xref: /freebsd/sys/net/iflib.c (revision 478d3005721019319c11a37980f8464ac22f529a)
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 /*
97  * enable accounting of every mbuf as it comes in to and goes out of
98  * iflib's software descriptor references
99  */
100 #define MEMORY_LOGGING 0
101 /*
102  * Enable mbuf vectors for compressing long mbuf chains
103  */
104 
105 /*
106  * NB:
107  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108  *   we prefetch needs to be determined by the time spent in m_free vis a vis
109  *   the cost of a prefetch. This will of course vary based on the workload:
110  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111  *        is quite expensive, thus suggesting very little prefetch.
112  *      - small packet forwarding which is just returning a single mbuf to
113  *        UMA will typically be very fast vis a vis the cost of a memory
114  *        access.
115  */
116 
117 
118 /*
119  * File organization:
120  *  - private structures
121  *  - iflib private utility functions
122  *  - ifnet functions
123  *  - vlan registry and other exported functions
124  *  - iflib public core functions
125  *
126  *
127  */
128 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
129 
130 struct iflib_txq;
131 typedef struct iflib_txq *iflib_txq_t;
132 struct iflib_rxq;
133 typedef struct iflib_rxq *iflib_rxq_t;
134 struct iflib_fl;
135 typedef struct iflib_fl *iflib_fl_t;
136 
137 struct iflib_ctx;
138 
139 typedef struct iflib_filter_info {
140 	driver_filter_t *ifi_filter;
141 	void *ifi_filter_arg;
142 	struct grouptask *ifi_task;
143 	void *ifi_ctx;
144 } *iflib_filter_info_t;
145 
146 struct iflib_ctx {
147 	KOBJ_FIELDS;
148    /*
149    * Pointer to hardware driver's softc
150    */
151 	void *ifc_softc;
152 	device_t ifc_dev;
153 	if_t ifc_ifp;
154 
155 	cpuset_t ifc_cpus;
156 	if_shared_ctx_t ifc_sctx;
157 	struct if_softc_ctx ifc_softc_ctx;
158 
159 	struct mtx ifc_mtx;
160 
161 	uint16_t ifc_nhwtxqs;
162 	uint16_t ifc_nhwrxqs;
163 
164 	iflib_txq_t ifc_txqs;
165 	iflib_rxq_t ifc_rxqs;
166 	uint32_t ifc_if_flags;
167 	uint32_t ifc_flags;
168 	uint32_t ifc_max_fl_buf_size;
169 	int ifc_in_detach;
170 
171 	int ifc_link_state;
172 	int ifc_link_irq;
173 	int ifc_watchdog_events;
174 	struct cdev *ifc_led_dev;
175 	struct resource *ifc_msix_mem;
176 
177 	struct if_irq ifc_legacy_irq;
178 	struct grouptask ifc_admin_task;
179 	struct grouptask ifc_vflr_task;
180 	struct iflib_filter_info ifc_filter_info;
181 	struct ifmedia	ifc_media;
182 
183 	struct sysctl_oid *ifc_sysctl_node;
184 	uint16_t ifc_sysctl_ntxqs;
185 	uint16_t ifc_sysctl_nrxqs;
186 	uint16_t ifc_sysctl_qs_eq_override;
187 
188 	qidx_t ifc_sysctl_ntxds[8];
189 	qidx_t ifc_sysctl_nrxds[8];
190 	struct if_txrx ifc_txrx;
191 #define isc_txd_encap  ifc_txrx.ift_txd_encap
192 #define isc_txd_flush  ifc_txrx.ift_txd_flush
193 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
194 #define isc_rxd_available ifc_txrx.ift_rxd_available
195 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
197 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
198 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
201 	eventhandler_tag ifc_vlan_attach_event;
202 	eventhandler_tag ifc_vlan_detach_event;
203 	uint8_t ifc_mac[ETHER_ADDR_LEN];
204 	char ifc_mtx_name[16];
205 };
206 
207 
208 void *
209 iflib_get_softc(if_ctx_t ctx)
210 {
211 
212 	return (ctx->ifc_softc);
213 }
214 
215 device_t
216 iflib_get_dev(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_dev);
220 }
221 
222 if_t
223 iflib_get_ifp(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_ifp);
227 }
228 
229 struct ifmedia *
230 iflib_get_media(if_ctx_t ctx)
231 {
232 
233 	return (&ctx->ifc_media);
234 }
235 
236 void
237 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
238 {
239 
240 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
241 }
242 
243 if_softc_ctx_t
244 iflib_get_softc_ctx(if_ctx_t ctx)
245 {
246 
247 	return (&ctx->ifc_softc_ctx);
248 }
249 
250 if_shared_ctx_t
251 iflib_get_sctx(if_ctx_t ctx)
252 {
253 
254 	return (ctx->ifc_sctx);
255 }
256 
257 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
258 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
259 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
260 
261 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
262 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
263 
264 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
265 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
266 #define RX_SW_DESC_INUSE        (1 << 3)
267 #define TX_SW_DESC_MAPPED       (1 << 4)
268 
269 typedef struct iflib_sw_rx_desc_array {
270 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
271 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
272 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
273 	uint8_t		*ifsd_flags;
274 } iflib_rxsd_array_t;
275 
276 typedef struct iflib_sw_tx_desc_array {
277 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
279 	uint8_t		*ifsd_flags;
280 } if_txsd_vec_t;
281 
282 
283 /* magic number that should be high enough for any hardware */
284 #define IFLIB_MAX_TX_SEGS		128
285 #define IFLIB_MAX_RX_SEGS		32
286 #define IFLIB_RX_COPY_THRESH		128
287 #define IFLIB_MAX_RX_REFRESH		32
288 /* The minimum descriptors per second before we start coalescing */
289 #define IFLIB_MIN_DESC_SEC		16384
290 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
291 #define IFLIB_QUEUE_IDLE		0
292 #define IFLIB_QUEUE_HUNG		1
293 #define IFLIB_QUEUE_WORKING		2
294 /* maximum number of txqs that can share an rx interrupt */
295 #define IFLIB_MAX_TX_SHARED_INTR	4
296 
297 /* this should really scale with ring size - this is a fairly arbitrary value */
298 #define TX_BATCH_SIZE			32
299 
300 #define IFLIB_RESTART_BUDGET		8
301 
302 #define	IFC_LEGACY		0x001
303 #define	IFC_QFLUSH		0x002
304 #define	IFC_MULTISEG		0x004
305 #define	IFC_DMAR		0x008
306 #define	IFC_SC_ALLOCATED	0x010
307 #define	IFC_INIT_DONE		0x020
308 #define	IFC_PREFETCH		0x040
309 #define	IFC_DO_RESET		0x080
310 #define	IFC_CHECK_HUNG		0x100
311 
312 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
313 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
314 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
315 struct iflib_txq {
316 	qidx_t		ift_in_use;
317 	qidx_t		ift_cidx;
318 	qidx_t		ift_cidx_processed;
319 	qidx_t		ift_pidx;
320 	uint8_t		ift_gen;
321 	uint8_t		ift_br_offset;
322 	uint16_t	ift_npending;
323 	uint16_t	ift_db_pending;
324 	uint16_t	ift_rs_pending;
325 	/* implicit pad */
326 	uint8_t		ift_txd_size[8];
327 	uint64_t	ift_processed;
328 	uint64_t	ift_cleaned;
329 	uint64_t	ift_cleaned_prev;
330 #if MEMORY_LOGGING
331 	uint64_t	ift_enqueued;
332 	uint64_t	ift_dequeued;
333 #endif
334 	uint64_t	ift_no_tx_dma_setup;
335 	uint64_t	ift_no_desc_avail;
336 	uint64_t	ift_mbuf_defrag_failed;
337 	uint64_t	ift_mbuf_defrag;
338 	uint64_t	ift_map_failed;
339 	uint64_t	ift_txd_encap_efbig;
340 	uint64_t	ift_pullups;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t		ift_desc_tag;
359 	bus_dma_tag_t		ift_tso_desc_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define MTX_NAME_LEN 16
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	char                    ift_db_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t           ifl_desc_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	/* If there is a separate completion queue -
425 	 * these are the cq cidx and pidx. Otherwise
426 	 * these are unused.
427 	 */
428 	qidx_t		ifr_size;
429 	qidx_t		ifr_cq_cidx;
430 	qidx_t		ifr_cq_pidx;
431 	uint8_t		ifr_cq_gen;
432 	uint8_t		ifr_fl_offset;
433 
434 	if_ctx_t	ifr_ctx;
435 	iflib_fl_t	ifr_fl;
436 	uint64_t	ifr_rx_irq;
437 	uint16_t	ifr_id;
438 	uint8_t		ifr_lro_enabled;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	struct lro_ctrl			ifr_lc;
443 	struct grouptask        ifr_task;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	struct mbuf **ifsd_m;
457 	iflib_fl_t ifsd_fl;
458 	qidx_t ifsd_cidx;
459 } *if_rxsd_t;
460 
461 /* multiple of word size */
462 #ifdef __LP64__
463 #define PKT_INFO_SIZE	6
464 #define RXD_INFO_SIZE	5
465 #define PKT_TYPE uint64_t
466 #else
467 #define PKT_INFO_SIZE	11
468 #define RXD_INFO_SIZE	8
469 #define PKT_TYPE uint32_t
470 #endif
471 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
472 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
473 
474 typedef struct if_pkt_info_pad {
475 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
476 } *if_pkt_info_pad_t;
477 typedef struct if_rxd_info_pad {
478 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
479 } *if_rxd_info_pad_t;
480 
481 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
482 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
483 
484 
485 static inline void
486 pkt_info_zero(if_pkt_info_t pi)
487 {
488 	if_pkt_info_pad_t pi_pad;
489 
490 	pi_pad = (if_pkt_info_pad_t)pi;
491 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
492 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
493 #ifndef __LP64__
494 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
495 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
496 #endif
497 }
498 
499 static inline void
500 rxd_info_zero(if_rxd_info_t ri)
501 {
502 	if_rxd_info_pad_t ri_pad;
503 	int i;
504 
505 	ri_pad = (if_rxd_info_pad_t)ri;
506 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
507 		ri_pad->rxd_val[i] = 0;
508 		ri_pad->rxd_val[i+1] = 0;
509 		ri_pad->rxd_val[i+2] = 0;
510 		ri_pad->rxd_val[i+3] = 0;
511 	}
512 #ifdef __LP64__
513 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
514 #endif
515 }
516 
517 /*
518  * Only allow a single packet to take up most 1/nth of the tx ring
519  */
520 #define MAX_SINGLE_PACKET_FRACTION 12
521 #define IF_BAD_DMA (bus_addr_t)-1
522 
523 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
524                    "iflib driver parameters");
525 
526 static int iflib_timer_int;
527 SYSCTL_INT(_net_iflib, OID_AUTO, timer_int, CTLFLAG_RW, &iflib_timer_int,
528     0, "interval at which to run per-queue timers (in ticks)");
529 
530 static int force_busdma = 1;
531 SYSCTL_INT(_net_iflib, OID_AUTO, force_busdma, CTLFLAG_RDTUN, &force_busdma,
532     1, "force busdma");
533 
534 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
535 
536 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
537 
538 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
539 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
540 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
541 
542 
543 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
544 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
545 
546 
547 /* Our boot-time initialization hook */
548 static int	iflib_module_event_handler(module_t, int, void *);
549 
550 static moduledata_t iflib_moduledata = {
551 	"iflib",
552 	iflib_module_event_handler,
553 	NULL
554 };
555 
556 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
557 MODULE_VERSION(iflib, 1);
558 
559 MODULE_DEPEND(iflib, pci, 1, 1, 1);
560 MODULE_DEPEND(iflib, ether, 1, 1, 1);
561 
562 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
563 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
564 
565 #ifndef IFLIB_DEBUG_COUNTERS
566 #ifdef INVARIANTS
567 #define IFLIB_DEBUG_COUNTERS 1
568 #else
569 #define IFLIB_DEBUG_COUNTERS 0
570 #endif /* !INVARIANTS */
571 #endif
572 
573 /*
574  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
575  */
576 static int iflib_min_tx_latency = 0;
577 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
578 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
579 static int iflib_no_tx_batch = 0;
580 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
581 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
582 
583 
584 #if IFLIB_DEBUG_COUNTERS
585 
586 static int iflib_tx_seen;
587 static int iflib_tx_sent;
588 static int iflib_tx_encap;
589 static int iflib_rx_allocs;
590 static int iflib_fl_refills;
591 static int iflib_fl_refills_large;
592 static int iflib_tx_frees;
593 
594 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
595 		   &iflib_tx_seen, 0, "# tx mbufs seen");
596 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
597 		   &iflib_tx_sent, 0, "# tx mbufs sent");
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
599 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
601 		   &iflib_tx_frees, 0, "# tx frees");
602 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
603 		   &iflib_rx_allocs, 0, "# rx allocations");
604 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
605 		   &iflib_fl_refills, 0, "# refills");
606 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
607 		   &iflib_fl_refills_large, 0, "# large refills");
608 
609 
610 static int iflib_txq_drain_flushing;
611 static int iflib_txq_drain_oactive;
612 static int iflib_txq_drain_notready;
613 static int iflib_txq_drain_encapfail;
614 
615 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
616 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
617 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
618 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
619 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
620 		   &iflib_txq_drain_notready, 0, "# drain notready");
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
622 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
623 
624 
625 static int iflib_encap_load_mbuf_fail;
626 static int iflib_encap_txq_avail_fail;
627 static int iflib_encap_txd_encap_fail;
628 
629 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
630 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
632 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
634 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
635 
636 static int iflib_task_fn_rxs;
637 static int iflib_rx_intr_enables;
638 static int iflib_fast_intrs;
639 static int iflib_intr_link;
640 static int iflib_intr_msix;
641 static int iflib_rx_unavail;
642 static int iflib_rx_ctx_inactive;
643 static int iflib_rx_zero_len;
644 static int iflib_rx_if_input;
645 static int iflib_rx_mbuf_null;
646 static int iflib_rxd_flush;
647 
648 static int iflib_verbose_debug;
649 
650 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
651 		   &iflib_intr_link, 0, "# intr link calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
653 		   &iflib_intr_msix, 0, "# intr msix calls");
654 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
655 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
657 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
658 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
659 		   &iflib_fast_intrs, 0, "# fast_intr calls");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
661 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
663 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
665 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
667 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
669 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
671 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
672 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
673 		   &iflib_verbose_debug, 0, "enable verbose debugging");
674 
675 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
676 static void
677 iflib_debug_reset(void)
678 {
679 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
680 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
681 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
682 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
683 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
684 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
685 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
686 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
687 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
688 }
689 
690 #else
691 #define DBG_COUNTER_INC(name)
692 static void iflib_debug_reset(void) {}
693 #endif
694 
695 
696 
697 #define IFLIB_DEBUG 0
698 
699 static void iflib_tx_structures_free(if_ctx_t ctx);
700 static void iflib_rx_structures_free(if_ctx_t ctx);
701 static int iflib_queues_alloc(if_ctx_t ctx);
702 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
703 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
704 static int iflib_qset_structures_setup(if_ctx_t ctx);
705 static int iflib_msix_init(if_ctx_t ctx);
706 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
707 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
708 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
709 static int iflib_register(if_ctx_t);
710 static void iflib_init_locked(if_ctx_t ctx);
711 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
713 static void iflib_ifmp_purge(iflib_txq_t txq);
714 static void _iflib_pre_assert(if_softc_ctx_t scctx);
715 static void iflib_stop(if_ctx_t ctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 #ifndef __NO_STRICT_ALIGNMENT
718 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
719 #endif
720 
721 #ifdef DEV_NETMAP
722 #include <sys/selinfo.h>
723 #include <net/netmap.h>
724 #include <dev/netmap/netmap_kern.h>
725 
726 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
727 
728 /*
729  * device-specific sysctl variables:
730  *
731  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
732  *	During regular operations the CRC is stripped, but on some
733  *	hardware reception of frames not multiple of 64 is slower,
734  *	so using crcstrip=0 helps in benchmarks.
735  *
736  * iflib_rx_miss, iflib_rx_miss_bufs:
737  *	count packets that might be missed due to lost interrupts.
738  */
739 SYSCTL_DECL(_dev_netmap);
740 /*
741  * The xl driver by default strips CRCs and we do not override it.
742  */
743 
744 int iflib_crcstrip = 1;
745 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
746     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
747 
748 int iflib_rx_miss, iflib_rx_miss_bufs;
749 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
750     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
751 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
752     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
753 
754 /*
755  * Register/unregister. We are already under netmap lock.
756  * Only called on the first register or the last unregister.
757  */
758 static int
759 iflib_netmap_register(struct netmap_adapter *na, int onoff)
760 {
761 	struct ifnet *ifp = na->ifp;
762 	if_ctx_t ctx = ifp->if_softc;
763 	int status;
764 
765 	CTX_LOCK(ctx);
766 	IFDI_INTR_DISABLE(ctx);
767 
768 	/* Tell the stack that the interface is no longer active */
769 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
770 
771 	if (!CTX_IS_VF(ctx))
772 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
773 
774 	/* enable or disable flags and callbacks in na and ifp */
775 	if (onoff) {
776 		nm_set_native_flags(na);
777 	} else {
778 		nm_clear_native_flags(na);
779 	}
780 	iflib_stop(ctx);
781 	iflib_init_locked(ctx);
782 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
783 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
784 	if (status)
785 		nm_clear_native_flags(na);
786 	CTX_UNLOCK(ctx);
787 	return (status);
788 }
789 
790 /*
791  * Reconcile kernel and user view of the transmit ring.
792  *
793  * All information is in the kring.
794  * Userspace wants to send packets up to the one before kring->rhead,
795  * kernel knows kring->nr_hwcur is the first unsent packet.
796  *
797  * Here we push packets out (as many as possible), and possibly
798  * reclaim buffers from previously completed transmission.
799  *
800  * The caller (netmap) guarantees that there is only one instance
801  * running at any time. Any interference with other driver
802  * methods should be handled by the individual drivers.
803  */
804 static int
805 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
806 {
807 	struct netmap_adapter *na = kring->na;
808 	struct ifnet *ifp = na->ifp;
809 	struct netmap_ring *ring = kring->ring;
810 	u_int nm_i;	/* index into the netmap ring */
811 	u_int nic_i;	/* index into the NIC ring */
812 	u_int n;
813 	u_int const lim = kring->nkr_num_slots - 1;
814 	u_int const head = kring->rhead;
815 	struct if_pkt_info pi;
816 
817 	/*
818 	 * interrupts on every tx packet are expensive so request
819 	 * them every half ring, or where NS_REPORT is set
820 	 */
821 	u_int report_frequency = kring->nkr_num_slots >> 1;
822 	/* device-specific */
823 	if_ctx_t ctx = ifp->if_softc;
824 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
825 
826 	if (txq->ift_sds.ifsd_map)
827 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
828 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
829 
830 
831 	/*
832 	 * First part: process new packets to send.
833 	 * nm_i is the current index in the netmap ring,
834 	 * nic_i is the corresponding index in the NIC ring.
835 	 *
836 	 * If we have packets to send (nm_i != head)
837 	 * iterate over the netmap ring, fetch length and update
838 	 * the corresponding slot in the NIC ring. Some drivers also
839 	 * need to update the buffer's physical address in the NIC slot
840 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
841 	 *
842 	 * The netmap_reload_map() calls is especially expensive,
843 	 * even when (as in this case) the tag is 0, so do only
844 	 * when the buffer has actually changed.
845 	 *
846 	 * If possible do not set the report/intr bit on all slots,
847 	 * but only a few times per ring or when NS_REPORT is set.
848 	 *
849 	 * Finally, on 10G and faster drivers, it might be useful
850 	 * to prefetch the next slot and txr entry.
851 	 */
852 
853 	nm_i = kring->nr_hwcur;
854 	pkt_info_zero(&pi);
855 	pi.ipi_segs = txq->ift_segs;
856 	pi.ipi_qsidx = kring->ring_id;
857 	if (nm_i != head) {	/* we have new packets to send */
858 		nic_i = netmap_idx_k2n(kring, nm_i);
859 
860 		__builtin_prefetch(&ring->slot[nm_i]);
861 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
862 		if (txq->ift_sds.ifsd_map)
863 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
864 
865 		for (n = 0; nm_i != head; n++) {
866 			struct netmap_slot *slot = &ring->slot[nm_i];
867 			u_int len = slot->len;
868 			uint64_t paddr;
869 			void *addr = PNMB(na, slot, &paddr);
870 			int flags = (slot->flags & NS_REPORT ||
871 				nic_i == 0 || nic_i == report_frequency) ?
872 				IPI_TX_INTR : 0;
873 
874 			/* device-specific */
875 			pi.ipi_len = len;
876 			pi.ipi_segs[0].ds_addr = paddr;
877 			pi.ipi_segs[0].ds_len = len;
878 			pi.ipi_nsegs = 1;
879 			pi.ipi_ndescs = 0;
880 			pi.ipi_pidx = nic_i;
881 			pi.ipi_flags = flags;
882 
883 			/* Fill the slot in the NIC ring. */
884 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
885 
886 			/* prefetch for next round */
887 			__builtin_prefetch(&ring->slot[nm_i + 1]);
888 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
889 			if (txq->ift_sds.ifsd_map) {
890 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
891 
892 				NM_CHECK_ADDR_LEN(na, addr, len);
893 
894 				if (slot->flags & NS_BUF_CHANGED) {
895 					/* buffer has changed, reload map */
896 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
897 				}
898 				/* make sure changes to the buffer are synced */
899 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
900 						BUS_DMASYNC_PREWRITE);
901 			}
902 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
903 			nm_i = nm_next(nm_i, lim);
904 			nic_i = nm_next(nic_i, lim);
905 		}
906 		kring->nr_hwcur = head;
907 
908 		/* synchronize the NIC ring */
909 		if (txq->ift_sds.ifsd_map)
910 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
911 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
912 
913 		/* (re)start the tx unit up to slot nic_i (excluded) */
914 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
915 	}
916 
917 	/*
918 	 * Second part: reclaim buffers for completed transmissions.
919 	 */
920 	if (iflib_tx_credits_update(ctx, txq)) {
921 		/* some tx completed, increment avail */
922 		nic_i = txq->ift_cidx_processed;
923 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
924 	}
925 	return (0);
926 }
927 
928 /*
929  * Reconcile kernel and user view of the receive ring.
930  * Same as for the txsync, this routine must be efficient.
931  * The caller guarantees a single invocations, but races against
932  * the rest of the driver should be handled here.
933  *
934  * On call, kring->rhead is the first packet that userspace wants
935  * to keep, and kring->rcur is the wakeup point.
936  * The kernel has previously reported packets up to kring->rtail.
937  *
938  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
939  * of whether or not we received an interrupt.
940  */
941 static int
942 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
943 {
944 	struct netmap_adapter *na = kring->na;
945 	struct netmap_ring *ring = kring->ring;
946 	uint32_t nm_i;	/* index into the netmap ring */
947 	uint32_t nic_i, nic_i_start;	/* index into the NIC ring */
948 	u_int i, n;
949 	u_int const lim = kring->nkr_num_slots - 1;
950 	u_int const head = kring->rhead;
951 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
952 	struct if_rxd_info ri;
953 	struct if_rxd_update iru;
954 
955 	struct ifnet *ifp = na->ifp;
956 	if_ctx_t ctx = ifp->if_softc;
957 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
958 	iflib_fl_t fl = rxq->ifr_fl;
959 	if (head > lim)
960 		return netmap_ring_reinit(kring);
961 
962 	/* XXX check sync modes */
963 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
964 		if (fl->ifl_sds.ifsd_map == NULL)
965 			continue;
966 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
967 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
968 	}
969 	/*
970 	 * First part: import newly received packets.
971 	 *
972 	 * nm_i is the index of the next free slot in the netmap ring,
973 	 * nic_i is the index of the next received packet in the NIC ring,
974 	 * and they may differ in case if_init() has been called while
975 	 * in netmap mode. For the receive ring we have
976 	 *
977 	 *	nic_i = rxr->next_check;
978 	 *	nm_i = kring->nr_hwtail (previous)
979 	 * and
980 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
981 	 *
982 	 * rxr->next_check is set to 0 on a ring reinit
983 	 */
984 	if (netmap_no_pendintr || force_update) {
985 		int crclen = iflib_crcstrip ? 0 : 4;
986 		int error, avail;
987 		uint16_t slot_flags = kring->nkr_slot_flags;
988 
989 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
990 			nic_i = fl->ifl_cidx;
991 			nm_i = netmap_idx_n2k(kring, nic_i);
992 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
993 			for (n = 0; avail > 0; n++, avail--) {
994 				rxd_info_zero(&ri);
995 				ri.iri_frags = rxq->ifr_frags;
996 				ri.iri_qsidx = kring->ring_id;
997 				ri.iri_ifp = ctx->ifc_ifp;
998 				ri.iri_cidx = nic_i;
999 
1000 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1001 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1002 				ring->slot[nm_i].flags = slot_flags;
1003 				if (fl->ifl_sds.ifsd_map)
1004 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1005 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1006 				nm_i = nm_next(nm_i, lim);
1007 				nic_i = nm_next(nic_i, lim);
1008 			}
1009 			if (n) { /* update the state variables */
1010 				if (netmap_no_pendintr && !force_update) {
1011 					/* diagnostics */
1012 					iflib_rx_miss ++;
1013 					iflib_rx_miss_bufs += n;
1014 				}
1015 				fl->ifl_cidx = nic_i;
1016 				kring->nr_hwtail = nm_i;
1017 			}
1018 			kring->nr_kflags &= ~NKR_PENDINTR;
1019 		}
1020 	}
1021 	/*
1022 	 * Second part: skip past packets that userspace has released.
1023 	 * (kring->nr_hwcur to head excluded),
1024 	 * and make the buffers available for reception.
1025 	 * As usual nm_i is the index in the netmap ring,
1026 	 * nic_i is the index in the NIC ring, and
1027 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1028 	 */
1029 	/* XXX not sure how this will work with multiple free lists */
1030 	nm_i = kring->nr_hwcur;
1031 	if (nm_i == head)
1032 		return (0);
1033 
1034 	iru.iru_paddrs = fl->ifl_bus_addrs;
1035 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1036 	iru.iru_idxs = fl->ifl_rxd_idxs;
1037 	iru.iru_qsidx = rxq->ifr_id;
1038 	iru.iru_buf_size = fl->ifl_buf_size;
1039 	iru.iru_flidx = fl->ifl_id;
1040 	nic_i_start = nic_i = netmap_idx_k2n(kring, nm_i);
1041 	for (i = 0; nm_i != head; i++) {
1042 		struct netmap_slot *slot = &ring->slot[nm_i];
1043 		void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
1044 
1045 		if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
1046 			goto ring_reset;
1047 
1048 		fl->ifl_vm_addrs[i] = addr;
1049 		if (fl->ifl_sds.ifsd_map && (slot->flags & NS_BUF_CHANGED)) {
1050 			/* buffer has changed, reload map */
1051 			netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
1052 		}
1053 		slot->flags &= ~NS_BUF_CHANGED;
1054 
1055 		nm_i = nm_next(nm_i, lim);
1056 		fl->ifl_rxd_idxs[i] = nic_i = nm_next(nic_i, lim);
1057 		if (nm_i != head && i < IFLIB_MAX_RX_REFRESH)
1058 			continue;
1059 
1060 		iru.iru_pidx = nic_i_start;
1061 		iru.iru_count = i;
1062 		i = 0;
1063 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1064 		if (fl->ifl_sds.ifsd_map == NULL) {
1065 			nic_i_start = nic_i;
1066 			continue;
1067 		}
1068 		nic_i = nic_i_start;
1069 		for (n = 0; n < iru.iru_count; n++) {
1070 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
1071 					BUS_DMASYNC_PREREAD);
1072 			nic_i = nm_next(nic_i, lim);
1073 		}
1074 		nic_i_start = nic_i;
1075 	}
1076 	kring->nr_hwcur = head;
1077 
1078 	if (fl->ifl_sds.ifsd_map)
1079 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1080 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1081 	/*
1082 	 * IMPORTANT: we must leave one free slot in the ring,
1083 	 * so move nic_i back by one unit
1084 	 */
1085 	nic_i = nm_prev(nic_i, lim);
1086 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
1087 	return 0;
1088 
1089 ring_reset:
1090 	return netmap_ring_reinit(kring);
1091 }
1092 
1093 static void
1094 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1095 {
1096 	struct ifnet *ifp = na->ifp;
1097 	if_ctx_t ctx = ifp->if_softc;
1098 
1099 	CTX_LOCK(ctx);
1100 	if (onoff) {
1101 		IFDI_INTR_ENABLE(ctx);
1102 	} else {
1103 		IFDI_INTR_DISABLE(ctx);
1104 	}
1105 	CTX_UNLOCK(ctx);
1106 }
1107 
1108 
1109 static int
1110 iflib_netmap_attach(if_ctx_t ctx)
1111 {
1112 	struct netmap_adapter na;
1113 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1114 
1115 	bzero(&na, sizeof(na));
1116 
1117 	na.ifp = ctx->ifc_ifp;
1118 	na.na_flags = NAF_BDG_MAYSLEEP;
1119 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1120 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1121 
1122 	na.num_tx_desc = scctx->isc_ntxd[0];
1123 	na.num_rx_desc = scctx->isc_nrxd[0];
1124 	na.nm_txsync = iflib_netmap_txsync;
1125 	na.nm_rxsync = iflib_netmap_rxsync;
1126 	na.nm_register = iflib_netmap_register;
1127 	na.nm_intr = iflib_netmap_intr;
1128 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1129 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1130 	return (netmap_attach(&na));
1131 }
1132 
1133 static void
1134 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1135 {
1136 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1137 	struct netmap_slot *slot;
1138 
1139 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1140 	if (slot == NULL)
1141 		return;
1142 	if (txq->ift_sds.ifsd_map == NULL)
1143 		return;
1144 
1145 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1146 
1147 		/*
1148 		 * In netmap mode, set the map for the packet buffer.
1149 		 * NOTE: Some drivers (not this one) also need to set
1150 		 * the physical buffer address in the NIC ring.
1151 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1152 		 * netmap slot index, si
1153 		 */
1154 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1155 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1156 	}
1157 }
1158 static void
1159 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1160 {
1161 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1162 	struct netmap_slot *slot;
1163 	struct if_rxd_update iru;
1164 	iflib_fl_t fl;
1165 	bus_dmamap_t *map;
1166 	int nrxd;
1167 	uint32_t i, j, pidx_start;
1168 
1169 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1170 	if (slot == NULL)
1171 		return;
1172 	fl = &rxq->ifr_fl[0];
1173 	map = fl->ifl_sds.ifsd_map;
1174 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1175 	iru.iru_paddrs = fl->ifl_bus_addrs;
1176 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1177 	iru.iru_idxs = fl->ifl_rxd_idxs;
1178 	iru.iru_qsidx = rxq->ifr_id;
1179 	iru.iru_buf_size = rxq->ifr_fl[0].ifl_buf_size;
1180 	iru.iru_flidx = 0;
1181 
1182 	for (pidx_start = i = j = 0; i < nrxd; i++, j++) {
1183 		int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1184 		void *addr;
1185 
1186 		fl->ifl_rxd_idxs[j] = i;
1187 		addr = fl->ifl_vm_addrs[j] = PNMB(na, slot + sj, &fl->ifl_bus_addrs[j]);
1188 		if (map) {
1189 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1190 			map++;
1191 		}
1192 
1193 		if (j < IFLIB_MAX_RX_REFRESH && i < nrxd - 1)
1194 			continue;
1195 
1196 		iru.iru_pidx = pidx_start;
1197 		pidx_start = i;
1198 		iru.iru_count = j;
1199 		j = 0;
1200 		MPASS(pidx_start + j <= nrxd);
1201 		/* Update descriptors and the cached value */
1202 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1203 	}
1204 	/* preserve queue */
1205 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1206 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1207 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1208 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1209 	} else
1210 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1211 }
1212 
1213 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1214 
1215 #else
1216 #define iflib_netmap_txq_init(ctx, txq)
1217 #define iflib_netmap_rxq_init(ctx, rxq)
1218 #define iflib_netmap_detach(ifp)
1219 
1220 #define iflib_netmap_attach(ctx) (0)
1221 #define netmap_rx_irq(ifp, qid, budget) (0)
1222 #define netmap_tx_irq(ifp, qid) do {} while (0)
1223 
1224 #endif
1225 
1226 #if defined(__i386__) || defined(__amd64__)
1227 static __inline void
1228 prefetch(void *x)
1229 {
1230 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1231 }
1232 #else
1233 #define prefetch(x)
1234 #endif
1235 
1236 static void
1237 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1238 {
1239 	if (err)
1240 		return;
1241 	*(bus_addr_t *) arg = segs[0].ds_addr;
1242 }
1243 
1244 int
1245 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1246 {
1247 	int err;
1248 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1249 	device_t dev = ctx->ifc_dev;
1250 
1251 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1252 
1253 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1254 				sctx->isc_q_align, 0,	/* alignment, bounds */
1255 				BUS_SPACE_MAXADDR,	/* lowaddr */
1256 				BUS_SPACE_MAXADDR,	/* highaddr */
1257 				NULL, NULL,		/* filter, filterarg */
1258 				size,			/* maxsize */
1259 				1,			/* nsegments */
1260 				size,			/* maxsegsize */
1261 				BUS_DMA_ALLOCNOW,	/* flags */
1262 				NULL,			/* lockfunc */
1263 				NULL,			/* lockarg */
1264 				&dma->idi_tag);
1265 	if (err) {
1266 		device_printf(dev,
1267 		    "%s: bus_dma_tag_create failed: %d\n",
1268 		    __func__, err);
1269 		goto fail_0;
1270 	}
1271 
1272 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1273 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1274 	if (err) {
1275 		device_printf(dev,
1276 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1277 		    __func__, (uintmax_t)size, err);
1278 		goto fail_1;
1279 	}
1280 
1281 	dma->idi_paddr = IF_BAD_DMA;
1282 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1283 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1284 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1285 		device_printf(dev,
1286 		    "%s: bus_dmamap_load failed: %d\n",
1287 		    __func__, err);
1288 		goto fail_2;
1289 	}
1290 
1291 	dma->idi_size = size;
1292 	return (0);
1293 
1294 fail_2:
1295 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1296 fail_1:
1297 	bus_dma_tag_destroy(dma->idi_tag);
1298 fail_0:
1299 	dma->idi_tag = NULL;
1300 
1301 	return (err);
1302 }
1303 
1304 int
1305 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1306 {
1307 	int i, err;
1308 	iflib_dma_info_t *dmaiter;
1309 
1310 	dmaiter = dmalist;
1311 	for (i = 0; i < count; i++, dmaiter++) {
1312 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1313 			break;
1314 	}
1315 	if (err)
1316 		iflib_dma_free_multi(dmalist, i);
1317 	return (err);
1318 }
1319 
1320 void
1321 iflib_dma_free(iflib_dma_info_t dma)
1322 {
1323 	if (dma->idi_tag == NULL)
1324 		return;
1325 	if (dma->idi_paddr != IF_BAD_DMA) {
1326 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1327 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1328 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1329 		dma->idi_paddr = IF_BAD_DMA;
1330 	}
1331 	if (dma->idi_vaddr != NULL) {
1332 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1333 		dma->idi_vaddr = NULL;
1334 	}
1335 	bus_dma_tag_destroy(dma->idi_tag);
1336 	dma->idi_tag = NULL;
1337 }
1338 
1339 void
1340 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1341 {
1342 	int i;
1343 	iflib_dma_info_t *dmaiter = dmalist;
1344 
1345 	for (i = 0; i < count; i++, dmaiter++)
1346 		iflib_dma_free(*dmaiter);
1347 }
1348 
1349 #ifdef EARLY_AP_STARTUP
1350 static const int iflib_started = 1;
1351 #else
1352 /*
1353  * We used to abuse the smp_started flag to decide if the queues have been
1354  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1355  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1356  * is set.  Run a SYSINIT() strictly after that to just set a usable
1357  * completion flag.
1358  */
1359 
1360 static int iflib_started;
1361 
1362 static void
1363 iflib_record_started(void *arg)
1364 {
1365 	iflib_started = 1;
1366 }
1367 
1368 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1369 	iflib_record_started, NULL);
1370 #endif
1371 
1372 static int
1373 iflib_fast_intr(void *arg)
1374 {
1375 	iflib_filter_info_t info = arg;
1376 	struct grouptask *gtask = info->ifi_task;
1377 	if (!iflib_started)
1378 		return (FILTER_HANDLED);
1379 
1380 	DBG_COUNTER_INC(fast_intrs);
1381 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1382 		return (FILTER_HANDLED);
1383 
1384 	GROUPTASK_ENQUEUE(gtask);
1385 	return (FILTER_HANDLED);
1386 }
1387 
1388 static int
1389 iflib_fast_intr_rxtx(void *arg)
1390 {
1391 	iflib_filter_info_t info = arg;
1392 	struct grouptask *gtask = info->ifi_task;
1393 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1394 	if_ctx_t ctx;
1395 	int i, cidx;
1396 
1397 	if (!iflib_started)
1398 		return (FILTER_HANDLED);
1399 
1400 	DBG_COUNTER_INC(fast_intrs);
1401 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1402 		return (FILTER_HANDLED);
1403 
1404 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1405 		qidx_t txqid = rxq->ifr_txqid[i];
1406 
1407 		ctx = rxq->ifr_ctx;
1408 
1409 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1410 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1411 			continue;
1412 		}
1413 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1414 	}
1415 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1416 		cidx = rxq->ifr_cq_cidx;
1417 	else
1418 		cidx = rxq->ifr_fl[0].ifl_cidx;
1419 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1420 		GROUPTASK_ENQUEUE(gtask);
1421 	else
1422 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1423 	return (FILTER_HANDLED);
1424 }
1425 
1426 
1427 static int
1428 iflib_fast_intr_ctx(void *arg)
1429 {
1430 	iflib_filter_info_t info = arg;
1431 	struct grouptask *gtask = info->ifi_task;
1432 
1433 	if (!iflib_started)
1434 		return (FILTER_HANDLED);
1435 
1436 	DBG_COUNTER_INC(fast_intrs);
1437 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1438 		return (FILTER_HANDLED);
1439 
1440 	GROUPTASK_ENQUEUE(gtask);
1441 	return (FILTER_HANDLED);
1442 }
1443 
1444 static int
1445 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1446 	driver_filter_t filter, driver_intr_t handler, void *arg,
1447 				 char *name)
1448 {
1449 	int rc, flags;
1450 	struct resource *res;
1451 	void *tag = NULL;
1452 	device_t dev = ctx->ifc_dev;
1453 
1454 	flags = RF_ACTIVE;
1455 	if (ctx->ifc_flags & IFC_LEGACY)
1456 		flags |= RF_SHAREABLE;
1457 	MPASS(rid < 512);
1458 	irq->ii_rid = rid;
1459 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1460 	if (res == NULL) {
1461 		device_printf(dev,
1462 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1463 		return (ENOMEM);
1464 	}
1465 	irq->ii_res = res;
1466 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1467 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1468 						filter, handler, arg, &tag);
1469 	if (rc != 0) {
1470 		device_printf(dev,
1471 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1472 					  rid, name ? name : "unknown", rc);
1473 		return (rc);
1474 	} else if (name)
1475 		bus_describe_intr(dev, res, tag, "%s", name);
1476 
1477 	irq->ii_tag = tag;
1478 	return (0);
1479 }
1480 
1481 
1482 /*********************************************************************
1483  *
1484  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1485  *  the information needed to transmit a packet on the wire. This is
1486  *  called only once at attach, setup is done every reset.
1487  *
1488  **********************************************************************/
1489 
1490 static int
1491 iflib_txsd_alloc(iflib_txq_t txq)
1492 {
1493 	if_ctx_t ctx = txq->ift_ctx;
1494 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1495 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1496 	device_t dev = ctx->ifc_dev;
1497 	int err, nsegments, ntsosegments;
1498 
1499 	nsegments = scctx->isc_tx_nsegments;
1500 	ntsosegments = scctx->isc_tx_tso_segments_max;
1501 	MPASS(scctx->isc_ntxd[0] > 0);
1502 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1503 	MPASS(nsegments > 0);
1504 	MPASS(ntsosegments > 0);
1505 	/*
1506 	 * Setup DMA descriptor areas.
1507 	 */
1508 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1509 			       1, 0,			/* alignment, bounds */
1510 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1511 			       BUS_SPACE_MAXADDR,	/* highaddr */
1512 			       NULL, NULL,		/* filter, filterarg */
1513 			       sctx->isc_tx_maxsize,		/* maxsize */
1514 			       nsegments,	/* nsegments */
1515 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1516 			       0,			/* flags */
1517 			       NULL,			/* lockfunc */
1518 			       NULL,			/* lockfuncarg */
1519 			       &txq->ift_desc_tag))) {
1520 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1521 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1522 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1523 		goto fail;
1524 	}
1525 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1526 			       1, 0,			/* alignment, bounds */
1527 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1528 			       BUS_SPACE_MAXADDR,	/* highaddr */
1529 			       NULL, NULL,		/* filter, filterarg */
1530 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1531 			       ntsosegments,	/* nsegments */
1532 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1533 			       0,			/* flags */
1534 			       NULL,			/* lockfunc */
1535 			       NULL,			/* lockfuncarg */
1536 			       &txq->ift_tso_desc_tag))) {
1537 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1538 
1539 		goto fail;
1540 	}
1541 	if (!(txq->ift_sds.ifsd_flags =
1542 	    (uint8_t *) malloc(sizeof(uint8_t) *
1543 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1544 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1545 		err = ENOMEM;
1546 		goto fail;
1547 	}
1548 	if (!(txq->ift_sds.ifsd_m =
1549 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1550 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1551 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1552 		err = ENOMEM;
1553 		goto fail;
1554 	}
1555 
1556         /* Create the descriptor buffer dma maps */
1557 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1558 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1559 		return (0);
1560 
1561 	if (!(txq->ift_sds.ifsd_map =
1562 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1563 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1564 		err = ENOMEM;
1565 		goto fail;
1566 	}
1567 
1568 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1569 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1570 		if (err != 0) {
1571 			device_printf(dev, "Unable to create TX DMA map\n");
1572 			goto fail;
1573 		}
1574 	}
1575 #endif
1576 	return (0);
1577 fail:
1578 	/* We free all, it handles case where we are in the middle */
1579 	iflib_tx_structures_free(ctx);
1580 	return (err);
1581 }
1582 
1583 static void
1584 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1585 {
1586 	bus_dmamap_t map;
1587 
1588 	map = NULL;
1589 	if (txq->ift_sds.ifsd_map != NULL)
1590 		map = txq->ift_sds.ifsd_map[i];
1591 	if (map != NULL) {
1592 		bus_dmamap_unload(txq->ift_desc_tag, map);
1593 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1594 		txq->ift_sds.ifsd_map[i] = NULL;
1595 	}
1596 }
1597 
1598 static void
1599 iflib_txq_destroy(iflib_txq_t txq)
1600 {
1601 	if_ctx_t ctx = txq->ift_ctx;
1602 
1603 	for (int i = 0; i < txq->ift_size; i++)
1604 		iflib_txsd_destroy(ctx, txq, i);
1605 	if (txq->ift_sds.ifsd_map != NULL) {
1606 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1607 		txq->ift_sds.ifsd_map = NULL;
1608 	}
1609 	if (txq->ift_sds.ifsd_m != NULL) {
1610 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1611 		txq->ift_sds.ifsd_m = NULL;
1612 	}
1613 	if (txq->ift_sds.ifsd_flags != NULL) {
1614 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1615 		txq->ift_sds.ifsd_flags = NULL;
1616 	}
1617 	if (txq->ift_desc_tag != NULL) {
1618 		bus_dma_tag_destroy(txq->ift_desc_tag);
1619 		txq->ift_desc_tag = NULL;
1620 	}
1621 	if (txq->ift_tso_desc_tag != NULL) {
1622 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1623 		txq->ift_tso_desc_tag = NULL;
1624 	}
1625 }
1626 
1627 static void
1628 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1629 {
1630 	struct mbuf **mp;
1631 
1632 	mp = &txq->ift_sds.ifsd_m[i];
1633 	if (*mp == NULL)
1634 		return;
1635 
1636 	if (txq->ift_sds.ifsd_map != NULL) {
1637 		bus_dmamap_sync(txq->ift_desc_tag,
1638 				txq->ift_sds.ifsd_map[i],
1639 				BUS_DMASYNC_POSTWRITE);
1640 		bus_dmamap_unload(txq->ift_desc_tag,
1641 				  txq->ift_sds.ifsd_map[i]);
1642 	}
1643 	m_free(*mp);
1644 	DBG_COUNTER_INC(tx_frees);
1645 	*mp = NULL;
1646 }
1647 
1648 static int
1649 iflib_txq_setup(iflib_txq_t txq)
1650 {
1651 	if_ctx_t ctx = txq->ift_ctx;
1652 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1653 	iflib_dma_info_t di;
1654 	int i;
1655 
1656 	/* Set number of descriptors available */
1657 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1658 	/* XXX make configurable */
1659 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1660 
1661 	/* Reset indices */
1662 	txq->ift_cidx_processed = 0;
1663 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1664 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1665 
1666 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1667 		bzero((void *)di->idi_vaddr, di->idi_size);
1668 
1669 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1670 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1671 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1672 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1673 	return (0);
1674 }
1675 
1676 /*********************************************************************
1677  *
1678  *  Allocate memory for rx_buffer structures. Since we use one
1679  *  rx_buffer per received packet, the maximum number of rx_buffer's
1680  *  that we'll need is equal to the number of receive descriptors
1681  *  that we've allocated.
1682  *
1683  **********************************************************************/
1684 static int
1685 iflib_rxsd_alloc(iflib_rxq_t rxq)
1686 {
1687 	if_ctx_t ctx = rxq->ifr_ctx;
1688 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1689 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1690 	device_t dev = ctx->ifc_dev;
1691 	iflib_fl_t fl;
1692 	int			err;
1693 
1694 	MPASS(scctx->isc_nrxd[0] > 0);
1695 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1696 
1697 	fl = rxq->ifr_fl;
1698 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1699 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1700 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1701 					 1, 0,			/* alignment, bounds */
1702 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1703 					 BUS_SPACE_MAXADDR,	/* highaddr */
1704 					 NULL, NULL,		/* filter, filterarg */
1705 					 sctx->isc_rx_maxsize,	/* maxsize */
1706 					 sctx->isc_rx_nsegments,	/* nsegments */
1707 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1708 					 0,			/* flags */
1709 					 NULL,			/* lockfunc */
1710 					 NULL,			/* lockarg */
1711 					 &fl->ifl_desc_tag);
1712 		if (err) {
1713 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1714 				__func__, err);
1715 			goto fail;
1716 		}
1717 		if (!(fl->ifl_sds.ifsd_flags =
1718 		      (uint8_t *) malloc(sizeof(uint8_t) *
1719 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1720 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1721 			err = ENOMEM;
1722 			goto fail;
1723 		}
1724 		if (!(fl->ifl_sds.ifsd_m =
1725 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1726 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1727 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1728 			err = ENOMEM;
1729 			goto fail;
1730 		}
1731 		if (!(fl->ifl_sds.ifsd_cl =
1732 		      (caddr_t *) malloc(sizeof(caddr_t) *
1733 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1734 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1735 			err = ENOMEM;
1736 			goto fail;
1737 		}
1738 
1739 		/* Create the descriptor buffer dma maps */
1740 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1741 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1742 			continue;
1743 
1744 		if (!(fl->ifl_sds.ifsd_map =
1745 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1746 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1747 			err = ENOMEM;
1748 			goto fail;
1749 		}
1750 
1751 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1752 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1753 			if (err != 0) {
1754 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1755 				goto fail;
1756 			}
1757 		}
1758 #endif
1759 	}
1760 	return (0);
1761 
1762 fail:
1763 	iflib_rx_structures_free(ctx);
1764 	return (err);
1765 }
1766 
1767 
1768 /*
1769  * Internal service routines
1770  */
1771 
1772 struct rxq_refill_cb_arg {
1773 	int               error;
1774 	bus_dma_segment_t seg;
1775 	int               nseg;
1776 };
1777 
1778 static void
1779 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1780 {
1781 	struct rxq_refill_cb_arg *cb_arg = arg;
1782 
1783 	cb_arg->error = error;
1784 	cb_arg->seg = segs[0];
1785 	cb_arg->nseg = nseg;
1786 }
1787 
1788 
1789 #ifdef ACPI_DMAR
1790 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1791 #else
1792 #define IS_DMAR(ctx) (0)
1793 #endif
1794 
1795 /**
1796  *	rxq_refill - refill an rxq  free-buffer list
1797  *	@ctx: the iflib context
1798  *	@rxq: the free-list to refill
1799  *	@n: the number of new buffers to allocate
1800  *
1801  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1802  *	The caller must assure that @n does not exceed the queue's capacity.
1803  */
1804 static void
1805 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1806 {
1807 	struct mbuf *m;
1808 	int idx, pidx = fl->ifl_pidx;
1809 	caddr_t cl, *sd_cl;
1810 	struct mbuf **sd_m;
1811 	uint8_t *sd_flags;
1812 	struct if_rxd_update iru;
1813 	bus_dmamap_t *sd_map;
1814 	int n, i = 0;
1815 	uint64_t bus_addr;
1816 	int err;
1817 
1818 	sd_m = fl->ifl_sds.ifsd_m;
1819 	sd_map = fl->ifl_sds.ifsd_map;
1820 	sd_cl = fl->ifl_sds.ifsd_cl;
1821 	sd_flags = fl->ifl_sds.ifsd_flags;
1822 	idx = pidx;
1823 
1824 	n  = count;
1825 	MPASS(n > 0);
1826 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1827 
1828 	if (pidx < fl->ifl_cidx)
1829 		MPASS(pidx + n <= fl->ifl_cidx);
1830 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1831 		MPASS(fl->ifl_gen == 0);
1832 	if (pidx > fl->ifl_cidx)
1833 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1834 
1835 	DBG_COUNTER_INC(fl_refills);
1836 	if (n > 8)
1837 		DBG_COUNTER_INC(fl_refills_large);
1838 	iru.iru_paddrs = fl->ifl_bus_addrs;
1839 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1840 	iru.iru_idxs = fl->ifl_rxd_idxs;
1841 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1842 	iru.iru_buf_size = fl->ifl_buf_size;
1843 	iru.iru_flidx = fl->ifl_id;
1844 	while (n--) {
1845 		/*
1846 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1847 		 * initialized after rx.
1848 		 *
1849 		 * If the cluster is still set then we know a minimum sized packet was received
1850 		 */
1851 		if ((cl = sd_cl[idx]) == NULL) {
1852 			if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1853 				break;
1854 #if MEMORY_LOGGING
1855 			fl->ifl_cl_enqueued++;
1856 #endif
1857 		}
1858 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1859 			break;
1860 		}
1861 #if MEMORY_LOGGING
1862 		fl->ifl_m_enqueued++;
1863 #endif
1864 
1865 		DBG_COUNTER_INC(rx_allocs);
1866 #if defined(__i386__) || defined(__amd64__)
1867 		if (!IS_DMAR(ctx)) {
1868 			bus_addr = pmap_kextract((vm_offset_t)cl);
1869 		} else
1870 #endif
1871 		{
1872 			struct rxq_refill_cb_arg cb_arg;
1873 			iflib_rxq_t q;
1874 
1875 			cb_arg.error = 0;
1876 			q = fl->ifl_rxq;
1877 			MPASS(sd_map != NULL);
1878 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1879 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1880 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[idx], BUS_DMASYNC_PREREAD);
1881 
1882 			if (err != 0 || cb_arg.error) {
1883 				/*
1884 				 * !zone_pack ?
1885 				 */
1886 				if (fl->ifl_zone == zone_pack)
1887 					uma_zfree(fl->ifl_zone, cl);
1888 				m_free(m);
1889 				n = 0;
1890 				goto done;
1891 			}
1892 			bus_addr = cb_arg.seg.ds_addr;
1893 		}
1894 		sd_flags[idx] |= RX_SW_DESC_INUSE;
1895 
1896 		MPASS(sd_m[idx] == NULL);
1897 		sd_cl[idx] = cl;
1898 		sd_m[idx] = m;
1899 		fl->ifl_rxd_idxs[i] = idx;
1900 		fl->ifl_bus_addrs[i] = bus_addr;
1901 		fl->ifl_vm_addrs[i] = cl;
1902 		fl->ifl_credits++;
1903 		i++;
1904 		MPASS(fl->ifl_credits <= fl->ifl_size);
1905 		if (++idx == fl->ifl_size) {
1906 			fl->ifl_gen = 1;
1907 			idx = 0;
1908 		}
1909 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1910 			iru.iru_pidx = pidx;
1911 			iru.iru_count = i;
1912 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1913 			i = 0;
1914 			pidx = idx;
1915 		}
1916 		fl->ifl_pidx = idx;
1917 
1918 	}
1919 done:
1920 	DBG_COUNTER_INC(rxd_flush);
1921 	if (fl->ifl_pidx == 0)
1922 		pidx = fl->ifl_size - 1;
1923 	else
1924 		pidx = fl->ifl_pidx - 1;
1925 
1926 	if (sd_map)
1927 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1928 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1929 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1930 }
1931 
1932 static __inline void
1933 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1934 {
1935 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1936 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1937 #ifdef INVARIANTS
1938 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1939 #endif
1940 
1941 	MPASS(fl->ifl_credits <= fl->ifl_size);
1942 	MPASS(reclaimable == delta);
1943 
1944 	if (reclaimable > 0)
1945 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1946 }
1947 
1948 static void
1949 iflib_fl_bufs_free(iflib_fl_t fl)
1950 {
1951 	iflib_dma_info_t idi = fl->ifl_ifdi;
1952 	uint32_t i;
1953 
1954 	for (i = 0; i < fl->ifl_size; i++) {
1955 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1956 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1957 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1958 
1959 		if (*sd_flags & RX_SW_DESC_INUSE) {
1960 			if (fl->ifl_sds.ifsd_map != NULL) {
1961 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1962 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1963 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1964 			}
1965 			if (*sd_m != NULL) {
1966 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1967 				uma_zfree(zone_mbuf, *sd_m);
1968 			}
1969 			if (*sd_cl != NULL)
1970 				uma_zfree(fl->ifl_zone, *sd_cl);
1971 			*sd_flags = 0;
1972 		} else {
1973 			MPASS(*sd_cl == NULL);
1974 			MPASS(*sd_m == NULL);
1975 		}
1976 #if MEMORY_LOGGING
1977 		fl->ifl_m_dequeued++;
1978 		fl->ifl_cl_dequeued++;
1979 #endif
1980 		*sd_cl = NULL;
1981 		*sd_m = NULL;
1982 	}
1983 #ifdef INVARIANTS
1984 	for (i = 0; i < fl->ifl_size; i++) {
1985 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1986 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1987 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1988 	}
1989 #endif
1990 	/*
1991 	 * Reset free list values
1992 	 */
1993 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1994 	bzero(idi->idi_vaddr, idi->idi_size);
1995 }
1996 
1997 /*********************************************************************
1998  *
1999  *  Initialize a receive ring and its buffers.
2000  *
2001  **********************************************************************/
2002 static int
2003 iflib_fl_setup(iflib_fl_t fl)
2004 {
2005 	iflib_rxq_t rxq = fl->ifl_rxq;
2006 	if_ctx_t ctx = rxq->ifr_ctx;
2007 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2008 
2009 	/*
2010 	** Free current RX buffer structs and their mbufs
2011 	*/
2012 	iflib_fl_bufs_free(fl);
2013 	/* Now replenish the mbufs */
2014 	MPASS(fl->ifl_credits == 0);
2015 	/*
2016 	 * XXX don't set the max_frame_size to larger
2017 	 * than the hardware can handle
2018 	 */
2019 	if (sctx->isc_max_frame_size <= 2048)
2020 		fl->ifl_buf_size = MCLBYTES;
2021 #ifndef CONTIGMALLOC_WORKS
2022 	else
2023 		fl->ifl_buf_size = MJUMPAGESIZE;
2024 #else
2025 	else if (sctx->isc_max_frame_size <= 4096)
2026 		fl->ifl_buf_size = MJUMPAGESIZE;
2027 	else if (sctx->isc_max_frame_size <= 9216)
2028 		fl->ifl_buf_size = MJUM9BYTES;
2029 	else
2030 		fl->ifl_buf_size = MJUM16BYTES;
2031 #endif
2032 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2033 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2034 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2035 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2036 
2037 
2038 	/* avoid pre-allocating zillions of clusters to an idle card
2039 	 * potentially speeding up attach
2040 	 */
2041 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2042 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2043 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2044 		return (ENOBUFS);
2045 	/*
2046 	 * handle failure
2047 	 */
2048 	MPASS(rxq != NULL);
2049 	MPASS(fl->ifl_ifdi != NULL);
2050 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2051 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2052 	return (0);
2053 }
2054 
2055 /*********************************************************************
2056  *
2057  *  Free receive ring data structures
2058  *
2059  **********************************************************************/
2060 static void
2061 iflib_rx_sds_free(iflib_rxq_t rxq)
2062 {
2063 	iflib_fl_t fl;
2064 	int i;
2065 
2066 	if (rxq->ifr_fl != NULL) {
2067 		for (i = 0; i < rxq->ifr_nfl; i++) {
2068 			fl = &rxq->ifr_fl[i];
2069 			if (fl->ifl_desc_tag != NULL) {
2070 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2071 				fl->ifl_desc_tag = NULL;
2072 			}
2073 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2074 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2075 			/* XXX destroy maps first */
2076 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2077 			fl->ifl_sds.ifsd_m = NULL;
2078 			fl->ifl_sds.ifsd_cl = NULL;
2079 			fl->ifl_sds.ifsd_map = NULL;
2080 		}
2081 		free(rxq->ifr_fl, M_IFLIB);
2082 		rxq->ifr_fl = NULL;
2083 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2084 	}
2085 }
2086 
2087 /*
2088  * MI independent logic
2089  *
2090  */
2091 static void
2092 iflib_timer(void *arg)
2093 {
2094 	iflib_txq_t txq = arg;
2095 	if_ctx_t ctx = txq->ift_ctx;
2096 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2097 
2098 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2099 		return;
2100 	/*
2101 	** Check on the state of the TX queue(s), this
2102 	** can be done without the lock because its RO
2103 	** and the HUNG state will be static if set.
2104 	*/
2105 	IFDI_TIMER(ctx, txq->ift_id);
2106 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2107 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2108 	     (sctx->isc_pause_frames == 0)))
2109 		goto hung;
2110 
2111 	if (ifmp_ring_is_stalled(txq->ift_br))
2112 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2113 	txq->ift_cleaned_prev = txq->ift_cleaned;
2114 	/* handle any laggards */
2115 	if (txq->ift_db_pending)
2116 		GROUPTASK_ENQUEUE(&txq->ift_task);
2117 
2118 	sctx->isc_pause_frames = 0;
2119 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2120 		callout_reset_on(&txq->ift_timer, iflib_timer_int, iflib_timer,
2121 		    txq, txq->ift_timer.c_cpu);
2122 	return;
2123 hung:
2124 	CTX_LOCK(ctx);
2125 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2126 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2127 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2128 
2129 	IFDI_WATCHDOG_RESET(ctx);
2130 	ctx->ifc_watchdog_events++;
2131 
2132 	ctx->ifc_flags |= IFC_DO_RESET;
2133 	iflib_admin_intr_deferred(ctx);
2134 	CTX_UNLOCK(ctx);
2135 }
2136 
2137 static void
2138 iflib_init_locked(if_ctx_t ctx)
2139 {
2140 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2141 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2142 	if_t ifp = ctx->ifc_ifp;
2143 	iflib_fl_t fl;
2144 	iflib_txq_t txq;
2145 	iflib_rxq_t rxq;
2146 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2147 
2148 
2149 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2150 	IFDI_INTR_DISABLE(ctx);
2151 
2152 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2153 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2154 	/* Set hardware offload abilities */
2155 	if_clearhwassist(ifp);
2156 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2157 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2158 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2159 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2160 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2161 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2162 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2163 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2164 
2165 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2166 		CALLOUT_LOCK(txq);
2167 		callout_stop(&txq->ift_timer);
2168 		CALLOUT_UNLOCK(txq);
2169 		iflib_netmap_txq_init(ctx, txq);
2170 	}
2171 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2172 		MPASS(rxq->ifr_id == i);
2173 		iflib_netmap_rxq_init(ctx, rxq);
2174 	}
2175 #ifdef INVARIANTS
2176 	i = if_getdrvflags(ifp);
2177 #endif
2178 	IFDI_INIT(ctx);
2179 	MPASS(if_getdrvflags(ifp) == i);
2180 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2181 		/* XXX this should really be done on a per-queue basis */
2182 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2183 			continue;
2184 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2185 			if (iflib_fl_setup(fl)) {
2186 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2187 				goto done;
2188 			}
2189 		}
2190 	}
2191 	done:
2192 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2193 	IFDI_INTR_ENABLE(ctx);
2194 	txq = ctx->ifc_txqs;
2195 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2196 		callout_reset_on(&txq->ift_timer, iflib_timer_int, iflib_timer,
2197 			txq, txq->ift_timer.c_cpu);
2198 }
2199 
2200 static int
2201 iflib_media_change(if_t ifp)
2202 {
2203 	if_ctx_t ctx = if_getsoftc(ifp);
2204 	int err;
2205 
2206 	CTX_LOCK(ctx);
2207 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2208 		iflib_init_locked(ctx);
2209 	CTX_UNLOCK(ctx);
2210 	return (err);
2211 }
2212 
2213 static void
2214 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2215 {
2216 	if_ctx_t ctx = if_getsoftc(ifp);
2217 
2218 	CTX_LOCK(ctx);
2219 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2220 	IFDI_MEDIA_STATUS(ctx, ifmr);
2221 	CTX_UNLOCK(ctx);
2222 }
2223 
2224 static void
2225 iflib_stop(if_ctx_t ctx)
2226 {
2227 	iflib_txq_t txq = ctx->ifc_txqs;
2228 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2229 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2230 	iflib_dma_info_t di;
2231 	iflib_fl_t fl;
2232 	int i, j;
2233 
2234 	/* Tell the stack that the interface is no longer active */
2235 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2236 
2237 	IFDI_INTR_DISABLE(ctx);
2238 	DELAY(1000);
2239 	IFDI_STOP(ctx);
2240 	DELAY(1000);
2241 
2242 	iflib_debug_reset();
2243 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2244 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2245 		/* make sure all transmitters have completed before proceeding XXX */
2246 
2247 		/* clean any enqueued buffers */
2248 		iflib_ifmp_purge(txq);
2249 		/* Free any existing tx buffers. */
2250 		for (j = 0; j < txq->ift_size; j++) {
2251 			iflib_txsd_free(ctx, txq, j);
2252 		}
2253 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2254 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2255 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2256 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2257 		txq->ift_pullups = 0;
2258 		ifmp_ring_reset_stats(txq->ift_br);
2259 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2260 			bzero((void *)di->idi_vaddr, di->idi_size);
2261 	}
2262 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2263 		/* make sure all transmitters have completed before proceeding XXX */
2264 
2265 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2266 			bzero((void *)di->idi_vaddr, di->idi_size);
2267 		/* also resets the free lists pidx/cidx */
2268 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2269 			iflib_fl_bufs_free(fl);
2270 	}
2271 }
2272 
2273 static inline caddr_t
2274 calc_next_rxd(iflib_fl_t fl, int cidx)
2275 {
2276 	qidx_t size;
2277 	int nrxd;
2278 	caddr_t start, end, cur, next;
2279 
2280 	nrxd = fl->ifl_size;
2281 	size = fl->ifl_rxd_size;
2282 	start = fl->ifl_ifdi->idi_vaddr;
2283 
2284 	if (__predict_false(size == 0))
2285 		return (start);
2286 	cur = start + size*cidx;
2287 	end = start + size*nrxd;
2288 	next = CACHE_PTR_NEXT(cur);
2289 	return (next < end ? next : start);
2290 }
2291 
2292 static inline void
2293 prefetch_pkts(iflib_fl_t fl, int cidx)
2294 {
2295 	int nextptr;
2296 	int nrxd = fl->ifl_size;
2297 	caddr_t next_rxd;
2298 
2299 
2300 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2301 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2302 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2303 	next_rxd = calc_next_rxd(fl, cidx);
2304 	prefetch(next_rxd);
2305 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2306 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2307 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2308 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2309 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2310 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2311 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2312 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2313 }
2314 
2315 static void
2316 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2317 {
2318 	int flid, cidx;
2319 	bus_dmamap_t map;
2320 	iflib_fl_t fl;
2321 	iflib_dma_info_t di;
2322 	int next;
2323 
2324 	map = NULL;
2325 	flid = irf->irf_flid;
2326 	cidx = irf->irf_idx;
2327 	fl = &rxq->ifr_fl[flid];
2328 	sd->ifsd_fl = fl;
2329 	sd->ifsd_cidx = cidx;
2330 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2331 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2332 	fl->ifl_credits--;
2333 #if MEMORY_LOGGING
2334 	fl->ifl_m_dequeued++;
2335 #endif
2336 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2337 		prefetch_pkts(fl, cidx);
2338 	if (fl->ifl_sds.ifsd_map != NULL) {
2339 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2340 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2341 		map = fl->ifl_sds.ifsd_map[cidx];
2342 		di = fl->ifl_ifdi;
2343 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2344 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2345 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2346 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2347 
2348 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2349 		MPASS(fl->ifl_cidx == cidx);
2350 		if (unload)
2351 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2352 	}
2353 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2354 	if (__predict_false(fl->ifl_cidx == 0))
2355 		fl->ifl_gen = 0;
2356 	if (map != NULL)
2357 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2358 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2359 }
2360 
2361 static struct mbuf *
2362 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2363 {
2364 	int i, padlen , flags;
2365 	struct mbuf *m, *mh, *mt;
2366 	caddr_t cl;
2367 
2368 	i = 0;
2369 	mh = NULL;
2370 	do {
2371 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2372 
2373 		MPASS(*sd->ifsd_cl != NULL);
2374 		MPASS(*sd->ifsd_m != NULL);
2375 
2376 		/* Don't include zero-length frags */
2377 		if (ri->iri_frags[i].irf_len == 0) {
2378 			/* XXX we can save the cluster here, but not the mbuf */
2379 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2380 			m_free(*sd->ifsd_m);
2381 			*sd->ifsd_m = NULL;
2382 			continue;
2383 		}
2384 		m = *sd->ifsd_m;
2385 		*sd->ifsd_m = NULL;
2386 		if (mh == NULL) {
2387 			flags = M_PKTHDR|M_EXT;
2388 			mh = mt = m;
2389 			padlen = ri->iri_pad;
2390 		} else {
2391 			flags = M_EXT;
2392 			mt->m_next = m;
2393 			mt = m;
2394 			/* assuming padding is only on the first fragment */
2395 			padlen = 0;
2396 		}
2397 		cl = *sd->ifsd_cl;
2398 		*sd->ifsd_cl = NULL;
2399 
2400 		/* Can these two be made one ? */
2401 		m_init(m, M_NOWAIT, MT_DATA, flags);
2402 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2403 		/*
2404 		 * These must follow m_init and m_cljset
2405 		 */
2406 		m->m_data += padlen;
2407 		ri->iri_len -= padlen;
2408 		m->m_len = ri->iri_frags[i].irf_len;
2409 	} while (++i < ri->iri_nfrags);
2410 
2411 	return (mh);
2412 }
2413 
2414 /*
2415  * Process one software descriptor
2416  */
2417 static struct mbuf *
2418 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2419 {
2420 	struct if_rxsd sd;
2421 	struct mbuf *m;
2422 
2423 	/* should I merge this back in now that the two paths are basically duplicated? */
2424 	if (ri->iri_nfrags == 1 &&
2425 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2426 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2427 		m = *sd.ifsd_m;
2428 		*sd.ifsd_m = NULL;
2429 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2430 #ifndef __NO_STRICT_ALIGNMENT
2431 		if (!IP_ALIGNED(m))
2432 			m->m_data += 2;
2433 #endif
2434 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2435 		m->m_len = ri->iri_frags[0].irf_len;
2436        } else {
2437 		m = assemble_segments(rxq, ri, &sd);
2438 	}
2439 	m->m_pkthdr.len = ri->iri_len;
2440 	m->m_pkthdr.rcvif = ri->iri_ifp;
2441 	m->m_flags |= ri->iri_flags;
2442 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2443 	m->m_pkthdr.flowid = ri->iri_flowid;
2444 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2445 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2446 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2447 	return (m);
2448 }
2449 
2450 static bool
2451 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2452 {
2453 	if_ctx_t ctx = rxq->ifr_ctx;
2454 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2455 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2456 	int avail, i;
2457 	qidx_t *cidxp;
2458 	struct if_rxd_info ri;
2459 	int err, budget_left, rx_bytes, rx_pkts;
2460 	iflib_fl_t fl;
2461 	struct ifnet *ifp;
2462 	int lro_enabled;
2463 
2464 	/*
2465 	 * XXX early demux data packets so that if_input processing only handles
2466 	 * acks in interrupt context
2467 	 */
2468 	struct mbuf *m, *mh, *mt;
2469 
2470 	ifp = ctx->ifc_ifp;
2471 #ifdef DEV_NETMAP
2472 	if (ifp->if_capenable & IFCAP_NETMAP) {
2473 		u_int work = 0;
2474 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2475 			return (FALSE);
2476 	}
2477 #endif
2478 
2479 	mh = mt = NULL;
2480 	MPASS(budget > 0);
2481 	rx_pkts	= rx_bytes = 0;
2482 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2483 		cidxp = &rxq->ifr_cq_cidx;
2484 	else
2485 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2486 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2487 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2488 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2489 		DBG_COUNTER_INC(rx_unavail);
2490 		return (false);
2491 	}
2492 
2493 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2494 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2495 			DBG_COUNTER_INC(rx_ctx_inactive);
2496 			break;
2497 		}
2498 		/*
2499 		 * Reset client set fields to their default values
2500 		 */
2501 		rxd_info_zero(&ri);
2502 		ri.iri_qsidx = rxq->ifr_id;
2503 		ri.iri_cidx = *cidxp;
2504 		ri.iri_ifp = ifp;
2505 		ri.iri_frags = rxq->ifr_frags;
2506 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2507 
2508 		if (err)
2509 			goto err;
2510 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2511 			*cidxp = ri.iri_cidx;
2512 			/* Update our consumer index */
2513 			/* XXX NB: shurd - check if this is still safe */
2514 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2515 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2516 				rxq->ifr_cq_gen = 0;
2517 			}
2518 			/* was this only a completion queue message? */
2519 			if (__predict_false(ri.iri_nfrags == 0))
2520 				continue;
2521 		}
2522 		MPASS(ri.iri_nfrags != 0);
2523 		MPASS(ri.iri_len != 0);
2524 
2525 		/* will advance the cidx on the corresponding free lists */
2526 		m = iflib_rxd_pkt_get(rxq, &ri);
2527 		if (avail == 0 && budget_left)
2528 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2529 
2530 		if (__predict_false(m == NULL)) {
2531 			DBG_COUNTER_INC(rx_mbuf_null);
2532 			continue;
2533 		}
2534 		/* imm_pkt: -- cxgb */
2535 		if (mh == NULL)
2536 			mh = mt = m;
2537 		else {
2538 			mt->m_nextpkt = m;
2539 			mt = m;
2540 		}
2541 	}
2542 	/* make sure that we can refill faster than drain */
2543 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2544 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2545 
2546 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2547 	while (mh != NULL) {
2548 		m = mh;
2549 		mh = mh->m_nextpkt;
2550 		m->m_nextpkt = NULL;
2551 #ifndef __NO_STRICT_ALIGNMENT
2552 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2553 			continue;
2554 #endif
2555 		rx_bytes += m->m_pkthdr.len;
2556 		rx_pkts++;
2557 #if defined(INET6) || defined(INET)
2558 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2559 			continue;
2560 #endif
2561 		DBG_COUNTER_INC(rx_if_input);
2562 		ifp->if_input(ifp, m);
2563 	}
2564 
2565 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2566 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2567 
2568 	/*
2569 	 * Flush any outstanding LRO work
2570 	 */
2571 #if defined(INET6) || defined(INET)
2572 	tcp_lro_flush_all(&rxq->ifr_lc);
2573 #endif
2574 	if (avail)
2575 		return true;
2576 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2577 err:
2578 	CTX_LOCK(ctx);
2579 	ctx->ifc_flags |= IFC_DO_RESET;
2580 	iflib_admin_intr_deferred(ctx);
2581 	CTX_UNLOCK(ctx);
2582 	return (false);
2583 }
2584 
2585 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2586 static inline qidx_t
2587 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2588 {
2589 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2590 	qidx_t minthresh = txq->ift_size / 8;
2591 	if (in_use > 4*minthresh)
2592 		return (notify_count);
2593 	if (in_use > 2*minthresh)
2594 		return (notify_count >> 1);
2595 	if (in_use > minthresh)
2596 		return (notify_count >> 3);
2597 	return (0);
2598 }
2599 
2600 static inline qidx_t
2601 txq_max_rs_deferred(iflib_txq_t txq)
2602 {
2603 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2604 	qidx_t minthresh = txq->ift_size / 8;
2605 	if (txq->ift_in_use > 4*minthresh)
2606 		return (notify_count);
2607 	if (txq->ift_in_use > 2*minthresh)
2608 		return (notify_count >> 1);
2609 	if (txq->ift_in_use > minthresh)
2610 		return (notify_count >> 2);
2611 	return (2);
2612 }
2613 
2614 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2615 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2616 
2617 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2618 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2619 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2620 
2621 /* forward compatibility for cxgb */
2622 #define FIRST_QSET(ctx) 0
2623 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2624 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2625 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2626 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2627 
2628 /* XXX we should be setting this to something other than zero */
2629 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2630 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2631 
2632 static inline bool
2633 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2634 {
2635 	qidx_t dbval, max;
2636 	bool rang;
2637 
2638 	rang = false;
2639 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2640 	if (ring || txq->ift_db_pending >= max) {
2641 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2642 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2643 		txq->ift_db_pending = txq->ift_npending = 0;
2644 		rang = true;
2645 	}
2646 	return (rang);
2647 }
2648 
2649 #ifdef PKT_DEBUG
2650 static void
2651 print_pkt(if_pkt_info_t pi)
2652 {
2653 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2654 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2655 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2656 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2657 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2658 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2659 }
2660 #endif
2661 
2662 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2663 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2664 
2665 static int
2666 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2667 {
2668 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2669 	struct ether_vlan_header *eh;
2670 	struct mbuf *m, *n;
2671 
2672 	n = m = *mp;
2673 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2674 	    M_WRITABLE(m) == 0) {
2675 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2676 			return (ENOMEM);
2677 		} else {
2678 			m_freem(*mp);
2679 			n = *mp = m;
2680 		}
2681 	}
2682 
2683 	/*
2684 	 * Determine where frame payload starts.
2685 	 * Jump over vlan headers if already present,
2686 	 * helpful for QinQ too.
2687 	 */
2688 	if (__predict_false(m->m_len < sizeof(*eh))) {
2689 		txq->ift_pullups++;
2690 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2691 			return (ENOMEM);
2692 	}
2693 	eh = mtod(m, struct ether_vlan_header *);
2694 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2695 		pi->ipi_etype = ntohs(eh->evl_proto);
2696 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2697 	} else {
2698 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2699 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2700 	}
2701 
2702 	switch (pi->ipi_etype) {
2703 #ifdef INET
2704 	case ETHERTYPE_IP:
2705 	{
2706 		struct ip *ip = NULL;
2707 		struct tcphdr *th = NULL;
2708 		int minthlen;
2709 
2710 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2711 		if (__predict_false(m->m_len < minthlen)) {
2712 			/*
2713 			 * if this code bloat is causing too much of a hit
2714 			 * move it to a separate function and mark it noinline
2715 			 */
2716 			if (m->m_len == pi->ipi_ehdrlen) {
2717 				n = m->m_next;
2718 				MPASS(n);
2719 				if (n->m_len >= sizeof(*ip))  {
2720 					ip = (struct ip *)n->m_data;
2721 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2722 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2723 				} else {
2724 					txq->ift_pullups++;
2725 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2726 						return (ENOMEM);
2727 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2728 				}
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 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2735 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2736 			}
2737 		} else {
2738 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2739 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2740 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2741 		}
2742 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2743 		pi->ipi_ipproto = ip->ip_p;
2744 		pi->ipi_flags |= IPI_TX_IPV4;
2745 
2746 		if (pi->ipi_csum_flags & CSUM_IP)
2747                        ip->ip_sum = 0;
2748 
2749 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2750 			if (__predict_false(th == NULL)) {
2751 				txq->ift_pullups++;
2752 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2753 					return (ENOMEM);
2754 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2755 			}
2756 			pi->ipi_tcp_hflags = th->th_flags;
2757 			pi->ipi_tcp_hlen = th->th_off << 2;
2758 			pi->ipi_tcp_seq = th->th_seq;
2759 		}
2760 		if (IS_TSO4(pi)) {
2761 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2762 				return (ENXIO);
2763 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2764 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2765 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2766 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2767 				ip->ip_sum = 0;
2768 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2769 			}
2770 		}
2771 		break;
2772 	}
2773 #endif
2774 #ifdef INET6
2775 	case ETHERTYPE_IPV6:
2776 	{
2777 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2778 		struct tcphdr *th;
2779 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2780 
2781 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2782 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2783 				return (ENOMEM);
2784 		}
2785 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2786 
2787 		/* XXX-BZ this will go badly in case of ext hdrs. */
2788 		pi->ipi_ipproto = ip6->ip6_nxt;
2789 		pi->ipi_flags |= IPI_TX_IPV6;
2790 
2791 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2792 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2793 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2794 					return (ENOMEM);
2795 			}
2796 			pi->ipi_tcp_hflags = th->th_flags;
2797 			pi->ipi_tcp_hlen = th->th_off << 2;
2798 		}
2799 		if (IS_TSO6(pi)) {
2800 
2801 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2802 				return (ENXIO);
2803 			/*
2804 			 * The corresponding flag is set by the stack in the IPv4
2805 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2806 			 * So, set it here because the rest of the flow requires it.
2807 			 */
2808 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2809 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2810 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2811 		}
2812 		break;
2813 	}
2814 #endif
2815 	default:
2816 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2817 		pi->ipi_ip_hlen = 0;
2818 		break;
2819 	}
2820 	*mp = m;
2821 
2822 	return (0);
2823 }
2824 
2825 static  __noinline  struct mbuf *
2826 collapse_pkthdr(struct mbuf *m0)
2827 {
2828 	struct mbuf *m, *m_next, *tmp;
2829 
2830 	m = m0;
2831 	m_next = m->m_next;
2832 	while (m_next != NULL && m_next->m_len == 0) {
2833 		m = m_next;
2834 		m->m_next = NULL;
2835 		m_free(m);
2836 		m_next = m_next->m_next;
2837 	}
2838 	m = m0;
2839 	m->m_next = m_next;
2840 	if ((m_next->m_flags & M_EXT) == 0) {
2841 		m = m_defrag(m, M_NOWAIT);
2842 	} else {
2843 		tmp = m_next->m_next;
2844 		memcpy(m_next, m, MPKTHSIZE);
2845 		m = m_next;
2846 		m->m_next = tmp;
2847 	}
2848 	return (m);
2849 }
2850 
2851 /*
2852  * If dodgy hardware rejects the scatter gather chain we've handed it
2853  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2854  * m_defrag'd mbufs
2855  */
2856 static __noinline struct mbuf *
2857 iflib_remove_mbuf(iflib_txq_t txq)
2858 {
2859 	int ntxd, i, pidx;
2860 	struct mbuf *m, *mh, **ifsd_m;
2861 
2862 	pidx = txq->ift_pidx;
2863 	ifsd_m = txq->ift_sds.ifsd_m;
2864 	ntxd = txq->ift_size;
2865 	mh = m = ifsd_m[pidx];
2866 	ifsd_m[pidx] = NULL;
2867 #if MEMORY_LOGGING
2868 	txq->ift_dequeued++;
2869 #endif
2870 	i = 1;
2871 
2872 	while (m) {
2873 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2874 #if MEMORY_LOGGING
2875 		txq->ift_dequeued++;
2876 #endif
2877 		m = m->m_next;
2878 		i++;
2879 	}
2880 	return (mh);
2881 }
2882 
2883 static int
2884 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2885 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2886 			  int max_segs, int flags)
2887 {
2888 	if_ctx_t ctx;
2889 	if_shared_ctx_t		sctx;
2890 	if_softc_ctx_t		scctx;
2891 	int i, next, pidx, err, maxsegsz, ntxd, count;
2892 	struct mbuf *m, *tmp, **ifsd_m;
2893 
2894 	m = *m0;
2895 
2896 	/*
2897 	 * Please don't ever do this
2898 	 */
2899 	if (__predict_false(m->m_len == 0))
2900 		*m0 = m = collapse_pkthdr(m);
2901 
2902 	ctx = txq->ift_ctx;
2903 	sctx = ctx->ifc_sctx;
2904 	scctx = &ctx->ifc_softc_ctx;
2905 	ifsd_m = txq->ift_sds.ifsd_m;
2906 	ntxd = txq->ift_size;
2907 	pidx = txq->ift_pidx;
2908 	if (force_busdma || map != NULL) {
2909 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2910 
2911 		err = bus_dmamap_load_mbuf_sg(tag, map,
2912 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2913 		if (err)
2914 			return (err);
2915 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2916 		count = 0;
2917 		m = *m0;
2918 		do {
2919 			if (__predict_false(m->m_len <= 0)) {
2920 				tmp = m;
2921 				m = m->m_next;
2922 				tmp->m_next = NULL;
2923 				m_free(tmp);
2924 				continue;
2925 			}
2926 			next = (pidx + count) & (ntxd-1);
2927 			MPASS(ifsd_m[next] == NULL);
2928 			ifsd_m[next] = m;
2929 			count++;
2930 			tmp = m;
2931 			m = m->m_next;
2932 		} while (m != NULL);
2933 	} else {
2934 		int buflen, sgsize, max_sgsize;
2935 		vm_offset_t vaddr;
2936 		vm_paddr_t curaddr;
2937 
2938 		count = i = 0;
2939 		maxsegsz = sctx->isc_tx_maxsize;
2940 		m = *m0;
2941 		do {
2942 			if (__predict_false(m->m_len <= 0)) {
2943 				tmp = m;
2944 				m = m->m_next;
2945 				tmp->m_next = NULL;
2946 				m_free(tmp);
2947 				continue;
2948 			}
2949 			buflen = m->m_len;
2950 			vaddr = (vm_offset_t)m->m_data;
2951 			/*
2952 			 * see if we can't be smarter about physically
2953 			 * contiguous mappings
2954 			 */
2955 			next = (pidx + count) & (ntxd-1);
2956 			MPASS(ifsd_m[next] == NULL);
2957 #if MEMORY_LOGGING
2958 			txq->ift_enqueued++;
2959 #endif
2960 			ifsd_m[next] = m;
2961 			while (buflen > 0) {
2962 				max_sgsize = MIN(buflen, maxsegsz);
2963 				curaddr = pmap_kextract(vaddr);
2964 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2965 				sgsize = MIN(sgsize, max_sgsize);
2966 				segs[i].ds_addr = curaddr;
2967 				segs[i].ds_len = sgsize;
2968 				vaddr += sgsize;
2969 				buflen -= sgsize;
2970 				i++;
2971 				if (i >= max_segs)
2972 					goto err;
2973 			}
2974 			count++;
2975 			tmp = m;
2976 			m = m->m_next;
2977 		} while (m != NULL);
2978 		*nsegs = i;
2979 	}
2980 	return (0);
2981 err:
2982 	*m0 = iflib_remove_mbuf(txq);
2983 	return (EFBIG);
2984 }
2985 
2986 static inline caddr_t
2987 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2988 {
2989 	qidx_t size;
2990 	int ntxd;
2991 	caddr_t start, end, cur, next;
2992 
2993 	ntxd = txq->ift_size;
2994 	size = txq->ift_txd_size[qid];
2995 	start = txq->ift_ifdi[qid].idi_vaddr;
2996 
2997 	if (__predict_false(size == 0))
2998 		return (start);
2999 	cur = start + size*cidx;
3000 	end = start + size*ntxd;
3001 	next = CACHE_PTR_NEXT(cur);
3002 	return (next < end ? next : start);
3003 }
3004 
3005 static int
3006 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3007 {
3008 	if_ctx_t		ctx;
3009 	if_shared_ctx_t		sctx;
3010 	if_softc_ctx_t		scctx;
3011 	bus_dma_segment_t	*segs;
3012 	struct mbuf		*m_head;
3013 	void			*next_txd;
3014 	bus_dmamap_t		map;
3015 	struct if_pkt_info	pi;
3016 	int remap = 0;
3017 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3018 	bus_dma_tag_t desc_tag;
3019 
3020 	segs = txq->ift_segs;
3021 	ctx = txq->ift_ctx;
3022 	sctx = ctx->ifc_sctx;
3023 	scctx = &ctx->ifc_softc_ctx;
3024 	segs = txq->ift_segs;
3025 	ntxd = txq->ift_size;
3026 	m_head = *m_headp;
3027 	map = NULL;
3028 
3029 	/*
3030 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3031 	 */
3032 	cidx = txq->ift_cidx;
3033 	pidx = txq->ift_pidx;
3034 	if (ctx->ifc_flags & IFC_PREFETCH) {
3035 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3036 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3037 			next_txd = calc_next_txd(txq, cidx, 0);
3038 			prefetch(next_txd);
3039 		}
3040 
3041 		/* prefetch the next cache line of mbuf pointers and flags */
3042 		prefetch(&txq->ift_sds.ifsd_m[next]);
3043 		if (txq->ift_sds.ifsd_map != NULL) {
3044 			prefetch(&txq->ift_sds.ifsd_map[next]);
3045 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3046 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3047 		}
3048 	}
3049 	if (txq->ift_sds.ifsd_map != NULL)
3050 		map = txq->ift_sds.ifsd_map[pidx];
3051 
3052 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3053 		desc_tag = txq->ift_tso_desc_tag;
3054 		max_segs = scctx->isc_tx_tso_segments_max;
3055 	} else {
3056 		desc_tag = txq->ift_desc_tag;
3057 		max_segs = scctx->isc_tx_nsegments;
3058 	}
3059 	m_head = *m_headp;
3060 
3061 	pkt_info_zero(&pi);
3062 	pi.ipi_len = m_head->m_pkthdr.len;
3063 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3064 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3065 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3066 	pi.ipi_pidx = pidx;
3067 	pi.ipi_qsidx = txq->ift_id;
3068 
3069 	/* deliberate bitwise OR to make one condition */
3070 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3071 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3072 			return (err);
3073 		m_head = *m_headp;
3074 	}
3075 
3076 retry:
3077 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3078 defrag:
3079 	if (__predict_false(err)) {
3080 		switch (err) {
3081 		case EFBIG:
3082 			/* try collapse once and defrag once */
3083 			if (remap == 0)
3084 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3085 			if (remap == 1)
3086 				m_head = m_defrag(*m_headp, M_NOWAIT);
3087 			remap++;
3088 			if (__predict_false(m_head == NULL))
3089 				goto defrag_failed;
3090 			txq->ift_mbuf_defrag++;
3091 			*m_headp = m_head;
3092 			goto retry;
3093 			break;
3094 		case ENOMEM:
3095 			txq->ift_no_tx_dma_setup++;
3096 			break;
3097 		default:
3098 			txq->ift_no_tx_dma_setup++;
3099 			m_freem(*m_headp);
3100 			DBG_COUNTER_INC(tx_frees);
3101 			*m_headp = NULL;
3102 			break;
3103 		}
3104 		txq->ift_map_failed++;
3105 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3106 		return (err);
3107 	}
3108 
3109 	/*
3110 	 * XXX assumes a 1 to 1 relationship between segments and
3111 	 *        descriptors - this does not hold true on all drivers, e.g.
3112 	 *        cxgb
3113 	 */
3114 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3115 		txq->ift_no_desc_avail++;
3116 		if (map != NULL)
3117 			bus_dmamap_unload(desc_tag, map);
3118 		DBG_COUNTER_INC(encap_txq_avail_fail);
3119 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3120 			GROUPTASK_ENQUEUE(&txq->ift_task);
3121 		return (ENOBUFS);
3122 	}
3123 	/*
3124 	 * On Intel cards we can greatly reduce the number of TX interrupts
3125 	 * we see by only setting report status on every Nth descriptor.
3126 	 * However, this also means that the driver will need to keep track
3127 	 * of the descriptors that RS was set on to check them for the DD bit.
3128 	 */
3129 	txq->ift_rs_pending += nsegs + 1;
3130 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3131 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3132 		pi.ipi_flags |= IPI_TX_INTR;
3133 		txq->ift_rs_pending = 0;
3134 	}
3135 
3136 	pi.ipi_segs = segs;
3137 	pi.ipi_nsegs = nsegs;
3138 
3139 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3140 #ifdef PKT_DEBUG
3141 	print_pkt(&pi);
3142 #endif
3143 	if (map != NULL)
3144 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3145 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3146 		if (map != NULL)
3147 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3148 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3149 		DBG_COUNTER_INC(tx_encap);
3150 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3151 
3152 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3153 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3154 			ndesc += txq->ift_size;
3155 			txq->ift_gen = 1;
3156 		}
3157 		/*
3158 		 * drivers can need as many as
3159 		 * two sentinels
3160 		 */
3161 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3162 		MPASS(pi.ipi_new_pidx != pidx);
3163 		MPASS(ndesc > 0);
3164 		txq->ift_in_use += ndesc;
3165 
3166 		/*
3167 		 * We update the last software descriptor again here because there may
3168 		 * be a sentinel and/or there may be more mbufs than segments
3169 		 */
3170 		txq->ift_pidx = pi.ipi_new_pidx;
3171 		txq->ift_npending += pi.ipi_ndescs;
3172 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3173 		*m_headp = m_head = iflib_remove_mbuf(txq);
3174 		remap = 1;
3175 		txq->ift_txd_encap_efbig++;
3176 		goto defrag;
3177 	} else
3178 		DBG_COUNTER_INC(encap_txd_encap_fail);
3179 	return (err);
3180 
3181 defrag_failed:
3182 	txq->ift_mbuf_defrag_failed++;
3183 	txq->ift_map_failed++;
3184 	m_freem(*m_headp);
3185 	DBG_COUNTER_INC(tx_frees);
3186 	*m_headp = NULL;
3187 	return (ENOMEM);
3188 }
3189 
3190 static void
3191 iflib_tx_desc_free(iflib_txq_t txq, int n)
3192 {
3193 	int hasmap;
3194 	uint32_t qsize, cidx, mask, gen;
3195 	struct mbuf *m, **ifsd_m;
3196 	uint8_t *ifsd_flags;
3197 	bus_dmamap_t *ifsd_map;
3198 	bool do_prefetch;
3199 
3200 	cidx = txq->ift_cidx;
3201 	gen = txq->ift_gen;
3202 	qsize = txq->ift_size;
3203 	mask = qsize-1;
3204 	hasmap = txq->ift_sds.ifsd_map != NULL;
3205 	ifsd_flags = txq->ift_sds.ifsd_flags;
3206 	ifsd_m = txq->ift_sds.ifsd_m;
3207 	ifsd_map = txq->ift_sds.ifsd_map;
3208 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3209 
3210 	while (n--) {
3211 		if (do_prefetch) {
3212 			prefetch(ifsd_m[(cidx + 3) & mask]);
3213 			prefetch(ifsd_m[(cidx + 4) & mask]);
3214 		}
3215 		if (ifsd_m[cidx] != NULL) {
3216 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3217 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3218 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3219 				/*
3220 				 * does it matter if it's not the TSO tag? If so we'll
3221 				 * have to add the type to flags
3222 				 */
3223 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3224 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3225 			}
3226 			if ((m = ifsd_m[cidx]) != NULL) {
3227 				/* XXX we don't support any drivers that batch packets yet */
3228 				MPASS(m->m_nextpkt == NULL);
3229 
3230 				m_free(m);
3231 				ifsd_m[cidx] = NULL;
3232 #if MEMORY_LOGGING
3233 				txq->ift_dequeued++;
3234 #endif
3235 				DBG_COUNTER_INC(tx_frees);
3236 			}
3237 		}
3238 		if (__predict_false(++cidx == qsize)) {
3239 			cidx = 0;
3240 			gen = 0;
3241 		}
3242 	}
3243 	txq->ift_cidx = cidx;
3244 	txq->ift_gen = gen;
3245 }
3246 
3247 static __inline int
3248 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3249 {
3250 	int reclaim;
3251 	if_ctx_t ctx = txq->ift_ctx;
3252 
3253 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3254 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3255 
3256 	/*
3257 	 * Need a rate-limiting check so that this isn't called every time
3258 	 */
3259 	iflib_tx_credits_update(ctx, txq);
3260 	reclaim = DESC_RECLAIMABLE(txq);
3261 
3262 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3263 #ifdef INVARIANTS
3264 		if (iflib_verbose_debug) {
3265 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3266 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3267 			       reclaim, thresh);
3268 
3269 		}
3270 #endif
3271 		return (0);
3272 	}
3273 	iflib_tx_desc_free(txq, reclaim);
3274 	txq->ift_cleaned += reclaim;
3275 	txq->ift_in_use -= reclaim;
3276 
3277 	return (reclaim);
3278 }
3279 
3280 static struct mbuf **
3281 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3282 {
3283 	int next, size;
3284 	struct mbuf **items;
3285 
3286 	size = r->size;
3287 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3288 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3289 
3290 	prefetch(items[(cidx + offset) & (size-1)]);
3291 	if (remaining > 1) {
3292 		prefetch(&items[next]);
3293 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3294 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3295 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3296 	}
3297 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3298 }
3299 
3300 static void
3301 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3302 {
3303 
3304 	ifmp_ring_check_drainage(txq->ift_br, budget);
3305 }
3306 
3307 static uint32_t
3308 iflib_txq_can_drain(struct ifmp_ring *r)
3309 {
3310 	iflib_txq_t txq = r->cookie;
3311 	if_ctx_t ctx = txq->ift_ctx;
3312 
3313 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3314 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3315 }
3316 
3317 static uint32_t
3318 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3319 {
3320 	iflib_txq_t txq = r->cookie;
3321 	if_ctx_t ctx = txq->ift_ctx;
3322 	struct ifnet *ifp = ctx->ifc_ifp;
3323 	struct mbuf **mp, *m;
3324 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3325 	int reclaimed, err, in_use_prev, desc_used;
3326 	bool do_prefetch, ring, rang;
3327 
3328 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3329 			    !LINK_ACTIVE(ctx))) {
3330 		DBG_COUNTER_INC(txq_drain_notready);
3331 		return (0);
3332 	}
3333 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3334 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3335 	avail = IDXDIFF(pidx, cidx, r->size);
3336 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3337 		DBG_COUNTER_INC(txq_drain_flushing);
3338 		for (i = 0; i < avail; i++) {
3339 			m_free(r->items[(cidx + i) & (r->size-1)]);
3340 			r->items[(cidx + i) & (r->size-1)] = NULL;
3341 		}
3342 		return (avail);
3343 	}
3344 
3345 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3346 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3347 		CALLOUT_LOCK(txq);
3348 		callout_stop(&txq->ift_timer);
3349 		CALLOUT_UNLOCK(txq);
3350 		DBG_COUNTER_INC(txq_drain_oactive);
3351 		return (0);
3352 	}
3353 	if (reclaimed)
3354 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3355 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3356 	count = MIN(avail, TX_BATCH_SIZE);
3357 #ifdef INVARIANTS
3358 	if (iflib_verbose_debug)
3359 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3360 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3361 #endif
3362 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3363 	avail = TXQ_AVAIL(txq);
3364 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3365 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3366 
3367 		mp = _ring_peek_one(r, cidx, i, rem);
3368 		MPASS(mp != NULL && *mp != NULL);
3369 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3370 			consumed++;
3371 			reclaimed++;
3372 			continue;
3373 		}
3374 		in_use_prev = txq->ift_in_use;
3375 		pidx_prev = txq->ift_pidx;
3376 		err = iflib_encap(txq, mp);
3377 		if (__predict_false(err)) {
3378 			DBG_COUNTER_INC(txq_drain_encapfail);
3379 			/* no room - bail out */
3380 			if (err == ENOBUFS)
3381 				break;
3382 			consumed++;
3383 			DBG_COUNTER_INC(txq_drain_encapfail);
3384 			/* we can't send this packet - skip it */
3385 			continue;
3386 		}
3387 		consumed++;
3388 		pkt_sent++;
3389 		m = *mp;
3390 		DBG_COUNTER_INC(tx_sent);
3391 		bytes_sent += m->m_pkthdr.len;
3392 		mcast_sent += !!(m->m_flags & M_MCAST);
3393 		avail = TXQ_AVAIL(txq);
3394 
3395 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3396 		desc_used += (txq->ift_in_use - in_use_prev);
3397 		ETHER_BPF_MTAP(ifp, m);
3398 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3399 			break;
3400 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3401 	}
3402 
3403 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3404 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3405 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3406 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3407 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3408 	if (mcast_sent)
3409 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3410 #ifdef INVARIANTS
3411 	if (iflib_verbose_debug)
3412 		printf("consumed=%d\n", consumed);
3413 #endif
3414 	return (consumed);
3415 }
3416 
3417 static uint32_t
3418 iflib_txq_drain_always(struct ifmp_ring *r)
3419 {
3420 	return (1);
3421 }
3422 
3423 static uint32_t
3424 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3425 {
3426 	int i, avail;
3427 	struct mbuf **mp;
3428 	iflib_txq_t txq;
3429 
3430 	txq = r->cookie;
3431 
3432 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3433 	CALLOUT_LOCK(txq);
3434 	callout_stop(&txq->ift_timer);
3435 	CALLOUT_UNLOCK(txq);
3436 
3437 	avail = IDXDIFF(pidx, cidx, r->size);
3438 	for (i = 0; i < avail; i++) {
3439 		mp = _ring_peek_one(r, cidx, i, avail - i);
3440 		if (__predict_false(*mp == (struct mbuf *)txq))
3441 			continue;
3442 		m_freem(*mp);
3443 	}
3444 	MPASS(ifmp_ring_is_stalled(r) == 0);
3445 	return (avail);
3446 }
3447 
3448 static void
3449 iflib_ifmp_purge(iflib_txq_t txq)
3450 {
3451 	struct ifmp_ring *r;
3452 
3453 	r = txq->ift_br;
3454 	r->drain = iflib_txq_drain_free;
3455 	r->can_drain = iflib_txq_drain_always;
3456 
3457 	ifmp_ring_check_drainage(r, r->size);
3458 
3459 	r->drain = iflib_txq_drain;
3460 	r->can_drain = iflib_txq_can_drain;
3461 }
3462 
3463 static void
3464 _task_fn_tx(void *context)
3465 {
3466 	iflib_txq_t txq = context;
3467 	if_ctx_t ctx = txq->ift_ctx;
3468 	struct ifnet *ifp = ctx->ifc_ifp;
3469 	int rc;
3470 
3471 #ifdef IFLIB_DIAGNOSTICS
3472 	txq->ift_cpu_exec_count[curcpu]++;
3473 #endif
3474 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3475 		return;
3476 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3477 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3478 			netmap_tx_irq(ifp, txq->ift_id);
3479 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3480 		return;
3481 	}
3482 	if (txq->ift_db_pending)
3483 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3484 	else
3485 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3486 	if (ctx->ifc_flags & IFC_LEGACY)
3487 		IFDI_INTR_ENABLE(ctx);
3488 	else {
3489 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3490 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3491 	}
3492 }
3493 
3494 static void
3495 _task_fn_rx(void *context)
3496 {
3497 	iflib_rxq_t rxq = context;
3498 	if_ctx_t ctx = rxq->ifr_ctx;
3499 	bool more;
3500 	int rc;
3501 
3502 #ifdef IFLIB_DIAGNOSTICS
3503 	rxq->ifr_cpu_exec_count[curcpu]++;
3504 #endif
3505 	DBG_COUNTER_INC(task_fn_rxs);
3506 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3507 		return;
3508 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3509 		if (ctx->ifc_flags & IFC_LEGACY)
3510 			IFDI_INTR_ENABLE(ctx);
3511 		else {
3512 			DBG_COUNTER_INC(rx_intr_enables);
3513 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3514 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3515 		}
3516 	}
3517 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3518 		return;
3519 	if (more)
3520 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3521 }
3522 
3523 static void
3524 _task_fn_admin(void *context)
3525 {
3526 	if_ctx_t ctx = context;
3527 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3528 	iflib_txq_t txq;
3529 	int i;
3530 
3531 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3532 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3533 			return;
3534 		}
3535 	}
3536 
3537 	CTX_LOCK(ctx);
3538 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3539 		CALLOUT_LOCK(txq);
3540 		callout_stop(&txq->ift_timer);
3541 		CALLOUT_UNLOCK(txq);
3542 	}
3543 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3544 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3545 		callout_reset_on(&txq->ift_timer, iflib_timer_int, iflib_timer,
3546 		    txq, txq->ift_timer.c_cpu);
3547 	IFDI_LINK_INTR_ENABLE(ctx);
3548 	if (ctx->ifc_flags & IFC_DO_RESET) {
3549 		ctx->ifc_flags &= ~IFC_DO_RESET;
3550 		iflib_if_init_locked(ctx);
3551 	}
3552 	CTX_UNLOCK(ctx);
3553 
3554 	if (LINK_ACTIVE(ctx) == 0)
3555 		return;
3556 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3557 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3558 }
3559 
3560 
3561 static void
3562 _task_fn_iov(void *context)
3563 {
3564 	if_ctx_t ctx = context;
3565 
3566 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3567 		return;
3568 
3569 	CTX_LOCK(ctx);
3570 	IFDI_VFLR_HANDLE(ctx);
3571 	CTX_UNLOCK(ctx);
3572 }
3573 
3574 static int
3575 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3576 {
3577 	int err;
3578 	if_int_delay_info_t info;
3579 	if_ctx_t ctx;
3580 
3581 	info = (if_int_delay_info_t)arg1;
3582 	ctx = info->iidi_ctx;
3583 	info->iidi_req = req;
3584 	info->iidi_oidp = oidp;
3585 	CTX_LOCK(ctx);
3586 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3587 	CTX_UNLOCK(ctx);
3588 	return (err);
3589 }
3590 
3591 /*********************************************************************
3592  *
3593  *  IFNET FUNCTIONS
3594  *
3595  **********************************************************************/
3596 
3597 static void
3598 iflib_if_init_locked(if_ctx_t ctx)
3599 {
3600 	iflib_stop(ctx);
3601 	iflib_init_locked(ctx);
3602 }
3603 
3604 
3605 static void
3606 iflib_if_init(void *arg)
3607 {
3608 	if_ctx_t ctx = arg;
3609 
3610 	CTX_LOCK(ctx);
3611 	iflib_if_init_locked(ctx);
3612 	CTX_UNLOCK(ctx);
3613 }
3614 
3615 static int
3616 iflib_if_transmit(if_t ifp, struct mbuf *m)
3617 {
3618 	if_ctx_t	ctx = if_getsoftc(ifp);
3619 
3620 	iflib_txq_t txq;
3621 	int err, qidx;
3622 
3623 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3624 		DBG_COUNTER_INC(tx_frees);
3625 		m_freem(m);
3626 		return (ENOBUFS);
3627 	}
3628 
3629 	MPASS(m->m_nextpkt == NULL);
3630 	qidx = 0;
3631 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3632 		qidx = QIDX(ctx, m);
3633 	/*
3634 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3635 	 */
3636 	txq = &ctx->ifc_txqs[qidx];
3637 
3638 #ifdef DRIVER_BACKPRESSURE
3639 	if (txq->ift_closed) {
3640 		while (m != NULL) {
3641 			next = m->m_nextpkt;
3642 			m->m_nextpkt = NULL;
3643 			m_freem(m);
3644 			m = next;
3645 		}
3646 		return (ENOBUFS);
3647 	}
3648 #endif
3649 #ifdef notyet
3650 	qidx = count = 0;
3651 	mp = marr;
3652 	next = m;
3653 	do {
3654 		count++;
3655 		next = next->m_nextpkt;
3656 	} while (next != NULL);
3657 
3658 	if (count > nitems(marr))
3659 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3660 			/* XXX check nextpkt */
3661 			m_freem(m);
3662 			/* XXX simplify for now */
3663 			DBG_COUNTER_INC(tx_frees);
3664 			return (ENOBUFS);
3665 		}
3666 	for (next = m, i = 0; next != NULL; i++) {
3667 		mp[i] = next;
3668 		next = next->m_nextpkt;
3669 		mp[i]->m_nextpkt = NULL;
3670 	}
3671 #endif
3672 	DBG_COUNTER_INC(tx_seen);
3673 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3674 
3675 	if (err) {
3676 		GROUPTASK_ENQUEUE(&txq->ift_task);
3677 		/* support forthcoming later */
3678 #ifdef DRIVER_BACKPRESSURE
3679 		txq->ift_closed = TRUE;
3680 #endif
3681 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3682 		m_freem(m);
3683 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3684 		GROUPTASK_ENQUEUE(&txq->ift_task);
3685 	}
3686 
3687 	return (err);
3688 }
3689 
3690 static void
3691 iflib_if_qflush(if_t ifp)
3692 {
3693 	if_ctx_t ctx = if_getsoftc(ifp);
3694 	iflib_txq_t txq = ctx->ifc_txqs;
3695 	int i;
3696 
3697 	CTX_LOCK(ctx);
3698 	ctx->ifc_flags |= IFC_QFLUSH;
3699 	CTX_UNLOCK(ctx);
3700 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3701 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3702 			iflib_txq_check_drain(txq, 0);
3703 	CTX_LOCK(ctx);
3704 	ctx->ifc_flags &= ~IFC_QFLUSH;
3705 	CTX_UNLOCK(ctx);
3706 
3707 	if_qflush(ifp);
3708 }
3709 
3710 
3711 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3712 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3713 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3714 
3715 static int
3716 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3717 {
3718 	if_ctx_t ctx = if_getsoftc(ifp);
3719 	struct ifreq	*ifr = (struct ifreq *)data;
3720 #if defined(INET) || defined(INET6)
3721 	struct ifaddr	*ifa = (struct ifaddr *)data;
3722 #endif
3723 	bool		avoid_reset = FALSE;
3724 	int		err = 0, reinit = 0, bits;
3725 
3726 	switch (command) {
3727 	case SIOCSIFADDR:
3728 #ifdef INET
3729 		if (ifa->ifa_addr->sa_family == AF_INET)
3730 			avoid_reset = TRUE;
3731 #endif
3732 #ifdef INET6
3733 		if (ifa->ifa_addr->sa_family == AF_INET6)
3734 			avoid_reset = TRUE;
3735 #endif
3736 		/*
3737 		** Calling init results in link renegotiation,
3738 		** so we avoid doing it when possible.
3739 		*/
3740 		if (avoid_reset) {
3741 			if_setflagbits(ifp, IFF_UP,0);
3742 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3743 				reinit = 1;
3744 #ifdef INET
3745 			if (!(if_getflags(ifp) & IFF_NOARP))
3746 				arp_ifinit(ifp, ifa);
3747 #endif
3748 		} else
3749 			err = ether_ioctl(ifp, command, data);
3750 		break;
3751 	case SIOCSIFMTU:
3752 		CTX_LOCK(ctx);
3753 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3754 			CTX_UNLOCK(ctx);
3755 			break;
3756 		}
3757 		bits = if_getdrvflags(ifp);
3758 		/* stop the driver and free any clusters before proceeding */
3759 		iflib_stop(ctx);
3760 
3761 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3762 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3763 				ctx->ifc_flags |= IFC_MULTISEG;
3764 			else
3765 				ctx->ifc_flags &= ~IFC_MULTISEG;
3766 			err = if_setmtu(ifp, ifr->ifr_mtu);
3767 		}
3768 		iflib_init_locked(ctx);
3769 		if_setdrvflags(ifp, bits);
3770 		CTX_UNLOCK(ctx);
3771 		break;
3772 	case SIOCSIFFLAGS:
3773 		CTX_LOCK(ctx);
3774 		if (if_getflags(ifp) & IFF_UP) {
3775 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3776 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3777 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3778 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3779 				}
3780 			} else
3781 				reinit = 1;
3782 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3783 			iflib_stop(ctx);
3784 		}
3785 		ctx->ifc_if_flags = if_getflags(ifp);
3786 		CTX_UNLOCK(ctx);
3787 		break;
3788 	case SIOCADDMULTI:
3789 	case SIOCDELMULTI:
3790 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3791 			CTX_LOCK(ctx);
3792 			IFDI_INTR_DISABLE(ctx);
3793 			IFDI_MULTI_SET(ctx);
3794 			IFDI_INTR_ENABLE(ctx);
3795 			CTX_UNLOCK(ctx);
3796 		}
3797 		break;
3798 	case SIOCSIFMEDIA:
3799 		CTX_LOCK(ctx);
3800 		IFDI_MEDIA_SET(ctx);
3801 		CTX_UNLOCK(ctx);
3802 		/* falls thru */
3803 	case SIOCGIFMEDIA:
3804 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3805 		break;
3806 	case SIOCGI2C:
3807 	{
3808 		struct ifi2creq i2c;
3809 
3810 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3811 		if (err != 0)
3812 			break;
3813 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3814 			err = EINVAL;
3815 			break;
3816 		}
3817 		if (i2c.len > sizeof(i2c.data)) {
3818 			err = EINVAL;
3819 			break;
3820 		}
3821 
3822 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3823 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3824 		break;
3825 	}
3826 	case SIOCSIFCAP:
3827 	{
3828 		int mask, setmask;
3829 
3830 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3831 		setmask = 0;
3832 #ifdef TCP_OFFLOAD
3833 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3834 #endif
3835 		setmask |= (mask & IFCAP_FLAGS);
3836 
3837 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3838 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3839 		if ((mask & IFCAP_WOL) &&
3840 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3841 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3842 		if_vlancap(ifp);
3843 		/*
3844 		 * want to ensure that traffic has stopped before we change any of the flags
3845 		 */
3846 		if (setmask) {
3847 			CTX_LOCK(ctx);
3848 			bits = if_getdrvflags(ifp);
3849 			if (bits & IFF_DRV_RUNNING)
3850 				iflib_stop(ctx);
3851 			if_togglecapenable(ifp, setmask);
3852 			if (bits & IFF_DRV_RUNNING)
3853 				iflib_init_locked(ctx);
3854 			if_setdrvflags(ifp, bits);
3855 			CTX_UNLOCK(ctx);
3856 		}
3857 		break;
3858 	    }
3859 	case SIOCGPRIVATE_0:
3860 	case SIOCSDRVSPEC:
3861 	case SIOCGDRVSPEC:
3862 		CTX_LOCK(ctx);
3863 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3864 		CTX_UNLOCK(ctx);
3865 		break;
3866 	default:
3867 		err = ether_ioctl(ifp, command, data);
3868 		break;
3869 	}
3870 	if (reinit)
3871 		iflib_if_init(ctx);
3872 	return (err);
3873 }
3874 
3875 static uint64_t
3876 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3877 {
3878 	if_ctx_t ctx = if_getsoftc(ifp);
3879 
3880 	return (IFDI_GET_COUNTER(ctx, cnt));
3881 }
3882 
3883 /*********************************************************************
3884  *
3885  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3886  *
3887  **********************************************************************/
3888 
3889 static void
3890 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3891 {
3892 	if_ctx_t ctx = if_getsoftc(ifp);
3893 
3894 	if ((void *)ctx != arg)
3895 		return;
3896 
3897 	if ((vtag == 0) || (vtag > 4095))
3898 		return;
3899 
3900 	CTX_LOCK(ctx);
3901 	IFDI_VLAN_REGISTER(ctx, vtag);
3902 	/* Re-init to load the changes */
3903 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3904 		iflib_init_locked(ctx);
3905 	CTX_UNLOCK(ctx);
3906 }
3907 
3908 static void
3909 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3910 {
3911 	if_ctx_t ctx = if_getsoftc(ifp);
3912 
3913 	if ((void *)ctx != arg)
3914 		return;
3915 
3916 	if ((vtag == 0) || (vtag > 4095))
3917 		return;
3918 
3919 	CTX_LOCK(ctx);
3920 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3921 	/* Re-init to load the changes */
3922 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3923 		iflib_init_locked(ctx);
3924 	CTX_UNLOCK(ctx);
3925 }
3926 
3927 static void
3928 iflib_led_func(void *arg, int onoff)
3929 {
3930 	if_ctx_t ctx = arg;
3931 
3932 	CTX_LOCK(ctx);
3933 	IFDI_LED_FUNC(ctx, onoff);
3934 	CTX_UNLOCK(ctx);
3935 }
3936 
3937 /*********************************************************************
3938  *
3939  *  BUS FUNCTION DEFINITIONS
3940  *
3941  **********************************************************************/
3942 
3943 int
3944 iflib_device_probe(device_t dev)
3945 {
3946 	pci_vendor_info_t *ent;
3947 
3948 	uint16_t	pci_vendor_id, pci_device_id;
3949 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3950 	uint16_t	pci_rev_id;
3951 	if_shared_ctx_t sctx;
3952 
3953 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3954 		return (ENOTSUP);
3955 
3956 	pci_vendor_id = pci_get_vendor(dev);
3957 	pci_device_id = pci_get_device(dev);
3958 	pci_subvendor_id = pci_get_subvendor(dev);
3959 	pci_subdevice_id = pci_get_subdevice(dev);
3960 	pci_rev_id = pci_get_revid(dev);
3961 	if (sctx->isc_parse_devinfo != NULL)
3962 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3963 
3964 	ent = sctx->isc_vendor_info;
3965 	while (ent->pvi_vendor_id != 0) {
3966 		if (pci_vendor_id != ent->pvi_vendor_id) {
3967 			ent++;
3968 			continue;
3969 		}
3970 		if ((pci_device_id == ent->pvi_device_id) &&
3971 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3972 		     (ent->pvi_subvendor_id == 0)) &&
3973 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3974 		     (ent->pvi_subdevice_id == 0)) &&
3975 		    ((pci_rev_id == ent->pvi_rev_id) ||
3976 		     (ent->pvi_rev_id == 0))) {
3977 
3978 			device_set_desc_copy(dev, ent->pvi_name);
3979 			/* this needs to be changed to zero if the bus probing code
3980 			 * ever stops re-probing on best match because the sctx
3981 			 * may have its values over written by register calls
3982 			 * in subsequent probes
3983 			 */
3984 			return (BUS_PROBE_DEFAULT);
3985 		}
3986 		ent++;
3987 	}
3988 	return (ENXIO);
3989 }
3990 
3991 int
3992 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3993 {
3994 	int err, rid, msix, msix_bar;
3995 	if_ctx_t ctx;
3996 	if_t ifp;
3997 	if_softc_ctx_t scctx;
3998 	int i;
3999 	uint16_t main_txq;
4000 	uint16_t main_rxq;
4001 
4002 
4003 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4004 
4005 	if (sc == NULL) {
4006 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4007 		device_set_softc(dev, ctx);
4008 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4009 	}
4010 
4011 	ctx->ifc_sctx = sctx;
4012 	ctx->ifc_dev = dev;
4013 	ctx->ifc_softc = sc;
4014 
4015 	if ((err = iflib_register(ctx)) != 0) {
4016 		device_printf(dev, "iflib_register failed %d\n", err);
4017 		return (err);
4018 	}
4019 	iflib_add_device_sysctl_pre(ctx);
4020 
4021 	scctx = &ctx->ifc_softc_ctx;
4022 	ifp = ctx->ifc_ifp;
4023 
4024 	/*
4025 	 * XXX sanity check that ntxd & nrxd are a power of 2
4026 	 */
4027 	if (ctx->ifc_sysctl_ntxqs != 0)
4028 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4029 	if (ctx->ifc_sysctl_nrxqs != 0)
4030 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4031 
4032 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4033 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4034 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4035 		else
4036 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4037 	}
4038 
4039 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4040 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4041 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4042 		else
4043 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4044 	}
4045 
4046 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4047 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4048 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4049 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4050 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4051 		}
4052 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4053 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4054 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4055 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4056 		}
4057 	}
4058 
4059 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4060 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4061 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4062 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4063 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4064 		}
4065 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4066 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4067 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4068 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4069 		}
4070 	}
4071 
4072 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4073 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4074 		return (err);
4075 	}
4076 	_iflib_pre_assert(scctx);
4077 	ctx->ifc_txrx = *scctx->isc_txrx;
4078 
4079 #ifdef INVARIANTS
4080 	MPASS(scctx->isc_capenable);
4081 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4082 		MPASS(scctx->isc_tx_csum_flags);
4083 #endif
4084 
4085 	if_setcapabilities(ifp, scctx->isc_capenable);
4086 	if_setcapenable(ifp, scctx->isc_capenable);
4087 
4088 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4089 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4090 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4091 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4092 
4093 #ifdef ACPI_DMAR
4094 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4095 		ctx->ifc_flags |= IFC_DMAR;
4096 #elif !(defined(__i386__) || defined(__amd64__))
4097 	/* set unconditionally for !x86 */
4098 	ctx->ifc_flags |= IFC_DMAR;
4099 #endif
4100 	if (force_busdma)
4101 		ctx->ifc_flags |= IFC_DMAR;
4102 
4103 	msix_bar = scctx->isc_msix_bar;
4104 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4105 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4106 
4107 	/* XXX change for per-queue sizes */
4108 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4109 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4110 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4111 		if (!powerof2(scctx->isc_nrxd[i])) {
4112 			/* round down instead? */
4113 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4114 			err = EINVAL;
4115 			goto fail;
4116 		}
4117 	}
4118 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4119 		if (!powerof2(scctx->isc_ntxd[i])) {
4120 			device_printf(dev,
4121 			    "# tx descriptors must be a power of 2");
4122 			err = EINVAL;
4123 			goto fail;
4124 		}
4125 	}
4126 
4127 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4128 	    MAX_SINGLE_PACKET_FRACTION)
4129 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4130 		    MAX_SINGLE_PACKET_FRACTION);
4131 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4132 	    MAX_SINGLE_PACKET_FRACTION)
4133 		scctx->isc_tx_tso_segments_max = max(1,
4134 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4135 
4136 	/*
4137 	 * Protect the stack against modern hardware
4138 	 */
4139 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4140 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4141 
4142 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4143 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4144 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4145 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4146 	if (scctx->isc_rss_table_size == 0)
4147 		scctx->isc_rss_table_size = 64;
4148 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4149 
4150 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4151 	/* XXX format name */
4152 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4153 	/*
4154 	** Now setup MSI or MSI/X, should
4155 	** return us the number of supported
4156 	** vectors. (Will be 1 for MSI)
4157 	*/
4158 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4159 		msix = scctx->isc_vectors;
4160 	} else if (scctx->isc_msix_bar != 0)
4161 	       /*
4162 		* The simple fact that isc_msix_bar is not 0 does not mean we
4163 		* we have a good value there that is known to work.
4164 		*/
4165 		msix = iflib_msix_init(ctx);
4166 	else {
4167 		scctx->isc_vectors = 1;
4168 		scctx->isc_ntxqsets = 1;
4169 		scctx->isc_nrxqsets = 1;
4170 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4171 		msix = 0;
4172 	}
4173 	/* Get memory for the station queues */
4174 	if ((err = iflib_queues_alloc(ctx))) {
4175 		device_printf(dev, "Unable to allocate queue memory\n");
4176 		goto fail;
4177 	}
4178 
4179 	if ((err = iflib_qset_structures_setup(ctx))) {
4180 		device_printf(dev, "qset structure setup failed %d\n", err);
4181 		goto fail_queues;
4182 	}
4183 
4184 	/*
4185 	 * Group taskqueues aren't properly set up until SMP is started,
4186 	 * so we disable interrupts until we can handle them post
4187 	 * SI_SUB_SMP.
4188 	 *
4189 	 * XXX: disabling interrupts doesn't actually work, at least for
4190 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4191 	 * we do null handling and depend on this not causing too large an
4192 	 * interrupt storm.
4193 	 */
4194 	IFDI_INTR_DISABLE(ctx);
4195 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4196 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4197 		goto fail_intr_free;
4198 	}
4199 	if (msix <= 1) {
4200 		rid = 0;
4201 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4202 			MPASS(msix == 1);
4203 			rid = 1;
4204 		}
4205 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4206 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4207 			goto fail_intr_free;
4208 		}
4209 	}
4210 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4211 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4212 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4213 		goto fail_detach;
4214 	}
4215 	if ((err = iflib_netmap_attach(ctx))) {
4216 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4217 		goto fail_detach;
4218 	}
4219 	*ctxp = ctx;
4220 
4221 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4222 	iflib_add_device_sysctl_post(ctx);
4223 	ctx->ifc_flags |= IFC_INIT_DONE;
4224 	return (0);
4225 fail_detach:
4226 	ether_ifdetach(ctx->ifc_ifp);
4227 fail_intr_free:
4228 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4229 		pci_release_msi(ctx->ifc_dev);
4230 fail_queues:
4231 	/* XXX free queues */
4232 fail:
4233 	IFDI_DETACH(ctx);
4234 	return (err);
4235 }
4236 
4237 int
4238 iflib_device_attach(device_t dev)
4239 {
4240 	if_ctx_t ctx;
4241 	if_shared_ctx_t sctx;
4242 
4243 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4244 		return (ENOTSUP);
4245 
4246 	pci_enable_busmaster(dev);
4247 
4248 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4249 }
4250 
4251 int
4252 iflib_device_deregister(if_ctx_t ctx)
4253 {
4254 	if_t ifp = ctx->ifc_ifp;
4255 	iflib_txq_t txq;
4256 	iflib_rxq_t rxq;
4257 	device_t dev = ctx->ifc_dev;
4258 	int i;
4259 	struct taskqgroup *tqg;
4260 
4261 	/* Make sure VLANS are not using driver */
4262 	if (if_vlantrunkinuse(ifp)) {
4263 		device_printf(dev,"Vlan in use, detach first\n");
4264 		return (EBUSY);
4265 	}
4266 
4267 	CTX_LOCK(ctx);
4268 	ctx->ifc_in_detach = 1;
4269 	iflib_stop(ctx);
4270 	CTX_UNLOCK(ctx);
4271 
4272 	/* Unregister VLAN events */
4273 	if (ctx->ifc_vlan_attach_event != NULL)
4274 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4275 	if (ctx->ifc_vlan_detach_event != NULL)
4276 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4277 
4278 	iflib_netmap_detach(ifp);
4279 	ether_ifdetach(ifp);
4280 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4281 	CTX_LOCK_DESTROY(ctx);
4282 	if (ctx->ifc_led_dev != NULL)
4283 		led_destroy(ctx->ifc_led_dev);
4284 	/* XXX drain any dependent tasks */
4285 	tqg = qgroup_if_io_tqg;
4286 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4287 		callout_drain(&txq->ift_timer);
4288 		if (txq->ift_task.gt_uniq != NULL)
4289 			taskqgroup_detach(tqg, &txq->ift_task);
4290 	}
4291 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4292 		if (rxq->ifr_task.gt_uniq != NULL)
4293 			taskqgroup_detach(tqg, &rxq->ifr_task);
4294 	}
4295 	tqg = qgroup_if_config_tqg;
4296 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4297 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4298 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4299 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4300 
4301 	IFDI_DETACH(ctx);
4302 	device_set_softc(ctx->ifc_dev, NULL);
4303 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4304 		pci_release_msi(dev);
4305 	}
4306 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4307 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4308 	}
4309 	if (ctx->ifc_msix_mem != NULL) {
4310 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4311 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4312 		ctx->ifc_msix_mem = NULL;
4313 	}
4314 
4315 	bus_generic_detach(dev);
4316 	if_free(ifp);
4317 
4318 	iflib_tx_structures_free(ctx);
4319 	iflib_rx_structures_free(ctx);
4320 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4321 		free(ctx->ifc_softc, M_IFLIB);
4322 	free(ctx, M_IFLIB);
4323 	return (0);
4324 }
4325 
4326 
4327 int
4328 iflib_device_detach(device_t dev)
4329 {
4330 	if_ctx_t ctx = device_get_softc(dev);
4331 
4332 	return (iflib_device_deregister(ctx));
4333 }
4334 
4335 int
4336 iflib_device_suspend(device_t dev)
4337 {
4338 	if_ctx_t ctx = device_get_softc(dev);
4339 
4340 	CTX_LOCK(ctx);
4341 	IFDI_SUSPEND(ctx);
4342 	CTX_UNLOCK(ctx);
4343 
4344 	return bus_generic_suspend(dev);
4345 }
4346 int
4347 iflib_device_shutdown(device_t dev)
4348 {
4349 	if_ctx_t ctx = device_get_softc(dev);
4350 
4351 	CTX_LOCK(ctx);
4352 	IFDI_SHUTDOWN(ctx);
4353 	CTX_UNLOCK(ctx);
4354 
4355 	return bus_generic_suspend(dev);
4356 }
4357 
4358 
4359 int
4360 iflib_device_resume(device_t dev)
4361 {
4362 	if_ctx_t ctx = device_get_softc(dev);
4363 	iflib_txq_t txq = ctx->ifc_txqs;
4364 
4365 	CTX_LOCK(ctx);
4366 	IFDI_RESUME(ctx);
4367 	iflib_init_locked(ctx);
4368 	CTX_UNLOCK(ctx);
4369 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4370 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4371 
4372 	return (bus_generic_resume(dev));
4373 }
4374 
4375 int
4376 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4377 {
4378 	int error;
4379 	if_ctx_t ctx = device_get_softc(dev);
4380 
4381 	CTX_LOCK(ctx);
4382 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4383 	CTX_UNLOCK(ctx);
4384 
4385 	return (error);
4386 }
4387 
4388 void
4389 iflib_device_iov_uninit(device_t dev)
4390 {
4391 	if_ctx_t ctx = device_get_softc(dev);
4392 
4393 	CTX_LOCK(ctx);
4394 	IFDI_IOV_UNINIT(ctx);
4395 	CTX_UNLOCK(ctx);
4396 }
4397 
4398 int
4399 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4400 {
4401 	int error;
4402 	if_ctx_t ctx = device_get_softc(dev);
4403 
4404 	CTX_LOCK(ctx);
4405 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4406 	CTX_UNLOCK(ctx);
4407 
4408 	return (error);
4409 }
4410 
4411 /*********************************************************************
4412  *
4413  *  MODULE FUNCTION DEFINITIONS
4414  *
4415  **********************************************************************/
4416 
4417 /*
4418  * - Start a fast taskqueue thread for each core
4419  * - Start a taskqueue for control operations
4420  */
4421 static int
4422 iflib_module_init(void)
4423 {
4424 
4425 	iflib_timer_int = hz / 2;
4426 	TUNABLE_INT_FETCH("net.iflib.timer_int", &iflib_timer_int);
4427 	return (0);
4428 }
4429 
4430 static int
4431 iflib_module_event_handler(module_t mod, int what, void *arg)
4432 {
4433 	int err;
4434 
4435 	switch (what) {
4436 	case MOD_LOAD:
4437 		if ((err = iflib_module_init()) != 0)
4438 			return (err);
4439 		break;
4440 	case MOD_UNLOAD:
4441 		return (EBUSY);
4442 	default:
4443 		return (EOPNOTSUPP);
4444 	}
4445 
4446 	return (0);
4447 }
4448 
4449 /*********************************************************************
4450  *
4451  *  PUBLIC FUNCTION DEFINITIONS
4452  *     ordered as in iflib.h
4453  *
4454  **********************************************************************/
4455 
4456 
4457 static void
4458 _iflib_assert(if_shared_ctx_t sctx)
4459 {
4460 	MPASS(sctx->isc_tx_maxsize);
4461 	MPASS(sctx->isc_tx_maxsegsize);
4462 
4463 	MPASS(sctx->isc_rx_maxsize);
4464 	MPASS(sctx->isc_rx_nsegments);
4465 	MPASS(sctx->isc_rx_maxsegsize);
4466 
4467 	MPASS(sctx->isc_nrxd_min[0]);
4468 	MPASS(sctx->isc_nrxd_max[0]);
4469 	MPASS(sctx->isc_nrxd_default[0]);
4470 	MPASS(sctx->isc_ntxd_min[0]);
4471 	MPASS(sctx->isc_ntxd_max[0]);
4472 	MPASS(sctx->isc_ntxd_default[0]);
4473 }
4474 
4475 static void
4476 _iflib_pre_assert(if_softc_ctx_t scctx)
4477 {
4478 
4479 	MPASS(scctx->isc_txrx->ift_txd_encap);
4480 	MPASS(scctx->isc_txrx->ift_txd_flush);
4481 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4482 	MPASS(scctx->isc_txrx->ift_rxd_available);
4483 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4484 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4485 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4486 }
4487 
4488 static int
4489 iflib_register(if_ctx_t ctx)
4490 {
4491 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4492 	driver_t *driver = sctx->isc_driver;
4493 	device_t dev = ctx->ifc_dev;
4494 	if_t ifp;
4495 
4496 	_iflib_assert(sctx);
4497 
4498 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4499 
4500 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4501 	if (ifp == NULL) {
4502 		device_printf(dev, "can not allocate ifnet structure\n");
4503 		return (ENOMEM);
4504 	}
4505 
4506 	/*
4507 	 * Initialize our context's device specific methods
4508 	 */
4509 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4510 	kobj_class_compile((kobj_class_t) driver);
4511 	driver->refs++;
4512 
4513 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4514 	if_setsoftc(ifp, ctx);
4515 	if_setdev(ifp, dev);
4516 	if_setinitfn(ifp, iflib_if_init);
4517 	if_setioctlfn(ifp, iflib_if_ioctl);
4518 	if_settransmitfn(ifp, iflib_if_transmit);
4519 	if_setqflushfn(ifp, iflib_if_qflush);
4520 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4521 
4522 	ctx->ifc_vlan_attach_event =
4523 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4524 							  EVENTHANDLER_PRI_FIRST);
4525 	ctx->ifc_vlan_detach_event =
4526 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4527 							  EVENTHANDLER_PRI_FIRST);
4528 
4529 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4530 					 iflib_media_change, iflib_media_status);
4531 
4532 	return (0);
4533 }
4534 
4535 
4536 static int
4537 iflib_queues_alloc(if_ctx_t ctx)
4538 {
4539 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4540 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4541 	device_t dev = ctx->ifc_dev;
4542 	int nrxqsets = scctx->isc_nrxqsets;
4543 	int ntxqsets = scctx->isc_ntxqsets;
4544 	iflib_txq_t txq;
4545 	iflib_rxq_t rxq;
4546 	iflib_fl_t fl = NULL;
4547 	int i, j, cpu, err, txconf, rxconf;
4548 	iflib_dma_info_t ifdip;
4549 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4550 	uint32_t *txqsizes = scctx->isc_txqsizes;
4551 	uint8_t nrxqs = sctx->isc_nrxqs;
4552 	uint8_t ntxqs = sctx->isc_ntxqs;
4553 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4554 	caddr_t *vaddrs;
4555 	uint64_t *paddrs;
4556 	struct ifmp_ring **brscp;
4557 
4558 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4559 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4560 
4561 	brscp = NULL;
4562 	txq = NULL;
4563 	rxq = NULL;
4564 
4565 /* Allocate the TX ring struct memory */
4566 	if (!(txq =
4567 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4568 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4569 		device_printf(dev, "Unable to allocate TX ring memory\n");
4570 		err = ENOMEM;
4571 		goto fail;
4572 	}
4573 
4574 	/* Now allocate the RX */
4575 	if (!(rxq =
4576 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4577 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4578 		device_printf(dev, "Unable to allocate RX ring memory\n");
4579 		err = ENOMEM;
4580 		goto rx_fail;
4581 	}
4582 
4583 	ctx->ifc_txqs = txq;
4584 	ctx->ifc_rxqs = rxq;
4585 
4586 	/*
4587 	 * XXX handle allocation failure
4588 	 */
4589 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4590 		/* Set up some basics */
4591 
4592 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4593 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4594 			err = ENOMEM;
4595 			goto err_tx_desc;
4596 		}
4597 		txq->ift_ifdi = ifdip;
4598 		for (j = 0; j < ntxqs; j++, ifdip++) {
4599 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4600 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4601 				err = ENOMEM;
4602 				goto err_tx_desc;
4603 			}
4604 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4605 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4606 		}
4607 		txq->ift_ctx = ctx;
4608 		txq->ift_id = i;
4609 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4610 			txq->ift_br_offset = 1;
4611 		} else {
4612 			txq->ift_br_offset = 0;
4613 		}
4614 		/* XXX fix this */
4615 		txq->ift_timer.c_cpu = cpu;
4616 
4617 		if (iflib_txsd_alloc(txq)) {
4618 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4619 			err = ENOMEM;
4620 			goto err_tx_desc;
4621 		}
4622 
4623 		/* Initialize the TX lock */
4624 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4625 		    device_get_nameunit(dev), txq->ift_id);
4626 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4627 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4628 
4629 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4630 			 device_get_nameunit(dev), txq->ift_id);
4631 
4632 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4633 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4634 		if (err) {
4635 			/* XXX free any allocated rings */
4636 			device_printf(dev, "Unable to allocate buf_ring\n");
4637 			goto err_tx_desc;
4638 		}
4639 	}
4640 
4641 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4642 		/* Set up some basics */
4643 
4644 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4645 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4646 			err = ENOMEM;
4647 			goto err_tx_desc;
4648 		}
4649 
4650 		rxq->ifr_ifdi = ifdip;
4651 		/* XXX this needs to be changed if #rx queues != #tx queues */
4652 		rxq->ifr_ntxqirq = 1;
4653 		rxq->ifr_txqid[0] = i;
4654 		for (j = 0; j < nrxqs; j++, ifdip++) {
4655 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4656 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4657 				err = ENOMEM;
4658 				goto err_tx_desc;
4659 			}
4660 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4661 		}
4662 		rxq->ifr_ctx = ctx;
4663 		rxq->ifr_id = i;
4664 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4665 			rxq->ifr_fl_offset = 1;
4666 		} else {
4667 			rxq->ifr_fl_offset = 0;
4668 		}
4669 		rxq->ifr_nfl = nfree_lists;
4670 		if (!(fl =
4671 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4672 			device_printf(dev, "Unable to allocate free list memory\n");
4673 			err = ENOMEM;
4674 			goto err_tx_desc;
4675 		}
4676 		rxq->ifr_fl = fl;
4677 		for (j = 0; j < nfree_lists; j++) {
4678 			fl[j].ifl_rxq = rxq;
4679 			fl[j].ifl_id = j;
4680 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4681 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4682 		}
4683         /* Allocate receive buffers for the ring*/
4684 		if (iflib_rxsd_alloc(rxq)) {
4685 			device_printf(dev,
4686 			    "Critical Failure setting up receive buffers\n");
4687 			err = ENOMEM;
4688 			goto err_rx_desc;
4689 		}
4690 	}
4691 
4692 	/* TXQs */
4693 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4694 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4695 	for (i = 0; i < ntxqsets; i++) {
4696 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4697 
4698 		for (j = 0; j < ntxqs; j++, di++) {
4699 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4700 			paddrs[i*ntxqs + j] = di->idi_paddr;
4701 		}
4702 	}
4703 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4704 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4705 		iflib_tx_structures_free(ctx);
4706 		free(vaddrs, M_IFLIB);
4707 		free(paddrs, M_IFLIB);
4708 		goto err_rx_desc;
4709 	}
4710 	free(vaddrs, M_IFLIB);
4711 	free(paddrs, M_IFLIB);
4712 
4713 	/* RXQs */
4714 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4715 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4716 	for (i = 0; i < nrxqsets; i++) {
4717 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4718 
4719 		for (j = 0; j < nrxqs; j++, di++) {
4720 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4721 			paddrs[i*nrxqs + j] = di->idi_paddr;
4722 		}
4723 	}
4724 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4725 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4726 		iflib_tx_structures_free(ctx);
4727 		free(vaddrs, M_IFLIB);
4728 		free(paddrs, M_IFLIB);
4729 		goto err_rx_desc;
4730 	}
4731 	free(vaddrs, M_IFLIB);
4732 	free(paddrs, M_IFLIB);
4733 
4734 	return (0);
4735 
4736 /* XXX handle allocation failure changes */
4737 err_rx_desc:
4738 err_tx_desc:
4739 	if (ctx->ifc_rxqs != NULL)
4740 		free(ctx->ifc_rxqs, M_IFLIB);
4741 	ctx->ifc_rxqs = NULL;
4742 	if (ctx->ifc_txqs != NULL)
4743 		free(ctx->ifc_txqs, M_IFLIB);
4744 	ctx->ifc_txqs = NULL;
4745 rx_fail:
4746 	if (brscp != NULL)
4747 		free(brscp, M_IFLIB);
4748 	if (rxq != NULL)
4749 		free(rxq, M_IFLIB);
4750 	if (txq != NULL)
4751 		free(txq, M_IFLIB);
4752 fail:
4753 	return (err);
4754 }
4755 
4756 static int
4757 iflib_tx_structures_setup(if_ctx_t ctx)
4758 {
4759 	iflib_txq_t txq = ctx->ifc_txqs;
4760 	int i;
4761 
4762 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4763 		iflib_txq_setup(txq);
4764 
4765 	return (0);
4766 }
4767 
4768 static void
4769 iflib_tx_structures_free(if_ctx_t ctx)
4770 {
4771 	iflib_txq_t txq = ctx->ifc_txqs;
4772 	int i, j;
4773 
4774 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4775 		iflib_txq_destroy(txq);
4776 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4777 			iflib_dma_free(&txq->ift_ifdi[j]);
4778 	}
4779 	free(ctx->ifc_txqs, M_IFLIB);
4780 	ctx->ifc_txqs = NULL;
4781 	IFDI_QUEUES_FREE(ctx);
4782 }
4783 
4784 /*********************************************************************
4785  *
4786  *  Initialize all receive rings.
4787  *
4788  **********************************************************************/
4789 static int
4790 iflib_rx_structures_setup(if_ctx_t ctx)
4791 {
4792 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4793 	int q;
4794 #if defined(INET6) || defined(INET)
4795 	int i, err;
4796 #endif
4797 
4798 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4799 #if defined(INET6) || defined(INET)
4800 		tcp_lro_free(&rxq->ifr_lc);
4801 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4802 		    TCP_LRO_ENTRIES, min(1024,
4803 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4804 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4805 			goto fail;
4806 		}
4807 		rxq->ifr_lro_enabled = TRUE;
4808 #endif
4809 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4810 	}
4811 	return (0);
4812 #if defined(INET6) || defined(INET)
4813 fail:
4814 	/*
4815 	 * Free RX software descriptors allocated so far, we will only handle
4816 	 * the rings that completed, the failing case will have
4817 	 * cleaned up for itself. 'q' failed, so its the terminus.
4818 	 */
4819 	rxq = ctx->ifc_rxqs;
4820 	for (i = 0; i < q; ++i, rxq++) {
4821 		iflib_rx_sds_free(rxq);
4822 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4823 	}
4824 	return (err);
4825 #endif
4826 }
4827 
4828 /*********************************************************************
4829  *
4830  *  Free all receive rings.
4831  *
4832  **********************************************************************/
4833 static void
4834 iflib_rx_structures_free(if_ctx_t ctx)
4835 {
4836 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4837 
4838 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4839 		iflib_rx_sds_free(rxq);
4840 	}
4841 }
4842 
4843 static int
4844 iflib_qset_structures_setup(if_ctx_t ctx)
4845 {
4846 	int err;
4847 
4848 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4849 		return (err);
4850 
4851 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4852 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4853 		iflib_tx_structures_free(ctx);
4854 		iflib_rx_structures_free(ctx);
4855 	}
4856 	return (err);
4857 }
4858 
4859 int
4860 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4861 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4862 {
4863 
4864 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4865 }
4866 
4867 static int
4868 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4869 {
4870 	int i, cpuid, eqid, count;
4871 
4872 	CPU_COPY(&ctx->ifc_cpus, cpus);
4873 	count = CPU_COUNT(&ctx->ifc_cpus);
4874 	eqid = qid % count;
4875 	/* clear up to the qid'th bit */
4876 	for (i = 0; i < eqid; i++) {
4877 		cpuid = CPU_FFS(cpus);
4878 		MPASS(cpuid != 0);
4879 		CPU_CLR(cpuid-1, cpus);
4880 	}
4881 	cpuid = CPU_FFS(cpus);
4882 	MPASS(cpuid != 0);
4883 	return (cpuid-1);
4884 }
4885 
4886 int
4887 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4888 						iflib_intr_type_t type, driver_filter_t *filter,
4889 						void *filter_arg, int qid, char *name)
4890 {
4891 	struct grouptask *gtask;
4892 	struct taskqgroup *tqg;
4893 	iflib_filter_info_t info;
4894 	cpuset_t cpus;
4895 	gtask_fn_t *fn;
4896 	int tqrid, err, cpuid;
4897 	driver_filter_t *intr_fast;
4898 	void *q;
4899 
4900 	info = &ctx->ifc_filter_info;
4901 	tqrid = rid;
4902 
4903 	switch (type) {
4904 	/* XXX merge tx/rx for netmap? */
4905 	case IFLIB_INTR_TX:
4906 		q = &ctx->ifc_txqs[qid];
4907 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4908 		gtask = &ctx->ifc_txqs[qid].ift_task;
4909 		tqg = qgroup_if_io_tqg;
4910 		fn = _task_fn_tx;
4911 		intr_fast = iflib_fast_intr;
4912 		GROUPTASK_INIT(gtask, 0, fn, q);
4913 		break;
4914 	case IFLIB_INTR_RX:
4915 		q = &ctx->ifc_rxqs[qid];
4916 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4917 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4918 		tqg = qgroup_if_io_tqg;
4919 		fn = _task_fn_rx;
4920 		intr_fast = iflib_fast_intr;
4921 		GROUPTASK_INIT(gtask, 0, fn, q);
4922 		break;
4923 	case IFLIB_INTR_RXTX:
4924 		q = &ctx->ifc_rxqs[qid];
4925 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4926 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4927 		tqg = qgroup_if_io_tqg;
4928 		fn = _task_fn_rx;
4929 		intr_fast = iflib_fast_intr_rxtx;
4930 		GROUPTASK_INIT(gtask, 0, fn, q);
4931 		break;
4932 	case IFLIB_INTR_ADMIN:
4933 		q = ctx;
4934 		tqrid = -1;
4935 		info = &ctx->ifc_filter_info;
4936 		gtask = &ctx->ifc_admin_task;
4937 		tqg = qgroup_if_config_tqg;
4938 		fn = _task_fn_admin;
4939 		intr_fast = iflib_fast_intr_ctx;
4940 		break;
4941 	default:
4942 		panic("unknown net intr type");
4943 	}
4944 
4945 	info->ifi_filter = filter;
4946 	info->ifi_filter_arg = filter_arg;
4947 	info->ifi_task = gtask;
4948 	info->ifi_ctx = q;
4949 
4950 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4951 	if (err != 0) {
4952 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4953 		return (err);
4954 	}
4955 	if (type == IFLIB_INTR_ADMIN)
4956 		return (0);
4957 
4958 	if (tqrid != -1) {
4959 		cpuid = find_nth(ctx, &cpus, qid);
4960 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4961 	} else {
4962 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4963 	}
4964 
4965 	return (0);
4966 }
4967 
4968 void
4969 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4970 {
4971 	struct grouptask *gtask;
4972 	struct taskqgroup *tqg;
4973 	gtask_fn_t *fn;
4974 	void *q;
4975 
4976 	switch (type) {
4977 	case IFLIB_INTR_TX:
4978 		q = &ctx->ifc_txqs[qid];
4979 		gtask = &ctx->ifc_txqs[qid].ift_task;
4980 		tqg = qgroup_if_io_tqg;
4981 		fn = _task_fn_tx;
4982 		break;
4983 	case IFLIB_INTR_RX:
4984 		q = &ctx->ifc_rxqs[qid];
4985 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4986 		tqg = qgroup_if_io_tqg;
4987 		fn = _task_fn_rx;
4988 		break;
4989 	case IFLIB_INTR_IOV:
4990 		q = ctx;
4991 		gtask = &ctx->ifc_vflr_task;
4992 		tqg = qgroup_if_config_tqg;
4993 		rid = -1;
4994 		fn = _task_fn_iov;
4995 		break;
4996 	default:
4997 		panic("unknown net intr type");
4998 	}
4999 	GROUPTASK_INIT(gtask, 0, fn, q);
5000 	taskqgroup_attach(tqg, gtask, q, rid, name);
5001 }
5002 
5003 void
5004 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5005 {
5006 	if (irq->ii_tag)
5007 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5008 
5009 	if (irq->ii_res)
5010 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5011 }
5012 
5013 static int
5014 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5015 {
5016 	iflib_txq_t txq = ctx->ifc_txqs;
5017 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5018 	if_irq_t irq = &ctx->ifc_legacy_irq;
5019 	iflib_filter_info_t info;
5020 	struct grouptask *gtask;
5021 	struct taskqgroup *tqg;
5022 	gtask_fn_t *fn;
5023 	int tqrid;
5024 	void *q;
5025 	int err;
5026 
5027 	q = &ctx->ifc_rxqs[0];
5028 	info = &rxq[0].ifr_filter_info;
5029 	gtask = &rxq[0].ifr_task;
5030 	tqg = qgroup_if_io_tqg;
5031 	tqrid = irq->ii_rid = *rid;
5032 	fn = _task_fn_rx;
5033 
5034 	ctx->ifc_flags |= IFC_LEGACY;
5035 	info->ifi_filter = filter;
5036 	info->ifi_filter_arg = filter_arg;
5037 	info->ifi_task = gtask;
5038 	info->ifi_ctx = ctx;
5039 
5040 	/* We allocate a single interrupt resource */
5041 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5042 		return (err);
5043 	GROUPTASK_INIT(gtask, 0, fn, q);
5044 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5045 
5046 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5047 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5048 	return (0);
5049 }
5050 
5051 void
5052 iflib_led_create(if_ctx_t ctx)
5053 {
5054 
5055 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5056 	    device_get_nameunit(ctx->ifc_dev));
5057 }
5058 
5059 void
5060 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5061 {
5062 
5063 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5064 }
5065 
5066 void
5067 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5068 {
5069 
5070 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5071 }
5072 
5073 void
5074 iflib_admin_intr_deferred(if_ctx_t ctx)
5075 {
5076 #ifdef INVARIANTS
5077 	struct grouptask *gtask;
5078 
5079 	gtask = &ctx->ifc_admin_task;
5080 	MPASS(gtask->gt_taskqueue != NULL);
5081 #endif
5082 
5083 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5084 }
5085 
5086 void
5087 iflib_iov_intr_deferred(if_ctx_t ctx)
5088 {
5089 
5090 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5091 }
5092 
5093 void
5094 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5095 {
5096 
5097 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5098 }
5099 
5100 void
5101 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5102 	char *name)
5103 {
5104 
5105 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5106 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5107 }
5108 
5109 void
5110 iflib_config_gtask_deinit(struct grouptask *gtask)
5111 {
5112 
5113 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5114 }
5115 
5116 void
5117 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5118 {
5119 	if_t ifp = ctx->ifc_ifp;
5120 	iflib_txq_t txq = ctx->ifc_txqs;
5121 
5122 	if_setbaudrate(ifp, baudrate);
5123 	if (baudrate >= IF_Gbps(10))
5124 		ctx->ifc_flags |= IFC_PREFETCH;
5125 
5126 	/* If link down, disable watchdog */
5127 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5128 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5129 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5130 	}
5131 	ctx->ifc_link_state = link_state;
5132 	if_link_state_change(ifp, link_state);
5133 }
5134 
5135 static int
5136 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5137 {
5138 	int credits;
5139 #ifdef INVARIANTS
5140 	int credits_pre = txq->ift_cidx_processed;
5141 #endif
5142 
5143 	if (ctx->isc_txd_credits_update == NULL)
5144 		return (0);
5145 
5146 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5147 		return (0);
5148 
5149 	txq->ift_processed += credits;
5150 	txq->ift_cidx_processed += credits;
5151 
5152 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5153 	if (txq->ift_cidx_processed >= txq->ift_size)
5154 		txq->ift_cidx_processed -= txq->ift_size;
5155 	return (credits);
5156 }
5157 
5158 static int
5159 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5160 {
5161 
5162 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5163 	    budget));
5164 }
5165 
5166 void
5167 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5168 	const char *description, if_int_delay_info_t info,
5169 	int offset, int value)
5170 {
5171 	info->iidi_ctx = ctx;
5172 	info->iidi_offset = offset;
5173 	info->iidi_value = value;
5174 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5175 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5176 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5177 	    info, 0, iflib_sysctl_int_delay, "I", description);
5178 }
5179 
5180 struct mtx *
5181 iflib_ctx_lock_get(if_ctx_t ctx)
5182 {
5183 
5184 	return (&ctx->ifc_mtx);
5185 }
5186 
5187 static int
5188 iflib_msix_init(if_ctx_t ctx)
5189 {
5190 	device_t dev = ctx->ifc_dev;
5191 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5192 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5193 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5194 	int iflib_num_tx_queues, iflib_num_rx_queues;
5195 	int err, admincnt, bar;
5196 
5197 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5198 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5199 
5200 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5201 
5202 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5203 	admincnt = sctx->isc_admin_intrcnt;
5204 	/* Override by tuneable */
5205 	if (scctx->isc_disable_msix)
5206 		goto msi;
5207 
5208 	/*
5209 	** When used in a virtualized environment
5210 	** PCI BUSMASTER capability may not be set
5211 	** so explicity set it here and rewrite
5212 	** the ENABLE in the MSIX control register
5213 	** at this point to cause the host to
5214 	** successfully initialize us.
5215 	*/
5216 	{
5217 		int msix_ctrl, rid;
5218 
5219  		pci_enable_busmaster(dev);
5220 		rid = 0;
5221 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5222 			rid += PCIR_MSIX_CTRL;
5223 			msix_ctrl = pci_read_config(dev, rid, 2);
5224 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5225 			pci_write_config(dev, rid, msix_ctrl, 2);
5226 		} else {
5227 			device_printf(dev, "PCIY_MSIX capability not found; "
5228 			                   "or rid %d == 0.\n", rid);
5229 			goto msi;
5230 		}
5231 	}
5232 
5233 	/*
5234 	 * bar == -1 => "trust me I know what I'm doing"
5235 	 * Some drivers are for hardware that is so shoddily
5236 	 * documented that no one knows which bars are which
5237 	 * so the developer has to map all bars. This hack
5238 	 * allows shoddy garbage to use msix in this framework.
5239 	 */
5240 	if (bar != -1) {
5241 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5242 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5243 		if (ctx->ifc_msix_mem == NULL) {
5244 			/* May not be enabled */
5245 			device_printf(dev, "Unable to map MSIX table \n");
5246 			goto msi;
5247 		}
5248 	}
5249 	/* First try MSI/X */
5250 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5251 		device_printf(dev, "System has MSIX disabled \n");
5252 		bus_release_resource(dev, SYS_RES_MEMORY,
5253 		    bar, ctx->ifc_msix_mem);
5254 		ctx->ifc_msix_mem = NULL;
5255 		goto msi;
5256 	}
5257 #if IFLIB_DEBUG
5258 	/* use only 1 qset in debug mode */
5259 	queuemsgs = min(msgs - admincnt, 1);
5260 #else
5261 	queuemsgs = msgs - admincnt;
5262 #endif
5263 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5264 #ifdef RSS
5265 		queues = imin(queuemsgs, rss_getnumbuckets());
5266 #else
5267 		queues = queuemsgs;
5268 #endif
5269 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5270 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5271 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5272 	} else {
5273 		device_printf(dev, "Unable to fetch CPU list\n");
5274 		/* Figure out a reasonable auto config value */
5275 		queues = min(queuemsgs, mp_ncpus);
5276 	}
5277 #ifdef  RSS
5278 	/* If we're doing RSS, clamp at the number of RSS buckets */
5279 	if (queues > rss_getnumbuckets())
5280 		queues = rss_getnumbuckets();
5281 #endif
5282 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5283 		rx_queues = iflib_num_rx_queues;
5284 	else
5285 		rx_queues = queues;
5286 	/*
5287 	 * We want this to be all logical CPUs by default
5288 	 */
5289 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5290 		tx_queues = iflib_num_tx_queues;
5291 	else
5292 		tx_queues = mp_ncpus;
5293 
5294 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5295 #ifdef INVARIANTS
5296 		if (tx_queues != rx_queues)
5297 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5298 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5299 #endif
5300 		tx_queues = min(rx_queues, tx_queues);
5301 		rx_queues = min(rx_queues, tx_queues);
5302 	}
5303 
5304 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5305 
5306 	vectors = rx_queues + admincnt;
5307 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5308 		device_printf(dev,
5309 					  "Using MSIX interrupts with %d vectors\n", vectors);
5310 		scctx->isc_vectors = vectors;
5311 		scctx->isc_nrxqsets = rx_queues;
5312 		scctx->isc_ntxqsets = tx_queues;
5313 		scctx->isc_intr = IFLIB_INTR_MSIX;
5314 
5315 		return (vectors);
5316 	} else {
5317 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5318 	}
5319 msi:
5320 	vectors = pci_msi_count(dev);
5321 	scctx->isc_nrxqsets = 1;
5322 	scctx->isc_ntxqsets = 1;
5323 	scctx->isc_vectors = vectors;
5324 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5325 		device_printf(dev,"Using an MSI interrupt\n");
5326 		scctx->isc_intr = IFLIB_INTR_MSI;
5327 	} else {
5328 		device_printf(dev,"Using a Legacy interrupt\n");
5329 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5330 	}
5331 
5332 	return (vectors);
5333 }
5334 
5335 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5336 
5337 static int
5338 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5339 {
5340 	int rc;
5341 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5342 	struct sbuf *sb;
5343 	char *ring_state = "UNKNOWN";
5344 
5345 	/* XXX needed ? */
5346 	rc = sysctl_wire_old_buffer(req, 0);
5347 	MPASS(rc == 0);
5348 	if (rc != 0)
5349 		return (rc);
5350 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5351 	MPASS(sb != NULL);
5352 	if (sb == NULL)
5353 		return (ENOMEM);
5354 	if (state[3] <= 3)
5355 		ring_state = ring_states[state[3]];
5356 
5357 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5358 		    state[0], state[1], state[2], ring_state);
5359 	rc = sbuf_finish(sb);
5360 	sbuf_delete(sb);
5361         return(rc);
5362 }
5363 
5364 enum iflib_ndesc_handler {
5365 	IFLIB_NTXD_HANDLER,
5366 	IFLIB_NRXD_HANDLER,
5367 };
5368 
5369 static int
5370 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5371 {
5372 	if_ctx_t ctx = (void *)arg1;
5373 	enum iflib_ndesc_handler type = arg2;
5374 	char buf[256] = {0};
5375 	qidx_t *ndesc;
5376 	char *p, *next;
5377 	int nqs, rc, i;
5378 
5379 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5380 
5381 	nqs = 8;
5382 	switch(type) {
5383 	case IFLIB_NTXD_HANDLER:
5384 		ndesc = ctx->ifc_sysctl_ntxds;
5385 		if (ctx->ifc_sctx)
5386 			nqs = ctx->ifc_sctx->isc_ntxqs;
5387 		break;
5388 	case IFLIB_NRXD_HANDLER:
5389 		ndesc = ctx->ifc_sysctl_nrxds;
5390 		if (ctx->ifc_sctx)
5391 			nqs = ctx->ifc_sctx->isc_nrxqs;
5392 		break;
5393 	}
5394 	if (nqs == 0)
5395 		nqs = 8;
5396 
5397 	for (i=0; i<8; i++) {
5398 		if (i >= nqs)
5399 			break;
5400 		if (i)
5401 			strcat(buf, ",");
5402 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5403 	}
5404 
5405 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5406 	if (rc || req->newptr == NULL)
5407 		return rc;
5408 
5409 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5410 	    i++, p = strsep(&next, " ,")) {
5411 		ndesc[i] = strtoul(p, NULL, 10);
5412 	}
5413 
5414 	return(rc);
5415 }
5416 
5417 #define NAME_BUFLEN 32
5418 static void
5419 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5420 {
5421         device_t dev = iflib_get_dev(ctx);
5422 	struct sysctl_oid_list *child, *oid_list;
5423 	struct sysctl_ctx_list *ctx_list;
5424 	struct sysctl_oid *node;
5425 
5426 	ctx_list = device_get_sysctl_ctx(dev);
5427 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5428 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5429 						      CTLFLAG_RD, NULL, "IFLIB fields");
5430 	oid_list = SYSCTL_CHILDREN(node);
5431 
5432 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5433 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5434 		       "driver version");
5435 
5436 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5437 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5438 			"# of txqs to use, 0 => use default #");
5439 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5440 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5441 			"# of rxqs to use, 0 => use default #");
5442 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5443 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5444                        "permit #txq != #rxq");
5445        SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5446                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5447                       "disable MSIX (default 0)");
5448 
5449 	/* XXX change for per-queue sizes */
5450 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5451 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5452                        mp_ndesc_handler, "A",
5453                        "list of # of tx descriptors to use, 0 = use default #");
5454 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5455 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5456                        mp_ndesc_handler, "A",
5457                        "list of # of rx descriptors to use, 0 = use default #");
5458 }
5459 
5460 static void
5461 iflib_add_device_sysctl_post(if_ctx_t ctx)
5462 {
5463 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5464 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5465         device_t dev = iflib_get_dev(ctx);
5466 	struct sysctl_oid_list *child;
5467 	struct sysctl_ctx_list *ctx_list;
5468 	iflib_fl_t fl;
5469 	iflib_txq_t txq;
5470 	iflib_rxq_t rxq;
5471 	int i, j;
5472 	char namebuf[NAME_BUFLEN];
5473 	char *qfmt;
5474 	struct sysctl_oid *queue_node, *fl_node, *node;
5475 	struct sysctl_oid_list *queue_list, *fl_list;
5476 	ctx_list = device_get_sysctl_ctx(dev);
5477 
5478 	node = ctx->ifc_sysctl_node;
5479 	child = SYSCTL_CHILDREN(node);
5480 
5481 	if (scctx->isc_ntxqsets > 100)
5482 		qfmt = "txq%03d";
5483 	else if (scctx->isc_ntxqsets > 10)
5484 		qfmt = "txq%02d";
5485 	else
5486 		qfmt = "txq%d";
5487 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5488 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5489 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5490 					     CTLFLAG_RD, NULL, "Queue Name");
5491 		queue_list = SYSCTL_CHILDREN(queue_node);
5492 #if MEMORY_LOGGING
5493 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5494 				CTLFLAG_RD,
5495 				&txq->ift_dequeued, "total mbufs freed");
5496 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5497 				CTLFLAG_RD,
5498 				&txq->ift_enqueued, "total mbufs enqueued");
5499 #endif
5500 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5501 				   CTLFLAG_RD,
5502 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5503 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5504 				   CTLFLAG_RD,
5505 				   &txq->ift_pullups, "# of times m_pullup was called");
5506 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5507 				   CTLFLAG_RD,
5508 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5509 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5510 				   CTLFLAG_RD,
5511 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5512 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5513 				   CTLFLAG_RD,
5514 				   &txq->ift_map_failed, "# of times dma map failed");
5515 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5516 				   CTLFLAG_RD,
5517 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5518 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5519 				   CTLFLAG_RD,
5520 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5521 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5522 				   CTLFLAG_RD,
5523 				   &txq->ift_pidx, 1, "Producer Index");
5524 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5525 				   CTLFLAG_RD,
5526 				   &txq->ift_cidx, 1, "Consumer Index");
5527 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5528 				   CTLFLAG_RD,
5529 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5530 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5531 				   CTLFLAG_RD,
5532 				   &txq->ift_in_use, 1, "descriptors in use");
5533 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5534 				   CTLFLAG_RD,
5535 				   &txq->ift_processed, "descriptors procesed for clean");
5536 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5537 				   CTLFLAG_RD,
5538 				   &txq->ift_cleaned, "total cleaned");
5539 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5540 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5541 				0, mp_ring_state_handler, "A", "soft ring state");
5542 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5543 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5544 				       "# of enqueues to the mp_ring for this queue");
5545 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5546 				       CTLFLAG_RD, &txq->ift_br->drops,
5547 				       "# of drops in the mp_ring for this queue");
5548 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5549 				       CTLFLAG_RD, &txq->ift_br->starts,
5550 				       "# of normal consumer starts in the mp_ring for this queue");
5551 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5552 				       CTLFLAG_RD, &txq->ift_br->stalls,
5553 					       "# of consumer stalls in the mp_ring for this queue");
5554 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5555 			       CTLFLAG_RD, &txq->ift_br->restarts,
5556 				       "# of consumer restarts in the mp_ring for this queue");
5557 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5558 				       CTLFLAG_RD, &txq->ift_br->abdications,
5559 				       "# of consumer abdications in the mp_ring for this queue");
5560 	}
5561 
5562 	if (scctx->isc_nrxqsets > 100)
5563 		qfmt = "rxq%03d";
5564 	else if (scctx->isc_nrxqsets > 10)
5565 		qfmt = "rxq%02d";
5566 	else
5567 		qfmt = "rxq%d";
5568 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5569 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5570 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5571 					     CTLFLAG_RD, NULL, "Queue Name");
5572 		queue_list = SYSCTL_CHILDREN(queue_node);
5573 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5574 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5575 				       CTLFLAG_RD,
5576 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5577 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5578 				       CTLFLAG_RD,
5579 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5580 		}
5581 
5582 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5583 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5584 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5585 						     CTLFLAG_RD, NULL, "freelist Name");
5586 			fl_list = SYSCTL_CHILDREN(fl_node);
5587 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5588 				       CTLFLAG_RD,
5589 				       &fl->ifl_pidx, 1, "Producer Index");
5590 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5591 				       CTLFLAG_RD,
5592 				       &fl->ifl_cidx, 1, "Consumer Index");
5593 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5594 				       CTLFLAG_RD,
5595 				       &fl->ifl_credits, 1, "credits available");
5596 #if MEMORY_LOGGING
5597 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5598 					CTLFLAG_RD,
5599 					&fl->ifl_m_enqueued, "mbufs allocated");
5600 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5601 					CTLFLAG_RD,
5602 					&fl->ifl_m_dequeued, "mbufs freed");
5603 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5604 					CTLFLAG_RD,
5605 					&fl->ifl_cl_enqueued, "clusters allocated");
5606 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5607 					CTLFLAG_RD,
5608 					&fl->ifl_cl_dequeued, "clusters freed");
5609 #endif
5610 
5611 		}
5612 	}
5613 
5614 }
5615 
5616 #ifndef __NO_STRICT_ALIGNMENT
5617 static struct mbuf *
5618 iflib_fixup_rx(struct mbuf *m)
5619 {
5620 	struct mbuf *n;
5621 
5622 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5623 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5624 		m->m_data += ETHER_HDR_LEN;
5625 		n = m;
5626 	} else {
5627 		MGETHDR(n, M_NOWAIT, MT_DATA);
5628 		if (n == NULL) {
5629 			m_freem(m);
5630 			return (NULL);
5631 		}
5632 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5633 		m->m_data += ETHER_HDR_LEN;
5634 		m->m_len -= ETHER_HDR_LEN;
5635 		n->m_len = ETHER_HDR_LEN;
5636 		M_MOVE_PKTHDR(n, m);
5637 		n->m_next = m;
5638 	}
5639 	return (n);
5640 }
5641 #endif
5642