xref: /freebsd/sys/net/iflib.c (revision 2f5cff37506aaa5d4ebf53560981899efadcc5e2)
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/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/netdump/netdump.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 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 
144 typedef struct iflib_filter_info {
145 	driver_filter_t *ifi_filter;
146 	void *ifi_filter_arg;
147 	struct grouptask *ifi_task;
148 	void *ifi_ctx;
149 } *iflib_filter_info_t;
150 
151 struct iflib_ctx {
152 	KOBJ_FIELDS;
153 	/*
154 	 * Pointer to hardware driver's softc
155 	 */
156 	void *ifc_softc;
157 	device_t ifc_dev;
158 	if_t ifc_ifp;
159 
160 	cpuset_t ifc_cpus;
161 	if_shared_ctx_t ifc_sctx;
162 	struct if_softc_ctx ifc_softc_ctx;
163 
164 	struct sx ifc_ctx_sx;
165 	struct mtx ifc_state_mtx;
166 
167 	iflib_txq_t ifc_txqs;
168 	iflib_rxq_t ifc_rxqs;
169 	uint32_t ifc_if_flags;
170 	uint32_t ifc_flags;
171 	uint32_t ifc_max_fl_buf_size;
172 	uint32_t ifc_rx_mbuf_sz;
173 
174 	int ifc_link_state;
175 	int ifc_link_irq;
176 	int ifc_watchdog_events;
177 	struct cdev *ifc_led_dev;
178 	struct resource *ifc_msix_mem;
179 
180 	struct if_irq ifc_legacy_irq;
181 	struct grouptask ifc_admin_task;
182 	struct grouptask ifc_vflr_task;
183 	struct iflib_filter_info ifc_filter_info;
184 	struct ifmedia	ifc_media;
185 
186 	struct sysctl_oid *ifc_sysctl_node;
187 	uint16_t ifc_sysctl_ntxqs;
188 	uint16_t ifc_sysctl_nrxqs;
189 	uint16_t ifc_sysctl_qs_eq_override;
190 	uint16_t ifc_sysctl_rx_budget;
191 	uint16_t ifc_sysctl_tx_abdicate;
192 	uint16_t ifc_sysctl_core_offset;
193 #define	CORE_OFFSET_UNSPECIFIED	0xffff
194 	uint8_t  ifc_sysctl_separate_txrx;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	struct ether_addr ifc_mac;
212 	char ifc_mtx_name[16];
213 };
214 
215 
216 void *
217 iflib_get_softc(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_softc);
221 }
222 
223 device_t
224 iflib_get_dev(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_dev);
228 }
229 
230 if_t
231 iflib_get_ifp(if_ctx_t ctx)
232 {
233 
234 	return (ctx->ifc_ifp);
235 }
236 
237 struct ifmedia *
238 iflib_get_media(if_ctx_t ctx)
239 {
240 
241 	return (&ctx->ifc_media);
242 }
243 
244 uint32_t
245 iflib_get_flags(if_ctx_t ctx)
246 {
247 	return (ctx->ifc_flags);
248 }
249 
250 void
251 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
252 {
253 
254 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
255 }
256 
257 if_softc_ctx_t
258 iflib_get_softc_ctx(if_ctx_t ctx)
259 {
260 
261 	return (&ctx->ifc_softc_ctx);
262 }
263 
264 if_shared_ctx_t
265 iflib_get_sctx(if_ctx_t ctx)
266 {
267 
268 	return (ctx->ifc_sctx);
269 }
270 
271 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
272 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
273 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
274 
275 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
276 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
277 
278 typedef struct iflib_sw_rx_desc_array {
279 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
280 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
281 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
282 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
283 } iflib_rxsd_array_t;
284 
285 typedef struct iflib_sw_tx_desc_array {
286 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
287 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
288 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
289 } if_txsd_vec_t;
290 
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 
311 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
312 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
313 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
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 	char                    ift_db_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 
384 	bitstr_t 	*ifl_rx_bitmap;
385 	qidx_t		ifl_fragidx;
386 	/* constant */
387 	qidx_t		ifl_size;
388 	uint16_t	ifl_buf_size;
389 	uint16_t	ifl_cltype;
390 	uma_zone_t	ifl_zone;
391 	iflib_rxsd_array_t	ifl_sds;
392 	iflib_rxq_t	ifl_rxq;
393 	uint8_t		ifl_id;
394 	bus_dma_tag_t	ifl_buf_tag;
395 	iflib_dma_info_t	ifl_ifdi;
396 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
397 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
398 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
399 }  __aligned(CACHE_LINE_SIZE);
400 
401 static inline qidx_t
402 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
403 {
404 	qidx_t used;
405 
406 	if (pidx > cidx)
407 		used = pidx - cidx;
408 	else if (pidx < cidx)
409 		used = size - cidx + pidx;
410 	else if (gen == 0 && pidx == cidx)
411 		used = 0;
412 	else if (gen == 1 && pidx == cidx)
413 		used = size;
414 	else
415 		panic("bad state");
416 
417 	return (used);
418 }
419 
420 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
421 
422 #define IDXDIFF(head, tail, wrap) \
423 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
424 
425 struct iflib_rxq {
426 	/* If there is a separate completion queue -
427 	 * these are the cq cidx and pidx. Otherwise
428 	 * these are unused.
429 	 */
430 	qidx_t		ifr_size;
431 	qidx_t		ifr_cq_cidx;
432 	qidx_t		ifr_cq_pidx;
433 	uint8_t		ifr_cq_gen;
434 	uint8_t		ifr_fl_offset;
435 
436 	if_ctx_t	ifr_ctx;
437 	iflib_fl_t	ifr_fl;
438 	uint64_t	ifr_rx_irq;
439 	struct pfil_head	*pfil;
440 	uint16_t	ifr_id;
441 	uint8_t		ifr_lro_enabled;
442 	uint8_t		ifr_nfl;
443 	uint8_t		ifr_ntxqirq;
444 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
445 	struct lro_ctrl			ifr_lc;
446 	struct grouptask        ifr_task;
447 	struct iflib_filter_info ifr_filter_info;
448 	iflib_dma_info_t		ifr_ifdi;
449 
450 	/* dynamically allocate if any drivers need a value substantially larger than this */
451 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
452 #ifdef IFLIB_DIAGNOSTICS
453 	uint64_t ifr_cpu_exec_count[256];
454 #endif
455 }  __aligned(CACHE_LINE_SIZE);
456 
457 typedef struct if_rxsd {
458 	caddr_t *ifsd_cl;
459 	iflib_fl_t ifsd_fl;
460 	qidx_t ifsd_cidx;
461 } *if_rxsd_t;
462 
463 /* multiple of word size */
464 #ifdef __LP64__
465 #define PKT_INFO_SIZE	6
466 #define RXD_INFO_SIZE	5
467 #define PKT_TYPE uint64_t
468 #else
469 #define PKT_INFO_SIZE	11
470 #define RXD_INFO_SIZE	8
471 #define PKT_TYPE uint32_t
472 #endif
473 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
474 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
475 
476 typedef struct if_pkt_info_pad {
477 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
478 } *if_pkt_info_pad_t;
479 typedef struct if_rxd_info_pad {
480 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
481 } *if_rxd_info_pad_t;
482 
483 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
484 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
485 
486 
487 static inline void
488 pkt_info_zero(if_pkt_info_t pi)
489 {
490 	if_pkt_info_pad_t pi_pad;
491 
492 	pi_pad = (if_pkt_info_pad_t)pi;
493 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
494 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
495 #ifndef __LP64__
496 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
497 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
498 #endif
499 }
500 
501 static device_method_t iflib_pseudo_methods[] = {
502 	DEVMETHOD(device_attach, noop_attach),
503 	DEVMETHOD(device_detach, iflib_pseudo_detach),
504 	DEVMETHOD_END
505 };
506 
507 driver_t iflib_pseudodriver = {
508 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
509 };
510 
511 static inline void
512 rxd_info_zero(if_rxd_info_t ri)
513 {
514 	if_rxd_info_pad_t ri_pad;
515 	int i;
516 
517 	ri_pad = (if_rxd_info_pad_t)ri;
518 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
519 		ri_pad->rxd_val[i] = 0;
520 		ri_pad->rxd_val[i+1] = 0;
521 		ri_pad->rxd_val[i+2] = 0;
522 		ri_pad->rxd_val[i+3] = 0;
523 	}
524 #ifdef __LP64__
525 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
526 #endif
527 }
528 
529 /*
530  * Only allow a single packet to take up most 1/nth of the tx ring
531  */
532 #define MAX_SINGLE_PACKET_FRACTION 12
533 #define IF_BAD_DMA (bus_addr_t)-1
534 
535 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
536 
537 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
538 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
539 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
540 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
541 
542 
543 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
544 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
545 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
546 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
547 
548 
549 
550 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
551 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
552 
553 void
554 iflib_set_detach(if_ctx_t ctx)
555 {
556 	STATE_LOCK(ctx);
557 	ctx->ifc_flags |= IFC_IN_DETACH;
558 	STATE_UNLOCK(ctx);
559 }
560 
561 /* Our boot-time initialization hook */
562 static int	iflib_module_event_handler(module_t, int, void *);
563 
564 static moduledata_t iflib_moduledata = {
565 	"iflib",
566 	iflib_module_event_handler,
567 	NULL
568 };
569 
570 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
571 MODULE_VERSION(iflib, 1);
572 
573 MODULE_DEPEND(iflib, pci, 1, 1, 1);
574 MODULE_DEPEND(iflib, ether, 1, 1, 1);
575 
576 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
577 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
578 
579 #ifndef IFLIB_DEBUG_COUNTERS
580 #ifdef INVARIANTS
581 #define IFLIB_DEBUG_COUNTERS 1
582 #else
583 #define IFLIB_DEBUG_COUNTERS 0
584 #endif /* !INVARIANTS */
585 #endif
586 
587 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
588                    "iflib driver parameters");
589 
590 /*
591  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
592  */
593 static int iflib_min_tx_latency = 0;
594 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
595 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
596 static int iflib_no_tx_batch = 0;
597 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
598 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
599 
600 
601 #if IFLIB_DEBUG_COUNTERS
602 
603 static int iflib_tx_seen;
604 static int iflib_tx_sent;
605 static int iflib_tx_encap;
606 static int iflib_rx_allocs;
607 static int iflib_fl_refills;
608 static int iflib_fl_refills_large;
609 static int iflib_tx_frees;
610 
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
612 		   &iflib_tx_seen, 0, "# tx mbufs seen");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
614 		   &iflib_tx_sent, 0, "# tx mbufs sent");
615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
616 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
617 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
618 		   &iflib_tx_frees, 0, "# tx frees");
619 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
620 		   &iflib_rx_allocs, 0, "# rx allocations");
621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
622 		   &iflib_fl_refills, 0, "# refills");
623 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
624 		   &iflib_fl_refills_large, 0, "# large refills");
625 
626 
627 static int iflib_txq_drain_flushing;
628 static int iflib_txq_drain_oactive;
629 static int iflib_txq_drain_notready;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
632 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
634 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
635 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
636 		   &iflib_txq_drain_notready, 0, "# drain notready");
637 
638 
639 static int iflib_encap_load_mbuf_fail;
640 static int iflib_encap_pad_mbuf_fail;
641 static int iflib_encap_txq_avail_fail;
642 static int iflib_encap_txd_encap_fail;
643 
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
645 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
647 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
649 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
650 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
651 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
652 
653 static int iflib_task_fn_rxs;
654 static int iflib_rx_intr_enables;
655 static int iflib_fast_intrs;
656 static int iflib_rx_unavail;
657 static int iflib_rx_ctx_inactive;
658 static int iflib_rx_if_input;
659 static int iflib_rxd_flush;
660 
661 static int iflib_verbose_debug;
662 
663 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
664 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
666 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
667 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
668 		   &iflib_fast_intrs, 0, "# fast_intr calls");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
670 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
671 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
672 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
673 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
674 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
675 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
676 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
677 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
678 		   &iflib_verbose_debug, 0, "enable verbose debugging");
679 
680 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
681 static void
682 iflib_debug_reset(void)
683 {
684 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
685 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
686 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
687 		iflib_txq_drain_notready =
688 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
689 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
690 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
691 		iflib_rx_unavail =
692 		iflib_rx_ctx_inactive = iflib_rx_if_input =
693 		iflib_rxd_flush = 0;
694 }
695 
696 #else
697 #define DBG_COUNTER_INC(name)
698 static void iflib_debug_reset(void) {}
699 #endif
700 
701 #define IFLIB_DEBUG 0
702 
703 static void iflib_tx_structures_free(if_ctx_t ctx);
704 static void iflib_rx_structures_free(if_ctx_t ctx);
705 static int iflib_queues_alloc(if_ctx_t ctx);
706 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
707 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
708 static int iflib_qset_structures_setup(if_ctx_t ctx);
709 static int iflib_msix_init(if_ctx_t ctx);
710 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
711 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
712 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
713 #ifdef ALTQ
714 static void iflib_altq_if_start(if_t ifp);
715 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
716 #endif
717 static int iflib_register(if_ctx_t);
718 static void iflib_init_locked(if_ctx_t ctx);
719 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
721 static void iflib_ifmp_purge(iflib_txq_t txq);
722 static void _iflib_pre_assert(if_softc_ctx_t scctx);
723 static void iflib_if_init_locked(if_ctx_t ctx);
724 static void iflib_free_intr_mem(if_ctx_t ctx);
725 #ifndef __NO_STRICT_ALIGNMENT
726 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
727 #endif
728 
729 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
730     SLIST_HEAD_INITIALIZER(cpu_offsets);
731 struct cpu_offset {
732 	SLIST_ENTRY(cpu_offset) entries;
733 	cpuset_t	set;
734 	unsigned int	refcount;
735 	uint16_t	offset;
736 };
737 static struct mtx cpu_offset_mtx;
738 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
739     MTX_DEF);
740 
741 NETDUMP_DEFINE(iflib);
742 
743 #ifdef DEV_NETMAP
744 #include <sys/selinfo.h>
745 #include <net/netmap.h>
746 #include <dev/netmap/netmap_kern.h>
747 
748 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
749 
750 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
751 
752 /*
753  * device-specific sysctl variables:
754  *
755  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
756  *	During regular operations the CRC is stripped, but on some
757  *	hardware reception of frames not multiple of 64 is slower,
758  *	so using crcstrip=0 helps in benchmarks.
759  *
760  * iflib_rx_miss, iflib_rx_miss_bufs:
761  *	count packets that might be missed due to lost interrupts.
762  */
763 SYSCTL_DECL(_dev_netmap);
764 /*
765  * The xl driver by default strips CRCs and we do not override it.
766  */
767 
768 int iflib_crcstrip = 1;
769 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
770     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
771 
772 int iflib_rx_miss, iflib_rx_miss_bufs;
773 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
774     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
775 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
776     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
777 
778 /*
779  * Register/unregister. We are already under netmap lock.
780  * Only called on the first register or the last unregister.
781  */
782 static int
783 iflib_netmap_register(struct netmap_adapter *na, int onoff)
784 {
785 	struct ifnet *ifp = na->ifp;
786 	if_ctx_t ctx = ifp->if_softc;
787 	int status;
788 
789 	CTX_LOCK(ctx);
790 	IFDI_INTR_DISABLE(ctx);
791 
792 	/* Tell the stack that the interface is no longer active */
793 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
794 
795 	if (!CTX_IS_VF(ctx))
796 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
797 
798 	/* enable or disable flags and callbacks in na and ifp */
799 	if (onoff) {
800 		nm_set_native_flags(na);
801 	} else {
802 		nm_clear_native_flags(na);
803 	}
804 	iflib_stop(ctx);
805 	iflib_init_locked(ctx);
806 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
807 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
808 	if (status)
809 		nm_clear_native_flags(na);
810 	CTX_UNLOCK(ctx);
811 	return (status);
812 }
813 
814 static int
815 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
816 {
817 	struct netmap_adapter *na = kring->na;
818 	u_int const lim = kring->nkr_num_slots - 1;
819 	u_int head = kring->rhead;
820 	struct netmap_ring *ring = kring->ring;
821 	bus_dmamap_t *map;
822 	struct if_rxd_update iru;
823 	if_ctx_t ctx = rxq->ifr_ctx;
824 	iflib_fl_t fl = &rxq->ifr_fl[0];
825 	uint32_t refill_pidx, nic_i;
826 #if IFLIB_DEBUG_COUNTERS
827 	int rf_count = 0;
828 #endif
829 
830 	if (nm_i == head && __predict_true(!init))
831 		return 0;
832 	iru_init(&iru, rxq, 0 /* flid */);
833 	map = fl->ifl_sds.ifsd_map;
834 	refill_pidx = netmap_idx_k2n(kring, nm_i);
835 	/*
836 	 * IMPORTANT: we must leave one free slot in the ring,
837 	 * so move head back by one unit
838 	 */
839 	head = nm_prev(head, lim);
840 	nic_i = UINT_MAX;
841 	DBG_COUNTER_INC(fl_refills);
842 	while (nm_i != head) {
843 #if IFLIB_DEBUG_COUNTERS
844 		if (++rf_count == 9)
845 			DBG_COUNTER_INC(fl_refills_large);
846 #endif
847 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
848 			struct netmap_slot *slot = &ring->slot[nm_i];
849 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
850 			uint32_t nic_i_dma = refill_pidx;
851 			nic_i = netmap_idx_k2n(kring, nm_i);
852 
853 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
854 
855 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
856 			        return netmap_ring_reinit(kring);
857 
858 			fl->ifl_vm_addrs[tmp_pidx] = addr;
859 			if (__predict_false(init)) {
860 				netmap_load_map(na, fl->ifl_buf_tag,
861 				    map[nic_i], addr);
862 			} else if (slot->flags & NS_BUF_CHANGED) {
863 				/* buffer has changed, reload map */
864 				netmap_reload_map(na, fl->ifl_buf_tag,
865 				    map[nic_i], addr);
866 			}
867 			slot->flags &= ~NS_BUF_CHANGED;
868 
869 			nm_i = nm_next(nm_i, lim);
870 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
871 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
872 				continue;
873 
874 			iru.iru_pidx = refill_pidx;
875 			iru.iru_count = tmp_pidx+1;
876 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
877 			refill_pidx = nic_i;
878 			for (int n = 0; n < iru.iru_count; n++) {
879 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
880 						BUS_DMASYNC_PREREAD);
881 				/* XXX - change this to not use the netmap func*/
882 				nic_i_dma = nm_next(nic_i_dma, lim);
883 			}
884 		}
885 	}
886 	kring->nr_hwcur = head;
887 
888 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
889 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
890 	if (__predict_true(nic_i != UINT_MAX)) {
891 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
892 		DBG_COUNTER_INC(rxd_flush);
893 	}
894 	return (0);
895 }
896 
897 /*
898  * Reconcile kernel and user view of the transmit ring.
899  *
900  * All information is in the kring.
901  * Userspace wants to send packets up to the one before kring->rhead,
902  * kernel knows kring->nr_hwcur is the first unsent packet.
903  *
904  * Here we push packets out (as many as possible), and possibly
905  * reclaim buffers from previously completed transmission.
906  *
907  * The caller (netmap) guarantees that there is only one instance
908  * running at any time. Any interference with other driver
909  * methods should be handled by the individual drivers.
910  */
911 static int
912 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
913 {
914 	struct netmap_adapter *na = kring->na;
915 	struct ifnet *ifp = na->ifp;
916 	struct netmap_ring *ring = kring->ring;
917 	u_int nm_i;	/* index into the netmap kring */
918 	u_int nic_i;	/* index into the NIC ring */
919 	u_int n;
920 	u_int const lim = kring->nkr_num_slots - 1;
921 	u_int const head = kring->rhead;
922 	struct if_pkt_info pi;
923 
924 	/*
925 	 * interrupts on every tx packet are expensive so request
926 	 * them every half ring, or where NS_REPORT is set
927 	 */
928 	u_int report_frequency = kring->nkr_num_slots >> 1;
929 	/* device-specific */
930 	if_ctx_t ctx = ifp->if_softc;
931 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
932 
933 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
934 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
935 
936 	/*
937 	 * First part: process new packets to send.
938 	 * nm_i is the current index in the netmap kring,
939 	 * nic_i is the corresponding index in the NIC ring.
940 	 *
941 	 * If we have packets to send (nm_i != head)
942 	 * iterate over the netmap ring, fetch length and update
943 	 * the corresponding slot in the NIC ring. Some drivers also
944 	 * need to update the buffer's physical address in the NIC slot
945 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
946 	 *
947 	 * The netmap_reload_map() calls is especially expensive,
948 	 * even when (as in this case) the tag is 0, so do only
949 	 * when the buffer has actually changed.
950 	 *
951 	 * If possible do not set the report/intr bit on all slots,
952 	 * but only a few times per ring or when NS_REPORT is set.
953 	 *
954 	 * Finally, on 10G and faster drivers, it might be useful
955 	 * to prefetch the next slot and txr entry.
956 	 */
957 
958 	nm_i = kring->nr_hwcur;
959 	if (nm_i != head) {	/* we have new packets to send */
960 		pkt_info_zero(&pi);
961 		pi.ipi_segs = txq->ift_segs;
962 		pi.ipi_qsidx = kring->ring_id;
963 		nic_i = netmap_idx_k2n(kring, nm_i);
964 
965 		__builtin_prefetch(&ring->slot[nm_i]);
966 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
967 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
968 
969 		for (n = 0; nm_i != head; n++) {
970 			struct netmap_slot *slot = &ring->slot[nm_i];
971 			u_int len = slot->len;
972 			uint64_t paddr;
973 			void *addr = PNMB(na, slot, &paddr);
974 			int flags = (slot->flags & NS_REPORT ||
975 				nic_i == 0 || nic_i == report_frequency) ?
976 				IPI_TX_INTR : 0;
977 
978 			/* device-specific */
979 			pi.ipi_len = len;
980 			pi.ipi_segs[0].ds_addr = paddr;
981 			pi.ipi_segs[0].ds_len = len;
982 			pi.ipi_nsegs = 1;
983 			pi.ipi_ndescs = 0;
984 			pi.ipi_pidx = nic_i;
985 			pi.ipi_flags = flags;
986 
987 			/* Fill the slot in the NIC ring. */
988 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
989 			DBG_COUNTER_INC(tx_encap);
990 
991 			/* prefetch for next round */
992 			__builtin_prefetch(&ring->slot[nm_i + 1]);
993 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
994 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
995 
996 			NM_CHECK_ADDR_LEN(na, addr, len);
997 
998 			if (slot->flags & NS_BUF_CHANGED) {
999 				/* buffer has changed, reload map */
1000 				netmap_reload_map(na, txq->ift_buf_tag,
1001 				    txq->ift_sds.ifsd_map[nic_i], addr);
1002 			}
1003 			/* make sure changes to the buffer are synced */
1004 			bus_dmamap_sync(txq->ift_buf_tag,
1005 			    txq->ift_sds.ifsd_map[nic_i],
1006 			    BUS_DMASYNC_PREWRITE);
1007 
1008 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1009 			nm_i = nm_next(nm_i, lim);
1010 			nic_i = nm_next(nic_i, lim);
1011 		}
1012 		kring->nr_hwcur = nm_i;
1013 
1014 		/* synchronize the NIC ring */
1015 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1016 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1017 
1018 		/* (re)start the tx unit up to slot nic_i (excluded) */
1019 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1020 	}
1021 
1022 	/*
1023 	 * Second part: reclaim buffers for completed transmissions.
1024 	 *
1025 	 * If there are unclaimed buffers, attempt to reclaim them.
1026 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1027 	 * minimal delay, then trigger the tx handler which will spin in the
1028 	 * group task queue.
1029 	 */
1030 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1031 		if (iflib_tx_credits_update(ctx, txq)) {
1032 			/* some tx completed, increment avail */
1033 			nic_i = txq->ift_cidx_processed;
1034 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1035 		}
1036 	}
1037 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1038 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1039 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1040 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1041 	}
1042 	return (0);
1043 }
1044 
1045 /*
1046  * Reconcile kernel and user view of the receive ring.
1047  * Same as for the txsync, this routine must be efficient.
1048  * The caller guarantees a single invocations, but races against
1049  * the rest of the driver should be handled here.
1050  *
1051  * On call, kring->rhead is the first packet that userspace wants
1052  * to keep, and kring->rcur is the wakeup point.
1053  * The kernel has previously reported packets up to kring->rtail.
1054  *
1055  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1056  * of whether or not we received an interrupt.
1057  */
1058 static int
1059 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1060 {
1061 	struct netmap_adapter *na = kring->na;
1062 	struct netmap_ring *ring = kring->ring;
1063 	iflib_fl_t fl;
1064 	uint32_t nm_i;	/* index into the netmap ring */
1065 	uint32_t nic_i;	/* index into the NIC ring */
1066 	u_int i, n;
1067 	u_int const lim = kring->nkr_num_slots - 1;
1068 	u_int const head = kring->rhead;
1069 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1070 	struct if_rxd_info ri;
1071 
1072 	struct ifnet *ifp = na->ifp;
1073 	if_ctx_t ctx = ifp->if_softc;
1074 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1075 	if (head > lim)
1076 		return netmap_ring_reinit(kring);
1077 
1078 	/*
1079 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1080 	 */
1081 
1082 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1083 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1084 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1085 	}
1086 
1087 	/*
1088 	 * First part: import newly received packets.
1089 	 *
1090 	 * nm_i is the index of the next free slot in the netmap ring,
1091 	 * nic_i is the index of the next received packet in the NIC ring,
1092 	 * and they may differ in case if_init() has been called while
1093 	 * in netmap mode. For the receive ring we have
1094 	 *
1095 	 *	nic_i = rxr->next_check;
1096 	 *	nm_i = kring->nr_hwtail (previous)
1097 	 * and
1098 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1099 	 *
1100 	 * rxr->next_check is set to 0 on a ring reinit
1101 	 */
1102 	if (netmap_no_pendintr || force_update) {
1103 		int crclen = iflib_crcstrip ? 0 : 4;
1104 		int error, avail;
1105 
1106 		for (i = 0; i < rxq->ifr_nfl; i++) {
1107 			fl = &rxq->ifr_fl[i];
1108 			nic_i = fl->ifl_cidx;
1109 			nm_i = netmap_idx_n2k(kring, nic_i);
1110 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1111 			    rxq->ifr_id, nic_i, USHRT_MAX);
1112 			for (n = 0; avail > 0; n++, avail--) {
1113 				rxd_info_zero(&ri);
1114 				ri.iri_frags = rxq->ifr_frags;
1115 				ri.iri_qsidx = kring->ring_id;
1116 				ri.iri_ifp = ctx->ifc_ifp;
1117 				ri.iri_cidx = nic_i;
1118 
1119 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1120 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1121 				ring->slot[nm_i].flags = 0;
1122 				bus_dmamap_sync(fl->ifl_buf_tag,
1123 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1124 				nm_i = nm_next(nm_i, lim);
1125 				nic_i = nm_next(nic_i, lim);
1126 			}
1127 			if (n) { /* update the state variables */
1128 				if (netmap_no_pendintr && !force_update) {
1129 					/* diagnostics */
1130 					iflib_rx_miss ++;
1131 					iflib_rx_miss_bufs += n;
1132 				}
1133 				fl->ifl_cidx = nic_i;
1134 				kring->nr_hwtail = nm_i;
1135 			}
1136 			kring->nr_kflags &= ~NKR_PENDINTR;
1137 		}
1138 	}
1139 	/*
1140 	 * Second part: skip past packets that userspace has released.
1141 	 * (kring->nr_hwcur to head excluded),
1142 	 * and make the buffers available for reception.
1143 	 * As usual nm_i is the index in the netmap ring,
1144 	 * nic_i is the index in the NIC ring, and
1145 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1146 	 */
1147 	/* XXX not sure how this will work with multiple free lists */
1148 	nm_i = kring->nr_hwcur;
1149 
1150 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1151 }
1152 
1153 static void
1154 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1155 {
1156 	struct ifnet *ifp = na->ifp;
1157 	if_ctx_t ctx = ifp->if_softc;
1158 
1159 	CTX_LOCK(ctx);
1160 	if (onoff) {
1161 		IFDI_INTR_ENABLE(ctx);
1162 	} else {
1163 		IFDI_INTR_DISABLE(ctx);
1164 	}
1165 	CTX_UNLOCK(ctx);
1166 }
1167 
1168 
1169 static int
1170 iflib_netmap_attach(if_ctx_t ctx)
1171 {
1172 	struct netmap_adapter na;
1173 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1174 
1175 	bzero(&na, sizeof(na));
1176 
1177 	na.ifp = ctx->ifc_ifp;
1178 	na.na_flags = NAF_BDG_MAYSLEEP;
1179 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1180 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1181 
1182 	na.num_tx_desc = scctx->isc_ntxd[0];
1183 	na.num_rx_desc = scctx->isc_nrxd[0];
1184 	na.nm_txsync = iflib_netmap_txsync;
1185 	na.nm_rxsync = iflib_netmap_rxsync;
1186 	na.nm_register = iflib_netmap_register;
1187 	na.nm_intr = iflib_netmap_intr;
1188 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1189 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1190 	return (netmap_attach(&na));
1191 }
1192 
1193 static void
1194 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1195 {
1196 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1197 	struct netmap_slot *slot;
1198 
1199 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1200 	if (slot == NULL)
1201 		return;
1202 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1203 
1204 		/*
1205 		 * In netmap mode, set the map for the packet buffer.
1206 		 * NOTE: Some drivers (not this one) also need to set
1207 		 * the physical buffer address in the NIC ring.
1208 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1209 		 * netmap slot index, si
1210 		 */
1211 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1212 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1213 		    NMB(na, slot + si));
1214 	}
1215 }
1216 
1217 static void
1218 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1219 {
1220 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1221 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1222 	struct netmap_slot *slot;
1223 	uint32_t nm_i;
1224 
1225 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1226 	if (slot == NULL)
1227 		return;
1228 	nm_i = netmap_idx_n2k(kring, 0);
1229 	netmap_fl_refill(rxq, kring, nm_i, true);
1230 }
1231 
1232 static void
1233 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1234 {
1235 	struct netmap_kring *kring;
1236 	uint16_t txqid;
1237 
1238 	txqid = txq->ift_id;
1239 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1240 
1241 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1242 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1243 		    BUS_DMASYNC_POSTREAD);
1244 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1245 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1246 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1247 			if (hz < 2000)
1248 				*reset_on = 1;
1249 			else
1250 				*reset_on = hz / 1000;
1251 		}
1252 	}
1253 }
1254 
1255 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1256 
1257 #else
1258 #define iflib_netmap_txq_init(ctx, txq)
1259 #define iflib_netmap_rxq_init(ctx, rxq)
1260 #define iflib_netmap_detach(ifp)
1261 
1262 #define iflib_netmap_attach(ctx) (0)
1263 #define netmap_rx_irq(ifp, qid, budget) (0)
1264 #define netmap_tx_irq(ifp, qid) do {} while (0)
1265 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1266 
1267 #endif
1268 
1269 #if defined(__i386__) || defined(__amd64__)
1270 static __inline void
1271 prefetch(void *x)
1272 {
1273 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1274 }
1275 static __inline void
1276 prefetch2cachelines(void *x)
1277 {
1278 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1279 #if (CACHE_LINE_SIZE < 128)
1280 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1281 #endif
1282 }
1283 #else
1284 #define prefetch(x)
1285 #define prefetch2cachelines(x)
1286 #endif
1287 
1288 static void
1289 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1290 {
1291 	iflib_fl_t fl;
1292 
1293 	fl = &rxq->ifr_fl[flid];
1294 	iru->iru_paddrs = fl->ifl_bus_addrs;
1295 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1296 	iru->iru_idxs = fl->ifl_rxd_idxs;
1297 	iru->iru_qsidx = rxq->ifr_id;
1298 	iru->iru_buf_size = fl->ifl_buf_size;
1299 	iru->iru_flidx = fl->ifl_id;
1300 }
1301 
1302 static void
1303 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1304 {
1305 	if (err)
1306 		return;
1307 	*(bus_addr_t *) arg = segs[0].ds_addr;
1308 }
1309 
1310 int
1311 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1312 {
1313 	int err;
1314 	device_t dev = ctx->ifc_dev;
1315 
1316 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1317 				align, 0,		/* alignment, bounds */
1318 				BUS_SPACE_MAXADDR,	/* lowaddr */
1319 				BUS_SPACE_MAXADDR,	/* highaddr */
1320 				NULL, NULL,		/* filter, filterarg */
1321 				size,			/* maxsize */
1322 				1,			/* nsegments */
1323 				size,			/* maxsegsize */
1324 				BUS_DMA_ALLOCNOW,	/* flags */
1325 				NULL,			/* lockfunc */
1326 				NULL,			/* lockarg */
1327 				&dma->idi_tag);
1328 	if (err) {
1329 		device_printf(dev,
1330 		    "%s: bus_dma_tag_create failed: %d\n",
1331 		    __func__, err);
1332 		goto fail_0;
1333 	}
1334 
1335 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1336 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1337 	if (err) {
1338 		device_printf(dev,
1339 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1340 		    __func__, (uintmax_t)size, err);
1341 		goto fail_1;
1342 	}
1343 
1344 	dma->idi_paddr = IF_BAD_DMA;
1345 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1346 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1347 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1348 		device_printf(dev,
1349 		    "%s: bus_dmamap_load failed: %d\n",
1350 		    __func__, err);
1351 		goto fail_2;
1352 	}
1353 
1354 	dma->idi_size = size;
1355 	return (0);
1356 
1357 fail_2:
1358 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1359 fail_1:
1360 	bus_dma_tag_destroy(dma->idi_tag);
1361 fail_0:
1362 	dma->idi_tag = NULL;
1363 
1364 	return (err);
1365 }
1366 
1367 int
1368 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1369 {
1370 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1371 
1372 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1373 
1374 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1375 }
1376 
1377 int
1378 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1379 {
1380 	int i, err;
1381 	iflib_dma_info_t *dmaiter;
1382 
1383 	dmaiter = dmalist;
1384 	for (i = 0; i < count; i++, dmaiter++) {
1385 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1386 			break;
1387 	}
1388 	if (err)
1389 		iflib_dma_free_multi(dmalist, i);
1390 	return (err);
1391 }
1392 
1393 void
1394 iflib_dma_free(iflib_dma_info_t dma)
1395 {
1396 	if (dma->idi_tag == NULL)
1397 		return;
1398 	if (dma->idi_paddr != IF_BAD_DMA) {
1399 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1400 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1401 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1402 		dma->idi_paddr = IF_BAD_DMA;
1403 	}
1404 	if (dma->idi_vaddr != NULL) {
1405 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1406 		dma->idi_vaddr = NULL;
1407 	}
1408 	bus_dma_tag_destroy(dma->idi_tag);
1409 	dma->idi_tag = NULL;
1410 }
1411 
1412 void
1413 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1414 {
1415 	int i;
1416 	iflib_dma_info_t *dmaiter = dmalist;
1417 
1418 	for (i = 0; i < count; i++, dmaiter++)
1419 		iflib_dma_free(*dmaiter);
1420 }
1421 
1422 #ifdef EARLY_AP_STARTUP
1423 static const int iflib_started = 1;
1424 #else
1425 /*
1426  * We used to abuse the smp_started flag to decide if the queues have been
1427  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1428  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1429  * is set.  Run a SYSINIT() strictly after that to just set a usable
1430  * completion flag.
1431  */
1432 
1433 static int iflib_started;
1434 
1435 static void
1436 iflib_record_started(void *arg)
1437 {
1438 	iflib_started = 1;
1439 }
1440 
1441 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1442 	iflib_record_started, NULL);
1443 #endif
1444 
1445 static int
1446 iflib_fast_intr(void *arg)
1447 {
1448 	iflib_filter_info_t info = arg;
1449 	struct grouptask *gtask = info->ifi_task;
1450 	int result;
1451 
1452 	if (!iflib_started)
1453 		return (FILTER_STRAY);
1454 
1455 	DBG_COUNTER_INC(fast_intrs);
1456 	if (info->ifi_filter != NULL) {
1457 		result = info->ifi_filter(info->ifi_filter_arg);
1458 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1459 			return (result);
1460 	}
1461 
1462 	GROUPTASK_ENQUEUE(gtask);
1463 	return (FILTER_HANDLED);
1464 }
1465 
1466 static int
1467 iflib_fast_intr_rxtx(void *arg)
1468 {
1469 	iflib_filter_info_t info = arg;
1470 	struct grouptask *gtask = info->ifi_task;
1471 	if_ctx_t ctx;
1472 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1473 	iflib_txq_t txq;
1474 	void *sc;
1475 	int i, cidx, result;
1476 	qidx_t txqid;
1477 
1478 	if (!iflib_started)
1479 		return (FILTER_STRAY);
1480 
1481 	DBG_COUNTER_INC(fast_intrs);
1482 	if (info->ifi_filter != NULL) {
1483 		result = info->ifi_filter(info->ifi_filter_arg);
1484 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1485 			return (result);
1486 	}
1487 
1488 	ctx = rxq->ifr_ctx;
1489 	sc = ctx->ifc_softc;
1490 	MPASS(rxq->ifr_ntxqirq);
1491 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1492 		txqid = rxq->ifr_txqid[i];
1493 		txq = &ctx->ifc_txqs[txqid];
1494 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1495 		    BUS_DMASYNC_POSTREAD);
1496 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1497 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1498 			continue;
1499 		}
1500 		GROUPTASK_ENQUEUE(&txq->ift_task);
1501 	}
1502 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1503 		cidx = rxq->ifr_cq_cidx;
1504 	else
1505 		cidx = rxq->ifr_fl[0].ifl_cidx;
1506 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1507 		GROUPTASK_ENQUEUE(gtask);
1508 	else {
1509 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1510 		DBG_COUNTER_INC(rx_intr_enables);
1511 	}
1512 	return (FILTER_HANDLED);
1513 }
1514 
1515 
1516 static int
1517 iflib_fast_intr_ctx(void *arg)
1518 {
1519 	iflib_filter_info_t info = arg;
1520 	struct grouptask *gtask = info->ifi_task;
1521 	int result;
1522 
1523 	if (!iflib_started)
1524 		return (FILTER_STRAY);
1525 
1526 	DBG_COUNTER_INC(fast_intrs);
1527 	if (info->ifi_filter != NULL) {
1528 		result = info->ifi_filter(info->ifi_filter_arg);
1529 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1530 			return (result);
1531 	}
1532 
1533 	GROUPTASK_ENQUEUE(gtask);
1534 	return (FILTER_HANDLED);
1535 }
1536 
1537 static int
1538 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1539 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1540 		 const char *name)
1541 {
1542 	int rc, flags;
1543 	struct resource *res;
1544 	void *tag = NULL;
1545 	device_t dev = ctx->ifc_dev;
1546 
1547 	flags = RF_ACTIVE;
1548 	if (ctx->ifc_flags & IFC_LEGACY)
1549 		flags |= RF_SHAREABLE;
1550 	MPASS(rid < 512);
1551 	irq->ii_rid = rid;
1552 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1553 	if (res == NULL) {
1554 		device_printf(dev,
1555 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1556 		return (ENOMEM);
1557 	}
1558 	irq->ii_res = res;
1559 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1560 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1561 						filter, handler, arg, &tag);
1562 	if (rc != 0) {
1563 		device_printf(dev,
1564 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1565 					  rid, name ? name : "unknown", rc);
1566 		return (rc);
1567 	} else if (name)
1568 		bus_describe_intr(dev, res, tag, "%s", name);
1569 
1570 	irq->ii_tag = tag;
1571 	return (0);
1572 }
1573 
1574 
1575 /*********************************************************************
1576  *
1577  *  Allocate DMA resources for TX buffers as well as memory for the TX
1578  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1579  *  iflib_sw_tx_desc_array structure, storing all the information that
1580  *  is needed to transmit a packet on the wire.  This is called only
1581  *  once at attach, setup is done every reset.
1582  *
1583  **********************************************************************/
1584 static int
1585 iflib_txsd_alloc(iflib_txq_t txq)
1586 {
1587 	if_ctx_t ctx = txq->ift_ctx;
1588 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1589 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1590 	device_t dev = ctx->ifc_dev;
1591 	bus_size_t tsomaxsize;
1592 	int err, nsegments, ntsosegments;
1593 	bool tso;
1594 
1595 	nsegments = scctx->isc_tx_nsegments;
1596 	ntsosegments = scctx->isc_tx_tso_segments_max;
1597 	tsomaxsize = scctx->isc_tx_tso_size_max;
1598 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1599 		tsomaxsize += sizeof(struct ether_vlan_header);
1600 	MPASS(scctx->isc_ntxd[0] > 0);
1601 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1602 	MPASS(nsegments > 0);
1603 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1604 		MPASS(ntsosegments > 0);
1605 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1606 	}
1607 
1608 	/*
1609 	 * Set up DMA tags for TX buffers.
1610 	 */
1611 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1612 			       1, 0,			/* alignment, bounds */
1613 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1614 			       BUS_SPACE_MAXADDR,	/* highaddr */
1615 			       NULL, NULL,		/* filter, filterarg */
1616 			       sctx->isc_tx_maxsize,		/* maxsize */
1617 			       nsegments,	/* nsegments */
1618 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1619 			       0,			/* flags */
1620 			       NULL,			/* lockfunc */
1621 			       NULL,			/* lockfuncarg */
1622 			       &txq->ift_buf_tag))) {
1623 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1624 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1625 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1626 		goto fail;
1627 	}
1628 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1629 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1630 			       1, 0,			/* alignment, bounds */
1631 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1632 			       BUS_SPACE_MAXADDR,	/* highaddr */
1633 			       NULL, NULL,		/* filter, filterarg */
1634 			       tsomaxsize,		/* maxsize */
1635 			       ntsosegments,	/* nsegments */
1636 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1637 			       0,			/* flags */
1638 			       NULL,			/* lockfunc */
1639 			       NULL,			/* lockfuncarg */
1640 			       &txq->ift_tso_buf_tag))) {
1641 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1642 		    err);
1643 		goto fail;
1644 	}
1645 
1646 	/* Allocate memory for the TX mbuf map. */
1647 	if (!(txq->ift_sds.ifsd_m =
1648 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1649 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1650 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1651 		err = ENOMEM;
1652 		goto fail;
1653 	}
1654 
1655 	/*
1656 	 * Create the DMA maps for TX buffers.
1657 	 */
1658 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1659 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1660 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1661 		device_printf(dev,
1662 		    "Unable to allocate TX buffer DMA map memory\n");
1663 		err = ENOMEM;
1664 		goto fail;
1665 	}
1666 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1667 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1668 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1669 		device_printf(dev,
1670 		    "Unable to allocate TSO TX buffer map memory\n");
1671 		err = ENOMEM;
1672 		goto fail;
1673 	}
1674 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1675 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1676 		    &txq->ift_sds.ifsd_map[i]);
1677 		if (err != 0) {
1678 			device_printf(dev, "Unable to create TX DMA map\n");
1679 			goto fail;
1680 		}
1681 		if (!tso)
1682 			continue;
1683 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1684 		    &txq->ift_sds.ifsd_tso_map[i]);
1685 		if (err != 0) {
1686 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1687 			goto fail;
1688 		}
1689 	}
1690 	return (0);
1691 fail:
1692 	/* We free all, it handles case where we are in the middle */
1693 	iflib_tx_structures_free(ctx);
1694 	return (err);
1695 }
1696 
1697 static void
1698 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1699 {
1700 	bus_dmamap_t map;
1701 
1702 	map = NULL;
1703 	if (txq->ift_sds.ifsd_map != NULL)
1704 		map = txq->ift_sds.ifsd_map[i];
1705 	if (map != NULL) {
1706 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1707 		bus_dmamap_unload(txq->ift_buf_tag, map);
1708 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1709 		txq->ift_sds.ifsd_map[i] = NULL;
1710 	}
1711 
1712 	map = NULL;
1713 	if (txq->ift_sds.ifsd_tso_map != NULL)
1714 		map = txq->ift_sds.ifsd_tso_map[i];
1715 	if (map != NULL) {
1716 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1717 		    BUS_DMASYNC_POSTWRITE);
1718 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1719 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1720 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1721 	}
1722 }
1723 
1724 static void
1725 iflib_txq_destroy(iflib_txq_t txq)
1726 {
1727 	if_ctx_t ctx = txq->ift_ctx;
1728 
1729 	for (int i = 0; i < txq->ift_size; i++)
1730 		iflib_txsd_destroy(ctx, txq, i);
1731 	if (txq->ift_sds.ifsd_map != NULL) {
1732 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1733 		txq->ift_sds.ifsd_map = NULL;
1734 	}
1735 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1736 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1737 		txq->ift_sds.ifsd_tso_map = NULL;
1738 	}
1739 	if (txq->ift_sds.ifsd_m != NULL) {
1740 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1741 		txq->ift_sds.ifsd_m = NULL;
1742 	}
1743 	if (txq->ift_buf_tag != NULL) {
1744 		bus_dma_tag_destroy(txq->ift_buf_tag);
1745 		txq->ift_buf_tag = NULL;
1746 	}
1747 	if (txq->ift_tso_buf_tag != NULL) {
1748 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1749 		txq->ift_tso_buf_tag = NULL;
1750 	}
1751 }
1752 
1753 static void
1754 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1755 {
1756 	struct mbuf **mp;
1757 
1758 	mp = &txq->ift_sds.ifsd_m[i];
1759 	if (*mp == NULL)
1760 		return;
1761 
1762 	if (txq->ift_sds.ifsd_map != NULL) {
1763 		bus_dmamap_sync(txq->ift_buf_tag,
1764 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1765 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1766 	}
1767 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1768 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1769 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1770 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1771 		    txq->ift_sds.ifsd_tso_map[i]);
1772 	}
1773 	m_free(*mp);
1774 	DBG_COUNTER_INC(tx_frees);
1775 	*mp = NULL;
1776 }
1777 
1778 static int
1779 iflib_txq_setup(iflib_txq_t txq)
1780 {
1781 	if_ctx_t ctx = txq->ift_ctx;
1782 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1783 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1784 	iflib_dma_info_t di;
1785 	int i;
1786 
1787 	/* Set number of descriptors available */
1788 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1789 	/* XXX make configurable */
1790 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1791 
1792 	/* Reset indices */
1793 	txq->ift_cidx_processed = 0;
1794 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1795 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1796 
1797 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1798 		bzero((void *)di->idi_vaddr, di->idi_size);
1799 
1800 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1801 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1802 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1803 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1804 	return (0);
1805 }
1806 
1807 /*********************************************************************
1808  *
1809  *  Allocate DMA resources for RX buffers as well as memory for the RX
1810  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1811  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1812  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1813  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1814  *  packet, the maximum number of entries we'll need is equal to the
1815  *  number of hardware receive descriptors that we've allocated.
1816  *
1817  **********************************************************************/
1818 static int
1819 iflib_rxsd_alloc(iflib_rxq_t rxq)
1820 {
1821 	if_ctx_t ctx = rxq->ifr_ctx;
1822 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1823 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1824 	device_t dev = ctx->ifc_dev;
1825 	iflib_fl_t fl;
1826 	int			err;
1827 
1828 	MPASS(scctx->isc_nrxd[0] > 0);
1829 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1830 
1831 	fl = rxq->ifr_fl;
1832 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1833 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1834 		/* Set up DMA tag for RX buffers. */
1835 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1836 					 1, 0,			/* alignment, bounds */
1837 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1838 					 BUS_SPACE_MAXADDR,	/* highaddr */
1839 					 NULL, NULL,		/* filter, filterarg */
1840 					 sctx->isc_rx_maxsize,	/* maxsize */
1841 					 sctx->isc_rx_nsegments,	/* nsegments */
1842 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1843 					 0,			/* flags */
1844 					 NULL,			/* lockfunc */
1845 					 NULL,			/* lockarg */
1846 					 &fl->ifl_buf_tag);
1847 		if (err) {
1848 			device_printf(dev,
1849 			    "Unable to allocate RX DMA tag: %d\n", err);
1850 			goto fail;
1851 		}
1852 
1853 		/* Allocate memory for the RX mbuf map. */
1854 		if (!(fl->ifl_sds.ifsd_m =
1855 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1856 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1857 			device_printf(dev,
1858 			    "Unable to allocate RX mbuf map memory\n");
1859 			err = ENOMEM;
1860 			goto fail;
1861 		}
1862 
1863 		/* Allocate memory for the direct RX cluster pointer map. */
1864 		if (!(fl->ifl_sds.ifsd_cl =
1865 		      (caddr_t *) malloc(sizeof(caddr_t) *
1866 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1867 			device_printf(dev,
1868 			    "Unable to allocate RX cluster map memory\n");
1869 			err = ENOMEM;
1870 			goto fail;
1871 		}
1872 
1873 		/* Allocate memory for the RX cluster bus address map. */
1874 		if (!(fl->ifl_sds.ifsd_ba =
1875 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1876 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1877 			device_printf(dev,
1878 			    "Unable to allocate RX bus address map memory\n");
1879 			err = ENOMEM;
1880 			goto fail;
1881 		}
1882 
1883 		/*
1884 		 * Create the DMA maps for RX buffers.
1885 		 */
1886 		if (!(fl->ifl_sds.ifsd_map =
1887 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1888 			device_printf(dev,
1889 			    "Unable to allocate RX buffer DMA map memory\n");
1890 			err = ENOMEM;
1891 			goto fail;
1892 		}
1893 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1894 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1895 			    &fl->ifl_sds.ifsd_map[i]);
1896 			if (err != 0) {
1897 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1898 				goto fail;
1899 			}
1900 		}
1901 	}
1902 	return (0);
1903 
1904 fail:
1905 	iflib_rx_structures_free(ctx);
1906 	return (err);
1907 }
1908 
1909 
1910 /*
1911  * Internal service routines
1912  */
1913 
1914 struct rxq_refill_cb_arg {
1915 	int               error;
1916 	bus_dma_segment_t seg;
1917 	int               nseg;
1918 };
1919 
1920 static void
1921 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1922 {
1923 	struct rxq_refill_cb_arg *cb_arg = arg;
1924 
1925 	cb_arg->error = error;
1926 	cb_arg->seg = segs[0];
1927 	cb_arg->nseg = nseg;
1928 }
1929 
1930 /**
1931  *	rxq_refill - refill an rxq  free-buffer list
1932  *	@ctx: the iflib context
1933  *	@rxq: the free-list to refill
1934  *	@n: the number of new buffers to allocate
1935  *
1936  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1937  *	The caller must assure that @n does not exceed the queue's capacity.
1938  */
1939 static void
1940 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1941 {
1942 	struct if_rxd_update iru;
1943 	struct rxq_refill_cb_arg cb_arg;
1944 	struct mbuf *m;
1945 	caddr_t cl, *sd_cl;
1946 	struct mbuf **sd_m;
1947 	bus_dmamap_t *sd_map;
1948 	bus_addr_t bus_addr, *sd_ba;
1949 	int err, frag_idx, i, idx, n, pidx;
1950 	qidx_t credits;
1951 
1952 	sd_m = fl->ifl_sds.ifsd_m;
1953 	sd_map = fl->ifl_sds.ifsd_map;
1954 	sd_cl = fl->ifl_sds.ifsd_cl;
1955 	sd_ba = fl->ifl_sds.ifsd_ba;
1956 	pidx = fl->ifl_pidx;
1957 	idx = pidx;
1958 	frag_idx = fl->ifl_fragidx;
1959 	credits = fl->ifl_credits;
1960 
1961 	i = 0;
1962 	n = count;
1963 	MPASS(n > 0);
1964 	MPASS(credits + n <= fl->ifl_size);
1965 
1966 	if (pidx < fl->ifl_cidx)
1967 		MPASS(pidx + n <= fl->ifl_cidx);
1968 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1969 		MPASS(fl->ifl_gen == 0);
1970 	if (pidx > fl->ifl_cidx)
1971 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1972 
1973 	DBG_COUNTER_INC(fl_refills);
1974 	if (n > 8)
1975 		DBG_COUNTER_INC(fl_refills_large);
1976 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1977 	while (n--) {
1978 		/*
1979 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1980 		 * initialized after rx.
1981 		 *
1982 		 * If the cluster is still set then we know a minimum sized packet was received
1983 		 */
1984 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1985 		    &frag_idx);
1986 		if (frag_idx < 0)
1987 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1988 		MPASS(frag_idx >= 0);
1989 		if ((cl = sd_cl[frag_idx]) == NULL) {
1990 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1991 				break;
1992 
1993 			cb_arg.error = 0;
1994 			MPASS(sd_map != NULL);
1995 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1996 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1997 			    BUS_DMA_NOWAIT);
1998 			if (err != 0 || cb_arg.error) {
1999 				/*
2000 				 * !zone_pack ?
2001 				 */
2002 				if (fl->ifl_zone == zone_pack)
2003 					uma_zfree(fl->ifl_zone, cl);
2004 				break;
2005 			}
2006 
2007 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2008 			sd_cl[frag_idx] = cl;
2009 #if MEMORY_LOGGING
2010 			fl->ifl_cl_enqueued++;
2011 #endif
2012 		} else {
2013 			bus_addr = sd_ba[frag_idx];
2014 		}
2015 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2016 		    BUS_DMASYNC_PREREAD);
2017 
2018 		if (sd_m[frag_idx] == NULL) {
2019 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2020 				break;
2021 			}
2022 			sd_m[frag_idx] = m;
2023 		}
2024 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2025 #if MEMORY_LOGGING
2026 		fl->ifl_m_enqueued++;
2027 #endif
2028 
2029 		DBG_COUNTER_INC(rx_allocs);
2030 		fl->ifl_rxd_idxs[i] = frag_idx;
2031 		fl->ifl_bus_addrs[i] = bus_addr;
2032 		fl->ifl_vm_addrs[i] = cl;
2033 		credits++;
2034 		i++;
2035 		MPASS(credits <= fl->ifl_size);
2036 		if (++idx == fl->ifl_size) {
2037 			fl->ifl_gen = 1;
2038 			idx = 0;
2039 		}
2040 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2041 			iru.iru_pidx = pidx;
2042 			iru.iru_count = i;
2043 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2044 			i = 0;
2045 			pidx = idx;
2046 			fl->ifl_pidx = idx;
2047 			fl->ifl_credits = credits;
2048 		}
2049 	}
2050 
2051 	if (i) {
2052 		iru.iru_pidx = pidx;
2053 		iru.iru_count = i;
2054 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2055 		fl->ifl_pidx = idx;
2056 		fl->ifl_credits = credits;
2057 	}
2058 	DBG_COUNTER_INC(rxd_flush);
2059 	if (fl->ifl_pidx == 0)
2060 		pidx = fl->ifl_size - 1;
2061 	else
2062 		pidx = fl->ifl_pidx - 1;
2063 
2064 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2065 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2066 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2067 	fl->ifl_fragidx = frag_idx;
2068 }
2069 
2070 static __inline void
2071 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2072 {
2073 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2074 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2075 #ifdef INVARIANTS
2076 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2077 #endif
2078 
2079 	MPASS(fl->ifl_credits <= fl->ifl_size);
2080 	MPASS(reclaimable == delta);
2081 
2082 	if (reclaimable > 0)
2083 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2084 }
2085 
2086 uint8_t
2087 iflib_in_detach(if_ctx_t ctx)
2088 {
2089 	bool in_detach;
2090 	STATE_LOCK(ctx);
2091 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2092 	STATE_UNLOCK(ctx);
2093 	return (in_detach);
2094 }
2095 
2096 static void
2097 iflib_fl_bufs_free(iflib_fl_t fl)
2098 {
2099 	iflib_dma_info_t idi = fl->ifl_ifdi;
2100 	bus_dmamap_t sd_map;
2101 	uint32_t i;
2102 
2103 	for (i = 0; i < fl->ifl_size; i++) {
2104 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2105 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2106 
2107 		if (*sd_cl != NULL) {
2108 			sd_map = fl->ifl_sds.ifsd_map[i];
2109 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2110 			    BUS_DMASYNC_POSTREAD);
2111 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2112 			if (*sd_cl != NULL)
2113 				uma_zfree(fl->ifl_zone, *sd_cl);
2114 			// XXX: Should this get moved out?
2115 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2116 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2117 			if (*sd_m != NULL) {
2118 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2119 				uma_zfree(zone_mbuf, *sd_m);
2120 			}
2121 		} else {
2122 			MPASS(*sd_cl == NULL);
2123 			MPASS(*sd_m == NULL);
2124 		}
2125 #if MEMORY_LOGGING
2126 		fl->ifl_m_dequeued++;
2127 		fl->ifl_cl_dequeued++;
2128 #endif
2129 		*sd_cl = NULL;
2130 		*sd_m = NULL;
2131 	}
2132 #ifdef INVARIANTS
2133 	for (i = 0; i < fl->ifl_size; i++) {
2134 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2135 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2136 	}
2137 #endif
2138 	/*
2139 	 * Reset free list values
2140 	 */
2141 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2142 	bzero(idi->idi_vaddr, idi->idi_size);
2143 }
2144 
2145 /*********************************************************************
2146  *
2147  *  Initialize a receive ring and its buffers.
2148  *
2149  **********************************************************************/
2150 static int
2151 iflib_fl_setup(iflib_fl_t fl)
2152 {
2153 	iflib_rxq_t rxq = fl->ifl_rxq;
2154 	if_ctx_t ctx = rxq->ifr_ctx;
2155 
2156 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2157 	/*
2158 	** Free current RX buffer structs and their mbufs
2159 	*/
2160 	iflib_fl_bufs_free(fl);
2161 	/* Now replenish the mbufs */
2162 	MPASS(fl->ifl_credits == 0);
2163 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2164 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2165 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2166 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2167 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2168 
2169 
2170 	/* avoid pre-allocating zillions of clusters to an idle card
2171 	 * potentially speeding up attach
2172 	 */
2173 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2174 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2175 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2176 		return (ENOBUFS);
2177 	/*
2178 	 * handle failure
2179 	 */
2180 	MPASS(rxq != NULL);
2181 	MPASS(fl->ifl_ifdi != NULL);
2182 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2183 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2184 	return (0);
2185 }
2186 
2187 /*********************************************************************
2188  *
2189  *  Free receive ring data structures
2190  *
2191  **********************************************************************/
2192 static void
2193 iflib_rx_sds_free(iflib_rxq_t rxq)
2194 {
2195 	iflib_fl_t fl;
2196 	int i, j;
2197 
2198 	if (rxq->ifr_fl != NULL) {
2199 		for (i = 0; i < rxq->ifr_nfl; i++) {
2200 			fl = &rxq->ifr_fl[i];
2201 			if (fl->ifl_buf_tag != NULL) {
2202 				if (fl->ifl_sds.ifsd_map != NULL) {
2203 					for (j = 0; j < fl->ifl_size; j++) {
2204 						if (fl->ifl_sds.ifsd_map[j] ==
2205 						    NULL)
2206 							continue;
2207 						bus_dmamap_sync(
2208 						    fl->ifl_buf_tag,
2209 						    fl->ifl_sds.ifsd_map[j],
2210 						    BUS_DMASYNC_POSTREAD);
2211 						bus_dmamap_unload(
2212 						    fl->ifl_buf_tag,
2213 						    fl->ifl_sds.ifsd_map[j]);
2214 					}
2215 				}
2216 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2217 				fl->ifl_buf_tag = NULL;
2218 			}
2219 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2220 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2221 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2222 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2223 			fl->ifl_sds.ifsd_m = NULL;
2224 			fl->ifl_sds.ifsd_cl = NULL;
2225 			fl->ifl_sds.ifsd_ba = NULL;
2226 			fl->ifl_sds.ifsd_map = NULL;
2227 		}
2228 		free(rxq->ifr_fl, M_IFLIB);
2229 		rxq->ifr_fl = NULL;
2230 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2231 	}
2232 }
2233 
2234 /*
2235  * MI independent logic
2236  *
2237  */
2238 static void
2239 iflib_timer(void *arg)
2240 {
2241 	iflib_txq_t txq = arg;
2242 	if_ctx_t ctx = txq->ift_ctx;
2243 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2244 	uint64_t this_tick = ticks;
2245 	uint32_t reset_on = hz / 2;
2246 
2247 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2248 		return;
2249 	/*
2250 	** Check on the state of the TX queue(s), this
2251 	** can be done without the lock because its RO
2252 	** and the HUNG state will be static if set.
2253 	*/
2254 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2255 		txq->ift_last_timer_tick = this_tick;
2256 		IFDI_TIMER(ctx, txq->ift_id);
2257 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2258 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2259 		     (sctx->isc_pause_frames == 0)))
2260 			goto hung;
2261 
2262 		if (ifmp_ring_is_stalled(txq->ift_br))
2263 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2264 		txq->ift_cleaned_prev = txq->ift_cleaned;
2265 	}
2266 #ifdef DEV_NETMAP
2267 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2268 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2269 #endif
2270 	/* handle any laggards */
2271 	if (txq->ift_db_pending)
2272 		GROUPTASK_ENQUEUE(&txq->ift_task);
2273 
2274 	sctx->isc_pause_frames = 0;
2275 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2276 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2277 	return;
2278  hung:
2279 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2280 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2281 	STATE_LOCK(ctx);
2282 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2283 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2284 	iflib_admin_intr_deferred(ctx);
2285 	STATE_UNLOCK(ctx);
2286 }
2287 
2288 static void
2289 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2290 {
2291 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2292 
2293 	/*
2294 	 * XXX don't set the max_frame_size to larger
2295 	 * than the hardware can handle
2296 	 */
2297 	if (sctx->isc_max_frame_size <= MCLBYTES)
2298 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2299 	else
2300 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2301 }
2302 
2303 uint32_t
2304 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2305 {
2306 	return (ctx->ifc_rx_mbuf_sz);
2307 }
2308 
2309 static void
2310 iflib_init_locked(if_ctx_t ctx)
2311 {
2312 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2313 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2314 	if_t ifp = ctx->ifc_ifp;
2315 	iflib_fl_t fl;
2316 	iflib_txq_t txq;
2317 	iflib_rxq_t rxq;
2318 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2319 
2320 
2321 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2322 	IFDI_INTR_DISABLE(ctx);
2323 
2324 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2325 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2326 	/* Set hardware offload abilities */
2327 	if_clearhwassist(ifp);
2328 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2329 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2330 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2331 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2332 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2333 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2334 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2335 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2336 
2337 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2338 		CALLOUT_LOCK(txq);
2339 		callout_stop(&txq->ift_timer);
2340 		CALLOUT_UNLOCK(txq);
2341 		iflib_netmap_txq_init(ctx, txq);
2342 	}
2343 
2344 	/*
2345 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2346 	 * that drivers can use the value when setting up the hardware receive
2347 	 * buffers.
2348 	 */
2349 	iflib_calc_rx_mbuf_sz(ctx);
2350 
2351 #ifdef INVARIANTS
2352 	i = if_getdrvflags(ifp);
2353 #endif
2354 	IFDI_INIT(ctx);
2355 	MPASS(if_getdrvflags(ifp) == i);
2356 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2357 		/* XXX this should really be done on a per-queue basis */
2358 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2359 			MPASS(rxq->ifr_id == i);
2360 			iflib_netmap_rxq_init(ctx, rxq);
2361 			continue;
2362 		}
2363 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2364 			if (iflib_fl_setup(fl)) {
2365 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2366 				goto done;
2367 			}
2368 		}
2369 	}
2370 done:
2371 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2372 	IFDI_INTR_ENABLE(ctx);
2373 	txq = ctx->ifc_txqs;
2374 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2375 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2376 			txq->ift_timer.c_cpu);
2377 }
2378 
2379 static int
2380 iflib_media_change(if_t ifp)
2381 {
2382 	if_ctx_t ctx = if_getsoftc(ifp);
2383 	int err;
2384 
2385 	CTX_LOCK(ctx);
2386 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2387 		iflib_init_locked(ctx);
2388 	CTX_UNLOCK(ctx);
2389 	return (err);
2390 }
2391 
2392 static void
2393 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2394 {
2395 	if_ctx_t ctx = if_getsoftc(ifp);
2396 
2397 	CTX_LOCK(ctx);
2398 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2399 	IFDI_MEDIA_STATUS(ctx, ifmr);
2400 	CTX_UNLOCK(ctx);
2401 }
2402 
2403 void
2404 iflib_stop(if_ctx_t ctx)
2405 {
2406 	iflib_txq_t txq = ctx->ifc_txqs;
2407 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2408 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2409 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2410 	iflib_dma_info_t di;
2411 	iflib_fl_t fl;
2412 	int i, j;
2413 
2414 	/* Tell the stack that the interface is no longer active */
2415 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2416 
2417 	IFDI_INTR_DISABLE(ctx);
2418 	DELAY(1000);
2419 	IFDI_STOP(ctx);
2420 	DELAY(1000);
2421 
2422 	iflib_debug_reset();
2423 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2424 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2425 		/* make sure all transmitters have completed before proceeding XXX */
2426 
2427 		CALLOUT_LOCK(txq);
2428 		callout_stop(&txq->ift_timer);
2429 		CALLOUT_UNLOCK(txq);
2430 
2431 		/* clean any enqueued buffers */
2432 		iflib_ifmp_purge(txq);
2433 		/* Free any existing tx buffers. */
2434 		for (j = 0; j < txq->ift_size; j++) {
2435 			iflib_txsd_free(ctx, txq, j);
2436 		}
2437 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2438 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2439 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2440 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2441 		txq->ift_pullups = 0;
2442 		ifmp_ring_reset_stats(txq->ift_br);
2443 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2444 			bzero((void *)di->idi_vaddr, di->idi_size);
2445 	}
2446 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2447 		/* make sure all transmitters have completed before proceeding XXX */
2448 
2449 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2450 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2451 			bzero((void *)di->idi_vaddr, di->idi_size);
2452 		/* also resets the free lists pidx/cidx */
2453 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2454 			iflib_fl_bufs_free(fl);
2455 	}
2456 }
2457 
2458 static inline caddr_t
2459 calc_next_rxd(iflib_fl_t fl, int cidx)
2460 {
2461 	qidx_t size;
2462 	int nrxd;
2463 	caddr_t start, end, cur, next;
2464 
2465 	nrxd = fl->ifl_size;
2466 	size = fl->ifl_rxd_size;
2467 	start = fl->ifl_ifdi->idi_vaddr;
2468 
2469 	if (__predict_false(size == 0))
2470 		return (start);
2471 	cur = start + size*cidx;
2472 	end = start + size*nrxd;
2473 	next = CACHE_PTR_NEXT(cur);
2474 	return (next < end ? next : start);
2475 }
2476 
2477 static inline void
2478 prefetch_pkts(iflib_fl_t fl, int cidx)
2479 {
2480 	int nextptr;
2481 	int nrxd = fl->ifl_size;
2482 	caddr_t next_rxd;
2483 
2484 
2485 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2486 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2487 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2488 	next_rxd = calc_next_rxd(fl, cidx);
2489 	prefetch(next_rxd);
2490 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2491 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2492 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2493 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2494 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2495 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2496 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2497 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2498 }
2499 
2500 static struct mbuf *
2501 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2502     int *pf_rv, if_rxd_info_t ri)
2503 {
2504 	bus_dmamap_t map;
2505 	iflib_fl_t fl;
2506 	caddr_t payload;
2507 	struct mbuf *m;
2508 	int flid, cidx, len, next;
2509 
2510 	map = NULL;
2511 	flid = irf->irf_flid;
2512 	cidx = irf->irf_idx;
2513 	fl = &rxq->ifr_fl[flid];
2514 	sd->ifsd_fl = fl;
2515 	sd->ifsd_cidx = cidx;
2516 	m = fl->ifl_sds.ifsd_m[cidx];
2517 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2518 	fl->ifl_credits--;
2519 #if MEMORY_LOGGING
2520 	fl->ifl_m_dequeued++;
2521 #endif
2522 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2523 		prefetch_pkts(fl, cidx);
2524 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2525 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2526 	map = fl->ifl_sds.ifsd_map[cidx];
2527 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2528 
2529 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2530 	MPASS(fl->ifl_cidx == cidx);
2531 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2532 
2533 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2534 		payload  = *sd->ifsd_cl;
2535 		payload +=  ri->iri_pad;
2536 		len = ri->iri_len - ri->iri_pad;
2537 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2538 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2539 		switch (*pf_rv) {
2540 		case PFIL_DROPPED:
2541 		case PFIL_CONSUMED:
2542 			/*
2543 			 * The filter ate it.  Everything is recycled.
2544 			 */
2545 			m = NULL;
2546 			unload = 0;
2547 			break;
2548 		case PFIL_REALLOCED:
2549 			/*
2550 			 * The filter copied it.  Everything is recycled.
2551 			 */
2552 			m = pfil_mem2mbuf(payload);
2553 			unload = 0;
2554 			break;
2555 		case PFIL_PASS:
2556 			/*
2557 			 * Filter said it was OK, so receive like
2558 			 * normal
2559 			 */
2560 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2561 			break;
2562 		default:
2563 			MPASS(0);
2564 		}
2565 	} else {
2566 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2567 		*pf_rv = PFIL_PASS;
2568 	}
2569 
2570 	if (unload)
2571 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2572 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2573 	if (__predict_false(fl->ifl_cidx == 0))
2574 		fl->ifl_gen = 0;
2575 	bit_clear(fl->ifl_rx_bitmap, cidx);
2576 	return (m);
2577 }
2578 
2579 static struct mbuf *
2580 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2581 {
2582 	struct mbuf *m, *mh, *mt;
2583 	caddr_t cl;
2584 	int  *pf_rv_ptr, flags, i, padlen;
2585 	bool consumed;
2586 
2587 	i = 0;
2588 	mh = NULL;
2589 	consumed = false;
2590 	*pf_rv = PFIL_PASS;
2591 	pf_rv_ptr = pf_rv;
2592 	do {
2593 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2594 		    pf_rv_ptr, ri);
2595 
2596 		MPASS(*sd->ifsd_cl != NULL);
2597 
2598 		/*
2599 		 * Exclude zero-length frags & frags from
2600 		 * packets the filter has consumed or dropped
2601 		 */
2602 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2603 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2604 			if (mh == NULL) {
2605 				/* everything saved here */
2606 				consumed = true;
2607 				pf_rv_ptr = NULL;
2608 				continue;
2609 			}
2610 			/* XXX we can save the cluster here, but not the mbuf */
2611 			m_init(m, M_NOWAIT, MT_DATA, 0);
2612 			m_free(m);
2613 			continue;
2614 		}
2615 		if (mh == NULL) {
2616 			flags = M_PKTHDR|M_EXT;
2617 			mh = mt = m;
2618 			padlen = ri->iri_pad;
2619 		} else {
2620 			flags = M_EXT;
2621 			mt->m_next = m;
2622 			mt = m;
2623 			/* assuming padding is only on the first fragment */
2624 			padlen = 0;
2625 		}
2626 		cl = *sd->ifsd_cl;
2627 		*sd->ifsd_cl = NULL;
2628 
2629 		/* Can these two be made one ? */
2630 		m_init(m, M_NOWAIT, MT_DATA, flags);
2631 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2632 		/*
2633 		 * These must follow m_init and m_cljset
2634 		 */
2635 		m->m_data += padlen;
2636 		ri->iri_len -= padlen;
2637 		m->m_len = ri->iri_frags[i].irf_len;
2638 	} while (++i < ri->iri_nfrags);
2639 
2640 	return (mh);
2641 }
2642 
2643 /*
2644  * Process one software descriptor
2645  */
2646 static struct mbuf *
2647 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2648 {
2649 	struct if_rxsd sd;
2650 	struct mbuf *m;
2651 	int pf_rv;
2652 
2653 	/* should I merge this back in now that the two paths are basically duplicated? */
2654 	if (ri->iri_nfrags == 1 &&
2655 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2656 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2657 		    &pf_rv, ri);
2658 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2659 			return (m);
2660 		if (pf_rv == PFIL_PASS) {
2661 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2662 #ifndef __NO_STRICT_ALIGNMENT
2663 			if (!IP_ALIGNED(m))
2664 				m->m_data += 2;
2665 #endif
2666 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2667 			m->m_len = ri->iri_frags[0].irf_len;
2668 		}
2669 	} else {
2670 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2671 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2672 			return (m);
2673 	}
2674 	m->m_pkthdr.len = ri->iri_len;
2675 	m->m_pkthdr.rcvif = ri->iri_ifp;
2676 	m->m_flags |= ri->iri_flags;
2677 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2678 	m->m_pkthdr.flowid = ri->iri_flowid;
2679 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2680 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2681 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2682 	return (m);
2683 }
2684 
2685 #if defined(INET6) || defined(INET)
2686 static void
2687 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2688 {
2689 	CURVNET_SET(lc->ifp->if_vnet);
2690 #if defined(INET6)
2691 	*v6 = VNET(ip6_forwarding);
2692 #endif
2693 #if defined(INET)
2694 	*v4 = VNET(ipforwarding);
2695 #endif
2696 	CURVNET_RESTORE();
2697 }
2698 
2699 /*
2700  * Returns true if it's possible this packet could be LROed.
2701  * if it returns false, it is guaranteed that tcp_lro_rx()
2702  * would not return zero.
2703  */
2704 static bool
2705 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2706 {
2707 	struct ether_header *eh;
2708 	uint16_t eh_type;
2709 
2710 	eh = mtod(m, struct ether_header *);
2711 	eh_type = ntohs(eh->ether_type);
2712 	switch (eh_type) {
2713 #if defined(INET6)
2714 		case ETHERTYPE_IPV6:
2715 			return !v6_forwarding;
2716 #endif
2717 #if defined (INET)
2718 		case ETHERTYPE_IP:
2719 			return !v4_forwarding;
2720 #endif
2721 	}
2722 
2723 	return false;
2724 }
2725 #else
2726 static void
2727 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2728 {
2729 }
2730 #endif
2731 
2732 static bool
2733 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2734 {
2735 	if_ctx_t ctx = rxq->ifr_ctx;
2736 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2737 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2738 	int avail, i;
2739 	qidx_t *cidxp;
2740 	struct if_rxd_info ri;
2741 	int err, budget_left, rx_bytes, rx_pkts;
2742 	iflib_fl_t fl;
2743 	struct ifnet *ifp;
2744 	int lro_enabled;
2745 	bool v4_forwarding, v6_forwarding, lro_possible;
2746 
2747 	/*
2748 	 * XXX early demux data packets so that if_input processing only handles
2749 	 * acks in interrupt context
2750 	 */
2751 	struct mbuf *m, *mh, *mt, *mf;
2752 
2753 	lro_possible = v4_forwarding = v6_forwarding = false;
2754 	ifp = ctx->ifc_ifp;
2755 	mh = mt = NULL;
2756 	MPASS(budget > 0);
2757 	rx_pkts	= rx_bytes = 0;
2758 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2759 		cidxp = &rxq->ifr_cq_cidx;
2760 	else
2761 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2762 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2763 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2764 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2765 		DBG_COUNTER_INC(rx_unavail);
2766 		return (false);
2767 	}
2768 
2769 	/* pfil needs the vnet to be set */
2770 	CURVNET_SET_QUIET(ifp->if_vnet);
2771 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2772 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2773 			DBG_COUNTER_INC(rx_ctx_inactive);
2774 			break;
2775 		}
2776 		/*
2777 		 * Reset client set fields to their default values
2778 		 */
2779 		rxd_info_zero(&ri);
2780 		ri.iri_qsidx = rxq->ifr_id;
2781 		ri.iri_cidx = *cidxp;
2782 		ri.iri_ifp = ifp;
2783 		ri.iri_frags = rxq->ifr_frags;
2784 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2785 
2786 		if (err)
2787 			goto err;
2788 		rx_pkts += 1;
2789 		rx_bytes += ri.iri_len;
2790 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2791 			*cidxp = ri.iri_cidx;
2792 			/* Update our consumer index */
2793 			/* XXX NB: shurd - check if this is still safe */
2794 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2795 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2796 				rxq->ifr_cq_gen = 0;
2797 			}
2798 			/* was this only a completion queue message? */
2799 			if (__predict_false(ri.iri_nfrags == 0))
2800 				continue;
2801 		}
2802 		MPASS(ri.iri_nfrags != 0);
2803 		MPASS(ri.iri_len != 0);
2804 
2805 		/* will advance the cidx on the corresponding free lists */
2806 		m = iflib_rxd_pkt_get(rxq, &ri);
2807 		avail--;
2808 		budget_left--;
2809 		if (avail == 0 && budget_left)
2810 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2811 
2812 		if (__predict_false(m == NULL))
2813 			continue;
2814 
2815 		/* imm_pkt: -- cxgb */
2816 		if (mh == NULL)
2817 			mh = mt = m;
2818 		else {
2819 			mt->m_nextpkt = m;
2820 			mt = m;
2821 		}
2822 	}
2823 	CURVNET_RESTORE();
2824 	/* make sure that we can refill faster than drain */
2825 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2826 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2827 
2828 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2829 	if (lro_enabled)
2830 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2831 	mt = mf = NULL;
2832 	while (mh != NULL) {
2833 		m = mh;
2834 		mh = mh->m_nextpkt;
2835 		m->m_nextpkt = NULL;
2836 #ifndef __NO_STRICT_ALIGNMENT
2837 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2838 			continue;
2839 #endif
2840 		rx_bytes += m->m_pkthdr.len;
2841 		rx_pkts++;
2842 #if defined(INET6) || defined(INET)
2843 		if (lro_enabled) {
2844 			if (!lro_possible) {
2845 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2846 				if (lro_possible && mf != NULL) {
2847 					ifp->if_input(ifp, mf);
2848 					DBG_COUNTER_INC(rx_if_input);
2849 					mt = mf = NULL;
2850 				}
2851 			}
2852 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2853 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2854 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2855 					continue;
2856 			}
2857 		}
2858 #endif
2859 		if (lro_possible) {
2860 			ifp->if_input(ifp, m);
2861 			DBG_COUNTER_INC(rx_if_input);
2862 			continue;
2863 		}
2864 
2865 		if (mf == NULL)
2866 			mf = m;
2867 		if (mt != NULL)
2868 			mt->m_nextpkt = m;
2869 		mt = m;
2870 	}
2871 	if (mf != NULL) {
2872 		ifp->if_input(ifp, mf);
2873 		DBG_COUNTER_INC(rx_if_input);
2874 	}
2875 
2876 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2877 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2878 
2879 	/*
2880 	 * Flush any outstanding LRO work
2881 	 */
2882 #if defined(INET6) || defined(INET)
2883 	tcp_lro_flush_all(&rxq->ifr_lc);
2884 #endif
2885 	if (avail)
2886 		return true;
2887 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2888 err:
2889 	STATE_LOCK(ctx);
2890 	ctx->ifc_flags |= IFC_DO_RESET;
2891 	iflib_admin_intr_deferred(ctx);
2892 	STATE_UNLOCK(ctx);
2893 	return (false);
2894 }
2895 
2896 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2897 static inline qidx_t
2898 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2899 {
2900 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2901 	qidx_t minthresh = txq->ift_size / 8;
2902 	if (in_use > 4*minthresh)
2903 		return (notify_count);
2904 	if (in_use > 2*minthresh)
2905 		return (notify_count >> 1);
2906 	if (in_use > minthresh)
2907 		return (notify_count >> 3);
2908 	return (0);
2909 }
2910 
2911 static inline qidx_t
2912 txq_max_rs_deferred(iflib_txq_t txq)
2913 {
2914 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2915 	qidx_t minthresh = txq->ift_size / 8;
2916 	if (txq->ift_in_use > 4*minthresh)
2917 		return (notify_count);
2918 	if (txq->ift_in_use > 2*minthresh)
2919 		return (notify_count >> 1);
2920 	if (txq->ift_in_use > minthresh)
2921 		return (notify_count >> 2);
2922 	return (2);
2923 }
2924 
2925 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2926 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2927 
2928 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2929 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2930 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2931 
2932 /* forward compatibility for cxgb */
2933 #define FIRST_QSET(ctx) 0
2934 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2935 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2936 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2937 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2938 
2939 /* XXX we should be setting this to something other than zero */
2940 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2941 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2942     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2943 
2944 static inline bool
2945 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2946 {
2947 	qidx_t dbval, max;
2948 	bool rang;
2949 
2950 	rang = false;
2951 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2952 	if (ring || txq->ift_db_pending >= max) {
2953 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2954 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2955 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2956 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2957 		txq->ift_db_pending = txq->ift_npending = 0;
2958 		rang = true;
2959 	}
2960 	return (rang);
2961 }
2962 
2963 #ifdef PKT_DEBUG
2964 static void
2965 print_pkt(if_pkt_info_t pi)
2966 {
2967 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2968 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2969 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2970 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2971 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2972 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2973 }
2974 #endif
2975 
2976 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2977 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2978 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2979 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2980 
2981 static int
2982 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2983 {
2984 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2985 	struct ether_vlan_header *eh;
2986 	struct mbuf *m;
2987 
2988 	m = *mp;
2989 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2990 	    M_WRITABLE(m) == 0) {
2991 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2992 			return (ENOMEM);
2993 		} else {
2994 			m_freem(*mp);
2995 			DBG_COUNTER_INC(tx_frees);
2996 			*mp = m;
2997 		}
2998 	}
2999 
3000 	/*
3001 	 * Determine where frame payload starts.
3002 	 * Jump over vlan headers if already present,
3003 	 * helpful for QinQ too.
3004 	 */
3005 	if (__predict_false(m->m_len < sizeof(*eh))) {
3006 		txq->ift_pullups++;
3007 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3008 			return (ENOMEM);
3009 	}
3010 	eh = mtod(m, struct ether_vlan_header *);
3011 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3012 		pi->ipi_etype = ntohs(eh->evl_proto);
3013 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3014 	} else {
3015 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3016 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3017 	}
3018 
3019 	switch (pi->ipi_etype) {
3020 #ifdef INET
3021 	case ETHERTYPE_IP:
3022 	{
3023 		struct mbuf *n;
3024 		struct ip *ip = NULL;
3025 		struct tcphdr *th = NULL;
3026 		int minthlen;
3027 
3028 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3029 		if (__predict_false(m->m_len < minthlen)) {
3030 			/*
3031 			 * if this code bloat is causing too much of a hit
3032 			 * move it to a separate function and mark it noinline
3033 			 */
3034 			if (m->m_len == pi->ipi_ehdrlen) {
3035 				n = m->m_next;
3036 				MPASS(n);
3037 				if (n->m_len >= sizeof(*ip))  {
3038 					ip = (struct ip *)n->m_data;
3039 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3040 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3041 				} else {
3042 					txq->ift_pullups++;
3043 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3044 						return (ENOMEM);
3045 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3046 				}
3047 			} else {
3048 				txq->ift_pullups++;
3049 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3050 					return (ENOMEM);
3051 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3052 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3053 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3054 			}
3055 		} else {
3056 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3057 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3058 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3059 		}
3060 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3061 		pi->ipi_ipproto = ip->ip_p;
3062 		pi->ipi_flags |= IPI_TX_IPV4;
3063 
3064 		/* TCP checksum offload may require TCP header length */
3065 		if (IS_TX_OFFLOAD4(pi)) {
3066 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3067 				if (__predict_false(th == NULL)) {
3068 					txq->ift_pullups++;
3069 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3070 						return (ENOMEM);
3071 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3072 				}
3073 				pi->ipi_tcp_hflags = th->th_flags;
3074 				pi->ipi_tcp_hlen = th->th_off << 2;
3075 				pi->ipi_tcp_seq = th->th_seq;
3076 			}
3077 			if (IS_TSO4(pi)) {
3078 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3079 					return (ENXIO);
3080 				/*
3081 				 * TSO always requires hardware checksum offload.
3082 				 */
3083 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3084 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3085 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3086 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3087 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3088 					ip->ip_sum = 0;
3089 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3090 				}
3091 			}
3092 		}
3093 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3094                        ip->ip_sum = 0;
3095 
3096 		break;
3097 	}
3098 #endif
3099 #ifdef INET6
3100 	case ETHERTYPE_IPV6:
3101 	{
3102 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3103 		struct tcphdr *th;
3104 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3105 
3106 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3107 			txq->ift_pullups++;
3108 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3109 				return (ENOMEM);
3110 		}
3111 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3112 
3113 		/* XXX-BZ this will go badly in case of ext hdrs. */
3114 		pi->ipi_ipproto = ip6->ip6_nxt;
3115 		pi->ipi_flags |= IPI_TX_IPV6;
3116 
3117 		/* TCP checksum offload may require TCP header length */
3118 		if (IS_TX_OFFLOAD6(pi)) {
3119 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3120 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3121 					txq->ift_pullups++;
3122 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3123 						return (ENOMEM);
3124 				}
3125 				pi->ipi_tcp_hflags = th->th_flags;
3126 				pi->ipi_tcp_hlen = th->th_off << 2;
3127 				pi->ipi_tcp_seq = th->th_seq;
3128 			}
3129 			if (IS_TSO6(pi)) {
3130 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3131 					return (ENXIO);
3132 				/*
3133 				 * TSO always requires hardware checksum offload.
3134 				 */
3135 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3136 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3137 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3138 			}
3139 		}
3140 		break;
3141 	}
3142 #endif
3143 	default:
3144 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3145 		pi->ipi_ip_hlen = 0;
3146 		break;
3147 	}
3148 	*mp = m;
3149 
3150 	return (0);
3151 }
3152 
3153 /*
3154  * If dodgy hardware rejects the scatter gather chain we've handed it
3155  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3156  * m_defrag'd mbufs
3157  */
3158 static __noinline struct mbuf *
3159 iflib_remove_mbuf(iflib_txq_t txq)
3160 {
3161 	int ntxd, pidx;
3162 	struct mbuf *m, **ifsd_m;
3163 
3164 	ifsd_m = txq->ift_sds.ifsd_m;
3165 	ntxd = txq->ift_size;
3166 	pidx = txq->ift_pidx & (ntxd - 1);
3167 	ifsd_m = txq->ift_sds.ifsd_m;
3168 	m = ifsd_m[pidx];
3169 	ifsd_m[pidx] = NULL;
3170 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3171 	if (txq->ift_sds.ifsd_tso_map != NULL)
3172 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3173 		    txq->ift_sds.ifsd_tso_map[pidx]);
3174 #if MEMORY_LOGGING
3175 	txq->ift_dequeued++;
3176 #endif
3177 	return (m);
3178 }
3179 
3180 static inline caddr_t
3181 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3182 {
3183 	qidx_t size;
3184 	int ntxd;
3185 	caddr_t start, end, cur, next;
3186 
3187 	ntxd = txq->ift_size;
3188 	size = txq->ift_txd_size[qid];
3189 	start = txq->ift_ifdi[qid].idi_vaddr;
3190 
3191 	if (__predict_false(size == 0))
3192 		return (start);
3193 	cur = start + size*cidx;
3194 	end = start + size*ntxd;
3195 	next = CACHE_PTR_NEXT(cur);
3196 	return (next < end ? next : start);
3197 }
3198 
3199 /*
3200  * Pad an mbuf to ensure a minimum ethernet frame size.
3201  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3202  */
3203 static __noinline int
3204 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3205 {
3206 	/*
3207 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3208 	 * and ARP message is the smallest common payload I can think of
3209 	 */
3210 	static char pad[18];	/* just zeros */
3211 	int n;
3212 	struct mbuf *new_head;
3213 
3214 	if (!M_WRITABLE(*m_head)) {
3215 		new_head = m_dup(*m_head, M_NOWAIT);
3216 		if (new_head == NULL) {
3217 			m_freem(*m_head);
3218 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3219 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3220 			DBG_COUNTER_INC(tx_frees);
3221 			return ENOMEM;
3222 		}
3223 		m_freem(*m_head);
3224 		*m_head = new_head;
3225 	}
3226 
3227 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3228 	     n > 0; n -= sizeof(pad))
3229 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3230 			break;
3231 
3232 	if (n > 0) {
3233 		m_freem(*m_head);
3234 		device_printf(dev, "cannot pad short frame\n");
3235 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3236 		DBG_COUNTER_INC(tx_frees);
3237 		return (ENOBUFS);
3238 	}
3239 
3240 	return 0;
3241 }
3242 
3243 static int
3244 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3245 {
3246 	if_ctx_t		ctx;
3247 	if_shared_ctx_t		sctx;
3248 	if_softc_ctx_t		scctx;
3249 	bus_dma_tag_t		buf_tag;
3250 	bus_dma_segment_t	*segs;
3251 	struct mbuf		*m_head, **ifsd_m;
3252 	void			*next_txd;
3253 	bus_dmamap_t		map;
3254 	struct if_pkt_info	pi;
3255 	int remap = 0;
3256 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3257 
3258 	ctx = txq->ift_ctx;
3259 	sctx = ctx->ifc_sctx;
3260 	scctx = &ctx->ifc_softc_ctx;
3261 	segs = txq->ift_segs;
3262 	ntxd = txq->ift_size;
3263 	m_head = *m_headp;
3264 	map = NULL;
3265 
3266 	/*
3267 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3268 	 */
3269 	cidx = txq->ift_cidx;
3270 	pidx = txq->ift_pidx;
3271 	if (ctx->ifc_flags & IFC_PREFETCH) {
3272 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3273 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3274 			next_txd = calc_next_txd(txq, cidx, 0);
3275 			prefetch(next_txd);
3276 		}
3277 
3278 		/* prefetch the next cache line of mbuf pointers and flags */
3279 		prefetch(&txq->ift_sds.ifsd_m[next]);
3280 		prefetch(&txq->ift_sds.ifsd_map[next]);
3281 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3282 	}
3283 	map = txq->ift_sds.ifsd_map[pidx];
3284 	ifsd_m = txq->ift_sds.ifsd_m;
3285 
3286 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3287 		buf_tag = txq->ift_tso_buf_tag;
3288 		max_segs = scctx->isc_tx_tso_segments_max;
3289 		map = txq->ift_sds.ifsd_tso_map[pidx];
3290 		MPASS(buf_tag != NULL);
3291 		MPASS(max_segs > 0);
3292 	} else {
3293 		buf_tag = txq->ift_buf_tag;
3294 		max_segs = scctx->isc_tx_nsegments;
3295 		map = txq->ift_sds.ifsd_map[pidx];
3296 	}
3297 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3298 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3299 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3300 		if (err) {
3301 			DBG_COUNTER_INC(encap_txd_encap_fail);
3302 			return err;
3303 		}
3304 	}
3305 	m_head = *m_headp;
3306 
3307 	pkt_info_zero(&pi);
3308 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3309 	pi.ipi_pidx = pidx;
3310 	pi.ipi_qsidx = txq->ift_id;
3311 	pi.ipi_len = m_head->m_pkthdr.len;
3312 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3313 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3314 
3315 	/* deliberate bitwise OR to make one condition */
3316 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3317 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3318 			DBG_COUNTER_INC(encap_txd_encap_fail);
3319 			return (err);
3320 		}
3321 		m_head = *m_headp;
3322 	}
3323 
3324 retry:
3325 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3326 	    BUS_DMA_NOWAIT);
3327 defrag:
3328 	if (__predict_false(err)) {
3329 		switch (err) {
3330 		case EFBIG:
3331 			/* try collapse once and defrag once */
3332 			if (remap == 0) {
3333 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3334 				/* try defrag if collapsing fails */
3335 				if (m_head == NULL)
3336 					remap++;
3337 			}
3338 			if (remap == 1) {
3339 				txq->ift_mbuf_defrag++;
3340 				m_head = m_defrag(*m_headp, M_NOWAIT);
3341 			}
3342 			/*
3343 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3344 			 * failed to map an mbuf that was run through m_defrag
3345 			 */
3346 			MPASS(remap <= 1);
3347 			if (__predict_false(m_head == NULL || remap > 1))
3348 				goto defrag_failed;
3349 			remap++;
3350 			*m_headp = m_head;
3351 			goto retry;
3352 			break;
3353 		case ENOMEM:
3354 			txq->ift_no_tx_dma_setup++;
3355 			break;
3356 		default:
3357 			txq->ift_no_tx_dma_setup++;
3358 			m_freem(*m_headp);
3359 			DBG_COUNTER_INC(tx_frees);
3360 			*m_headp = NULL;
3361 			break;
3362 		}
3363 		txq->ift_map_failed++;
3364 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3365 		DBG_COUNTER_INC(encap_txd_encap_fail);
3366 		return (err);
3367 	}
3368 	ifsd_m[pidx] = m_head;
3369 	/*
3370 	 * XXX assumes a 1 to 1 relationship between segments and
3371 	 *        descriptors - this does not hold true on all drivers, e.g.
3372 	 *        cxgb
3373 	 */
3374 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3375 		txq->ift_no_desc_avail++;
3376 		bus_dmamap_unload(buf_tag, map);
3377 		DBG_COUNTER_INC(encap_txq_avail_fail);
3378 		DBG_COUNTER_INC(encap_txd_encap_fail);
3379 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3380 			GROUPTASK_ENQUEUE(&txq->ift_task);
3381 		return (ENOBUFS);
3382 	}
3383 	/*
3384 	 * On Intel cards we can greatly reduce the number of TX interrupts
3385 	 * we see by only setting report status on every Nth descriptor.
3386 	 * However, this also means that the driver will need to keep track
3387 	 * of the descriptors that RS was set on to check them for the DD bit.
3388 	 */
3389 	txq->ift_rs_pending += nsegs + 1;
3390 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3391 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3392 		pi.ipi_flags |= IPI_TX_INTR;
3393 		txq->ift_rs_pending = 0;
3394 	}
3395 
3396 	pi.ipi_segs = segs;
3397 	pi.ipi_nsegs = nsegs;
3398 
3399 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3400 #ifdef PKT_DEBUG
3401 	print_pkt(&pi);
3402 #endif
3403 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3404 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3405 		DBG_COUNTER_INC(tx_encap);
3406 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3407 
3408 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3409 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3410 			ndesc += txq->ift_size;
3411 			txq->ift_gen = 1;
3412 		}
3413 		/*
3414 		 * drivers can need as many as
3415 		 * two sentinels
3416 		 */
3417 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3418 		MPASS(pi.ipi_new_pidx != pidx);
3419 		MPASS(ndesc > 0);
3420 		txq->ift_in_use += ndesc;
3421 
3422 		/*
3423 		 * We update the last software descriptor again here because there may
3424 		 * be a sentinel and/or there may be more mbufs than segments
3425 		 */
3426 		txq->ift_pidx = pi.ipi_new_pidx;
3427 		txq->ift_npending += pi.ipi_ndescs;
3428 	} else {
3429 		*m_headp = m_head = iflib_remove_mbuf(txq);
3430 		if (err == EFBIG) {
3431 			txq->ift_txd_encap_efbig++;
3432 			if (remap < 2) {
3433 				remap = 1;
3434 				goto defrag;
3435 			}
3436 		}
3437 		goto defrag_failed;
3438 	}
3439 	/*
3440 	 * err can't possibly be non-zero here, so we don't neet to test it
3441 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3442 	 */
3443 	return (err);
3444 
3445 defrag_failed:
3446 	txq->ift_mbuf_defrag_failed++;
3447 	txq->ift_map_failed++;
3448 	m_freem(*m_headp);
3449 	DBG_COUNTER_INC(tx_frees);
3450 	*m_headp = NULL;
3451 	DBG_COUNTER_INC(encap_txd_encap_fail);
3452 	return (ENOMEM);
3453 }
3454 
3455 static void
3456 iflib_tx_desc_free(iflib_txq_t txq, int n)
3457 {
3458 	uint32_t qsize, cidx, mask, gen;
3459 	struct mbuf *m, **ifsd_m;
3460 	bool do_prefetch;
3461 
3462 	cidx = txq->ift_cidx;
3463 	gen = txq->ift_gen;
3464 	qsize = txq->ift_size;
3465 	mask = qsize-1;
3466 	ifsd_m = txq->ift_sds.ifsd_m;
3467 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3468 
3469 	while (n-- > 0) {
3470 		if (do_prefetch) {
3471 			prefetch(ifsd_m[(cidx + 3) & mask]);
3472 			prefetch(ifsd_m[(cidx + 4) & mask]);
3473 		}
3474 		if ((m = ifsd_m[cidx]) != NULL) {
3475 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3476 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3477 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3478 				    txq->ift_sds.ifsd_tso_map[cidx],
3479 				    BUS_DMASYNC_POSTWRITE);
3480 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3481 				    txq->ift_sds.ifsd_tso_map[cidx]);
3482 			} else {
3483 				bus_dmamap_sync(txq->ift_buf_tag,
3484 				    txq->ift_sds.ifsd_map[cidx],
3485 				    BUS_DMASYNC_POSTWRITE);
3486 				bus_dmamap_unload(txq->ift_buf_tag,
3487 				    txq->ift_sds.ifsd_map[cidx]);
3488 			}
3489 			/* XXX we don't support any drivers that batch packets yet */
3490 			MPASS(m->m_nextpkt == NULL);
3491 			m_freem(m);
3492 			ifsd_m[cidx] = NULL;
3493 #if MEMORY_LOGGING
3494 			txq->ift_dequeued++;
3495 #endif
3496 			DBG_COUNTER_INC(tx_frees);
3497 		}
3498 		if (__predict_false(++cidx == qsize)) {
3499 			cidx = 0;
3500 			gen = 0;
3501 		}
3502 	}
3503 	txq->ift_cidx = cidx;
3504 	txq->ift_gen = gen;
3505 }
3506 
3507 static __inline int
3508 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3509 {
3510 	int reclaim;
3511 	if_ctx_t ctx = txq->ift_ctx;
3512 
3513 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3514 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3515 
3516 	/*
3517 	 * Need a rate-limiting check so that this isn't called every time
3518 	 */
3519 	iflib_tx_credits_update(ctx, txq);
3520 	reclaim = DESC_RECLAIMABLE(txq);
3521 
3522 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3523 #ifdef INVARIANTS
3524 		if (iflib_verbose_debug) {
3525 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3526 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3527 			       reclaim, thresh);
3528 
3529 		}
3530 #endif
3531 		return (0);
3532 	}
3533 	iflib_tx_desc_free(txq, reclaim);
3534 	txq->ift_cleaned += reclaim;
3535 	txq->ift_in_use -= reclaim;
3536 
3537 	return (reclaim);
3538 }
3539 
3540 static struct mbuf **
3541 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3542 {
3543 	int next, size;
3544 	struct mbuf **items;
3545 
3546 	size = r->size;
3547 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3548 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3549 
3550 	prefetch(items[(cidx + offset) & (size-1)]);
3551 	if (remaining > 1) {
3552 		prefetch2cachelines(&items[next]);
3553 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3554 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3555 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3556 	}
3557 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3558 }
3559 
3560 static void
3561 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3562 {
3563 
3564 	ifmp_ring_check_drainage(txq->ift_br, budget);
3565 }
3566 
3567 static uint32_t
3568 iflib_txq_can_drain(struct ifmp_ring *r)
3569 {
3570 	iflib_txq_t txq = r->cookie;
3571 	if_ctx_t ctx = txq->ift_ctx;
3572 
3573 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3574 		return (1);
3575 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3576 	    BUS_DMASYNC_POSTREAD);
3577 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3578 	    false));
3579 }
3580 
3581 static uint32_t
3582 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3583 {
3584 	iflib_txq_t txq = r->cookie;
3585 	if_ctx_t ctx = txq->ift_ctx;
3586 	struct ifnet *ifp = ctx->ifc_ifp;
3587 	struct mbuf **mp, *m;
3588 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3589 	int reclaimed, err, in_use_prev, desc_used;
3590 	bool do_prefetch, ring, rang;
3591 
3592 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3593 			    !LINK_ACTIVE(ctx))) {
3594 		DBG_COUNTER_INC(txq_drain_notready);
3595 		return (0);
3596 	}
3597 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3598 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3599 	avail = IDXDIFF(pidx, cidx, r->size);
3600 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3601 		DBG_COUNTER_INC(txq_drain_flushing);
3602 		for (i = 0; i < avail; i++) {
3603 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3604 				m_free(r->items[(cidx + i) & (r->size-1)]);
3605 			r->items[(cidx + i) & (r->size-1)] = NULL;
3606 		}
3607 		return (avail);
3608 	}
3609 
3610 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3611 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3612 		CALLOUT_LOCK(txq);
3613 		callout_stop(&txq->ift_timer);
3614 		CALLOUT_UNLOCK(txq);
3615 		DBG_COUNTER_INC(txq_drain_oactive);
3616 		return (0);
3617 	}
3618 	if (reclaimed)
3619 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3620 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3621 	count = MIN(avail, TX_BATCH_SIZE);
3622 #ifdef INVARIANTS
3623 	if (iflib_verbose_debug)
3624 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3625 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3626 #endif
3627 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3628 	avail = TXQ_AVAIL(txq);
3629 	err = 0;
3630 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3631 		int rem = do_prefetch ? count - i : 0;
3632 
3633 		mp = _ring_peek_one(r, cidx, i, rem);
3634 		MPASS(mp != NULL && *mp != NULL);
3635 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3636 			consumed++;
3637 			reclaimed++;
3638 			continue;
3639 		}
3640 		in_use_prev = txq->ift_in_use;
3641 		err = iflib_encap(txq, mp);
3642 		if (__predict_false(err)) {
3643 			/* no room - bail out */
3644 			if (err == ENOBUFS)
3645 				break;
3646 			consumed++;
3647 			/* we can't send this packet - skip it */
3648 			continue;
3649 		}
3650 		consumed++;
3651 		pkt_sent++;
3652 		m = *mp;
3653 		DBG_COUNTER_INC(tx_sent);
3654 		bytes_sent += m->m_pkthdr.len;
3655 		mcast_sent += !!(m->m_flags & M_MCAST);
3656 		avail = TXQ_AVAIL(txq);
3657 
3658 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3659 		desc_used += (txq->ift_in_use - in_use_prev);
3660 		ETHER_BPF_MTAP(ifp, m);
3661 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3662 			break;
3663 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3664 	}
3665 
3666 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3667 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3668 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3669 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3670 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3671 	if (mcast_sent)
3672 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3673 #ifdef INVARIANTS
3674 	if (iflib_verbose_debug)
3675 		printf("consumed=%d\n", consumed);
3676 #endif
3677 	return (consumed);
3678 }
3679 
3680 static uint32_t
3681 iflib_txq_drain_always(struct ifmp_ring *r)
3682 {
3683 	return (1);
3684 }
3685 
3686 static uint32_t
3687 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3688 {
3689 	int i, avail;
3690 	struct mbuf **mp;
3691 	iflib_txq_t txq;
3692 
3693 	txq = r->cookie;
3694 
3695 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3696 	CALLOUT_LOCK(txq);
3697 	callout_stop(&txq->ift_timer);
3698 	CALLOUT_UNLOCK(txq);
3699 
3700 	avail = IDXDIFF(pidx, cidx, r->size);
3701 	for (i = 0; i < avail; i++) {
3702 		mp = _ring_peek_one(r, cidx, i, avail - i);
3703 		if (__predict_false(*mp == (struct mbuf *)txq))
3704 			continue;
3705 		m_freem(*mp);
3706 		DBG_COUNTER_INC(tx_frees);
3707 	}
3708 	MPASS(ifmp_ring_is_stalled(r) == 0);
3709 	return (avail);
3710 }
3711 
3712 static void
3713 iflib_ifmp_purge(iflib_txq_t txq)
3714 {
3715 	struct ifmp_ring *r;
3716 
3717 	r = txq->ift_br;
3718 	r->drain = iflib_txq_drain_free;
3719 	r->can_drain = iflib_txq_drain_always;
3720 
3721 	ifmp_ring_check_drainage(r, r->size);
3722 
3723 	r->drain = iflib_txq_drain;
3724 	r->can_drain = iflib_txq_can_drain;
3725 }
3726 
3727 static void
3728 _task_fn_tx(void *context)
3729 {
3730 	iflib_txq_t txq = context;
3731 	if_ctx_t ctx = txq->ift_ctx;
3732 #if defined(ALTQ) || defined(DEV_NETMAP)
3733 	if_t ifp = ctx->ifc_ifp;
3734 #endif
3735 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3736 
3737 #ifdef IFLIB_DIAGNOSTICS
3738 	txq->ift_cpu_exec_count[curcpu]++;
3739 #endif
3740 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3741 		return;
3742 #ifdef DEV_NETMAP
3743 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3744 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3745 		    BUS_DMASYNC_POSTREAD);
3746 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3747 			netmap_tx_irq(ifp, txq->ift_id);
3748 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3749 		return;
3750 	}
3751 #endif
3752 #ifdef ALTQ
3753 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3754 		iflib_altq_if_start(ifp);
3755 #endif
3756 	if (txq->ift_db_pending)
3757 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3758 	else if (!abdicate)
3759 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3760 	/*
3761 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3762 	 */
3763 	if (abdicate)
3764 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3765 	if (ctx->ifc_flags & IFC_LEGACY)
3766 		IFDI_INTR_ENABLE(ctx);
3767 	else {
3768 #ifdef INVARIANTS
3769 		int rc =
3770 #endif
3771 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3772 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3773 	}
3774 }
3775 
3776 static void
3777 _task_fn_rx(void *context)
3778 {
3779 	iflib_rxq_t rxq = context;
3780 	if_ctx_t ctx = rxq->ifr_ctx;
3781 	bool more;
3782 	uint16_t budget;
3783 
3784 #ifdef IFLIB_DIAGNOSTICS
3785 	rxq->ifr_cpu_exec_count[curcpu]++;
3786 #endif
3787 	DBG_COUNTER_INC(task_fn_rxs);
3788 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3789 		return;
3790 	more = true;
3791 #ifdef DEV_NETMAP
3792 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3793 		u_int work = 0;
3794 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3795 			more = false;
3796 		}
3797 	}
3798 #endif
3799 	budget = ctx->ifc_sysctl_rx_budget;
3800 	if (budget == 0)
3801 		budget = 16;	/* XXX */
3802 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3803 		if (ctx->ifc_flags & IFC_LEGACY)
3804 			IFDI_INTR_ENABLE(ctx);
3805 		else {
3806 #ifdef INVARIANTS
3807 			int rc =
3808 #endif
3809 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3810 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3811 			DBG_COUNTER_INC(rx_intr_enables);
3812 		}
3813 	}
3814 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3815 		return;
3816 	if (more)
3817 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3818 }
3819 
3820 static void
3821 _task_fn_admin(void *context)
3822 {
3823 	if_ctx_t ctx = context;
3824 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3825 	iflib_txq_t txq;
3826 	int i;
3827 	bool oactive, running, do_reset, do_watchdog, in_detach;
3828 	uint32_t reset_on = hz / 2;
3829 
3830 	STATE_LOCK(ctx);
3831 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3832 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3833 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3834 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3835 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3836 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3837 	STATE_UNLOCK(ctx);
3838 
3839 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3840 		return;
3841 	if (in_detach)
3842 		return;
3843 
3844 	CTX_LOCK(ctx);
3845 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3846 		CALLOUT_LOCK(txq);
3847 		callout_stop(&txq->ift_timer);
3848 		CALLOUT_UNLOCK(txq);
3849 	}
3850 	if (do_watchdog) {
3851 		ctx->ifc_watchdog_events++;
3852 		IFDI_WATCHDOG_RESET(ctx);
3853 	}
3854 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3855 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3856 #ifdef DEV_NETMAP
3857 		reset_on = hz / 2;
3858 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3859 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3860 #endif
3861 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3862 	}
3863 	IFDI_LINK_INTR_ENABLE(ctx);
3864 	if (do_reset)
3865 		iflib_if_init_locked(ctx);
3866 	CTX_UNLOCK(ctx);
3867 
3868 	if (LINK_ACTIVE(ctx) == 0)
3869 		return;
3870 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3871 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3872 }
3873 
3874 
3875 static void
3876 _task_fn_iov(void *context)
3877 {
3878 	if_ctx_t ctx = context;
3879 
3880 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3881 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3882 		return;
3883 
3884 	CTX_LOCK(ctx);
3885 	IFDI_VFLR_HANDLE(ctx);
3886 	CTX_UNLOCK(ctx);
3887 }
3888 
3889 static int
3890 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3891 {
3892 	int err;
3893 	if_int_delay_info_t info;
3894 	if_ctx_t ctx;
3895 
3896 	info = (if_int_delay_info_t)arg1;
3897 	ctx = info->iidi_ctx;
3898 	info->iidi_req = req;
3899 	info->iidi_oidp = oidp;
3900 	CTX_LOCK(ctx);
3901 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3902 	CTX_UNLOCK(ctx);
3903 	return (err);
3904 }
3905 
3906 /*********************************************************************
3907  *
3908  *  IFNET FUNCTIONS
3909  *
3910  **********************************************************************/
3911 
3912 static void
3913 iflib_if_init_locked(if_ctx_t ctx)
3914 {
3915 	iflib_stop(ctx);
3916 	iflib_init_locked(ctx);
3917 }
3918 
3919 
3920 static void
3921 iflib_if_init(void *arg)
3922 {
3923 	if_ctx_t ctx = arg;
3924 
3925 	CTX_LOCK(ctx);
3926 	iflib_if_init_locked(ctx);
3927 	CTX_UNLOCK(ctx);
3928 }
3929 
3930 static int
3931 iflib_if_transmit(if_t ifp, struct mbuf *m)
3932 {
3933 	if_ctx_t	ctx = if_getsoftc(ifp);
3934 
3935 	iflib_txq_t txq;
3936 	int err, qidx;
3937 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3938 
3939 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3940 		DBG_COUNTER_INC(tx_frees);
3941 		m_freem(m);
3942 		return (ENETDOWN);
3943 	}
3944 
3945 	MPASS(m->m_nextpkt == NULL);
3946 	/* ALTQ-enabled interfaces always use queue 0. */
3947 	qidx = 0;
3948 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3949 		qidx = QIDX(ctx, m);
3950 	/*
3951 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3952 	 */
3953 	txq = &ctx->ifc_txqs[qidx];
3954 
3955 #ifdef DRIVER_BACKPRESSURE
3956 	if (txq->ift_closed) {
3957 		while (m != NULL) {
3958 			next = m->m_nextpkt;
3959 			m->m_nextpkt = NULL;
3960 			m_freem(m);
3961 			DBG_COUNTER_INC(tx_frees);
3962 			m = next;
3963 		}
3964 		return (ENOBUFS);
3965 	}
3966 #endif
3967 #ifdef notyet
3968 	qidx = count = 0;
3969 	mp = marr;
3970 	next = m;
3971 	do {
3972 		count++;
3973 		next = next->m_nextpkt;
3974 	} while (next != NULL);
3975 
3976 	if (count > nitems(marr))
3977 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3978 			/* XXX check nextpkt */
3979 			m_freem(m);
3980 			/* XXX simplify for now */
3981 			DBG_COUNTER_INC(tx_frees);
3982 			return (ENOBUFS);
3983 		}
3984 	for (next = m, i = 0; next != NULL; i++) {
3985 		mp[i] = next;
3986 		next = next->m_nextpkt;
3987 		mp[i]->m_nextpkt = NULL;
3988 	}
3989 #endif
3990 	DBG_COUNTER_INC(tx_seen);
3991 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3992 
3993 	if (abdicate)
3994 		GROUPTASK_ENQUEUE(&txq->ift_task);
3995  	if (err) {
3996 		if (!abdicate)
3997 			GROUPTASK_ENQUEUE(&txq->ift_task);
3998 		/* support forthcoming later */
3999 #ifdef DRIVER_BACKPRESSURE
4000 		txq->ift_closed = TRUE;
4001 #endif
4002 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4003 		m_freem(m);
4004 		DBG_COUNTER_INC(tx_frees);
4005 	}
4006 
4007 	return (err);
4008 }
4009 
4010 #ifdef ALTQ
4011 /*
4012  * The overall approach to integrating iflib with ALTQ is to continue to use
4013  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4014  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4015  * is redundant/unnecessary, but doing so minimizes the amount of
4016  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4017  * redundantly queueing to an intermediate mp_ring is swamped by the
4018  * performance limitations inherent in using ALTQ.
4019  *
4020  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4021  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4022  * given interface.  If ALTQ is enabled for an interface, then all
4023  * transmitted packets for that interface will be submitted to the ALTQ
4024  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4025  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4026  * update stats that the iflib machinery handles, and which is sensitve to
4027  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4028  * will be installed as the start routine for use by ALTQ facilities that
4029  * need to trigger queue drains on a scheduled basis.
4030  *
4031  */
4032 static void
4033 iflib_altq_if_start(if_t ifp)
4034 {
4035 	struct ifaltq *ifq = &ifp->if_snd;
4036 	struct mbuf *m;
4037 
4038 	IFQ_LOCK(ifq);
4039 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4040 	while (m != NULL) {
4041 		iflib_if_transmit(ifp, m);
4042 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4043 	}
4044 	IFQ_UNLOCK(ifq);
4045 }
4046 
4047 static int
4048 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4049 {
4050 	int err;
4051 
4052 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4053 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4054 		if (err == 0)
4055 			iflib_altq_if_start(ifp);
4056 	} else
4057 		err = iflib_if_transmit(ifp, m);
4058 
4059 	return (err);
4060 }
4061 #endif /* ALTQ */
4062 
4063 static void
4064 iflib_if_qflush(if_t ifp)
4065 {
4066 	if_ctx_t ctx = if_getsoftc(ifp);
4067 	iflib_txq_t txq = ctx->ifc_txqs;
4068 	int i;
4069 
4070 	STATE_LOCK(ctx);
4071 	ctx->ifc_flags |= IFC_QFLUSH;
4072 	STATE_UNLOCK(ctx);
4073 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4074 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4075 			iflib_txq_check_drain(txq, 0);
4076 	STATE_LOCK(ctx);
4077 	ctx->ifc_flags &= ~IFC_QFLUSH;
4078 	STATE_UNLOCK(ctx);
4079 
4080 	/*
4081 	 * When ALTQ is enabled, this will also take care of purging the
4082 	 * ALTQ queue(s).
4083 	 */
4084 	if_qflush(ifp);
4085 }
4086 
4087 
4088 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4089 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4090 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4091 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4092 
4093 static int
4094 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4095 {
4096 	if_ctx_t ctx = if_getsoftc(ifp);
4097 	struct ifreq	*ifr = (struct ifreq *)data;
4098 #if defined(INET) || defined(INET6)
4099 	struct ifaddr	*ifa = (struct ifaddr *)data;
4100 #endif
4101 	bool		avoid_reset = FALSE;
4102 	int		err = 0, reinit = 0, bits;
4103 
4104 	switch (command) {
4105 	case SIOCSIFADDR:
4106 #ifdef INET
4107 		if (ifa->ifa_addr->sa_family == AF_INET)
4108 			avoid_reset = TRUE;
4109 #endif
4110 #ifdef INET6
4111 		if (ifa->ifa_addr->sa_family == AF_INET6)
4112 			avoid_reset = TRUE;
4113 #endif
4114 		/*
4115 		** Calling init results in link renegotiation,
4116 		** so we avoid doing it when possible.
4117 		*/
4118 		if (avoid_reset) {
4119 			if_setflagbits(ifp, IFF_UP,0);
4120 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4121 				reinit = 1;
4122 #ifdef INET
4123 			if (!(if_getflags(ifp) & IFF_NOARP))
4124 				arp_ifinit(ifp, ifa);
4125 #endif
4126 		} else
4127 			err = ether_ioctl(ifp, command, data);
4128 		break;
4129 	case SIOCSIFMTU:
4130 		CTX_LOCK(ctx);
4131 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4132 			CTX_UNLOCK(ctx);
4133 			break;
4134 		}
4135 		bits = if_getdrvflags(ifp);
4136 		/* stop the driver and free any clusters before proceeding */
4137 		iflib_stop(ctx);
4138 
4139 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4140 			STATE_LOCK(ctx);
4141 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4142 				ctx->ifc_flags |= IFC_MULTISEG;
4143 			else
4144 				ctx->ifc_flags &= ~IFC_MULTISEG;
4145 			STATE_UNLOCK(ctx);
4146 			err = if_setmtu(ifp, ifr->ifr_mtu);
4147 		}
4148 		iflib_init_locked(ctx);
4149 		STATE_LOCK(ctx);
4150 		if_setdrvflags(ifp, bits);
4151 		STATE_UNLOCK(ctx);
4152 		CTX_UNLOCK(ctx);
4153 		break;
4154 	case SIOCSIFFLAGS:
4155 		CTX_LOCK(ctx);
4156 		if (if_getflags(ifp) & IFF_UP) {
4157 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4158 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4159 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4160 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4161 				}
4162 			} else
4163 				reinit = 1;
4164 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4165 			iflib_stop(ctx);
4166 		}
4167 		ctx->ifc_if_flags = if_getflags(ifp);
4168 		CTX_UNLOCK(ctx);
4169 		break;
4170 	case SIOCADDMULTI:
4171 	case SIOCDELMULTI:
4172 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4173 			CTX_LOCK(ctx);
4174 			IFDI_INTR_DISABLE(ctx);
4175 			IFDI_MULTI_SET(ctx);
4176 			IFDI_INTR_ENABLE(ctx);
4177 			CTX_UNLOCK(ctx);
4178 		}
4179 		break;
4180 	case SIOCSIFMEDIA:
4181 		CTX_LOCK(ctx);
4182 		IFDI_MEDIA_SET(ctx);
4183 		CTX_UNLOCK(ctx);
4184 		/* falls thru */
4185 	case SIOCGIFMEDIA:
4186 	case SIOCGIFXMEDIA:
4187 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4188 		break;
4189 	case SIOCGI2C:
4190 	{
4191 		struct ifi2creq i2c;
4192 
4193 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4194 		if (err != 0)
4195 			break;
4196 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4197 			err = EINVAL;
4198 			break;
4199 		}
4200 		if (i2c.len > sizeof(i2c.data)) {
4201 			err = EINVAL;
4202 			break;
4203 		}
4204 
4205 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4206 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4207 			    sizeof(i2c));
4208 		break;
4209 	}
4210 	case SIOCSIFCAP:
4211 	{
4212 		int mask, setmask, oldmask;
4213 
4214 		oldmask = if_getcapenable(ifp);
4215 		mask = ifr->ifr_reqcap ^ oldmask;
4216 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4217 		setmask = 0;
4218 #ifdef TCP_OFFLOAD
4219 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4220 #endif
4221 		setmask |= (mask & IFCAP_FLAGS);
4222 		setmask |= (mask & IFCAP_WOL);
4223 
4224 		/*
4225 		 * If any RX csum has changed, change all the ones that
4226 		 * are supported by the driver.
4227 		 */
4228 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4229 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4230 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4231 		}
4232 
4233 		/*
4234 		 * want to ensure that traffic has stopped before we change any of the flags
4235 		 */
4236 		if (setmask) {
4237 			CTX_LOCK(ctx);
4238 			bits = if_getdrvflags(ifp);
4239 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4240 				iflib_stop(ctx);
4241 			STATE_LOCK(ctx);
4242 			if_togglecapenable(ifp, setmask);
4243 			STATE_UNLOCK(ctx);
4244 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4245 				iflib_init_locked(ctx);
4246 			STATE_LOCK(ctx);
4247 			if_setdrvflags(ifp, bits);
4248 			STATE_UNLOCK(ctx);
4249 			CTX_UNLOCK(ctx);
4250 		}
4251 		if_vlancap(ifp);
4252 		break;
4253 	}
4254 	case SIOCGPRIVATE_0:
4255 	case SIOCSDRVSPEC:
4256 	case SIOCGDRVSPEC:
4257 		CTX_LOCK(ctx);
4258 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4259 		CTX_UNLOCK(ctx);
4260 		break;
4261 	default:
4262 		err = ether_ioctl(ifp, command, data);
4263 		break;
4264 	}
4265 	if (reinit)
4266 		iflib_if_init(ctx);
4267 	return (err);
4268 }
4269 
4270 static uint64_t
4271 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4272 {
4273 	if_ctx_t ctx = if_getsoftc(ifp);
4274 
4275 	return (IFDI_GET_COUNTER(ctx, cnt));
4276 }
4277 
4278 /*********************************************************************
4279  *
4280  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4281  *
4282  **********************************************************************/
4283 
4284 static void
4285 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4286 {
4287 	if_ctx_t ctx = if_getsoftc(ifp);
4288 
4289 	if ((void *)ctx != arg)
4290 		return;
4291 
4292 	if ((vtag == 0) || (vtag > 4095))
4293 		return;
4294 
4295 	CTX_LOCK(ctx);
4296 	IFDI_VLAN_REGISTER(ctx, vtag);
4297 	/* Re-init to load the changes */
4298 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4299 		iflib_if_init_locked(ctx);
4300 	CTX_UNLOCK(ctx);
4301 }
4302 
4303 static void
4304 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4305 {
4306 	if_ctx_t ctx = if_getsoftc(ifp);
4307 
4308 	if ((void *)ctx != arg)
4309 		return;
4310 
4311 	if ((vtag == 0) || (vtag > 4095))
4312 		return;
4313 
4314 	CTX_LOCK(ctx);
4315 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4316 	/* Re-init to load the changes */
4317 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4318 		iflib_if_init_locked(ctx);
4319 	CTX_UNLOCK(ctx);
4320 }
4321 
4322 static void
4323 iflib_led_func(void *arg, int onoff)
4324 {
4325 	if_ctx_t ctx = arg;
4326 
4327 	CTX_LOCK(ctx);
4328 	IFDI_LED_FUNC(ctx, onoff);
4329 	CTX_UNLOCK(ctx);
4330 }
4331 
4332 /*********************************************************************
4333  *
4334  *  BUS FUNCTION DEFINITIONS
4335  *
4336  **********************************************************************/
4337 
4338 int
4339 iflib_device_probe(device_t dev)
4340 {
4341 	pci_vendor_info_t *ent;
4342 
4343 	uint16_t	pci_vendor_id, pci_device_id;
4344 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4345 	uint16_t	pci_rev_id;
4346 	if_shared_ctx_t sctx;
4347 
4348 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4349 		return (ENOTSUP);
4350 
4351 	pci_vendor_id = pci_get_vendor(dev);
4352 	pci_device_id = pci_get_device(dev);
4353 	pci_subvendor_id = pci_get_subvendor(dev);
4354 	pci_subdevice_id = pci_get_subdevice(dev);
4355 	pci_rev_id = pci_get_revid(dev);
4356 	if (sctx->isc_parse_devinfo != NULL)
4357 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4358 
4359 	ent = sctx->isc_vendor_info;
4360 	while (ent->pvi_vendor_id != 0) {
4361 		if (pci_vendor_id != ent->pvi_vendor_id) {
4362 			ent++;
4363 			continue;
4364 		}
4365 		if ((pci_device_id == ent->pvi_device_id) &&
4366 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4367 		     (ent->pvi_subvendor_id == 0)) &&
4368 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4369 		     (ent->pvi_subdevice_id == 0)) &&
4370 		    ((pci_rev_id == ent->pvi_rev_id) ||
4371 		     (ent->pvi_rev_id == 0))) {
4372 
4373 			device_set_desc_copy(dev, ent->pvi_name);
4374 			/* this needs to be changed to zero if the bus probing code
4375 			 * ever stops re-probing on best match because the sctx
4376 			 * may have its values over written by register calls
4377 			 * in subsequent probes
4378 			 */
4379 			return (BUS_PROBE_DEFAULT);
4380 		}
4381 		ent++;
4382 	}
4383 	return (ENXIO);
4384 }
4385 
4386 static void
4387 iflib_reset_qvalues(if_ctx_t ctx)
4388 {
4389 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4390 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4391 	device_t dev = ctx->ifc_dev;
4392 	int i;
4393 
4394 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4395 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4396 	/*
4397 	 * XXX sanity check that ntxd & nrxd are a power of 2
4398 	 */
4399 	if (ctx->ifc_sysctl_ntxqs != 0)
4400 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4401 	if (ctx->ifc_sysctl_nrxqs != 0)
4402 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4403 
4404 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4405 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4406 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4407 		else
4408 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4409 	}
4410 
4411 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4412 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4413 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4414 		else
4415 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4416 	}
4417 
4418 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4419 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4420 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4421 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4422 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4423 		}
4424 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4425 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4426 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4427 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4428 		}
4429 	}
4430 
4431 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4432 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4433 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4434 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4435 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4436 		}
4437 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4438 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4439 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4440 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4441 		}
4442 	}
4443 }
4444 
4445 static void
4446 iflib_add_pfil(if_ctx_t ctx)
4447 {
4448 	struct pfil_head *pfil;
4449 	struct pfil_head_args pa;
4450 	iflib_rxq_t rxq;
4451 	int i;
4452 
4453 	pa.pa_version = PFIL_VERSION;
4454 	pa.pa_flags = PFIL_IN;
4455 	pa.pa_type = PFIL_TYPE_ETHERNET;
4456 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4457 	pfil = pfil_head_register(&pa);
4458 
4459 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4460 		rxq->pfil = pfil;
4461 	}
4462 }
4463 
4464 static void
4465 iflib_rem_pfil(if_ctx_t ctx)
4466 {
4467 	struct pfil_head *pfil;
4468 	iflib_rxq_t rxq;
4469 	int i;
4470 
4471 	rxq = ctx->ifc_rxqs;
4472 	pfil = rxq->pfil;
4473 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4474 		rxq->pfil = NULL;
4475 	}
4476 	pfil_head_unregister(pfil);
4477 }
4478 
4479 static uint16_t
4480 get_ctx_core_offset(if_ctx_t ctx)
4481 {
4482 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4483 	struct cpu_offset *op;
4484 	uint16_t qc;
4485 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4486 
4487 	if (ret != CORE_OFFSET_UNSPECIFIED)
4488 		return (ret);
4489 
4490 	if (ctx->ifc_sysctl_separate_txrx)
4491 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4492 	else
4493 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4494 
4495 	mtx_lock(&cpu_offset_mtx);
4496 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4497 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4498 			ret = op->offset;
4499 			op->offset += qc;
4500 			MPASS(op->refcount < UINT_MAX);
4501 			op->refcount++;
4502 			break;
4503 		}
4504 	}
4505 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4506 		ret = 0;
4507 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4508 		    M_NOWAIT | M_ZERO);
4509 		if (op == NULL) {
4510 			device_printf(ctx->ifc_dev,
4511 			    "allocation for cpu offset failed.\n");
4512 		} else {
4513 			op->offset = qc;
4514 			op->refcount = 1;
4515 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4516 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4517 		}
4518 	}
4519 	mtx_unlock(&cpu_offset_mtx);
4520 
4521 	return (ret);
4522 }
4523 
4524 static void
4525 unref_ctx_core_offset(if_ctx_t ctx)
4526 {
4527 	struct cpu_offset *op, *top;
4528 
4529 	mtx_lock(&cpu_offset_mtx);
4530 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4531 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4532 			MPASS(op->refcount > 0);
4533 			op->refcount--;
4534 			if (op->refcount == 0) {
4535 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4536 				free(op, M_IFLIB);
4537 			}
4538 			break;
4539 		}
4540 	}
4541 	mtx_unlock(&cpu_offset_mtx);
4542 }
4543 
4544 int
4545 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4546 {
4547 	int err, rid, msix;
4548 	if_ctx_t ctx;
4549 	if_t ifp;
4550 	if_softc_ctx_t scctx;
4551 	int i;
4552 	uint16_t main_txq;
4553 	uint16_t main_rxq;
4554 
4555 
4556 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4557 
4558 	if (sc == NULL) {
4559 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4560 		device_set_softc(dev, ctx);
4561 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4562 	}
4563 
4564 	ctx->ifc_sctx = sctx;
4565 	ctx->ifc_dev = dev;
4566 	ctx->ifc_softc = sc;
4567 
4568 	if ((err = iflib_register(ctx)) != 0) {
4569 		device_printf(dev, "iflib_register failed %d\n", err);
4570 		goto fail_ctx_free;
4571 	}
4572 	iflib_add_device_sysctl_pre(ctx);
4573 
4574 	scctx = &ctx->ifc_softc_ctx;
4575 	ifp = ctx->ifc_ifp;
4576 
4577 	iflib_reset_qvalues(ctx);
4578 	CTX_LOCK(ctx);
4579 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4580 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4581 		goto fail_unlock;
4582 	}
4583 	_iflib_pre_assert(scctx);
4584 	ctx->ifc_txrx = *scctx->isc_txrx;
4585 
4586 #ifdef INVARIANTS
4587 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4588 		MPASS(scctx->isc_tx_csum_flags);
4589 #endif
4590 
4591 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4592 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4593 
4594 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4595 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4596 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4597 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4598 
4599 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4600 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4601 
4602 	/* XXX change for per-queue sizes */
4603 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4604 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4605 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4606 		if (!powerof2(scctx->isc_nrxd[i])) {
4607 			/* round down instead? */
4608 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4609 			err = EINVAL;
4610 			goto fail_iflib_detach;
4611 		}
4612 	}
4613 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4614 		if (!powerof2(scctx->isc_ntxd[i])) {
4615 			device_printf(dev,
4616 			    "# tx descriptors must be a power of 2");
4617 			err = EINVAL;
4618 			goto fail_iflib_detach;
4619 		}
4620 	}
4621 
4622 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4623 	    MAX_SINGLE_PACKET_FRACTION)
4624 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4625 		    MAX_SINGLE_PACKET_FRACTION);
4626 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4627 	    MAX_SINGLE_PACKET_FRACTION)
4628 		scctx->isc_tx_tso_segments_max = max(1,
4629 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4630 
4631 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4632 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4633 		/*
4634 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4635 		 * but some MACs do.
4636 		 */
4637 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4638 		    IP_MAXPACKET));
4639 		/*
4640 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4641 		 * into account.  In the worst case, each of these calls will
4642 		 * add another mbuf and, thus, the requirement for another DMA
4643 		 * segment.  So for best performance, it doesn't make sense to
4644 		 * advertize a maximum of TSO segments that typically will
4645 		 * require defragmentation in iflib_encap().
4646 		 */
4647 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4648 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4649 	}
4650 	if (scctx->isc_rss_table_size == 0)
4651 		scctx->isc_rss_table_size = 64;
4652 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4653 
4654 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4655 	/* XXX format name */
4656 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4657 	    NULL, NULL, "admin");
4658 
4659 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4660 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4661 		device_printf(dev, "Unable to fetch CPU list\n");
4662 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4663 	}
4664 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4665 
4666 	/*
4667 	** Now set up MSI or MSI-X, should return us the number of supported
4668 	** vectors (will be 1 for a legacy interrupt and MSI).
4669 	*/
4670 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4671 		msix = scctx->isc_vectors;
4672 	} else if (scctx->isc_msix_bar != 0)
4673 	       /*
4674 		* The simple fact that isc_msix_bar is not 0 does not mean we
4675 		* we have a good value there that is known to work.
4676 		*/
4677 		msix = iflib_msix_init(ctx);
4678 	else {
4679 		scctx->isc_vectors = 1;
4680 		scctx->isc_ntxqsets = 1;
4681 		scctx->isc_nrxqsets = 1;
4682 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4683 		msix = 0;
4684 	}
4685 	/* Get memory for the station queues */
4686 	if ((err = iflib_queues_alloc(ctx))) {
4687 		device_printf(dev, "Unable to allocate queue memory\n");
4688 		goto fail_intr_free;
4689 	}
4690 
4691 	if ((err = iflib_qset_structures_setup(ctx)))
4692 		goto fail_queues;
4693 
4694 	/*
4695 	 * Now that we know how many queues there are, get the core offset.
4696 	 */
4697 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4698 
4699 	/*
4700 	 * Group taskqueues aren't properly set up until SMP is started,
4701 	 * so we disable interrupts until we can handle them post
4702 	 * SI_SUB_SMP.
4703 	 *
4704 	 * XXX: disabling interrupts doesn't actually work, at least for
4705 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4706 	 * we do null handling and depend on this not causing too large an
4707 	 * interrupt storm.
4708 	 */
4709 	IFDI_INTR_DISABLE(ctx);
4710 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4711 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4712 		goto fail_queues;
4713 	}
4714 	if (msix <= 1) {
4715 		rid = 0;
4716 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4717 			MPASS(msix == 1);
4718 			rid = 1;
4719 		}
4720 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4721 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4722 			goto fail_queues;
4723 		}
4724 	}
4725 
4726 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4727 
4728 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4729 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4730 		goto fail_detach;
4731 	}
4732 
4733 	/*
4734 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4735 	 * This must appear after the call to ether_ifattach() because
4736 	 * ether_ifattach() sets if_hdrlen to the default value.
4737 	 */
4738 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4739 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4740 
4741 	if ((err = iflib_netmap_attach(ctx))) {
4742 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4743 		goto fail_detach;
4744 	}
4745 	*ctxp = ctx;
4746 
4747 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4748 
4749 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4750 	iflib_add_device_sysctl_post(ctx);
4751 	iflib_add_pfil(ctx);
4752 	ctx->ifc_flags |= IFC_INIT_DONE;
4753 	CTX_UNLOCK(ctx);
4754 	return (0);
4755 
4756 fail_detach:
4757 	ether_ifdetach(ctx->ifc_ifp);
4758 fail_intr_free:
4759 	iflib_free_intr_mem(ctx);
4760 fail_queues:
4761 	iflib_tx_structures_free(ctx);
4762 	iflib_rx_structures_free(ctx);
4763 fail_iflib_detach:
4764 	IFDI_DETACH(ctx);
4765 fail_unlock:
4766 	CTX_UNLOCK(ctx);
4767 fail_ctx_free:
4768         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4769                 free(ctx->ifc_softc, M_IFLIB);
4770         free(ctx, M_IFLIB);
4771 	return (err);
4772 }
4773 
4774 int
4775 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4776 					  struct iflib_cloneattach_ctx *clctx)
4777 {
4778 	int err;
4779 	if_ctx_t ctx;
4780 	if_t ifp;
4781 	if_softc_ctx_t scctx;
4782 	int i;
4783 	void *sc;
4784 	uint16_t main_txq;
4785 	uint16_t main_rxq;
4786 
4787 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4788 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4789 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4790 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4791 		ctx->ifc_flags |= IFC_PSEUDO;
4792 
4793 	ctx->ifc_sctx = sctx;
4794 	ctx->ifc_softc = sc;
4795 	ctx->ifc_dev = dev;
4796 
4797 	if ((err = iflib_register(ctx)) != 0) {
4798 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4799 		goto fail_ctx_free;
4800 	}
4801 	iflib_add_device_sysctl_pre(ctx);
4802 
4803 	scctx = &ctx->ifc_softc_ctx;
4804 	ifp = ctx->ifc_ifp;
4805 
4806 	/*
4807 	 * XXX sanity check that ntxd & nrxd are a power of 2
4808 	 */
4809 	iflib_reset_qvalues(ctx);
4810 	CTX_LOCK(ctx);
4811 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4812 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4813 		goto fail_unlock;
4814 	}
4815 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4816 		ether_gen_addr(ifp, &ctx->ifc_mac);
4817 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4818 								clctx->cc_params)) != 0) {
4819 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4820 		goto fail_ctx_free;
4821 	}
4822 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4823 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4824 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4825 
4826 #ifdef INVARIANTS
4827 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4828 		MPASS(scctx->isc_tx_csum_flags);
4829 #endif
4830 
4831 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4832 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4833 
4834 	ifp->if_flags |= IFF_NOGROUP;
4835 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4836 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4837 
4838 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4839 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4840 			goto fail_detach;
4841 		}
4842 		*ctxp = ctx;
4843 
4844 		/*
4845 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4846 		 * This must appear after the call to ether_ifattach() because
4847 		 * ether_ifattach() sets if_hdrlen to the default value.
4848 		 */
4849 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4850 			if_setifheaderlen(ifp,
4851 			    sizeof(struct ether_vlan_header));
4852 
4853 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4854 		iflib_add_device_sysctl_post(ctx);
4855 		ctx->ifc_flags |= IFC_INIT_DONE;
4856 		return (0);
4857 	}
4858 	_iflib_pre_assert(scctx);
4859 	ctx->ifc_txrx = *scctx->isc_txrx;
4860 
4861 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4862 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4863 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4864 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4865 
4866 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4867 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4868 
4869 	/* XXX change for per-queue sizes */
4870 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4871 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4872 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4873 		if (!powerof2(scctx->isc_nrxd[i])) {
4874 			/* round down instead? */
4875 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4876 			err = EINVAL;
4877 			goto fail_iflib_detach;
4878 		}
4879 	}
4880 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4881 		if (!powerof2(scctx->isc_ntxd[i])) {
4882 			device_printf(dev,
4883 			    "# tx descriptors must be a power of 2");
4884 			err = EINVAL;
4885 			goto fail_iflib_detach;
4886 		}
4887 	}
4888 
4889 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4890 	    MAX_SINGLE_PACKET_FRACTION)
4891 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4892 		    MAX_SINGLE_PACKET_FRACTION);
4893 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4894 	    MAX_SINGLE_PACKET_FRACTION)
4895 		scctx->isc_tx_tso_segments_max = max(1,
4896 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4897 
4898 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4899 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4900 		/*
4901 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4902 		 * but some MACs do.
4903 		 */
4904 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4905 		    IP_MAXPACKET));
4906 		/*
4907 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4908 		 * into account.  In the worst case, each of these calls will
4909 		 * add another mbuf and, thus, the requirement for another DMA
4910 		 * segment.  So for best performance, it doesn't make sense to
4911 		 * advertize a maximum of TSO segments that typically will
4912 		 * require defragmentation in iflib_encap().
4913 		 */
4914 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4915 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4916 	}
4917 	if (scctx->isc_rss_table_size == 0)
4918 		scctx->isc_rss_table_size = 64;
4919 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4920 
4921 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4922 	/* XXX format name */
4923 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4924 	    NULL, NULL, "admin");
4925 
4926 	/* XXX --- can support > 1 -- but keep it simple for now */
4927 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4928 
4929 	/* Get memory for the station queues */
4930 	if ((err = iflib_queues_alloc(ctx))) {
4931 		device_printf(dev, "Unable to allocate queue memory\n");
4932 		goto fail_iflib_detach;
4933 	}
4934 
4935 	if ((err = iflib_qset_structures_setup(ctx))) {
4936 		device_printf(dev, "qset structure setup failed %d\n", err);
4937 		goto fail_queues;
4938 	}
4939 
4940 	/*
4941 	 * XXX What if anything do we want to do about interrupts?
4942 	 */
4943 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4944 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4945 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4946 		goto fail_detach;
4947 	}
4948 
4949 	/*
4950 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4951 	 * This must appear after the call to ether_ifattach() because
4952 	 * ether_ifattach() sets if_hdrlen to the default value.
4953 	 */
4954 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4955 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4956 
4957 	/* XXX handle more than one queue */
4958 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4959 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4960 
4961 	*ctxp = ctx;
4962 
4963 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4964 	iflib_add_device_sysctl_post(ctx);
4965 	ctx->ifc_flags |= IFC_INIT_DONE;
4966 	CTX_UNLOCK(ctx);
4967 	return (0);
4968 fail_detach:
4969 	ether_ifdetach(ctx->ifc_ifp);
4970 fail_queues:
4971 	iflib_tx_structures_free(ctx);
4972 	iflib_rx_structures_free(ctx);
4973 fail_iflib_detach:
4974 	IFDI_DETACH(ctx);
4975 fail_unlock:
4976 	CTX_UNLOCK(ctx);
4977 fail_ctx_free:
4978 	free(ctx->ifc_softc, M_IFLIB);
4979 	free(ctx, M_IFLIB);
4980 	return (err);
4981 }
4982 
4983 int
4984 iflib_pseudo_deregister(if_ctx_t ctx)
4985 {
4986 	if_t ifp = ctx->ifc_ifp;
4987 	iflib_txq_t txq;
4988 	iflib_rxq_t rxq;
4989 	int i, j;
4990 	struct taskqgroup *tqg;
4991 	iflib_fl_t fl;
4992 
4993 	/* Unregister VLAN events */
4994 	if (ctx->ifc_vlan_attach_event != NULL)
4995 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4996 	if (ctx->ifc_vlan_detach_event != NULL)
4997 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4998 
4999 	ether_ifdetach(ifp);
5000 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5001 	CTX_LOCK_DESTROY(ctx);
5002 	/* XXX drain any dependent tasks */
5003 	tqg = qgroup_if_io_tqg;
5004 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5005 		callout_drain(&txq->ift_timer);
5006 		if (txq->ift_task.gt_uniq != NULL)
5007 			taskqgroup_detach(tqg, &txq->ift_task);
5008 	}
5009 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5010 		if (rxq->ifr_task.gt_uniq != NULL)
5011 			taskqgroup_detach(tqg, &rxq->ifr_task);
5012 
5013 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5014 			free(fl->ifl_rx_bitmap, M_IFLIB);
5015 	}
5016 	tqg = qgroup_if_config_tqg;
5017 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5018 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5019 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5020 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5021 
5022 	if_free(ifp);
5023 
5024 	iflib_tx_structures_free(ctx);
5025 	iflib_rx_structures_free(ctx);
5026 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5027 		free(ctx->ifc_softc, M_IFLIB);
5028 	free(ctx, M_IFLIB);
5029 	return (0);
5030 }
5031 
5032 int
5033 iflib_device_attach(device_t dev)
5034 {
5035 	if_ctx_t ctx;
5036 	if_shared_ctx_t sctx;
5037 
5038 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5039 		return (ENOTSUP);
5040 
5041 	pci_enable_busmaster(dev);
5042 
5043 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5044 }
5045 
5046 int
5047 iflib_device_deregister(if_ctx_t ctx)
5048 {
5049 	if_t ifp = ctx->ifc_ifp;
5050 	iflib_txq_t txq;
5051 	iflib_rxq_t rxq;
5052 	device_t dev = ctx->ifc_dev;
5053 	int i, j;
5054 	struct taskqgroup *tqg;
5055 	iflib_fl_t fl;
5056 
5057 	/* Make sure VLANS are not using driver */
5058 	if (if_vlantrunkinuse(ifp)) {
5059 		device_printf(dev, "Vlan in use, detach first\n");
5060 		return (EBUSY);
5061 	}
5062 #ifdef PCI_IOV
5063 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5064 		device_printf(dev, "SR-IOV in use; detach first.\n");
5065 		return (EBUSY);
5066 	}
5067 #endif
5068 
5069 	STATE_LOCK(ctx);
5070 	ctx->ifc_flags |= IFC_IN_DETACH;
5071 	STATE_UNLOCK(ctx);
5072 
5073 	CTX_LOCK(ctx);
5074 	iflib_stop(ctx);
5075 	CTX_UNLOCK(ctx);
5076 
5077 	/* Unregister VLAN events */
5078 	if (ctx->ifc_vlan_attach_event != NULL)
5079 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5080 	if (ctx->ifc_vlan_detach_event != NULL)
5081 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5082 
5083 	iflib_netmap_detach(ifp);
5084 	ether_ifdetach(ifp);
5085 	iflib_rem_pfil(ctx);
5086 	if (ctx->ifc_led_dev != NULL)
5087 		led_destroy(ctx->ifc_led_dev);
5088 	/* XXX drain any dependent tasks */
5089 	tqg = qgroup_if_io_tqg;
5090 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5091 		callout_drain(&txq->ift_timer);
5092 		if (txq->ift_task.gt_uniq != NULL)
5093 			taskqgroup_detach(tqg, &txq->ift_task);
5094 	}
5095 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5096 		if (rxq->ifr_task.gt_uniq != NULL)
5097 			taskqgroup_detach(tqg, &rxq->ifr_task);
5098 
5099 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5100 			free(fl->ifl_rx_bitmap, M_IFLIB);
5101 	}
5102 	tqg = qgroup_if_config_tqg;
5103 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5104 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5105 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5106 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5107 	CTX_LOCK(ctx);
5108 	IFDI_DETACH(ctx);
5109 	CTX_UNLOCK(ctx);
5110 
5111 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5112 	CTX_LOCK_DESTROY(ctx);
5113 	device_set_softc(ctx->ifc_dev, NULL);
5114 	iflib_free_intr_mem(ctx);
5115 
5116 	bus_generic_detach(dev);
5117 	if_free(ifp);
5118 
5119 	iflib_tx_structures_free(ctx);
5120 	iflib_rx_structures_free(ctx);
5121 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5122 		free(ctx->ifc_softc, M_IFLIB);
5123 	unref_ctx_core_offset(ctx);
5124 	STATE_LOCK_DESTROY(ctx);
5125 	free(ctx, M_IFLIB);
5126 	return (0);
5127 }
5128 
5129 static void
5130 iflib_free_intr_mem(if_ctx_t ctx)
5131 {
5132 
5133 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5134 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5135 	}
5136 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5137 		pci_release_msi(ctx->ifc_dev);
5138 	}
5139 	if (ctx->ifc_msix_mem != NULL) {
5140 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5141 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5142 		ctx->ifc_msix_mem = NULL;
5143 	}
5144 }
5145 
5146 int
5147 iflib_device_detach(device_t dev)
5148 {
5149 	if_ctx_t ctx = device_get_softc(dev);
5150 
5151 	return (iflib_device_deregister(ctx));
5152 }
5153 
5154 int
5155 iflib_device_suspend(device_t dev)
5156 {
5157 	if_ctx_t ctx = device_get_softc(dev);
5158 
5159 	CTX_LOCK(ctx);
5160 	IFDI_SUSPEND(ctx);
5161 	CTX_UNLOCK(ctx);
5162 
5163 	return bus_generic_suspend(dev);
5164 }
5165 int
5166 iflib_device_shutdown(device_t dev)
5167 {
5168 	if_ctx_t ctx = device_get_softc(dev);
5169 
5170 	CTX_LOCK(ctx);
5171 	IFDI_SHUTDOWN(ctx);
5172 	CTX_UNLOCK(ctx);
5173 
5174 	return bus_generic_suspend(dev);
5175 }
5176 
5177 
5178 int
5179 iflib_device_resume(device_t dev)
5180 {
5181 	if_ctx_t ctx = device_get_softc(dev);
5182 	iflib_txq_t txq = ctx->ifc_txqs;
5183 
5184 	CTX_LOCK(ctx);
5185 	IFDI_RESUME(ctx);
5186 	iflib_if_init_locked(ctx);
5187 	CTX_UNLOCK(ctx);
5188 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5189 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5190 
5191 	return (bus_generic_resume(dev));
5192 }
5193 
5194 int
5195 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5196 {
5197 	int error;
5198 	if_ctx_t ctx = device_get_softc(dev);
5199 
5200 	CTX_LOCK(ctx);
5201 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5202 	CTX_UNLOCK(ctx);
5203 
5204 	return (error);
5205 }
5206 
5207 void
5208 iflib_device_iov_uninit(device_t dev)
5209 {
5210 	if_ctx_t ctx = device_get_softc(dev);
5211 
5212 	CTX_LOCK(ctx);
5213 	IFDI_IOV_UNINIT(ctx);
5214 	CTX_UNLOCK(ctx);
5215 }
5216 
5217 int
5218 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5219 {
5220 	int error;
5221 	if_ctx_t ctx = device_get_softc(dev);
5222 
5223 	CTX_LOCK(ctx);
5224 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5225 	CTX_UNLOCK(ctx);
5226 
5227 	return (error);
5228 }
5229 
5230 /*********************************************************************
5231  *
5232  *  MODULE FUNCTION DEFINITIONS
5233  *
5234  **********************************************************************/
5235 
5236 /*
5237  * - Start a fast taskqueue thread for each core
5238  * - Start a taskqueue for control operations
5239  */
5240 static int
5241 iflib_module_init(void)
5242 {
5243 	return (0);
5244 }
5245 
5246 static int
5247 iflib_module_event_handler(module_t mod, int what, void *arg)
5248 {
5249 	int err;
5250 
5251 	switch (what) {
5252 	case MOD_LOAD:
5253 		if ((err = iflib_module_init()) != 0)
5254 			return (err);
5255 		break;
5256 	case MOD_UNLOAD:
5257 		return (EBUSY);
5258 	default:
5259 		return (EOPNOTSUPP);
5260 	}
5261 
5262 	return (0);
5263 }
5264 
5265 /*********************************************************************
5266  *
5267  *  PUBLIC FUNCTION DEFINITIONS
5268  *     ordered as in iflib.h
5269  *
5270  **********************************************************************/
5271 
5272 
5273 static void
5274 _iflib_assert(if_shared_ctx_t sctx)
5275 {
5276 	MPASS(sctx->isc_tx_maxsize);
5277 	MPASS(sctx->isc_tx_maxsegsize);
5278 
5279 	MPASS(sctx->isc_rx_maxsize);
5280 	MPASS(sctx->isc_rx_nsegments);
5281 	MPASS(sctx->isc_rx_maxsegsize);
5282 
5283 	MPASS(sctx->isc_nrxd_min[0]);
5284 	MPASS(sctx->isc_nrxd_max[0]);
5285 	MPASS(sctx->isc_nrxd_default[0]);
5286 	MPASS(sctx->isc_ntxd_min[0]);
5287 	MPASS(sctx->isc_ntxd_max[0]);
5288 	MPASS(sctx->isc_ntxd_default[0]);
5289 }
5290 
5291 static void
5292 _iflib_pre_assert(if_softc_ctx_t scctx)
5293 {
5294 
5295 	MPASS(scctx->isc_txrx->ift_txd_encap);
5296 	MPASS(scctx->isc_txrx->ift_txd_flush);
5297 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5298 	MPASS(scctx->isc_txrx->ift_rxd_available);
5299 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5300 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5301 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5302 }
5303 
5304 static int
5305 iflib_register(if_ctx_t ctx)
5306 {
5307 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5308 	driver_t *driver = sctx->isc_driver;
5309 	device_t dev = ctx->ifc_dev;
5310 	if_t ifp;
5311 
5312 	_iflib_assert(sctx);
5313 
5314 	CTX_LOCK_INIT(ctx);
5315 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5316 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5317 	if (ifp == NULL) {
5318 		device_printf(dev, "can not allocate ifnet structure\n");
5319 		return (ENOMEM);
5320 	}
5321 
5322 	/*
5323 	 * Initialize our context's device specific methods
5324 	 */
5325 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5326 	kobj_class_compile((kobj_class_t) driver);
5327 	driver->refs++;
5328 
5329 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5330 	if_setsoftc(ifp, ctx);
5331 	if_setdev(ifp, dev);
5332 	if_setinitfn(ifp, iflib_if_init);
5333 	if_setioctlfn(ifp, iflib_if_ioctl);
5334 #ifdef ALTQ
5335 	if_setstartfn(ifp, iflib_altq_if_start);
5336 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5337 	if_setsendqready(ifp);
5338 #else
5339 	if_settransmitfn(ifp, iflib_if_transmit);
5340 #endif
5341 	if_setqflushfn(ifp, iflib_if_qflush);
5342 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5343 
5344 	ctx->ifc_vlan_attach_event =
5345 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5346 							  EVENTHANDLER_PRI_FIRST);
5347 	ctx->ifc_vlan_detach_event =
5348 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5349 							  EVENTHANDLER_PRI_FIRST);
5350 
5351 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5352 					 iflib_media_change, iflib_media_status);
5353 
5354 	return (0);
5355 }
5356 
5357 
5358 static int
5359 iflib_queues_alloc(if_ctx_t ctx)
5360 {
5361 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5362 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5363 	device_t dev = ctx->ifc_dev;
5364 	int nrxqsets = scctx->isc_nrxqsets;
5365 	int ntxqsets = scctx->isc_ntxqsets;
5366 	iflib_txq_t txq;
5367 	iflib_rxq_t rxq;
5368 	iflib_fl_t fl = NULL;
5369 	int i, j, cpu, err, txconf, rxconf;
5370 	iflib_dma_info_t ifdip;
5371 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5372 	uint32_t *txqsizes = scctx->isc_txqsizes;
5373 	uint8_t nrxqs = sctx->isc_nrxqs;
5374 	uint8_t ntxqs = sctx->isc_ntxqs;
5375 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5376 	caddr_t *vaddrs;
5377 	uint64_t *paddrs;
5378 
5379 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5380 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5381 
5382 	/* Allocate the TX ring struct memory */
5383 	if (!(ctx->ifc_txqs =
5384 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5385 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5386 		device_printf(dev, "Unable to allocate TX ring memory\n");
5387 		err = ENOMEM;
5388 		goto fail;
5389 	}
5390 
5391 	/* Now allocate the RX */
5392 	if (!(ctx->ifc_rxqs =
5393 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5394 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5395 		device_printf(dev, "Unable to allocate RX ring memory\n");
5396 		err = ENOMEM;
5397 		goto rx_fail;
5398 	}
5399 
5400 	txq = ctx->ifc_txqs;
5401 	rxq = ctx->ifc_rxqs;
5402 
5403 	/*
5404 	 * XXX handle allocation failure
5405 	 */
5406 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5407 		/* Set up some basics */
5408 
5409 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5410 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5411 			device_printf(dev,
5412 			    "Unable to allocate TX DMA info memory\n");
5413 			err = ENOMEM;
5414 			goto err_tx_desc;
5415 		}
5416 		txq->ift_ifdi = ifdip;
5417 		for (j = 0; j < ntxqs; j++, ifdip++) {
5418 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5419 				device_printf(dev,
5420 				    "Unable to allocate TX descriptors\n");
5421 				err = ENOMEM;
5422 				goto err_tx_desc;
5423 			}
5424 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5425 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5426 		}
5427 		txq->ift_ctx = ctx;
5428 		txq->ift_id = i;
5429 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5430 			txq->ift_br_offset = 1;
5431 		} else {
5432 			txq->ift_br_offset = 0;
5433 		}
5434 		/* XXX fix this */
5435 		txq->ift_timer.c_cpu = cpu;
5436 
5437 		if (iflib_txsd_alloc(txq)) {
5438 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5439 			err = ENOMEM;
5440 			goto err_tx_desc;
5441 		}
5442 
5443 		/* Initialize the TX lock */
5444 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5445 		    device_get_nameunit(dev), txq->ift_id);
5446 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5447 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5448 
5449 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5450 			 device_get_nameunit(dev), txq->ift_id);
5451 
5452 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5453 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5454 		if (err) {
5455 			/* XXX free any allocated rings */
5456 			device_printf(dev, "Unable to allocate buf_ring\n");
5457 			goto err_tx_desc;
5458 		}
5459 	}
5460 
5461 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5462 		/* Set up some basics */
5463 
5464 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5465 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5466 			device_printf(dev,
5467 			    "Unable to allocate RX DMA info memory\n");
5468 			err = ENOMEM;
5469 			goto err_tx_desc;
5470 		}
5471 
5472 		rxq->ifr_ifdi = ifdip;
5473 		/* XXX this needs to be changed if #rx queues != #tx queues */
5474 		rxq->ifr_ntxqirq = 1;
5475 		rxq->ifr_txqid[0] = i;
5476 		for (j = 0; j < nrxqs; j++, ifdip++) {
5477 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5478 				device_printf(dev,
5479 				    "Unable to allocate RX descriptors\n");
5480 				err = ENOMEM;
5481 				goto err_tx_desc;
5482 			}
5483 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5484 		}
5485 		rxq->ifr_ctx = ctx;
5486 		rxq->ifr_id = i;
5487 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5488 			rxq->ifr_fl_offset = 1;
5489 		} else {
5490 			rxq->ifr_fl_offset = 0;
5491 		}
5492 		rxq->ifr_nfl = nfree_lists;
5493 		if (!(fl =
5494 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5495 			device_printf(dev, "Unable to allocate free list memory\n");
5496 			err = ENOMEM;
5497 			goto err_tx_desc;
5498 		}
5499 		rxq->ifr_fl = fl;
5500 		for (j = 0; j < nfree_lists; j++) {
5501 			fl[j].ifl_rxq = rxq;
5502 			fl[j].ifl_id = j;
5503 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5504 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5505 		}
5506 		/* Allocate receive buffers for the ring */
5507 		if (iflib_rxsd_alloc(rxq)) {
5508 			device_printf(dev,
5509 			    "Critical Failure setting up receive buffers\n");
5510 			err = ENOMEM;
5511 			goto err_rx_desc;
5512 		}
5513 
5514 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5515 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5516 			    M_WAITOK);
5517 	}
5518 
5519 	/* TXQs */
5520 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5521 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5522 	for (i = 0; i < ntxqsets; i++) {
5523 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5524 
5525 		for (j = 0; j < ntxqs; j++, di++) {
5526 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5527 			paddrs[i*ntxqs + j] = di->idi_paddr;
5528 		}
5529 	}
5530 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5531 		device_printf(ctx->ifc_dev,
5532 		    "Unable to allocate device TX queue\n");
5533 		iflib_tx_structures_free(ctx);
5534 		free(vaddrs, M_IFLIB);
5535 		free(paddrs, M_IFLIB);
5536 		goto err_rx_desc;
5537 	}
5538 	free(vaddrs, M_IFLIB);
5539 	free(paddrs, M_IFLIB);
5540 
5541 	/* RXQs */
5542 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5543 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5544 	for (i = 0; i < nrxqsets; i++) {
5545 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5546 
5547 		for (j = 0; j < nrxqs; j++, di++) {
5548 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5549 			paddrs[i*nrxqs + j] = di->idi_paddr;
5550 		}
5551 	}
5552 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5553 		device_printf(ctx->ifc_dev,
5554 		    "Unable to allocate device RX queue\n");
5555 		iflib_tx_structures_free(ctx);
5556 		free(vaddrs, M_IFLIB);
5557 		free(paddrs, M_IFLIB);
5558 		goto err_rx_desc;
5559 	}
5560 	free(vaddrs, M_IFLIB);
5561 	free(paddrs, M_IFLIB);
5562 
5563 	return (0);
5564 
5565 /* XXX handle allocation failure changes */
5566 err_rx_desc:
5567 err_tx_desc:
5568 rx_fail:
5569 	if (ctx->ifc_rxqs != NULL)
5570 		free(ctx->ifc_rxqs, M_IFLIB);
5571 	ctx->ifc_rxqs = NULL;
5572 	if (ctx->ifc_txqs != NULL)
5573 		free(ctx->ifc_txqs, M_IFLIB);
5574 	ctx->ifc_txqs = NULL;
5575 fail:
5576 	return (err);
5577 }
5578 
5579 static int
5580 iflib_tx_structures_setup(if_ctx_t ctx)
5581 {
5582 	iflib_txq_t txq = ctx->ifc_txqs;
5583 	int i;
5584 
5585 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5586 		iflib_txq_setup(txq);
5587 
5588 	return (0);
5589 }
5590 
5591 static void
5592 iflib_tx_structures_free(if_ctx_t ctx)
5593 {
5594 	iflib_txq_t txq = ctx->ifc_txqs;
5595 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5596 	int i, j;
5597 
5598 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5599 		iflib_txq_destroy(txq);
5600 		for (j = 0; j < sctx->isc_ntxqs; j++)
5601 			iflib_dma_free(&txq->ift_ifdi[j]);
5602 	}
5603 	free(ctx->ifc_txqs, M_IFLIB);
5604 	ctx->ifc_txqs = NULL;
5605 	IFDI_QUEUES_FREE(ctx);
5606 }
5607 
5608 /*********************************************************************
5609  *
5610  *  Initialize all receive rings.
5611  *
5612  **********************************************************************/
5613 static int
5614 iflib_rx_structures_setup(if_ctx_t ctx)
5615 {
5616 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5617 	int q;
5618 #if defined(INET6) || defined(INET)
5619 	int i, err;
5620 #endif
5621 
5622 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5623 #if defined(INET6) || defined(INET)
5624 		tcp_lro_free(&rxq->ifr_lc);
5625 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5626 		    TCP_LRO_ENTRIES, min(1024,
5627 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5628 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5629 			goto fail;
5630 		}
5631 		rxq->ifr_lro_enabled = TRUE;
5632 #endif
5633 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5634 	}
5635 	return (0);
5636 #if defined(INET6) || defined(INET)
5637 fail:
5638 	/*
5639 	 * Free RX software descriptors allocated so far, we will only handle
5640 	 * the rings that completed, the failing case will have
5641 	 * cleaned up for itself. 'q' failed, so its the terminus.
5642 	 */
5643 	rxq = ctx->ifc_rxqs;
5644 	for (i = 0; i < q; ++i, rxq++) {
5645 		iflib_rx_sds_free(rxq);
5646 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5647 	}
5648 	return (err);
5649 #endif
5650 }
5651 
5652 /*********************************************************************
5653  *
5654  *  Free all receive rings.
5655  *
5656  **********************************************************************/
5657 static void
5658 iflib_rx_structures_free(if_ctx_t ctx)
5659 {
5660 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5661 
5662 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5663 		iflib_rx_sds_free(rxq);
5664 	}
5665 	free(ctx->ifc_rxqs, M_IFLIB);
5666 	ctx->ifc_rxqs = NULL;
5667 }
5668 
5669 static int
5670 iflib_qset_structures_setup(if_ctx_t ctx)
5671 {
5672 	int err;
5673 
5674 	/*
5675 	 * It is expected that the caller takes care of freeing queues if this
5676 	 * fails.
5677 	 */
5678 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5679 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5680 		return (err);
5681 	}
5682 
5683 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5684 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5685 
5686 	return (err);
5687 }
5688 
5689 int
5690 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5691 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5692 {
5693 
5694 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5695 }
5696 
5697 #ifdef SMP
5698 static int
5699 find_nth(if_ctx_t ctx, int qid)
5700 {
5701 	cpuset_t cpus;
5702 	int i, cpuid, eqid, count;
5703 
5704 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5705 	count = CPU_COUNT(&cpus);
5706 	eqid = qid % count;
5707 	/* clear up to the qid'th bit */
5708 	for (i = 0; i < eqid; i++) {
5709 		cpuid = CPU_FFS(&cpus);
5710 		MPASS(cpuid != 0);
5711 		CPU_CLR(cpuid-1, &cpus);
5712 	}
5713 	cpuid = CPU_FFS(&cpus);
5714 	MPASS(cpuid != 0);
5715 	return (cpuid-1);
5716 }
5717 
5718 #ifdef SCHED_ULE
5719 extern struct cpu_group *cpu_top;              /* CPU topology */
5720 
5721 static int
5722 find_child_with_core(int cpu, struct cpu_group *grp)
5723 {
5724 	int i;
5725 
5726 	if (grp->cg_children == 0)
5727 		return -1;
5728 
5729 	MPASS(grp->cg_child);
5730 	for (i = 0; i < grp->cg_children; i++) {
5731 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5732 			return i;
5733 	}
5734 
5735 	return -1;
5736 }
5737 
5738 /*
5739  * Find the nth "close" core to the specified core
5740  * "close" is defined as the deepest level that shares
5741  * at least an L2 cache.  With threads, this will be
5742  * threads on the same core.  If the shared cache is L3
5743  * or higher, simply returns the same core.
5744  */
5745 static int
5746 find_close_core(int cpu, int core_offset)
5747 {
5748 	struct cpu_group *grp;
5749 	int i;
5750 	int fcpu;
5751 	cpuset_t cs;
5752 
5753 	grp = cpu_top;
5754 	if (grp == NULL)
5755 		return cpu;
5756 	i = 0;
5757 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5758 		/* If the child only has one cpu, don't descend */
5759 		if (grp->cg_child[i].cg_count <= 1)
5760 			break;
5761 		grp = &grp->cg_child[i];
5762 	}
5763 
5764 	/* If they don't share at least an L2 cache, use the same CPU */
5765 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5766 		return cpu;
5767 
5768 	/* Now pick one */
5769 	CPU_COPY(&grp->cg_mask, &cs);
5770 
5771 	/* Add the selected CPU offset to core offset. */
5772 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5773 		if (fcpu - 1 == cpu)
5774 			break;
5775 		CPU_CLR(fcpu - 1, &cs);
5776 	}
5777 	MPASS(fcpu);
5778 
5779 	core_offset += i;
5780 
5781 	CPU_COPY(&grp->cg_mask, &cs);
5782 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5783 		MPASS(CPU_FFS(&cs));
5784 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5785 	}
5786 	MPASS(CPU_FFS(&cs));
5787 	return CPU_FFS(&cs) - 1;
5788 }
5789 #else
5790 static int
5791 find_close_core(int cpu, int core_offset __unused)
5792 {
5793 	return cpu;
5794 }
5795 #endif
5796 
5797 static int
5798 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5799 {
5800 	switch (type) {
5801 	case IFLIB_INTR_TX:
5802 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5803 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5804 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5805 	case IFLIB_INTR_RX:
5806 	case IFLIB_INTR_RXTX:
5807 		/* RX queues get the specified core */
5808 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5809 	default:
5810 		return -1;
5811 	}
5812 }
5813 #else
5814 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5815 #define find_close_core(cpuid, tid)	CPU_FIRST()
5816 #define find_nth(ctx, gid)		CPU_FIRST()
5817 #endif
5818 
5819 /* Just to avoid copy/paste */
5820 static inline int
5821 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5822     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5823     const char *name)
5824 {
5825 	device_t dev;
5826 	int co, cpuid, err, tid;
5827 
5828 	dev = ctx->ifc_dev;
5829 	co = ctx->ifc_sysctl_core_offset;
5830 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5831 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5832 	cpuid = find_nth(ctx, qid + co);
5833 	tid = get_core_offset(ctx, type, qid);
5834 	MPASS(tid >= 0);
5835 	cpuid = find_close_core(cpuid, tid);
5836 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5837 	    name);
5838 	if (err) {
5839 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5840 		return (err);
5841 	}
5842 #ifdef notyet
5843 	if (cpuid > ctx->ifc_cpuid_highest)
5844 		ctx->ifc_cpuid_highest = cpuid;
5845 #endif
5846 	return 0;
5847 }
5848 
5849 int
5850 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5851 			iflib_intr_type_t type, driver_filter_t *filter,
5852 			void *filter_arg, int qid, const char *name)
5853 {
5854 	device_t dev;
5855 	struct grouptask *gtask;
5856 	struct taskqgroup *tqg;
5857 	iflib_filter_info_t info;
5858 	gtask_fn_t *fn;
5859 	int tqrid, err;
5860 	driver_filter_t *intr_fast;
5861 	void *q;
5862 
5863 	info = &ctx->ifc_filter_info;
5864 	tqrid = rid;
5865 
5866 	switch (type) {
5867 	/* XXX merge tx/rx for netmap? */
5868 	case IFLIB_INTR_TX:
5869 		q = &ctx->ifc_txqs[qid];
5870 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5871 		gtask = &ctx->ifc_txqs[qid].ift_task;
5872 		tqg = qgroup_if_io_tqg;
5873 		fn = _task_fn_tx;
5874 		intr_fast = iflib_fast_intr;
5875 		GROUPTASK_INIT(gtask, 0, fn, q);
5876 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5877 		break;
5878 	case IFLIB_INTR_RX:
5879 		q = &ctx->ifc_rxqs[qid];
5880 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5881 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5882 		tqg = qgroup_if_io_tqg;
5883 		fn = _task_fn_rx;
5884 		intr_fast = iflib_fast_intr;
5885 		GROUPTASK_INIT(gtask, 0, fn, q);
5886 		break;
5887 	case IFLIB_INTR_RXTX:
5888 		q = &ctx->ifc_rxqs[qid];
5889 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5890 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5891 		tqg = qgroup_if_io_tqg;
5892 		fn = _task_fn_rx;
5893 		intr_fast = iflib_fast_intr_rxtx;
5894 		GROUPTASK_INIT(gtask, 0, fn, q);
5895 		break;
5896 	case IFLIB_INTR_ADMIN:
5897 		q = ctx;
5898 		tqrid = -1;
5899 		info = &ctx->ifc_filter_info;
5900 		gtask = &ctx->ifc_admin_task;
5901 		tqg = qgroup_if_config_tqg;
5902 		fn = _task_fn_admin;
5903 		intr_fast = iflib_fast_intr_ctx;
5904 		break;
5905 	default:
5906 		panic("unknown net intr type");
5907 	}
5908 
5909 	info->ifi_filter = filter;
5910 	info->ifi_filter_arg = filter_arg;
5911 	info->ifi_task = gtask;
5912 	info->ifi_ctx = q;
5913 
5914 	dev = ctx->ifc_dev;
5915 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5916 	if (err != 0) {
5917 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5918 		return (err);
5919 	}
5920 	if (type == IFLIB_INTR_ADMIN)
5921 		return (0);
5922 
5923 	if (tqrid != -1) {
5924 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5925 		    q, name);
5926 		if (err)
5927 			return (err);
5928 	} else {
5929 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5930 	}
5931 
5932 	return (0);
5933 }
5934 
5935 void
5936 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5937 {
5938 	struct grouptask *gtask;
5939 	struct taskqgroup *tqg;
5940 	gtask_fn_t *fn;
5941 	void *q;
5942 	int err;
5943 
5944 	switch (type) {
5945 	case IFLIB_INTR_TX:
5946 		q = &ctx->ifc_txqs[qid];
5947 		gtask = &ctx->ifc_txqs[qid].ift_task;
5948 		tqg = qgroup_if_io_tqg;
5949 		fn = _task_fn_tx;
5950 		break;
5951 	case IFLIB_INTR_RX:
5952 		q = &ctx->ifc_rxqs[qid];
5953 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5954 		tqg = qgroup_if_io_tqg;
5955 		fn = _task_fn_rx;
5956 		break;
5957 	case IFLIB_INTR_IOV:
5958 		q = ctx;
5959 		gtask = &ctx->ifc_vflr_task;
5960 		tqg = qgroup_if_config_tqg;
5961 		fn = _task_fn_iov;
5962 		break;
5963 	default:
5964 		panic("unknown net intr type");
5965 	}
5966 	GROUPTASK_INIT(gtask, 0, fn, q);
5967 	if (irq != NULL) {
5968 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5969 		    q, name);
5970 		if (err)
5971 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5972 			    irq->ii_res, name);
5973 	} else {
5974 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5975 	}
5976 }
5977 
5978 void
5979 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5980 {
5981 
5982 	if (irq->ii_tag)
5983 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5984 
5985 	if (irq->ii_res)
5986 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5987 		    rman_get_rid(irq->ii_res), irq->ii_res);
5988 }
5989 
5990 static int
5991 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5992 {
5993 	iflib_txq_t txq = ctx->ifc_txqs;
5994 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5995 	if_irq_t irq = &ctx->ifc_legacy_irq;
5996 	iflib_filter_info_t info;
5997 	device_t dev;
5998 	struct grouptask *gtask;
5999 	struct resource *res;
6000 	struct taskqgroup *tqg;
6001 	gtask_fn_t *fn;
6002 	int tqrid;
6003 	void *q;
6004 	int err;
6005 
6006 	q = &ctx->ifc_rxqs[0];
6007 	info = &rxq[0].ifr_filter_info;
6008 	gtask = &rxq[0].ifr_task;
6009 	tqg = qgroup_if_io_tqg;
6010 	tqrid = irq->ii_rid = *rid;
6011 	fn = _task_fn_rx;
6012 
6013 	ctx->ifc_flags |= IFC_LEGACY;
6014 	info->ifi_filter = filter;
6015 	info->ifi_filter_arg = filter_arg;
6016 	info->ifi_task = gtask;
6017 	info->ifi_ctx = ctx;
6018 
6019 	dev = ctx->ifc_dev;
6020 	/* We allocate a single interrupt resource */
6021 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
6022 		return (err);
6023 	GROUPTASK_INIT(gtask, 0, fn, q);
6024 	res = irq->ii_res;
6025 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6026 
6027 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6028 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6029 	    "tx");
6030 	return (0);
6031 }
6032 
6033 void
6034 iflib_led_create(if_ctx_t ctx)
6035 {
6036 
6037 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6038 	    device_get_nameunit(ctx->ifc_dev));
6039 }
6040 
6041 void
6042 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6043 {
6044 
6045 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6046 }
6047 
6048 void
6049 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6050 {
6051 
6052 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6053 }
6054 
6055 void
6056 iflib_admin_intr_deferred(if_ctx_t ctx)
6057 {
6058 #ifdef INVARIANTS
6059 	struct grouptask *gtask;
6060 
6061 	gtask = &ctx->ifc_admin_task;
6062 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6063 #endif
6064 
6065 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6066 }
6067 
6068 void
6069 iflib_iov_intr_deferred(if_ctx_t ctx)
6070 {
6071 
6072 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6073 }
6074 
6075 void
6076 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
6077 {
6078 
6079 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6080 	    name);
6081 }
6082 
6083 void
6084 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6085 	const char *name)
6086 {
6087 
6088 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6089 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6090 	    name);
6091 }
6092 
6093 void
6094 iflib_config_gtask_deinit(struct grouptask *gtask)
6095 {
6096 
6097 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6098 }
6099 
6100 void
6101 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6102 {
6103 	if_t ifp = ctx->ifc_ifp;
6104 	iflib_txq_t txq = ctx->ifc_txqs;
6105 
6106 	if_setbaudrate(ifp, baudrate);
6107 	if (baudrate >= IF_Gbps(10)) {
6108 		STATE_LOCK(ctx);
6109 		ctx->ifc_flags |= IFC_PREFETCH;
6110 		STATE_UNLOCK(ctx);
6111 	}
6112 	/* If link down, disable watchdog */
6113 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6114 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6115 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6116 	}
6117 	ctx->ifc_link_state = link_state;
6118 	if_link_state_change(ifp, link_state);
6119 }
6120 
6121 static int
6122 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6123 {
6124 	int credits;
6125 #ifdef INVARIANTS
6126 	int credits_pre = txq->ift_cidx_processed;
6127 #endif
6128 
6129 	if (ctx->isc_txd_credits_update == NULL)
6130 		return (0);
6131 
6132 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6133 	    BUS_DMASYNC_POSTREAD);
6134 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6135 		return (0);
6136 
6137 	txq->ift_processed += credits;
6138 	txq->ift_cidx_processed += credits;
6139 
6140 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6141 	if (txq->ift_cidx_processed >= txq->ift_size)
6142 		txq->ift_cidx_processed -= txq->ift_size;
6143 	return (credits);
6144 }
6145 
6146 static int
6147 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6148 {
6149 	iflib_fl_t fl;
6150 	u_int i;
6151 
6152 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6153 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6154 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6155 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6156 	    budget));
6157 }
6158 
6159 void
6160 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6161 	const char *description, if_int_delay_info_t info,
6162 	int offset, int value)
6163 {
6164 	info->iidi_ctx = ctx;
6165 	info->iidi_offset = offset;
6166 	info->iidi_value = value;
6167 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6168 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6169 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6170 	    info, 0, iflib_sysctl_int_delay, "I", description);
6171 }
6172 
6173 struct sx *
6174 iflib_ctx_lock_get(if_ctx_t ctx)
6175 {
6176 
6177 	return (&ctx->ifc_ctx_sx);
6178 }
6179 
6180 static int
6181 iflib_msix_init(if_ctx_t ctx)
6182 {
6183 	device_t dev = ctx->ifc_dev;
6184 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6185 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6186 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6187 	int iflib_num_tx_queues, iflib_num_rx_queues;
6188 	int err, admincnt, bar;
6189 
6190 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6191 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6192 
6193 	if (bootverbose)
6194 		device_printf(dev, "msix_init qsets capped at %d\n",
6195 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6196 
6197 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6198 	admincnt = sctx->isc_admin_intrcnt;
6199 	/* Override by tuneable */
6200 	if (scctx->isc_disable_msix)
6201 		goto msi;
6202 
6203 	/* First try MSI-X */
6204 	if ((msgs = pci_msix_count(dev)) == 0) {
6205 		if (bootverbose)
6206 			device_printf(dev, "MSI-X not supported or disabled\n");
6207 		goto msi;
6208 	}
6209 	/*
6210 	 * bar == -1 => "trust me I know what I'm doing"
6211 	 * Some drivers are for hardware that is so shoddily
6212 	 * documented that no one knows which bars are which
6213 	 * so the developer has to map all bars. This hack
6214 	 * allows shoddy garbage to use MSI-X in this framework.
6215 	 */
6216 	if (bar != -1) {
6217 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6218 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6219 		if (ctx->ifc_msix_mem == NULL) {
6220 			device_printf(dev, "Unable to map MSI-X table\n");
6221 			goto msi;
6222 		}
6223 	}
6224 #if IFLIB_DEBUG
6225 	/* use only 1 qset in debug mode */
6226 	queuemsgs = min(msgs - admincnt, 1);
6227 #else
6228 	queuemsgs = msgs - admincnt;
6229 #endif
6230 #ifdef RSS
6231 	queues = imin(queuemsgs, rss_getnumbuckets());
6232 #else
6233 	queues = queuemsgs;
6234 #endif
6235 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6236 	if (bootverbose)
6237 		device_printf(dev,
6238 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6239 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6240 #ifdef  RSS
6241 	/* If we're doing RSS, clamp at the number of RSS buckets */
6242 	if (queues > rss_getnumbuckets())
6243 		queues = rss_getnumbuckets();
6244 #endif
6245 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6246 		rx_queues = iflib_num_rx_queues;
6247 	else
6248 		rx_queues = queues;
6249 
6250 	if (rx_queues > scctx->isc_nrxqsets)
6251 		rx_queues = scctx->isc_nrxqsets;
6252 
6253 	/*
6254 	 * We want this to be all logical CPUs by default
6255 	 */
6256 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6257 		tx_queues = iflib_num_tx_queues;
6258 	else
6259 		tx_queues = mp_ncpus;
6260 
6261 	if (tx_queues > scctx->isc_ntxqsets)
6262 		tx_queues = scctx->isc_ntxqsets;
6263 
6264 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6265 #ifdef INVARIANTS
6266 		if (tx_queues != rx_queues)
6267 			device_printf(dev,
6268 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6269 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6270 #endif
6271 		tx_queues = min(rx_queues, tx_queues);
6272 		rx_queues = min(rx_queues, tx_queues);
6273 	}
6274 
6275 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6276 	    rx_queues, tx_queues);
6277 
6278 	vectors = rx_queues + admincnt;
6279 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6280 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6281 		    vectors);
6282 		scctx->isc_vectors = vectors;
6283 		scctx->isc_nrxqsets = rx_queues;
6284 		scctx->isc_ntxqsets = tx_queues;
6285 		scctx->isc_intr = IFLIB_INTR_MSIX;
6286 
6287 		return (vectors);
6288 	} else {
6289 		device_printf(dev,
6290 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6291 		    vectors, err);
6292 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6293 		    ctx->ifc_msix_mem);
6294 		ctx->ifc_msix_mem = NULL;
6295 	}
6296 msi:
6297 	vectors = pci_msi_count(dev);
6298 	scctx->isc_nrxqsets = 1;
6299 	scctx->isc_ntxqsets = 1;
6300 	scctx->isc_vectors = vectors;
6301 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6302 		device_printf(dev,"Using an MSI interrupt\n");
6303 		scctx->isc_intr = IFLIB_INTR_MSI;
6304 	} else {
6305 		scctx->isc_vectors = 1;
6306 		device_printf(dev,"Using a Legacy interrupt\n");
6307 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6308 	}
6309 
6310 	return (vectors);
6311 }
6312 
6313 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6314 
6315 static int
6316 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6317 {
6318 	int rc;
6319 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6320 	struct sbuf *sb;
6321 	const char *ring_state = "UNKNOWN";
6322 
6323 	/* XXX needed ? */
6324 	rc = sysctl_wire_old_buffer(req, 0);
6325 	MPASS(rc == 0);
6326 	if (rc != 0)
6327 		return (rc);
6328 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6329 	MPASS(sb != NULL);
6330 	if (sb == NULL)
6331 		return (ENOMEM);
6332 	if (state[3] <= 3)
6333 		ring_state = ring_states[state[3]];
6334 
6335 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6336 		    state[0], state[1], state[2], ring_state);
6337 	rc = sbuf_finish(sb);
6338 	sbuf_delete(sb);
6339         return(rc);
6340 }
6341 
6342 enum iflib_ndesc_handler {
6343 	IFLIB_NTXD_HANDLER,
6344 	IFLIB_NRXD_HANDLER,
6345 };
6346 
6347 static int
6348 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6349 {
6350 	if_ctx_t ctx = (void *)arg1;
6351 	enum iflib_ndesc_handler type = arg2;
6352 	char buf[256] = {0};
6353 	qidx_t *ndesc;
6354 	char *p, *next;
6355 	int nqs, rc, i;
6356 
6357 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6358 
6359 	nqs = 8;
6360 	switch(type) {
6361 	case IFLIB_NTXD_HANDLER:
6362 		ndesc = ctx->ifc_sysctl_ntxds;
6363 		if (ctx->ifc_sctx)
6364 			nqs = ctx->ifc_sctx->isc_ntxqs;
6365 		break;
6366 	case IFLIB_NRXD_HANDLER:
6367 		ndesc = ctx->ifc_sysctl_nrxds;
6368 		if (ctx->ifc_sctx)
6369 			nqs = ctx->ifc_sctx->isc_nrxqs;
6370 		break;
6371 	default:
6372 			panic("unhandled type");
6373 	}
6374 	if (nqs == 0)
6375 		nqs = 8;
6376 
6377 	for (i=0; i<8; i++) {
6378 		if (i >= nqs)
6379 			break;
6380 		if (i)
6381 			strcat(buf, ",");
6382 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6383 	}
6384 
6385 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6386 	if (rc || req->newptr == NULL)
6387 		return rc;
6388 
6389 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6390 	    i++, p = strsep(&next, " ,")) {
6391 		ndesc[i] = strtoul(p, NULL, 10);
6392 	}
6393 
6394 	return(rc);
6395 }
6396 
6397 #define NAME_BUFLEN 32
6398 static void
6399 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6400 {
6401         device_t dev = iflib_get_dev(ctx);
6402 	struct sysctl_oid_list *child, *oid_list;
6403 	struct sysctl_ctx_list *ctx_list;
6404 	struct sysctl_oid *node;
6405 
6406 	ctx_list = device_get_sysctl_ctx(dev);
6407 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6408 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6409 						      CTLFLAG_RD, NULL, "IFLIB fields");
6410 	oid_list = SYSCTL_CHILDREN(node);
6411 
6412 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6413 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6414 		       "driver version");
6415 
6416 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6417 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6418 			"# of txqs to use, 0 => use default #");
6419 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6420 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6421 			"# of rxqs to use, 0 => use default #");
6422 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6423 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6424                        "permit #txq != #rxq");
6425 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6426                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6427                       "disable MSI-X (default 0)");
6428 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6429 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6430                        "set the rx budget");
6431 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6432 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6433 		       "cause tx to abdicate instead of running to completion");
6434 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6435 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6436 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6437 		       "offset to start using cores at");
6438 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6439 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6440 		       "use separate cores for TX and RX");
6441 
6442 	/* XXX change for per-queue sizes */
6443 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6444 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6445                        mp_ndesc_handler, "A",
6446                        "list of # of tx descriptors to use, 0 = use default #");
6447 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6448 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6449                        mp_ndesc_handler, "A",
6450                        "list of # of rx descriptors to use, 0 = use default #");
6451 }
6452 
6453 static void
6454 iflib_add_device_sysctl_post(if_ctx_t ctx)
6455 {
6456 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6457 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6458         device_t dev = iflib_get_dev(ctx);
6459 	struct sysctl_oid_list *child;
6460 	struct sysctl_ctx_list *ctx_list;
6461 	iflib_fl_t fl;
6462 	iflib_txq_t txq;
6463 	iflib_rxq_t rxq;
6464 	int i, j;
6465 	char namebuf[NAME_BUFLEN];
6466 	char *qfmt;
6467 	struct sysctl_oid *queue_node, *fl_node, *node;
6468 	struct sysctl_oid_list *queue_list, *fl_list;
6469 	ctx_list = device_get_sysctl_ctx(dev);
6470 
6471 	node = ctx->ifc_sysctl_node;
6472 	child = SYSCTL_CHILDREN(node);
6473 
6474 	if (scctx->isc_ntxqsets > 100)
6475 		qfmt = "txq%03d";
6476 	else if (scctx->isc_ntxqsets > 10)
6477 		qfmt = "txq%02d";
6478 	else
6479 		qfmt = "txq%d";
6480 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6481 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6482 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6483 					     CTLFLAG_RD, NULL, "Queue Name");
6484 		queue_list = SYSCTL_CHILDREN(queue_node);
6485 #if MEMORY_LOGGING
6486 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6487 				CTLFLAG_RD,
6488 				&txq->ift_dequeued, "total mbufs freed");
6489 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6490 				CTLFLAG_RD,
6491 				&txq->ift_enqueued, "total mbufs enqueued");
6492 #endif
6493 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6494 				   CTLFLAG_RD,
6495 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6496 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6497 				   CTLFLAG_RD,
6498 				   &txq->ift_pullups, "# of times m_pullup was called");
6499 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6500 				   CTLFLAG_RD,
6501 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6502 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6503 				   CTLFLAG_RD,
6504 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6505 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6506 				   CTLFLAG_RD,
6507 				   &txq->ift_map_failed, "# of times dma map failed");
6508 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6509 				   CTLFLAG_RD,
6510 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6511 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6512 				   CTLFLAG_RD,
6513 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6514 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6515 				   CTLFLAG_RD,
6516 				   &txq->ift_pidx, 1, "Producer Index");
6517 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6518 				   CTLFLAG_RD,
6519 				   &txq->ift_cidx, 1, "Consumer Index");
6520 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6521 				   CTLFLAG_RD,
6522 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6523 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6524 				   CTLFLAG_RD,
6525 				   &txq->ift_in_use, 1, "descriptors in use");
6526 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6527 				   CTLFLAG_RD,
6528 				   &txq->ift_processed, "descriptors procesed for clean");
6529 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6530 				   CTLFLAG_RD,
6531 				   &txq->ift_cleaned, "total cleaned");
6532 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6533 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6534 				0, mp_ring_state_handler, "A", "soft ring state");
6535 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6536 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6537 				       "# of enqueues to the mp_ring for this queue");
6538 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6539 				       CTLFLAG_RD, &txq->ift_br->drops,
6540 				       "# of drops in the mp_ring for this queue");
6541 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6542 				       CTLFLAG_RD, &txq->ift_br->starts,
6543 				       "# of normal consumer starts in the mp_ring for this queue");
6544 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6545 				       CTLFLAG_RD, &txq->ift_br->stalls,
6546 					       "# of consumer stalls in the mp_ring for this queue");
6547 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6548 			       CTLFLAG_RD, &txq->ift_br->restarts,
6549 				       "# of consumer restarts in the mp_ring for this queue");
6550 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6551 				       CTLFLAG_RD, &txq->ift_br->abdications,
6552 				       "# of consumer abdications in the mp_ring for this queue");
6553 	}
6554 
6555 	if (scctx->isc_nrxqsets > 100)
6556 		qfmt = "rxq%03d";
6557 	else if (scctx->isc_nrxqsets > 10)
6558 		qfmt = "rxq%02d";
6559 	else
6560 		qfmt = "rxq%d";
6561 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6562 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6563 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6564 					     CTLFLAG_RD, NULL, "Queue Name");
6565 		queue_list = SYSCTL_CHILDREN(queue_node);
6566 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6567 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6568 				       CTLFLAG_RD,
6569 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6570 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6571 				       CTLFLAG_RD,
6572 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6573 		}
6574 
6575 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6576 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6577 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6578 						     CTLFLAG_RD, NULL, "freelist Name");
6579 			fl_list = SYSCTL_CHILDREN(fl_node);
6580 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6581 				       CTLFLAG_RD,
6582 				       &fl->ifl_pidx, 1, "Producer Index");
6583 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6584 				       CTLFLAG_RD,
6585 				       &fl->ifl_cidx, 1, "Consumer Index");
6586 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6587 				       CTLFLAG_RD,
6588 				       &fl->ifl_credits, 1, "credits available");
6589 #if MEMORY_LOGGING
6590 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6591 					CTLFLAG_RD,
6592 					&fl->ifl_m_enqueued, "mbufs allocated");
6593 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6594 					CTLFLAG_RD,
6595 					&fl->ifl_m_dequeued, "mbufs freed");
6596 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6597 					CTLFLAG_RD,
6598 					&fl->ifl_cl_enqueued, "clusters allocated");
6599 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6600 					CTLFLAG_RD,
6601 					&fl->ifl_cl_dequeued, "clusters freed");
6602 #endif
6603 
6604 		}
6605 	}
6606 
6607 }
6608 
6609 void
6610 iflib_request_reset(if_ctx_t ctx)
6611 {
6612 
6613 	STATE_LOCK(ctx);
6614 	ctx->ifc_flags |= IFC_DO_RESET;
6615 	STATE_UNLOCK(ctx);
6616 }
6617 
6618 #ifndef __NO_STRICT_ALIGNMENT
6619 static struct mbuf *
6620 iflib_fixup_rx(struct mbuf *m)
6621 {
6622 	struct mbuf *n;
6623 
6624 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6625 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6626 		m->m_data += ETHER_HDR_LEN;
6627 		n = m;
6628 	} else {
6629 		MGETHDR(n, M_NOWAIT, MT_DATA);
6630 		if (n == NULL) {
6631 			m_freem(m);
6632 			return (NULL);
6633 		}
6634 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6635 		m->m_data += ETHER_HDR_LEN;
6636 		m->m_len -= ETHER_HDR_LEN;
6637 		n->m_len = ETHER_HDR_LEN;
6638 		M_MOVE_PKTHDR(n, m);
6639 		n->m_next = m;
6640 	}
6641 	return (n);
6642 }
6643 #endif
6644 
6645 #ifdef NETDUMP
6646 static void
6647 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6648 {
6649 	if_ctx_t ctx;
6650 
6651 	ctx = if_getsoftc(ifp);
6652 	CTX_LOCK(ctx);
6653 	*nrxr = NRXQSETS(ctx);
6654 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6655 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6656 	CTX_UNLOCK(ctx);
6657 }
6658 
6659 static void
6660 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6661 {
6662 	if_ctx_t ctx;
6663 	if_softc_ctx_t scctx;
6664 	iflib_fl_t fl;
6665 	iflib_rxq_t rxq;
6666 	int i, j;
6667 
6668 	ctx = if_getsoftc(ifp);
6669 	scctx = &ctx->ifc_softc_ctx;
6670 
6671 	switch (event) {
6672 	case NETDUMP_START:
6673 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6674 			rxq = &ctx->ifc_rxqs[i];
6675 			for (j = 0; j < rxq->ifr_nfl; j++) {
6676 				fl = rxq->ifr_fl;
6677 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6678 			}
6679 		}
6680 		iflib_no_tx_batch = 1;
6681 		break;
6682 	default:
6683 		break;
6684 	}
6685 }
6686 
6687 static int
6688 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6689 {
6690 	if_ctx_t ctx;
6691 	iflib_txq_t txq;
6692 	int error;
6693 
6694 	ctx = if_getsoftc(ifp);
6695 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6696 	    IFF_DRV_RUNNING)
6697 		return (EBUSY);
6698 
6699 	txq = &ctx->ifc_txqs[0];
6700 	error = iflib_encap(txq, &m);
6701 	if (error == 0)
6702 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6703 	return (error);
6704 }
6705 
6706 static int
6707 iflib_netdump_poll(struct ifnet *ifp, int count)
6708 {
6709 	if_ctx_t ctx;
6710 	if_softc_ctx_t scctx;
6711 	iflib_txq_t txq;
6712 	int i;
6713 
6714 	ctx = if_getsoftc(ifp);
6715 	scctx = &ctx->ifc_softc_ctx;
6716 
6717 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6718 	    IFF_DRV_RUNNING)
6719 		return (EBUSY);
6720 
6721 	txq = &ctx->ifc_txqs[0];
6722 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6723 
6724 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6725 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6726 	return (0);
6727 }
6728 #endif /* NETDUMP */
6729