xref: /freebsd/sys/net/iflib.c (revision 70a8349311934f20fcb65e763b8ac80134446796)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
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 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.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/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
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 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 #include <sys/bitstring.h>
97 /*
98  * enable accounting of every mbuf as it comes in to and goes out of
99  * iflib's software descriptor references
100  */
101 #define MEMORY_LOGGING 0
102 /*
103  * Enable mbuf vectors for compressing long mbuf chains
104  */
105 
106 /*
107  * NB:
108  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
109  *   we prefetch needs to be determined by the time spent in m_free vis a vis
110  *   the cost of a prefetch. This will of course vary based on the workload:
111  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
112  *        is quite expensive, thus suggesting very little prefetch.
113  *      - small packet forwarding which is just returning a single mbuf to
114  *        UMA will typically be very fast vis a vis the cost of a memory
115  *        access.
116  */
117 
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 struct iflib_txq;
132 typedef struct iflib_txq *iflib_txq_t;
133 struct iflib_rxq;
134 typedef struct iflib_rxq *iflib_rxq_t;
135 struct iflib_fl;
136 typedef struct iflib_fl *iflib_fl_t;
137 
138 struct iflib_ctx;
139 
140 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
141 
142 typedef struct iflib_filter_info {
143 	driver_filter_t *ifi_filter;
144 	void *ifi_filter_arg;
145 	struct grouptask *ifi_task;
146 	void *ifi_ctx;
147 } *iflib_filter_info_t;
148 
149 struct iflib_ctx {
150 	KOBJ_FIELDS;
151    /*
152    * Pointer to hardware driver's softc
153    */
154 	void *ifc_softc;
155 	device_t ifc_dev;
156 	if_t ifc_ifp;
157 
158 	cpuset_t ifc_cpus;
159 	if_shared_ctx_t ifc_sctx;
160 	struct if_softc_ctx ifc_softc_ctx;
161 
162 	struct mtx ifc_mtx;
163 
164 	uint16_t ifc_nhwtxqs;
165 	uint16_t ifc_nhwrxqs;
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 	int ifc_in_detach;
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 
192 	qidx_t ifc_sysctl_ntxds[8];
193 	qidx_t ifc_sysctl_nrxds[8];
194 	struct if_txrx ifc_txrx;
195 #define isc_txd_encap  ifc_txrx.ift_txd_encap
196 #define isc_txd_flush  ifc_txrx.ift_txd_flush
197 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
198 #define isc_rxd_available ifc_txrx.ift_rxd_available
199 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
200 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
201 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
204 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
205 	eventhandler_tag ifc_vlan_attach_event;
206 	eventhandler_tag ifc_vlan_detach_event;
207 	uint8_t ifc_mac[ETHER_ADDR_LEN];
208 	char ifc_mtx_name[16];
209 };
210 
211 
212 void *
213 iflib_get_softc(if_ctx_t ctx)
214 {
215 
216 	return (ctx->ifc_softc);
217 }
218 
219 device_t
220 iflib_get_dev(if_ctx_t ctx)
221 {
222 
223 	return (ctx->ifc_dev);
224 }
225 
226 if_t
227 iflib_get_ifp(if_ctx_t ctx)
228 {
229 
230 	return (ctx->ifc_ifp);
231 }
232 
233 struct ifmedia *
234 iflib_get_media(if_ctx_t ctx)
235 {
236 
237 	return (&ctx->ifc_media);
238 }
239 
240 void
241 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
242 {
243 
244 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
245 }
246 
247 if_softc_ctx_t
248 iflib_get_softc_ctx(if_ctx_t ctx)
249 {
250 
251 	return (&ctx->ifc_softc_ctx);
252 }
253 
254 if_shared_ctx_t
255 iflib_get_sctx(if_ctx_t ctx)
256 {
257 
258 	return (ctx->ifc_sctx);
259 }
260 
261 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
262 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
263 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
264 
265 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
266 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
267 
268 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
269 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
270 #define RX_SW_DESC_INUSE        (1 << 3)
271 #define TX_SW_DESC_MAPPED       (1 << 4)
272 
273 #define	M_TOOBIG		M_PROTO1
274 
275 typedef struct iflib_sw_rx_desc_array {
276 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
277 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
278 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
279 	uint8_t		*ifsd_flags;
280 } iflib_rxsd_array_t;
281 
282 typedef struct iflib_sw_tx_desc_array {
283 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
284 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
285 	uint8_t		*ifsd_flags;
286 } if_txsd_vec_t;
287 
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 /* bnxt supports 64 with hardware LRO enabled */
292 #define IFLIB_MAX_RX_SEGS		64
293 #define IFLIB_RX_COPY_THRESH		128
294 #define IFLIB_MAX_RX_REFRESH		32
295 /* The minimum descriptors per second before we start coalescing */
296 #define IFLIB_MIN_DESC_SEC		16384
297 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
298 #define IFLIB_QUEUE_IDLE		0
299 #define IFLIB_QUEUE_HUNG		1
300 #define IFLIB_QUEUE_WORKING		2
301 /* maximum number of txqs that can share an rx interrupt */
302 #define IFLIB_MAX_TX_SHARED_INTR	4
303 
304 /* this should really scale with ring size - this is a fairly arbitrary value */
305 #define TX_BATCH_SIZE			32
306 
307 #define IFLIB_RESTART_BUDGET		8
308 
309 #define	IFC_LEGACY		0x001
310 #define	IFC_QFLUSH		0x002
311 #define	IFC_MULTISEG		0x004
312 #define	IFC_DMAR		0x008
313 #define	IFC_SC_ALLOCATED	0x010
314 #define	IFC_INIT_DONE		0x020
315 #define	IFC_PREFETCH		0x040
316 #define	IFC_DO_RESET		0x080
317 #define	IFC_CHECK_HUNG		0x100
318 
319 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
320 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
321 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
322 struct iflib_txq {
323 	qidx_t		ift_in_use;
324 	qidx_t		ift_cidx;
325 	qidx_t		ift_cidx_processed;
326 	qidx_t		ift_pidx;
327 	uint8_t		ift_gen;
328 	uint8_t		ift_br_offset;
329 	uint16_t	ift_npending;
330 	uint16_t	ift_db_pending;
331 	uint16_t	ift_rs_pending;
332 	/* implicit pad */
333 	uint8_t		ift_txd_size[8];
334 	uint64_t	ift_processed;
335 	uint64_t	ift_cleaned;
336 	uint64_t	ift_cleaned_prev;
337 #if MEMORY_LOGGING
338 	uint64_t	ift_enqueued;
339 	uint64_t	ift_dequeued;
340 #endif
341 	uint64_t	ift_no_tx_dma_setup;
342 	uint64_t	ift_no_desc_avail;
343 	uint64_t	ift_mbuf_defrag_failed;
344 	uint64_t	ift_mbuf_defrag;
345 	uint64_t	ift_map_failed;
346 	uint64_t	ift_txd_encap_efbig;
347 	uint64_t	ift_pullups;
348 
349 	struct mtx	ift_mtx;
350 	struct mtx	ift_db_mtx;
351 
352 	/* constant values */
353 	if_ctx_t	ift_ctx;
354 	struct ifmp_ring        *ift_br;
355 	struct grouptask	ift_task;
356 	qidx_t		ift_size;
357 	uint16_t	ift_id;
358 	struct callout	ift_timer;
359 
360 	if_txsd_vec_t	ift_sds;
361 	uint8_t		ift_qstatus;
362 	uint8_t		ift_closed;
363 	uint8_t		ift_update_freq;
364 	struct iflib_filter_info ift_filter_info;
365 	bus_dma_tag_t		ift_desc_tag;
366 	bus_dma_tag_t		ift_tso_desc_tag;
367 	iflib_dma_info_t	ift_ifdi;
368 #define MTX_NAME_LEN 16
369 	char                    ift_mtx_name[MTX_NAME_LEN];
370 	char                    ift_db_mtx_name[MTX_NAME_LEN];
371 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
372 #ifdef IFLIB_DIAGNOSTICS
373 	uint64_t ift_cpu_exec_count[256];
374 #endif
375 } __aligned(CACHE_LINE_SIZE);
376 
377 struct iflib_fl {
378 	qidx_t		ifl_cidx;
379 	qidx_t		ifl_pidx;
380 	qidx_t		ifl_credits;
381 	uint8_t		ifl_gen;
382 	uint8_t		ifl_rxd_size;
383 #if MEMORY_LOGGING
384 	uint64_t	ifl_m_enqueued;
385 	uint64_t	ifl_m_dequeued;
386 	uint64_t	ifl_cl_enqueued;
387 	uint64_t	ifl_cl_dequeued;
388 #endif
389 	/* implicit pad */
390 
391 	bitstr_t 	*ifl_rx_bitmap;
392 	qidx_t		ifl_fragidx;
393 	/* constant */
394 	qidx_t		ifl_size;
395 	uint16_t	ifl_buf_size;
396 	uint16_t	ifl_cltype;
397 	uma_zone_t	ifl_zone;
398 	iflib_rxsd_array_t	ifl_sds;
399 	iflib_rxq_t	ifl_rxq;
400 	uint8_t		ifl_id;
401 	bus_dma_tag_t           ifl_desc_tag;
402 	iflib_dma_info_t	ifl_ifdi;
403 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
404 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
405 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
406 }  __aligned(CACHE_LINE_SIZE);
407 
408 static inline qidx_t
409 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
410 {
411 	qidx_t used;
412 
413 	if (pidx > cidx)
414 		used = pidx - cidx;
415 	else if (pidx < cidx)
416 		used = size - cidx + pidx;
417 	else if (gen == 0 && pidx == cidx)
418 		used = 0;
419 	else if (gen == 1 && pidx == cidx)
420 		used = size;
421 	else
422 		panic("bad state");
423 
424 	return (used);
425 }
426 
427 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
428 
429 #define IDXDIFF(head, tail, wrap) \
430 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
431 
432 struct iflib_rxq {
433 	/* If there is a separate completion queue -
434 	 * these are the cq cidx and pidx. Otherwise
435 	 * these are unused.
436 	 */
437 	qidx_t		ifr_size;
438 	qidx_t		ifr_cq_cidx;
439 	qidx_t		ifr_cq_pidx;
440 	uint8_t		ifr_cq_gen;
441 	uint8_t		ifr_fl_offset;
442 
443 	if_ctx_t	ifr_ctx;
444 	iflib_fl_t	ifr_fl;
445 	uint64_t	ifr_rx_irq;
446 	uint16_t	ifr_id;
447 	uint8_t		ifr_lro_enabled;
448 	uint8_t		ifr_nfl;
449 	uint8_t		ifr_ntxqirq;
450 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
451 	struct lro_ctrl			ifr_lc;
452 	struct grouptask        ifr_task;
453 	struct iflib_filter_info ifr_filter_info;
454 	iflib_dma_info_t		ifr_ifdi;
455 
456 	/* dynamically allocate if any drivers need a value substantially larger than this */
457 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
458 #ifdef IFLIB_DIAGNOSTICS
459 	uint64_t ifr_cpu_exec_count[256];
460 #endif
461 }  __aligned(CACHE_LINE_SIZE);
462 
463 typedef struct if_rxsd {
464 	caddr_t *ifsd_cl;
465 	struct mbuf **ifsd_m;
466 	iflib_fl_t ifsd_fl;
467 	qidx_t ifsd_cidx;
468 } *if_rxsd_t;
469 
470 /* multiple of word size */
471 #ifdef __LP64__
472 #define PKT_INFO_SIZE	6
473 #define RXD_INFO_SIZE	5
474 #define PKT_TYPE uint64_t
475 #else
476 #define PKT_INFO_SIZE	11
477 #define RXD_INFO_SIZE	8
478 #define PKT_TYPE uint32_t
479 #endif
480 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
481 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
482 
483 typedef struct if_pkt_info_pad {
484 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
485 } *if_pkt_info_pad_t;
486 typedef struct if_rxd_info_pad {
487 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
488 } *if_rxd_info_pad_t;
489 
490 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
491 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
492 
493 
494 static inline void
495 pkt_info_zero(if_pkt_info_t pi)
496 {
497 	if_pkt_info_pad_t pi_pad;
498 
499 	pi_pad = (if_pkt_info_pad_t)pi;
500 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
501 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
502 #ifndef __LP64__
503 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
504 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
505 #endif
506 }
507 
508 static inline void
509 rxd_info_zero(if_rxd_info_t ri)
510 {
511 	if_rxd_info_pad_t ri_pad;
512 	int i;
513 
514 	ri_pad = (if_rxd_info_pad_t)ri;
515 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
516 		ri_pad->rxd_val[i] = 0;
517 		ri_pad->rxd_val[i+1] = 0;
518 		ri_pad->rxd_val[i+2] = 0;
519 		ri_pad->rxd_val[i+3] = 0;
520 	}
521 #ifdef __LP64__
522 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
523 #endif
524 }
525 
526 /*
527  * Only allow a single packet to take up most 1/nth of the tx ring
528  */
529 #define MAX_SINGLE_PACKET_FRACTION 12
530 #define IF_BAD_DMA (bus_addr_t)-1
531 
532 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
533 
534 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
535 
536 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
537 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
538 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
539 
540 
541 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
542 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
543 
544 
545 /* Our boot-time initialization hook */
546 static int	iflib_module_event_handler(module_t, int, void *);
547 
548 static moduledata_t iflib_moduledata = {
549 	"iflib",
550 	iflib_module_event_handler,
551 	NULL
552 };
553 
554 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
555 MODULE_VERSION(iflib, 1);
556 
557 MODULE_DEPEND(iflib, pci, 1, 1, 1);
558 MODULE_DEPEND(iflib, ether, 1, 1, 1);
559 
560 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
561 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
562 
563 #ifndef IFLIB_DEBUG_COUNTERS
564 #ifdef INVARIANTS
565 #define IFLIB_DEBUG_COUNTERS 1
566 #else
567 #define IFLIB_DEBUG_COUNTERS 0
568 #endif /* !INVARIANTS */
569 #endif
570 
571 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
572                    "iflib driver parameters");
573 
574 /*
575  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
576  */
577 static int iflib_min_tx_latency = 0;
578 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
579 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
580 static int iflib_no_tx_batch = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
582 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
583 
584 
585 #if IFLIB_DEBUG_COUNTERS
586 
587 static int iflib_tx_seen;
588 static int iflib_tx_sent;
589 static int iflib_tx_encap;
590 static int iflib_rx_allocs;
591 static int iflib_fl_refills;
592 static int iflib_fl_refills_large;
593 static int iflib_tx_frees;
594 
595 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
596 		   &iflib_tx_seen, 0, "# tx mbufs seen");
597 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
598 		   &iflib_tx_sent, 0, "# tx mbufs sent");
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
600 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
602 		   &iflib_tx_frees, 0, "# tx frees");
603 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
604 		   &iflib_rx_allocs, 0, "# rx allocations");
605 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
606 		   &iflib_fl_refills, 0, "# refills");
607 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
608 		   &iflib_fl_refills_large, 0, "# large refills");
609 
610 
611 static int iflib_txq_drain_flushing;
612 static int iflib_txq_drain_oactive;
613 static int iflib_txq_drain_notready;
614 static int iflib_txq_drain_encapfail;
615 
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
617 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
619 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
621 		   &iflib_txq_drain_notready, 0, "# drain notready");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
623 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
624 
625 
626 static int iflib_encap_load_mbuf_fail;
627 static int iflib_encap_txq_avail_fail;
628 static int iflib_encap_txd_encap_fail;
629 
630 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
631 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
632 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
633 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
635 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
636 
637 static int iflib_task_fn_rxs;
638 static int iflib_rx_intr_enables;
639 static int iflib_fast_intrs;
640 static int iflib_intr_link;
641 static int iflib_intr_msix;
642 static int iflib_rx_unavail;
643 static int iflib_rx_ctx_inactive;
644 static int iflib_rx_zero_len;
645 static int iflib_rx_if_input;
646 static int iflib_rx_mbuf_null;
647 static int iflib_rxd_flush;
648 
649 static int iflib_verbose_debug;
650 
651 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
652 		   &iflib_intr_link, 0, "# intr link calls");
653 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
654 		   &iflib_intr_msix, 0, "# intr msix calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
656 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
658 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
659 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
660 		   &iflib_fast_intrs, 0, "# fast_intr calls");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
662 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
664 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
666 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
668 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
670 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
671 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
672 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
673 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
674 		   &iflib_verbose_debug, 0, "enable verbose debugging");
675 
676 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
677 static void
678 iflib_debug_reset(void)
679 {
680 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
681 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
682 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
683 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
684 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
685 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
686 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
687 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
688 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
689 }
690 
691 #else
692 #define DBG_COUNTER_INC(name)
693 static void iflib_debug_reset(void) {}
694 #endif
695 
696 
697 
698 #define IFLIB_DEBUG 0
699 
700 static void iflib_tx_structures_free(if_ctx_t ctx);
701 static void iflib_rx_structures_free(if_ctx_t ctx);
702 static int iflib_queues_alloc(if_ctx_t ctx);
703 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
704 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
705 static int iflib_qset_structures_setup(if_ctx_t ctx);
706 static int iflib_msix_init(if_ctx_t ctx);
707 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
708 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
709 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
710 static int iflib_register(if_ctx_t);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_stop(if_ctx_t ctx);
717 static void iflib_if_init_locked(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 #ifdef DEV_NETMAP
723 #include <sys/selinfo.h>
724 #include <net/netmap.h>
725 #include <dev/netmap/netmap_kern.h>
726 
727 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
728 
729 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
730 
731 /*
732  * device-specific sysctl variables:
733  *
734  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
735  *	During regular operations the CRC is stripped, but on some
736  *	hardware reception of frames not multiple of 64 is slower,
737  *	so using crcstrip=0 helps in benchmarks.
738  *
739  * iflib_rx_miss, iflib_rx_miss_bufs:
740  *	count packets that might be missed due to lost interrupts.
741  */
742 SYSCTL_DECL(_dev_netmap);
743 /*
744  * The xl driver by default strips CRCs and we do not override it.
745  */
746 
747 int iflib_crcstrip = 1;
748 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
749     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
750 
751 int iflib_rx_miss, iflib_rx_miss_bufs;
752 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
753     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
754 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
755     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
756 
757 /*
758  * Register/unregister. We are already under netmap lock.
759  * Only called on the first register or the last unregister.
760  */
761 static int
762 iflib_netmap_register(struct netmap_adapter *na, int onoff)
763 {
764 	struct ifnet *ifp = na->ifp;
765 	if_ctx_t ctx = ifp->if_softc;
766 	int status;
767 
768 	CTX_LOCK(ctx);
769 	IFDI_INTR_DISABLE(ctx);
770 
771 	/* Tell the stack that the interface is no longer active */
772 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
773 
774 	if (!CTX_IS_VF(ctx))
775 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
776 
777 	/* enable or disable flags and callbacks in na and ifp */
778 	if (onoff) {
779 		nm_set_native_flags(na);
780 	} else {
781 		nm_clear_native_flags(na);
782 	}
783 	iflib_stop(ctx);
784 	iflib_init_locked(ctx);
785 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
786 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
787 	if (status)
788 		nm_clear_native_flags(na);
789 	CTX_UNLOCK(ctx);
790 	return (status);
791 }
792 
793 static int
794 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
795 {
796 	struct netmap_adapter *na = kring->na;
797 	u_int const lim = kring->nkr_num_slots - 1;
798 	u_int head = kring->rhead;
799 	struct netmap_ring *ring = kring->ring;
800 	bus_dmamap_t *map;
801 	struct if_rxd_update iru;
802 	if_ctx_t ctx = rxq->ifr_ctx;
803 	iflib_fl_t fl = &rxq->ifr_fl[0];
804 	uint32_t refill_pidx, nic_i;
805 
806 	if (nm_i == head && __predict_true(!init))
807 		return 0;
808 	iru_init(&iru, rxq, 0 /* flid */);
809 	map = fl->ifl_sds.ifsd_map;
810 	refill_pidx = netmap_idx_k2n(kring, nm_i);
811 	/*
812 	 * IMPORTANT: we must leave one free slot in the ring,
813 	 * so move head back by one unit
814 	 */
815 	head = nm_prev(head, lim);
816 	while (nm_i != head) {
817 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
818 			struct netmap_slot *slot = &ring->slot[nm_i];
819 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
820 			uint32_t nic_i_dma = refill_pidx;
821 			nic_i = netmap_idx_k2n(kring, nm_i);
822 
823 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
824 
825 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
826 			        return netmap_ring_reinit(kring);
827 
828 			fl->ifl_vm_addrs[tmp_pidx] = addr;
829 			if (__predict_false(init) && map) {
830 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
831 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
832 				/* buffer has changed, reload map */
833 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
834 			}
835 			slot->flags &= ~NS_BUF_CHANGED;
836 
837 			nm_i = nm_next(nm_i, lim);
838 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
839 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
840 				continue;
841 
842 			iru.iru_pidx = refill_pidx;
843 			iru.iru_count = tmp_pidx+1;
844 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
845 
846 			refill_pidx = nic_i;
847 			if (map == NULL)
848 				continue;
849 
850 			for (int n = 0; n < iru.iru_count; n++) {
851 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
852 						BUS_DMASYNC_PREREAD);
853 				/* XXX - change this to not use the netmap func*/
854 				nic_i_dma = nm_next(nic_i_dma, lim);
855 			}
856 		}
857 	}
858 	kring->nr_hwcur = head;
859 
860 	if (map)
861 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
862 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
863 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
864 	return (0);
865 }
866 
867 /*
868  * Reconcile kernel and user view of the transmit ring.
869  *
870  * All information is in the kring.
871  * Userspace wants to send packets up to the one before kring->rhead,
872  * kernel knows kring->nr_hwcur is the first unsent packet.
873  *
874  * Here we push packets out (as many as possible), and possibly
875  * reclaim buffers from previously completed transmission.
876  *
877  * The caller (netmap) guarantees that there is only one instance
878  * running at any time. Any interference with other driver
879  * methods should be handled by the individual drivers.
880  */
881 static int
882 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
883 {
884 	struct netmap_adapter *na = kring->na;
885 	struct ifnet *ifp = na->ifp;
886 	struct netmap_ring *ring = kring->ring;
887 	u_int nm_i;	/* index into the netmap ring */
888 	u_int nic_i;	/* index into the NIC ring */
889 	u_int n;
890 	u_int const lim = kring->nkr_num_slots - 1;
891 	u_int const head = kring->rhead;
892 	struct if_pkt_info pi;
893 
894 	/*
895 	 * interrupts on every tx packet are expensive so request
896 	 * them every half ring, or where NS_REPORT is set
897 	 */
898 	u_int report_frequency = kring->nkr_num_slots >> 1;
899 	/* device-specific */
900 	if_ctx_t ctx = ifp->if_softc;
901 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
902 
903 	if (txq->ift_sds.ifsd_map)
904 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
905 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
906 
907 
908 	/*
909 	 * First part: process new packets to send.
910 	 * nm_i is the current index in the netmap ring,
911 	 * nic_i is the corresponding index in the NIC ring.
912 	 *
913 	 * If we have packets to send (nm_i != head)
914 	 * iterate over the netmap ring, fetch length and update
915 	 * the corresponding slot in the NIC ring. Some drivers also
916 	 * need to update the buffer's physical address in the NIC slot
917 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
918 	 *
919 	 * The netmap_reload_map() calls is especially expensive,
920 	 * even when (as in this case) the tag is 0, so do only
921 	 * when the buffer has actually changed.
922 	 *
923 	 * If possible do not set the report/intr bit on all slots,
924 	 * but only a few times per ring or when NS_REPORT is set.
925 	 *
926 	 * Finally, on 10G and faster drivers, it might be useful
927 	 * to prefetch the next slot and txr entry.
928 	 */
929 
930 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
931 	pkt_info_zero(&pi);
932 	pi.ipi_segs = txq->ift_segs;
933 	pi.ipi_qsidx = kring->ring_id;
934 	if (nm_i != head) {	/* we have new packets to send */
935 		nic_i = netmap_idx_k2n(kring, nm_i);
936 
937 		__builtin_prefetch(&ring->slot[nm_i]);
938 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
939 		if (txq->ift_sds.ifsd_map)
940 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
941 
942 		for (n = 0; nm_i != head; n++) {
943 			struct netmap_slot *slot = &ring->slot[nm_i];
944 			u_int len = slot->len;
945 			uint64_t paddr;
946 			void *addr = PNMB(na, slot, &paddr);
947 			int flags = (slot->flags & NS_REPORT ||
948 				nic_i == 0 || nic_i == report_frequency) ?
949 				IPI_TX_INTR : 0;
950 
951 			/* device-specific */
952 			pi.ipi_len = len;
953 			pi.ipi_segs[0].ds_addr = paddr;
954 			pi.ipi_segs[0].ds_len = len;
955 			pi.ipi_nsegs = 1;
956 			pi.ipi_ndescs = 0;
957 			pi.ipi_pidx = nic_i;
958 			pi.ipi_flags = flags;
959 
960 			/* Fill the slot in the NIC ring. */
961 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
962 
963 			/* prefetch for next round */
964 			__builtin_prefetch(&ring->slot[nm_i + 1]);
965 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
966 			if (txq->ift_sds.ifsd_map) {
967 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
968 
969 				NM_CHECK_ADDR_LEN(na, addr, len);
970 
971 				if (slot->flags & NS_BUF_CHANGED) {
972 					/* buffer has changed, reload map */
973 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
974 				}
975 				/* make sure changes to the buffer are synced */
976 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
977 						BUS_DMASYNC_PREWRITE);
978 			}
979 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
980 			nm_i = nm_next(nm_i, lim);
981 			nic_i = nm_next(nic_i, lim);
982 		}
983 		kring->nr_hwcur = head;
984 
985 		/* synchronize the NIC ring */
986 		if (txq->ift_sds.ifsd_map)
987 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
988 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
989 
990 		/* (re)start the tx unit up to slot nic_i (excluded) */
991 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
992 	}
993 
994 	/*
995 	 * Second part: reclaim buffers for completed transmissions.
996 	 */
997 	if (iflib_tx_credits_update(ctx, txq)) {
998 		/* some tx completed, increment avail */
999 		nic_i = txq->ift_cidx_processed;
1000 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1001 	}
1002 	return (0);
1003 }
1004 
1005 /*
1006  * Reconcile kernel and user view of the receive ring.
1007  * Same as for the txsync, this routine must be efficient.
1008  * The caller guarantees a single invocations, but races against
1009  * the rest of the driver should be handled here.
1010  *
1011  * On call, kring->rhead is the first packet that userspace wants
1012  * to keep, and kring->rcur is the wakeup point.
1013  * The kernel has previously reported packets up to kring->rtail.
1014  *
1015  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1016  * of whether or not we received an interrupt.
1017  */
1018 static int
1019 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1020 {
1021 	struct netmap_adapter *na = kring->na;
1022 	struct netmap_ring *ring = kring->ring;
1023 	uint32_t nm_i;	/* index into the netmap ring */
1024 	uint32_t nic_i;	/* index into the NIC ring */
1025 	u_int i, n;
1026 	u_int const lim = kring->nkr_num_slots - 1;
1027 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1028 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1029 	struct if_rxd_info ri;
1030 
1031 	struct ifnet *ifp = na->ifp;
1032 	if_ctx_t ctx = ifp->if_softc;
1033 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1034 	iflib_fl_t fl = rxq->ifr_fl;
1035 	if (head > lim)
1036 		return netmap_ring_reinit(kring);
1037 
1038 	/* XXX check sync modes */
1039 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1040 		if (fl->ifl_sds.ifsd_map == NULL)
1041 			continue;
1042 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1043 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1044 	}
1045 	/*
1046 	 * First part: import newly received packets.
1047 	 *
1048 	 * nm_i is the index of the next free slot in the netmap ring,
1049 	 * nic_i is the index of the next received packet in the NIC ring,
1050 	 * and they may differ in case if_init() has been called while
1051 	 * in netmap mode. For the receive ring we have
1052 	 *
1053 	 *	nic_i = rxr->next_check;
1054 	 *	nm_i = kring->nr_hwtail (previous)
1055 	 * and
1056 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1057 	 *
1058 	 * rxr->next_check is set to 0 on a ring reinit
1059 	 */
1060 	if (netmap_no_pendintr || force_update) {
1061 		int crclen = iflib_crcstrip ? 0 : 4;
1062 		int error, avail;
1063 		uint16_t slot_flags = kring->nkr_slot_flags;
1064 
1065 		for (i = 0; i < rxq->ifr_nfl; i++) {
1066 			fl = &rxq->ifr_fl[i];
1067 			nic_i = fl->ifl_cidx;
1068 			nm_i = netmap_idx_n2k(kring, nic_i);
1069 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1070 			for (n = 0; avail > 0; n++, avail--) {
1071 				rxd_info_zero(&ri);
1072 				ri.iri_frags = rxq->ifr_frags;
1073 				ri.iri_qsidx = kring->ring_id;
1074 				ri.iri_ifp = ctx->ifc_ifp;
1075 				ri.iri_cidx = nic_i;
1076 
1077 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1078 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1079 				ring->slot[nm_i].flags = slot_flags;
1080 				if (fl->ifl_sds.ifsd_map)
1081 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1082 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1083 				nm_i = nm_next(nm_i, lim);
1084 				nic_i = nm_next(nic_i, lim);
1085 			}
1086 			if (n) { /* update the state variables */
1087 				if (netmap_no_pendintr && !force_update) {
1088 					/* diagnostics */
1089 					iflib_rx_miss ++;
1090 					iflib_rx_miss_bufs += n;
1091 				}
1092 				fl->ifl_cidx = nic_i;
1093 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1094 			}
1095 			kring->nr_kflags &= ~NKR_PENDINTR;
1096 		}
1097 	}
1098 	/*
1099 	 * Second part: skip past packets that userspace has released.
1100 	 * (kring->nr_hwcur to head excluded),
1101 	 * and make the buffers available for reception.
1102 	 * As usual nm_i is the index in the netmap ring,
1103 	 * nic_i is the index in the NIC ring, and
1104 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1105 	 */
1106 	/* XXX not sure how this will work with multiple free lists */
1107 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1108 
1109 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1110 }
1111 
1112 static void
1113 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1114 {
1115 	struct ifnet *ifp = na->ifp;
1116 	if_ctx_t ctx = ifp->if_softc;
1117 
1118 	CTX_LOCK(ctx);
1119 	if (onoff) {
1120 		IFDI_INTR_ENABLE(ctx);
1121 	} else {
1122 		IFDI_INTR_DISABLE(ctx);
1123 	}
1124 	CTX_UNLOCK(ctx);
1125 }
1126 
1127 
1128 static int
1129 iflib_netmap_attach(if_ctx_t ctx)
1130 {
1131 	struct netmap_adapter na;
1132 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1133 
1134 	bzero(&na, sizeof(na));
1135 
1136 	na.ifp = ctx->ifc_ifp;
1137 	na.na_flags = NAF_BDG_MAYSLEEP;
1138 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1139 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1140 
1141 	na.num_tx_desc = scctx->isc_ntxd[0];
1142 	na.num_rx_desc = scctx->isc_nrxd[0];
1143 	na.nm_txsync = iflib_netmap_txsync;
1144 	na.nm_rxsync = iflib_netmap_rxsync;
1145 	na.nm_register = iflib_netmap_register;
1146 	na.nm_intr = iflib_netmap_intr;
1147 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1148 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1149 	return (netmap_attach(&na));
1150 }
1151 
1152 static void
1153 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1154 {
1155 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1156 	struct netmap_slot *slot;
1157 
1158 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1159 	if (slot == NULL)
1160 		return;
1161 	if (txq->ift_sds.ifsd_map == NULL)
1162 		return;
1163 
1164 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1165 
1166 		/*
1167 		 * In netmap mode, set the map for the packet buffer.
1168 		 * NOTE: Some drivers (not this one) also need to set
1169 		 * the physical buffer address in the NIC ring.
1170 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1171 		 * netmap slot index, si
1172 		 */
1173 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1174 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1175 	}
1176 }
1177 
1178 static void
1179 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1180 {
1181 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1182 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1183 	struct netmap_slot *slot;
1184 	uint32_t nm_i;
1185 
1186 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1187 	if (slot == NULL)
1188 		return;
1189 	nm_i = netmap_idx_n2k(kring, 0);
1190 	netmap_fl_refill(rxq, kring, nm_i, true);
1191 }
1192 
1193 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1194 
1195 #else
1196 #define iflib_netmap_txq_init(ctx, txq)
1197 #define iflib_netmap_rxq_init(ctx, rxq)
1198 #define iflib_netmap_detach(ifp)
1199 
1200 #define iflib_netmap_attach(ctx) (0)
1201 #define netmap_rx_irq(ifp, qid, budget) (0)
1202 #define netmap_tx_irq(ifp, qid) do {} while (0)
1203 
1204 #endif
1205 
1206 #if defined(__i386__) || defined(__amd64__)
1207 static __inline void
1208 prefetch(void *x)
1209 {
1210 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1211 }
1212 static __inline void
1213 prefetch2cachelines(void *x)
1214 {
1215 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1216 #if (CACHE_LINE_SIZE < 128)
1217 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1218 #endif
1219 }
1220 #else
1221 #define prefetch(x)
1222 #define prefetch2cachelines(x)
1223 #endif
1224 
1225 static void
1226 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1227 {
1228 	iflib_fl_t fl;
1229 
1230 	fl = &rxq->ifr_fl[flid];
1231 	iru->iru_paddrs = fl->ifl_bus_addrs;
1232 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1233 	iru->iru_idxs = fl->ifl_rxd_idxs;
1234 	iru->iru_qsidx = rxq->ifr_id;
1235 	iru->iru_buf_size = fl->ifl_buf_size;
1236 	iru->iru_flidx = fl->ifl_id;
1237 }
1238 
1239 static void
1240 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1241 {
1242 	if (err)
1243 		return;
1244 	*(bus_addr_t *) arg = segs[0].ds_addr;
1245 }
1246 
1247 int
1248 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1249 {
1250 	int err;
1251 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1252 	device_t dev = ctx->ifc_dev;
1253 
1254 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1255 
1256 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1257 				sctx->isc_q_align, 0,	/* alignment, bounds */
1258 				BUS_SPACE_MAXADDR,	/* lowaddr */
1259 				BUS_SPACE_MAXADDR,	/* highaddr */
1260 				NULL, NULL,		/* filter, filterarg */
1261 				size,			/* maxsize */
1262 				1,			/* nsegments */
1263 				size,			/* maxsegsize */
1264 				BUS_DMA_ALLOCNOW,	/* flags */
1265 				NULL,			/* lockfunc */
1266 				NULL,			/* lockarg */
1267 				&dma->idi_tag);
1268 	if (err) {
1269 		device_printf(dev,
1270 		    "%s: bus_dma_tag_create failed: %d\n",
1271 		    __func__, err);
1272 		goto fail_0;
1273 	}
1274 
1275 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1276 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1277 	if (err) {
1278 		device_printf(dev,
1279 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1280 		    __func__, (uintmax_t)size, err);
1281 		goto fail_1;
1282 	}
1283 
1284 	dma->idi_paddr = IF_BAD_DMA;
1285 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1286 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1287 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1288 		device_printf(dev,
1289 		    "%s: bus_dmamap_load failed: %d\n",
1290 		    __func__, err);
1291 		goto fail_2;
1292 	}
1293 
1294 	dma->idi_size = size;
1295 	return (0);
1296 
1297 fail_2:
1298 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1299 fail_1:
1300 	bus_dma_tag_destroy(dma->idi_tag);
1301 fail_0:
1302 	dma->idi_tag = NULL;
1303 
1304 	return (err);
1305 }
1306 
1307 int
1308 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1309 {
1310 	int i, err;
1311 	iflib_dma_info_t *dmaiter;
1312 
1313 	dmaiter = dmalist;
1314 	for (i = 0; i < count; i++, dmaiter++) {
1315 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1316 			break;
1317 	}
1318 	if (err)
1319 		iflib_dma_free_multi(dmalist, i);
1320 	return (err);
1321 }
1322 
1323 void
1324 iflib_dma_free(iflib_dma_info_t dma)
1325 {
1326 	if (dma->idi_tag == NULL)
1327 		return;
1328 	if (dma->idi_paddr != IF_BAD_DMA) {
1329 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1330 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1331 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1332 		dma->idi_paddr = IF_BAD_DMA;
1333 	}
1334 	if (dma->idi_vaddr != NULL) {
1335 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1336 		dma->idi_vaddr = NULL;
1337 	}
1338 	bus_dma_tag_destroy(dma->idi_tag);
1339 	dma->idi_tag = NULL;
1340 }
1341 
1342 void
1343 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1344 {
1345 	int i;
1346 	iflib_dma_info_t *dmaiter = dmalist;
1347 
1348 	for (i = 0; i < count; i++, dmaiter++)
1349 		iflib_dma_free(*dmaiter);
1350 }
1351 
1352 #ifdef EARLY_AP_STARTUP
1353 static const int iflib_started = 1;
1354 #else
1355 /*
1356  * We used to abuse the smp_started flag to decide if the queues have been
1357  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1358  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1359  * is set.  Run a SYSINIT() strictly after that to just set a usable
1360  * completion flag.
1361  */
1362 
1363 static int iflib_started;
1364 
1365 static void
1366 iflib_record_started(void *arg)
1367 {
1368 	iflib_started = 1;
1369 }
1370 
1371 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1372 	iflib_record_started, NULL);
1373 #endif
1374 
1375 static int
1376 iflib_fast_intr(void *arg)
1377 {
1378 	iflib_filter_info_t info = arg;
1379 	struct grouptask *gtask = info->ifi_task;
1380 	if (!iflib_started)
1381 		return (FILTER_HANDLED);
1382 
1383 	DBG_COUNTER_INC(fast_intrs);
1384 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1385 		return (FILTER_HANDLED);
1386 
1387 	GROUPTASK_ENQUEUE(gtask);
1388 	return (FILTER_HANDLED);
1389 }
1390 
1391 static int
1392 iflib_fast_intr_rxtx(void *arg)
1393 {
1394 	iflib_filter_info_t info = arg;
1395 	struct grouptask *gtask = info->ifi_task;
1396 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1397 	if_ctx_t ctx;
1398 	int i, cidx;
1399 
1400 	if (!iflib_started)
1401 		return (FILTER_HANDLED);
1402 
1403 	DBG_COUNTER_INC(fast_intrs);
1404 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1405 		return (FILTER_HANDLED);
1406 
1407 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1408 		qidx_t txqid = rxq->ifr_txqid[i];
1409 
1410 		ctx = rxq->ifr_ctx;
1411 
1412 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1413 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1414 			continue;
1415 		}
1416 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1417 	}
1418 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1419 		cidx = rxq->ifr_cq_cidx;
1420 	else
1421 		cidx = rxq->ifr_fl[0].ifl_cidx;
1422 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1423 		GROUPTASK_ENQUEUE(gtask);
1424 	else
1425 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1426 	return (FILTER_HANDLED);
1427 }
1428 
1429 
1430 static int
1431 iflib_fast_intr_ctx(void *arg)
1432 {
1433 	iflib_filter_info_t info = arg;
1434 	struct grouptask *gtask = info->ifi_task;
1435 
1436 	if (!iflib_started)
1437 		return (FILTER_HANDLED);
1438 
1439 	DBG_COUNTER_INC(fast_intrs);
1440 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1441 		return (FILTER_HANDLED);
1442 
1443 	GROUPTASK_ENQUEUE(gtask);
1444 	return (FILTER_HANDLED);
1445 }
1446 
1447 static int
1448 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1449 	driver_filter_t filter, driver_intr_t handler, void *arg,
1450 				 char *name)
1451 {
1452 	int rc, flags;
1453 	struct resource *res;
1454 	void *tag = NULL;
1455 	device_t dev = ctx->ifc_dev;
1456 
1457 	flags = RF_ACTIVE;
1458 	if (ctx->ifc_flags & IFC_LEGACY)
1459 		flags |= RF_SHAREABLE;
1460 	MPASS(rid < 512);
1461 	irq->ii_rid = rid;
1462 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1463 	if (res == NULL) {
1464 		device_printf(dev,
1465 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1466 		return (ENOMEM);
1467 	}
1468 	irq->ii_res = res;
1469 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1470 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1471 						filter, handler, arg, &tag);
1472 	if (rc != 0) {
1473 		device_printf(dev,
1474 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1475 					  rid, name ? name : "unknown", rc);
1476 		return (rc);
1477 	} else if (name)
1478 		bus_describe_intr(dev, res, tag, "%s", name);
1479 
1480 	irq->ii_tag = tag;
1481 	return (0);
1482 }
1483 
1484 
1485 /*********************************************************************
1486  *
1487  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1488  *  the information needed to transmit a packet on the wire. This is
1489  *  called only once at attach, setup is done every reset.
1490  *
1491  **********************************************************************/
1492 
1493 static int
1494 iflib_txsd_alloc(iflib_txq_t txq)
1495 {
1496 	if_ctx_t ctx = txq->ift_ctx;
1497 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1498 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1499 	device_t dev = ctx->ifc_dev;
1500 	int err, nsegments, ntsosegments;
1501 
1502 	nsegments = scctx->isc_tx_nsegments;
1503 	ntsosegments = scctx->isc_tx_tso_segments_max;
1504 	MPASS(scctx->isc_ntxd[0] > 0);
1505 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1506 	MPASS(nsegments > 0);
1507 	MPASS(ntsosegments > 0);
1508 	/*
1509 	 * Setup DMA descriptor areas.
1510 	 */
1511 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1512 			       1, 0,			/* alignment, bounds */
1513 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1514 			       BUS_SPACE_MAXADDR,	/* highaddr */
1515 			       NULL, NULL,		/* filter, filterarg */
1516 			       sctx->isc_tx_maxsize,		/* maxsize */
1517 			       nsegments,	/* nsegments */
1518 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1519 			       0,			/* flags */
1520 			       NULL,			/* lockfunc */
1521 			       NULL,			/* lockfuncarg */
1522 			       &txq->ift_desc_tag))) {
1523 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1524 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1525 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1526 		goto fail;
1527 	}
1528 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1529 			       1, 0,			/* alignment, bounds */
1530 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1531 			       BUS_SPACE_MAXADDR,	/* highaddr */
1532 			       NULL, NULL,		/* filter, filterarg */
1533 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1534 			       ntsosegments,	/* nsegments */
1535 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1536 			       0,			/* flags */
1537 			       NULL,			/* lockfunc */
1538 			       NULL,			/* lockfuncarg */
1539 			       &txq->ift_tso_desc_tag))) {
1540 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1541 
1542 		goto fail;
1543 	}
1544 	if (!(txq->ift_sds.ifsd_flags =
1545 	    (uint8_t *) malloc(sizeof(uint8_t) *
1546 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1547 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1548 		err = ENOMEM;
1549 		goto fail;
1550 	}
1551 	if (!(txq->ift_sds.ifsd_m =
1552 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1553 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1554 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1555 		err = ENOMEM;
1556 		goto fail;
1557 	}
1558 
1559         /* Create the descriptor buffer dma maps */
1560 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1561 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1562 		return (0);
1563 
1564 	if (!(txq->ift_sds.ifsd_map =
1565 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1566 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1567 		err = ENOMEM;
1568 		goto fail;
1569 	}
1570 
1571 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1572 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1573 		if (err != 0) {
1574 			device_printf(dev, "Unable to create TX DMA map\n");
1575 			goto fail;
1576 		}
1577 	}
1578 #endif
1579 	return (0);
1580 fail:
1581 	/* We free all, it handles case where we are in the middle */
1582 	iflib_tx_structures_free(ctx);
1583 	return (err);
1584 }
1585 
1586 static void
1587 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1588 {
1589 	bus_dmamap_t map;
1590 
1591 	map = NULL;
1592 	if (txq->ift_sds.ifsd_map != NULL)
1593 		map = txq->ift_sds.ifsd_map[i];
1594 	if (map != NULL) {
1595 		bus_dmamap_unload(txq->ift_desc_tag, map);
1596 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1597 		txq->ift_sds.ifsd_map[i] = NULL;
1598 	}
1599 }
1600 
1601 static void
1602 iflib_txq_destroy(iflib_txq_t txq)
1603 {
1604 	if_ctx_t ctx = txq->ift_ctx;
1605 
1606 	for (int i = 0; i < txq->ift_size; i++)
1607 		iflib_txsd_destroy(ctx, txq, i);
1608 	if (txq->ift_sds.ifsd_map != NULL) {
1609 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1610 		txq->ift_sds.ifsd_map = NULL;
1611 	}
1612 	if (txq->ift_sds.ifsd_m != NULL) {
1613 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1614 		txq->ift_sds.ifsd_m = NULL;
1615 	}
1616 	if (txq->ift_sds.ifsd_flags != NULL) {
1617 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1618 		txq->ift_sds.ifsd_flags = NULL;
1619 	}
1620 	if (txq->ift_desc_tag != NULL) {
1621 		bus_dma_tag_destroy(txq->ift_desc_tag);
1622 		txq->ift_desc_tag = NULL;
1623 	}
1624 	if (txq->ift_tso_desc_tag != NULL) {
1625 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1626 		txq->ift_tso_desc_tag = NULL;
1627 	}
1628 }
1629 
1630 static void
1631 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1632 {
1633 	struct mbuf **mp;
1634 
1635 	mp = &txq->ift_sds.ifsd_m[i];
1636 	if (*mp == NULL)
1637 		return;
1638 
1639 	if (txq->ift_sds.ifsd_map != NULL) {
1640 		bus_dmamap_sync(txq->ift_desc_tag,
1641 				txq->ift_sds.ifsd_map[i],
1642 				BUS_DMASYNC_POSTWRITE);
1643 		bus_dmamap_unload(txq->ift_desc_tag,
1644 				  txq->ift_sds.ifsd_map[i]);
1645 	}
1646 	m_free(*mp);
1647 	DBG_COUNTER_INC(tx_frees);
1648 	*mp = NULL;
1649 }
1650 
1651 static int
1652 iflib_txq_setup(iflib_txq_t txq)
1653 {
1654 	if_ctx_t ctx = txq->ift_ctx;
1655 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1656 	iflib_dma_info_t di;
1657 	int i;
1658 
1659 	/* Set number of descriptors available */
1660 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1661 	/* XXX make configurable */
1662 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1663 
1664 	/* Reset indices */
1665 	txq->ift_cidx_processed = 0;
1666 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1667 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1668 
1669 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1670 		bzero((void *)di->idi_vaddr, di->idi_size);
1671 
1672 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1673 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1674 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1675 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1676 	return (0);
1677 }
1678 
1679 /*********************************************************************
1680  *
1681  *  Allocate memory for rx_buffer structures. Since we use one
1682  *  rx_buffer per received packet, the maximum number of rx_buffer's
1683  *  that we'll need is equal to the number of receive descriptors
1684  *  that we've allocated.
1685  *
1686  **********************************************************************/
1687 static int
1688 iflib_rxsd_alloc(iflib_rxq_t rxq)
1689 {
1690 	if_ctx_t ctx = rxq->ifr_ctx;
1691 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1692 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1693 	device_t dev = ctx->ifc_dev;
1694 	iflib_fl_t fl;
1695 	int			err;
1696 
1697 	MPASS(scctx->isc_nrxd[0] > 0);
1698 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1699 
1700 	fl = rxq->ifr_fl;
1701 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1702 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1703 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1704 					 1, 0,			/* alignment, bounds */
1705 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1706 					 BUS_SPACE_MAXADDR,	/* highaddr */
1707 					 NULL, NULL,		/* filter, filterarg */
1708 					 sctx->isc_rx_maxsize,	/* maxsize */
1709 					 sctx->isc_rx_nsegments,	/* nsegments */
1710 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1711 					 0,			/* flags */
1712 					 NULL,			/* lockfunc */
1713 					 NULL,			/* lockarg */
1714 					 &fl->ifl_desc_tag);
1715 		if (err) {
1716 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1717 				__func__, err);
1718 			goto fail;
1719 		}
1720 		if (!(fl->ifl_sds.ifsd_flags =
1721 		      (uint8_t *) malloc(sizeof(uint8_t) *
1722 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1723 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1724 			err = ENOMEM;
1725 			goto fail;
1726 		}
1727 		if (!(fl->ifl_sds.ifsd_m =
1728 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1729 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1730 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1731 			err = ENOMEM;
1732 			goto fail;
1733 		}
1734 		if (!(fl->ifl_sds.ifsd_cl =
1735 		      (caddr_t *) malloc(sizeof(caddr_t) *
1736 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1737 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1738 			err = ENOMEM;
1739 			goto fail;
1740 		}
1741 
1742 		/* Create the descriptor buffer dma maps */
1743 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1744 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1745 			continue;
1746 
1747 		if (!(fl->ifl_sds.ifsd_map =
1748 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1749 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1750 			err = ENOMEM;
1751 			goto fail;
1752 		}
1753 
1754 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1755 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1756 			if (err != 0) {
1757 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1758 				goto fail;
1759 			}
1760 		}
1761 #endif
1762 	}
1763 	return (0);
1764 
1765 fail:
1766 	iflib_rx_structures_free(ctx);
1767 	return (err);
1768 }
1769 
1770 
1771 /*
1772  * Internal service routines
1773  */
1774 
1775 struct rxq_refill_cb_arg {
1776 	int               error;
1777 	bus_dma_segment_t seg;
1778 	int               nseg;
1779 };
1780 
1781 static void
1782 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1783 {
1784 	struct rxq_refill_cb_arg *cb_arg = arg;
1785 
1786 	cb_arg->error = error;
1787 	cb_arg->seg = segs[0];
1788 	cb_arg->nseg = nseg;
1789 }
1790 
1791 
1792 #ifdef ACPI_DMAR
1793 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1794 #else
1795 #define IS_DMAR(ctx) (0)
1796 #endif
1797 
1798 /**
1799  *	rxq_refill - refill an rxq  free-buffer list
1800  *	@ctx: the iflib context
1801  *	@rxq: the free-list to refill
1802  *	@n: the number of new buffers to allocate
1803  *
1804  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1805  *	The caller must assure that @n does not exceed the queue's capacity.
1806  */
1807 static void
1808 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1809 {
1810 	struct mbuf *m;
1811 	int idx, frag_idx = fl->ifl_fragidx;
1812         int pidx = fl->ifl_pidx;
1813 	caddr_t cl, *sd_cl;
1814 	struct mbuf **sd_m;
1815 	uint8_t *sd_flags;
1816 	struct if_rxd_update iru;
1817 	bus_dmamap_t *sd_map;
1818 	int n, i = 0;
1819 	uint64_t bus_addr;
1820 	int err;
1821 	qidx_t credits;
1822 
1823 	sd_m = fl->ifl_sds.ifsd_m;
1824 	sd_map = fl->ifl_sds.ifsd_map;
1825 	sd_cl = fl->ifl_sds.ifsd_cl;
1826 	sd_flags = fl->ifl_sds.ifsd_flags;
1827 	idx = pidx;
1828 	credits = fl->ifl_credits;
1829 
1830 	n  = count;
1831 	MPASS(n > 0);
1832 	MPASS(credits + n <= fl->ifl_size);
1833 
1834 	if (pidx < fl->ifl_cidx)
1835 		MPASS(pidx + n <= fl->ifl_cidx);
1836 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1837 		MPASS(fl->ifl_gen == 0);
1838 	if (pidx > fl->ifl_cidx)
1839 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1840 
1841 	DBG_COUNTER_INC(fl_refills);
1842 	if (n > 8)
1843 		DBG_COUNTER_INC(fl_refills_large);
1844 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1845 	while (n--) {
1846 		/*
1847 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1848 		 * initialized after rx.
1849 		 *
1850 		 * If the cluster is still set then we know a minimum sized packet was received
1851 		 */
1852 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1853 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1854                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1855 		if ((cl = sd_cl[frag_idx]) == NULL) {
1856                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1857 				break;
1858 #if MEMORY_LOGGING
1859 			fl->ifl_cl_enqueued++;
1860 #endif
1861 		}
1862 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1863 			break;
1864 		}
1865 #if MEMORY_LOGGING
1866 		fl->ifl_m_enqueued++;
1867 #endif
1868 
1869 		DBG_COUNTER_INC(rx_allocs);
1870 #if defined(__i386__) || defined(__amd64__)
1871 		if (!IS_DMAR(ctx)) {
1872 			bus_addr = pmap_kextract((vm_offset_t)cl);
1873 		} else
1874 #endif
1875 		{
1876 			struct rxq_refill_cb_arg cb_arg;
1877 			iflib_rxq_t q;
1878 
1879 			cb_arg.error = 0;
1880 			q = fl->ifl_rxq;
1881 			MPASS(sd_map != NULL);
1882 			MPASS(sd_map[frag_idx] != NULL);
1883 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1884 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1885 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1886 					BUS_DMASYNC_PREREAD);
1887 
1888 			if (err != 0 || cb_arg.error) {
1889 				/*
1890 				 * !zone_pack ?
1891 				 */
1892 				if (fl->ifl_zone == zone_pack)
1893 					uma_zfree(fl->ifl_zone, cl);
1894 				m_free(m);
1895 				n = 0;
1896 				goto done;
1897 			}
1898 			bus_addr = cb_arg.seg.ds_addr;
1899 		}
1900                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1901 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1902 
1903 		MPASS(sd_m[frag_idx] == NULL);
1904 		sd_cl[frag_idx] = cl;
1905 		sd_m[frag_idx] = m;
1906 		fl->ifl_rxd_idxs[i] = frag_idx;
1907 		fl->ifl_bus_addrs[i] = bus_addr;
1908 		fl->ifl_vm_addrs[i] = cl;
1909 		credits++;
1910 		i++;
1911 		MPASS(credits <= fl->ifl_size);
1912 		if (++idx == fl->ifl_size) {
1913 			fl->ifl_gen = 1;
1914 			idx = 0;
1915 		}
1916 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1917 			iru.iru_pidx = pidx;
1918 			iru.iru_count = i;
1919 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1920 			i = 0;
1921 			pidx = idx;
1922 			fl->ifl_pidx = idx;
1923 			fl->ifl_credits = credits;
1924 		}
1925 
1926 	}
1927 done:
1928 	if (i) {
1929 		iru.iru_pidx = pidx;
1930 		iru.iru_count = i;
1931 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1932 		fl->ifl_pidx = idx;
1933 		fl->ifl_credits = credits;
1934 	}
1935 	DBG_COUNTER_INC(rxd_flush);
1936 	if (fl->ifl_pidx == 0)
1937 		pidx = fl->ifl_size - 1;
1938 	else
1939 		pidx = fl->ifl_pidx - 1;
1940 
1941 	if (sd_map)
1942 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1943 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1944 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1945 	fl->ifl_fragidx = frag_idx;
1946 }
1947 
1948 static __inline void
1949 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1950 {
1951 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1952 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1953 #ifdef INVARIANTS
1954 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1955 #endif
1956 
1957 	MPASS(fl->ifl_credits <= fl->ifl_size);
1958 	MPASS(reclaimable == delta);
1959 
1960 	if (reclaimable > 0)
1961 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1962 }
1963 
1964 static void
1965 iflib_fl_bufs_free(iflib_fl_t fl)
1966 {
1967 	iflib_dma_info_t idi = fl->ifl_ifdi;
1968 	uint32_t i;
1969 
1970 	for (i = 0; i < fl->ifl_size; i++) {
1971 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1972 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1973 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1974 
1975 		if (*sd_flags & RX_SW_DESC_INUSE) {
1976 			if (fl->ifl_sds.ifsd_map != NULL) {
1977 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1978 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1979 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1980 			}
1981 			if (*sd_m != NULL) {
1982 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1983 				uma_zfree(zone_mbuf, *sd_m);
1984 			}
1985 			if (*sd_cl != NULL)
1986 				uma_zfree(fl->ifl_zone, *sd_cl);
1987 			*sd_flags = 0;
1988 		} else {
1989 			MPASS(*sd_cl == NULL);
1990 			MPASS(*sd_m == NULL);
1991 		}
1992 #if MEMORY_LOGGING
1993 		fl->ifl_m_dequeued++;
1994 		fl->ifl_cl_dequeued++;
1995 #endif
1996 		*sd_cl = NULL;
1997 		*sd_m = NULL;
1998 	}
1999 #ifdef INVARIANTS
2000 	for (i = 0; i < fl->ifl_size; i++) {
2001 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2002 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2003 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2004 	}
2005 #endif
2006 	/*
2007 	 * Reset free list values
2008 	 */
2009 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2010 	bzero(idi->idi_vaddr, idi->idi_size);
2011 }
2012 
2013 /*********************************************************************
2014  *
2015  *  Initialize a receive ring and its buffers.
2016  *
2017  **********************************************************************/
2018 static int
2019 iflib_fl_setup(iflib_fl_t fl)
2020 {
2021 	iflib_rxq_t rxq = fl->ifl_rxq;
2022 	if_ctx_t ctx = rxq->ifr_ctx;
2023 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2024 
2025 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size);
2026 	/*
2027 	** Free current RX buffer structs and their mbufs
2028 	*/
2029 	iflib_fl_bufs_free(fl);
2030 	/* Now replenish the mbufs */
2031 	MPASS(fl->ifl_credits == 0);
2032 	/*
2033 	 * XXX don't set the max_frame_size to larger
2034 	 * than the hardware can handle
2035 	 */
2036 	if (sctx->isc_max_frame_size <= 2048)
2037 		fl->ifl_buf_size = MCLBYTES;
2038 #ifndef CONTIGMALLOC_WORKS
2039 	else
2040 		fl->ifl_buf_size = MJUMPAGESIZE;
2041 #else
2042 	else if (sctx->isc_max_frame_size <= 4096)
2043 		fl->ifl_buf_size = MJUMPAGESIZE;
2044 	else if (sctx->isc_max_frame_size <= 9216)
2045 		fl->ifl_buf_size = MJUM9BYTES;
2046 	else
2047 		fl->ifl_buf_size = MJUM16BYTES;
2048 #endif
2049 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2050 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2051 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2052 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2053 
2054 
2055 	/* avoid pre-allocating zillions of clusters to an idle card
2056 	 * potentially speeding up attach
2057 	 */
2058 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2059 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2060 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2061 		return (ENOBUFS);
2062 	/*
2063 	 * handle failure
2064 	 */
2065 	MPASS(rxq != NULL);
2066 	MPASS(fl->ifl_ifdi != NULL);
2067 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2068 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2069 	return (0);
2070 }
2071 
2072 /*********************************************************************
2073  *
2074  *  Free receive ring data structures
2075  *
2076  **********************************************************************/
2077 static void
2078 iflib_rx_sds_free(iflib_rxq_t rxq)
2079 {
2080 	iflib_fl_t fl;
2081 	int i;
2082 
2083 	if (rxq->ifr_fl != NULL) {
2084 		for (i = 0; i < rxq->ifr_nfl; i++) {
2085 			fl = &rxq->ifr_fl[i];
2086 			if (fl->ifl_desc_tag != NULL) {
2087 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2088 				fl->ifl_desc_tag = NULL;
2089 			}
2090 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2091 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2092 			/* XXX destroy maps first */
2093 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2094 			fl->ifl_sds.ifsd_m = NULL;
2095 			fl->ifl_sds.ifsd_cl = NULL;
2096 			fl->ifl_sds.ifsd_map = NULL;
2097 		}
2098 		free(rxq->ifr_fl, M_IFLIB);
2099 		rxq->ifr_fl = NULL;
2100 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2101 	}
2102 }
2103 
2104 /*
2105  * MI independent logic
2106  *
2107  */
2108 static void
2109 iflib_timer(void *arg)
2110 {
2111 	iflib_txq_t txq = arg;
2112 	if_ctx_t ctx = txq->ift_ctx;
2113 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2114 
2115 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2116 		return;
2117 	/*
2118 	** Check on the state of the TX queue(s), this
2119 	** can be done without the lock because its RO
2120 	** and the HUNG state will be static if set.
2121 	*/
2122 	IFDI_TIMER(ctx, txq->ift_id);
2123 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2124 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2125 	     (sctx->isc_pause_frames == 0)))
2126 		goto hung;
2127 
2128 	if (ifmp_ring_is_stalled(txq->ift_br))
2129 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2130 	txq->ift_cleaned_prev = txq->ift_cleaned;
2131 	/* handle any laggards */
2132 	if (txq->ift_db_pending)
2133 		GROUPTASK_ENQUEUE(&txq->ift_task);
2134 
2135 	sctx->isc_pause_frames = 0;
2136 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2137 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2138 	return;
2139 hung:
2140 	CTX_LOCK(ctx);
2141 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2142 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2143 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2144 
2145 	IFDI_WATCHDOG_RESET(ctx);
2146 	ctx->ifc_watchdog_events++;
2147 
2148 	ctx->ifc_flags |= IFC_DO_RESET;
2149 	iflib_admin_intr_deferred(ctx);
2150 	CTX_UNLOCK(ctx);
2151 }
2152 
2153 static void
2154 iflib_init_locked(if_ctx_t ctx)
2155 {
2156 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2157 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2158 	if_t ifp = ctx->ifc_ifp;
2159 	iflib_fl_t fl;
2160 	iflib_txq_t txq;
2161 	iflib_rxq_t rxq;
2162 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2163 
2164 
2165 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2166 	IFDI_INTR_DISABLE(ctx);
2167 
2168 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2169 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2170 	/* Set hardware offload abilities */
2171 	if_clearhwassist(ifp);
2172 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2173 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2174 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2175 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2176 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2177 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2178 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2179 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2180 
2181 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2182 		CALLOUT_LOCK(txq);
2183 		callout_stop(&txq->ift_timer);
2184 		CALLOUT_UNLOCK(txq);
2185 		iflib_netmap_txq_init(ctx, txq);
2186 	}
2187 #ifdef INVARIANTS
2188 	i = if_getdrvflags(ifp);
2189 #endif
2190 	IFDI_INIT(ctx);
2191 	MPASS(if_getdrvflags(ifp) == i);
2192 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2193 		/* XXX this should really be done on a per-queue basis */
2194 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2195 			MPASS(rxq->ifr_id == i);
2196 			iflib_netmap_rxq_init(ctx, rxq);
2197 			continue;
2198 		}
2199 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2200 			if (iflib_fl_setup(fl)) {
2201 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2202 				goto done;
2203 			}
2204 		}
2205 	}
2206 	done:
2207 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2208 	IFDI_INTR_ENABLE(ctx);
2209 	txq = ctx->ifc_txqs;
2210 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2211 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2212 			txq->ift_timer.c_cpu);
2213 }
2214 
2215 static int
2216 iflib_media_change(if_t ifp)
2217 {
2218 	if_ctx_t ctx = if_getsoftc(ifp);
2219 	int err;
2220 
2221 	CTX_LOCK(ctx);
2222 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2223 		iflib_init_locked(ctx);
2224 	CTX_UNLOCK(ctx);
2225 	return (err);
2226 }
2227 
2228 static void
2229 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2230 {
2231 	if_ctx_t ctx = if_getsoftc(ifp);
2232 
2233 	CTX_LOCK(ctx);
2234 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2235 	IFDI_MEDIA_STATUS(ctx, ifmr);
2236 	CTX_UNLOCK(ctx);
2237 }
2238 
2239 static void
2240 iflib_stop(if_ctx_t ctx)
2241 {
2242 	iflib_txq_t txq = ctx->ifc_txqs;
2243 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2244 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2245 	iflib_dma_info_t di;
2246 	iflib_fl_t fl;
2247 	int i, j;
2248 
2249 	/* Tell the stack that the interface is no longer active */
2250 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2251 
2252 	IFDI_INTR_DISABLE(ctx);
2253 	DELAY(1000);
2254 	IFDI_STOP(ctx);
2255 	DELAY(1000);
2256 
2257 	iflib_debug_reset();
2258 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2259 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2260 		/* make sure all transmitters have completed before proceeding XXX */
2261 
2262 		/* clean any enqueued buffers */
2263 		iflib_ifmp_purge(txq);
2264 		/* Free any existing tx buffers. */
2265 		for (j = 0; j < txq->ift_size; j++) {
2266 			iflib_txsd_free(ctx, txq, j);
2267 		}
2268 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2269 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2270 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2271 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2272 		txq->ift_pullups = 0;
2273 		ifmp_ring_reset_stats(txq->ift_br);
2274 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2275 			bzero((void *)di->idi_vaddr, di->idi_size);
2276 	}
2277 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2278 		/* make sure all transmitters have completed before proceeding XXX */
2279 
2280 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2281 			bzero((void *)di->idi_vaddr, di->idi_size);
2282 		/* also resets the free lists pidx/cidx */
2283 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2284 			iflib_fl_bufs_free(fl);
2285 	}
2286 }
2287 
2288 static inline caddr_t
2289 calc_next_rxd(iflib_fl_t fl, int cidx)
2290 {
2291 	qidx_t size;
2292 	int nrxd;
2293 	caddr_t start, end, cur, next;
2294 
2295 	nrxd = fl->ifl_size;
2296 	size = fl->ifl_rxd_size;
2297 	start = fl->ifl_ifdi->idi_vaddr;
2298 
2299 	if (__predict_false(size == 0))
2300 		return (start);
2301 	cur = start + size*cidx;
2302 	end = start + size*nrxd;
2303 	next = CACHE_PTR_NEXT(cur);
2304 	return (next < end ? next : start);
2305 }
2306 
2307 static inline void
2308 prefetch_pkts(iflib_fl_t fl, int cidx)
2309 {
2310 	int nextptr;
2311 	int nrxd = fl->ifl_size;
2312 	caddr_t next_rxd;
2313 
2314 
2315 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2316 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2317 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2318 	next_rxd = calc_next_rxd(fl, cidx);
2319 	prefetch(next_rxd);
2320 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2321 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2322 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2323 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2324 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2325 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2326 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2327 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2328 }
2329 
2330 static void
2331 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2332 {
2333 	int flid, cidx;
2334 	bus_dmamap_t map;
2335 	iflib_fl_t fl;
2336 	iflib_dma_info_t di;
2337 	int next;
2338 
2339 	map = NULL;
2340 	flid = irf->irf_flid;
2341 	cidx = irf->irf_idx;
2342 	fl = &rxq->ifr_fl[flid];
2343 	sd->ifsd_fl = fl;
2344 	sd->ifsd_cidx = cidx;
2345 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2346 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2347 	fl->ifl_credits--;
2348 #if MEMORY_LOGGING
2349 	fl->ifl_m_dequeued++;
2350 #endif
2351 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2352 		prefetch_pkts(fl, cidx);
2353 	if (fl->ifl_sds.ifsd_map != NULL) {
2354 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2355 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2356 		map = fl->ifl_sds.ifsd_map[cidx];
2357 		di = fl->ifl_ifdi;
2358 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2359 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2360 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2361 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2362 
2363 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2364 		MPASS(fl->ifl_cidx == cidx);
2365 		if (unload)
2366 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2367 	}
2368 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2369 	if (__predict_false(fl->ifl_cidx == 0))
2370 		fl->ifl_gen = 0;
2371 	if (map != NULL)
2372 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2373 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2374         bit_clear(fl->ifl_rx_bitmap, cidx);
2375 }
2376 
2377 static struct mbuf *
2378 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2379 {
2380 	int i, padlen , flags;
2381 	struct mbuf *m, *mh, *mt;
2382 	caddr_t cl;
2383 
2384 	i = 0;
2385 	mh = NULL;
2386 	do {
2387 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2388 
2389 		MPASS(*sd->ifsd_cl != NULL);
2390 		MPASS(*sd->ifsd_m != NULL);
2391 
2392 		/* Don't include zero-length frags */
2393 		if (ri->iri_frags[i].irf_len == 0) {
2394 			/* XXX we can save the cluster here, but not the mbuf */
2395 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2396 			m_free(*sd->ifsd_m);
2397 			*sd->ifsd_m = NULL;
2398 			continue;
2399 		}
2400 		m = *sd->ifsd_m;
2401 		*sd->ifsd_m = NULL;
2402 		if (mh == NULL) {
2403 			flags = M_PKTHDR|M_EXT;
2404 			mh = mt = m;
2405 			padlen = ri->iri_pad;
2406 		} else {
2407 			flags = M_EXT;
2408 			mt->m_next = m;
2409 			mt = m;
2410 			/* assuming padding is only on the first fragment */
2411 			padlen = 0;
2412 		}
2413 		cl = *sd->ifsd_cl;
2414 		*sd->ifsd_cl = NULL;
2415 
2416 		/* Can these two be made one ? */
2417 		m_init(m, M_NOWAIT, MT_DATA, flags);
2418 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2419 		/*
2420 		 * These must follow m_init and m_cljset
2421 		 */
2422 		m->m_data += padlen;
2423 		ri->iri_len -= padlen;
2424 		m->m_len = ri->iri_frags[i].irf_len;
2425 	} while (++i < ri->iri_nfrags);
2426 
2427 	return (mh);
2428 }
2429 
2430 /*
2431  * Process one software descriptor
2432  */
2433 static struct mbuf *
2434 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2435 {
2436 	struct if_rxsd sd;
2437 	struct mbuf *m;
2438 
2439 	/* should I merge this back in now that the two paths are basically duplicated? */
2440 	if (ri->iri_nfrags == 1 &&
2441 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2442 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2443 		m = *sd.ifsd_m;
2444 		*sd.ifsd_m = NULL;
2445 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2446 #ifndef __NO_STRICT_ALIGNMENT
2447 		if (!IP_ALIGNED(m))
2448 			m->m_data += 2;
2449 #endif
2450 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2451 		m->m_len = ri->iri_frags[0].irf_len;
2452        } else {
2453 		m = assemble_segments(rxq, ri, &sd);
2454 	}
2455 	m->m_pkthdr.len = ri->iri_len;
2456 	m->m_pkthdr.rcvif = ri->iri_ifp;
2457 	m->m_flags |= ri->iri_flags;
2458 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2459 	m->m_pkthdr.flowid = ri->iri_flowid;
2460 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2461 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2462 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2463 	return (m);
2464 }
2465 
2466 static bool
2467 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2468 {
2469 	if_ctx_t ctx = rxq->ifr_ctx;
2470 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2471 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2472 	int avail, i;
2473 	qidx_t *cidxp;
2474 	struct if_rxd_info ri;
2475 	int err, budget_left, rx_bytes, rx_pkts;
2476 	iflib_fl_t fl;
2477 	struct ifnet *ifp;
2478 	int lro_enabled;
2479 
2480 	/*
2481 	 * XXX early demux data packets so that if_input processing only handles
2482 	 * acks in interrupt context
2483 	 */
2484 	struct mbuf *m, *mh, *mt, *mf;
2485 
2486 	ifp = ctx->ifc_ifp;
2487 	mh = mt = NULL;
2488 	MPASS(budget > 0);
2489 	rx_pkts	= rx_bytes = 0;
2490 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2491 		cidxp = &rxq->ifr_cq_cidx;
2492 	else
2493 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2494 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2495 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2496 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2497 		DBG_COUNTER_INC(rx_unavail);
2498 		return (false);
2499 	}
2500 
2501 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2502 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2503 			DBG_COUNTER_INC(rx_ctx_inactive);
2504 			break;
2505 		}
2506 		/*
2507 		 * Reset client set fields to their default values
2508 		 */
2509 		rxd_info_zero(&ri);
2510 		ri.iri_qsidx = rxq->ifr_id;
2511 		ri.iri_cidx = *cidxp;
2512 		ri.iri_ifp = ifp;
2513 		ri.iri_frags = rxq->ifr_frags;
2514 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2515 
2516 		if (err)
2517 			goto err;
2518 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2519 			*cidxp = ri.iri_cidx;
2520 			/* Update our consumer index */
2521 			/* XXX NB: shurd - check if this is still safe */
2522 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2523 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2524 				rxq->ifr_cq_gen = 0;
2525 			}
2526 			/* was this only a completion queue message? */
2527 			if (__predict_false(ri.iri_nfrags == 0))
2528 				continue;
2529 		}
2530 		MPASS(ri.iri_nfrags != 0);
2531 		MPASS(ri.iri_len != 0);
2532 
2533 		/* will advance the cidx on the corresponding free lists */
2534 		m = iflib_rxd_pkt_get(rxq, &ri);
2535 		if (avail == 0 && budget_left)
2536 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2537 
2538 		if (__predict_false(m == NULL)) {
2539 			DBG_COUNTER_INC(rx_mbuf_null);
2540 			continue;
2541 		}
2542 		/* imm_pkt: -- cxgb */
2543 		if (mh == NULL)
2544 			mh = mt = m;
2545 		else {
2546 			mt->m_nextpkt = m;
2547 			mt = m;
2548 		}
2549 	}
2550 	/* make sure that we can refill faster than drain */
2551 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2552 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2553 
2554 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2555 	mt = mf = NULL;
2556 	while (mh != NULL) {
2557 		m = mh;
2558 		if (mf == NULL)
2559 			mf = m;
2560 		mh = mh->m_nextpkt;
2561 		m->m_nextpkt = NULL;
2562 #ifndef __NO_STRICT_ALIGNMENT
2563 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2564 			continue;
2565 #endif
2566 		rx_bytes += m->m_pkthdr.len;
2567 		rx_pkts++;
2568 #if defined(INET6) || defined(INET)
2569 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0) {
2570 			if (mf == m)
2571 				mf = NULL;
2572 			continue;
2573 		}
2574 #endif
2575 		if (mt != NULL)
2576 			mt->m_nextpkt = m;
2577 		mt = m;
2578 	}
2579 	if (mf != NULL) {
2580 		ifp->if_input(ifp, mf);
2581 		DBG_COUNTER_INC(rx_if_input);
2582 	}
2583 
2584 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2585 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2586 
2587 	/*
2588 	 * Flush any outstanding LRO work
2589 	 */
2590 #if defined(INET6) || defined(INET)
2591 	tcp_lro_flush_all(&rxq->ifr_lc);
2592 #endif
2593 	if (avail)
2594 		return true;
2595 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2596 err:
2597 	CTX_LOCK(ctx);
2598 	ctx->ifc_flags |= IFC_DO_RESET;
2599 	iflib_admin_intr_deferred(ctx);
2600 	CTX_UNLOCK(ctx);
2601 	return (false);
2602 }
2603 
2604 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2605 static inline qidx_t
2606 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2607 {
2608 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2609 	qidx_t minthresh = txq->ift_size / 8;
2610 	if (in_use > 4*minthresh)
2611 		return (notify_count);
2612 	if (in_use > 2*minthresh)
2613 		return (notify_count >> 1);
2614 	if (in_use > minthresh)
2615 		return (notify_count >> 3);
2616 	return (0);
2617 }
2618 
2619 static inline qidx_t
2620 txq_max_rs_deferred(iflib_txq_t txq)
2621 {
2622 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2623 	qidx_t minthresh = txq->ift_size / 8;
2624 	if (txq->ift_in_use > 4*minthresh)
2625 		return (notify_count);
2626 	if (txq->ift_in_use > 2*minthresh)
2627 		return (notify_count >> 1);
2628 	if (txq->ift_in_use > minthresh)
2629 		return (notify_count >> 2);
2630 	return (2);
2631 }
2632 
2633 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2634 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2635 
2636 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2637 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2638 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2639 
2640 /* forward compatibility for cxgb */
2641 #define FIRST_QSET(ctx) 0
2642 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2643 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2644 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2645 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2646 
2647 /* XXX we should be setting this to something other than zero */
2648 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2649 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2650 
2651 static inline bool
2652 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2653 {
2654 	qidx_t dbval, max;
2655 	bool rang;
2656 
2657 	rang = false;
2658 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2659 	if (ring || txq->ift_db_pending >= max) {
2660 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2661 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2662 		txq->ift_db_pending = txq->ift_npending = 0;
2663 		rang = true;
2664 	}
2665 	return (rang);
2666 }
2667 
2668 #ifdef PKT_DEBUG
2669 static void
2670 print_pkt(if_pkt_info_t pi)
2671 {
2672 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2673 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2674 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2675 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2676 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2677 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2678 }
2679 #endif
2680 
2681 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2682 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2683 
2684 static int
2685 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2686 {
2687 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2688 	struct ether_vlan_header *eh;
2689 	struct mbuf *m, *n;
2690 
2691 	n = m = *mp;
2692 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2693 	    M_WRITABLE(m) == 0) {
2694 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2695 			return (ENOMEM);
2696 		} else {
2697 			m_freem(*mp);
2698 			n = *mp = m;
2699 		}
2700 	}
2701 
2702 	/*
2703 	 * Determine where frame payload starts.
2704 	 * Jump over vlan headers if already present,
2705 	 * helpful for QinQ too.
2706 	 */
2707 	if (__predict_false(m->m_len < sizeof(*eh))) {
2708 		txq->ift_pullups++;
2709 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2710 			return (ENOMEM);
2711 	}
2712 	eh = mtod(m, struct ether_vlan_header *);
2713 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2714 		pi->ipi_etype = ntohs(eh->evl_proto);
2715 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2716 	} else {
2717 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2718 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2719 	}
2720 
2721 	switch (pi->ipi_etype) {
2722 #ifdef INET
2723 	case ETHERTYPE_IP:
2724 	{
2725 		struct ip *ip = NULL;
2726 		struct tcphdr *th = NULL;
2727 		int minthlen;
2728 
2729 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2730 		if (__predict_false(m->m_len < minthlen)) {
2731 			/*
2732 			 * if this code bloat is causing too much of a hit
2733 			 * move it to a separate function and mark it noinline
2734 			 */
2735 			if (m->m_len == pi->ipi_ehdrlen) {
2736 				n = m->m_next;
2737 				MPASS(n);
2738 				if (n->m_len >= sizeof(*ip))  {
2739 					ip = (struct ip *)n->m_data;
2740 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2741 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2742 				} else {
2743 					txq->ift_pullups++;
2744 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2745 						return (ENOMEM);
2746 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2747 				}
2748 			} else {
2749 				txq->ift_pullups++;
2750 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2751 					return (ENOMEM);
2752 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2753 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2754 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2755 			}
2756 		} else {
2757 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2758 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2759 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2760 		}
2761 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2762 		pi->ipi_ipproto = ip->ip_p;
2763 		pi->ipi_flags |= IPI_TX_IPV4;
2764 
2765 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2766                        ip->ip_sum = 0;
2767 
2768 		if (IS_TSO4(pi)) {
2769 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2770 				if (__predict_false(th == NULL)) {
2771 					txq->ift_pullups++;
2772 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2773 						return (ENOMEM);
2774 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2775 				}
2776 				pi->ipi_tcp_hflags = th->th_flags;
2777 				pi->ipi_tcp_hlen = th->th_off << 2;
2778 				pi->ipi_tcp_seq = th->th_seq;
2779 			}
2780 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2781 				return (ENXIO);
2782 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2783 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2784 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2785 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2786 				ip->ip_sum = 0;
2787 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2788 			}
2789 		}
2790 		break;
2791 	}
2792 #endif
2793 #ifdef INET6
2794 	case ETHERTYPE_IPV6:
2795 	{
2796 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2797 		struct tcphdr *th;
2798 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2799 
2800 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2801 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2802 				return (ENOMEM);
2803 		}
2804 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2805 
2806 		/* XXX-BZ this will go badly in case of ext hdrs. */
2807 		pi->ipi_ipproto = ip6->ip6_nxt;
2808 		pi->ipi_flags |= IPI_TX_IPV6;
2809 
2810 		if (IS_TSO6(pi)) {
2811 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2812 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2813 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2814 						return (ENOMEM);
2815 				}
2816 				pi->ipi_tcp_hflags = th->th_flags;
2817 				pi->ipi_tcp_hlen = th->th_off << 2;
2818 			}
2819 
2820 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2821 				return (ENXIO);
2822 			/*
2823 			 * The corresponding flag is set by the stack in the IPv4
2824 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2825 			 * So, set it here because the rest of the flow requires it.
2826 			 */
2827 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2828 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2829 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2830 		}
2831 		break;
2832 	}
2833 #endif
2834 	default:
2835 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2836 		pi->ipi_ip_hlen = 0;
2837 		break;
2838 	}
2839 	*mp = m;
2840 
2841 	return (0);
2842 }
2843 
2844 static  __noinline  struct mbuf *
2845 collapse_pkthdr(struct mbuf *m0)
2846 {
2847 	struct mbuf *m, *m_next, *tmp;
2848 
2849 	m = m0;
2850 	m_next = m->m_next;
2851 	while (m_next != NULL && m_next->m_len == 0) {
2852 		m = m_next;
2853 		m->m_next = NULL;
2854 		m_free(m);
2855 		m_next = m_next->m_next;
2856 	}
2857 	m = m0;
2858 	m->m_next = m_next;
2859 	if ((m_next->m_flags & M_EXT) == 0) {
2860 		m = m_defrag(m, M_NOWAIT);
2861 	} else {
2862 		tmp = m_next->m_next;
2863 		memcpy(m_next, m, MPKTHSIZE);
2864 		m = m_next;
2865 		m->m_next = tmp;
2866 	}
2867 	return (m);
2868 }
2869 
2870 /*
2871  * If dodgy hardware rejects the scatter gather chain we've handed it
2872  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2873  * m_defrag'd mbufs
2874  */
2875 static __noinline struct mbuf *
2876 iflib_remove_mbuf(iflib_txq_t txq)
2877 {
2878 	int ntxd, i, pidx;
2879 	struct mbuf *m, *mh, **ifsd_m;
2880 
2881 	pidx = txq->ift_pidx;
2882 	ifsd_m = txq->ift_sds.ifsd_m;
2883 	ntxd = txq->ift_size;
2884 	mh = m = ifsd_m[pidx];
2885 	ifsd_m[pidx] = NULL;
2886 #if MEMORY_LOGGING
2887 	txq->ift_dequeued++;
2888 #endif
2889 	i = 1;
2890 
2891 	while (m) {
2892 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2893 #if MEMORY_LOGGING
2894 		txq->ift_dequeued++;
2895 #endif
2896 		m = m->m_next;
2897 		i++;
2898 	}
2899 	return (mh);
2900 }
2901 
2902 static int
2903 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2904 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2905 			  int max_segs, int flags)
2906 {
2907 	if_ctx_t ctx;
2908 	if_shared_ctx_t		sctx;
2909 	if_softc_ctx_t		scctx;
2910 	int i, next, pidx, err, ntxd, count;
2911 	struct mbuf *m, *tmp, **ifsd_m;
2912 
2913 	m = *m0;
2914 
2915 	/*
2916 	 * Please don't ever do this
2917 	 */
2918 	if (__predict_false(m->m_len == 0))
2919 		*m0 = m = collapse_pkthdr(m);
2920 
2921 	ctx = txq->ift_ctx;
2922 	sctx = ctx->ifc_sctx;
2923 	scctx = &ctx->ifc_softc_ctx;
2924 	ifsd_m = txq->ift_sds.ifsd_m;
2925 	ntxd = txq->ift_size;
2926 	pidx = txq->ift_pidx;
2927 	if (map != NULL) {
2928 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2929 
2930 		err = bus_dmamap_load_mbuf_sg(tag, map,
2931 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2932 		if (err)
2933 			return (err);
2934 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2935 		count = 0;
2936 		m = *m0;
2937 		do {
2938 			if (__predict_false(m->m_len <= 0)) {
2939 				tmp = m;
2940 				m = m->m_next;
2941 				tmp->m_next = NULL;
2942 				m_free(tmp);
2943 				continue;
2944 			}
2945 			m = m->m_next;
2946 			count++;
2947 		} while (m != NULL);
2948 		if (count > *nsegs) {
2949 			ifsd_m[pidx] = *m0;
2950 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
2951 			return (0);
2952 		}
2953 		m = *m0;
2954 		count = 0;
2955 		do {
2956 			next = (pidx + count) & (ntxd-1);
2957 			MPASS(ifsd_m[next] == NULL);
2958 			ifsd_m[next] = m;
2959 			count++;
2960 			tmp = m;
2961 			m = m->m_next;
2962 		} while (m != NULL);
2963 	} else {
2964 		int buflen, sgsize, maxsegsz, max_sgsize;
2965 		vm_offset_t vaddr;
2966 		vm_paddr_t curaddr;
2967 
2968 		count = i = 0;
2969 		m = *m0;
2970 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
2971 			maxsegsz = scctx->isc_tx_tso_segsize_max;
2972 		else
2973 			maxsegsz = sctx->isc_tx_maxsegsize;
2974 
2975 		do {
2976 			if (__predict_false(m->m_len <= 0)) {
2977 				tmp = m;
2978 				m = m->m_next;
2979 				tmp->m_next = NULL;
2980 				m_free(tmp);
2981 				continue;
2982 			}
2983 			buflen = m->m_len;
2984 			vaddr = (vm_offset_t)m->m_data;
2985 			/*
2986 			 * see if we can't be smarter about physically
2987 			 * contiguous mappings
2988 			 */
2989 			next = (pidx + count) & (ntxd-1);
2990 			MPASS(ifsd_m[next] == NULL);
2991 #if MEMORY_LOGGING
2992 			txq->ift_enqueued++;
2993 #endif
2994 			ifsd_m[next] = m;
2995 			while (buflen > 0) {
2996 				if (i >= max_segs)
2997 					goto err;
2998 				max_sgsize = MIN(buflen, maxsegsz);
2999 				curaddr = pmap_kextract(vaddr);
3000 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3001 				sgsize = MIN(sgsize, max_sgsize);
3002 				segs[i].ds_addr = curaddr;
3003 				segs[i].ds_len = sgsize;
3004 				vaddr += sgsize;
3005 				buflen -= sgsize;
3006 				i++;
3007 			}
3008 			count++;
3009 			tmp = m;
3010 			m = m->m_next;
3011 		} while (m != NULL);
3012 		*nsegs = i;
3013 	}
3014 	return (0);
3015 err:
3016 	*m0 = iflib_remove_mbuf(txq);
3017 	return (EFBIG);
3018 }
3019 
3020 static inline caddr_t
3021 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3022 {
3023 	qidx_t size;
3024 	int ntxd;
3025 	caddr_t start, end, cur, next;
3026 
3027 	ntxd = txq->ift_size;
3028 	size = txq->ift_txd_size[qid];
3029 	start = txq->ift_ifdi[qid].idi_vaddr;
3030 
3031 	if (__predict_false(size == 0))
3032 		return (start);
3033 	cur = start + size*cidx;
3034 	end = start + size*ntxd;
3035 	next = CACHE_PTR_NEXT(cur);
3036 	return (next < end ? next : start);
3037 }
3038 
3039 static int
3040 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3041 {
3042 	if_ctx_t		ctx;
3043 	if_shared_ctx_t		sctx;
3044 	if_softc_ctx_t		scctx;
3045 	bus_dma_segment_t	*segs;
3046 	struct mbuf		*m_head;
3047 	void			*next_txd;
3048 	bus_dmamap_t		map;
3049 	struct if_pkt_info	pi;
3050 	int remap = 0;
3051 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3052 	bus_dma_tag_t desc_tag;
3053 
3054 	segs = txq->ift_segs;
3055 	ctx = txq->ift_ctx;
3056 	sctx = ctx->ifc_sctx;
3057 	scctx = &ctx->ifc_softc_ctx;
3058 	segs = txq->ift_segs;
3059 	ntxd = txq->ift_size;
3060 	m_head = *m_headp;
3061 	map = NULL;
3062 
3063 	/*
3064 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3065 	 */
3066 	cidx = txq->ift_cidx;
3067 	pidx = txq->ift_pidx;
3068 	if (ctx->ifc_flags & IFC_PREFETCH) {
3069 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3070 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3071 			next_txd = calc_next_txd(txq, cidx, 0);
3072 			prefetch(next_txd);
3073 		}
3074 
3075 		/* prefetch the next cache line of mbuf pointers and flags */
3076 		prefetch(&txq->ift_sds.ifsd_m[next]);
3077 		if (txq->ift_sds.ifsd_map != NULL) {
3078 			prefetch(&txq->ift_sds.ifsd_map[next]);
3079 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3080 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3081 		}
3082 	} else if (txq->ift_sds.ifsd_map != NULL)
3083 		map = txq->ift_sds.ifsd_map[pidx];
3084 
3085 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3086 		desc_tag = txq->ift_tso_desc_tag;
3087 		max_segs = scctx->isc_tx_tso_segments_max;
3088 	} else {
3089 		desc_tag = txq->ift_desc_tag;
3090 		max_segs = scctx->isc_tx_nsegments;
3091 	}
3092 	m_head = *m_headp;
3093 
3094 	pkt_info_zero(&pi);
3095 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3096 	pi.ipi_pidx = pidx;
3097 	pi.ipi_qsidx = txq->ift_id;
3098 	pi.ipi_len = m_head->m_pkthdr.len;
3099 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3100 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3101 
3102 	/* deliberate bitwise OR to make one condition */
3103 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3104 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3105 			return (err);
3106 		m_head = *m_headp;
3107 	}
3108 
3109 retry:
3110 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3111 defrag:
3112 	if (__predict_false(err)) {
3113 		switch (err) {
3114 		case EFBIG:
3115 			/* try collapse once and defrag once */
3116 			if (remap == 0)
3117 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3118 			if (remap == 1)
3119 				m_head = m_defrag(*m_headp, M_NOWAIT);
3120 			remap++;
3121 			if (__predict_false(m_head == NULL))
3122 				goto defrag_failed;
3123 			txq->ift_mbuf_defrag++;
3124 			*m_headp = m_head;
3125 			goto retry;
3126 			break;
3127 		case ENOMEM:
3128 			txq->ift_no_tx_dma_setup++;
3129 			break;
3130 		default:
3131 			txq->ift_no_tx_dma_setup++;
3132 			m_freem(*m_headp);
3133 			DBG_COUNTER_INC(tx_frees);
3134 			*m_headp = NULL;
3135 			break;
3136 		}
3137 		txq->ift_map_failed++;
3138 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3139 		return (err);
3140 	}
3141 
3142 	/*
3143 	 * XXX assumes a 1 to 1 relationship between segments and
3144 	 *        descriptors - this does not hold true on all drivers, e.g.
3145 	 *        cxgb
3146 	 */
3147 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3148 		txq->ift_no_desc_avail++;
3149 		if (map != NULL)
3150 			bus_dmamap_unload(desc_tag, map);
3151 		DBG_COUNTER_INC(encap_txq_avail_fail);
3152 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3153 			GROUPTASK_ENQUEUE(&txq->ift_task);
3154 		return (ENOBUFS);
3155 	}
3156 	/*
3157 	 * On Intel cards we can greatly reduce the number of TX interrupts
3158 	 * we see by only setting report status on every Nth descriptor.
3159 	 * However, this also means that the driver will need to keep track
3160 	 * of the descriptors that RS was set on to check them for the DD bit.
3161 	 */
3162 	txq->ift_rs_pending += nsegs + 1;
3163 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3164 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3165 		pi.ipi_flags |= IPI_TX_INTR;
3166 		txq->ift_rs_pending = 0;
3167 	}
3168 
3169 	pi.ipi_segs = segs;
3170 	pi.ipi_nsegs = nsegs;
3171 
3172 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3173 #ifdef PKT_DEBUG
3174 	print_pkt(&pi);
3175 #endif
3176 	if (map != NULL)
3177 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3178 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3179 		if (map != NULL)
3180 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3181 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3182 		DBG_COUNTER_INC(tx_encap);
3183 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3184 
3185 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3186 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3187 			ndesc += txq->ift_size;
3188 			txq->ift_gen = 1;
3189 		}
3190 		/*
3191 		 * drivers can need as many as
3192 		 * two sentinels
3193 		 */
3194 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3195 		MPASS(pi.ipi_new_pidx != pidx);
3196 		MPASS(ndesc > 0);
3197 		txq->ift_in_use += ndesc;
3198 
3199 		/*
3200 		 * We update the last software descriptor again here because there may
3201 		 * be a sentinel and/or there may be more mbufs than segments
3202 		 */
3203 		txq->ift_pidx = pi.ipi_new_pidx;
3204 		txq->ift_npending += pi.ipi_ndescs;
3205 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3206 		*m_headp = m_head = iflib_remove_mbuf(txq);
3207 		remap = 1;
3208 		txq->ift_txd_encap_efbig++;
3209 		goto defrag;
3210 	} else
3211 		DBG_COUNTER_INC(encap_txd_encap_fail);
3212 	return (err);
3213 
3214 defrag_failed:
3215 	txq->ift_mbuf_defrag_failed++;
3216 	txq->ift_map_failed++;
3217 	m_freem(*m_headp);
3218 	DBG_COUNTER_INC(tx_frees);
3219 	*m_headp = NULL;
3220 	return (ENOMEM);
3221 }
3222 
3223 static void
3224 iflib_tx_desc_free(iflib_txq_t txq, int n)
3225 {
3226 	int hasmap;
3227 	uint32_t qsize, cidx, mask, gen;
3228 	struct mbuf *m, **ifsd_m;
3229 	uint8_t *ifsd_flags;
3230 	bus_dmamap_t *ifsd_map;
3231 	bool do_prefetch;
3232 
3233 	cidx = txq->ift_cidx;
3234 	gen = txq->ift_gen;
3235 	qsize = txq->ift_size;
3236 	mask = qsize-1;
3237 	hasmap = txq->ift_sds.ifsd_map != NULL;
3238 	ifsd_flags = txq->ift_sds.ifsd_flags;
3239 	ifsd_m = txq->ift_sds.ifsd_m;
3240 	ifsd_map = txq->ift_sds.ifsd_map;
3241 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3242 
3243 	while (n--) {
3244 		if (do_prefetch) {
3245 			prefetch(ifsd_m[(cidx + 3) & mask]);
3246 			prefetch(ifsd_m[(cidx + 4) & mask]);
3247 		}
3248 		if (ifsd_m[cidx] != NULL) {
3249 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3250 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3251 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3252 				/*
3253 				 * does it matter if it's not the TSO tag? If so we'll
3254 				 * have to add the type to flags
3255 				 */
3256 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3257 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3258 			}
3259 			if ((m = ifsd_m[cidx]) != NULL) {
3260 				/* XXX we don't support any drivers that batch packets yet */
3261 				MPASS(m->m_nextpkt == NULL);
3262 				/* if the number of clusters exceeds the number of segments
3263 				 * there won't be space on the ring to save a pointer to each
3264 				 * cluster so we simply free the list here
3265 				 */
3266 				if (m->m_flags & M_TOOBIG) {
3267 					m_freem(m);
3268 				} else {
3269 					m_free(m);
3270 				}
3271 				ifsd_m[cidx] = NULL;
3272 #if MEMORY_LOGGING
3273 				txq->ift_dequeued++;
3274 #endif
3275 				DBG_COUNTER_INC(tx_frees);
3276 			}
3277 		}
3278 		if (__predict_false(++cidx == qsize)) {
3279 			cidx = 0;
3280 			gen = 0;
3281 		}
3282 	}
3283 	txq->ift_cidx = cidx;
3284 	txq->ift_gen = gen;
3285 }
3286 
3287 static __inline int
3288 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3289 {
3290 	int reclaim;
3291 	if_ctx_t ctx = txq->ift_ctx;
3292 
3293 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3294 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3295 
3296 	/*
3297 	 * Need a rate-limiting check so that this isn't called every time
3298 	 */
3299 	iflib_tx_credits_update(ctx, txq);
3300 	reclaim = DESC_RECLAIMABLE(txq);
3301 
3302 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3303 #ifdef INVARIANTS
3304 		if (iflib_verbose_debug) {
3305 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3306 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3307 			       reclaim, thresh);
3308 
3309 		}
3310 #endif
3311 		return (0);
3312 	}
3313 	iflib_tx_desc_free(txq, reclaim);
3314 	txq->ift_cleaned += reclaim;
3315 	txq->ift_in_use -= reclaim;
3316 
3317 	return (reclaim);
3318 }
3319 
3320 static struct mbuf **
3321 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3322 {
3323 	int next, size;
3324 	struct mbuf **items;
3325 
3326 	size = r->size;
3327 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3328 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3329 
3330 	prefetch(items[(cidx + offset) & (size-1)]);
3331 	if (remaining > 1) {
3332 		prefetch2cachelines(&items[next]);
3333 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3334 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3335 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3336 	}
3337 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3338 }
3339 
3340 static void
3341 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3342 {
3343 
3344 	ifmp_ring_check_drainage(txq->ift_br, budget);
3345 }
3346 
3347 static uint32_t
3348 iflib_txq_can_drain(struct ifmp_ring *r)
3349 {
3350 	iflib_txq_t txq = r->cookie;
3351 	if_ctx_t ctx = txq->ift_ctx;
3352 
3353 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3354 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3355 }
3356 
3357 static uint32_t
3358 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3359 {
3360 	iflib_txq_t txq = r->cookie;
3361 	if_ctx_t ctx = txq->ift_ctx;
3362 	struct ifnet *ifp = ctx->ifc_ifp;
3363 	struct mbuf **mp, *m;
3364 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3365 	int reclaimed, err, in_use_prev, desc_used;
3366 	bool do_prefetch, ring, rang;
3367 
3368 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3369 			    !LINK_ACTIVE(ctx))) {
3370 		DBG_COUNTER_INC(txq_drain_notready);
3371 		return (0);
3372 	}
3373 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3374 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3375 	avail = IDXDIFF(pidx, cidx, r->size);
3376 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3377 		DBG_COUNTER_INC(txq_drain_flushing);
3378 		for (i = 0; i < avail; i++) {
3379 			m_free(r->items[(cidx + i) & (r->size-1)]);
3380 			r->items[(cidx + i) & (r->size-1)] = NULL;
3381 		}
3382 		return (avail);
3383 	}
3384 
3385 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3386 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3387 		CALLOUT_LOCK(txq);
3388 		callout_stop(&txq->ift_timer);
3389 		CALLOUT_UNLOCK(txq);
3390 		DBG_COUNTER_INC(txq_drain_oactive);
3391 		return (0);
3392 	}
3393 	if (reclaimed)
3394 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3395 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3396 	count = MIN(avail, TX_BATCH_SIZE);
3397 #ifdef INVARIANTS
3398 	if (iflib_verbose_debug)
3399 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3400 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3401 #endif
3402 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3403 	avail = TXQ_AVAIL(txq);
3404 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3405 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3406 
3407 		mp = _ring_peek_one(r, cidx, i, rem);
3408 		MPASS(mp != NULL && *mp != NULL);
3409 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3410 			consumed++;
3411 			reclaimed++;
3412 			continue;
3413 		}
3414 		in_use_prev = txq->ift_in_use;
3415 		pidx_prev = txq->ift_pidx;
3416 		err = iflib_encap(txq, mp);
3417 		if (__predict_false(err)) {
3418 			DBG_COUNTER_INC(txq_drain_encapfail);
3419 			/* no room - bail out */
3420 			if (err == ENOBUFS)
3421 				break;
3422 			consumed++;
3423 			DBG_COUNTER_INC(txq_drain_encapfail);
3424 			/* we can't send this packet - skip it */
3425 			continue;
3426 		}
3427 		consumed++;
3428 		pkt_sent++;
3429 		m = *mp;
3430 		DBG_COUNTER_INC(tx_sent);
3431 		bytes_sent += m->m_pkthdr.len;
3432 		mcast_sent += !!(m->m_flags & M_MCAST);
3433 		avail = TXQ_AVAIL(txq);
3434 
3435 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3436 		desc_used += (txq->ift_in_use - in_use_prev);
3437 		ETHER_BPF_MTAP(ifp, m);
3438 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3439 			break;
3440 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3441 	}
3442 
3443 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3444 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3445 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3446 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3447 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3448 	if (mcast_sent)
3449 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3450 #ifdef INVARIANTS
3451 	if (iflib_verbose_debug)
3452 		printf("consumed=%d\n", consumed);
3453 #endif
3454 	return (consumed);
3455 }
3456 
3457 static uint32_t
3458 iflib_txq_drain_always(struct ifmp_ring *r)
3459 {
3460 	return (1);
3461 }
3462 
3463 static uint32_t
3464 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3465 {
3466 	int i, avail;
3467 	struct mbuf **mp;
3468 	iflib_txq_t txq;
3469 
3470 	txq = r->cookie;
3471 
3472 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3473 	CALLOUT_LOCK(txq);
3474 	callout_stop(&txq->ift_timer);
3475 	CALLOUT_UNLOCK(txq);
3476 
3477 	avail = IDXDIFF(pidx, cidx, r->size);
3478 	for (i = 0; i < avail; i++) {
3479 		mp = _ring_peek_one(r, cidx, i, avail - i);
3480 		if (__predict_false(*mp == (struct mbuf *)txq))
3481 			continue;
3482 		m_freem(*mp);
3483 	}
3484 	MPASS(ifmp_ring_is_stalled(r) == 0);
3485 	return (avail);
3486 }
3487 
3488 static void
3489 iflib_ifmp_purge(iflib_txq_t txq)
3490 {
3491 	struct ifmp_ring *r;
3492 
3493 	r = txq->ift_br;
3494 	r->drain = iflib_txq_drain_free;
3495 	r->can_drain = iflib_txq_drain_always;
3496 
3497 	ifmp_ring_check_drainage(r, r->size);
3498 
3499 	r->drain = iflib_txq_drain;
3500 	r->can_drain = iflib_txq_can_drain;
3501 }
3502 
3503 static void
3504 _task_fn_tx(void *context)
3505 {
3506 	iflib_txq_t txq = context;
3507 	if_ctx_t ctx = txq->ift_ctx;
3508 	struct ifnet *ifp = ctx->ifc_ifp;
3509 	int rc;
3510 
3511 #ifdef IFLIB_DIAGNOSTICS
3512 	txq->ift_cpu_exec_count[curcpu]++;
3513 #endif
3514 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3515 		return;
3516 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3517 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3518 			netmap_tx_irq(ifp, txq->ift_id);
3519 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3520 		return;
3521 	}
3522 	if (txq->ift_db_pending)
3523 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3524 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3525 	if (ctx->ifc_flags & IFC_LEGACY)
3526 		IFDI_INTR_ENABLE(ctx);
3527 	else {
3528 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3529 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3530 	}
3531 }
3532 
3533 static void
3534 _task_fn_rx(void *context)
3535 {
3536 	iflib_rxq_t rxq = context;
3537 	if_ctx_t ctx = rxq->ifr_ctx;
3538 	bool more;
3539 	int rc;
3540 	uint16_t budget;
3541 
3542 #ifdef IFLIB_DIAGNOSTICS
3543 	rxq->ifr_cpu_exec_count[curcpu]++;
3544 #endif
3545 	DBG_COUNTER_INC(task_fn_rxs);
3546 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3547 		return;
3548 	more = true;
3549 #ifdef DEV_NETMAP
3550 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3551 		u_int work = 0;
3552 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3553 			more = false;
3554 		}
3555 	}
3556 #endif
3557 	budget = ctx->ifc_sysctl_rx_budget;
3558 	if (budget == 0)
3559 		budget = 16;	/* XXX */
3560 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3561 		if (ctx->ifc_flags & IFC_LEGACY)
3562 			IFDI_INTR_ENABLE(ctx);
3563 		else {
3564 			DBG_COUNTER_INC(rx_intr_enables);
3565 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3566 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3567 		}
3568 	}
3569 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3570 		return;
3571 	if (more)
3572 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3573 }
3574 
3575 static void
3576 _task_fn_admin(void *context)
3577 {
3578 	if_ctx_t ctx = context;
3579 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3580 	iflib_txq_t txq;
3581 	int i;
3582 
3583 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3584 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3585 			return;
3586 		}
3587 	}
3588 
3589 	CTX_LOCK(ctx);
3590 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3591 		CALLOUT_LOCK(txq);
3592 		callout_stop(&txq->ift_timer);
3593 		CALLOUT_UNLOCK(txq);
3594 	}
3595 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3596 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3597 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3598 	IFDI_LINK_INTR_ENABLE(ctx);
3599 	if (ctx->ifc_flags & IFC_DO_RESET) {
3600 		ctx->ifc_flags &= ~IFC_DO_RESET;
3601 		iflib_if_init_locked(ctx);
3602 	}
3603 	CTX_UNLOCK(ctx);
3604 
3605 	if (LINK_ACTIVE(ctx) == 0)
3606 		return;
3607 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3608 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3609 }
3610 
3611 
3612 static void
3613 _task_fn_iov(void *context)
3614 {
3615 	if_ctx_t ctx = context;
3616 
3617 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3618 		return;
3619 
3620 	CTX_LOCK(ctx);
3621 	IFDI_VFLR_HANDLE(ctx);
3622 	CTX_UNLOCK(ctx);
3623 }
3624 
3625 static int
3626 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3627 {
3628 	int err;
3629 	if_int_delay_info_t info;
3630 	if_ctx_t ctx;
3631 
3632 	info = (if_int_delay_info_t)arg1;
3633 	ctx = info->iidi_ctx;
3634 	info->iidi_req = req;
3635 	info->iidi_oidp = oidp;
3636 	CTX_LOCK(ctx);
3637 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3638 	CTX_UNLOCK(ctx);
3639 	return (err);
3640 }
3641 
3642 /*********************************************************************
3643  *
3644  *  IFNET FUNCTIONS
3645  *
3646  **********************************************************************/
3647 
3648 static void
3649 iflib_if_init_locked(if_ctx_t ctx)
3650 {
3651 	iflib_stop(ctx);
3652 	iflib_init_locked(ctx);
3653 }
3654 
3655 
3656 static void
3657 iflib_if_init(void *arg)
3658 {
3659 	if_ctx_t ctx = arg;
3660 
3661 	CTX_LOCK(ctx);
3662 	iflib_if_init_locked(ctx);
3663 	CTX_UNLOCK(ctx);
3664 }
3665 
3666 static int
3667 iflib_if_transmit(if_t ifp, struct mbuf *m)
3668 {
3669 	if_ctx_t	ctx = if_getsoftc(ifp);
3670 
3671 	iflib_txq_t txq;
3672 	int err, qidx;
3673 
3674 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3675 		DBG_COUNTER_INC(tx_frees);
3676 		m_freem(m);
3677 		return (ENOBUFS);
3678 	}
3679 
3680 	MPASS(m->m_nextpkt == NULL);
3681 	qidx = 0;
3682 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3683 		qidx = QIDX(ctx, m);
3684 	/*
3685 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3686 	 */
3687 	txq = &ctx->ifc_txqs[qidx];
3688 
3689 #ifdef DRIVER_BACKPRESSURE
3690 	if (txq->ift_closed) {
3691 		while (m != NULL) {
3692 			next = m->m_nextpkt;
3693 			m->m_nextpkt = NULL;
3694 			m_freem(m);
3695 			m = next;
3696 		}
3697 		return (ENOBUFS);
3698 	}
3699 #endif
3700 #ifdef notyet
3701 	qidx = count = 0;
3702 	mp = marr;
3703 	next = m;
3704 	do {
3705 		count++;
3706 		next = next->m_nextpkt;
3707 	} while (next != NULL);
3708 
3709 	if (count > nitems(marr))
3710 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3711 			/* XXX check nextpkt */
3712 			m_freem(m);
3713 			/* XXX simplify for now */
3714 			DBG_COUNTER_INC(tx_frees);
3715 			return (ENOBUFS);
3716 		}
3717 	for (next = m, i = 0; next != NULL; i++) {
3718 		mp[i] = next;
3719 		next = next->m_nextpkt;
3720 		mp[i]->m_nextpkt = NULL;
3721 	}
3722 #endif
3723 	DBG_COUNTER_INC(tx_seen);
3724 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3725 
3726 	GROUPTASK_ENQUEUE(&txq->ift_task);
3727 	if (err) {
3728 		/* support forthcoming later */
3729 #ifdef DRIVER_BACKPRESSURE
3730 		txq->ift_closed = TRUE;
3731 #endif
3732 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3733 		m_freem(m);
3734 	}
3735 
3736 	return (err);
3737 }
3738 
3739 static void
3740 iflib_if_qflush(if_t ifp)
3741 {
3742 	if_ctx_t ctx = if_getsoftc(ifp);
3743 	iflib_txq_t txq = ctx->ifc_txqs;
3744 	int i;
3745 
3746 	CTX_LOCK(ctx);
3747 	ctx->ifc_flags |= IFC_QFLUSH;
3748 	CTX_UNLOCK(ctx);
3749 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3750 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3751 			iflib_txq_check_drain(txq, 0);
3752 	CTX_LOCK(ctx);
3753 	ctx->ifc_flags &= ~IFC_QFLUSH;
3754 	CTX_UNLOCK(ctx);
3755 
3756 	if_qflush(ifp);
3757 }
3758 
3759 
3760 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3761 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3762 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3763 
3764 static int
3765 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3766 {
3767 	if_ctx_t ctx = if_getsoftc(ifp);
3768 	struct ifreq	*ifr = (struct ifreq *)data;
3769 #if defined(INET) || defined(INET6)
3770 	struct ifaddr	*ifa = (struct ifaddr *)data;
3771 #endif
3772 	bool		avoid_reset = FALSE;
3773 	int		err = 0, reinit = 0, bits;
3774 
3775 	switch (command) {
3776 	case SIOCSIFADDR:
3777 #ifdef INET
3778 		if (ifa->ifa_addr->sa_family == AF_INET)
3779 			avoid_reset = TRUE;
3780 #endif
3781 #ifdef INET6
3782 		if (ifa->ifa_addr->sa_family == AF_INET6)
3783 			avoid_reset = TRUE;
3784 #endif
3785 		/*
3786 		** Calling init results in link renegotiation,
3787 		** so we avoid doing it when possible.
3788 		*/
3789 		if (avoid_reset) {
3790 			if_setflagbits(ifp, IFF_UP,0);
3791 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3792 				reinit = 1;
3793 #ifdef INET
3794 			if (!(if_getflags(ifp) & IFF_NOARP))
3795 				arp_ifinit(ifp, ifa);
3796 #endif
3797 		} else
3798 			err = ether_ioctl(ifp, command, data);
3799 		break;
3800 	case SIOCSIFMTU:
3801 		CTX_LOCK(ctx);
3802 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3803 			CTX_UNLOCK(ctx);
3804 			break;
3805 		}
3806 		bits = if_getdrvflags(ifp);
3807 		/* stop the driver and free any clusters before proceeding */
3808 		iflib_stop(ctx);
3809 
3810 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3811 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3812 				ctx->ifc_flags |= IFC_MULTISEG;
3813 			else
3814 				ctx->ifc_flags &= ~IFC_MULTISEG;
3815 			err = if_setmtu(ifp, ifr->ifr_mtu);
3816 		}
3817 		iflib_init_locked(ctx);
3818 		if_setdrvflags(ifp, bits);
3819 		CTX_UNLOCK(ctx);
3820 		break;
3821 	case SIOCSIFFLAGS:
3822 		CTX_LOCK(ctx);
3823 		if (if_getflags(ifp) & IFF_UP) {
3824 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3825 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3826 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3827 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3828 				}
3829 			} else
3830 				reinit = 1;
3831 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3832 			iflib_stop(ctx);
3833 		}
3834 		ctx->ifc_if_flags = if_getflags(ifp);
3835 		CTX_UNLOCK(ctx);
3836 		break;
3837 	case SIOCADDMULTI:
3838 	case SIOCDELMULTI:
3839 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3840 			CTX_LOCK(ctx);
3841 			IFDI_INTR_DISABLE(ctx);
3842 			IFDI_MULTI_SET(ctx);
3843 			IFDI_INTR_ENABLE(ctx);
3844 			CTX_UNLOCK(ctx);
3845 		}
3846 		break;
3847 	case SIOCSIFMEDIA:
3848 		CTX_LOCK(ctx);
3849 		IFDI_MEDIA_SET(ctx);
3850 		CTX_UNLOCK(ctx);
3851 		/* falls thru */
3852 	case SIOCGIFMEDIA:
3853 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3854 		break;
3855 	case SIOCGI2C:
3856 	{
3857 		struct ifi2creq i2c;
3858 
3859 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3860 		if (err != 0)
3861 			break;
3862 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3863 			err = EINVAL;
3864 			break;
3865 		}
3866 		if (i2c.len > sizeof(i2c.data)) {
3867 			err = EINVAL;
3868 			break;
3869 		}
3870 
3871 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3872 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3873 		break;
3874 	}
3875 	case SIOCSIFCAP:
3876 	{
3877 		int mask, setmask;
3878 
3879 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3880 		setmask = 0;
3881 #ifdef TCP_OFFLOAD
3882 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3883 #endif
3884 		setmask |= (mask & IFCAP_FLAGS);
3885 
3886 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3887 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3888 		if ((mask & IFCAP_WOL) &&
3889 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3890 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3891 		if_vlancap(ifp);
3892 		/*
3893 		 * want to ensure that traffic has stopped before we change any of the flags
3894 		 */
3895 		if (setmask) {
3896 			CTX_LOCK(ctx);
3897 			bits = if_getdrvflags(ifp);
3898 			if (bits & IFF_DRV_RUNNING)
3899 				iflib_stop(ctx);
3900 			if_togglecapenable(ifp, setmask);
3901 			if (bits & IFF_DRV_RUNNING)
3902 				iflib_init_locked(ctx);
3903 			if_setdrvflags(ifp, bits);
3904 			CTX_UNLOCK(ctx);
3905 		}
3906 		break;
3907 	    }
3908 	case SIOCGPRIVATE_0:
3909 	case SIOCSDRVSPEC:
3910 	case SIOCGDRVSPEC:
3911 		CTX_LOCK(ctx);
3912 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3913 		CTX_UNLOCK(ctx);
3914 		break;
3915 	default:
3916 		err = ether_ioctl(ifp, command, data);
3917 		break;
3918 	}
3919 	if (reinit)
3920 		iflib_if_init(ctx);
3921 	return (err);
3922 }
3923 
3924 static uint64_t
3925 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3926 {
3927 	if_ctx_t ctx = if_getsoftc(ifp);
3928 
3929 	return (IFDI_GET_COUNTER(ctx, cnt));
3930 }
3931 
3932 /*********************************************************************
3933  *
3934  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3935  *
3936  **********************************************************************/
3937 
3938 static void
3939 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3940 {
3941 	if_ctx_t ctx = if_getsoftc(ifp);
3942 
3943 	if ((void *)ctx != arg)
3944 		return;
3945 
3946 	if ((vtag == 0) || (vtag > 4095))
3947 		return;
3948 
3949 	CTX_LOCK(ctx);
3950 	IFDI_VLAN_REGISTER(ctx, vtag);
3951 	/* Re-init to load the changes */
3952 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3953 		iflib_if_init_locked(ctx);
3954 	CTX_UNLOCK(ctx);
3955 }
3956 
3957 static void
3958 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3959 {
3960 	if_ctx_t ctx = if_getsoftc(ifp);
3961 
3962 	if ((void *)ctx != arg)
3963 		return;
3964 
3965 	if ((vtag == 0) || (vtag > 4095))
3966 		return;
3967 
3968 	CTX_LOCK(ctx);
3969 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3970 	/* Re-init to load the changes */
3971 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3972 		iflib_if_init_locked(ctx);
3973 	CTX_UNLOCK(ctx);
3974 }
3975 
3976 static void
3977 iflib_led_func(void *arg, int onoff)
3978 {
3979 	if_ctx_t ctx = arg;
3980 
3981 	CTX_LOCK(ctx);
3982 	IFDI_LED_FUNC(ctx, onoff);
3983 	CTX_UNLOCK(ctx);
3984 }
3985 
3986 /*********************************************************************
3987  *
3988  *  BUS FUNCTION DEFINITIONS
3989  *
3990  **********************************************************************/
3991 
3992 int
3993 iflib_device_probe(device_t dev)
3994 {
3995 	pci_vendor_info_t *ent;
3996 
3997 	uint16_t	pci_vendor_id, pci_device_id;
3998 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3999 	uint16_t	pci_rev_id;
4000 	if_shared_ctx_t sctx;
4001 
4002 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4003 		return (ENOTSUP);
4004 
4005 	pci_vendor_id = pci_get_vendor(dev);
4006 	pci_device_id = pci_get_device(dev);
4007 	pci_subvendor_id = pci_get_subvendor(dev);
4008 	pci_subdevice_id = pci_get_subdevice(dev);
4009 	pci_rev_id = pci_get_revid(dev);
4010 	if (sctx->isc_parse_devinfo != NULL)
4011 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4012 
4013 	ent = sctx->isc_vendor_info;
4014 	while (ent->pvi_vendor_id != 0) {
4015 		if (pci_vendor_id != ent->pvi_vendor_id) {
4016 			ent++;
4017 			continue;
4018 		}
4019 		if ((pci_device_id == ent->pvi_device_id) &&
4020 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4021 		     (ent->pvi_subvendor_id == 0)) &&
4022 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4023 		     (ent->pvi_subdevice_id == 0)) &&
4024 		    ((pci_rev_id == ent->pvi_rev_id) ||
4025 		     (ent->pvi_rev_id == 0))) {
4026 
4027 			device_set_desc_copy(dev, ent->pvi_name);
4028 			/* this needs to be changed to zero if the bus probing code
4029 			 * ever stops re-probing on best match because the sctx
4030 			 * may have its values over written by register calls
4031 			 * in subsequent probes
4032 			 */
4033 			return (BUS_PROBE_DEFAULT);
4034 		}
4035 		ent++;
4036 	}
4037 	return (ENXIO);
4038 }
4039 
4040 int
4041 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4042 {
4043 	int err, rid, msix, msix_bar;
4044 	if_ctx_t ctx;
4045 	if_t ifp;
4046 	if_softc_ctx_t scctx;
4047 	int i;
4048 	uint16_t main_txq;
4049 	uint16_t main_rxq;
4050 
4051 
4052 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4053 
4054 	if (sc == NULL) {
4055 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4056 		device_set_softc(dev, ctx);
4057 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4058 	}
4059 
4060 	ctx->ifc_sctx = sctx;
4061 	ctx->ifc_dev = dev;
4062 	ctx->ifc_softc = sc;
4063 
4064 	if ((err = iflib_register(ctx)) != 0) {
4065 		device_printf(dev, "iflib_register failed %d\n", err);
4066 		return (err);
4067 	}
4068 	iflib_add_device_sysctl_pre(ctx);
4069 
4070 	scctx = &ctx->ifc_softc_ctx;
4071 	ifp = ctx->ifc_ifp;
4072 
4073 	/*
4074 	 * XXX sanity check that ntxd & nrxd are a power of 2
4075 	 */
4076 	if (ctx->ifc_sysctl_ntxqs != 0)
4077 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4078 	if (ctx->ifc_sysctl_nrxqs != 0)
4079 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4080 
4081 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4082 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4083 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4084 		else
4085 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4086 	}
4087 
4088 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4089 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4090 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4091 		else
4092 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4093 	}
4094 
4095 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4096 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4097 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4098 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4099 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4100 		}
4101 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4102 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4103 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4104 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4105 		}
4106 	}
4107 
4108 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4109 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4110 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4111 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4112 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4113 		}
4114 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4115 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4116 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4117 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4118 		}
4119 	}
4120 
4121 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4122 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4123 		return (err);
4124 	}
4125 	_iflib_pre_assert(scctx);
4126 	ctx->ifc_txrx = *scctx->isc_txrx;
4127 
4128 #ifdef INVARIANTS
4129 	MPASS(scctx->isc_capenable);
4130 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4131 		MPASS(scctx->isc_tx_csum_flags);
4132 #endif
4133 
4134 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4135 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4136 
4137 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4138 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4139 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4140 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4141 
4142 #ifdef ACPI_DMAR
4143 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4144 		ctx->ifc_flags |= IFC_DMAR;
4145 #elif !(defined(__i386__) || defined(__amd64__))
4146 	/* set unconditionally for !x86 */
4147 	ctx->ifc_flags |= IFC_DMAR;
4148 #endif
4149 
4150 	msix_bar = scctx->isc_msix_bar;
4151 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4152 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4153 
4154 	/* XXX change for per-queue sizes */
4155 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4156 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4157 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4158 		if (!powerof2(scctx->isc_nrxd[i])) {
4159 			/* round down instead? */
4160 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4161 			err = EINVAL;
4162 			goto fail;
4163 		}
4164 	}
4165 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4166 		if (!powerof2(scctx->isc_ntxd[i])) {
4167 			device_printf(dev,
4168 			    "# tx descriptors must be a power of 2");
4169 			err = EINVAL;
4170 			goto fail;
4171 		}
4172 	}
4173 
4174 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4175 	    MAX_SINGLE_PACKET_FRACTION)
4176 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4177 		    MAX_SINGLE_PACKET_FRACTION);
4178 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4179 	    MAX_SINGLE_PACKET_FRACTION)
4180 		scctx->isc_tx_tso_segments_max = max(1,
4181 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4182 
4183 	/*
4184 	 * Protect the stack against modern hardware
4185 	 */
4186 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4187 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4188 
4189 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4190 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4191 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4192 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4193 	if (scctx->isc_rss_table_size == 0)
4194 		scctx->isc_rss_table_size = 64;
4195 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4196 
4197 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4198 	/* XXX format name */
4199 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4200 	/*
4201 	** Now setup MSI or MSI/X, should
4202 	** return us the number of supported
4203 	** vectors. (Will be 1 for MSI)
4204 	*/
4205 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4206 		msix = scctx->isc_vectors;
4207 	} else if (scctx->isc_msix_bar != 0)
4208 	       /*
4209 		* The simple fact that isc_msix_bar is not 0 does not mean we
4210 		* we have a good value there that is known to work.
4211 		*/
4212 		msix = iflib_msix_init(ctx);
4213 	else {
4214 		scctx->isc_vectors = 1;
4215 		scctx->isc_ntxqsets = 1;
4216 		scctx->isc_nrxqsets = 1;
4217 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4218 		msix = 0;
4219 	}
4220 	/* Get memory for the station queues */
4221 	if ((err = iflib_queues_alloc(ctx))) {
4222 		device_printf(dev, "Unable to allocate queue memory\n");
4223 		goto fail;
4224 	}
4225 
4226 	if ((err = iflib_qset_structures_setup(ctx))) {
4227 		device_printf(dev, "qset structure setup failed %d\n", err);
4228 		goto fail_queues;
4229 	}
4230 
4231 	/*
4232 	 * Group taskqueues aren't properly set up until SMP is started,
4233 	 * so we disable interrupts until we can handle them post
4234 	 * SI_SUB_SMP.
4235 	 *
4236 	 * XXX: disabling interrupts doesn't actually work, at least for
4237 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4238 	 * we do null handling and depend on this not causing too large an
4239 	 * interrupt storm.
4240 	 */
4241 	IFDI_INTR_DISABLE(ctx);
4242 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4243 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4244 		goto fail_intr_free;
4245 	}
4246 	if (msix <= 1) {
4247 		rid = 0;
4248 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4249 			MPASS(msix == 1);
4250 			rid = 1;
4251 		}
4252 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4253 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4254 			goto fail_intr_free;
4255 		}
4256 	}
4257 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4258 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4259 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4260 		goto fail_detach;
4261 	}
4262 	if ((err = iflib_netmap_attach(ctx))) {
4263 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4264 		goto fail_detach;
4265 	}
4266 	*ctxp = ctx;
4267 
4268 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4269 	iflib_add_device_sysctl_post(ctx);
4270 	ctx->ifc_flags |= IFC_INIT_DONE;
4271 	return (0);
4272 fail_detach:
4273 	ether_ifdetach(ctx->ifc_ifp);
4274 fail_intr_free:
4275 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4276 		pci_release_msi(ctx->ifc_dev);
4277 fail_queues:
4278 	/* XXX free queues */
4279 fail:
4280 	IFDI_DETACH(ctx);
4281 	return (err);
4282 }
4283 
4284 int
4285 iflib_device_attach(device_t dev)
4286 {
4287 	if_ctx_t ctx;
4288 	if_shared_ctx_t sctx;
4289 
4290 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4291 		return (ENOTSUP);
4292 
4293 	pci_enable_busmaster(dev);
4294 
4295 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4296 }
4297 
4298 int
4299 iflib_device_deregister(if_ctx_t ctx)
4300 {
4301 	if_t ifp = ctx->ifc_ifp;
4302 	iflib_txq_t txq;
4303 	iflib_rxq_t rxq;
4304 	device_t dev = ctx->ifc_dev;
4305 	int i, j;
4306 	struct taskqgroup *tqg;
4307 	iflib_fl_t fl;
4308 
4309 	/* Make sure VLANS are not using driver */
4310 	if (if_vlantrunkinuse(ifp)) {
4311 		device_printf(dev,"Vlan in use, detach first\n");
4312 		return (EBUSY);
4313 	}
4314 
4315 	CTX_LOCK(ctx);
4316 	ctx->ifc_in_detach = 1;
4317 	iflib_stop(ctx);
4318 	CTX_UNLOCK(ctx);
4319 
4320 	/* Unregister VLAN events */
4321 	if (ctx->ifc_vlan_attach_event != NULL)
4322 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4323 	if (ctx->ifc_vlan_detach_event != NULL)
4324 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4325 
4326 	iflib_netmap_detach(ifp);
4327 	ether_ifdetach(ifp);
4328 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4329 	CTX_LOCK_DESTROY(ctx);
4330 	if (ctx->ifc_led_dev != NULL)
4331 		led_destroy(ctx->ifc_led_dev);
4332 	/* XXX drain any dependent tasks */
4333 	tqg = qgroup_if_io_tqg;
4334 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4335 		callout_drain(&txq->ift_timer);
4336 		if (txq->ift_task.gt_uniq != NULL)
4337 			taskqgroup_detach(tqg, &txq->ift_task);
4338 	}
4339 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4340 		if (rxq->ifr_task.gt_uniq != NULL)
4341 			taskqgroup_detach(tqg, &rxq->ifr_task);
4342 
4343 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4344 			free(fl->ifl_rx_bitmap, M_IFLIB);
4345 
4346 	}
4347 	tqg = qgroup_if_config_tqg;
4348 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4349 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4350 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4351 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4352 
4353 	IFDI_DETACH(ctx);
4354 	device_set_softc(ctx->ifc_dev, NULL);
4355 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4356 		pci_release_msi(dev);
4357 	}
4358 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4359 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4360 	}
4361 	if (ctx->ifc_msix_mem != NULL) {
4362 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4363 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4364 		ctx->ifc_msix_mem = NULL;
4365 	}
4366 
4367 	bus_generic_detach(dev);
4368 	if_free(ifp);
4369 
4370 	iflib_tx_structures_free(ctx);
4371 	iflib_rx_structures_free(ctx);
4372 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4373 		free(ctx->ifc_softc, M_IFLIB);
4374 	free(ctx, M_IFLIB);
4375 	return (0);
4376 }
4377 
4378 
4379 int
4380 iflib_device_detach(device_t dev)
4381 {
4382 	if_ctx_t ctx = device_get_softc(dev);
4383 
4384 	return (iflib_device_deregister(ctx));
4385 }
4386 
4387 int
4388 iflib_device_suspend(device_t dev)
4389 {
4390 	if_ctx_t ctx = device_get_softc(dev);
4391 
4392 	CTX_LOCK(ctx);
4393 	IFDI_SUSPEND(ctx);
4394 	CTX_UNLOCK(ctx);
4395 
4396 	return bus_generic_suspend(dev);
4397 }
4398 int
4399 iflib_device_shutdown(device_t dev)
4400 {
4401 	if_ctx_t ctx = device_get_softc(dev);
4402 
4403 	CTX_LOCK(ctx);
4404 	IFDI_SHUTDOWN(ctx);
4405 	CTX_UNLOCK(ctx);
4406 
4407 	return bus_generic_suspend(dev);
4408 }
4409 
4410 
4411 int
4412 iflib_device_resume(device_t dev)
4413 {
4414 	if_ctx_t ctx = device_get_softc(dev);
4415 	iflib_txq_t txq = ctx->ifc_txqs;
4416 
4417 	CTX_LOCK(ctx);
4418 	IFDI_RESUME(ctx);
4419 	iflib_init_locked(ctx);
4420 	CTX_UNLOCK(ctx);
4421 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4422 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4423 
4424 	return (bus_generic_resume(dev));
4425 }
4426 
4427 int
4428 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4429 {
4430 	int error;
4431 	if_ctx_t ctx = device_get_softc(dev);
4432 
4433 	CTX_LOCK(ctx);
4434 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4435 	CTX_UNLOCK(ctx);
4436 
4437 	return (error);
4438 }
4439 
4440 void
4441 iflib_device_iov_uninit(device_t dev)
4442 {
4443 	if_ctx_t ctx = device_get_softc(dev);
4444 
4445 	CTX_LOCK(ctx);
4446 	IFDI_IOV_UNINIT(ctx);
4447 	CTX_UNLOCK(ctx);
4448 }
4449 
4450 int
4451 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4452 {
4453 	int error;
4454 	if_ctx_t ctx = device_get_softc(dev);
4455 
4456 	CTX_LOCK(ctx);
4457 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4458 	CTX_UNLOCK(ctx);
4459 
4460 	return (error);
4461 }
4462 
4463 /*********************************************************************
4464  *
4465  *  MODULE FUNCTION DEFINITIONS
4466  *
4467  **********************************************************************/
4468 
4469 /*
4470  * - Start a fast taskqueue thread for each core
4471  * - Start a taskqueue for control operations
4472  */
4473 static int
4474 iflib_module_init(void)
4475 {
4476 	return (0);
4477 }
4478 
4479 static int
4480 iflib_module_event_handler(module_t mod, int what, void *arg)
4481 {
4482 	int err;
4483 
4484 	switch (what) {
4485 	case MOD_LOAD:
4486 		if ((err = iflib_module_init()) != 0)
4487 			return (err);
4488 		break;
4489 	case MOD_UNLOAD:
4490 		return (EBUSY);
4491 	default:
4492 		return (EOPNOTSUPP);
4493 	}
4494 
4495 	return (0);
4496 }
4497 
4498 /*********************************************************************
4499  *
4500  *  PUBLIC FUNCTION DEFINITIONS
4501  *     ordered as in iflib.h
4502  *
4503  **********************************************************************/
4504 
4505 
4506 static void
4507 _iflib_assert(if_shared_ctx_t sctx)
4508 {
4509 	MPASS(sctx->isc_tx_maxsize);
4510 	MPASS(sctx->isc_tx_maxsegsize);
4511 
4512 	MPASS(sctx->isc_rx_maxsize);
4513 	MPASS(sctx->isc_rx_nsegments);
4514 	MPASS(sctx->isc_rx_maxsegsize);
4515 
4516 	MPASS(sctx->isc_nrxd_min[0]);
4517 	MPASS(sctx->isc_nrxd_max[0]);
4518 	MPASS(sctx->isc_nrxd_default[0]);
4519 	MPASS(sctx->isc_ntxd_min[0]);
4520 	MPASS(sctx->isc_ntxd_max[0]);
4521 	MPASS(sctx->isc_ntxd_default[0]);
4522 }
4523 
4524 static void
4525 _iflib_pre_assert(if_softc_ctx_t scctx)
4526 {
4527 
4528 	MPASS(scctx->isc_txrx->ift_txd_encap);
4529 	MPASS(scctx->isc_txrx->ift_txd_flush);
4530 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4531 	MPASS(scctx->isc_txrx->ift_rxd_available);
4532 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4533 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4534 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4535 }
4536 
4537 static int
4538 iflib_register(if_ctx_t ctx)
4539 {
4540 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4541 	driver_t *driver = sctx->isc_driver;
4542 	device_t dev = ctx->ifc_dev;
4543 	if_t ifp;
4544 
4545 	_iflib_assert(sctx);
4546 
4547 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4548 
4549 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4550 	if (ifp == NULL) {
4551 		device_printf(dev, "can not allocate ifnet structure\n");
4552 		return (ENOMEM);
4553 	}
4554 
4555 	/*
4556 	 * Initialize our context's device specific methods
4557 	 */
4558 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4559 	kobj_class_compile((kobj_class_t) driver);
4560 	driver->refs++;
4561 
4562 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4563 	if_setsoftc(ifp, ctx);
4564 	if_setdev(ifp, dev);
4565 	if_setinitfn(ifp, iflib_if_init);
4566 	if_setioctlfn(ifp, iflib_if_ioctl);
4567 	if_settransmitfn(ifp, iflib_if_transmit);
4568 	if_setqflushfn(ifp, iflib_if_qflush);
4569 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4570 
4571 	ctx->ifc_vlan_attach_event =
4572 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4573 							  EVENTHANDLER_PRI_FIRST);
4574 	ctx->ifc_vlan_detach_event =
4575 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4576 							  EVENTHANDLER_PRI_FIRST);
4577 
4578 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4579 					 iflib_media_change, iflib_media_status);
4580 
4581 	return (0);
4582 }
4583 
4584 
4585 static int
4586 iflib_queues_alloc(if_ctx_t ctx)
4587 {
4588 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4589 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4590 	device_t dev = ctx->ifc_dev;
4591 	int nrxqsets = scctx->isc_nrxqsets;
4592 	int ntxqsets = scctx->isc_ntxqsets;
4593 	iflib_txq_t txq;
4594 	iflib_rxq_t rxq;
4595 	iflib_fl_t fl = NULL;
4596 	int i, j, cpu, err, txconf, rxconf;
4597 	iflib_dma_info_t ifdip;
4598 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4599 	uint32_t *txqsizes = scctx->isc_txqsizes;
4600 	uint8_t nrxqs = sctx->isc_nrxqs;
4601 	uint8_t ntxqs = sctx->isc_ntxqs;
4602 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4603 	caddr_t *vaddrs;
4604 	uint64_t *paddrs;
4605 	struct ifmp_ring **brscp;
4606 
4607 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4608 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4609 
4610 	brscp = NULL;
4611 	txq = NULL;
4612 	rxq = NULL;
4613 
4614 /* Allocate the TX ring struct memory */
4615 	if (!(txq =
4616 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4617 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4618 		device_printf(dev, "Unable to allocate TX ring memory\n");
4619 		err = ENOMEM;
4620 		goto fail;
4621 	}
4622 
4623 	/* Now allocate the RX */
4624 	if (!(rxq =
4625 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4626 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4627 		device_printf(dev, "Unable to allocate RX ring memory\n");
4628 		err = ENOMEM;
4629 		goto rx_fail;
4630 	}
4631 
4632 	ctx->ifc_txqs = txq;
4633 	ctx->ifc_rxqs = rxq;
4634 
4635 	/*
4636 	 * XXX handle allocation failure
4637 	 */
4638 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4639 		/* Set up some basics */
4640 
4641 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4642 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4643 			err = ENOMEM;
4644 			goto err_tx_desc;
4645 		}
4646 		txq->ift_ifdi = ifdip;
4647 		for (j = 0; j < ntxqs; j++, ifdip++) {
4648 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4649 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4650 				err = ENOMEM;
4651 				goto err_tx_desc;
4652 			}
4653 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4654 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4655 		}
4656 		txq->ift_ctx = ctx;
4657 		txq->ift_id = i;
4658 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4659 			txq->ift_br_offset = 1;
4660 		} else {
4661 			txq->ift_br_offset = 0;
4662 		}
4663 		/* XXX fix this */
4664 		txq->ift_timer.c_cpu = cpu;
4665 
4666 		if (iflib_txsd_alloc(txq)) {
4667 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4668 			err = ENOMEM;
4669 			goto err_tx_desc;
4670 		}
4671 
4672 		/* Initialize the TX lock */
4673 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4674 		    device_get_nameunit(dev), txq->ift_id);
4675 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4676 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4677 
4678 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4679 			 device_get_nameunit(dev), txq->ift_id);
4680 
4681 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4682 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4683 		if (err) {
4684 			/* XXX free any allocated rings */
4685 			device_printf(dev, "Unable to allocate buf_ring\n");
4686 			goto err_tx_desc;
4687 		}
4688 	}
4689 
4690 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4691 		/* Set up some basics */
4692 
4693 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4694 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4695 			err = ENOMEM;
4696 			goto err_tx_desc;
4697 		}
4698 
4699 		rxq->ifr_ifdi = ifdip;
4700 		/* XXX this needs to be changed if #rx queues != #tx queues */
4701 		rxq->ifr_ntxqirq = 1;
4702 		rxq->ifr_txqid[0] = i;
4703 		for (j = 0; j < nrxqs; j++, ifdip++) {
4704 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4705 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4706 				err = ENOMEM;
4707 				goto err_tx_desc;
4708 			}
4709 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4710 		}
4711 		rxq->ifr_ctx = ctx;
4712 		rxq->ifr_id = i;
4713 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4714 			rxq->ifr_fl_offset = 1;
4715 		} else {
4716 			rxq->ifr_fl_offset = 0;
4717 		}
4718 		rxq->ifr_nfl = nfree_lists;
4719 		if (!(fl =
4720 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4721 			device_printf(dev, "Unable to allocate free list memory\n");
4722 			err = ENOMEM;
4723 			goto err_tx_desc;
4724 		}
4725 		rxq->ifr_fl = fl;
4726 		for (j = 0; j < nfree_lists; j++) {
4727 			fl[j].ifl_rxq = rxq;
4728 			fl[j].ifl_id = j;
4729 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4730 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4731 		}
4732         /* Allocate receive buffers for the ring*/
4733 		if (iflib_rxsd_alloc(rxq)) {
4734 			device_printf(dev,
4735 			    "Critical Failure setting up receive buffers\n");
4736 			err = ENOMEM;
4737 			goto err_rx_desc;
4738 		}
4739 
4740 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4741 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4742 	}
4743 
4744 	/* TXQs */
4745 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4746 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4747 	for (i = 0; i < ntxqsets; i++) {
4748 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4749 
4750 		for (j = 0; j < ntxqs; j++, di++) {
4751 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4752 			paddrs[i*ntxqs + j] = di->idi_paddr;
4753 		}
4754 	}
4755 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4756 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4757 		iflib_tx_structures_free(ctx);
4758 		free(vaddrs, M_IFLIB);
4759 		free(paddrs, M_IFLIB);
4760 		goto err_rx_desc;
4761 	}
4762 	free(vaddrs, M_IFLIB);
4763 	free(paddrs, M_IFLIB);
4764 
4765 	/* RXQs */
4766 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4767 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4768 	for (i = 0; i < nrxqsets; i++) {
4769 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4770 
4771 		for (j = 0; j < nrxqs; j++, di++) {
4772 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4773 			paddrs[i*nrxqs + j] = di->idi_paddr;
4774 		}
4775 	}
4776 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4777 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4778 		iflib_tx_structures_free(ctx);
4779 		free(vaddrs, M_IFLIB);
4780 		free(paddrs, M_IFLIB);
4781 		goto err_rx_desc;
4782 	}
4783 	free(vaddrs, M_IFLIB);
4784 	free(paddrs, M_IFLIB);
4785 
4786 	return (0);
4787 
4788 /* XXX handle allocation failure changes */
4789 err_rx_desc:
4790 err_tx_desc:
4791 	if (ctx->ifc_rxqs != NULL)
4792 		free(ctx->ifc_rxqs, M_IFLIB);
4793 	ctx->ifc_rxqs = NULL;
4794 	if (ctx->ifc_txqs != NULL)
4795 		free(ctx->ifc_txqs, M_IFLIB);
4796 	ctx->ifc_txqs = NULL;
4797 rx_fail:
4798 	if (brscp != NULL)
4799 		free(brscp, M_IFLIB);
4800 	if (rxq != NULL)
4801 		free(rxq, M_IFLIB);
4802 	if (txq != NULL)
4803 		free(txq, M_IFLIB);
4804 fail:
4805 	return (err);
4806 }
4807 
4808 static int
4809 iflib_tx_structures_setup(if_ctx_t ctx)
4810 {
4811 	iflib_txq_t txq = ctx->ifc_txqs;
4812 	int i;
4813 
4814 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4815 		iflib_txq_setup(txq);
4816 
4817 	return (0);
4818 }
4819 
4820 static void
4821 iflib_tx_structures_free(if_ctx_t ctx)
4822 {
4823 	iflib_txq_t txq = ctx->ifc_txqs;
4824 	int i, j;
4825 
4826 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4827 		iflib_txq_destroy(txq);
4828 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4829 			iflib_dma_free(&txq->ift_ifdi[j]);
4830 	}
4831 	free(ctx->ifc_txqs, M_IFLIB);
4832 	ctx->ifc_txqs = NULL;
4833 	IFDI_QUEUES_FREE(ctx);
4834 }
4835 
4836 /*********************************************************************
4837  *
4838  *  Initialize all receive rings.
4839  *
4840  **********************************************************************/
4841 static int
4842 iflib_rx_structures_setup(if_ctx_t ctx)
4843 {
4844 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4845 	int q;
4846 #if defined(INET6) || defined(INET)
4847 	int i, err;
4848 #endif
4849 
4850 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4851 #if defined(INET6) || defined(INET)
4852 		tcp_lro_free(&rxq->ifr_lc);
4853 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4854 		    TCP_LRO_ENTRIES, min(1024,
4855 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4856 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4857 			goto fail;
4858 		}
4859 		rxq->ifr_lro_enabled = TRUE;
4860 #endif
4861 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4862 	}
4863 	return (0);
4864 #if defined(INET6) || defined(INET)
4865 fail:
4866 	/*
4867 	 * Free RX software descriptors allocated so far, we will only handle
4868 	 * the rings that completed, the failing case will have
4869 	 * cleaned up for itself. 'q' failed, so its the terminus.
4870 	 */
4871 	rxq = ctx->ifc_rxqs;
4872 	for (i = 0; i < q; ++i, rxq++) {
4873 		iflib_rx_sds_free(rxq);
4874 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4875 	}
4876 	return (err);
4877 #endif
4878 }
4879 
4880 /*********************************************************************
4881  *
4882  *  Free all receive rings.
4883  *
4884  **********************************************************************/
4885 static void
4886 iflib_rx_structures_free(if_ctx_t ctx)
4887 {
4888 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4889 
4890 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4891 		iflib_rx_sds_free(rxq);
4892 	}
4893 }
4894 
4895 static int
4896 iflib_qset_structures_setup(if_ctx_t ctx)
4897 {
4898 	int err;
4899 
4900 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4901 		return (err);
4902 
4903 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4904 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4905 		iflib_tx_structures_free(ctx);
4906 		iflib_rx_structures_free(ctx);
4907 	}
4908 	return (err);
4909 }
4910 
4911 int
4912 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4913 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4914 {
4915 
4916 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4917 }
4918 
4919 static int
4920 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4921 {
4922 	int i, cpuid, eqid, count;
4923 
4924 	CPU_COPY(&ctx->ifc_cpus, cpus);
4925 	count = CPU_COUNT(&ctx->ifc_cpus);
4926 	eqid = qid % count;
4927 	/* clear up to the qid'th bit */
4928 	for (i = 0; i < eqid; i++) {
4929 		cpuid = CPU_FFS(cpus);
4930 		MPASS(cpuid != 0);
4931 		CPU_CLR(cpuid-1, cpus);
4932 	}
4933 	cpuid = CPU_FFS(cpus);
4934 	MPASS(cpuid != 0);
4935 	return (cpuid-1);
4936 }
4937 
4938 int
4939 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4940 						iflib_intr_type_t type, driver_filter_t *filter,
4941 						void *filter_arg, int qid, char *name)
4942 {
4943 	struct grouptask *gtask;
4944 	struct taskqgroup *tqg;
4945 	iflib_filter_info_t info;
4946 	cpuset_t cpus;
4947 	gtask_fn_t *fn;
4948 	int tqrid, err, cpuid;
4949 	driver_filter_t *intr_fast;
4950 	void *q;
4951 
4952 	info = &ctx->ifc_filter_info;
4953 	tqrid = rid;
4954 
4955 	switch (type) {
4956 	/* XXX merge tx/rx for netmap? */
4957 	case IFLIB_INTR_TX:
4958 		q = &ctx->ifc_txqs[qid];
4959 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4960 		gtask = &ctx->ifc_txqs[qid].ift_task;
4961 		tqg = qgroup_if_io_tqg;
4962 		fn = _task_fn_tx;
4963 		intr_fast = iflib_fast_intr;
4964 		GROUPTASK_INIT(gtask, 0, fn, q);
4965 		break;
4966 	case IFLIB_INTR_RX:
4967 		q = &ctx->ifc_rxqs[qid];
4968 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4969 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4970 		tqg = qgroup_if_io_tqg;
4971 		fn = _task_fn_rx;
4972 		intr_fast = iflib_fast_intr;
4973 		GROUPTASK_INIT(gtask, 0, fn, q);
4974 		break;
4975 	case IFLIB_INTR_RXTX:
4976 		q = &ctx->ifc_rxqs[qid];
4977 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4978 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4979 		tqg = qgroup_if_io_tqg;
4980 		fn = _task_fn_rx;
4981 		intr_fast = iflib_fast_intr_rxtx;
4982 		GROUPTASK_INIT(gtask, 0, fn, q);
4983 		break;
4984 	case IFLIB_INTR_ADMIN:
4985 		q = ctx;
4986 		tqrid = -1;
4987 		info = &ctx->ifc_filter_info;
4988 		gtask = &ctx->ifc_admin_task;
4989 		tqg = qgroup_if_config_tqg;
4990 		fn = _task_fn_admin;
4991 		intr_fast = iflib_fast_intr_ctx;
4992 		break;
4993 	default:
4994 		panic("unknown net intr type");
4995 	}
4996 
4997 	info->ifi_filter = filter;
4998 	info->ifi_filter_arg = filter_arg;
4999 	info->ifi_task = gtask;
5000 	info->ifi_ctx = q;
5001 
5002 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5003 	if (err != 0) {
5004 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5005 		return (err);
5006 	}
5007 	if (type == IFLIB_INTR_ADMIN)
5008 		return (0);
5009 
5010 	if (tqrid != -1) {
5011 		cpuid = find_nth(ctx, &cpus, qid);
5012 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, rman_get_start(irq->ii_res), name);
5013 	} else {
5014 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5015 	}
5016 
5017 	return (0);
5018 }
5019 
5020 void
5021 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5022 {
5023 	struct grouptask *gtask;
5024 	struct taskqgroup *tqg;
5025 	gtask_fn_t *fn;
5026 	void *q;
5027 	int irq_num = -1;
5028 
5029 	switch (type) {
5030 	case IFLIB_INTR_TX:
5031 		q = &ctx->ifc_txqs[qid];
5032 		gtask = &ctx->ifc_txqs[qid].ift_task;
5033 		tqg = qgroup_if_io_tqg;
5034 		fn = _task_fn_tx;
5035 		if (irq != NULL)
5036 			irq_num = rman_get_start(irq->ii_res);
5037 		break;
5038 	case IFLIB_INTR_RX:
5039 		q = &ctx->ifc_rxqs[qid];
5040 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5041 		tqg = qgroup_if_io_tqg;
5042 		fn = _task_fn_rx;
5043 		if (irq != NULL)
5044 			irq_num = rman_get_start(irq->ii_res);
5045 		break;
5046 	case IFLIB_INTR_IOV:
5047 		q = ctx;
5048 		gtask = &ctx->ifc_vflr_task;
5049 		tqg = qgroup_if_config_tqg;
5050 		fn = _task_fn_iov;
5051 		break;
5052 	default:
5053 		panic("unknown net intr type");
5054 	}
5055 	GROUPTASK_INIT(gtask, 0, fn, q);
5056 	taskqgroup_attach(tqg, gtask, q, irq_num, name);
5057 }
5058 
5059 void
5060 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5061 {
5062 	if (irq->ii_tag)
5063 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5064 
5065 	if (irq->ii_res)
5066 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5067 }
5068 
5069 static int
5070 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5071 {
5072 	iflib_txq_t txq = ctx->ifc_txqs;
5073 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5074 	if_irq_t irq = &ctx->ifc_legacy_irq;
5075 	iflib_filter_info_t info;
5076 	struct grouptask *gtask;
5077 	struct taskqgroup *tqg;
5078 	gtask_fn_t *fn;
5079 	int tqrid;
5080 	void *q;
5081 	int err;
5082 
5083 	q = &ctx->ifc_rxqs[0];
5084 	info = &rxq[0].ifr_filter_info;
5085 	gtask = &rxq[0].ifr_task;
5086 	tqg = qgroup_if_io_tqg;
5087 	tqrid = irq->ii_rid = *rid;
5088 	fn = _task_fn_rx;
5089 
5090 	ctx->ifc_flags |= IFC_LEGACY;
5091 	info->ifi_filter = filter;
5092 	info->ifi_filter_arg = filter_arg;
5093 	info->ifi_task = gtask;
5094 	info->ifi_ctx = ctx;
5095 
5096 	/* We allocate a single interrupt resource */
5097 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5098 		return (err);
5099 	GROUPTASK_INIT(gtask, 0, fn, q);
5100 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5101 
5102 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5103 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5104 	return (0);
5105 }
5106 
5107 void
5108 iflib_led_create(if_ctx_t ctx)
5109 {
5110 
5111 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5112 	    device_get_nameunit(ctx->ifc_dev));
5113 }
5114 
5115 void
5116 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5117 {
5118 
5119 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5120 }
5121 
5122 void
5123 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5124 {
5125 
5126 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5127 }
5128 
5129 void
5130 iflib_admin_intr_deferred(if_ctx_t ctx)
5131 {
5132 #ifdef INVARIANTS
5133 	struct grouptask *gtask;
5134 
5135 	gtask = &ctx->ifc_admin_task;
5136 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5137 #endif
5138 
5139 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5140 }
5141 
5142 void
5143 iflib_iov_intr_deferred(if_ctx_t ctx)
5144 {
5145 
5146 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5147 }
5148 
5149 void
5150 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5151 {
5152 
5153 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5154 }
5155 
5156 void
5157 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5158 	char *name)
5159 {
5160 
5161 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5162 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5163 }
5164 
5165 void
5166 iflib_config_gtask_deinit(struct grouptask *gtask)
5167 {
5168 
5169 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5170 }
5171 
5172 void
5173 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5174 {
5175 	if_t ifp = ctx->ifc_ifp;
5176 	iflib_txq_t txq = ctx->ifc_txqs;
5177 
5178 	if_setbaudrate(ifp, baudrate);
5179 	if (baudrate >= IF_Gbps(10))
5180 		ctx->ifc_flags |= IFC_PREFETCH;
5181 
5182 	/* If link down, disable watchdog */
5183 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5184 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5185 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5186 	}
5187 	ctx->ifc_link_state = link_state;
5188 	if_link_state_change(ifp, link_state);
5189 }
5190 
5191 static int
5192 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5193 {
5194 	int credits;
5195 #ifdef INVARIANTS
5196 	int credits_pre = txq->ift_cidx_processed;
5197 #endif
5198 
5199 	if (ctx->isc_txd_credits_update == NULL)
5200 		return (0);
5201 
5202 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5203 		return (0);
5204 
5205 	txq->ift_processed += credits;
5206 	txq->ift_cidx_processed += credits;
5207 
5208 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5209 	if (txq->ift_cidx_processed >= txq->ift_size)
5210 		txq->ift_cidx_processed -= txq->ift_size;
5211 	return (credits);
5212 }
5213 
5214 static int
5215 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5216 {
5217 
5218 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5219 	    budget));
5220 }
5221 
5222 void
5223 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5224 	const char *description, if_int_delay_info_t info,
5225 	int offset, int value)
5226 {
5227 	info->iidi_ctx = ctx;
5228 	info->iidi_offset = offset;
5229 	info->iidi_value = value;
5230 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5231 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5232 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5233 	    info, 0, iflib_sysctl_int_delay, "I", description);
5234 }
5235 
5236 struct mtx *
5237 iflib_ctx_lock_get(if_ctx_t ctx)
5238 {
5239 
5240 	return (&ctx->ifc_mtx);
5241 }
5242 
5243 static int
5244 iflib_msix_init(if_ctx_t ctx)
5245 {
5246 	device_t dev = ctx->ifc_dev;
5247 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5248 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5249 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5250 	int iflib_num_tx_queues, iflib_num_rx_queues;
5251 	int err, admincnt, bar;
5252 
5253 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5254 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5255 
5256 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5257 
5258 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5259 	admincnt = sctx->isc_admin_intrcnt;
5260 	/* Override by global tuneable */
5261 	{
5262 		int i;
5263 		size_t len = sizeof(i);
5264 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5265 		if (err == 0) {
5266 			if (i == 0)
5267 				goto msi;
5268 		}
5269 		else {
5270 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5271 		}
5272 	}
5273 	/* Override by tuneable */
5274 	if (scctx->isc_disable_msix)
5275 		goto msi;
5276 
5277 	/*
5278 	** When used in a virtualized environment
5279 	** PCI BUSMASTER capability may not be set
5280 	** so explicity set it here and rewrite
5281 	** the ENABLE in the MSIX control register
5282 	** at this point to cause the host to
5283 	** successfully initialize us.
5284 	*/
5285 	{
5286 		int msix_ctrl, rid;
5287 
5288  		pci_enable_busmaster(dev);
5289 		rid = 0;
5290 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5291 			rid += PCIR_MSIX_CTRL;
5292 			msix_ctrl = pci_read_config(dev, rid, 2);
5293 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5294 			pci_write_config(dev, rid, msix_ctrl, 2);
5295 		} else {
5296 			device_printf(dev, "PCIY_MSIX capability not found; "
5297 			                   "or rid %d == 0.\n", rid);
5298 			goto msi;
5299 		}
5300 	}
5301 
5302 	/*
5303 	 * bar == -1 => "trust me I know what I'm doing"
5304 	 * Some drivers are for hardware that is so shoddily
5305 	 * documented that no one knows which bars are which
5306 	 * so the developer has to map all bars. This hack
5307 	 * allows shoddy garbage to use msix in this framework.
5308 	 */
5309 	if (bar != -1) {
5310 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5311 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5312 		if (ctx->ifc_msix_mem == NULL) {
5313 			/* May not be enabled */
5314 			device_printf(dev, "Unable to map MSIX table \n");
5315 			goto msi;
5316 		}
5317 	}
5318 	/* First try MSI/X */
5319 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5320 		device_printf(dev, "System has MSIX disabled \n");
5321 		bus_release_resource(dev, SYS_RES_MEMORY,
5322 		    bar, ctx->ifc_msix_mem);
5323 		ctx->ifc_msix_mem = NULL;
5324 		goto msi;
5325 	}
5326 #if IFLIB_DEBUG
5327 	/* use only 1 qset in debug mode */
5328 	queuemsgs = min(msgs - admincnt, 1);
5329 #else
5330 	queuemsgs = msgs - admincnt;
5331 #endif
5332 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5333 #ifdef RSS
5334 		queues = imin(queuemsgs, rss_getnumbuckets());
5335 #else
5336 		queues = queuemsgs;
5337 #endif
5338 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5339 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5340 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5341 	} else {
5342 		device_printf(dev, "Unable to fetch CPU list\n");
5343 		/* Figure out a reasonable auto config value */
5344 		queues = min(queuemsgs, mp_ncpus);
5345 	}
5346 #ifdef  RSS
5347 	/* If we're doing RSS, clamp at the number of RSS buckets */
5348 	if (queues > rss_getnumbuckets())
5349 		queues = rss_getnumbuckets();
5350 #endif
5351 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5352 		rx_queues = iflib_num_rx_queues;
5353 	else
5354 		rx_queues = queues;
5355 	/*
5356 	 * We want this to be all logical CPUs by default
5357 	 */
5358 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5359 		tx_queues = iflib_num_tx_queues;
5360 	else
5361 		tx_queues = mp_ncpus;
5362 
5363 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5364 #ifdef INVARIANTS
5365 		if (tx_queues != rx_queues)
5366 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5367 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5368 #endif
5369 		tx_queues = min(rx_queues, tx_queues);
5370 		rx_queues = min(rx_queues, tx_queues);
5371 	}
5372 
5373 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5374 
5375 	vectors = rx_queues + admincnt;
5376 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5377 		device_printf(dev,
5378 					  "Using MSIX interrupts with %d vectors\n", vectors);
5379 		scctx->isc_vectors = vectors;
5380 		scctx->isc_nrxqsets = rx_queues;
5381 		scctx->isc_ntxqsets = tx_queues;
5382 		scctx->isc_intr = IFLIB_INTR_MSIX;
5383 
5384 		return (vectors);
5385 	} else {
5386 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5387 	}
5388 msi:
5389 	vectors = pci_msi_count(dev);
5390 	scctx->isc_nrxqsets = 1;
5391 	scctx->isc_ntxqsets = 1;
5392 	scctx->isc_vectors = vectors;
5393 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5394 		device_printf(dev,"Using an MSI interrupt\n");
5395 		scctx->isc_intr = IFLIB_INTR_MSI;
5396 	} else {
5397 		device_printf(dev,"Using a Legacy interrupt\n");
5398 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5399 	}
5400 
5401 	return (vectors);
5402 }
5403 
5404 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5405 
5406 static int
5407 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5408 {
5409 	int rc;
5410 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5411 	struct sbuf *sb;
5412 	char *ring_state = "UNKNOWN";
5413 
5414 	/* XXX needed ? */
5415 	rc = sysctl_wire_old_buffer(req, 0);
5416 	MPASS(rc == 0);
5417 	if (rc != 0)
5418 		return (rc);
5419 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5420 	MPASS(sb != NULL);
5421 	if (sb == NULL)
5422 		return (ENOMEM);
5423 	if (state[3] <= 3)
5424 		ring_state = ring_states[state[3]];
5425 
5426 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5427 		    state[0], state[1], state[2], ring_state);
5428 	rc = sbuf_finish(sb);
5429 	sbuf_delete(sb);
5430         return(rc);
5431 }
5432 
5433 enum iflib_ndesc_handler {
5434 	IFLIB_NTXD_HANDLER,
5435 	IFLIB_NRXD_HANDLER,
5436 };
5437 
5438 static int
5439 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5440 {
5441 	if_ctx_t ctx = (void *)arg1;
5442 	enum iflib_ndesc_handler type = arg2;
5443 	char buf[256] = {0};
5444 	qidx_t *ndesc;
5445 	char *p, *next;
5446 	int nqs, rc, i;
5447 
5448 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5449 
5450 	nqs = 8;
5451 	switch(type) {
5452 	case IFLIB_NTXD_HANDLER:
5453 		ndesc = ctx->ifc_sysctl_ntxds;
5454 		if (ctx->ifc_sctx)
5455 			nqs = ctx->ifc_sctx->isc_ntxqs;
5456 		break;
5457 	case IFLIB_NRXD_HANDLER:
5458 		ndesc = ctx->ifc_sysctl_nrxds;
5459 		if (ctx->ifc_sctx)
5460 			nqs = ctx->ifc_sctx->isc_nrxqs;
5461 		break;
5462 	}
5463 	if (nqs == 0)
5464 		nqs = 8;
5465 
5466 	for (i=0; i<8; i++) {
5467 		if (i >= nqs)
5468 			break;
5469 		if (i)
5470 			strcat(buf, ",");
5471 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5472 	}
5473 
5474 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5475 	if (rc || req->newptr == NULL)
5476 		return rc;
5477 
5478 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5479 	    i++, p = strsep(&next, " ,")) {
5480 		ndesc[i] = strtoul(p, NULL, 10);
5481 	}
5482 
5483 	return(rc);
5484 }
5485 
5486 #define NAME_BUFLEN 32
5487 static void
5488 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5489 {
5490         device_t dev = iflib_get_dev(ctx);
5491 	struct sysctl_oid_list *child, *oid_list;
5492 	struct sysctl_ctx_list *ctx_list;
5493 	struct sysctl_oid *node;
5494 
5495 	ctx_list = device_get_sysctl_ctx(dev);
5496 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5497 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5498 						      CTLFLAG_RD, NULL, "IFLIB fields");
5499 	oid_list = SYSCTL_CHILDREN(node);
5500 
5501 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5502 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5503 		       "driver version");
5504 
5505 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5506 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5507 			"# of txqs to use, 0 => use default #");
5508 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5509 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5510 			"# of rxqs to use, 0 => use default #");
5511 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5512 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5513                        "permit #txq != #rxq");
5514 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5515                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5516                       "disable MSIX (default 0)");
5517 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5518 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5519                        "set the rx budget");
5520 
5521 	/* XXX change for per-queue sizes */
5522 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5523 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5524                        mp_ndesc_handler, "A",
5525                        "list of # of tx descriptors to use, 0 = use default #");
5526 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5527 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5528                        mp_ndesc_handler, "A",
5529                        "list of # of rx descriptors to use, 0 = use default #");
5530 }
5531 
5532 static void
5533 iflib_add_device_sysctl_post(if_ctx_t ctx)
5534 {
5535 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5536 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5537         device_t dev = iflib_get_dev(ctx);
5538 	struct sysctl_oid_list *child;
5539 	struct sysctl_ctx_list *ctx_list;
5540 	iflib_fl_t fl;
5541 	iflib_txq_t txq;
5542 	iflib_rxq_t rxq;
5543 	int i, j;
5544 	char namebuf[NAME_BUFLEN];
5545 	char *qfmt;
5546 	struct sysctl_oid *queue_node, *fl_node, *node;
5547 	struct sysctl_oid_list *queue_list, *fl_list;
5548 	ctx_list = device_get_sysctl_ctx(dev);
5549 
5550 	node = ctx->ifc_sysctl_node;
5551 	child = SYSCTL_CHILDREN(node);
5552 
5553 	if (scctx->isc_ntxqsets > 100)
5554 		qfmt = "txq%03d";
5555 	else if (scctx->isc_ntxqsets > 10)
5556 		qfmt = "txq%02d";
5557 	else
5558 		qfmt = "txq%d";
5559 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5560 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5561 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5562 					     CTLFLAG_RD, NULL, "Queue Name");
5563 		queue_list = SYSCTL_CHILDREN(queue_node);
5564 #if MEMORY_LOGGING
5565 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5566 				CTLFLAG_RD,
5567 				&txq->ift_dequeued, "total mbufs freed");
5568 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5569 				CTLFLAG_RD,
5570 				&txq->ift_enqueued, "total mbufs enqueued");
5571 #endif
5572 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5573 				   CTLFLAG_RD,
5574 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5575 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5576 				   CTLFLAG_RD,
5577 				   &txq->ift_pullups, "# of times m_pullup was called");
5578 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5579 				   CTLFLAG_RD,
5580 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5581 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5582 				   CTLFLAG_RD,
5583 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5584 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5585 				   CTLFLAG_RD,
5586 				   &txq->ift_map_failed, "# of times dma map failed");
5587 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5588 				   CTLFLAG_RD,
5589 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5590 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5591 				   CTLFLAG_RD,
5592 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5593 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5594 				   CTLFLAG_RD,
5595 				   &txq->ift_pidx, 1, "Producer Index");
5596 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5597 				   CTLFLAG_RD,
5598 				   &txq->ift_cidx, 1, "Consumer Index");
5599 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5600 				   CTLFLAG_RD,
5601 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5602 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5603 				   CTLFLAG_RD,
5604 				   &txq->ift_in_use, 1, "descriptors in use");
5605 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5606 				   CTLFLAG_RD,
5607 				   &txq->ift_processed, "descriptors procesed for clean");
5608 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5609 				   CTLFLAG_RD,
5610 				   &txq->ift_cleaned, "total cleaned");
5611 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5612 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5613 				0, mp_ring_state_handler, "A", "soft ring state");
5614 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5615 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5616 				       "# of enqueues to the mp_ring for this queue");
5617 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5618 				       CTLFLAG_RD, &txq->ift_br->drops,
5619 				       "# of drops in the mp_ring for this queue");
5620 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5621 				       CTLFLAG_RD, &txq->ift_br->starts,
5622 				       "# of normal consumer starts in the mp_ring for this queue");
5623 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5624 				       CTLFLAG_RD, &txq->ift_br->stalls,
5625 					       "# of consumer stalls in the mp_ring for this queue");
5626 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5627 			       CTLFLAG_RD, &txq->ift_br->restarts,
5628 				       "# of consumer restarts in the mp_ring for this queue");
5629 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5630 				       CTLFLAG_RD, &txq->ift_br->abdications,
5631 				       "# of consumer abdications in the mp_ring for this queue");
5632 	}
5633 
5634 	if (scctx->isc_nrxqsets > 100)
5635 		qfmt = "rxq%03d";
5636 	else if (scctx->isc_nrxqsets > 10)
5637 		qfmt = "rxq%02d";
5638 	else
5639 		qfmt = "rxq%d";
5640 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5641 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5642 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5643 					     CTLFLAG_RD, NULL, "Queue Name");
5644 		queue_list = SYSCTL_CHILDREN(queue_node);
5645 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5646 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5647 				       CTLFLAG_RD,
5648 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5649 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5650 				       CTLFLAG_RD,
5651 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5652 		}
5653 
5654 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5655 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5656 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5657 						     CTLFLAG_RD, NULL, "freelist Name");
5658 			fl_list = SYSCTL_CHILDREN(fl_node);
5659 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5660 				       CTLFLAG_RD,
5661 				       &fl->ifl_pidx, 1, "Producer Index");
5662 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5663 				       CTLFLAG_RD,
5664 				       &fl->ifl_cidx, 1, "Consumer Index");
5665 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5666 				       CTLFLAG_RD,
5667 				       &fl->ifl_credits, 1, "credits available");
5668 #if MEMORY_LOGGING
5669 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5670 					CTLFLAG_RD,
5671 					&fl->ifl_m_enqueued, "mbufs allocated");
5672 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5673 					CTLFLAG_RD,
5674 					&fl->ifl_m_dequeued, "mbufs freed");
5675 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5676 					CTLFLAG_RD,
5677 					&fl->ifl_cl_enqueued, "clusters allocated");
5678 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5679 					CTLFLAG_RD,
5680 					&fl->ifl_cl_dequeued, "clusters freed");
5681 #endif
5682 
5683 		}
5684 	}
5685 
5686 }
5687 
5688 #ifndef __NO_STRICT_ALIGNMENT
5689 static struct mbuf *
5690 iflib_fixup_rx(struct mbuf *m)
5691 {
5692 	struct mbuf *n;
5693 
5694 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5695 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5696 		m->m_data += ETHER_HDR_LEN;
5697 		n = m;
5698 	} else {
5699 		MGETHDR(n, M_NOWAIT, MT_DATA);
5700 		if (n == NULL) {
5701 			m_freem(m);
5702 			return (NULL);
5703 		}
5704 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5705 		m->m_data += ETHER_HDR_LEN;
5706 		m->m_len -= ETHER_HDR_LEN;
5707 		n->m_len = ETHER_HDR_LEN;
5708 		M_MOVE_PKTHDR(n, m);
5709 		n->m_next = m;
5710 	}
5711 	return (n);
5712 }
5713 #endif
5714