xref: /freebsd/sys/net/iflib.c (revision 732a02b4e77866604a120a275c082bb6221bd2ff)
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 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 
480 static inline void
481 pkt_info_zero(if_pkt_info_t pi)
482 {
483 	if_pkt_info_pad_t pi_pad;
484 
485 	pi_pad = (if_pkt_info_pad_t)pi;
486 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
487 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
488 #ifndef __LP64__
489 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
490 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
491 #endif
492 }
493 
494 static device_method_t iflib_pseudo_methods[] = {
495 	DEVMETHOD(device_attach, noop_attach),
496 	DEVMETHOD(device_detach, iflib_pseudo_detach),
497 	DEVMETHOD_END
498 };
499 
500 driver_t iflib_pseudodriver = {
501 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
502 };
503 
504 static inline void
505 rxd_info_zero(if_rxd_info_t ri)
506 {
507 	if_rxd_info_pad_t ri_pad;
508 	int i;
509 
510 	ri_pad = (if_rxd_info_pad_t)ri;
511 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
512 		ri_pad->rxd_val[i] = 0;
513 		ri_pad->rxd_val[i+1] = 0;
514 		ri_pad->rxd_val[i+2] = 0;
515 		ri_pad->rxd_val[i+3] = 0;
516 	}
517 #ifdef __LP64__
518 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
519 #endif
520 }
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA (bus_addr_t)-1
527 
528 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
531 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
534 
535 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
536 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
537 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
538 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
539 
540 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
541 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
542 
543 void
544 iflib_set_detach(if_ctx_t ctx)
545 {
546 	STATE_LOCK(ctx);
547 	ctx->ifc_flags |= IFC_IN_DETACH;
548 	STATE_UNLOCK(ctx);
549 }
550 
551 /* Our boot-time initialization hook */
552 static int	iflib_module_event_handler(module_t, int, void *);
553 
554 static moduledata_t iflib_moduledata = {
555 	"iflib",
556 	iflib_module_event_handler,
557 	NULL
558 };
559 
560 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
561 MODULE_VERSION(iflib, 1);
562 
563 MODULE_DEPEND(iflib, pci, 1, 1, 1);
564 MODULE_DEPEND(iflib, ether, 1, 1, 1);
565 
566 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
567 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
568 
569 #ifndef IFLIB_DEBUG_COUNTERS
570 #ifdef INVARIANTS
571 #define IFLIB_DEBUG_COUNTERS 1
572 #else
573 #define IFLIB_DEBUG_COUNTERS 0
574 #endif /* !INVARIANTS */
575 #endif
576 
577 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
578     "iflib driver parameters");
579 
580 /*
581  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
582  */
583 static int iflib_min_tx_latency = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
585 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
586 static int iflib_no_tx_batch = 0;
587 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
588 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
589 
590 
591 #if IFLIB_DEBUG_COUNTERS
592 
593 static int iflib_tx_seen;
594 static int iflib_tx_sent;
595 static int iflib_tx_encap;
596 static int iflib_rx_allocs;
597 static int iflib_fl_refills;
598 static int iflib_fl_refills_large;
599 static int iflib_tx_frees;
600 
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
602 		   &iflib_tx_seen, 0, "# TX mbufs seen");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
604 		   &iflib_tx_sent, 0, "# TX mbufs sent");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
606 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
608 		   &iflib_tx_frees, 0, "# TX frees");
609 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
610 		   &iflib_rx_allocs, 0, "# RX allocations");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
612 		   &iflib_fl_refills, 0, "# refills");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
614 		   &iflib_fl_refills_large, 0, "# large refills");
615 
616 
617 static int iflib_txq_drain_flushing;
618 static int iflib_txq_drain_oactive;
619 static int iflib_txq_drain_notready;
620 
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
622 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
623 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
624 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
626 		   &iflib_txq_drain_notready, 0, "# drain notready");
627 
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
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 static int
1414 iflib_fast_intr(void *arg)
1415 {
1416 	iflib_filter_info_t info = arg;
1417 	struct grouptask *gtask = info->ifi_task;
1418 	int result;
1419 
1420 	DBG_COUNTER_INC(fast_intrs);
1421 	if (info->ifi_filter != NULL) {
1422 		result = info->ifi_filter(info->ifi_filter_arg);
1423 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1424 			return (result);
1425 	}
1426 
1427 	GROUPTASK_ENQUEUE(gtask);
1428 	return (FILTER_HANDLED);
1429 }
1430 
1431 static int
1432 iflib_fast_intr_rxtx(void *arg)
1433 {
1434 	iflib_filter_info_t info = arg;
1435 	struct grouptask *gtask = info->ifi_task;
1436 	if_ctx_t ctx;
1437 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1438 	iflib_txq_t txq;
1439 	void *sc;
1440 	int i, cidx, result;
1441 	qidx_t txqid;
1442 	bool intr_enable, intr_legacy;
1443 
1444 	DBG_COUNTER_INC(fast_intrs);
1445 	if (info->ifi_filter != NULL) {
1446 		result = info->ifi_filter(info->ifi_filter_arg);
1447 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1448 			return (result);
1449 	}
1450 
1451 	ctx = rxq->ifr_ctx;
1452 	sc = ctx->ifc_softc;
1453 	intr_enable = false;
1454 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1455 	MPASS(rxq->ifr_ntxqirq);
1456 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1457 		txqid = rxq->ifr_txqid[i];
1458 		txq = &ctx->ifc_txqs[txqid];
1459 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1460 		    BUS_DMASYNC_POSTREAD);
1461 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1462 			if (intr_legacy)
1463 				intr_enable = true;
1464 			else
1465 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1466 			continue;
1467 		}
1468 		GROUPTASK_ENQUEUE(&txq->ift_task);
1469 	}
1470 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1471 		cidx = rxq->ifr_cq_cidx;
1472 	else
1473 		cidx = rxq->ifr_fl[0].ifl_cidx;
1474 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1475 		GROUPTASK_ENQUEUE(gtask);
1476 	else {
1477 		if (intr_legacy)
1478 			intr_enable = true;
1479 		else
1480 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1481 		DBG_COUNTER_INC(rx_intr_enables);
1482 	}
1483 	if (intr_enable)
1484 		IFDI_INTR_ENABLE(ctx);
1485 	return (FILTER_HANDLED);
1486 }
1487 
1488 
1489 static int
1490 iflib_fast_intr_ctx(void *arg)
1491 {
1492 	iflib_filter_info_t info = arg;
1493 	struct grouptask *gtask = info->ifi_task;
1494 	int result;
1495 
1496 	DBG_COUNTER_INC(fast_intrs);
1497 	if (info->ifi_filter != NULL) {
1498 		result = info->ifi_filter(info->ifi_filter_arg);
1499 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1500 			return (result);
1501 	}
1502 
1503 	GROUPTASK_ENQUEUE(gtask);
1504 	return (FILTER_HANDLED);
1505 }
1506 
1507 static int
1508 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1509 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1510 		 const char *name)
1511 {
1512 	struct resource *res;
1513 	void *tag = NULL;
1514 	device_t dev = ctx->ifc_dev;
1515 	int flags, i, rc;
1516 
1517 	flags = RF_ACTIVE;
1518 	if (ctx->ifc_flags & IFC_LEGACY)
1519 		flags |= RF_SHAREABLE;
1520 	MPASS(rid < 512);
1521 	i = rid;
1522 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1523 	if (res == NULL) {
1524 		device_printf(dev,
1525 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1526 		return (ENOMEM);
1527 	}
1528 	irq->ii_res = res;
1529 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1530 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1531 						filter, handler, arg, &tag);
1532 	if (rc != 0) {
1533 		device_printf(dev,
1534 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1535 					  rid, name ? name : "unknown", rc);
1536 		return (rc);
1537 	} else if (name)
1538 		bus_describe_intr(dev, res, tag, "%s", name);
1539 
1540 	irq->ii_tag = tag;
1541 	return (0);
1542 }
1543 
1544 /*********************************************************************
1545  *
1546  *  Allocate DMA resources for TX buffers as well as memory for the TX
1547  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1548  *  iflib_sw_tx_desc_array structure, storing all the information that
1549  *  is needed to transmit a packet on the wire.  This is called only
1550  *  once at attach, setup is done every reset.
1551  *
1552  **********************************************************************/
1553 static int
1554 iflib_txsd_alloc(iflib_txq_t txq)
1555 {
1556 	if_ctx_t ctx = txq->ift_ctx;
1557 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1558 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1559 	device_t dev = ctx->ifc_dev;
1560 	bus_size_t tsomaxsize;
1561 	int err, nsegments, ntsosegments;
1562 	bool tso;
1563 
1564 	nsegments = scctx->isc_tx_nsegments;
1565 	ntsosegments = scctx->isc_tx_tso_segments_max;
1566 	tsomaxsize = scctx->isc_tx_tso_size_max;
1567 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1568 		tsomaxsize += sizeof(struct ether_vlan_header);
1569 	MPASS(scctx->isc_ntxd[0] > 0);
1570 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1571 	MPASS(nsegments > 0);
1572 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1573 		MPASS(ntsosegments > 0);
1574 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1575 	}
1576 
1577 	/*
1578 	 * Set up DMA tags for TX buffers.
1579 	 */
1580 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1581 			       1, 0,			/* alignment, bounds */
1582 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1583 			       BUS_SPACE_MAXADDR,	/* highaddr */
1584 			       NULL, NULL,		/* filter, filterarg */
1585 			       sctx->isc_tx_maxsize,		/* maxsize */
1586 			       nsegments,	/* nsegments */
1587 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1588 			       0,			/* flags */
1589 			       NULL,			/* lockfunc */
1590 			       NULL,			/* lockfuncarg */
1591 			       &txq->ift_buf_tag))) {
1592 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1593 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1594 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1595 		goto fail;
1596 	}
1597 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1598 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1599 			       1, 0,			/* alignment, bounds */
1600 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1601 			       BUS_SPACE_MAXADDR,	/* highaddr */
1602 			       NULL, NULL,		/* filter, filterarg */
1603 			       tsomaxsize,		/* maxsize */
1604 			       ntsosegments,	/* nsegments */
1605 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1606 			       0,			/* flags */
1607 			       NULL,			/* lockfunc */
1608 			       NULL,			/* lockfuncarg */
1609 			       &txq->ift_tso_buf_tag))) {
1610 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1611 		    err);
1612 		goto fail;
1613 	}
1614 
1615 	/* Allocate memory for the TX mbuf map. */
1616 	if (!(txq->ift_sds.ifsd_m =
1617 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1618 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1619 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1620 		err = ENOMEM;
1621 		goto fail;
1622 	}
1623 
1624 	/*
1625 	 * Create the DMA maps for TX buffers.
1626 	 */
1627 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1628 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1629 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1630 		device_printf(dev,
1631 		    "Unable to allocate TX buffer DMA map memory\n");
1632 		err = ENOMEM;
1633 		goto fail;
1634 	}
1635 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1636 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1637 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1638 		device_printf(dev,
1639 		    "Unable to allocate TSO TX buffer map memory\n");
1640 		err = ENOMEM;
1641 		goto fail;
1642 	}
1643 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1644 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1645 		    &txq->ift_sds.ifsd_map[i]);
1646 		if (err != 0) {
1647 			device_printf(dev, "Unable to create TX DMA map\n");
1648 			goto fail;
1649 		}
1650 		if (!tso)
1651 			continue;
1652 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1653 		    &txq->ift_sds.ifsd_tso_map[i]);
1654 		if (err != 0) {
1655 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1656 			goto fail;
1657 		}
1658 	}
1659 	return (0);
1660 fail:
1661 	/* We free all, it handles case where we are in the middle */
1662 	iflib_tx_structures_free(ctx);
1663 	return (err);
1664 }
1665 
1666 static void
1667 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1668 {
1669 	bus_dmamap_t map;
1670 
1671 	if (txq->ift_sds.ifsd_map != NULL) {
1672 		map = txq->ift_sds.ifsd_map[i];
1673 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1674 		bus_dmamap_unload(txq->ift_buf_tag, map);
1675 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1676 		txq->ift_sds.ifsd_map[i] = NULL;
1677 	}
1678 
1679 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1680 		map = txq->ift_sds.ifsd_tso_map[i];
1681 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1682 		    BUS_DMASYNC_POSTWRITE);
1683 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1684 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1685 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1686 	}
1687 }
1688 
1689 static void
1690 iflib_txq_destroy(iflib_txq_t txq)
1691 {
1692 	if_ctx_t ctx = txq->ift_ctx;
1693 
1694 	for (int i = 0; i < txq->ift_size; i++)
1695 		iflib_txsd_destroy(ctx, txq, i);
1696 
1697 	if (txq->ift_br != NULL) {
1698 		ifmp_ring_free(txq->ift_br);
1699 		txq->ift_br = NULL;
1700 	}
1701 
1702 	mtx_destroy(&txq->ift_mtx);
1703 
1704 	if (txq->ift_sds.ifsd_map != NULL) {
1705 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1706 		txq->ift_sds.ifsd_map = NULL;
1707 	}
1708 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1709 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1710 		txq->ift_sds.ifsd_tso_map = NULL;
1711 	}
1712 	if (txq->ift_sds.ifsd_m != NULL) {
1713 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1714 		txq->ift_sds.ifsd_m = NULL;
1715 	}
1716 	if (txq->ift_buf_tag != NULL) {
1717 		bus_dma_tag_destroy(txq->ift_buf_tag);
1718 		txq->ift_buf_tag = NULL;
1719 	}
1720 	if (txq->ift_tso_buf_tag != NULL) {
1721 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1722 		txq->ift_tso_buf_tag = NULL;
1723 	}
1724 	if (txq->ift_ifdi != NULL) {
1725 		free(txq->ift_ifdi, M_IFLIB);
1726 	}
1727 }
1728 
1729 static void
1730 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1731 {
1732 	struct mbuf **mp;
1733 
1734 	mp = &txq->ift_sds.ifsd_m[i];
1735 	if (*mp == NULL)
1736 		return;
1737 
1738 	if (txq->ift_sds.ifsd_map != NULL) {
1739 		bus_dmamap_sync(txq->ift_buf_tag,
1740 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1741 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1742 	}
1743 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1744 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1745 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1746 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1747 		    txq->ift_sds.ifsd_tso_map[i]);
1748 	}
1749 	m_free(*mp);
1750 	DBG_COUNTER_INC(tx_frees);
1751 	*mp = NULL;
1752 }
1753 
1754 static int
1755 iflib_txq_setup(iflib_txq_t txq)
1756 {
1757 	if_ctx_t ctx = txq->ift_ctx;
1758 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1759 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1760 	iflib_dma_info_t di;
1761 	int i;
1762 
1763 	/* Set number of descriptors available */
1764 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1765 	/* XXX make configurable */
1766 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1767 
1768 	/* Reset indices */
1769 	txq->ift_cidx_processed = 0;
1770 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1771 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1772 
1773 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1774 		bzero((void *)di->idi_vaddr, di->idi_size);
1775 
1776 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1777 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1778 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1779 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1780 	return (0);
1781 }
1782 
1783 /*********************************************************************
1784  *
1785  *  Allocate DMA resources for RX buffers as well as memory for the RX
1786  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1787  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1788  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1789  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1790  *  packet, the maximum number of entries we'll need is equal to the
1791  *  number of hardware receive descriptors that we've allocated.
1792  *
1793  **********************************************************************/
1794 static int
1795 iflib_rxsd_alloc(iflib_rxq_t rxq)
1796 {
1797 	if_ctx_t ctx = rxq->ifr_ctx;
1798 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1799 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1800 	device_t dev = ctx->ifc_dev;
1801 	iflib_fl_t fl;
1802 	int			err;
1803 
1804 	MPASS(scctx->isc_nrxd[0] > 0);
1805 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1806 
1807 	fl = rxq->ifr_fl;
1808 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1809 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1810 		/* Set up DMA tag for RX buffers. */
1811 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1812 					 1, 0,			/* alignment, bounds */
1813 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1814 					 BUS_SPACE_MAXADDR,	/* highaddr */
1815 					 NULL, NULL,		/* filter, filterarg */
1816 					 sctx->isc_rx_maxsize,	/* maxsize */
1817 					 sctx->isc_rx_nsegments,	/* nsegments */
1818 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1819 					 0,			/* flags */
1820 					 NULL,			/* lockfunc */
1821 					 NULL,			/* lockarg */
1822 					 &fl->ifl_buf_tag);
1823 		if (err) {
1824 			device_printf(dev,
1825 			    "Unable to allocate RX DMA tag: %d\n", err);
1826 			goto fail;
1827 		}
1828 
1829 		/* Allocate memory for the RX mbuf map. */
1830 		if (!(fl->ifl_sds.ifsd_m =
1831 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1832 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1833 			device_printf(dev,
1834 			    "Unable to allocate RX mbuf map memory\n");
1835 			err = ENOMEM;
1836 			goto fail;
1837 		}
1838 
1839 		/* Allocate memory for the direct RX cluster pointer map. */
1840 		if (!(fl->ifl_sds.ifsd_cl =
1841 		      (caddr_t *) malloc(sizeof(caddr_t) *
1842 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1843 			device_printf(dev,
1844 			    "Unable to allocate RX cluster map memory\n");
1845 			err = ENOMEM;
1846 			goto fail;
1847 		}
1848 
1849 		/* Allocate memory for the RX cluster bus address map. */
1850 		if (!(fl->ifl_sds.ifsd_ba =
1851 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1852 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1853 			device_printf(dev,
1854 			    "Unable to allocate RX bus address map memory\n");
1855 			err = ENOMEM;
1856 			goto fail;
1857 		}
1858 
1859 		/*
1860 		 * Create the DMA maps for RX buffers.
1861 		 */
1862 		if (!(fl->ifl_sds.ifsd_map =
1863 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1864 			device_printf(dev,
1865 			    "Unable to allocate RX buffer DMA map memory\n");
1866 			err = ENOMEM;
1867 			goto fail;
1868 		}
1869 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1870 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1871 			    &fl->ifl_sds.ifsd_map[i]);
1872 			if (err != 0) {
1873 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1874 				goto fail;
1875 			}
1876 		}
1877 	}
1878 	return (0);
1879 
1880 fail:
1881 	iflib_rx_structures_free(ctx);
1882 	return (err);
1883 }
1884 
1885 
1886 /*
1887  * Internal service routines
1888  */
1889 
1890 struct rxq_refill_cb_arg {
1891 	int               error;
1892 	bus_dma_segment_t seg;
1893 	int               nseg;
1894 };
1895 
1896 static void
1897 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1898 {
1899 	struct rxq_refill_cb_arg *cb_arg = arg;
1900 
1901 	cb_arg->error = error;
1902 	cb_arg->seg = segs[0];
1903 	cb_arg->nseg = nseg;
1904 }
1905 
1906 /**
1907  * _iflib_fl_refill - refill an rxq free-buffer list
1908  * @ctx: the iflib context
1909  * @fl: the free list to refill
1910  * @count: the number of new buffers to allocate
1911  *
1912  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1913  * The caller must assure that @count does not exceed the queue's capacity.
1914  */
1915 static uint8_t
1916 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1917 {
1918 	struct if_rxd_update iru;
1919 	struct rxq_refill_cb_arg cb_arg;
1920 	struct mbuf *m;
1921 	caddr_t cl, *sd_cl;
1922 	struct mbuf **sd_m;
1923 	bus_dmamap_t *sd_map;
1924 	bus_addr_t bus_addr, *sd_ba;
1925 	int err, frag_idx, i, idx, n, pidx;
1926 	qidx_t credits;
1927 
1928 	sd_m = fl->ifl_sds.ifsd_m;
1929 	sd_map = fl->ifl_sds.ifsd_map;
1930 	sd_cl = fl->ifl_sds.ifsd_cl;
1931 	sd_ba = fl->ifl_sds.ifsd_ba;
1932 	pidx = fl->ifl_pidx;
1933 	idx = pidx;
1934 	frag_idx = fl->ifl_fragidx;
1935 	credits = fl->ifl_credits;
1936 
1937 	i = 0;
1938 	n = count;
1939 	MPASS(n > 0);
1940 	MPASS(credits + n <= fl->ifl_size);
1941 
1942 	if (pidx < fl->ifl_cidx)
1943 		MPASS(pidx + n <= fl->ifl_cidx);
1944 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1945 		MPASS(fl->ifl_gen == 0);
1946 	if (pidx > fl->ifl_cidx)
1947 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1948 
1949 	DBG_COUNTER_INC(fl_refills);
1950 	if (n > 8)
1951 		DBG_COUNTER_INC(fl_refills_large);
1952 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1953 	while (n--) {
1954 		/*
1955 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1956 		 * initialized after rx.
1957 		 *
1958 		 * If the cluster is still set then we know a minimum sized packet was received
1959 		 */
1960 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1961 		    &frag_idx);
1962 		if (frag_idx < 0)
1963 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1964 		MPASS(frag_idx >= 0);
1965 		if ((cl = sd_cl[frag_idx]) == NULL) {
1966 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1967 				break;
1968 
1969 			cb_arg.error = 0;
1970 			MPASS(sd_map != NULL);
1971 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1972 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1973 			    BUS_DMA_NOWAIT);
1974 			if (err != 0 || cb_arg.error) {
1975 				/*
1976 				 * !zone_pack ?
1977 				 */
1978 				if (fl->ifl_zone == zone_pack)
1979 					uma_zfree(fl->ifl_zone, cl);
1980 				break;
1981 			}
1982 
1983 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1984 			sd_cl[frag_idx] = cl;
1985 #if MEMORY_LOGGING
1986 			fl->ifl_cl_enqueued++;
1987 #endif
1988 		} else {
1989 			bus_addr = sd_ba[frag_idx];
1990 		}
1991 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
1992 		    BUS_DMASYNC_PREREAD);
1993 
1994 		if (sd_m[frag_idx] == NULL) {
1995 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1996 				break;
1997 			}
1998 			sd_m[frag_idx] = m;
1999 		}
2000 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2001 #if MEMORY_LOGGING
2002 		fl->ifl_m_enqueued++;
2003 #endif
2004 
2005 		DBG_COUNTER_INC(rx_allocs);
2006 		fl->ifl_rxd_idxs[i] = frag_idx;
2007 		fl->ifl_bus_addrs[i] = bus_addr;
2008 		fl->ifl_vm_addrs[i] = cl;
2009 		credits++;
2010 		i++;
2011 		MPASS(credits <= fl->ifl_size);
2012 		if (++idx == fl->ifl_size) {
2013 			fl->ifl_gen = 1;
2014 			idx = 0;
2015 		}
2016 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2017 			iru.iru_pidx = pidx;
2018 			iru.iru_count = i;
2019 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2020 			i = 0;
2021 			pidx = idx;
2022 			fl->ifl_pidx = idx;
2023 			fl->ifl_credits = credits;
2024 		}
2025 	}
2026 
2027 	if (i) {
2028 		iru.iru_pidx = pidx;
2029 		iru.iru_count = i;
2030 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2031 		fl->ifl_pidx = idx;
2032 		fl->ifl_credits = credits;
2033 	}
2034 	DBG_COUNTER_INC(rxd_flush);
2035 	if (fl->ifl_pidx == 0)
2036 		pidx = fl->ifl_size - 1;
2037 	else
2038 		pidx = fl->ifl_pidx - 1;
2039 
2040 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2041 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2042 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2043 	fl->ifl_fragidx = frag_idx + 1;
2044 	if (fl->ifl_fragidx == fl->ifl_size)
2045 		fl->ifl_fragidx = 0;
2046 
2047 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2048 }
2049 
2050 static __inline uint8_t
2051 __iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2052 {
2053 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2054 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2055 #ifdef INVARIANTS
2056 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2057 #endif
2058 
2059 	MPASS(fl->ifl_credits <= fl->ifl_size);
2060 	MPASS(reclaimable == delta);
2061 
2062 	if (reclaimable > 0)
2063 		return (_iflib_fl_refill(ctx, fl, reclaimable));
2064 	return (0);
2065 }
2066 
2067 uint8_t
2068 iflib_in_detach(if_ctx_t ctx)
2069 {
2070 	bool in_detach;
2071 
2072 	STATE_LOCK(ctx);
2073 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2074 	STATE_UNLOCK(ctx);
2075 	return (in_detach);
2076 }
2077 
2078 static void
2079 iflib_fl_bufs_free(iflib_fl_t fl)
2080 {
2081 	iflib_dma_info_t idi = fl->ifl_ifdi;
2082 	bus_dmamap_t sd_map;
2083 	uint32_t i;
2084 
2085 	for (i = 0; i < fl->ifl_size; i++) {
2086 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2087 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2088 
2089 		if (*sd_cl != NULL) {
2090 			sd_map = fl->ifl_sds.ifsd_map[i];
2091 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2092 			    BUS_DMASYNC_POSTREAD);
2093 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2094 			if (*sd_cl != NULL)
2095 				uma_zfree(fl->ifl_zone, *sd_cl);
2096 			if (*sd_m != NULL) {
2097 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2098 				uma_zfree(zone_mbuf, *sd_m);
2099 			}
2100 		} else {
2101 			MPASS(*sd_cl == NULL);
2102 			MPASS(*sd_m == NULL);
2103 		}
2104 #if MEMORY_LOGGING
2105 		fl->ifl_m_dequeued++;
2106 		fl->ifl_cl_dequeued++;
2107 #endif
2108 		*sd_cl = NULL;
2109 		*sd_m = NULL;
2110 	}
2111 #ifdef INVARIANTS
2112 	for (i = 0; i < fl->ifl_size; i++) {
2113 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2114 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2115 	}
2116 #endif
2117 	/*
2118 	 * Reset free list values
2119 	 */
2120 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2121 	bzero(idi->idi_vaddr, idi->idi_size);
2122 }
2123 
2124 /*********************************************************************
2125  *
2126  *  Initialize a free list and its buffers.
2127  *
2128  **********************************************************************/
2129 static int
2130 iflib_fl_setup(iflib_fl_t fl)
2131 {
2132 	iflib_rxq_t rxq = fl->ifl_rxq;
2133 	if_ctx_t ctx = rxq->ifr_ctx;
2134 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2135 	int qidx;
2136 
2137 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2138 	/*
2139 	** Free current RX buffer structs and their mbufs
2140 	*/
2141 	iflib_fl_bufs_free(fl);
2142 	/* Now replenish the mbufs */
2143 	MPASS(fl->ifl_credits == 0);
2144 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2145 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2146 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2147 	else
2148 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2149 	/*
2150 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2151 	 * to the selected mbuf size.
2152 	 */
2153 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2154 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2155 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2156 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2157 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2158 
2159 
2160 	/* avoid pre-allocating zillions of clusters to an idle card
2161 	 * potentially speeding up attach
2162 	 */
2163 	(void) _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2164 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2165 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2166 		return (ENOBUFS);
2167 	/*
2168 	 * handle failure
2169 	 */
2170 	MPASS(rxq != NULL);
2171 	MPASS(fl->ifl_ifdi != NULL);
2172 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2173 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2174 	return (0);
2175 }
2176 
2177 /*********************************************************************
2178  *
2179  *  Free receive ring data structures
2180  *
2181  **********************************************************************/
2182 static void
2183 iflib_rx_sds_free(iflib_rxq_t rxq)
2184 {
2185 	iflib_fl_t fl;
2186 	int i, j;
2187 
2188 	if (rxq->ifr_fl != NULL) {
2189 		for (i = 0; i < rxq->ifr_nfl; i++) {
2190 			fl = &rxq->ifr_fl[i];
2191 			if (fl->ifl_buf_tag != NULL) {
2192 				if (fl->ifl_sds.ifsd_map != NULL) {
2193 					for (j = 0; j < fl->ifl_size; j++) {
2194 						bus_dmamap_sync(
2195 						    fl->ifl_buf_tag,
2196 						    fl->ifl_sds.ifsd_map[j],
2197 						    BUS_DMASYNC_POSTREAD);
2198 						bus_dmamap_unload(
2199 						    fl->ifl_buf_tag,
2200 						    fl->ifl_sds.ifsd_map[j]);
2201 						bus_dmamap_destroy(
2202 						    fl->ifl_buf_tag,
2203 						    fl->ifl_sds.ifsd_map[j]);
2204 					}
2205 				}
2206 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2207 				fl->ifl_buf_tag = NULL;
2208 			}
2209 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2210 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2211 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2212 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2213 			fl->ifl_sds.ifsd_m = NULL;
2214 			fl->ifl_sds.ifsd_cl = NULL;
2215 			fl->ifl_sds.ifsd_ba = NULL;
2216 			fl->ifl_sds.ifsd_map = NULL;
2217 		}
2218 		free(rxq->ifr_fl, M_IFLIB);
2219 		rxq->ifr_fl = NULL;
2220 		free(rxq->ifr_ifdi, M_IFLIB);
2221 		rxq->ifr_ifdi = NULL;
2222 		rxq->ifr_cq_cidx = 0;
2223 	}
2224 }
2225 
2226 /*
2227  * Timer routine
2228  */
2229 static void
2230 iflib_timer(void *arg)
2231 {
2232 	iflib_txq_t txq = arg;
2233 	if_ctx_t ctx = txq->ift_ctx;
2234 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2235 	uint64_t this_tick = ticks;
2236 	uint32_t reset_on = hz / 2;
2237 
2238 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2239 		return;
2240 
2241 	/*
2242 	** Check on the state of the TX queue(s), this
2243 	** can be done without the lock because its RO
2244 	** and the HUNG state will be static if set.
2245 	*/
2246 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2247 		txq->ift_last_timer_tick = this_tick;
2248 		IFDI_TIMER(ctx, txq->ift_id);
2249 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2250 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2251 		     (sctx->isc_pause_frames == 0)))
2252 			goto hung;
2253 
2254 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2255 		    ifmp_ring_is_stalled(txq->ift_br)) {
2256 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2257 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2258 		}
2259 		txq->ift_cleaned_prev = txq->ift_cleaned;
2260 	}
2261 #ifdef DEV_NETMAP
2262 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2263 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2264 #endif
2265 	/* handle any laggards */
2266 	if (txq->ift_db_pending)
2267 		GROUPTASK_ENQUEUE(&txq->ift_task);
2268 
2269 	sctx->isc_pause_frames = 0;
2270 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2271 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2272 	return;
2273 
2274  hung:
2275 	device_printf(ctx->ifc_dev,
2276 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2277 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2278 	STATE_LOCK(ctx);
2279 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2280 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2281 	iflib_admin_intr_deferred(ctx);
2282 	STATE_UNLOCK(ctx);
2283 }
2284 
2285 static uint16_t
2286 iflib_get_mbuf_size_for(unsigned int size)
2287 {
2288 
2289 	if (size <= MCLBYTES)
2290 		return (MCLBYTES);
2291 	else
2292 		return (MJUMPAGESIZE);
2293 }
2294 
2295 static void
2296 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2297 {
2298 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2299 
2300 	/*
2301 	 * XXX don't set the max_frame_size to larger
2302 	 * than the hardware can handle
2303 	 */
2304 	ctx->ifc_rx_mbuf_sz =
2305 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2306 }
2307 
2308 uint32_t
2309 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2310 {
2311 
2312 	return (ctx->ifc_rx_mbuf_sz);
2313 }
2314 
2315 static void
2316 iflib_init_locked(if_ctx_t ctx)
2317 {
2318 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2319 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2320 	if_t ifp = ctx->ifc_ifp;
2321 	iflib_fl_t fl;
2322 	iflib_txq_t txq;
2323 	iflib_rxq_t rxq;
2324 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2325 
2326 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2327 	IFDI_INTR_DISABLE(ctx);
2328 
2329 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2330 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2331 	/* Set hardware offload abilities */
2332 	if_clearhwassist(ifp);
2333 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2334 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2335 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2336 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2337 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2338 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2339 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2340 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2341 
2342 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2343 		CALLOUT_LOCK(txq);
2344 		callout_stop(&txq->ift_timer);
2345 		CALLOUT_UNLOCK(txq);
2346 		iflib_netmap_txq_init(ctx, txq);
2347 	}
2348 
2349 	/*
2350 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2351 	 * that drivers can use the value when setting up the hardware receive
2352 	 * buffers.
2353 	 */
2354 	iflib_calc_rx_mbuf_sz(ctx);
2355 
2356 #ifdef INVARIANTS
2357 	i = if_getdrvflags(ifp);
2358 #endif
2359 	IFDI_INIT(ctx);
2360 	MPASS(if_getdrvflags(ifp) == i);
2361 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2362 		/* XXX this should really be done on a per-queue basis */
2363 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2364 			MPASS(rxq->ifr_id == i);
2365 			iflib_netmap_rxq_init(ctx, rxq);
2366 			continue;
2367 		}
2368 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2369 			if (iflib_fl_setup(fl)) {
2370 				device_printf(ctx->ifc_dev,
2371 				    "setting up free list %d failed - "
2372 				    "check cluster settings\n", j);
2373 				goto done;
2374 			}
2375 		}
2376 	}
2377 done:
2378 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2379 	IFDI_INTR_ENABLE(ctx);
2380 	txq = ctx->ifc_txqs;
2381 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2382 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2383 			txq->ift_timer.c_cpu);
2384 }
2385 
2386 static int
2387 iflib_media_change(if_t ifp)
2388 {
2389 	if_ctx_t ctx = if_getsoftc(ifp);
2390 	int err;
2391 
2392 	CTX_LOCK(ctx);
2393 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2394 		iflib_init_locked(ctx);
2395 	CTX_UNLOCK(ctx);
2396 	return (err);
2397 }
2398 
2399 static void
2400 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2401 {
2402 	if_ctx_t ctx = if_getsoftc(ifp);
2403 
2404 	CTX_LOCK(ctx);
2405 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2406 	IFDI_MEDIA_STATUS(ctx, ifmr);
2407 	CTX_UNLOCK(ctx);
2408 }
2409 
2410 void
2411 iflib_stop(if_ctx_t ctx)
2412 {
2413 	iflib_txq_t txq = ctx->ifc_txqs;
2414 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2415 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2416 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2417 	iflib_dma_info_t di;
2418 	iflib_fl_t fl;
2419 	int i, j;
2420 
2421 	/* Tell the stack that the interface is no longer active */
2422 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2423 
2424 	IFDI_INTR_DISABLE(ctx);
2425 	DELAY(1000);
2426 	IFDI_STOP(ctx);
2427 	DELAY(1000);
2428 
2429 	iflib_debug_reset();
2430 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2431 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2432 		/* make sure all transmitters have completed before proceeding XXX */
2433 
2434 		CALLOUT_LOCK(txq);
2435 		callout_stop(&txq->ift_timer);
2436 		CALLOUT_UNLOCK(txq);
2437 
2438 		/* clean any enqueued buffers */
2439 		iflib_ifmp_purge(txq);
2440 		/* Free any existing tx buffers. */
2441 		for (j = 0; j < txq->ift_size; j++) {
2442 			iflib_txsd_free(ctx, txq, j);
2443 		}
2444 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2445 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2446 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2447 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2448 		txq->ift_pullups = 0;
2449 		ifmp_ring_reset_stats(txq->ift_br);
2450 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2451 			bzero((void *)di->idi_vaddr, di->idi_size);
2452 	}
2453 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2454 		/* make sure all transmitters have completed before proceeding XXX */
2455 
2456 		rxq->ifr_cq_cidx = 0;
2457 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2458 			bzero((void *)di->idi_vaddr, di->idi_size);
2459 		/* also resets the free lists pidx/cidx */
2460 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2461 			iflib_fl_bufs_free(fl);
2462 	}
2463 }
2464 
2465 static inline caddr_t
2466 calc_next_rxd(iflib_fl_t fl, int cidx)
2467 {
2468 	qidx_t size;
2469 	int nrxd;
2470 	caddr_t start, end, cur, next;
2471 
2472 	nrxd = fl->ifl_size;
2473 	size = fl->ifl_rxd_size;
2474 	start = fl->ifl_ifdi->idi_vaddr;
2475 
2476 	if (__predict_false(size == 0))
2477 		return (start);
2478 	cur = start + size*cidx;
2479 	end = start + size*nrxd;
2480 	next = CACHE_PTR_NEXT(cur);
2481 	return (next < end ? next : start);
2482 }
2483 
2484 static inline void
2485 prefetch_pkts(iflib_fl_t fl, int cidx)
2486 {
2487 	int nextptr;
2488 	int nrxd = fl->ifl_size;
2489 	caddr_t next_rxd;
2490 
2491 
2492 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2493 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2494 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2495 	next_rxd = calc_next_rxd(fl, cidx);
2496 	prefetch(next_rxd);
2497 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2498 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2499 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2500 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2501 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2502 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2503 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2504 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2505 }
2506 
2507 static struct mbuf *
2508 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2509     int *pf_rv, if_rxd_info_t ri)
2510 {
2511 	bus_dmamap_t map;
2512 	iflib_fl_t fl;
2513 	caddr_t payload;
2514 	struct mbuf *m;
2515 	int flid, cidx, len, next;
2516 
2517 	map = NULL;
2518 	flid = irf->irf_flid;
2519 	cidx = irf->irf_idx;
2520 	fl = &rxq->ifr_fl[flid];
2521 	sd->ifsd_fl = fl;
2522 	m = fl->ifl_sds.ifsd_m[cidx];
2523 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2524 	fl->ifl_credits--;
2525 #if MEMORY_LOGGING
2526 	fl->ifl_m_dequeued++;
2527 #endif
2528 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2529 		prefetch_pkts(fl, cidx);
2530 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2531 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2532 	map = fl->ifl_sds.ifsd_map[cidx];
2533 
2534 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2535 
2536 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2537 	    irf->irf_len != 0) {
2538 		payload  = *sd->ifsd_cl;
2539 		payload +=  ri->iri_pad;
2540 		len = ri->iri_len - ri->iri_pad;
2541 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2542 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2543 		switch (*pf_rv) {
2544 		case PFIL_DROPPED:
2545 		case PFIL_CONSUMED:
2546 			/*
2547 			 * The filter ate it.  Everything is recycled.
2548 			 */
2549 			m = NULL;
2550 			unload = 0;
2551 			break;
2552 		case PFIL_REALLOCED:
2553 			/*
2554 			 * The filter copied it.  Everything is recycled.
2555 			 */
2556 			m = pfil_mem2mbuf(payload);
2557 			unload = 0;
2558 			break;
2559 		case PFIL_PASS:
2560 			/*
2561 			 * Filter said it was OK, so receive like
2562 			 * normal
2563 			 */
2564 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2565 			break;
2566 		default:
2567 			MPASS(0);
2568 		}
2569 	} else {
2570 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2571 		*pf_rv = PFIL_PASS;
2572 	}
2573 
2574 	if (unload && irf->irf_len != 0)
2575 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2576 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2577 	if (__predict_false(fl->ifl_cidx == 0))
2578 		fl->ifl_gen = 0;
2579 	bit_clear(fl->ifl_rx_bitmap, cidx);
2580 	return (m);
2581 }
2582 
2583 static struct mbuf *
2584 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2585 {
2586 	struct mbuf *m, *mh, *mt;
2587 	caddr_t cl;
2588 	int  *pf_rv_ptr, flags, i, padlen;
2589 	bool consumed;
2590 
2591 	i = 0;
2592 	mh = NULL;
2593 	consumed = false;
2594 	*pf_rv = PFIL_PASS;
2595 	pf_rv_ptr = pf_rv;
2596 	do {
2597 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2598 		    pf_rv_ptr, ri);
2599 
2600 		MPASS(*sd->ifsd_cl != NULL);
2601 
2602 		/*
2603 		 * Exclude zero-length frags & frags from
2604 		 * packets the filter has consumed or dropped
2605 		 */
2606 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2607 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2608 			if (mh == NULL) {
2609 				/* everything saved here */
2610 				consumed = true;
2611 				pf_rv_ptr = NULL;
2612 				continue;
2613 			}
2614 			/* XXX we can save the cluster here, but not the mbuf */
2615 			m_init(m, M_NOWAIT, MT_DATA, 0);
2616 			m_free(m);
2617 			continue;
2618 		}
2619 		if (mh == NULL) {
2620 			flags = M_PKTHDR|M_EXT;
2621 			mh = mt = m;
2622 			padlen = ri->iri_pad;
2623 		} else {
2624 			flags = M_EXT;
2625 			mt->m_next = m;
2626 			mt = m;
2627 			/* assuming padding is only on the first fragment */
2628 			padlen = 0;
2629 		}
2630 		cl = *sd->ifsd_cl;
2631 		*sd->ifsd_cl = NULL;
2632 
2633 		/* Can these two be made one ? */
2634 		m_init(m, M_NOWAIT, MT_DATA, flags);
2635 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2636 		/*
2637 		 * These must follow m_init and m_cljset
2638 		 */
2639 		m->m_data += padlen;
2640 		ri->iri_len -= padlen;
2641 		m->m_len = ri->iri_frags[i].irf_len;
2642 	} while (++i < ri->iri_nfrags);
2643 
2644 	return (mh);
2645 }
2646 
2647 /*
2648  * Process one software descriptor
2649  */
2650 static struct mbuf *
2651 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2652 {
2653 	struct if_rxsd sd;
2654 	struct mbuf *m;
2655 	int pf_rv;
2656 
2657 	/* should I merge this back in now that the two paths are basically duplicated? */
2658 	if (ri->iri_nfrags == 1 &&
2659 	    ri->iri_frags[0].irf_len != 0 &&
2660 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2661 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2662 		    &pf_rv, ri);
2663 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2664 			return (m);
2665 		if (pf_rv == PFIL_PASS) {
2666 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2667 #ifndef __NO_STRICT_ALIGNMENT
2668 			if (!IP_ALIGNED(m))
2669 				m->m_data += 2;
2670 #endif
2671 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2672 			m->m_len = ri->iri_frags[0].irf_len;
2673 		}
2674 	} else {
2675 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2676 		if (m == NULL)
2677 			return (NULL);
2678 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2679 			return (m);
2680 	}
2681 	m->m_pkthdr.len = ri->iri_len;
2682 	m->m_pkthdr.rcvif = ri->iri_ifp;
2683 	m->m_flags |= ri->iri_flags;
2684 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2685 	m->m_pkthdr.flowid = ri->iri_flowid;
2686 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2687 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2688 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2689 	return (m);
2690 }
2691 
2692 #if defined(INET6) || defined(INET)
2693 static void
2694 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2695 {
2696 	CURVNET_SET(lc->ifp->if_vnet);
2697 #if defined(INET6)
2698 	*v6 = V_ip6_forwarding;
2699 #endif
2700 #if defined(INET)
2701 	*v4 = V_ipforwarding;
2702 #endif
2703 	CURVNET_RESTORE();
2704 }
2705 
2706 /*
2707  * Returns true if it's possible this packet could be LROed.
2708  * if it returns false, it is guaranteed that tcp_lro_rx()
2709  * would not return zero.
2710  */
2711 static bool
2712 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2713 {
2714 	struct ether_header *eh;
2715 
2716 	eh = mtod(m, struct ether_header *);
2717 	switch (eh->ether_type) {
2718 #if defined(INET6)
2719 		case htons(ETHERTYPE_IPV6):
2720 			return (!v6_forwarding);
2721 #endif
2722 #if defined (INET)
2723 		case htons(ETHERTYPE_IP):
2724 			return (!v4_forwarding);
2725 #endif
2726 	}
2727 
2728 	return false;
2729 }
2730 #else
2731 static void
2732 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2733 {
2734 }
2735 #endif
2736 
2737 static void
2738 _task_fn_rx_watchdog(void *context)
2739 {
2740 	iflib_rxq_t rxq = context;
2741 
2742 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2743 }
2744 
2745 static uint8_t
2746 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2747 {
2748 	if_t ifp;
2749 	if_ctx_t ctx = rxq->ifr_ctx;
2750 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2751 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2752 	int avail, i;
2753 	qidx_t *cidxp;
2754 	struct if_rxd_info ri;
2755 	int err, budget_left, rx_bytes, rx_pkts;
2756 	iflib_fl_t fl;
2757 	int lro_enabled;
2758 	bool v4_forwarding, v6_forwarding, lro_possible;
2759 	uint8_t retval = 0;
2760 
2761 	/*
2762 	 * XXX early demux data packets so that if_input processing only handles
2763 	 * acks in interrupt context
2764 	 */
2765 	struct mbuf *m, *mh, *mt, *mf;
2766 
2767 	NET_EPOCH_ASSERT();
2768 
2769 	lro_possible = v4_forwarding = v6_forwarding = false;
2770 	ifp = ctx->ifc_ifp;
2771 	mh = mt = NULL;
2772 	MPASS(budget > 0);
2773 	rx_pkts	= rx_bytes = 0;
2774 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2775 		cidxp = &rxq->ifr_cq_cidx;
2776 	else
2777 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2778 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2779 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2780 			retval |= __iflib_fl_refill_all(ctx, fl);
2781 		DBG_COUNTER_INC(rx_unavail);
2782 		return (retval);
2783 	}
2784 
2785 	/* pfil needs the vnet to be set */
2786 	CURVNET_SET_QUIET(ifp->if_vnet);
2787 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2788 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2789 			DBG_COUNTER_INC(rx_ctx_inactive);
2790 			break;
2791 		}
2792 		/*
2793 		 * Reset client set fields to their default values
2794 		 */
2795 		rxd_info_zero(&ri);
2796 		ri.iri_qsidx = rxq->ifr_id;
2797 		ri.iri_cidx = *cidxp;
2798 		ri.iri_ifp = ifp;
2799 		ri.iri_frags = rxq->ifr_frags;
2800 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2801 
2802 		if (err)
2803 			goto err;
2804 		rx_pkts += 1;
2805 		rx_bytes += ri.iri_len;
2806 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2807 			*cidxp = ri.iri_cidx;
2808 			/* Update our consumer index */
2809 			/* XXX NB: shurd - check if this is still safe */
2810 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2811 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2812 			/* was this only a completion queue message? */
2813 			if (__predict_false(ri.iri_nfrags == 0))
2814 				continue;
2815 		}
2816 		MPASS(ri.iri_nfrags != 0);
2817 		MPASS(ri.iri_len != 0);
2818 
2819 		/* will advance the cidx on the corresponding free lists */
2820 		m = iflib_rxd_pkt_get(rxq, &ri);
2821 		avail--;
2822 		budget_left--;
2823 		if (avail == 0 && budget_left)
2824 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2825 
2826 		if (__predict_false(m == NULL))
2827 			continue;
2828 
2829 		/* imm_pkt: -- cxgb */
2830 		if (mh == NULL)
2831 			mh = mt = m;
2832 		else {
2833 			mt->m_nextpkt = m;
2834 			mt = m;
2835 		}
2836 	}
2837 	CURVNET_RESTORE();
2838 	/* make sure that we can refill faster than drain */
2839 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2840 		retval |= __iflib_fl_refill_all(ctx, fl);
2841 
2842 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2843 	if (lro_enabled)
2844 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2845 	mt = mf = NULL;
2846 	while (mh != NULL) {
2847 		m = mh;
2848 		mh = mh->m_nextpkt;
2849 		m->m_nextpkt = NULL;
2850 #ifndef __NO_STRICT_ALIGNMENT
2851 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2852 			continue;
2853 #endif
2854 		rx_bytes += m->m_pkthdr.len;
2855 		rx_pkts++;
2856 #if defined(INET6) || defined(INET)
2857 		if (lro_enabled) {
2858 			if (!lro_possible) {
2859 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2860 				if (lro_possible && mf != NULL) {
2861 					ifp->if_input(ifp, mf);
2862 					DBG_COUNTER_INC(rx_if_input);
2863 					mt = mf = NULL;
2864 				}
2865 			}
2866 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2867 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2868 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2869 					continue;
2870 			}
2871 		}
2872 #endif
2873 		if (lro_possible) {
2874 			ifp->if_input(ifp, m);
2875 			DBG_COUNTER_INC(rx_if_input);
2876 			continue;
2877 		}
2878 
2879 		if (mf == NULL)
2880 			mf = m;
2881 		if (mt != NULL)
2882 			mt->m_nextpkt = m;
2883 		mt = m;
2884 	}
2885 	if (mf != NULL) {
2886 		ifp->if_input(ifp, mf);
2887 		DBG_COUNTER_INC(rx_if_input);
2888 	}
2889 
2890 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2891 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2892 
2893 	/*
2894 	 * Flush any outstanding LRO work
2895 	 */
2896 #if defined(INET6) || defined(INET)
2897 	tcp_lro_flush_all(&rxq->ifr_lc);
2898 #endif
2899 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2900 		retval |= IFLIB_RXEOF_MORE;
2901 	return (retval);
2902 err:
2903 	STATE_LOCK(ctx);
2904 	ctx->ifc_flags |= IFC_DO_RESET;
2905 	iflib_admin_intr_deferred(ctx);
2906 	STATE_UNLOCK(ctx);
2907 	return (0);
2908 }
2909 
2910 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2911 static inline qidx_t
2912 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2913 {
2914 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2915 	qidx_t minthresh = txq->ift_size / 8;
2916 	if (in_use > 4*minthresh)
2917 		return (notify_count);
2918 	if (in_use > 2*minthresh)
2919 		return (notify_count >> 1);
2920 	if (in_use > minthresh)
2921 		return (notify_count >> 3);
2922 	return (0);
2923 }
2924 
2925 static inline qidx_t
2926 txq_max_rs_deferred(iflib_txq_t txq)
2927 {
2928 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2929 	qidx_t minthresh = txq->ift_size / 8;
2930 	if (txq->ift_in_use > 4*minthresh)
2931 		return (notify_count);
2932 	if (txq->ift_in_use > 2*minthresh)
2933 		return (notify_count >> 1);
2934 	if (txq->ift_in_use > minthresh)
2935 		return (notify_count >> 2);
2936 	return (2);
2937 }
2938 
2939 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2940 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2941 
2942 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2943 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2944 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2945 
2946 /* forward compatibility for cxgb */
2947 #define FIRST_QSET(ctx) 0
2948 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2949 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2950 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2951 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2952 
2953 /* XXX we should be setting this to something other than zero */
2954 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2955 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2956     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2957 
2958 static inline bool
2959 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2960 {
2961 	qidx_t dbval, max;
2962 	bool rang;
2963 
2964 	rang = false;
2965 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2966 	if (ring || txq->ift_db_pending >= max) {
2967 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2968 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2969 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2970 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2971 		txq->ift_db_pending = txq->ift_npending = 0;
2972 		rang = true;
2973 	}
2974 	return (rang);
2975 }
2976 
2977 #ifdef PKT_DEBUG
2978 static void
2979 print_pkt(if_pkt_info_t pi)
2980 {
2981 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2982 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2983 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2984 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2985 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2986 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2987 }
2988 #endif
2989 
2990 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2991 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2992 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2993 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2994 
2995 static int
2996 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2997 {
2998 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2999 	struct ether_vlan_header *eh;
3000 	struct mbuf *m;
3001 
3002 	m = *mp;
3003 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3004 	    M_WRITABLE(m) == 0) {
3005 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3006 			return (ENOMEM);
3007 		} else {
3008 			m_freem(*mp);
3009 			DBG_COUNTER_INC(tx_frees);
3010 			*mp = m;
3011 		}
3012 	}
3013 
3014 	/*
3015 	 * Determine where frame payload starts.
3016 	 * Jump over vlan headers if already present,
3017 	 * helpful for QinQ too.
3018 	 */
3019 	if (__predict_false(m->m_len < sizeof(*eh))) {
3020 		txq->ift_pullups++;
3021 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3022 			return (ENOMEM);
3023 	}
3024 	eh = mtod(m, struct ether_vlan_header *);
3025 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3026 		pi->ipi_etype = ntohs(eh->evl_proto);
3027 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3028 	} else {
3029 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3030 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3031 	}
3032 
3033 	switch (pi->ipi_etype) {
3034 #ifdef INET
3035 	case ETHERTYPE_IP:
3036 	{
3037 		struct mbuf *n;
3038 		struct ip *ip = NULL;
3039 		struct tcphdr *th = NULL;
3040 		int minthlen;
3041 
3042 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3043 		if (__predict_false(m->m_len < minthlen)) {
3044 			/*
3045 			 * if this code bloat is causing too much of a hit
3046 			 * move it to a separate function and mark it noinline
3047 			 */
3048 			if (m->m_len == pi->ipi_ehdrlen) {
3049 				n = m->m_next;
3050 				MPASS(n);
3051 				if (n->m_len >= sizeof(*ip))  {
3052 					ip = (struct ip *)n->m_data;
3053 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3054 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3055 				} else {
3056 					txq->ift_pullups++;
3057 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3058 						return (ENOMEM);
3059 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3060 				}
3061 			} else {
3062 				txq->ift_pullups++;
3063 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3064 					return (ENOMEM);
3065 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3066 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3067 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3068 			}
3069 		} else {
3070 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3071 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3072 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3073 		}
3074 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3075 		pi->ipi_ipproto = ip->ip_p;
3076 		pi->ipi_flags |= IPI_TX_IPV4;
3077 
3078 		/* TCP checksum offload may require TCP header length */
3079 		if (IS_TX_OFFLOAD4(pi)) {
3080 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3081 				if (__predict_false(th == NULL)) {
3082 					txq->ift_pullups++;
3083 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3084 						return (ENOMEM);
3085 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3086 				}
3087 				pi->ipi_tcp_hflags = th->th_flags;
3088 				pi->ipi_tcp_hlen = th->th_off << 2;
3089 				pi->ipi_tcp_seq = th->th_seq;
3090 			}
3091 			if (IS_TSO4(pi)) {
3092 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3093 					return (ENXIO);
3094 				/*
3095 				 * TSO always requires hardware checksum offload.
3096 				 */
3097 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3098 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3099 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3100 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3101 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3102 					ip->ip_sum = 0;
3103 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3104 				}
3105 			}
3106 		}
3107 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3108                        ip->ip_sum = 0;
3109 
3110 		break;
3111 	}
3112 #endif
3113 #ifdef INET6
3114 	case ETHERTYPE_IPV6:
3115 	{
3116 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3117 		struct tcphdr *th;
3118 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3119 
3120 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3121 			txq->ift_pullups++;
3122 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3123 				return (ENOMEM);
3124 		}
3125 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3126 
3127 		/* XXX-BZ this will go badly in case of ext hdrs. */
3128 		pi->ipi_ipproto = ip6->ip6_nxt;
3129 		pi->ipi_flags |= IPI_TX_IPV6;
3130 
3131 		/* TCP checksum offload may require TCP header length */
3132 		if (IS_TX_OFFLOAD6(pi)) {
3133 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3134 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3135 					txq->ift_pullups++;
3136 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3137 						return (ENOMEM);
3138 				}
3139 				pi->ipi_tcp_hflags = th->th_flags;
3140 				pi->ipi_tcp_hlen = th->th_off << 2;
3141 				pi->ipi_tcp_seq = th->th_seq;
3142 			}
3143 			if (IS_TSO6(pi)) {
3144 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3145 					return (ENXIO);
3146 				/*
3147 				 * TSO always requires hardware checksum offload.
3148 				 */
3149 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3150 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3151 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3152 			}
3153 		}
3154 		break;
3155 	}
3156 #endif
3157 	default:
3158 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3159 		pi->ipi_ip_hlen = 0;
3160 		break;
3161 	}
3162 	*mp = m;
3163 
3164 	return (0);
3165 }
3166 
3167 /*
3168  * If dodgy hardware rejects the scatter gather chain we've handed it
3169  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3170  * m_defrag'd mbufs
3171  */
3172 static __noinline struct mbuf *
3173 iflib_remove_mbuf(iflib_txq_t txq)
3174 {
3175 	int ntxd, pidx;
3176 	struct mbuf *m, **ifsd_m;
3177 
3178 	ifsd_m = txq->ift_sds.ifsd_m;
3179 	ntxd = txq->ift_size;
3180 	pidx = txq->ift_pidx & (ntxd - 1);
3181 	ifsd_m = txq->ift_sds.ifsd_m;
3182 	m = ifsd_m[pidx];
3183 	ifsd_m[pidx] = NULL;
3184 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3185 	if (txq->ift_sds.ifsd_tso_map != NULL)
3186 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3187 		    txq->ift_sds.ifsd_tso_map[pidx]);
3188 #if MEMORY_LOGGING
3189 	txq->ift_dequeued++;
3190 #endif
3191 	return (m);
3192 }
3193 
3194 static inline caddr_t
3195 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3196 {
3197 	qidx_t size;
3198 	int ntxd;
3199 	caddr_t start, end, cur, next;
3200 
3201 	ntxd = txq->ift_size;
3202 	size = txq->ift_txd_size[qid];
3203 	start = txq->ift_ifdi[qid].idi_vaddr;
3204 
3205 	if (__predict_false(size == 0))
3206 		return (start);
3207 	cur = start + size*cidx;
3208 	end = start + size*ntxd;
3209 	next = CACHE_PTR_NEXT(cur);
3210 	return (next < end ? next : start);
3211 }
3212 
3213 /*
3214  * Pad an mbuf to ensure a minimum ethernet frame size.
3215  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3216  */
3217 static __noinline int
3218 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3219 {
3220 	/*
3221 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3222 	 * and ARP message is the smallest common payload I can think of
3223 	 */
3224 	static char pad[18];	/* just zeros */
3225 	int n;
3226 	struct mbuf *new_head;
3227 
3228 	if (!M_WRITABLE(*m_head)) {
3229 		new_head = m_dup(*m_head, M_NOWAIT);
3230 		if (new_head == NULL) {
3231 			m_freem(*m_head);
3232 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3233 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3234 			DBG_COUNTER_INC(tx_frees);
3235 			return ENOMEM;
3236 		}
3237 		m_freem(*m_head);
3238 		*m_head = new_head;
3239 	}
3240 
3241 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3242 	     n > 0; n -= sizeof(pad))
3243 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3244 			break;
3245 
3246 	if (n > 0) {
3247 		m_freem(*m_head);
3248 		device_printf(dev, "cannot pad short frame\n");
3249 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3250 		DBG_COUNTER_INC(tx_frees);
3251 		return (ENOBUFS);
3252 	}
3253 
3254 	return 0;
3255 }
3256 
3257 static int
3258 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3259 {
3260 	if_ctx_t		ctx;
3261 	if_shared_ctx_t		sctx;
3262 	if_softc_ctx_t		scctx;
3263 	bus_dma_tag_t		buf_tag;
3264 	bus_dma_segment_t	*segs;
3265 	struct mbuf		*m_head, **ifsd_m;
3266 	void			*next_txd;
3267 	bus_dmamap_t		map;
3268 	struct if_pkt_info	pi;
3269 	int remap = 0;
3270 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3271 
3272 	ctx = txq->ift_ctx;
3273 	sctx = ctx->ifc_sctx;
3274 	scctx = &ctx->ifc_softc_ctx;
3275 	segs = txq->ift_segs;
3276 	ntxd = txq->ift_size;
3277 	m_head = *m_headp;
3278 	map = NULL;
3279 
3280 	/*
3281 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3282 	 */
3283 	cidx = txq->ift_cidx;
3284 	pidx = txq->ift_pidx;
3285 	if (ctx->ifc_flags & IFC_PREFETCH) {
3286 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3287 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3288 			next_txd = calc_next_txd(txq, cidx, 0);
3289 			prefetch(next_txd);
3290 		}
3291 
3292 		/* prefetch the next cache line of mbuf pointers and flags */
3293 		prefetch(&txq->ift_sds.ifsd_m[next]);
3294 		prefetch(&txq->ift_sds.ifsd_map[next]);
3295 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3296 	}
3297 	map = txq->ift_sds.ifsd_map[pidx];
3298 	ifsd_m = txq->ift_sds.ifsd_m;
3299 
3300 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3301 		buf_tag = txq->ift_tso_buf_tag;
3302 		max_segs = scctx->isc_tx_tso_segments_max;
3303 		map = txq->ift_sds.ifsd_tso_map[pidx];
3304 		MPASS(buf_tag != NULL);
3305 		MPASS(max_segs > 0);
3306 	} else {
3307 		buf_tag = txq->ift_buf_tag;
3308 		max_segs = scctx->isc_tx_nsegments;
3309 		map = txq->ift_sds.ifsd_map[pidx];
3310 	}
3311 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3312 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3313 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3314 		if (err) {
3315 			DBG_COUNTER_INC(encap_txd_encap_fail);
3316 			return err;
3317 		}
3318 	}
3319 	m_head = *m_headp;
3320 
3321 	pkt_info_zero(&pi);
3322 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3323 	pi.ipi_pidx = pidx;
3324 	pi.ipi_qsidx = txq->ift_id;
3325 	pi.ipi_len = m_head->m_pkthdr.len;
3326 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3327 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3328 
3329 	/* deliberate bitwise OR to make one condition */
3330 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3331 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3332 			DBG_COUNTER_INC(encap_txd_encap_fail);
3333 			return (err);
3334 		}
3335 		m_head = *m_headp;
3336 	}
3337 
3338 retry:
3339 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3340 	    BUS_DMA_NOWAIT);
3341 defrag:
3342 	if (__predict_false(err)) {
3343 		switch (err) {
3344 		case EFBIG:
3345 			/* try collapse once and defrag once */
3346 			if (remap == 0) {
3347 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3348 				/* try defrag if collapsing fails */
3349 				if (m_head == NULL)
3350 					remap++;
3351 			}
3352 			if (remap == 1) {
3353 				txq->ift_mbuf_defrag++;
3354 				m_head = m_defrag(*m_headp, M_NOWAIT);
3355 			}
3356 			/*
3357 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3358 			 * failed to map an mbuf that was run through m_defrag
3359 			 */
3360 			MPASS(remap <= 1);
3361 			if (__predict_false(m_head == NULL || remap > 1))
3362 				goto defrag_failed;
3363 			remap++;
3364 			*m_headp = m_head;
3365 			goto retry;
3366 			break;
3367 		case ENOMEM:
3368 			txq->ift_no_tx_dma_setup++;
3369 			break;
3370 		default:
3371 			txq->ift_no_tx_dma_setup++;
3372 			m_freem(*m_headp);
3373 			DBG_COUNTER_INC(tx_frees);
3374 			*m_headp = NULL;
3375 			break;
3376 		}
3377 		txq->ift_map_failed++;
3378 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3379 		DBG_COUNTER_INC(encap_txd_encap_fail);
3380 		return (err);
3381 	}
3382 	ifsd_m[pidx] = m_head;
3383 	/*
3384 	 * XXX assumes a 1 to 1 relationship between segments and
3385 	 *        descriptors - this does not hold true on all drivers, e.g.
3386 	 *        cxgb
3387 	 */
3388 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3389 		txq->ift_no_desc_avail++;
3390 		bus_dmamap_unload(buf_tag, map);
3391 		DBG_COUNTER_INC(encap_txq_avail_fail);
3392 		DBG_COUNTER_INC(encap_txd_encap_fail);
3393 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3394 			GROUPTASK_ENQUEUE(&txq->ift_task);
3395 		return (ENOBUFS);
3396 	}
3397 	/*
3398 	 * On Intel cards we can greatly reduce the number of TX interrupts
3399 	 * we see by only setting report status on every Nth descriptor.
3400 	 * However, this also means that the driver will need to keep track
3401 	 * of the descriptors that RS was set on to check them for the DD bit.
3402 	 */
3403 	txq->ift_rs_pending += nsegs + 1;
3404 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3405 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3406 		pi.ipi_flags |= IPI_TX_INTR;
3407 		txq->ift_rs_pending = 0;
3408 	}
3409 
3410 	pi.ipi_segs = segs;
3411 	pi.ipi_nsegs = nsegs;
3412 
3413 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3414 #ifdef PKT_DEBUG
3415 	print_pkt(&pi);
3416 #endif
3417 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3418 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3419 		DBG_COUNTER_INC(tx_encap);
3420 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3421 
3422 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3423 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3424 			ndesc += txq->ift_size;
3425 			txq->ift_gen = 1;
3426 		}
3427 		/*
3428 		 * drivers can need as many as
3429 		 * two sentinels
3430 		 */
3431 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3432 		MPASS(pi.ipi_new_pidx != pidx);
3433 		MPASS(ndesc > 0);
3434 		txq->ift_in_use += ndesc;
3435 
3436 		/*
3437 		 * We update the last software descriptor again here because there may
3438 		 * be a sentinel and/or there may be more mbufs than segments
3439 		 */
3440 		txq->ift_pidx = pi.ipi_new_pidx;
3441 		txq->ift_npending += pi.ipi_ndescs;
3442 	} else {
3443 		*m_headp = m_head = iflib_remove_mbuf(txq);
3444 		if (err == EFBIG) {
3445 			txq->ift_txd_encap_efbig++;
3446 			if (remap < 2) {
3447 				remap = 1;
3448 				goto defrag;
3449 			}
3450 		}
3451 		goto defrag_failed;
3452 	}
3453 	/*
3454 	 * err can't possibly be non-zero here, so we don't neet to test it
3455 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3456 	 */
3457 	return (err);
3458 
3459 defrag_failed:
3460 	txq->ift_mbuf_defrag_failed++;
3461 	txq->ift_map_failed++;
3462 	m_freem(*m_headp);
3463 	DBG_COUNTER_INC(tx_frees);
3464 	*m_headp = NULL;
3465 	DBG_COUNTER_INC(encap_txd_encap_fail);
3466 	return (ENOMEM);
3467 }
3468 
3469 static void
3470 iflib_tx_desc_free(iflib_txq_t txq, int n)
3471 {
3472 	uint32_t qsize, cidx, mask, gen;
3473 	struct mbuf *m, **ifsd_m;
3474 	bool do_prefetch;
3475 
3476 	cidx = txq->ift_cidx;
3477 	gen = txq->ift_gen;
3478 	qsize = txq->ift_size;
3479 	mask = qsize-1;
3480 	ifsd_m = txq->ift_sds.ifsd_m;
3481 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3482 
3483 	while (n-- > 0) {
3484 		if (do_prefetch) {
3485 			prefetch(ifsd_m[(cidx + 3) & mask]);
3486 			prefetch(ifsd_m[(cidx + 4) & mask]);
3487 		}
3488 		if ((m = ifsd_m[cidx]) != NULL) {
3489 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3490 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3491 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3492 				    txq->ift_sds.ifsd_tso_map[cidx],
3493 				    BUS_DMASYNC_POSTWRITE);
3494 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3495 				    txq->ift_sds.ifsd_tso_map[cidx]);
3496 			} else {
3497 				bus_dmamap_sync(txq->ift_buf_tag,
3498 				    txq->ift_sds.ifsd_map[cidx],
3499 				    BUS_DMASYNC_POSTWRITE);
3500 				bus_dmamap_unload(txq->ift_buf_tag,
3501 				    txq->ift_sds.ifsd_map[cidx]);
3502 			}
3503 			/* XXX we don't support any drivers that batch packets yet */
3504 			MPASS(m->m_nextpkt == NULL);
3505 			m_freem(m);
3506 			ifsd_m[cidx] = NULL;
3507 #if MEMORY_LOGGING
3508 			txq->ift_dequeued++;
3509 #endif
3510 			DBG_COUNTER_INC(tx_frees);
3511 		}
3512 		if (__predict_false(++cidx == qsize)) {
3513 			cidx = 0;
3514 			gen = 0;
3515 		}
3516 	}
3517 	txq->ift_cidx = cidx;
3518 	txq->ift_gen = gen;
3519 }
3520 
3521 static __inline int
3522 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3523 {
3524 	int reclaim;
3525 	if_ctx_t ctx = txq->ift_ctx;
3526 
3527 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3528 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3529 
3530 	/*
3531 	 * Need a rate-limiting check so that this isn't called every time
3532 	 */
3533 	iflib_tx_credits_update(ctx, txq);
3534 	reclaim = DESC_RECLAIMABLE(txq);
3535 
3536 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3537 #ifdef INVARIANTS
3538 		if (iflib_verbose_debug) {
3539 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3540 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3541 			       reclaim, thresh);
3542 
3543 		}
3544 #endif
3545 		return (0);
3546 	}
3547 	iflib_tx_desc_free(txq, reclaim);
3548 	txq->ift_cleaned += reclaim;
3549 	txq->ift_in_use -= reclaim;
3550 
3551 	return (reclaim);
3552 }
3553 
3554 static struct mbuf **
3555 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3556 {
3557 	int next, size;
3558 	struct mbuf **items;
3559 
3560 	size = r->size;
3561 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3562 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3563 
3564 	prefetch(items[(cidx + offset) & (size-1)]);
3565 	if (remaining > 1) {
3566 		prefetch2cachelines(&items[next]);
3567 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3568 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3569 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3570 	}
3571 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3572 }
3573 
3574 static void
3575 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3576 {
3577 
3578 	ifmp_ring_check_drainage(txq->ift_br, budget);
3579 }
3580 
3581 static uint32_t
3582 iflib_txq_can_drain(struct ifmp_ring *r)
3583 {
3584 	iflib_txq_t txq = r->cookie;
3585 	if_ctx_t ctx = txq->ift_ctx;
3586 
3587 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3588 		return (1);
3589 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3590 	    BUS_DMASYNC_POSTREAD);
3591 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3592 	    false));
3593 }
3594 
3595 static uint32_t
3596 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3597 {
3598 	iflib_txq_t txq = r->cookie;
3599 	if_ctx_t ctx = txq->ift_ctx;
3600 	if_t ifp = ctx->ifc_ifp;
3601 	struct mbuf *m, **mp;
3602 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3603 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3604 	bool do_prefetch, rang, ring;
3605 
3606 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3607 			    !LINK_ACTIVE(ctx))) {
3608 		DBG_COUNTER_INC(txq_drain_notready);
3609 		return (0);
3610 	}
3611 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3612 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3613 	avail = IDXDIFF(pidx, cidx, r->size);
3614 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3615 		DBG_COUNTER_INC(txq_drain_flushing);
3616 		for (i = 0; i < avail; i++) {
3617 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3618 				m_free(r->items[(cidx + i) & (r->size-1)]);
3619 			r->items[(cidx + i) & (r->size-1)] = NULL;
3620 		}
3621 		return (avail);
3622 	}
3623 
3624 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3625 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3626 		CALLOUT_LOCK(txq);
3627 		callout_stop(&txq->ift_timer);
3628 		CALLOUT_UNLOCK(txq);
3629 		DBG_COUNTER_INC(txq_drain_oactive);
3630 		return (0);
3631 	}
3632 	if (reclaimed)
3633 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3634 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3635 	count = MIN(avail, TX_BATCH_SIZE);
3636 #ifdef INVARIANTS
3637 	if (iflib_verbose_debug)
3638 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3639 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3640 #endif
3641 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3642 	txq_avail = TXQ_AVAIL(txq);
3643 	err = 0;
3644 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3645 		int rem = do_prefetch ? count - i : 0;
3646 
3647 		mp = _ring_peek_one(r, cidx, i, rem);
3648 		MPASS(mp != NULL && *mp != NULL);
3649 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3650 			consumed++;
3651 			continue;
3652 		}
3653 		in_use_prev = txq->ift_in_use;
3654 		err = iflib_encap(txq, mp);
3655 		if (__predict_false(err)) {
3656 			/* no room - bail out */
3657 			if (err == ENOBUFS)
3658 				break;
3659 			consumed++;
3660 			/* we can't send this packet - skip it */
3661 			continue;
3662 		}
3663 		consumed++;
3664 		pkt_sent++;
3665 		m = *mp;
3666 		DBG_COUNTER_INC(tx_sent);
3667 		bytes_sent += m->m_pkthdr.len;
3668 		mcast_sent += !!(m->m_flags & M_MCAST);
3669 		txq_avail = TXQ_AVAIL(txq);
3670 
3671 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3672 		ETHER_BPF_MTAP(ifp, m);
3673 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3674 			break;
3675 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3676 	}
3677 
3678 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3679 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3680 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3681 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3682 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3683 	if (mcast_sent)
3684 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3685 #ifdef INVARIANTS
3686 	if (iflib_verbose_debug)
3687 		printf("consumed=%d\n", consumed);
3688 #endif
3689 	return (consumed);
3690 }
3691 
3692 static uint32_t
3693 iflib_txq_drain_always(struct ifmp_ring *r)
3694 {
3695 	return (1);
3696 }
3697 
3698 static uint32_t
3699 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3700 {
3701 	int i, avail;
3702 	struct mbuf **mp;
3703 	iflib_txq_t txq;
3704 
3705 	txq = r->cookie;
3706 
3707 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3708 	CALLOUT_LOCK(txq);
3709 	callout_stop(&txq->ift_timer);
3710 	CALLOUT_UNLOCK(txq);
3711 
3712 	avail = IDXDIFF(pidx, cidx, r->size);
3713 	for (i = 0; i < avail; i++) {
3714 		mp = _ring_peek_one(r, cidx, i, avail - i);
3715 		if (__predict_false(*mp == (struct mbuf *)txq))
3716 			continue;
3717 		m_freem(*mp);
3718 		DBG_COUNTER_INC(tx_frees);
3719 	}
3720 	MPASS(ifmp_ring_is_stalled(r) == 0);
3721 	return (avail);
3722 }
3723 
3724 static void
3725 iflib_ifmp_purge(iflib_txq_t txq)
3726 {
3727 	struct ifmp_ring *r;
3728 
3729 	r = txq->ift_br;
3730 	r->drain = iflib_txq_drain_free;
3731 	r->can_drain = iflib_txq_drain_always;
3732 
3733 	ifmp_ring_check_drainage(r, r->size);
3734 
3735 	r->drain = iflib_txq_drain;
3736 	r->can_drain = iflib_txq_can_drain;
3737 }
3738 
3739 static void
3740 _task_fn_tx(void *context)
3741 {
3742 	iflib_txq_t txq = context;
3743 	if_ctx_t ctx = txq->ift_ctx;
3744 #if defined(ALTQ) || defined(DEV_NETMAP)
3745 	if_t ifp = ctx->ifc_ifp;
3746 #endif
3747 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3748 
3749 #ifdef IFLIB_DIAGNOSTICS
3750 	txq->ift_cpu_exec_count[curcpu]++;
3751 #endif
3752 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3753 		return;
3754 #ifdef DEV_NETMAP
3755 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3756 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3757 		    BUS_DMASYNC_POSTREAD);
3758 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3759 			netmap_tx_irq(ifp, txq->ift_id);
3760 		if (ctx->ifc_flags & IFC_LEGACY)
3761 			IFDI_INTR_ENABLE(ctx);
3762 		else
3763 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3764 		return;
3765 	}
3766 #endif
3767 #ifdef ALTQ
3768 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3769 		iflib_altq_if_start(ifp);
3770 #endif
3771 	if (txq->ift_db_pending)
3772 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3773 	else if (!abdicate)
3774 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3775 	/*
3776 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3777 	 */
3778 	if (abdicate)
3779 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3780 	if (ctx->ifc_flags & IFC_LEGACY)
3781 		IFDI_INTR_ENABLE(ctx);
3782 	else
3783 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3784 }
3785 
3786 static void
3787 _task_fn_rx(void *context)
3788 {
3789 	iflib_rxq_t rxq = context;
3790 	if_ctx_t ctx = rxq->ifr_ctx;
3791 	uint8_t more;
3792 	uint16_t budget;
3793 
3794 #ifdef IFLIB_DIAGNOSTICS
3795 	rxq->ifr_cpu_exec_count[curcpu]++;
3796 #endif
3797 	DBG_COUNTER_INC(task_fn_rxs);
3798 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3799 		return;
3800 #ifdef DEV_NETMAP
3801 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3802 		u_int work = 0;
3803 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3804 			more = 0;
3805 			goto skip_rxeof;
3806 		}
3807 	}
3808 #endif
3809 	budget = ctx->ifc_sysctl_rx_budget;
3810 	if (budget == 0)
3811 		budget = 16;	/* XXX */
3812 	more = iflib_rxeof(rxq, budget);
3813 #ifdef DEV_NETMAP
3814 skip_rxeof:
3815 #endif
3816 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3817 		if (ctx->ifc_flags & IFC_LEGACY)
3818 			IFDI_INTR_ENABLE(ctx);
3819 		else
3820 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3821 		DBG_COUNTER_INC(rx_intr_enables);
3822 	}
3823 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3824 		return;
3825 
3826 	if (more & IFLIB_RXEOF_MORE)
3827 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3828 	else if (more & IFLIB_RXEOF_EMPTY)
3829 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3830 }
3831 
3832 static void
3833 _task_fn_admin(void *context)
3834 {
3835 	if_ctx_t ctx = context;
3836 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3837 	iflib_txq_t txq;
3838 	int i;
3839 	bool oactive, running, do_reset, do_watchdog, in_detach;
3840 	uint32_t reset_on = hz / 2;
3841 
3842 	STATE_LOCK(ctx);
3843 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3844 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3845 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3846 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3847 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3848 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3849 	STATE_UNLOCK(ctx);
3850 
3851 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3852 		return;
3853 	if (in_detach)
3854 		return;
3855 
3856 	CTX_LOCK(ctx);
3857 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3858 		CALLOUT_LOCK(txq);
3859 		callout_stop(&txq->ift_timer);
3860 		CALLOUT_UNLOCK(txq);
3861 	}
3862 	if (do_watchdog) {
3863 		ctx->ifc_watchdog_events++;
3864 		IFDI_WATCHDOG_RESET(ctx);
3865 	}
3866 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3867 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3868 #ifdef DEV_NETMAP
3869 		reset_on = hz / 2;
3870 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3871 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3872 #endif
3873 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3874 	}
3875 	IFDI_LINK_INTR_ENABLE(ctx);
3876 	if (do_reset)
3877 		iflib_if_init_locked(ctx);
3878 	CTX_UNLOCK(ctx);
3879 
3880 	if (LINK_ACTIVE(ctx) == 0)
3881 		return;
3882 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3883 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3884 }
3885 
3886 
3887 static void
3888 _task_fn_iov(void *context)
3889 {
3890 	if_ctx_t ctx = context;
3891 
3892 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3893 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3894 		return;
3895 
3896 	CTX_LOCK(ctx);
3897 	IFDI_VFLR_HANDLE(ctx);
3898 	CTX_UNLOCK(ctx);
3899 }
3900 
3901 static int
3902 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3903 {
3904 	int err;
3905 	if_int_delay_info_t info;
3906 	if_ctx_t ctx;
3907 
3908 	info = (if_int_delay_info_t)arg1;
3909 	ctx = info->iidi_ctx;
3910 	info->iidi_req = req;
3911 	info->iidi_oidp = oidp;
3912 	CTX_LOCK(ctx);
3913 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3914 	CTX_UNLOCK(ctx);
3915 	return (err);
3916 }
3917 
3918 /*********************************************************************
3919  *
3920  *  IFNET FUNCTIONS
3921  *
3922  **********************************************************************/
3923 
3924 static void
3925 iflib_if_init_locked(if_ctx_t ctx)
3926 {
3927 	iflib_stop(ctx);
3928 	iflib_init_locked(ctx);
3929 }
3930 
3931 
3932 static void
3933 iflib_if_init(void *arg)
3934 {
3935 	if_ctx_t ctx = arg;
3936 
3937 	CTX_LOCK(ctx);
3938 	iflib_if_init_locked(ctx);
3939 	CTX_UNLOCK(ctx);
3940 }
3941 
3942 static int
3943 iflib_if_transmit(if_t ifp, struct mbuf *m)
3944 {
3945 	if_ctx_t	ctx = if_getsoftc(ifp);
3946 
3947 	iflib_txq_t txq;
3948 	int err, qidx;
3949 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3950 
3951 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3952 		DBG_COUNTER_INC(tx_frees);
3953 		m_freem(m);
3954 		return (ENETDOWN);
3955 	}
3956 
3957 	MPASS(m->m_nextpkt == NULL);
3958 	/* ALTQ-enabled interfaces always use queue 0. */
3959 	qidx = 0;
3960 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3961 		qidx = QIDX(ctx, m);
3962 	/*
3963 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3964 	 */
3965 	txq = &ctx->ifc_txqs[qidx];
3966 
3967 #ifdef DRIVER_BACKPRESSURE
3968 	if (txq->ift_closed) {
3969 		while (m != NULL) {
3970 			next = m->m_nextpkt;
3971 			m->m_nextpkt = NULL;
3972 			m_freem(m);
3973 			DBG_COUNTER_INC(tx_frees);
3974 			m = next;
3975 		}
3976 		return (ENOBUFS);
3977 	}
3978 #endif
3979 #ifdef notyet
3980 	qidx = count = 0;
3981 	mp = marr;
3982 	next = m;
3983 	do {
3984 		count++;
3985 		next = next->m_nextpkt;
3986 	} while (next != NULL);
3987 
3988 	if (count > nitems(marr))
3989 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3990 			/* XXX check nextpkt */
3991 			m_freem(m);
3992 			/* XXX simplify for now */
3993 			DBG_COUNTER_INC(tx_frees);
3994 			return (ENOBUFS);
3995 		}
3996 	for (next = m, i = 0; next != NULL; i++) {
3997 		mp[i] = next;
3998 		next = next->m_nextpkt;
3999 		mp[i]->m_nextpkt = NULL;
4000 	}
4001 #endif
4002 	DBG_COUNTER_INC(tx_seen);
4003 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4004 
4005 	if (abdicate)
4006 		GROUPTASK_ENQUEUE(&txq->ift_task);
4007  	if (err) {
4008 		if (!abdicate)
4009 			GROUPTASK_ENQUEUE(&txq->ift_task);
4010 		/* support forthcoming later */
4011 #ifdef DRIVER_BACKPRESSURE
4012 		txq->ift_closed = TRUE;
4013 #endif
4014 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4015 		m_freem(m);
4016 		DBG_COUNTER_INC(tx_frees);
4017 	}
4018 
4019 	return (err);
4020 }
4021 
4022 #ifdef ALTQ
4023 /*
4024  * The overall approach to integrating iflib with ALTQ is to continue to use
4025  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4026  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4027  * is redundant/unnecessary, but doing so minimizes the amount of
4028  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4029  * redundantly queueing to an intermediate mp_ring is swamped by the
4030  * performance limitations inherent in using ALTQ.
4031  *
4032  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4033  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4034  * given interface.  If ALTQ is enabled for an interface, then all
4035  * transmitted packets for that interface will be submitted to the ALTQ
4036  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4037  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4038  * update stats that the iflib machinery handles, and which is sensitve to
4039  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4040  * will be installed as the start routine for use by ALTQ facilities that
4041  * need to trigger queue drains on a scheduled basis.
4042  *
4043  */
4044 static void
4045 iflib_altq_if_start(if_t ifp)
4046 {
4047 	struct ifaltq *ifq = &ifp->if_snd;
4048 	struct mbuf *m;
4049 
4050 	IFQ_LOCK(ifq);
4051 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4052 	while (m != NULL) {
4053 		iflib_if_transmit(ifp, m);
4054 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4055 	}
4056 	IFQ_UNLOCK(ifq);
4057 }
4058 
4059 static int
4060 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4061 {
4062 	int err;
4063 
4064 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4065 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4066 		if (err == 0)
4067 			iflib_altq_if_start(ifp);
4068 	} else
4069 		err = iflib_if_transmit(ifp, m);
4070 
4071 	return (err);
4072 }
4073 #endif /* ALTQ */
4074 
4075 static void
4076 iflib_if_qflush(if_t ifp)
4077 {
4078 	if_ctx_t ctx = if_getsoftc(ifp);
4079 	iflib_txq_t txq = ctx->ifc_txqs;
4080 	int i;
4081 
4082 	STATE_LOCK(ctx);
4083 	ctx->ifc_flags |= IFC_QFLUSH;
4084 	STATE_UNLOCK(ctx);
4085 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4086 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4087 			iflib_txq_check_drain(txq, 0);
4088 	STATE_LOCK(ctx);
4089 	ctx->ifc_flags &= ~IFC_QFLUSH;
4090 	STATE_UNLOCK(ctx);
4091 
4092 	/*
4093 	 * When ALTQ is enabled, this will also take care of purging the
4094 	 * ALTQ queue(s).
4095 	 */
4096 	if_qflush(ifp);
4097 }
4098 
4099 
4100 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4101 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4102 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4103 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4104 
4105 static int
4106 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4107 {
4108 	if_ctx_t ctx = if_getsoftc(ifp);
4109 	struct ifreq	*ifr = (struct ifreq *)data;
4110 #if defined(INET) || defined(INET6)
4111 	struct ifaddr	*ifa = (struct ifaddr *)data;
4112 #endif
4113 	bool		avoid_reset = false;
4114 	int		err = 0, reinit = 0, bits;
4115 
4116 	switch (command) {
4117 	case SIOCSIFADDR:
4118 #ifdef INET
4119 		if (ifa->ifa_addr->sa_family == AF_INET)
4120 			avoid_reset = true;
4121 #endif
4122 #ifdef INET6
4123 		if (ifa->ifa_addr->sa_family == AF_INET6)
4124 			avoid_reset = true;
4125 #endif
4126 		/*
4127 		** Calling init results in link renegotiation,
4128 		** so we avoid doing it when possible.
4129 		*/
4130 		if (avoid_reset) {
4131 			if_setflagbits(ifp, IFF_UP,0);
4132 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4133 				reinit = 1;
4134 #ifdef INET
4135 			if (!(if_getflags(ifp) & IFF_NOARP))
4136 				arp_ifinit(ifp, ifa);
4137 #endif
4138 		} else
4139 			err = ether_ioctl(ifp, command, data);
4140 		break;
4141 	case SIOCSIFMTU:
4142 		CTX_LOCK(ctx);
4143 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4144 			CTX_UNLOCK(ctx);
4145 			break;
4146 		}
4147 		bits = if_getdrvflags(ifp);
4148 		/* stop the driver and free any clusters before proceeding */
4149 		iflib_stop(ctx);
4150 
4151 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4152 			STATE_LOCK(ctx);
4153 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4154 				ctx->ifc_flags |= IFC_MULTISEG;
4155 			else
4156 				ctx->ifc_flags &= ~IFC_MULTISEG;
4157 			STATE_UNLOCK(ctx);
4158 			err = if_setmtu(ifp, ifr->ifr_mtu);
4159 		}
4160 		iflib_init_locked(ctx);
4161 		STATE_LOCK(ctx);
4162 		if_setdrvflags(ifp, bits);
4163 		STATE_UNLOCK(ctx);
4164 		CTX_UNLOCK(ctx);
4165 		break;
4166 	case SIOCSIFFLAGS:
4167 		CTX_LOCK(ctx);
4168 		if (if_getflags(ifp) & IFF_UP) {
4169 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4170 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4171 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4172 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4173 				}
4174 			} else
4175 				reinit = 1;
4176 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4177 			iflib_stop(ctx);
4178 		}
4179 		ctx->ifc_if_flags = if_getflags(ifp);
4180 		CTX_UNLOCK(ctx);
4181 		break;
4182 	case SIOCADDMULTI:
4183 	case SIOCDELMULTI:
4184 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4185 			CTX_LOCK(ctx);
4186 			IFDI_INTR_DISABLE(ctx);
4187 			IFDI_MULTI_SET(ctx);
4188 			IFDI_INTR_ENABLE(ctx);
4189 			CTX_UNLOCK(ctx);
4190 		}
4191 		break;
4192 	case SIOCSIFMEDIA:
4193 		CTX_LOCK(ctx);
4194 		IFDI_MEDIA_SET(ctx);
4195 		CTX_UNLOCK(ctx);
4196 		/* FALLTHROUGH */
4197 	case SIOCGIFMEDIA:
4198 	case SIOCGIFXMEDIA:
4199 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4200 		break;
4201 	case SIOCGI2C:
4202 	{
4203 		struct ifi2creq i2c;
4204 
4205 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4206 		if (err != 0)
4207 			break;
4208 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4209 			err = EINVAL;
4210 			break;
4211 		}
4212 		if (i2c.len > sizeof(i2c.data)) {
4213 			err = EINVAL;
4214 			break;
4215 		}
4216 
4217 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4218 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4219 			    sizeof(i2c));
4220 		break;
4221 	}
4222 	case SIOCSIFCAP:
4223 	{
4224 		int mask, setmask, oldmask;
4225 
4226 		oldmask = if_getcapenable(ifp);
4227 		mask = ifr->ifr_reqcap ^ oldmask;
4228 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4229 		setmask = 0;
4230 #ifdef TCP_OFFLOAD
4231 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4232 #endif
4233 		setmask |= (mask & IFCAP_FLAGS);
4234 		setmask |= (mask & IFCAP_WOL);
4235 
4236 		/*
4237 		 * If any RX csum has changed, change all the ones that
4238 		 * are supported by the driver.
4239 		 */
4240 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4241 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4242 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4243 		}
4244 
4245 		/*
4246 		 * want to ensure that traffic has stopped before we change any of the flags
4247 		 */
4248 		if (setmask) {
4249 			CTX_LOCK(ctx);
4250 			bits = if_getdrvflags(ifp);
4251 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4252 				iflib_stop(ctx);
4253 			STATE_LOCK(ctx);
4254 			if_togglecapenable(ifp, setmask);
4255 			STATE_UNLOCK(ctx);
4256 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4257 				iflib_init_locked(ctx);
4258 			STATE_LOCK(ctx);
4259 			if_setdrvflags(ifp, bits);
4260 			STATE_UNLOCK(ctx);
4261 			CTX_UNLOCK(ctx);
4262 		}
4263 		if_vlancap(ifp);
4264 		break;
4265 	}
4266 	case SIOCGPRIVATE_0:
4267 	case SIOCSDRVSPEC:
4268 	case SIOCGDRVSPEC:
4269 		CTX_LOCK(ctx);
4270 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4271 		CTX_UNLOCK(ctx);
4272 		break;
4273 	default:
4274 		err = ether_ioctl(ifp, command, data);
4275 		break;
4276 	}
4277 	if (reinit)
4278 		iflib_if_init(ctx);
4279 	return (err);
4280 }
4281 
4282 static uint64_t
4283 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4284 {
4285 	if_ctx_t ctx = if_getsoftc(ifp);
4286 
4287 	return (IFDI_GET_COUNTER(ctx, cnt));
4288 }
4289 
4290 /*********************************************************************
4291  *
4292  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4293  *
4294  **********************************************************************/
4295 
4296 static void
4297 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4298 {
4299 	if_ctx_t ctx = if_getsoftc(ifp);
4300 
4301 	if ((void *)ctx != arg)
4302 		return;
4303 
4304 	if ((vtag == 0) || (vtag > 4095))
4305 		return;
4306 
4307 	if (iflib_in_detach(ctx))
4308 		return;
4309 
4310 	CTX_LOCK(ctx);
4311 	IFDI_VLAN_REGISTER(ctx, vtag);
4312 	/* Re-init to load the changes */
4313 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4314 		iflib_if_init_locked(ctx);
4315 	CTX_UNLOCK(ctx);
4316 }
4317 
4318 static void
4319 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4320 {
4321 	if_ctx_t ctx = if_getsoftc(ifp);
4322 
4323 	if ((void *)ctx != arg)
4324 		return;
4325 
4326 	if ((vtag == 0) || (vtag > 4095))
4327 		return;
4328 
4329 	CTX_LOCK(ctx);
4330 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4331 	/* Re-init to load the changes */
4332 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4333 		iflib_if_init_locked(ctx);
4334 	CTX_UNLOCK(ctx);
4335 }
4336 
4337 static void
4338 iflib_led_func(void *arg, int onoff)
4339 {
4340 	if_ctx_t ctx = arg;
4341 
4342 	CTX_LOCK(ctx);
4343 	IFDI_LED_FUNC(ctx, onoff);
4344 	CTX_UNLOCK(ctx);
4345 }
4346 
4347 /*********************************************************************
4348  *
4349  *  BUS FUNCTION DEFINITIONS
4350  *
4351  **********************************************************************/
4352 
4353 int
4354 iflib_device_probe(device_t dev)
4355 {
4356 	const pci_vendor_info_t *ent;
4357 	if_shared_ctx_t sctx;
4358 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4359 	uint16_t pci_vendor_id;
4360 
4361 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4362 		return (ENOTSUP);
4363 
4364 	pci_vendor_id = pci_get_vendor(dev);
4365 	pci_device_id = pci_get_device(dev);
4366 	pci_subvendor_id = pci_get_subvendor(dev);
4367 	pci_subdevice_id = pci_get_subdevice(dev);
4368 	pci_rev_id = pci_get_revid(dev);
4369 	if (sctx->isc_parse_devinfo != NULL)
4370 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4371 
4372 	ent = sctx->isc_vendor_info;
4373 	while (ent->pvi_vendor_id != 0) {
4374 		if (pci_vendor_id != ent->pvi_vendor_id) {
4375 			ent++;
4376 			continue;
4377 		}
4378 		if ((pci_device_id == ent->pvi_device_id) &&
4379 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4380 		     (ent->pvi_subvendor_id == 0)) &&
4381 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4382 		     (ent->pvi_subdevice_id == 0)) &&
4383 		    ((pci_rev_id == ent->pvi_rev_id) ||
4384 		     (ent->pvi_rev_id == 0))) {
4385 
4386 			device_set_desc_copy(dev, ent->pvi_name);
4387 			/* this needs to be changed to zero if the bus probing code
4388 			 * ever stops re-probing on best match because the sctx
4389 			 * may have its values over written by register calls
4390 			 * in subsequent probes
4391 			 */
4392 			return (BUS_PROBE_DEFAULT);
4393 		}
4394 		ent++;
4395 	}
4396 	return (ENXIO);
4397 }
4398 
4399 int
4400 iflib_device_probe_vendor(device_t dev)
4401 {
4402 	int probe;
4403 
4404 	probe = iflib_device_probe(dev);
4405 	if (probe == BUS_PROBE_DEFAULT)
4406 		return (BUS_PROBE_VENDOR);
4407 	else
4408 		return (probe);
4409 }
4410 
4411 static void
4412 iflib_reset_qvalues(if_ctx_t ctx)
4413 {
4414 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4415 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4416 	device_t dev = ctx->ifc_dev;
4417 	int i;
4418 
4419 	if (ctx->ifc_sysctl_ntxqs != 0)
4420 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4421 	if (ctx->ifc_sysctl_nrxqs != 0)
4422 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4423 
4424 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4425 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4426 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4427 		else
4428 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4429 	}
4430 
4431 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4432 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4433 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4434 		else
4435 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4436 	}
4437 
4438 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4439 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4440 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4441 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4442 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4443 		}
4444 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4445 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4446 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4447 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4448 		}
4449 		if (!powerof2(scctx->isc_nrxd[i])) {
4450 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4451 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4452 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4453 		}
4454 	}
4455 
4456 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4457 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4458 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4459 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4460 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4461 		}
4462 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4463 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4464 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4465 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4466 		}
4467 		if (!powerof2(scctx->isc_ntxd[i])) {
4468 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4469 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4470 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4471 		}
4472 	}
4473 }
4474 
4475 static void
4476 iflib_add_pfil(if_ctx_t ctx)
4477 {
4478 	struct pfil_head *pfil;
4479 	struct pfil_head_args pa;
4480 	iflib_rxq_t rxq;
4481 	int i;
4482 
4483 	pa.pa_version = PFIL_VERSION;
4484 	pa.pa_flags = PFIL_IN;
4485 	pa.pa_type = PFIL_TYPE_ETHERNET;
4486 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4487 	pfil = pfil_head_register(&pa);
4488 
4489 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4490 		rxq->pfil = pfil;
4491 	}
4492 }
4493 
4494 static void
4495 iflib_rem_pfil(if_ctx_t ctx)
4496 {
4497 	struct pfil_head *pfil;
4498 	iflib_rxq_t rxq;
4499 	int i;
4500 
4501 	rxq = ctx->ifc_rxqs;
4502 	pfil = rxq->pfil;
4503 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4504 		rxq->pfil = NULL;
4505 	}
4506 	pfil_head_unregister(pfil);
4507 }
4508 
4509 static uint16_t
4510 get_ctx_core_offset(if_ctx_t ctx)
4511 {
4512 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4513 	struct cpu_offset *op;
4514 	uint16_t qc;
4515 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4516 
4517 	if (ret != CORE_OFFSET_UNSPECIFIED)
4518 		return (ret);
4519 
4520 	if (ctx->ifc_sysctl_separate_txrx)
4521 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4522 	else
4523 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4524 
4525 	mtx_lock(&cpu_offset_mtx);
4526 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4527 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4528 			ret = op->offset;
4529 			op->offset += qc;
4530 			MPASS(op->refcount < UINT_MAX);
4531 			op->refcount++;
4532 			break;
4533 		}
4534 	}
4535 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4536 		ret = 0;
4537 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4538 		    M_NOWAIT | M_ZERO);
4539 		if (op == NULL) {
4540 			device_printf(ctx->ifc_dev,
4541 			    "allocation for cpu offset failed.\n");
4542 		} else {
4543 			op->offset = qc;
4544 			op->refcount = 1;
4545 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4546 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4547 		}
4548 	}
4549 	mtx_unlock(&cpu_offset_mtx);
4550 
4551 	return (ret);
4552 }
4553 
4554 static void
4555 unref_ctx_core_offset(if_ctx_t ctx)
4556 {
4557 	struct cpu_offset *op, *top;
4558 
4559 	mtx_lock(&cpu_offset_mtx);
4560 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4561 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4562 			MPASS(op->refcount > 0);
4563 			op->refcount--;
4564 			if (op->refcount == 0) {
4565 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4566 				free(op, M_IFLIB);
4567 			}
4568 			break;
4569 		}
4570 	}
4571 	mtx_unlock(&cpu_offset_mtx);
4572 }
4573 
4574 int
4575 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4576 {
4577 	if_ctx_t ctx;
4578 	if_t ifp;
4579 	if_softc_ctx_t scctx;
4580 	kobjop_desc_t kobj_desc;
4581 	kobj_method_t *kobj_method;
4582 	int err, msix, rid;
4583 	uint16_t main_rxq, main_txq;
4584 
4585 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4586 
4587 	if (sc == NULL) {
4588 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4589 		device_set_softc(dev, ctx);
4590 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4591 	}
4592 
4593 	ctx->ifc_sctx = sctx;
4594 	ctx->ifc_dev = dev;
4595 	ctx->ifc_softc = sc;
4596 
4597 	if ((err = iflib_register(ctx)) != 0) {
4598 		device_printf(dev, "iflib_register failed %d\n", err);
4599 		goto fail_ctx_free;
4600 	}
4601 	iflib_add_device_sysctl_pre(ctx);
4602 
4603 	scctx = &ctx->ifc_softc_ctx;
4604 	ifp = ctx->ifc_ifp;
4605 
4606 	iflib_reset_qvalues(ctx);
4607 	CTX_LOCK(ctx);
4608 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4609 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4610 		goto fail_unlock;
4611 	}
4612 	_iflib_pre_assert(scctx);
4613 	ctx->ifc_txrx = *scctx->isc_txrx;
4614 
4615 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4616 		ctx->ifc_mediap = scctx->isc_media;
4617 
4618 #ifdef INVARIANTS
4619 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4620 		MPASS(scctx->isc_tx_csum_flags);
4621 #endif
4622 
4623 	if_setcapabilities(ifp,
4624 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4625 	if_setcapenable(ifp,
4626 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4627 
4628 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4629 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4630 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4631 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4632 
4633 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4634 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4635 
4636 	/* XXX change for per-queue sizes */
4637 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4638 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4639 
4640 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4641 	    MAX_SINGLE_PACKET_FRACTION)
4642 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4643 		    MAX_SINGLE_PACKET_FRACTION);
4644 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4645 	    MAX_SINGLE_PACKET_FRACTION)
4646 		scctx->isc_tx_tso_segments_max = max(1,
4647 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4648 
4649 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4650 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4651 		/*
4652 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4653 		 * but some MACs do.
4654 		 */
4655 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4656 		    IP_MAXPACKET));
4657 		/*
4658 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4659 		 * into account.  In the worst case, each of these calls will
4660 		 * add another mbuf and, thus, the requirement for another DMA
4661 		 * segment.  So for best performance, it doesn't make sense to
4662 		 * advertize a maximum of TSO segments that typically will
4663 		 * require defragmentation in iflib_encap().
4664 		 */
4665 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4666 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4667 	}
4668 	if (scctx->isc_rss_table_size == 0)
4669 		scctx->isc_rss_table_size = 64;
4670 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4671 
4672 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4673 	/* XXX format name */
4674 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4675 	    NULL, NULL, "admin");
4676 
4677 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4678 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4679 		device_printf(dev, "Unable to fetch CPU list\n");
4680 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4681 	}
4682 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4683 
4684 	/*
4685 	** Now set up MSI or MSI-X, should return us the number of supported
4686 	** vectors (will be 1 for a legacy interrupt and MSI).
4687 	*/
4688 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4689 		msix = scctx->isc_vectors;
4690 	} else if (scctx->isc_msix_bar != 0)
4691 	       /*
4692 		* The simple fact that isc_msix_bar is not 0 does not mean we
4693 		* we have a good value there that is known to work.
4694 		*/
4695 		msix = iflib_msix_init(ctx);
4696 	else {
4697 		scctx->isc_vectors = 1;
4698 		scctx->isc_ntxqsets = 1;
4699 		scctx->isc_nrxqsets = 1;
4700 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4701 		msix = 0;
4702 	}
4703 	/* Get memory for the station queues */
4704 	if ((err = iflib_queues_alloc(ctx))) {
4705 		device_printf(dev, "Unable to allocate queue memory\n");
4706 		goto fail_intr_free;
4707 	}
4708 
4709 	if ((err = iflib_qset_structures_setup(ctx)))
4710 		goto fail_queues;
4711 
4712 	/*
4713 	 * Now that we know how many queues there are, get the core offset.
4714 	 */
4715 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4716 
4717 	if (msix > 1) {
4718 		/*
4719 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4720 		 * aren't the default NULL implementation.
4721 		 */
4722 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4723 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4724 		    kobj_desc);
4725 		if (kobj_method == &kobj_desc->deflt) {
4726 			device_printf(dev,
4727 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4728 			err = EOPNOTSUPP;
4729 			goto fail_queues;
4730 		}
4731 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4732 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4733 		    kobj_desc);
4734 		if (kobj_method == &kobj_desc->deflt) {
4735 			device_printf(dev,
4736 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4737 			err = EOPNOTSUPP;
4738 			goto fail_queues;
4739 		}
4740 
4741 		/*
4742 		 * Assign the MSI-X vectors.
4743 		 * Note that the default NULL ifdi_msix_intr_assign method will
4744 		 * fail here, too.
4745 		 */
4746 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4747 		if (err != 0) {
4748 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4749 			    err);
4750 			goto fail_queues;
4751 		}
4752 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4753 		rid = 0;
4754 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4755 			MPASS(msix == 1);
4756 			rid = 1;
4757 		}
4758 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4759 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4760 			goto fail_queues;
4761 		}
4762 	} else {
4763 		device_printf(dev,
4764 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4765 		err = ENODEV;
4766 		goto fail_intr_free;
4767 	}
4768 
4769 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4770 
4771 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4772 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4773 		goto fail_detach;
4774 	}
4775 
4776 	/*
4777 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4778 	 * This must appear after the call to ether_ifattach() because
4779 	 * ether_ifattach() sets if_hdrlen to the default value.
4780 	 */
4781 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4782 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4783 
4784 	if ((err = iflib_netmap_attach(ctx))) {
4785 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4786 		goto fail_detach;
4787 	}
4788 	*ctxp = ctx;
4789 
4790 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4791 
4792 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4793 	iflib_add_device_sysctl_post(ctx);
4794 	iflib_add_pfil(ctx);
4795 	ctx->ifc_flags |= IFC_INIT_DONE;
4796 	CTX_UNLOCK(ctx);
4797 
4798 	return (0);
4799 
4800 fail_detach:
4801 	ether_ifdetach(ctx->ifc_ifp);
4802 fail_intr_free:
4803 	iflib_free_intr_mem(ctx);
4804 fail_queues:
4805 	iflib_tx_structures_free(ctx);
4806 	iflib_rx_structures_free(ctx);
4807 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4808 	IFDI_DETACH(ctx);
4809 fail_unlock:
4810 	CTX_UNLOCK(ctx);
4811 	iflib_deregister(ctx);
4812 fail_ctx_free:
4813 	device_set_softc(ctx->ifc_dev, NULL);
4814         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4815                 free(ctx->ifc_softc, M_IFLIB);
4816         free(ctx, M_IFLIB);
4817 	return (err);
4818 }
4819 
4820 int
4821 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4822 					  struct iflib_cloneattach_ctx *clctx)
4823 {
4824 	int err;
4825 	if_ctx_t ctx;
4826 	if_t ifp;
4827 	if_softc_ctx_t scctx;
4828 	int i;
4829 	void *sc;
4830 	uint16_t main_txq;
4831 	uint16_t main_rxq;
4832 
4833 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4834 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4835 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4836 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4837 		ctx->ifc_flags |= IFC_PSEUDO;
4838 
4839 	ctx->ifc_sctx = sctx;
4840 	ctx->ifc_softc = sc;
4841 	ctx->ifc_dev = dev;
4842 
4843 	if ((err = iflib_register(ctx)) != 0) {
4844 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4845 		goto fail_ctx_free;
4846 	}
4847 	iflib_add_device_sysctl_pre(ctx);
4848 
4849 	scctx = &ctx->ifc_softc_ctx;
4850 	ifp = ctx->ifc_ifp;
4851 
4852 	iflib_reset_qvalues(ctx);
4853 	CTX_LOCK(ctx);
4854 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4855 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4856 		goto fail_unlock;
4857 	}
4858 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4859 		ether_gen_addr(ifp, &ctx->ifc_mac);
4860 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4861 								clctx->cc_params)) != 0) {
4862 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4863 		goto fail_ctx_free;
4864 	}
4865 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4866 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4867 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4868 
4869 #ifdef INVARIANTS
4870 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4871 		MPASS(scctx->isc_tx_csum_flags);
4872 #endif
4873 
4874 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4875 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4876 
4877 	ifp->if_flags |= IFF_NOGROUP;
4878 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4879 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4880 
4881 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4882 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4883 			goto fail_detach;
4884 		}
4885 		*ctxp = ctx;
4886 
4887 		/*
4888 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4889 		 * This must appear after the call to ether_ifattach() because
4890 		 * ether_ifattach() sets if_hdrlen to the default value.
4891 		 */
4892 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4893 			if_setifheaderlen(ifp,
4894 			    sizeof(struct ether_vlan_header));
4895 
4896 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4897 		iflib_add_device_sysctl_post(ctx);
4898 		ctx->ifc_flags |= IFC_INIT_DONE;
4899 		return (0);
4900 	}
4901 	_iflib_pre_assert(scctx);
4902 	ctx->ifc_txrx = *scctx->isc_txrx;
4903 
4904 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4905 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4906 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4907 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4908 
4909 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4910 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4911 
4912 	/* XXX change for per-queue sizes */
4913 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4914 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4915 
4916 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4917 	    MAX_SINGLE_PACKET_FRACTION)
4918 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4919 		    MAX_SINGLE_PACKET_FRACTION);
4920 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4921 	    MAX_SINGLE_PACKET_FRACTION)
4922 		scctx->isc_tx_tso_segments_max = max(1,
4923 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4924 
4925 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4926 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4927 		/*
4928 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4929 		 * but some MACs do.
4930 		 */
4931 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4932 		    IP_MAXPACKET));
4933 		/*
4934 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4935 		 * into account.  In the worst case, each of these calls will
4936 		 * add another mbuf and, thus, the requirement for another DMA
4937 		 * segment.  So for best performance, it doesn't make sense to
4938 		 * advertize a maximum of TSO segments that typically will
4939 		 * require defragmentation in iflib_encap().
4940 		 */
4941 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4942 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4943 	}
4944 	if (scctx->isc_rss_table_size == 0)
4945 		scctx->isc_rss_table_size = 64;
4946 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4947 
4948 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4949 	/* XXX format name */
4950 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4951 	    NULL, NULL, "admin");
4952 
4953 	/* XXX --- can support > 1 -- but keep it simple for now */
4954 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4955 
4956 	/* Get memory for the station queues */
4957 	if ((err = iflib_queues_alloc(ctx))) {
4958 		device_printf(dev, "Unable to allocate queue memory\n");
4959 		goto fail_iflib_detach;
4960 	}
4961 
4962 	if ((err = iflib_qset_structures_setup(ctx))) {
4963 		device_printf(dev, "qset structure setup failed %d\n", err);
4964 		goto fail_queues;
4965 	}
4966 
4967 	/*
4968 	 * XXX What if anything do we want to do about interrupts?
4969 	 */
4970 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4971 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4972 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4973 		goto fail_detach;
4974 	}
4975 
4976 	/*
4977 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4978 	 * This must appear after the call to ether_ifattach() because
4979 	 * ether_ifattach() sets if_hdrlen to the default value.
4980 	 */
4981 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4982 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4983 
4984 	/* XXX handle more than one queue */
4985 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4986 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4987 
4988 	*ctxp = ctx;
4989 
4990 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4991 	iflib_add_device_sysctl_post(ctx);
4992 	ctx->ifc_flags |= IFC_INIT_DONE;
4993 	CTX_UNLOCK(ctx);
4994 
4995 	return (0);
4996 fail_detach:
4997 	ether_ifdetach(ctx->ifc_ifp);
4998 fail_queues:
4999 	iflib_tx_structures_free(ctx);
5000 	iflib_rx_structures_free(ctx);
5001 fail_iflib_detach:
5002 	IFDI_DETACH(ctx);
5003 fail_unlock:
5004 	CTX_UNLOCK(ctx);
5005 	iflib_deregister(ctx);
5006 fail_ctx_free:
5007 	free(ctx->ifc_softc, M_IFLIB);
5008 	free(ctx, M_IFLIB);
5009 	return (err);
5010 }
5011 
5012 int
5013 iflib_pseudo_deregister(if_ctx_t ctx)
5014 {
5015 	if_t ifp = ctx->ifc_ifp;
5016 	iflib_txq_t txq;
5017 	iflib_rxq_t rxq;
5018 	int i, j;
5019 	struct taskqgroup *tqg;
5020 	iflib_fl_t fl;
5021 
5022 	/* Unregister VLAN event handlers early */
5023 	iflib_unregister_vlan_handlers(ctx);
5024 
5025 	ether_ifdetach(ifp);
5026 	/* XXX drain any dependent tasks */
5027 	tqg = qgroup_if_io_tqg;
5028 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5029 		callout_drain(&txq->ift_timer);
5030 		if (txq->ift_task.gt_uniq != NULL)
5031 			taskqgroup_detach(tqg, &txq->ift_task);
5032 	}
5033 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5034 		callout_drain(&rxq->ifr_watchdog);
5035 		if (rxq->ifr_task.gt_uniq != NULL)
5036 			taskqgroup_detach(tqg, &rxq->ifr_task);
5037 
5038 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5039 			free(fl->ifl_rx_bitmap, M_IFLIB);
5040 	}
5041 	tqg = qgroup_if_config_tqg;
5042 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5043 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5044 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5045 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5046 
5047 	iflib_tx_structures_free(ctx);
5048 	iflib_rx_structures_free(ctx);
5049 
5050 	iflib_deregister(ctx);
5051 
5052 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5053 		free(ctx->ifc_softc, M_IFLIB);
5054 	free(ctx, M_IFLIB);
5055 	return (0);
5056 }
5057 
5058 int
5059 iflib_device_attach(device_t dev)
5060 {
5061 	if_ctx_t ctx;
5062 	if_shared_ctx_t sctx;
5063 
5064 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5065 		return (ENOTSUP);
5066 
5067 	pci_enable_busmaster(dev);
5068 
5069 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5070 }
5071 
5072 int
5073 iflib_device_deregister(if_ctx_t ctx)
5074 {
5075 	if_t ifp = ctx->ifc_ifp;
5076 	iflib_txq_t txq;
5077 	iflib_rxq_t rxq;
5078 	device_t dev = ctx->ifc_dev;
5079 	int i, j;
5080 	struct taskqgroup *tqg;
5081 	iflib_fl_t fl;
5082 
5083 	/* Make sure VLANS are not using driver */
5084 	if (if_vlantrunkinuse(ifp)) {
5085 		device_printf(dev, "Vlan in use, detach first\n");
5086 		return (EBUSY);
5087 	}
5088 #ifdef PCI_IOV
5089 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5090 		device_printf(dev, "SR-IOV in use; detach first.\n");
5091 		return (EBUSY);
5092 	}
5093 #endif
5094 
5095 	STATE_LOCK(ctx);
5096 	ctx->ifc_flags |= IFC_IN_DETACH;
5097 	STATE_UNLOCK(ctx);
5098 
5099 	/* Unregister VLAN handlers before calling iflib_stop() */
5100 	iflib_unregister_vlan_handlers(ctx);
5101 
5102 	iflib_netmap_detach(ifp);
5103 	ether_ifdetach(ifp);
5104 
5105 	CTX_LOCK(ctx);
5106 	iflib_stop(ctx);
5107 	CTX_UNLOCK(ctx);
5108 
5109 	iflib_rem_pfil(ctx);
5110 	if (ctx->ifc_led_dev != NULL)
5111 		led_destroy(ctx->ifc_led_dev);
5112 	/* XXX drain any dependent tasks */
5113 	tqg = qgroup_if_io_tqg;
5114 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5115 		callout_drain(&txq->ift_timer);
5116 		if (txq->ift_task.gt_uniq != NULL)
5117 			taskqgroup_detach(tqg, &txq->ift_task);
5118 	}
5119 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5120 		if (rxq->ifr_task.gt_uniq != NULL)
5121 			taskqgroup_detach(tqg, &rxq->ifr_task);
5122 
5123 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5124 			free(fl->ifl_rx_bitmap, M_IFLIB);
5125 	}
5126 	tqg = qgroup_if_config_tqg;
5127 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5128 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5129 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5130 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5131 	CTX_LOCK(ctx);
5132 	IFDI_DETACH(ctx);
5133 	CTX_UNLOCK(ctx);
5134 
5135 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5136 	iflib_free_intr_mem(ctx);
5137 
5138 	bus_generic_detach(dev);
5139 
5140 	iflib_tx_structures_free(ctx);
5141 	iflib_rx_structures_free(ctx);
5142 
5143 	iflib_deregister(ctx);
5144 
5145 	device_set_softc(ctx->ifc_dev, NULL);
5146 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5147 		free(ctx->ifc_softc, M_IFLIB);
5148 	unref_ctx_core_offset(ctx);
5149 	free(ctx, M_IFLIB);
5150 	return (0);
5151 }
5152 
5153 static void
5154 iflib_free_intr_mem(if_ctx_t ctx)
5155 {
5156 
5157 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5158 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5159 	}
5160 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5161 		pci_release_msi(ctx->ifc_dev);
5162 	}
5163 	if (ctx->ifc_msix_mem != NULL) {
5164 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5165 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5166 		ctx->ifc_msix_mem = NULL;
5167 	}
5168 }
5169 
5170 int
5171 iflib_device_detach(device_t dev)
5172 {
5173 	if_ctx_t ctx = device_get_softc(dev);
5174 
5175 	return (iflib_device_deregister(ctx));
5176 }
5177 
5178 int
5179 iflib_device_suspend(device_t dev)
5180 {
5181 	if_ctx_t ctx = device_get_softc(dev);
5182 
5183 	CTX_LOCK(ctx);
5184 	IFDI_SUSPEND(ctx);
5185 	CTX_UNLOCK(ctx);
5186 
5187 	return bus_generic_suspend(dev);
5188 }
5189 int
5190 iflib_device_shutdown(device_t dev)
5191 {
5192 	if_ctx_t ctx = device_get_softc(dev);
5193 
5194 	CTX_LOCK(ctx);
5195 	IFDI_SHUTDOWN(ctx);
5196 	CTX_UNLOCK(ctx);
5197 
5198 	return bus_generic_suspend(dev);
5199 }
5200 
5201 
5202 int
5203 iflib_device_resume(device_t dev)
5204 {
5205 	if_ctx_t ctx = device_get_softc(dev);
5206 	iflib_txq_t txq = ctx->ifc_txqs;
5207 
5208 	CTX_LOCK(ctx);
5209 	IFDI_RESUME(ctx);
5210 	iflib_if_init_locked(ctx);
5211 	CTX_UNLOCK(ctx);
5212 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5213 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5214 
5215 	return (bus_generic_resume(dev));
5216 }
5217 
5218 int
5219 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5220 {
5221 	int error;
5222 	if_ctx_t ctx = device_get_softc(dev);
5223 
5224 	CTX_LOCK(ctx);
5225 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5226 	CTX_UNLOCK(ctx);
5227 
5228 	return (error);
5229 }
5230 
5231 void
5232 iflib_device_iov_uninit(device_t dev)
5233 {
5234 	if_ctx_t ctx = device_get_softc(dev);
5235 
5236 	CTX_LOCK(ctx);
5237 	IFDI_IOV_UNINIT(ctx);
5238 	CTX_UNLOCK(ctx);
5239 }
5240 
5241 int
5242 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, 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_VF_ADD(ctx, vfnum, params);
5249 	CTX_UNLOCK(ctx);
5250 
5251 	return (error);
5252 }
5253 
5254 /*********************************************************************
5255  *
5256  *  MODULE FUNCTION DEFINITIONS
5257  *
5258  **********************************************************************/
5259 
5260 /*
5261  * - Start a fast taskqueue thread for each core
5262  * - Start a taskqueue for control operations
5263  */
5264 static int
5265 iflib_module_init(void)
5266 {
5267 	return (0);
5268 }
5269 
5270 static int
5271 iflib_module_event_handler(module_t mod, int what, void *arg)
5272 {
5273 	int err;
5274 
5275 	switch (what) {
5276 	case MOD_LOAD:
5277 		if ((err = iflib_module_init()) != 0)
5278 			return (err);
5279 		break;
5280 	case MOD_UNLOAD:
5281 		return (EBUSY);
5282 	default:
5283 		return (EOPNOTSUPP);
5284 	}
5285 
5286 	return (0);
5287 }
5288 
5289 /*********************************************************************
5290  *
5291  *  PUBLIC FUNCTION DEFINITIONS
5292  *     ordered as in iflib.h
5293  *
5294  **********************************************************************/
5295 
5296 
5297 static void
5298 _iflib_assert(if_shared_ctx_t sctx)
5299 {
5300 	int i;
5301 
5302 	MPASS(sctx->isc_tx_maxsize);
5303 	MPASS(sctx->isc_tx_maxsegsize);
5304 
5305 	MPASS(sctx->isc_rx_maxsize);
5306 	MPASS(sctx->isc_rx_nsegments);
5307 	MPASS(sctx->isc_rx_maxsegsize);
5308 
5309 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5310 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5311 		MPASS(sctx->isc_nrxd_min[i]);
5312 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5313 		MPASS(sctx->isc_nrxd_max[i]);
5314 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5315 		MPASS(sctx->isc_nrxd_default[i]);
5316 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5317 	}
5318 
5319 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5320 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5321 		MPASS(sctx->isc_ntxd_min[i]);
5322 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5323 		MPASS(sctx->isc_ntxd_max[i]);
5324 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5325 		MPASS(sctx->isc_ntxd_default[i]);
5326 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5327 	}
5328 }
5329 
5330 static void
5331 _iflib_pre_assert(if_softc_ctx_t scctx)
5332 {
5333 
5334 	MPASS(scctx->isc_txrx->ift_txd_encap);
5335 	MPASS(scctx->isc_txrx->ift_txd_flush);
5336 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5337 	MPASS(scctx->isc_txrx->ift_rxd_available);
5338 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5339 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5340 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5341 }
5342 
5343 static int
5344 iflib_register(if_ctx_t ctx)
5345 {
5346 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5347 	driver_t *driver = sctx->isc_driver;
5348 	device_t dev = ctx->ifc_dev;
5349 	if_t ifp;
5350 
5351 	_iflib_assert(sctx);
5352 
5353 	CTX_LOCK_INIT(ctx);
5354 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5355 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5356 	if (ifp == NULL) {
5357 		device_printf(dev, "can not allocate ifnet structure\n");
5358 		return (ENOMEM);
5359 	}
5360 
5361 	/*
5362 	 * Initialize our context's device specific methods
5363 	 */
5364 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5365 	kobj_class_compile((kobj_class_t) driver);
5366 
5367 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5368 	if_setsoftc(ifp, ctx);
5369 	if_setdev(ifp, dev);
5370 	if_setinitfn(ifp, iflib_if_init);
5371 	if_setioctlfn(ifp, iflib_if_ioctl);
5372 #ifdef ALTQ
5373 	if_setstartfn(ifp, iflib_altq_if_start);
5374 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5375 	if_setsendqready(ifp);
5376 #else
5377 	if_settransmitfn(ifp, iflib_if_transmit);
5378 #endif
5379 	if_setqflushfn(ifp, iflib_if_qflush);
5380 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST |
5381 	    IFF_KNOWSEPOCH);
5382 
5383 	ctx->ifc_vlan_attach_event =
5384 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5385 							  EVENTHANDLER_PRI_FIRST);
5386 	ctx->ifc_vlan_detach_event =
5387 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5388 							  EVENTHANDLER_PRI_FIRST);
5389 
5390 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5391 		ctx->ifc_mediap = &ctx->ifc_media;
5392 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5393 		    iflib_media_change, iflib_media_status);
5394 	}
5395 	return (0);
5396 }
5397 
5398 static void
5399 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5400 {
5401 	/* Unregister VLAN events */
5402 	if (ctx->ifc_vlan_attach_event != NULL) {
5403 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5404 		ctx->ifc_vlan_attach_event = NULL;
5405 	}
5406 	if (ctx->ifc_vlan_detach_event != NULL) {
5407 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5408 		ctx->ifc_vlan_detach_event = NULL;
5409 	}
5410 
5411 }
5412 
5413 static void
5414 iflib_deregister(if_ctx_t ctx)
5415 {
5416 	if_t ifp = ctx->ifc_ifp;
5417 
5418 	/* Remove all media */
5419 	ifmedia_removeall(&ctx->ifc_media);
5420 
5421 	/* Ensure that VLAN event handlers are unregistered */
5422 	iflib_unregister_vlan_handlers(ctx);
5423 
5424 	/* Release kobject reference */
5425 	kobj_delete((kobj_t) ctx, NULL);
5426 
5427 	/* Free the ifnet structure */
5428 	if_free(ifp);
5429 
5430 	STATE_LOCK_DESTROY(ctx);
5431 
5432 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5433 	CTX_LOCK_DESTROY(ctx);
5434 }
5435 
5436 static int
5437 iflib_queues_alloc(if_ctx_t ctx)
5438 {
5439 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5440 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5441 	device_t dev = ctx->ifc_dev;
5442 	int nrxqsets = scctx->isc_nrxqsets;
5443 	int ntxqsets = scctx->isc_ntxqsets;
5444 	iflib_txq_t txq;
5445 	iflib_rxq_t rxq;
5446 	iflib_fl_t fl = NULL;
5447 	int i, j, cpu, err, txconf, rxconf;
5448 	iflib_dma_info_t ifdip;
5449 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5450 	uint32_t *txqsizes = scctx->isc_txqsizes;
5451 	uint8_t nrxqs = sctx->isc_nrxqs;
5452 	uint8_t ntxqs = sctx->isc_ntxqs;
5453 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5454 	caddr_t *vaddrs;
5455 	uint64_t *paddrs;
5456 
5457 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5458 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5459 
5460 	/* Allocate the TX ring struct memory */
5461 	if (!(ctx->ifc_txqs =
5462 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5463 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5464 		device_printf(dev, "Unable to allocate TX ring memory\n");
5465 		err = ENOMEM;
5466 		goto fail;
5467 	}
5468 
5469 	/* Now allocate the RX */
5470 	if (!(ctx->ifc_rxqs =
5471 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5472 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5473 		device_printf(dev, "Unable to allocate RX ring memory\n");
5474 		err = ENOMEM;
5475 		goto rx_fail;
5476 	}
5477 
5478 	txq = ctx->ifc_txqs;
5479 	rxq = ctx->ifc_rxqs;
5480 
5481 	/*
5482 	 * XXX handle allocation failure
5483 	 */
5484 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5485 		/* Set up some basics */
5486 
5487 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5488 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5489 			device_printf(dev,
5490 			    "Unable to allocate TX DMA info memory\n");
5491 			err = ENOMEM;
5492 			goto err_tx_desc;
5493 		}
5494 		txq->ift_ifdi = ifdip;
5495 		for (j = 0; j < ntxqs; j++, ifdip++) {
5496 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5497 				device_printf(dev,
5498 				    "Unable to allocate TX descriptors\n");
5499 				err = ENOMEM;
5500 				goto err_tx_desc;
5501 			}
5502 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5503 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5504 		}
5505 		txq->ift_ctx = ctx;
5506 		txq->ift_id = i;
5507 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5508 			txq->ift_br_offset = 1;
5509 		} else {
5510 			txq->ift_br_offset = 0;
5511 		}
5512 		/* XXX fix this */
5513 		txq->ift_timer.c_cpu = cpu;
5514 
5515 		if (iflib_txsd_alloc(txq)) {
5516 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5517 			err = ENOMEM;
5518 			goto err_tx_desc;
5519 		}
5520 
5521 		/* Initialize the TX lock */
5522 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5523 		    device_get_nameunit(dev), txq->ift_id);
5524 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5525 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5526 
5527 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5528 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5529 		if (err) {
5530 			/* XXX free any allocated rings */
5531 			device_printf(dev, "Unable to allocate buf_ring\n");
5532 			goto err_tx_desc;
5533 		}
5534 	}
5535 
5536 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5537 		/* Set up some basics */
5538 		callout_init(&rxq->ifr_watchdog, 1);
5539 
5540 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5541 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5542 			device_printf(dev,
5543 			    "Unable to allocate RX DMA info memory\n");
5544 			err = ENOMEM;
5545 			goto err_tx_desc;
5546 		}
5547 
5548 		rxq->ifr_ifdi = ifdip;
5549 		/* XXX this needs to be changed if #rx queues != #tx queues */
5550 		rxq->ifr_ntxqirq = 1;
5551 		rxq->ifr_txqid[0] = i;
5552 		for (j = 0; j < nrxqs; j++, ifdip++) {
5553 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5554 				device_printf(dev,
5555 				    "Unable to allocate RX descriptors\n");
5556 				err = ENOMEM;
5557 				goto err_tx_desc;
5558 			}
5559 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5560 		}
5561 		rxq->ifr_ctx = ctx;
5562 		rxq->ifr_id = i;
5563 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5564 			rxq->ifr_fl_offset = 1;
5565 		} else {
5566 			rxq->ifr_fl_offset = 0;
5567 		}
5568 		rxq->ifr_nfl = nfree_lists;
5569 		if (!(fl =
5570 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5571 			device_printf(dev, "Unable to allocate free list memory\n");
5572 			err = ENOMEM;
5573 			goto err_tx_desc;
5574 		}
5575 		rxq->ifr_fl = fl;
5576 		for (j = 0; j < nfree_lists; j++) {
5577 			fl[j].ifl_rxq = rxq;
5578 			fl[j].ifl_id = j;
5579 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5580 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5581 		}
5582 		/* Allocate receive buffers for the ring */
5583 		if (iflib_rxsd_alloc(rxq)) {
5584 			device_printf(dev,
5585 			    "Critical Failure setting up receive buffers\n");
5586 			err = ENOMEM;
5587 			goto err_rx_desc;
5588 		}
5589 
5590 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5591 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5592 			    M_WAITOK);
5593 	}
5594 
5595 	/* TXQs */
5596 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5597 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5598 	for (i = 0; i < ntxqsets; i++) {
5599 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5600 
5601 		for (j = 0; j < ntxqs; j++, di++) {
5602 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5603 			paddrs[i*ntxqs + j] = di->idi_paddr;
5604 		}
5605 	}
5606 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5607 		device_printf(ctx->ifc_dev,
5608 		    "Unable to allocate device TX queue\n");
5609 		iflib_tx_structures_free(ctx);
5610 		free(vaddrs, M_IFLIB);
5611 		free(paddrs, M_IFLIB);
5612 		goto err_rx_desc;
5613 	}
5614 	free(vaddrs, M_IFLIB);
5615 	free(paddrs, M_IFLIB);
5616 
5617 	/* RXQs */
5618 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5619 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5620 	for (i = 0; i < nrxqsets; i++) {
5621 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5622 
5623 		for (j = 0; j < nrxqs; j++, di++) {
5624 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5625 			paddrs[i*nrxqs + j] = di->idi_paddr;
5626 		}
5627 	}
5628 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5629 		device_printf(ctx->ifc_dev,
5630 		    "Unable to allocate device RX 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 	return (0);
5640 
5641 /* XXX handle allocation failure changes */
5642 err_rx_desc:
5643 err_tx_desc:
5644 rx_fail:
5645 	if (ctx->ifc_rxqs != NULL)
5646 		free(ctx->ifc_rxqs, M_IFLIB);
5647 	ctx->ifc_rxqs = NULL;
5648 	if (ctx->ifc_txqs != NULL)
5649 		free(ctx->ifc_txqs, M_IFLIB);
5650 	ctx->ifc_txqs = NULL;
5651 fail:
5652 	return (err);
5653 }
5654 
5655 static int
5656 iflib_tx_structures_setup(if_ctx_t ctx)
5657 {
5658 	iflib_txq_t txq = ctx->ifc_txqs;
5659 	int i;
5660 
5661 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5662 		iflib_txq_setup(txq);
5663 
5664 	return (0);
5665 }
5666 
5667 static void
5668 iflib_tx_structures_free(if_ctx_t ctx)
5669 {
5670 	iflib_txq_t txq = ctx->ifc_txqs;
5671 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5672 	int i, j;
5673 
5674 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5675 		for (j = 0; j < sctx->isc_ntxqs; j++)
5676 			iflib_dma_free(&txq->ift_ifdi[j]);
5677 		iflib_txq_destroy(txq);
5678 	}
5679 	free(ctx->ifc_txqs, M_IFLIB);
5680 	ctx->ifc_txqs = NULL;
5681 	IFDI_QUEUES_FREE(ctx);
5682 }
5683 
5684 /*********************************************************************
5685  *
5686  *  Initialize all receive rings.
5687  *
5688  **********************************************************************/
5689 static int
5690 iflib_rx_structures_setup(if_ctx_t ctx)
5691 {
5692 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5693 	int q;
5694 #if defined(INET6) || defined(INET)
5695 	int err, i;
5696 #endif
5697 
5698 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5699 #if defined(INET6) || defined(INET)
5700 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5701 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5702 			    TCP_LRO_ENTRIES, min(1024,
5703 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5704 			if (err != 0) {
5705 				device_printf(ctx->ifc_dev,
5706 				    "LRO Initialization failed!\n");
5707 				goto fail;
5708 			}
5709 		}
5710 #endif
5711 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5712 	}
5713 	return (0);
5714 #if defined(INET6) || defined(INET)
5715 fail:
5716 	/*
5717 	 * Free LRO resources allocated so far, we will only handle
5718 	 * the rings that completed, the failing case will have
5719 	 * cleaned up for itself.  'q' failed, so its the terminus.
5720 	 */
5721 	rxq = ctx->ifc_rxqs;
5722 	for (i = 0; i < q; ++i, rxq++) {
5723 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5724 			tcp_lro_free(&rxq->ifr_lc);
5725 	}
5726 	return (err);
5727 #endif
5728 }
5729 
5730 /*********************************************************************
5731  *
5732  *  Free all receive rings.
5733  *
5734  **********************************************************************/
5735 static void
5736 iflib_rx_structures_free(if_ctx_t ctx)
5737 {
5738 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5739 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5740 	int i, j;
5741 
5742 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5743 		for (j = 0; j < sctx->isc_nrxqs; j++)
5744 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5745 		iflib_rx_sds_free(rxq);
5746 #if defined(INET6) || defined(INET)
5747 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5748 			tcp_lro_free(&rxq->ifr_lc);
5749 #endif
5750 	}
5751 	free(ctx->ifc_rxqs, M_IFLIB);
5752 	ctx->ifc_rxqs = NULL;
5753 }
5754 
5755 static int
5756 iflib_qset_structures_setup(if_ctx_t ctx)
5757 {
5758 	int err;
5759 
5760 	/*
5761 	 * It is expected that the caller takes care of freeing queues if this
5762 	 * fails.
5763 	 */
5764 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5765 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5766 		return (err);
5767 	}
5768 
5769 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5770 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5771 
5772 	return (err);
5773 }
5774 
5775 int
5776 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5777 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5778 {
5779 
5780 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5781 }
5782 
5783 #ifdef SMP
5784 static int
5785 find_nth(if_ctx_t ctx, int qid)
5786 {
5787 	cpuset_t cpus;
5788 	int i, cpuid, eqid, count;
5789 
5790 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5791 	count = CPU_COUNT(&cpus);
5792 	eqid = qid % count;
5793 	/* clear up to the qid'th bit */
5794 	for (i = 0; i < eqid; i++) {
5795 		cpuid = CPU_FFS(&cpus);
5796 		MPASS(cpuid != 0);
5797 		CPU_CLR(cpuid-1, &cpus);
5798 	}
5799 	cpuid = CPU_FFS(&cpus);
5800 	MPASS(cpuid != 0);
5801 	return (cpuid-1);
5802 }
5803 
5804 #ifdef SCHED_ULE
5805 extern struct cpu_group *cpu_top;              /* CPU topology */
5806 
5807 static int
5808 find_child_with_core(int cpu, struct cpu_group *grp)
5809 {
5810 	int i;
5811 
5812 	if (grp->cg_children == 0)
5813 		return -1;
5814 
5815 	MPASS(grp->cg_child);
5816 	for (i = 0; i < grp->cg_children; i++) {
5817 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5818 			return i;
5819 	}
5820 
5821 	return -1;
5822 }
5823 
5824 /*
5825  * Find the nth "close" core to the specified core
5826  * "close" is defined as the deepest level that shares
5827  * at least an L2 cache.  With threads, this will be
5828  * threads on the same core.  If the shared cache is L3
5829  * or higher, simply returns the same core.
5830  */
5831 static int
5832 find_close_core(int cpu, int core_offset)
5833 {
5834 	struct cpu_group *grp;
5835 	int i;
5836 	int fcpu;
5837 	cpuset_t cs;
5838 
5839 	grp = cpu_top;
5840 	if (grp == NULL)
5841 		return cpu;
5842 	i = 0;
5843 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5844 		/* If the child only has one cpu, don't descend */
5845 		if (grp->cg_child[i].cg_count <= 1)
5846 			break;
5847 		grp = &grp->cg_child[i];
5848 	}
5849 
5850 	/* If they don't share at least an L2 cache, use the same CPU */
5851 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5852 		return cpu;
5853 
5854 	/* Now pick one */
5855 	CPU_COPY(&grp->cg_mask, &cs);
5856 
5857 	/* Add the selected CPU offset to core offset. */
5858 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5859 		if (fcpu - 1 == cpu)
5860 			break;
5861 		CPU_CLR(fcpu - 1, &cs);
5862 	}
5863 	MPASS(fcpu);
5864 
5865 	core_offset += i;
5866 
5867 	CPU_COPY(&grp->cg_mask, &cs);
5868 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5869 		MPASS(CPU_FFS(&cs));
5870 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5871 	}
5872 	MPASS(CPU_FFS(&cs));
5873 	return CPU_FFS(&cs) - 1;
5874 }
5875 #else
5876 static int
5877 find_close_core(int cpu, int core_offset __unused)
5878 {
5879 	return cpu;
5880 }
5881 #endif
5882 
5883 static int
5884 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5885 {
5886 	switch (type) {
5887 	case IFLIB_INTR_TX:
5888 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5889 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5890 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5891 	case IFLIB_INTR_RX:
5892 	case IFLIB_INTR_RXTX:
5893 		/* RX queues get the specified core */
5894 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5895 	default:
5896 		return -1;
5897 	}
5898 }
5899 #else
5900 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5901 #define find_close_core(cpuid, tid)	CPU_FIRST()
5902 #define find_nth(ctx, gid)		CPU_FIRST()
5903 #endif
5904 
5905 /* Just to avoid copy/paste */
5906 static inline int
5907 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5908     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5909     const char *name)
5910 {
5911 	device_t dev;
5912 	int co, cpuid, err, tid;
5913 
5914 	dev = ctx->ifc_dev;
5915 	co = ctx->ifc_sysctl_core_offset;
5916 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5917 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5918 	cpuid = find_nth(ctx, qid + co);
5919 	tid = get_core_offset(ctx, type, qid);
5920 	if (tid < 0) {
5921 		device_printf(dev, "get_core_offset failed\n");
5922 		return (EOPNOTSUPP);
5923 	}
5924 	cpuid = find_close_core(cpuid, tid);
5925 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5926 	    name);
5927 	if (err) {
5928 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5929 		return (err);
5930 	}
5931 #ifdef notyet
5932 	if (cpuid > ctx->ifc_cpuid_highest)
5933 		ctx->ifc_cpuid_highest = cpuid;
5934 #endif
5935 	return (0);
5936 }
5937 
5938 int
5939 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5940 			iflib_intr_type_t type, driver_filter_t *filter,
5941 			void *filter_arg, int qid, const char *name)
5942 {
5943 	device_t dev;
5944 	struct grouptask *gtask;
5945 	struct taskqgroup *tqg;
5946 	iflib_filter_info_t info;
5947 	gtask_fn_t *fn;
5948 	int tqrid, err;
5949 	driver_filter_t *intr_fast;
5950 	void *q;
5951 
5952 	info = &ctx->ifc_filter_info;
5953 	tqrid = rid;
5954 
5955 	switch (type) {
5956 	/* XXX merge tx/rx for netmap? */
5957 	case IFLIB_INTR_TX:
5958 		q = &ctx->ifc_txqs[qid];
5959 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5960 		gtask = &ctx->ifc_txqs[qid].ift_task;
5961 		tqg = qgroup_if_io_tqg;
5962 		fn = _task_fn_tx;
5963 		intr_fast = iflib_fast_intr;
5964 		GROUPTASK_INIT(gtask, 0, fn, q);
5965 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5966 		break;
5967 	case IFLIB_INTR_RX:
5968 		q = &ctx->ifc_rxqs[qid];
5969 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5970 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5971 		tqg = qgroup_if_io_tqg;
5972 		fn = _task_fn_rx;
5973 		intr_fast = iflib_fast_intr;
5974 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
5975 		break;
5976 	case IFLIB_INTR_RXTX:
5977 		q = &ctx->ifc_rxqs[qid];
5978 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5979 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5980 		tqg = qgroup_if_io_tqg;
5981 		fn = _task_fn_rx;
5982 		intr_fast = iflib_fast_intr_rxtx;
5983 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
5984 		break;
5985 	case IFLIB_INTR_ADMIN:
5986 		q = ctx;
5987 		tqrid = -1;
5988 		info = &ctx->ifc_filter_info;
5989 		gtask = &ctx->ifc_admin_task;
5990 		tqg = qgroup_if_config_tqg;
5991 		fn = _task_fn_admin;
5992 		intr_fast = iflib_fast_intr_ctx;
5993 		break;
5994 	default:
5995 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5996 		    __func__);
5997 		return (EINVAL);
5998 	}
5999 
6000 	info->ifi_filter = filter;
6001 	info->ifi_filter_arg = filter_arg;
6002 	info->ifi_task = gtask;
6003 	info->ifi_ctx = q;
6004 
6005 	dev = ctx->ifc_dev;
6006 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6007 	if (err != 0) {
6008 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6009 		return (err);
6010 	}
6011 	if (type == IFLIB_INTR_ADMIN)
6012 		return (0);
6013 
6014 	if (tqrid != -1) {
6015 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6016 		    q, name);
6017 		if (err)
6018 			return (err);
6019 	} else {
6020 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6021 	}
6022 
6023 	return (0);
6024 }
6025 
6026 void
6027 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6028 {
6029 	struct grouptask *gtask;
6030 	struct taskqgroup *tqg;
6031 	gtask_fn_t *fn;
6032 	void *q;
6033 	int err;
6034 
6035 	switch (type) {
6036 	case IFLIB_INTR_TX:
6037 		q = &ctx->ifc_txqs[qid];
6038 		gtask = &ctx->ifc_txqs[qid].ift_task;
6039 		tqg = qgroup_if_io_tqg;
6040 		fn = _task_fn_tx;
6041 		GROUPTASK_INIT(gtask, 0, fn, q);
6042 		break;
6043 	case IFLIB_INTR_RX:
6044 		q = &ctx->ifc_rxqs[qid];
6045 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6046 		tqg = qgroup_if_io_tqg;
6047 		fn = _task_fn_rx;
6048 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6049 		break;
6050 	case IFLIB_INTR_IOV:
6051 		q = ctx;
6052 		gtask = &ctx->ifc_vflr_task;
6053 		tqg = qgroup_if_config_tqg;
6054 		fn = _task_fn_iov;
6055 		GROUPTASK_INIT(gtask, 0, fn, q);
6056 		break;
6057 	default:
6058 		panic("unknown net intr type");
6059 	}
6060 	if (irq != NULL) {
6061 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6062 		    q, name);
6063 		if (err)
6064 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6065 			    irq->ii_res, name);
6066 	} else {
6067 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6068 	}
6069 }
6070 
6071 void
6072 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6073 {
6074 
6075 	if (irq->ii_tag)
6076 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6077 
6078 	if (irq->ii_res)
6079 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6080 		    rman_get_rid(irq->ii_res), irq->ii_res);
6081 }
6082 
6083 static int
6084 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6085 {
6086 	iflib_txq_t txq = ctx->ifc_txqs;
6087 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6088 	if_irq_t irq = &ctx->ifc_legacy_irq;
6089 	iflib_filter_info_t info;
6090 	device_t dev;
6091 	struct grouptask *gtask;
6092 	struct resource *res;
6093 	struct taskqgroup *tqg;
6094 	void *q;
6095 	int err, tqrid;
6096 	bool rx_only;
6097 
6098 	q = &ctx->ifc_rxqs[0];
6099 	info = &rxq[0].ifr_filter_info;
6100 	gtask = &rxq[0].ifr_task;
6101 	tqg = qgroup_if_io_tqg;
6102 	tqrid = *rid;
6103 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6104 
6105 	ctx->ifc_flags |= IFC_LEGACY;
6106 	info->ifi_filter = filter;
6107 	info->ifi_filter_arg = filter_arg;
6108 	info->ifi_task = gtask;
6109 	info->ifi_ctx = rx_only ? ctx : q;
6110 
6111 	dev = ctx->ifc_dev;
6112 	/* We allocate a single interrupt resource */
6113 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6114 	    iflib_fast_intr_rxtx, NULL, info, name);
6115 	if (err != 0)
6116 		return (err);
6117 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6118 	res = irq->ii_res;
6119 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6120 
6121 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6122 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6123 	    "tx");
6124 	return (0);
6125 }
6126 
6127 void
6128 iflib_led_create(if_ctx_t ctx)
6129 {
6130 
6131 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6132 	    device_get_nameunit(ctx->ifc_dev));
6133 }
6134 
6135 void
6136 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6137 {
6138 
6139 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6140 }
6141 
6142 void
6143 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6144 {
6145 
6146 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6147 }
6148 
6149 void
6150 iflib_admin_intr_deferred(if_ctx_t ctx)
6151 {
6152 
6153 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6154 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6155 }
6156 
6157 void
6158 iflib_iov_intr_deferred(if_ctx_t ctx)
6159 {
6160 
6161 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6162 }
6163 
6164 void
6165 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6166 {
6167 
6168 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6169 	    name);
6170 }
6171 
6172 void
6173 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6174 	const char *name)
6175 {
6176 
6177 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6178 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6179 	    name);
6180 }
6181 
6182 void
6183 iflib_config_gtask_deinit(struct grouptask *gtask)
6184 {
6185 
6186 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6187 }
6188 
6189 void
6190 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6191 {
6192 	if_t ifp = ctx->ifc_ifp;
6193 	iflib_txq_t txq = ctx->ifc_txqs;
6194 
6195 	if_setbaudrate(ifp, baudrate);
6196 	if (baudrate >= IF_Gbps(10)) {
6197 		STATE_LOCK(ctx);
6198 		ctx->ifc_flags |= IFC_PREFETCH;
6199 		STATE_UNLOCK(ctx);
6200 	}
6201 	/* If link down, disable watchdog */
6202 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6203 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6204 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6205 	}
6206 	ctx->ifc_link_state = link_state;
6207 	if_link_state_change(ifp, link_state);
6208 }
6209 
6210 static int
6211 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6212 {
6213 	int credits;
6214 #ifdef INVARIANTS
6215 	int credits_pre = txq->ift_cidx_processed;
6216 #endif
6217 
6218 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6219 	    BUS_DMASYNC_POSTREAD);
6220 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6221 		return (0);
6222 
6223 	txq->ift_processed += credits;
6224 	txq->ift_cidx_processed += credits;
6225 
6226 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6227 	if (txq->ift_cidx_processed >= txq->ift_size)
6228 		txq->ift_cidx_processed -= txq->ift_size;
6229 	return (credits);
6230 }
6231 
6232 static int
6233 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6234 {
6235 	iflib_fl_t fl;
6236 	u_int i;
6237 
6238 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6239 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6240 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6241 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6242 	    budget));
6243 }
6244 
6245 void
6246 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6247 	const char *description, if_int_delay_info_t info,
6248 	int offset, int value)
6249 {
6250 	info->iidi_ctx = ctx;
6251 	info->iidi_offset = offset;
6252 	info->iidi_value = value;
6253 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6254 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6255 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6256 	    info, 0, iflib_sysctl_int_delay, "I", description);
6257 }
6258 
6259 struct sx *
6260 iflib_ctx_lock_get(if_ctx_t ctx)
6261 {
6262 
6263 	return (&ctx->ifc_ctx_sx);
6264 }
6265 
6266 static int
6267 iflib_msix_init(if_ctx_t ctx)
6268 {
6269 	device_t dev = ctx->ifc_dev;
6270 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6271 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6272 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6273 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6274 
6275 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6276 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6277 
6278 	if (bootverbose)
6279 		device_printf(dev, "msix_init qsets capped at %d\n",
6280 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6281 
6282 	/* Override by tuneable */
6283 	if (scctx->isc_disable_msix)
6284 		goto msi;
6285 
6286 	/* First try MSI-X */
6287 	if ((msgs = pci_msix_count(dev)) == 0) {
6288 		if (bootverbose)
6289 			device_printf(dev, "MSI-X not supported or disabled\n");
6290 		goto msi;
6291 	}
6292 
6293 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6294 	/*
6295 	 * bar == -1 => "trust me I know what I'm doing"
6296 	 * Some drivers are for hardware that is so shoddily
6297 	 * documented that no one knows which bars are which
6298 	 * so the developer has to map all bars. This hack
6299 	 * allows shoddy garbage to use MSI-X in this framework.
6300 	 */
6301 	if (bar != -1) {
6302 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6303 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6304 		if (ctx->ifc_msix_mem == NULL) {
6305 			device_printf(dev, "Unable to map MSI-X table\n");
6306 			goto msi;
6307 		}
6308 	}
6309 
6310 	admincnt = sctx->isc_admin_intrcnt;
6311 #if IFLIB_DEBUG
6312 	/* use only 1 qset in debug mode */
6313 	queuemsgs = min(msgs - admincnt, 1);
6314 #else
6315 	queuemsgs = msgs - admincnt;
6316 #endif
6317 #ifdef RSS
6318 	queues = imin(queuemsgs, rss_getnumbuckets());
6319 #else
6320 	queues = queuemsgs;
6321 #endif
6322 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6323 	if (bootverbose)
6324 		device_printf(dev,
6325 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6326 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6327 #ifdef  RSS
6328 	/* If we're doing RSS, clamp at the number of RSS buckets */
6329 	if (queues > rss_getnumbuckets())
6330 		queues = rss_getnumbuckets();
6331 #endif
6332 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6333 		rx_queues = iflib_num_rx_queues;
6334 	else
6335 		rx_queues = queues;
6336 
6337 	if (rx_queues > scctx->isc_nrxqsets)
6338 		rx_queues = scctx->isc_nrxqsets;
6339 
6340 	/*
6341 	 * We want this to be all logical CPUs by default
6342 	 */
6343 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6344 		tx_queues = iflib_num_tx_queues;
6345 	else
6346 		tx_queues = mp_ncpus;
6347 
6348 	if (tx_queues > scctx->isc_ntxqsets)
6349 		tx_queues = scctx->isc_ntxqsets;
6350 
6351 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6352 #ifdef INVARIANTS
6353 		if (tx_queues != rx_queues)
6354 			device_printf(dev,
6355 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6356 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6357 #endif
6358 		tx_queues = min(rx_queues, tx_queues);
6359 		rx_queues = min(rx_queues, tx_queues);
6360 	}
6361 
6362 	vectors = rx_queues + admincnt;
6363 	if (msgs < vectors) {
6364 		device_printf(dev,
6365 		    "insufficient number of MSI-X vectors "
6366 		    "(supported %d, need %d)\n", msgs, vectors);
6367 		goto msi;
6368 	}
6369 
6370 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6371 	    tx_queues);
6372 	msgs = vectors;
6373 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6374 		if (vectors != msgs) {
6375 			device_printf(dev,
6376 			    "Unable to allocate sufficient MSI-X vectors "
6377 			    "(got %d, need %d)\n", vectors, msgs);
6378 			pci_release_msi(dev);
6379 			if (bar != -1) {
6380 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6381 				    ctx->ifc_msix_mem);
6382 				ctx->ifc_msix_mem = NULL;
6383 			}
6384 			goto msi;
6385 		}
6386 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6387 		    vectors);
6388 		scctx->isc_vectors = vectors;
6389 		scctx->isc_nrxqsets = rx_queues;
6390 		scctx->isc_ntxqsets = tx_queues;
6391 		scctx->isc_intr = IFLIB_INTR_MSIX;
6392 
6393 		return (vectors);
6394 	} else {
6395 		device_printf(dev,
6396 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6397 		    err);
6398 		if (bar != -1) {
6399 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6400 			    ctx->ifc_msix_mem);
6401 			ctx->ifc_msix_mem = NULL;
6402 		}
6403 	}
6404 
6405 msi:
6406 	vectors = pci_msi_count(dev);
6407 	scctx->isc_nrxqsets = 1;
6408 	scctx->isc_ntxqsets = 1;
6409 	scctx->isc_vectors = vectors;
6410 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6411 		device_printf(dev,"Using an MSI interrupt\n");
6412 		scctx->isc_intr = IFLIB_INTR_MSI;
6413 	} else {
6414 		scctx->isc_vectors = 1;
6415 		device_printf(dev,"Using a Legacy interrupt\n");
6416 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6417 	}
6418 
6419 	return (vectors);
6420 }
6421 
6422 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6423 
6424 static int
6425 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6426 {
6427 	int rc;
6428 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6429 	struct sbuf *sb;
6430 	const char *ring_state = "UNKNOWN";
6431 
6432 	/* XXX needed ? */
6433 	rc = sysctl_wire_old_buffer(req, 0);
6434 	MPASS(rc == 0);
6435 	if (rc != 0)
6436 		return (rc);
6437 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6438 	MPASS(sb != NULL);
6439 	if (sb == NULL)
6440 		return (ENOMEM);
6441 	if (state[3] <= 3)
6442 		ring_state = ring_states[state[3]];
6443 
6444 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6445 		    state[0], state[1], state[2], ring_state);
6446 	rc = sbuf_finish(sb);
6447 	sbuf_delete(sb);
6448         return(rc);
6449 }
6450 
6451 enum iflib_ndesc_handler {
6452 	IFLIB_NTXD_HANDLER,
6453 	IFLIB_NRXD_HANDLER,
6454 };
6455 
6456 static int
6457 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6458 {
6459 	if_ctx_t ctx = (void *)arg1;
6460 	enum iflib_ndesc_handler type = arg2;
6461 	char buf[256] = {0};
6462 	qidx_t *ndesc;
6463 	char *p, *next;
6464 	int nqs, rc, i;
6465 
6466 	nqs = 8;
6467 	switch(type) {
6468 	case IFLIB_NTXD_HANDLER:
6469 		ndesc = ctx->ifc_sysctl_ntxds;
6470 		if (ctx->ifc_sctx)
6471 			nqs = ctx->ifc_sctx->isc_ntxqs;
6472 		break;
6473 	case IFLIB_NRXD_HANDLER:
6474 		ndesc = ctx->ifc_sysctl_nrxds;
6475 		if (ctx->ifc_sctx)
6476 			nqs = ctx->ifc_sctx->isc_nrxqs;
6477 		break;
6478 	default:
6479 		printf("%s: unhandled type\n", __func__);
6480 		return (EINVAL);
6481 	}
6482 	if (nqs == 0)
6483 		nqs = 8;
6484 
6485 	for (i=0; i<8; i++) {
6486 		if (i >= nqs)
6487 			break;
6488 		if (i)
6489 			strcat(buf, ",");
6490 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6491 	}
6492 
6493 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6494 	if (rc || req->newptr == NULL)
6495 		return rc;
6496 
6497 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6498 	    i++, p = strsep(&next, " ,")) {
6499 		ndesc[i] = strtoul(p, NULL, 10);
6500 	}
6501 
6502 	return(rc);
6503 }
6504 
6505 #define NAME_BUFLEN 32
6506 static void
6507 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6508 {
6509         device_t dev = iflib_get_dev(ctx);
6510 	struct sysctl_oid_list *child, *oid_list;
6511 	struct sysctl_ctx_list *ctx_list;
6512 	struct sysctl_oid *node;
6513 
6514 	ctx_list = device_get_sysctl_ctx(dev);
6515 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6516 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6517 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6518 	oid_list = SYSCTL_CHILDREN(node);
6519 
6520 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6521 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6522 		       "driver version");
6523 
6524 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6525 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6526 			"# of txqs to use, 0 => use default #");
6527 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6528 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6529 			"# of rxqs to use, 0 => use default #");
6530 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6531 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6532                        "permit #txq != #rxq");
6533 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6534                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6535                       "disable MSI-X (default 0)");
6536 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6537 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6538 		       "set the RX budget");
6539 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6540 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6541 		       "cause TX to abdicate instead of running to completion");
6542 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6543 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6544 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6545 		       "offset to start using cores at");
6546 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6547 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6548 		       "use separate cores for TX and RX");
6549 
6550 	/* XXX change for per-queue sizes */
6551 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6552 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6553 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6554 	    "list of # of TX descriptors to use, 0 = use default #");
6555 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6556 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6557 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6558 	    "list of # of RX descriptors to use, 0 = use default #");
6559 }
6560 
6561 static void
6562 iflib_add_device_sysctl_post(if_ctx_t ctx)
6563 {
6564 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6565 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6566         device_t dev = iflib_get_dev(ctx);
6567 	struct sysctl_oid_list *child;
6568 	struct sysctl_ctx_list *ctx_list;
6569 	iflib_fl_t fl;
6570 	iflib_txq_t txq;
6571 	iflib_rxq_t rxq;
6572 	int i, j;
6573 	char namebuf[NAME_BUFLEN];
6574 	char *qfmt;
6575 	struct sysctl_oid *queue_node, *fl_node, *node;
6576 	struct sysctl_oid_list *queue_list, *fl_list;
6577 	ctx_list = device_get_sysctl_ctx(dev);
6578 
6579 	node = ctx->ifc_sysctl_node;
6580 	child = SYSCTL_CHILDREN(node);
6581 
6582 	if (scctx->isc_ntxqsets > 100)
6583 		qfmt = "txq%03d";
6584 	else if (scctx->isc_ntxqsets > 10)
6585 		qfmt = "txq%02d";
6586 	else
6587 		qfmt = "txq%d";
6588 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6589 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6590 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6591 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6592 		queue_list = SYSCTL_CHILDREN(queue_node);
6593 #if MEMORY_LOGGING
6594 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6595 				CTLFLAG_RD,
6596 				&txq->ift_dequeued, "total mbufs freed");
6597 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6598 				CTLFLAG_RD,
6599 				&txq->ift_enqueued, "total mbufs enqueued");
6600 #endif
6601 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6602 				   CTLFLAG_RD,
6603 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6604 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6605 				   CTLFLAG_RD,
6606 				   &txq->ift_pullups, "# of times m_pullup was called");
6607 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6608 				   CTLFLAG_RD,
6609 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6610 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6611 				   CTLFLAG_RD,
6612 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6613 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6614 				   CTLFLAG_RD,
6615 				   &txq->ift_map_failed, "# of times DMA map failed");
6616 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6617 				   CTLFLAG_RD,
6618 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6619 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6620 				   CTLFLAG_RD,
6621 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6622 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6623 				   CTLFLAG_RD,
6624 				   &txq->ift_pidx, 1, "Producer Index");
6625 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6626 				   CTLFLAG_RD,
6627 				   &txq->ift_cidx, 1, "Consumer Index");
6628 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6629 				   CTLFLAG_RD,
6630 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6631 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6632 				   CTLFLAG_RD,
6633 				   &txq->ift_in_use, 1, "descriptors in use");
6634 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6635 				   CTLFLAG_RD,
6636 				   &txq->ift_processed, "descriptors procesed for clean");
6637 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6638 				   CTLFLAG_RD,
6639 				   &txq->ift_cleaned, "total cleaned");
6640 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6641 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6642 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6643 		    mp_ring_state_handler, "A", "soft ring state");
6644 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6645 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6646 				       "# of enqueues to the mp_ring for this queue");
6647 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6648 				       CTLFLAG_RD, &txq->ift_br->drops,
6649 				       "# of drops in the mp_ring for this queue");
6650 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6651 				       CTLFLAG_RD, &txq->ift_br->starts,
6652 				       "# of normal consumer starts in the mp_ring for this queue");
6653 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6654 				       CTLFLAG_RD, &txq->ift_br->stalls,
6655 					       "# of consumer stalls in the mp_ring for this queue");
6656 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6657 			       CTLFLAG_RD, &txq->ift_br->restarts,
6658 				       "# of consumer restarts in the mp_ring for this queue");
6659 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6660 				       CTLFLAG_RD, &txq->ift_br->abdications,
6661 				       "# of consumer abdications in the mp_ring for this queue");
6662 	}
6663 
6664 	if (scctx->isc_nrxqsets > 100)
6665 		qfmt = "rxq%03d";
6666 	else if (scctx->isc_nrxqsets > 10)
6667 		qfmt = "rxq%02d";
6668 	else
6669 		qfmt = "rxq%d";
6670 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6671 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6672 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6673 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6674 		queue_list = SYSCTL_CHILDREN(queue_node);
6675 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6676 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6677 				       CTLFLAG_RD,
6678 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6679 		}
6680 
6681 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6682 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6683 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6684 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6685 			fl_list = SYSCTL_CHILDREN(fl_node);
6686 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6687 				       CTLFLAG_RD,
6688 				       &fl->ifl_pidx, 1, "Producer Index");
6689 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6690 				       CTLFLAG_RD,
6691 				       &fl->ifl_cidx, 1, "Consumer Index");
6692 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6693 				       CTLFLAG_RD,
6694 				       &fl->ifl_credits, 1, "credits available");
6695 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6696 				       CTLFLAG_RD,
6697 				       &fl->ifl_buf_size, 1, "buffer size");
6698 #if MEMORY_LOGGING
6699 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6700 					CTLFLAG_RD,
6701 					&fl->ifl_m_enqueued, "mbufs allocated");
6702 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6703 					CTLFLAG_RD,
6704 					&fl->ifl_m_dequeued, "mbufs freed");
6705 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6706 					CTLFLAG_RD,
6707 					&fl->ifl_cl_enqueued, "clusters allocated");
6708 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6709 					CTLFLAG_RD,
6710 					&fl->ifl_cl_dequeued, "clusters freed");
6711 #endif
6712 
6713 		}
6714 	}
6715 
6716 }
6717 
6718 void
6719 iflib_request_reset(if_ctx_t ctx)
6720 {
6721 
6722 	STATE_LOCK(ctx);
6723 	ctx->ifc_flags |= IFC_DO_RESET;
6724 	STATE_UNLOCK(ctx);
6725 }
6726 
6727 #ifndef __NO_STRICT_ALIGNMENT
6728 static struct mbuf *
6729 iflib_fixup_rx(struct mbuf *m)
6730 {
6731 	struct mbuf *n;
6732 
6733 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6734 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6735 		m->m_data += ETHER_HDR_LEN;
6736 		n = m;
6737 	} else {
6738 		MGETHDR(n, M_NOWAIT, MT_DATA);
6739 		if (n == NULL) {
6740 			m_freem(m);
6741 			return (NULL);
6742 		}
6743 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6744 		m->m_data += ETHER_HDR_LEN;
6745 		m->m_len -= ETHER_HDR_LEN;
6746 		n->m_len = ETHER_HDR_LEN;
6747 		M_MOVE_PKTHDR(n, m);
6748 		n->m_next = m;
6749 	}
6750 	return (n);
6751 }
6752 #endif
6753 
6754 #ifdef DEBUGNET
6755 static void
6756 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6757 {
6758 	if_ctx_t ctx;
6759 
6760 	ctx = if_getsoftc(ifp);
6761 	CTX_LOCK(ctx);
6762 	*nrxr = NRXQSETS(ctx);
6763 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6764 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6765 	CTX_UNLOCK(ctx);
6766 }
6767 
6768 static void
6769 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6770 {
6771 	if_ctx_t ctx;
6772 	if_softc_ctx_t scctx;
6773 	iflib_fl_t fl;
6774 	iflib_rxq_t rxq;
6775 	int i, j;
6776 
6777 	ctx = if_getsoftc(ifp);
6778 	scctx = &ctx->ifc_softc_ctx;
6779 
6780 	switch (event) {
6781 	case DEBUGNET_START:
6782 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6783 			rxq = &ctx->ifc_rxqs[i];
6784 			for (j = 0; j < rxq->ifr_nfl; j++) {
6785 				fl = rxq->ifr_fl;
6786 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6787 			}
6788 		}
6789 		iflib_no_tx_batch = 1;
6790 		break;
6791 	default:
6792 		break;
6793 	}
6794 }
6795 
6796 static int
6797 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6798 {
6799 	if_ctx_t ctx;
6800 	iflib_txq_t txq;
6801 	int error;
6802 
6803 	ctx = if_getsoftc(ifp);
6804 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6805 	    IFF_DRV_RUNNING)
6806 		return (EBUSY);
6807 
6808 	txq = &ctx->ifc_txqs[0];
6809 	error = iflib_encap(txq, &m);
6810 	if (error == 0)
6811 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6812 	return (error);
6813 }
6814 
6815 static int
6816 iflib_debugnet_poll(if_t ifp, int count)
6817 {
6818 	struct epoch_tracker et;
6819 	if_ctx_t ctx;
6820 	if_softc_ctx_t scctx;
6821 	iflib_txq_t txq;
6822 	int i;
6823 
6824 	ctx = if_getsoftc(ifp);
6825 	scctx = &ctx->ifc_softc_ctx;
6826 
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 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6833 
6834 	NET_EPOCH_ENTER(et);
6835 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6836 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6837 	NET_EPOCH_EXIT(et);
6838 	return (0);
6839 }
6840 #endif /* DEBUGNET */
6841