xref: /freebsd/sys/net/iflib.c (revision 13138422bc354a1ec35f53a27c4efeccdffc5639)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t	ift_buf_tag;
359 	bus_dma_tag_t	ift_tso_buf_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define MTX_NAME_LEN 16
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the command queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 	qidx_t ifsd_cidx;
455 } *if_rxsd_t;
456 
457 /* multiple of word size */
458 #ifdef __LP64__
459 #define PKT_INFO_SIZE	6
460 #define RXD_INFO_SIZE	5
461 #define PKT_TYPE uint64_t
462 #else
463 #define PKT_INFO_SIZE	11
464 #define RXD_INFO_SIZE	8
465 #define PKT_TYPE uint32_t
466 #endif
467 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
468 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
469 
470 typedef struct if_pkt_info_pad {
471 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
472 } *if_pkt_info_pad_t;
473 typedef struct if_rxd_info_pad {
474 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
475 } *if_rxd_info_pad_t;
476 
477 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
478 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
479 
480 
481 static inline void
482 pkt_info_zero(if_pkt_info_t pi)
483 {
484 	if_pkt_info_pad_t pi_pad;
485 
486 	pi_pad = (if_pkt_info_pad_t)pi;
487 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
488 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
489 #ifndef __LP64__
490 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
491 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
492 #endif
493 }
494 
495 static device_method_t iflib_pseudo_methods[] = {
496 	DEVMETHOD(device_attach, noop_attach),
497 	DEVMETHOD(device_detach, iflib_pseudo_detach),
498 	DEVMETHOD_END
499 };
500 
501 driver_t iflib_pseudodriver = {
502 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
503 };
504 
505 static inline void
506 rxd_info_zero(if_rxd_info_t ri)
507 {
508 	if_rxd_info_pad_t ri_pad;
509 	int i;
510 
511 	ri_pad = (if_rxd_info_pad_t)ri;
512 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
513 		ri_pad->rxd_val[i] = 0;
514 		ri_pad->rxd_val[i+1] = 0;
515 		ri_pad->rxd_val[i+2] = 0;
516 		ri_pad->rxd_val[i+3] = 0;
517 	}
518 #ifdef __LP64__
519 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
520 #endif
521 }
522 
523 /*
524  * Only allow a single packet to take up most 1/nth of the tx ring
525  */
526 #define MAX_SINGLE_PACKET_FRACTION 12
527 #define IF_BAD_DMA (bus_addr_t)-1
528 
529 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
530 
531 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
532 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
535 
536 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
537 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
538 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
539 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
540 
541 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
542 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
543 
544 void
545 iflib_set_detach(if_ctx_t ctx)
546 {
547 	STATE_LOCK(ctx);
548 	ctx->ifc_flags |= IFC_IN_DETACH;
549 	STATE_UNLOCK(ctx);
550 }
551 
552 /* Our boot-time initialization hook */
553 static int	iflib_module_event_handler(module_t, int, void *);
554 
555 static moduledata_t iflib_moduledata = {
556 	"iflib",
557 	iflib_module_event_handler,
558 	NULL
559 };
560 
561 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
562 MODULE_VERSION(iflib, 1);
563 
564 MODULE_DEPEND(iflib, pci, 1, 1, 1);
565 MODULE_DEPEND(iflib, ether, 1, 1, 1);
566 
567 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
568 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
569 
570 #ifndef IFLIB_DEBUG_COUNTERS
571 #ifdef INVARIANTS
572 #define IFLIB_DEBUG_COUNTERS 1
573 #else
574 #define IFLIB_DEBUG_COUNTERS 0
575 #endif /* !INVARIANTS */
576 #endif
577 
578 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
579                    "iflib driver parameters");
580 
581 /*
582  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
583  */
584 static int iflib_min_tx_latency = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
586 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
587 static int iflib_no_tx_batch = 0;
588 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
589 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
590 
591 
592 #if IFLIB_DEBUG_COUNTERS
593 
594 static int iflib_tx_seen;
595 static int iflib_tx_sent;
596 static int iflib_tx_encap;
597 static int iflib_rx_allocs;
598 static int iflib_fl_refills;
599 static int iflib_fl_refills_large;
600 static int iflib_tx_frees;
601 
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
603 		   &iflib_tx_seen, 0, "# TX mbufs seen");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
605 		   &iflib_tx_sent, 0, "# TX mbufs sent");
606 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
607 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
609 		   &iflib_tx_frees, 0, "# TX frees");
610 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
611 		   &iflib_rx_allocs, 0, "# RX allocations");
612 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
613 		   &iflib_fl_refills, 0, "# refills");
614 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
615 		   &iflib_fl_refills_large, 0, "# large refills");
616 
617 
618 static int iflib_txq_drain_flushing;
619 static int iflib_txq_drain_oactive;
620 static int iflib_txq_drain_notready;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
623 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
625 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
627 		   &iflib_txq_drain_notready, 0, "# drain notready");
628 
629 
630 static int iflib_encap_load_mbuf_fail;
631 static int iflib_encap_pad_mbuf_fail;
632 static int iflib_encap_txq_avail_fail;
633 static int iflib_encap_txd_encap_fail;
634 
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
636 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
638 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
640 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
642 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
643 
644 static int iflib_task_fn_rxs;
645 static int iflib_rx_intr_enables;
646 static int iflib_fast_intrs;
647 static int iflib_rx_unavail;
648 static int iflib_rx_ctx_inactive;
649 static int iflib_rx_if_input;
650 static int iflib_rxd_flush;
651 
652 static int iflib_verbose_debug;
653 
654 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
655 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
657 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
658 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
659 		   &iflib_fast_intrs, 0, "# fast_intr calls");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
661 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
663 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
665 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
667 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
668 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
669 		   &iflib_verbose_debug, 0, "enable verbose debugging");
670 
671 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
672 static void
673 iflib_debug_reset(void)
674 {
675 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
676 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
677 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
678 		iflib_txq_drain_notready =
679 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
680 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
681 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
682 		iflib_rx_unavail =
683 		iflib_rx_ctx_inactive = iflib_rx_if_input =
684 		iflib_rxd_flush = 0;
685 }
686 
687 #else
688 #define DBG_COUNTER_INC(name)
689 static void iflib_debug_reset(void) {}
690 #endif
691 
692 #define IFLIB_DEBUG 0
693 
694 static void iflib_tx_structures_free(if_ctx_t ctx);
695 static void iflib_rx_structures_free(if_ctx_t ctx);
696 static int iflib_queues_alloc(if_ctx_t ctx);
697 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
698 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
699 static int iflib_qset_structures_setup(if_ctx_t ctx);
700 static int iflib_msix_init(if_ctx_t ctx);
701 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
702 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
703 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
704 #ifdef ALTQ
705 static void iflib_altq_if_start(if_t ifp);
706 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
707 #endif
708 static int iflib_register(if_ctx_t);
709 static void iflib_deregister(if_ctx_t);
710 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 #ifdef DEV_NETMAP
737 #include <sys/selinfo.h>
738 #include <net/netmap.h>
739 #include <dev/netmap/netmap_kern.h>
740 
741 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
742 
743 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
744 
745 /*
746  * device-specific sysctl variables:
747  *
748  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
749  *	During regular operations the CRC is stripped, but on some
750  *	hardware reception of frames not multiple of 64 is slower,
751  *	so using crcstrip=0 helps in benchmarks.
752  *
753  * iflib_rx_miss, iflib_rx_miss_bufs:
754  *	count packets that might be missed due to lost interrupts.
755  */
756 SYSCTL_DECL(_dev_netmap);
757 /*
758  * The xl driver by default strips CRCs and we do not override it.
759  */
760 
761 int iflib_crcstrip = 1;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
763     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
764 
765 int iflib_rx_miss, iflib_rx_miss_bufs;
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
767     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
768 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
769     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
770 
771 /*
772  * Register/unregister. We are already under netmap lock.
773  * Only called on the first register or the last unregister.
774  */
775 static int
776 iflib_netmap_register(struct netmap_adapter *na, int onoff)
777 {
778 	if_t ifp = na->ifp;
779 	if_ctx_t ctx = ifp->if_softc;
780 	int status;
781 
782 	CTX_LOCK(ctx);
783 	IFDI_INTR_DISABLE(ctx);
784 
785 	/* Tell the stack that the interface is no longer active */
786 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
787 
788 	if (!CTX_IS_VF(ctx))
789 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
790 
791 	/* enable or disable flags and callbacks in na and ifp */
792 	if (onoff) {
793 		nm_set_native_flags(na);
794 	} else {
795 		nm_clear_native_flags(na);
796 	}
797 	iflib_stop(ctx);
798 	iflib_init_locked(ctx);
799 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
800 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
801 	if (status)
802 		nm_clear_native_flags(na);
803 	CTX_UNLOCK(ctx);
804 	return (status);
805 }
806 
807 static int
808 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
809 {
810 	struct netmap_adapter *na = kring->na;
811 	u_int const lim = kring->nkr_num_slots - 1;
812 	u_int head = kring->rhead;
813 	struct netmap_ring *ring = kring->ring;
814 	bus_dmamap_t *map;
815 	struct if_rxd_update iru;
816 	if_ctx_t ctx = rxq->ifr_ctx;
817 	iflib_fl_t fl = &rxq->ifr_fl[0];
818 	uint32_t refill_pidx, nic_i;
819 #if IFLIB_DEBUG_COUNTERS
820 	int rf_count = 0;
821 #endif
822 
823 	if (nm_i == head && __predict_true(!init))
824 		return 0;
825 	iru_init(&iru, rxq, 0 /* flid */);
826 	map = fl->ifl_sds.ifsd_map;
827 	refill_pidx = netmap_idx_k2n(kring, nm_i);
828 	/*
829 	 * IMPORTANT: we must leave one free slot in the ring,
830 	 * so move head back by one unit
831 	 */
832 	head = nm_prev(head, lim);
833 	nic_i = UINT_MAX;
834 	DBG_COUNTER_INC(fl_refills);
835 	while (nm_i != head) {
836 #if IFLIB_DEBUG_COUNTERS
837 		if (++rf_count == 9)
838 			DBG_COUNTER_INC(fl_refills_large);
839 #endif
840 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
841 			struct netmap_slot *slot = &ring->slot[nm_i];
842 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
843 			uint32_t nic_i_dma = refill_pidx;
844 			nic_i = netmap_idx_k2n(kring, nm_i);
845 
846 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
847 
848 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
849 			        return netmap_ring_reinit(kring);
850 
851 			fl->ifl_vm_addrs[tmp_pidx] = addr;
852 			if (__predict_false(init)) {
853 				netmap_load_map(na, fl->ifl_buf_tag,
854 				    map[nic_i], addr);
855 			} else if (slot->flags & NS_BUF_CHANGED) {
856 				/* buffer has changed, reload map */
857 				netmap_reload_map(na, fl->ifl_buf_tag,
858 				    map[nic_i], addr);
859 			}
860 			slot->flags &= ~NS_BUF_CHANGED;
861 
862 			nm_i = nm_next(nm_i, lim);
863 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
864 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
865 				continue;
866 
867 			iru.iru_pidx = refill_pidx;
868 			iru.iru_count = tmp_pidx+1;
869 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
870 			refill_pidx = nic_i;
871 			for (int n = 0; n < iru.iru_count; n++) {
872 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
873 						BUS_DMASYNC_PREREAD);
874 				/* XXX - change this to not use the netmap func*/
875 				nic_i_dma = nm_next(nic_i_dma, lim);
876 			}
877 		}
878 	}
879 	kring->nr_hwcur = head;
880 
881 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
882 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
883 	if (__predict_true(nic_i != UINT_MAX)) {
884 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
885 		DBG_COUNTER_INC(rxd_flush);
886 	}
887 	return (0);
888 }
889 
890 /*
891  * Reconcile kernel and user view of the transmit ring.
892  *
893  * All information is in the kring.
894  * Userspace wants to send packets up to the one before kring->rhead,
895  * kernel knows kring->nr_hwcur is the first unsent packet.
896  *
897  * Here we push packets out (as many as possible), and possibly
898  * reclaim buffers from previously completed transmission.
899  *
900  * The caller (netmap) guarantees that there is only one instance
901  * running at any time. Any interference with other driver
902  * methods should be handled by the individual drivers.
903  */
904 static int
905 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
906 {
907 	struct netmap_adapter *na = kring->na;
908 	if_t ifp = na->ifp;
909 	struct netmap_ring *ring = kring->ring;
910 	u_int nm_i;	/* index into the netmap kring */
911 	u_int nic_i;	/* index into the NIC ring */
912 	u_int n;
913 	u_int const lim = kring->nkr_num_slots - 1;
914 	u_int const head = kring->rhead;
915 	struct if_pkt_info pi;
916 
917 	/*
918 	 * interrupts on every tx packet are expensive so request
919 	 * them every half ring, or where NS_REPORT is set
920 	 */
921 	u_int report_frequency = kring->nkr_num_slots >> 1;
922 	/* device-specific */
923 	if_ctx_t ctx = ifp->if_softc;
924 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
925 
926 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
927 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
928 
929 	/*
930 	 * First part: process new packets to send.
931 	 * nm_i is the current index in the netmap kring,
932 	 * nic_i is the corresponding index in the NIC ring.
933 	 *
934 	 * If we have packets to send (nm_i != head)
935 	 * iterate over the netmap ring, fetch length and update
936 	 * the corresponding slot in the NIC ring. Some drivers also
937 	 * need to update the buffer's physical address in the NIC slot
938 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
939 	 *
940 	 * The netmap_reload_map() calls is especially expensive,
941 	 * even when (as in this case) the tag is 0, so do only
942 	 * when the buffer has actually changed.
943 	 *
944 	 * If possible do not set the report/intr bit on all slots,
945 	 * but only a few times per ring or when NS_REPORT is set.
946 	 *
947 	 * Finally, on 10G and faster drivers, it might be useful
948 	 * to prefetch the next slot and txr entry.
949 	 */
950 
951 	nm_i = kring->nr_hwcur;
952 	if (nm_i != head) {	/* we have new packets to send */
953 		pkt_info_zero(&pi);
954 		pi.ipi_segs = txq->ift_segs;
955 		pi.ipi_qsidx = kring->ring_id;
956 		nic_i = netmap_idx_k2n(kring, nm_i);
957 
958 		__builtin_prefetch(&ring->slot[nm_i]);
959 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
960 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
961 
962 		for (n = 0; nm_i != head; n++) {
963 			struct netmap_slot *slot = &ring->slot[nm_i];
964 			u_int len = slot->len;
965 			uint64_t paddr;
966 			void *addr = PNMB(na, slot, &paddr);
967 			int flags = (slot->flags & NS_REPORT ||
968 				nic_i == 0 || nic_i == report_frequency) ?
969 				IPI_TX_INTR : 0;
970 
971 			/* device-specific */
972 			pi.ipi_len = len;
973 			pi.ipi_segs[0].ds_addr = paddr;
974 			pi.ipi_segs[0].ds_len = len;
975 			pi.ipi_nsegs = 1;
976 			pi.ipi_ndescs = 0;
977 			pi.ipi_pidx = nic_i;
978 			pi.ipi_flags = flags;
979 
980 			/* Fill the slot in the NIC ring. */
981 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
982 			DBG_COUNTER_INC(tx_encap);
983 
984 			/* prefetch for next round */
985 			__builtin_prefetch(&ring->slot[nm_i + 1]);
986 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
987 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
988 
989 			NM_CHECK_ADDR_LEN(na, addr, len);
990 
991 			if (slot->flags & NS_BUF_CHANGED) {
992 				/* buffer has changed, reload map */
993 				netmap_reload_map(na, txq->ift_buf_tag,
994 				    txq->ift_sds.ifsd_map[nic_i], addr);
995 			}
996 			/* make sure changes to the buffer are synced */
997 			bus_dmamap_sync(txq->ift_buf_tag,
998 			    txq->ift_sds.ifsd_map[nic_i],
999 			    BUS_DMASYNC_PREWRITE);
1000 
1001 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1002 			nm_i = nm_next(nm_i, lim);
1003 			nic_i = nm_next(nic_i, lim);
1004 		}
1005 		kring->nr_hwcur = nm_i;
1006 
1007 		/* synchronize the NIC ring */
1008 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1009 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1010 
1011 		/* (re)start the tx unit up to slot nic_i (excluded) */
1012 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1013 	}
1014 
1015 	/*
1016 	 * Second part: reclaim buffers for completed transmissions.
1017 	 *
1018 	 * If there are unclaimed buffers, attempt to reclaim them.
1019 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1020 	 * minimal delay, then trigger the tx handler which will spin in the
1021 	 * group task queue.
1022 	 */
1023 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1024 		if (iflib_tx_credits_update(ctx, txq)) {
1025 			/* some tx completed, increment avail */
1026 			nic_i = txq->ift_cidx_processed;
1027 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1028 		}
1029 	}
1030 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1031 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1032 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1033 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1034 	}
1035 	return (0);
1036 }
1037 
1038 /*
1039  * Reconcile kernel and user view of the receive ring.
1040  * Same as for the txsync, this routine must be efficient.
1041  * The caller guarantees a single invocations, but races against
1042  * the rest of the driver should be handled here.
1043  *
1044  * On call, kring->rhead is the first packet that userspace wants
1045  * to keep, and kring->rcur is the wakeup point.
1046  * The kernel has previously reported packets up to kring->rtail.
1047  *
1048  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1049  * of whether or not we received an interrupt.
1050  */
1051 static int
1052 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1053 {
1054 	struct netmap_adapter *na = kring->na;
1055 	struct netmap_ring *ring = kring->ring;
1056 	if_t ifp = na->ifp;
1057 	iflib_fl_t fl;
1058 	uint32_t nm_i;	/* index into the netmap ring */
1059 	uint32_t nic_i;	/* index into the NIC ring */
1060 	u_int i, n;
1061 	u_int const lim = kring->nkr_num_slots - 1;
1062 	u_int const head = kring->rhead;
1063 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1064 	struct if_rxd_info ri;
1065 
1066 	if_ctx_t ctx = ifp->if_softc;
1067 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1068 	if (head > lim)
1069 		return netmap_ring_reinit(kring);
1070 
1071 	/*
1072 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1073 	 */
1074 
1075 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1076 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1077 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1078 	}
1079 
1080 	/*
1081 	 * First part: import newly received packets.
1082 	 *
1083 	 * nm_i is the index of the next free slot in the netmap ring,
1084 	 * nic_i is the index of the next received packet in the NIC ring,
1085 	 * and they may differ in case if_init() has been called while
1086 	 * in netmap mode. For the receive ring we have
1087 	 *
1088 	 *	nic_i = rxr->next_check;
1089 	 *	nm_i = kring->nr_hwtail (previous)
1090 	 * and
1091 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1092 	 *
1093 	 * rxr->next_check is set to 0 on a ring reinit
1094 	 */
1095 	if (netmap_no_pendintr || force_update) {
1096 		int crclen = iflib_crcstrip ? 0 : 4;
1097 		int error, avail;
1098 
1099 		for (i = 0; i < rxq->ifr_nfl; i++) {
1100 			fl = &rxq->ifr_fl[i];
1101 			nic_i = fl->ifl_cidx;
1102 			nm_i = netmap_idx_n2k(kring, nic_i);
1103 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1104 			    rxq->ifr_id, nic_i, USHRT_MAX);
1105 			for (n = 0; avail > 0; n++, avail--) {
1106 				rxd_info_zero(&ri);
1107 				ri.iri_frags = rxq->ifr_frags;
1108 				ri.iri_qsidx = kring->ring_id;
1109 				ri.iri_ifp = ctx->ifc_ifp;
1110 				ri.iri_cidx = nic_i;
1111 
1112 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1113 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1114 				ring->slot[nm_i].flags = 0;
1115 				bus_dmamap_sync(fl->ifl_buf_tag,
1116 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1117 				nm_i = nm_next(nm_i, lim);
1118 				nic_i = nm_next(nic_i, lim);
1119 			}
1120 			if (n) { /* update the state variables */
1121 				if (netmap_no_pendintr && !force_update) {
1122 					/* diagnostics */
1123 					iflib_rx_miss ++;
1124 					iflib_rx_miss_bufs += n;
1125 				}
1126 				fl->ifl_cidx = nic_i;
1127 				kring->nr_hwtail = nm_i;
1128 			}
1129 			kring->nr_kflags &= ~NKR_PENDINTR;
1130 		}
1131 	}
1132 	/*
1133 	 * Second part: skip past packets that userspace has released.
1134 	 * (kring->nr_hwcur to head excluded),
1135 	 * and make the buffers available for reception.
1136 	 * As usual nm_i is the index in the netmap ring,
1137 	 * nic_i is the index in the NIC ring, and
1138 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1139 	 */
1140 	/* XXX not sure how this will work with multiple free lists */
1141 	nm_i = kring->nr_hwcur;
1142 
1143 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1144 }
1145 
1146 static void
1147 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1148 {
1149 	if_ctx_t ctx = na->ifp->if_softc;
1150 
1151 	CTX_LOCK(ctx);
1152 	if (onoff) {
1153 		IFDI_INTR_ENABLE(ctx);
1154 	} else {
1155 		IFDI_INTR_DISABLE(ctx);
1156 	}
1157 	CTX_UNLOCK(ctx);
1158 }
1159 
1160 
1161 static int
1162 iflib_netmap_attach(if_ctx_t ctx)
1163 {
1164 	struct netmap_adapter na;
1165 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1166 
1167 	bzero(&na, sizeof(na));
1168 
1169 	na.ifp = ctx->ifc_ifp;
1170 	na.na_flags = NAF_BDG_MAYSLEEP;
1171 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1172 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1173 
1174 	na.num_tx_desc = scctx->isc_ntxd[0];
1175 	na.num_rx_desc = scctx->isc_nrxd[0];
1176 	na.nm_txsync = iflib_netmap_txsync;
1177 	na.nm_rxsync = iflib_netmap_rxsync;
1178 	na.nm_register = iflib_netmap_register;
1179 	na.nm_intr = iflib_netmap_intr;
1180 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1181 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1182 	return (netmap_attach(&na));
1183 }
1184 
1185 static void
1186 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1187 {
1188 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1189 	struct netmap_slot *slot;
1190 
1191 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1192 	if (slot == NULL)
1193 		return;
1194 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1195 
1196 		/*
1197 		 * In netmap mode, set the map for the packet buffer.
1198 		 * NOTE: Some drivers (not this one) also need to set
1199 		 * the physical buffer address in the NIC ring.
1200 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1201 		 * netmap slot index, si
1202 		 */
1203 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1204 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1205 		    NMB(na, slot + si));
1206 	}
1207 }
1208 
1209 static void
1210 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1211 {
1212 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1213 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1214 	struct netmap_slot *slot;
1215 	uint32_t nm_i;
1216 
1217 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1218 	if (slot == NULL)
1219 		return;
1220 	nm_i = netmap_idx_n2k(kring, 0);
1221 	netmap_fl_refill(rxq, kring, nm_i, true);
1222 }
1223 
1224 static void
1225 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1226 {
1227 	struct netmap_kring *kring;
1228 	uint16_t txqid;
1229 
1230 	txqid = txq->ift_id;
1231 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1232 
1233 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1234 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1235 		    BUS_DMASYNC_POSTREAD);
1236 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1237 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1238 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1239 			if (hz < 2000)
1240 				*reset_on = 1;
1241 			else
1242 				*reset_on = hz / 1000;
1243 		}
1244 	}
1245 }
1246 
1247 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1248 
1249 #else
1250 #define iflib_netmap_txq_init(ctx, txq)
1251 #define iflib_netmap_rxq_init(ctx, rxq)
1252 #define iflib_netmap_detach(ifp)
1253 
1254 #define iflib_netmap_attach(ctx) (0)
1255 #define netmap_rx_irq(ifp, qid, budget) (0)
1256 #define netmap_tx_irq(ifp, qid) do {} while (0)
1257 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1258 #endif
1259 
1260 #if defined(__i386__) || defined(__amd64__)
1261 static __inline void
1262 prefetch(void *x)
1263 {
1264 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1265 }
1266 static __inline void
1267 prefetch2cachelines(void *x)
1268 {
1269 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1270 #if (CACHE_LINE_SIZE < 128)
1271 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1272 #endif
1273 }
1274 #else
1275 #define prefetch(x)
1276 #define prefetch2cachelines(x)
1277 #endif
1278 
1279 static void
1280 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1281 {
1282 	iflib_fl_t fl;
1283 
1284 	fl = &rxq->ifr_fl[flid];
1285 	iru->iru_paddrs = fl->ifl_bus_addrs;
1286 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1287 	iru->iru_idxs = fl->ifl_rxd_idxs;
1288 	iru->iru_qsidx = rxq->ifr_id;
1289 	iru->iru_buf_size = fl->ifl_buf_size;
1290 	iru->iru_flidx = fl->ifl_id;
1291 }
1292 
1293 static void
1294 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1295 {
1296 	if (err)
1297 		return;
1298 	*(bus_addr_t *) arg = segs[0].ds_addr;
1299 }
1300 
1301 int
1302 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1303 {
1304 	int err;
1305 	device_t dev = ctx->ifc_dev;
1306 
1307 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1308 				align, 0,		/* alignment, bounds */
1309 				BUS_SPACE_MAXADDR,	/* lowaddr */
1310 				BUS_SPACE_MAXADDR,	/* highaddr */
1311 				NULL, NULL,		/* filter, filterarg */
1312 				size,			/* maxsize */
1313 				1,			/* nsegments */
1314 				size,			/* maxsegsize */
1315 				BUS_DMA_ALLOCNOW,	/* flags */
1316 				NULL,			/* lockfunc */
1317 				NULL,			/* lockarg */
1318 				&dma->idi_tag);
1319 	if (err) {
1320 		device_printf(dev,
1321 		    "%s: bus_dma_tag_create failed: %d\n",
1322 		    __func__, err);
1323 		goto fail_0;
1324 	}
1325 
1326 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1327 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1328 	if (err) {
1329 		device_printf(dev,
1330 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1331 		    __func__, (uintmax_t)size, err);
1332 		goto fail_1;
1333 	}
1334 
1335 	dma->idi_paddr = IF_BAD_DMA;
1336 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1337 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1338 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1339 		device_printf(dev,
1340 		    "%s: bus_dmamap_load failed: %d\n",
1341 		    __func__, err);
1342 		goto fail_2;
1343 	}
1344 
1345 	dma->idi_size = size;
1346 	return (0);
1347 
1348 fail_2:
1349 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1350 fail_1:
1351 	bus_dma_tag_destroy(dma->idi_tag);
1352 fail_0:
1353 	dma->idi_tag = NULL;
1354 
1355 	return (err);
1356 }
1357 
1358 int
1359 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1360 {
1361 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1362 
1363 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1364 
1365 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1366 }
1367 
1368 int
1369 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1370 {
1371 	int i, err;
1372 	iflib_dma_info_t *dmaiter;
1373 
1374 	dmaiter = dmalist;
1375 	for (i = 0; i < count; i++, dmaiter++) {
1376 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1377 			break;
1378 	}
1379 	if (err)
1380 		iflib_dma_free_multi(dmalist, i);
1381 	return (err);
1382 }
1383 
1384 void
1385 iflib_dma_free(iflib_dma_info_t dma)
1386 {
1387 	if (dma->idi_tag == NULL)
1388 		return;
1389 	if (dma->idi_paddr != IF_BAD_DMA) {
1390 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1391 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1392 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1393 		dma->idi_paddr = IF_BAD_DMA;
1394 	}
1395 	if (dma->idi_vaddr != NULL) {
1396 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1397 		dma->idi_vaddr = NULL;
1398 	}
1399 	bus_dma_tag_destroy(dma->idi_tag);
1400 	dma->idi_tag = NULL;
1401 }
1402 
1403 void
1404 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1405 {
1406 	int i;
1407 	iflib_dma_info_t *dmaiter = dmalist;
1408 
1409 	for (i = 0; i < count; i++, dmaiter++)
1410 		iflib_dma_free(*dmaiter);
1411 }
1412 
1413 #ifdef EARLY_AP_STARTUP
1414 static const int iflib_started = 1;
1415 #else
1416 /*
1417  * We used to abuse the smp_started flag to decide if the queues have been
1418  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1419  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1420  * is set.  Run a SYSINIT() strictly after that to just set a usable
1421  * completion flag.
1422  */
1423 
1424 static int iflib_started;
1425 
1426 static void
1427 iflib_record_started(void *arg)
1428 {
1429 	iflib_started = 1;
1430 }
1431 
1432 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1433 	iflib_record_started, NULL);
1434 #endif
1435 
1436 static int
1437 iflib_fast_intr(void *arg)
1438 {
1439 	iflib_filter_info_t info = arg;
1440 	struct grouptask *gtask = info->ifi_task;
1441 	int result;
1442 
1443 	if (!iflib_started)
1444 		return (FILTER_STRAY);
1445 
1446 	DBG_COUNTER_INC(fast_intrs);
1447 	if (info->ifi_filter != NULL) {
1448 		result = info->ifi_filter(info->ifi_filter_arg);
1449 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1450 			return (result);
1451 	}
1452 
1453 	GROUPTASK_ENQUEUE(gtask);
1454 	return (FILTER_HANDLED);
1455 }
1456 
1457 static int
1458 iflib_fast_intr_rxtx(void *arg)
1459 {
1460 	iflib_filter_info_t info = arg;
1461 	struct grouptask *gtask = info->ifi_task;
1462 	if_ctx_t ctx;
1463 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1464 	iflib_txq_t txq;
1465 	void *sc;
1466 	int i, cidx, result;
1467 	qidx_t txqid;
1468 	bool intr_enable, intr_legacy;
1469 
1470 	if (!iflib_started)
1471 		return (FILTER_STRAY);
1472 
1473 	DBG_COUNTER_INC(fast_intrs);
1474 	if (info->ifi_filter != NULL) {
1475 		result = info->ifi_filter(info->ifi_filter_arg);
1476 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1477 			return (result);
1478 	}
1479 
1480 	ctx = rxq->ifr_ctx;
1481 	sc = ctx->ifc_softc;
1482 	intr_enable = false;
1483 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1484 	MPASS(rxq->ifr_ntxqirq);
1485 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1486 		txqid = rxq->ifr_txqid[i];
1487 		txq = &ctx->ifc_txqs[txqid];
1488 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1489 		    BUS_DMASYNC_POSTREAD);
1490 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1491 			if (intr_legacy)
1492 				intr_enable = true;
1493 			else
1494 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1495 			continue;
1496 		}
1497 		GROUPTASK_ENQUEUE(&txq->ift_task);
1498 	}
1499 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1500 		cidx = rxq->ifr_cq_cidx;
1501 	else
1502 		cidx = rxq->ifr_fl[0].ifl_cidx;
1503 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1504 		GROUPTASK_ENQUEUE(gtask);
1505 	else {
1506 		if (intr_legacy)
1507 			intr_enable = true;
1508 		else
1509 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1510 		DBG_COUNTER_INC(rx_intr_enables);
1511 	}
1512 	if (intr_enable)
1513 		IFDI_INTR_ENABLE(ctx);
1514 	return (FILTER_HANDLED);
1515 }
1516 
1517 
1518 static int
1519 iflib_fast_intr_ctx(void *arg)
1520 {
1521 	iflib_filter_info_t info = arg;
1522 	struct grouptask *gtask = info->ifi_task;
1523 	int result;
1524 
1525 	if (!iflib_started)
1526 		return (FILTER_STRAY);
1527 
1528 	DBG_COUNTER_INC(fast_intrs);
1529 	if (info->ifi_filter != NULL) {
1530 		result = info->ifi_filter(info->ifi_filter_arg);
1531 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1532 			return (result);
1533 	}
1534 
1535 	GROUPTASK_ENQUEUE(gtask);
1536 	return (FILTER_HANDLED);
1537 }
1538 
1539 static int
1540 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1541 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1542 		 const char *name)
1543 {
1544 	struct resource *res;
1545 	void *tag = NULL;
1546 	device_t dev = ctx->ifc_dev;
1547 	int flags, i, rc;
1548 
1549 	flags = RF_ACTIVE;
1550 	if (ctx->ifc_flags & IFC_LEGACY)
1551 		flags |= RF_SHAREABLE;
1552 	MPASS(rid < 512);
1553 	i = rid;
1554 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1555 	if (res == NULL) {
1556 		device_printf(dev,
1557 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1558 		return (ENOMEM);
1559 	}
1560 	irq->ii_res = res;
1561 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1562 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1563 						filter, handler, arg, &tag);
1564 	if (rc != 0) {
1565 		device_printf(dev,
1566 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1567 					  rid, name ? name : "unknown", rc);
1568 		return (rc);
1569 	} else if (name)
1570 		bus_describe_intr(dev, res, tag, "%s", name);
1571 
1572 	irq->ii_tag = tag;
1573 	return (0);
1574 }
1575 
1576 /*********************************************************************
1577  *
1578  *  Allocate DMA resources for TX buffers as well as memory for the TX
1579  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1580  *  iflib_sw_tx_desc_array structure, storing all the information that
1581  *  is needed to transmit a packet on the wire.  This is called only
1582  *  once at attach, setup is done every reset.
1583  *
1584  **********************************************************************/
1585 static int
1586 iflib_txsd_alloc(iflib_txq_t txq)
1587 {
1588 	if_ctx_t ctx = txq->ift_ctx;
1589 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1590 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1591 	device_t dev = ctx->ifc_dev;
1592 	bus_size_t tsomaxsize;
1593 	int err, nsegments, ntsosegments;
1594 	bool tso;
1595 
1596 	nsegments = scctx->isc_tx_nsegments;
1597 	ntsosegments = scctx->isc_tx_tso_segments_max;
1598 	tsomaxsize = scctx->isc_tx_tso_size_max;
1599 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1600 		tsomaxsize += sizeof(struct ether_vlan_header);
1601 	MPASS(scctx->isc_ntxd[0] > 0);
1602 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1603 	MPASS(nsegments > 0);
1604 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1605 		MPASS(ntsosegments > 0);
1606 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1607 	}
1608 
1609 	/*
1610 	 * Set up DMA tags for TX buffers.
1611 	 */
1612 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1613 			       1, 0,			/* alignment, bounds */
1614 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1615 			       BUS_SPACE_MAXADDR,	/* highaddr */
1616 			       NULL, NULL,		/* filter, filterarg */
1617 			       sctx->isc_tx_maxsize,		/* maxsize */
1618 			       nsegments,	/* nsegments */
1619 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1620 			       0,			/* flags */
1621 			       NULL,			/* lockfunc */
1622 			       NULL,			/* lockfuncarg */
1623 			       &txq->ift_buf_tag))) {
1624 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1625 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1626 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1627 		goto fail;
1628 	}
1629 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1630 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1631 			       1, 0,			/* alignment, bounds */
1632 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1633 			       BUS_SPACE_MAXADDR,	/* highaddr */
1634 			       NULL, NULL,		/* filter, filterarg */
1635 			       tsomaxsize,		/* maxsize */
1636 			       ntsosegments,	/* nsegments */
1637 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1638 			       0,			/* flags */
1639 			       NULL,			/* lockfunc */
1640 			       NULL,			/* lockfuncarg */
1641 			       &txq->ift_tso_buf_tag))) {
1642 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1643 		    err);
1644 		goto fail;
1645 	}
1646 
1647 	/* Allocate memory for the TX mbuf map. */
1648 	if (!(txq->ift_sds.ifsd_m =
1649 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1650 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1651 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1652 		err = ENOMEM;
1653 		goto fail;
1654 	}
1655 
1656 	/*
1657 	 * Create the DMA maps for TX buffers.
1658 	 */
1659 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1660 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1661 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1662 		device_printf(dev,
1663 		    "Unable to allocate TX buffer DMA map memory\n");
1664 		err = ENOMEM;
1665 		goto fail;
1666 	}
1667 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1668 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1669 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1670 		device_printf(dev,
1671 		    "Unable to allocate TSO TX buffer map memory\n");
1672 		err = ENOMEM;
1673 		goto fail;
1674 	}
1675 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1676 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1677 		    &txq->ift_sds.ifsd_map[i]);
1678 		if (err != 0) {
1679 			device_printf(dev, "Unable to create TX DMA map\n");
1680 			goto fail;
1681 		}
1682 		if (!tso)
1683 			continue;
1684 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1685 		    &txq->ift_sds.ifsd_tso_map[i]);
1686 		if (err != 0) {
1687 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1688 			goto fail;
1689 		}
1690 	}
1691 	return (0);
1692 fail:
1693 	/* We free all, it handles case where we are in the middle */
1694 	iflib_tx_structures_free(ctx);
1695 	return (err);
1696 }
1697 
1698 static void
1699 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1700 {
1701 	bus_dmamap_t map;
1702 
1703 	if (txq->ift_sds.ifsd_map != NULL) {
1704 		map = txq->ift_sds.ifsd_map[i];
1705 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1706 		bus_dmamap_unload(txq->ift_buf_tag, map);
1707 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1708 		txq->ift_sds.ifsd_map[i] = NULL;
1709 	}
1710 
1711 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1712 		map = txq->ift_sds.ifsd_tso_map[i];
1713 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1714 		    BUS_DMASYNC_POSTWRITE);
1715 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1716 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1717 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1718 	}
1719 }
1720 
1721 static void
1722 iflib_txq_destroy(iflib_txq_t txq)
1723 {
1724 	if_ctx_t ctx = txq->ift_ctx;
1725 
1726 	for (int i = 0; i < txq->ift_size; i++)
1727 		iflib_txsd_destroy(ctx, txq, i);
1728 
1729 	if (txq->ift_br != NULL) {
1730 		ifmp_ring_free(txq->ift_br);
1731 		txq->ift_br = NULL;
1732 	}
1733 
1734 	mtx_destroy(&txq->ift_mtx);
1735 
1736 	if (txq->ift_sds.ifsd_map != NULL) {
1737 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1738 		txq->ift_sds.ifsd_map = NULL;
1739 	}
1740 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1741 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1742 		txq->ift_sds.ifsd_tso_map = NULL;
1743 	}
1744 	if (txq->ift_sds.ifsd_m != NULL) {
1745 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1746 		txq->ift_sds.ifsd_m = NULL;
1747 	}
1748 	if (txq->ift_buf_tag != NULL) {
1749 		bus_dma_tag_destroy(txq->ift_buf_tag);
1750 		txq->ift_buf_tag = NULL;
1751 	}
1752 	if (txq->ift_tso_buf_tag != NULL) {
1753 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1754 		txq->ift_tso_buf_tag = NULL;
1755 	}
1756 	if (txq->ift_ifdi != NULL) {
1757 		free(txq->ift_ifdi, M_IFLIB);
1758 	}
1759 }
1760 
1761 static void
1762 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1763 {
1764 	struct mbuf **mp;
1765 
1766 	mp = &txq->ift_sds.ifsd_m[i];
1767 	if (*mp == NULL)
1768 		return;
1769 
1770 	if (txq->ift_sds.ifsd_map != NULL) {
1771 		bus_dmamap_sync(txq->ift_buf_tag,
1772 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1773 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1774 	}
1775 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1776 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1777 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1778 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1779 		    txq->ift_sds.ifsd_tso_map[i]);
1780 	}
1781 	m_free(*mp);
1782 	DBG_COUNTER_INC(tx_frees);
1783 	*mp = NULL;
1784 }
1785 
1786 static int
1787 iflib_txq_setup(iflib_txq_t txq)
1788 {
1789 	if_ctx_t ctx = txq->ift_ctx;
1790 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1791 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1792 	iflib_dma_info_t di;
1793 	int i;
1794 
1795 	/* Set number of descriptors available */
1796 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1797 	/* XXX make configurable */
1798 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1799 
1800 	/* Reset indices */
1801 	txq->ift_cidx_processed = 0;
1802 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1803 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1804 
1805 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1806 		bzero((void *)di->idi_vaddr, di->idi_size);
1807 
1808 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1809 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1810 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1811 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1812 	return (0);
1813 }
1814 
1815 /*********************************************************************
1816  *
1817  *  Allocate DMA resources for RX buffers as well as memory for the RX
1818  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1819  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1820  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1821  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1822  *  packet, the maximum number of entries we'll need is equal to the
1823  *  number of hardware receive descriptors that we've allocated.
1824  *
1825  **********************************************************************/
1826 static int
1827 iflib_rxsd_alloc(iflib_rxq_t rxq)
1828 {
1829 	if_ctx_t ctx = rxq->ifr_ctx;
1830 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1831 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1832 	device_t dev = ctx->ifc_dev;
1833 	iflib_fl_t fl;
1834 	int			err;
1835 
1836 	MPASS(scctx->isc_nrxd[0] > 0);
1837 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1838 
1839 	fl = rxq->ifr_fl;
1840 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1841 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1842 		/* Set up DMA tag for RX buffers. */
1843 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1844 					 1, 0,			/* alignment, bounds */
1845 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1846 					 BUS_SPACE_MAXADDR,	/* highaddr */
1847 					 NULL, NULL,		/* filter, filterarg */
1848 					 sctx->isc_rx_maxsize,	/* maxsize */
1849 					 sctx->isc_rx_nsegments,	/* nsegments */
1850 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1851 					 0,			/* flags */
1852 					 NULL,			/* lockfunc */
1853 					 NULL,			/* lockarg */
1854 					 &fl->ifl_buf_tag);
1855 		if (err) {
1856 			device_printf(dev,
1857 			    "Unable to allocate RX DMA tag: %d\n", err);
1858 			goto fail;
1859 		}
1860 
1861 		/* Allocate memory for the RX mbuf map. */
1862 		if (!(fl->ifl_sds.ifsd_m =
1863 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1864 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1865 			device_printf(dev,
1866 			    "Unable to allocate RX mbuf map memory\n");
1867 			err = ENOMEM;
1868 			goto fail;
1869 		}
1870 
1871 		/* Allocate memory for the direct RX cluster pointer map. */
1872 		if (!(fl->ifl_sds.ifsd_cl =
1873 		      (caddr_t *) malloc(sizeof(caddr_t) *
1874 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1875 			device_printf(dev,
1876 			    "Unable to allocate RX cluster map memory\n");
1877 			err = ENOMEM;
1878 			goto fail;
1879 		}
1880 
1881 		/* Allocate memory for the RX cluster bus address map. */
1882 		if (!(fl->ifl_sds.ifsd_ba =
1883 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1884 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1885 			device_printf(dev,
1886 			    "Unable to allocate RX bus address map memory\n");
1887 			err = ENOMEM;
1888 			goto fail;
1889 		}
1890 
1891 		/*
1892 		 * Create the DMA maps for RX buffers.
1893 		 */
1894 		if (!(fl->ifl_sds.ifsd_map =
1895 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1896 			device_printf(dev,
1897 			    "Unable to allocate RX buffer DMA map memory\n");
1898 			err = ENOMEM;
1899 			goto fail;
1900 		}
1901 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1902 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1903 			    &fl->ifl_sds.ifsd_map[i]);
1904 			if (err != 0) {
1905 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1906 				goto fail;
1907 			}
1908 		}
1909 	}
1910 	return (0);
1911 
1912 fail:
1913 	iflib_rx_structures_free(ctx);
1914 	return (err);
1915 }
1916 
1917 
1918 /*
1919  * Internal service routines
1920  */
1921 
1922 struct rxq_refill_cb_arg {
1923 	int               error;
1924 	bus_dma_segment_t seg;
1925 	int               nseg;
1926 };
1927 
1928 static void
1929 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1930 {
1931 	struct rxq_refill_cb_arg *cb_arg = arg;
1932 
1933 	cb_arg->error = error;
1934 	cb_arg->seg = segs[0];
1935 	cb_arg->nseg = nseg;
1936 }
1937 
1938 /**
1939  * _iflib_fl_refill - refill an rxq free-buffer list
1940  * @ctx: the iflib context
1941  * @fl: the free list to refill
1942  * @count: the number of new buffers to allocate
1943  *
1944  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1945  * The caller must assure that @count does not exceed the queue's capacity.
1946  */
1947 static uint8_t
1948 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1949 {
1950 	struct if_rxd_update iru;
1951 	struct rxq_refill_cb_arg cb_arg;
1952 	struct mbuf *m;
1953 	caddr_t cl, *sd_cl;
1954 	struct mbuf **sd_m;
1955 	bus_dmamap_t *sd_map;
1956 	bus_addr_t bus_addr, *sd_ba;
1957 	int err, frag_idx, i, idx, n, pidx;
1958 	qidx_t credits;
1959 
1960 	sd_m = fl->ifl_sds.ifsd_m;
1961 	sd_map = fl->ifl_sds.ifsd_map;
1962 	sd_cl = fl->ifl_sds.ifsd_cl;
1963 	sd_ba = fl->ifl_sds.ifsd_ba;
1964 	pidx = fl->ifl_pidx;
1965 	idx = pidx;
1966 	frag_idx = fl->ifl_fragidx;
1967 	credits = fl->ifl_credits;
1968 
1969 	i = 0;
1970 	n = count;
1971 	MPASS(n > 0);
1972 	MPASS(credits + n <= fl->ifl_size);
1973 
1974 	if (pidx < fl->ifl_cidx)
1975 		MPASS(pidx + n <= fl->ifl_cidx);
1976 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1977 		MPASS(fl->ifl_gen == 0);
1978 	if (pidx > fl->ifl_cidx)
1979 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1980 
1981 	DBG_COUNTER_INC(fl_refills);
1982 	if (n > 8)
1983 		DBG_COUNTER_INC(fl_refills_large);
1984 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1985 	while (n--) {
1986 		/*
1987 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1988 		 * initialized after rx.
1989 		 *
1990 		 * If the cluster is still set then we know a minimum sized packet was received
1991 		 */
1992 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1993 		    &frag_idx);
1994 		if (frag_idx < 0)
1995 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1996 		MPASS(frag_idx >= 0);
1997 		if ((cl = sd_cl[frag_idx]) == NULL) {
1998 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1999 				break;
2000 
2001 			cb_arg.error = 0;
2002 			MPASS(sd_map != NULL);
2003 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2004 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2005 			    BUS_DMA_NOWAIT);
2006 			if (err != 0 || cb_arg.error) {
2007 				/*
2008 				 * !zone_pack ?
2009 				 */
2010 				if (fl->ifl_zone == zone_pack)
2011 					uma_zfree(fl->ifl_zone, cl);
2012 				break;
2013 			}
2014 
2015 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2016 			sd_cl[frag_idx] = cl;
2017 #if MEMORY_LOGGING
2018 			fl->ifl_cl_enqueued++;
2019 #endif
2020 		} else {
2021 			bus_addr = sd_ba[frag_idx];
2022 		}
2023 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2024 		    BUS_DMASYNC_PREREAD);
2025 
2026 		if (sd_m[frag_idx] == NULL) {
2027 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2028 				break;
2029 			}
2030 			sd_m[frag_idx] = m;
2031 		}
2032 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2033 #if MEMORY_LOGGING
2034 		fl->ifl_m_enqueued++;
2035 #endif
2036 
2037 		DBG_COUNTER_INC(rx_allocs);
2038 		fl->ifl_rxd_idxs[i] = frag_idx;
2039 		fl->ifl_bus_addrs[i] = bus_addr;
2040 		fl->ifl_vm_addrs[i] = cl;
2041 		credits++;
2042 		i++;
2043 		MPASS(credits <= fl->ifl_size);
2044 		if (++idx == fl->ifl_size) {
2045 			fl->ifl_gen = 1;
2046 			idx = 0;
2047 		}
2048 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2049 			iru.iru_pidx = pidx;
2050 			iru.iru_count = i;
2051 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2052 			i = 0;
2053 			pidx = idx;
2054 			fl->ifl_pidx = idx;
2055 			fl->ifl_credits = credits;
2056 		}
2057 	}
2058 
2059 	if (i) {
2060 		iru.iru_pidx = pidx;
2061 		iru.iru_count = i;
2062 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2063 		fl->ifl_pidx = idx;
2064 		fl->ifl_credits = credits;
2065 	}
2066 	DBG_COUNTER_INC(rxd_flush);
2067 	if (fl->ifl_pidx == 0)
2068 		pidx = fl->ifl_size - 1;
2069 	else
2070 		pidx = fl->ifl_pidx - 1;
2071 
2072 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2073 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2074 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2075 	fl->ifl_fragidx = frag_idx;
2076 
2077 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2078 }
2079 
2080 static __inline uint8_t
2081 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2082 {
2083 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2084 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2085 #ifdef INVARIANTS
2086 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2087 #endif
2088 
2089 	MPASS(fl->ifl_credits <= fl->ifl_size);
2090 	MPASS(reclaimable == delta);
2091 
2092 	if (reclaimable > 0)
2093 		return (_iflib_fl_refill(ctx, fl, min(max, reclaimable)));
2094 	return (0);
2095 }
2096 
2097 uint8_t
2098 iflib_in_detach(if_ctx_t ctx)
2099 {
2100 	bool in_detach;
2101 
2102 	STATE_LOCK(ctx);
2103 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2104 	STATE_UNLOCK(ctx);
2105 	return (in_detach);
2106 }
2107 
2108 static void
2109 iflib_fl_bufs_free(iflib_fl_t fl)
2110 {
2111 	iflib_dma_info_t idi = fl->ifl_ifdi;
2112 	bus_dmamap_t sd_map;
2113 	uint32_t i;
2114 
2115 	for (i = 0; i < fl->ifl_size; i++) {
2116 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2117 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2118 
2119 		if (*sd_cl != NULL) {
2120 			sd_map = fl->ifl_sds.ifsd_map[i];
2121 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2122 			    BUS_DMASYNC_POSTREAD);
2123 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2124 			if (*sd_cl != NULL)
2125 				uma_zfree(fl->ifl_zone, *sd_cl);
2126 			if (*sd_m != NULL) {
2127 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2128 				uma_zfree(zone_mbuf, *sd_m);
2129 			}
2130 		} else {
2131 			MPASS(*sd_cl == NULL);
2132 			MPASS(*sd_m == NULL);
2133 		}
2134 #if MEMORY_LOGGING
2135 		fl->ifl_m_dequeued++;
2136 		fl->ifl_cl_dequeued++;
2137 #endif
2138 		*sd_cl = NULL;
2139 		*sd_m = NULL;
2140 	}
2141 #ifdef INVARIANTS
2142 	for (i = 0; i < fl->ifl_size; i++) {
2143 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2144 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2145 	}
2146 #endif
2147 	/*
2148 	 * Reset free list values
2149 	 */
2150 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2151 	bzero(idi->idi_vaddr, idi->idi_size);
2152 }
2153 
2154 /*********************************************************************
2155  *
2156  *  Initialize a free list and its buffers.
2157  *
2158  **********************************************************************/
2159 static int
2160 iflib_fl_setup(iflib_fl_t fl)
2161 {
2162 	iflib_rxq_t rxq = fl->ifl_rxq;
2163 	if_ctx_t ctx = rxq->ifr_ctx;
2164 
2165 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2166 	/*
2167 	** Free current RX buffer structs and their mbufs
2168 	*/
2169 	iflib_fl_bufs_free(fl);
2170 	/* Now replenish the mbufs */
2171 	MPASS(fl->ifl_credits == 0);
2172 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2173 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2174 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2175 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2176 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2177 
2178 
2179 	/* avoid pre-allocating zillions of clusters to an idle card
2180 	 * potentially speeding up attach
2181 	 */
2182 	(void) _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2183 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2184 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2185 		return (ENOBUFS);
2186 	/*
2187 	 * handle failure
2188 	 */
2189 	MPASS(rxq != NULL);
2190 	MPASS(fl->ifl_ifdi != NULL);
2191 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2192 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2193 	return (0);
2194 }
2195 
2196 /*********************************************************************
2197  *
2198  *  Free receive ring data structures
2199  *
2200  **********************************************************************/
2201 static void
2202 iflib_rx_sds_free(iflib_rxq_t rxq)
2203 {
2204 	iflib_fl_t fl;
2205 	int i, j;
2206 
2207 	if (rxq->ifr_fl != NULL) {
2208 		for (i = 0; i < rxq->ifr_nfl; i++) {
2209 			fl = &rxq->ifr_fl[i];
2210 			if (fl->ifl_buf_tag != NULL) {
2211 				if (fl->ifl_sds.ifsd_map != NULL) {
2212 					for (j = 0; j < fl->ifl_size; j++) {
2213 						bus_dmamap_sync(
2214 						    fl->ifl_buf_tag,
2215 						    fl->ifl_sds.ifsd_map[j],
2216 						    BUS_DMASYNC_POSTREAD);
2217 						bus_dmamap_unload(
2218 						    fl->ifl_buf_tag,
2219 						    fl->ifl_sds.ifsd_map[j]);
2220 						bus_dmamap_destroy(
2221 						    fl->ifl_buf_tag,
2222 						    fl->ifl_sds.ifsd_map[j]);
2223 					}
2224 				}
2225 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2226 				fl->ifl_buf_tag = NULL;
2227 			}
2228 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2229 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2230 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2231 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2232 			fl->ifl_sds.ifsd_m = NULL;
2233 			fl->ifl_sds.ifsd_cl = NULL;
2234 			fl->ifl_sds.ifsd_ba = NULL;
2235 			fl->ifl_sds.ifsd_map = NULL;
2236 		}
2237 		free(rxq->ifr_fl, M_IFLIB);
2238 		rxq->ifr_fl = NULL;
2239 		free(rxq->ifr_ifdi, M_IFLIB);
2240 		rxq->ifr_ifdi = NULL;
2241 		rxq->ifr_cq_cidx = 0;
2242 	}
2243 }
2244 
2245 /*
2246  * Timer routine
2247  */
2248 static void
2249 iflib_timer(void *arg)
2250 {
2251 	iflib_txq_t txq = arg;
2252 	if_ctx_t ctx = txq->ift_ctx;
2253 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2254 	uint64_t this_tick = ticks;
2255 	uint32_t reset_on = hz / 2;
2256 
2257 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2258 		return;
2259 
2260 	/*
2261 	** Check on the state of the TX queue(s), this
2262 	** can be done without the lock because its RO
2263 	** and the HUNG state will be static if set.
2264 	*/
2265 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2266 		txq->ift_last_timer_tick = this_tick;
2267 		IFDI_TIMER(ctx, txq->ift_id);
2268 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2269 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2270 		     (sctx->isc_pause_frames == 0)))
2271 			goto hung;
2272 
2273 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2274 		    ifmp_ring_is_stalled(txq->ift_br)) {
2275 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2276 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2277 		}
2278 		txq->ift_cleaned_prev = txq->ift_cleaned;
2279 	}
2280 #ifdef DEV_NETMAP
2281 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2282 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2283 #endif
2284 	/* handle any laggards */
2285 	if (txq->ift_db_pending)
2286 		GROUPTASK_ENQUEUE(&txq->ift_task);
2287 
2288 	sctx->isc_pause_frames = 0;
2289 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2290 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2291 	return;
2292 
2293  hung:
2294 	device_printf(ctx->ifc_dev,
2295 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2296 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2297 	STATE_LOCK(ctx);
2298 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2299 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2300 	iflib_admin_intr_deferred(ctx);
2301 	STATE_UNLOCK(ctx);
2302 }
2303 
2304 static void
2305 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2306 {
2307 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2308 
2309 	/*
2310 	 * XXX don't set the max_frame_size to larger
2311 	 * than the hardware can handle
2312 	 */
2313 	if (sctx->isc_max_frame_size <= MCLBYTES)
2314 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2315 	else
2316 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2317 }
2318 
2319 uint32_t
2320 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2321 {
2322 
2323 	return (ctx->ifc_rx_mbuf_sz);
2324 }
2325 
2326 static void
2327 iflib_init_locked(if_ctx_t ctx)
2328 {
2329 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2330 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2331 	if_t ifp = ctx->ifc_ifp;
2332 	iflib_fl_t fl;
2333 	iflib_txq_t txq;
2334 	iflib_rxq_t rxq;
2335 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2336 
2337 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2338 	IFDI_INTR_DISABLE(ctx);
2339 
2340 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2341 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2342 	/* Set hardware offload abilities */
2343 	if_clearhwassist(ifp);
2344 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2345 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2346 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2347 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2348 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2349 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2350 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2351 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2352 
2353 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2354 		CALLOUT_LOCK(txq);
2355 		callout_stop(&txq->ift_timer);
2356 		CALLOUT_UNLOCK(txq);
2357 		iflib_netmap_txq_init(ctx, txq);
2358 	}
2359 
2360 	/*
2361 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2362 	 * that drivers can use the value when setting up the hardware receive
2363 	 * buffers.
2364 	 */
2365 	iflib_calc_rx_mbuf_sz(ctx);
2366 
2367 #ifdef INVARIANTS
2368 	i = if_getdrvflags(ifp);
2369 #endif
2370 	IFDI_INIT(ctx);
2371 	MPASS(if_getdrvflags(ifp) == i);
2372 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2373 		/* XXX this should really be done on a per-queue basis */
2374 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2375 			MPASS(rxq->ifr_id == i);
2376 			iflib_netmap_rxq_init(ctx, rxq);
2377 			continue;
2378 		}
2379 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2380 			if (iflib_fl_setup(fl)) {
2381 				device_printf(ctx->ifc_dev,
2382 				    "setting up free list %d failed - "
2383 				    "check cluster settings\n", j);
2384 				goto done;
2385 			}
2386 		}
2387 	}
2388 done:
2389 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2390 	IFDI_INTR_ENABLE(ctx);
2391 	txq = ctx->ifc_txqs;
2392 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2393 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2394 			txq->ift_timer.c_cpu);
2395 }
2396 
2397 static int
2398 iflib_media_change(if_t ifp)
2399 {
2400 	if_ctx_t ctx = if_getsoftc(ifp);
2401 	int err;
2402 
2403 	CTX_LOCK(ctx);
2404 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2405 		iflib_init_locked(ctx);
2406 	CTX_UNLOCK(ctx);
2407 	return (err);
2408 }
2409 
2410 static void
2411 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2412 {
2413 	if_ctx_t ctx = if_getsoftc(ifp);
2414 
2415 	CTX_LOCK(ctx);
2416 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2417 	IFDI_MEDIA_STATUS(ctx, ifmr);
2418 	CTX_UNLOCK(ctx);
2419 }
2420 
2421 void
2422 iflib_stop(if_ctx_t ctx)
2423 {
2424 	iflib_txq_t txq = ctx->ifc_txqs;
2425 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2426 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2427 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2428 	iflib_dma_info_t di;
2429 	iflib_fl_t fl;
2430 	int i, j;
2431 
2432 	/* Tell the stack that the interface is no longer active */
2433 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2434 
2435 	IFDI_INTR_DISABLE(ctx);
2436 	DELAY(1000);
2437 	IFDI_STOP(ctx);
2438 	DELAY(1000);
2439 
2440 	iflib_debug_reset();
2441 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2442 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2443 		/* make sure all transmitters have completed before proceeding XXX */
2444 
2445 		CALLOUT_LOCK(txq);
2446 		callout_stop(&txq->ift_timer);
2447 		CALLOUT_UNLOCK(txq);
2448 
2449 		/* clean any enqueued buffers */
2450 		iflib_ifmp_purge(txq);
2451 		/* Free any existing tx buffers. */
2452 		for (j = 0; j < txq->ift_size; j++) {
2453 			iflib_txsd_free(ctx, txq, j);
2454 		}
2455 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2456 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2457 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2458 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2459 		txq->ift_pullups = 0;
2460 		ifmp_ring_reset_stats(txq->ift_br);
2461 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2462 			bzero((void *)di->idi_vaddr, di->idi_size);
2463 	}
2464 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2465 		/* make sure all transmitters have completed before proceeding XXX */
2466 
2467 		rxq->ifr_cq_cidx = 0;
2468 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2469 			bzero((void *)di->idi_vaddr, di->idi_size);
2470 		/* also resets the free lists pidx/cidx */
2471 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2472 			iflib_fl_bufs_free(fl);
2473 	}
2474 }
2475 
2476 static inline caddr_t
2477 calc_next_rxd(iflib_fl_t fl, int cidx)
2478 {
2479 	qidx_t size;
2480 	int nrxd;
2481 	caddr_t start, end, cur, next;
2482 
2483 	nrxd = fl->ifl_size;
2484 	size = fl->ifl_rxd_size;
2485 	start = fl->ifl_ifdi->idi_vaddr;
2486 
2487 	if (__predict_false(size == 0))
2488 		return (start);
2489 	cur = start + size*cidx;
2490 	end = start + size*nrxd;
2491 	next = CACHE_PTR_NEXT(cur);
2492 	return (next < end ? next : start);
2493 }
2494 
2495 static inline void
2496 prefetch_pkts(iflib_fl_t fl, int cidx)
2497 {
2498 	int nextptr;
2499 	int nrxd = fl->ifl_size;
2500 	caddr_t next_rxd;
2501 
2502 
2503 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2504 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2505 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2506 	next_rxd = calc_next_rxd(fl, cidx);
2507 	prefetch(next_rxd);
2508 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2509 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2510 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2511 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2512 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2513 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2514 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2515 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2516 }
2517 
2518 static struct mbuf *
2519 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2520     int *pf_rv, if_rxd_info_t ri)
2521 {
2522 	bus_dmamap_t map;
2523 	iflib_fl_t fl;
2524 	caddr_t payload;
2525 	struct mbuf *m;
2526 	int flid, cidx, len, next;
2527 
2528 	map = NULL;
2529 	flid = irf->irf_flid;
2530 	cidx = irf->irf_idx;
2531 	fl = &rxq->ifr_fl[flid];
2532 	sd->ifsd_fl = fl;
2533 	sd->ifsd_cidx = cidx;
2534 	m = fl->ifl_sds.ifsd_m[cidx];
2535 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2536 	fl->ifl_credits--;
2537 #if MEMORY_LOGGING
2538 	fl->ifl_m_dequeued++;
2539 #endif
2540 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2541 		prefetch_pkts(fl, cidx);
2542 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2543 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2544 	map = fl->ifl_sds.ifsd_map[cidx];
2545 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2546 
2547 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2548 	MPASS(fl->ifl_cidx == cidx);
2549 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2550 
2551 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2552 		payload  = *sd->ifsd_cl;
2553 		payload +=  ri->iri_pad;
2554 		len = ri->iri_len - ri->iri_pad;
2555 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2556 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2557 		switch (*pf_rv) {
2558 		case PFIL_DROPPED:
2559 		case PFIL_CONSUMED:
2560 			/*
2561 			 * The filter ate it.  Everything is recycled.
2562 			 */
2563 			m = NULL;
2564 			unload = 0;
2565 			break;
2566 		case PFIL_REALLOCED:
2567 			/*
2568 			 * The filter copied it.  Everything is recycled.
2569 			 */
2570 			m = pfil_mem2mbuf(payload);
2571 			unload = 0;
2572 			break;
2573 		case PFIL_PASS:
2574 			/*
2575 			 * Filter said it was OK, so receive like
2576 			 * normal
2577 			 */
2578 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2579 			break;
2580 		default:
2581 			MPASS(0);
2582 		}
2583 	} else {
2584 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2585 		*pf_rv = PFIL_PASS;
2586 	}
2587 
2588 	if (unload)
2589 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2590 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2591 	if (__predict_false(fl->ifl_cidx == 0))
2592 		fl->ifl_gen = 0;
2593 	bit_clear(fl->ifl_rx_bitmap, cidx);
2594 	return (m);
2595 }
2596 
2597 static struct mbuf *
2598 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2599 {
2600 	struct mbuf *m, *mh, *mt;
2601 	caddr_t cl;
2602 	int  *pf_rv_ptr, flags, i, padlen;
2603 	bool consumed;
2604 
2605 	i = 0;
2606 	mh = NULL;
2607 	consumed = false;
2608 	*pf_rv = PFIL_PASS;
2609 	pf_rv_ptr = pf_rv;
2610 	do {
2611 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2612 		    pf_rv_ptr, ri);
2613 
2614 		MPASS(*sd->ifsd_cl != NULL);
2615 
2616 		/*
2617 		 * Exclude zero-length frags & frags from
2618 		 * packets the filter has consumed or dropped
2619 		 */
2620 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2621 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2622 			if (mh == NULL) {
2623 				/* everything saved here */
2624 				consumed = true;
2625 				pf_rv_ptr = NULL;
2626 				continue;
2627 			}
2628 			/* XXX we can save the cluster here, but not the mbuf */
2629 			m_init(m, M_NOWAIT, MT_DATA, 0);
2630 			m_free(m);
2631 			continue;
2632 		}
2633 		if (mh == NULL) {
2634 			flags = M_PKTHDR|M_EXT;
2635 			mh = mt = m;
2636 			padlen = ri->iri_pad;
2637 		} else {
2638 			flags = M_EXT;
2639 			mt->m_next = m;
2640 			mt = m;
2641 			/* assuming padding is only on the first fragment */
2642 			padlen = 0;
2643 		}
2644 		cl = *sd->ifsd_cl;
2645 		*sd->ifsd_cl = NULL;
2646 
2647 		/* Can these two be made one ? */
2648 		m_init(m, M_NOWAIT, MT_DATA, flags);
2649 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2650 		/*
2651 		 * These must follow m_init and m_cljset
2652 		 */
2653 		m->m_data += padlen;
2654 		ri->iri_len -= padlen;
2655 		m->m_len = ri->iri_frags[i].irf_len;
2656 	} while (++i < ri->iri_nfrags);
2657 
2658 	return (mh);
2659 }
2660 
2661 /*
2662  * Process one software descriptor
2663  */
2664 static struct mbuf *
2665 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2666 {
2667 	struct if_rxsd sd;
2668 	struct mbuf *m;
2669 	int pf_rv;
2670 
2671 	/* should I merge this back in now that the two paths are basically duplicated? */
2672 	if (ri->iri_nfrags == 1 &&
2673 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2674 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2675 		    &pf_rv, ri);
2676 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2677 			return (m);
2678 		if (pf_rv == PFIL_PASS) {
2679 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2680 #ifndef __NO_STRICT_ALIGNMENT
2681 			if (!IP_ALIGNED(m))
2682 				m->m_data += 2;
2683 #endif
2684 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2685 			m->m_len = ri->iri_frags[0].irf_len;
2686 		}
2687 	} else {
2688 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2689 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2690 			return (m);
2691 	}
2692 	m->m_pkthdr.len = ri->iri_len;
2693 	m->m_pkthdr.rcvif = ri->iri_ifp;
2694 	m->m_flags |= ri->iri_flags;
2695 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2696 	m->m_pkthdr.flowid = ri->iri_flowid;
2697 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2698 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2699 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2700 	return (m);
2701 }
2702 
2703 #if defined(INET6) || defined(INET)
2704 static void
2705 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2706 {
2707 	CURVNET_SET(lc->ifp->if_vnet);
2708 #if defined(INET6)
2709 	*v6 = V_ip6_forwarding;
2710 #endif
2711 #if defined(INET)
2712 	*v4 = V_ipforwarding;
2713 #endif
2714 	CURVNET_RESTORE();
2715 }
2716 
2717 /*
2718  * Returns true if it's possible this packet could be LROed.
2719  * if it returns false, it is guaranteed that tcp_lro_rx()
2720  * would not return zero.
2721  */
2722 static bool
2723 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2724 {
2725 	struct ether_header *eh;
2726 
2727 	eh = mtod(m, struct ether_header *);
2728 	switch (eh->ether_type) {
2729 #if defined(INET6)
2730 		case htons(ETHERTYPE_IPV6):
2731 			return (!v6_forwarding);
2732 #endif
2733 #if defined (INET)
2734 		case htons(ETHERTYPE_IP):
2735 			return (!v4_forwarding);
2736 #endif
2737 	}
2738 
2739 	return false;
2740 }
2741 #else
2742 static void
2743 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2744 {
2745 }
2746 #endif
2747 
2748 static void
2749 _task_fn_rx_watchdog(void *context)
2750 {
2751 	iflib_rxq_t rxq = context;
2752 
2753 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2754 }
2755 
2756 static uint8_t
2757 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2758 {
2759 	if_t ifp;
2760 	if_ctx_t ctx = rxq->ifr_ctx;
2761 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2762 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2763 	int avail, i;
2764 	qidx_t *cidxp;
2765 	struct if_rxd_info ri;
2766 	int err, budget_left, rx_bytes, rx_pkts;
2767 	iflib_fl_t fl;
2768 	int lro_enabled;
2769 	bool v4_forwarding, v6_forwarding, lro_possible;
2770 	uint8_t retval = 0;
2771 
2772 	/*
2773 	 * XXX early demux data packets so that if_input processing only handles
2774 	 * acks in interrupt context
2775 	 */
2776 	struct mbuf *m, *mh, *mt, *mf;
2777 
2778 	NET_EPOCH_ASSERT();
2779 
2780 	lro_possible = v4_forwarding = v6_forwarding = false;
2781 	ifp = ctx->ifc_ifp;
2782 	mh = mt = NULL;
2783 	MPASS(budget > 0);
2784 	rx_pkts	= rx_bytes = 0;
2785 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2786 		cidxp = &rxq->ifr_cq_cidx;
2787 	else
2788 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2789 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2790 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2791 			retval |= __iflib_fl_refill_lt(ctx, fl, budget + 8);
2792 		DBG_COUNTER_INC(rx_unavail);
2793 		return (retval);
2794 	}
2795 
2796 	/* pfil needs the vnet to be set */
2797 	CURVNET_SET_QUIET(ifp->if_vnet);
2798 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2799 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2800 			DBG_COUNTER_INC(rx_ctx_inactive);
2801 			break;
2802 		}
2803 		/*
2804 		 * Reset client set fields to their default values
2805 		 */
2806 		rxd_info_zero(&ri);
2807 		ri.iri_qsidx = rxq->ifr_id;
2808 		ri.iri_cidx = *cidxp;
2809 		ri.iri_ifp = ifp;
2810 		ri.iri_frags = rxq->ifr_frags;
2811 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2812 
2813 		if (err)
2814 			goto err;
2815 		rx_pkts += 1;
2816 		rx_bytes += ri.iri_len;
2817 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2818 			*cidxp = ri.iri_cidx;
2819 			/* Update our consumer index */
2820 			/* XXX NB: shurd - check if this is still safe */
2821 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2822 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2823 			/* was this only a completion queue message? */
2824 			if (__predict_false(ri.iri_nfrags == 0))
2825 				continue;
2826 		}
2827 		MPASS(ri.iri_nfrags != 0);
2828 		MPASS(ri.iri_len != 0);
2829 
2830 		/* will advance the cidx on the corresponding free lists */
2831 		m = iflib_rxd_pkt_get(rxq, &ri);
2832 		avail--;
2833 		budget_left--;
2834 		if (avail == 0 && budget_left)
2835 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2836 
2837 		if (__predict_false(m == NULL))
2838 			continue;
2839 
2840 		/* imm_pkt: -- cxgb */
2841 		if (mh == NULL)
2842 			mh = mt = m;
2843 		else {
2844 			mt->m_nextpkt = m;
2845 			mt = m;
2846 		}
2847 	}
2848 	CURVNET_RESTORE();
2849 	/* make sure that we can refill faster than drain */
2850 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2851 		retval |= __iflib_fl_refill_lt(ctx, fl, budget + 8);
2852 
2853 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2854 	if (lro_enabled)
2855 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2856 	mt = mf = NULL;
2857 	while (mh != NULL) {
2858 		m = mh;
2859 		mh = mh->m_nextpkt;
2860 		m->m_nextpkt = NULL;
2861 #ifndef __NO_STRICT_ALIGNMENT
2862 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2863 			continue;
2864 #endif
2865 		rx_bytes += m->m_pkthdr.len;
2866 		rx_pkts++;
2867 #if defined(INET6) || defined(INET)
2868 		if (lro_enabled) {
2869 			if (!lro_possible) {
2870 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2871 				if (lro_possible && mf != NULL) {
2872 					ifp->if_input(ifp, mf);
2873 					DBG_COUNTER_INC(rx_if_input);
2874 					mt = mf = NULL;
2875 				}
2876 			}
2877 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2878 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2879 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2880 					continue;
2881 			}
2882 		}
2883 #endif
2884 		if (lro_possible) {
2885 			ifp->if_input(ifp, m);
2886 			DBG_COUNTER_INC(rx_if_input);
2887 			continue;
2888 		}
2889 
2890 		if (mf == NULL)
2891 			mf = m;
2892 		if (mt != NULL)
2893 			mt->m_nextpkt = m;
2894 		mt = m;
2895 	}
2896 	if (mf != NULL) {
2897 		ifp->if_input(ifp, mf);
2898 		DBG_COUNTER_INC(rx_if_input);
2899 	}
2900 
2901 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2902 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2903 
2904 	/*
2905 	 * Flush any outstanding LRO work
2906 	 */
2907 #if defined(INET6) || defined(INET)
2908 	tcp_lro_flush_all(&rxq->ifr_lc);
2909 #endif
2910 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2911 		retval |= IFLIB_RXEOF_MORE;
2912 	return (retval);
2913 err:
2914 	STATE_LOCK(ctx);
2915 	ctx->ifc_flags |= IFC_DO_RESET;
2916 	iflib_admin_intr_deferred(ctx);
2917 	STATE_UNLOCK(ctx);
2918 	return (0);
2919 }
2920 
2921 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2922 static inline qidx_t
2923 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2924 {
2925 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2926 	qidx_t minthresh = txq->ift_size / 8;
2927 	if (in_use > 4*minthresh)
2928 		return (notify_count);
2929 	if (in_use > 2*minthresh)
2930 		return (notify_count >> 1);
2931 	if (in_use > minthresh)
2932 		return (notify_count >> 3);
2933 	return (0);
2934 }
2935 
2936 static inline qidx_t
2937 txq_max_rs_deferred(iflib_txq_t txq)
2938 {
2939 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2940 	qidx_t minthresh = txq->ift_size / 8;
2941 	if (txq->ift_in_use > 4*minthresh)
2942 		return (notify_count);
2943 	if (txq->ift_in_use > 2*minthresh)
2944 		return (notify_count >> 1);
2945 	if (txq->ift_in_use > minthresh)
2946 		return (notify_count >> 2);
2947 	return (2);
2948 }
2949 
2950 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2951 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2952 
2953 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2954 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2955 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2956 
2957 /* forward compatibility for cxgb */
2958 #define FIRST_QSET(ctx) 0
2959 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2960 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2961 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2962 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2963 
2964 /* XXX we should be setting this to something other than zero */
2965 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2966 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2967     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2968 
2969 static inline bool
2970 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2971 {
2972 	qidx_t dbval, max;
2973 	bool rang;
2974 
2975 	rang = false;
2976 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2977 	if (ring || txq->ift_db_pending >= max) {
2978 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2979 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2980 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2981 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2982 		txq->ift_db_pending = txq->ift_npending = 0;
2983 		rang = true;
2984 	}
2985 	return (rang);
2986 }
2987 
2988 #ifdef PKT_DEBUG
2989 static void
2990 print_pkt(if_pkt_info_t pi)
2991 {
2992 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2993 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2994 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2995 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2996 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2997 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2998 }
2999 #endif
3000 
3001 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3002 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3003 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3004 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3005 
3006 static int
3007 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3008 {
3009 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3010 	struct ether_vlan_header *eh;
3011 	struct mbuf *m;
3012 
3013 	m = *mp;
3014 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3015 	    M_WRITABLE(m) == 0) {
3016 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3017 			return (ENOMEM);
3018 		} else {
3019 			m_freem(*mp);
3020 			DBG_COUNTER_INC(tx_frees);
3021 			*mp = m;
3022 		}
3023 	}
3024 
3025 	/*
3026 	 * Determine where frame payload starts.
3027 	 * Jump over vlan headers if already present,
3028 	 * helpful for QinQ too.
3029 	 */
3030 	if (__predict_false(m->m_len < sizeof(*eh))) {
3031 		txq->ift_pullups++;
3032 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3033 			return (ENOMEM);
3034 	}
3035 	eh = mtod(m, struct ether_vlan_header *);
3036 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3037 		pi->ipi_etype = ntohs(eh->evl_proto);
3038 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3039 	} else {
3040 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3041 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3042 	}
3043 
3044 	switch (pi->ipi_etype) {
3045 #ifdef INET
3046 	case ETHERTYPE_IP:
3047 	{
3048 		struct mbuf *n;
3049 		struct ip *ip = NULL;
3050 		struct tcphdr *th = NULL;
3051 		int minthlen;
3052 
3053 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3054 		if (__predict_false(m->m_len < minthlen)) {
3055 			/*
3056 			 * if this code bloat is causing too much of a hit
3057 			 * move it to a separate function and mark it noinline
3058 			 */
3059 			if (m->m_len == pi->ipi_ehdrlen) {
3060 				n = m->m_next;
3061 				MPASS(n);
3062 				if (n->m_len >= sizeof(*ip))  {
3063 					ip = (struct ip *)n->m_data;
3064 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3065 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3066 				} else {
3067 					txq->ift_pullups++;
3068 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3069 						return (ENOMEM);
3070 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3071 				}
3072 			} else {
3073 				txq->ift_pullups++;
3074 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3075 					return (ENOMEM);
3076 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3077 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3078 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3079 			}
3080 		} else {
3081 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3082 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3083 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3084 		}
3085 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3086 		pi->ipi_ipproto = ip->ip_p;
3087 		pi->ipi_flags |= IPI_TX_IPV4;
3088 
3089 		/* TCP checksum offload may require TCP header length */
3090 		if (IS_TX_OFFLOAD4(pi)) {
3091 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3092 				if (__predict_false(th == NULL)) {
3093 					txq->ift_pullups++;
3094 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3095 						return (ENOMEM);
3096 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3097 				}
3098 				pi->ipi_tcp_hflags = th->th_flags;
3099 				pi->ipi_tcp_hlen = th->th_off << 2;
3100 				pi->ipi_tcp_seq = th->th_seq;
3101 			}
3102 			if (IS_TSO4(pi)) {
3103 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3104 					return (ENXIO);
3105 				/*
3106 				 * TSO always requires hardware checksum offload.
3107 				 */
3108 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3109 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3110 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3111 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3112 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3113 					ip->ip_sum = 0;
3114 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3115 				}
3116 			}
3117 		}
3118 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3119                        ip->ip_sum = 0;
3120 
3121 		break;
3122 	}
3123 #endif
3124 #ifdef INET6
3125 	case ETHERTYPE_IPV6:
3126 	{
3127 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3128 		struct tcphdr *th;
3129 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3130 
3131 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3132 			txq->ift_pullups++;
3133 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3134 				return (ENOMEM);
3135 		}
3136 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3137 
3138 		/* XXX-BZ this will go badly in case of ext hdrs. */
3139 		pi->ipi_ipproto = ip6->ip6_nxt;
3140 		pi->ipi_flags |= IPI_TX_IPV6;
3141 
3142 		/* TCP checksum offload may require TCP header length */
3143 		if (IS_TX_OFFLOAD6(pi)) {
3144 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3145 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3146 					txq->ift_pullups++;
3147 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3148 						return (ENOMEM);
3149 				}
3150 				pi->ipi_tcp_hflags = th->th_flags;
3151 				pi->ipi_tcp_hlen = th->th_off << 2;
3152 				pi->ipi_tcp_seq = th->th_seq;
3153 			}
3154 			if (IS_TSO6(pi)) {
3155 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3156 					return (ENXIO);
3157 				/*
3158 				 * TSO always requires hardware checksum offload.
3159 				 */
3160 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3161 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3162 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3163 			}
3164 		}
3165 		break;
3166 	}
3167 #endif
3168 	default:
3169 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3170 		pi->ipi_ip_hlen = 0;
3171 		break;
3172 	}
3173 	*mp = m;
3174 
3175 	return (0);
3176 }
3177 
3178 /*
3179  * If dodgy hardware rejects the scatter gather chain we've handed it
3180  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3181  * m_defrag'd mbufs
3182  */
3183 static __noinline struct mbuf *
3184 iflib_remove_mbuf(iflib_txq_t txq)
3185 {
3186 	int ntxd, pidx;
3187 	struct mbuf *m, **ifsd_m;
3188 
3189 	ifsd_m = txq->ift_sds.ifsd_m;
3190 	ntxd = txq->ift_size;
3191 	pidx = txq->ift_pidx & (ntxd - 1);
3192 	ifsd_m = txq->ift_sds.ifsd_m;
3193 	m = ifsd_m[pidx];
3194 	ifsd_m[pidx] = NULL;
3195 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3196 	if (txq->ift_sds.ifsd_tso_map != NULL)
3197 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3198 		    txq->ift_sds.ifsd_tso_map[pidx]);
3199 #if MEMORY_LOGGING
3200 	txq->ift_dequeued++;
3201 #endif
3202 	return (m);
3203 }
3204 
3205 static inline caddr_t
3206 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3207 {
3208 	qidx_t size;
3209 	int ntxd;
3210 	caddr_t start, end, cur, next;
3211 
3212 	ntxd = txq->ift_size;
3213 	size = txq->ift_txd_size[qid];
3214 	start = txq->ift_ifdi[qid].idi_vaddr;
3215 
3216 	if (__predict_false(size == 0))
3217 		return (start);
3218 	cur = start + size*cidx;
3219 	end = start + size*ntxd;
3220 	next = CACHE_PTR_NEXT(cur);
3221 	return (next < end ? next : start);
3222 }
3223 
3224 /*
3225  * Pad an mbuf to ensure a minimum ethernet frame size.
3226  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3227  */
3228 static __noinline int
3229 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3230 {
3231 	/*
3232 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3233 	 * and ARP message is the smallest common payload I can think of
3234 	 */
3235 	static char pad[18];	/* just zeros */
3236 	int n;
3237 	struct mbuf *new_head;
3238 
3239 	if (!M_WRITABLE(*m_head)) {
3240 		new_head = m_dup(*m_head, M_NOWAIT);
3241 		if (new_head == NULL) {
3242 			m_freem(*m_head);
3243 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3244 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3245 			DBG_COUNTER_INC(tx_frees);
3246 			return ENOMEM;
3247 		}
3248 		m_freem(*m_head);
3249 		*m_head = new_head;
3250 	}
3251 
3252 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3253 	     n > 0; n -= sizeof(pad))
3254 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3255 			break;
3256 
3257 	if (n > 0) {
3258 		m_freem(*m_head);
3259 		device_printf(dev, "cannot pad short frame\n");
3260 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3261 		DBG_COUNTER_INC(tx_frees);
3262 		return (ENOBUFS);
3263 	}
3264 
3265 	return 0;
3266 }
3267 
3268 static int
3269 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3270 {
3271 	if_ctx_t		ctx;
3272 	if_shared_ctx_t		sctx;
3273 	if_softc_ctx_t		scctx;
3274 	bus_dma_tag_t		buf_tag;
3275 	bus_dma_segment_t	*segs;
3276 	struct mbuf		*m_head, **ifsd_m;
3277 	void			*next_txd;
3278 	bus_dmamap_t		map;
3279 	struct if_pkt_info	pi;
3280 	int remap = 0;
3281 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3282 
3283 	ctx = txq->ift_ctx;
3284 	sctx = ctx->ifc_sctx;
3285 	scctx = &ctx->ifc_softc_ctx;
3286 	segs = txq->ift_segs;
3287 	ntxd = txq->ift_size;
3288 	m_head = *m_headp;
3289 	map = NULL;
3290 
3291 	/*
3292 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3293 	 */
3294 	cidx = txq->ift_cidx;
3295 	pidx = txq->ift_pidx;
3296 	if (ctx->ifc_flags & IFC_PREFETCH) {
3297 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3298 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3299 			next_txd = calc_next_txd(txq, cidx, 0);
3300 			prefetch(next_txd);
3301 		}
3302 
3303 		/* prefetch the next cache line of mbuf pointers and flags */
3304 		prefetch(&txq->ift_sds.ifsd_m[next]);
3305 		prefetch(&txq->ift_sds.ifsd_map[next]);
3306 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3307 	}
3308 	map = txq->ift_sds.ifsd_map[pidx];
3309 	ifsd_m = txq->ift_sds.ifsd_m;
3310 
3311 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3312 		buf_tag = txq->ift_tso_buf_tag;
3313 		max_segs = scctx->isc_tx_tso_segments_max;
3314 		map = txq->ift_sds.ifsd_tso_map[pidx];
3315 		MPASS(buf_tag != NULL);
3316 		MPASS(max_segs > 0);
3317 	} else {
3318 		buf_tag = txq->ift_buf_tag;
3319 		max_segs = scctx->isc_tx_nsegments;
3320 		map = txq->ift_sds.ifsd_map[pidx];
3321 	}
3322 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3323 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3324 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3325 		if (err) {
3326 			DBG_COUNTER_INC(encap_txd_encap_fail);
3327 			return err;
3328 		}
3329 	}
3330 	m_head = *m_headp;
3331 
3332 	pkt_info_zero(&pi);
3333 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3334 	pi.ipi_pidx = pidx;
3335 	pi.ipi_qsidx = txq->ift_id;
3336 	pi.ipi_len = m_head->m_pkthdr.len;
3337 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3338 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3339 
3340 	/* deliberate bitwise OR to make one condition */
3341 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3342 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3343 			DBG_COUNTER_INC(encap_txd_encap_fail);
3344 			return (err);
3345 		}
3346 		m_head = *m_headp;
3347 	}
3348 
3349 retry:
3350 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3351 	    BUS_DMA_NOWAIT);
3352 defrag:
3353 	if (__predict_false(err)) {
3354 		switch (err) {
3355 		case EFBIG:
3356 			/* try collapse once and defrag once */
3357 			if (remap == 0) {
3358 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3359 				/* try defrag if collapsing fails */
3360 				if (m_head == NULL)
3361 					remap++;
3362 			}
3363 			if (remap == 1) {
3364 				txq->ift_mbuf_defrag++;
3365 				m_head = m_defrag(*m_headp, M_NOWAIT);
3366 			}
3367 			/*
3368 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3369 			 * failed to map an mbuf that was run through m_defrag
3370 			 */
3371 			MPASS(remap <= 1);
3372 			if (__predict_false(m_head == NULL || remap > 1))
3373 				goto defrag_failed;
3374 			remap++;
3375 			*m_headp = m_head;
3376 			goto retry;
3377 			break;
3378 		case ENOMEM:
3379 			txq->ift_no_tx_dma_setup++;
3380 			break;
3381 		default:
3382 			txq->ift_no_tx_dma_setup++;
3383 			m_freem(*m_headp);
3384 			DBG_COUNTER_INC(tx_frees);
3385 			*m_headp = NULL;
3386 			break;
3387 		}
3388 		txq->ift_map_failed++;
3389 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3390 		DBG_COUNTER_INC(encap_txd_encap_fail);
3391 		return (err);
3392 	}
3393 	ifsd_m[pidx] = m_head;
3394 	/*
3395 	 * XXX assumes a 1 to 1 relationship between segments and
3396 	 *        descriptors - this does not hold true on all drivers, e.g.
3397 	 *        cxgb
3398 	 */
3399 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3400 		txq->ift_no_desc_avail++;
3401 		bus_dmamap_unload(buf_tag, map);
3402 		DBG_COUNTER_INC(encap_txq_avail_fail);
3403 		DBG_COUNTER_INC(encap_txd_encap_fail);
3404 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3405 			GROUPTASK_ENQUEUE(&txq->ift_task);
3406 		return (ENOBUFS);
3407 	}
3408 	/*
3409 	 * On Intel cards we can greatly reduce the number of TX interrupts
3410 	 * we see by only setting report status on every Nth descriptor.
3411 	 * However, this also means that the driver will need to keep track
3412 	 * of the descriptors that RS was set on to check them for the DD bit.
3413 	 */
3414 	txq->ift_rs_pending += nsegs + 1;
3415 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3416 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3417 		pi.ipi_flags |= IPI_TX_INTR;
3418 		txq->ift_rs_pending = 0;
3419 	}
3420 
3421 	pi.ipi_segs = segs;
3422 	pi.ipi_nsegs = nsegs;
3423 
3424 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3425 #ifdef PKT_DEBUG
3426 	print_pkt(&pi);
3427 #endif
3428 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3429 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3430 		DBG_COUNTER_INC(tx_encap);
3431 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3432 
3433 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3434 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3435 			ndesc += txq->ift_size;
3436 			txq->ift_gen = 1;
3437 		}
3438 		/*
3439 		 * drivers can need as many as
3440 		 * two sentinels
3441 		 */
3442 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3443 		MPASS(pi.ipi_new_pidx != pidx);
3444 		MPASS(ndesc > 0);
3445 		txq->ift_in_use += ndesc;
3446 
3447 		/*
3448 		 * We update the last software descriptor again here because there may
3449 		 * be a sentinel and/or there may be more mbufs than segments
3450 		 */
3451 		txq->ift_pidx = pi.ipi_new_pidx;
3452 		txq->ift_npending += pi.ipi_ndescs;
3453 	} else {
3454 		*m_headp = m_head = iflib_remove_mbuf(txq);
3455 		if (err == EFBIG) {
3456 			txq->ift_txd_encap_efbig++;
3457 			if (remap < 2) {
3458 				remap = 1;
3459 				goto defrag;
3460 			}
3461 		}
3462 		goto defrag_failed;
3463 	}
3464 	/*
3465 	 * err can't possibly be non-zero here, so we don't neet to test it
3466 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3467 	 */
3468 	return (err);
3469 
3470 defrag_failed:
3471 	txq->ift_mbuf_defrag_failed++;
3472 	txq->ift_map_failed++;
3473 	m_freem(*m_headp);
3474 	DBG_COUNTER_INC(tx_frees);
3475 	*m_headp = NULL;
3476 	DBG_COUNTER_INC(encap_txd_encap_fail);
3477 	return (ENOMEM);
3478 }
3479 
3480 static void
3481 iflib_tx_desc_free(iflib_txq_t txq, int n)
3482 {
3483 	uint32_t qsize, cidx, mask, gen;
3484 	struct mbuf *m, **ifsd_m;
3485 	bool do_prefetch;
3486 
3487 	cidx = txq->ift_cidx;
3488 	gen = txq->ift_gen;
3489 	qsize = txq->ift_size;
3490 	mask = qsize-1;
3491 	ifsd_m = txq->ift_sds.ifsd_m;
3492 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3493 
3494 	while (n-- > 0) {
3495 		if (do_prefetch) {
3496 			prefetch(ifsd_m[(cidx + 3) & mask]);
3497 			prefetch(ifsd_m[(cidx + 4) & mask]);
3498 		}
3499 		if ((m = ifsd_m[cidx]) != NULL) {
3500 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3501 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3502 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3503 				    txq->ift_sds.ifsd_tso_map[cidx],
3504 				    BUS_DMASYNC_POSTWRITE);
3505 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3506 				    txq->ift_sds.ifsd_tso_map[cidx]);
3507 			} else {
3508 				bus_dmamap_sync(txq->ift_buf_tag,
3509 				    txq->ift_sds.ifsd_map[cidx],
3510 				    BUS_DMASYNC_POSTWRITE);
3511 				bus_dmamap_unload(txq->ift_buf_tag,
3512 				    txq->ift_sds.ifsd_map[cidx]);
3513 			}
3514 			/* XXX we don't support any drivers that batch packets yet */
3515 			MPASS(m->m_nextpkt == NULL);
3516 			m_freem(m);
3517 			ifsd_m[cidx] = NULL;
3518 #if MEMORY_LOGGING
3519 			txq->ift_dequeued++;
3520 #endif
3521 			DBG_COUNTER_INC(tx_frees);
3522 		}
3523 		if (__predict_false(++cidx == qsize)) {
3524 			cidx = 0;
3525 			gen = 0;
3526 		}
3527 	}
3528 	txq->ift_cidx = cidx;
3529 	txq->ift_gen = gen;
3530 }
3531 
3532 static __inline int
3533 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3534 {
3535 	int reclaim;
3536 	if_ctx_t ctx = txq->ift_ctx;
3537 
3538 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3539 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3540 
3541 	/*
3542 	 * Need a rate-limiting check so that this isn't called every time
3543 	 */
3544 	iflib_tx_credits_update(ctx, txq);
3545 	reclaim = DESC_RECLAIMABLE(txq);
3546 
3547 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3548 #ifdef INVARIANTS
3549 		if (iflib_verbose_debug) {
3550 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3551 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3552 			       reclaim, thresh);
3553 
3554 		}
3555 #endif
3556 		return (0);
3557 	}
3558 	iflib_tx_desc_free(txq, reclaim);
3559 	txq->ift_cleaned += reclaim;
3560 	txq->ift_in_use -= reclaim;
3561 
3562 	return (reclaim);
3563 }
3564 
3565 static struct mbuf **
3566 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3567 {
3568 	int next, size;
3569 	struct mbuf **items;
3570 
3571 	size = r->size;
3572 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3573 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3574 
3575 	prefetch(items[(cidx + offset) & (size-1)]);
3576 	if (remaining > 1) {
3577 		prefetch2cachelines(&items[next]);
3578 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3579 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3580 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3581 	}
3582 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3583 }
3584 
3585 static void
3586 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3587 {
3588 
3589 	ifmp_ring_check_drainage(txq->ift_br, budget);
3590 }
3591 
3592 static uint32_t
3593 iflib_txq_can_drain(struct ifmp_ring *r)
3594 {
3595 	iflib_txq_t txq = r->cookie;
3596 	if_ctx_t ctx = txq->ift_ctx;
3597 
3598 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3599 		return (1);
3600 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3601 	    BUS_DMASYNC_POSTREAD);
3602 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3603 	    false));
3604 }
3605 
3606 static uint32_t
3607 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3608 {
3609 	iflib_txq_t txq = r->cookie;
3610 	if_ctx_t ctx = txq->ift_ctx;
3611 	if_t ifp = ctx->ifc_ifp;
3612 	struct mbuf *m, **mp;
3613 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3614 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3615 	bool do_prefetch, rang, ring;
3616 
3617 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3618 			    !LINK_ACTIVE(ctx))) {
3619 		DBG_COUNTER_INC(txq_drain_notready);
3620 		return (0);
3621 	}
3622 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3623 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3624 	avail = IDXDIFF(pidx, cidx, r->size);
3625 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3626 		DBG_COUNTER_INC(txq_drain_flushing);
3627 		for (i = 0; i < avail; i++) {
3628 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3629 				m_free(r->items[(cidx + i) & (r->size-1)]);
3630 			r->items[(cidx + i) & (r->size-1)] = NULL;
3631 		}
3632 		return (avail);
3633 	}
3634 
3635 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3636 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3637 		CALLOUT_LOCK(txq);
3638 		callout_stop(&txq->ift_timer);
3639 		CALLOUT_UNLOCK(txq);
3640 		DBG_COUNTER_INC(txq_drain_oactive);
3641 		return (0);
3642 	}
3643 	if (reclaimed)
3644 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3645 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3646 	count = MIN(avail, TX_BATCH_SIZE);
3647 #ifdef INVARIANTS
3648 	if (iflib_verbose_debug)
3649 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3650 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3651 #endif
3652 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3653 	txq_avail = TXQ_AVAIL(txq);
3654 	err = 0;
3655 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3656 		int rem = do_prefetch ? count - i : 0;
3657 
3658 		mp = _ring_peek_one(r, cidx, i, rem);
3659 		MPASS(mp != NULL && *mp != NULL);
3660 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3661 			consumed++;
3662 			continue;
3663 		}
3664 		in_use_prev = txq->ift_in_use;
3665 		err = iflib_encap(txq, mp);
3666 		if (__predict_false(err)) {
3667 			/* no room - bail out */
3668 			if (err == ENOBUFS)
3669 				break;
3670 			consumed++;
3671 			/* we can't send this packet - skip it */
3672 			continue;
3673 		}
3674 		consumed++;
3675 		pkt_sent++;
3676 		m = *mp;
3677 		DBG_COUNTER_INC(tx_sent);
3678 		bytes_sent += m->m_pkthdr.len;
3679 		mcast_sent += !!(m->m_flags & M_MCAST);
3680 		txq_avail = TXQ_AVAIL(txq);
3681 
3682 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3683 		ETHER_BPF_MTAP(ifp, m);
3684 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3685 			break;
3686 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3687 	}
3688 
3689 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3690 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3691 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3692 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3693 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3694 	if (mcast_sent)
3695 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3696 #ifdef INVARIANTS
3697 	if (iflib_verbose_debug)
3698 		printf("consumed=%d\n", consumed);
3699 #endif
3700 	return (consumed);
3701 }
3702 
3703 static uint32_t
3704 iflib_txq_drain_always(struct ifmp_ring *r)
3705 {
3706 	return (1);
3707 }
3708 
3709 static uint32_t
3710 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3711 {
3712 	int i, avail;
3713 	struct mbuf **mp;
3714 	iflib_txq_t txq;
3715 
3716 	txq = r->cookie;
3717 
3718 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3719 	CALLOUT_LOCK(txq);
3720 	callout_stop(&txq->ift_timer);
3721 	CALLOUT_UNLOCK(txq);
3722 
3723 	avail = IDXDIFF(pidx, cidx, r->size);
3724 	for (i = 0; i < avail; i++) {
3725 		mp = _ring_peek_one(r, cidx, i, avail - i);
3726 		if (__predict_false(*mp == (struct mbuf *)txq))
3727 			continue;
3728 		m_freem(*mp);
3729 		DBG_COUNTER_INC(tx_frees);
3730 	}
3731 	MPASS(ifmp_ring_is_stalled(r) == 0);
3732 	return (avail);
3733 }
3734 
3735 static void
3736 iflib_ifmp_purge(iflib_txq_t txq)
3737 {
3738 	struct ifmp_ring *r;
3739 
3740 	r = txq->ift_br;
3741 	r->drain = iflib_txq_drain_free;
3742 	r->can_drain = iflib_txq_drain_always;
3743 
3744 	ifmp_ring_check_drainage(r, r->size);
3745 
3746 	r->drain = iflib_txq_drain;
3747 	r->can_drain = iflib_txq_can_drain;
3748 }
3749 
3750 static void
3751 _task_fn_tx(void *context)
3752 {
3753 	iflib_txq_t txq = context;
3754 	if_ctx_t ctx = txq->ift_ctx;
3755 #if defined(ALTQ) || defined(DEV_NETMAP)
3756 	if_t ifp = ctx->ifc_ifp;
3757 #endif
3758 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3759 
3760 #ifdef IFLIB_DIAGNOSTICS
3761 	txq->ift_cpu_exec_count[curcpu]++;
3762 #endif
3763 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3764 		return;
3765 #ifdef DEV_NETMAP
3766 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3767 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3768 		    BUS_DMASYNC_POSTREAD);
3769 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3770 			netmap_tx_irq(ifp, txq->ift_id);
3771 		if (ctx->ifc_flags & IFC_LEGACY)
3772 			IFDI_INTR_ENABLE(ctx);
3773 		else
3774 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3775 		return;
3776 	}
3777 #endif
3778 #ifdef ALTQ
3779 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3780 		iflib_altq_if_start(ifp);
3781 #endif
3782 	if (txq->ift_db_pending)
3783 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3784 	else if (!abdicate)
3785 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3786 	/*
3787 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3788 	 */
3789 	if (abdicate)
3790 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3791 	if (ctx->ifc_flags & IFC_LEGACY)
3792 		IFDI_INTR_ENABLE(ctx);
3793 	else
3794 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3795 }
3796 
3797 static void
3798 _task_fn_rx(void *context)
3799 {
3800 	iflib_rxq_t rxq = context;
3801 	if_ctx_t ctx = rxq->ifr_ctx;
3802 	uint8_t more;
3803 	uint16_t budget;
3804 
3805 #ifdef IFLIB_DIAGNOSTICS
3806 	rxq->ifr_cpu_exec_count[curcpu]++;
3807 #endif
3808 	DBG_COUNTER_INC(task_fn_rxs);
3809 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3810 		return;
3811 #ifdef DEV_NETMAP
3812 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3813 		u_int work = 0;
3814 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3815 			more = 0;
3816 			goto skip_rxeof;
3817 		}
3818 	}
3819 #endif
3820 	budget = ctx->ifc_sysctl_rx_budget;
3821 	if (budget == 0)
3822 		budget = 16;	/* XXX */
3823 	more = iflib_rxeof(rxq, budget);
3824 #ifdef DEV_NETMAP
3825 skip_rxeof:
3826 #endif
3827 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3828 		if (ctx->ifc_flags & IFC_LEGACY)
3829 			IFDI_INTR_ENABLE(ctx);
3830 		else
3831 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3832 		DBG_COUNTER_INC(rx_intr_enables);
3833 	}
3834 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3835 		return;
3836 
3837 	if (more & IFLIB_RXEOF_MORE)
3838 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3839 	else if (more & IFLIB_RXEOF_EMPTY)
3840 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3841 }
3842 
3843 static void
3844 _task_fn_admin(void *context)
3845 {
3846 	if_ctx_t ctx = context;
3847 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3848 	iflib_txq_t txq;
3849 	int i;
3850 	bool oactive, running, do_reset, do_watchdog, in_detach;
3851 	uint32_t reset_on = hz / 2;
3852 
3853 	STATE_LOCK(ctx);
3854 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3855 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3856 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3857 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3858 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3859 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3860 	STATE_UNLOCK(ctx);
3861 
3862 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3863 		return;
3864 	if (in_detach)
3865 		return;
3866 
3867 	CTX_LOCK(ctx);
3868 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3869 		CALLOUT_LOCK(txq);
3870 		callout_stop(&txq->ift_timer);
3871 		CALLOUT_UNLOCK(txq);
3872 	}
3873 	if (do_watchdog) {
3874 		ctx->ifc_watchdog_events++;
3875 		IFDI_WATCHDOG_RESET(ctx);
3876 	}
3877 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3878 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3879 #ifdef DEV_NETMAP
3880 		reset_on = hz / 2;
3881 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3882 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3883 #endif
3884 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3885 	}
3886 	IFDI_LINK_INTR_ENABLE(ctx);
3887 	if (do_reset)
3888 		iflib_if_init_locked(ctx);
3889 	CTX_UNLOCK(ctx);
3890 
3891 	if (LINK_ACTIVE(ctx) == 0)
3892 		return;
3893 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3894 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3895 }
3896 
3897 
3898 static void
3899 _task_fn_iov(void *context)
3900 {
3901 	if_ctx_t ctx = context;
3902 
3903 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3904 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3905 		return;
3906 
3907 	CTX_LOCK(ctx);
3908 	IFDI_VFLR_HANDLE(ctx);
3909 	CTX_UNLOCK(ctx);
3910 }
3911 
3912 static int
3913 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3914 {
3915 	int err;
3916 	if_int_delay_info_t info;
3917 	if_ctx_t ctx;
3918 
3919 	info = (if_int_delay_info_t)arg1;
3920 	ctx = info->iidi_ctx;
3921 	info->iidi_req = req;
3922 	info->iidi_oidp = oidp;
3923 	CTX_LOCK(ctx);
3924 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3925 	CTX_UNLOCK(ctx);
3926 	return (err);
3927 }
3928 
3929 /*********************************************************************
3930  *
3931  *  IFNET FUNCTIONS
3932  *
3933  **********************************************************************/
3934 
3935 static void
3936 iflib_if_init_locked(if_ctx_t ctx)
3937 {
3938 	iflib_stop(ctx);
3939 	iflib_init_locked(ctx);
3940 }
3941 
3942 
3943 static void
3944 iflib_if_init(void *arg)
3945 {
3946 	if_ctx_t ctx = arg;
3947 
3948 	CTX_LOCK(ctx);
3949 	iflib_if_init_locked(ctx);
3950 	CTX_UNLOCK(ctx);
3951 }
3952 
3953 static int
3954 iflib_if_transmit(if_t ifp, struct mbuf *m)
3955 {
3956 	if_ctx_t	ctx = if_getsoftc(ifp);
3957 
3958 	iflib_txq_t txq;
3959 	int err, qidx;
3960 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3961 
3962 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3963 		DBG_COUNTER_INC(tx_frees);
3964 		m_freem(m);
3965 		return (ENETDOWN);
3966 	}
3967 
3968 	MPASS(m->m_nextpkt == NULL);
3969 	/* ALTQ-enabled interfaces always use queue 0. */
3970 	qidx = 0;
3971 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3972 		qidx = QIDX(ctx, m);
3973 	/*
3974 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3975 	 */
3976 	txq = &ctx->ifc_txqs[qidx];
3977 
3978 #ifdef DRIVER_BACKPRESSURE
3979 	if (txq->ift_closed) {
3980 		while (m != NULL) {
3981 			next = m->m_nextpkt;
3982 			m->m_nextpkt = NULL;
3983 			m_freem(m);
3984 			DBG_COUNTER_INC(tx_frees);
3985 			m = next;
3986 		}
3987 		return (ENOBUFS);
3988 	}
3989 #endif
3990 #ifdef notyet
3991 	qidx = count = 0;
3992 	mp = marr;
3993 	next = m;
3994 	do {
3995 		count++;
3996 		next = next->m_nextpkt;
3997 	} while (next != NULL);
3998 
3999 	if (count > nitems(marr))
4000 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4001 			/* XXX check nextpkt */
4002 			m_freem(m);
4003 			/* XXX simplify for now */
4004 			DBG_COUNTER_INC(tx_frees);
4005 			return (ENOBUFS);
4006 		}
4007 	for (next = m, i = 0; next != NULL; i++) {
4008 		mp[i] = next;
4009 		next = next->m_nextpkt;
4010 		mp[i]->m_nextpkt = NULL;
4011 	}
4012 #endif
4013 	DBG_COUNTER_INC(tx_seen);
4014 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4015 
4016 	if (abdicate)
4017 		GROUPTASK_ENQUEUE(&txq->ift_task);
4018  	if (err) {
4019 		if (!abdicate)
4020 			GROUPTASK_ENQUEUE(&txq->ift_task);
4021 		/* support forthcoming later */
4022 #ifdef DRIVER_BACKPRESSURE
4023 		txq->ift_closed = TRUE;
4024 #endif
4025 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4026 		m_freem(m);
4027 		DBG_COUNTER_INC(tx_frees);
4028 	}
4029 
4030 	return (err);
4031 }
4032 
4033 #ifdef ALTQ
4034 /*
4035  * The overall approach to integrating iflib with ALTQ is to continue to use
4036  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4037  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4038  * is redundant/unnecessary, but doing so minimizes the amount of
4039  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4040  * redundantly queueing to an intermediate mp_ring is swamped by the
4041  * performance limitations inherent in using ALTQ.
4042  *
4043  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4044  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4045  * given interface.  If ALTQ is enabled for an interface, then all
4046  * transmitted packets for that interface will be submitted to the ALTQ
4047  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4048  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4049  * update stats that the iflib machinery handles, and which is sensitve to
4050  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4051  * will be installed as the start routine for use by ALTQ facilities that
4052  * need to trigger queue drains on a scheduled basis.
4053  *
4054  */
4055 static void
4056 iflib_altq_if_start(if_t ifp)
4057 {
4058 	struct ifaltq *ifq = &ifp->if_snd;
4059 	struct mbuf *m;
4060 
4061 	IFQ_LOCK(ifq);
4062 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4063 	while (m != NULL) {
4064 		iflib_if_transmit(ifp, m);
4065 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4066 	}
4067 	IFQ_UNLOCK(ifq);
4068 }
4069 
4070 static int
4071 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4072 {
4073 	int err;
4074 
4075 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4076 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4077 		if (err == 0)
4078 			iflib_altq_if_start(ifp);
4079 	} else
4080 		err = iflib_if_transmit(ifp, m);
4081 
4082 	return (err);
4083 }
4084 #endif /* ALTQ */
4085 
4086 static void
4087 iflib_if_qflush(if_t ifp)
4088 {
4089 	if_ctx_t ctx = if_getsoftc(ifp);
4090 	iflib_txq_t txq = ctx->ifc_txqs;
4091 	int i;
4092 
4093 	STATE_LOCK(ctx);
4094 	ctx->ifc_flags |= IFC_QFLUSH;
4095 	STATE_UNLOCK(ctx);
4096 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4097 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4098 			iflib_txq_check_drain(txq, 0);
4099 	STATE_LOCK(ctx);
4100 	ctx->ifc_flags &= ~IFC_QFLUSH;
4101 	STATE_UNLOCK(ctx);
4102 
4103 	/*
4104 	 * When ALTQ is enabled, this will also take care of purging the
4105 	 * ALTQ queue(s).
4106 	 */
4107 	if_qflush(ifp);
4108 }
4109 
4110 
4111 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4112 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4113 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4114 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4115 
4116 static int
4117 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4118 {
4119 	if_ctx_t ctx = if_getsoftc(ifp);
4120 	struct ifreq	*ifr = (struct ifreq *)data;
4121 #if defined(INET) || defined(INET6)
4122 	struct ifaddr	*ifa = (struct ifaddr *)data;
4123 #endif
4124 	bool		avoid_reset = false;
4125 	int		err = 0, reinit = 0, bits;
4126 
4127 	switch (command) {
4128 	case SIOCSIFADDR:
4129 #ifdef INET
4130 		if (ifa->ifa_addr->sa_family == AF_INET)
4131 			avoid_reset = true;
4132 #endif
4133 #ifdef INET6
4134 		if (ifa->ifa_addr->sa_family == AF_INET6)
4135 			avoid_reset = true;
4136 #endif
4137 		/*
4138 		** Calling init results in link renegotiation,
4139 		** so we avoid doing it when possible.
4140 		*/
4141 		if (avoid_reset) {
4142 			if_setflagbits(ifp, IFF_UP,0);
4143 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4144 				reinit = 1;
4145 #ifdef INET
4146 			if (!(if_getflags(ifp) & IFF_NOARP))
4147 				arp_ifinit(ifp, ifa);
4148 #endif
4149 		} else
4150 			err = ether_ioctl(ifp, command, data);
4151 		break;
4152 	case SIOCSIFMTU:
4153 		CTX_LOCK(ctx);
4154 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4155 			CTX_UNLOCK(ctx);
4156 			break;
4157 		}
4158 		bits = if_getdrvflags(ifp);
4159 		/* stop the driver and free any clusters before proceeding */
4160 		iflib_stop(ctx);
4161 
4162 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4163 			STATE_LOCK(ctx);
4164 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4165 				ctx->ifc_flags |= IFC_MULTISEG;
4166 			else
4167 				ctx->ifc_flags &= ~IFC_MULTISEG;
4168 			STATE_UNLOCK(ctx);
4169 			err = if_setmtu(ifp, ifr->ifr_mtu);
4170 		}
4171 		iflib_init_locked(ctx);
4172 		STATE_LOCK(ctx);
4173 		if_setdrvflags(ifp, bits);
4174 		STATE_UNLOCK(ctx);
4175 		CTX_UNLOCK(ctx);
4176 		break;
4177 	case SIOCSIFFLAGS:
4178 		CTX_LOCK(ctx);
4179 		if (if_getflags(ifp) & IFF_UP) {
4180 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4181 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4182 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4183 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4184 				}
4185 			} else
4186 				reinit = 1;
4187 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4188 			iflib_stop(ctx);
4189 		}
4190 		ctx->ifc_if_flags = if_getflags(ifp);
4191 		CTX_UNLOCK(ctx);
4192 		break;
4193 	case SIOCADDMULTI:
4194 	case SIOCDELMULTI:
4195 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4196 			CTX_LOCK(ctx);
4197 			IFDI_INTR_DISABLE(ctx);
4198 			IFDI_MULTI_SET(ctx);
4199 			IFDI_INTR_ENABLE(ctx);
4200 			CTX_UNLOCK(ctx);
4201 		}
4202 		break;
4203 	case SIOCSIFMEDIA:
4204 		CTX_LOCK(ctx);
4205 		IFDI_MEDIA_SET(ctx);
4206 		CTX_UNLOCK(ctx);
4207 		/* FALLTHROUGH */
4208 	case SIOCGIFMEDIA:
4209 	case SIOCGIFXMEDIA:
4210 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4211 		break;
4212 	case SIOCGI2C:
4213 	{
4214 		struct ifi2creq i2c;
4215 
4216 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4217 		if (err != 0)
4218 			break;
4219 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4220 			err = EINVAL;
4221 			break;
4222 		}
4223 		if (i2c.len > sizeof(i2c.data)) {
4224 			err = EINVAL;
4225 			break;
4226 		}
4227 
4228 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4229 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4230 			    sizeof(i2c));
4231 		break;
4232 	}
4233 	case SIOCSIFCAP:
4234 	{
4235 		int mask, setmask, oldmask;
4236 
4237 		oldmask = if_getcapenable(ifp);
4238 		mask = ifr->ifr_reqcap ^ oldmask;
4239 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4240 		setmask = 0;
4241 #ifdef TCP_OFFLOAD
4242 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4243 #endif
4244 		setmask |= (mask & IFCAP_FLAGS);
4245 		setmask |= (mask & IFCAP_WOL);
4246 
4247 		/*
4248 		 * If any RX csum has changed, change all the ones that
4249 		 * are supported by the driver.
4250 		 */
4251 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4252 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4253 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4254 		}
4255 
4256 		/*
4257 		 * want to ensure that traffic has stopped before we change any of the flags
4258 		 */
4259 		if (setmask) {
4260 			CTX_LOCK(ctx);
4261 			bits = if_getdrvflags(ifp);
4262 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4263 				iflib_stop(ctx);
4264 			STATE_LOCK(ctx);
4265 			if_togglecapenable(ifp, setmask);
4266 			STATE_UNLOCK(ctx);
4267 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4268 				iflib_init_locked(ctx);
4269 			STATE_LOCK(ctx);
4270 			if_setdrvflags(ifp, bits);
4271 			STATE_UNLOCK(ctx);
4272 			CTX_UNLOCK(ctx);
4273 		}
4274 		if_vlancap(ifp);
4275 		break;
4276 	}
4277 	case SIOCGPRIVATE_0:
4278 	case SIOCSDRVSPEC:
4279 	case SIOCGDRVSPEC:
4280 		CTX_LOCK(ctx);
4281 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4282 		CTX_UNLOCK(ctx);
4283 		break;
4284 	default:
4285 		err = ether_ioctl(ifp, command, data);
4286 		break;
4287 	}
4288 	if (reinit)
4289 		iflib_if_init(ctx);
4290 	return (err);
4291 }
4292 
4293 static uint64_t
4294 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4295 {
4296 	if_ctx_t ctx = if_getsoftc(ifp);
4297 
4298 	return (IFDI_GET_COUNTER(ctx, cnt));
4299 }
4300 
4301 /*********************************************************************
4302  *
4303  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4304  *
4305  **********************************************************************/
4306 
4307 static void
4308 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4309 {
4310 	if_ctx_t ctx = if_getsoftc(ifp);
4311 
4312 	if ((void *)ctx != arg)
4313 		return;
4314 
4315 	if ((vtag == 0) || (vtag > 4095))
4316 		return;
4317 
4318 	if (iflib_in_detach(ctx))
4319 		return;
4320 
4321 	CTX_LOCK(ctx);
4322 	IFDI_VLAN_REGISTER(ctx, vtag);
4323 	/* Re-init to load the changes */
4324 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4325 		iflib_if_init_locked(ctx);
4326 	CTX_UNLOCK(ctx);
4327 }
4328 
4329 static void
4330 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4331 {
4332 	if_ctx_t ctx = if_getsoftc(ifp);
4333 
4334 	if ((void *)ctx != arg)
4335 		return;
4336 
4337 	if ((vtag == 0) || (vtag > 4095))
4338 		return;
4339 
4340 	CTX_LOCK(ctx);
4341 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4342 	/* Re-init to load the changes */
4343 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4344 		iflib_if_init_locked(ctx);
4345 	CTX_UNLOCK(ctx);
4346 }
4347 
4348 static void
4349 iflib_led_func(void *arg, int onoff)
4350 {
4351 	if_ctx_t ctx = arg;
4352 
4353 	CTX_LOCK(ctx);
4354 	IFDI_LED_FUNC(ctx, onoff);
4355 	CTX_UNLOCK(ctx);
4356 }
4357 
4358 /*********************************************************************
4359  *
4360  *  BUS FUNCTION DEFINITIONS
4361  *
4362  **********************************************************************/
4363 
4364 int
4365 iflib_device_probe(device_t dev)
4366 {
4367 	const pci_vendor_info_t *ent;
4368 	if_shared_ctx_t sctx;
4369 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4370 	uint16_t pci_vendor_id;
4371 
4372 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4373 		return (ENOTSUP);
4374 
4375 	pci_vendor_id = pci_get_vendor(dev);
4376 	pci_device_id = pci_get_device(dev);
4377 	pci_subvendor_id = pci_get_subvendor(dev);
4378 	pci_subdevice_id = pci_get_subdevice(dev);
4379 	pci_rev_id = pci_get_revid(dev);
4380 	if (sctx->isc_parse_devinfo != NULL)
4381 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4382 
4383 	ent = sctx->isc_vendor_info;
4384 	while (ent->pvi_vendor_id != 0) {
4385 		if (pci_vendor_id != ent->pvi_vendor_id) {
4386 			ent++;
4387 			continue;
4388 		}
4389 		if ((pci_device_id == ent->pvi_device_id) &&
4390 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4391 		     (ent->pvi_subvendor_id == 0)) &&
4392 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4393 		     (ent->pvi_subdevice_id == 0)) &&
4394 		    ((pci_rev_id == ent->pvi_rev_id) ||
4395 		     (ent->pvi_rev_id == 0))) {
4396 
4397 			device_set_desc_copy(dev, ent->pvi_name);
4398 			/* this needs to be changed to zero if the bus probing code
4399 			 * ever stops re-probing on best match because the sctx
4400 			 * may have its values over written by register calls
4401 			 * in subsequent probes
4402 			 */
4403 			return (BUS_PROBE_DEFAULT);
4404 		}
4405 		ent++;
4406 	}
4407 	return (ENXIO);
4408 }
4409 
4410 int
4411 iflib_device_probe_vendor(device_t dev)
4412 {
4413 	int probe;
4414 
4415 	probe = iflib_device_probe(dev);
4416 	if (probe == BUS_PROBE_DEFAULT)
4417 		return (BUS_PROBE_VENDOR);
4418 	else
4419 		return (probe);
4420 }
4421 
4422 static void
4423 iflib_reset_qvalues(if_ctx_t ctx)
4424 {
4425 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4426 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4427 	device_t dev = ctx->ifc_dev;
4428 	int i;
4429 
4430 	if (ctx->ifc_sysctl_ntxqs != 0)
4431 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4432 	if (ctx->ifc_sysctl_nrxqs != 0)
4433 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4434 
4435 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4436 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4437 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4438 		else
4439 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4440 	}
4441 
4442 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4443 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4444 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4445 		else
4446 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4447 	}
4448 
4449 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4450 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4451 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4452 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4453 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4454 		}
4455 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4456 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4457 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4458 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4459 		}
4460 		if (!powerof2(scctx->isc_nrxd[i])) {
4461 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4462 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4463 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4464 		}
4465 	}
4466 
4467 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4468 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4469 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4470 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4471 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4472 		}
4473 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4474 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4475 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4476 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4477 		}
4478 		if (!powerof2(scctx->isc_ntxd[i])) {
4479 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4480 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4481 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4482 		}
4483 	}
4484 }
4485 
4486 static void
4487 iflib_add_pfil(if_ctx_t ctx)
4488 {
4489 	struct pfil_head *pfil;
4490 	struct pfil_head_args pa;
4491 	iflib_rxq_t rxq;
4492 	int i;
4493 
4494 	pa.pa_version = PFIL_VERSION;
4495 	pa.pa_flags = PFIL_IN;
4496 	pa.pa_type = PFIL_TYPE_ETHERNET;
4497 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4498 	pfil = pfil_head_register(&pa);
4499 
4500 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4501 		rxq->pfil = pfil;
4502 	}
4503 }
4504 
4505 static void
4506 iflib_rem_pfil(if_ctx_t ctx)
4507 {
4508 	struct pfil_head *pfil;
4509 	iflib_rxq_t rxq;
4510 	int i;
4511 
4512 	rxq = ctx->ifc_rxqs;
4513 	pfil = rxq->pfil;
4514 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4515 		rxq->pfil = NULL;
4516 	}
4517 	pfil_head_unregister(pfil);
4518 }
4519 
4520 static uint16_t
4521 get_ctx_core_offset(if_ctx_t ctx)
4522 {
4523 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4524 	struct cpu_offset *op;
4525 	uint16_t qc;
4526 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4527 
4528 	if (ret != CORE_OFFSET_UNSPECIFIED)
4529 		return (ret);
4530 
4531 	if (ctx->ifc_sysctl_separate_txrx)
4532 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4533 	else
4534 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4535 
4536 	mtx_lock(&cpu_offset_mtx);
4537 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4538 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4539 			ret = op->offset;
4540 			op->offset += qc;
4541 			MPASS(op->refcount < UINT_MAX);
4542 			op->refcount++;
4543 			break;
4544 		}
4545 	}
4546 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4547 		ret = 0;
4548 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4549 		    M_NOWAIT | M_ZERO);
4550 		if (op == NULL) {
4551 			device_printf(ctx->ifc_dev,
4552 			    "allocation for cpu offset failed.\n");
4553 		} else {
4554 			op->offset = qc;
4555 			op->refcount = 1;
4556 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4557 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4558 		}
4559 	}
4560 	mtx_unlock(&cpu_offset_mtx);
4561 
4562 	return (ret);
4563 }
4564 
4565 static void
4566 unref_ctx_core_offset(if_ctx_t ctx)
4567 {
4568 	struct cpu_offset *op, *top;
4569 
4570 	mtx_lock(&cpu_offset_mtx);
4571 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4572 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4573 			MPASS(op->refcount > 0);
4574 			op->refcount--;
4575 			if (op->refcount == 0) {
4576 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4577 				free(op, M_IFLIB);
4578 			}
4579 			break;
4580 		}
4581 	}
4582 	mtx_unlock(&cpu_offset_mtx);
4583 }
4584 
4585 int
4586 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4587 {
4588 	if_ctx_t ctx;
4589 	if_t ifp;
4590 	if_softc_ctx_t scctx;
4591 	kobjop_desc_t kobj_desc;
4592 	kobj_method_t *kobj_method;
4593 	int err, msix, rid;
4594 	uint16_t main_rxq, main_txq;
4595 
4596 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4597 
4598 	if (sc == NULL) {
4599 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4600 		device_set_softc(dev, ctx);
4601 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4602 	}
4603 
4604 	ctx->ifc_sctx = sctx;
4605 	ctx->ifc_dev = dev;
4606 	ctx->ifc_softc = sc;
4607 
4608 	if ((err = iflib_register(ctx)) != 0) {
4609 		device_printf(dev, "iflib_register failed %d\n", err);
4610 		goto fail_ctx_free;
4611 	}
4612 	iflib_add_device_sysctl_pre(ctx);
4613 
4614 	scctx = &ctx->ifc_softc_ctx;
4615 	ifp = ctx->ifc_ifp;
4616 
4617 	iflib_reset_qvalues(ctx);
4618 	CTX_LOCK(ctx);
4619 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4620 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4621 		goto fail_unlock;
4622 	}
4623 	_iflib_pre_assert(scctx);
4624 	ctx->ifc_txrx = *scctx->isc_txrx;
4625 
4626 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4627 		ctx->ifc_mediap = scctx->isc_media;
4628 
4629 #ifdef INVARIANTS
4630 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4631 		MPASS(scctx->isc_tx_csum_flags);
4632 #endif
4633 
4634 	if_setcapabilities(ifp,
4635 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4636 	if_setcapenable(ifp,
4637 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4638 
4639 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4640 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4641 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4642 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4643 
4644 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4645 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4646 
4647 	/* XXX change for per-queue sizes */
4648 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4649 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4650 
4651 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4652 	    MAX_SINGLE_PACKET_FRACTION)
4653 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4654 		    MAX_SINGLE_PACKET_FRACTION);
4655 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4656 	    MAX_SINGLE_PACKET_FRACTION)
4657 		scctx->isc_tx_tso_segments_max = max(1,
4658 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4659 
4660 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4661 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4662 		/*
4663 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4664 		 * but some MACs do.
4665 		 */
4666 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4667 		    IP_MAXPACKET));
4668 		/*
4669 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4670 		 * into account.  In the worst case, each of these calls will
4671 		 * add another mbuf and, thus, the requirement for another DMA
4672 		 * segment.  So for best performance, it doesn't make sense to
4673 		 * advertize a maximum of TSO segments that typically will
4674 		 * require defragmentation in iflib_encap().
4675 		 */
4676 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4677 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4678 	}
4679 	if (scctx->isc_rss_table_size == 0)
4680 		scctx->isc_rss_table_size = 64;
4681 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4682 
4683 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4684 	/* XXX format name */
4685 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4686 	    NULL, NULL, "admin");
4687 
4688 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4689 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4690 		device_printf(dev, "Unable to fetch CPU list\n");
4691 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4692 	}
4693 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4694 
4695 	/*
4696 	** Now set up MSI or MSI-X, should return us the number of supported
4697 	** vectors (will be 1 for a legacy interrupt and MSI).
4698 	*/
4699 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4700 		msix = scctx->isc_vectors;
4701 	} else if (scctx->isc_msix_bar != 0)
4702 	       /*
4703 		* The simple fact that isc_msix_bar is not 0 does not mean we
4704 		* we have a good value there that is known to work.
4705 		*/
4706 		msix = iflib_msix_init(ctx);
4707 	else {
4708 		scctx->isc_vectors = 1;
4709 		scctx->isc_ntxqsets = 1;
4710 		scctx->isc_nrxqsets = 1;
4711 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4712 		msix = 0;
4713 	}
4714 	/* Get memory for the station queues */
4715 	if ((err = iflib_queues_alloc(ctx))) {
4716 		device_printf(dev, "Unable to allocate queue memory\n");
4717 		goto fail_intr_free;
4718 	}
4719 
4720 	if ((err = iflib_qset_structures_setup(ctx)))
4721 		goto fail_queues;
4722 
4723 	/*
4724 	 * Now that we know how many queues there are, get the core offset.
4725 	 */
4726 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4727 
4728 	/*
4729 	 * Group taskqueues aren't properly set up until SMP is started,
4730 	 * so we disable interrupts until we can handle them post
4731 	 * SI_SUB_SMP.
4732 	 *
4733 	 * XXX: disabling interrupts doesn't actually work, at least for
4734 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4735 	 * we do null handling and depend on this not causing too large an
4736 	 * interrupt storm.
4737 	 */
4738 	IFDI_INTR_DISABLE(ctx);
4739 
4740 	if (msix > 1) {
4741 		/*
4742 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4743 		 * aren't the default NULL implementation.
4744 		 */
4745 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4746 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4747 		    kobj_desc);
4748 		if (kobj_method == &kobj_desc->deflt) {
4749 			device_printf(dev,
4750 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4751 			err = EOPNOTSUPP;
4752 			goto fail_queues;
4753 		}
4754 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4755 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4756 		    kobj_desc);
4757 		if (kobj_method == &kobj_desc->deflt) {
4758 			device_printf(dev,
4759 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4760 			err = EOPNOTSUPP;
4761 			goto fail_queues;
4762 		}
4763 
4764 		/*
4765 		 * Assign the MSI-X vectors.
4766 		 * Note that the default NULL ifdi_msix_intr_assign method will
4767 		 * fail here, too.
4768 		 */
4769 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4770 		if (err != 0) {
4771 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4772 			    err);
4773 			goto fail_queues;
4774 		}
4775 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4776 		rid = 0;
4777 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4778 			MPASS(msix == 1);
4779 			rid = 1;
4780 		}
4781 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4782 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4783 			goto fail_queues;
4784 		}
4785 	} else {
4786 		device_printf(dev,
4787 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4788 		err = ENODEV;
4789 		goto fail_intr_free;
4790 	}
4791 
4792 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4793 
4794 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4795 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4796 		goto fail_detach;
4797 	}
4798 
4799 	/*
4800 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4801 	 * This must appear after the call to ether_ifattach() because
4802 	 * ether_ifattach() sets if_hdrlen to the default value.
4803 	 */
4804 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4805 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4806 
4807 	if ((err = iflib_netmap_attach(ctx))) {
4808 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4809 		goto fail_detach;
4810 	}
4811 	*ctxp = ctx;
4812 
4813 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4814 
4815 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4816 	iflib_add_device_sysctl_post(ctx);
4817 	iflib_add_pfil(ctx);
4818 	ctx->ifc_flags |= IFC_INIT_DONE;
4819 	CTX_UNLOCK(ctx);
4820 
4821 	return (0);
4822 
4823 fail_detach:
4824 	ether_ifdetach(ctx->ifc_ifp);
4825 fail_intr_free:
4826 	iflib_free_intr_mem(ctx);
4827 fail_queues:
4828 	iflib_tx_structures_free(ctx);
4829 	iflib_rx_structures_free(ctx);
4830 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4831 	IFDI_DETACH(ctx);
4832 fail_unlock:
4833 	CTX_UNLOCK(ctx);
4834 	iflib_deregister(ctx);
4835 fail_ctx_free:
4836 	device_set_softc(ctx->ifc_dev, NULL);
4837         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4838                 free(ctx->ifc_softc, M_IFLIB);
4839         free(ctx, M_IFLIB);
4840 	return (err);
4841 }
4842 
4843 int
4844 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4845 					  struct iflib_cloneattach_ctx *clctx)
4846 {
4847 	int err;
4848 	if_ctx_t ctx;
4849 	if_t ifp;
4850 	if_softc_ctx_t scctx;
4851 	int i;
4852 	void *sc;
4853 	uint16_t main_txq;
4854 	uint16_t main_rxq;
4855 
4856 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4857 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4858 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4859 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4860 		ctx->ifc_flags |= IFC_PSEUDO;
4861 
4862 	ctx->ifc_sctx = sctx;
4863 	ctx->ifc_softc = sc;
4864 	ctx->ifc_dev = dev;
4865 
4866 	if ((err = iflib_register(ctx)) != 0) {
4867 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4868 		goto fail_ctx_free;
4869 	}
4870 	iflib_add_device_sysctl_pre(ctx);
4871 
4872 	scctx = &ctx->ifc_softc_ctx;
4873 	ifp = ctx->ifc_ifp;
4874 
4875 	iflib_reset_qvalues(ctx);
4876 	CTX_LOCK(ctx);
4877 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4878 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4879 		goto fail_unlock;
4880 	}
4881 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4882 		ether_gen_addr(ifp, &ctx->ifc_mac);
4883 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4884 								clctx->cc_params)) != 0) {
4885 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4886 		goto fail_ctx_free;
4887 	}
4888 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4889 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4890 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4891 
4892 #ifdef INVARIANTS
4893 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4894 		MPASS(scctx->isc_tx_csum_flags);
4895 #endif
4896 
4897 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4898 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4899 
4900 	ifp->if_flags |= IFF_NOGROUP;
4901 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4902 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4903 
4904 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4905 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4906 			goto fail_detach;
4907 		}
4908 		*ctxp = ctx;
4909 
4910 		/*
4911 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4912 		 * This must appear after the call to ether_ifattach() because
4913 		 * ether_ifattach() sets if_hdrlen to the default value.
4914 		 */
4915 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4916 			if_setifheaderlen(ifp,
4917 			    sizeof(struct ether_vlan_header));
4918 
4919 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4920 		iflib_add_device_sysctl_post(ctx);
4921 		ctx->ifc_flags |= IFC_INIT_DONE;
4922 		return (0);
4923 	}
4924 	_iflib_pre_assert(scctx);
4925 	ctx->ifc_txrx = *scctx->isc_txrx;
4926 
4927 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4928 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4929 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4930 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4931 
4932 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4933 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4934 
4935 	/* XXX change for per-queue sizes */
4936 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4937 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4938 
4939 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4940 	    MAX_SINGLE_PACKET_FRACTION)
4941 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4942 		    MAX_SINGLE_PACKET_FRACTION);
4943 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4944 	    MAX_SINGLE_PACKET_FRACTION)
4945 		scctx->isc_tx_tso_segments_max = max(1,
4946 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4947 
4948 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4949 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4950 		/*
4951 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4952 		 * but some MACs do.
4953 		 */
4954 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4955 		    IP_MAXPACKET));
4956 		/*
4957 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4958 		 * into account.  In the worst case, each of these calls will
4959 		 * add another mbuf and, thus, the requirement for another DMA
4960 		 * segment.  So for best performance, it doesn't make sense to
4961 		 * advertize a maximum of TSO segments that typically will
4962 		 * require defragmentation in iflib_encap().
4963 		 */
4964 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4965 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4966 	}
4967 	if (scctx->isc_rss_table_size == 0)
4968 		scctx->isc_rss_table_size = 64;
4969 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4970 
4971 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4972 	/* XXX format name */
4973 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4974 	    NULL, NULL, "admin");
4975 
4976 	/* XXX --- can support > 1 -- but keep it simple for now */
4977 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4978 
4979 	/* Get memory for the station queues */
4980 	if ((err = iflib_queues_alloc(ctx))) {
4981 		device_printf(dev, "Unable to allocate queue memory\n");
4982 		goto fail_iflib_detach;
4983 	}
4984 
4985 	if ((err = iflib_qset_structures_setup(ctx))) {
4986 		device_printf(dev, "qset structure setup failed %d\n", err);
4987 		goto fail_queues;
4988 	}
4989 
4990 	/*
4991 	 * XXX What if anything do we want to do about interrupts?
4992 	 */
4993 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4994 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4995 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4996 		goto fail_detach;
4997 	}
4998 
4999 	/*
5000 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5001 	 * This must appear after the call to ether_ifattach() because
5002 	 * ether_ifattach() sets if_hdrlen to the default value.
5003 	 */
5004 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5005 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5006 
5007 	/* XXX handle more than one queue */
5008 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5009 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5010 
5011 	*ctxp = ctx;
5012 
5013 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5014 	iflib_add_device_sysctl_post(ctx);
5015 	ctx->ifc_flags |= IFC_INIT_DONE;
5016 	CTX_UNLOCK(ctx);
5017 
5018 	return (0);
5019 fail_detach:
5020 	ether_ifdetach(ctx->ifc_ifp);
5021 fail_queues:
5022 	iflib_tx_structures_free(ctx);
5023 	iflib_rx_structures_free(ctx);
5024 fail_iflib_detach:
5025 	IFDI_DETACH(ctx);
5026 fail_unlock:
5027 	CTX_UNLOCK(ctx);
5028 	iflib_deregister(ctx);
5029 fail_ctx_free:
5030 	free(ctx->ifc_softc, M_IFLIB);
5031 	free(ctx, M_IFLIB);
5032 	return (err);
5033 }
5034 
5035 int
5036 iflib_pseudo_deregister(if_ctx_t ctx)
5037 {
5038 	if_t ifp = ctx->ifc_ifp;
5039 	iflib_txq_t txq;
5040 	iflib_rxq_t rxq;
5041 	int i, j;
5042 	struct taskqgroup *tqg;
5043 	iflib_fl_t fl;
5044 
5045 	/* Unregister VLAN event handlers early */
5046 	iflib_unregister_vlan_handlers(ctx);
5047 
5048 	ether_ifdetach(ifp);
5049 	/* XXX drain any dependent tasks */
5050 	tqg = qgroup_if_io_tqg;
5051 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5052 		callout_drain(&txq->ift_timer);
5053 		if (txq->ift_task.gt_uniq != NULL)
5054 			taskqgroup_detach(tqg, &txq->ift_task);
5055 	}
5056 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5057 		callout_drain(&rxq->ifr_watchdog);
5058 		if (rxq->ifr_task.gt_uniq != NULL)
5059 			taskqgroup_detach(tqg, &rxq->ifr_task);
5060 
5061 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5062 			free(fl->ifl_rx_bitmap, M_IFLIB);
5063 	}
5064 	tqg = qgroup_if_config_tqg;
5065 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5066 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5067 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5068 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5069 
5070 	iflib_tx_structures_free(ctx);
5071 	iflib_rx_structures_free(ctx);
5072 
5073 	iflib_deregister(ctx);
5074 
5075 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5076 		free(ctx->ifc_softc, M_IFLIB);
5077 	free(ctx, M_IFLIB);
5078 	return (0);
5079 }
5080 
5081 int
5082 iflib_device_attach(device_t dev)
5083 {
5084 	if_ctx_t ctx;
5085 	if_shared_ctx_t sctx;
5086 
5087 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5088 		return (ENOTSUP);
5089 
5090 	pci_enable_busmaster(dev);
5091 
5092 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5093 }
5094 
5095 int
5096 iflib_device_deregister(if_ctx_t ctx)
5097 {
5098 	if_t ifp = ctx->ifc_ifp;
5099 	iflib_txq_t txq;
5100 	iflib_rxq_t rxq;
5101 	device_t dev = ctx->ifc_dev;
5102 	int i, j;
5103 	struct taskqgroup *tqg;
5104 	iflib_fl_t fl;
5105 
5106 	/* Make sure VLANS are not using driver */
5107 	if (if_vlantrunkinuse(ifp)) {
5108 		device_printf(dev, "Vlan in use, detach first\n");
5109 		return (EBUSY);
5110 	}
5111 #ifdef PCI_IOV
5112 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5113 		device_printf(dev, "SR-IOV in use; detach first.\n");
5114 		return (EBUSY);
5115 	}
5116 #endif
5117 
5118 	STATE_LOCK(ctx);
5119 	ctx->ifc_flags |= IFC_IN_DETACH;
5120 	STATE_UNLOCK(ctx);
5121 
5122 	/* Unregister VLAN handlers before calling iflib_stop() */
5123 	iflib_unregister_vlan_handlers(ctx);
5124 
5125 	iflib_netmap_detach(ifp);
5126 	ether_ifdetach(ifp);
5127 
5128 	CTX_LOCK(ctx);
5129 	iflib_stop(ctx);
5130 	CTX_UNLOCK(ctx);
5131 
5132 	iflib_rem_pfil(ctx);
5133 	if (ctx->ifc_led_dev != NULL)
5134 		led_destroy(ctx->ifc_led_dev);
5135 	/* XXX drain any dependent tasks */
5136 	tqg = qgroup_if_io_tqg;
5137 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5138 		callout_drain(&txq->ift_timer);
5139 		if (txq->ift_task.gt_uniq != NULL)
5140 			taskqgroup_detach(tqg, &txq->ift_task);
5141 	}
5142 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5143 		if (rxq->ifr_task.gt_uniq != NULL)
5144 			taskqgroup_detach(tqg, &rxq->ifr_task);
5145 
5146 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5147 			free(fl->ifl_rx_bitmap, M_IFLIB);
5148 	}
5149 	tqg = qgroup_if_config_tqg;
5150 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5151 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5152 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5153 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5154 	CTX_LOCK(ctx);
5155 	IFDI_DETACH(ctx);
5156 	CTX_UNLOCK(ctx);
5157 
5158 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5159 	iflib_free_intr_mem(ctx);
5160 
5161 	bus_generic_detach(dev);
5162 
5163 	iflib_tx_structures_free(ctx);
5164 	iflib_rx_structures_free(ctx);
5165 
5166 	iflib_deregister(ctx);
5167 
5168 	device_set_softc(ctx->ifc_dev, NULL);
5169 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5170 		free(ctx->ifc_softc, M_IFLIB);
5171 	unref_ctx_core_offset(ctx);
5172 	free(ctx, M_IFLIB);
5173 	return (0);
5174 }
5175 
5176 static void
5177 iflib_free_intr_mem(if_ctx_t ctx)
5178 {
5179 
5180 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5181 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5182 	}
5183 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5184 		pci_release_msi(ctx->ifc_dev);
5185 	}
5186 	if (ctx->ifc_msix_mem != NULL) {
5187 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5188 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5189 		ctx->ifc_msix_mem = NULL;
5190 	}
5191 }
5192 
5193 int
5194 iflib_device_detach(device_t dev)
5195 {
5196 	if_ctx_t ctx = device_get_softc(dev);
5197 
5198 	return (iflib_device_deregister(ctx));
5199 }
5200 
5201 int
5202 iflib_device_suspend(device_t dev)
5203 {
5204 	if_ctx_t ctx = device_get_softc(dev);
5205 
5206 	CTX_LOCK(ctx);
5207 	IFDI_SUSPEND(ctx);
5208 	CTX_UNLOCK(ctx);
5209 
5210 	return bus_generic_suspend(dev);
5211 }
5212 int
5213 iflib_device_shutdown(device_t dev)
5214 {
5215 	if_ctx_t ctx = device_get_softc(dev);
5216 
5217 	CTX_LOCK(ctx);
5218 	IFDI_SHUTDOWN(ctx);
5219 	CTX_UNLOCK(ctx);
5220 
5221 	return bus_generic_suspend(dev);
5222 }
5223 
5224 
5225 int
5226 iflib_device_resume(device_t dev)
5227 {
5228 	if_ctx_t ctx = device_get_softc(dev);
5229 	iflib_txq_t txq = ctx->ifc_txqs;
5230 
5231 	CTX_LOCK(ctx);
5232 	IFDI_RESUME(ctx);
5233 	iflib_if_init_locked(ctx);
5234 	CTX_UNLOCK(ctx);
5235 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5236 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5237 
5238 	return (bus_generic_resume(dev));
5239 }
5240 
5241 int
5242 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5243 {
5244 	int error;
5245 	if_ctx_t ctx = device_get_softc(dev);
5246 
5247 	CTX_LOCK(ctx);
5248 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5249 	CTX_UNLOCK(ctx);
5250 
5251 	return (error);
5252 }
5253 
5254 void
5255 iflib_device_iov_uninit(device_t dev)
5256 {
5257 	if_ctx_t ctx = device_get_softc(dev);
5258 
5259 	CTX_LOCK(ctx);
5260 	IFDI_IOV_UNINIT(ctx);
5261 	CTX_UNLOCK(ctx);
5262 }
5263 
5264 int
5265 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5266 {
5267 	int error;
5268 	if_ctx_t ctx = device_get_softc(dev);
5269 
5270 	CTX_LOCK(ctx);
5271 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5272 	CTX_UNLOCK(ctx);
5273 
5274 	return (error);
5275 }
5276 
5277 /*********************************************************************
5278  *
5279  *  MODULE FUNCTION DEFINITIONS
5280  *
5281  **********************************************************************/
5282 
5283 /*
5284  * - Start a fast taskqueue thread for each core
5285  * - Start a taskqueue for control operations
5286  */
5287 static int
5288 iflib_module_init(void)
5289 {
5290 	return (0);
5291 }
5292 
5293 static int
5294 iflib_module_event_handler(module_t mod, int what, void *arg)
5295 {
5296 	int err;
5297 
5298 	switch (what) {
5299 	case MOD_LOAD:
5300 		if ((err = iflib_module_init()) != 0)
5301 			return (err);
5302 		break;
5303 	case MOD_UNLOAD:
5304 		return (EBUSY);
5305 	default:
5306 		return (EOPNOTSUPP);
5307 	}
5308 
5309 	return (0);
5310 }
5311 
5312 /*********************************************************************
5313  *
5314  *  PUBLIC FUNCTION DEFINITIONS
5315  *     ordered as in iflib.h
5316  *
5317  **********************************************************************/
5318 
5319 
5320 static void
5321 _iflib_assert(if_shared_ctx_t sctx)
5322 {
5323 	int i;
5324 
5325 	MPASS(sctx->isc_tx_maxsize);
5326 	MPASS(sctx->isc_tx_maxsegsize);
5327 
5328 	MPASS(sctx->isc_rx_maxsize);
5329 	MPASS(sctx->isc_rx_nsegments);
5330 	MPASS(sctx->isc_rx_maxsegsize);
5331 
5332 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5333 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5334 		MPASS(sctx->isc_nrxd_min[i]);
5335 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5336 		MPASS(sctx->isc_nrxd_max[i]);
5337 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5338 		MPASS(sctx->isc_nrxd_default[i]);
5339 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5340 	}
5341 
5342 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5343 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5344 		MPASS(sctx->isc_ntxd_min[i]);
5345 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5346 		MPASS(sctx->isc_ntxd_max[i]);
5347 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5348 		MPASS(sctx->isc_ntxd_default[i]);
5349 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5350 	}
5351 }
5352 
5353 static void
5354 _iflib_pre_assert(if_softc_ctx_t scctx)
5355 {
5356 
5357 	MPASS(scctx->isc_txrx->ift_txd_encap);
5358 	MPASS(scctx->isc_txrx->ift_txd_flush);
5359 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5360 	MPASS(scctx->isc_txrx->ift_rxd_available);
5361 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5362 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5363 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5364 }
5365 
5366 static int
5367 iflib_register(if_ctx_t ctx)
5368 {
5369 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5370 	driver_t *driver = sctx->isc_driver;
5371 	device_t dev = ctx->ifc_dev;
5372 	if_t ifp;
5373 
5374 	_iflib_assert(sctx);
5375 
5376 	CTX_LOCK_INIT(ctx);
5377 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5378 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5379 	if (ifp == NULL) {
5380 		device_printf(dev, "can not allocate ifnet structure\n");
5381 		return (ENOMEM);
5382 	}
5383 
5384 	/*
5385 	 * Initialize our context's device specific methods
5386 	 */
5387 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5388 	kobj_class_compile((kobj_class_t) driver);
5389 
5390 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5391 	if_setsoftc(ifp, ctx);
5392 	if_setdev(ifp, dev);
5393 	if_setinitfn(ifp, iflib_if_init);
5394 	if_setioctlfn(ifp, iflib_if_ioctl);
5395 #ifdef ALTQ
5396 	if_setstartfn(ifp, iflib_altq_if_start);
5397 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5398 	if_setsendqready(ifp);
5399 #else
5400 	if_settransmitfn(ifp, iflib_if_transmit);
5401 #endif
5402 	if_setqflushfn(ifp, iflib_if_qflush);
5403 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5404 
5405 	ctx->ifc_vlan_attach_event =
5406 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5407 							  EVENTHANDLER_PRI_FIRST);
5408 	ctx->ifc_vlan_detach_event =
5409 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5410 							  EVENTHANDLER_PRI_FIRST);
5411 
5412 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5413 		ctx->ifc_mediap = &ctx->ifc_media;
5414 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5415 		    iflib_media_change, iflib_media_status);
5416 	}
5417 	return (0);
5418 }
5419 
5420 static void
5421 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5422 {
5423 	/* Unregister VLAN events */
5424 	if (ctx->ifc_vlan_attach_event != NULL) {
5425 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5426 		ctx->ifc_vlan_attach_event = NULL;
5427 	}
5428 	if (ctx->ifc_vlan_detach_event != NULL) {
5429 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5430 		ctx->ifc_vlan_detach_event = NULL;
5431 	}
5432 
5433 }
5434 
5435 static void
5436 iflib_deregister(if_ctx_t ctx)
5437 {
5438 	if_t ifp = ctx->ifc_ifp;
5439 
5440 	/* Remove all media */
5441 	ifmedia_removeall(&ctx->ifc_media);
5442 
5443 	/* Ensure that VLAN event handlers are unregistered */
5444 	iflib_unregister_vlan_handlers(ctx);
5445 
5446 	/* Release kobject reference */
5447 	kobj_delete((kobj_t) ctx, NULL);
5448 
5449 	/* Free the ifnet structure */
5450 	if_free(ifp);
5451 
5452 	STATE_LOCK_DESTROY(ctx);
5453 
5454 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5455 	CTX_LOCK_DESTROY(ctx);
5456 }
5457 
5458 static int
5459 iflib_queues_alloc(if_ctx_t ctx)
5460 {
5461 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5462 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5463 	device_t dev = ctx->ifc_dev;
5464 	int nrxqsets = scctx->isc_nrxqsets;
5465 	int ntxqsets = scctx->isc_ntxqsets;
5466 	iflib_txq_t txq;
5467 	iflib_rxq_t rxq;
5468 	iflib_fl_t fl = NULL;
5469 	int i, j, cpu, err, txconf, rxconf;
5470 	iflib_dma_info_t ifdip;
5471 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5472 	uint32_t *txqsizes = scctx->isc_txqsizes;
5473 	uint8_t nrxqs = sctx->isc_nrxqs;
5474 	uint8_t ntxqs = sctx->isc_ntxqs;
5475 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5476 	caddr_t *vaddrs;
5477 	uint64_t *paddrs;
5478 
5479 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5480 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5481 
5482 	/* Allocate the TX ring struct memory */
5483 	if (!(ctx->ifc_txqs =
5484 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5485 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5486 		device_printf(dev, "Unable to allocate TX ring memory\n");
5487 		err = ENOMEM;
5488 		goto fail;
5489 	}
5490 
5491 	/* Now allocate the RX */
5492 	if (!(ctx->ifc_rxqs =
5493 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5494 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5495 		device_printf(dev, "Unable to allocate RX ring memory\n");
5496 		err = ENOMEM;
5497 		goto rx_fail;
5498 	}
5499 
5500 	txq = ctx->ifc_txqs;
5501 	rxq = ctx->ifc_rxqs;
5502 
5503 	/*
5504 	 * XXX handle allocation failure
5505 	 */
5506 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5507 		/* Set up some basics */
5508 
5509 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5510 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5511 			device_printf(dev,
5512 			    "Unable to allocate TX DMA info memory\n");
5513 			err = ENOMEM;
5514 			goto err_tx_desc;
5515 		}
5516 		txq->ift_ifdi = ifdip;
5517 		for (j = 0; j < ntxqs; j++, ifdip++) {
5518 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5519 				device_printf(dev,
5520 				    "Unable to allocate TX descriptors\n");
5521 				err = ENOMEM;
5522 				goto err_tx_desc;
5523 			}
5524 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5525 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5526 		}
5527 		txq->ift_ctx = ctx;
5528 		txq->ift_id = i;
5529 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5530 			txq->ift_br_offset = 1;
5531 		} else {
5532 			txq->ift_br_offset = 0;
5533 		}
5534 		/* XXX fix this */
5535 		txq->ift_timer.c_cpu = cpu;
5536 
5537 		if (iflib_txsd_alloc(txq)) {
5538 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5539 			err = ENOMEM;
5540 			goto err_tx_desc;
5541 		}
5542 
5543 		/* Initialize the TX lock */
5544 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5545 		    device_get_nameunit(dev), txq->ift_id);
5546 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5547 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5548 
5549 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5550 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5551 		if (err) {
5552 			/* XXX free any allocated rings */
5553 			device_printf(dev, "Unable to allocate buf_ring\n");
5554 			goto err_tx_desc;
5555 		}
5556 	}
5557 
5558 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5559 		/* Set up some basics */
5560 		callout_init(&rxq->ifr_watchdog, 1);
5561 
5562 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5563 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5564 			device_printf(dev,
5565 			    "Unable to allocate RX DMA info memory\n");
5566 			err = ENOMEM;
5567 			goto err_tx_desc;
5568 		}
5569 
5570 		rxq->ifr_ifdi = ifdip;
5571 		/* XXX this needs to be changed if #rx queues != #tx queues */
5572 		rxq->ifr_ntxqirq = 1;
5573 		rxq->ifr_txqid[0] = i;
5574 		for (j = 0; j < nrxqs; j++, ifdip++) {
5575 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5576 				device_printf(dev,
5577 				    "Unable to allocate RX descriptors\n");
5578 				err = ENOMEM;
5579 				goto err_tx_desc;
5580 			}
5581 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5582 		}
5583 		rxq->ifr_ctx = ctx;
5584 		rxq->ifr_id = i;
5585 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5586 			rxq->ifr_fl_offset = 1;
5587 		} else {
5588 			rxq->ifr_fl_offset = 0;
5589 		}
5590 		rxq->ifr_nfl = nfree_lists;
5591 		if (!(fl =
5592 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5593 			device_printf(dev, "Unable to allocate free list memory\n");
5594 			err = ENOMEM;
5595 			goto err_tx_desc;
5596 		}
5597 		rxq->ifr_fl = fl;
5598 		for (j = 0; j < nfree_lists; j++) {
5599 			fl[j].ifl_rxq = rxq;
5600 			fl[j].ifl_id = j;
5601 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5602 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5603 		}
5604 		/* Allocate receive buffers for the ring */
5605 		if (iflib_rxsd_alloc(rxq)) {
5606 			device_printf(dev,
5607 			    "Critical Failure setting up receive buffers\n");
5608 			err = ENOMEM;
5609 			goto err_rx_desc;
5610 		}
5611 
5612 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5613 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5614 			    M_WAITOK);
5615 	}
5616 
5617 	/* TXQs */
5618 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5619 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5620 	for (i = 0; i < ntxqsets; i++) {
5621 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5622 
5623 		for (j = 0; j < ntxqs; j++, di++) {
5624 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5625 			paddrs[i*ntxqs + j] = di->idi_paddr;
5626 		}
5627 	}
5628 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5629 		device_printf(ctx->ifc_dev,
5630 		    "Unable to allocate device TX queue\n");
5631 		iflib_tx_structures_free(ctx);
5632 		free(vaddrs, M_IFLIB);
5633 		free(paddrs, M_IFLIB);
5634 		goto err_rx_desc;
5635 	}
5636 	free(vaddrs, M_IFLIB);
5637 	free(paddrs, M_IFLIB);
5638 
5639 	/* RXQs */
5640 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5641 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5642 	for (i = 0; i < nrxqsets; i++) {
5643 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5644 
5645 		for (j = 0; j < nrxqs; j++, di++) {
5646 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5647 			paddrs[i*nrxqs + j] = di->idi_paddr;
5648 		}
5649 	}
5650 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5651 		device_printf(ctx->ifc_dev,
5652 		    "Unable to allocate device RX queue\n");
5653 		iflib_tx_structures_free(ctx);
5654 		free(vaddrs, M_IFLIB);
5655 		free(paddrs, M_IFLIB);
5656 		goto err_rx_desc;
5657 	}
5658 	free(vaddrs, M_IFLIB);
5659 	free(paddrs, M_IFLIB);
5660 
5661 	return (0);
5662 
5663 /* XXX handle allocation failure changes */
5664 err_rx_desc:
5665 err_tx_desc:
5666 rx_fail:
5667 	if (ctx->ifc_rxqs != NULL)
5668 		free(ctx->ifc_rxqs, M_IFLIB);
5669 	ctx->ifc_rxqs = NULL;
5670 	if (ctx->ifc_txqs != NULL)
5671 		free(ctx->ifc_txqs, M_IFLIB);
5672 	ctx->ifc_txqs = NULL;
5673 fail:
5674 	return (err);
5675 }
5676 
5677 static int
5678 iflib_tx_structures_setup(if_ctx_t ctx)
5679 {
5680 	iflib_txq_t txq = ctx->ifc_txqs;
5681 	int i;
5682 
5683 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5684 		iflib_txq_setup(txq);
5685 
5686 	return (0);
5687 }
5688 
5689 static void
5690 iflib_tx_structures_free(if_ctx_t ctx)
5691 {
5692 	iflib_txq_t txq = ctx->ifc_txqs;
5693 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5694 	int i, j;
5695 
5696 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5697 		for (j = 0; j < sctx->isc_ntxqs; j++)
5698 			iflib_dma_free(&txq->ift_ifdi[j]);
5699 		iflib_txq_destroy(txq);
5700 	}
5701 	free(ctx->ifc_txqs, M_IFLIB);
5702 	ctx->ifc_txqs = NULL;
5703 	IFDI_QUEUES_FREE(ctx);
5704 }
5705 
5706 /*********************************************************************
5707  *
5708  *  Initialize all receive rings.
5709  *
5710  **********************************************************************/
5711 static int
5712 iflib_rx_structures_setup(if_ctx_t ctx)
5713 {
5714 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5715 	int q;
5716 #if defined(INET6) || defined(INET)
5717 	int err, i;
5718 #endif
5719 
5720 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5721 #if defined(INET6) || defined(INET)
5722 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5723 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5724 			    TCP_LRO_ENTRIES, min(1024,
5725 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5726 			if (err != 0) {
5727 				device_printf(ctx->ifc_dev,
5728 				    "LRO Initialization failed!\n");
5729 				goto fail;
5730 			}
5731 		}
5732 #endif
5733 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5734 	}
5735 	return (0);
5736 #if defined(INET6) || defined(INET)
5737 fail:
5738 	/*
5739 	 * Free LRO resources allocated so far, we will only handle
5740 	 * the rings that completed, the failing case will have
5741 	 * cleaned up for itself.  'q' failed, so its the terminus.
5742 	 */
5743 	rxq = ctx->ifc_rxqs;
5744 	for (i = 0; i < q; ++i, rxq++) {
5745 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5746 			tcp_lro_free(&rxq->ifr_lc);
5747 	}
5748 	return (err);
5749 #endif
5750 }
5751 
5752 /*********************************************************************
5753  *
5754  *  Free all receive rings.
5755  *
5756  **********************************************************************/
5757 static void
5758 iflib_rx_structures_free(if_ctx_t ctx)
5759 {
5760 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5761 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5762 	int i, j;
5763 
5764 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5765 		for (j = 0; j < sctx->isc_nrxqs; j++)
5766 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5767 		iflib_rx_sds_free(rxq);
5768 #if defined(INET6) || defined(INET)
5769 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5770 			tcp_lro_free(&rxq->ifr_lc);
5771 #endif
5772 	}
5773 	free(ctx->ifc_rxqs, M_IFLIB);
5774 	ctx->ifc_rxqs = NULL;
5775 }
5776 
5777 static int
5778 iflib_qset_structures_setup(if_ctx_t ctx)
5779 {
5780 	int err;
5781 
5782 	/*
5783 	 * It is expected that the caller takes care of freeing queues if this
5784 	 * fails.
5785 	 */
5786 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5787 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5788 		return (err);
5789 	}
5790 
5791 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5792 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5793 
5794 	return (err);
5795 }
5796 
5797 int
5798 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5799 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5800 {
5801 
5802 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5803 }
5804 
5805 #ifdef SMP
5806 static int
5807 find_nth(if_ctx_t ctx, int qid)
5808 {
5809 	cpuset_t cpus;
5810 	int i, cpuid, eqid, count;
5811 
5812 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5813 	count = CPU_COUNT(&cpus);
5814 	eqid = qid % count;
5815 	/* clear up to the qid'th bit */
5816 	for (i = 0; i < eqid; i++) {
5817 		cpuid = CPU_FFS(&cpus);
5818 		MPASS(cpuid != 0);
5819 		CPU_CLR(cpuid-1, &cpus);
5820 	}
5821 	cpuid = CPU_FFS(&cpus);
5822 	MPASS(cpuid != 0);
5823 	return (cpuid-1);
5824 }
5825 
5826 #ifdef SCHED_ULE
5827 extern struct cpu_group *cpu_top;              /* CPU topology */
5828 
5829 static int
5830 find_child_with_core(int cpu, struct cpu_group *grp)
5831 {
5832 	int i;
5833 
5834 	if (grp->cg_children == 0)
5835 		return -1;
5836 
5837 	MPASS(grp->cg_child);
5838 	for (i = 0; i < grp->cg_children; i++) {
5839 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5840 			return i;
5841 	}
5842 
5843 	return -1;
5844 }
5845 
5846 /*
5847  * Find the nth "close" core to the specified core
5848  * "close" is defined as the deepest level that shares
5849  * at least an L2 cache.  With threads, this will be
5850  * threads on the same core.  If the shared cache is L3
5851  * or higher, simply returns the same core.
5852  */
5853 static int
5854 find_close_core(int cpu, int core_offset)
5855 {
5856 	struct cpu_group *grp;
5857 	int i;
5858 	int fcpu;
5859 	cpuset_t cs;
5860 
5861 	grp = cpu_top;
5862 	if (grp == NULL)
5863 		return cpu;
5864 	i = 0;
5865 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5866 		/* If the child only has one cpu, don't descend */
5867 		if (grp->cg_child[i].cg_count <= 1)
5868 			break;
5869 		grp = &grp->cg_child[i];
5870 	}
5871 
5872 	/* If they don't share at least an L2 cache, use the same CPU */
5873 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5874 		return cpu;
5875 
5876 	/* Now pick one */
5877 	CPU_COPY(&grp->cg_mask, &cs);
5878 
5879 	/* Add the selected CPU offset to core offset. */
5880 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5881 		if (fcpu - 1 == cpu)
5882 			break;
5883 		CPU_CLR(fcpu - 1, &cs);
5884 	}
5885 	MPASS(fcpu);
5886 
5887 	core_offset += i;
5888 
5889 	CPU_COPY(&grp->cg_mask, &cs);
5890 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5891 		MPASS(CPU_FFS(&cs));
5892 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5893 	}
5894 	MPASS(CPU_FFS(&cs));
5895 	return CPU_FFS(&cs) - 1;
5896 }
5897 #else
5898 static int
5899 find_close_core(int cpu, int core_offset __unused)
5900 {
5901 	return cpu;
5902 }
5903 #endif
5904 
5905 static int
5906 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5907 {
5908 	switch (type) {
5909 	case IFLIB_INTR_TX:
5910 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5911 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5912 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5913 	case IFLIB_INTR_RX:
5914 	case IFLIB_INTR_RXTX:
5915 		/* RX queues get the specified core */
5916 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5917 	default:
5918 		return -1;
5919 	}
5920 }
5921 #else
5922 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5923 #define find_close_core(cpuid, tid)	CPU_FIRST()
5924 #define find_nth(ctx, gid)		CPU_FIRST()
5925 #endif
5926 
5927 /* Just to avoid copy/paste */
5928 static inline int
5929 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5930     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5931     const char *name)
5932 {
5933 	device_t dev;
5934 	int co, cpuid, err, tid;
5935 
5936 	dev = ctx->ifc_dev;
5937 	co = ctx->ifc_sysctl_core_offset;
5938 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5939 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5940 	cpuid = find_nth(ctx, qid + co);
5941 	tid = get_core_offset(ctx, type, qid);
5942 	if (tid < 0) {
5943 		device_printf(dev, "get_core_offset failed\n");
5944 		return (EOPNOTSUPP);
5945 	}
5946 	cpuid = find_close_core(cpuid, tid);
5947 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5948 	    name);
5949 	if (err) {
5950 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5951 		return (err);
5952 	}
5953 #ifdef notyet
5954 	if (cpuid > ctx->ifc_cpuid_highest)
5955 		ctx->ifc_cpuid_highest = cpuid;
5956 #endif
5957 	return (0);
5958 }
5959 
5960 int
5961 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5962 			iflib_intr_type_t type, driver_filter_t *filter,
5963 			void *filter_arg, int qid, const char *name)
5964 {
5965 	device_t dev;
5966 	struct grouptask *gtask;
5967 	struct taskqgroup *tqg;
5968 	iflib_filter_info_t info;
5969 	gtask_fn_t *fn;
5970 	int tqrid, err;
5971 	driver_filter_t *intr_fast;
5972 	void *q;
5973 
5974 	info = &ctx->ifc_filter_info;
5975 	tqrid = rid;
5976 
5977 	switch (type) {
5978 	/* XXX merge tx/rx for netmap? */
5979 	case IFLIB_INTR_TX:
5980 		q = &ctx->ifc_txqs[qid];
5981 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5982 		gtask = &ctx->ifc_txqs[qid].ift_task;
5983 		tqg = qgroup_if_io_tqg;
5984 		fn = _task_fn_tx;
5985 		intr_fast = iflib_fast_intr;
5986 		GROUPTASK_INIT(gtask, 0, fn, q);
5987 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5988 		break;
5989 	case IFLIB_INTR_RX:
5990 		q = &ctx->ifc_rxqs[qid];
5991 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5992 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5993 		tqg = qgroup_if_io_tqg;
5994 		fn = _task_fn_rx;
5995 		intr_fast = iflib_fast_intr;
5996 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
5997 		break;
5998 	case IFLIB_INTR_RXTX:
5999 		q = &ctx->ifc_rxqs[qid];
6000 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6001 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6002 		tqg = qgroup_if_io_tqg;
6003 		fn = _task_fn_rx;
6004 		intr_fast = iflib_fast_intr_rxtx;
6005 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6006 		break;
6007 	case IFLIB_INTR_ADMIN:
6008 		q = ctx;
6009 		tqrid = -1;
6010 		info = &ctx->ifc_filter_info;
6011 		gtask = &ctx->ifc_admin_task;
6012 		tqg = qgroup_if_config_tqg;
6013 		fn = _task_fn_admin;
6014 		intr_fast = iflib_fast_intr_ctx;
6015 		break;
6016 	default:
6017 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6018 		    __func__);
6019 		return (EINVAL);
6020 	}
6021 
6022 	info->ifi_filter = filter;
6023 	info->ifi_filter_arg = filter_arg;
6024 	info->ifi_task = gtask;
6025 	info->ifi_ctx = q;
6026 
6027 	dev = ctx->ifc_dev;
6028 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6029 	if (err != 0) {
6030 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6031 		return (err);
6032 	}
6033 	if (type == IFLIB_INTR_ADMIN)
6034 		return (0);
6035 
6036 	if (tqrid != -1) {
6037 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6038 		    q, name);
6039 		if (err)
6040 			return (err);
6041 	} else {
6042 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6043 	}
6044 
6045 	return (0);
6046 }
6047 
6048 void
6049 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6050 {
6051 	struct grouptask *gtask;
6052 	struct taskqgroup *tqg;
6053 	gtask_fn_t *fn;
6054 	void *q;
6055 	int err;
6056 
6057 	switch (type) {
6058 	case IFLIB_INTR_TX:
6059 		q = &ctx->ifc_txqs[qid];
6060 		gtask = &ctx->ifc_txqs[qid].ift_task;
6061 		tqg = qgroup_if_io_tqg;
6062 		fn = _task_fn_tx;
6063 		GROUPTASK_INIT(gtask, 0, fn, q);
6064 		break;
6065 	case IFLIB_INTR_RX:
6066 		q = &ctx->ifc_rxqs[qid];
6067 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6068 		tqg = qgroup_if_io_tqg;
6069 		fn = _task_fn_rx;
6070 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6071 		break;
6072 	case IFLIB_INTR_IOV:
6073 		q = ctx;
6074 		gtask = &ctx->ifc_vflr_task;
6075 		tqg = qgroup_if_config_tqg;
6076 		fn = _task_fn_iov;
6077 		GROUPTASK_INIT(gtask, 0, fn, q);
6078 		break;
6079 	default:
6080 		panic("unknown net intr type");
6081 	}
6082 	if (irq != NULL) {
6083 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6084 		    q, name);
6085 		if (err)
6086 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6087 			    irq->ii_res, name);
6088 	} else {
6089 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6090 	}
6091 }
6092 
6093 void
6094 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6095 {
6096 
6097 	if (irq->ii_tag)
6098 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6099 
6100 	if (irq->ii_res)
6101 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6102 		    rman_get_rid(irq->ii_res), irq->ii_res);
6103 }
6104 
6105 static int
6106 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6107 {
6108 	iflib_txq_t txq = ctx->ifc_txqs;
6109 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6110 	if_irq_t irq = &ctx->ifc_legacy_irq;
6111 	iflib_filter_info_t info;
6112 	device_t dev;
6113 	struct grouptask *gtask;
6114 	struct resource *res;
6115 	struct taskqgroup *tqg;
6116 	void *q;
6117 	int err, tqrid;
6118 	bool rx_only;
6119 
6120 	q = &ctx->ifc_rxqs[0];
6121 	info = &rxq[0].ifr_filter_info;
6122 	gtask = &rxq[0].ifr_task;
6123 	tqg = qgroup_if_io_tqg;
6124 	tqrid = *rid;
6125 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6126 
6127 	ctx->ifc_flags |= IFC_LEGACY;
6128 	info->ifi_filter = filter;
6129 	info->ifi_filter_arg = filter_arg;
6130 	info->ifi_task = gtask;
6131 	info->ifi_ctx = rx_only ? ctx : q;
6132 
6133 	dev = ctx->ifc_dev;
6134 	/* We allocate a single interrupt resource */
6135 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6136 	    iflib_fast_intr_rxtx, NULL, info, name);
6137 	if (err != 0)
6138 		return (err);
6139 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6140 	res = irq->ii_res;
6141 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6142 
6143 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6144 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6145 	    "tx");
6146 	return (0);
6147 }
6148 
6149 void
6150 iflib_led_create(if_ctx_t ctx)
6151 {
6152 
6153 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6154 	    device_get_nameunit(ctx->ifc_dev));
6155 }
6156 
6157 void
6158 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6159 {
6160 
6161 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6162 }
6163 
6164 void
6165 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6166 {
6167 
6168 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6169 }
6170 
6171 void
6172 iflib_admin_intr_deferred(if_ctx_t ctx)
6173 {
6174 #ifdef INVARIANTS
6175 	struct grouptask *gtask;
6176 
6177 	gtask = &ctx->ifc_admin_task;
6178 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6179 #endif
6180 
6181 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6182 }
6183 
6184 void
6185 iflib_iov_intr_deferred(if_ctx_t ctx)
6186 {
6187 
6188 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6189 }
6190 
6191 void
6192 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6193 {
6194 
6195 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6196 	    name);
6197 }
6198 
6199 void
6200 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6201 	const char *name)
6202 {
6203 
6204 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6205 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6206 	    name);
6207 }
6208 
6209 void
6210 iflib_config_gtask_deinit(struct grouptask *gtask)
6211 {
6212 
6213 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6214 }
6215 
6216 void
6217 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6218 {
6219 	if_t ifp = ctx->ifc_ifp;
6220 	iflib_txq_t txq = ctx->ifc_txqs;
6221 
6222 	if_setbaudrate(ifp, baudrate);
6223 	if (baudrate >= IF_Gbps(10)) {
6224 		STATE_LOCK(ctx);
6225 		ctx->ifc_flags |= IFC_PREFETCH;
6226 		STATE_UNLOCK(ctx);
6227 	}
6228 	/* If link down, disable watchdog */
6229 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6230 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6231 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6232 	}
6233 	ctx->ifc_link_state = link_state;
6234 	if_link_state_change(ifp, link_state);
6235 }
6236 
6237 static int
6238 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6239 {
6240 	int credits;
6241 #ifdef INVARIANTS
6242 	int credits_pre = txq->ift_cidx_processed;
6243 #endif
6244 
6245 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6246 	    BUS_DMASYNC_POSTREAD);
6247 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6248 		return (0);
6249 
6250 	txq->ift_processed += credits;
6251 	txq->ift_cidx_processed += credits;
6252 
6253 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6254 	if (txq->ift_cidx_processed >= txq->ift_size)
6255 		txq->ift_cidx_processed -= txq->ift_size;
6256 	return (credits);
6257 }
6258 
6259 static int
6260 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6261 {
6262 	iflib_fl_t fl;
6263 	u_int i;
6264 
6265 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6266 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6267 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6268 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6269 	    budget));
6270 }
6271 
6272 void
6273 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6274 	const char *description, if_int_delay_info_t info,
6275 	int offset, int value)
6276 {
6277 	info->iidi_ctx = ctx;
6278 	info->iidi_offset = offset;
6279 	info->iidi_value = value;
6280 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6281 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6282 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6283 	    info, 0, iflib_sysctl_int_delay, "I", description);
6284 }
6285 
6286 struct sx *
6287 iflib_ctx_lock_get(if_ctx_t ctx)
6288 {
6289 
6290 	return (&ctx->ifc_ctx_sx);
6291 }
6292 
6293 static int
6294 iflib_msix_init(if_ctx_t ctx)
6295 {
6296 	device_t dev = ctx->ifc_dev;
6297 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6298 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6299 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6300 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6301 
6302 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6303 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6304 
6305 	if (bootverbose)
6306 		device_printf(dev, "msix_init qsets capped at %d\n",
6307 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6308 
6309 	/* Override by tuneable */
6310 	if (scctx->isc_disable_msix)
6311 		goto msi;
6312 
6313 	/* First try MSI-X */
6314 	if ((msgs = pci_msix_count(dev)) == 0) {
6315 		if (bootverbose)
6316 			device_printf(dev, "MSI-X not supported or disabled\n");
6317 		goto msi;
6318 	}
6319 
6320 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6321 	/*
6322 	 * bar == -1 => "trust me I know what I'm doing"
6323 	 * Some drivers are for hardware that is so shoddily
6324 	 * documented that no one knows which bars are which
6325 	 * so the developer has to map all bars. This hack
6326 	 * allows shoddy garbage to use MSI-X in this framework.
6327 	 */
6328 	if (bar != -1) {
6329 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6330 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6331 		if (ctx->ifc_msix_mem == NULL) {
6332 			device_printf(dev, "Unable to map MSI-X table\n");
6333 			goto msi;
6334 		}
6335 	}
6336 
6337 	admincnt = sctx->isc_admin_intrcnt;
6338 #if IFLIB_DEBUG
6339 	/* use only 1 qset in debug mode */
6340 	queuemsgs = min(msgs - admincnt, 1);
6341 #else
6342 	queuemsgs = msgs - admincnt;
6343 #endif
6344 #ifdef RSS
6345 	queues = imin(queuemsgs, rss_getnumbuckets());
6346 #else
6347 	queues = queuemsgs;
6348 #endif
6349 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6350 	if (bootverbose)
6351 		device_printf(dev,
6352 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6353 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6354 #ifdef  RSS
6355 	/* If we're doing RSS, clamp at the number of RSS buckets */
6356 	if (queues > rss_getnumbuckets())
6357 		queues = rss_getnumbuckets();
6358 #endif
6359 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6360 		rx_queues = iflib_num_rx_queues;
6361 	else
6362 		rx_queues = queues;
6363 
6364 	if (rx_queues > scctx->isc_nrxqsets)
6365 		rx_queues = scctx->isc_nrxqsets;
6366 
6367 	/*
6368 	 * We want this to be all logical CPUs by default
6369 	 */
6370 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6371 		tx_queues = iflib_num_tx_queues;
6372 	else
6373 		tx_queues = mp_ncpus;
6374 
6375 	if (tx_queues > scctx->isc_ntxqsets)
6376 		tx_queues = scctx->isc_ntxqsets;
6377 
6378 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6379 #ifdef INVARIANTS
6380 		if (tx_queues != rx_queues)
6381 			device_printf(dev,
6382 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6383 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6384 #endif
6385 		tx_queues = min(rx_queues, tx_queues);
6386 		rx_queues = min(rx_queues, tx_queues);
6387 	}
6388 
6389 	vectors = rx_queues + admincnt;
6390 	if (msgs < vectors) {
6391 		device_printf(dev,
6392 		    "insufficient number of MSI-X vectors "
6393 		    "(supported %d, need %d)\n", msgs, vectors);
6394 		goto msi;
6395 	}
6396 
6397 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6398 	    tx_queues);
6399 	msgs = vectors;
6400 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6401 		if (vectors != msgs) {
6402 			device_printf(dev,
6403 			    "Unable to allocate sufficient MSI-X vectors "
6404 			    "(got %d, need %d)\n", vectors, msgs);
6405 			pci_release_msi(dev);
6406 			if (bar != -1) {
6407 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6408 				    ctx->ifc_msix_mem);
6409 				ctx->ifc_msix_mem = NULL;
6410 			}
6411 			goto msi;
6412 		}
6413 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6414 		    vectors);
6415 		scctx->isc_vectors = vectors;
6416 		scctx->isc_nrxqsets = rx_queues;
6417 		scctx->isc_ntxqsets = tx_queues;
6418 		scctx->isc_intr = IFLIB_INTR_MSIX;
6419 
6420 		return (vectors);
6421 	} else {
6422 		device_printf(dev,
6423 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6424 		    err);
6425 		if (bar != -1) {
6426 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6427 			    ctx->ifc_msix_mem);
6428 			ctx->ifc_msix_mem = NULL;
6429 		}
6430 	}
6431 
6432 msi:
6433 	vectors = pci_msi_count(dev);
6434 	scctx->isc_nrxqsets = 1;
6435 	scctx->isc_ntxqsets = 1;
6436 	scctx->isc_vectors = vectors;
6437 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6438 		device_printf(dev,"Using an MSI interrupt\n");
6439 		scctx->isc_intr = IFLIB_INTR_MSI;
6440 	} else {
6441 		scctx->isc_vectors = 1;
6442 		device_printf(dev,"Using a Legacy interrupt\n");
6443 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6444 	}
6445 
6446 	return (vectors);
6447 }
6448 
6449 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6450 
6451 static int
6452 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6453 {
6454 	int rc;
6455 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6456 	struct sbuf *sb;
6457 	const char *ring_state = "UNKNOWN";
6458 
6459 	/* XXX needed ? */
6460 	rc = sysctl_wire_old_buffer(req, 0);
6461 	MPASS(rc == 0);
6462 	if (rc != 0)
6463 		return (rc);
6464 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6465 	MPASS(sb != NULL);
6466 	if (sb == NULL)
6467 		return (ENOMEM);
6468 	if (state[3] <= 3)
6469 		ring_state = ring_states[state[3]];
6470 
6471 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6472 		    state[0], state[1], state[2], ring_state);
6473 	rc = sbuf_finish(sb);
6474 	sbuf_delete(sb);
6475         return(rc);
6476 }
6477 
6478 enum iflib_ndesc_handler {
6479 	IFLIB_NTXD_HANDLER,
6480 	IFLIB_NRXD_HANDLER,
6481 };
6482 
6483 static int
6484 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6485 {
6486 	if_ctx_t ctx = (void *)arg1;
6487 	enum iflib_ndesc_handler type = arg2;
6488 	char buf[256] = {0};
6489 	qidx_t *ndesc;
6490 	char *p, *next;
6491 	int nqs, rc, i;
6492 
6493 	nqs = 8;
6494 	switch(type) {
6495 	case IFLIB_NTXD_HANDLER:
6496 		ndesc = ctx->ifc_sysctl_ntxds;
6497 		if (ctx->ifc_sctx)
6498 			nqs = ctx->ifc_sctx->isc_ntxqs;
6499 		break;
6500 	case IFLIB_NRXD_HANDLER:
6501 		ndesc = ctx->ifc_sysctl_nrxds;
6502 		if (ctx->ifc_sctx)
6503 			nqs = ctx->ifc_sctx->isc_nrxqs;
6504 		break;
6505 	default:
6506 		printf("%s: unhandled type\n", __func__);
6507 		return (EINVAL);
6508 	}
6509 	if (nqs == 0)
6510 		nqs = 8;
6511 
6512 	for (i=0; i<8; i++) {
6513 		if (i >= nqs)
6514 			break;
6515 		if (i)
6516 			strcat(buf, ",");
6517 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6518 	}
6519 
6520 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6521 	if (rc || req->newptr == NULL)
6522 		return rc;
6523 
6524 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6525 	    i++, p = strsep(&next, " ,")) {
6526 		ndesc[i] = strtoul(p, NULL, 10);
6527 	}
6528 
6529 	return(rc);
6530 }
6531 
6532 #define NAME_BUFLEN 32
6533 static void
6534 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6535 {
6536         device_t dev = iflib_get_dev(ctx);
6537 	struct sysctl_oid_list *child, *oid_list;
6538 	struct sysctl_ctx_list *ctx_list;
6539 	struct sysctl_oid *node;
6540 
6541 	ctx_list = device_get_sysctl_ctx(dev);
6542 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6543 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6544 						      CTLFLAG_RD, NULL, "IFLIB fields");
6545 	oid_list = SYSCTL_CHILDREN(node);
6546 
6547 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6548 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6549 		       "driver version");
6550 
6551 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6552 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6553 			"# of txqs to use, 0 => use default #");
6554 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6555 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6556 			"# of rxqs to use, 0 => use default #");
6557 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6558 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6559                        "permit #txq != #rxq");
6560 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6561                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6562                       "disable MSI-X (default 0)");
6563 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6564 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6565 		       "set the RX budget");
6566 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6567 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6568 		       "cause TX to abdicate instead of running to completion");
6569 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6570 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6571 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6572 		       "offset to start using cores at");
6573 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6574 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6575 		       "use separate cores for TX and RX");
6576 
6577 	/* XXX change for per-queue sizes */
6578 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6579 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6580                        mp_ndesc_handler, "A",
6581 		       "list of # of TX descriptors to use, 0 = use default #");
6582 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6583 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6584                        mp_ndesc_handler, "A",
6585 		       "list of # of RX descriptors to use, 0 = use default #");
6586 }
6587 
6588 static void
6589 iflib_add_device_sysctl_post(if_ctx_t ctx)
6590 {
6591 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6592 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6593         device_t dev = iflib_get_dev(ctx);
6594 	struct sysctl_oid_list *child;
6595 	struct sysctl_ctx_list *ctx_list;
6596 	iflib_fl_t fl;
6597 	iflib_txq_t txq;
6598 	iflib_rxq_t rxq;
6599 	int i, j;
6600 	char namebuf[NAME_BUFLEN];
6601 	char *qfmt;
6602 	struct sysctl_oid *queue_node, *fl_node, *node;
6603 	struct sysctl_oid_list *queue_list, *fl_list;
6604 	ctx_list = device_get_sysctl_ctx(dev);
6605 
6606 	node = ctx->ifc_sysctl_node;
6607 	child = SYSCTL_CHILDREN(node);
6608 
6609 	if (scctx->isc_ntxqsets > 100)
6610 		qfmt = "txq%03d";
6611 	else if (scctx->isc_ntxqsets > 10)
6612 		qfmt = "txq%02d";
6613 	else
6614 		qfmt = "txq%d";
6615 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6616 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6617 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6618 					     CTLFLAG_RD, NULL, "Queue Name");
6619 		queue_list = SYSCTL_CHILDREN(queue_node);
6620 #if MEMORY_LOGGING
6621 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6622 				CTLFLAG_RD,
6623 				&txq->ift_dequeued, "total mbufs freed");
6624 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6625 				CTLFLAG_RD,
6626 				&txq->ift_enqueued, "total mbufs enqueued");
6627 #endif
6628 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6629 				   CTLFLAG_RD,
6630 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6631 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6632 				   CTLFLAG_RD,
6633 				   &txq->ift_pullups, "# of times m_pullup was called");
6634 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6635 				   CTLFLAG_RD,
6636 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6637 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6638 				   CTLFLAG_RD,
6639 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6640 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6641 				   CTLFLAG_RD,
6642 				   &txq->ift_map_failed, "# of times DMA map failed");
6643 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6644 				   CTLFLAG_RD,
6645 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6646 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6647 				   CTLFLAG_RD,
6648 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6649 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6650 				   CTLFLAG_RD,
6651 				   &txq->ift_pidx, 1, "Producer Index");
6652 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6653 				   CTLFLAG_RD,
6654 				   &txq->ift_cidx, 1, "Consumer Index");
6655 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6656 				   CTLFLAG_RD,
6657 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6658 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6659 				   CTLFLAG_RD,
6660 				   &txq->ift_in_use, 1, "descriptors in use");
6661 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6662 				   CTLFLAG_RD,
6663 				   &txq->ift_processed, "descriptors procesed for clean");
6664 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6665 				   CTLFLAG_RD,
6666 				   &txq->ift_cleaned, "total cleaned");
6667 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6668 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6669 				0, mp_ring_state_handler, "A", "soft ring state");
6670 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6671 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6672 				       "# of enqueues to the mp_ring for this queue");
6673 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6674 				       CTLFLAG_RD, &txq->ift_br->drops,
6675 				       "# of drops in the mp_ring for this queue");
6676 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6677 				       CTLFLAG_RD, &txq->ift_br->starts,
6678 				       "# of normal consumer starts in the mp_ring for this queue");
6679 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6680 				       CTLFLAG_RD, &txq->ift_br->stalls,
6681 					       "# of consumer stalls in the mp_ring for this queue");
6682 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6683 			       CTLFLAG_RD, &txq->ift_br->restarts,
6684 				       "# of consumer restarts in the mp_ring for this queue");
6685 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6686 				       CTLFLAG_RD, &txq->ift_br->abdications,
6687 				       "# of consumer abdications in the mp_ring for this queue");
6688 	}
6689 
6690 	if (scctx->isc_nrxqsets > 100)
6691 		qfmt = "rxq%03d";
6692 	else if (scctx->isc_nrxqsets > 10)
6693 		qfmt = "rxq%02d";
6694 	else
6695 		qfmt = "rxq%d";
6696 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6697 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6698 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6699 					     CTLFLAG_RD, NULL, "Queue Name");
6700 		queue_list = SYSCTL_CHILDREN(queue_node);
6701 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6702 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6703 				       CTLFLAG_RD,
6704 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6705 		}
6706 
6707 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6708 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6709 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6710 						     CTLFLAG_RD, NULL, "freelist Name");
6711 			fl_list = SYSCTL_CHILDREN(fl_node);
6712 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6713 				       CTLFLAG_RD,
6714 				       &fl->ifl_pidx, 1, "Producer Index");
6715 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6716 				       CTLFLAG_RD,
6717 				       &fl->ifl_cidx, 1, "Consumer Index");
6718 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6719 				       CTLFLAG_RD,
6720 				       &fl->ifl_credits, 1, "credits available");
6721 #if MEMORY_LOGGING
6722 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6723 					CTLFLAG_RD,
6724 					&fl->ifl_m_enqueued, "mbufs allocated");
6725 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6726 					CTLFLAG_RD,
6727 					&fl->ifl_m_dequeued, "mbufs freed");
6728 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6729 					CTLFLAG_RD,
6730 					&fl->ifl_cl_enqueued, "clusters allocated");
6731 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6732 					CTLFLAG_RD,
6733 					&fl->ifl_cl_dequeued, "clusters freed");
6734 #endif
6735 
6736 		}
6737 	}
6738 
6739 }
6740 
6741 void
6742 iflib_request_reset(if_ctx_t ctx)
6743 {
6744 
6745 	STATE_LOCK(ctx);
6746 	ctx->ifc_flags |= IFC_DO_RESET;
6747 	STATE_UNLOCK(ctx);
6748 }
6749 
6750 #ifndef __NO_STRICT_ALIGNMENT
6751 static struct mbuf *
6752 iflib_fixup_rx(struct mbuf *m)
6753 {
6754 	struct mbuf *n;
6755 
6756 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6757 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6758 		m->m_data += ETHER_HDR_LEN;
6759 		n = m;
6760 	} else {
6761 		MGETHDR(n, M_NOWAIT, MT_DATA);
6762 		if (n == NULL) {
6763 			m_freem(m);
6764 			return (NULL);
6765 		}
6766 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6767 		m->m_data += ETHER_HDR_LEN;
6768 		m->m_len -= ETHER_HDR_LEN;
6769 		n->m_len = ETHER_HDR_LEN;
6770 		M_MOVE_PKTHDR(n, m);
6771 		n->m_next = m;
6772 	}
6773 	return (n);
6774 }
6775 #endif
6776 
6777 #ifdef DEBUGNET
6778 static void
6779 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6780 {
6781 	if_ctx_t ctx;
6782 
6783 	ctx = if_getsoftc(ifp);
6784 	CTX_LOCK(ctx);
6785 	*nrxr = NRXQSETS(ctx);
6786 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6787 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6788 	CTX_UNLOCK(ctx);
6789 }
6790 
6791 static void
6792 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6793 {
6794 	if_ctx_t ctx;
6795 	if_softc_ctx_t scctx;
6796 	iflib_fl_t fl;
6797 	iflib_rxq_t rxq;
6798 	int i, j;
6799 
6800 	ctx = if_getsoftc(ifp);
6801 	scctx = &ctx->ifc_softc_ctx;
6802 
6803 	switch (event) {
6804 	case DEBUGNET_START:
6805 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6806 			rxq = &ctx->ifc_rxqs[i];
6807 			for (j = 0; j < rxq->ifr_nfl; j++) {
6808 				fl = rxq->ifr_fl;
6809 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6810 			}
6811 		}
6812 		iflib_no_tx_batch = 1;
6813 		break;
6814 	default:
6815 		break;
6816 	}
6817 }
6818 
6819 static int
6820 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6821 {
6822 	if_ctx_t ctx;
6823 	iflib_txq_t txq;
6824 	int error;
6825 
6826 	ctx = if_getsoftc(ifp);
6827 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6828 	    IFF_DRV_RUNNING)
6829 		return (EBUSY);
6830 
6831 	txq = &ctx->ifc_txqs[0];
6832 	error = iflib_encap(txq, &m);
6833 	if (error == 0)
6834 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6835 	return (error);
6836 }
6837 
6838 static int
6839 iflib_debugnet_poll(if_t ifp, int count)
6840 {
6841 	struct epoch_tracker et;
6842 	if_ctx_t ctx;
6843 	if_softc_ctx_t scctx;
6844 	iflib_txq_t txq;
6845 	int i;
6846 
6847 	ctx = if_getsoftc(ifp);
6848 	scctx = &ctx->ifc_softc_ctx;
6849 
6850 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6851 	    IFF_DRV_RUNNING)
6852 		return (EBUSY);
6853 
6854 	txq = &ctx->ifc_txqs[0];
6855 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6856 
6857 	NET_EPOCH_ENTER(et);
6858 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6859 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6860 	NET_EPOCH_EXIT(et);
6861 	return (0);
6862 }
6863 #endif /* DEBUGNET */
6864