xref: /freebsd/sys/net/iflib.c (revision 7c1b51d6dc2e165ae7333373513b080f17cf79bd)
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 typedef struct iflib_filter_info {
141 	driver_filter_t *ifi_filter;
142 	void *ifi_filter_arg;
143 	struct grouptask *ifi_task;
144 	void *ifi_ctx;
145 } *iflib_filter_info_t;
146 
147 struct iflib_ctx {
148 	KOBJ_FIELDS;
149    /*
150    * Pointer to hardware driver's softc
151    */
152 	void *ifc_softc;
153 	device_t ifc_dev;
154 	if_t ifc_ifp;
155 
156 	cpuset_t ifc_cpus;
157 	if_shared_ctx_t ifc_sctx;
158 	struct if_softc_ctx ifc_softc_ctx;
159 
160 	struct mtx ifc_mtx;
161 
162 	uint16_t ifc_nhwtxqs;
163 	uint16_t ifc_nhwrxqs;
164 
165 	iflib_txq_t ifc_txqs;
166 	iflib_rxq_t ifc_rxqs;
167 	uint32_t ifc_if_flags;
168 	uint32_t ifc_flags;
169 	uint32_t ifc_max_fl_buf_size;
170 	int ifc_in_detach;
171 
172 	int ifc_link_state;
173 	int ifc_link_irq;
174 	int ifc_watchdog_events;
175 	struct cdev *ifc_led_dev;
176 	struct resource *ifc_msix_mem;
177 
178 	struct if_irq ifc_legacy_irq;
179 	struct grouptask ifc_admin_task;
180 	struct grouptask ifc_vflr_task;
181 	struct iflib_filter_info ifc_filter_info;
182 	struct ifmedia	ifc_media;
183 
184 	struct sysctl_oid *ifc_sysctl_node;
185 	uint16_t ifc_sysctl_ntxqs;
186 	uint16_t ifc_sysctl_nrxqs;
187 	uint16_t ifc_sysctl_qs_eq_override;
188 
189 	qidx_t ifc_sysctl_ntxds[8];
190 	qidx_t ifc_sysctl_nrxds[8];
191 	struct if_txrx ifc_txrx;
192 #define isc_txd_encap  ifc_txrx.ift_txd_encap
193 #define isc_txd_flush  ifc_txrx.ift_txd_flush
194 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
195 #define isc_rxd_available ifc_txrx.ift_rxd_available
196 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
197 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
198 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
201 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
202 	eventhandler_tag ifc_vlan_attach_event;
203 	eventhandler_tag ifc_vlan_detach_event;
204 	uint8_t ifc_mac[ETHER_ADDR_LEN];
205 	char ifc_mtx_name[16];
206 };
207 
208 
209 void *
210 iflib_get_softc(if_ctx_t ctx)
211 {
212 
213 	return (ctx->ifc_softc);
214 }
215 
216 device_t
217 iflib_get_dev(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_dev);
221 }
222 
223 if_t
224 iflib_get_ifp(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_ifp);
228 }
229 
230 struct ifmedia *
231 iflib_get_media(if_ctx_t ctx)
232 {
233 
234 	return (&ctx->ifc_media);
235 }
236 
237 void
238 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
239 {
240 
241 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
242 }
243 
244 if_softc_ctx_t
245 iflib_get_softc_ctx(if_ctx_t ctx)
246 {
247 
248 	return (&ctx->ifc_softc_ctx);
249 }
250 
251 if_shared_ctx_t
252 iflib_get_sctx(if_ctx_t ctx)
253 {
254 
255 	return (ctx->ifc_sctx);
256 }
257 
258 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
259 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
260 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
261 
262 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
263 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
264 
265 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
266 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
267 #define RX_SW_DESC_INUSE        (1 << 3)
268 #define TX_SW_DESC_MAPPED       (1 << 4)
269 
270 #define	M_TOOBIG		M_PROTO1
271 
272 typedef struct iflib_sw_rx_desc_array {
273 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
274 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
275 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
276 	uint8_t		*ifsd_flags;
277 } iflib_rxsd_array_t;
278 
279 typedef struct iflib_sw_tx_desc_array {
280 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
282 	uint8_t		*ifsd_flags;
283 } if_txsd_vec_t;
284 
285 
286 /* magic number that should be high enough for any hardware */
287 #define IFLIB_MAX_TX_SEGS		128
288 #define IFLIB_MAX_RX_SEGS		32
289 #define IFLIB_RX_COPY_THRESH		128
290 #define IFLIB_MAX_RX_REFRESH		32
291 /* The minimum descriptors per second before we start coalescing */
292 #define IFLIB_MIN_DESC_SEC		16384
293 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
294 #define IFLIB_QUEUE_IDLE		0
295 #define IFLIB_QUEUE_HUNG		1
296 #define IFLIB_QUEUE_WORKING		2
297 /* maximum number of txqs that can share an rx interrupt */
298 #define IFLIB_MAX_TX_SHARED_INTR	4
299 
300 /* this should really scale with ring size - this is a fairly arbitrary value */
301 #define TX_BATCH_SIZE			32
302 
303 #define IFLIB_RESTART_BUDGET		8
304 
305 #define	IFC_LEGACY		0x001
306 #define	IFC_QFLUSH		0x002
307 #define	IFC_MULTISEG		0x004
308 #define	IFC_DMAR		0x008
309 #define	IFC_SC_ALLOCATED	0x010
310 #define	IFC_INIT_DONE		0x020
311 #define	IFC_PREFETCH		0x040
312 #define	IFC_DO_RESET		0x080
313 #define	IFC_CHECK_HUNG		0x100
314 
315 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
316 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
317 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
318 struct iflib_txq {
319 	qidx_t		ift_in_use;
320 	qidx_t		ift_cidx;
321 	qidx_t		ift_cidx_processed;
322 	qidx_t		ift_pidx;
323 	uint8_t		ift_gen;
324 	uint8_t		ift_br_offset;
325 	uint16_t	ift_npending;
326 	uint16_t	ift_db_pending;
327 	uint16_t	ift_rs_pending;
328 	/* implicit pad */
329 	uint8_t		ift_txd_size[8];
330 	uint64_t	ift_processed;
331 	uint64_t	ift_cleaned;
332 	uint64_t	ift_cleaned_prev;
333 #if MEMORY_LOGGING
334 	uint64_t	ift_enqueued;
335 	uint64_t	ift_dequeued;
336 #endif
337 	uint64_t	ift_no_tx_dma_setup;
338 	uint64_t	ift_no_desc_avail;
339 	uint64_t	ift_mbuf_defrag_failed;
340 	uint64_t	ift_mbuf_defrag;
341 	uint64_t	ift_map_failed;
342 	uint64_t	ift_txd_encap_efbig;
343 	uint64_t	ift_pullups;
344 
345 	struct mtx	ift_mtx;
346 	struct mtx	ift_db_mtx;
347 
348 	/* constant values */
349 	if_ctx_t	ift_ctx;
350 	struct ifmp_ring        *ift_br;
351 	struct grouptask	ift_task;
352 	qidx_t		ift_size;
353 	uint16_t	ift_id;
354 	struct callout	ift_timer;
355 
356 	if_txsd_vec_t	ift_sds;
357 	uint8_t		ift_qstatus;
358 	uint8_t		ift_closed;
359 	uint8_t		ift_update_freq;
360 	struct iflib_filter_info ift_filter_info;
361 	bus_dma_tag_t		ift_desc_tag;
362 	bus_dma_tag_t		ift_tso_desc_tag;
363 	iflib_dma_info_t	ift_ifdi;
364 #define MTX_NAME_LEN 16
365 	char                    ift_mtx_name[MTX_NAME_LEN];
366 	char                    ift_db_mtx_name[MTX_NAME_LEN];
367 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
368 #ifdef IFLIB_DIAGNOSTICS
369 	uint64_t ift_cpu_exec_count[256];
370 #endif
371 } __aligned(CACHE_LINE_SIZE);
372 
373 struct iflib_fl {
374 	qidx_t		ifl_cidx;
375 	qidx_t		ifl_pidx;
376 	qidx_t		ifl_credits;
377 	uint8_t		ifl_gen;
378 	uint8_t		ifl_rxd_size;
379 #if MEMORY_LOGGING
380 	uint64_t	ifl_m_enqueued;
381 	uint64_t	ifl_m_dequeued;
382 	uint64_t	ifl_cl_enqueued;
383 	uint64_t	ifl_cl_dequeued;
384 #endif
385 	/* implicit pad */
386 
387 	bitstr_t 	*ifl_rx_bitmap;
388 	qidx_t		ifl_fragidx;
389 	/* constant */
390 	qidx_t		ifl_size;
391 	uint16_t	ifl_buf_size;
392 	uint16_t	ifl_cltype;
393 	uma_zone_t	ifl_zone;
394 	iflib_rxsd_array_t	ifl_sds;
395 	iflib_rxq_t	ifl_rxq;
396 	uint8_t		ifl_id;
397 	bus_dma_tag_t           ifl_desc_tag;
398 	iflib_dma_info_t	ifl_ifdi;
399 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
400 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
401 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
402 }  __aligned(CACHE_LINE_SIZE);
403 
404 static inline qidx_t
405 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
406 {
407 	qidx_t used;
408 
409 	if (pidx > cidx)
410 		used = pidx - cidx;
411 	else if (pidx < cidx)
412 		used = size - cidx + pidx;
413 	else if (gen == 0 && pidx == cidx)
414 		used = 0;
415 	else if (gen == 1 && pidx == cidx)
416 		used = size;
417 	else
418 		panic("bad state");
419 
420 	return (used);
421 }
422 
423 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
424 
425 #define IDXDIFF(head, tail, wrap) \
426 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
427 
428 struct iflib_rxq {
429 	/* If there is a separate completion queue -
430 	 * these are the cq cidx and pidx. Otherwise
431 	 * these are unused.
432 	 */
433 	qidx_t		ifr_size;
434 	qidx_t		ifr_cq_cidx;
435 	qidx_t		ifr_cq_pidx;
436 	uint8_t		ifr_cq_gen;
437 	uint8_t		ifr_fl_offset;
438 
439 	if_ctx_t	ifr_ctx;
440 	iflib_fl_t	ifr_fl;
441 	uint64_t	ifr_rx_irq;
442 	uint16_t	ifr_id;
443 	uint8_t		ifr_lro_enabled;
444 	uint8_t		ifr_nfl;
445 	uint8_t		ifr_ntxqirq;
446 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
447 	struct lro_ctrl			ifr_lc;
448 	struct grouptask        ifr_task;
449 	struct iflib_filter_info ifr_filter_info;
450 	iflib_dma_info_t		ifr_ifdi;
451 
452 	/* dynamically allocate if any drivers need a value substantially larger than this */
453 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
454 #ifdef IFLIB_DIAGNOSTICS
455 	uint64_t ifr_cpu_exec_count[256];
456 #endif
457 }  __aligned(CACHE_LINE_SIZE);
458 
459 typedef struct if_rxsd {
460 	caddr_t *ifsd_cl;
461 	struct mbuf **ifsd_m;
462 	iflib_fl_t ifsd_fl;
463 	qidx_t ifsd_cidx;
464 } *if_rxsd_t;
465 
466 /* multiple of word size */
467 #ifdef __LP64__
468 #define PKT_INFO_SIZE	6
469 #define RXD_INFO_SIZE	5
470 #define PKT_TYPE uint64_t
471 #else
472 #define PKT_INFO_SIZE	11
473 #define RXD_INFO_SIZE	8
474 #define PKT_TYPE uint32_t
475 #endif
476 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
477 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
478 
479 typedef struct if_pkt_info_pad {
480 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
481 } *if_pkt_info_pad_t;
482 typedef struct if_rxd_info_pad {
483 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
484 } *if_rxd_info_pad_t;
485 
486 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
487 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
488 
489 
490 static inline void
491 pkt_info_zero(if_pkt_info_t pi)
492 {
493 	if_pkt_info_pad_t pi_pad;
494 
495 	pi_pad = (if_pkt_info_pad_t)pi;
496 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
497 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
498 #ifndef __LP64__
499 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
500 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
501 #endif
502 }
503 
504 static inline void
505 rxd_info_zero(if_rxd_info_t ri)
506 {
507 	if_rxd_info_pad_t ri_pad;
508 	int i;
509 
510 	ri_pad = (if_rxd_info_pad_t)ri;
511 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
512 		ri_pad->rxd_val[i] = 0;
513 		ri_pad->rxd_val[i+1] = 0;
514 		ri_pad->rxd_val[i+2] = 0;
515 		ri_pad->rxd_val[i+3] = 0;
516 	}
517 #ifdef __LP64__
518 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
519 #endif
520 }
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA (bus_addr_t)-1
527 
528 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
531 
532 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
533 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
534 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
535 
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
539 
540 
541 /* Our boot-time initialization hook */
542 static int	iflib_module_event_handler(module_t, int, void *);
543 
544 static moduledata_t iflib_moduledata = {
545 	"iflib",
546 	iflib_module_event_handler,
547 	NULL
548 };
549 
550 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
551 MODULE_VERSION(iflib, 1);
552 
553 MODULE_DEPEND(iflib, pci, 1, 1, 1);
554 MODULE_DEPEND(iflib, ether, 1, 1, 1);
555 
556 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
557 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
558 
559 #ifndef IFLIB_DEBUG_COUNTERS
560 #ifdef INVARIANTS
561 #define IFLIB_DEBUG_COUNTERS 1
562 #else
563 #define IFLIB_DEBUG_COUNTERS 0
564 #endif /* !INVARIANTS */
565 #endif
566 
567 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
568                    "iflib driver parameters");
569 
570 /*
571  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
572  */
573 static int iflib_min_tx_latency = 0;
574 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
575 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
576 static int iflib_no_tx_batch = 0;
577 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
578 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
579 
580 
581 #if IFLIB_DEBUG_COUNTERS
582 
583 static int iflib_tx_seen;
584 static int iflib_tx_sent;
585 static int iflib_tx_encap;
586 static int iflib_rx_allocs;
587 static int iflib_fl_refills;
588 static int iflib_fl_refills_large;
589 static int iflib_tx_frees;
590 
591 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
592 		   &iflib_tx_seen, 0, "# tx mbufs seen");
593 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
594 		   &iflib_tx_sent, 0, "# tx mbufs sent");
595 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
596 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
597 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
598 		   &iflib_tx_frees, 0, "# tx frees");
599 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
600 		   &iflib_rx_allocs, 0, "# rx allocations");
601 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
602 		   &iflib_fl_refills, 0, "# refills");
603 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
604 		   &iflib_fl_refills_large, 0, "# large refills");
605 
606 
607 static int iflib_txq_drain_flushing;
608 static int iflib_txq_drain_oactive;
609 static int iflib_txq_drain_notready;
610 static int iflib_txq_drain_encapfail;
611 
612 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
613 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
614 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
615 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
617 		   &iflib_txq_drain_notready, 0, "# drain notready");
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
619 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
620 
621 
622 static int iflib_encap_load_mbuf_fail;
623 static int iflib_encap_txq_avail_fail;
624 static int iflib_encap_txd_encap_fail;
625 
626 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
627 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
628 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
629 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
630 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
631 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
632 
633 static int iflib_task_fn_rxs;
634 static int iflib_rx_intr_enables;
635 static int iflib_fast_intrs;
636 static int iflib_intr_link;
637 static int iflib_intr_msix;
638 static int iflib_rx_unavail;
639 static int iflib_rx_ctx_inactive;
640 static int iflib_rx_zero_len;
641 static int iflib_rx_if_input;
642 static int iflib_rx_mbuf_null;
643 static int iflib_rxd_flush;
644 
645 static int iflib_verbose_debug;
646 
647 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
648 		   &iflib_intr_link, 0, "# intr link calls");
649 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
650 		   &iflib_intr_msix, 0, "# intr msix calls");
651 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
652 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
653 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
654 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
655 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
656 		   &iflib_fast_intrs, 0, "# fast_intr calls");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
658 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
660 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
662 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
666 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
668 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
669 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
670 		   &iflib_verbose_debug, 0, "enable verbose debugging");
671 
672 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
673 static void
674 iflib_debug_reset(void)
675 {
676 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
677 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
678 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
679 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
680 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
681 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
682 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
683 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
684 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
685 }
686 
687 #else
688 #define DBG_COUNTER_INC(name)
689 static void iflib_debug_reset(void) {}
690 #endif
691 
692 
693 
694 #define IFLIB_DEBUG 0
695 
696 static void iflib_tx_structures_free(if_ctx_t ctx);
697 static void iflib_rx_structures_free(if_ctx_t ctx);
698 static int iflib_queues_alloc(if_ctx_t ctx);
699 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
700 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
701 static int iflib_qset_structures_setup(if_ctx_t ctx);
702 static int iflib_msix_init(if_ctx_t ctx);
703 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
704 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
705 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
706 static int iflib_register(if_ctx_t);
707 static void iflib_init_locked(if_ctx_t ctx);
708 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
709 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
710 static void iflib_ifmp_purge(iflib_txq_t txq);
711 static void _iflib_pre_assert(if_softc_ctx_t scctx);
712 static void iflib_stop(if_ctx_t ctx);
713 static void iflib_if_init_locked(if_ctx_t ctx);
714 #ifndef __NO_STRICT_ALIGNMENT
715 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
716 #endif
717 
718 #ifdef DEV_NETMAP
719 #include <sys/selinfo.h>
720 #include <net/netmap.h>
721 #include <dev/netmap/netmap_kern.h>
722 
723 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
724 
725 /*
726  * device-specific sysctl variables:
727  *
728  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
729  *	During regular operations the CRC is stripped, but on some
730  *	hardware reception of frames not multiple of 64 is slower,
731  *	so using crcstrip=0 helps in benchmarks.
732  *
733  * iflib_rx_miss, iflib_rx_miss_bufs:
734  *	count packets that might be missed due to lost interrupts.
735  */
736 SYSCTL_DECL(_dev_netmap);
737 /*
738  * The xl driver by default strips CRCs and we do not override it.
739  */
740 
741 int iflib_crcstrip = 1;
742 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
743     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
744 
745 int iflib_rx_miss, iflib_rx_miss_bufs;
746 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
747     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
748 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
749     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
750 
751 /*
752  * Register/unregister. We are already under netmap lock.
753  * Only called on the first register or the last unregister.
754  */
755 static int
756 iflib_netmap_register(struct netmap_adapter *na, int onoff)
757 {
758 	struct ifnet *ifp = na->ifp;
759 	if_ctx_t ctx = ifp->if_softc;
760 	int status;
761 
762 	CTX_LOCK(ctx);
763 	IFDI_INTR_DISABLE(ctx);
764 
765 	/* Tell the stack that the interface is no longer active */
766 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
767 
768 	if (!CTX_IS_VF(ctx))
769 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
770 
771 	/* enable or disable flags and callbacks in na and ifp */
772 	if (onoff) {
773 		nm_set_native_flags(na);
774 	} else {
775 		nm_clear_native_flags(na);
776 	}
777 	iflib_stop(ctx);
778 	iflib_init_locked(ctx);
779 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
780 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
781 	if (status)
782 		nm_clear_native_flags(na);
783 	CTX_UNLOCK(ctx);
784 	return (status);
785 }
786 
787 /*
788  * Reconcile kernel and user view of the transmit ring.
789  *
790  * All information is in the kring.
791  * Userspace wants to send packets up to the one before kring->rhead,
792  * kernel knows kring->nr_hwcur is the first unsent packet.
793  *
794  * Here we push packets out (as many as possible), and possibly
795  * reclaim buffers from previously completed transmission.
796  *
797  * The caller (netmap) guarantees that there is only one instance
798  * running at any time. Any interference with other driver
799  * methods should be handled by the individual drivers.
800  */
801 static int
802 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	struct ifnet *ifp = na->ifp;
806 	struct netmap_ring *ring = kring->ring;
807 	u_int nm_i;	/* index into the netmap ring */
808 	u_int nic_i;	/* index into the NIC ring */
809 	u_int n;
810 	u_int const lim = kring->nkr_num_slots - 1;
811 	u_int const head = kring->rhead;
812 	struct if_pkt_info pi;
813 
814 	/*
815 	 * interrupts on every tx packet are expensive so request
816 	 * them every half ring, or where NS_REPORT is set
817 	 */
818 	u_int report_frequency = kring->nkr_num_slots >> 1;
819 	/* device-specific */
820 	if_ctx_t ctx = ifp->if_softc;
821 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
822 
823 	if (txq->ift_sds.ifsd_map)
824 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
825 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
826 
827 
828 	/*
829 	 * First part: process new packets to send.
830 	 * nm_i is the current index in the netmap ring,
831 	 * nic_i is the corresponding index in the NIC ring.
832 	 *
833 	 * If we have packets to send (nm_i != head)
834 	 * iterate over the netmap ring, fetch length and update
835 	 * the corresponding slot in the NIC ring. Some drivers also
836 	 * need to update the buffer's physical address in the NIC slot
837 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
838 	 *
839 	 * The netmap_reload_map() calls is especially expensive,
840 	 * even when (as in this case) the tag is 0, so do only
841 	 * when the buffer has actually changed.
842 	 *
843 	 * If possible do not set the report/intr bit on all slots,
844 	 * but only a few times per ring or when NS_REPORT is set.
845 	 *
846 	 * Finally, on 10G and faster drivers, it might be useful
847 	 * to prefetch the next slot and txr entry.
848 	 */
849 
850 	nm_i = kring->nr_hwcur;
851 	pkt_info_zero(&pi);
852 	pi.ipi_segs = txq->ift_segs;
853 	pi.ipi_qsidx = kring->ring_id;
854 	if (nm_i != head) {	/* we have new packets to send */
855 		nic_i = netmap_idx_k2n(kring, nm_i);
856 
857 		__builtin_prefetch(&ring->slot[nm_i]);
858 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
859 		if (txq->ift_sds.ifsd_map)
860 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
861 
862 		for (n = 0; nm_i != head; n++) {
863 			struct netmap_slot *slot = &ring->slot[nm_i];
864 			u_int len = slot->len;
865 			uint64_t paddr;
866 			void *addr = PNMB(na, slot, &paddr);
867 			int flags = (slot->flags & NS_REPORT ||
868 				nic_i == 0 || nic_i == report_frequency) ?
869 				IPI_TX_INTR : 0;
870 
871 			/* device-specific */
872 			pi.ipi_len = len;
873 			pi.ipi_segs[0].ds_addr = paddr;
874 			pi.ipi_segs[0].ds_len = len;
875 			pi.ipi_nsegs = 1;
876 			pi.ipi_ndescs = 0;
877 			pi.ipi_pidx = nic_i;
878 			pi.ipi_flags = flags;
879 
880 			/* Fill the slot in the NIC ring. */
881 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
882 
883 			/* prefetch for next round */
884 			__builtin_prefetch(&ring->slot[nm_i + 1]);
885 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
886 			if (txq->ift_sds.ifsd_map) {
887 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
888 
889 				NM_CHECK_ADDR_LEN(na, addr, len);
890 
891 				if (slot->flags & NS_BUF_CHANGED) {
892 					/* buffer has changed, reload map */
893 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
894 				}
895 				/* make sure changes to the buffer are synced */
896 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
897 						BUS_DMASYNC_PREWRITE);
898 			}
899 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
900 			nm_i = nm_next(nm_i, lim);
901 			nic_i = nm_next(nic_i, lim);
902 		}
903 		kring->nr_hwcur = head;
904 
905 		/* synchronize the NIC ring */
906 		if (txq->ift_sds.ifsd_map)
907 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
908 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
909 
910 		/* (re)start the tx unit up to slot nic_i (excluded) */
911 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
912 	}
913 
914 	/*
915 	 * Second part: reclaim buffers for completed transmissions.
916 	 */
917 	if (iflib_tx_credits_update(ctx, txq)) {
918 		/* some tx completed, increment avail */
919 		nic_i = txq->ift_cidx_processed;
920 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
921 	}
922 	return (0);
923 }
924 
925 /*
926  * Reconcile kernel and user view of the receive ring.
927  * Same as for the txsync, this routine must be efficient.
928  * The caller guarantees a single invocations, but races against
929  * the rest of the driver should be handled here.
930  *
931  * On call, kring->rhead is the first packet that userspace wants
932  * to keep, and kring->rcur is the wakeup point.
933  * The kernel has previously reported packets up to kring->rtail.
934  *
935  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
936  * of whether or not we received an interrupt.
937  */
938 static int
939 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
940 {
941 	struct netmap_adapter *na = kring->na;
942 	struct netmap_ring *ring = kring->ring;
943 	uint32_t nm_i;	/* index into the netmap ring */
944 	uint32_t nic_i, nic_i_start;	/* index into the NIC ring */
945 	u_int i, n;
946 	u_int const lim = kring->nkr_num_slots - 1;
947 	u_int const head = kring->rhead;
948 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
949 	struct if_rxd_info ri;
950 	struct if_rxd_update iru;
951 
952 	struct ifnet *ifp = na->ifp;
953 	if_ctx_t ctx = ifp->if_softc;
954 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
955 	iflib_fl_t fl = rxq->ifr_fl;
956 	if (head > lim)
957 		return netmap_ring_reinit(kring);
958 
959 	/* XXX check sync modes */
960 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
961 		if (fl->ifl_sds.ifsd_map == NULL)
962 			continue;
963 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
964 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
965 	}
966 	/*
967 	 * First part: import newly received packets.
968 	 *
969 	 * nm_i is the index of the next free slot in the netmap ring,
970 	 * nic_i is the index of the next received packet in the NIC ring,
971 	 * and they may differ in case if_init() has been called while
972 	 * in netmap mode. For the receive ring we have
973 	 *
974 	 *	nic_i = rxr->next_check;
975 	 *	nm_i = kring->nr_hwtail (previous)
976 	 * and
977 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
978 	 *
979 	 * rxr->next_check is set to 0 on a ring reinit
980 	 */
981 	if (netmap_no_pendintr || force_update) {
982 		int crclen = iflib_crcstrip ? 0 : 4;
983 		int error, avail;
984 		uint16_t slot_flags = kring->nkr_slot_flags;
985 
986 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
987 			nic_i = fl->ifl_cidx;
988 			nm_i = netmap_idx_n2k(kring, nic_i);
989 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
990 			for (n = 0; avail > 0; n++, avail--) {
991 				rxd_info_zero(&ri);
992 				ri.iri_frags = rxq->ifr_frags;
993 				ri.iri_qsidx = kring->ring_id;
994 				ri.iri_ifp = ctx->ifc_ifp;
995 				ri.iri_cidx = nic_i;
996 
997 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
998 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
999 				ring->slot[nm_i].flags = slot_flags;
1000 				if (fl->ifl_sds.ifsd_map)
1001 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1002 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1003 				nm_i = nm_next(nm_i, lim);
1004 				nic_i = nm_next(nic_i, lim);
1005 			}
1006 			if (n) { /* update the state variables */
1007 				if (netmap_no_pendintr && !force_update) {
1008 					/* diagnostics */
1009 					iflib_rx_miss ++;
1010 					iflib_rx_miss_bufs += n;
1011 				}
1012 				fl->ifl_cidx = nic_i;
1013 				kring->nr_hwtail = nm_i;
1014 			}
1015 			kring->nr_kflags &= ~NKR_PENDINTR;
1016 		}
1017 	}
1018 	/*
1019 	 * Second part: skip past packets that userspace has released.
1020 	 * (kring->nr_hwcur to head excluded),
1021 	 * and make the buffers available for reception.
1022 	 * As usual nm_i is the index in the netmap ring,
1023 	 * nic_i is the index in the NIC ring, and
1024 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1025 	 */
1026 	/* XXX not sure how this will work with multiple free lists */
1027 	nm_i = kring->nr_hwcur;
1028 	if (nm_i == head)
1029 		return (0);
1030 
1031 	iru.iru_paddrs = fl->ifl_bus_addrs;
1032 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1033 	iru.iru_idxs = fl->ifl_rxd_idxs;
1034 	iru.iru_qsidx = rxq->ifr_id;
1035 	iru.iru_buf_size = fl->ifl_buf_size;
1036 	iru.iru_flidx = fl->ifl_id;
1037 	nic_i_start = nic_i = netmap_idx_k2n(kring, nm_i);
1038 	for (i = 0; nm_i != head; i++) {
1039 		struct netmap_slot *slot = &ring->slot[nm_i];
1040 		void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
1041 
1042 		if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
1043 			goto ring_reset;
1044 
1045 		fl->ifl_vm_addrs[i] = addr;
1046 		if (fl->ifl_sds.ifsd_map && (slot->flags & NS_BUF_CHANGED)) {
1047 			/* buffer has changed, reload map */
1048 			netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
1049 		}
1050 		slot->flags &= ~NS_BUF_CHANGED;
1051 
1052 		nm_i = nm_next(nm_i, lim);
1053 		fl->ifl_rxd_idxs[i] = nic_i = nm_next(nic_i, lim);
1054 		if (nm_i != head && i < IFLIB_MAX_RX_REFRESH)
1055 			continue;
1056 
1057 		iru.iru_pidx = nic_i_start;
1058 		iru.iru_count = i;
1059 		i = 0;
1060 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1061 		if (fl->ifl_sds.ifsd_map == NULL) {
1062 			nic_i_start = nic_i;
1063 			continue;
1064 		}
1065 		nic_i = nic_i_start;
1066 		for (n = 0; n < iru.iru_count; n++) {
1067 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
1068 					BUS_DMASYNC_PREREAD);
1069 			nic_i = nm_next(nic_i, lim);
1070 		}
1071 		nic_i_start = nic_i;
1072 	}
1073 	kring->nr_hwcur = head;
1074 
1075 	if (fl->ifl_sds.ifsd_map)
1076 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1077 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1078 	/*
1079 	 * IMPORTANT: we must leave one free slot in the ring,
1080 	 * so move nic_i back by one unit
1081 	 */
1082 	nic_i = nm_prev(nic_i, lim);
1083 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
1084 	return 0;
1085 
1086 ring_reset:
1087 	return netmap_ring_reinit(kring);
1088 }
1089 
1090 static void
1091 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1092 {
1093 	struct ifnet *ifp = na->ifp;
1094 	if_ctx_t ctx = ifp->if_softc;
1095 
1096 	CTX_LOCK(ctx);
1097 	if (onoff) {
1098 		IFDI_INTR_ENABLE(ctx);
1099 	} else {
1100 		IFDI_INTR_DISABLE(ctx);
1101 	}
1102 	CTX_UNLOCK(ctx);
1103 }
1104 
1105 
1106 static int
1107 iflib_netmap_attach(if_ctx_t ctx)
1108 {
1109 	struct netmap_adapter na;
1110 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1111 
1112 	bzero(&na, sizeof(na));
1113 
1114 	na.ifp = ctx->ifc_ifp;
1115 	na.na_flags = NAF_BDG_MAYSLEEP;
1116 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1117 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1118 
1119 	na.num_tx_desc = scctx->isc_ntxd[0];
1120 	na.num_rx_desc = scctx->isc_nrxd[0];
1121 	na.nm_txsync = iflib_netmap_txsync;
1122 	na.nm_rxsync = iflib_netmap_rxsync;
1123 	na.nm_register = iflib_netmap_register;
1124 	na.nm_intr = iflib_netmap_intr;
1125 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1126 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1127 	return (netmap_attach(&na));
1128 }
1129 
1130 static void
1131 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1132 {
1133 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1134 	struct netmap_slot *slot;
1135 
1136 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1137 	if (slot == NULL)
1138 		return;
1139 	if (txq->ift_sds.ifsd_map == NULL)
1140 		return;
1141 
1142 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1143 
1144 		/*
1145 		 * In netmap mode, set the map for the packet buffer.
1146 		 * NOTE: Some drivers (not this one) also need to set
1147 		 * the physical buffer address in the NIC ring.
1148 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1149 		 * netmap slot index, si
1150 		 */
1151 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1152 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1153 	}
1154 }
1155 static void
1156 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1157 {
1158 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1159 	struct netmap_slot *slot;
1160 	struct if_rxd_update iru;
1161 	iflib_fl_t fl;
1162 	bus_dmamap_t *map;
1163 	int nrxd;
1164 	uint32_t i, j, pidx_start;
1165 
1166 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1167 	if (slot == NULL)
1168 		return;
1169 	fl = &rxq->ifr_fl[0];
1170 	map = fl->ifl_sds.ifsd_map;
1171 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1172 	iru.iru_paddrs = fl->ifl_bus_addrs;
1173 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1174 	iru.iru_idxs = fl->ifl_rxd_idxs;
1175 	iru.iru_qsidx = rxq->ifr_id;
1176 	iru.iru_buf_size = rxq->ifr_fl[0].ifl_buf_size;
1177 	iru.iru_flidx = 0;
1178 
1179 	for (pidx_start = i = j = 0; i < nrxd; i++, j++) {
1180 		int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1181 		void *addr;
1182 
1183 		fl->ifl_rxd_idxs[j] = i;
1184 		addr = fl->ifl_vm_addrs[j] = PNMB(na, slot + sj, &fl->ifl_bus_addrs[j]);
1185 		if (map) {
1186 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1187 			map++;
1188 		}
1189 
1190 		if (j < IFLIB_MAX_RX_REFRESH && i < nrxd - 1)
1191 			continue;
1192 
1193 		iru.iru_pidx = pidx_start;
1194 		pidx_start = i;
1195 		iru.iru_count = j;
1196 		j = 0;
1197 		MPASS(pidx_start + j <= nrxd);
1198 		/* Update descriptors and the cached value */
1199 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1200 	}
1201 	/* preserve queue */
1202 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1203 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1204 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1205 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1206 	} else
1207 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1208 }
1209 
1210 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1211 
1212 #else
1213 #define iflib_netmap_txq_init(ctx, txq)
1214 #define iflib_netmap_rxq_init(ctx, rxq)
1215 #define iflib_netmap_detach(ifp)
1216 
1217 #define iflib_netmap_attach(ctx) (0)
1218 #define netmap_rx_irq(ifp, qid, budget) (0)
1219 #define netmap_tx_irq(ifp, qid) do {} while (0)
1220 
1221 #endif
1222 
1223 #if defined(__i386__) || defined(__amd64__)
1224 static __inline void
1225 prefetch(void *x)
1226 {
1227 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1228 }
1229 #else
1230 #define prefetch(x)
1231 #endif
1232 
1233 static void
1234 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1235 {
1236 	if (err)
1237 		return;
1238 	*(bus_addr_t *) arg = segs[0].ds_addr;
1239 }
1240 
1241 int
1242 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1243 {
1244 	int err;
1245 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1246 	device_t dev = ctx->ifc_dev;
1247 
1248 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1249 
1250 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1251 				sctx->isc_q_align, 0,	/* alignment, bounds */
1252 				BUS_SPACE_MAXADDR,	/* lowaddr */
1253 				BUS_SPACE_MAXADDR,	/* highaddr */
1254 				NULL, NULL,		/* filter, filterarg */
1255 				size,			/* maxsize */
1256 				1,			/* nsegments */
1257 				size,			/* maxsegsize */
1258 				BUS_DMA_ALLOCNOW,	/* flags */
1259 				NULL,			/* lockfunc */
1260 				NULL,			/* lockarg */
1261 				&dma->idi_tag);
1262 	if (err) {
1263 		device_printf(dev,
1264 		    "%s: bus_dma_tag_create failed: %d\n",
1265 		    __func__, err);
1266 		goto fail_0;
1267 	}
1268 
1269 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1270 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1271 	if (err) {
1272 		device_printf(dev,
1273 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1274 		    __func__, (uintmax_t)size, err);
1275 		goto fail_1;
1276 	}
1277 
1278 	dma->idi_paddr = IF_BAD_DMA;
1279 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1280 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1281 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1282 		device_printf(dev,
1283 		    "%s: bus_dmamap_load failed: %d\n",
1284 		    __func__, err);
1285 		goto fail_2;
1286 	}
1287 
1288 	dma->idi_size = size;
1289 	return (0);
1290 
1291 fail_2:
1292 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1293 fail_1:
1294 	bus_dma_tag_destroy(dma->idi_tag);
1295 fail_0:
1296 	dma->idi_tag = NULL;
1297 
1298 	return (err);
1299 }
1300 
1301 int
1302 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1303 {
1304 	int i, err;
1305 	iflib_dma_info_t *dmaiter;
1306 
1307 	dmaiter = dmalist;
1308 	for (i = 0; i < count; i++, dmaiter++) {
1309 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1310 			break;
1311 	}
1312 	if (err)
1313 		iflib_dma_free_multi(dmalist, i);
1314 	return (err);
1315 }
1316 
1317 void
1318 iflib_dma_free(iflib_dma_info_t dma)
1319 {
1320 	if (dma->idi_tag == NULL)
1321 		return;
1322 	if (dma->idi_paddr != IF_BAD_DMA) {
1323 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1324 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1325 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1326 		dma->idi_paddr = IF_BAD_DMA;
1327 	}
1328 	if (dma->idi_vaddr != NULL) {
1329 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1330 		dma->idi_vaddr = NULL;
1331 	}
1332 	bus_dma_tag_destroy(dma->idi_tag);
1333 	dma->idi_tag = NULL;
1334 }
1335 
1336 void
1337 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1338 {
1339 	int i;
1340 	iflib_dma_info_t *dmaiter = dmalist;
1341 
1342 	for (i = 0; i < count; i++, dmaiter++)
1343 		iflib_dma_free(*dmaiter);
1344 }
1345 
1346 #ifdef EARLY_AP_STARTUP
1347 static const int iflib_started = 1;
1348 #else
1349 /*
1350  * We used to abuse the smp_started flag to decide if the queues have been
1351  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1352  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1353  * is set.  Run a SYSINIT() strictly after that to just set a usable
1354  * completion flag.
1355  */
1356 
1357 static int iflib_started;
1358 
1359 static void
1360 iflib_record_started(void *arg)
1361 {
1362 	iflib_started = 1;
1363 }
1364 
1365 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1366 	iflib_record_started, NULL);
1367 #endif
1368 
1369 static int
1370 iflib_fast_intr(void *arg)
1371 {
1372 	iflib_filter_info_t info = arg;
1373 	struct grouptask *gtask = info->ifi_task;
1374 	if (!iflib_started)
1375 		return (FILTER_HANDLED);
1376 
1377 	DBG_COUNTER_INC(fast_intrs);
1378 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1379 		return (FILTER_HANDLED);
1380 
1381 	GROUPTASK_ENQUEUE(gtask);
1382 	return (FILTER_HANDLED);
1383 }
1384 
1385 static int
1386 iflib_fast_intr_rxtx(void *arg)
1387 {
1388 	iflib_filter_info_t info = arg;
1389 	struct grouptask *gtask = info->ifi_task;
1390 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1391 	if_ctx_t ctx;
1392 	int i, cidx;
1393 
1394 	if (!iflib_started)
1395 		return (FILTER_HANDLED);
1396 
1397 	DBG_COUNTER_INC(fast_intrs);
1398 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1399 		return (FILTER_HANDLED);
1400 
1401 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1402 		qidx_t txqid = rxq->ifr_txqid[i];
1403 
1404 		ctx = rxq->ifr_ctx;
1405 
1406 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1407 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1408 			continue;
1409 		}
1410 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1411 	}
1412 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1413 		cidx = rxq->ifr_cq_cidx;
1414 	else
1415 		cidx = rxq->ifr_fl[0].ifl_cidx;
1416 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1417 		GROUPTASK_ENQUEUE(gtask);
1418 	else
1419 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1420 	return (FILTER_HANDLED);
1421 }
1422 
1423 
1424 static int
1425 iflib_fast_intr_ctx(void *arg)
1426 {
1427 	iflib_filter_info_t info = arg;
1428 	struct grouptask *gtask = info->ifi_task;
1429 
1430 	if (!iflib_started)
1431 		return (FILTER_HANDLED);
1432 
1433 	DBG_COUNTER_INC(fast_intrs);
1434 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1435 		return (FILTER_HANDLED);
1436 
1437 	GROUPTASK_ENQUEUE(gtask);
1438 	return (FILTER_HANDLED);
1439 }
1440 
1441 static int
1442 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1443 	driver_filter_t filter, driver_intr_t handler, void *arg,
1444 				 char *name)
1445 {
1446 	int rc, flags;
1447 	struct resource *res;
1448 	void *tag = NULL;
1449 	device_t dev = ctx->ifc_dev;
1450 
1451 	flags = RF_ACTIVE;
1452 	if (ctx->ifc_flags & IFC_LEGACY)
1453 		flags |= RF_SHAREABLE;
1454 	MPASS(rid < 512);
1455 	irq->ii_rid = rid;
1456 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1457 	if (res == NULL) {
1458 		device_printf(dev,
1459 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1460 		return (ENOMEM);
1461 	}
1462 	irq->ii_res = res;
1463 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1464 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1465 						filter, handler, arg, &tag);
1466 	if (rc != 0) {
1467 		device_printf(dev,
1468 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1469 					  rid, name ? name : "unknown", rc);
1470 		return (rc);
1471 	} else if (name)
1472 		bus_describe_intr(dev, res, tag, "%s", name);
1473 
1474 	irq->ii_tag = tag;
1475 	return (0);
1476 }
1477 
1478 
1479 /*********************************************************************
1480  *
1481  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1482  *  the information needed to transmit a packet on the wire. This is
1483  *  called only once at attach, setup is done every reset.
1484  *
1485  **********************************************************************/
1486 
1487 static int
1488 iflib_txsd_alloc(iflib_txq_t txq)
1489 {
1490 	if_ctx_t ctx = txq->ift_ctx;
1491 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1492 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1493 	device_t dev = ctx->ifc_dev;
1494 	int err, nsegments, ntsosegments;
1495 
1496 	nsegments = scctx->isc_tx_nsegments;
1497 	ntsosegments = scctx->isc_tx_tso_segments_max;
1498 	MPASS(scctx->isc_ntxd[0] > 0);
1499 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1500 	MPASS(nsegments > 0);
1501 	MPASS(ntsosegments > 0);
1502 	/*
1503 	 * Setup DMA descriptor areas.
1504 	 */
1505 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1506 			       1, 0,			/* alignment, bounds */
1507 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1508 			       BUS_SPACE_MAXADDR,	/* highaddr */
1509 			       NULL, NULL,		/* filter, filterarg */
1510 			       sctx->isc_tx_maxsize,		/* maxsize */
1511 			       nsegments,	/* nsegments */
1512 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1513 			       0,			/* flags */
1514 			       NULL,			/* lockfunc */
1515 			       NULL,			/* lockfuncarg */
1516 			       &txq->ift_desc_tag))) {
1517 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1518 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1519 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1520 		goto fail;
1521 	}
1522 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1523 			       1, 0,			/* alignment, bounds */
1524 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1525 			       BUS_SPACE_MAXADDR,	/* highaddr */
1526 			       NULL, NULL,		/* filter, filterarg */
1527 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1528 			       ntsosegments,	/* nsegments */
1529 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1530 			       0,			/* flags */
1531 			       NULL,			/* lockfunc */
1532 			       NULL,			/* lockfuncarg */
1533 			       &txq->ift_tso_desc_tag))) {
1534 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1535 
1536 		goto fail;
1537 	}
1538 	if (!(txq->ift_sds.ifsd_flags =
1539 	    (uint8_t *) malloc(sizeof(uint8_t) *
1540 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1541 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1542 		err = ENOMEM;
1543 		goto fail;
1544 	}
1545 	if (!(txq->ift_sds.ifsd_m =
1546 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1547 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1548 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1549 		err = ENOMEM;
1550 		goto fail;
1551 	}
1552 
1553         /* Create the descriptor buffer dma maps */
1554 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1555 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1556 		return (0);
1557 
1558 	if (!(txq->ift_sds.ifsd_map =
1559 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1560 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1561 		err = ENOMEM;
1562 		goto fail;
1563 	}
1564 
1565 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1566 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1567 		if (err != 0) {
1568 			device_printf(dev, "Unable to create TX DMA map\n");
1569 			goto fail;
1570 		}
1571 	}
1572 #endif
1573 	return (0);
1574 fail:
1575 	/* We free all, it handles case where we are in the middle */
1576 	iflib_tx_structures_free(ctx);
1577 	return (err);
1578 }
1579 
1580 static void
1581 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1582 {
1583 	bus_dmamap_t map;
1584 
1585 	map = NULL;
1586 	if (txq->ift_sds.ifsd_map != NULL)
1587 		map = txq->ift_sds.ifsd_map[i];
1588 	if (map != NULL) {
1589 		bus_dmamap_unload(txq->ift_desc_tag, map);
1590 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1591 		txq->ift_sds.ifsd_map[i] = NULL;
1592 	}
1593 }
1594 
1595 static void
1596 iflib_txq_destroy(iflib_txq_t txq)
1597 {
1598 	if_ctx_t ctx = txq->ift_ctx;
1599 
1600 	for (int i = 0; i < txq->ift_size; i++)
1601 		iflib_txsd_destroy(ctx, txq, i);
1602 	if (txq->ift_sds.ifsd_map != NULL) {
1603 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1604 		txq->ift_sds.ifsd_map = NULL;
1605 	}
1606 	if (txq->ift_sds.ifsd_m != NULL) {
1607 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1608 		txq->ift_sds.ifsd_m = NULL;
1609 	}
1610 	if (txq->ift_sds.ifsd_flags != NULL) {
1611 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1612 		txq->ift_sds.ifsd_flags = NULL;
1613 	}
1614 	if (txq->ift_desc_tag != NULL) {
1615 		bus_dma_tag_destroy(txq->ift_desc_tag);
1616 		txq->ift_desc_tag = NULL;
1617 	}
1618 	if (txq->ift_tso_desc_tag != NULL) {
1619 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1620 		txq->ift_tso_desc_tag = NULL;
1621 	}
1622 }
1623 
1624 static void
1625 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1626 {
1627 	struct mbuf **mp;
1628 
1629 	mp = &txq->ift_sds.ifsd_m[i];
1630 	if (*mp == NULL)
1631 		return;
1632 
1633 	if (txq->ift_sds.ifsd_map != NULL) {
1634 		bus_dmamap_sync(txq->ift_desc_tag,
1635 				txq->ift_sds.ifsd_map[i],
1636 				BUS_DMASYNC_POSTWRITE);
1637 		bus_dmamap_unload(txq->ift_desc_tag,
1638 				  txq->ift_sds.ifsd_map[i]);
1639 	}
1640 	m_free(*mp);
1641 	DBG_COUNTER_INC(tx_frees);
1642 	*mp = NULL;
1643 }
1644 
1645 static int
1646 iflib_txq_setup(iflib_txq_t txq)
1647 {
1648 	if_ctx_t ctx = txq->ift_ctx;
1649 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1650 	iflib_dma_info_t di;
1651 	int i;
1652 
1653 	/* Set number of descriptors available */
1654 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1655 	/* XXX make configurable */
1656 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1657 
1658 	/* Reset indices */
1659 	txq->ift_cidx_processed = 0;
1660 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1661 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1662 
1663 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1664 		bzero((void *)di->idi_vaddr, di->idi_size);
1665 
1666 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1667 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1668 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1669 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1670 	return (0);
1671 }
1672 
1673 /*********************************************************************
1674  *
1675  *  Allocate memory for rx_buffer structures. Since we use one
1676  *  rx_buffer per received packet, the maximum number of rx_buffer's
1677  *  that we'll need is equal to the number of receive descriptors
1678  *  that we've allocated.
1679  *
1680  **********************************************************************/
1681 static int
1682 iflib_rxsd_alloc(iflib_rxq_t rxq)
1683 {
1684 	if_ctx_t ctx = rxq->ifr_ctx;
1685 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1686 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1687 	device_t dev = ctx->ifc_dev;
1688 	iflib_fl_t fl;
1689 	int			err;
1690 
1691 	MPASS(scctx->isc_nrxd[0] > 0);
1692 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1693 
1694 	fl = rxq->ifr_fl;
1695 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1696 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1697 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1698 					 1, 0,			/* alignment, bounds */
1699 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1700 					 BUS_SPACE_MAXADDR,	/* highaddr */
1701 					 NULL, NULL,		/* filter, filterarg */
1702 					 sctx->isc_rx_maxsize,	/* maxsize */
1703 					 sctx->isc_rx_nsegments,	/* nsegments */
1704 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1705 					 0,			/* flags */
1706 					 NULL,			/* lockfunc */
1707 					 NULL,			/* lockarg */
1708 					 &fl->ifl_desc_tag);
1709 		if (err) {
1710 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1711 				__func__, err);
1712 			goto fail;
1713 		}
1714 		if (!(fl->ifl_sds.ifsd_flags =
1715 		      (uint8_t *) malloc(sizeof(uint8_t) *
1716 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1717 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1718 			err = ENOMEM;
1719 			goto fail;
1720 		}
1721 		if (!(fl->ifl_sds.ifsd_m =
1722 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1723 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1724 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1725 			err = ENOMEM;
1726 			goto fail;
1727 		}
1728 		if (!(fl->ifl_sds.ifsd_cl =
1729 		      (caddr_t *) malloc(sizeof(caddr_t) *
1730 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1731 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1732 			err = ENOMEM;
1733 			goto fail;
1734 		}
1735 
1736 		/* Create the descriptor buffer dma maps */
1737 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1738 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1739 			continue;
1740 
1741 		if (!(fl->ifl_sds.ifsd_map =
1742 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1743 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1744 			err = ENOMEM;
1745 			goto fail;
1746 		}
1747 
1748 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1749 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1750 			if (err != 0) {
1751 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1752 				goto fail;
1753 			}
1754 		}
1755 #endif
1756 	}
1757 	return (0);
1758 
1759 fail:
1760 	iflib_rx_structures_free(ctx);
1761 	return (err);
1762 }
1763 
1764 
1765 /*
1766  * Internal service routines
1767  */
1768 
1769 struct rxq_refill_cb_arg {
1770 	int               error;
1771 	bus_dma_segment_t seg;
1772 	int               nseg;
1773 };
1774 
1775 static void
1776 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1777 {
1778 	struct rxq_refill_cb_arg *cb_arg = arg;
1779 
1780 	cb_arg->error = error;
1781 	cb_arg->seg = segs[0];
1782 	cb_arg->nseg = nseg;
1783 }
1784 
1785 
1786 #ifdef ACPI_DMAR
1787 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1788 #else
1789 #define IS_DMAR(ctx) (0)
1790 #endif
1791 
1792 /**
1793  *	rxq_refill - refill an rxq  free-buffer list
1794  *	@ctx: the iflib context
1795  *	@rxq: the free-list to refill
1796  *	@n: the number of new buffers to allocate
1797  *
1798  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1799  *	The caller must assure that @n does not exceed the queue's capacity.
1800  */
1801 static void
1802 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1803 {
1804 	struct mbuf *m;
1805 	int idx, frag_idx = fl->ifl_fragidx;
1806         int pidx = fl->ifl_pidx;
1807 	caddr_t cl, *sd_cl;
1808 	struct mbuf **sd_m;
1809 	uint8_t *sd_flags;
1810 	struct if_rxd_update iru;
1811 	bus_dmamap_t *sd_map;
1812 	int n, i = 0;
1813 	uint64_t bus_addr;
1814 	int err;
1815 
1816 	sd_m = fl->ifl_sds.ifsd_m;
1817 	sd_map = fl->ifl_sds.ifsd_map;
1818 	sd_cl = fl->ifl_sds.ifsd_cl;
1819 	sd_flags = fl->ifl_sds.ifsd_flags;
1820 	idx = pidx;
1821 
1822 	n  = count;
1823 	MPASS(n > 0);
1824 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1825 
1826 	if (pidx < fl->ifl_cidx)
1827 		MPASS(pidx + n <= fl->ifl_cidx);
1828 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1829 		MPASS(fl->ifl_gen == 0);
1830 	if (pidx > fl->ifl_cidx)
1831 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1832 
1833 	DBG_COUNTER_INC(fl_refills);
1834 	if (n > 8)
1835 		DBG_COUNTER_INC(fl_refills_large);
1836 	iru.iru_paddrs = fl->ifl_bus_addrs;
1837 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1838 	iru.iru_idxs = fl->ifl_rxd_idxs;
1839 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1840 	iru.iru_buf_size = fl->ifl_buf_size;
1841 	iru.iru_flidx = fl->ifl_id;
1842 	while (n--) {
1843 		/*
1844 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1845 		 * initialized after rx.
1846 		 *
1847 		 * If the cluster is still set then we know a minimum sized packet was received
1848 		 */
1849 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1850 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1851                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1852 		if ((cl = sd_cl[frag_idx]) == NULL) {
1853                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1854 				break;
1855 #if MEMORY_LOGGING
1856 			fl->ifl_cl_enqueued++;
1857 #endif
1858 		}
1859 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1860 			break;
1861 		}
1862 #if MEMORY_LOGGING
1863 		fl->ifl_m_enqueued++;
1864 #endif
1865 
1866 		DBG_COUNTER_INC(rx_allocs);
1867 #if defined(__i386__) || defined(__amd64__)
1868 		if (!IS_DMAR(ctx)) {
1869 			bus_addr = pmap_kextract((vm_offset_t)cl);
1870 		} else
1871 #endif
1872 		{
1873 			struct rxq_refill_cb_arg cb_arg;
1874 			iflib_rxq_t q;
1875 
1876 			cb_arg.error = 0;
1877 			q = fl->ifl_rxq;
1878 			MPASS(sd_map != NULL);
1879 			MPASS(sd_map[frag_idx] != NULL);
1880 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1881 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1882 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1883 					BUS_DMASYNC_PREREAD);
1884 
1885 			if (err != 0 || cb_arg.error) {
1886 				/*
1887 				 * !zone_pack ?
1888 				 */
1889 				if (fl->ifl_zone == zone_pack)
1890 					uma_zfree(fl->ifl_zone, cl);
1891 				m_free(m);
1892 				n = 0;
1893 				goto done;
1894 			}
1895 			bus_addr = cb_arg.seg.ds_addr;
1896 		}
1897                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1898 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1899 
1900 		MPASS(sd_m[frag_idx] == NULL);
1901 		sd_cl[frag_idx] = cl;
1902 		sd_m[frag_idx] = m;
1903 		fl->ifl_rxd_idxs[i] = frag_idx;
1904 		fl->ifl_bus_addrs[i] = bus_addr;
1905 		fl->ifl_vm_addrs[i] = cl;
1906 		fl->ifl_credits++;
1907 		i++;
1908 		MPASS(fl->ifl_credits <= fl->ifl_size);
1909 		if (++idx == fl->ifl_size) {
1910 			fl->ifl_gen = 1;
1911 			idx = 0;
1912 		}
1913 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1914 			iru.iru_pidx = pidx;
1915 			iru.iru_count = i;
1916 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1917 			i = 0;
1918 			pidx = idx;
1919 			fl->ifl_pidx = idx;
1920 		}
1921 
1922 	}
1923 done:
1924 	DBG_COUNTER_INC(rxd_flush);
1925 	if (fl->ifl_pidx == 0)
1926 		pidx = fl->ifl_size - 1;
1927 	else
1928 		pidx = fl->ifl_pidx - 1;
1929 
1930 	if (sd_map)
1931 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1932 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1933 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1934 	fl->ifl_fragidx = frag_idx;
1935 }
1936 
1937 static __inline void
1938 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1939 {
1940 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1941 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1942 #ifdef INVARIANTS
1943 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1944 #endif
1945 
1946 	MPASS(fl->ifl_credits <= fl->ifl_size);
1947 	MPASS(reclaimable == delta);
1948 
1949 	if (reclaimable > 0)
1950 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1951 }
1952 
1953 static void
1954 iflib_fl_bufs_free(iflib_fl_t fl)
1955 {
1956 	iflib_dma_info_t idi = fl->ifl_ifdi;
1957 	uint32_t i;
1958 
1959 	for (i = 0; i < fl->ifl_size; i++) {
1960 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1961 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1962 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1963 
1964 		if (*sd_flags & RX_SW_DESC_INUSE) {
1965 			if (fl->ifl_sds.ifsd_map != NULL) {
1966 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1967 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1968 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1969 			}
1970 			if (*sd_m != NULL) {
1971 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1972 				uma_zfree(zone_mbuf, *sd_m);
1973 			}
1974 			if (*sd_cl != NULL)
1975 				uma_zfree(fl->ifl_zone, *sd_cl);
1976 			*sd_flags = 0;
1977 		} else {
1978 			MPASS(*sd_cl == NULL);
1979 			MPASS(*sd_m == NULL);
1980 		}
1981 #if MEMORY_LOGGING
1982 		fl->ifl_m_dequeued++;
1983 		fl->ifl_cl_dequeued++;
1984 #endif
1985 		*sd_cl = NULL;
1986 		*sd_m = NULL;
1987 	}
1988 #ifdef INVARIANTS
1989 	for (i = 0; i < fl->ifl_size; i++) {
1990 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1991 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1992 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1993 	}
1994 #endif
1995 	/*
1996 	 * Reset free list values
1997 	 */
1998 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
1999 	bzero(idi->idi_vaddr, idi->idi_size);
2000 }
2001 
2002 /*********************************************************************
2003  *
2004  *  Initialize a receive ring and its buffers.
2005  *
2006  **********************************************************************/
2007 static int
2008 iflib_fl_setup(iflib_fl_t fl)
2009 {
2010 	iflib_rxq_t rxq = fl->ifl_rxq;
2011 	if_ctx_t ctx = rxq->ifr_ctx;
2012 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2013 
2014 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size);
2015 	/*
2016 	** Free current RX buffer structs and their mbufs
2017 	*/
2018 	iflib_fl_bufs_free(fl);
2019 	/* Now replenish the mbufs */
2020 	MPASS(fl->ifl_credits == 0);
2021 	/*
2022 	 * XXX don't set the max_frame_size to larger
2023 	 * than the hardware can handle
2024 	 */
2025 	if (sctx->isc_max_frame_size <= 2048)
2026 		fl->ifl_buf_size = MCLBYTES;
2027 #ifndef CONTIGMALLOC_WORKS
2028 	else
2029 		fl->ifl_buf_size = MJUMPAGESIZE;
2030 #else
2031 	else if (sctx->isc_max_frame_size <= 4096)
2032 		fl->ifl_buf_size = MJUMPAGESIZE;
2033 	else if (sctx->isc_max_frame_size <= 9216)
2034 		fl->ifl_buf_size = MJUM9BYTES;
2035 	else
2036 		fl->ifl_buf_size = MJUM16BYTES;
2037 #endif
2038 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2039 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2040 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2041 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2042 
2043 
2044 	/* avoid pre-allocating zillions of clusters to an idle card
2045 	 * potentially speeding up attach
2046 	 */
2047 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2048 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2049 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2050 		return (ENOBUFS);
2051 	/*
2052 	 * handle failure
2053 	 */
2054 	MPASS(rxq != NULL);
2055 	MPASS(fl->ifl_ifdi != NULL);
2056 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2057 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2058 	return (0);
2059 }
2060 
2061 /*********************************************************************
2062  *
2063  *  Free receive ring data structures
2064  *
2065  **********************************************************************/
2066 static void
2067 iflib_rx_sds_free(iflib_rxq_t rxq)
2068 {
2069 	iflib_fl_t fl;
2070 	int i;
2071 
2072 	if (rxq->ifr_fl != NULL) {
2073 		for (i = 0; i < rxq->ifr_nfl; i++) {
2074 			fl = &rxq->ifr_fl[i];
2075 			if (fl->ifl_desc_tag != NULL) {
2076 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2077 				fl->ifl_desc_tag = NULL;
2078 			}
2079 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2080 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2081 			/* XXX destroy maps first */
2082 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2083 			fl->ifl_sds.ifsd_m = NULL;
2084 			fl->ifl_sds.ifsd_cl = NULL;
2085 			fl->ifl_sds.ifsd_map = NULL;
2086 		}
2087 		free(rxq->ifr_fl, M_IFLIB);
2088 		rxq->ifr_fl = NULL;
2089 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2090 	}
2091 }
2092 
2093 /*
2094  * MI independent logic
2095  *
2096  */
2097 static void
2098 iflib_timer(void *arg)
2099 {
2100 	iflib_txq_t txq = arg;
2101 	if_ctx_t ctx = txq->ift_ctx;
2102 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2103 
2104 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2105 		return;
2106 	/*
2107 	** Check on the state of the TX queue(s), this
2108 	** can be done without the lock because its RO
2109 	** and the HUNG state will be static if set.
2110 	*/
2111 	IFDI_TIMER(ctx, txq->ift_id);
2112 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2113 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2114 	     (sctx->isc_pause_frames == 0)))
2115 		goto hung;
2116 
2117 	if (ifmp_ring_is_stalled(txq->ift_br))
2118 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2119 	txq->ift_cleaned_prev = txq->ift_cleaned;
2120 	/* handle any laggards */
2121 	if (txq->ift_db_pending)
2122 		GROUPTASK_ENQUEUE(&txq->ift_task);
2123 
2124 	sctx->isc_pause_frames = 0;
2125 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2126 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2127 	return;
2128 hung:
2129 	CTX_LOCK(ctx);
2130 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2131 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2132 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2133 
2134 	IFDI_WATCHDOG_RESET(ctx);
2135 	ctx->ifc_watchdog_events++;
2136 
2137 	ctx->ifc_flags |= IFC_DO_RESET;
2138 	iflib_admin_intr_deferred(ctx);
2139 	CTX_UNLOCK(ctx);
2140 }
2141 
2142 static void
2143 iflib_init_locked(if_ctx_t ctx)
2144 {
2145 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2146 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2147 	if_t ifp = ctx->ifc_ifp;
2148 	iflib_fl_t fl;
2149 	iflib_txq_t txq;
2150 	iflib_rxq_t rxq;
2151 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2152 
2153 
2154 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2155 	IFDI_INTR_DISABLE(ctx);
2156 
2157 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2158 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2159 	/* Set hardware offload abilities */
2160 	if_clearhwassist(ifp);
2161 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2162 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2163 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2164 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2165 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2166 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2167 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2168 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2169 
2170 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2171 		CALLOUT_LOCK(txq);
2172 		callout_stop(&txq->ift_timer);
2173 		CALLOUT_UNLOCK(txq);
2174 		iflib_netmap_txq_init(ctx, txq);
2175 	}
2176 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2177 		MPASS(rxq->ifr_id == i);
2178 		iflib_netmap_rxq_init(ctx, rxq);
2179 	}
2180 #ifdef INVARIANTS
2181 	i = if_getdrvflags(ifp);
2182 #endif
2183 	IFDI_INIT(ctx);
2184 	MPASS(if_getdrvflags(ifp) == i);
2185 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2186 		/* XXX this should really be done on a per-queue basis */
2187 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2188 			continue;
2189 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2190 			if (iflib_fl_setup(fl)) {
2191 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2192 				goto done;
2193 			}
2194 		}
2195 	}
2196 	done:
2197 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2198 	IFDI_INTR_ENABLE(ctx);
2199 	txq = ctx->ifc_txqs;
2200 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2201 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2202 			txq->ift_timer.c_cpu);
2203 }
2204 
2205 static int
2206 iflib_media_change(if_t ifp)
2207 {
2208 	if_ctx_t ctx = if_getsoftc(ifp);
2209 	int err;
2210 
2211 	CTX_LOCK(ctx);
2212 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2213 		iflib_init_locked(ctx);
2214 	CTX_UNLOCK(ctx);
2215 	return (err);
2216 }
2217 
2218 static void
2219 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2220 {
2221 	if_ctx_t ctx = if_getsoftc(ifp);
2222 
2223 	CTX_LOCK(ctx);
2224 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2225 	IFDI_MEDIA_STATUS(ctx, ifmr);
2226 	CTX_UNLOCK(ctx);
2227 }
2228 
2229 static void
2230 iflib_stop(if_ctx_t ctx)
2231 {
2232 	iflib_txq_t txq = ctx->ifc_txqs;
2233 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2234 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2235 	iflib_dma_info_t di;
2236 	iflib_fl_t fl;
2237 	int i, j;
2238 
2239 	/* Tell the stack that the interface is no longer active */
2240 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2241 
2242 	IFDI_INTR_DISABLE(ctx);
2243 	DELAY(1000);
2244 	IFDI_STOP(ctx);
2245 	DELAY(1000);
2246 
2247 	iflib_debug_reset();
2248 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2249 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2250 		/* make sure all transmitters have completed before proceeding XXX */
2251 
2252 		/* clean any enqueued buffers */
2253 		iflib_ifmp_purge(txq);
2254 		/* Free any existing tx buffers. */
2255 		for (j = 0; j < txq->ift_size; j++) {
2256 			iflib_txsd_free(ctx, txq, j);
2257 		}
2258 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2259 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2260 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2261 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2262 		txq->ift_pullups = 0;
2263 		ifmp_ring_reset_stats(txq->ift_br);
2264 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2265 			bzero((void *)di->idi_vaddr, di->idi_size);
2266 	}
2267 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2268 		/* make sure all transmitters have completed before proceeding XXX */
2269 
2270 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2271 			bzero((void *)di->idi_vaddr, di->idi_size);
2272 		/* also resets the free lists pidx/cidx */
2273 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2274 			iflib_fl_bufs_free(fl);
2275 	}
2276 }
2277 
2278 static inline caddr_t
2279 calc_next_rxd(iflib_fl_t fl, int cidx)
2280 {
2281 	qidx_t size;
2282 	int nrxd;
2283 	caddr_t start, end, cur, next;
2284 
2285 	nrxd = fl->ifl_size;
2286 	size = fl->ifl_rxd_size;
2287 	start = fl->ifl_ifdi->idi_vaddr;
2288 
2289 	if (__predict_false(size == 0))
2290 		return (start);
2291 	cur = start + size*cidx;
2292 	end = start + size*nrxd;
2293 	next = CACHE_PTR_NEXT(cur);
2294 	return (next < end ? next : start);
2295 }
2296 
2297 static inline void
2298 prefetch_pkts(iflib_fl_t fl, int cidx)
2299 {
2300 	int nextptr;
2301 	int nrxd = fl->ifl_size;
2302 	caddr_t next_rxd;
2303 
2304 
2305 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2306 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2307 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2308 	next_rxd = calc_next_rxd(fl, cidx);
2309 	prefetch(next_rxd);
2310 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2311 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2312 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2313 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2314 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2315 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2316 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2317 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2318 }
2319 
2320 static void
2321 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2322 {
2323 	int flid, cidx;
2324 	bus_dmamap_t map;
2325 	iflib_fl_t fl;
2326 	iflib_dma_info_t di;
2327 	int next;
2328 
2329 	map = NULL;
2330 	flid = irf->irf_flid;
2331 	cidx = irf->irf_idx;
2332 	fl = &rxq->ifr_fl[flid];
2333 	sd->ifsd_fl = fl;
2334 	sd->ifsd_cidx = cidx;
2335 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2336 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2337 	fl->ifl_credits--;
2338 #if MEMORY_LOGGING
2339 	fl->ifl_m_dequeued++;
2340 #endif
2341 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2342 		prefetch_pkts(fl, cidx);
2343 	if (fl->ifl_sds.ifsd_map != NULL) {
2344 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2345 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2346 		map = fl->ifl_sds.ifsd_map[cidx];
2347 		di = fl->ifl_ifdi;
2348 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2349 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2350 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2351 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2352 
2353 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2354 		MPASS(fl->ifl_cidx == cidx);
2355 		if (unload)
2356 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2357 	}
2358 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2359 	if (__predict_false(fl->ifl_cidx == 0))
2360 		fl->ifl_gen = 0;
2361 	if (map != NULL)
2362 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2363 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2364         bit_clear(fl->ifl_rx_bitmap, cidx);
2365 }
2366 
2367 static struct mbuf *
2368 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2369 {
2370 	int i, padlen , flags;
2371 	struct mbuf *m, *mh, *mt;
2372 	caddr_t cl;
2373 
2374 	i = 0;
2375 	mh = NULL;
2376 	do {
2377 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2378 
2379 		MPASS(*sd->ifsd_cl != NULL);
2380 		MPASS(*sd->ifsd_m != NULL);
2381 
2382 		/* Don't include zero-length frags */
2383 		if (ri->iri_frags[i].irf_len == 0) {
2384 			/* XXX we can save the cluster here, but not the mbuf */
2385 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2386 			m_free(*sd->ifsd_m);
2387 			*sd->ifsd_m = NULL;
2388 			continue;
2389 		}
2390 		m = *sd->ifsd_m;
2391 		*sd->ifsd_m = NULL;
2392 		if (mh == NULL) {
2393 			flags = M_PKTHDR|M_EXT;
2394 			mh = mt = m;
2395 			padlen = ri->iri_pad;
2396 		} else {
2397 			flags = M_EXT;
2398 			mt->m_next = m;
2399 			mt = m;
2400 			/* assuming padding is only on the first fragment */
2401 			padlen = 0;
2402 		}
2403 		cl = *sd->ifsd_cl;
2404 		*sd->ifsd_cl = NULL;
2405 
2406 		/* Can these two be made one ? */
2407 		m_init(m, M_NOWAIT, MT_DATA, flags);
2408 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2409 		/*
2410 		 * These must follow m_init and m_cljset
2411 		 */
2412 		m->m_data += padlen;
2413 		ri->iri_len -= padlen;
2414 		m->m_len = ri->iri_frags[i].irf_len;
2415 	} while (++i < ri->iri_nfrags);
2416 
2417 	return (mh);
2418 }
2419 
2420 /*
2421  * Process one software descriptor
2422  */
2423 static struct mbuf *
2424 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2425 {
2426 	struct if_rxsd sd;
2427 	struct mbuf *m;
2428 
2429 	/* should I merge this back in now that the two paths are basically duplicated? */
2430 	if (ri->iri_nfrags == 1 &&
2431 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2432 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2433 		m = *sd.ifsd_m;
2434 		*sd.ifsd_m = NULL;
2435 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2436 #ifndef __NO_STRICT_ALIGNMENT
2437 		if (!IP_ALIGNED(m))
2438 			m->m_data += 2;
2439 #endif
2440 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2441 		m->m_len = ri->iri_frags[0].irf_len;
2442        } else {
2443 		m = assemble_segments(rxq, ri, &sd);
2444 	}
2445 	m->m_pkthdr.len = ri->iri_len;
2446 	m->m_pkthdr.rcvif = ri->iri_ifp;
2447 	m->m_flags |= ri->iri_flags;
2448 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2449 	m->m_pkthdr.flowid = ri->iri_flowid;
2450 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2451 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2452 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2453 	return (m);
2454 }
2455 
2456 static bool
2457 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2458 {
2459 	if_ctx_t ctx = rxq->ifr_ctx;
2460 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2461 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2462 	int avail, i;
2463 	qidx_t *cidxp;
2464 	struct if_rxd_info ri;
2465 	int err, budget_left, rx_bytes, rx_pkts;
2466 	iflib_fl_t fl;
2467 	struct ifnet *ifp;
2468 	int lro_enabled;
2469 
2470 	/*
2471 	 * XXX early demux data packets so that if_input processing only handles
2472 	 * acks in interrupt context
2473 	 */
2474 	struct mbuf *m, *mh, *mt;
2475 
2476 	ifp = ctx->ifc_ifp;
2477 #ifdef DEV_NETMAP
2478 	if (ifp->if_capenable & IFCAP_NETMAP) {
2479 		u_int work = 0;
2480 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2481 			return (FALSE);
2482 	}
2483 #endif
2484 
2485 	mh = mt = NULL;
2486 	MPASS(budget > 0);
2487 	rx_pkts	= rx_bytes = 0;
2488 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2489 		cidxp = &rxq->ifr_cq_cidx;
2490 	else
2491 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2492 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2493 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2494 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2495 		DBG_COUNTER_INC(rx_unavail);
2496 		return (false);
2497 	}
2498 
2499 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2500 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2501 			DBG_COUNTER_INC(rx_ctx_inactive);
2502 			break;
2503 		}
2504 		/*
2505 		 * Reset client set fields to their default values
2506 		 */
2507 		rxd_info_zero(&ri);
2508 		ri.iri_qsidx = rxq->ifr_id;
2509 		ri.iri_cidx = *cidxp;
2510 		ri.iri_ifp = ifp;
2511 		ri.iri_frags = rxq->ifr_frags;
2512 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2513 
2514 		if (err)
2515 			goto err;
2516 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2517 			*cidxp = ri.iri_cidx;
2518 			/* Update our consumer index */
2519 			/* XXX NB: shurd - check if this is still safe */
2520 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2521 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2522 				rxq->ifr_cq_gen = 0;
2523 			}
2524 			/* was this only a completion queue message? */
2525 			if (__predict_false(ri.iri_nfrags == 0))
2526 				continue;
2527 		}
2528 		MPASS(ri.iri_nfrags != 0);
2529 		MPASS(ri.iri_len != 0);
2530 
2531 		/* will advance the cidx on the corresponding free lists */
2532 		m = iflib_rxd_pkt_get(rxq, &ri);
2533 		if (avail == 0 && budget_left)
2534 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2535 
2536 		if (__predict_false(m == NULL)) {
2537 			DBG_COUNTER_INC(rx_mbuf_null);
2538 			continue;
2539 		}
2540 		/* imm_pkt: -- cxgb */
2541 		if (mh == NULL)
2542 			mh = mt = m;
2543 		else {
2544 			mt->m_nextpkt = m;
2545 			mt = m;
2546 		}
2547 	}
2548 	/* make sure that we can refill faster than drain */
2549 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2550 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2551 
2552 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2553 	while (mh != NULL) {
2554 		m = mh;
2555 		mh = mh->m_nextpkt;
2556 		m->m_nextpkt = NULL;
2557 #ifndef __NO_STRICT_ALIGNMENT
2558 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2559 			continue;
2560 #endif
2561 		rx_bytes += m->m_pkthdr.len;
2562 		rx_pkts++;
2563 #if defined(INET6) || defined(INET)
2564 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2565 			continue;
2566 #endif
2567 		DBG_COUNTER_INC(rx_if_input);
2568 		ifp->if_input(ifp, m);
2569 	}
2570 
2571 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2572 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2573 
2574 	/*
2575 	 * Flush any outstanding LRO work
2576 	 */
2577 #if defined(INET6) || defined(INET)
2578 	tcp_lro_flush_all(&rxq->ifr_lc);
2579 #endif
2580 	if (avail)
2581 		return true;
2582 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2583 err:
2584 	CTX_LOCK(ctx);
2585 	ctx->ifc_flags |= IFC_DO_RESET;
2586 	iflib_admin_intr_deferred(ctx);
2587 	CTX_UNLOCK(ctx);
2588 	return (false);
2589 }
2590 
2591 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2592 static inline qidx_t
2593 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2594 {
2595 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2596 	qidx_t minthresh = txq->ift_size / 8;
2597 	if (in_use > 4*minthresh)
2598 		return (notify_count);
2599 	if (in_use > 2*minthresh)
2600 		return (notify_count >> 1);
2601 	if (in_use > minthresh)
2602 		return (notify_count >> 3);
2603 	return (0);
2604 }
2605 
2606 static inline qidx_t
2607 txq_max_rs_deferred(iflib_txq_t txq)
2608 {
2609 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2610 	qidx_t minthresh = txq->ift_size / 8;
2611 	if (txq->ift_in_use > 4*minthresh)
2612 		return (notify_count);
2613 	if (txq->ift_in_use > 2*minthresh)
2614 		return (notify_count >> 1);
2615 	if (txq->ift_in_use > minthresh)
2616 		return (notify_count >> 2);
2617 	return (2);
2618 }
2619 
2620 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2621 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2622 
2623 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2624 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2625 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2626 
2627 /* forward compatibility for cxgb */
2628 #define FIRST_QSET(ctx) 0
2629 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2630 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2631 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2632 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2633 
2634 /* XXX we should be setting this to something other than zero */
2635 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2636 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2637 
2638 static inline bool
2639 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2640 {
2641 	qidx_t dbval, max;
2642 	bool rang;
2643 
2644 	rang = false;
2645 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2646 	if (ring || txq->ift_db_pending >= max) {
2647 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2648 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2649 		txq->ift_db_pending = txq->ift_npending = 0;
2650 		rang = true;
2651 	}
2652 	return (rang);
2653 }
2654 
2655 #ifdef PKT_DEBUG
2656 static void
2657 print_pkt(if_pkt_info_t pi)
2658 {
2659 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2660 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2661 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2662 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2663 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2664 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2665 }
2666 #endif
2667 
2668 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2669 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2670 
2671 static int
2672 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2673 {
2674 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2675 	struct ether_vlan_header *eh;
2676 	struct mbuf *m, *n;
2677 
2678 	n = m = *mp;
2679 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2680 	    M_WRITABLE(m) == 0) {
2681 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2682 			return (ENOMEM);
2683 		} else {
2684 			m_freem(*mp);
2685 			n = *mp = m;
2686 		}
2687 	}
2688 
2689 	/*
2690 	 * Determine where frame payload starts.
2691 	 * Jump over vlan headers if already present,
2692 	 * helpful for QinQ too.
2693 	 */
2694 	if (__predict_false(m->m_len < sizeof(*eh))) {
2695 		txq->ift_pullups++;
2696 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2697 			return (ENOMEM);
2698 	}
2699 	eh = mtod(m, struct ether_vlan_header *);
2700 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2701 		pi->ipi_etype = ntohs(eh->evl_proto);
2702 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2703 	} else {
2704 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2705 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2706 	}
2707 
2708 	switch (pi->ipi_etype) {
2709 #ifdef INET
2710 	case ETHERTYPE_IP:
2711 	{
2712 		struct ip *ip = NULL;
2713 		struct tcphdr *th = NULL;
2714 		int minthlen;
2715 
2716 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2717 		if (__predict_false(m->m_len < minthlen)) {
2718 			/*
2719 			 * if this code bloat is causing too much of a hit
2720 			 * move it to a separate function and mark it noinline
2721 			 */
2722 			if (m->m_len == pi->ipi_ehdrlen) {
2723 				n = m->m_next;
2724 				MPASS(n);
2725 				if (n->m_len >= sizeof(*ip))  {
2726 					ip = (struct ip *)n->m_data;
2727 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2728 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2729 				} else {
2730 					txq->ift_pullups++;
2731 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2732 						return (ENOMEM);
2733 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2734 				}
2735 			} else {
2736 				txq->ift_pullups++;
2737 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2738 					return (ENOMEM);
2739 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2740 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2741 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2742 			}
2743 		} else {
2744 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2745 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2746 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2747 		}
2748 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2749 		pi->ipi_ipproto = ip->ip_p;
2750 		pi->ipi_flags |= IPI_TX_IPV4;
2751 
2752 		if (pi->ipi_csum_flags & CSUM_IP)
2753                        ip->ip_sum = 0;
2754 
2755 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2756 			if (__predict_false(th == NULL)) {
2757 				txq->ift_pullups++;
2758 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2759 					return (ENOMEM);
2760 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2761 			}
2762 			pi->ipi_tcp_hflags = th->th_flags;
2763 			pi->ipi_tcp_hlen = th->th_off << 2;
2764 			pi->ipi_tcp_seq = th->th_seq;
2765 		}
2766 		if (IS_TSO4(pi)) {
2767 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2768 				return (ENXIO);
2769 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2770 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2771 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2772 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2773 				ip->ip_sum = 0;
2774 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2775 			}
2776 		}
2777 		break;
2778 	}
2779 #endif
2780 #ifdef INET6
2781 	case ETHERTYPE_IPV6:
2782 	{
2783 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2784 		struct tcphdr *th;
2785 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2786 
2787 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2788 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2789 				return (ENOMEM);
2790 		}
2791 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2792 
2793 		/* XXX-BZ this will go badly in case of ext hdrs. */
2794 		pi->ipi_ipproto = ip6->ip6_nxt;
2795 		pi->ipi_flags |= IPI_TX_IPV6;
2796 
2797 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2798 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2799 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2800 					return (ENOMEM);
2801 			}
2802 			pi->ipi_tcp_hflags = th->th_flags;
2803 			pi->ipi_tcp_hlen = th->th_off << 2;
2804 		}
2805 		if (IS_TSO6(pi)) {
2806 
2807 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2808 				return (ENXIO);
2809 			/*
2810 			 * The corresponding flag is set by the stack in the IPv4
2811 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2812 			 * So, set it here because the rest of the flow requires it.
2813 			 */
2814 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2815 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2816 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2817 		}
2818 		break;
2819 	}
2820 #endif
2821 	default:
2822 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2823 		pi->ipi_ip_hlen = 0;
2824 		break;
2825 	}
2826 	*mp = m;
2827 
2828 	return (0);
2829 }
2830 
2831 static  __noinline  struct mbuf *
2832 collapse_pkthdr(struct mbuf *m0)
2833 {
2834 	struct mbuf *m, *m_next, *tmp;
2835 
2836 	m = m0;
2837 	m_next = m->m_next;
2838 	while (m_next != NULL && m_next->m_len == 0) {
2839 		m = m_next;
2840 		m->m_next = NULL;
2841 		m_free(m);
2842 		m_next = m_next->m_next;
2843 	}
2844 	m = m0;
2845 	m->m_next = m_next;
2846 	if ((m_next->m_flags & M_EXT) == 0) {
2847 		m = m_defrag(m, M_NOWAIT);
2848 	} else {
2849 		tmp = m_next->m_next;
2850 		memcpy(m_next, m, MPKTHSIZE);
2851 		m = m_next;
2852 		m->m_next = tmp;
2853 	}
2854 	return (m);
2855 }
2856 
2857 /*
2858  * If dodgy hardware rejects the scatter gather chain we've handed it
2859  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2860  * m_defrag'd mbufs
2861  */
2862 static __noinline struct mbuf *
2863 iflib_remove_mbuf(iflib_txq_t txq)
2864 {
2865 	int ntxd, i, pidx;
2866 	struct mbuf *m, *mh, **ifsd_m;
2867 
2868 	pidx = txq->ift_pidx;
2869 	ifsd_m = txq->ift_sds.ifsd_m;
2870 	ntxd = txq->ift_size;
2871 	mh = m = ifsd_m[pidx];
2872 	ifsd_m[pidx] = NULL;
2873 #if MEMORY_LOGGING
2874 	txq->ift_dequeued++;
2875 #endif
2876 	i = 1;
2877 
2878 	while (m) {
2879 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2880 #if MEMORY_LOGGING
2881 		txq->ift_dequeued++;
2882 #endif
2883 		m = m->m_next;
2884 		i++;
2885 	}
2886 	return (mh);
2887 }
2888 
2889 static int
2890 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2891 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2892 			  int max_segs, int flags)
2893 {
2894 	if_ctx_t ctx;
2895 	if_shared_ctx_t		sctx;
2896 	if_softc_ctx_t		scctx;
2897 	int i, next, pidx, err, ntxd, count;
2898 	struct mbuf *m, *tmp, **ifsd_m;
2899 
2900 	m = *m0;
2901 
2902 	/*
2903 	 * Please don't ever do this
2904 	 */
2905 	if (__predict_false(m->m_len == 0))
2906 		*m0 = m = collapse_pkthdr(m);
2907 
2908 	ctx = txq->ift_ctx;
2909 	sctx = ctx->ifc_sctx;
2910 	scctx = &ctx->ifc_softc_ctx;
2911 	ifsd_m = txq->ift_sds.ifsd_m;
2912 	ntxd = txq->ift_size;
2913 	pidx = txq->ift_pidx;
2914 	if (map != NULL) {
2915 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2916 
2917 		err = bus_dmamap_load_mbuf_sg(tag, map,
2918 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2919 		if (err)
2920 			return (err);
2921 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2922 		count = 0;
2923 		m = *m0;
2924 		do {
2925 			if (__predict_false(m->m_len <= 0)) {
2926 				tmp = m;
2927 				m = m->m_next;
2928 				tmp->m_next = NULL;
2929 				m_free(tmp);
2930 				continue;
2931 			}
2932 			m = m->m_next;
2933 			count++;
2934 		} while (m != NULL);
2935 		if (count > *nsegs) {
2936 			ifsd_m[pidx] = *m0;
2937 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
2938 			return (0);
2939 		}
2940 		m = *m0;
2941 		count = 0;
2942 		do {
2943 			next = (pidx + count) & (ntxd-1);
2944 			MPASS(ifsd_m[next] == NULL);
2945 			ifsd_m[next] = m;
2946 			count++;
2947 			tmp = m;
2948 			m = m->m_next;
2949 		} while (m != NULL);
2950 	} else {
2951 		int buflen, sgsize, maxsegsz, max_sgsize;
2952 		vm_offset_t vaddr;
2953 		vm_paddr_t curaddr;
2954 
2955 		count = i = 0;
2956 		m = *m0;
2957 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
2958 			maxsegsz = scctx->isc_tx_tso_segsize_max;
2959 		else
2960 			maxsegsz = sctx->isc_tx_maxsegsize;
2961 
2962 		do {
2963 			if (__predict_false(m->m_len <= 0)) {
2964 				tmp = m;
2965 				m = m->m_next;
2966 				tmp->m_next = NULL;
2967 				m_free(tmp);
2968 				continue;
2969 			}
2970 			buflen = m->m_len;
2971 			vaddr = (vm_offset_t)m->m_data;
2972 			/*
2973 			 * see if we can't be smarter about physically
2974 			 * contiguous mappings
2975 			 */
2976 			next = (pidx + count) & (ntxd-1);
2977 			MPASS(ifsd_m[next] == NULL);
2978 #if MEMORY_LOGGING
2979 			txq->ift_enqueued++;
2980 #endif
2981 			ifsd_m[next] = m;
2982 			while (buflen > 0) {
2983 				if (i >= max_segs)
2984 					goto err;
2985 				max_sgsize = MIN(buflen, maxsegsz);
2986 				curaddr = pmap_kextract(vaddr);
2987 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2988 				sgsize = MIN(sgsize, max_sgsize);
2989 				segs[i].ds_addr = curaddr;
2990 				segs[i].ds_len = sgsize;
2991 				vaddr += sgsize;
2992 				buflen -= sgsize;
2993 				i++;
2994 			}
2995 			count++;
2996 			tmp = m;
2997 			m = m->m_next;
2998 		} while (m != NULL);
2999 		*nsegs = i;
3000 	}
3001 	return (0);
3002 err:
3003 	*m0 = iflib_remove_mbuf(txq);
3004 	return (EFBIG);
3005 }
3006 
3007 static inline caddr_t
3008 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3009 {
3010 	qidx_t size;
3011 	int ntxd;
3012 	caddr_t start, end, cur, next;
3013 
3014 	ntxd = txq->ift_size;
3015 	size = txq->ift_txd_size[qid];
3016 	start = txq->ift_ifdi[qid].idi_vaddr;
3017 
3018 	if (__predict_false(size == 0))
3019 		return (start);
3020 	cur = start + size*cidx;
3021 	end = start + size*ntxd;
3022 	next = CACHE_PTR_NEXT(cur);
3023 	return (next < end ? next : start);
3024 }
3025 
3026 static int
3027 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3028 {
3029 	if_ctx_t		ctx;
3030 	if_shared_ctx_t		sctx;
3031 	if_softc_ctx_t		scctx;
3032 	bus_dma_segment_t	*segs;
3033 	struct mbuf		*m_head;
3034 	void			*next_txd;
3035 	bus_dmamap_t		map;
3036 	struct if_pkt_info	pi;
3037 	int remap = 0;
3038 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3039 	bus_dma_tag_t desc_tag;
3040 
3041 	segs = txq->ift_segs;
3042 	ctx = txq->ift_ctx;
3043 	sctx = ctx->ifc_sctx;
3044 	scctx = &ctx->ifc_softc_ctx;
3045 	segs = txq->ift_segs;
3046 	ntxd = txq->ift_size;
3047 	m_head = *m_headp;
3048 	map = NULL;
3049 
3050 	/*
3051 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3052 	 */
3053 	cidx = txq->ift_cidx;
3054 	pidx = txq->ift_pidx;
3055 	if (ctx->ifc_flags & IFC_PREFETCH) {
3056 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3057 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3058 			next_txd = calc_next_txd(txq, cidx, 0);
3059 			prefetch(next_txd);
3060 		}
3061 
3062 		/* prefetch the next cache line of mbuf pointers and flags */
3063 		prefetch(&txq->ift_sds.ifsd_m[next]);
3064 		if (txq->ift_sds.ifsd_map != NULL) {
3065 			prefetch(&txq->ift_sds.ifsd_map[next]);
3066 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3067 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3068 		}
3069 	} else if (txq->ift_sds.ifsd_map != NULL)
3070 		map = txq->ift_sds.ifsd_map[pidx];
3071 
3072 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3073 		desc_tag = txq->ift_tso_desc_tag;
3074 		max_segs = scctx->isc_tx_tso_segments_max;
3075 	} else {
3076 		desc_tag = txq->ift_desc_tag;
3077 		max_segs = scctx->isc_tx_nsegments;
3078 	}
3079 	m_head = *m_headp;
3080 
3081 	pkt_info_zero(&pi);
3082 	pi.ipi_len = m_head->m_pkthdr.len;
3083 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3084 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3085 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3086 	pi.ipi_pidx = pidx;
3087 	pi.ipi_qsidx = txq->ift_id;
3088 
3089 	/* deliberate bitwise OR to make one condition */
3090 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3091 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3092 			return (err);
3093 		m_head = *m_headp;
3094 	}
3095 
3096 retry:
3097 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3098 defrag:
3099 	if (__predict_false(err)) {
3100 		switch (err) {
3101 		case EFBIG:
3102 			/* try collapse once and defrag once */
3103 			if (remap == 0)
3104 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3105 			if (remap == 1)
3106 				m_head = m_defrag(*m_headp, M_NOWAIT);
3107 			remap++;
3108 			if (__predict_false(m_head == NULL))
3109 				goto defrag_failed;
3110 			txq->ift_mbuf_defrag++;
3111 			*m_headp = m_head;
3112 			goto retry;
3113 			break;
3114 		case ENOMEM:
3115 			txq->ift_no_tx_dma_setup++;
3116 			break;
3117 		default:
3118 			txq->ift_no_tx_dma_setup++;
3119 			m_freem(*m_headp);
3120 			DBG_COUNTER_INC(tx_frees);
3121 			*m_headp = NULL;
3122 			break;
3123 		}
3124 		txq->ift_map_failed++;
3125 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3126 		return (err);
3127 	}
3128 
3129 	/*
3130 	 * XXX assumes a 1 to 1 relationship between segments and
3131 	 *        descriptors - this does not hold true on all drivers, e.g.
3132 	 *        cxgb
3133 	 */
3134 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3135 		txq->ift_no_desc_avail++;
3136 		if (map != NULL)
3137 			bus_dmamap_unload(desc_tag, map);
3138 		DBG_COUNTER_INC(encap_txq_avail_fail);
3139 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3140 			GROUPTASK_ENQUEUE(&txq->ift_task);
3141 		return (ENOBUFS);
3142 	}
3143 	/*
3144 	 * On Intel cards we can greatly reduce the number of TX interrupts
3145 	 * we see by only setting report status on every Nth descriptor.
3146 	 * However, this also means that the driver will need to keep track
3147 	 * of the descriptors that RS was set on to check them for the DD bit.
3148 	 */
3149 	txq->ift_rs_pending += nsegs + 1;
3150 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3151 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3152 		pi.ipi_flags |= IPI_TX_INTR;
3153 		txq->ift_rs_pending = 0;
3154 	}
3155 
3156 	pi.ipi_segs = segs;
3157 	pi.ipi_nsegs = nsegs;
3158 
3159 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3160 #ifdef PKT_DEBUG
3161 	print_pkt(&pi);
3162 #endif
3163 	if (map != NULL)
3164 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3165 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3166 		if (map != NULL)
3167 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3168 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3169 		DBG_COUNTER_INC(tx_encap);
3170 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3171 
3172 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3173 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3174 			ndesc += txq->ift_size;
3175 			txq->ift_gen = 1;
3176 		}
3177 		/*
3178 		 * drivers can need as many as
3179 		 * two sentinels
3180 		 */
3181 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3182 		MPASS(pi.ipi_new_pidx != pidx);
3183 		MPASS(ndesc > 0);
3184 		txq->ift_in_use += ndesc;
3185 
3186 		/*
3187 		 * We update the last software descriptor again here because there may
3188 		 * be a sentinel and/or there may be more mbufs than segments
3189 		 */
3190 		txq->ift_pidx = pi.ipi_new_pidx;
3191 		txq->ift_npending += pi.ipi_ndescs;
3192 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3193 		*m_headp = m_head = iflib_remove_mbuf(txq);
3194 		remap = 1;
3195 		txq->ift_txd_encap_efbig++;
3196 		goto defrag;
3197 	} else
3198 		DBG_COUNTER_INC(encap_txd_encap_fail);
3199 	return (err);
3200 
3201 defrag_failed:
3202 	txq->ift_mbuf_defrag_failed++;
3203 	txq->ift_map_failed++;
3204 	m_freem(*m_headp);
3205 	DBG_COUNTER_INC(tx_frees);
3206 	*m_headp = NULL;
3207 	return (ENOMEM);
3208 }
3209 
3210 static void
3211 iflib_tx_desc_free(iflib_txq_t txq, int n)
3212 {
3213 	int hasmap;
3214 	uint32_t qsize, cidx, mask, gen;
3215 	struct mbuf *m, **ifsd_m;
3216 	uint8_t *ifsd_flags;
3217 	bus_dmamap_t *ifsd_map;
3218 	bool do_prefetch;
3219 
3220 	cidx = txq->ift_cidx;
3221 	gen = txq->ift_gen;
3222 	qsize = txq->ift_size;
3223 	mask = qsize-1;
3224 	hasmap = txq->ift_sds.ifsd_map != NULL;
3225 	ifsd_flags = txq->ift_sds.ifsd_flags;
3226 	ifsd_m = txq->ift_sds.ifsd_m;
3227 	ifsd_map = txq->ift_sds.ifsd_map;
3228 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3229 
3230 	while (n--) {
3231 		if (do_prefetch) {
3232 			prefetch(ifsd_m[(cidx + 3) & mask]);
3233 			prefetch(ifsd_m[(cidx + 4) & mask]);
3234 		}
3235 		if (ifsd_m[cidx] != NULL) {
3236 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3237 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3238 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3239 				/*
3240 				 * does it matter if it's not the TSO tag? If so we'll
3241 				 * have to add the type to flags
3242 				 */
3243 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3244 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3245 			}
3246 			if ((m = ifsd_m[cidx]) != NULL) {
3247 				/* XXX we don't support any drivers that batch packets yet */
3248 				MPASS(m->m_nextpkt == NULL);
3249 				/* if the number of clusters exceeds the number of segments
3250 				 * there won't be space on the ring to save a pointer to each
3251 				 * cluster so we simply free the list here
3252 				 */
3253 				if (m->m_flags & M_TOOBIG) {
3254 					m_freem(m);
3255 				} else {
3256 					m_free(m);
3257 				}
3258 				ifsd_m[cidx] = NULL;
3259 #if MEMORY_LOGGING
3260 				txq->ift_dequeued++;
3261 #endif
3262 				DBG_COUNTER_INC(tx_frees);
3263 			}
3264 		}
3265 		if (__predict_false(++cidx == qsize)) {
3266 			cidx = 0;
3267 			gen = 0;
3268 		}
3269 	}
3270 	txq->ift_cidx = cidx;
3271 	txq->ift_gen = gen;
3272 }
3273 
3274 static __inline int
3275 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3276 {
3277 	int reclaim;
3278 	if_ctx_t ctx = txq->ift_ctx;
3279 
3280 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3281 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3282 
3283 	/*
3284 	 * Need a rate-limiting check so that this isn't called every time
3285 	 */
3286 	iflib_tx_credits_update(ctx, txq);
3287 	reclaim = DESC_RECLAIMABLE(txq);
3288 
3289 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3290 #ifdef INVARIANTS
3291 		if (iflib_verbose_debug) {
3292 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3293 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3294 			       reclaim, thresh);
3295 
3296 		}
3297 #endif
3298 		return (0);
3299 	}
3300 	iflib_tx_desc_free(txq, reclaim);
3301 	txq->ift_cleaned += reclaim;
3302 	txq->ift_in_use -= reclaim;
3303 
3304 	return (reclaim);
3305 }
3306 
3307 static struct mbuf **
3308 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3309 {
3310 	int next, size;
3311 	struct mbuf **items;
3312 
3313 	size = r->size;
3314 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3315 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3316 
3317 	prefetch(items[(cidx + offset) & (size-1)]);
3318 	if (remaining > 1) {
3319 		prefetch(&items[next]);
3320 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3321 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3322 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3323 	}
3324 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3325 }
3326 
3327 static void
3328 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3329 {
3330 
3331 	ifmp_ring_check_drainage(txq->ift_br, budget);
3332 }
3333 
3334 static uint32_t
3335 iflib_txq_can_drain(struct ifmp_ring *r)
3336 {
3337 	iflib_txq_t txq = r->cookie;
3338 	if_ctx_t ctx = txq->ift_ctx;
3339 
3340 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3341 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3342 }
3343 
3344 static uint32_t
3345 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3346 {
3347 	iflib_txq_t txq = r->cookie;
3348 	if_ctx_t ctx = txq->ift_ctx;
3349 	struct ifnet *ifp = ctx->ifc_ifp;
3350 	struct mbuf **mp, *m;
3351 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3352 	int reclaimed, err, in_use_prev, desc_used;
3353 	bool do_prefetch, ring, rang;
3354 
3355 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3356 			    !LINK_ACTIVE(ctx))) {
3357 		DBG_COUNTER_INC(txq_drain_notready);
3358 		return (0);
3359 	}
3360 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3361 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3362 	avail = IDXDIFF(pidx, cidx, r->size);
3363 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3364 		DBG_COUNTER_INC(txq_drain_flushing);
3365 		for (i = 0; i < avail; i++) {
3366 			m_free(r->items[(cidx + i) & (r->size-1)]);
3367 			r->items[(cidx + i) & (r->size-1)] = NULL;
3368 		}
3369 		return (avail);
3370 	}
3371 
3372 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3373 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3374 		CALLOUT_LOCK(txq);
3375 		callout_stop(&txq->ift_timer);
3376 		CALLOUT_UNLOCK(txq);
3377 		DBG_COUNTER_INC(txq_drain_oactive);
3378 		return (0);
3379 	}
3380 	if (reclaimed)
3381 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3382 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3383 	count = MIN(avail, TX_BATCH_SIZE);
3384 #ifdef INVARIANTS
3385 	if (iflib_verbose_debug)
3386 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3387 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3388 #endif
3389 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3390 	avail = TXQ_AVAIL(txq);
3391 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3392 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3393 
3394 		mp = _ring_peek_one(r, cidx, i, rem);
3395 		MPASS(mp != NULL && *mp != NULL);
3396 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3397 			consumed++;
3398 			reclaimed++;
3399 			continue;
3400 		}
3401 		in_use_prev = txq->ift_in_use;
3402 		pidx_prev = txq->ift_pidx;
3403 		err = iflib_encap(txq, mp);
3404 		if (__predict_false(err)) {
3405 			DBG_COUNTER_INC(txq_drain_encapfail);
3406 			/* no room - bail out */
3407 			if (err == ENOBUFS)
3408 				break;
3409 			consumed++;
3410 			DBG_COUNTER_INC(txq_drain_encapfail);
3411 			/* we can't send this packet - skip it */
3412 			continue;
3413 		}
3414 		consumed++;
3415 		pkt_sent++;
3416 		m = *mp;
3417 		DBG_COUNTER_INC(tx_sent);
3418 		bytes_sent += m->m_pkthdr.len;
3419 		mcast_sent += !!(m->m_flags & M_MCAST);
3420 		avail = TXQ_AVAIL(txq);
3421 
3422 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3423 		desc_used += (txq->ift_in_use - in_use_prev);
3424 		ETHER_BPF_MTAP(ifp, m);
3425 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3426 			break;
3427 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3428 	}
3429 
3430 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3431 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3432 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3433 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3434 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3435 	if (mcast_sent)
3436 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3437 #ifdef INVARIANTS
3438 	if (iflib_verbose_debug)
3439 		printf("consumed=%d\n", consumed);
3440 #endif
3441 	return (consumed);
3442 }
3443 
3444 static uint32_t
3445 iflib_txq_drain_always(struct ifmp_ring *r)
3446 {
3447 	return (1);
3448 }
3449 
3450 static uint32_t
3451 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3452 {
3453 	int i, avail;
3454 	struct mbuf **mp;
3455 	iflib_txq_t txq;
3456 
3457 	txq = r->cookie;
3458 
3459 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3460 	CALLOUT_LOCK(txq);
3461 	callout_stop(&txq->ift_timer);
3462 	CALLOUT_UNLOCK(txq);
3463 
3464 	avail = IDXDIFF(pidx, cidx, r->size);
3465 	for (i = 0; i < avail; i++) {
3466 		mp = _ring_peek_one(r, cidx, i, avail - i);
3467 		if (__predict_false(*mp == (struct mbuf *)txq))
3468 			continue;
3469 		m_freem(*mp);
3470 	}
3471 	MPASS(ifmp_ring_is_stalled(r) == 0);
3472 	return (avail);
3473 }
3474 
3475 static void
3476 iflib_ifmp_purge(iflib_txq_t txq)
3477 {
3478 	struct ifmp_ring *r;
3479 
3480 	r = txq->ift_br;
3481 	r->drain = iflib_txq_drain_free;
3482 	r->can_drain = iflib_txq_drain_always;
3483 
3484 	ifmp_ring_check_drainage(r, r->size);
3485 
3486 	r->drain = iflib_txq_drain;
3487 	r->can_drain = iflib_txq_can_drain;
3488 }
3489 
3490 static void
3491 _task_fn_tx(void *context)
3492 {
3493 	iflib_txq_t txq = context;
3494 	if_ctx_t ctx = txq->ift_ctx;
3495 	struct ifnet *ifp = ctx->ifc_ifp;
3496 	int rc;
3497 
3498 #ifdef IFLIB_DIAGNOSTICS
3499 	txq->ift_cpu_exec_count[curcpu]++;
3500 #endif
3501 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3502 		return;
3503 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3504 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3505 			netmap_tx_irq(ifp, txq->ift_id);
3506 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3507 		return;
3508 	}
3509 	if (txq->ift_db_pending)
3510 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3511 	else
3512 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3513 	if (ctx->ifc_flags & IFC_LEGACY)
3514 		IFDI_INTR_ENABLE(ctx);
3515 	else {
3516 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3517 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3518 	}
3519 }
3520 
3521 static void
3522 _task_fn_rx(void *context)
3523 {
3524 	iflib_rxq_t rxq = context;
3525 	if_ctx_t ctx = rxq->ifr_ctx;
3526 	bool more;
3527 	int rc;
3528 
3529 #ifdef IFLIB_DIAGNOSTICS
3530 	rxq->ifr_cpu_exec_count[curcpu]++;
3531 #endif
3532 	DBG_COUNTER_INC(task_fn_rxs);
3533 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3534 		return;
3535 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3536 		if (ctx->ifc_flags & IFC_LEGACY)
3537 			IFDI_INTR_ENABLE(ctx);
3538 		else {
3539 			DBG_COUNTER_INC(rx_intr_enables);
3540 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3541 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3542 		}
3543 	}
3544 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3545 		return;
3546 	if (more)
3547 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3548 }
3549 
3550 static void
3551 _task_fn_admin(void *context)
3552 {
3553 	if_ctx_t ctx = context;
3554 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3555 	iflib_txq_t txq;
3556 	int i;
3557 
3558 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3559 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3560 			return;
3561 		}
3562 	}
3563 
3564 	CTX_LOCK(ctx);
3565 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3566 		CALLOUT_LOCK(txq);
3567 		callout_stop(&txq->ift_timer);
3568 		CALLOUT_UNLOCK(txq);
3569 	}
3570 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3571 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3572 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3573 	IFDI_LINK_INTR_ENABLE(ctx);
3574 	if (ctx->ifc_flags & IFC_DO_RESET) {
3575 		ctx->ifc_flags &= ~IFC_DO_RESET;
3576 		iflib_if_init_locked(ctx);
3577 	}
3578 	CTX_UNLOCK(ctx);
3579 
3580 	if (LINK_ACTIVE(ctx) == 0)
3581 		return;
3582 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3583 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3584 }
3585 
3586 
3587 static void
3588 _task_fn_iov(void *context)
3589 {
3590 	if_ctx_t ctx = context;
3591 
3592 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3593 		return;
3594 
3595 	CTX_LOCK(ctx);
3596 	IFDI_VFLR_HANDLE(ctx);
3597 	CTX_UNLOCK(ctx);
3598 }
3599 
3600 static int
3601 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3602 {
3603 	int err;
3604 	if_int_delay_info_t info;
3605 	if_ctx_t ctx;
3606 
3607 	info = (if_int_delay_info_t)arg1;
3608 	ctx = info->iidi_ctx;
3609 	info->iidi_req = req;
3610 	info->iidi_oidp = oidp;
3611 	CTX_LOCK(ctx);
3612 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3613 	CTX_UNLOCK(ctx);
3614 	return (err);
3615 }
3616 
3617 /*********************************************************************
3618  *
3619  *  IFNET FUNCTIONS
3620  *
3621  **********************************************************************/
3622 
3623 static void
3624 iflib_if_init_locked(if_ctx_t ctx)
3625 {
3626 	iflib_stop(ctx);
3627 	iflib_init_locked(ctx);
3628 }
3629 
3630 
3631 static void
3632 iflib_if_init(void *arg)
3633 {
3634 	if_ctx_t ctx = arg;
3635 
3636 	CTX_LOCK(ctx);
3637 	iflib_if_init_locked(ctx);
3638 	CTX_UNLOCK(ctx);
3639 }
3640 
3641 static int
3642 iflib_if_transmit(if_t ifp, struct mbuf *m)
3643 {
3644 	if_ctx_t	ctx = if_getsoftc(ifp);
3645 
3646 	iflib_txq_t txq;
3647 	int err, qidx;
3648 
3649 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3650 		DBG_COUNTER_INC(tx_frees);
3651 		m_freem(m);
3652 		return (ENOBUFS);
3653 	}
3654 
3655 	MPASS(m->m_nextpkt == NULL);
3656 	qidx = 0;
3657 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3658 		qidx = QIDX(ctx, m);
3659 	/*
3660 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3661 	 */
3662 	txq = &ctx->ifc_txqs[qidx];
3663 
3664 #ifdef DRIVER_BACKPRESSURE
3665 	if (txq->ift_closed) {
3666 		while (m != NULL) {
3667 			next = m->m_nextpkt;
3668 			m->m_nextpkt = NULL;
3669 			m_freem(m);
3670 			m = next;
3671 		}
3672 		return (ENOBUFS);
3673 	}
3674 #endif
3675 #ifdef notyet
3676 	qidx = count = 0;
3677 	mp = marr;
3678 	next = m;
3679 	do {
3680 		count++;
3681 		next = next->m_nextpkt;
3682 	} while (next != NULL);
3683 
3684 	if (count > nitems(marr))
3685 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3686 			/* XXX check nextpkt */
3687 			m_freem(m);
3688 			/* XXX simplify for now */
3689 			DBG_COUNTER_INC(tx_frees);
3690 			return (ENOBUFS);
3691 		}
3692 	for (next = m, i = 0; next != NULL; i++) {
3693 		mp[i] = next;
3694 		next = next->m_nextpkt;
3695 		mp[i]->m_nextpkt = NULL;
3696 	}
3697 #endif
3698 	DBG_COUNTER_INC(tx_seen);
3699 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3700 
3701 	if (err) {
3702 		GROUPTASK_ENQUEUE(&txq->ift_task);
3703 		/* support forthcoming later */
3704 #ifdef DRIVER_BACKPRESSURE
3705 		txq->ift_closed = TRUE;
3706 #endif
3707 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3708 		m_freem(m);
3709 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3710 		GROUPTASK_ENQUEUE(&txq->ift_task);
3711 	}
3712 
3713 	return (err);
3714 }
3715 
3716 static void
3717 iflib_if_qflush(if_t ifp)
3718 {
3719 	if_ctx_t ctx = if_getsoftc(ifp);
3720 	iflib_txq_t txq = ctx->ifc_txqs;
3721 	int i;
3722 
3723 	CTX_LOCK(ctx);
3724 	ctx->ifc_flags |= IFC_QFLUSH;
3725 	CTX_UNLOCK(ctx);
3726 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3727 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3728 			iflib_txq_check_drain(txq, 0);
3729 	CTX_LOCK(ctx);
3730 	ctx->ifc_flags &= ~IFC_QFLUSH;
3731 	CTX_UNLOCK(ctx);
3732 
3733 	if_qflush(ifp);
3734 }
3735 
3736 
3737 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3738 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3739 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3740 
3741 static int
3742 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3743 {
3744 	if_ctx_t ctx = if_getsoftc(ifp);
3745 	struct ifreq	*ifr = (struct ifreq *)data;
3746 #if defined(INET) || defined(INET6)
3747 	struct ifaddr	*ifa = (struct ifaddr *)data;
3748 #endif
3749 	bool		avoid_reset = FALSE;
3750 	int		err = 0, reinit = 0, bits;
3751 
3752 	switch (command) {
3753 	case SIOCSIFADDR:
3754 #ifdef INET
3755 		if (ifa->ifa_addr->sa_family == AF_INET)
3756 			avoid_reset = TRUE;
3757 #endif
3758 #ifdef INET6
3759 		if (ifa->ifa_addr->sa_family == AF_INET6)
3760 			avoid_reset = TRUE;
3761 #endif
3762 		/*
3763 		** Calling init results in link renegotiation,
3764 		** so we avoid doing it when possible.
3765 		*/
3766 		if (avoid_reset) {
3767 			if_setflagbits(ifp, IFF_UP,0);
3768 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3769 				reinit = 1;
3770 #ifdef INET
3771 			if (!(if_getflags(ifp) & IFF_NOARP))
3772 				arp_ifinit(ifp, ifa);
3773 #endif
3774 		} else
3775 			err = ether_ioctl(ifp, command, data);
3776 		break;
3777 	case SIOCSIFMTU:
3778 		CTX_LOCK(ctx);
3779 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3780 			CTX_UNLOCK(ctx);
3781 			break;
3782 		}
3783 		bits = if_getdrvflags(ifp);
3784 		/* stop the driver and free any clusters before proceeding */
3785 		iflib_stop(ctx);
3786 
3787 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3788 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3789 				ctx->ifc_flags |= IFC_MULTISEG;
3790 			else
3791 				ctx->ifc_flags &= ~IFC_MULTISEG;
3792 			err = if_setmtu(ifp, ifr->ifr_mtu);
3793 		}
3794 		iflib_init_locked(ctx);
3795 		if_setdrvflags(ifp, bits);
3796 		CTX_UNLOCK(ctx);
3797 		break;
3798 	case SIOCSIFFLAGS:
3799 		CTX_LOCK(ctx);
3800 		if (if_getflags(ifp) & IFF_UP) {
3801 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3802 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3803 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3804 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3805 				}
3806 			} else
3807 				reinit = 1;
3808 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3809 			iflib_stop(ctx);
3810 		}
3811 		ctx->ifc_if_flags = if_getflags(ifp);
3812 		CTX_UNLOCK(ctx);
3813 		break;
3814 	case SIOCADDMULTI:
3815 	case SIOCDELMULTI:
3816 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3817 			CTX_LOCK(ctx);
3818 			IFDI_INTR_DISABLE(ctx);
3819 			IFDI_MULTI_SET(ctx);
3820 			IFDI_INTR_ENABLE(ctx);
3821 			CTX_UNLOCK(ctx);
3822 		}
3823 		break;
3824 	case SIOCSIFMEDIA:
3825 		CTX_LOCK(ctx);
3826 		IFDI_MEDIA_SET(ctx);
3827 		CTX_UNLOCK(ctx);
3828 		/* falls thru */
3829 	case SIOCGIFMEDIA:
3830 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3831 		break;
3832 	case SIOCGI2C:
3833 	{
3834 		struct ifi2creq i2c;
3835 
3836 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3837 		if (err != 0)
3838 			break;
3839 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3840 			err = EINVAL;
3841 			break;
3842 		}
3843 		if (i2c.len > sizeof(i2c.data)) {
3844 			err = EINVAL;
3845 			break;
3846 		}
3847 
3848 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3849 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3850 		break;
3851 	}
3852 	case SIOCSIFCAP:
3853 	{
3854 		int mask, setmask;
3855 
3856 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3857 		setmask = 0;
3858 #ifdef TCP_OFFLOAD
3859 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3860 #endif
3861 		setmask |= (mask & IFCAP_FLAGS);
3862 
3863 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3864 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3865 		if ((mask & IFCAP_WOL) &&
3866 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3867 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3868 		if_vlancap(ifp);
3869 		/*
3870 		 * want to ensure that traffic has stopped before we change any of the flags
3871 		 */
3872 		if (setmask) {
3873 			CTX_LOCK(ctx);
3874 			bits = if_getdrvflags(ifp);
3875 			if (bits & IFF_DRV_RUNNING)
3876 				iflib_stop(ctx);
3877 			if_togglecapenable(ifp, setmask);
3878 			if (bits & IFF_DRV_RUNNING)
3879 				iflib_init_locked(ctx);
3880 			if_setdrvflags(ifp, bits);
3881 			CTX_UNLOCK(ctx);
3882 		}
3883 		break;
3884 	    }
3885 	case SIOCGPRIVATE_0:
3886 	case SIOCSDRVSPEC:
3887 	case SIOCGDRVSPEC:
3888 		CTX_LOCK(ctx);
3889 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3890 		CTX_UNLOCK(ctx);
3891 		break;
3892 	default:
3893 		err = ether_ioctl(ifp, command, data);
3894 		break;
3895 	}
3896 	if (reinit)
3897 		iflib_if_init(ctx);
3898 	return (err);
3899 }
3900 
3901 static uint64_t
3902 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3903 {
3904 	if_ctx_t ctx = if_getsoftc(ifp);
3905 
3906 	return (IFDI_GET_COUNTER(ctx, cnt));
3907 }
3908 
3909 /*********************************************************************
3910  *
3911  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3912  *
3913  **********************************************************************/
3914 
3915 static void
3916 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3917 {
3918 	if_ctx_t ctx = if_getsoftc(ifp);
3919 
3920 	if ((void *)ctx != arg)
3921 		return;
3922 
3923 	if ((vtag == 0) || (vtag > 4095))
3924 		return;
3925 
3926 	CTX_LOCK(ctx);
3927 	IFDI_VLAN_REGISTER(ctx, vtag);
3928 	/* Re-init to load the changes */
3929 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3930 		iflib_if_init_locked(ctx);
3931 	CTX_UNLOCK(ctx);
3932 }
3933 
3934 static void
3935 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3936 {
3937 	if_ctx_t ctx = if_getsoftc(ifp);
3938 
3939 	if ((void *)ctx != arg)
3940 		return;
3941 
3942 	if ((vtag == 0) || (vtag > 4095))
3943 		return;
3944 
3945 	CTX_LOCK(ctx);
3946 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3947 	/* Re-init to load the changes */
3948 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3949 		iflib_if_init_locked(ctx);
3950 	CTX_UNLOCK(ctx);
3951 }
3952 
3953 static void
3954 iflib_led_func(void *arg, int onoff)
3955 {
3956 	if_ctx_t ctx = arg;
3957 
3958 	CTX_LOCK(ctx);
3959 	IFDI_LED_FUNC(ctx, onoff);
3960 	CTX_UNLOCK(ctx);
3961 }
3962 
3963 /*********************************************************************
3964  *
3965  *  BUS FUNCTION DEFINITIONS
3966  *
3967  **********************************************************************/
3968 
3969 int
3970 iflib_device_probe(device_t dev)
3971 {
3972 	pci_vendor_info_t *ent;
3973 
3974 	uint16_t	pci_vendor_id, pci_device_id;
3975 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3976 	uint16_t	pci_rev_id;
3977 	if_shared_ctx_t sctx;
3978 
3979 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3980 		return (ENOTSUP);
3981 
3982 	pci_vendor_id = pci_get_vendor(dev);
3983 	pci_device_id = pci_get_device(dev);
3984 	pci_subvendor_id = pci_get_subvendor(dev);
3985 	pci_subdevice_id = pci_get_subdevice(dev);
3986 	pci_rev_id = pci_get_revid(dev);
3987 	if (sctx->isc_parse_devinfo != NULL)
3988 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3989 
3990 	ent = sctx->isc_vendor_info;
3991 	while (ent->pvi_vendor_id != 0) {
3992 		if (pci_vendor_id != ent->pvi_vendor_id) {
3993 			ent++;
3994 			continue;
3995 		}
3996 		if ((pci_device_id == ent->pvi_device_id) &&
3997 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3998 		     (ent->pvi_subvendor_id == 0)) &&
3999 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4000 		     (ent->pvi_subdevice_id == 0)) &&
4001 		    ((pci_rev_id == ent->pvi_rev_id) ||
4002 		     (ent->pvi_rev_id == 0))) {
4003 
4004 			device_set_desc_copy(dev, ent->pvi_name);
4005 			/* this needs to be changed to zero if the bus probing code
4006 			 * ever stops re-probing on best match because the sctx
4007 			 * may have its values over written by register calls
4008 			 * in subsequent probes
4009 			 */
4010 			return (BUS_PROBE_DEFAULT);
4011 		}
4012 		ent++;
4013 	}
4014 	return (ENXIO);
4015 }
4016 
4017 int
4018 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4019 {
4020 	int err, rid, msix, msix_bar;
4021 	if_ctx_t ctx;
4022 	if_t ifp;
4023 	if_softc_ctx_t scctx;
4024 	int i;
4025 	uint16_t main_txq;
4026 	uint16_t main_rxq;
4027 
4028 
4029 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4030 
4031 	if (sc == NULL) {
4032 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4033 		device_set_softc(dev, ctx);
4034 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4035 	}
4036 
4037 	ctx->ifc_sctx = sctx;
4038 	ctx->ifc_dev = dev;
4039 	ctx->ifc_softc = sc;
4040 
4041 	if ((err = iflib_register(ctx)) != 0) {
4042 		device_printf(dev, "iflib_register failed %d\n", err);
4043 		return (err);
4044 	}
4045 	iflib_add_device_sysctl_pre(ctx);
4046 
4047 	scctx = &ctx->ifc_softc_ctx;
4048 	ifp = ctx->ifc_ifp;
4049 
4050 	/*
4051 	 * XXX sanity check that ntxd & nrxd are a power of 2
4052 	 */
4053 	if (ctx->ifc_sysctl_ntxqs != 0)
4054 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4055 	if (ctx->ifc_sysctl_nrxqs != 0)
4056 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4057 
4058 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4059 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4060 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4061 		else
4062 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4063 	}
4064 
4065 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4066 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4067 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4068 		else
4069 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4070 	}
4071 
4072 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4073 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4074 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4075 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4076 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4077 		}
4078 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4079 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4080 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4081 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4082 		}
4083 	}
4084 
4085 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4086 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4087 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4088 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4089 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4090 		}
4091 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4092 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4093 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4094 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4095 		}
4096 	}
4097 
4098 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4099 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4100 		return (err);
4101 	}
4102 	_iflib_pre_assert(scctx);
4103 	ctx->ifc_txrx = *scctx->isc_txrx;
4104 
4105 #ifdef INVARIANTS
4106 	MPASS(scctx->isc_capenable);
4107 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4108 		MPASS(scctx->isc_tx_csum_flags);
4109 #endif
4110 
4111 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4112 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4113 
4114 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4115 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4116 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4117 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4118 
4119 #ifdef ACPI_DMAR
4120 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4121 		ctx->ifc_flags |= IFC_DMAR;
4122 #elif !(defined(__i386__) || defined(__amd64__))
4123 	/* set unconditionally for !x86 */
4124 	ctx->ifc_flags |= IFC_DMAR;
4125 #endif
4126 
4127 	msix_bar = scctx->isc_msix_bar;
4128 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4129 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4130 
4131 	/* XXX change for per-queue sizes */
4132 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4133 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4134 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4135 		if (!powerof2(scctx->isc_nrxd[i])) {
4136 			/* round down instead? */
4137 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4138 			err = EINVAL;
4139 			goto fail;
4140 		}
4141 	}
4142 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4143 		if (!powerof2(scctx->isc_ntxd[i])) {
4144 			device_printf(dev,
4145 			    "# tx descriptors must be a power of 2");
4146 			err = EINVAL;
4147 			goto fail;
4148 		}
4149 	}
4150 
4151 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4152 	    MAX_SINGLE_PACKET_FRACTION)
4153 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4154 		    MAX_SINGLE_PACKET_FRACTION);
4155 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4156 	    MAX_SINGLE_PACKET_FRACTION)
4157 		scctx->isc_tx_tso_segments_max = max(1,
4158 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4159 
4160 	/*
4161 	 * Protect the stack against modern hardware
4162 	 */
4163 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4164 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4165 
4166 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4167 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4168 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4169 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4170 	if (scctx->isc_rss_table_size == 0)
4171 		scctx->isc_rss_table_size = 64;
4172 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4173 
4174 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4175 	/* XXX format name */
4176 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4177 	/*
4178 	** Now setup MSI or MSI/X, should
4179 	** return us the number of supported
4180 	** vectors. (Will be 1 for MSI)
4181 	*/
4182 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4183 		msix = scctx->isc_vectors;
4184 	} else if (scctx->isc_msix_bar != 0)
4185 	       /*
4186 		* The simple fact that isc_msix_bar is not 0 does not mean we
4187 		* we have a good value there that is known to work.
4188 		*/
4189 		msix = iflib_msix_init(ctx);
4190 	else {
4191 		scctx->isc_vectors = 1;
4192 		scctx->isc_ntxqsets = 1;
4193 		scctx->isc_nrxqsets = 1;
4194 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4195 		msix = 0;
4196 	}
4197 	/* Get memory for the station queues */
4198 	if ((err = iflib_queues_alloc(ctx))) {
4199 		device_printf(dev, "Unable to allocate queue memory\n");
4200 		goto fail;
4201 	}
4202 
4203 	if ((err = iflib_qset_structures_setup(ctx))) {
4204 		device_printf(dev, "qset structure setup failed %d\n", err);
4205 		goto fail_queues;
4206 	}
4207 
4208 	/*
4209 	 * Group taskqueues aren't properly set up until SMP is started,
4210 	 * so we disable interrupts until we can handle them post
4211 	 * SI_SUB_SMP.
4212 	 *
4213 	 * XXX: disabling interrupts doesn't actually work, at least for
4214 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4215 	 * we do null handling and depend on this not causing too large an
4216 	 * interrupt storm.
4217 	 */
4218 	IFDI_INTR_DISABLE(ctx);
4219 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4220 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4221 		goto fail_intr_free;
4222 	}
4223 	if (msix <= 1) {
4224 		rid = 0;
4225 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4226 			MPASS(msix == 1);
4227 			rid = 1;
4228 		}
4229 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4230 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4231 			goto fail_intr_free;
4232 		}
4233 	}
4234 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4235 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4236 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4237 		goto fail_detach;
4238 	}
4239 	if ((err = iflib_netmap_attach(ctx))) {
4240 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4241 		goto fail_detach;
4242 	}
4243 	*ctxp = ctx;
4244 
4245 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4246 	iflib_add_device_sysctl_post(ctx);
4247 	ctx->ifc_flags |= IFC_INIT_DONE;
4248 	return (0);
4249 fail_detach:
4250 	ether_ifdetach(ctx->ifc_ifp);
4251 fail_intr_free:
4252 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4253 		pci_release_msi(ctx->ifc_dev);
4254 fail_queues:
4255 	/* XXX free queues */
4256 fail:
4257 	IFDI_DETACH(ctx);
4258 	return (err);
4259 }
4260 
4261 int
4262 iflib_device_attach(device_t dev)
4263 {
4264 	if_ctx_t ctx;
4265 	if_shared_ctx_t sctx;
4266 
4267 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4268 		return (ENOTSUP);
4269 
4270 	pci_enable_busmaster(dev);
4271 
4272 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4273 }
4274 
4275 int
4276 iflib_device_deregister(if_ctx_t ctx)
4277 {
4278 	if_t ifp = ctx->ifc_ifp;
4279 	iflib_txq_t txq;
4280 	iflib_rxq_t rxq;
4281 	device_t dev = ctx->ifc_dev;
4282 	int i, j;
4283 	struct taskqgroup *tqg;
4284 	iflib_fl_t fl;
4285 
4286 	/* Make sure VLANS are not using driver */
4287 	if (if_vlantrunkinuse(ifp)) {
4288 		device_printf(dev,"Vlan in use, detach first\n");
4289 		return (EBUSY);
4290 	}
4291 
4292 	CTX_LOCK(ctx);
4293 	ctx->ifc_in_detach = 1;
4294 	iflib_stop(ctx);
4295 	CTX_UNLOCK(ctx);
4296 
4297 	/* Unregister VLAN events */
4298 	if (ctx->ifc_vlan_attach_event != NULL)
4299 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4300 	if (ctx->ifc_vlan_detach_event != NULL)
4301 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4302 
4303 	iflib_netmap_detach(ifp);
4304 	ether_ifdetach(ifp);
4305 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4306 	CTX_LOCK_DESTROY(ctx);
4307 	if (ctx->ifc_led_dev != NULL)
4308 		led_destroy(ctx->ifc_led_dev);
4309 	/* XXX drain any dependent tasks */
4310 	tqg = qgroup_if_io_tqg;
4311 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4312 		callout_drain(&txq->ift_timer);
4313 		if (txq->ift_task.gt_uniq != NULL)
4314 			taskqgroup_detach(tqg, &txq->ift_task);
4315 	}
4316 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4317 		if (rxq->ifr_task.gt_uniq != NULL)
4318 			taskqgroup_detach(tqg, &rxq->ifr_task);
4319 
4320 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4321 			free(fl->ifl_rx_bitmap, M_IFLIB);
4322 
4323 	}
4324 	tqg = qgroup_if_config_tqg;
4325 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4326 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4327 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4328 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4329 
4330 	IFDI_DETACH(ctx);
4331 	device_set_softc(ctx->ifc_dev, NULL);
4332 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4333 		pci_release_msi(dev);
4334 	}
4335 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4336 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4337 	}
4338 	if (ctx->ifc_msix_mem != NULL) {
4339 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4340 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4341 		ctx->ifc_msix_mem = NULL;
4342 	}
4343 
4344 	bus_generic_detach(dev);
4345 	if_free(ifp);
4346 
4347 	iflib_tx_structures_free(ctx);
4348 	iflib_rx_structures_free(ctx);
4349 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4350 		free(ctx->ifc_softc, M_IFLIB);
4351 	free(ctx, M_IFLIB);
4352 	return (0);
4353 }
4354 
4355 
4356 int
4357 iflib_device_detach(device_t dev)
4358 {
4359 	if_ctx_t ctx = device_get_softc(dev);
4360 
4361 	return (iflib_device_deregister(ctx));
4362 }
4363 
4364 int
4365 iflib_device_suspend(device_t dev)
4366 {
4367 	if_ctx_t ctx = device_get_softc(dev);
4368 
4369 	CTX_LOCK(ctx);
4370 	IFDI_SUSPEND(ctx);
4371 	CTX_UNLOCK(ctx);
4372 
4373 	return bus_generic_suspend(dev);
4374 }
4375 int
4376 iflib_device_shutdown(device_t dev)
4377 {
4378 	if_ctx_t ctx = device_get_softc(dev);
4379 
4380 	CTX_LOCK(ctx);
4381 	IFDI_SHUTDOWN(ctx);
4382 	CTX_UNLOCK(ctx);
4383 
4384 	return bus_generic_suspend(dev);
4385 }
4386 
4387 
4388 int
4389 iflib_device_resume(device_t dev)
4390 {
4391 	if_ctx_t ctx = device_get_softc(dev);
4392 	iflib_txq_t txq = ctx->ifc_txqs;
4393 
4394 	CTX_LOCK(ctx);
4395 	IFDI_RESUME(ctx);
4396 	iflib_init_locked(ctx);
4397 	CTX_UNLOCK(ctx);
4398 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4399 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4400 
4401 	return (bus_generic_resume(dev));
4402 }
4403 
4404 int
4405 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4406 {
4407 	int error;
4408 	if_ctx_t ctx = device_get_softc(dev);
4409 
4410 	CTX_LOCK(ctx);
4411 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4412 	CTX_UNLOCK(ctx);
4413 
4414 	return (error);
4415 }
4416 
4417 void
4418 iflib_device_iov_uninit(device_t dev)
4419 {
4420 	if_ctx_t ctx = device_get_softc(dev);
4421 
4422 	CTX_LOCK(ctx);
4423 	IFDI_IOV_UNINIT(ctx);
4424 	CTX_UNLOCK(ctx);
4425 }
4426 
4427 int
4428 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, 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_VF_ADD(ctx, vfnum, params);
4435 	CTX_UNLOCK(ctx);
4436 
4437 	return (error);
4438 }
4439 
4440 /*********************************************************************
4441  *
4442  *  MODULE FUNCTION DEFINITIONS
4443  *
4444  **********************************************************************/
4445 
4446 /*
4447  * - Start a fast taskqueue thread for each core
4448  * - Start a taskqueue for control operations
4449  */
4450 static int
4451 iflib_module_init(void)
4452 {
4453 	return (0);
4454 }
4455 
4456 static int
4457 iflib_module_event_handler(module_t mod, int what, void *arg)
4458 {
4459 	int err;
4460 
4461 	switch (what) {
4462 	case MOD_LOAD:
4463 		if ((err = iflib_module_init()) != 0)
4464 			return (err);
4465 		break;
4466 	case MOD_UNLOAD:
4467 		return (EBUSY);
4468 	default:
4469 		return (EOPNOTSUPP);
4470 	}
4471 
4472 	return (0);
4473 }
4474 
4475 /*********************************************************************
4476  *
4477  *  PUBLIC FUNCTION DEFINITIONS
4478  *     ordered as in iflib.h
4479  *
4480  **********************************************************************/
4481 
4482 
4483 static void
4484 _iflib_assert(if_shared_ctx_t sctx)
4485 {
4486 	MPASS(sctx->isc_tx_maxsize);
4487 	MPASS(sctx->isc_tx_maxsegsize);
4488 
4489 	MPASS(sctx->isc_rx_maxsize);
4490 	MPASS(sctx->isc_rx_nsegments);
4491 	MPASS(sctx->isc_rx_maxsegsize);
4492 
4493 	MPASS(sctx->isc_nrxd_min[0]);
4494 	MPASS(sctx->isc_nrxd_max[0]);
4495 	MPASS(sctx->isc_nrxd_default[0]);
4496 	MPASS(sctx->isc_ntxd_min[0]);
4497 	MPASS(sctx->isc_ntxd_max[0]);
4498 	MPASS(sctx->isc_ntxd_default[0]);
4499 }
4500 
4501 static void
4502 _iflib_pre_assert(if_softc_ctx_t scctx)
4503 {
4504 
4505 	MPASS(scctx->isc_txrx->ift_txd_encap);
4506 	MPASS(scctx->isc_txrx->ift_txd_flush);
4507 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4508 	MPASS(scctx->isc_txrx->ift_rxd_available);
4509 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4510 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4511 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4512 }
4513 
4514 static int
4515 iflib_register(if_ctx_t ctx)
4516 {
4517 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4518 	driver_t *driver = sctx->isc_driver;
4519 	device_t dev = ctx->ifc_dev;
4520 	if_t ifp;
4521 
4522 	_iflib_assert(sctx);
4523 
4524 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4525 
4526 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4527 	if (ifp == NULL) {
4528 		device_printf(dev, "can not allocate ifnet structure\n");
4529 		return (ENOMEM);
4530 	}
4531 
4532 	/*
4533 	 * Initialize our context's device specific methods
4534 	 */
4535 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4536 	kobj_class_compile((kobj_class_t) driver);
4537 	driver->refs++;
4538 
4539 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4540 	if_setsoftc(ifp, ctx);
4541 	if_setdev(ifp, dev);
4542 	if_setinitfn(ifp, iflib_if_init);
4543 	if_setioctlfn(ifp, iflib_if_ioctl);
4544 	if_settransmitfn(ifp, iflib_if_transmit);
4545 	if_setqflushfn(ifp, iflib_if_qflush);
4546 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4547 
4548 	ctx->ifc_vlan_attach_event =
4549 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4550 							  EVENTHANDLER_PRI_FIRST);
4551 	ctx->ifc_vlan_detach_event =
4552 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4553 							  EVENTHANDLER_PRI_FIRST);
4554 
4555 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4556 					 iflib_media_change, iflib_media_status);
4557 
4558 	return (0);
4559 }
4560 
4561 
4562 static int
4563 iflib_queues_alloc(if_ctx_t ctx)
4564 {
4565 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4566 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4567 	device_t dev = ctx->ifc_dev;
4568 	int nrxqsets = scctx->isc_nrxqsets;
4569 	int ntxqsets = scctx->isc_ntxqsets;
4570 	iflib_txq_t txq;
4571 	iflib_rxq_t rxq;
4572 	iflib_fl_t fl = NULL;
4573 	int i, j, cpu, err, txconf, rxconf;
4574 	iflib_dma_info_t ifdip;
4575 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4576 	uint32_t *txqsizes = scctx->isc_txqsizes;
4577 	uint8_t nrxqs = sctx->isc_nrxqs;
4578 	uint8_t ntxqs = sctx->isc_ntxqs;
4579 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4580 	caddr_t *vaddrs;
4581 	uint64_t *paddrs;
4582 	struct ifmp_ring **brscp;
4583 
4584 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4585 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4586 
4587 	brscp = NULL;
4588 	txq = NULL;
4589 	rxq = NULL;
4590 
4591 /* Allocate the TX ring struct memory */
4592 	if (!(txq =
4593 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4594 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4595 		device_printf(dev, "Unable to allocate TX ring memory\n");
4596 		err = ENOMEM;
4597 		goto fail;
4598 	}
4599 
4600 	/* Now allocate the RX */
4601 	if (!(rxq =
4602 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4603 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4604 		device_printf(dev, "Unable to allocate RX ring memory\n");
4605 		err = ENOMEM;
4606 		goto rx_fail;
4607 	}
4608 
4609 	ctx->ifc_txqs = txq;
4610 	ctx->ifc_rxqs = rxq;
4611 
4612 	/*
4613 	 * XXX handle allocation failure
4614 	 */
4615 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4616 		/* Set up some basics */
4617 
4618 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4619 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4620 			err = ENOMEM;
4621 			goto err_tx_desc;
4622 		}
4623 		txq->ift_ifdi = ifdip;
4624 		for (j = 0; j < ntxqs; j++, ifdip++) {
4625 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4626 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4627 				err = ENOMEM;
4628 				goto err_tx_desc;
4629 			}
4630 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4631 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4632 		}
4633 		txq->ift_ctx = ctx;
4634 		txq->ift_id = i;
4635 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4636 			txq->ift_br_offset = 1;
4637 		} else {
4638 			txq->ift_br_offset = 0;
4639 		}
4640 		/* XXX fix this */
4641 		txq->ift_timer.c_cpu = cpu;
4642 
4643 		if (iflib_txsd_alloc(txq)) {
4644 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4645 			err = ENOMEM;
4646 			goto err_tx_desc;
4647 		}
4648 
4649 		/* Initialize the TX lock */
4650 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4651 		    device_get_nameunit(dev), txq->ift_id);
4652 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4653 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4654 
4655 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4656 			 device_get_nameunit(dev), txq->ift_id);
4657 
4658 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4659 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4660 		if (err) {
4661 			/* XXX free any allocated rings */
4662 			device_printf(dev, "Unable to allocate buf_ring\n");
4663 			goto err_tx_desc;
4664 		}
4665 	}
4666 
4667 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4668 		/* Set up some basics */
4669 
4670 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4671 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4672 			err = ENOMEM;
4673 			goto err_tx_desc;
4674 		}
4675 
4676 		rxq->ifr_ifdi = ifdip;
4677 		/* XXX this needs to be changed if #rx queues != #tx queues */
4678 		rxq->ifr_ntxqirq = 1;
4679 		rxq->ifr_txqid[0] = i;
4680 		for (j = 0; j < nrxqs; j++, ifdip++) {
4681 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4682 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4683 				err = ENOMEM;
4684 				goto err_tx_desc;
4685 			}
4686 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4687 		}
4688 		rxq->ifr_ctx = ctx;
4689 		rxq->ifr_id = i;
4690 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4691 			rxq->ifr_fl_offset = 1;
4692 		} else {
4693 			rxq->ifr_fl_offset = 0;
4694 		}
4695 		rxq->ifr_nfl = nfree_lists;
4696 		if (!(fl =
4697 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4698 			device_printf(dev, "Unable to allocate free list memory\n");
4699 			err = ENOMEM;
4700 			goto err_tx_desc;
4701 		}
4702 		rxq->ifr_fl = fl;
4703 		for (j = 0; j < nfree_lists; j++) {
4704 			fl[j].ifl_rxq = rxq;
4705 			fl[j].ifl_id = j;
4706 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4707 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4708 		}
4709         /* Allocate receive buffers for the ring*/
4710 		if (iflib_rxsd_alloc(rxq)) {
4711 			device_printf(dev,
4712 			    "Critical Failure setting up receive buffers\n");
4713 			err = ENOMEM;
4714 			goto err_rx_desc;
4715 		}
4716 
4717 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4718 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4719 	}
4720 
4721 	/* TXQs */
4722 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4723 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4724 	for (i = 0; i < ntxqsets; i++) {
4725 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4726 
4727 		for (j = 0; j < ntxqs; j++, di++) {
4728 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4729 			paddrs[i*ntxqs + j] = di->idi_paddr;
4730 		}
4731 	}
4732 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4733 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4734 		iflib_tx_structures_free(ctx);
4735 		free(vaddrs, M_IFLIB);
4736 		free(paddrs, M_IFLIB);
4737 		goto err_rx_desc;
4738 	}
4739 	free(vaddrs, M_IFLIB);
4740 	free(paddrs, M_IFLIB);
4741 
4742 	/* RXQs */
4743 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4744 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4745 	for (i = 0; i < nrxqsets; i++) {
4746 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4747 
4748 		for (j = 0; j < nrxqs; j++, di++) {
4749 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4750 			paddrs[i*nrxqs + j] = di->idi_paddr;
4751 		}
4752 	}
4753 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4754 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4755 		iflib_tx_structures_free(ctx);
4756 		free(vaddrs, M_IFLIB);
4757 		free(paddrs, M_IFLIB);
4758 		goto err_rx_desc;
4759 	}
4760 	free(vaddrs, M_IFLIB);
4761 	free(paddrs, M_IFLIB);
4762 
4763 	return (0);
4764 
4765 /* XXX handle allocation failure changes */
4766 err_rx_desc:
4767 err_tx_desc:
4768 	if (ctx->ifc_rxqs != NULL)
4769 		free(ctx->ifc_rxqs, M_IFLIB);
4770 	ctx->ifc_rxqs = NULL;
4771 	if (ctx->ifc_txqs != NULL)
4772 		free(ctx->ifc_txqs, M_IFLIB);
4773 	ctx->ifc_txqs = NULL;
4774 rx_fail:
4775 	if (brscp != NULL)
4776 		free(brscp, M_IFLIB);
4777 	if (rxq != NULL)
4778 		free(rxq, M_IFLIB);
4779 	if (txq != NULL)
4780 		free(txq, M_IFLIB);
4781 fail:
4782 	return (err);
4783 }
4784 
4785 static int
4786 iflib_tx_structures_setup(if_ctx_t ctx)
4787 {
4788 	iflib_txq_t txq = ctx->ifc_txqs;
4789 	int i;
4790 
4791 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4792 		iflib_txq_setup(txq);
4793 
4794 	return (0);
4795 }
4796 
4797 static void
4798 iflib_tx_structures_free(if_ctx_t ctx)
4799 {
4800 	iflib_txq_t txq = ctx->ifc_txqs;
4801 	int i, j;
4802 
4803 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4804 		iflib_txq_destroy(txq);
4805 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4806 			iflib_dma_free(&txq->ift_ifdi[j]);
4807 	}
4808 	free(ctx->ifc_txqs, M_IFLIB);
4809 	ctx->ifc_txqs = NULL;
4810 	IFDI_QUEUES_FREE(ctx);
4811 }
4812 
4813 /*********************************************************************
4814  *
4815  *  Initialize all receive rings.
4816  *
4817  **********************************************************************/
4818 static int
4819 iflib_rx_structures_setup(if_ctx_t ctx)
4820 {
4821 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4822 	int q;
4823 #if defined(INET6) || defined(INET)
4824 	int i, err;
4825 #endif
4826 
4827 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4828 #if defined(INET6) || defined(INET)
4829 		tcp_lro_free(&rxq->ifr_lc);
4830 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4831 		    TCP_LRO_ENTRIES, min(1024,
4832 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4833 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4834 			goto fail;
4835 		}
4836 		rxq->ifr_lro_enabled = TRUE;
4837 #endif
4838 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4839 	}
4840 	return (0);
4841 #if defined(INET6) || defined(INET)
4842 fail:
4843 	/*
4844 	 * Free RX software descriptors allocated so far, we will only handle
4845 	 * the rings that completed, the failing case will have
4846 	 * cleaned up for itself. 'q' failed, so its the terminus.
4847 	 */
4848 	rxq = ctx->ifc_rxqs;
4849 	for (i = 0; i < q; ++i, rxq++) {
4850 		iflib_rx_sds_free(rxq);
4851 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4852 	}
4853 	return (err);
4854 #endif
4855 }
4856 
4857 /*********************************************************************
4858  *
4859  *  Free all receive rings.
4860  *
4861  **********************************************************************/
4862 static void
4863 iflib_rx_structures_free(if_ctx_t ctx)
4864 {
4865 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4866 
4867 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4868 		iflib_rx_sds_free(rxq);
4869 	}
4870 }
4871 
4872 static int
4873 iflib_qset_structures_setup(if_ctx_t ctx)
4874 {
4875 	int err;
4876 
4877 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4878 		return (err);
4879 
4880 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4881 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4882 		iflib_tx_structures_free(ctx);
4883 		iflib_rx_structures_free(ctx);
4884 	}
4885 	return (err);
4886 }
4887 
4888 int
4889 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4890 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4891 {
4892 
4893 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4894 }
4895 
4896 static int
4897 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4898 {
4899 	int i, cpuid, eqid, count;
4900 
4901 	CPU_COPY(&ctx->ifc_cpus, cpus);
4902 	count = CPU_COUNT(&ctx->ifc_cpus);
4903 	eqid = qid % count;
4904 	/* clear up to the qid'th bit */
4905 	for (i = 0; i < eqid; i++) {
4906 		cpuid = CPU_FFS(cpus);
4907 		MPASS(cpuid != 0);
4908 		CPU_CLR(cpuid-1, cpus);
4909 	}
4910 	cpuid = CPU_FFS(cpus);
4911 	MPASS(cpuid != 0);
4912 	return (cpuid-1);
4913 }
4914 
4915 int
4916 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4917 						iflib_intr_type_t type, driver_filter_t *filter,
4918 						void *filter_arg, int qid, char *name)
4919 {
4920 	struct grouptask *gtask;
4921 	struct taskqgroup *tqg;
4922 	iflib_filter_info_t info;
4923 	cpuset_t cpus;
4924 	gtask_fn_t *fn;
4925 	int tqrid, err, cpuid;
4926 	driver_filter_t *intr_fast;
4927 	void *q;
4928 
4929 	info = &ctx->ifc_filter_info;
4930 	tqrid = rid;
4931 
4932 	switch (type) {
4933 	/* XXX merge tx/rx for netmap? */
4934 	case IFLIB_INTR_TX:
4935 		q = &ctx->ifc_txqs[qid];
4936 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4937 		gtask = &ctx->ifc_txqs[qid].ift_task;
4938 		tqg = qgroup_if_io_tqg;
4939 		fn = _task_fn_tx;
4940 		intr_fast = iflib_fast_intr;
4941 		GROUPTASK_INIT(gtask, 0, fn, q);
4942 		break;
4943 	case IFLIB_INTR_RX:
4944 		q = &ctx->ifc_rxqs[qid];
4945 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4946 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4947 		tqg = qgroup_if_io_tqg;
4948 		fn = _task_fn_rx;
4949 		intr_fast = iflib_fast_intr;
4950 		GROUPTASK_INIT(gtask, 0, fn, q);
4951 		break;
4952 	case IFLIB_INTR_RXTX:
4953 		q = &ctx->ifc_rxqs[qid];
4954 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4955 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4956 		tqg = qgroup_if_io_tqg;
4957 		fn = _task_fn_rx;
4958 		intr_fast = iflib_fast_intr_rxtx;
4959 		GROUPTASK_INIT(gtask, 0, fn, q);
4960 		break;
4961 	case IFLIB_INTR_ADMIN:
4962 		q = ctx;
4963 		tqrid = -1;
4964 		info = &ctx->ifc_filter_info;
4965 		gtask = &ctx->ifc_admin_task;
4966 		tqg = qgroup_if_config_tqg;
4967 		fn = _task_fn_admin;
4968 		intr_fast = iflib_fast_intr_ctx;
4969 		break;
4970 	default:
4971 		panic("unknown net intr type");
4972 	}
4973 
4974 	info->ifi_filter = filter;
4975 	info->ifi_filter_arg = filter_arg;
4976 	info->ifi_task = gtask;
4977 	info->ifi_ctx = q;
4978 
4979 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4980 	if (err != 0) {
4981 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4982 		return (err);
4983 	}
4984 	if (type == IFLIB_INTR_ADMIN)
4985 		return (0);
4986 
4987 	if (tqrid != -1) {
4988 		cpuid = find_nth(ctx, &cpus, qid);
4989 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4990 	} else {
4991 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4992 	}
4993 
4994 	return (0);
4995 }
4996 
4997 void
4998 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4999 {
5000 	struct grouptask *gtask;
5001 	struct taskqgroup *tqg;
5002 	gtask_fn_t *fn;
5003 	void *q;
5004 
5005 	switch (type) {
5006 	case IFLIB_INTR_TX:
5007 		q = &ctx->ifc_txqs[qid];
5008 		gtask = &ctx->ifc_txqs[qid].ift_task;
5009 		tqg = qgroup_if_io_tqg;
5010 		fn = _task_fn_tx;
5011 		break;
5012 	case IFLIB_INTR_RX:
5013 		q = &ctx->ifc_rxqs[qid];
5014 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5015 		tqg = qgroup_if_io_tqg;
5016 		fn = _task_fn_rx;
5017 		break;
5018 	case IFLIB_INTR_IOV:
5019 		q = ctx;
5020 		gtask = &ctx->ifc_vflr_task;
5021 		tqg = qgroup_if_config_tqg;
5022 		rid = -1;
5023 		fn = _task_fn_iov;
5024 		break;
5025 	default:
5026 		panic("unknown net intr type");
5027 	}
5028 	GROUPTASK_INIT(gtask, 0, fn, q);
5029 	taskqgroup_attach(tqg, gtask, q, rid, name);
5030 }
5031 
5032 void
5033 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5034 {
5035 	if (irq->ii_tag)
5036 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5037 
5038 	if (irq->ii_res)
5039 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5040 }
5041 
5042 static int
5043 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5044 {
5045 	iflib_txq_t txq = ctx->ifc_txqs;
5046 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5047 	if_irq_t irq = &ctx->ifc_legacy_irq;
5048 	iflib_filter_info_t info;
5049 	struct grouptask *gtask;
5050 	struct taskqgroup *tqg;
5051 	gtask_fn_t *fn;
5052 	int tqrid;
5053 	void *q;
5054 	int err;
5055 
5056 	q = &ctx->ifc_rxqs[0];
5057 	info = &rxq[0].ifr_filter_info;
5058 	gtask = &rxq[0].ifr_task;
5059 	tqg = qgroup_if_io_tqg;
5060 	tqrid = irq->ii_rid = *rid;
5061 	fn = _task_fn_rx;
5062 
5063 	ctx->ifc_flags |= IFC_LEGACY;
5064 	info->ifi_filter = filter;
5065 	info->ifi_filter_arg = filter_arg;
5066 	info->ifi_task = gtask;
5067 	info->ifi_ctx = ctx;
5068 
5069 	/* We allocate a single interrupt resource */
5070 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5071 		return (err);
5072 	GROUPTASK_INIT(gtask, 0, fn, q);
5073 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5074 
5075 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5076 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5077 	return (0);
5078 }
5079 
5080 void
5081 iflib_led_create(if_ctx_t ctx)
5082 {
5083 
5084 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5085 	    device_get_nameunit(ctx->ifc_dev));
5086 }
5087 
5088 void
5089 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5090 {
5091 
5092 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5093 }
5094 
5095 void
5096 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5097 {
5098 
5099 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5100 }
5101 
5102 void
5103 iflib_admin_intr_deferred(if_ctx_t ctx)
5104 {
5105 #ifdef INVARIANTS
5106 	struct grouptask *gtask;
5107 
5108 	gtask = &ctx->ifc_admin_task;
5109 	MPASS(gtask->gt_taskqueue != NULL);
5110 #endif
5111 
5112 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5113 }
5114 
5115 void
5116 iflib_iov_intr_deferred(if_ctx_t ctx)
5117 {
5118 
5119 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5120 }
5121 
5122 void
5123 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5124 {
5125 
5126 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5127 }
5128 
5129 void
5130 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5131 	char *name)
5132 {
5133 
5134 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5135 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5136 }
5137 
5138 void
5139 iflib_config_gtask_deinit(struct grouptask *gtask)
5140 {
5141 
5142 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5143 }
5144 
5145 void
5146 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5147 {
5148 	if_t ifp = ctx->ifc_ifp;
5149 	iflib_txq_t txq = ctx->ifc_txqs;
5150 
5151 	if_setbaudrate(ifp, baudrate);
5152 	if (baudrate >= IF_Gbps(10))
5153 		ctx->ifc_flags |= IFC_PREFETCH;
5154 
5155 	/* If link down, disable watchdog */
5156 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5157 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5158 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5159 	}
5160 	ctx->ifc_link_state = link_state;
5161 	if_link_state_change(ifp, link_state);
5162 }
5163 
5164 static int
5165 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5166 {
5167 	int credits;
5168 #ifdef INVARIANTS
5169 	int credits_pre = txq->ift_cidx_processed;
5170 #endif
5171 
5172 	if (ctx->isc_txd_credits_update == NULL)
5173 		return (0);
5174 
5175 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5176 		return (0);
5177 
5178 	txq->ift_processed += credits;
5179 	txq->ift_cidx_processed += credits;
5180 
5181 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5182 	if (txq->ift_cidx_processed >= txq->ift_size)
5183 		txq->ift_cidx_processed -= txq->ift_size;
5184 	return (credits);
5185 }
5186 
5187 static int
5188 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5189 {
5190 
5191 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5192 	    budget));
5193 }
5194 
5195 void
5196 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5197 	const char *description, if_int_delay_info_t info,
5198 	int offset, int value)
5199 {
5200 	info->iidi_ctx = ctx;
5201 	info->iidi_offset = offset;
5202 	info->iidi_value = value;
5203 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5204 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5205 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5206 	    info, 0, iflib_sysctl_int_delay, "I", description);
5207 }
5208 
5209 struct mtx *
5210 iflib_ctx_lock_get(if_ctx_t ctx)
5211 {
5212 
5213 	return (&ctx->ifc_mtx);
5214 }
5215 
5216 static int
5217 iflib_msix_init(if_ctx_t ctx)
5218 {
5219 	device_t dev = ctx->ifc_dev;
5220 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5221 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5222 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5223 	int iflib_num_tx_queues, iflib_num_rx_queues;
5224 	int err, admincnt, bar;
5225 
5226 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5227 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5228 
5229 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5230 
5231 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5232 	admincnt = sctx->isc_admin_intrcnt;
5233 	/* Override by tuneable */
5234 	if (scctx->isc_disable_msix)
5235 		goto msi;
5236 
5237 	/*
5238 	** When used in a virtualized environment
5239 	** PCI BUSMASTER capability may not be set
5240 	** so explicity set it here and rewrite
5241 	** the ENABLE in the MSIX control register
5242 	** at this point to cause the host to
5243 	** successfully initialize us.
5244 	*/
5245 	{
5246 		int msix_ctrl, rid;
5247 
5248  		pci_enable_busmaster(dev);
5249 		rid = 0;
5250 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5251 			rid += PCIR_MSIX_CTRL;
5252 			msix_ctrl = pci_read_config(dev, rid, 2);
5253 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5254 			pci_write_config(dev, rid, msix_ctrl, 2);
5255 		} else {
5256 			device_printf(dev, "PCIY_MSIX capability not found; "
5257 			                   "or rid %d == 0.\n", rid);
5258 			goto msi;
5259 		}
5260 	}
5261 
5262 	/*
5263 	 * bar == -1 => "trust me I know what I'm doing"
5264 	 * Some drivers are for hardware that is so shoddily
5265 	 * documented that no one knows which bars are which
5266 	 * so the developer has to map all bars. This hack
5267 	 * allows shoddy garbage to use msix in this framework.
5268 	 */
5269 	if (bar != -1) {
5270 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5271 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5272 		if (ctx->ifc_msix_mem == NULL) {
5273 			/* May not be enabled */
5274 			device_printf(dev, "Unable to map MSIX table \n");
5275 			goto msi;
5276 		}
5277 	}
5278 	/* First try MSI/X */
5279 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5280 		device_printf(dev, "System has MSIX disabled \n");
5281 		bus_release_resource(dev, SYS_RES_MEMORY,
5282 		    bar, ctx->ifc_msix_mem);
5283 		ctx->ifc_msix_mem = NULL;
5284 		goto msi;
5285 	}
5286 #if IFLIB_DEBUG
5287 	/* use only 1 qset in debug mode */
5288 	queuemsgs = min(msgs - admincnt, 1);
5289 #else
5290 	queuemsgs = msgs - admincnt;
5291 #endif
5292 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5293 #ifdef RSS
5294 		queues = imin(queuemsgs, rss_getnumbuckets());
5295 #else
5296 		queues = queuemsgs;
5297 #endif
5298 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5299 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5300 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5301 	} else {
5302 		device_printf(dev, "Unable to fetch CPU list\n");
5303 		/* Figure out a reasonable auto config value */
5304 		queues = min(queuemsgs, mp_ncpus);
5305 	}
5306 #ifdef  RSS
5307 	/* If we're doing RSS, clamp at the number of RSS buckets */
5308 	if (queues > rss_getnumbuckets())
5309 		queues = rss_getnumbuckets();
5310 #endif
5311 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5312 		rx_queues = iflib_num_rx_queues;
5313 	else
5314 		rx_queues = queues;
5315 	/*
5316 	 * We want this to be all logical CPUs by default
5317 	 */
5318 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5319 		tx_queues = iflib_num_tx_queues;
5320 	else
5321 		tx_queues = mp_ncpus;
5322 
5323 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5324 #ifdef INVARIANTS
5325 		if (tx_queues != rx_queues)
5326 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5327 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5328 #endif
5329 		tx_queues = min(rx_queues, tx_queues);
5330 		rx_queues = min(rx_queues, tx_queues);
5331 	}
5332 
5333 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5334 
5335 	vectors = rx_queues + admincnt;
5336 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5337 		device_printf(dev,
5338 					  "Using MSIX interrupts with %d vectors\n", vectors);
5339 		scctx->isc_vectors = vectors;
5340 		scctx->isc_nrxqsets = rx_queues;
5341 		scctx->isc_ntxqsets = tx_queues;
5342 		scctx->isc_intr = IFLIB_INTR_MSIX;
5343 
5344 		return (vectors);
5345 	} else {
5346 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5347 	}
5348 msi:
5349 	vectors = pci_msi_count(dev);
5350 	scctx->isc_nrxqsets = 1;
5351 	scctx->isc_ntxqsets = 1;
5352 	scctx->isc_vectors = vectors;
5353 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5354 		device_printf(dev,"Using an MSI interrupt\n");
5355 		scctx->isc_intr = IFLIB_INTR_MSI;
5356 	} else {
5357 		device_printf(dev,"Using a Legacy interrupt\n");
5358 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5359 	}
5360 
5361 	return (vectors);
5362 }
5363 
5364 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5365 
5366 static int
5367 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5368 {
5369 	int rc;
5370 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5371 	struct sbuf *sb;
5372 	char *ring_state = "UNKNOWN";
5373 
5374 	/* XXX needed ? */
5375 	rc = sysctl_wire_old_buffer(req, 0);
5376 	MPASS(rc == 0);
5377 	if (rc != 0)
5378 		return (rc);
5379 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5380 	MPASS(sb != NULL);
5381 	if (sb == NULL)
5382 		return (ENOMEM);
5383 	if (state[3] <= 3)
5384 		ring_state = ring_states[state[3]];
5385 
5386 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5387 		    state[0], state[1], state[2], ring_state);
5388 	rc = sbuf_finish(sb);
5389 	sbuf_delete(sb);
5390         return(rc);
5391 }
5392 
5393 enum iflib_ndesc_handler {
5394 	IFLIB_NTXD_HANDLER,
5395 	IFLIB_NRXD_HANDLER,
5396 };
5397 
5398 static int
5399 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5400 {
5401 	if_ctx_t ctx = (void *)arg1;
5402 	enum iflib_ndesc_handler type = arg2;
5403 	char buf[256] = {0};
5404 	qidx_t *ndesc;
5405 	char *p, *next;
5406 	int nqs, rc, i;
5407 
5408 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5409 
5410 	nqs = 8;
5411 	switch(type) {
5412 	case IFLIB_NTXD_HANDLER:
5413 		ndesc = ctx->ifc_sysctl_ntxds;
5414 		if (ctx->ifc_sctx)
5415 			nqs = ctx->ifc_sctx->isc_ntxqs;
5416 		break;
5417 	case IFLIB_NRXD_HANDLER:
5418 		ndesc = ctx->ifc_sysctl_nrxds;
5419 		if (ctx->ifc_sctx)
5420 			nqs = ctx->ifc_sctx->isc_nrxqs;
5421 		break;
5422 	}
5423 	if (nqs == 0)
5424 		nqs = 8;
5425 
5426 	for (i=0; i<8; i++) {
5427 		if (i >= nqs)
5428 			break;
5429 		if (i)
5430 			strcat(buf, ",");
5431 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5432 	}
5433 
5434 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5435 	if (rc || req->newptr == NULL)
5436 		return rc;
5437 
5438 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5439 	    i++, p = strsep(&next, " ,")) {
5440 		ndesc[i] = strtoul(p, NULL, 10);
5441 	}
5442 
5443 	return(rc);
5444 }
5445 
5446 #define NAME_BUFLEN 32
5447 static void
5448 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5449 {
5450         device_t dev = iflib_get_dev(ctx);
5451 	struct sysctl_oid_list *child, *oid_list;
5452 	struct sysctl_ctx_list *ctx_list;
5453 	struct sysctl_oid *node;
5454 
5455 	ctx_list = device_get_sysctl_ctx(dev);
5456 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5457 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5458 						      CTLFLAG_RD, NULL, "IFLIB fields");
5459 	oid_list = SYSCTL_CHILDREN(node);
5460 
5461 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5462 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5463 		       "driver version");
5464 
5465 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5466 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5467 			"# of txqs to use, 0 => use default #");
5468 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5469 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5470 			"# of rxqs to use, 0 => use default #");
5471 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5472 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5473                        "permit #txq != #rxq");
5474        SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5475                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5476                       "disable MSIX (default 0)");
5477 
5478 	/* XXX change for per-queue sizes */
5479 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5480 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5481                        mp_ndesc_handler, "A",
5482                        "list of # of tx descriptors to use, 0 = use default #");
5483 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5484 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5485                        mp_ndesc_handler, "A",
5486                        "list of # of rx descriptors to use, 0 = use default #");
5487 }
5488 
5489 static void
5490 iflib_add_device_sysctl_post(if_ctx_t ctx)
5491 {
5492 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5493 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5494         device_t dev = iflib_get_dev(ctx);
5495 	struct sysctl_oid_list *child;
5496 	struct sysctl_ctx_list *ctx_list;
5497 	iflib_fl_t fl;
5498 	iflib_txq_t txq;
5499 	iflib_rxq_t rxq;
5500 	int i, j;
5501 	char namebuf[NAME_BUFLEN];
5502 	char *qfmt;
5503 	struct sysctl_oid *queue_node, *fl_node, *node;
5504 	struct sysctl_oid_list *queue_list, *fl_list;
5505 	ctx_list = device_get_sysctl_ctx(dev);
5506 
5507 	node = ctx->ifc_sysctl_node;
5508 	child = SYSCTL_CHILDREN(node);
5509 
5510 	if (scctx->isc_ntxqsets > 100)
5511 		qfmt = "txq%03d";
5512 	else if (scctx->isc_ntxqsets > 10)
5513 		qfmt = "txq%02d";
5514 	else
5515 		qfmt = "txq%d";
5516 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5517 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5518 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5519 					     CTLFLAG_RD, NULL, "Queue Name");
5520 		queue_list = SYSCTL_CHILDREN(queue_node);
5521 #if MEMORY_LOGGING
5522 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5523 				CTLFLAG_RD,
5524 				&txq->ift_dequeued, "total mbufs freed");
5525 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5526 				CTLFLAG_RD,
5527 				&txq->ift_enqueued, "total mbufs enqueued");
5528 #endif
5529 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5530 				   CTLFLAG_RD,
5531 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5532 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5533 				   CTLFLAG_RD,
5534 				   &txq->ift_pullups, "# of times m_pullup was called");
5535 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5536 				   CTLFLAG_RD,
5537 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5538 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5539 				   CTLFLAG_RD,
5540 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5541 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5542 				   CTLFLAG_RD,
5543 				   &txq->ift_map_failed, "# of times dma map failed");
5544 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5545 				   CTLFLAG_RD,
5546 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5547 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5548 				   CTLFLAG_RD,
5549 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5550 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5551 				   CTLFLAG_RD,
5552 				   &txq->ift_pidx, 1, "Producer Index");
5553 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5554 				   CTLFLAG_RD,
5555 				   &txq->ift_cidx, 1, "Consumer Index");
5556 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5557 				   CTLFLAG_RD,
5558 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5559 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5560 				   CTLFLAG_RD,
5561 				   &txq->ift_in_use, 1, "descriptors in use");
5562 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5563 				   CTLFLAG_RD,
5564 				   &txq->ift_processed, "descriptors procesed for clean");
5565 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5566 				   CTLFLAG_RD,
5567 				   &txq->ift_cleaned, "total cleaned");
5568 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5569 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5570 				0, mp_ring_state_handler, "A", "soft ring state");
5571 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5572 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5573 				       "# of enqueues to the mp_ring for this queue");
5574 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5575 				       CTLFLAG_RD, &txq->ift_br->drops,
5576 				       "# of drops in the mp_ring for this queue");
5577 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5578 				       CTLFLAG_RD, &txq->ift_br->starts,
5579 				       "# of normal consumer starts in the mp_ring for this queue");
5580 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5581 				       CTLFLAG_RD, &txq->ift_br->stalls,
5582 					       "# of consumer stalls in the mp_ring for this queue");
5583 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5584 			       CTLFLAG_RD, &txq->ift_br->restarts,
5585 				       "# of consumer restarts in the mp_ring for this queue");
5586 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5587 				       CTLFLAG_RD, &txq->ift_br->abdications,
5588 				       "# of consumer abdications in the mp_ring for this queue");
5589 	}
5590 
5591 	if (scctx->isc_nrxqsets > 100)
5592 		qfmt = "rxq%03d";
5593 	else if (scctx->isc_nrxqsets > 10)
5594 		qfmt = "rxq%02d";
5595 	else
5596 		qfmt = "rxq%d";
5597 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5598 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5599 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5600 					     CTLFLAG_RD, NULL, "Queue Name");
5601 		queue_list = SYSCTL_CHILDREN(queue_node);
5602 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5603 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5604 				       CTLFLAG_RD,
5605 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5606 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5607 				       CTLFLAG_RD,
5608 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5609 		}
5610 
5611 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5612 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5613 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5614 						     CTLFLAG_RD, NULL, "freelist Name");
5615 			fl_list = SYSCTL_CHILDREN(fl_node);
5616 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5617 				       CTLFLAG_RD,
5618 				       &fl->ifl_pidx, 1, "Producer Index");
5619 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5620 				       CTLFLAG_RD,
5621 				       &fl->ifl_cidx, 1, "Consumer Index");
5622 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5623 				       CTLFLAG_RD,
5624 				       &fl->ifl_credits, 1, "credits available");
5625 #if MEMORY_LOGGING
5626 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5627 					CTLFLAG_RD,
5628 					&fl->ifl_m_enqueued, "mbufs allocated");
5629 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5630 					CTLFLAG_RD,
5631 					&fl->ifl_m_dequeued, "mbufs freed");
5632 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5633 					CTLFLAG_RD,
5634 					&fl->ifl_cl_enqueued, "clusters allocated");
5635 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5636 					CTLFLAG_RD,
5637 					&fl->ifl_cl_dequeued, "clusters freed");
5638 #endif
5639 
5640 		}
5641 	}
5642 
5643 }
5644 
5645 #ifndef __NO_STRICT_ALIGNMENT
5646 static struct mbuf *
5647 iflib_fixup_rx(struct mbuf *m)
5648 {
5649 	struct mbuf *n;
5650 
5651 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5652 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5653 		m->m_data += ETHER_HDR_LEN;
5654 		n = m;
5655 	} else {
5656 		MGETHDR(n, M_NOWAIT, MT_DATA);
5657 		if (n == NULL) {
5658 			m_freem(m);
5659 			return (NULL);
5660 		}
5661 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5662 		m->m_data += ETHER_HDR_LEN;
5663 		m->m_len -= ETHER_HDR_LEN;
5664 		n->m_len = ETHER_HDR_LEN;
5665 		M_MOVE_PKTHDR(n, m);
5666 		n->m_next = m;
5667 	}
5668 	return (n);
5669 }
5670 #endif
5671