift_size / 8;
3110 	if (in_use > 4*minthresh)
3111 		return (notify_count);
3112 	if (in_use > 2*minthresh)
3113 		return (notify_count >> 1);
3114 	if (in_use > minthresh)
3115 		return (notify_count >> 3);
3116 	return (0);
3117 }
3118 
3119 static inline qidx_t
3120 txq_max_rs_deferred(iflib_txq_t txq)
3121 {
3122 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3123 	qidx_t minthresh = txq->ift_size / 8;
3124 	if (txq->ift_in_use > 4*minthresh)
3125 		return (notify_count);
3126 	if (txq->ift_in_use > 2*minthresh)
3127 		return (notify_count >> 1);
3128 	if (txq->ift_in_use > minthresh)
3129 		return (notify_count >> 2);
3130 	return (2);
3131 }
3132 
3133 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3134 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3135 
3136 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3137 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3138 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3139 
3140 /* forward compatibility for cxgb */
3141 #define FIRST_QSET(ctx) 0
3142 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3143 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3144 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3145 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3146 
3147 /* XXX we should be setting this to something other than zero */
3148 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3149 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3150     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3151 
3152 static inline bool
3153 iflib_txd_db_check(iflib_txq_t txq, int ring)
3154 {
3155 	if_ctx_t ctx = txq->ift_ctx;
3156 	qidx_t dbval, max;
3157 
3158 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3159 
3160 	/* force || threshold exceeded || at the edge of the ring */
3161 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3162 
3163 		/*
3164 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3165 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3166 		 * producer index explicitly (INTC).
3167 		 */
3168 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3169 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3170 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3171 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3172 
3173 		/*
3174 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3175 		 */
3176 		txq->ift_db_pending = txq->ift_npending = 0;
3177 		return (true);
3178 	}
3179 	return (false);
3180 }
3181 
3182 #ifdef PKT_DEBUG
3183 static void
3184 print_pkt(if_pkt_info_t pi)
3185 {
3186 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3187 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3188 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3189 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3190 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3191 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3192 }
3193 #endif
3194 
3195 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3196 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3197 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3198 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3199 
3200 /**
3201  * Parses out ethernet header information in the given mbuf.
3202  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3203  *
3204  * This will account for the VLAN header if present.
3205  *
3206  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3207  * types of packets.
3208  */
3209 static int
3210 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3211 {
3212 	struct ether_vlan_header *eh;
3213 	struct mbuf *m;
3214 
3215 	m = *mp;
3216 	if (__predict_false(m->m_len < sizeof(*eh))) {
3217 		(*pullups)++;
3218 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3219 			return (ENOMEM);
3220 	}
3221 	eh = mtod(m, struct ether_vlan_header *);
3222 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3223 		pi->ipi_etype = ntohs(eh->evl_proto);
3224 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3225 	} else {
3226 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3227 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3228 	}
3229 	*mp = m;
3230 
3231 	return (0);
3232 }
3233 
3234 /**
3235  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3236  * Currently this information includes: IP ToS value, IP header version/presence
3237  *
3238  * This is missing some checks and doesn't edit the packet content as it goes,
3239  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3240  */
3241 static int
3242 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3243 {
3244 	struct mbuf *m;
3245 	int err;
3246 
3247 	*pullups = 0;
3248 	m = *mp;
3249 	if (!M_WRITABLE(m)) {
3250 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3251 			return (ENOMEM);
3252 		} else {
3253 			m_freem(*mp);
3254 			DBG_COUNTER_INC(tx_frees);
3255 			*mp = m;
3256 		}
3257 	}
3258 
3259 	/* Fills out pi->ipi_etype */
3260 	err = iflib_parse_ether_header(pi, mp, pullups);
3261 	if (err)
3262 		return (err);
3263 	m = *mp;
3264 
3265 	switch (pi->ipi_etype) {
3266 #ifdef INET
3267 	case ETHERTYPE_IP:
3268 	{
3269 		struct mbuf *n;
3270 		struct ip *ip = NULL;
3271 		int miniplen;
3272 
3273 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3274 		if (__predict_false(m->m_len < miniplen)) {
3275 			/*
3276 			 * Check for common case where the first mbuf only contains
3277 			 * the Ethernet header
3278 			 */
3279 			if (m->m_len == pi->ipi_ehdrlen) {
3280 				n = m->m_next;
3281 				MPASS(n);
3282 				/* If next mbuf contains at least the minimal IP header, then stop */
3283 				if (n->m_len >= sizeof(*ip)) {
3284 					ip = (struct ip *)n->m_data;
3285 				} else {
3286 					(*pullups)++;
3287 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3288 						return (ENOMEM);
3289 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3290 				}
3291 			} else {
3292 				(*pullups)++;
3293 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3294 					return (ENOMEM);
3295 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3296 			}
3297 		} else {
3298 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3299 		}
3300 
3301 		/* Have the IPv4 header w/ no options here */
3302 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3303 		pi->ipi_ipproto = ip->ip_p;
3304 		pi->ipi_ip_tos = ip->ip_tos;
3305 		pi->ipi_flags |= IPI_TX_IPV4;
3306 
3307 		break;
3308 	}
3309 #endif
3310 #ifdef INET6
3311 	case ETHERTYPE_IPV6:
3312 	{
3313 		struct ip6_hdr *ip6;
3314 
3315 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3316 			(*pullups)++;
3317 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3318 				return (ENOMEM);
3319 		}
3320 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3321 
3322 		/* Have the IPv6 fixed header here */
3323 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3324 		pi->ipi_ipproto = ip6->ip6_nxt;
3325 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3326 		pi->ipi_flags |= IPI_TX_IPV6;
3327 
3328 		break;
3329 	}
3330 #endif
3331 	default:
3332 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3333 		pi->ipi_ip_hlen = 0;
3334 		break;
3335 	}
3336 	*mp = m;
3337 
3338 	return (0);
3339 
3340 }
3341 
3342 static int
3343 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3344 {
3345 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3346 	struct mbuf *m;
3347 	int err;
3348 
3349 	m = *mp;
3350 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3351 	    M_WRITABLE(m) == 0) {
3352 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3353 			return (ENOMEM);
3354 		} else {
3355 			m_freem(*mp);
3356 			DBG_COUNTER_INC(tx_frees);
3357 			*mp = m;
3358 		}
3359 	}
3360 
3361 	/* Fills out pi->ipi_etype */
3362 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3363 	if (__predict_false(err))
3364 		return (err);
3365 	m = *mp;
3366 
3367 	switch (pi->ipi_etype) {
3368 #ifdef INET
3369 	case ETHERTYPE_IP:
3370 	{
3371 		struct mbuf *n;
3372 		struct ip *ip = NULL;
3373 		struct tcphdr *th = NULL;
3374 		int minthlen;
3375 
3376 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3377 		if (__predict_false(m->m_len < minthlen)) {
3378 			/*
3379 			 * if this code bloat is causing too much of a hit
3380 			 * move it to a separate function and mark it noinline
3381 			 */
3382 			if (m->m_len == pi->ipi_ehdrlen) {
3383 				n = m->m_next;
3384 				MPASS(n);
3385 				if (n->m_len >= sizeof(*ip))  {
3386 					ip = (struct ip *)n->m_data;
3387 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3388 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3389 				} else {
3390 					txq->ift_pullups++;
3391 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3392 						return (ENOMEM);
3393 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3394 				}
3395 			} else {
3396 				txq->ift_pullups++;
3397 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3398 					return (ENOMEM);
3399 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3400 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3401 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3402 			}
3403 		} else {
3404 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3405 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3406 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3407 		}
3408 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3409 		pi->ipi_ipproto = ip->ip_p;
3410 		pi->ipi_ip_tos = ip->ip_tos;
3411 		pi->ipi_flags |= IPI_TX_IPV4;
3412 
3413 		/* TCP checksum offload may require TCP header length */
3414 		if (IS_TX_OFFLOAD4(pi)) {
3415 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3416 				if (__predict_false(th == NULL)) {
3417 					txq->ift_pullups++;
3418 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3419 						return (ENOMEM);
3420 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3421 				}
3422 				pi->ipi_tcp_hflags = th->th_flags;
3423 				pi->ipi_tcp_hlen = th->th_off << 2;
3424 				pi->ipi_tcp_seq = th->th_seq;
3425 			}
3426 			if (IS_TSO4(pi)) {
3427 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3428 					return (ENXIO);
3429 				/*
3430 				 * TSO always requires hardware checksum offload.
3431 				 */
3432 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3433 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3434 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3435 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3436 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3437 					ip->ip_sum = 0;
3438 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3439 				}
3440 			}
3441 		}
3442 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3443                        ip->ip_sum = 0;
3444 
3445 		break;
3446 	}
3447 #endif
3448 #ifdef INET6
3449 	case ETHERTYPE_IPV6:
3450 	{
3451 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3452 		struct tcphdr *th;
3453 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3454 
3455 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3456 			txq->ift_pullups++;
3457 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3458 				return (ENOMEM);
3459 		}
3460 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3461 
3462 		/* XXX-BZ this will go badly in case of ext hdrs. */
3463 		pi->ipi_ipproto = ip6->ip6_nxt;
3464 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3465 		pi->ipi_flags |= IPI_TX_IPV6;
3466 
3467 		/* TCP checksum offload may require TCP header length */
3468 		if (IS_TX_OFFLOAD6(pi)) {
3469 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3470 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3471 					txq->ift_pullups++;
3472 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3473 						return (ENOMEM);
3474 				}
3475 				pi->ipi_tcp_hflags = th->th_flags;
3476 				pi->ipi_tcp_hlen = th->th_off << 2;
3477 				pi->ipi_tcp_seq = th->th_seq;
3478 			}
3479 			if (IS_TSO6(pi)) {
3480 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3481 					return (ENXIO);
3482 				/*
3483 				 * TSO always requires hardware checksum offload.
3484 				 */
3485 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3486 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3487 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3488 			}
3489 		}
3490 		break;
3491 	}
3492 #endif
3493 	default:
3494 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3495 		pi->ipi_ip_hlen = 0;
3496 		break;
3497 	}
3498 	*mp = m;
3499 
3500 	return (0);
3501 }
3502 
3503 /*
3504  * If dodgy hardware rejects the scatter gather chain we've handed it
3505  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3506  * m_defrag'd mbufs
3507  */
3508 static __noinline struct mbuf *
3509 iflib_remove_mbuf(iflib_txq_t txq)
3510 {
3511 	int ntxd, pidx;
3512 	struct mbuf *m, **ifsd_m;
3513 
3514 	ifsd_m = txq->ift_sds.ifsd_m;
3515 	ntxd = txq->ift_size;
3516 	pidx = txq->ift_pidx & (ntxd - 1);
3517 	ifsd_m = txq->ift_sds.ifsd_m;
3518 	m = ifsd_m[pidx];
3519 	ifsd_m[pidx] = NULL;
3520 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3521 	if (txq->ift_sds.ifsd_tso_map != NULL)
3522 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3523 		    txq->ift_sds.ifsd_tso_map[pidx]);
3524 #if MEMORY_LOGGING
3525 	txq->ift_dequeued++;
3526 #endif
3527 	return (m);
3528 }
3529 
3530 static inline caddr_t
3531 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3532 {
3533 	qidx_t size;
3534 	int ntxd;
3535 	caddr_t start, end, cur, next;
3536 
3537 	ntxd = txq->ift_size;
3538 	size = txq->ift_txd_size[qid];
3539 	start = txq->ift_ifdi[qid].idi_vaddr;
3540 
3541 	if (__predict_false(size == 0))
3542 		return (start);
3543 	cur = start + size*cidx;
3544 	end = start + size*ntxd;
3545 	next = CACHE_PTR_NEXT(cur);
3546 	return (next < end ? next : start);
3547 }
3548 
3549 /*
3550  * Pad an mbuf to ensure a minimum ethernet frame size.
3551  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3552  */
3553 static __noinline int
3554 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3555 {
3556 	/*
3557 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3558 	 * and ARP message is the smallest common payload I can think of
3559 	 */
3560 	static char pad[18];	/* just zeros */
3561 	int n;
3562 	struct mbuf *new_head;
3563 
3564 	if (!M_WRITABLE(*m_head)) {
3565 		new_head = m_dup(*m_head, M_NOWAIT);
3566 		if (new_head == NULL) {
3567 			m_freem(*m_head);
3568 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3569 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3570 			DBG_COUNTER_INC(tx_frees);
3571 			return ENOMEM;
3572 		}
3573 		m_freem(*m_head);
3574 		*m_head = new_head;
3575 	}
3576 
3577 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3578 	     n > 0; n -= sizeof(pad))
3579 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3580 			break;
3581 
3582 	if (n > 0) {
3583 		m_freem(*m_head);
3584 		device_printf(dev, "cannot pad short frame\n");
3585 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3586 		DBG_COUNTER_INC(tx_frees);
3587 		return (ENOBUFS);
3588 	}
3589 
3590 	return 0;
3591 }
3592 
3593 static int
3594 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3595 {
3596 	if_ctx_t		ctx;
3597 	if_shared_ctx_t		sctx;
3598 	if_softc_ctx_t		scctx;
3599 	bus_dma_tag_t		buf_tag;
3600 	bus_dma_segment_t	*segs;
3601 	struct mbuf		*m_head, **ifsd_m;
3602 	void			*next_txd;
3603 	bus_dmamap_t		map;
3604 	struct if_pkt_info	pi;
3605 	int remap = 0;
3606 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3607 
3608 	ctx = txq->ift_ctx;
3609 	sctx = ctx->ifc_sctx;
3610 	scctx = &ctx->ifc_softc_ctx;
3611 	segs = txq->ift_segs;
3612 	ntxd = txq->ift_size;
3613 	m_head = *m_headp;
3614 	map = NULL;
3615 
3616 	/*
3617 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3618 	 */
3619 	cidx = txq->ift_cidx;
3620 	pidx = txq->ift_pidx;
3621 	if (ctx->ifc_flags & IFC_PREFETCH) {
3622 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3623 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3624 			next_txd = calc_next_txd(txq, cidx, 0);
3625 			prefetch(next_txd);
3626 		}
3627 
3628 		/* prefetch the next cache line of mbuf pointers and flags */
3629 		prefetch(&txq->ift_sds.ifsd_m[next]);
3630 		prefetch(&txq->ift_sds.ifsd_map[next]);
3631 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3632 	}
3633 	map = txq->ift_sds.ifsd_map[pidx];
3634 	ifsd_m = txq->ift_sds.ifsd_m;
3635 
3636 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3637 		buf_tag = txq->ift_tso_buf_tag;
3638 		max_segs = scctx->isc_tx_tso_segments_max;
3639 		map = txq->ift_sds.ifsd_tso_map[pidx];
3640 		MPASS(buf_tag != NULL);
3641 		MPASS(max_segs > 0);
3642 	} else {
3643 		buf_tag = txq->ift_buf_tag;
3644 		max_segs = scctx->isc_tx_nsegments;
3645 		map = txq->ift_sds.ifsd_map[pidx];
3646 	}
3647 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3648 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3649 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3650 		if (err) {
3651 			DBG_COUNTER_INC(encap_txd_encap_fail);
3652 			return err;
3653 		}
3654 	}
3655 	m_head = *m_headp;
3656 
3657 	pkt_info_zero(&pi);
3658 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3659 	pi.ipi_pidx = pidx;
3660 	pi.ipi_qsidx = txq->ift_id;
3661 	pi.ipi_len = m_head->m_pkthdr.len;
3662 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3663 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3664 
3665 	/* deliberate bitwise OR to make one condition */
3666 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3667 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3668 			DBG_COUNTER_INC(encap_txd_encap_fail);
3669 			return (err);
3670 		}
3671 		m_head = *m_headp;
3672 	}
3673 
3674 retry:
3675 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3676 	    BUS_DMA_NOWAIT);
3677 defrag:
3678 	if (__predict_false(err)) {
3679 		switch (err) {
3680 		case EFBIG:
3681 			/* try collapse once and defrag once */
3682 			if (remap == 0) {
3683 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3684 				/* try defrag if collapsing fails */
3685 				if (m_head == NULL)
3686 					remap++;
3687 			}
3688 			if (remap == 1) {
3689 				txq->ift_mbuf_defrag++;
3690 				m_head = m_defrag(*m_headp, M_NOWAIT);
3691 			}
3692 			/*
3693 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3694 			 * failed to map an mbuf that was run through m_defrag
3695 			 */
3696 			MPASS(remap <= 1);
3697 			if (__predict_false(m_head == NULL || remap > 1))
3698 				goto defrag_failed;
3699 			remap++;
3700 			*m_headp = m_head;
3701 			goto retry;
3702 			break;
3703 		case ENOMEM:
3704 			txq->ift_no_tx_dma_setup++;
3705 			break;
3706 		default:
3707 			txq->ift_no_tx_dma_setup++;
3708 			m_freem(*m_headp);
3709 			DBG_COUNTER_INC(tx_frees);
3710 			*m_headp = NULL;
3711 			break;
3712 		}
3713 		txq->ift_map_failed++;
3714 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3715 		DBG_COUNTER_INC(encap_txd_encap_fail);
3716 		return (err);
3717 	}
3718 	ifsd_m[pidx] = m_head;
3719 	/*
3720 	 * XXX assumes a 1 to 1 relationship between segments and
3721 	 *        descriptors - this does not hold true on all drivers, e.g.
3722 	 *        cxgb
3723 	 */
3724 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3725 		txq->ift_no_desc_avail++;
3726 		bus_dmamap_unload(buf_tag, map);
3727 		DBG_COUNTER_INC(encap_txq_avail_fail);
3728 		DBG_COUNTER_INC(encap_txd_encap_fail);
3729 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3730 			GROUPTASK_ENQUEUE(&txq->ift_task);
3731 		return (ENOBUFS);
3732 	}
3733 	/*
3734 	 * On Intel cards we can greatly reduce the number of TX interrupts
3735 	 * we see by only setting report status on every Nth descriptor.
3736 	 * However, this also means that the driver will need to keep track
3737 	 * of the descriptors that RS was set on to check them for the DD bit.
3738 	 */
3739 	txq->ift_rs_pending += nsegs + 1;
3740 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3741 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3742 		pi.ipi_flags |= IPI_TX_INTR;
3743 		txq->ift_rs_pending = 0;
3744 	}
3745 
3746 	pi.ipi_segs = segs;
3747 	pi.ipi_nsegs = nsegs;
3748 
3749 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3750 #ifdef PKT_DEBUG
3751 	print_pkt(&pi);
3752 #endif
3753 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3754 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3755 		DBG_COUNTER_INC(tx_encap);
3756 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3757 
3758 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3759 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3760 			ndesc += txq->ift_size;
3761 			txq->ift_gen = 1;
3762 		}
3763 		/*
3764 		 * drivers can need as many as
3765 		 * two sentinels
3766 		 */
3767 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3768 		MPASS(pi.ipi_new_pidx != pidx);
3769 		MPASS(ndesc > 0);
3770 		txq->ift_in_use += ndesc;
3771 		txq->ift_db_pending += ndesc;
3772 
3773 		/*
3774 		 * We update the last software descriptor again here because there may
3775 		 * be a sentinel and/or there may be more mbufs than segments
3776 		 */
3777 		txq->ift_pidx = pi.ipi_new_pidx;
3778 		txq->ift_npending += pi.ipi_ndescs;
3779 	} else {
3780 		*m_headp = m_head = iflib_remove_mbuf(txq);
3781 		if (err == EFBIG) {
3782 			txq->ift_txd_encap_efbig++;
3783 			if (remap < 2) {
3784 				remap = 1;
3785 				goto defrag;
3786 			}
3787 		}
3788 		goto defrag_failed;
3789 	}
3790 	/*
3791 	 * err can't possibly be non-zero here, so we don't neet to test it
3792 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3793 	 */
3794 	return (err);
3795 
3796 defrag_failed:
3797 	txq->ift_mbuf_defrag_failed++;
3798 	txq->ift_map_failed++;
3799 	m_freem(*m_headp);
3800 	DBG_COUNTER_INC(tx_frees);
3801 	*m_headp = NULL;
3802 	DBG_COUNTER_INC(encap_txd_encap_fail);
3803 	return (ENOMEM);
3804 }
3805 
3806 static void
3807 iflib_tx_desc_free(iflib_txq_t txq, int n)
3808 {
3809 	uint32_t qsize, cidx, mask, gen;
3810 	struct mbuf *m, **ifsd_m;
3811 	bool do_prefetch;
3812 
3813 	cidx = txq->ift_cidx;
3814 	gen = txq->ift_gen;
3815 	qsize = txq->ift_size;
3816 	mask = qsize-1;
3817 	ifsd_m = txq->ift_sds.ifsd_m;
3818 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3819 
3820 	while (n-- > 0) {
3821 		if (do_prefetch) {
3822 			prefetch(ifsd_m[(cidx + 3) & mask]);
3823 			prefetch(ifsd_m[(cidx + 4) & mask]);
3824 		}
3825 		if ((m = ifsd_m[cidx]) != NULL) {
3826 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3827 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3828 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3829 				    txq->ift_sds.ifsd_tso_map[cidx],
3830 				    BUS_DMASYNC_POSTWRITE);
3831 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3832 				    txq->ift_sds.ifsd_tso_map[cidx]);
3833 			} else {
3834 				bus_dmamap_sync(txq->ift_buf_tag,
3835 				    txq->ift_sds.ifsd_map[cidx],
3836 				    BUS_DMASYNC_POSTWRITE);
3837 				bus_dmamap_unload(txq->ift_buf_tag,
3838 				    txq->ift_sds.ifsd_map[cidx]);
3839 			}
3840 			/* XXX we don't support any drivers that batch packets yet */
3841 			MPASS(m->m_nextpkt == NULL);
3842 			m_freem(m);
3843 			ifsd_m[cidx] = NULL;
3844 #if MEMORY_LOGGING
3845 			txq->ift_dequeued++;
3846 #endif
3847 			DBG_COUNTER_INC(tx_frees);
3848 		}
3849 		if (__predict_false(++cidx == qsize)) {
3850 			cidx = 0;
3851 			gen = 0;
3852 		}
3853 	}
3854 	txq->ift_cidx = cidx;
3855 	txq->ift_gen = gen;
3856 }
3857 
3858 static __inline int
3859 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3860 {
3861 	int reclaim;
3862 	if_ctx_t ctx = txq->ift_ctx;
3863 
3864 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3865 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3866 
3867 	/*
3868 	 * Need a rate-limiting check so that this isn't called every time
3869 	 */
3870 	iflib_tx_credits_update(ctx, txq);
3871 	reclaim = DESC_RECLAIMABLE(txq);
3872 
3873 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3874 #ifdef INVARIANTS
3875 		if (iflib_verbose_debug) {
3876 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3877 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3878 			       reclaim, thresh);
3879 		}
3880 #endif
3881 		return (0);
3882 	}
3883 	iflib_tx_desc_free(txq, reclaim);
3884 	txq->ift_cleaned += reclaim;
3885 	txq->ift_in_use -= reclaim;
3886 
3887 	return (reclaim);
3888 }
3889 
3890 static struct mbuf **
3891 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3892 {
3893 	int next, size;
3894 	struct mbuf **items;
3895 
3896 	size = r->size;
3897 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3898 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3899 
3900 	prefetch(items[(cidx + offset) & (size-1)]);
3901 	if (remaining > 1) {
3902 		prefetch2cachelines(&items[next]);
3903 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3904 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3905 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3906 	}
3907 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3908 }
3909 
3910 static void
3911 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3912 {
3913 
3914 	ifmp_ring_check_drainage(txq->ift_br, budget);
3915 }
3916 
3917 static uint32_t
3918 iflib_txq_can_drain(struct ifmp_ring *r)
3919 {
3920 	iflib_txq_t txq = r->cookie;
3921 	if_ctx_t ctx = txq->ift_ctx;
3922 
3923 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3924 		return (1);
3925 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3926 	    BUS_DMASYNC_POSTREAD);
3927 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3928 	    false));
3929 }
3930 
3931 static uint32_t
3932 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3933 {
3934 	iflib_txq_t txq = r->cookie;
3935 	if_ctx_t ctx = txq->ift_ctx;
3936 	if_t ifp = ctx->ifc_ifp;
3937 	struct mbuf *m, **mp;
3938 	int avail, bytes_sent, skipped, count, err, i;
3939 	int mcast_sent, pkt_sent, reclaimed;
3940 	bool do_prefetch, rang, ring;
3941 
3942 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3943 			    !LINK_ACTIVE(ctx))) {
3944 		DBG_COUNTER_INC(txq_drain_notready);
3945 		return (0);
3946 	}
3947 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3948 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3949 	avail = IDXDIFF(pidx, cidx, r->size);
3950 
3951 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3952 		/*
3953 		 * The driver is unloading so we need to free all pending packets.
3954 		 */
3955 		DBG_COUNTER_INC(txq_drain_flushing);
3956 		for (i = 0; i < avail; i++) {
3957 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3958 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3959 			r->items[(cidx + i) & (r->size-1)] = NULL;
3960 		}
3961 		return (avail);
3962 	}
3963 
3964 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3965 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3966 		CALLOUT_LOCK(txq);
3967 		callout_stop(&txq->ift_timer);
3968 		CALLOUT_UNLOCK(txq);
3969 		DBG_COUNTER_INC(txq_drain_oactive);
3970 		return (0);
3971 	}
3972 
3973 	/*
3974 	 * If we've reclaimed any packets this queue cannot be hung.
3975 	 */
3976 	if (reclaimed)
3977 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3978 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3979 	count = MIN(avail, TX_BATCH_SIZE);
3980 #ifdef INVARIANTS
3981 	if (iflib_verbose_debug)
3982 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3983 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3984 #endif
3985 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3986 	err = 0;
3987 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3988 		int rem = do_prefetch ? count - i : 0;
3989 
3990 		mp = _ring_peek_one(r, cidx, i, rem);
3991 		MPASS(mp != NULL && *mp != NULL);
3992 
3993 		/*
3994 		 * Completion interrupts will use the address of the txq
3995 		 * as a sentinel to enqueue _something_ in order to acquire
3996 		 * the lock on the mp_ring (there's no direct lock call).
3997 		 * We obviously whave to check for these sentinel cases
3998 		 * and skip them.
3999 		 */
4000 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4001 			skipped++;
4002 			continue;
4003 		}
4004 		err = iflib_encap(txq, mp);
4005 		if (__predict_false(err)) {
4006 			/* no room - bail out */
4007 			if (err == ENOBUFS)
4008 				break;
4009 			skipped++;
4010 			/* we can't send this packet - skip it */
4011 			continue;
4012 		}
4013 		pkt_sent++;
4014 		m = *mp;
4015 		DBG_COUNTER_INC(tx_sent);
4016 		bytes_sent += m->m_pkthdr.len;
4017 		mcast_sent += !!(m->m_flags & M_MCAST);
4018 
4019 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4020 			break;
4021 		ETHER_BPF_MTAP(ifp, m);
4022 		rang = iflib_txd_db_check(txq, false);
4023 	}
4024 
4025 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4026 	ring = rang ? false  : (iflib_min_tx_latency | err);
4027 	iflib_txd_db_check(txq, ring);
4028 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4029 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4030 	if (mcast_sent)
4031 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4032 #ifdef INVARIANTS
4033 	if (iflib_verbose_debug)
4034 		printf("consumed=%d\n", skipped + pkt_sent);
4035 #endif
4036 	return (skipped + pkt_sent);
4037 }
4038 
4039 static uint32_t
4040 iflib_txq_drain_always(struct ifmp_ring *r)
4041 {
4042 	return (1);
4043 }
4044 
4045 static uint32_t
4046 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4047 {
4048 	int i, avail;
4049 	struct mbuf **mp;
4050 	iflib_txq_t txq;
4051 
4052 	txq = r->cookie;
4053 
4054 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4055 	CALLOUT_LOCK(txq);
4056 	callout_stop(&txq->ift_timer);
4057 	CALLOUT_UNLOCK(txq);
4058 
4059 	avail = IDXDIFF(pidx, cidx, r->size);
4060 	for (i = 0; i < avail; i++) {
4061 		mp = _ring_peek_one(r, cidx, i, avail - i);
4062 		if (__predict_false(*mp == (struct mbuf *)txq))
4063 			continue;
4064 		m_freem(*mp);
4065 		DBG_COUNTER_INC(tx_frees);
4066 	}
4067 	MPASS(ifmp_ring_is_stalled(r) == 0);
4068 	return (avail);
4069 }
4070 
4071 static void
4072 iflib_ifmp_purge(iflib_txq_t txq)
4073 {
4074 	struct ifmp_ring *r;
4075 
4076 	r = txq->ift_br;
4077 	r->drain = iflib_txq_drain_free;
4078 	r->can_drain = iflib_txq_drain_always;
4079 
4080 	ifmp_ring_check_drainage(r, r->size);
4081 
4082 	r->drain = iflib_txq_drain;
4083 	r->can_drain = iflib_txq_can_drain;
4084 }
4085 
4086 static void
4087 _task_fn_tx(void *context)
4088 {
4089 	iflib_txq_t txq = context;
4090 	if_ctx_t ctx = txq->ift_ctx;
4091 	if_t ifp = ctx->ifc_ifp;
4092 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4093 
4094 #ifdef IFLIB_DIAGNOSTICS
4095 	txq->ift_cpu_exec_count[curcpu]++;
4096 #endif
4097 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4098 		return;
4099 #ifdef DEV_NETMAP
4100 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4101 	    netmap_tx_irq(ifp, txq->ift_id))
4102 		goto skip_ifmp;
4103 #endif
4104 #ifdef ALTQ
4105 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) /* XXX - DRVAPI */
4106 		iflib_altq_if_start(ifp);
4107 #endif
4108 	if (txq->ift_db_pending)
4109 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4110 	else if (!abdicate)
4111 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4112 	/*
4113 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4114 	 */
4115 	if (abdicate)
4116 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4117 #ifdef DEV_NETMAP
4118 skip_ifmp:
4119 #endif
4120 	if (ctx->ifc_flags & IFC_LEGACY)
4121 		IFDI_INTR_ENABLE(ctx);
4122 	else
4123 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4124 }
4125 
4126 static void
4127 _task_fn_rx(void *context)
4128 {
4129 	iflib_rxq_t rxq = context;
4130 	if_ctx_t ctx = rxq->ifr_ctx;
4131 	uint8_t more;
4132 	uint16_t budget;
4133 #ifdef DEV_NETMAP
4134 	u_int work = 0;
4135 	int nmirq;
4136 #endif
4137 
4138 #ifdef IFLIB_DIAGNOSTICS
4139 	rxq->ifr_cpu_exec_count[curcpu]++;
4140 #endif
4141 	DBG_COUNTER_INC(task_fn_rxs);
4142 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4143 		return;
4144 #ifdef DEV_NETMAP
4145 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4146 	if (nmirq != NM_IRQ_PASS) {
4147 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4148 		goto skip_rxeof;
4149 	}
4150 #endif
4151 	budget = ctx->ifc_sysctl_rx_budget;
4152 	if (budget == 0)
4153 		budget = 16;	/* XXX */
4154 	more = iflib_rxeof(rxq, budget);
4155 #ifdef DEV_NETMAP
4156 skip_rxeof:
4157 #endif
4158 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4159 		if (ctx->ifc_flags & IFC_LEGACY)
4160 			IFDI_INTR_ENABLE(ctx);
4161 		else
4162 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4163 		DBG_COUNTER_INC(rx_intr_enables);
4164 	}
4165 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4166 		return;
4167 
4168 	if (more & IFLIB_RXEOF_MORE)
4169 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4170 	else if (more & IFLIB_RXEOF_EMPTY)
4171 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4172 }
4173 
4174 static void
4175 _task_fn_admin(void *context)
4176 {
4177 	if_ctx_t ctx = context;
4178 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4179 	iflib_txq_t txq;
4180 	int i;
4181 	bool oactive, running, do_reset, do_watchdog, in_detach;
4182 
4183 	STATE_LOCK(ctx);
4184 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4185 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4186 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4187 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4188 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4189 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4190 	STATE_UNLOCK(ctx);
4191 
4192 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4193 		return;
4194 	if (in_detach)
4195 		return;
4196 
4197 	CTX_LOCK(ctx);
4198 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4199 		CALLOUT_LOCK(txq);
4200 		callout_stop(&txq->ift_timer);
4201 		CALLOUT_UNLOCK(txq);
4202 	}
4203 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4204 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4205 	if (do_watchdog) {
4206 		ctx->ifc_watchdog_events++;
4207 		IFDI_WATCHDOG_RESET(ctx);
4208 	}
4209 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4210 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4211 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4212 		    txq->ift_timer.c_cpu);
4213 	}
4214 	IFDI_LINK_INTR_ENABLE(ctx);
4215 	if (do_reset)
4216 		iflib_if_init_locked(ctx);
4217 	CTX_UNLOCK(ctx);
4218 
4219 	if (LINK_ACTIVE(ctx) == 0)
4220 		return;
4221 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4222 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4223 }
4224 
4225 static void
4226 _task_fn_iov(void *context)
4227 {
4228 	if_ctx_t ctx = context;
4229 
4230 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4231 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4232 		return;
4233 
4234 	CTX_LOCK(ctx);
4235 	IFDI_VFLR_HANDLE(ctx);
4236 	CTX_UNLOCK(ctx);
4237 }
4238 
4239 static int
4240 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4241 {
4242 	int err;
4243 	if_int_delay_info_t info;
4244 	if_ctx_t ctx;
4245 
4246 	info = (if_int_delay_info_t)arg1;
4247 	ctx = info->iidi_ctx;
4248 	info->iidi_req = req;
4249 	info->iidi_oidp = oidp;
4250 	CTX_LOCK(ctx);
4251 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4252 	CTX_UNLOCK(ctx);
4253 	return (err);
4254 }
4255 
4256 /*********************************************************************
4257  *
4258  *  IFNET FUNCTIONS
4259  *
4260  **********************************************************************/
4261 
4262 static void
4263 iflib_if_init_locked(if_ctx_t ctx)
4264 {
4265 	iflib_stop(ctx);
4266 	iflib_init_locked(ctx);
4267 }
4268 
4269 static void
4270 iflib_if_init(void *arg)
4271 {
4272 	if_ctx_t ctx = arg;
4273 
4274 	CTX_LOCK(ctx);
4275 	iflib_if_init_locked(ctx);
4276 	CTX_UNLOCK(ctx);
4277 }
4278 
4279 static int
4280 iflib_if_transmit(if_t ifp, struct mbuf *m)
4281 {
4282 	if_ctx_t ctx = if_getsoftc(ifp);
4283 	iflib_txq_t txq;
4284 	int err, qidx;
4285 	int abdicate;
4286 
4287 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4288 		DBG_COUNTER_INC(tx_frees);
4289 		m_freem(m);
4290 		return (ENETDOWN);
4291 	}
4292 
4293 	MPASS(m->m_nextpkt == NULL);
4294 	/* ALTQ-enabled interfaces always use queue 0. */
4295 	qidx = 0;
4296 	/* Use driver-supplied queue selection method if it exists */
4297 	if (ctx->isc_txq_select_v2) {
4298 		struct if_pkt_info pi;
4299 		uint64_t early_pullups = 0;
4300 		pkt_info_zero(&pi);
4301 
4302 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4303 		if (__predict_false(err != 0)) {
4304 			/* Assign pullups for bad pkts to default queue */
4305 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4306 			DBG_COUNTER_INC(encap_txd_encap_fail);
4307 			return (err);
4308 		}
4309 		/* Let driver make queueing decision */
4310 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4311 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4312 	}
4313 	/* Backwards compatibility w/ simpler queue select */
4314 	else if (ctx->isc_txq_select)
4315 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4316 	/* If not, use iflib's standard method */
4317 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4318 		qidx = QIDX(ctx, m);
4319 
4320 	/* Set TX queue */
4321 	txq = &ctx->ifc_txqs[qidx];
4322 
4323 #ifdef DRIVER_BACKPRESSURE
4324 	if (txq->ift_closed) {
4325 		while (m != NULL) {
4326 			next = m->m_nextpkt;
4327 			m->m_nextpkt = NULL;
4328 			m_freem(m);
4329 			DBG_COUNTER_INC(tx_frees);
4330 			m = next;
4331 		}
4332 		return (ENOBUFS);
4333 	}
4334 #endif
4335 #ifdef notyet
4336 	qidx = count = 0;
4337 	mp = marr;
4338 	next = m;
4339 	do {
4340 		count++;
4341 		next = next->m_nextpkt;
4342 	} while (next != NULL);
4343 
4344 	if (count > nitems(marr))
4345 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4346 			/* XXX check nextpkt */
4347 			m_freem(m);
4348 			/* XXX simplify for now */
4349 			DBG_COUNTER_INC(tx_frees);
4350 			return (ENOBUFS);
4351 		}
4352 	for (next = m, i = 0; next != NULL; i++) {
4353 		mp[i] = next;
4354 		next = next->m_nextpkt;
4355 		mp[i]->m_nextpkt = NULL;
4356 	}
4357 #endif
4358 	DBG_COUNTER_INC(tx_seen);
4359 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4360 
4361 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4362 
4363 	if (abdicate)
4364 		GROUPTASK_ENQUEUE(&txq->ift_task);
4365  	if (err) {
4366 		if (!abdicate)
4367 			GROUPTASK_ENQUEUE(&txq->ift_task);
4368 		/* support forthcoming later */
4369 #ifdef DRIVER_BACKPRESSURE
4370 		txq->ift_closed = TRUE;
4371 #endif
4372 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4373 		m_freem(m);
4374 		DBG_COUNTER_INC(tx_frees);
4375 	}
4376 
4377 	return (err);
4378 }
4379 
4380 #ifdef ALTQ
4381 /*
4382  * The overall approach to integrating iflib with ALTQ is to continue to use
4383  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4384  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4385  * is redundant/unnecessary, but doing so minimizes the amount of
4386  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4387  * redundantly queueing to an intermediate mp_ring is swamped by the
4388  * performance limitations inherent in using ALTQ.
4389  *
4390  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4391  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4392  * given interface.  If ALTQ is enabled for an interface, then all
4393  * transmitted packets for that interface will be submitted to the ALTQ
4394  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4395  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4396  * update stats that the iflib machinery handles, and which is sensitve to
4397  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4398  * will be installed as the start routine for use by ALTQ facilities that
4399  * need to trigger queue drains on a scheduled basis.
4400  *
4401  */
4402 static void
4403 iflib_altq_if_start(if_t ifp)
4404 {
4405 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4406 	struct mbuf *m;
4407 
4408 	IFQ_LOCK(ifq);
4409 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4410 	while (m != NULL) {
4411 		iflib_if_transmit(ifp, m);
4412 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4413 	}
4414 	IFQ_UNLOCK(ifq);
4415 }
4416 
4417 static int
4418 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4419 {
4420 	int err;
4421 
4422 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) { /* XXX - DRVAPI */
4423 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4424 		if (err == 0)
4425 			iflib_altq_if_start(ifp);
4426 	} else
4427 		err = iflib_if_transmit(ifp, m);
4428 
4429 	return (err);
4430 }
4431 #endif /* ALTQ */
4432 
4433 static void
4434 iflib_if_qflush(if_t ifp)
4435 {
4436 	if_ctx_t ctx = if_getsoftc(ifp);
4437 	iflib_txq_t txq = ctx->ifc_txqs;
4438 	int i;
4439 
4440 	STATE_LOCK(ctx);
4441 	ctx->ifc_flags |= IFC_QFLUSH;
4442 	STATE_UNLOCK(ctx);
4443 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4444 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4445 			iflib_txq_check_drain(txq, 0);
4446 	STATE_LOCK(ctx);
4447 	ctx->ifc_flags &= ~IFC_QFLUSH;
4448 	STATE_UNLOCK(ctx);
4449 
4450 	/*
4451 	 * When ALTQ is enabled, this will also take care of purging the
4452 	 * ALTQ queue(s).
4453 	 */
4454 	if_qflush(ifp);
4455 }
4456 
4457 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4458 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4459 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4460 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4461 
4462 static int
4463 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4464 {
4465 	if_ctx_t ctx = if_getsoftc(ifp);
4466 	struct ifreq	*ifr = (struct ifreq *)data;
4467 #if defined(INET) || defined(INET6)
4468 	struct ifaddr	*ifa = (struct ifaddr *)data;
4469 #endif
4470 	bool		avoid_reset = false;
4471 	int		err = 0, reinit = 0, bits;
4472 
4473 	switch (command) {
4474 	case SIOCSIFADDR:
4475 #ifdef INET
4476 		if (ifa->ifa_addr->sa_family == AF_INET)
4477 			avoid_reset = true;
4478 #endif
4479 #ifdef INET6
4480 		if (ifa->ifa_addr->sa_family == AF_INET6)
4481 			avoid_reset = true;
4482 #endif
4483 		/*
4484 		** Calling init results in link renegotiation,
4485 		** so we avoid doing it when possible.
4486 		*/
4487 		if (avoid_reset) {
4488 			if_setflagbits(ifp, IFF_UP,0);
4489 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4490 				reinit = 1;
4491 #ifdef INET
4492 			if (!(if_getflags(ifp) & IFF_NOARP))
4493 				arp_ifinit(ifp, ifa);
4494 #endif
4495 		} else
4496 			err = ether_ioctl(ifp, command, data);
4497 		break;
4498 	case SIOCSIFMTU:
4499 		CTX_LOCK(ctx);
4500 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4501 			CTX_UNLOCK(ctx);
4502 			break;
4503 		}
4504 		bits = if_getdrvflags(ifp);
4505 		/* stop the driver and free any clusters before proceeding */
4506 		iflib_stop(ctx);
4507 
4508 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4509 			STATE_LOCK(ctx);
4510 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4511 				ctx->ifc_flags |= IFC_MULTISEG;
4512 			else
4513 				ctx->ifc_flags &= ~IFC_MULTISEG;
4514 			STATE_UNLOCK(ctx);
4515 			err = if_setmtu(ifp, ifr->ifr_mtu);
4516 		}
4517 		iflib_init_locked(ctx);
4518 		STATE_LOCK(ctx);
4519 		if_setdrvflags(ifp, bits);
4520 		STATE_UNLOCK(ctx);
4521 		CTX_UNLOCK(ctx);
4522 		break;
4523 	case SIOCSIFFLAGS:
4524 		CTX_LOCK(ctx);
4525 		if (if_getflags(ifp) & IFF_UP) {
4526 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4527 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4528 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4529 					CTX_UNLOCK(ctx);
4530 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4531 					CTX_LOCK(ctx);
4532 				}
4533 			} else
4534 				reinit = 1;
4535 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4536 			iflib_stop(ctx);
4537 		}
4538 		ctx->ifc_if_flags = if_getflags(ifp);
4539 		CTX_UNLOCK(ctx);
4540 		break;
4541 	case SIOCADDMULTI:
4542 	case SIOCDELMULTI:
4543 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4544 			CTX_LOCK(ctx);
4545 			IFDI_INTR_DISABLE(ctx);
4546 			IFDI_MULTI_SET(ctx);
4547 			IFDI_INTR_ENABLE(ctx);
4548 			CTX_UNLOCK(ctx);
4549 		}
4550 		break;
4551 	case SIOCSIFMEDIA:
4552 		CTX_LOCK(ctx);
4553 		IFDI_MEDIA_SET(ctx);
4554 		CTX_UNLOCK(ctx);
4555 		/* FALLTHROUGH */
4556 	case SIOCGIFMEDIA:
4557 	case SIOCGIFXMEDIA:
4558 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4559 		break;
4560 	case SIOCGI2C:
4561 	{
4562 		struct ifi2creq i2c;
4563 
4564 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4565 		if (err != 0)
4566 			break;
4567 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4568 			err = EINVAL;
4569 			break;
4570 		}
4571 		if (i2c.len > sizeof(i2c.data)) {
4572 			err = EINVAL;
4573 			break;
4574 		}
4575 
4576 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4577 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4578 			    sizeof(i2c));
4579 		break;
4580 	}
4581 	case SIOCSIFCAP:
4582 	{
4583 		int mask, setmask, oldmask;
4584 
4585 		oldmask = if_getcapenable(ifp);
4586 		mask = ifr->ifr_reqcap ^ oldmask;
4587 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4588 		setmask = 0;
4589 #ifdef TCP_OFFLOAD
4590 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4591 #endif
4592 		setmask |= (mask & IFCAP_FLAGS);
4593 		setmask |= (mask & IFCAP_WOL);
4594 
4595 		/*
4596 		 * If any RX csum has changed, change all the ones that
4597 		 * are supported by the driver.
4598 		 */
4599 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4600 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4601 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4602 		}
4603 
4604 		/*
4605 		 * want to ensure that traffic has stopped before we change any of the flags
4606 		 */
4607 		if (setmask) {
4608 			CTX_LOCK(ctx);
4609 			bits = if_getdrvflags(ifp);
4610 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4611 				iflib_stop(ctx);
4612 			STATE_LOCK(ctx);
4613 			if_togglecapenable(ifp, setmask);
4614 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4615 			STATE_UNLOCK(ctx);
4616 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4617 				iflib_init_locked(ctx);
4618 			STATE_LOCK(ctx);
4619 			if_setdrvflags(ifp, bits);
4620 			STATE_UNLOCK(ctx);
4621 			CTX_UNLOCK(ctx);
4622 		}
4623 		if_vlancap(ifp);
4624 		break;
4625 	}
4626 	case SIOCGPRIVATE_0:
4627 	case SIOCSDRVSPEC:
4628 	case SIOCGDRVSPEC:
4629 		CTX_LOCK(ctx);
4630 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4631 		CTX_UNLOCK(ctx);
4632 		break;
4633 	default:
4634 		err = ether_ioctl(ifp, command, data);
4635 		break;
4636 	}
4637 	if (reinit)
4638 		iflib_if_init(ctx);
4639 	return (err);
4640 }
4641 
4642 static uint64_t
4643 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4644 {
4645 	if_ctx_t ctx = if_getsoftc(ifp);
4646 
4647 	return (IFDI_GET_COUNTER(ctx, cnt));
4648 }
4649 
4650 /*********************************************************************
4651  *
4652  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4653  *
4654  **********************************************************************/
4655 
4656 static void
4657 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4658 {
4659 	if_ctx_t ctx = if_getsoftc(ifp);
4660 
4661 	if ((void *)ctx != arg)
4662 		return;
4663 
4664 	if ((vtag == 0) || (vtag > 4095))
4665 		return;
4666 
4667 	if (iflib_in_detach(ctx))
4668 		return;
4669 
4670 	CTX_LOCK(ctx);
4671 	/* Driver may need all untagged packets to be flushed */
4672 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4673 		iflib_stop(ctx);
4674 	IFDI_VLAN_REGISTER(ctx, vtag);
4675 	/* Re-init to load the changes, if required */
4676 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4677 		iflib_init_locked(ctx);
4678 	CTX_UNLOCK(ctx);
4679 }
4680 
4681 static void
4682 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4683 {
4684 	if_ctx_t ctx = if_getsoftc(ifp);
4685 
4686 	if ((void *)ctx != arg)
4687 		return;
4688 
4689 	if ((vtag == 0) || (vtag > 4095))
4690 		return;
4691 
4692 	CTX_LOCK(ctx);
4693 	/* Driver may need all tagged packets to be flushed */
4694 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4695 		iflib_stop(ctx);
4696 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4697 	/* Re-init to load the changes, if required */
4698 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4699 		iflib_init_locked(ctx);
4700 	CTX_UNLOCK(ctx);
4701 }
4702 
4703 static void
4704 iflib_led_func(void *arg, int onoff)
4705 {
4706 	if_ctx_t ctx = arg;
4707 
4708 	CTX_LOCK(ctx);
4709 	IFDI_LED_FUNC(ctx, onoff);
4710 	CTX_UNLOCK(ctx);
4711 }
4712 
4713 /*********************************************************************
4714  *
4715  *  BUS FUNCTION DEFINITIONS
4716  *
4717  **********************************************************************/
4718 
4719 int
4720 iflib_device_probe(device_t dev)
4721 {
4722 	const pci_vendor_info_t *ent;
4723 	if_shared_ctx_t sctx;
4724 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4725 	uint16_t pci_vendor_id;
4726 
4727 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4728 		return (ENOTSUP);
4729 
4730 	pci_vendor_id = pci_get_vendor(dev);
4731 	pci_device_id = pci_get_device(dev);
4732 	pci_subvendor_id = pci_get_subvendor(dev);
4733 	pci_subdevice_id = pci_get_subdevice(dev);
4734 	pci_rev_id = pci_get_revid(dev);
4735 	if (sctx->isc_parse_devinfo != NULL)
4736 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4737 
4738 	ent = sctx->isc_vendor_info;
4739 	while (ent->pvi_vendor_id != 0) {
4740 		if (pci_vendor_id != ent->pvi_vendor_id) {
4741 			ent++;
4742 			continue;
4743 		}
4744 		if ((pci_device_id == ent->pvi_device_id) &&
4745 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4746 		     (ent->pvi_subvendor_id == 0)) &&
4747 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4748 		     (ent->pvi_subdevice_id == 0)) &&
4749 		    ((pci_rev_id == ent->pvi_rev_id) ||
4750 		     (ent->pvi_rev_id == 0))) {
4751 			device_set_desc_copy(dev, ent->pvi_name);
4752 			/* this needs to be changed to zero if the bus probing code
4753 			 * ever stops re-probing on best match because the sctx
4754 			 * may have its values over written by register calls
4755 			 * in subsequent probes
4756 			 */
4757 			return (BUS_PROBE_DEFAULT);
4758 		}
4759 		ent++;
4760 	}
4761 	return (ENXIO);
4762 }
4763 
4764 int
4765 iflib_device_probe_vendor(device_t dev)
4766 {
4767 	int probe;
4768 
4769 	probe = iflib_device_probe(dev);
4770 	if (probe == BUS_PROBE_DEFAULT)
4771 		return (BUS_PROBE_VENDOR);
4772 	else
4773 		return (probe);
4774 }
4775 
4776 static void
4777 iflib_reset_qvalues(if_ctx_t ctx)
4778 {
4779 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4780 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4781 	device_t dev = ctx->ifc_dev;
4782 	int i;
4783 
4784 	if (ctx->ifc_sysctl_ntxqs != 0)
4785 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4786 	if (ctx->ifc_sysctl_nrxqs != 0)
4787 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4788 
4789 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4790 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4791 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4792 		else
4793 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4794 	}
4795 
4796 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4797 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4798 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4799 		else
4800 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4801 	}
4802 
4803 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4804 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4805 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4806 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4807 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4808 		}
4809 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4810 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4811 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4812 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4813 		}
4814 		if (!powerof2(scctx->isc_nrxd[i])) {
4815 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4816 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4817 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4818 		}
4819 	}
4820 
4821 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4822 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4823 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4824 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4825 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4826 		}
4827 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4828 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4829 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4830 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4831 		}
4832 		if (!powerof2(scctx->isc_ntxd[i])) {
4833 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4834 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4835 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4836 		}
4837 	}
4838 }
4839 
4840 static void
4841 iflib_add_pfil(if_ctx_t ctx)
4842 {
4843 	struct pfil_head *pfil;
4844 	struct pfil_head_args pa;
4845 	iflib_rxq_t rxq;
4846 	int i;
4847 
4848 	pa.pa_version = PFIL_VERSION;
4849 	pa.pa_flags = PFIL_IN;
4850 	pa.pa_type = PFIL_TYPE_ETHERNET;
4851 	pa.pa_headname = if_name(ctx->ifc_ifp);
4852 	pfil = pfil_head_register(&pa);
4853 
4854 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4855 		rxq->pfil = pfil;
4856 	}
4857 }
4858 
4859 static void
4860 iflib_rem_pfil(if_ctx_t ctx)
4861 {
4862 	struct pfil_head *pfil;
4863 	iflib_rxq_t rxq;
4864 	int i;
4865 
4866 	rxq = ctx->ifc_rxqs;
4867 	pfil = rxq->pfil;
4868 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4869 		rxq->pfil = NULL;
4870 	}
4871 	pfil_head_unregister(pfil);
4872 }
4873 
4874 
4875 /*
4876  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4877  * cpuid and wrapping as necessary.
4878  */
4879 static unsigned int
4880 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4881 {
4882 	unsigned int first_valid;
4883 	unsigned int last_valid;
4884 
4885 	/* cpuid should always be in the valid set */
4886 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4887 
4888 	/* valid set should never be empty */
4889 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4890 
4891 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4892 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4893 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4894 	while (n > 0) {
4895 		do {
4896 			cpuid++;
4897 			if (cpuid > last_valid)
4898 				cpuid = first_valid;
4899 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4900 		n--;
4901 	}
4902 
4903 	return (cpuid);
4904 }
4905 
4906 #if defined(SMP) && defined(SCHED_ULE)
4907 extern struct cpu_group *cpu_top;              /* CPU topology */
4908 
4909 static int
4910 find_child_with_core(int cpu, struct cpu_group *grp)
4911 {
4912 	int i;
4913 
4914 	if (grp->cg_children == 0)
4915 		return -1;
4916 
4917 	MPASS(grp->cg_child);
4918 	for (i = 0; i < grp->cg_children; i++) {
4919 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4920 			return i;
4921 	}
4922 
4923 	return -1;
4924 }
4925 
4926 
4927 /*
4928  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4929  * does not distinguish among multiple L2 neighbors if the given CPU has
4930  * more than one (it will always return the same result in that case).
4931  */
4932 static int
4933 find_l2_neighbor(int cpu)
4934 {
4935 	struct cpu_group *grp;
4936 	int i;
4937 
4938 	grp = cpu_top;
4939 	if (grp == NULL)
4940 		return -1;
4941 
4942 	/*
4943 	 * Find the smallest CPU group that contains the given core.
4944 	 */
4945 	i = 0;
4946 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4947 		/*
4948 		 * If the smallest group containing the given CPU has less
4949 		 * than two members, we conclude the given CPU has no
4950 		 * L2 neighbor.
4951 		 */
4952 		if (grp->cg_child[i].cg_count <= 1)
4953 			return (-1);
4954 		grp = &grp->cg_child[i];
4955 	}
4956 
4957 	/* Must share L2. */
4958 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4959 		return -1;
4960 
4961 	/*
4962 	 * Select the first member of the set that isn't the reference
4963 	 * CPU, which at this point is guaranteed to exist.
4964 	 */
4965 	for (i = 0; i < CPU_SETSIZE; i++) {
4966 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4967 			return (i);
4968 	}
4969 
4970 	/* Should never be reached */
4971 	return (-1);
4972 }
4973 
4974 #else
4975 static int
4976 find_l2_neighbor(int cpu)
4977 {
4978 
4979 	return (-1);
4980 }
4981 #endif
4982 
4983 /*
4984  * CPU mapping behaviors
4985  * ---------------------
4986  * 'separate txrx' refers to the separate_txrx sysctl
4987  * 'use logical' refers to the use_logical_cores sysctl
4988  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4989  *
4990  *  separate     use     INTR
4991  *    txrx     logical   CPUS   result
4992  * ---------- --------- ------ ------------------------------------------------
4993  *     -          -       X     RX and TX queues mapped to consecutive physical
4994  *                              cores with RX/TX pairs on same core and excess
4995  *                              of either following
4996  *     -          X       X     RX and TX queues mapped to consecutive cores
4997  *                              of any type with RX/TX pairs on same core and
4998  *                              excess of either following
4999  *     X          -       X     RX and TX queues mapped to consecutive physical
5000  *                              cores; all RX then all TX
5001  *     X          X       X     RX queues mapped to consecutive physical cores
5002  *                              first, then TX queues mapped to L2 neighbor of
5003  *                              the corresponding RX queue if one exists,
5004  *                              otherwise to consecutive physical cores
5005  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5006  *                              any type with RX/TX pairs on same core and excess
5007  *                              of either following
5008  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5009  *                              any type; all RX then all TX
5010  */
5011 static unsigned int
5012 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5013     bool is_tx)
5014 {
5015 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5016 	unsigned int core_index;
5017 
5018 	if (ctx->ifc_sysctl_separate_txrx) {
5019 		/*
5020 		 * When using separate CPUs for TX and RX, the assignment
5021 		 * will always be of a consecutive CPU out of the set of
5022 		 * context CPUs, except for the specific case where the
5023 		 * context CPUs are phsyical cores, the use of logical cores
5024 		 * has been enabled, the assignment is for TX, the TX qid
5025 		 * corresponds to an RX qid, and the CPU assigned to the
5026 		 * corresponding RX queue has an L2 neighbor.
5027 		 */
5028 		if (ctx->ifc_sysctl_use_logical_cores &&
5029 		    ctx->ifc_cpus_are_physical_cores &&
5030 		    is_tx && qid < scctx->isc_nrxqsets) {
5031 			int l2_neighbor;
5032 			unsigned int rx_cpuid;
5033 
5034 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5035 			l2_neighbor = find_l2_neighbor(rx_cpuid);
5036 			if (l2_neighbor != -1) {
5037 				return (l2_neighbor);
5038 			}
5039 			/*
5040 			 * ... else fall through to the normal
5041 			 * consecutive-after-RX assignment scheme.
5042 			 *
5043 			 * Note that we are assuming that all RX queue CPUs
5044 			 * have an L2 neighbor, or all do not.  If a mixed
5045 			 * scenario is possible, we will have to keep track
5046 			 * separately of how many queues prior to this one
5047 			 * were not able to be assigned to an L2 neighbor.
5048 			 */
5049 		}
5050 		if (is_tx)
5051 			core_index = scctx->isc_nrxqsets + qid;
5052 		else
5053 			core_index = qid;
5054 	} else {
5055 		core_index = qid;
5056 	}
5057 
5058 	return (cpuid_advance(ctx, base_cpuid, core_index));
5059 }
5060 
5061 static uint16_t
5062 get_ctx_core_offset(if_ctx_t ctx)
5063 {
5064 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5065 	struct cpu_offset *op;
5066 	cpuset_t assigned_cpus;
5067 	unsigned int cores_consumed;
5068 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5069 	unsigned int first_valid;
5070 	unsigned int last_valid;
5071 	unsigned int i;
5072 
5073 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5074 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5075 
5076 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5077 		/*
5078 		 * Align the user-chosen base CPU ID to the next valid CPU
5079 		 * for this device.  If the chosen base CPU ID is smaller
5080 		 * than the first valid CPU or larger than the last valid
5081 		 * CPU, we assume the user does not know what the valid
5082 		 * range is for this device and is thinking in terms of a
5083 		 * zero-based reference frame, and so we shift the given
5084 		 * value into the valid range (and wrap accordingly) so the
5085 		 * intent is translated to the proper frame of reference.
5086 		 * If the base CPU ID is within the valid first/last, but
5087 		 * does not correspond to a valid CPU, it is advanced to the
5088 		 * next valid CPU (wrapping if necessary).
5089 		 */
5090 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5091 			/* shift from zero-based to first_valid-based */
5092 			base_cpuid += first_valid;
5093 			/* wrap to range [first_valid, last_valid] */
5094 			base_cpuid = (base_cpuid - first_valid) %
5095 			    (last_valid - first_valid + 1);
5096 		}
5097 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5098 			/*
5099 			 * base_cpuid is in [first_valid, last_valid], but
5100 			 * not a member of the valid set.  In this case,
5101 			 * there will always be a member of the valid set
5102 			 * with a CPU ID that is greater than base_cpuid,
5103 			 * and we simply advance to it.
5104 			 */
5105 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5106 				base_cpuid++;
5107 		}
5108 		return (base_cpuid);
5109 	}
5110 
5111 	/*
5112 	 * Determine how many cores will be consumed by performing the CPU
5113 	 * assignments and counting how many of the assigned CPUs correspond
5114 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5115 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5116 	 * the set of context CPUs.
5117 	 *
5118 	 * Note not all assigned CPUs will be in the set of context CPUs
5119 	 * when separate CPUs are being allocated to TX and RX queues,
5120 	 * assignment to logical cores has been enabled, the set of context
5121 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5122 	 * neighbors of CPUs that RX queues have been mapped to - in this
5123 	 * case we do only want to count how many CPUs in the set of context
5124 	 * CPUs have been consumed, as that determines the next CPU in that
5125 	 * set to start allocating at for the next device for which
5126 	 * core_offset is not set.
5127 	 */
5128 	CPU_ZERO(&assigned_cpus);
5129 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5130 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5131 		    &assigned_cpus);
5132 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5133 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5134 		    &assigned_cpus);
5135 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5136 	cores_consumed = CPU_COUNT(&assigned_cpus);
5137 
5138 	mtx_lock(&cpu_offset_mtx);
5139 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5140 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5141 			base_cpuid = op->next_cpuid;
5142 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5143 			    cores_consumed);
5144 			MPASS(op->refcount < UINT_MAX);
5145 			op->refcount++;
5146 			break;
5147 		}
5148 	}
5149 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5150 		base_cpuid = first_valid;
5151 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5152 		    M_NOWAIT | M_ZERO);
5153 		if (op == NULL) {
5154 			device_printf(ctx->ifc_dev,
5155 			    "allocation for cpu offset failed.\n");
5156 		} else {
5157 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5158 			    cores_consumed);
5159 			op->refcount = 1;
5160 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5161 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5162 		}
5163 	}
5164 	mtx_unlock(&cpu_offset_mtx);
5165 
5166 	return (base_cpuid);
5167 }
5168 
5169 static void
5170 unref_ctx_core_offset(if_ctx_t ctx)
5171 {
5172 	struct cpu_offset *op, *top;
5173 
5174 	mtx_lock(&cpu_offset_mtx);
5175 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5176 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5177 			MPASS(op->refcount > 0);
5178 			op->refcount--;
5179 			if (op->refcount == 0) {
5180 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5181 				free(op, M_IFLIB);
5182 			}
5183 			break;
5184 		}
5185 	}
5186 	mtx_unlock(&cpu_offset_mtx);
5187 }
5188 
5189 int
5190 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5191 {
5192 	if_ctx_t ctx;
5193 	if_t ifp;
5194 	if_softc_ctx_t scctx;
5195 	kobjop_desc_t kobj_desc;
5196 	kobj_method_t *kobj_method;
5197 	int err, msix, rid;
5198 	int num_txd, num_rxd;
5199 
5200 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5201 
5202 	if (sc == NULL) {
5203 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5204 		device_set_softc(dev, ctx);
5205 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5206 	}
5207 
5208 	ctx->ifc_sctx = sctx;
5209 	ctx->ifc_dev = dev;
5210 	ctx->ifc_softc = sc;
5211 
5212 	if ((err = iflib_register(ctx)) != 0) {
5213 		device_printf(dev, "iflib_register failed %d\n", err);
5214 		goto fail_ctx_free;
5215 	}
5216 	iflib_add_device_sysctl_pre(ctx);
5217 
5218 	scctx = &ctx->ifc_softc_ctx;
5219 	ifp = ctx->ifc_ifp;
5220 
5221 	iflib_reset_qvalues(ctx);
5222 	IFNET_WLOCK();
5223 	CTX_LOCK(ctx);
5224 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5225 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5226 		goto fail_unlock;
5227 	}
5228 	_iflib_pre_assert(scctx);
5229 	ctx->ifc_txrx = *scctx->isc_txrx;
5230 
5231 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5232 
5233 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5234 		ctx->ifc_mediap = scctx->isc_media;
5235 
5236 #ifdef INVARIANTS
5237 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5238 		MPASS(scctx->isc_tx_csum_flags);
5239 #endif
5240 
5241 	if_setcapabilities(ifp,
5242 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5243 	if_setcapenable(ifp,
5244 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5245 
5246 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5247 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5248 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5249 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5250 
5251 	num_txd = iflib_num_tx_descs(ctx);
5252 	num_rxd = iflib_num_rx_descs(ctx);
5253 
5254 	/* XXX change for per-queue sizes */
5255 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5256 	    num_txd, num_rxd);
5257 
5258 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5259 		scctx->isc_tx_nsegments = max(1, num_txd /
5260 		    MAX_SINGLE_PACKET_FRACTION);
5261 	if (scctx->isc_tx_tso_segments_max > num_txd /
5262 	    MAX_SINGLE_PACKET_FRACTION)
5263 		scctx->isc_tx_tso_segments_max = max(1,
5264 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5265 
5266 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5267 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5268 		/*
5269 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5270 		 * but some MACs do.
5271 		 */
5272 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5273 		    IP_MAXPACKET));
5274 		/*
5275 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5276 		 * into account.  In the worst case, each of these calls will
5277 		 * add another mbuf and, thus, the requirement for another DMA
5278 		 * segment.  So for best performance, it doesn't make sense to
5279 		 * advertize a maximum of TSO segments that typically will
5280 		 * require defragmentation in iflib_encap().
5281 		 */
5282 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5283 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5284 	}
5285 	if (scctx->isc_rss_table_size == 0)
5286 		scctx->isc_rss_table_size = 64;
5287 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5288 
5289 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5290 	/* XXX format name */
5291 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5292 	    NULL, NULL, "admin");
5293 
5294 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5295 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5296 		device_printf(dev, "Unable to fetch CPU list\n");
5297 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5298 		ctx->ifc_cpus_are_physical_cores = false;
5299 	} else
5300 		ctx->ifc_cpus_are_physical_cores = true;
5301 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5302 
5303 	/*
5304 	** Now set up MSI or MSI-X, should return us the number of supported
5305 	** vectors (will be 1 for a legacy interrupt and MSI).
5306 	*/
5307 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5308 		msix = scctx->isc_vectors;
5309 	} else if (scctx->isc_msix_bar != 0)
5310 	       /*
5311 		* The simple fact that isc_msix_bar is not 0 does not mean we
5312 		* we have a good value there that is known to work.
5313 		*/
5314 		msix = iflib_msix_init(ctx);
5315 	else {
5316 		scctx->isc_vectors = 1;
5317 		scctx->isc_ntxqsets = 1;
5318 		scctx->isc_nrxqsets = 1;
5319 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5320 		msix = 0;
5321 	}
5322 	/* Get memory for the station queues */
5323 	if ((err = iflib_queues_alloc(ctx))) {
5324 		device_printf(dev, "Unable to allocate queue memory\n");
5325 		goto fail_intr_free;
5326 	}
5327 
5328 	if ((err = iflib_qset_structures_setup(ctx)))
5329 		goto fail_queues;
5330 
5331 	/*
5332 	 * Now that we know how many queues there are, get the core offset.
5333 	 */
5334 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5335 
5336 	if (msix > 1) {
5337 		/*
5338 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5339 		 * aren't the default NULL implementation.
5340 		 */
5341 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5342 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5343 		    kobj_desc);
5344 		if (kobj_method == &kobj_desc->deflt) {
5345 			device_printf(dev,
5346 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5347 			err = EOPNOTSUPP;
5348 			goto fail_queues;
5349 		}
5350 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5351 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5352 		    kobj_desc);
5353 		if (kobj_method == &kobj_desc->deflt) {
5354 			device_printf(dev,
5355 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5356 			err = EOPNOTSUPP;
5357 			goto fail_queues;
5358 		}
5359 
5360 		/*
5361 		 * Assign the MSI-X vectors.
5362 		 * Note that the default NULL ifdi_msix_intr_assign method will
5363 		 * fail here, too.
5364 		 */
5365 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5366 		if (err != 0) {
5367 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5368 			    err);
5369 			goto fail_queues;
5370 		}
5371 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5372 		rid = 0;
5373 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5374 			MPASS(msix == 1);
5375 			rid = 1;
5376 		}
5377 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5378 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5379 			goto fail_queues;
5380 		}
5381 	} else {
5382 		device_printf(dev,
5383 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5384 		err = ENODEV;
5385 		goto fail_queues;
5386 	}
5387 
5388 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5389 
5390 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5391 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5392 		goto fail_detach;
5393 	}
5394 
5395 	/*
5396 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5397 	 * This must appear after the call to ether_ifattach() because
5398 	 * ether_ifattach() sets if_hdrlen to the default value.
5399 	 */
5400 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5401 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5402 
5403 	if ((err = iflib_netmap_attach(ctx))) {
5404 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5405 		goto fail_detach;
5406 	}
5407 	*ctxp = ctx;
5408 
5409 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5410 
5411 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5412 	iflib_add_device_sysctl_post(ctx);
5413 	iflib_add_pfil(ctx);
5414 	ctx->ifc_flags |= IFC_INIT_DONE;
5415 	CTX_UNLOCK(ctx);
5416 	IFNET_WUNLOCK();
5417 
5418 	return (0);
5419 
5420 fail_detach:
5421 	ether_ifdetach(ctx->ifc_ifp);
5422 fail_queues:
5423 	iflib_tqg_detach(ctx);
5424 	iflib_tx_structures_free(ctx);
5425 	iflib_rx_structures_free(ctx);
5426 	IFDI_DETACH(ctx);
5427 	IFDI_QUEUES_FREE(ctx);
5428 fail_intr_free:
5429 	iflib_free_intr_mem(ctx);
5430 fail_unlock:
5431 	CTX_UNLOCK(ctx);
5432 	IFNET_WUNLOCK();
5433 	iflib_deregister(ctx);
5434 fail_ctx_free:
5435 	device_set_softc(ctx->ifc_dev, NULL);
5436         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5437                 free(ctx->ifc_softc, M_IFLIB);
5438         free(ctx, M_IFLIB);
5439 	return (err);
5440 }
5441 
5442 int
5443 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5444 					  struct iflib_cloneattach_ctx *clctx)
5445 {
5446 	int num_txd, num_rxd;
5447 	int err;
5448 	if_ctx_t ctx;
5449 	if_t ifp;
5450 	if_softc_ctx_t scctx;
5451 	int i;
5452 	void *sc;
5453 
5454 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5455 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5456 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5457 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5458 		ctx->ifc_flags |= IFC_PSEUDO;
5459 
5460 	ctx->ifc_sctx = sctx;
5461 	ctx->ifc_softc = sc;
5462 	ctx->ifc_dev = dev;
5463 
5464 	if ((err = iflib_register(ctx)) != 0) {
5465 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5466 		goto fail_ctx_free;
5467 	}
5468 	iflib_add_device_sysctl_pre(ctx);
5469 
5470 	scctx = &ctx->ifc_softc_ctx;
5471 	ifp = ctx->ifc_ifp;
5472 
5473 	iflib_reset_qvalues(ctx);
5474 	CTX_LOCK(ctx);
5475 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5476 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5477 		goto fail_unlock;
5478 	}
5479 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5480 		ether_gen_addr(ifp, &ctx->ifc_mac);
5481 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5482 								clctx->cc_params)) != 0) {
5483 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5484 		goto fail_unlock;
5485 	}
5486 #ifdef INVARIANTS
5487 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5488 		MPASS(scctx->isc_tx_csum_flags);
5489 #endif
5490 
5491 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5492 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5493 
5494 	if_setflagbits(ifp, IFF_NOGROUP, 0);
5495 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5496 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5497 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5498 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5499 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5500 		} else {
5501 			if_attach(ctx->ifc_ifp);
5502 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5503 		}
5504 
5505 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5506 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5507 			goto fail_detach;
5508 		}
5509 		*ctxp = ctx;
5510 
5511 		/*
5512 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5513 		 * This must appear after the call to ether_ifattach() because
5514 		 * ether_ifattach() sets if_hdrlen to the default value.
5515 		 */
5516 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5517 			if_setifheaderlen(ifp,
5518 			    sizeof(struct ether_vlan_header));
5519 
5520 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5521 		iflib_add_device_sysctl_post(ctx);
5522 		ctx->ifc_flags |= IFC_INIT_DONE;
5523 		CTX_UNLOCK(ctx);
5524 		return (0);
5525 	}
5526 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5527 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5528 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5529 
5530 	_iflib_pre_assert(scctx);
5531 	ctx->ifc_txrx = *scctx->isc_txrx;
5532 
5533 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5534 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5535 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5536 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5537 
5538 	num_txd = iflib_num_tx_descs(ctx);
5539 	num_rxd = iflib_num_rx_descs(ctx);
5540 
5541 	/* XXX change for per-queue sizes */
5542 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5543 	    num_txd, num_rxd);
5544 
5545 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5546 		scctx->isc_tx_nsegments = max(1, num_txd /
5547 		    MAX_SINGLE_PACKET_FRACTION);
5548 	if (scctx->isc_tx_tso_segments_max > num_txd /
5549 	    MAX_SINGLE_PACKET_FRACTION)
5550 		scctx->isc_tx_tso_segments_max = max(1,
5551 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5552 
5553 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5554 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5555 		/*
5556 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5557 		 * but some MACs do.
5558 		 */
5559 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5560 		    IP_MAXPACKET));
5561 		/*
5562 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5563 		 * into account.  In the worst case, each of these calls will
5564 		 * add another mbuf and, thus, the requirement for another DMA
5565 		 * segment.  So for best performance, it doesn't make sense to
5566 		 * advertize a maximum of TSO segments that typically will
5567 		 * require defragmentation in iflib_encap().
5568 		 */
5569 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5570 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5571 	}
5572 	if (scctx->isc_rss_table_size == 0)
5573 		scctx->isc_rss_table_size = 64;
5574 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5575 
5576 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5577 	/* XXX format name */
5578 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5579 	    NULL, NULL, "admin");
5580 
5581 	/* XXX --- can support > 1 -- but keep it simple for now */
5582 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5583 
5584 	/* Get memory for the station queues */
5585 	if ((err = iflib_queues_alloc(ctx))) {
5586 		device_printf(dev, "Unable to allocate queue memory\n");
5587 		goto fail_iflib_detach;
5588 	}
5589 
5590 	if ((err = iflib_qset_structures_setup(ctx))) {
5591 		device_printf(dev, "qset structure setup failed %d\n", err);
5592 		goto fail_queues;
5593 	}
5594 
5595 	/*
5596 	 * XXX What if anything do we want to do about interrupts?
5597 	 */
5598 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5599 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5600 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5601 		goto fail_detach;
5602 	}
5603 
5604 	/*
5605 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5606 	 * This must appear after the call to ether_ifattach() because
5607 	 * ether_ifattach() sets if_hdrlen to the default value.
5608 	 */
5609 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5610 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5611 
5612 	/* XXX handle more than one queue */
5613 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5614 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5615 
5616 	*ctxp = ctx;
5617 
5618 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5619 	iflib_add_device_sysctl_post(ctx);
5620 	ctx->ifc_flags |= IFC_INIT_DONE;
5621 	CTX_UNLOCK(ctx);
5622 
5623 	return (0);
5624 fail_detach:
5625 	ether_ifdetach(ctx->ifc_ifp);
5626 fail_queues:
5627 	iflib_tqg_detach(ctx);
5628 	iflib_tx_structures_free(ctx);
5629 	iflib_rx_structures_free(ctx);
5630 fail_iflib_detach:
5631 	IFDI_DETACH(ctx);
5632 	IFDI_QUEUES_FREE(ctx);
5633 fail_unlock:
5634 	CTX_UNLOCK(ctx);
5635 	iflib_deregister(ctx);
5636 fail_ctx_free:
5637 	free(ctx->ifc_softc, M_IFLIB);
5638 	free(ctx, M_IFLIB);
5639 	return (err);
5640 }
5641 
5642 int
5643 iflib_pseudo_deregister(if_ctx_t ctx)
5644 {
5645 	if_t ifp = ctx->ifc_ifp;
5646 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5647 
5648 	/* Unregister VLAN event handlers early */
5649 	iflib_unregister_vlan_handlers(ctx);
5650 
5651 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5652 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5653 		bpfdetach(ifp);
5654 		if_detach(ifp);
5655 	} else {
5656 		ether_ifdetach(ifp);
5657 	}
5658 
5659 	iflib_tqg_detach(ctx);
5660 	iflib_tx_structures_free(ctx);
5661 	iflib_rx_structures_free(ctx);
5662 	IFDI_DETACH(ctx);
5663 	IFDI_QUEUES_FREE(ctx);
5664 
5665 	iflib_deregister(ctx);
5666 
5667 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5668 		free(ctx->ifc_softc, M_IFLIB);
5669 	free(ctx, M_IFLIB);
5670 	return (0);
5671 }
5672 
5673 int
5674 iflib_device_attach(device_t dev)
5675 {
5676 	if_ctx_t ctx;
5677 	if_shared_ctx_t sctx;
5678 
5679 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5680 		return (ENOTSUP);
5681 
5682 	pci_enable_busmaster(dev);
5683 
5684 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5685 }
5686 
5687 int
5688 iflib_device_deregister(if_ctx_t ctx)
5689 {
5690 	if_t ifp = ctx->ifc_ifp;
5691 	device_t dev = ctx->ifc_dev;
5692 
5693 	/* Make sure VLANS are not using driver */
5694 	if (if_vlantrunkinuse(ifp)) {
5695 		device_printf(dev, "Vlan in use, detach first\n");
5696 		return (EBUSY);
5697 	}
5698 #ifdef PCI_IOV
5699 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5700 		device_printf(dev, "SR-IOV in use; detach first.\n");
5701 		return (EBUSY);
5702 	}
5703 #endif
5704 
5705 	STATE_LOCK(ctx);
5706 	ctx->ifc_flags |= IFC_IN_DETACH;
5707 	STATE_UNLOCK(ctx);
5708 
5709 	/* Unregister VLAN handlers before calling iflib_stop() */
5710 	iflib_unregister_vlan_handlers(ctx);
5711 
5712 	iflib_netmap_detach(ifp);
5713 	ether_ifdetach(ifp);
5714 
5715 	CTX_LOCK(ctx);
5716 	iflib_stop(ctx);
5717 	CTX_UNLOCK(ctx);
5718 
5719 	iflib_rem_pfil(ctx);
5720 	if (ctx->ifc_led_dev != NULL)
5721 		led_destroy(ctx->ifc_led_dev);
5722 
5723 	iflib_tqg_detach(ctx);
5724 	iflib_tx_structures_free(ctx);
5725 	iflib_rx_structures_free(ctx);
5726 
5727 	CTX_LOCK(ctx);
5728 	IFDI_DETACH(ctx);
5729 	IFDI_QUEUES_FREE(ctx);
5730 	CTX_UNLOCK(ctx);
5731 
5732 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5733 	iflib_free_intr_mem(ctx);
5734 
5735 	bus_generic_detach(dev);
5736 
5737 	iflib_deregister(ctx);
5738 
5739 	device_set_softc(ctx->ifc_dev, NULL);
5740 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5741 		free(ctx->ifc_softc, M_IFLIB);
5742 	unref_ctx_core_offset(ctx);
5743 	free(ctx, M_IFLIB);
5744 	return (0);
5745 }
5746 
5747 static void
5748 iflib_tqg_detach(if_ctx_t ctx)
5749 {
5750 	iflib_txq_t txq;
5751 	iflib_rxq_t rxq;
5752 	int i;
5753 	struct taskqgroup *tqg;
5754 
5755 	/* XXX drain any dependent tasks */
5756 	tqg = qgroup_if_io_tqg;
5757 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5758 		callout_drain(&txq->ift_timer);
5759 #ifdef DEV_NETMAP
5760 		callout_drain(&txq->ift_netmap_timer);
5761 #endif /* DEV_NETMAP */
5762 		if (txq->ift_task.gt_uniq != NULL)
5763 			taskqgroup_detach(tqg, &txq->ift_task);
5764 	}
5765 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5766 		if (rxq->ifr_task.gt_uniq != NULL)
5767 			taskqgroup_detach(tqg, &rxq->ifr_task);
5768 	}
5769 	tqg = qgroup_if_config_tqg;
5770 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5771 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5772 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5773 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5774 }
5775 
5776 static void
5777 iflib_free_intr_mem(if_ctx_t ctx)
5778 {
5779 
5780 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5781 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5782 	}
5783 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5784 		pci_release_msi(ctx->ifc_dev);
5785 	}
5786 	if (ctx->ifc_msix_mem != NULL) {
5787 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5788 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5789 		ctx->ifc_msix_mem = NULL;
5790 	}
5791 }
5792 
5793 int
5794 iflib_device_detach(device_t dev)
5795 {
5796 	if_ctx_t ctx = device_get_softc(dev);
5797 
5798 	return (iflib_device_deregister(ctx));
5799 }
5800 
5801 int
5802 iflib_device_suspend(device_t dev)
5803 {
5804 	if_ctx_t ctx = device_get_softc(dev);
5805 
5806 	CTX_LOCK(ctx);
5807 	IFDI_SUSPEND(ctx);
5808 	CTX_UNLOCK(ctx);
5809 
5810 	return bus_generic_suspend(dev);
5811 }
5812 int
5813 iflib_device_shutdown(device_t dev)
5814 {
5815 	if_ctx_t ctx = device_get_softc(dev);
5816 
5817 	CTX_LOCK(ctx);
5818 	IFDI_SHUTDOWN(ctx);
5819 	CTX_UNLOCK(ctx);
5820 
5821 	return bus_generic_suspend(dev);
5822 }
5823 
5824 int
5825 iflib_device_resume(device_t dev)
5826 {
5827 	if_ctx_t ctx = device_get_softc(dev);
5828 	iflib_txq_t txq = ctx->ifc_txqs;
5829 
5830 	CTX_LOCK(ctx);
5831 	IFDI_RESUME(ctx);
5832 	iflib_if_init_locked(ctx);
5833 	CTX_UNLOCK(ctx);
5834 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5835 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5836 
5837 	return (bus_generic_resume(dev));
5838 }
5839 
5840 int
5841 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5842 {
5843 	int error;
5844 	if_ctx_t ctx = device_get_softc(dev);
5845 
5846 	CTX_LOCK(ctx);
5847 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5848 	CTX_UNLOCK(ctx);
5849 
5850 	return (error);
5851 }
5852 
5853 void
5854 iflib_device_iov_uninit(device_t dev)
5855 {
5856 	if_ctx_t ctx = device_get_softc(dev);
5857 
5858 	CTX_LOCK(ctx);
5859 	IFDI_IOV_UNINIT(ctx);
5860 	CTX_UNLOCK(ctx);
5861 }
5862 
5863 int
5864 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5865 {
5866 	int error;
5867 	if_ctx_t ctx = device_get_softc(dev);
5868 
5869 	CTX_LOCK(ctx);
5870 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5871 	CTX_UNLOCK(ctx);
5872 
5873 	return (error);
5874 }
5875 
5876 /*********************************************************************
5877  *
5878  *  MODULE FUNCTION DEFINITIONS
5879  *
5880  **********************************************************************/
5881 
5882 /*
5883  * - Start a fast taskqueue thread for each core
5884  * - Start a taskqueue for control operations
5885  */
5886 static int
5887 iflib_module_init(void)
5888 {
5889 	iflib_timer_default = hz / 2;
5890 	return (0);
5891 }
5892 
5893 static int
5894 iflib_module_event_handler(module_t mod, int what, void *arg)
5895 {
5896 	int err;
5897 
5898 	switch (what) {
5899 	case MOD_LOAD:
5900 		if ((err = iflib_module_init()) != 0)
5901 			return (err);
5902 		break;
5903 	case MOD_UNLOAD:
5904 		return (EBUSY);
5905 	default:
5906 		return (EOPNOTSUPP);
5907 	}
5908 
5909 	return (0);
5910 }
5911 
5912 /*********************************************************************
5913  *
5914  *  PUBLIC FUNCTION DEFINITIONS
5915  *     ordered as in iflib.h
5916  *
5917  **********************************************************************/
5918 
5919 static void
5920 _iflib_assert(if_shared_ctx_t sctx)
5921 {
5922 	int i;
5923 
5924 	MPASS(sctx->isc_tx_maxsize);
5925 	MPASS(sctx->isc_tx_maxsegsize);
5926 
5927 	MPASS(sctx->isc_rx_maxsize);
5928 	MPASS(sctx->isc_rx_nsegments);
5929 	MPASS(sctx->isc_rx_maxsegsize);
5930 
5931 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5932 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5933 		MPASS(sctx->isc_nrxd_min[i]);
5934 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5935 		MPASS(sctx->isc_nrxd_max[i]);
5936 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5937 		MPASS(sctx->isc_nrxd_default[i]);
5938 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5939 	}
5940 
5941 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5942 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5943 		MPASS(sctx->isc_ntxd_min[i]);
5944 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5945 		MPASS(sctx->isc_ntxd_max[i]);
5946 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5947 		MPASS(sctx->isc_ntxd_default[i]);
5948 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5949 	}
5950 }
5951 
5952 static void
5953 _iflib_pre_assert(if_softc_ctx_t scctx)
5954 {
5955 
5956 	MPASS(scctx->isc_txrx->ift_txd_encap);
5957 	MPASS(scctx->isc_txrx->ift_txd_flush);
5958 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5959 	MPASS(scctx->isc_txrx->ift_rxd_available);
5960 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5961 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5962 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5963 }
5964 
5965 static int
5966 iflib_register(if_ctx_t ctx)
5967 {
5968 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5969 	driver_t *driver = sctx->isc_driver;
5970 	device_t dev = ctx->ifc_dev;
5971 	if_t ifp;
5972 	u_char type;
5973 	int iflags;
5974 
5975 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5976 		_iflib_assert(sctx);
5977 
5978 	CTX_LOCK_INIT(ctx);
5979 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5980 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5981 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5982 			type = IFT_ETHER;
5983 		else
5984 			type = IFT_PPP;
5985 	} else
5986 		type = IFT_ETHER;
5987 	ifp = ctx->ifc_ifp = if_alloc(type);
5988 	if (ifp == NULL) {
5989 		device_printf(dev, "can not allocate ifnet structure\n");
5990 		return (ENOMEM);
5991 	}
5992 
5993 	/*
5994 	 * Initialize our context's device specific methods
5995 	 */
5996 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5997 	kobj_class_compile((kobj_class_t) driver);
5998 
5999 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
6000 	if_setsoftc(ifp, ctx);
6001 	if_setdev(ifp, dev);
6002 	if_setinitfn(ifp, iflib_if_init);
6003 	if_setioctlfn(ifp, iflib_if_ioctl);
6004 #ifdef ALTQ
6005 	if_setstartfn(ifp, iflib_altq_if_start);
6006 	if_settransmitfn(ifp, iflib_altq_if_transmit);
6007 	if_setsendqready(ifp);
6008 #else
6009 	if_settransmitfn(ifp, iflib_if_transmit);
6010 #endif
6011 	if_setqflushfn(ifp, iflib_if_qflush);
6012 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
6013 
6014 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
6015 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
6016 		iflags |= IFF_POINTOPOINT;
6017 	else
6018 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
6019 	if_setflags(ifp, iflags);
6020 	ctx->ifc_vlan_attach_event =
6021 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
6022 							  EVENTHANDLER_PRI_FIRST);
6023 	ctx->ifc_vlan_detach_event =
6024 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
6025 							  EVENTHANDLER_PRI_FIRST);
6026 
6027 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
6028 		ctx->ifc_mediap = &ctx->ifc_media;
6029 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
6030 		    iflib_media_change, iflib_media_status);
6031 	}
6032 	return (0);
6033 }
6034 
6035 static void
6036 iflib_unregister_vlan_handlers(if_ctx_t ctx)
6037 {
6038 	/* Unregister VLAN events */
6039 	if (ctx->ifc_vlan_attach_event != NULL) {
6040 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
6041 		ctx->ifc_vlan_attach_event = NULL;
6042 	}
6043 	if (ctx->ifc_vlan_detach_event != NULL) {
6044 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
6045 		ctx->ifc_vlan_detach_event = NULL;
6046 	}
6047 
6048 }
6049 
6050 static void
6051 iflib_deregister(if_ctx_t ctx)
6052 {
6053 	if_t ifp = ctx->ifc_ifp;
6054 
6055 	/* Remove all media */
6056 	ifmedia_removeall(&ctx->ifc_media);
6057 
6058 	/* Ensure that VLAN event handlers are unregistered */
6059 	iflib_unregister_vlan_handlers(ctx);
6060 
6061 	/* Release kobject reference */
6062 	kobj_delete((kobj_t) ctx, NULL);
6063 
6064 	/* Free the ifnet structure */
6065 	if_free(ifp);
6066 
6067 	STATE_LOCK_DESTROY(ctx);
6068 
6069 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
6070 	CTX_LOCK_DESTROY(ctx);
6071 }
6072 
6073 static int
6074 iflib_queues_alloc(if_ctx_t ctx)
6075 {
6076 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6077 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6078 	device_t dev = ctx->ifc_dev;
6079 	int nrxqsets = scctx->isc_nrxqsets;
6080 	int ntxqsets = scctx->isc_ntxqsets;
6081 	iflib_txq_t txq;
6082 	iflib_rxq_t rxq;
6083 	iflib_fl_t fl = NULL;
6084 	int i, j, cpu, err, txconf, rxconf;
6085 	iflib_dma_info_t ifdip;
6086 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
6087 	uint32_t *txqsizes = scctx->isc_txqsizes;
6088 	uint8_t nrxqs = sctx->isc_nrxqs;
6089 	uint8_t ntxqs = sctx->isc_ntxqs;
6090 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
6091 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
6092 	caddr_t *vaddrs;
6093 	uint64_t *paddrs;
6094 
6095 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
6096 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
6097 	KASSERT(nrxqs >= fl_offset + nfree_lists,
6098            ("there must be at least a rxq for each free list"));
6099 
6100 	/* Allocate the TX ring struct memory */
6101 	if (!(ctx->ifc_txqs =
6102 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
6103 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6104 		device_printf(dev, "Unable to allocate TX ring memory\n");
6105 		err = ENOMEM;
6106 		goto fail;
6107 	}
6108 
6109 	/* Now allocate the RX */
6110 	if (!(ctx->ifc_rxqs =
6111 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
6112 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6113 		device_printf(dev, "Unable to allocate RX ring memory\n");
6114 		err = ENOMEM;
6115 		goto rx_fail;
6116 	}
6117 
6118 	txq = ctx->ifc_txqs;
6119 	rxq = ctx->ifc_rxqs;
6120 
6121 	/*
6122 	 * XXX handle allocation failure
6123 	 */
6124 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
6125 		/* Set up some basics */
6126 
6127 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
6128 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6129 			device_printf(dev,
6130 			    "Unable to allocate TX DMA info memory\n");
6131 			err = ENOMEM;
6132 			goto err_tx_desc;
6133 		}
6134 		txq->ift_ifdi = ifdip;
6135 		for (j = 0; j < ntxqs; j++, ifdip++) {
6136 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
6137 				device_printf(dev,
6138 				    "Unable to allocate TX descriptors\n");
6139 				err = ENOMEM;
6140 				goto err_tx_desc;
6141 			}
6142 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
6143 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
6144 		}
6145 		txq->ift_ctx = ctx;
6146 		txq->ift_id = i;
6147 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
6148 			txq->ift_br_offset = 1;
6149 		} else {
6150 			txq->ift_br_offset = 0;
6151 		}
6152 
6153 		if (iflib_txsd_alloc(txq)) {
6154 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6155 			err = ENOMEM;
6156 			goto err_tx_desc;
6157 		}
6158 
6159 		/* Initialize the TX lock */
6160 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6161 		    device_get_nameunit(dev), txq->ift_id);
6162 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6163 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6164 		txq->ift_timer.c_cpu = cpu;
6165 #ifdef DEV_NETMAP
6166 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6167 		txq->ift_netmap_timer.c_cpu = cpu;
6168 #endif /* DEV_NETMAP */
6169 
6170 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6171 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6172 		if (err) {
6173 			/* XXX free any allocated rings */
6174 			device_printf(dev, "Unable to allocate buf_ring\n");
6175 			goto err_tx_desc;
6176 		}
6177 	}
6178 
6179 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6180 		/* Set up some basics */
6181 		callout_init(&rxq->ifr_watchdog, 1);
6182 
6183 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6184 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6185 			device_printf(dev,
6186 			    "Unable to allocate RX DMA info memory\n");
6187 			err = ENOMEM;
6188 			goto err_tx_desc;
6189 		}
6190 
6191 		rxq->ifr_ifdi = ifdip;
6192 		/* XXX this needs to be changed if #rx queues != #tx queues */
6193 		rxq->ifr_ntxqirq = 1;
6194 		rxq->ifr_txqid[0] = i;
6195 		for (j = 0; j < nrxqs; j++, ifdip++) {
6196 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6197 				device_printf(dev,
6198 				    "Unable to allocate RX descriptors\n");
6199 				err = ENOMEM;
6200 				goto err_tx_desc;
6201 			}
6202 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6203 		}
6204 		rxq->ifr_ctx = ctx;
6205 		rxq->ifr_id = i;
6206 		rxq->ifr_fl_offset = fl_offset;
6207 		rxq->ifr_nfl = nfree_lists;
6208 		if (!(fl =
6209 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6210 			device_printf(dev, "Unable to allocate free list memory\n");
6211 			err = ENOMEM;
6212 			goto err_tx_desc;
6213 		}
6214 		rxq->ifr_fl = fl;
6215 		for (j = 0; j < nfree_lists; j++) {
6216 			fl[j].ifl_rxq = rxq;
6217 			fl[j].ifl_id = j;
6218 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6219 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6220 		}
6221 		/* Allocate receive buffers for the ring */
6222 		if (iflib_rxsd_alloc(rxq)) {
6223 			device_printf(dev,
6224 			    "Critical Failure setting up receive buffers\n");
6225 			err = ENOMEM;
6226 			goto err_rx_desc;
6227 		}
6228 
6229 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6230 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6231 			    M_WAITOK);
6232 	}
6233 
6234 	/* TXQs */
6235 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6236 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6237 	for (i = 0; i < ntxqsets; i++) {
6238 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6239 
6240 		for (j = 0; j < ntxqs; j++, di++) {
6241 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6242 			paddrs[i*ntxqs + j] = di->idi_paddr;
6243 		}
6244 	}
6245 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6246 		device_printf(ctx->ifc_dev,
6247 		    "Unable to allocate device TX queue\n");
6248 		iflib_tx_structures_free(ctx);
6249 		free(vaddrs, M_IFLIB);
6250 		free(paddrs, M_IFLIB);
6251 		goto err_rx_desc;
6252 	}
6253 	free(vaddrs, M_IFLIB);
6254 	free(paddrs, M_IFLIB);
6255 
6256 	/* RXQs */
6257 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6258 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6259 	for (i = 0; i < nrxqsets; i++) {
6260 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6261 
6262 		for (j = 0; j < nrxqs; j++, di++) {
6263 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6264 			paddrs[i*nrxqs + j] = di->idi_paddr;
6265 		}
6266 	}
6267 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6268 		device_printf(ctx->ifc_dev,
6269 		    "Unable to allocate device RX queue\n");
6270 		iflib_tx_structures_free(ctx);
6271 		free(vaddrs, M_IFLIB);
6272 		free(paddrs, M_IFLIB);
6273 		goto err_rx_desc;
6274 	}
6275 	free(vaddrs, M_IFLIB);
6276 	free(paddrs, M_IFLIB);
6277 
6278 	return (0);
6279 
6280 /* XXX handle allocation failure changes */
6281 err_rx_desc:
6282 err_tx_desc:
6283 rx_fail:
6284 	if (ctx->ifc_rxqs != NULL)
6285 		free(ctx->ifc_rxqs, M_IFLIB);
6286 	ctx->ifc_rxqs = NULL;
6287 	if (ctx->ifc_txqs != NULL)
6288 		free(ctx->ifc_txqs, M_IFLIB);
6289 	ctx->ifc_txqs = NULL;
6290 fail:
6291 	return (err);
6292 }
6293 
6294 static int
6295 iflib_tx_structures_setup(if_ctx_t ctx)
6296 {
6297 	iflib_txq_t txq = ctx->ifc_txqs;
6298 	int i;
6299 
6300 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6301 		iflib_txq_setup(txq);
6302 
6303 	return (0);
6304 }
6305 
6306 static void
6307 iflib_tx_structures_free(if_ctx_t ctx)
6308 {
6309 	iflib_txq_t txq = ctx->ifc_txqs;
6310 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6311 	int i, j;
6312 
6313 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6314 		for (j = 0; j < sctx->isc_ntxqs; j++)
6315 			iflib_dma_free(&txq->ift_ifdi[j]);
6316 		iflib_txq_destroy(txq);
6317 	}
6318 	free(ctx->ifc_txqs, M_IFLIB);
6319 	ctx->ifc_txqs = NULL;
6320 }
6321 
6322 /*********************************************************************
6323  *
6324  *  Initialize all receive rings.
6325  *
6326  **********************************************************************/
6327 static int
6328 iflib_rx_structures_setup(if_ctx_t ctx)
6329 {
6330 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6331 	int q;
6332 #if defined(INET6) || defined(INET)
6333 	int err, i;
6334 #endif
6335 
6336 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6337 #if defined(INET6) || defined(INET)
6338 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6339 		    TCP_LRO_ENTRIES, min(1024,
6340 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6341 		if (err != 0) {
6342 			device_printf(ctx->ifc_dev,
6343 			    "LRO Initialization failed!\n");
6344 			goto fail;
6345 		}
6346 #endif
6347 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6348 	}
6349 	return (0);
6350 #if defined(INET6) || defined(INET)
6351 fail:
6352 	/*
6353 	 * Free LRO resources allocated so far, we will only handle
6354 	 * the rings that completed, the failing case will have
6355 	 * cleaned up for itself.  'q' failed, so its the terminus.
6356 	 */
6357 	rxq = ctx->ifc_rxqs;
6358 	for (i = 0; i < q; ++i, rxq++) {
6359 		tcp_lro_free(&rxq->ifr_lc);
6360 	}
6361 	return (err);
6362 #endif
6363 }
6364 
6365 /*********************************************************************
6366  *
6367  *  Free all receive rings.
6368  *
6369  **********************************************************************/
6370 static void
6371 iflib_rx_structures_free(if_ctx_t ctx)
6372 {
6373 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6374 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6375 	int i, j;
6376 
6377 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6378 		for (j = 0; j < sctx->isc_nrxqs; j++)
6379 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6380 		iflib_rx_sds_free(rxq);
6381 #if defined(INET6) || defined(INET)
6382 		tcp_lro_free(&rxq->ifr_lc);
6383 #endif
6384 	}
6385 	free(ctx->ifc_rxqs, M_IFLIB);
6386 	ctx->ifc_rxqs = NULL;
6387 }
6388 
6389 static int
6390 iflib_qset_structures_setup(if_ctx_t ctx)
6391 {
6392 	int err;
6393 
6394 	/*
6395 	 * It is expected that the caller takes care of freeing queues if this
6396 	 * fails.
6397 	 */
6398 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6399 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6400 		return (err);
6401 	}
6402 
6403 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6404 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6405 
6406 	return (err);
6407 }
6408 
6409 int
6410 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6411 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6412 {
6413 
6414 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6415 }
6416 
6417 /* Just to avoid copy/paste */
6418 static inline int
6419 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6420     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6421     const char *name)
6422 {
6423 	device_t dev;
6424 	unsigned int base_cpuid, cpuid;
6425 	int err;
6426 
6427 	dev = ctx->ifc_dev;
6428 	base_cpuid = ctx->ifc_sysctl_core_offset;
6429 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6430 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6431 	    irq ? irq->ii_res : NULL, name);
6432 	if (err) {
6433 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6434 		return (err);
6435 	}
6436 #ifdef notyet
6437 	if (cpuid > ctx->ifc_cpuid_highest)
6438 		ctx->ifc_cpuid_highest = cpuid;
6439 #endif
6440 	return (0);
6441 }
6442 
6443 int
6444 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6445 			iflib_intr_type_t type, driver_filter_t *filter,
6446 			void *filter_arg, int qid, const char *name)
6447 {
6448 	device_t dev;
6449 	struct grouptask *gtask;
6450 	struct taskqgroup *tqg;
6451 	iflib_filter_info_t info;
6452 	gtask_fn_t *fn;
6453 	int tqrid, err;
6454 	driver_filter_t *intr_fast;
6455 	void *q;
6456 
6457 	info = &ctx->ifc_filter_info;
6458 	tqrid = rid;
6459 
6460 	switch (type) {
6461 	/* XXX merge tx/rx for netmap? */
6462 	case IFLIB_INTR_TX:
6463 		q = &ctx->ifc_txqs[qid];
6464 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6465 		gtask = &ctx->ifc_txqs[qid].ift_task;
6466 		tqg = qgroup_if_io_tqg;
6467 		fn = _task_fn_tx;
6468 		intr_fast = iflib_fast_intr;
6469 		GROUPTASK_INIT(gtask, 0, fn, q);
6470 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6471 		break;
6472 	case IFLIB_INTR_RX:
6473 		q = &ctx->ifc_rxqs[qid];
6474 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6475 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6476 		tqg = qgroup_if_io_tqg;
6477 		fn = _task_fn_rx;
6478 		intr_fast = iflib_fast_intr;
6479 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6480 		break;
6481 	case IFLIB_INTR_RXTX:
6482 		q = &ctx->ifc_rxqs[qid];
6483 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6484 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6485 		tqg = qgroup_if_io_tqg;
6486 		fn = _task_fn_rx;
6487 		intr_fast = iflib_fast_intr_rxtx;
6488 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6489 		break;
6490 	case IFLIB_INTR_ADMIN:
6491 		q = ctx;
6492 		tqrid = -1;
6493 		info = &ctx->ifc_filter_info;
6494 		gtask = &ctx->ifc_admin_task;
6495 		tqg = qgroup_if_config_tqg;
6496 		fn = _task_fn_admin;
6497 		intr_fast = iflib_fast_intr_ctx;
6498 		break;
6499 	default:
6500 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6501 		    __func__);
6502 		return (EINVAL);
6503 	}
6504 
6505 	info->ifi_filter = filter;
6506 	info->ifi_filter_arg = filter_arg;
6507 	info->ifi_task = gtask;
6508 	info->ifi_ctx = q;
6509 
6510 	dev = ctx->ifc_dev;
6511 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6512 	if (err != 0) {
6513 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6514 		return (err);
6515 	}
6516 	if (type == IFLIB_INTR_ADMIN)
6517 		return (0);
6518 
6519 	if (tqrid != -1) {
6520 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6521 		    name);
6522 		if (err)
6523 			return (err);
6524 	} else {
6525 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6526 	}
6527 
6528 	return (0);
6529 }
6530 
6531 void
6532 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6533 {
6534 	device_t dev;
6535 	struct grouptask *gtask;
6536 	struct taskqgroup *tqg;
6537 	gtask_fn_t *fn;
6538 	void *q;
6539 	int err;
6540 
6541 	switch (type) {
6542 	case IFLIB_INTR_TX:
6543 		q = &ctx->ifc_txqs[qid];
6544 		gtask = &ctx->ifc_txqs[qid].ift_task;
6545 		tqg = qgroup_if_io_tqg;
6546 		fn = _task_fn_tx;
6547 		GROUPTASK_INIT(gtask, 0, fn, q);
6548 		break;
6549 	case IFLIB_INTR_RX:
6550 		q = &ctx->ifc_rxqs[qid];
6551 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6552 		tqg = qgroup_if_io_tqg;
6553 		fn = _task_fn_rx;
6554 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6555 		break;
6556 	case IFLIB_INTR_IOV:
6557 		q = ctx;
6558 		gtask = &ctx->ifc_vflr_task;
6559 		tqg = qgroup_if_config_tqg;
6560 		fn = _task_fn_iov;
6561 		GROUPTASK_INIT(gtask, 0, fn, q);
6562 		break;
6563 	default:
6564 		panic("unknown net intr type");
6565 	}
6566 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6567 	if (err) {
6568 		dev = ctx->ifc_dev;
6569 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6570 		    name);
6571 	}
6572 }
6573 
6574 void
6575 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6576 {
6577 
6578 	if (irq->ii_tag)
6579 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6580 
6581 	if (irq->ii_res)
6582 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6583 		    rman_get_rid(irq->ii_res), irq->ii_res);
6584 }
6585 
6586 static int
6587 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6588 {
6589 	iflib_txq_t txq = ctx->ifc_txqs;
6590 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6591 	if_irq_t irq = &ctx->ifc_legacy_irq;
6592 	iflib_filter_info_t info;
6593 	device_t dev;
6594 	struct grouptask *gtask;
6595 	struct resource *res;
6596 	struct taskqgroup *tqg;
6597 	void *q;
6598 	int err, tqrid;
6599 	bool rx_only;
6600 
6601 	q = &ctx->ifc_rxqs[0];
6602 	info = &rxq[0].ifr_filter_info;
6603 	gtask = &rxq[0].ifr_task;
6604 	tqg = qgroup_if_io_tqg;
6605 	tqrid = *rid;
6606 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6607 
6608 	ctx->ifc_flags |= IFC_LEGACY;
6609 	info->ifi_filter = filter;
6610 	info->ifi_filter_arg = filter_arg;
6611 	info->ifi_task = gtask;
6612 	info->ifi_ctx = rx_only ? ctx : q;
6613 
6614 	dev = ctx->ifc_dev;
6615 	/* We allocate a single interrupt resource */
6616 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6617 	    iflib_fast_intr_rxtx, NULL, info, name);
6618 	if (err != 0)
6619 		return (err);
6620 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6621 	res = irq->ii_res;
6622 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6623 
6624 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6625 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6626 	    "tx");
6627 	return (0);
6628 }
6629 
6630 void
6631 iflib_led_create(if_ctx_t ctx)
6632 {
6633 
6634 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6635 	    device_get_nameunit(ctx->ifc_dev));
6636 }
6637 
6638 void
6639 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6640 {
6641 
6642 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6643 }
6644 
6645 void
6646 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6647 {
6648 
6649 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6650 }
6651 
6652 void
6653 iflib_admin_intr_deferred(if_ctx_t ctx)
6654 {
6655 
6656 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6657 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6658 }
6659 
6660 void
6661 iflib_iov_intr_deferred(if_ctx_t ctx)
6662 {
6663 
6664 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6665 }
6666 
6667 void
6668 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6669 {
6670 
6671 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6672 	    name);
6673 }
6674 
6675 void
6676 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6677 	const char *name)
6678 {
6679 
6680 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6681 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6682 	    name);
6683 }
6684 
6685 void
6686 iflib_config_gtask_deinit(struct grouptask *gtask)
6687 {
6688 
6689 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6690 }
6691 
6692 void
6693 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6694 {
6695 	if_t ifp = ctx->ifc_ifp;
6696 	iflib_txq_t txq = ctx->ifc_txqs;
6697 
6698 	if_setbaudrate(ifp, baudrate);
6699 	if (baudrate >= IF_Gbps(10)) {
6700 		STATE_LOCK(ctx);
6701 		ctx->ifc_flags |= IFC_PREFETCH;
6702 		STATE_UNLOCK(ctx);
6703 	}
6704 	/* If link down, disable watchdog */
6705 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6706 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6707 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6708 	}
6709 	ctx->ifc_link_state = link_state;
6710 	if_link_state_change(ifp, link_state);
6711 }
6712 
6713 static int
6714 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6715 {
6716 	int credits;
6717 #ifdef INVARIANTS
6718 	int credits_pre = txq->ift_cidx_processed;
6719 #endif
6720 
6721 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6722 	    BUS_DMASYNC_POSTREAD);
6723 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6724 		return (0);
6725 
6726 	txq->ift_processed += credits;
6727 	txq->ift_cidx_processed += credits;
6728 
6729 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6730 	if (txq->ift_cidx_processed >= txq->ift_size)
6731 		txq->ift_cidx_processed -= txq->ift_size;
6732 	return (credits);
6733 }
6734 
6735 static int
6736 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6737 {
6738 	iflib_fl_t fl;
6739 	u_int i;
6740 
6741 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6742 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6743 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6744 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6745 	    budget));
6746 }
6747 
6748 void
6749 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6750 	const char *description, if_int_delay_info_t info,
6751 	int offset, int value)
6752 {
6753 	info->iidi_ctx = ctx;
6754 	info->iidi_offset = offset;
6755 	info->iidi_value = value;
6756 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6757 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6758 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6759 	    info, 0, iflib_sysctl_int_delay, "I", description);
6760 }
6761 
6762 struct sx *
6763 iflib_ctx_lock_get(if_ctx_t ctx)
6764 {
6765 
6766 	return (&ctx->ifc_ctx_sx);
6767 }
6768 
6769 static int
6770 iflib_msix_init(if_ctx_t ctx)
6771 {
6772 	device_t dev = ctx->ifc_dev;
6773 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6774 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6775 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6776 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6777 
6778 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6779 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6780 
6781 	if (bootverbose)
6782 		device_printf(dev, "msix_init qsets capped at %d\n",
6783 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6784 
6785 	/* Override by tuneable */
6786 	if (scctx->isc_disable_msix)
6787 		goto msi;
6788 
6789 	/* First try MSI-X */
6790 	if ((msgs = pci_msix_count(dev)) == 0) {
6791 		if (bootverbose)
6792 			device_printf(dev, "MSI-X not supported or disabled\n");
6793 		goto msi;
6794 	}
6795 
6796 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6797 	/*
6798 	 * bar == -1 => "trust me I know what I'm doing"
6799 	 * Some drivers are for hardware that is so shoddily
6800 	 * documented that no one knows which bars are which
6801 	 * so the developer has to map all bars. This hack
6802 	 * allows shoddy garbage to use MSI-X in this framework.
6803 	 */
6804 	if (bar != -1) {
6805 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6806 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6807 		if (ctx->ifc_msix_mem == NULL) {
6808 			device_printf(dev, "Unable to map MSI-X table\n");
6809 			goto msi;
6810 		}
6811 	}
6812 
6813 	admincnt = sctx->isc_admin_intrcnt;
6814 #if IFLIB_DEBUG
6815 	/* use only 1 qset in debug mode */
6816 	queuemsgs = min(msgs - admincnt, 1);
6817 #else
6818 	queuemsgs = msgs - admincnt;
6819 #endif
6820 #ifdef RSS
6821 	queues = imin(queuemsgs, rss_getnumbuckets());
6822 #else
6823 	queues = queuemsgs;
6824 #endif
6825 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6826 	if (bootverbose)
6827 		device_printf(dev,
6828 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6829 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6830 #ifdef  RSS
6831 	/* If we're doing RSS, clamp at the number of RSS buckets */
6832 	if (queues > rss_getnumbuckets())
6833 		queues = rss_getnumbuckets();
6834 #endif
6835 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6836 		rx_queues = iflib_num_rx_queues;
6837 	else
6838 		rx_queues = queues;
6839 
6840 	if (rx_queues > scctx->isc_nrxqsets)
6841 		rx_queues = scctx->isc_nrxqsets;
6842 
6843 	/*
6844 	 * We want this to be all logical CPUs by default
6845 	 */
6846 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6847 		tx_queues = iflib_num_tx_queues;
6848 	else
6849 		tx_queues = mp_ncpus;
6850 
6851 	if (tx_queues > scctx->isc_ntxqsets)
6852 		tx_queues = scctx->isc_ntxqsets;
6853 
6854 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6855 #ifdef INVARIANTS
6856 		if (tx_queues != rx_queues)
6857 			device_printf(dev,
6858 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6859 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6860 #endif
6861 		tx_queues = min(rx_queues, tx_queues);
6862 		rx_queues = min(rx_queues, tx_queues);
6863 	}
6864 
6865 	vectors = rx_queues + admincnt;
6866 	if (msgs < vectors) {
6867 		device_printf(dev,
6868 		    "insufficient number of MSI-X vectors "
6869 		    "(supported %d, need %d)\n", msgs, vectors);
6870 		goto msi;
6871 	}
6872 
6873 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6874 	    tx_queues);
6875 	msgs = vectors;
6876 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6877 		if (vectors != msgs) {
6878 			device_printf(dev,
6879 			    "Unable to allocate sufficient MSI-X vectors "
6880 			    "(got %d, need %d)\n", vectors, msgs);
6881 			pci_release_msi(dev);
6882 			if (bar != -1) {
6883 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6884 				    ctx->ifc_msix_mem);
6885 				ctx->ifc_msix_mem = NULL;
6886 			}
6887 			goto msi;
6888 		}
6889 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6890 		    vectors);
6891 		scctx->isc_vectors = vectors;
6892 		scctx->isc_nrxqsets = rx_queues;
6893 		scctx->isc_ntxqsets = tx_queues;
6894 		scctx->isc_intr = IFLIB_INTR_MSIX;
6895 
6896 		return (vectors);
6897 	} else {
6898 		device_printf(dev,
6899 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6900 		    err);
6901 		if (bar != -1) {
6902 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6903 			    ctx->ifc_msix_mem);
6904 			ctx->ifc_msix_mem = NULL;
6905 		}
6906 	}
6907 
6908 msi:
6909 	vectors = pci_msi_count(dev);
6910 	scctx->isc_nrxqsets = 1;
6911 	scctx->isc_ntxqsets = 1;
6912 	scctx->isc_vectors = vectors;
6913 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6914 		device_printf(dev,"Using an MSI interrupt\n");
6915 		scctx->isc_intr = IFLIB_INTR_MSI;
6916 	} else {
6917 		scctx->isc_vectors = 1;
6918 		device_printf(dev,"Using a Legacy interrupt\n");
6919 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6920 	}
6921 
6922 	return (vectors);
6923 }
6924 
6925 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6926 
6927 static int
6928 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6929 {
6930 	int rc;
6931 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6932 	struct sbuf *sb;
6933 	const char *ring_state = "UNKNOWN";
6934 
6935 	/* XXX needed ? */
6936 	rc = sysctl_wire_old_buffer(req, 0);
6937 	MPASS(rc == 0);
6938 	if (rc != 0)
6939 		return (rc);
6940 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6941 	MPASS(sb != NULL);
6942 	if (sb == NULL)
6943 		return (ENOMEM);
6944 	if (state[3] <= 3)
6945 		ring_state = ring_states[state[3]];
6946 
6947 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6948 		    state[0], state[1], state[2], ring_state);
6949 	rc = sbuf_finish(sb);
6950 	sbuf_delete(sb);
6951         return(rc);
6952 }
6953 
6954 enum iflib_ndesc_handler {
6955 	IFLIB_NTXD_HANDLER,
6956 	IFLIB_NRXD_HANDLER,
6957 };
6958 
6959 static int
6960 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6961 {
6962 	if_ctx_t ctx = (void *)arg1;
6963 	enum iflib_ndesc_handler type = arg2;
6964 	char buf[256] = {0};
6965 	qidx_t *ndesc;
6966 	char *p, *next;
6967 	int nqs, rc, i;
6968 
6969 	nqs = 8;
6970 	switch(type) {
6971 	case IFLIB_NTXD_HANDLER:
6972 		ndesc = ctx->ifc_sysctl_ntxds;
6973 		if (ctx->ifc_sctx)
6974 			nqs = ctx->ifc_sctx->isc_ntxqs;
6975 		break;
6976 	case IFLIB_NRXD_HANDLER:
6977 		ndesc = ctx->ifc_sysctl_nrxds;
6978 		if (ctx->ifc_sctx)
6979 			nqs = ctx->ifc_sctx->isc_nrxqs;
6980 		break;
6981 	default:
6982 		printf("%s: unhandled type\n", __func__);
6983 		return (EINVAL);
6984 	}
6985 	if (nqs == 0)
6986 		nqs = 8;
6987 
6988 	for (i=0; i<8; i++) {
6989 		if (i >= nqs)
6990 			break;
6991 		if (i)
6992 			strcat(buf, ",");
6993 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6994 	}
6995 
6996 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6997 	if (rc || req->newptr == NULL)
6998 		return rc;
6999 
7000 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
7001 	    i++, p = strsep(&next, " ,")) {
7002 		ndesc[i] = strtoul(p, NULL, 10);
7003 	}
7004 
7005 	return(rc);
7006 }
7007 
7008 #define NAME_BUFLEN 32
7009 static void
7010 iflib_add_device_sysctl_pre(if_ctx_t ctx)
7011 {
7012         device_t dev = iflib_get_dev(ctx);
7013 	struct sysctl_oid_list *child, *oid_list;
7014 	struct sysctl_ctx_list *ctx_list;
7015 	struct sysctl_oid *node;
7016 
7017 	ctx_list = device_get_sysctl_ctx(dev);
7018 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
7019 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
7020 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
7021 	oid_list = SYSCTL_CHILDREN(node);
7022 
7023 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
7024 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
7025 		       "driver version");
7026 
7027 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
7028 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
7029 			"# of txqs to use, 0 => use default #");
7030 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
7031 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
7032 			"# of rxqs to use, 0 => use default #");
7033 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
7034 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
7035                        "permit #txq != #rxq");
7036 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
7037                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
7038                       "disable MSI-X (default 0)");
7039 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
7040 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
7041 		       "set the RX budget");
7042 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
7043 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7044 		       "cause TX to abdicate instead of running to completion");
7045 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7046 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
7047 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7048 		       "offset to start using cores at");
7049 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
7050 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7051 		       "use separate cores for TX and RX");
7052 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
7053 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7054 		      "try to make use of logical cores for TX and RX");
7055 
7056 	/* XXX change for per-queue sizes */
7057 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
7058 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7059 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7060 	    "list of # of TX descriptors to use, 0 = use default #");
7061 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
7062 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7063 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7064 	    "list of # of RX descriptors to use, 0 = use default #");
7065 }
7066 
7067 static void
7068 iflib_add_device_sysctl_post(if_ctx_t ctx)
7069 {
7070 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7071 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7072         device_t dev = iflib_get_dev(ctx);
7073 	struct sysctl_oid_list *child;
7074 	struct sysctl_ctx_list *ctx_list;
7075 	iflib_fl_t fl;
7076 	iflib_txq_t txq;
7077 	iflib_rxq_t rxq;
7078 	int i, j;
7079 	char namebuf[NAME_BUFLEN];
7080 	char *qfmt;
7081 	struct sysctl_oid *queue_node, *fl_node, *node;
7082 	struct sysctl_oid_list *queue_list, *fl_list;
7083 	ctx_list = device_get_sysctl_ctx(dev);
7084 
7085 	node = ctx->ifc_sysctl_node;
7086 	child = SYSCTL_CHILDREN(node);
7087 
7088 	if (scctx->isc_ntxqsets > 100)
7089 		qfmt = "txq%03d";
7090 	else if (scctx->isc_ntxqsets > 10)
7091 		qfmt = "txq%02d";
7092 	else
7093 		qfmt = "txq%d";
7094 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7095 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7096 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7097 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7098 		queue_list = SYSCTL_CHILDREN(queue_node);
7099 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7100 			       CTLFLAG_RD,
7101 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
7102 #if MEMORY_LOGGING
7103 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7104 				CTLFLAG_RD,
7105 				&txq->ift_dequeued, "total mbufs freed");
7106 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7107 				CTLFLAG_RD,
7108 				&txq->ift_enqueued, "total mbufs enqueued");
7109 #endif
7110 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7111 				   CTLFLAG_RD,
7112 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
7113 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7114 				   CTLFLAG_RD,
7115 				   &txq->ift_pullups, "# of times m_pullup was called");
7116 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
7117 				   CTLFLAG_RD,
7118 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7119 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
7120 				   CTLFLAG_RD,
7121 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
7122 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
7123 				   CTLFLAG_RD,
7124 				   &txq->ift_map_failed, "# of times DMA map failed");
7125 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
7126 				   CTLFLAG_RD,
7127 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
7128 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
7129 				   CTLFLAG_RD,
7130 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
7131 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7132 				   CTLFLAG_RD,
7133 				   &txq->ift_pidx, 1, "Producer Index");
7134 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7135 				   CTLFLAG_RD,
7136 				   &txq->ift_cidx, 1, "Consumer Index");
7137 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
7138 				   CTLFLAG_RD,
7139 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
7140 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7141 				   CTLFLAG_RD,
7142 				   &txq->ift_in_use, 1, "descriptors in use");
7143 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
7144 				   CTLFLAG_RD,
7145 				   &txq->ift_processed, "descriptors procesed for clean");
7146 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7147 				   CTLFLAG_RD,
7148 				   &txq->ift_cleaned, "total cleaned");
7149 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7150 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7151 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7152 		    mp_ring_state_handler, "A", "soft ring state");
7153 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
7154 				       CTLFLAG_RD, &txq->ift_br->enqueues,
7155 				       "# of enqueues to the mp_ring for this queue");
7156 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
7157 				       CTLFLAG_RD, &txq->ift_br->drops,
7158 				       "# of drops in the mp_ring for this queue");
7159 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
7160 				       CTLFLAG_RD, &txq->ift_br->starts,
7161 				       "# of normal consumer starts in the mp_ring for this queue");
7162 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
7163 				       CTLFLAG_RD, &txq->ift_br->stalls,
7164 					       "# of consumer stalls in the mp_ring for this queue");
7165 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7166 			       CTLFLAG_RD, &txq->ift_br->restarts,
7167 				       "# of consumer restarts in the mp_ring for this queue");
7168 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7169 				       CTLFLAG_RD, &txq->ift_br->abdications,
7170 				       "# of consumer abdications in the mp_ring for this queue");
7171 	}
7172 
7173 	if (scctx->isc_nrxqsets > 100)
7174 		qfmt = "rxq%03d";
7175 	else if (scctx->isc_nrxqsets > 10)
7176 		qfmt = "rxq%02d";
7177 	else
7178 		qfmt = "rxq%d";
7179 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7180 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7181 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7182 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7183 		queue_list = SYSCTL_CHILDREN(queue_node);
7184 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7185 			       CTLFLAG_RD,
7186 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7187 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7188 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7189 				       CTLFLAG_RD,
7190 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7191 		}
7192 
7193 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7194 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7195 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7196 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7197 			fl_list = SYSCTL_CHILDREN(fl_node);
7198 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7199 				       CTLFLAG_RD,
7200 				       &fl->ifl_pidx, 1, "Producer Index");
7201 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7202 				       CTLFLAG_RD,
7203 				       &fl->ifl_cidx, 1, "Consumer Index");
7204 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7205 				       CTLFLAG_RD,
7206 				       &fl->ifl_credits, 1, "credits available");
7207 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7208 				       CTLFLAG_RD,
7209 				       &fl->ifl_buf_size, 1, "buffer size");
7210 #if MEMORY_LOGGING
7211 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7212 					CTLFLAG_RD,
7213 					&fl->ifl_m_enqueued, "mbufs allocated");
7214 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7215 					CTLFLAG_RD,
7216 					&fl->ifl_m_dequeued, "mbufs freed");
7217 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7218 					CTLFLAG_RD,
7219 					&fl->ifl_cl_enqueued, "clusters allocated");
7220 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7221 					CTLFLAG_RD,
7222 					&fl->ifl_cl_dequeued, "clusters freed");
7223 #endif
7224 		}
7225 	}
7226 
7227 }
7228 
7229 void
7230 iflib_request_reset(if_ctx_t ctx)
7231 {
7232 
7233 	STATE_LOCK(ctx);
7234 	ctx->ifc_flags |= IFC_DO_RESET;
7235 	STATE_UNLOCK(ctx);
7236 }
7237 
7238 #ifndef __NO_STRICT_ALIGNMENT
7239 static struct mbuf *
7240 iflib_fixup_rx(struct mbuf *m)
7241 {
7242 	struct mbuf *n;
7243 
7244 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7245 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7246 		m->m_data += ETHER_HDR_LEN;
7247 		n = m;
7248 	} else {
7249 		MGETHDR(n, M_NOWAIT, MT_DATA);
7250 		if (n == NULL) {
7251 			m_freem(m);
7252 			return (NULL);
7253 		}
7254 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7255 		m->m_data += ETHER_HDR_LEN;
7256 		m->m_len -= ETHER_HDR_LEN;
7257 		n->m_len = ETHER_HDR_LEN;
7258 		M_MOVE_PKTHDR(n, m);
7259 		n->m_next = m;
7260 	}
7261 	return (n);
7262 }
7263 #endif
7264 
7265 #ifdef DEBUGNET
7266 static void
7267 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7268 {
7269 	if_ctx_t ctx;
7270 
7271 	ctx = if_getsoftc(ifp);
7272 	CTX_LOCK(ctx);
7273 	*nrxr = NRXQSETS(ctx);
7274 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7275 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7276 	CTX_UNLOCK(ctx);
7277 }
7278 
7279 static void
7280 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7281 {
7282 	if_ctx_t ctx;
7283 	if_softc_ctx_t scctx;
7284 	iflib_fl_t fl;
7285 	iflib_rxq_t rxq;
7286 	int i, j;
7287 
7288 	ctx = if_getsoftc(ifp);
7289 	scctx = &ctx->ifc_softc_ctx;
7290 
7291 	switch (event) {
7292 	case DEBUGNET_START:
7293 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7294 			rxq = &ctx->ifc_rxqs[i];
7295 			for (j = 0; j < rxq->ifr_nfl; j++) {
7296 				fl = rxq->ifr_fl;
7297 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7298 			}
7299 		}
7300 		iflib_no_tx_batch = 1;
7301 		break;
7302 	default:
7303 		break;
7304 	}
7305 }
7306 
7307 static int
7308 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7309 {
7310 	if_ctx_t ctx;
7311 	iflib_txq_t txq;
7312 	int error;
7313 
7314 	ctx = if_getsoftc(ifp);
7315 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7316 	    IFF_DRV_RUNNING)
7317 		return (EBUSY);
7318 
7319 	txq = &ctx->ifc_txqs[0];
7320 	error = iflib_encap(txq, &m);
7321 	if (error == 0)
7322 		(void)iflib_txd_db_check(txq, true);
7323 	return (error);
7324 }
7325 
7326 static int
7327 iflib_debugnet_poll(if_t ifp, int count)
7328 {
7329 	struct epoch_tracker et;
7330 	if_ctx_t ctx;
7331 	if_softc_ctx_t scctx;
7332 	iflib_txq_t txq;
7333 	int i;
7334 
7335 	ctx = if_getsoftc(ifp);
7336 	scctx = &ctx->ifc_softc_ctx;
7337 
7338 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7339 	    IFF_DRV_RUNNING)
7340 		return (EBUSY);
7341 
7342 	txq = &ctx->ifc_txqs[0];
7343 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7344 
7345 	NET_EPOCH_ENTER(et);
7346 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7347 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7348 	NET_EPOCH_EXIT(et);
7349 	return (0);
7350 }
7351 #endif /* DEBUGNET */
7352