xref: /freebsd/sys/net/iflib.c (revision 264104f26834fdb27974e0c5fdedf8f2f5a90383)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 #include <sys/bitstring.h>
97 /*
98  * enable accounting of every mbuf as it comes in to and goes out of
99  * iflib's software descriptor references
100  */
101 #define MEMORY_LOGGING 0
102 /*
103  * Enable mbuf vectors for compressing long mbuf chains
104  */
105 
106 /*
107  * NB:
108  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
109  *   we prefetch needs to be determined by the time spent in m_free vis a vis
110  *   the cost of a prefetch. This will of course vary based on the workload:
111  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
112  *        is quite expensive, thus suggesting very little prefetch.
113  *      - small packet forwarding which is just returning a single mbuf to
114  *        UMA will typically be very fast vis a vis the cost of a memory
115  *        access.
116  */
117 
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 struct iflib_txq;
132 typedef struct iflib_txq *iflib_txq_t;
133 struct iflib_rxq;
134 typedef struct iflib_rxq *iflib_rxq_t;
135 struct iflib_fl;
136 typedef struct iflib_fl *iflib_fl_t;
137 
138 struct iflib_ctx;
139 
140 typedef struct iflib_filter_info {
141 	driver_filter_t *ifi_filter;
142 	void *ifi_filter_arg;
143 	struct grouptask *ifi_task;
144 	void *ifi_ctx;
145 } *iflib_filter_info_t;
146 
147 struct iflib_ctx {
148 	KOBJ_FIELDS;
149    /*
150    * Pointer to hardware driver's softc
151    */
152 	void *ifc_softc;
153 	device_t ifc_dev;
154 	if_t ifc_ifp;
155 
156 	cpuset_t ifc_cpus;
157 	if_shared_ctx_t ifc_sctx;
158 	struct if_softc_ctx ifc_softc_ctx;
159 
160 	struct mtx ifc_mtx;
161 
162 	uint16_t ifc_nhwtxqs;
163 	uint16_t ifc_nhwrxqs;
164 
165 	iflib_txq_t ifc_txqs;
166 	iflib_rxq_t ifc_rxqs;
167 	uint32_t ifc_if_flags;
168 	uint32_t ifc_flags;
169 	uint32_t ifc_max_fl_buf_size;
170 	int ifc_in_detach;
171 
172 	int ifc_link_state;
173 	int ifc_link_irq;
174 	int ifc_watchdog_events;
175 	struct cdev *ifc_led_dev;
176 	struct resource *ifc_msix_mem;
177 
178 	struct if_irq ifc_legacy_irq;
179 	struct grouptask ifc_admin_task;
180 	struct grouptask ifc_vflr_task;
181 	struct iflib_filter_info ifc_filter_info;
182 	struct ifmedia	ifc_media;
183 
184 	struct sysctl_oid *ifc_sysctl_node;
185 	uint16_t ifc_sysctl_ntxqs;
186 	uint16_t ifc_sysctl_nrxqs;
187 	uint16_t ifc_sysctl_qs_eq_override;
188 
189 	qidx_t ifc_sysctl_ntxds[8];
190 	qidx_t ifc_sysctl_nrxds[8];
191 	struct if_txrx ifc_txrx;
192 #define isc_txd_encap  ifc_txrx.ift_txd_encap
193 #define isc_txd_flush  ifc_txrx.ift_txd_flush
194 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
195 #define isc_rxd_available ifc_txrx.ift_rxd_available
196 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
197 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
198 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
201 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
202 	eventhandler_tag ifc_vlan_attach_event;
203 	eventhandler_tag ifc_vlan_detach_event;
204 	uint8_t ifc_mac[ETHER_ADDR_LEN];
205 	char ifc_mtx_name[16];
206 };
207 
208 
209 void *
210 iflib_get_softc(if_ctx_t ctx)
211 {
212 
213 	return (ctx->ifc_softc);
214 }
215 
216 device_t
217 iflib_get_dev(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_dev);
221 }
222 
223 if_t
224 iflib_get_ifp(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_ifp);
228 }
229 
230 struct ifmedia *
231 iflib_get_media(if_ctx_t ctx)
232 {
233 
234 	return (&ctx->ifc_media);
235 }
236 
237 void
238 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
239 {
240 
241 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
242 }
243 
244 if_softc_ctx_t
245 iflib_get_softc_ctx(if_ctx_t ctx)
246 {
247 
248 	return (&ctx->ifc_softc_ctx);
249 }
250 
251 if_shared_ctx_t
252 iflib_get_sctx(if_ctx_t ctx)
253 {
254 
255 	return (ctx->ifc_sctx);
256 }
257 
258 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
259 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
260 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
261 
262 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
263 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
264 
265 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
266 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
267 #define RX_SW_DESC_INUSE        (1 << 3)
268 #define TX_SW_DESC_MAPPED       (1 << 4)
269 
270 typedef struct iflib_sw_rx_desc_array {
271 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
272 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
273 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
274 	uint8_t		*ifsd_flags;
275 } iflib_rxsd_array_t;
276 
277 typedef struct iflib_sw_tx_desc_array {
278 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
280 	uint8_t		*ifsd_flags;
281 } if_txsd_vec_t;
282 
283 
284 /* magic number that should be high enough for any hardware */
285 #define IFLIB_MAX_TX_SEGS		128
286 #define IFLIB_MAX_RX_SEGS		32
287 #define IFLIB_RX_COPY_THRESH		128
288 #define IFLIB_MAX_RX_REFRESH		32
289 /* The minimum descriptors per second before we start coalescing */
290 #define IFLIB_MIN_DESC_SEC		16384
291 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
292 #define IFLIB_QUEUE_IDLE		0
293 #define IFLIB_QUEUE_HUNG		1
294 #define IFLIB_QUEUE_WORKING		2
295 /* maximum number of txqs that can share an rx interrupt */
296 #define IFLIB_MAX_TX_SHARED_INTR	4
297 
298 /* this should really scale with ring size - this is a fairly arbitrary value */
299 #define TX_BATCH_SIZE			32
300 
301 #define IFLIB_RESTART_BUDGET		8
302 
303 #define	IFC_LEGACY		0x001
304 #define	IFC_QFLUSH		0x002
305 #define	IFC_MULTISEG		0x004
306 #define	IFC_DMAR		0x008
307 #define	IFC_SC_ALLOCATED	0x010
308 #define	IFC_INIT_DONE		0x020
309 #define	IFC_PREFETCH		0x040
310 #define	IFC_DO_RESET		0x080
311 #define	IFC_CHECK_HUNG		0x100
312 
313 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
314 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
315 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
316 struct iflib_txq {
317 	qidx_t		ift_in_use;
318 	qidx_t		ift_cidx;
319 	qidx_t		ift_cidx_processed;
320 	qidx_t		ift_pidx;
321 	uint8_t		ift_gen;
322 	uint8_t		ift_br_offset;
323 	uint16_t	ift_npending;
324 	uint16_t	ift_db_pending;
325 	uint16_t	ift_rs_pending;
326 	/* implicit pad */
327 	uint8_t		ift_txd_size[8];
328 	uint64_t	ift_processed;
329 	uint64_t	ift_cleaned;
330 	uint64_t	ift_cleaned_prev;
331 #if MEMORY_LOGGING
332 	uint64_t	ift_enqueued;
333 	uint64_t	ift_dequeued;
334 #endif
335 	uint64_t	ift_no_tx_dma_setup;
336 	uint64_t	ift_no_desc_avail;
337 	uint64_t	ift_mbuf_defrag_failed;
338 	uint64_t	ift_mbuf_defrag;
339 	uint64_t	ift_map_failed;
340 	uint64_t	ift_txd_encap_efbig;
341 	uint64_t	ift_pullups;
342 
343 	struct mtx	ift_mtx;
344 	struct mtx	ift_db_mtx;
345 
346 	/* constant values */
347 	if_ctx_t	ift_ctx;
348 	struct ifmp_ring        *ift_br;
349 	struct grouptask	ift_task;
350 	qidx_t		ift_size;
351 	uint16_t	ift_id;
352 	struct callout	ift_timer;
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t		ift_desc_tag;
360 	bus_dma_tag_t		ift_tso_desc_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define MTX_NAME_LEN 16
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	char                    ift_db_mtx_name[MTX_NAME_LEN];
365 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
366 #ifdef IFLIB_DIAGNOSTICS
367 	uint64_t ift_cpu_exec_count[256];
368 #endif
369 } __aligned(CACHE_LINE_SIZE);
370 
371 struct iflib_fl {
372 	qidx_t		ifl_cidx;
373 	qidx_t		ifl_pidx;
374 	qidx_t		ifl_credits;
375 	uint8_t		ifl_gen;
376 	uint8_t		ifl_rxd_size;
377 #if MEMORY_LOGGING
378 	uint64_t	ifl_m_enqueued;
379 	uint64_t	ifl_m_dequeued;
380 	uint64_t	ifl_cl_enqueued;
381 	uint64_t	ifl_cl_dequeued;
382 #endif
383 	/* implicit pad */
384 
385 	bitstr_t 	*ifl_rx_bitmap;;
386 	qidx_t		ifl_fragidx;
387 	/* constant */
388 	qidx_t		ifl_size;
389 	uint16_t	ifl_buf_size;
390 	uint16_t	ifl_cltype;
391 	uma_zone_t	ifl_zone;
392 	iflib_rxsd_array_t	ifl_sds;
393 	iflib_rxq_t	ifl_rxq;
394 	uint8_t		ifl_id;
395 	bus_dma_tag_t           ifl_desc_tag;
396 	iflib_dma_info_t	ifl_ifdi;
397 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
398 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
399 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
400 }  __aligned(CACHE_LINE_SIZE);
401 
402 static inline qidx_t
403 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
404 {
405 	qidx_t used;
406 
407 	if (pidx > cidx)
408 		used = pidx - cidx;
409 	else if (pidx < cidx)
410 		used = size - cidx + pidx;
411 	else if (gen == 0 && pidx == cidx)
412 		used = 0;
413 	else if (gen == 1 && pidx == cidx)
414 		used = size;
415 	else
416 		panic("bad state");
417 
418 	return (used);
419 }
420 
421 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 
423 #define IDXDIFF(head, tail, wrap) \
424 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
425 
426 struct iflib_rxq {
427 	/* If there is a separate completion queue -
428 	 * these are the cq cidx and pidx. Otherwise
429 	 * these are unused.
430 	 */
431 	qidx_t		ifr_size;
432 	qidx_t		ifr_cq_cidx;
433 	qidx_t		ifr_cq_pidx;
434 	uint8_t		ifr_cq_gen;
435 	uint8_t		ifr_fl_offset;
436 
437 	if_ctx_t	ifr_ctx;
438 	iflib_fl_t	ifr_fl;
439 	uint64_t	ifr_rx_irq;
440 	uint16_t	ifr_id;
441 	uint8_t		ifr_lro_enabled;
442 	uint8_t		ifr_nfl;
443 	uint8_t		ifr_ntxqirq;
444 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
445 	struct lro_ctrl			ifr_lc;
446 	struct grouptask        ifr_task;
447 	struct iflib_filter_info ifr_filter_info;
448 	iflib_dma_info_t		ifr_ifdi;
449 
450 	/* dynamically allocate if any drivers need a value substantially larger than this */
451 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
452 #ifdef IFLIB_DIAGNOSTICS
453 	uint64_t ifr_cpu_exec_count[256];
454 #endif
455 }  __aligned(CACHE_LINE_SIZE);
456 
457 typedef struct if_rxsd {
458 	caddr_t *ifsd_cl;
459 	struct mbuf **ifsd_m;
460 	iflib_fl_t ifsd_fl;
461 	qidx_t ifsd_cidx;
462 } *if_rxsd_t;
463 
464 /* multiple of word size */
465 #ifdef __LP64__
466 #define PKT_INFO_SIZE	6
467 #define RXD_INFO_SIZE	5
468 #define PKT_TYPE uint64_t
469 #else
470 #define PKT_INFO_SIZE	11
471 #define RXD_INFO_SIZE	8
472 #define PKT_TYPE uint32_t
473 #endif
474 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
475 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
476 
477 typedef struct if_pkt_info_pad {
478 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
479 } *if_pkt_info_pad_t;
480 typedef struct if_rxd_info_pad {
481 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
482 } *if_rxd_info_pad_t;
483 
484 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
485 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
486 
487 
488 static inline void
489 pkt_info_zero(if_pkt_info_t pi)
490 {
491 	if_pkt_info_pad_t pi_pad;
492 
493 	pi_pad = (if_pkt_info_pad_t)pi;
494 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
495 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
496 #ifndef __LP64__
497 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
498 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
499 #endif
500 }
501 
502 static inline void
503 rxd_info_zero(if_rxd_info_t ri)
504 {
505 	if_rxd_info_pad_t ri_pad;
506 	int i;
507 
508 	ri_pad = (if_rxd_info_pad_t)ri;
509 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
510 		ri_pad->rxd_val[i] = 0;
511 		ri_pad->rxd_val[i+1] = 0;
512 		ri_pad->rxd_val[i+2] = 0;
513 		ri_pad->rxd_val[i+3] = 0;
514 	}
515 #ifdef __LP64__
516 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
517 #endif
518 }
519 
520 /*
521  * Only allow a single packet to take up most 1/nth of the tx ring
522  */
523 #define MAX_SINGLE_PACKET_FRACTION 12
524 #define IF_BAD_DMA (bus_addr_t)-1
525 
526 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
527 
528 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
529 
530 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
531 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
532 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
533 
534 
535 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
536 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
537 
538 
539 /* Our boot-time initialization hook */
540 static int	iflib_module_event_handler(module_t, int, void *);
541 
542 static moduledata_t iflib_moduledata = {
543 	"iflib",
544 	iflib_module_event_handler,
545 	NULL
546 };
547 
548 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
549 MODULE_VERSION(iflib, 1);
550 
551 MODULE_DEPEND(iflib, pci, 1, 1, 1);
552 MODULE_DEPEND(iflib, ether, 1, 1, 1);
553 
554 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
555 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
556 
557 #ifndef IFLIB_DEBUG_COUNTERS
558 #ifdef INVARIANTS
559 #define IFLIB_DEBUG_COUNTERS 1
560 #else
561 #define IFLIB_DEBUG_COUNTERS 0
562 #endif /* !INVARIANTS */
563 #endif
564 
565 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
566                    "iflib driver parameters");
567 
568 /*
569  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
570  */
571 static int iflib_min_tx_latency = 0;
572 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
573 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
574 static int iflib_no_tx_batch = 0;
575 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
576 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
577 
578 
579 #if IFLIB_DEBUG_COUNTERS
580 
581 static int iflib_tx_seen;
582 static int iflib_tx_sent;
583 static int iflib_tx_encap;
584 static int iflib_rx_allocs;
585 static int iflib_fl_refills;
586 static int iflib_fl_refills_large;
587 static int iflib_tx_frees;
588 
589 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
590 		   &iflib_tx_seen, 0, "# tx mbufs seen");
591 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
592 		   &iflib_tx_sent, 0, "# tx mbufs sent");
593 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
594 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
595 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
596 		   &iflib_tx_frees, 0, "# tx frees");
597 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
598 		   &iflib_rx_allocs, 0, "# rx allocations");
599 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
600 		   &iflib_fl_refills, 0, "# refills");
601 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
602 		   &iflib_fl_refills_large, 0, "# large refills");
603 
604 
605 static int iflib_txq_drain_flushing;
606 static int iflib_txq_drain_oactive;
607 static int iflib_txq_drain_notready;
608 static int iflib_txq_drain_encapfail;
609 
610 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
611 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
612 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
613 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
614 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
615 		   &iflib_txq_drain_notready, 0, "# drain notready");
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
617 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
618 
619 
620 static int iflib_encap_load_mbuf_fail;
621 static int iflib_encap_txq_avail_fail;
622 static int iflib_encap_txd_encap_fail;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
625 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
626 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
627 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
628 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
629 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
630 
631 static int iflib_task_fn_rxs;
632 static int iflib_rx_intr_enables;
633 static int iflib_fast_intrs;
634 static int iflib_intr_link;
635 static int iflib_intr_msix;
636 static int iflib_rx_unavail;
637 static int iflib_rx_ctx_inactive;
638 static int iflib_rx_zero_len;
639 static int iflib_rx_if_input;
640 static int iflib_rx_mbuf_null;
641 static int iflib_rxd_flush;
642 
643 static int iflib_verbose_debug;
644 
645 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
646 		   &iflib_intr_link, 0, "# intr link calls");
647 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
648 		   &iflib_intr_msix, 0, "# intr msix calls");
649 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
650 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
651 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
652 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
653 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
654 		   &iflib_fast_intrs, 0, "# fast_intr calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
656 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
658 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
660 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
662 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
664 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
678 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
679 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
680 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
681 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
682 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
683 }
684 
685 #else
686 #define DBG_COUNTER_INC(name)
687 static void iflib_debug_reset(void) {}
688 #endif
689 
690 
691 
692 #define IFLIB_DEBUG 0
693 
694 static void iflib_tx_structures_free(if_ctx_t ctx);
695 static void iflib_rx_structures_free(if_ctx_t ctx);
696 static int iflib_queues_alloc(if_ctx_t ctx);
697 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
698 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
699 static int iflib_qset_structures_setup(if_ctx_t ctx);
700 static int iflib_msix_init(if_ctx_t ctx);
701 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
702 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
703 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
704 static int iflib_register(if_ctx_t);
705 static void iflib_init_locked(if_ctx_t ctx);
706 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
707 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
708 static void iflib_ifmp_purge(iflib_txq_t txq);
709 static void _iflib_pre_assert(if_softc_ctx_t scctx);
710 static void iflib_stop(if_ctx_t ctx);
711 static void iflib_if_init_locked(if_ctx_t ctx);
712 #ifndef __NO_STRICT_ALIGNMENT
713 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
714 #endif
715 
716 #ifdef DEV_NETMAP
717 #include <sys/selinfo.h>
718 #include <net/netmap.h>
719 #include <dev/netmap/netmap_kern.h>
720 
721 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
722 
723 /*
724  * device-specific sysctl variables:
725  *
726  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
727  *	During regular operations the CRC is stripped, but on some
728  *	hardware reception of frames not multiple of 64 is slower,
729  *	so using crcstrip=0 helps in benchmarks.
730  *
731  * iflib_rx_miss, iflib_rx_miss_bufs:
732  *	count packets that might be missed due to lost interrupts.
733  */
734 SYSCTL_DECL(_dev_netmap);
735 /*
736  * The xl driver by default strips CRCs and we do not override it.
737  */
738 
739 int iflib_crcstrip = 1;
740 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
741     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
742 
743 int iflib_rx_miss, iflib_rx_miss_bufs;
744 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
745     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
746 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
747     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
748 
749 /*
750  * Register/unregister. We are already under netmap lock.
751  * Only called on the first register or the last unregister.
752  */
753 static int
754 iflib_netmap_register(struct netmap_adapter *na, int onoff)
755 {
756 	struct ifnet *ifp = na->ifp;
757 	if_ctx_t ctx = ifp->if_softc;
758 	int status;
759 
760 	CTX_LOCK(ctx);
761 	IFDI_INTR_DISABLE(ctx);
762 
763 	/* Tell the stack that the interface is no longer active */
764 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
765 
766 	if (!CTX_IS_VF(ctx))
767 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
768 
769 	/* enable or disable flags and callbacks in na and ifp */
770 	if (onoff) {
771 		nm_set_native_flags(na);
772 	} else {
773 		nm_clear_native_flags(na);
774 	}
775 	iflib_stop(ctx);
776 	iflib_init_locked(ctx);
777 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
778 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
779 	if (status)
780 		nm_clear_native_flags(na);
781 	CTX_UNLOCK(ctx);
782 	return (status);
783 }
784 
785 /*
786  * Reconcile kernel and user view of the transmit ring.
787  *
788  * All information is in the kring.
789  * Userspace wants to send packets up to the one before kring->rhead,
790  * kernel knows kring->nr_hwcur is the first unsent packet.
791  *
792  * Here we push packets out (as many as possible), and possibly
793  * reclaim buffers from previously completed transmission.
794  *
795  * The caller (netmap) guarantees that there is only one instance
796  * running at any time. Any interference with other driver
797  * methods should be handled by the individual drivers.
798  */
799 static int
800 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
801 {
802 	struct netmap_adapter *na = kring->na;
803 	struct ifnet *ifp = na->ifp;
804 	struct netmap_ring *ring = kring->ring;
805 	u_int nm_i;	/* index into the netmap ring */
806 	u_int nic_i;	/* index into the NIC ring */
807 	u_int n;
808 	u_int const lim = kring->nkr_num_slots - 1;
809 	u_int const head = kring->rhead;
810 	struct if_pkt_info pi;
811 
812 	/*
813 	 * interrupts on every tx packet are expensive so request
814 	 * them every half ring, or where NS_REPORT is set
815 	 */
816 	u_int report_frequency = kring->nkr_num_slots >> 1;
817 	/* device-specific */
818 	if_ctx_t ctx = ifp->if_softc;
819 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
820 
821 	if (txq->ift_sds.ifsd_map)
822 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
823 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
824 
825 
826 	/*
827 	 * First part: process new packets to send.
828 	 * nm_i is the current index in the netmap ring,
829 	 * nic_i is the corresponding index in the NIC ring.
830 	 *
831 	 * If we have packets to send (nm_i != head)
832 	 * iterate over the netmap ring, fetch length and update
833 	 * the corresponding slot in the NIC ring. Some drivers also
834 	 * need to update the buffer's physical address in the NIC slot
835 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
836 	 *
837 	 * The netmap_reload_map() calls is especially expensive,
838 	 * even when (as in this case) the tag is 0, so do only
839 	 * when the buffer has actually changed.
840 	 *
841 	 * If possible do not set the report/intr bit on all slots,
842 	 * but only a few times per ring or when NS_REPORT is set.
843 	 *
844 	 * Finally, on 10G and faster drivers, it might be useful
845 	 * to prefetch the next slot and txr entry.
846 	 */
847 
848 	nm_i = kring->nr_hwcur;
849 	pkt_info_zero(&pi);
850 	pi.ipi_segs = txq->ift_segs;
851 	pi.ipi_qsidx = kring->ring_id;
852 	if (nm_i != head) {	/* we have new packets to send */
853 		nic_i = netmap_idx_k2n(kring, nm_i);
854 
855 		__builtin_prefetch(&ring->slot[nm_i]);
856 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
857 		if (txq->ift_sds.ifsd_map)
858 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
859 
860 		for (n = 0; nm_i != head; n++) {
861 			struct netmap_slot *slot = &ring->slot[nm_i];
862 			u_int len = slot->len;
863 			uint64_t paddr;
864 			void *addr = PNMB(na, slot, &paddr);
865 			int flags = (slot->flags & NS_REPORT ||
866 				nic_i == 0 || nic_i == report_frequency) ?
867 				IPI_TX_INTR : 0;
868 
869 			/* device-specific */
870 			pi.ipi_len = len;
871 			pi.ipi_segs[0].ds_addr = paddr;
872 			pi.ipi_segs[0].ds_len = len;
873 			pi.ipi_nsegs = 1;
874 			pi.ipi_ndescs = 0;
875 			pi.ipi_pidx = nic_i;
876 			pi.ipi_flags = flags;
877 
878 			/* Fill the slot in the NIC ring. */
879 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
880 
881 			/* prefetch for next round */
882 			__builtin_prefetch(&ring->slot[nm_i + 1]);
883 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
884 			if (txq->ift_sds.ifsd_map) {
885 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
886 
887 				NM_CHECK_ADDR_LEN(na, addr, len);
888 
889 				if (slot->flags & NS_BUF_CHANGED) {
890 					/* buffer has changed, reload map */
891 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
892 				}
893 				/* make sure changes to the buffer are synced */
894 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
895 						BUS_DMASYNC_PREWRITE);
896 			}
897 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
898 			nm_i = nm_next(nm_i, lim);
899 			nic_i = nm_next(nic_i, lim);
900 		}
901 		kring->nr_hwcur = head;
902 
903 		/* synchronize the NIC ring */
904 		if (txq->ift_sds.ifsd_map)
905 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
906 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
907 
908 		/* (re)start the tx unit up to slot nic_i (excluded) */
909 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
910 	}
911 
912 	/*
913 	 * Second part: reclaim buffers for completed transmissions.
914 	 */
915 	if (iflib_tx_credits_update(ctx, txq)) {
916 		/* some tx completed, increment avail */
917 		nic_i = txq->ift_cidx_processed;
918 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
919 	}
920 	return (0);
921 }
922 
923 /*
924  * Reconcile kernel and user view of the receive ring.
925  * Same as for the txsync, this routine must be efficient.
926  * The caller guarantees a single invocations, but races against
927  * the rest of the driver should be handled here.
928  *
929  * On call, kring->rhead is the first packet that userspace wants
930  * to keep, and kring->rcur is the wakeup point.
931  * The kernel has previously reported packets up to kring->rtail.
932  *
933  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
934  * of whether or not we received an interrupt.
935  */
936 static int
937 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
938 {
939 	struct netmap_adapter *na = kring->na;
940 	struct netmap_ring *ring = kring->ring;
941 	uint32_t nm_i;	/* index into the netmap ring */
942 	uint32_t nic_i, nic_i_start;	/* index into the NIC ring */
943 	u_int i, n;
944 	u_int const lim = kring->nkr_num_slots - 1;
945 	u_int const head = kring->rhead;
946 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
947 	struct if_rxd_info ri;
948 	struct if_rxd_update iru;
949 
950 	struct ifnet *ifp = na->ifp;
951 	if_ctx_t ctx = ifp->if_softc;
952 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
953 	iflib_fl_t fl = rxq->ifr_fl;
954 	if (head > lim)
955 		return netmap_ring_reinit(kring);
956 
957 	/* XXX check sync modes */
958 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
959 		if (fl->ifl_sds.ifsd_map == NULL)
960 			continue;
961 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
962 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
963 	}
964 	/*
965 	 * First part: import newly received packets.
966 	 *
967 	 * nm_i is the index of the next free slot in the netmap ring,
968 	 * nic_i is the index of the next received packet in the NIC ring,
969 	 * and they may differ in case if_init() has been called while
970 	 * in netmap mode. For the receive ring we have
971 	 *
972 	 *	nic_i = rxr->next_check;
973 	 *	nm_i = kring->nr_hwtail (previous)
974 	 * and
975 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
976 	 *
977 	 * rxr->next_check is set to 0 on a ring reinit
978 	 */
979 	if (netmap_no_pendintr || force_update) {
980 		int crclen = iflib_crcstrip ? 0 : 4;
981 		int error, avail;
982 		uint16_t slot_flags = kring->nkr_slot_flags;
983 
984 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
985 			nic_i = fl->ifl_cidx;
986 			nm_i = netmap_idx_n2k(kring, nic_i);
987 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
988 			for (n = 0; avail > 0; n++, avail--) {
989 				rxd_info_zero(&ri);
990 				ri.iri_frags = rxq->ifr_frags;
991 				ri.iri_qsidx = kring->ring_id;
992 				ri.iri_ifp = ctx->ifc_ifp;
993 				ri.iri_cidx = nic_i;
994 
995 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
996 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
997 				ring->slot[nm_i].flags = slot_flags;
998 				if (fl->ifl_sds.ifsd_map)
999 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1000 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1001 				nm_i = nm_next(nm_i, lim);
1002 				nic_i = nm_next(nic_i, lim);
1003 			}
1004 			if (n) { /* update the state variables */
1005 				if (netmap_no_pendintr && !force_update) {
1006 					/* diagnostics */
1007 					iflib_rx_miss ++;
1008 					iflib_rx_miss_bufs += n;
1009 				}
1010 				fl->ifl_cidx = nic_i;
1011 				kring->nr_hwtail = nm_i;
1012 			}
1013 			kring->nr_kflags &= ~NKR_PENDINTR;
1014 		}
1015 	}
1016 	/*
1017 	 * Second part: skip past packets that userspace has released.
1018 	 * (kring->nr_hwcur to head excluded),
1019 	 * and make the buffers available for reception.
1020 	 * As usual nm_i is the index in the netmap ring,
1021 	 * nic_i is the index in the NIC ring, and
1022 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1023 	 */
1024 	/* XXX not sure how this will work with multiple free lists */
1025 	nm_i = kring->nr_hwcur;
1026 	if (nm_i == head)
1027 		return (0);
1028 
1029 	iru.iru_paddrs = fl->ifl_bus_addrs;
1030 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1031 	iru.iru_idxs = fl->ifl_rxd_idxs;
1032 	iru.iru_qsidx = rxq->ifr_id;
1033 	iru.iru_buf_size = fl->ifl_buf_size;
1034 	iru.iru_flidx = fl->ifl_id;
1035 	nic_i_start = nic_i = netmap_idx_k2n(kring, nm_i);
1036 	for (i = 0; nm_i != head; i++) {
1037 		struct netmap_slot *slot = &ring->slot[nm_i];
1038 		void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
1039 
1040 		if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
1041 			goto ring_reset;
1042 
1043 		fl->ifl_vm_addrs[i] = addr;
1044 		if (fl->ifl_sds.ifsd_map && (slot->flags & NS_BUF_CHANGED)) {
1045 			/* buffer has changed, reload map */
1046 			netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
1047 		}
1048 		slot->flags &= ~NS_BUF_CHANGED;
1049 
1050 		nm_i = nm_next(nm_i, lim);
1051 		fl->ifl_rxd_idxs[i] = nic_i = nm_next(nic_i, lim);
1052 		if (nm_i != head && i < IFLIB_MAX_RX_REFRESH)
1053 			continue;
1054 
1055 		iru.iru_pidx = nic_i_start;
1056 		iru.iru_count = i;
1057 		i = 0;
1058 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1059 		if (fl->ifl_sds.ifsd_map == NULL) {
1060 			nic_i_start = nic_i;
1061 			continue;
1062 		}
1063 		nic_i = nic_i_start;
1064 		for (n = 0; n < iru.iru_count; n++) {
1065 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
1066 					BUS_DMASYNC_PREREAD);
1067 			nic_i = nm_next(nic_i, lim);
1068 		}
1069 		nic_i_start = nic_i;
1070 	}
1071 	kring->nr_hwcur = head;
1072 
1073 	if (fl->ifl_sds.ifsd_map)
1074 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1075 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1076 	/*
1077 	 * IMPORTANT: we must leave one free slot in the ring,
1078 	 * so move nic_i back by one unit
1079 	 */
1080 	nic_i = nm_prev(nic_i, lim);
1081 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
1082 	return 0;
1083 
1084 ring_reset:
1085 	return netmap_ring_reinit(kring);
1086 }
1087 
1088 static void
1089 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1090 {
1091 	struct ifnet *ifp = na->ifp;
1092 	if_ctx_t ctx = ifp->if_softc;
1093 
1094 	CTX_LOCK(ctx);
1095 	if (onoff) {
1096 		IFDI_INTR_ENABLE(ctx);
1097 	} else {
1098 		IFDI_INTR_DISABLE(ctx);
1099 	}
1100 	CTX_UNLOCK(ctx);
1101 }
1102 
1103 
1104 static int
1105 iflib_netmap_attach(if_ctx_t ctx)
1106 {
1107 	struct netmap_adapter na;
1108 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1109 
1110 	bzero(&na, sizeof(na));
1111 
1112 	na.ifp = ctx->ifc_ifp;
1113 	na.na_flags = NAF_BDG_MAYSLEEP;
1114 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1115 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1116 
1117 	na.num_tx_desc = scctx->isc_ntxd[0];
1118 	na.num_rx_desc = scctx->isc_nrxd[0];
1119 	na.nm_txsync = iflib_netmap_txsync;
1120 	na.nm_rxsync = iflib_netmap_rxsync;
1121 	na.nm_register = iflib_netmap_register;
1122 	na.nm_intr = iflib_netmap_intr;
1123 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1124 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1125 	return (netmap_attach(&na));
1126 }
1127 
1128 static void
1129 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1130 {
1131 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1132 	struct netmap_slot *slot;
1133 
1134 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1135 	if (slot == NULL)
1136 		return;
1137 	if (txq->ift_sds.ifsd_map == NULL)
1138 		return;
1139 
1140 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1141 
1142 		/*
1143 		 * In netmap mode, set the map for the packet buffer.
1144 		 * NOTE: Some drivers (not this one) also need to set
1145 		 * the physical buffer address in the NIC ring.
1146 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1147 		 * netmap slot index, si
1148 		 */
1149 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1150 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1151 	}
1152 }
1153 static void
1154 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1155 {
1156 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1157 	struct netmap_slot *slot;
1158 	struct if_rxd_update iru;
1159 	iflib_fl_t fl;
1160 	bus_dmamap_t *map;
1161 	int nrxd;
1162 	uint32_t i, j, pidx_start;
1163 
1164 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1165 	if (slot == NULL)
1166 		return;
1167 	fl = &rxq->ifr_fl[0];
1168 	map = fl->ifl_sds.ifsd_map;
1169 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1170 	iru.iru_paddrs = fl->ifl_bus_addrs;
1171 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1172 	iru.iru_idxs = fl->ifl_rxd_idxs;
1173 	iru.iru_qsidx = rxq->ifr_id;
1174 	iru.iru_buf_size = rxq->ifr_fl[0].ifl_buf_size;
1175 	iru.iru_flidx = 0;
1176 
1177 	for (pidx_start = i = j = 0; i < nrxd; i++, j++) {
1178 		int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1179 		void *addr;
1180 
1181 		fl->ifl_rxd_idxs[j] = i;
1182 		addr = fl->ifl_vm_addrs[j] = PNMB(na, slot + sj, &fl->ifl_bus_addrs[j]);
1183 		if (map) {
1184 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1185 			map++;
1186 		}
1187 
1188 		if (j < IFLIB_MAX_RX_REFRESH && i < nrxd - 1)
1189 			continue;
1190 
1191 		iru.iru_pidx = pidx_start;
1192 		pidx_start = i;
1193 		iru.iru_count = j;
1194 		j = 0;
1195 		MPASS(pidx_start + j <= nrxd);
1196 		/* Update descriptors and the cached value */
1197 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1198 	}
1199 	/* preserve queue */
1200 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1201 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1202 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1203 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1204 	} else
1205 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1206 }
1207 
1208 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1209 
1210 #else
1211 #define iflib_netmap_txq_init(ctx, txq)
1212 #define iflib_netmap_rxq_init(ctx, rxq)
1213 #define iflib_netmap_detach(ifp)
1214 
1215 #define iflib_netmap_attach(ctx) (0)
1216 #define netmap_rx_irq(ifp, qid, budget) (0)
1217 #define netmap_tx_irq(ifp, qid) do {} while (0)
1218 
1219 #endif
1220 
1221 #if defined(__i386__) || defined(__amd64__)
1222 static __inline void
1223 prefetch(void *x)
1224 {
1225 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1226 }
1227 #else
1228 #define prefetch(x)
1229 #endif
1230 
1231 static void
1232 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1233 {
1234 	if (err)
1235 		return;
1236 	*(bus_addr_t *) arg = segs[0].ds_addr;
1237 }
1238 
1239 int
1240 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1241 {
1242 	int err;
1243 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1244 	device_t dev = ctx->ifc_dev;
1245 
1246 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1247 
1248 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1249 				sctx->isc_q_align, 0,	/* alignment, bounds */
1250 				BUS_SPACE_MAXADDR,	/* lowaddr */
1251 				BUS_SPACE_MAXADDR,	/* highaddr */
1252 				NULL, NULL,		/* filter, filterarg */
1253 				size,			/* maxsize */
1254 				1,			/* nsegments */
1255 				size,			/* maxsegsize */
1256 				BUS_DMA_ALLOCNOW,	/* flags */
1257 				NULL,			/* lockfunc */
1258 				NULL,			/* lockarg */
1259 				&dma->idi_tag);
1260 	if (err) {
1261 		device_printf(dev,
1262 		    "%s: bus_dma_tag_create failed: %d\n",
1263 		    __func__, err);
1264 		goto fail_0;
1265 	}
1266 
1267 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1268 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1269 	if (err) {
1270 		device_printf(dev,
1271 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1272 		    __func__, (uintmax_t)size, err);
1273 		goto fail_1;
1274 	}
1275 
1276 	dma->idi_paddr = IF_BAD_DMA;
1277 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1278 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1279 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1280 		device_printf(dev,
1281 		    "%s: bus_dmamap_load failed: %d\n",
1282 		    __func__, err);
1283 		goto fail_2;
1284 	}
1285 
1286 	dma->idi_size = size;
1287 	return (0);
1288 
1289 fail_2:
1290 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1291 fail_1:
1292 	bus_dma_tag_destroy(dma->idi_tag);
1293 fail_0:
1294 	dma->idi_tag = NULL;
1295 
1296 	return (err);
1297 }
1298 
1299 int
1300 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1301 {
1302 	int i, err;
1303 	iflib_dma_info_t *dmaiter;
1304 
1305 	dmaiter = dmalist;
1306 	for (i = 0; i < count; i++, dmaiter++) {
1307 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1308 			break;
1309 	}
1310 	if (err)
1311 		iflib_dma_free_multi(dmalist, i);
1312 	return (err);
1313 }
1314 
1315 void
1316 iflib_dma_free(iflib_dma_info_t dma)
1317 {
1318 	if (dma->idi_tag == NULL)
1319 		return;
1320 	if (dma->idi_paddr != IF_BAD_DMA) {
1321 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1322 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1323 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1324 		dma->idi_paddr = IF_BAD_DMA;
1325 	}
1326 	if (dma->idi_vaddr != NULL) {
1327 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1328 		dma->idi_vaddr = NULL;
1329 	}
1330 	bus_dma_tag_destroy(dma->idi_tag);
1331 	dma->idi_tag = NULL;
1332 }
1333 
1334 void
1335 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1336 {
1337 	int i;
1338 	iflib_dma_info_t *dmaiter = dmalist;
1339 
1340 	for (i = 0; i < count; i++, dmaiter++)
1341 		iflib_dma_free(*dmaiter);
1342 }
1343 
1344 #ifdef EARLY_AP_STARTUP
1345 static const int iflib_started = 1;
1346 #else
1347 /*
1348  * We used to abuse the smp_started flag to decide if the queues have been
1349  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1350  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1351  * is set.  Run a SYSINIT() strictly after that to just set a usable
1352  * completion flag.
1353  */
1354 
1355 static int iflib_started;
1356 
1357 static void
1358 iflib_record_started(void *arg)
1359 {
1360 	iflib_started = 1;
1361 }
1362 
1363 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1364 	iflib_record_started, NULL);
1365 #endif
1366 
1367 static int
1368 iflib_fast_intr(void *arg)
1369 {
1370 	iflib_filter_info_t info = arg;
1371 	struct grouptask *gtask = info->ifi_task;
1372 	if (!iflib_started)
1373 		return (FILTER_HANDLED);
1374 
1375 	DBG_COUNTER_INC(fast_intrs);
1376 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1377 		return (FILTER_HANDLED);
1378 
1379 	GROUPTASK_ENQUEUE(gtask);
1380 	return (FILTER_HANDLED);
1381 }
1382 
1383 static int
1384 iflib_fast_intr_rxtx(void *arg)
1385 {
1386 	iflib_filter_info_t info = arg;
1387 	struct grouptask *gtask = info->ifi_task;
1388 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1389 	if_ctx_t ctx;
1390 	int i, cidx;
1391 
1392 	if (!iflib_started)
1393 		return (FILTER_HANDLED);
1394 
1395 	DBG_COUNTER_INC(fast_intrs);
1396 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1397 		return (FILTER_HANDLED);
1398 
1399 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1400 		qidx_t txqid = rxq->ifr_txqid[i];
1401 
1402 		ctx = rxq->ifr_ctx;
1403 
1404 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1405 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1406 			continue;
1407 		}
1408 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1409 	}
1410 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1411 		cidx = rxq->ifr_cq_cidx;
1412 	else
1413 		cidx = rxq->ifr_fl[0].ifl_cidx;
1414 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1415 		GROUPTASK_ENQUEUE(gtask);
1416 	else
1417 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1418 	return (FILTER_HANDLED);
1419 }
1420 
1421 
1422 static int
1423 iflib_fast_intr_ctx(void *arg)
1424 {
1425 	iflib_filter_info_t info = arg;
1426 	struct grouptask *gtask = info->ifi_task;
1427 
1428 	if (!iflib_started)
1429 		return (FILTER_HANDLED);
1430 
1431 	DBG_COUNTER_INC(fast_intrs);
1432 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1433 		return (FILTER_HANDLED);
1434 
1435 	GROUPTASK_ENQUEUE(gtask);
1436 	return (FILTER_HANDLED);
1437 }
1438 
1439 static int
1440 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1441 	driver_filter_t filter, driver_intr_t handler, void *arg,
1442 				 char *name)
1443 {
1444 	int rc, flags;
1445 	struct resource *res;
1446 	void *tag = NULL;
1447 	device_t dev = ctx->ifc_dev;
1448 
1449 	flags = RF_ACTIVE;
1450 	if (ctx->ifc_flags & IFC_LEGACY)
1451 		flags |= RF_SHAREABLE;
1452 	MPASS(rid < 512);
1453 	irq->ii_rid = rid;
1454 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1455 	if (res == NULL) {
1456 		device_printf(dev,
1457 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1458 		return (ENOMEM);
1459 	}
1460 	irq->ii_res = res;
1461 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1462 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1463 						filter, handler, arg, &tag);
1464 	if (rc != 0) {
1465 		device_printf(dev,
1466 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1467 					  rid, name ? name : "unknown", rc);
1468 		return (rc);
1469 	} else if (name)
1470 		bus_describe_intr(dev, res, tag, "%s", name);
1471 
1472 	irq->ii_tag = tag;
1473 	return (0);
1474 }
1475 
1476 
1477 /*********************************************************************
1478  *
1479  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1480  *  the information needed to transmit a packet on the wire. This is
1481  *  called only once at attach, setup is done every reset.
1482  *
1483  **********************************************************************/
1484 
1485 static int
1486 iflib_txsd_alloc(iflib_txq_t txq)
1487 {
1488 	if_ctx_t ctx = txq->ift_ctx;
1489 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1490 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1491 	device_t dev = ctx->ifc_dev;
1492 	int err, nsegments, ntsosegments;
1493 
1494 	nsegments = scctx->isc_tx_nsegments;
1495 	ntsosegments = scctx->isc_tx_tso_segments_max;
1496 	MPASS(scctx->isc_ntxd[0] > 0);
1497 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1498 	MPASS(nsegments > 0);
1499 	MPASS(ntsosegments > 0);
1500 	/*
1501 	 * Setup DMA descriptor areas.
1502 	 */
1503 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1504 			       1, 0,			/* alignment, bounds */
1505 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1506 			       BUS_SPACE_MAXADDR,	/* highaddr */
1507 			       NULL, NULL,		/* filter, filterarg */
1508 			       sctx->isc_tx_maxsize,		/* maxsize */
1509 			       nsegments,	/* nsegments */
1510 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1511 			       0,			/* flags */
1512 			       NULL,			/* lockfunc */
1513 			       NULL,			/* lockfuncarg */
1514 			       &txq->ift_desc_tag))) {
1515 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1516 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1517 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1518 		goto fail;
1519 	}
1520 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1521 			       1, 0,			/* alignment, bounds */
1522 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1523 			       BUS_SPACE_MAXADDR,	/* highaddr */
1524 			       NULL, NULL,		/* filter, filterarg */
1525 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1526 			       ntsosegments,	/* nsegments */
1527 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1528 			       0,			/* flags */
1529 			       NULL,			/* lockfunc */
1530 			       NULL,			/* lockfuncarg */
1531 			       &txq->ift_tso_desc_tag))) {
1532 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1533 
1534 		goto fail;
1535 	}
1536 	if (!(txq->ift_sds.ifsd_flags =
1537 	    (uint8_t *) malloc(sizeof(uint8_t) *
1538 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1539 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1540 		err = ENOMEM;
1541 		goto fail;
1542 	}
1543 	if (!(txq->ift_sds.ifsd_m =
1544 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1545 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1546 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1547 		err = ENOMEM;
1548 		goto fail;
1549 	}
1550 
1551         /* Create the descriptor buffer dma maps */
1552 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1553 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1554 		return (0);
1555 
1556 	if (!(txq->ift_sds.ifsd_map =
1557 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1558 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1559 		err = ENOMEM;
1560 		goto fail;
1561 	}
1562 
1563 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1564 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1565 		if (err != 0) {
1566 			device_printf(dev, "Unable to create TX DMA map\n");
1567 			goto fail;
1568 		}
1569 	}
1570 #endif
1571 	return (0);
1572 fail:
1573 	/* We free all, it handles case where we are in the middle */
1574 	iflib_tx_structures_free(ctx);
1575 	return (err);
1576 }
1577 
1578 static void
1579 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1580 {
1581 	bus_dmamap_t map;
1582 
1583 	map = NULL;
1584 	if (txq->ift_sds.ifsd_map != NULL)
1585 		map = txq->ift_sds.ifsd_map[i];
1586 	if (map != NULL) {
1587 		bus_dmamap_unload(txq->ift_desc_tag, map);
1588 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1589 		txq->ift_sds.ifsd_map[i] = NULL;
1590 	}
1591 }
1592 
1593 static void
1594 iflib_txq_destroy(iflib_txq_t txq)
1595 {
1596 	if_ctx_t ctx = txq->ift_ctx;
1597 
1598 	for (int i = 0; i < txq->ift_size; i++)
1599 		iflib_txsd_destroy(ctx, txq, i);
1600 	if (txq->ift_sds.ifsd_map != NULL) {
1601 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1602 		txq->ift_sds.ifsd_map = NULL;
1603 	}
1604 	if (txq->ift_sds.ifsd_m != NULL) {
1605 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1606 		txq->ift_sds.ifsd_m = NULL;
1607 	}
1608 	if (txq->ift_sds.ifsd_flags != NULL) {
1609 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1610 		txq->ift_sds.ifsd_flags = NULL;
1611 	}
1612 	if (txq->ift_desc_tag != NULL) {
1613 		bus_dma_tag_destroy(txq->ift_desc_tag);
1614 		txq->ift_desc_tag = NULL;
1615 	}
1616 	if (txq->ift_tso_desc_tag != NULL) {
1617 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1618 		txq->ift_tso_desc_tag = NULL;
1619 	}
1620 }
1621 
1622 static void
1623 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1624 {
1625 	struct mbuf **mp;
1626 
1627 	mp = &txq->ift_sds.ifsd_m[i];
1628 	if (*mp == NULL)
1629 		return;
1630 
1631 	if (txq->ift_sds.ifsd_map != NULL) {
1632 		bus_dmamap_sync(txq->ift_desc_tag,
1633 				txq->ift_sds.ifsd_map[i],
1634 				BUS_DMASYNC_POSTWRITE);
1635 		bus_dmamap_unload(txq->ift_desc_tag,
1636 				  txq->ift_sds.ifsd_map[i]);
1637 	}
1638 	m_free(*mp);
1639 	DBG_COUNTER_INC(tx_frees);
1640 	*mp = NULL;
1641 }
1642 
1643 static int
1644 iflib_txq_setup(iflib_txq_t txq)
1645 {
1646 	if_ctx_t ctx = txq->ift_ctx;
1647 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1648 	iflib_dma_info_t di;
1649 	int i;
1650 
1651 	/* Set number of descriptors available */
1652 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1653 	/* XXX make configurable */
1654 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1655 
1656 	/* Reset indices */
1657 	txq->ift_cidx_processed = 0;
1658 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1659 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1660 
1661 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1662 		bzero((void *)di->idi_vaddr, di->idi_size);
1663 
1664 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1665 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1666 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1667 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1668 	return (0);
1669 }
1670 
1671 /*********************************************************************
1672  *
1673  *  Allocate memory for rx_buffer structures. Since we use one
1674  *  rx_buffer per received packet, the maximum number of rx_buffer's
1675  *  that we'll need is equal to the number of receive descriptors
1676  *  that we've allocated.
1677  *
1678  **********************************************************************/
1679 static int
1680 iflib_rxsd_alloc(iflib_rxq_t rxq)
1681 {
1682 	if_ctx_t ctx = rxq->ifr_ctx;
1683 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1684 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1685 	device_t dev = ctx->ifc_dev;
1686 	iflib_fl_t fl;
1687 	int			err;
1688 
1689 	MPASS(scctx->isc_nrxd[0] > 0);
1690 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1691 
1692 	fl = rxq->ifr_fl;
1693 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1694 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1695 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1696 					 1, 0,			/* alignment, bounds */
1697 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1698 					 BUS_SPACE_MAXADDR,	/* highaddr */
1699 					 NULL, NULL,		/* filter, filterarg */
1700 					 sctx->isc_rx_maxsize,	/* maxsize */
1701 					 sctx->isc_rx_nsegments,	/* nsegments */
1702 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1703 					 0,			/* flags */
1704 					 NULL,			/* lockfunc */
1705 					 NULL,			/* lockarg */
1706 					 &fl->ifl_desc_tag);
1707 		if (err) {
1708 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1709 				__func__, err);
1710 			goto fail;
1711 		}
1712 		if (!(fl->ifl_sds.ifsd_flags =
1713 		      (uint8_t *) malloc(sizeof(uint8_t) *
1714 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1715 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1716 			err = ENOMEM;
1717 			goto fail;
1718 		}
1719 		if (!(fl->ifl_sds.ifsd_m =
1720 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1721 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1722 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1723 			err = ENOMEM;
1724 			goto fail;
1725 		}
1726 		if (!(fl->ifl_sds.ifsd_cl =
1727 		      (caddr_t *) malloc(sizeof(caddr_t) *
1728 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1729 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1730 			err = ENOMEM;
1731 			goto fail;
1732 		}
1733 
1734 		/* Create the descriptor buffer dma maps */
1735 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1736 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1737 			continue;
1738 
1739 		if (!(fl->ifl_sds.ifsd_map =
1740 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1741 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1742 			err = ENOMEM;
1743 			goto fail;
1744 		}
1745 
1746 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1747 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1748 			if (err != 0) {
1749 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1750 				goto fail;
1751 			}
1752 		}
1753 #endif
1754 	}
1755 	return (0);
1756 
1757 fail:
1758 	iflib_rx_structures_free(ctx);
1759 	return (err);
1760 }
1761 
1762 
1763 /*
1764  * Internal service routines
1765  */
1766 
1767 struct rxq_refill_cb_arg {
1768 	int               error;
1769 	bus_dma_segment_t seg;
1770 	int               nseg;
1771 };
1772 
1773 static void
1774 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1775 {
1776 	struct rxq_refill_cb_arg *cb_arg = arg;
1777 
1778 	cb_arg->error = error;
1779 	cb_arg->seg = segs[0];
1780 	cb_arg->nseg = nseg;
1781 }
1782 
1783 
1784 #ifdef ACPI_DMAR
1785 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1786 #else
1787 #define IS_DMAR(ctx) (0)
1788 #endif
1789 
1790 /**
1791  *	rxq_refill - refill an rxq  free-buffer list
1792  *	@ctx: the iflib context
1793  *	@rxq: the free-list to refill
1794  *	@n: the number of new buffers to allocate
1795  *
1796  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1797  *	The caller must assure that @n does not exceed the queue's capacity.
1798  */
1799 static void
1800 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1801 {
1802 	struct mbuf *m;
1803 	int idx, frag_idx = fl->ifl_fragidx, pidx = fl->ifl_pidx;
1804 	caddr_t cl, *sd_cl;
1805 	struct mbuf **sd_m;
1806 	uint8_t *sd_flags;
1807 	struct if_rxd_update iru;
1808 	bus_dmamap_t *sd_map;
1809 	int n, i = 0;
1810 	uint64_t bus_addr;
1811 	int err;
1812 
1813 	sd_m = fl->ifl_sds.ifsd_m;
1814 	sd_map = fl->ifl_sds.ifsd_map;
1815 	sd_cl = fl->ifl_sds.ifsd_cl;
1816 	sd_flags = fl->ifl_sds.ifsd_flags;
1817 	idx = pidx;
1818 
1819 	n  = count;
1820 	MPASS(n > 0);
1821 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1822 
1823 	if (pidx < fl->ifl_cidx)
1824 		MPASS(pidx + n <= fl->ifl_cidx);
1825 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1826 		MPASS(fl->ifl_gen == 0);
1827 	if (pidx > fl->ifl_cidx)
1828 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1829 
1830 	DBG_COUNTER_INC(fl_refills);
1831 	if (n > 8)
1832 		DBG_COUNTER_INC(fl_refills_large);
1833 	iru.iru_paddrs = fl->ifl_bus_addrs;
1834 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1835 	iru.iru_idxs = fl->ifl_rxd_idxs;
1836 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1837 	iru.iru_buf_size = fl->ifl_buf_size;
1838 	iru.iru_flidx = fl->ifl_id;
1839 	while (n--) {
1840 		/*
1841 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1842 		 * initialized after rx.
1843 		 *
1844 		 * If the cluster is still set then we know a minimum sized packet was received
1845 		 */
1846 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1847 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1848                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1849 		if ((cl = sd_cl[frag_idx]) == NULL) {
1850                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1851 				break;
1852 #if MEMORY_LOGGING
1853 			fl->ifl_cl_enqueued++;
1854 #endif
1855 		}
1856 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1857 			break;
1858 		}
1859 #if MEMORY_LOGGING
1860 		fl->ifl_m_enqueued++;
1861 #endif
1862 
1863 		DBG_COUNTER_INC(rx_allocs);
1864 #if defined(__i386__) || defined(__amd64__)
1865 		if (!IS_DMAR(ctx)) {
1866 			bus_addr = pmap_kextract((vm_offset_t)cl);
1867 		} else
1868 #endif
1869 		{
1870 			struct rxq_refill_cb_arg cb_arg;
1871 			iflib_rxq_t q;
1872 
1873 			cb_arg.error = 0;
1874 			q = fl->ifl_rxq;
1875 			MPASS(sd_map != NULL);
1876 			MPASS(sd_map[frag_idx] != NULL);
1877 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1878 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1879 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1880 					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                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1895 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1896 
1897 		MPASS(sd_m[frag_idx] == NULL);
1898 		sd_cl[frag_idx] = cl;
1899 		sd_m[frag_idx] = m;
1900 		fl->ifl_rxd_idxs[i] = frag_idx;
1901 		fl->ifl_bus_addrs[i] = bus_addr;
1902 		fl->ifl_vm_addrs[i] = cl;
1903 		fl->ifl_credits++;
1904 		i++;
1905 		MPASS(fl->ifl_credits <= fl->ifl_size);
1906 		if (++idx == fl->ifl_size) {
1907 			fl->ifl_gen = 1;
1908 			idx = 0;
1909 		}
1910 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1911 			iru.iru_pidx = pidx;
1912 			iru.iru_count = i;
1913 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1914 			i = 0;
1915 			pidx = idx;
1916 			fl->ifl_pidx = idx;
1917 		}
1918 
1919 	}
1920 done:
1921 	DBG_COUNTER_INC(rxd_flush);
1922 	if (fl->ifl_pidx == 0)
1923 		pidx = fl->ifl_size - 1;
1924 	else
1925 		pidx = fl->ifl_pidx - 1;
1926 
1927 	if (sd_map)
1928 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1929 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1930 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1931 	fl->ifl_fragidx = frag_idx;
1932 }
1933 
1934 static __inline void
1935 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1936 {
1937 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1938 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1939 #ifdef INVARIANTS
1940 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1941 #endif
1942 
1943 	MPASS(fl->ifl_credits <= fl->ifl_size);
1944 	MPASS(reclaimable == delta);
1945 
1946 	if (reclaimable > 0)
1947 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1948 }
1949 
1950 static void
1951 iflib_fl_bufs_free(iflib_fl_t fl)
1952 {
1953 	iflib_dma_info_t idi = fl->ifl_ifdi;
1954 	uint32_t i;
1955 
1956 	for (i = 0; i < fl->ifl_size; i++) {
1957 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1958 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1959 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1960 
1961 		if (*sd_flags & RX_SW_DESC_INUSE) {
1962 			if (fl->ifl_sds.ifsd_map != NULL) {
1963 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1964 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1965 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1966 			}
1967 			if (*sd_m != NULL) {
1968 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1969 				uma_zfree(zone_mbuf, *sd_m);
1970 			}
1971 			if (*sd_cl != NULL)
1972 				uma_zfree(fl->ifl_zone, *sd_cl);
1973 			*sd_flags = 0;
1974 		} else {
1975 			MPASS(*sd_cl == NULL);
1976 			MPASS(*sd_m == NULL);
1977 		}
1978 #if MEMORY_LOGGING
1979 		fl->ifl_m_dequeued++;
1980 		fl->ifl_cl_dequeued++;
1981 #endif
1982 		*sd_cl = NULL;
1983 		*sd_m = NULL;
1984 	}
1985 #ifdef INVARIANTS
1986 	for (i = 0; i < fl->ifl_size; i++) {
1987 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1988 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1989 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1990 	}
1991 #endif
1992 	/*
1993 	 * Reset free list values
1994 	 */
1995 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1996 	bzero(idi->idi_vaddr, idi->idi_size);
1997 }
1998 
1999 /*********************************************************************
2000  *
2001  *  Initialize a receive ring and its buffers.
2002  *
2003  **********************************************************************/
2004 static int
2005 iflib_fl_setup(iflib_fl_t fl)
2006 {
2007 	iflib_rxq_t rxq = fl->ifl_rxq;
2008 	if_ctx_t ctx = rxq->ifr_ctx;
2009 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2010 
2011        fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
2012 	/*
2013 	** Free current RX buffer structs and their mbufs
2014 	*/
2015 	iflib_fl_bufs_free(fl);
2016 	/* Now replenish the mbufs */
2017 	MPASS(fl->ifl_credits == 0);
2018 	/*
2019 	 * XXX don't set the max_frame_size to larger
2020 	 * than the hardware can handle
2021 	 */
2022 	if (sctx->isc_max_frame_size <= 2048)
2023 		fl->ifl_buf_size = MCLBYTES;
2024 #ifndef CONTIGMALLOC_WORKS
2025 	else
2026 		fl->ifl_buf_size = MJUMPAGESIZE;
2027 #else
2028 	else if (sctx->isc_max_frame_size <= 4096)
2029 		fl->ifl_buf_size = MJUMPAGESIZE;
2030 	else if (sctx->isc_max_frame_size <= 9216)
2031 		fl->ifl_buf_size = MJUM9BYTES;
2032 	else
2033 		fl->ifl_buf_size = MJUM16BYTES;
2034 #endif
2035 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2036 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2037 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2038 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2039 
2040 
2041 	/* avoid pre-allocating zillions of clusters to an idle card
2042 	 * potentially speeding up attach
2043 	 */
2044 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2045 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2046 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2047 		return (ENOBUFS);
2048 	/*
2049 	 * handle failure
2050 	 */
2051 	MPASS(rxq != NULL);
2052 	MPASS(fl->ifl_ifdi != NULL);
2053 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2054 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2055 	return (0);
2056 }
2057 
2058 /*********************************************************************
2059  *
2060  *  Free receive ring data structures
2061  *
2062  **********************************************************************/
2063 static void
2064 iflib_rx_sds_free(iflib_rxq_t rxq)
2065 {
2066 	iflib_fl_t fl;
2067 	int i;
2068 
2069 	if (rxq->ifr_fl != NULL) {
2070 		for (i = 0; i < rxq->ifr_nfl; i++) {
2071 			fl = &rxq->ifr_fl[i];
2072 			if (fl->ifl_desc_tag != NULL) {
2073 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2074 				fl->ifl_desc_tag = NULL;
2075 			}
2076 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2077 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2078 			/* XXX destroy maps first */
2079 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2080 			fl->ifl_sds.ifsd_m = NULL;
2081 			fl->ifl_sds.ifsd_cl = NULL;
2082 			fl->ifl_sds.ifsd_map = NULL;
2083 		}
2084 		free(rxq->ifr_fl, M_IFLIB);
2085 		rxq->ifr_fl = NULL;
2086 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2087 	}
2088 }
2089 
2090 /*
2091  * MI independent logic
2092  *
2093  */
2094 static void
2095 iflib_timer(void *arg)
2096 {
2097 	iflib_txq_t txq = arg;
2098 	if_ctx_t ctx = txq->ift_ctx;
2099 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2100 
2101 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2102 		return;
2103 	/*
2104 	** Check on the state of the TX queue(s), this
2105 	** can be done without the lock because its RO
2106 	** and the HUNG state will be static if set.
2107 	*/
2108 	IFDI_TIMER(ctx, txq->ift_id);
2109 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2110 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2111 	     (sctx->isc_pause_frames == 0)))
2112 		goto hung;
2113 
2114 	if (ifmp_ring_is_stalled(txq->ift_br))
2115 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2116 	txq->ift_cleaned_prev = txq->ift_cleaned;
2117 	/* handle any laggards */
2118 	if (txq->ift_db_pending)
2119 		GROUPTASK_ENQUEUE(&txq->ift_task);
2120 
2121 	sctx->isc_pause_frames = 0;
2122 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2123 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2124 	return;
2125 hung:
2126 	CTX_LOCK(ctx);
2127 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2128 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2129 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2130 
2131 	IFDI_WATCHDOG_RESET(ctx);
2132 	ctx->ifc_watchdog_events++;
2133 
2134 	ctx->ifc_flags |= IFC_DO_RESET;
2135 	iflib_admin_intr_deferred(ctx);
2136 	CTX_UNLOCK(ctx);
2137 }
2138 
2139 static void
2140 iflib_init_locked(if_ctx_t ctx)
2141 {
2142 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2143 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2144 	if_t ifp = ctx->ifc_ifp;
2145 	iflib_fl_t fl;
2146 	iflib_txq_t txq;
2147 	iflib_rxq_t rxq;
2148 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2149 
2150 
2151 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2152 	IFDI_INTR_DISABLE(ctx);
2153 
2154 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2155 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2156 	/* Set hardware offload abilities */
2157 	if_clearhwassist(ifp);
2158 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2159 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2160 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2161 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2162 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2163 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2164 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2165 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2166 
2167 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2168 		CALLOUT_LOCK(txq);
2169 		callout_stop(&txq->ift_timer);
2170 		CALLOUT_UNLOCK(txq);
2171 		iflib_netmap_txq_init(ctx, txq);
2172 	}
2173 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2174 		MPASS(rxq->ifr_id == i);
2175 		iflib_netmap_rxq_init(ctx, rxq);
2176 	}
2177 #ifdef INVARIANTS
2178 	i = if_getdrvflags(ifp);
2179 #endif
2180 	IFDI_INIT(ctx);
2181 	MPASS(if_getdrvflags(ifp) == i);
2182 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2183 		/* XXX this should really be done on a per-queue basis */
2184 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2185 			continue;
2186 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2187 			if (iflib_fl_setup(fl)) {
2188 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2189 				goto done;
2190 			}
2191 		}
2192 	}
2193 	done:
2194 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2195 	IFDI_INTR_ENABLE(ctx);
2196 	txq = ctx->ifc_txqs;
2197 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2198 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2199 			txq->ift_timer.c_cpu);
2200 }
2201 
2202 static int
2203 iflib_media_change(if_t ifp)
2204 {
2205 	if_ctx_t ctx = if_getsoftc(ifp);
2206 	int err;
2207 
2208 	CTX_LOCK(ctx);
2209 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2210 		iflib_init_locked(ctx);
2211 	CTX_UNLOCK(ctx);
2212 	return (err);
2213 }
2214 
2215 static void
2216 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2217 {
2218 	if_ctx_t ctx = if_getsoftc(ifp);
2219 
2220 	CTX_LOCK(ctx);
2221 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2222 	IFDI_MEDIA_STATUS(ctx, ifmr);
2223 	CTX_UNLOCK(ctx);
2224 }
2225 
2226 static void
2227 iflib_stop(if_ctx_t ctx)
2228 {
2229 	iflib_txq_t txq = ctx->ifc_txqs;
2230 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2231 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2232 	iflib_dma_info_t di;
2233 	iflib_fl_t fl;
2234 	int i, j;
2235 
2236 	/* Tell the stack that the interface is no longer active */
2237 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2238 
2239 	IFDI_INTR_DISABLE(ctx);
2240 	DELAY(1000);
2241 	IFDI_STOP(ctx);
2242 	DELAY(1000);
2243 
2244 	iflib_debug_reset();
2245 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2246 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2247 		/* make sure all transmitters have completed before proceeding XXX */
2248 
2249 		/* clean any enqueued buffers */
2250 		iflib_ifmp_purge(txq);
2251 		/* Free any existing tx buffers. */
2252 		for (j = 0; j < txq->ift_size; j++) {
2253 			iflib_txsd_free(ctx, txq, j);
2254 		}
2255 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2256 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2257 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2258 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2259 		txq->ift_pullups = 0;
2260 		ifmp_ring_reset_stats(txq->ift_br);
2261 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2262 			bzero((void *)di->idi_vaddr, di->idi_size);
2263 	}
2264 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2265 		/* make sure all transmitters have completed before proceeding XXX */
2266 
2267 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2268 			bzero((void *)di->idi_vaddr, di->idi_size);
2269 		/* also resets the free lists pidx/cidx */
2270 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2271 			iflib_fl_bufs_free(fl);
2272 	}
2273 }
2274 
2275 static inline caddr_t
2276 calc_next_rxd(iflib_fl_t fl, int cidx)
2277 {
2278 	qidx_t size;
2279 	int nrxd;
2280 	caddr_t start, end, cur, next;
2281 
2282 	nrxd = fl->ifl_size;
2283 	size = fl->ifl_rxd_size;
2284 	start = fl->ifl_ifdi->idi_vaddr;
2285 
2286 	if (__predict_false(size == 0))
2287 		return (start);
2288 	cur = start + size*cidx;
2289 	end = start + size*nrxd;
2290 	next = CACHE_PTR_NEXT(cur);
2291 	return (next < end ? next : start);
2292 }
2293 
2294 static inline void
2295 prefetch_pkts(iflib_fl_t fl, int cidx)
2296 {
2297 	int nextptr;
2298 	int nrxd = fl->ifl_size;
2299 	caddr_t next_rxd;
2300 
2301 
2302 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2303 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2304 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2305 	next_rxd = calc_next_rxd(fl, cidx);
2306 	prefetch(next_rxd);
2307 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2308 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2309 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2310 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2311 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2312 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2313 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2314 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2315 }
2316 
2317 static void
2318 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2319 {
2320 	int flid, cidx;
2321 	bus_dmamap_t map;
2322 	iflib_fl_t fl;
2323 	iflib_dma_info_t di;
2324 	int next;
2325 
2326 	map = NULL;
2327 	flid = irf->irf_flid;
2328 	cidx = irf->irf_idx;
2329 	fl = &rxq->ifr_fl[flid];
2330 	sd->ifsd_fl = fl;
2331 	sd->ifsd_cidx = cidx;
2332 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2333 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2334 	fl->ifl_credits--;
2335 #if MEMORY_LOGGING
2336 	fl->ifl_m_dequeued++;
2337 #endif
2338 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2339 		prefetch_pkts(fl, cidx);
2340 	if (fl->ifl_sds.ifsd_map != NULL) {
2341 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2342 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2343 		map = fl->ifl_sds.ifsd_map[cidx];
2344 		di = fl->ifl_ifdi;
2345 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2346 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2347 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2348 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2349 
2350 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2351 		MPASS(fl->ifl_cidx == cidx);
2352 		if (unload)
2353 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2354 	}
2355 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2356 	if (__predict_false(fl->ifl_cidx == 0))
2357 		fl->ifl_gen = 0;
2358 	if (map != NULL)
2359 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2360 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2361         bit_clear(fl->ifl_rx_bitmap, cidx);
2362 }
2363 
2364 static struct mbuf *
2365 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2366 {
2367 	int i, padlen , flags;
2368 	struct mbuf *m, *mh, *mt;
2369 	caddr_t cl;
2370 
2371 	i = 0;
2372 	mh = NULL;
2373 	do {
2374 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2375 
2376 		MPASS(*sd->ifsd_cl != NULL);
2377 		MPASS(*sd->ifsd_m != NULL);
2378 
2379 		/* Don't include zero-length frags */
2380 		if (ri->iri_frags[i].irf_len == 0) {
2381 			/* XXX we can save the cluster here, but not the mbuf */
2382 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2383 			m_free(*sd->ifsd_m);
2384 			*sd->ifsd_m = NULL;
2385 			continue;
2386 		}
2387 		m = *sd->ifsd_m;
2388 		*sd->ifsd_m = NULL;
2389 		if (mh == NULL) {
2390 			flags = M_PKTHDR|M_EXT;
2391 			mh = mt = m;
2392 			padlen = ri->iri_pad;
2393 		} else {
2394 			flags = M_EXT;
2395 			mt->m_next = m;
2396 			mt = m;
2397 			/* assuming padding is only on the first fragment */
2398 			padlen = 0;
2399 		}
2400 		cl = *sd->ifsd_cl;
2401 		*sd->ifsd_cl = NULL;
2402 
2403 		/* Can these two be made one ? */
2404 		m_init(m, M_NOWAIT, MT_DATA, flags);
2405 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2406 		/*
2407 		 * These must follow m_init and m_cljset
2408 		 */
2409 		m->m_data += padlen;
2410 		ri->iri_len -= padlen;
2411 		m->m_len = ri->iri_frags[i].irf_len;
2412 	} while (++i < ri->iri_nfrags);
2413 
2414 	return (mh);
2415 }
2416 
2417 /*
2418  * Process one software descriptor
2419  */
2420 static struct mbuf *
2421 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2422 {
2423 	struct if_rxsd sd;
2424 	struct mbuf *m;
2425 
2426 	/* should I merge this back in now that the two paths are basically duplicated? */
2427 	if (ri->iri_nfrags == 1 &&
2428 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2429 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2430 		m = *sd.ifsd_m;
2431 		*sd.ifsd_m = NULL;
2432 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2433 #ifndef __NO_STRICT_ALIGNMENT
2434 		if (!IP_ALIGNED(m))
2435 			m->m_data += 2;
2436 #endif
2437 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2438 		m->m_len = ri->iri_frags[0].irf_len;
2439        } else {
2440 		m = assemble_segments(rxq, ri, &sd);
2441 	}
2442 	m->m_pkthdr.len = ri->iri_len;
2443 	m->m_pkthdr.rcvif = ri->iri_ifp;
2444 	m->m_flags |= ri->iri_flags;
2445 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2446 	m->m_pkthdr.flowid = ri->iri_flowid;
2447 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2448 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2449 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2450 	return (m);
2451 }
2452 
2453 static bool
2454 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2455 {
2456 	if_ctx_t ctx = rxq->ifr_ctx;
2457 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2458 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2459 	int avail, i;
2460 	qidx_t *cidxp;
2461 	struct if_rxd_info ri;
2462 	int err, budget_left, rx_bytes, rx_pkts;
2463 	iflib_fl_t fl;
2464 	struct ifnet *ifp;
2465 	int lro_enabled;
2466 
2467 	/*
2468 	 * XXX early demux data packets so that if_input processing only handles
2469 	 * acks in interrupt context
2470 	 */
2471 	struct mbuf *m, *mh, *mt;
2472 
2473 	ifp = ctx->ifc_ifp;
2474 #ifdef DEV_NETMAP
2475 	if (ifp->if_capenable & IFCAP_NETMAP) {
2476 		u_int work = 0;
2477 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2478 			return (FALSE);
2479 	}
2480 #endif
2481 
2482 	mh = mt = NULL;
2483 	MPASS(budget > 0);
2484 	rx_pkts	= rx_bytes = 0;
2485 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2486 		cidxp = &rxq->ifr_cq_cidx;
2487 	else
2488 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2489 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2490 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2491 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2492 		DBG_COUNTER_INC(rx_unavail);
2493 		return (false);
2494 	}
2495 
2496 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2497 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2498 			DBG_COUNTER_INC(rx_ctx_inactive);
2499 			break;
2500 		}
2501 		/*
2502 		 * Reset client set fields to their default values
2503 		 */
2504 		rxd_info_zero(&ri);
2505 		ri.iri_qsidx = rxq->ifr_id;
2506 		ri.iri_cidx = *cidxp;
2507 		ri.iri_ifp = ifp;
2508 		ri.iri_frags = rxq->ifr_frags;
2509 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2510 
2511 		if (err)
2512 			goto err;
2513 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2514 			*cidxp = ri.iri_cidx;
2515 			/* Update our consumer index */
2516 			/* XXX NB: shurd - check if this is still safe */
2517 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2518 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2519 				rxq->ifr_cq_gen = 0;
2520 			}
2521 			/* was this only a completion queue message? */
2522 			if (__predict_false(ri.iri_nfrags == 0))
2523 				continue;
2524 		}
2525 		MPASS(ri.iri_nfrags != 0);
2526 		MPASS(ri.iri_len != 0);
2527 
2528 		/* will advance the cidx on the corresponding free lists */
2529 		m = iflib_rxd_pkt_get(rxq, &ri);
2530 		if (avail == 0 && budget_left)
2531 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2532 
2533 		if (__predict_false(m == NULL)) {
2534 			DBG_COUNTER_INC(rx_mbuf_null);
2535 			continue;
2536 		}
2537 		/* imm_pkt: -- cxgb */
2538 		if (mh == NULL)
2539 			mh = mt = m;
2540 		else {
2541 			mt->m_nextpkt = m;
2542 			mt = m;
2543 		}
2544 	}
2545 	/* make sure that we can refill faster than drain */
2546 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2547 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2548 
2549 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2550 	while (mh != NULL) {
2551 		m = mh;
2552 		mh = mh->m_nextpkt;
2553 		m->m_nextpkt = NULL;
2554 #ifndef __NO_STRICT_ALIGNMENT
2555 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2556 			continue;
2557 #endif
2558 		rx_bytes += m->m_pkthdr.len;
2559 		rx_pkts++;
2560 #if defined(INET6) || defined(INET)
2561 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2562 			continue;
2563 #endif
2564 		DBG_COUNTER_INC(rx_if_input);
2565 		ifp->if_input(ifp, m);
2566 	}
2567 
2568 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2569 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2570 
2571 	/*
2572 	 * Flush any outstanding LRO work
2573 	 */
2574 #if defined(INET6) || defined(INET)
2575 	tcp_lro_flush_all(&rxq->ifr_lc);
2576 #endif
2577 	if (avail)
2578 		return true;
2579 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2580 err:
2581 	CTX_LOCK(ctx);
2582 	ctx->ifc_flags |= IFC_DO_RESET;
2583 	iflib_admin_intr_deferred(ctx);
2584 	CTX_UNLOCK(ctx);
2585 	return (false);
2586 }
2587 
2588 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2589 static inline qidx_t
2590 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2591 {
2592 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2593 	qidx_t minthresh = txq->ift_size / 8;
2594 	if (in_use > 4*minthresh)
2595 		return (notify_count);
2596 	if (in_use > 2*minthresh)
2597 		return (notify_count >> 1);
2598 	if (in_use > minthresh)
2599 		return (notify_count >> 3);
2600 	return (0);
2601 }
2602 
2603 static inline qidx_t
2604 txq_max_rs_deferred(iflib_txq_t txq)
2605 {
2606 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2607 	qidx_t minthresh = txq->ift_size / 8;
2608 	if (txq->ift_in_use > 4*minthresh)
2609 		return (notify_count);
2610 	if (txq->ift_in_use > 2*minthresh)
2611 		return (notify_count >> 1);
2612 	if (txq->ift_in_use > minthresh)
2613 		return (notify_count >> 2);
2614 	return (2);
2615 }
2616 
2617 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2618 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2619 
2620 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2621 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2622 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2623 
2624 /* forward compatibility for cxgb */
2625 #define FIRST_QSET(ctx) 0
2626 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2627 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2628 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2629 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2630 
2631 /* XXX we should be setting this to something other than zero */
2632 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2633 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2634 
2635 static inline bool
2636 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2637 {
2638 	qidx_t dbval, max;
2639 	bool rang;
2640 
2641 	rang = false;
2642 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2643 	if (ring || txq->ift_db_pending >= max) {
2644 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2645 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2646 		txq->ift_db_pending = txq->ift_npending = 0;
2647 		rang = true;
2648 	}
2649 	return (rang);
2650 }
2651 
2652 #ifdef PKT_DEBUG
2653 static void
2654 print_pkt(if_pkt_info_t pi)
2655 {
2656 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2657 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2658 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2659 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2660 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2661 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2662 }
2663 #endif
2664 
2665 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2666 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2667 
2668 static int
2669 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2670 {
2671 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2672 	struct ether_vlan_header *eh;
2673 	struct mbuf *m, *n;
2674 
2675 	n = m = *mp;
2676 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2677 	    M_WRITABLE(m) == 0) {
2678 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2679 			return (ENOMEM);
2680 		} else {
2681 			m_freem(*mp);
2682 			n = *mp = m;
2683 		}
2684 	}
2685 
2686 	/*
2687 	 * Determine where frame payload starts.
2688 	 * Jump over vlan headers if already present,
2689 	 * helpful for QinQ too.
2690 	 */
2691 	if (__predict_false(m->m_len < sizeof(*eh))) {
2692 		txq->ift_pullups++;
2693 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2694 			return (ENOMEM);
2695 	}
2696 	eh = mtod(m, struct ether_vlan_header *);
2697 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2698 		pi->ipi_etype = ntohs(eh->evl_proto);
2699 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2700 	} else {
2701 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2702 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2703 	}
2704 
2705 	switch (pi->ipi_etype) {
2706 #ifdef INET
2707 	case ETHERTYPE_IP:
2708 	{
2709 		struct ip *ip = NULL;
2710 		struct tcphdr *th = NULL;
2711 		int minthlen;
2712 
2713 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2714 		if (__predict_false(m->m_len < minthlen)) {
2715 			/*
2716 			 * if this code bloat is causing too much of a hit
2717 			 * move it to a separate function and mark it noinline
2718 			 */
2719 			if (m->m_len == pi->ipi_ehdrlen) {
2720 				n = m->m_next;
2721 				MPASS(n);
2722 				if (n->m_len >= sizeof(*ip))  {
2723 					ip = (struct ip *)n->m_data;
2724 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2725 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2726 				} else {
2727 					txq->ift_pullups++;
2728 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2729 						return (ENOMEM);
2730 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2731 				}
2732 			} else {
2733 				txq->ift_pullups++;
2734 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2735 					return (ENOMEM);
2736 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2737 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2738 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2739 			}
2740 		} else {
2741 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2742 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2743 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2744 		}
2745 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2746 		pi->ipi_ipproto = ip->ip_p;
2747 		pi->ipi_flags |= IPI_TX_IPV4;
2748 
2749 		if (pi->ipi_csum_flags & CSUM_IP)
2750                        ip->ip_sum = 0;
2751 
2752 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2753 			if (__predict_false(th == NULL)) {
2754 				txq->ift_pullups++;
2755 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2756 					return (ENOMEM);
2757 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2758 			}
2759 			pi->ipi_tcp_hflags = th->th_flags;
2760 			pi->ipi_tcp_hlen = th->th_off << 2;
2761 			pi->ipi_tcp_seq = th->th_seq;
2762 		}
2763 		if (IS_TSO4(pi)) {
2764 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2765 				return (ENXIO);
2766 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2767 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2768 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2769 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2770 				ip->ip_sum = 0;
2771 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2772 			}
2773 		}
2774 		break;
2775 	}
2776 #endif
2777 #ifdef INET6
2778 	case ETHERTYPE_IPV6:
2779 	{
2780 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2781 		struct tcphdr *th;
2782 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2783 
2784 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2785 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2786 				return (ENOMEM);
2787 		}
2788 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2789 
2790 		/* XXX-BZ this will go badly in case of ext hdrs. */
2791 		pi->ipi_ipproto = ip6->ip6_nxt;
2792 		pi->ipi_flags |= IPI_TX_IPV6;
2793 
2794 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2795 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2796 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2797 					return (ENOMEM);
2798 			}
2799 			pi->ipi_tcp_hflags = th->th_flags;
2800 			pi->ipi_tcp_hlen = th->th_off << 2;
2801 		}
2802 		if (IS_TSO6(pi)) {
2803 
2804 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2805 				return (ENXIO);
2806 			/*
2807 			 * The corresponding flag is set by the stack in the IPv4
2808 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2809 			 * So, set it here because the rest of the flow requires it.
2810 			 */
2811 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2812 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2813 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2814 		}
2815 		break;
2816 	}
2817 #endif
2818 	default:
2819 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2820 		pi->ipi_ip_hlen = 0;
2821 		break;
2822 	}
2823 	*mp = m;
2824 
2825 	return (0);
2826 }
2827 
2828 static  __noinline  struct mbuf *
2829 collapse_pkthdr(struct mbuf *m0)
2830 {
2831 	struct mbuf *m, *m_next, *tmp;
2832 
2833 	m = m0;
2834 	m_next = m->m_next;
2835 	while (m_next != NULL && m_next->m_len == 0) {
2836 		m = m_next;
2837 		m->m_next = NULL;
2838 		m_free(m);
2839 		m_next = m_next->m_next;
2840 	}
2841 	m = m0;
2842 	m->m_next = m_next;
2843 	if ((m_next->m_flags & M_EXT) == 0) {
2844 		m = m_defrag(m, M_NOWAIT);
2845 	} else {
2846 		tmp = m_next->m_next;
2847 		memcpy(m_next, m, MPKTHSIZE);
2848 		m = m_next;
2849 		m->m_next = tmp;
2850 	}
2851 	return (m);
2852 }
2853 
2854 /*
2855  * If dodgy hardware rejects the scatter gather chain we've handed it
2856  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2857  * m_defrag'd mbufs
2858  */
2859 static __noinline struct mbuf *
2860 iflib_remove_mbuf(iflib_txq_t txq)
2861 {
2862 	int ntxd, i, pidx;
2863 	struct mbuf *m, *mh, **ifsd_m;
2864 
2865 	pidx = txq->ift_pidx;
2866 	ifsd_m = txq->ift_sds.ifsd_m;
2867 	ntxd = txq->ift_size;
2868 	mh = m = ifsd_m[pidx];
2869 	ifsd_m[pidx] = NULL;
2870 #if MEMORY_LOGGING
2871 	txq->ift_dequeued++;
2872 #endif
2873 	i = 1;
2874 
2875 	while (m) {
2876 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2877 #if MEMORY_LOGGING
2878 		txq->ift_dequeued++;
2879 #endif
2880 		m = m->m_next;
2881 		i++;
2882 	}
2883 	return (mh);
2884 }
2885 
2886 static int
2887 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2888 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2889 			  int max_segs, int flags)
2890 {
2891 	if_ctx_t ctx;
2892 	if_shared_ctx_t		sctx;
2893 	if_softc_ctx_t		scctx;
2894 	int i, next, pidx, err, maxsegsz, ntxd, count;
2895 	struct mbuf *m, *tmp, **ifsd_m;
2896 
2897 	m = *m0;
2898 
2899 	/*
2900 	 * Please don't ever do this
2901 	 */
2902 	if (__predict_false(m->m_len == 0))
2903 		*m0 = m = collapse_pkthdr(m);
2904 
2905 	ctx = txq->ift_ctx;
2906 	sctx = ctx->ifc_sctx;
2907 	scctx = &ctx->ifc_softc_ctx;
2908 	ifsd_m = txq->ift_sds.ifsd_m;
2909 	ntxd = txq->ift_size;
2910 	pidx = txq->ift_pidx;
2911 	if (map != NULL) {
2912 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2913 
2914 		err = bus_dmamap_load_mbuf_sg(tag, map,
2915 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2916 		if (err)
2917 			return (err);
2918 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2919 		count = 0;
2920 		m = *m0;
2921 		do {
2922 			if (__predict_false(m->m_len <= 0)) {
2923 				tmp = m;
2924 				m = m->m_next;
2925 				tmp->m_next = NULL;
2926 				m_free(tmp);
2927 				continue;
2928 			}
2929 			next = (pidx + count) & (ntxd-1);
2930 			MPASS(ifsd_m[next] == NULL);
2931 			ifsd_m[next] = m;
2932 			count++;
2933 			tmp = m;
2934 			m = m->m_next;
2935 		} while (m != NULL);
2936 	} else {
2937 		int buflen, sgsize, max_sgsize;
2938 		vm_offset_t vaddr;
2939 		vm_paddr_t curaddr;
2940 
2941 		count = i = 0;
2942 		maxsegsz = sctx->isc_tx_maxsize;
2943 		m = *m0;
2944 		do {
2945 			if (__predict_false(m->m_len <= 0)) {
2946 				tmp = m;
2947 				m = m->m_next;
2948 				tmp->m_next = NULL;
2949 				m_free(tmp);
2950 				continue;
2951 			}
2952 			buflen = m->m_len;
2953 			vaddr = (vm_offset_t)m->m_data;
2954 			/*
2955 			 * see if we can't be smarter about physically
2956 			 * contiguous mappings
2957 			 */
2958 			next = (pidx + count) & (ntxd-1);
2959 			MPASS(ifsd_m[next] == NULL);
2960 #if MEMORY_LOGGING
2961 			txq->ift_enqueued++;
2962 #endif
2963 			ifsd_m[next] = m;
2964 			while (buflen > 0) {
2965 				max_sgsize = MIN(buflen, maxsegsz);
2966 				curaddr = pmap_kextract(vaddr);
2967 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2968 				sgsize = MIN(sgsize, max_sgsize);
2969 				segs[i].ds_addr = curaddr;
2970 				segs[i].ds_len = sgsize;
2971 				vaddr += sgsize;
2972 				buflen -= sgsize;
2973 				i++;
2974 				if (i >= max_segs)
2975 					goto err;
2976 			}
2977 			count++;
2978 			tmp = m;
2979 			m = m->m_next;
2980 		} while (m != NULL);
2981 		*nsegs = i;
2982 	}
2983 	return (0);
2984 err:
2985 	*m0 = iflib_remove_mbuf(txq);
2986 	return (EFBIG);
2987 }
2988 
2989 static inline caddr_t
2990 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2991 {
2992 	qidx_t size;
2993 	int ntxd;
2994 	caddr_t start, end, cur, next;
2995 
2996 	ntxd = txq->ift_size;
2997 	size = txq->ift_txd_size[qid];
2998 	start = txq->ift_ifdi[qid].idi_vaddr;
2999 
3000 	if (__predict_false(size == 0))
3001 		return (start);
3002 	cur = start + size*cidx;
3003 	end = start + size*ntxd;
3004 	next = CACHE_PTR_NEXT(cur);
3005 	return (next < end ? next : start);
3006 }
3007 
3008 static int
3009 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3010 {
3011 	if_ctx_t		ctx;
3012 	if_shared_ctx_t		sctx;
3013 	if_softc_ctx_t		scctx;
3014 	bus_dma_segment_t	*segs;
3015 	struct mbuf		*m_head;
3016 	void			*next_txd;
3017 	bus_dmamap_t		map;
3018 	struct if_pkt_info	pi;
3019 	int remap = 0;
3020 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3021 	bus_dma_tag_t desc_tag;
3022 
3023 	segs = txq->ift_segs;
3024 	ctx = txq->ift_ctx;
3025 	sctx = ctx->ifc_sctx;
3026 	scctx = &ctx->ifc_softc_ctx;
3027 	segs = txq->ift_segs;
3028 	ntxd = txq->ift_size;
3029 	m_head = *m_headp;
3030 	map = NULL;
3031 
3032 	/*
3033 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3034 	 */
3035 	cidx = txq->ift_cidx;
3036 	pidx = txq->ift_pidx;
3037 	if (ctx->ifc_flags & IFC_PREFETCH) {
3038 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3039 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3040 			next_txd = calc_next_txd(txq, cidx, 0);
3041 			prefetch(next_txd);
3042 		}
3043 
3044 		/* prefetch the next cache line of mbuf pointers and flags */
3045 		prefetch(&txq->ift_sds.ifsd_m[next]);
3046 		if (txq->ift_sds.ifsd_map != NULL) {
3047 			prefetch(&txq->ift_sds.ifsd_map[next]);
3048 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3049 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3050 		}
3051 	} else if (txq->ift_sds.ifsd_map != NULL)
3052 		map = txq->ift_sds.ifsd_map[pidx];
3053 
3054 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3055 		desc_tag = txq->ift_tso_desc_tag;
3056 		max_segs = scctx->isc_tx_tso_segments_max;
3057 	} else {
3058 		desc_tag = txq->ift_desc_tag;
3059 		max_segs = scctx->isc_tx_nsegments;
3060 	}
3061 	m_head = *m_headp;
3062 
3063 	pkt_info_zero(&pi);
3064 	pi.ipi_len = m_head->m_pkthdr.len;
3065 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3066 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3067 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3068 	pi.ipi_pidx = pidx;
3069 	pi.ipi_qsidx = txq->ift_id;
3070 
3071 	/* deliberate bitwise OR to make one condition */
3072 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3073 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3074 			return (err);
3075 		m_head = *m_headp;
3076 	}
3077 
3078 retry:
3079 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3080 defrag:
3081 	if (__predict_false(err)) {
3082 		switch (err) {
3083 		case EFBIG:
3084 			/* try collapse once and defrag once */
3085 			if (remap == 0)
3086 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3087 			if (remap == 1)
3088 				m_head = m_defrag(*m_headp, M_NOWAIT);
3089 			remap++;
3090 			if (__predict_false(m_head == NULL))
3091 				goto defrag_failed;
3092 			txq->ift_mbuf_defrag++;
3093 			*m_headp = m_head;
3094 			goto retry;
3095 			break;
3096 		case ENOMEM:
3097 			txq->ift_no_tx_dma_setup++;
3098 			break;
3099 		default:
3100 			txq->ift_no_tx_dma_setup++;
3101 			m_freem(*m_headp);
3102 			DBG_COUNTER_INC(tx_frees);
3103 			*m_headp = NULL;
3104 			break;
3105 		}
3106 		txq->ift_map_failed++;
3107 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3108 		return (err);
3109 	}
3110 
3111 	/*
3112 	 * XXX assumes a 1 to 1 relationship between segments and
3113 	 *        descriptors - this does not hold true on all drivers, e.g.
3114 	 *        cxgb
3115 	 */
3116 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3117 		txq->ift_no_desc_avail++;
3118 		if (map != NULL)
3119 			bus_dmamap_unload(desc_tag, map);
3120 		DBG_COUNTER_INC(encap_txq_avail_fail);
3121 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3122 			GROUPTASK_ENQUEUE(&txq->ift_task);
3123 		return (ENOBUFS);
3124 	}
3125 	/*
3126 	 * On Intel cards we can greatly reduce the number of TX interrupts
3127 	 * we see by only setting report status on every Nth descriptor.
3128 	 * However, this also means that the driver will need to keep track
3129 	 * of the descriptors that RS was set on to check them for the DD bit.
3130 	 */
3131 	txq->ift_rs_pending += nsegs + 1;
3132 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3133 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3134 		pi.ipi_flags |= IPI_TX_INTR;
3135 		txq->ift_rs_pending = 0;
3136 	}
3137 
3138 	pi.ipi_segs = segs;
3139 	pi.ipi_nsegs = nsegs;
3140 
3141 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3142 #ifdef PKT_DEBUG
3143 	print_pkt(&pi);
3144 #endif
3145 	if (map != NULL)
3146 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3147 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3148 		if (map != NULL)
3149 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3150 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3151 		DBG_COUNTER_INC(tx_encap);
3152 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3153 
3154 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3155 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3156 			ndesc += txq->ift_size;
3157 			txq->ift_gen = 1;
3158 		}
3159 		/*
3160 		 * drivers can need as many as
3161 		 * two sentinels
3162 		 */
3163 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3164 		MPASS(pi.ipi_new_pidx != pidx);
3165 		MPASS(ndesc > 0);
3166 		txq->ift_in_use += ndesc;
3167 
3168 		/*
3169 		 * We update the last software descriptor again here because there may
3170 		 * be a sentinel and/or there may be more mbufs than segments
3171 		 */
3172 		txq->ift_pidx = pi.ipi_new_pidx;
3173 		txq->ift_npending += pi.ipi_ndescs;
3174 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3175 		*m_headp = m_head = iflib_remove_mbuf(txq);
3176 		remap = 1;
3177 		txq->ift_txd_encap_efbig++;
3178 		goto defrag;
3179 	} else
3180 		DBG_COUNTER_INC(encap_txd_encap_fail);
3181 	return (err);
3182 
3183 defrag_failed:
3184 	txq->ift_mbuf_defrag_failed++;
3185 	txq->ift_map_failed++;
3186 	m_freem(*m_headp);
3187 	DBG_COUNTER_INC(tx_frees);
3188 	*m_headp = NULL;
3189 	return (ENOMEM);
3190 }
3191 
3192 static void
3193 iflib_tx_desc_free(iflib_txq_t txq, int n)
3194 {
3195 	int hasmap;
3196 	uint32_t qsize, cidx, mask, gen;
3197 	struct mbuf *m, **ifsd_m;
3198 	uint8_t *ifsd_flags;
3199 	bus_dmamap_t *ifsd_map;
3200 	bool do_prefetch;
3201 
3202 	cidx = txq->ift_cidx;
3203 	gen = txq->ift_gen;
3204 	qsize = txq->ift_size;
3205 	mask = qsize-1;
3206 	hasmap = txq->ift_sds.ifsd_map != NULL;
3207 	ifsd_flags = txq->ift_sds.ifsd_flags;
3208 	ifsd_m = txq->ift_sds.ifsd_m;
3209 	ifsd_map = txq->ift_sds.ifsd_map;
3210 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3211 
3212 	while (n--) {
3213 		if (do_prefetch) {
3214 			prefetch(ifsd_m[(cidx + 3) & mask]);
3215 			prefetch(ifsd_m[(cidx + 4) & mask]);
3216 		}
3217 		if (ifsd_m[cidx] != NULL) {
3218 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3219 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3220 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3221 				/*
3222 				 * does it matter if it's not the TSO tag? If so we'll
3223 				 * have to add the type to flags
3224 				 */
3225 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3226 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3227 			}
3228 			if ((m = ifsd_m[cidx]) != NULL) {
3229 				/* XXX we don't support any drivers that batch packets yet */
3230 				MPASS(m->m_nextpkt == NULL);
3231 
3232 				m_free(m);
3233 				ifsd_m[cidx] = NULL;
3234 #if MEMORY_LOGGING
3235 				txq->ift_dequeued++;
3236 #endif
3237 				DBG_COUNTER_INC(tx_frees);
3238 			}
3239 		}
3240 		if (__predict_false(++cidx == qsize)) {
3241 			cidx = 0;
3242 			gen = 0;
3243 		}
3244 	}
3245 	txq->ift_cidx = cidx;
3246 	txq->ift_gen = gen;
3247 }
3248 
3249 static __inline int
3250 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3251 {
3252 	int reclaim;
3253 	if_ctx_t ctx = txq->ift_ctx;
3254 
3255 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3256 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3257 
3258 	/*
3259 	 * Need a rate-limiting check so that this isn't called every time
3260 	 */
3261 	iflib_tx_credits_update(ctx, txq);
3262 	reclaim = DESC_RECLAIMABLE(txq);
3263 
3264 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3265 #ifdef INVARIANTS
3266 		if (iflib_verbose_debug) {
3267 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3268 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3269 			       reclaim, thresh);
3270 
3271 		}
3272 #endif
3273 		return (0);
3274 	}
3275 	iflib_tx_desc_free(txq, reclaim);
3276 	txq->ift_cleaned += reclaim;
3277 	txq->ift_in_use -= reclaim;
3278 
3279 	return (reclaim);
3280 }
3281 
3282 static struct mbuf **
3283 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3284 {
3285 	int next, size;
3286 	struct mbuf **items;
3287 
3288 	size = r->size;
3289 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3290 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3291 
3292 	prefetch(items[(cidx + offset) & (size-1)]);
3293 	if (remaining > 1) {
3294 		prefetch(&items[next]);
3295 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3296 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3297 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3298 	}
3299 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3300 }
3301 
3302 static void
3303 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3304 {
3305 
3306 	ifmp_ring_check_drainage(txq->ift_br, budget);
3307 }
3308 
3309 static uint32_t
3310 iflib_txq_can_drain(struct ifmp_ring *r)
3311 {
3312 	iflib_txq_t txq = r->cookie;
3313 	if_ctx_t ctx = txq->ift_ctx;
3314 
3315 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3316 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3317 }
3318 
3319 static uint32_t
3320 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3321 {
3322 	iflib_txq_t txq = r->cookie;
3323 	if_ctx_t ctx = txq->ift_ctx;
3324 	struct ifnet *ifp = ctx->ifc_ifp;
3325 	struct mbuf **mp, *m;
3326 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3327 	int reclaimed, err, in_use_prev, desc_used;
3328 	bool do_prefetch, ring, rang;
3329 
3330 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3331 			    !LINK_ACTIVE(ctx))) {
3332 		DBG_COUNTER_INC(txq_drain_notready);
3333 		return (0);
3334 	}
3335 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3336 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3337 	avail = IDXDIFF(pidx, cidx, r->size);
3338 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3339 		DBG_COUNTER_INC(txq_drain_flushing);
3340 		for (i = 0; i < avail; i++) {
3341 			m_free(r->items[(cidx + i) & (r->size-1)]);
3342 			r->items[(cidx + i) & (r->size-1)] = NULL;
3343 		}
3344 		return (avail);
3345 	}
3346 
3347 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3348 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3349 		CALLOUT_LOCK(txq);
3350 		callout_stop(&txq->ift_timer);
3351 		CALLOUT_UNLOCK(txq);
3352 		DBG_COUNTER_INC(txq_drain_oactive);
3353 		return (0);
3354 	}
3355 	if (reclaimed)
3356 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3357 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3358 	count = MIN(avail, TX_BATCH_SIZE);
3359 #ifdef INVARIANTS
3360 	if (iflib_verbose_debug)
3361 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3362 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3363 #endif
3364 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3365 	avail = TXQ_AVAIL(txq);
3366 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3367 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3368 
3369 		mp = _ring_peek_one(r, cidx, i, rem);
3370 		MPASS(mp != NULL && *mp != NULL);
3371 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3372 			consumed++;
3373 			reclaimed++;
3374 			continue;
3375 		}
3376 		in_use_prev = txq->ift_in_use;
3377 		pidx_prev = txq->ift_pidx;
3378 		err = iflib_encap(txq, mp);
3379 		if (__predict_false(err)) {
3380 			DBG_COUNTER_INC(txq_drain_encapfail);
3381 			/* no room - bail out */
3382 			if (err == ENOBUFS)
3383 				break;
3384 			consumed++;
3385 			DBG_COUNTER_INC(txq_drain_encapfail);
3386 			/* we can't send this packet - skip it */
3387 			continue;
3388 		}
3389 		consumed++;
3390 		pkt_sent++;
3391 		m = *mp;
3392 		DBG_COUNTER_INC(tx_sent);
3393 		bytes_sent += m->m_pkthdr.len;
3394 		mcast_sent += !!(m->m_flags & M_MCAST);
3395 		avail = TXQ_AVAIL(txq);
3396 
3397 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3398 		desc_used += (txq->ift_in_use - in_use_prev);
3399 		ETHER_BPF_MTAP(ifp, m);
3400 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3401 			break;
3402 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3403 	}
3404 
3405 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3406 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3407 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3408 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3409 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3410 	if (mcast_sent)
3411 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3412 #ifdef INVARIANTS
3413 	if (iflib_verbose_debug)
3414 		printf("consumed=%d\n", consumed);
3415 #endif
3416 	return (consumed);
3417 }
3418 
3419 static uint32_t
3420 iflib_txq_drain_always(struct ifmp_ring *r)
3421 {
3422 	return (1);
3423 }
3424 
3425 static uint32_t
3426 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3427 {
3428 	int i, avail;
3429 	struct mbuf **mp;
3430 	iflib_txq_t txq;
3431 
3432 	txq = r->cookie;
3433 
3434 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3435 	CALLOUT_LOCK(txq);
3436 	callout_stop(&txq->ift_timer);
3437 	CALLOUT_UNLOCK(txq);
3438 
3439 	avail = IDXDIFF(pidx, cidx, r->size);
3440 	for (i = 0; i < avail; i++) {
3441 		mp = _ring_peek_one(r, cidx, i, avail - i);
3442 		if (__predict_false(*mp == (struct mbuf *)txq))
3443 			continue;
3444 		m_freem(*mp);
3445 	}
3446 	MPASS(ifmp_ring_is_stalled(r) == 0);
3447 	return (avail);
3448 }
3449 
3450 static void
3451 iflib_ifmp_purge(iflib_txq_t txq)
3452 {
3453 	struct ifmp_ring *r;
3454 
3455 	r = txq->ift_br;
3456 	r->drain = iflib_txq_drain_free;
3457 	r->can_drain = iflib_txq_drain_always;
3458 
3459 	ifmp_ring_check_drainage(r, r->size);
3460 
3461 	r->drain = iflib_txq_drain;
3462 	r->can_drain = iflib_txq_can_drain;
3463 }
3464 
3465 static void
3466 _task_fn_tx(void *context)
3467 {
3468 	iflib_txq_t txq = context;
3469 	if_ctx_t ctx = txq->ift_ctx;
3470 	struct ifnet *ifp = ctx->ifc_ifp;
3471 	int rc;
3472 
3473 #ifdef IFLIB_DIAGNOSTICS
3474 	txq->ift_cpu_exec_count[curcpu]++;
3475 #endif
3476 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3477 		return;
3478 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3479 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3480 			netmap_tx_irq(ifp, txq->ift_id);
3481 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3482 		return;
3483 	}
3484 	if (txq->ift_db_pending)
3485 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3486 	else
3487 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3488 	if (ctx->ifc_flags & IFC_LEGACY)
3489 		IFDI_INTR_ENABLE(ctx);
3490 	else {
3491 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3492 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3493 	}
3494 }
3495 
3496 static void
3497 _task_fn_rx(void *context)
3498 {
3499 	iflib_rxq_t rxq = context;
3500 	if_ctx_t ctx = rxq->ifr_ctx;
3501 	bool more;
3502 	int rc;
3503 
3504 #ifdef IFLIB_DIAGNOSTICS
3505 	rxq->ifr_cpu_exec_count[curcpu]++;
3506 #endif
3507 	DBG_COUNTER_INC(task_fn_rxs);
3508 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3509 		return;
3510 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3511 		if (ctx->ifc_flags & IFC_LEGACY)
3512 			IFDI_INTR_ENABLE(ctx);
3513 		else {
3514 			DBG_COUNTER_INC(rx_intr_enables);
3515 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3516 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3517 		}
3518 	}
3519 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3520 		return;
3521 	if (more)
3522 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3523 }
3524 
3525 static void
3526 _task_fn_admin(void *context)
3527 {
3528 	if_ctx_t ctx = context;
3529 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3530 	iflib_txq_t txq;
3531 	int i;
3532 
3533 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3534 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3535 			return;
3536 		}
3537 	}
3538 
3539 	CTX_LOCK(ctx);
3540 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3541 		CALLOUT_LOCK(txq);
3542 		callout_stop(&txq->ift_timer);
3543 		CALLOUT_UNLOCK(txq);
3544 	}
3545 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3546 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3547 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3548 	IFDI_LINK_INTR_ENABLE(ctx);
3549 	if (ctx->ifc_flags & IFC_DO_RESET) {
3550 		ctx->ifc_flags &= ~IFC_DO_RESET;
3551 		iflib_if_init_locked(ctx);
3552 	}
3553 	CTX_UNLOCK(ctx);
3554 
3555 	if (LINK_ACTIVE(ctx) == 0)
3556 		return;
3557 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3558 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3559 }
3560 
3561 
3562 static void
3563 _task_fn_iov(void *context)
3564 {
3565 	if_ctx_t ctx = context;
3566 
3567 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3568 		return;
3569 
3570 	CTX_LOCK(ctx);
3571 	IFDI_VFLR_HANDLE(ctx);
3572 	CTX_UNLOCK(ctx);
3573 }
3574 
3575 static int
3576 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3577 {
3578 	int err;
3579 	if_int_delay_info_t info;
3580 	if_ctx_t ctx;
3581 
3582 	info = (if_int_delay_info_t)arg1;
3583 	ctx = info->iidi_ctx;
3584 	info->iidi_req = req;
3585 	info->iidi_oidp = oidp;
3586 	CTX_LOCK(ctx);
3587 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3588 	CTX_UNLOCK(ctx);
3589 	return (err);
3590 }
3591 
3592 /*********************************************************************
3593  *
3594  *  IFNET FUNCTIONS
3595  *
3596  **********************************************************************/
3597 
3598 static void
3599 iflib_if_init_locked(if_ctx_t ctx)
3600 {
3601 	iflib_stop(ctx);
3602 	iflib_init_locked(ctx);
3603 }
3604 
3605 
3606 static void
3607 iflib_if_init(void *arg)
3608 {
3609 	if_ctx_t ctx = arg;
3610 
3611 	CTX_LOCK(ctx);
3612 	iflib_if_init_locked(ctx);
3613 	CTX_UNLOCK(ctx);
3614 }
3615 
3616 static int
3617 iflib_if_transmit(if_t ifp, struct mbuf *m)
3618 {
3619 	if_ctx_t	ctx = if_getsoftc(ifp);
3620 
3621 	iflib_txq_t txq;
3622 	int err, qidx;
3623 
3624 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3625 		DBG_COUNTER_INC(tx_frees);
3626 		m_freem(m);
3627 		return (ENOBUFS);
3628 	}
3629 
3630 	MPASS(m->m_nextpkt == NULL);
3631 	qidx = 0;
3632 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3633 		qidx = QIDX(ctx, m);
3634 	/*
3635 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3636 	 */
3637 	txq = &ctx->ifc_txqs[qidx];
3638 
3639 #ifdef DRIVER_BACKPRESSURE
3640 	if (txq->ift_closed) {
3641 		while (m != NULL) {
3642 			next = m->m_nextpkt;
3643 			m->m_nextpkt = NULL;
3644 			m_freem(m);
3645 			m = next;
3646 		}
3647 		return (ENOBUFS);
3648 	}
3649 #endif
3650 #ifdef notyet
3651 	qidx = count = 0;
3652 	mp = marr;
3653 	next = m;
3654 	do {
3655 		count++;
3656 		next = next->m_nextpkt;
3657 	} while (next != NULL);
3658 
3659 	if (count > nitems(marr))
3660 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3661 			/* XXX check nextpkt */
3662 			m_freem(m);
3663 			/* XXX simplify for now */
3664 			DBG_COUNTER_INC(tx_frees);
3665 			return (ENOBUFS);
3666 		}
3667 	for (next = m, i = 0; next != NULL; i++) {
3668 		mp[i] = next;
3669 		next = next->m_nextpkt;
3670 		mp[i]->m_nextpkt = NULL;
3671 	}
3672 #endif
3673 	DBG_COUNTER_INC(tx_seen);
3674 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3675 
3676 	if (err) {
3677 		GROUPTASK_ENQUEUE(&txq->ift_task);
3678 		/* support forthcoming later */
3679 #ifdef DRIVER_BACKPRESSURE
3680 		txq->ift_closed = TRUE;
3681 #endif
3682 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3683 		m_freem(m);
3684 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3685 		GROUPTASK_ENQUEUE(&txq->ift_task);
3686 	}
3687 
3688 	return (err);
3689 }
3690 
3691 static void
3692 iflib_if_qflush(if_t ifp)
3693 {
3694 	if_ctx_t ctx = if_getsoftc(ifp);
3695 	iflib_txq_t txq = ctx->ifc_txqs;
3696 	int i;
3697 
3698 	CTX_LOCK(ctx);
3699 	ctx->ifc_flags |= IFC_QFLUSH;
3700 	CTX_UNLOCK(ctx);
3701 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3702 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3703 			iflib_txq_check_drain(txq, 0);
3704 	CTX_LOCK(ctx);
3705 	ctx->ifc_flags &= ~IFC_QFLUSH;
3706 	CTX_UNLOCK(ctx);
3707 
3708 	if_qflush(ifp);
3709 }
3710 
3711 
3712 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3713 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3714 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3715 
3716 static int
3717 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3718 {
3719 	if_ctx_t ctx = if_getsoftc(ifp);
3720 	struct ifreq	*ifr = (struct ifreq *)data;
3721 #if defined(INET) || defined(INET6)
3722 	struct ifaddr	*ifa = (struct ifaddr *)data;
3723 #endif
3724 	bool		avoid_reset = FALSE;
3725 	int		err = 0, reinit = 0, bits;
3726 
3727 	switch (command) {
3728 	case SIOCSIFADDR:
3729 #ifdef INET
3730 		if (ifa->ifa_addr->sa_family == AF_INET)
3731 			avoid_reset = TRUE;
3732 #endif
3733 #ifdef INET6
3734 		if (ifa->ifa_addr->sa_family == AF_INET6)
3735 			avoid_reset = TRUE;
3736 #endif
3737 		/*
3738 		** Calling init results in link renegotiation,
3739 		** so we avoid doing it when possible.
3740 		*/
3741 		if (avoid_reset) {
3742 			if_setflagbits(ifp, IFF_UP,0);
3743 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3744 				reinit = 1;
3745 #ifdef INET
3746 			if (!(if_getflags(ifp) & IFF_NOARP))
3747 				arp_ifinit(ifp, ifa);
3748 #endif
3749 		} else
3750 			err = ether_ioctl(ifp, command, data);
3751 		break;
3752 	case SIOCSIFMTU:
3753 		CTX_LOCK(ctx);
3754 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3755 			CTX_UNLOCK(ctx);
3756 			break;
3757 		}
3758 		bits = if_getdrvflags(ifp);
3759 		/* stop the driver and free any clusters before proceeding */
3760 		iflib_stop(ctx);
3761 
3762 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3763 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3764 				ctx->ifc_flags |= IFC_MULTISEG;
3765 			else
3766 				ctx->ifc_flags &= ~IFC_MULTISEG;
3767 			err = if_setmtu(ifp, ifr->ifr_mtu);
3768 		}
3769 		iflib_init_locked(ctx);
3770 		if_setdrvflags(ifp, bits);
3771 		CTX_UNLOCK(ctx);
3772 		break;
3773 	case SIOCSIFFLAGS:
3774 		CTX_LOCK(ctx);
3775 		if (if_getflags(ifp) & IFF_UP) {
3776 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3777 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3778 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3779 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3780 				}
3781 			} else
3782 				reinit = 1;
3783 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3784 			iflib_stop(ctx);
3785 		}
3786 		ctx->ifc_if_flags = if_getflags(ifp);
3787 		CTX_UNLOCK(ctx);
3788 		break;
3789 	case SIOCADDMULTI:
3790 	case SIOCDELMULTI:
3791 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3792 			CTX_LOCK(ctx);
3793 			IFDI_INTR_DISABLE(ctx);
3794 			IFDI_MULTI_SET(ctx);
3795 			IFDI_INTR_ENABLE(ctx);
3796 			CTX_UNLOCK(ctx);
3797 		}
3798 		break;
3799 	case SIOCSIFMEDIA:
3800 		CTX_LOCK(ctx);
3801 		IFDI_MEDIA_SET(ctx);
3802 		CTX_UNLOCK(ctx);
3803 		/* falls thru */
3804 	case SIOCGIFMEDIA:
3805 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3806 		break;
3807 	case SIOCGI2C:
3808 	{
3809 		struct ifi2creq i2c;
3810 
3811 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3812 		if (err != 0)
3813 			break;
3814 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3815 			err = EINVAL;
3816 			break;
3817 		}
3818 		if (i2c.len > sizeof(i2c.data)) {
3819 			err = EINVAL;
3820 			break;
3821 		}
3822 
3823 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3824 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3825 		break;
3826 	}
3827 	case SIOCSIFCAP:
3828 	{
3829 		int mask, setmask;
3830 
3831 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3832 		setmask = 0;
3833 #ifdef TCP_OFFLOAD
3834 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3835 #endif
3836 		setmask |= (mask & IFCAP_FLAGS);
3837 
3838 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3839 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3840 		if ((mask & IFCAP_WOL) &&
3841 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3842 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3843 		if_vlancap(ifp);
3844 		/*
3845 		 * want to ensure that traffic has stopped before we change any of the flags
3846 		 */
3847 		if (setmask) {
3848 			CTX_LOCK(ctx);
3849 			bits = if_getdrvflags(ifp);
3850 			if (bits & IFF_DRV_RUNNING)
3851 				iflib_stop(ctx);
3852 			if_togglecapenable(ifp, setmask);
3853 			if (bits & IFF_DRV_RUNNING)
3854 				iflib_init_locked(ctx);
3855 			if_setdrvflags(ifp, bits);
3856 			CTX_UNLOCK(ctx);
3857 		}
3858 		break;
3859 	    }
3860 	case SIOCGPRIVATE_0:
3861 	case SIOCSDRVSPEC:
3862 	case SIOCGDRVSPEC:
3863 		CTX_LOCK(ctx);
3864 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3865 		CTX_UNLOCK(ctx);
3866 		break;
3867 	default:
3868 		err = ether_ioctl(ifp, command, data);
3869 		break;
3870 	}
3871 	if (reinit)
3872 		iflib_if_init(ctx);
3873 	return (err);
3874 }
3875 
3876 static uint64_t
3877 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3878 {
3879 	if_ctx_t ctx = if_getsoftc(ifp);
3880 
3881 	return (IFDI_GET_COUNTER(ctx, cnt));
3882 }
3883 
3884 /*********************************************************************
3885  *
3886  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3887  *
3888  **********************************************************************/
3889 
3890 static void
3891 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3892 {
3893 	if_ctx_t ctx = if_getsoftc(ifp);
3894 
3895 	if ((void *)ctx != arg)
3896 		return;
3897 
3898 	if ((vtag == 0) || (vtag > 4095))
3899 		return;
3900 
3901 	CTX_LOCK(ctx);
3902 	IFDI_VLAN_REGISTER(ctx, vtag);
3903 	/* Re-init to load the changes */
3904 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3905 		iflib_init_locked(ctx);
3906 	CTX_UNLOCK(ctx);
3907 }
3908 
3909 static void
3910 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3911 {
3912 	if_ctx_t ctx = if_getsoftc(ifp);
3913 
3914 	if ((void *)ctx != arg)
3915 		return;
3916 
3917 	if ((vtag == 0) || (vtag > 4095))
3918 		return;
3919 
3920 	CTX_LOCK(ctx);
3921 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3922 	/* Re-init to load the changes */
3923 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3924 		iflib_init_locked(ctx);
3925 	CTX_UNLOCK(ctx);
3926 }
3927 
3928 static void
3929 iflib_led_func(void *arg, int onoff)
3930 {
3931 	if_ctx_t ctx = arg;
3932 
3933 	CTX_LOCK(ctx);
3934 	IFDI_LED_FUNC(ctx, onoff);
3935 	CTX_UNLOCK(ctx);
3936 }
3937 
3938 /*********************************************************************
3939  *
3940  *  BUS FUNCTION DEFINITIONS
3941  *
3942  **********************************************************************/
3943 
3944 int
3945 iflib_device_probe(device_t dev)
3946 {
3947 	pci_vendor_info_t *ent;
3948 
3949 	uint16_t	pci_vendor_id, pci_device_id;
3950 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3951 	uint16_t	pci_rev_id;
3952 	if_shared_ctx_t sctx;
3953 
3954 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3955 		return (ENOTSUP);
3956 
3957 	pci_vendor_id = pci_get_vendor(dev);
3958 	pci_device_id = pci_get_device(dev);
3959 	pci_subvendor_id = pci_get_subvendor(dev);
3960 	pci_subdevice_id = pci_get_subdevice(dev);
3961 	pci_rev_id = pci_get_revid(dev);
3962 	if (sctx->isc_parse_devinfo != NULL)
3963 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3964 
3965 	ent = sctx->isc_vendor_info;
3966 	while (ent->pvi_vendor_id != 0) {
3967 		if (pci_vendor_id != ent->pvi_vendor_id) {
3968 			ent++;
3969 			continue;
3970 		}
3971 		if ((pci_device_id == ent->pvi_device_id) &&
3972 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3973 		     (ent->pvi_subvendor_id == 0)) &&
3974 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3975 		     (ent->pvi_subdevice_id == 0)) &&
3976 		    ((pci_rev_id == ent->pvi_rev_id) ||
3977 		     (ent->pvi_rev_id == 0))) {
3978 
3979 			device_set_desc_copy(dev, ent->pvi_name);
3980 			/* this needs to be changed to zero if the bus probing code
3981 			 * ever stops re-probing on best match because the sctx
3982 			 * may have its values over written by register calls
3983 			 * in subsequent probes
3984 			 */
3985 			return (BUS_PROBE_DEFAULT);
3986 		}
3987 		ent++;
3988 	}
3989 	return (ENXIO);
3990 }
3991 
3992 int
3993 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3994 {
3995 	int err, rid, msix, msix_bar;
3996 	if_ctx_t ctx;
3997 	if_t ifp;
3998 	if_softc_ctx_t scctx;
3999 	int i;
4000 	uint16_t main_txq;
4001 	uint16_t main_rxq;
4002 
4003 
4004 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4005 
4006 	if (sc == NULL) {
4007 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4008 		device_set_softc(dev, ctx);
4009 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4010 	}
4011 
4012 	ctx->ifc_sctx = sctx;
4013 	ctx->ifc_dev = dev;
4014 	ctx->ifc_softc = sc;
4015 
4016 	if ((err = iflib_register(ctx)) != 0) {
4017 		device_printf(dev, "iflib_register failed %d\n", err);
4018 		return (err);
4019 	}
4020 	iflib_add_device_sysctl_pre(ctx);
4021 
4022 	scctx = &ctx->ifc_softc_ctx;
4023 	ifp = ctx->ifc_ifp;
4024 
4025 	/*
4026 	 * XXX sanity check that ntxd & nrxd are a power of 2
4027 	 */
4028 	if (ctx->ifc_sysctl_ntxqs != 0)
4029 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4030 	if (ctx->ifc_sysctl_nrxqs != 0)
4031 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4032 
4033 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4034 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4035 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4036 		else
4037 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4038 	}
4039 
4040 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4041 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4042 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4043 		else
4044 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4045 	}
4046 
4047 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4048 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4049 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4050 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4051 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4052 		}
4053 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4054 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4055 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4056 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4057 		}
4058 	}
4059 
4060 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4061 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4062 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4063 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4064 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4065 		}
4066 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4067 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4068 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4069 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4070 		}
4071 	}
4072 
4073 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4074 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4075 		return (err);
4076 	}
4077 	_iflib_pre_assert(scctx);
4078 	ctx->ifc_txrx = *scctx->isc_txrx;
4079 
4080 #ifdef INVARIANTS
4081 	MPASS(scctx->isc_capenable);
4082 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4083 		MPASS(scctx->isc_tx_csum_flags);
4084 #endif
4085 
4086 	if_setcapabilities(ifp, scctx->isc_capenable);
4087 	if_setcapenable(ifp, scctx->isc_capenable);
4088 
4089 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4090 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4091 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4092 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4093 
4094 #ifdef ACPI_DMAR
4095 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4096 		ctx->ifc_flags |= IFC_DMAR;
4097 #elif !(defined(__i386__) || defined(__amd64__))
4098 	/* set unconditionally for !x86 */
4099 	ctx->ifc_flags |= IFC_DMAR;
4100 #endif
4101 
4102 	msix_bar = scctx->isc_msix_bar;
4103 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4104 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4105 
4106 	/* XXX change for per-queue sizes */
4107 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4108 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4109 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4110 		if (!powerof2(scctx->isc_nrxd[i])) {
4111 			/* round down instead? */
4112 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4113 			err = EINVAL;
4114 			goto fail;
4115 		}
4116 	}
4117 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4118 		if (!powerof2(scctx->isc_ntxd[i])) {
4119 			device_printf(dev,
4120 			    "# tx descriptors must be a power of 2");
4121 			err = EINVAL;
4122 			goto fail;
4123 		}
4124 	}
4125 
4126 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4127 	    MAX_SINGLE_PACKET_FRACTION)
4128 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4129 		    MAX_SINGLE_PACKET_FRACTION);
4130 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4131 	    MAX_SINGLE_PACKET_FRACTION)
4132 		scctx->isc_tx_tso_segments_max = max(1,
4133 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4134 
4135 	/*
4136 	 * Protect the stack against modern hardware
4137 	 */
4138 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4139 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4140 
4141 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4142 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4143 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4144 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4145 	if (scctx->isc_rss_table_size == 0)
4146 		scctx->isc_rss_table_size = 64;
4147 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4148 
4149 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4150 	/* XXX format name */
4151 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4152 	/*
4153 	** Now setup MSI or MSI/X, should
4154 	** return us the number of supported
4155 	** vectors. (Will be 1 for MSI)
4156 	*/
4157 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4158 		msix = scctx->isc_vectors;
4159 	} else if (scctx->isc_msix_bar != 0)
4160 	       /*
4161 		* The simple fact that isc_msix_bar is not 0 does not mean we
4162 		* we have a good value there that is known to work.
4163 		*/
4164 		msix = iflib_msix_init(ctx);
4165 	else {
4166 		scctx->isc_vectors = 1;
4167 		scctx->isc_ntxqsets = 1;
4168 		scctx->isc_nrxqsets = 1;
4169 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4170 		msix = 0;
4171 	}
4172 	/* Get memory for the station queues */
4173 	if ((err = iflib_queues_alloc(ctx))) {
4174 		device_printf(dev, "Unable to allocate queue memory\n");
4175 		goto fail;
4176 	}
4177 
4178 	if ((err = iflib_qset_structures_setup(ctx))) {
4179 		device_printf(dev, "qset structure setup failed %d\n", err);
4180 		goto fail_queues;
4181 	}
4182 
4183 	/*
4184 	 * Group taskqueues aren't properly set up until SMP is started,
4185 	 * so we disable interrupts until we can handle them post
4186 	 * SI_SUB_SMP.
4187 	 *
4188 	 * XXX: disabling interrupts doesn't actually work, at least for
4189 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4190 	 * we do null handling and depend on this not causing too large an
4191 	 * interrupt storm.
4192 	 */
4193 	IFDI_INTR_DISABLE(ctx);
4194 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4195 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4196 		goto fail_intr_free;
4197 	}
4198 	if (msix <= 1) {
4199 		rid = 0;
4200 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4201 			MPASS(msix == 1);
4202 			rid = 1;
4203 		}
4204 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4205 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4206 			goto fail_intr_free;
4207 		}
4208 	}
4209 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4210 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4211 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4212 		goto fail_detach;
4213 	}
4214 	if ((err = iflib_netmap_attach(ctx))) {
4215 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4216 		goto fail_detach;
4217 	}
4218 	*ctxp = ctx;
4219 
4220 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4221 	iflib_add_device_sysctl_post(ctx);
4222 	ctx->ifc_flags |= IFC_INIT_DONE;
4223 	return (0);
4224 fail_detach:
4225 	ether_ifdetach(ctx->ifc_ifp);
4226 fail_intr_free:
4227 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4228 		pci_release_msi(ctx->ifc_dev);
4229 fail_queues:
4230 	/* XXX free queues */
4231 fail:
4232 	IFDI_DETACH(ctx);
4233 	return (err);
4234 }
4235 
4236 int
4237 iflib_device_attach(device_t dev)
4238 {
4239 	if_ctx_t ctx;
4240 	if_shared_ctx_t sctx;
4241 
4242 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4243 		return (ENOTSUP);
4244 
4245 	pci_enable_busmaster(dev);
4246 
4247 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4248 }
4249 
4250 int
4251 iflib_device_deregister(if_ctx_t ctx)
4252 {
4253 	if_t ifp = ctx->ifc_ifp;
4254 	iflib_txq_t txq;
4255 	iflib_rxq_t rxq;
4256 	device_t dev = ctx->ifc_dev;
4257 	int i, j;
4258 	struct taskqgroup *tqg;
4259 	iflib_fl_t fl;
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 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4296 			free(fl->ifl_rx_bitmap, M_IFLIB);
4297 
4298 	}
4299 	tqg = qgroup_if_config_tqg;
4300 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4301 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4302 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4303 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4304 
4305 	IFDI_DETACH(ctx);
4306 	device_set_softc(ctx->ifc_dev, NULL);
4307 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4308 		pci_release_msi(dev);
4309 	}
4310 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4311 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4312 	}
4313 	if (ctx->ifc_msix_mem != NULL) {
4314 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4315 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4316 		ctx->ifc_msix_mem = NULL;
4317 	}
4318 
4319 	bus_generic_detach(dev);
4320 	if_free(ifp);
4321 
4322 	iflib_tx_structures_free(ctx);
4323 	iflib_rx_structures_free(ctx);
4324 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4325 		free(ctx->ifc_softc, M_IFLIB);
4326 	free(ctx, M_IFLIB);
4327 	return (0);
4328 }
4329 
4330 
4331 int
4332 iflib_device_detach(device_t dev)
4333 {
4334 	if_ctx_t ctx = device_get_softc(dev);
4335 
4336 	return (iflib_device_deregister(ctx));
4337 }
4338 
4339 int
4340 iflib_device_suspend(device_t dev)
4341 {
4342 	if_ctx_t ctx = device_get_softc(dev);
4343 
4344 	CTX_LOCK(ctx);
4345 	IFDI_SUSPEND(ctx);
4346 	CTX_UNLOCK(ctx);
4347 
4348 	return bus_generic_suspend(dev);
4349 }
4350 int
4351 iflib_device_shutdown(device_t dev)
4352 {
4353 	if_ctx_t ctx = device_get_softc(dev);
4354 
4355 	CTX_LOCK(ctx);
4356 	IFDI_SHUTDOWN(ctx);
4357 	CTX_UNLOCK(ctx);
4358 
4359 	return bus_generic_suspend(dev);
4360 }
4361 
4362 
4363 int
4364 iflib_device_resume(device_t dev)
4365 {
4366 	if_ctx_t ctx = device_get_softc(dev);
4367 	iflib_txq_t txq = ctx->ifc_txqs;
4368 
4369 	CTX_LOCK(ctx);
4370 	IFDI_RESUME(ctx);
4371 	iflib_init_locked(ctx);
4372 	CTX_UNLOCK(ctx);
4373 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4374 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4375 
4376 	return (bus_generic_resume(dev));
4377 }
4378 
4379 int
4380 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4381 {
4382 	int error;
4383 	if_ctx_t ctx = device_get_softc(dev);
4384 
4385 	CTX_LOCK(ctx);
4386 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4387 	CTX_UNLOCK(ctx);
4388 
4389 	return (error);
4390 }
4391 
4392 void
4393 iflib_device_iov_uninit(device_t dev)
4394 {
4395 	if_ctx_t ctx = device_get_softc(dev);
4396 
4397 	CTX_LOCK(ctx);
4398 	IFDI_IOV_UNINIT(ctx);
4399 	CTX_UNLOCK(ctx);
4400 }
4401 
4402 int
4403 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4404 {
4405 	int error;
4406 	if_ctx_t ctx = device_get_softc(dev);
4407 
4408 	CTX_LOCK(ctx);
4409 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4410 	CTX_UNLOCK(ctx);
4411 
4412 	return (error);
4413 }
4414 
4415 /*********************************************************************
4416  *
4417  *  MODULE FUNCTION DEFINITIONS
4418  *
4419  **********************************************************************/
4420 
4421 /*
4422  * - Start a fast taskqueue thread for each core
4423  * - Start a taskqueue for control operations
4424  */
4425 static int
4426 iflib_module_init(void)
4427 {
4428 	return (0);
4429 }
4430 
4431 static int
4432 iflib_module_event_handler(module_t mod, int what, void *arg)
4433 {
4434 	int err;
4435 
4436 	switch (what) {
4437 	case MOD_LOAD:
4438 		if ((err = iflib_module_init()) != 0)
4439 			return (err);
4440 		break;
4441 	case MOD_UNLOAD:
4442 		return (EBUSY);
4443 	default:
4444 		return (EOPNOTSUPP);
4445 	}
4446 
4447 	return (0);
4448 }
4449 
4450 /*********************************************************************
4451  *
4452  *  PUBLIC FUNCTION DEFINITIONS
4453  *     ordered as in iflib.h
4454  *
4455  **********************************************************************/
4456 
4457 
4458 static void
4459 _iflib_assert(if_shared_ctx_t sctx)
4460 {
4461 	MPASS(sctx->isc_tx_maxsize);
4462 	MPASS(sctx->isc_tx_maxsegsize);
4463 
4464 	MPASS(sctx->isc_rx_maxsize);
4465 	MPASS(sctx->isc_rx_nsegments);
4466 	MPASS(sctx->isc_rx_maxsegsize);
4467 
4468 	MPASS(sctx->isc_nrxd_min[0]);
4469 	MPASS(sctx->isc_nrxd_max[0]);
4470 	MPASS(sctx->isc_nrxd_default[0]);
4471 	MPASS(sctx->isc_ntxd_min[0]);
4472 	MPASS(sctx->isc_ntxd_max[0]);
4473 	MPASS(sctx->isc_ntxd_default[0]);
4474 }
4475 
4476 static void
4477 _iflib_pre_assert(if_softc_ctx_t scctx)
4478 {
4479 
4480 	MPASS(scctx->isc_txrx->ift_txd_encap);
4481 	MPASS(scctx->isc_txrx->ift_txd_flush);
4482 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4483 	MPASS(scctx->isc_txrx->ift_rxd_available);
4484 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4485 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4486 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4487 }
4488 
4489 static int
4490 iflib_register(if_ctx_t ctx)
4491 {
4492 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4493 	driver_t *driver = sctx->isc_driver;
4494 	device_t dev = ctx->ifc_dev;
4495 	if_t ifp;
4496 
4497 	_iflib_assert(sctx);
4498 
4499 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4500 
4501 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4502 	if (ifp == NULL) {
4503 		device_printf(dev, "can not allocate ifnet structure\n");
4504 		return (ENOMEM);
4505 	}
4506 
4507 	/*
4508 	 * Initialize our context's device specific methods
4509 	 */
4510 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4511 	kobj_class_compile((kobj_class_t) driver);
4512 	driver->refs++;
4513 
4514 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4515 	if_setsoftc(ifp, ctx);
4516 	if_setdev(ifp, dev);
4517 	if_setinitfn(ifp, iflib_if_init);
4518 	if_setioctlfn(ifp, iflib_if_ioctl);
4519 	if_settransmitfn(ifp, iflib_if_transmit);
4520 	if_setqflushfn(ifp, iflib_if_qflush);
4521 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4522 
4523 	ctx->ifc_vlan_attach_event =
4524 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4525 							  EVENTHANDLER_PRI_FIRST);
4526 	ctx->ifc_vlan_detach_event =
4527 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4528 							  EVENTHANDLER_PRI_FIRST);
4529 
4530 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4531 					 iflib_media_change, iflib_media_status);
4532 
4533 	return (0);
4534 }
4535 
4536 
4537 static int
4538 iflib_queues_alloc(if_ctx_t ctx)
4539 {
4540 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4541 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4542 	device_t dev = ctx->ifc_dev;
4543 	int nrxqsets = scctx->isc_nrxqsets;
4544 	int ntxqsets = scctx->isc_ntxqsets;
4545 	iflib_txq_t txq;
4546 	iflib_rxq_t rxq;
4547 	iflib_fl_t fl = NULL;
4548 	int i, j, cpu, err, txconf, rxconf;
4549 	iflib_dma_info_t ifdip;
4550 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4551 	uint32_t *txqsizes = scctx->isc_txqsizes;
4552 	uint8_t nrxqs = sctx->isc_nrxqs;
4553 	uint8_t ntxqs = sctx->isc_ntxqs;
4554 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4555 	caddr_t *vaddrs;
4556 	uint64_t *paddrs;
4557 	struct ifmp_ring **brscp;
4558 
4559 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4560 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4561 
4562 	brscp = NULL;
4563 	txq = NULL;
4564 	rxq = NULL;
4565 
4566 /* Allocate the TX ring struct memory */
4567 	if (!(txq =
4568 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4569 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4570 		device_printf(dev, "Unable to allocate TX ring memory\n");
4571 		err = ENOMEM;
4572 		goto fail;
4573 	}
4574 
4575 	/* Now allocate the RX */
4576 	if (!(rxq =
4577 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4578 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4579 		device_printf(dev, "Unable to allocate RX ring memory\n");
4580 		err = ENOMEM;
4581 		goto rx_fail;
4582 	}
4583 
4584 	ctx->ifc_txqs = txq;
4585 	ctx->ifc_rxqs = rxq;
4586 
4587 	/*
4588 	 * XXX handle allocation failure
4589 	 */
4590 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4591 		/* Set up some basics */
4592 
4593 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4594 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4595 			err = ENOMEM;
4596 			goto err_tx_desc;
4597 		}
4598 		txq->ift_ifdi = ifdip;
4599 		for (j = 0; j < ntxqs; j++, ifdip++) {
4600 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4601 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4602 				err = ENOMEM;
4603 				goto err_tx_desc;
4604 			}
4605 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4606 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4607 		}
4608 		txq->ift_ctx = ctx;
4609 		txq->ift_id = i;
4610 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4611 			txq->ift_br_offset = 1;
4612 		} else {
4613 			txq->ift_br_offset = 0;
4614 		}
4615 		/* XXX fix this */
4616 		txq->ift_timer.c_cpu = cpu;
4617 
4618 		if (iflib_txsd_alloc(txq)) {
4619 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4620 			err = ENOMEM;
4621 			goto err_tx_desc;
4622 		}
4623 
4624 		/* Initialize the TX lock */
4625 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4626 		    device_get_nameunit(dev), txq->ift_id);
4627 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4628 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4629 
4630 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4631 			 device_get_nameunit(dev), txq->ift_id);
4632 
4633 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4634 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4635 		if (err) {
4636 			/* XXX free any allocated rings */
4637 			device_printf(dev, "Unable to allocate buf_ring\n");
4638 			goto err_tx_desc;
4639 		}
4640 	}
4641 
4642 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4643 		/* Set up some basics */
4644 
4645 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4646 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4647 			err = ENOMEM;
4648 			goto err_tx_desc;
4649 		}
4650 
4651 		rxq->ifr_ifdi = ifdip;
4652 		/* XXX this needs to be changed if #rx queues != #tx queues */
4653 		rxq->ifr_ntxqirq = 1;
4654 		rxq->ifr_txqid[0] = i;
4655 		for (j = 0; j < nrxqs; j++, ifdip++) {
4656 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4657 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4658 				err = ENOMEM;
4659 				goto err_tx_desc;
4660 			}
4661 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4662 		}
4663 		rxq->ifr_ctx = ctx;
4664 		rxq->ifr_id = i;
4665 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4666 			rxq->ifr_fl_offset = 1;
4667 		} else {
4668 			rxq->ifr_fl_offset = 0;
4669 		}
4670 		rxq->ifr_nfl = nfree_lists;
4671 		if (!(fl =
4672 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4673 			device_printf(dev, "Unable to allocate free list memory\n");
4674 			err = ENOMEM;
4675 			goto err_tx_desc;
4676 		}
4677 		rxq->ifr_fl = fl;
4678 		for (j = 0; j < nfree_lists; j++) {
4679 			fl[j].ifl_rxq = rxq;
4680 			fl[j].ifl_id = j;
4681 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4682 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4683 		}
4684         /* Allocate receive buffers for the ring*/
4685 		if (iflib_rxsd_alloc(rxq)) {
4686 			device_printf(dev,
4687 			    "Critical Failure setting up receive buffers\n");
4688 			err = ENOMEM;
4689 			goto err_rx_desc;
4690 		}
4691 	}
4692 
4693 	/* TXQs */
4694 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4695 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4696 	for (i = 0; i < ntxqsets; i++) {
4697 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4698 
4699 		for (j = 0; j < ntxqs; j++, di++) {
4700 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4701 			paddrs[i*ntxqs + j] = di->idi_paddr;
4702 		}
4703 	}
4704 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4705 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4706 		iflib_tx_structures_free(ctx);
4707 		free(vaddrs, M_IFLIB);
4708 		free(paddrs, M_IFLIB);
4709 		goto err_rx_desc;
4710 	}
4711 	free(vaddrs, M_IFLIB);
4712 	free(paddrs, M_IFLIB);
4713 
4714 	/* RXQs */
4715 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4716 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4717 	for (i = 0; i < nrxqsets; i++) {
4718 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4719 
4720 		for (j = 0; j < nrxqs; j++, di++) {
4721 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4722 			paddrs[i*nrxqs + j] = di->idi_paddr;
4723 		}
4724 	}
4725 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4726 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4727 		iflib_tx_structures_free(ctx);
4728 		free(vaddrs, M_IFLIB);
4729 		free(paddrs, M_IFLIB);
4730 		goto err_rx_desc;
4731 	}
4732 	free(vaddrs, M_IFLIB);
4733 	free(paddrs, M_IFLIB);
4734 
4735 	return (0);
4736 
4737 /* XXX handle allocation failure changes */
4738 err_rx_desc:
4739 err_tx_desc:
4740 	if (ctx->ifc_rxqs != NULL)
4741 		free(ctx->ifc_rxqs, M_IFLIB);
4742 	ctx->ifc_rxqs = NULL;
4743 	if (ctx->ifc_txqs != NULL)
4744 		free(ctx->ifc_txqs, M_IFLIB);
4745 	ctx->ifc_txqs = NULL;
4746 rx_fail:
4747 	if (brscp != NULL)
4748 		free(brscp, M_IFLIB);
4749 	if (rxq != NULL)
4750 		free(rxq, M_IFLIB);
4751 	if (txq != NULL)
4752 		free(txq, M_IFLIB);
4753 fail:
4754 	return (err);
4755 }
4756 
4757 static int
4758 iflib_tx_structures_setup(if_ctx_t ctx)
4759 {
4760 	iflib_txq_t txq = ctx->ifc_txqs;
4761 	int i;
4762 
4763 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4764 		iflib_txq_setup(txq);
4765 
4766 	return (0);
4767 }
4768 
4769 static void
4770 iflib_tx_structures_free(if_ctx_t ctx)
4771 {
4772 	iflib_txq_t txq = ctx->ifc_txqs;
4773 	int i, j;
4774 
4775 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4776 		iflib_txq_destroy(txq);
4777 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4778 			iflib_dma_free(&txq->ift_ifdi[j]);
4779 	}
4780 	free(ctx->ifc_txqs, M_IFLIB);
4781 	ctx->ifc_txqs = NULL;
4782 	IFDI_QUEUES_FREE(ctx);
4783 }
4784 
4785 /*********************************************************************
4786  *
4787  *  Initialize all receive rings.
4788  *
4789  **********************************************************************/
4790 static int
4791 iflib_rx_structures_setup(if_ctx_t ctx)
4792 {
4793 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4794 	int q;
4795 #if defined(INET6) || defined(INET)
4796 	int i, err;
4797 #endif
4798 
4799 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4800 #if defined(INET6) || defined(INET)
4801 		tcp_lro_free(&rxq->ifr_lc);
4802 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4803 		    TCP_LRO_ENTRIES, min(1024,
4804 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4805 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4806 			goto fail;
4807 		}
4808 		rxq->ifr_lro_enabled = TRUE;
4809 #endif
4810 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4811 	}
4812 	return (0);
4813 #if defined(INET6) || defined(INET)
4814 fail:
4815 	/*
4816 	 * Free RX software descriptors allocated so far, we will only handle
4817 	 * the rings that completed, the failing case will have
4818 	 * cleaned up for itself. 'q' failed, so its the terminus.
4819 	 */
4820 	rxq = ctx->ifc_rxqs;
4821 	for (i = 0; i < q; ++i, rxq++) {
4822 		iflib_rx_sds_free(rxq);
4823 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4824 	}
4825 	return (err);
4826 #endif
4827 }
4828 
4829 /*********************************************************************
4830  *
4831  *  Free all receive rings.
4832  *
4833  **********************************************************************/
4834 static void
4835 iflib_rx_structures_free(if_ctx_t ctx)
4836 {
4837 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4838 
4839 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4840 		iflib_rx_sds_free(rxq);
4841 	}
4842 }
4843 
4844 static int
4845 iflib_qset_structures_setup(if_ctx_t ctx)
4846 {
4847 	int err;
4848 
4849 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4850 		return (err);
4851 
4852 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4853 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4854 		iflib_tx_structures_free(ctx);
4855 		iflib_rx_structures_free(ctx);
4856 	}
4857 	return (err);
4858 }
4859 
4860 int
4861 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4862 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4863 {
4864 
4865 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4866 }
4867 
4868 static int
4869 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4870 {
4871 	int i, cpuid, eqid, count;
4872 
4873 	CPU_COPY(&ctx->ifc_cpus, cpus);
4874 	count = CPU_COUNT(&ctx->ifc_cpus);
4875 	eqid = qid % count;
4876 	/* clear up to the qid'th bit */
4877 	for (i = 0; i < eqid; i++) {
4878 		cpuid = CPU_FFS(cpus);
4879 		MPASS(cpuid != 0);
4880 		CPU_CLR(cpuid-1, cpus);
4881 	}
4882 	cpuid = CPU_FFS(cpus);
4883 	MPASS(cpuid != 0);
4884 	return (cpuid-1);
4885 }
4886 
4887 int
4888 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4889 						iflib_intr_type_t type, driver_filter_t *filter,
4890 						void *filter_arg, int qid, char *name)
4891 {
4892 	struct grouptask *gtask;
4893 	struct taskqgroup *tqg;
4894 	iflib_filter_info_t info;
4895 	cpuset_t cpus;
4896 	gtask_fn_t *fn;
4897 	int tqrid, err, cpuid;
4898 	driver_filter_t *intr_fast;
4899 	void *q;
4900 
4901 	info = &ctx->ifc_filter_info;
4902 	tqrid = rid;
4903 
4904 	switch (type) {
4905 	/* XXX merge tx/rx for netmap? */
4906 	case IFLIB_INTR_TX:
4907 		q = &ctx->ifc_txqs[qid];
4908 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4909 		gtask = &ctx->ifc_txqs[qid].ift_task;
4910 		tqg = qgroup_if_io_tqg;
4911 		fn = _task_fn_tx;
4912 		intr_fast = iflib_fast_intr;
4913 		GROUPTASK_INIT(gtask, 0, fn, q);
4914 		break;
4915 	case IFLIB_INTR_RX:
4916 		q = &ctx->ifc_rxqs[qid];
4917 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4918 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4919 		tqg = qgroup_if_io_tqg;
4920 		fn = _task_fn_rx;
4921 		intr_fast = iflib_fast_intr;
4922 		GROUPTASK_INIT(gtask, 0, fn, q);
4923 		break;
4924 	case IFLIB_INTR_RXTX:
4925 		q = &ctx->ifc_rxqs[qid];
4926 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4927 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4928 		tqg = qgroup_if_io_tqg;
4929 		fn = _task_fn_rx;
4930 		intr_fast = iflib_fast_intr_rxtx;
4931 		GROUPTASK_INIT(gtask, 0, fn, q);
4932 		break;
4933 	case IFLIB_INTR_ADMIN:
4934 		q = ctx;
4935 		tqrid = -1;
4936 		info = &ctx->ifc_filter_info;
4937 		gtask = &ctx->ifc_admin_task;
4938 		tqg = qgroup_if_config_tqg;
4939 		fn = _task_fn_admin;
4940 		intr_fast = iflib_fast_intr_ctx;
4941 		break;
4942 	default:
4943 		panic("unknown net intr type");
4944 	}
4945 
4946 	info->ifi_filter = filter;
4947 	info->ifi_filter_arg = filter_arg;
4948 	info->ifi_task = gtask;
4949 	info->ifi_ctx = q;
4950 
4951 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4952 	if (err != 0) {
4953 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4954 		return (err);
4955 	}
4956 	if (type == IFLIB_INTR_ADMIN)
4957 		return (0);
4958 
4959 	if (tqrid != -1) {
4960 		cpuid = find_nth(ctx, &cpus, qid);
4961 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4962 	} else {
4963 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4964 	}
4965 
4966 	return (0);
4967 }
4968 
4969 void
4970 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4971 {
4972 	struct grouptask *gtask;
4973 	struct taskqgroup *tqg;
4974 	gtask_fn_t *fn;
4975 	void *q;
4976 
4977 	switch (type) {
4978 	case IFLIB_INTR_TX:
4979 		q = &ctx->ifc_txqs[qid];
4980 		gtask = &ctx->ifc_txqs[qid].ift_task;
4981 		tqg = qgroup_if_io_tqg;
4982 		fn = _task_fn_tx;
4983 		break;
4984 	case IFLIB_INTR_RX:
4985 		q = &ctx->ifc_rxqs[qid];
4986 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4987 		tqg = qgroup_if_io_tqg;
4988 		fn = _task_fn_rx;
4989 		break;
4990 	case IFLIB_INTR_IOV:
4991 		q = ctx;
4992 		gtask = &ctx->ifc_vflr_task;
4993 		tqg = qgroup_if_config_tqg;
4994 		rid = -1;
4995 		fn = _task_fn_iov;
4996 		break;
4997 	default:
4998 		panic("unknown net intr type");
4999 	}
5000 	GROUPTASK_INIT(gtask, 0, fn, q);
5001 	taskqgroup_attach(tqg, gtask, q, rid, name);
5002 }
5003 
5004 void
5005 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5006 {
5007 	if (irq->ii_tag)
5008 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5009 
5010 	if (irq->ii_res)
5011 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5012 }
5013 
5014 static int
5015 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5016 {
5017 	iflib_txq_t txq = ctx->ifc_txqs;
5018 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5019 	if_irq_t irq = &ctx->ifc_legacy_irq;
5020 	iflib_filter_info_t info;
5021 	struct grouptask *gtask;
5022 	struct taskqgroup *tqg;
5023 	gtask_fn_t *fn;
5024 	int tqrid;
5025 	void *q;
5026 	int err;
5027 
5028 	q = &ctx->ifc_rxqs[0];
5029 	info = &rxq[0].ifr_filter_info;
5030 	gtask = &rxq[0].ifr_task;
5031 	tqg = qgroup_if_io_tqg;
5032 	tqrid = irq->ii_rid = *rid;
5033 	fn = _task_fn_rx;
5034 
5035 	ctx->ifc_flags |= IFC_LEGACY;
5036 	info->ifi_filter = filter;
5037 	info->ifi_filter_arg = filter_arg;
5038 	info->ifi_task = gtask;
5039 	info->ifi_ctx = ctx;
5040 
5041 	/* We allocate a single interrupt resource */
5042 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5043 		return (err);
5044 	GROUPTASK_INIT(gtask, 0, fn, q);
5045 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5046 
5047 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5048 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5049 	return (0);
5050 }
5051 
5052 void
5053 iflib_led_create(if_ctx_t ctx)
5054 {
5055 
5056 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5057 	    device_get_nameunit(ctx->ifc_dev));
5058 }
5059 
5060 void
5061 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5062 {
5063 
5064 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5065 }
5066 
5067 void
5068 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5069 {
5070 
5071 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5072 }
5073 
5074 void
5075 iflib_admin_intr_deferred(if_ctx_t ctx)
5076 {
5077 #ifdef INVARIANTS
5078 	struct grouptask *gtask;
5079 
5080 	gtask = &ctx->ifc_admin_task;
5081 	MPASS(gtask->gt_taskqueue != NULL);
5082 #endif
5083 
5084 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5085 }
5086 
5087 void
5088 iflib_iov_intr_deferred(if_ctx_t ctx)
5089 {
5090 
5091 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5092 }
5093 
5094 void
5095 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5096 {
5097 
5098 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5099 }
5100 
5101 void
5102 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5103 	char *name)
5104 {
5105 
5106 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5107 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5108 }
5109 
5110 void
5111 iflib_config_gtask_deinit(struct grouptask *gtask)
5112 {
5113 
5114 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5115 }
5116 
5117 void
5118 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5119 {
5120 	if_t ifp = ctx->ifc_ifp;
5121 	iflib_txq_t txq = ctx->ifc_txqs;
5122 
5123 	if_setbaudrate(ifp, baudrate);
5124 	if (baudrate >= IF_Gbps(10))
5125 		ctx->ifc_flags |= IFC_PREFETCH;
5126 
5127 	/* If link down, disable watchdog */
5128 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5129 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5130 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5131 	}
5132 	ctx->ifc_link_state = link_state;
5133 	if_link_state_change(ifp, link_state);
5134 }
5135 
5136 static int
5137 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5138 {
5139 	int credits;
5140 #ifdef INVARIANTS
5141 	int credits_pre = txq->ift_cidx_processed;
5142 #endif
5143 
5144 	if (ctx->isc_txd_credits_update == NULL)
5145 		return (0);
5146 
5147 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5148 		return (0);
5149 
5150 	txq->ift_processed += credits;
5151 	txq->ift_cidx_processed += credits;
5152 
5153 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5154 	if (txq->ift_cidx_processed >= txq->ift_size)
5155 		txq->ift_cidx_processed -= txq->ift_size;
5156 	return (credits);
5157 }
5158 
5159 static int
5160 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5161 {
5162 
5163 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5164 	    budget));
5165 }
5166 
5167 void
5168 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5169 	const char *description, if_int_delay_info_t info,
5170 	int offset, int value)
5171 {
5172 	info->iidi_ctx = ctx;
5173 	info->iidi_offset = offset;
5174 	info->iidi_value = value;
5175 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5176 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5177 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5178 	    info, 0, iflib_sysctl_int_delay, "I", description);
5179 }
5180 
5181 struct mtx *
5182 iflib_ctx_lock_get(if_ctx_t ctx)
5183 {
5184 
5185 	return (&ctx->ifc_mtx);
5186 }
5187 
5188 static int
5189 iflib_msix_init(if_ctx_t ctx)
5190 {
5191 	device_t dev = ctx->ifc_dev;
5192 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5193 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5194 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5195 	int iflib_num_tx_queues, iflib_num_rx_queues;
5196 	int err, admincnt, bar;
5197 
5198 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5199 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5200 
5201 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5202 
5203 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5204 	admincnt = sctx->isc_admin_intrcnt;
5205 	/* Override by tuneable */
5206 	if (scctx->isc_disable_msix)
5207 		goto msi;
5208 
5209 	/*
5210 	** When used in a virtualized environment
5211 	** PCI BUSMASTER capability may not be set
5212 	** so explicity set it here and rewrite
5213 	** the ENABLE in the MSIX control register
5214 	** at this point to cause the host to
5215 	** successfully initialize us.
5216 	*/
5217 	{
5218 		int msix_ctrl, rid;
5219 
5220  		pci_enable_busmaster(dev);
5221 		rid = 0;
5222 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5223 			rid += PCIR_MSIX_CTRL;
5224 			msix_ctrl = pci_read_config(dev, rid, 2);
5225 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5226 			pci_write_config(dev, rid, msix_ctrl, 2);
5227 		} else {
5228 			device_printf(dev, "PCIY_MSIX capability not found; "
5229 			                   "or rid %d == 0.\n", rid);
5230 			goto msi;
5231 		}
5232 	}
5233 
5234 	/*
5235 	 * bar == -1 => "trust me I know what I'm doing"
5236 	 * Some drivers are for hardware that is so shoddily
5237 	 * documented that no one knows which bars are which
5238 	 * so the developer has to map all bars. This hack
5239 	 * allows shoddy garbage to use msix in this framework.
5240 	 */
5241 	if (bar != -1) {
5242 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5243 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5244 		if (ctx->ifc_msix_mem == NULL) {
5245 			/* May not be enabled */
5246 			device_printf(dev, "Unable to map MSIX table \n");
5247 			goto msi;
5248 		}
5249 	}
5250 	/* First try MSI/X */
5251 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5252 		device_printf(dev, "System has MSIX disabled \n");
5253 		bus_release_resource(dev, SYS_RES_MEMORY,
5254 		    bar, ctx->ifc_msix_mem);
5255 		ctx->ifc_msix_mem = NULL;
5256 		goto msi;
5257 	}
5258 #if IFLIB_DEBUG
5259 	/* use only 1 qset in debug mode */
5260 	queuemsgs = min(msgs - admincnt, 1);
5261 #else
5262 	queuemsgs = msgs - admincnt;
5263 #endif
5264 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5265 #ifdef RSS
5266 		queues = imin(queuemsgs, rss_getnumbuckets());
5267 #else
5268 		queues = queuemsgs;
5269 #endif
5270 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5271 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5272 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5273 	} else {
5274 		device_printf(dev, "Unable to fetch CPU list\n");
5275 		/* Figure out a reasonable auto config value */
5276 		queues = min(queuemsgs, mp_ncpus);
5277 	}
5278 #ifdef  RSS
5279 	/* If we're doing RSS, clamp at the number of RSS buckets */
5280 	if (queues > rss_getnumbuckets())
5281 		queues = rss_getnumbuckets();
5282 #endif
5283 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5284 		rx_queues = iflib_num_rx_queues;
5285 	else
5286 		rx_queues = queues;
5287 	/*
5288 	 * We want this to be all logical CPUs by default
5289 	 */
5290 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5291 		tx_queues = iflib_num_tx_queues;
5292 	else
5293 		tx_queues = mp_ncpus;
5294 
5295 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5296 #ifdef INVARIANTS
5297 		if (tx_queues != rx_queues)
5298 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5299 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5300 #endif
5301 		tx_queues = min(rx_queues, tx_queues);
5302 		rx_queues = min(rx_queues, tx_queues);
5303 	}
5304 
5305 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5306 
5307 	vectors = rx_queues + admincnt;
5308 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5309 		device_printf(dev,
5310 					  "Using MSIX interrupts with %d vectors\n", vectors);
5311 		scctx->isc_vectors = vectors;
5312 		scctx->isc_nrxqsets = rx_queues;
5313 		scctx->isc_ntxqsets = tx_queues;
5314 		scctx->isc_intr = IFLIB_INTR_MSIX;
5315 
5316 		return (vectors);
5317 	} else {
5318 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5319 	}
5320 msi:
5321 	vectors = pci_msi_count(dev);
5322 	scctx->isc_nrxqsets = 1;
5323 	scctx->isc_ntxqsets = 1;
5324 	scctx->isc_vectors = vectors;
5325 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5326 		device_printf(dev,"Using an MSI interrupt\n");
5327 		scctx->isc_intr = IFLIB_INTR_MSI;
5328 	} else {
5329 		device_printf(dev,"Using a Legacy interrupt\n");
5330 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5331 	}
5332 
5333 	return (vectors);
5334 }
5335 
5336 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5337 
5338 static int
5339 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5340 {
5341 	int rc;
5342 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5343 	struct sbuf *sb;
5344 	char *ring_state = "UNKNOWN";
5345 
5346 	/* XXX needed ? */
5347 	rc = sysctl_wire_old_buffer(req, 0);
5348 	MPASS(rc == 0);
5349 	if (rc != 0)
5350 		return (rc);
5351 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5352 	MPASS(sb != NULL);
5353 	if (sb == NULL)
5354 		return (ENOMEM);
5355 	if (state[3] <= 3)
5356 		ring_state = ring_states[state[3]];
5357 
5358 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5359 		    state[0], state[1], state[2], ring_state);
5360 	rc = sbuf_finish(sb);
5361 	sbuf_delete(sb);
5362         return(rc);
5363 }
5364 
5365 enum iflib_ndesc_handler {
5366 	IFLIB_NTXD_HANDLER,
5367 	IFLIB_NRXD_HANDLER,
5368 };
5369 
5370 static int
5371 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5372 {
5373 	if_ctx_t ctx = (void *)arg1;
5374 	enum iflib_ndesc_handler type = arg2;
5375 	char buf[256] = {0};
5376 	qidx_t *ndesc;
5377 	char *p, *next;
5378 	int nqs, rc, i;
5379 
5380 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5381 
5382 	nqs = 8;
5383 	switch(type) {
5384 	case IFLIB_NTXD_HANDLER:
5385 		ndesc = ctx->ifc_sysctl_ntxds;
5386 		if (ctx->ifc_sctx)
5387 			nqs = ctx->ifc_sctx->isc_ntxqs;
5388 		break;
5389 	case IFLIB_NRXD_HANDLER:
5390 		ndesc = ctx->ifc_sysctl_nrxds;
5391 		if (ctx->ifc_sctx)
5392 			nqs = ctx->ifc_sctx->isc_nrxqs;
5393 		break;
5394 	}
5395 	if (nqs == 0)
5396 		nqs = 8;
5397 
5398 	for (i=0; i<8; i++) {
5399 		if (i >= nqs)
5400 			break;
5401 		if (i)
5402 			strcat(buf, ",");
5403 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5404 	}
5405 
5406 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5407 	if (rc || req->newptr == NULL)
5408 		return rc;
5409 
5410 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5411 	    i++, p = strsep(&next, " ,")) {
5412 		ndesc[i] = strtoul(p, NULL, 10);
5413 	}
5414 
5415 	return(rc);
5416 }
5417 
5418 #define NAME_BUFLEN 32
5419 static void
5420 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5421 {
5422         device_t dev = iflib_get_dev(ctx);
5423 	struct sysctl_oid_list *child, *oid_list;
5424 	struct sysctl_ctx_list *ctx_list;
5425 	struct sysctl_oid *node;
5426 
5427 	ctx_list = device_get_sysctl_ctx(dev);
5428 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5429 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5430 						      CTLFLAG_RD, NULL, "IFLIB fields");
5431 	oid_list = SYSCTL_CHILDREN(node);
5432 
5433 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5434 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5435 		       "driver version");
5436 
5437 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5438 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5439 			"# of txqs to use, 0 => use default #");
5440 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5441 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5442 			"# of rxqs to use, 0 => use default #");
5443 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5444 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5445                        "permit #txq != #rxq");
5446        SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5447                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5448                       "disable MSIX (default 0)");
5449 
5450 	/* XXX change for per-queue sizes */
5451 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5452 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5453                        mp_ndesc_handler, "A",
5454                        "list of # of tx descriptors to use, 0 = use default #");
5455 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5456 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5457                        mp_ndesc_handler, "A",
5458                        "list of # of rx descriptors to use, 0 = use default #");
5459 }
5460 
5461 static void
5462 iflib_add_device_sysctl_post(if_ctx_t ctx)
5463 {
5464 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5465 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5466         device_t dev = iflib_get_dev(ctx);
5467 	struct sysctl_oid_list *child;
5468 	struct sysctl_ctx_list *ctx_list;
5469 	iflib_fl_t fl;
5470 	iflib_txq_t txq;
5471 	iflib_rxq_t rxq;
5472 	int i, j;
5473 	char namebuf[NAME_BUFLEN];
5474 	char *qfmt;
5475 	struct sysctl_oid *queue_node, *fl_node, *node;
5476 	struct sysctl_oid_list *queue_list, *fl_list;
5477 	ctx_list = device_get_sysctl_ctx(dev);
5478 
5479 	node = ctx->ifc_sysctl_node;
5480 	child = SYSCTL_CHILDREN(node);
5481 
5482 	if (scctx->isc_ntxqsets > 100)
5483 		qfmt = "txq%03d";
5484 	else if (scctx->isc_ntxqsets > 10)
5485 		qfmt = "txq%02d";
5486 	else
5487 		qfmt = "txq%d";
5488 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5489 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5490 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5491 					     CTLFLAG_RD, NULL, "Queue Name");
5492 		queue_list = SYSCTL_CHILDREN(queue_node);
5493 #if MEMORY_LOGGING
5494 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5495 				CTLFLAG_RD,
5496 				&txq->ift_dequeued, "total mbufs freed");
5497 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5498 				CTLFLAG_RD,
5499 				&txq->ift_enqueued, "total mbufs enqueued");
5500 #endif
5501 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5502 				   CTLFLAG_RD,
5503 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5504 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5505 				   CTLFLAG_RD,
5506 				   &txq->ift_pullups, "# of times m_pullup was called");
5507 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5508 				   CTLFLAG_RD,
5509 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5510 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5511 				   CTLFLAG_RD,
5512 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5513 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5514 				   CTLFLAG_RD,
5515 				   &txq->ift_map_failed, "# of times dma map failed");
5516 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5517 				   CTLFLAG_RD,
5518 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5519 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5520 				   CTLFLAG_RD,
5521 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5522 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5523 				   CTLFLAG_RD,
5524 				   &txq->ift_pidx, 1, "Producer Index");
5525 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5526 				   CTLFLAG_RD,
5527 				   &txq->ift_cidx, 1, "Consumer Index");
5528 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5529 				   CTLFLAG_RD,
5530 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5531 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5532 				   CTLFLAG_RD,
5533 				   &txq->ift_in_use, 1, "descriptors in use");
5534 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5535 				   CTLFLAG_RD,
5536 				   &txq->ift_processed, "descriptors procesed for clean");
5537 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5538 				   CTLFLAG_RD,
5539 				   &txq->ift_cleaned, "total cleaned");
5540 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5541 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5542 				0, mp_ring_state_handler, "A", "soft ring state");
5543 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5544 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5545 				       "# of enqueues to the mp_ring for this queue");
5546 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5547 				       CTLFLAG_RD, &txq->ift_br->drops,
5548 				       "# of drops in the mp_ring for this queue");
5549 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5550 				       CTLFLAG_RD, &txq->ift_br->starts,
5551 				       "# of normal consumer starts in the mp_ring for this queue");
5552 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5553 				       CTLFLAG_RD, &txq->ift_br->stalls,
5554 					       "# of consumer stalls in the mp_ring for this queue");
5555 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5556 			       CTLFLAG_RD, &txq->ift_br->restarts,
5557 				       "# of consumer restarts in the mp_ring for this queue");
5558 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5559 				       CTLFLAG_RD, &txq->ift_br->abdications,
5560 				       "# of consumer abdications in the mp_ring for this queue");
5561 	}
5562 
5563 	if (scctx->isc_nrxqsets > 100)
5564 		qfmt = "rxq%03d";
5565 	else if (scctx->isc_nrxqsets > 10)
5566 		qfmt = "rxq%02d";
5567 	else
5568 		qfmt = "rxq%d";
5569 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5570 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5571 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5572 					     CTLFLAG_RD, NULL, "Queue Name");
5573 		queue_list = SYSCTL_CHILDREN(queue_node);
5574 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5575 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5576 				       CTLFLAG_RD,
5577 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5578 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5579 				       CTLFLAG_RD,
5580 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5581 		}
5582 
5583 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5584 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5585 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5586 						     CTLFLAG_RD, NULL, "freelist Name");
5587 			fl_list = SYSCTL_CHILDREN(fl_node);
5588 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5589 				       CTLFLAG_RD,
5590 				       &fl->ifl_pidx, 1, "Producer Index");
5591 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5592 				       CTLFLAG_RD,
5593 				       &fl->ifl_cidx, 1, "Consumer Index");
5594 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5595 				       CTLFLAG_RD,
5596 				       &fl->ifl_credits, 1, "credits available");
5597 #if MEMORY_LOGGING
5598 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5599 					CTLFLAG_RD,
5600 					&fl->ifl_m_enqueued, "mbufs allocated");
5601 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5602 					CTLFLAG_RD,
5603 					&fl->ifl_m_dequeued, "mbufs freed");
5604 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5605 					CTLFLAG_RD,
5606 					&fl->ifl_cl_enqueued, "clusters allocated");
5607 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5608 					CTLFLAG_RD,
5609 					&fl->ifl_cl_dequeued, "clusters freed");
5610 #endif
5611 
5612 		}
5613 	}
5614 
5615 }
5616 
5617 #ifndef __NO_STRICT_ALIGNMENT
5618 static struct mbuf *
5619 iflib_fixup_rx(struct mbuf *m)
5620 {
5621 	struct mbuf *n;
5622 
5623 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5624 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5625 		m->m_data += ETHER_HDR_LEN;
5626 		n = m;
5627 	} else {
5628 		MGETHDR(n, M_NOWAIT, MT_DATA);
5629 		if (n == NULL) {
5630 			m_freem(m);
5631 			return (NULL);
5632 		}
5633 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5634 		m->m_data += ETHER_HDR_LEN;
5635 		m->m_len -= ETHER_HDR_LEN;
5636 		n->m_len = ETHER_HDR_LEN;
5637 		M_MOVE_PKTHDR(n, m);
5638 		n->m_next = m;
5639 	}
5640 	return (n);
5641 }
5642 #endif
5643