xref: /freebsd/sys/net/iflib.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of
98  * iflib's software descriptor references
99  */
100 #define MEMORY_LOGGING 0
101 /*
102  * Enable mbuf vectors for compressing long mbuf chains
103  */
104 
105 /*
106  * NB:
107  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108  *   we prefetch needs to be determined by the time spent in m_free vis a vis
109  *   the cost of a prefetch. This will of course vary based on the workload:
110  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111  *        is quite expensive, thus suggesting very little prefetch.
112  *      - small packet forwarding which is just returning a single mbuf to
113  *        UMA will typically be very fast vis a vis the cost of a memory
114  *        access.
115  */
116 
117 
118 /*
119  * File organization:
120  *  - private structures
121  *  - iflib private utility functions
122  *  - ifnet functions
123  *  - vlan registry and other exported functions
124  *  - iflib public core functions
125  *
126  *
127  */
128 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
129 
130 struct iflib_txq;
131 typedef struct iflib_txq *iflib_txq_t;
132 struct iflib_rxq;
133 typedef struct iflib_rxq *iflib_rxq_t;
134 struct iflib_fl;
135 typedef struct iflib_fl *iflib_fl_t;
136 
137 struct iflib_ctx;
138 
139 typedef struct iflib_filter_info {
140 	driver_filter_t *ifi_filter;
141 	void *ifi_filter_arg;
142 	struct grouptask *ifi_task;
143 	void *ifi_ctx;
144 } *iflib_filter_info_t;
145 
146 struct iflib_ctx {
147 	KOBJ_FIELDS;
148    /*
149    * Pointer to hardware driver's softc
150    */
151 	void *ifc_softc;
152 	device_t ifc_dev;
153 	if_t ifc_ifp;
154 
155 	cpuset_t ifc_cpus;
156 	if_shared_ctx_t ifc_sctx;
157 	struct if_softc_ctx ifc_softc_ctx;
158 
159 	struct mtx ifc_mtx;
160 
161 	uint16_t ifc_nhwtxqs;
162 	uint16_t ifc_nhwrxqs;
163 
164 	iflib_txq_t ifc_txqs;
165 	iflib_rxq_t ifc_rxqs;
166 	uint32_t ifc_if_flags;
167 	uint32_t ifc_flags;
168 	uint32_t ifc_max_fl_buf_size;
169 	int ifc_in_detach;
170 
171 	int ifc_link_state;
172 	int ifc_link_irq;
173 	int ifc_watchdog_events;
174 	struct cdev *ifc_led_dev;
175 	struct resource *ifc_msix_mem;
176 
177 	struct if_irq ifc_legacy_irq;
178 	struct grouptask ifc_admin_task;
179 	struct grouptask ifc_vflr_task;
180 	struct iflib_filter_info ifc_filter_info;
181 	struct ifmedia	ifc_media;
182 
183 	struct sysctl_oid *ifc_sysctl_node;
184 	uint16_t ifc_sysctl_ntxqs;
185 	uint16_t ifc_sysctl_nrxqs;
186 	uint16_t ifc_sysctl_qs_eq_override;
187 
188 	qidx_t ifc_sysctl_ntxds[8];
189 	qidx_t ifc_sysctl_nrxds[8];
190 	struct if_txrx ifc_txrx;
191 #define isc_txd_encap  ifc_txrx.ift_txd_encap
192 #define isc_txd_flush  ifc_txrx.ift_txd_flush
193 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
194 #define isc_rxd_available ifc_txrx.ift_rxd_available
195 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
197 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
198 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
201 	eventhandler_tag ifc_vlan_attach_event;
202 	eventhandler_tag ifc_vlan_detach_event;
203 	uint8_t ifc_mac[ETHER_ADDR_LEN];
204 	char ifc_mtx_name[16];
205 };
206 
207 
208 void *
209 iflib_get_softc(if_ctx_t ctx)
210 {
211 
212 	return (ctx->ifc_softc);
213 }
214 
215 device_t
216 iflib_get_dev(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_dev);
220 }
221 
222 if_t
223 iflib_get_ifp(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_ifp);
227 }
228 
229 struct ifmedia *
230 iflib_get_media(if_ctx_t ctx)
231 {
232 
233 	return (&ctx->ifc_media);
234 }
235 
236 void
237 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
238 {
239 
240 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
241 }
242 
243 if_softc_ctx_t
244 iflib_get_softc_ctx(if_ctx_t ctx)
245 {
246 
247 	return (&ctx->ifc_softc_ctx);
248 }
249 
250 if_shared_ctx_t
251 iflib_get_sctx(if_ctx_t ctx)
252 {
253 
254 	return (ctx->ifc_sctx);
255 }
256 
257 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
258 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
259 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
260 
261 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
262 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
263 
264 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
265 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
266 #define RX_SW_DESC_INUSE        (1 << 3)
267 #define TX_SW_DESC_MAPPED       (1 << 4)
268 
269 typedef struct iflib_sw_rx_desc_array {
270 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
271 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
272 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
273 	uint8_t		*ifsd_flags;
274 } iflib_rxsd_array_t;
275 
276 typedef struct iflib_sw_tx_desc_array {
277 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
279 	uint8_t		*ifsd_flags;
280 } if_txsd_vec_t;
281 
282 
283 /* magic number that should be high enough for any hardware */
284 #define IFLIB_MAX_TX_SEGS		128
285 #define IFLIB_MAX_RX_SEGS		32
286 #define IFLIB_RX_COPY_THRESH		128
287 #define IFLIB_MAX_RX_REFRESH		32
288 /* The minimum descriptors per second before we start coalescing */
289 #define IFLIB_MIN_DESC_SEC		16384
290 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
291 #define IFLIB_QUEUE_IDLE		0
292 #define IFLIB_QUEUE_HUNG		1
293 #define IFLIB_QUEUE_WORKING		2
294 /* maximum number of txqs that can share an rx interrupt */
295 #define IFLIB_MAX_TX_SHARED_INTR	4
296 
297 /* this should really scale with ring size - this is a fairly arbitrary value */
298 #define TX_BATCH_SIZE			32
299 
300 #define IFLIB_RESTART_BUDGET		8
301 
302 #define	IFC_LEGACY		0x001
303 #define	IFC_QFLUSH		0x002
304 #define	IFC_MULTISEG		0x004
305 #define	IFC_DMAR		0x008
306 #define	IFC_SC_ALLOCATED	0x010
307 #define	IFC_INIT_DONE		0x020
308 #define	IFC_PREFETCH		0x040
309 #define	IFC_DO_RESET		0x080
310 #define	IFC_CHECK_HUNG		0x100
311 
312 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
313 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
314 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
315 struct iflib_txq {
316 	qidx_t		ift_in_use;
317 	qidx_t		ift_cidx;
318 	qidx_t		ift_cidx_processed;
319 	qidx_t		ift_pidx;
320 	uint8_t		ift_gen;
321 	uint8_t		ift_br_offset;
322 	uint16_t	ift_npending;
323 	uint16_t	ift_db_pending;
324 	uint16_t	ift_rs_pending;
325 	/* implicit pad */
326 	uint8_t		ift_txd_size[8];
327 	uint64_t	ift_processed;
328 	uint64_t	ift_cleaned;
329 	uint64_t	ift_cleaned_prev;
330 #if MEMORY_LOGGING
331 	uint64_t	ift_enqueued;
332 	uint64_t	ift_dequeued;
333 #endif
334 	uint64_t	ift_no_tx_dma_setup;
335 	uint64_t	ift_no_desc_avail;
336 	uint64_t	ift_mbuf_defrag_failed;
337 	uint64_t	ift_mbuf_defrag;
338 	uint64_t	ift_map_failed;
339 	uint64_t	ift_txd_encap_efbig;
340 	uint64_t	ift_pullups;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t		ift_desc_tag;
359 	bus_dma_tag_t		ift_tso_desc_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define MTX_NAME_LEN 16
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	char                    ift_db_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t           ifl_desc_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	/* If there is a separate completion queue -
425 	 * these are the cq cidx and pidx. Otherwise
426 	 * these are unused.
427 	 */
428 	qidx_t		ifr_size;
429 	qidx_t		ifr_cq_cidx;
430 	qidx_t		ifr_cq_pidx;
431 	uint8_t		ifr_cq_gen;
432 	uint8_t		ifr_fl_offset;
433 
434 	if_ctx_t	ifr_ctx;
435 	iflib_fl_t	ifr_fl;
436 	uint64_t	ifr_rx_irq;
437 	uint16_t	ifr_id;
438 	uint8_t		ifr_lro_enabled;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	struct lro_ctrl			ifr_lc;
443 	struct grouptask        ifr_task;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	struct mbuf **ifsd_m;
457 	iflib_fl_t ifsd_fl;
458 	qidx_t ifsd_cidx;
459 } *if_rxsd_t;
460 
461 /* multiple of word size */
462 #ifdef __LP64__
463 #define PKT_INFO_SIZE	6
464 #define RXD_INFO_SIZE	5
465 #define PKT_TYPE uint64_t
466 #else
467 #define PKT_INFO_SIZE	11
468 #define RXD_INFO_SIZE	8
469 #define PKT_TYPE uint32_t
470 #endif
471 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
472 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
473 
474 typedef struct if_pkt_info_pad {
475 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
476 } *if_pkt_info_pad_t;
477 typedef struct if_rxd_info_pad {
478 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
479 } *if_rxd_info_pad_t;
480 
481 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
482 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
483 
484 
485 static inline void
486 pkt_info_zero(if_pkt_info_t pi)
487 {
488 	if_pkt_info_pad_t pi_pad;
489 
490 	pi_pad = (if_pkt_info_pad_t)pi;
491 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
492 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
493 #ifndef __LP64__
494 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
495 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
496 #endif
497 }
498 
499 static inline void
500 rxd_info_zero(if_rxd_info_t ri)
501 {
502 	if_rxd_info_pad_t ri_pad;
503 	int i;
504 
505 	ri_pad = (if_rxd_info_pad_t)ri;
506 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
507 		ri_pad->rxd_val[i] = 0;
508 		ri_pad->rxd_val[i+1] = 0;
509 		ri_pad->rxd_val[i+2] = 0;
510 		ri_pad->rxd_val[i+3] = 0;
511 	}
512 #ifdef __LP64__
513 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
514 #endif
515 }
516 
517 /*
518  * Only allow a single packet to take up most 1/nth of the tx ring
519  */
520 #define MAX_SINGLE_PACKET_FRACTION 12
521 #define IF_BAD_DMA (bus_addr_t)-1
522 
523 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
524 
525 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
526 
527 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
528 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
529 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
530 
531 
532 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
533 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
534 
535 
536 /* Our boot-time initialization hook */
537 static int	iflib_module_event_handler(module_t, int, void *);
538 
539 static moduledata_t iflib_moduledata = {
540 	"iflib",
541 	iflib_module_event_handler,
542 	NULL
543 };
544 
545 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
546 MODULE_VERSION(iflib, 1);
547 
548 MODULE_DEPEND(iflib, pci, 1, 1, 1);
549 MODULE_DEPEND(iflib, ether, 1, 1, 1);
550 
551 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
552 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
553 
554 #ifndef IFLIB_DEBUG_COUNTERS
555 #ifdef INVARIANTS
556 #define IFLIB_DEBUG_COUNTERS 1
557 #else
558 #define IFLIB_DEBUG_COUNTERS 0
559 #endif /* !INVARIANTS */
560 #endif
561 
562 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
563                    "iflib driver parameters");
564 
565 /*
566  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
567  */
568 static int iflib_min_tx_latency = 0;
569 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
570 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
571 static int iflib_no_tx_batch = 0;
572 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
573 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
574 
575 
576 #if IFLIB_DEBUG_COUNTERS
577 
578 static int iflib_tx_seen;
579 static int iflib_tx_sent;
580 static int iflib_tx_encap;
581 static int iflib_rx_allocs;
582 static int iflib_fl_refills;
583 static int iflib_fl_refills_large;
584 static int iflib_tx_frees;
585 
586 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
587 		   &iflib_tx_seen, 0, "# tx mbufs seen");
588 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
589 		   &iflib_tx_sent, 0, "# tx mbufs sent");
590 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
591 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
592 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
593 		   &iflib_tx_frees, 0, "# tx frees");
594 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
595 		   &iflib_rx_allocs, 0, "# rx allocations");
596 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
597 		   &iflib_fl_refills, 0, "# refills");
598 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
599 		   &iflib_fl_refills_large, 0, "# large refills");
600 
601 
602 static int iflib_txq_drain_flushing;
603 static int iflib_txq_drain_oactive;
604 static int iflib_txq_drain_notready;
605 static int iflib_txq_drain_encapfail;
606 
607 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
608 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
609 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
610 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
611 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
612 		   &iflib_txq_drain_notready, 0, "# drain notready");
613 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
614 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
615 
616 
617 static int iflib_encap_load_mbuf_fail;
618 static int iflib_encap_txq_avail_fail;
619 static int iflib_encap_txd_encap_fail;
620 
621 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
622 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
623 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
624 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
625 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
626 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
627 
628 static int iflib_task_fn_rxs;
629 static int iflib_rx_intr_enables;
630 static int iflib_fast_intrs;
631 static int iflib_intr_link;
632 static int iflib_intr_msix;
633 static int iflib_rx_unavail;
634 static int iflib_rx_ctx_inactive;
635 static int iflib_rx_zero_len;
636 static int iflib_rx_if_input;
637 static int iflib_rx_mbuf_null;
638 static int iflib_rxd_flush;
639 
640 static int iflib_verbose_debug;
641 
642 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
643 		   &iflib_intr_link, 0, "# intr link calls");
644 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
645 		   &iflib_intr_msix, 0, "# intr msix calls");
646 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
647 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
648 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
649 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
650 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
651 		   &iflib_fast_intrs, 0, "# fast_intr calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
653 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
654 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
655 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
657 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
659 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
661 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
663 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
664 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
665 		   &iflib_verbose_debug, 0, "enable verbose debugging");
666 
667 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
668 static void
669 iflib_debug_reset(void)
670 {
671 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
672 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
673 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
674 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
675 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
676 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
677 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
678 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
679 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
680 }
681 
682 #else
683 #define DBG_COUNTER_INC(name)
684 static void iflib_debug_reset(void) {}
685 #endif
686 
687 
688 
689 #define IFLIB_DEBUG 0
690 
691 static void iflib_tx_structures_free(if_ctx_t ctx);
692 static void iflib_rx_structures_free(if_ctx_t ctx);
693 static int iflib_queues_alloc(if_ctx_t ctx);
694 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
695 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
696 static int iflib_qset_structures_setup(if_ctx_t ctx);
697 static int iflib_msix_init(if_ctx_t ctx);
698 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
699 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
700 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
701 static int iflib_register(if_ctx_t);
702 static void iflib_init_locked(if_ctx_t ctx);
703 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
704 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
705 static void iflib_ifmp_purge(iflib_txq_t txq);
706 static void _iflib_pre_assert(if_softc_ctx_t scctx);
707 static void iflib_stop(if_ctx_t ctx);
708 static void iflib_if_init_locked(if_ctx_t ctx);
709 #ifndef __NO_STRICT_ALIGNMENT
710 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
711 #endif
712 
713 #ifdef DEV_NETMAP
714 #include <sys/selinfo.h>
715 #include <net/netmap.h>
716 #include <dev/netmap/netmap_kern.h>
717 
718 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
719 
720 /*
721  * device-specific sysctl variables:
722  *
723  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
724  *	During regular operations the CRC is stripped, but on some
725  *	hardware reception of frames not multiple of 64 is slower,
726  *	so using crcstrip=0 helps in benchmarks.
727  *
728  * iflib_rx_miss, iflib_rx_miss_bufs:
729  *	count packets that might be missed due to lost interrupts.
730  */
731 SYSCTL_DECL(_dev_netmap);
732 /*
733  * The xl driver by default strips CRCs and we do not override it.
734  */
735 
736 int iflib_crcstrip = 1;
737 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
738     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
739 
740 int iflib_rx_miss, iflib_rx_miss_bufs;
741 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
742     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
743 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
744     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
745 
746 /*
747  * Register/unregister. We are already under netmap lock.
748  * Only called on the first register or the last unregister.
749  */
750 static int
751 iflib_netmap_register(struct netmap_adapter *na, int onoff)
752 {
753 	struct ifnet *ifp = na->ifp;
754 	if_ctx_t ctx = ifp->if_softc;
755 	int status;
756 
757 	CTX_LOCK(ctx);
758 	IFDI_INTR_DISABLE(ctx);
759 
760 	/* Tell the stack that the interface is no longer active */
761 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
762 
763 	if (!CTX_IS_VF(ctx))
764 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
765 
766 	/* enable or disable flags and callbacks in na and ifp */
767 	if (onoff) {
768 		nm_set_native_flags(na);
769 	} else {
770 		nm_clear_native_flags(na);
771 	}
772 	iflib_stop(ctx);
773 	iflib_init_locked(ctx);
774 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
775 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
776 	if (status)
777 		nm_clear_native_flags(na);
778 	CTX_UNLOCK(ctx);
779 	return (status);
780 }
781 
782 /*
783  * Reconcile kernel and user view of the transmit ring.
784  *
785  * All information is in the kring.
786  * Userspace wants to send packets up to the one before kring->rhead,
787  * kernel knows kring->nr_hwcur is the first unsent packet.
788  *
789  * Here we push packets out (as many as possible), and possibly
790  * reclaim buffers from previously completed transmission.
791  *
792  * The caller (netmap) guarantees that there is only one instance
793  * running at any time. Any interference with other driver
794  * methods should be handled by the individual drivers.
795  */
796 static int
797 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
798 {
799 	struct netmap_adapter *na = kring->na;
800 	struct ifnet *ifp = na->ifp;
801 	struct netmap_ring *ring = kring->ring;
802 	u_int nm_i;	/* index into the netmap ring */
803 	u_int nic_i;	/* index into the NIC ring */
804 	u_int n;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int const head = kring->rhead;
807 	struct if_pkt_info pi;
808 
809 	/*
810 	 * interrupts on every tx packet are expensive so request
811 	 * them every half ring, or where NS_REPORT is set
812 	 */
813 	u_int report_frequency = kring->nkr_num_slots >> 1;
814 	/* device-specific */
815 	if_ctx_t ctx = ifp->if_softc;
816 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
817 
818 	if (txq->ift_sds.ifsd_map)
819 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
820 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
821 
822 
823 	/*
824 	 * First part: process new packets to send.
825 	 * nm_i is the current index in the netmap ring,
826 	 * nic_i is the corresponding index in the NIC ring.
827 	 *
828 	 * If we have packets to send (nm_i != head)
829 	 * iterate over the netmap ring, fetch length and update
830 	 * the corresponding slot in the NIC ring. Some drivers also
831 	 * need to update the buffer's physical address in the NIC slot
832 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
833 	 *
834 	 * The netmap_reload_map() calls is especially expensive,
835 	 * even when (as in this case) the tag is 0, so do only
836 	 * when the buffer has actually changed.
837 	 *
838 	 * If possible do not set the report/intr bit on all slots,
839 	 * but only a few times per ring or when NS_REPORT is set.
840 	 *
841 	 * Finally, on 10G and faster drivers, it might be useful
842 	 * to prefetch the next slot and txr entry.
843 	 */
844 
845 	nm_i = kring->nr_hwcur;
846 	pkt_info_zero(&pi);
847 	pi.ipi_segs = txq->ift_segs;
848 	pi.ipi_qsidx = kring->ring_id;
849 	if (nm_i != head) {	/* we have new packets to send */
850 		nic_i = netmap_idx_k2n(kring, nm_i);
851 
852 		__builtin_prefetch(&ring->slot[nm_i]);
853 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
854 		if (txq->ift_sds.ifsd_map)
855 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
856 
857 		for (n = 0; nm_i != head; n++) {
858 			struct netmap_slot *slot = &ring->slot[nm_i];
859 			u_int len = slot->len;
860 			uint64_t paddr;
861 			void *addr = PNMB(na, slot, &paddr);
862 			int flags = (slot->flags & NS_REPORT ||
863 				nic_i == 0 || nic_i == report_frequency) ?
864 				IPI_TX_INTR : 0;
865 
866 			/* device-specific */
867 			pi.ipi_len = len;
868 			pi.ipi_segs[0].ds_addr = paddr;
869 			pi.ipi_segs[0].ds_len = len;
870 			pi.ipi_nsegs = 1;
871 			pi.ipi_ndescs = 0;
872 			pi.ipi_pidx = nic_i;
873 			pi.ipi_flags = flags;
874 
875 			/* Fill the slot in the NIC ring. */
876 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
877 
878 			/* prefetch for next round */
879 			__builtin_prefetch(&ring->slot[nm_i + 1]);
880 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
881 			if (txq->ift_sds.ifsd_map) {
882 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
883 
884 				NM_CHECK_ADDR_LEN(na, addr, len);
885 
886 				if (slot->flags & NS_BUF_CHANGED) {
887 					/* buffer has changed, reload map */
888 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
889 				}
890 				/* make sure changes to the buffer are synced */
891 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
892 						BUS_DMASYNC_PREWRITE);
893 			}
894 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
895 			nm_i = nm_next(nm_i, lim);
896 			nic_i = nm_next(nic_i, lim);
897 		}
898 		kring->nr_hwcur = head;
899 
900 		/* synchronize the NIC ring */
901 		if (txq->ift_sds.ifsd_map)
902 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
903 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
904 
905 		/* (re)start the tx unit up to slot nic_i (excluded) */
906 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
907 	}
908 
909 	/*
910 	 * Second part: reclaim buffers for completed transmissions.
911 	 */
912 	if (iflib_tx_credits_update(ctx, txq)) {
913 		/* some tx completed, increment avail */
914 		nic_i = txq->ift_cidx_processed;
915 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
916 	}
917 	return (0);
918 }
919 
920 /*
921  * Reconcile kernel and user view of the receive ring.
922  * Same as for the txsync, this routine must be efficient.
923  * The caller guarantees a single invocations, but races against
924  * the rest of the driver should be handled here.
925  *
926  * On call, kring->rhead is the first packet that userspace wants
927  * to keep, and kring->rcur is the wakeup point.
928  * The kernel has previously reported packets up to kring->rtail.
929  *
930  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
931  * of whether or not we received an interrupt.
932  */
933 static int
934 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
935 {
936 	struct netmap_adapter *na = kring->na;
937 	struct netmap_ring *ring = kring->ring;
938 	uint32_t nm_i;	/* index into the netmap ring */
939 	uint32_t nic_i, nic_i_start;	/* index into the NIC ring */
940 	u_int i, n;
941 	u_int const lim = kring->nkr_num_slots - 1;
942 	u_int const head = kring->rhead;
943 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
944 	struct if_rxd_info ri;
945 	struct if_rxd_update iru;
946 
947 	struct ifnet *ifp = na->ifp;
948 	if_ctx_t ctx = ifp->if_softc;
949 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
950 	iflib_fl_t fl = rxq->ifr_fl;
951 	if (head > lim)
952 		return netmap_ring_reinit(kring);
953 
954 	/* XXX check sync modes */
955 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
956 		if (fl->ifl_sds.ifsd_map == NULL)
957 			continue;
958 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
959 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
960 	}
961 	/*
962 	 * First part: import newly received packets.
963 	 *
964 	 * nm_i is the index of the next free slot in the netmap ring,
965 	 * nic_i is the index of the next received packet in the NIC ring,
966 	 * and they may differ in case if_init() has been called while
967 	 * in netmap mode. For the receive ring we have
968 	 *
969 	 *	nic_i = rxr->next_check;
970 	 *	nm_i = kring->nr_hwtail (previous)
971 	 * and
972 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
973 	 *
974 	 * rxr->next_check is set to 0 on a ring reinit
975 	 */
976 	if (netmap_no_pendintr || force_update) {
977 		int crclen = iflib_crcstrip ? 0 : 4;
978 		int error, avail;
979 		uint16_t slot_flags = kring->nkr_slot_flags;
980 
981 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
982 			nic_i = fl->ifl_cidx;
983 			nm_i = netmap_idx_n2k(kring, nic_i);
984 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
985 			for (n = 0; avail > 0; n++, avail--) {
986 				rxd_info_zero(&ri);
987 				ri.iri_frags = rxq->ifr_frags;
988 				ri.iri_qsidx = kring->ring_id;
989 				ri.iri_ifp = ctx->ifc_ifp;
990 				ri.iri_cidx = nic_i;
991 
992 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
993 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
994 				ring->slot[nm_i].flags = slot_flags;
995 				if (fl->ifl_sds.ifsd_map)
996 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
997 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
998 				nm_i = nm_next(nm_i, lim);
999 				nic_i = nm_next(nic_i, lim);
1000 			}
1001 			if (n) { /* update the state variables */
1002 				if (netmap_no_pendintr && !force_update) {
1003 					/* diagnostics */
1004 					iflib_rx_miss ++;
1005 					iflib_rx_miss_bufs += n;
1006 				}
1007 				fl->ifl_cidx = nic_i;
1008 				kring->nr_hwtail = nm_i;
1009 			}
1010 			kring->nr_kflags &= ~NKR_PENDINTR;
1011 		}
1012 	}
1013 	/*
1014 	 * Second part: skip past packets that userspace has released.
1015 	 * (kring->nr_hwcur to head excluded),
1016 	 * and make the buffers available for reception.
1017 	 * As usual nm_i is the index in the netmap ring,
1018 	 * nic_i is the index in the NIC ring, and
1019 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1020 	 */
1021 	/* XXX not sure how this will work with multiple free lists */
1022 	nm_i = kring->nr_hwcur;
1023 	if (nm_i == head)
1024 		return (0);
1025 
1026 	iru.iru_paddrs = fl->ifl_bus_addrs;
1027 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1028 	iru.iru_idxs = fl->ifl_rxd_idxs;
1029 	iru.iru_qsidx = rxq->ifr_id;
1030 	iru.iru_buf_size = fl->ifl_buf_size;
1031 	iru.iru_flidx = fl->ifl_id;
1032 	nic_i_start = nic_i = netmap_idx_k2n(kring, nm_i);
1033 	for (i = 0; nm_i != head; i++) {
1034 		struct netmap_slot *slot = &ring->slot[nm_i];
1035 		void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
1036 
1037 		if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
1038 			goto ring_reset;
1039 
1040 		fl->ifl_vm_addrs[i] = addr;
1041 		if (fl->ifl_sds.ifsd_map && (slot->flags & NS_BUF_CHANGED)) {
1042 			/* buffer has changed, reload map */
1043 			netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
1044 		}
1045 		slot->flags &= ~NS_BUF_CHANGED;
1046 
1047 		nm_i = nm_next(nm_i, lim);
1048 		fl->ifl_rxd_idxs[i] = nic_i = nm_next(nic_i, lim);
1049 		if (nm_i != head && i < IFLIB_MAX_RX_REFRESH)
1050 			continue;
1051 
1052 		iru.iru_pidx = nic_i_start;
1053 		iru.iru_count = i;
1054 		i = 0;
1055 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1056 		if (fl->ifl_sds.ifsd_map == NULL) {
1057 			nic_i_start = nic_i;
1058 			continue;
1059 		}
1060 		nic_i = nic_i_start;
1061 		for (n = 0; n < iru.iru_count; n++) {
1062 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
1063 					BUS_DMASYNC_PREREAD);
1064 			nic_i = nm_next(nic_i, lim);
1065 		}
1066 		nic_i_start = nic_i;
1067 	}
1068 	kring->nr_hwcur = head;
1069 
1070 	if (fl->ifl_sds.ifsd_map)
1071 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1072 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1073 	/*
1074 	 * IMPORTANT: we must leave one free slot in the ring,
1075 	 * so move nic_i back by one unit
1076 	 */
1077 	nic_i = nm_prev(nic_i, lim);
1078 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
1079 	return 0;
1080 
1081 ring_reset:
1082 	return netmap_ring_reinit(kring);
1083 }
1084 
1085 static void
1086 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1087 {
1088 	struct ifnet *ifp = na->ifp;
1089 	if_ctx_t ctx = ifp->if_softc;
1090 
1091 	CTX_LOCK(ctx);
1092 	if (onoff) {
1093 		IFDI_INTR_ENABLE(ctx);
1094 	} else {
1095 		IFDI_INTR_DISABLE(ctx);
1096 	}
1097 	CTX_UNLOCK(ctx);
1098 }
1099 
1100 
1101 static int
1102 iflib_netmap_attach(if_ctx_t ctx)
1103 {
1104 	struct netmap_adapter na;
1105 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1106 
1107 	bzero(&na, sizeof(na));
1108 
1109 	na.ifp = ctx->ifc_ifp;
1110 	na.na_flags = NAF_BDG_MAYSLEEP;
1111 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1112 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1113 
1114 	na.num_tx_desc = scctx->isc_ntxd[0];
1115 	na.num_rx_desc = scctx->isc_nrxd[0];
1116 	na.nm_txsync = iflib_netmap_txsync;
1117 	na.nm_rxsync = iflib_netmap_rxsync;
1118 	na.nm_register = iflib_netmap_register;
1119 	na.nm_intr = iflib_netmap_intr;
1120 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1121 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1122 	return (netmap_attach(&na));
1123 }
1124 
1125 static void
1126 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1127 {
1128 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1129 	struct netmap_slot *slot;
1130 
1131 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1132 	if (slot == NULL)
1133 		return;
1134 	if (txq->ift_sds.ifsd_map == NULL)
1135 		return;
1136 
1137 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1138 
1139 		/*
1140 		 * In netmap mode, set the map for the packet buffer.
1141 		 * NOTE: Some drivers (not this one) also need to set
1142 		 * the physical buffer address in the NIC ring.
1143 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1144 		 * netmap slot index, si
1145 		 */
1146 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1147 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1148 	}
1149 }
1150 static void
1151 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1152 {
1153 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1154 	struct netmap_slot *slot;
1155 	struct if_rxd_update iru;
1156 	iflib_fl_t fl;
1157 	bus_dmamap_t *map;
1158 	int nrxd;
1159 	uint32_t i, j, pidx_start;
1160 
1161 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1162 	if (slot == NULL)
1163 		return;
1164 	fl = &rxq->ifr_fl[0];
1165 	map = fl->ifl_sds.ifsd_map;
1166 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1167 	iru.iru_paddrs = fl->ifl_bus_addrs;
1168 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1169 	iru.iru_idxs = fl->ifl_rxd_idxs;
1170 	iru.iru_qsidx = rxq->ifr_id;
1171 	iru.iru_buf_size = rxq->ifr_fl[0].ifl_buf_size;
1172 	iru.iru_flidx = 0;
1173 
1174 	for (pidx_start = i = j = 0; i < nrxd; i++, j++) {
1175 		int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1176 		void *addr;
1177 
1178 		fl->ifl_rxd_idxs[j] = i;
1179 		addr = fl->ifl_vm_addrs[j] = PNMB(na, slot + sj, &fl->ifl_bus_addrs[j]);
1180 		if (map) {
1181 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1182 			map++;
1183 		}
1184 
1185 		if (j < IFLIB_MAX_RX_REFRESH && i < nrxd - 1)
1186 			continue;
1187 
1188 		iru.iru_pidx = pidx_start;
1189 		pidx_start = i;
1190 		iru.iru_count = j;
1191 		j = 0;
1192 		MPASS(pidx_start + j <= nrxd);
1193 		/* Update descriptors and the cached value */
1194 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1195 	}
1196 	/* preserve queue */
1197 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1198 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1199 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1200 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1201 	} else
1202 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1203 }
1204 
1205 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1206 
1207 #else
1208 #define iflib_netmap_txq_init(ctx, txq)
1209 #define iflib_netmap_rxq_init(ctx, rxq)
1210 #define iflib_netmap_detach(ifp)
1211 
1212 #define iflib_netmap_attach(ctx) (0)
1213 #define netmap_rx_irq(ifp, qid, budget) (0)
1214 #define netmap_tx_irq(ifp, qid) do {} while (0)
1215 
1216 #endif
1217 
1218 #if defined(__i386__) || defined(__amd64__)
1219 static __inline void
1220 prefetch(void *x)
1221 {
1222 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1223 }
1224 #else
1225 #define prefetch(x)
1226 #endif
1227 
1228 static void
1229 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1230 {
1231 	if (err)
1232 		return;
1233 	*(bus_addr_t *) arg = segs[0].ds_addr;
1234 }
1235 
1236 int
1237 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1238 {
1239 	int err;
1240 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1241 	device_t dev = ctx->ifc_dev;
1242 
1243 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1244 
1245 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1246 				sctx->isc_q_align, 0,	/* alignment, bounds */
1247 				BUS_SPACE_MAXADDR,	/* lowaddr */
1248 				BUS_SPACE_MAXADDR,	/* highaddr */
1249 				NULL, NULL,		/* filter, filterarg */
1250 				size,			/* maxsize */
1251 				1,			/* nsegments */
1252 				size,			/* maxsegsize */
1253 				BUS_DMA_ALLOCNOW,	/* flags */
1254 				NULL,			/* lockfunc */
1255 				NULL,			/* lockarg */
1256 				&dma->idi_tag);
1257 	if (err) {
1258 		device_printf(dev,
1259 		    "%s: bus_dma_tag_create failed: %d\n",
1260 		    __func__, err);
1261 		goto fail_0;
1262 	}
1263 
1264 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1265 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1266 	if (err) {
1267 		device_printf(dev,
1268 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1269 		    __func__, (uintmax_t)size, err);
1270 		goto fail_1;
1271 	}
1272 
1273 	dma->idi_paddr = IF_BAD_DMA;
1274 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1275 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1276 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1277 		device_printf(dev,
1278 		    "%s: bus_dmamap_load failed: %d\n",
1279 		    __func__, err);
1280 		goto fail_2;
1281 	}
1282 
1283 	dma->idi_size = size;
1284 	return (0);
1285 
1286 fail_2:
1287 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1288 fail_1:
1289 	bus_dma_tag_destroy(dma->idi_tag);
1290 fail_0:
1291 	dma->idi_tag = NULL;
1292 
1293 	return (err);
1294 }
1295 
1296 int
1297 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1298 {
1299 	int i, err;
1300 	iflib_dma_info_t *dmaiter;
1301 
1302 	dmaiter = dmalist;
1303 	for (i = 0; i < count; i++, dmaiter++) {
1304 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1305 			break;
1306 	}
1307 	if (err)
1308 		iflib_dma_free_multi(dmalist, i);
1309 	return (err);
1310 }
1311 
1312 void
1313 iflib_dma_free(iflib_dma_info_t dma)
1314 {
1315 	if (dma->idi_tag == NULL)
1316 		return;
1317 	if (dma->idi_paddr != IF_BAD_DMA) {
1318 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1319 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1320 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1321 		dma->idi_paddr = IF_BAD_DMA;
1322 	}
1323 	if (dma->idi_vaddr != NULL) {
1324 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1325 		dma->idi_vaddr = NULL;
1326 	}
1327 	bus_dma_tag_destroy(dma->idi_tag);
1328 	dma->idi_tag = NULL;
1329 }
1330 
1331 void
1332 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1333 {
1334 	int i;
1335 	iflib_dma_info_t *dmaiter = dmalist;
1336 
1337 	for (i = 0; i < count; i++, dmaiter++)
1338 		iflib_dma_free(*dmaiter);
1339 }
1340 
1341 #ifdef EARLY_AP_STARTUP
1342 static const int iflib_started = 1;
1343 #else
1344 /*
1345  * We used to abuse the smp_started flag to decide if the queues have been
1346  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1347  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1348  * is set.  Run a SYSINIT() strictly after that to just set a usable
1349  * completion flag.
1350  */
1351 
1352 static int iflib_started;
1353 
1354 static void
1355 iflib_record_started(void *arg)
1356 {
1357 	iflib_started = 1;
1358 }
1359 
1360 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1361 	iflib_record_started, NULL);
1362 #endif
1363 
1364 static int
1365 iflib_fast_intr(void *arg)
1366 {
1367 	iflib_filter_info_t info = arg;
1368 	struct grouptask *gtask = info->ifi_task;
1369 	if (!iflib_started)
1370 		return (FILTER_HANDLED);
1371 
1372 	DBG_COUNTER_INC(fast_intrs);
1373 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1374 		return (FILTER_HANDLED);
1375 
1376 	GROUPTASK_ENQUEUE(gtask);
1377 	return (FILTER_HANDLED);
1378 }
1379 
1380 static int
1381 iflib_fast_intr_rxtx(void *arg)
1382 {
1383 	iflib_filter_info_t info = arg;
1384 	struct grouptask *gtask = info->ifi_task;
1385 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1386 	if_ctx_t ctx;
1387 	int i, cidx;
1388 
1389 	if (!iflib_started)
1390 		return (FILTER_HANDLED);
1391 
1392 	DBG_COUNTER_INC(fast_intrs);
1393 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1394 		return (FILTER_HANDLED);
1395 
1396 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1397 		qidx_t txqid = rxq->ifr_txqid[i];
1398 
1399 		ctx = rxq->ifr_ctx;
1400 
1401 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1402 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1403 			continue;
1404 		}
1405 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1406 	}
1407 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1408 		cidx = rxq->ifr_cq_cidx;
1409 	else
1410 		cidx = rxq->ifr_fl[0].ifl_cidx;
1411 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1412 		GROUPTASK_ENQUEUE(gtask);
1413 	else
1414 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1415 	return (FILTER_HANDLED);
1416 }
1417 
1418 
1419 static int
1420 iflib_fast_intr_ctx(void *arg)
1421 {
1422 	iflib_filter_info_t info = arg;
1423 	struct grouptask *gtask = info->ifi_task;
1424 
1425 	if (!iflib_started)
1426 		return (FILTER_HANDLED);
1427 
1428 	DBG_COUNTER_INC(fast_intrs);
1429 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1430 		return (FILTER_HANDLED);
1431 
1432 	GROUPTASK_ENQUEUE(gtask);
1433 	return (FILTER_HANDLED);
1434 }
1435 
1436 static int
1437 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1438 	driver_filter_t filter, driver_intr_t handler, void *arg,
1439 				 char *name)
1440 {
1441 	int rc, flags;
1442 	struct resource *res;
1443 	void *tag = NULL;
1444 	device_t dev = ctx->ifc_dev;
1445 
1446 	flags = RF_ACTIVE;
1447 	if (ctx->ifc_flags & IFC_LEGACY)
1448 		flags |= RF_SHAREABLE;
1449 	MPASS(rid < 512);
1450 	irq->ii_rid = rid;
1451 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1452 	if (res == NULL) {
1453 		device_printf(dev,
1454 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1455 		return (ENOMEM);
1456 	}
1457 	irq->ii_res = res;
1458 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1459 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1460 						filter, handler, arg, &tag);
1461 	if (rc != 0) {
1462 		device_printf(dev,
1463 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1464 					  rid, name ? name : "unknown", rc);
1465 		return (rc);
1466 	} else if (name)
1467 		bus_describe_intr(dev, res, tag, "%s", name);
1468 
1469 	irq->ii_tag = tag;
1470 	return (0);
1471 }
1472 
1473 
1474 /*********************************************************************
1475  *
1476  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1477  *  the information needed to transmit a packet on the wire. This is
1478  *  called only once at attach, setup is done every reset.
1479  *
1480  **********************************************************************/
1481 
1482 static int
1483 iflib_txsd_alloc(iflib_txq_t txq)
1484 {
1485 	if_ctx_t ctx = txq->ift_ctx;
1486 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1487 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1488 	device_t dev = ctx->ifc_dev;
1489 	int err, nsegments, ntsosegments;
1490 
1491 	nsegments = scctx->isc_tx_nsegments;
1492 	ntsosegments = scctx->isc_tx_tso_segments_max;
1493 	MPASS(scctx->isc_ntxd[0] > 0);
1494 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1495 	MPASS(nsegments > 0);
1496 	MPASS(ntsosegments > 0);
1497 	/*
1498 	 * Setup DMA descriptor areas.
1499 	 */
1500 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1501 			       1, 0,			/* alignment, bounds */
1502 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1503 			       BUS_SPACE_MAXADDR,	/* highaddr */
1504 			       NULL, NULL,		/* filter, filterarg */
1505 			       sctx->isc_tx_maxsize,		/* maxsize */
1506 			       nsegments,	/* nsegments */
1507 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1508 			       0,			/* flags */
1509 			       NULL,			/* lockfunc */
1510 			       NULL,			/* lockfuncarg */
1511 			       &txq->ift_desc_tag))) {
1512 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1513 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1514 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1515 		goto fail;
1516 	}
1517 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1518 			       1, 0,			/* alignment, bounds */
1519 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1520 			       BUS_SPACE_MAXADDR,	/* highaddr */
1521 			       NULL, NULL,		/* filter, filterarg */
1522 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1523 			       ntsosegments,	/* nsegments */
1524 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1525 			       0,			/* flags */
1526 			       NULL,			/* lockfunc */
1527 			       NULL,			/* lockfuncarg */
1528 			       &txq->ift_tso_desc_tag))) {
1529 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1530 
1531 		goto fail;
1532 	}
1533 	if (!(txq->ift_sds.ifsd_flags =
1534 	    (uint8_t *) malloc(sizeof(uint8_t) *
1535 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1536 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1537 		err = ENOMEM;
1538 		goto fail;
1539 	}
1540 	if (!(txq->ift_sds.ifsd_m =
1541 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1542 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1543 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1544 		err = ENOMEM;
1545 		goto fail;
1546 	}
1547 
1548         /* Create the descriptor buffer dma maps */
1549 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1550 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1551 		return (0);
1552 
1553 	if (!(txq->ift_sds.ifsd_map =
1554 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1555 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1556 		err = ENOMEM;
1557 		goto fail;
1558 	}
1559 
1560 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1561 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1562 		if (err != 0) {
1563 			device_printf(dev, "Unable to create TX DMA map\n");
1564 			goto fail;
1565 		}
1566 	}
1567 #endif
1568 	return (0);
1569 fail:
1570 	/* We free all, it handles case where we are in the middle */
1571 	iflib_tx_structures_free(ctx);
1572 	return (err);
1573 }
1574 
1575 static void
1576 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1577 {
1578 	bus_dmamap_t map;
1579 
1580 	map = NULL;
1581 	if (txq->ift_sds.ifsd_map != NULL)
1582 		map = txq->ift_sds.ifsd_map[i];
1583 	if (map != NULL) {
1584 		bus_dmamap_unload(txq->ift_desc_tag, map);
1585 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1586 		txq->ift_sds.ifsd_map[i] = NULL;
1587 	}
1588 }
1589 
1590 static void
1591 iflib_txq_destroy(iflib_txq_t txq)
1592 {
1593 	if_ctx_t ctx = txq->ift_ctx;
1594 
1595 	for (int i = 0; i < txq->ift_size; i++)
1596 		iflib_txsd_destroy(ctx, txq, i);
1597 	if (txq->ift_sds.ifsd_map != NULL) {
1598 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1599 		txq->ift_sds.ifsd_map = NULL;
1600 	}
1601 	if (txq->ift_sds.ifsd_m != NULL) {
1602 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1603 		txq->ift_sds.ifsd_m = NULL;
1604 	}
1605 	if (txq->ift_sds.ifsd_flags != NULL) {
1606 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1607 		txq->ift_sds.ifsd_flags = NULL;
1608 	}
1609 	if (txq->ift_desc_tag != NULL) {
1610 		bus_dma_tag_destroy(txq->ift_desc_tag);
1611 		txq->ift_desc_tag = NULL;
1612 	}
1613 	if (txq->ift_tso_desc_tag != NULL) {
1614 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1615 		txq->ift_tso_desc_tag = NULL;
1616 	}
1617 }
1618 
1619 static void
1620 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1621 {
1622 	struct mbuf **mp;
1623 
1624 	mp = &txq->ift_sds.ifsd_m[i];
1625 	if (*mp == NULL)
1626 		return;
1627 
1628 	if (txq->ift_sds.ifsd_map != NULL) {
1629 		bus_dmamap_sync(txq->ift_desc_tag,
1630 				txq->ift_sds.ifsd_map[i],
1631 				BUS_DMASYNC_POSTWRITE);
1632 		bus_dmamap_unload(txq->ift_desc_tag,
1633 				  txq->ift_sds.ifsd_map[i]);
1634 	}
1635 	m_free(*mp);
1636 	DBG_COUNTER_INC(tx_frees);
1637 	*mp = NULL;
1638 }
1639 
1640 static int
1641 iflib_txq_setup(iflib_txq_t txq)
1642 {
1643 	if_ctx_t ctx = txq->ift_ctx;
1644 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1645 	iflib_dma_info_t di;
1646 	int i;
1647 
1648 	/* Set number of descriptors available */
1649 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1650 	/* XXX make configurable */
1651 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1652 
1653 	/* Reset indices */
1654 	txq->ift_cidx_processed = 0;
1655 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1656 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1657 
1658 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1659 		bzero((void *)di->idi_vaddr, di->idi_size);
1660 
1661 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1662 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1663 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1664 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1665 	return (0);
1666 }
1667 
1668 /*********************************************************************
1669  *
1670  *  Allocate memory for rx_buffer structures. Since we use one
1671  *  rx_buffer per received packet, the maximum number of rx_buffer's
1672  *  that we'll need is equal to the number of receive descriptors
1673  *  that we've allocated.
1674  *
1675  **********************************************************************/
1676 static int
1677 iflib_rxsd_alloc(iflib_rxq_t rxq)
1678 {
1679 	if_ctx_t ctx = rxq->ifr_ctx;
1680 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1681 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1682 	device_t dev = ctx->ifc_dev;
1683 	iflib_fl_t fl;
1684 	int			err;
1685 
1686 	MPASS(scctx->isc_nrxd[0] > 0);
1687 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1688 
1689 	fl = rxq->ifr_fl;
1690 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1691 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1692 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1693 					 1, 0,			/* alignment, bounds */
1694 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1695 					 BUS_SPACE_MAXADDR,	/* highaddr */
1696 					 NULL, NULL,		/* filter, filterarg */
1697 					 sctx->isc_rx_maxsize,	/* maxsize */
1698 					 sctx->isc_rx_nsegments,	/* nsegments */
1699 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1700 					 0,			/* flags */
1701 					 NULL,			/* lockfunc */
1702 					 NULL,			/* lockarg */
1703 					 &fl->ifl_desc_tag);
1704 		if (err) {
1705 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1706 				__func__, err);
1707 			goto fail;
1708 		}
1709 		if (!(fl->ifl_sds.ifsd_flags =
1710 		      (uint8_t *) malloc(sizeof(uint8_t) *
1711 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1712 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1713 			err = ENOMEM;
1714 			goto fail;
1715 		}
1716 		if (!(fl->ifl_sds.ifsd_m =
1717 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1718 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1719 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1720 			err = ENOMEM;
1721 			goto fail;
1722 		}
1723 		if (!(fl->ifl_sds.ifsd_cl =
1724 		      (caddr_t *) malloc(sizeof(caddr_t) *
1725 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1726 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1727 			err = ENOMEM;
1728 			goto fail;
1729 		}
1730 
1731 		/* Create the descriptor buffer dma maps */
1732 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1733 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1734 			continue;
1735 
1736 		if (!(fl->ifl_sds.ifsd_map =
1737 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1738 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1739 			err = ENOMEM;
1740 			goto fail;
1741 		}
1742 
1743 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1744 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1745 			if (err != 0) {
1746 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1747 				goto fail;
1748 			}
1749 		}
1750 #endif
1751 	}
1752 	return (0);
1753 
1754 fail:
1755 	iflib_rx_structures_free(ctx);
1756 	return (err);
1757 }
1758 
1759 
1760 /*
1761  * Internal service routines
1762  */
1763 
1764 struct rxq_refill_cb_arg {
1765 	int               error;
1766 	bus_dma_segment_t seg;
1767 	int               nseg;
1768 };
1769 
1770 static void
1771 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1772 {
1773 	struct rxq_refill_cb_arg *cb_arg = arg;
1774 
1775 	cb_arg->error = error;
1776 	cb_arg->seg = segs[0];
1777 	cb_arg->nseg = nseg;
1778 }
1779 
1780 
1781 #ifdef ACPI_DMAR
1782 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1783 #else
1784 #define IS_DMAR(ctx) (0)
1785 #endif
1786 
1787 /**
1788  *	rxq_refill - refill an rxq  free-buffer list
1789  *	@ctx: the iflib context
1790  *	@rxq: the free-list to refill
1791  *	@n: the number of new buffers to allocate
1792  *
1793  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1794  *	The caller must assure that @n does not exceed the queue's capacity.
1795  */
1796 static void
1797 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1798 {
1799 	struct mbuf *m;
1800 	int idx, pidx = fl->ifl_pidx;
1801 	caddr_t cl, *sd_cl;
1802 	struct mbuf **sd_m;
1803 	uint8_t *sd_flags;
1804 	struct if_rxd_update iru;
1805 	bus_dmamap_t *sd_map;
1806 	int n, i = 0;
1807 	uint64_t bus_addr;
1808 	int err;
1809 
1810 	sd_m = fl->ifl_sds.ifsd_m;
1811 	sd_map = fl->ifl_sds.ifsd_map;
1812 	sd_cl = fl->ifl_sds.ifsd_cl;
1813 	sd_flags = fl->ifl_sds.ifsd_flags;
1814 	idx = pidx;
1815 
1816 	n  = count;
1817 	MPASS(n > 0);
1818 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1819 
1820 	if (pidx < fl->ifl_cidx)
1821 		MPASS(pidx + n <= fl->ifl_cidx);
1822 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1823 		MPASS(fl->ifl_gen == 0);
1824 	if (pidx > fl->ifl_cidx)
1825 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1826 
1827 	DBG_COUNTER_INC(fl_refills);
1828 	if (n > 8)
1829 		DBG_COUNTER_INC(fl_refills_large);
1830 	iru.iru_paddrs = fl->ifl_bus_addrs;
1831 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1832 	iru.iru_idxs = fl->ifl_rxd_idxs;
1833 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1834 	iru.iru_buf_size = fl->ifl_buf_size;
1835 	iru.iru_flidx = fl->ifl_id;
1836 	while (n--) {
1837 		/*
1838 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1839 		 * initialized after rx.
1840 		 *
1841 		 * If the cluster is still set then we know a minimum sized packet was received
1842 		 */
1843 		if ((cl = sd_cl[idx]) == NULL) {
1844 			if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1845 				break;
1846 #if MEMORY_LOGGING
1847 			fl->ifl_cl_enqueued++;
1848 #endif
1849 		}
1850 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1851 			break;
1852 		}
1853 #if MEMORY_LOGGING
1854 		fl->ifl_m_enqueued++;
1855 #endif
1856 
1857 		DBG_COUNTER_INC(rx_allocs);
1858 #if defined(__i386__) || defined(__amd64__)
1859 		if (!IS_DMAR(ctx)) {
1860 			bus_addr = pmap_kextract((vm_offset_t)cl);
1861 		} else
1862 #endif
1863 		{
1864 			struct rxq_refill_cb_arg cb_arg;
1865 			iflib_rxq_t q;
1866 
1867 			cb_arg.error = 0;
1868 			q = fl->ifl_rxq;
1869 			MPASS(sd_map != NULL);
1870 			MPASS(sd_map[idx] != NULL);
1871 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1872 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1873 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[idx], BUS_DMASYNC_PREREAD);
1874 
1875 			if (err != 0 || cb_arg.error) {
1876 				/*
1877 				 * !zone_pack ?
1878 				 */
1879 				if (fl->ifl_zone == zone_pack)
1880 					uma_zfree(fl->ifl_zone, cl);
1881 				m_free(m);
1882 				n = 0;
1883 				goto done;
1884 			}
1885 			bus_addr = cb_arg.seg.ds_addr;
1886 		}
1887 		sd_flags[idx] |= RX_SW_DESC_INUSE;
1888 
1889 		MPASS(sd_m[idx] == NULL);
1890 		sd_cl[idx] = cl;
1891 		sd_m[idx] = m;
1892 		fl->ifl_rxd_idxs[i] = idx;
1893 		fl->ifl_bus_addrs[i] = bus_addr;
1894 		fl->ifl_vm_addrs[i] = cl;
1895 		fl->ifl_credits++;
1896 		i++;
1897 		MPASS(fl->ifl_credits <= fl->ifl_size);
1898 		if (++idx == fl->ifl_size) {
1899 			fl->ifl_gen = 1;
1900 			idx = 0;
1901 		}
1902 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1903 			iru.iru_pidx = pidx;
1904 			iru.iru_count = i;
1905 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1906 			i = 0;
1907 			pidx = idx;
1908 		}
1909 		fl->ifl_pidx = idx;
1910 
1911 	}
1912 done:
1913 	DBG_COUNTER_INC(rxd_flush);
1914 	if (fl->ifl_pidx == 0)
1915 		pidx = fl->ifl_size - 1;
1916 	else
1917 		pidx = fl->ifl_pidx - 1;
1918 
1919 	if (sd_map)
1920 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1921 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1922 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1923 }
1924 
1925 static __inline void
1926 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1927 {
1928 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1929 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1930 #ifdef INVARIANTS
1931 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1932 #endif
1933 
1934 	MPASS(fl->ifl_credits <= fl->ifl_size);
1935 	MPASS(reclaimable == delta);
1936 
1937 	if (reclaimable > 0)
1938 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1939 }
1940 
1941 static void
1942 iflib_fl_bufs_free(iflib_fl_t fl)
1943 {
1944 	iflib_dma_info_t idi = fl->ifl_ifdi;
1945 	uint32_t i;
1946 
1947 	for (i = 0; i < fl->ifl_size; i++) {
1948 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1949 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1950 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1951 
1952 		if (*sd_flags & RX_SW_DESC_INUSE) {
1953 			if (fl->ifl_sds.ifsd_map != NULL) {
1954 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1955 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1956 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1957 			}
1958 			if (*sd_m != NULL) {
1959 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1960 				uma_zfree(zone_mbuf, *sd_m);
1961 			}
1962 			if (*sd_cl != NULL)
1963 				uma_zfree(fl->ifl_zone, *sd_cl);
1964 			*sd_flags = 0;
1965 		} else {
1966 			MPASS(*sd_cl == NULL);
1967 			MPASS(*sd_m == NULL);
1968 		}
1969 #if MEMORY_LOGGING
1970 		fl->ifl_m_dequeued++;
1971 		fl->ifl_cl_dequeued++;
1972 #endif
1973 		*sd_cl = NULL;
1974 		*sd_m = NULL;
1975 	}
1976 #ifdef INVARIANTS
1977 	for (i = 0; i < fl->ifl_size; i++) {
1978 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1979 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1980 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1981 	}
1982 #endif
1983 	/*
1984 	 * Reset free list values
1985 	 */
1986 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1987 	bzero(idi->idi_vaddr, idi->idi_size);
1988 }
1989 
1990 /*********************************************************************
1991  *
1992  *  Initialize a receive ring and its buffers.
1993  *
1994  **********************************************************************/
1995 static int
1996 iflib_fl_setup(iflib_fl_t fl)
1997 {
1998 	iflib_rxq_t rxq = fl->ifl_rxq;
1999 	if_ctx_t ctx = rxq->ifr_ctx;
2000 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2001 
2002 	/*
2003 	** Free current RX buffer structs and their mbufs
2004 	*/
2005 	iflib_fl_bufs_free(fl);
2006 	/* Now replenish the mbufs */
2007 	MPASS(fl->ifl_credits == 0);
2008 	/*
2009 	 * XXX don't set the max_frame_size to larger
2010 	 * than the hardware can handle
2011 	 */
2012 	if (sctx->isc_max_frame_size <= 2048)
2013 		fl->ifl_buf_size = MCLBYTES;
2014 #ifndef CONTIGMALLOC_WORKS
2015 	else
2016 		fl->ifl_buf_size = MJUMPAGESIZE;
2017 #else
2018 	else if (sctx->isc_max_frame_size <= 4096)
2019 		fl->ifl_buf_size = MJUMPAGESIZE;
2020 	else if (sctx->isc_max_frame_size <= 9216)
2021 		fl->ifl_buf_size = MJUM9BYTES;
2022 	else
2023 		fl->ifl_buf_size = MJUM16BYTES;
2024 #endif
2025 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2026 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2027 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2028 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2029 
2030 
2031 	/* avoid pre-allocating zillions of clusters to an idle card
2032 	 * potentially speeding up attach
2033 	 */
2034 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2035 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2036 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2037 		return (ENOBUFS);
2038 	/*
2039 	 * handle failure
2040 	 */
2041 	MPASS(rxq != NULL);
2042 	MPASS(fl->ifl_ifdi != NULL);
2043 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2044 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2045 	return (0);
2046 }
2047 
2048 /*********************************************************************
2049  *
2050  *  Free receive ring data structures
2051  *
2052  **********************************************************************/
2053 static void
2054 iflib_rx_sds_free(iflib_rxq_t rxq)
2055 {
2056 	iflib_fl_t fl;
2057 	int i;
2058 
2059 	if (rxq->ifr_fl != NULL) {
2060 		for (i = 0; i < rxq->ifr_nfl; i++) {
2061 			fl = &rxq->ifr_fl[i];
2062 			if (fl->ifl_desc_tag != NULL) {
2063 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2064 				fl->ifl_desc_tag = NULL;
2065 			}
2066 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2067 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2068 			/* XXX destroy maps first */
2069 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2070 			fl->ifl_sds.ifsd_m = NULL;
2071 			fl->ifl_sds.ifsd_cl = NULL;
2072 			fl->ifl_sds.ifsd_map = NULL;
2073 		}
2074 		free(rxq->ifr_fl, M_IFLIB);
2075 		rxq->ifr_fl = NULL;
2076 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2077 	}
2078 }
2079 
2080 /*
2081  * MI independent logic
2082  *
2083  */
2084 static void
2085 iflib_timer(void *arg)
2086 {
2087 	iflib_txq_t txq = arg;
2088 	if_ctx_t ctx = txq->ift_ctx;
2089 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2090 
2091 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2092 		return;
2093 	/*
2094 	** Check on the state of the TX queue(s), this
2095 	** can be done without the lock because its RO
2096 	** and the HUNG state will be static if set.
2097 	*/
2098 	IFDI_TIMER(ctx, txq->ift_id);
2099 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2100 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2101 	     (sctx->isc_pause_frames == 0)))
2102 		goto hung;
2103 
2104 	if (ifmp_ring_is_stalled(txq->ift_br))
2105 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2106 	txq->ift_cleaned_prev = txq->ift_cleaned;
2107 	/* handle any laggards */
2108 	if (txq->ift_db_pending)
2109 		GROUPTASK_ENQUEUE(&txq->ift_task);
2110 
2111 	sctx->isc_pause_frames = 0;
2112 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2113 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2114 	return;
2115 hung:
2116 	CTX_LOCK(ctx);
2117 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2118 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2119 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2120 
2121 	IFDI_WATCHDOG_RESET(ctx);
2122 	ctx->ifc_watchdog_events++;
2123 
2124 	ctx->ifc_flags |= IFC_DO_RESET;
2125 	iflib_admin_intr_deferred(ctx);
2126 	CTX_UNLOCK(ctx);
2127 }
2128 
2129 static void
2130 iflib_init_locked(if_ctx_t ctx)
2131 {
2132 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2133 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2134 	if_t ifp = ctx->ifc_ifp;
2135 	iflib_fl_t fl;
2136 	iflib_txq_t txq;
2137 	iflib_rxq_t rxq;
2138 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2139 
2140 
2141 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2142 	IFDI_INTR_DISABLE(ctx);
2143 
2144 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2145 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2146 	/* Set hardware offload abilities */
2147 	if_clearhwassist(ifp);
2148 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2149 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2150 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2151 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2152 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2153 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2154 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2155 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2156 
2157 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2158 		CALLOUT_LOCK(txq);
2159 		callout_stop(&txq->ift_timer);
2160 		CALLOUT_UNLOCK(txq);
2161 		iflib_netmap_txq_init(ctx, txq);
2162 	}
2163 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2164 		MPASS(rxq->ifr_id == i);
2165 		iflib_netmap_rxq_init(ctx, rxq);
2166 	}
2167 #ifdef INVARIANTS
2168 	i = if_getdrvflags(ifp);
2169 #endif
2170 	IFDI_INIT(ctx);
2171 	MPASS(if_getdrvflags(ifp) == i);
2172 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2173 		/* XXX this should really be done on a per-queue basis */
2174 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2175 			continue;
2176 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2177 			if (iflib_fl_setup(fl)) {
2178 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2179 				goto done;
2180 			}
2181 		}
2182 	}
2183 	done:
2184 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2185 	IFDI_INTR_ENABLE(ctx);
2186 	txq = ctx->ifc_txqs;
2187 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2188 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2189 			txq->ift_timer.c_cpu);
2190 }
2191 
2192 static int
2193 iflib_media_change(if_t ifp)
2194 {
2195 	if_ctx_t ctx = if_getsoftc(ifp);
2196 	int err;
2197 
2198 	CTX_LOCK(ctx);
2199 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2200 		iflib_init_locked(ctx);
2201 	CTX_UNLOCK(ctx);
2202 	return (err);
2203 }
2204 
2205 static void
2206 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2207 {
2208 	if_ctx_t ctx = if_getsoftc(ifp);
2209 
2210 	CTX_LOCK(ctx);
2211 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2212 	IFDI_MEDIA_STATUS(ctx, ifmr);
2213 	CTX_UNLOCK(ctx);
2214 }
2215 
2216 static void
2217 iflib_stop(if_ctx_t ctx)
2218 {
2219 	iflib_txq_t txq = ctx->ifc_txqs;
2220 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2221 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2222 	iflib_dma_info_t di;
2223 	iflib_fl_t fl;
2224 	int i, j;
2225 
2226 	/* Tell the stack that the interface is no longer active */
2227 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2228 
2229 	IFDI_INTR_DISABLE(ctx);
2230 	DELAY(1000);
2231 	IFDI_STOP(ctx);
2232 	DELAY(1000);
2233 
2234 	iflib_debug_reset();
2235 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2236 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2237 		/* make sure all transmitters have completed before proceeding XXX */
2238 
2239 		/* clean any enqueued buffers */
2240 		iflib_ifmp_purge(txq);
2241 		/* Free any existing tx buffers. */
2242 		for (j = 0; j < txq->ift_size; j++) {
2243 			iflib_txsd_free(ctx, txq, j);
2244 		}
2245 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2246 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2247 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2248 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2249 		txq->ift_pullups = 0;
2250 		ifmp_ring_reset_stats(txq->ift_br);
2251 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2252 			bzero((void *)di->idi_vaddr, di->idi_size);
2253 	}
2254 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2255 		/* make sure all transmitters have completed before proceeding XXX */
2256 
2257 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2258 			bzero((void *)di->idi_vaddr, di->idi_size);
2259 		/* also resets the free lists pidx/cidx */
2260 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2261 			iflib_fl_bufs_free(fl);
2262 	}
2263 }
2264 
2265 static inline caddr_t
2266 calc_next_rxd(iflib_fl_t fl, int cidx)
2267 {
2268 	qidx_t size;
2269 	int nrxd;
2270 	caddr_t start, end, cur, next;
2271 
2272 	nrxd = fl->ifl_size;
2273 	size = fl->ifl_rxd_size;
2274 	start = fl->ifl_ifdi->idi_vaddr;
2275 
2276 	if (__predict_false(size == 0))
2277 		return (start);
2278 	cur = start + size*cidx;
2279 	end = start + size*nrxd;
2280 	next = CACHE_PTR_NEXT(cur);
2281 	return (next < end ? next : start);
2282 }
2283 
2284 static inline void
2285 prefetch_pkts(iflib_fl_t fl, int cidx)
2286 {
2287 	int nextptr;
2288 	int nrxd = fl->ifl_size;
2289 	caddr_t next_rxd;
2290 
2291 
2292 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2293 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2294 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2295 	next_rxd = calc_next_rxd(fl, cidx);
2296 	prefetch(next_rxd);
2297 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2298 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2299 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2300 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2301 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2302 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2303 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2304 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2305 }
2306 
2307 static void
2308 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2309 {
2310 	int flid, cidx;
2311 	bus_dmamap_t map;
2312 	iflib_fl_t fl;
2313 	iflib_dma_info_t di;
2314 	int next;
2315 
2316 	map = NULL;
2317 	flid = irf->irf_flid;
2318 	cidx = irf->irf_idx;
2319 	fl = &rxq->ifr_fl[flid];
2320 	sd->ifsd_fl = fl;
2321 	sd->ifsd_cidx = cidx;
2322 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2323 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2324 	fl->ifl_credits--;
2325 #if MEMORY_LOGGING
2326 	fl->ifl_m_dequeued++;
2327 #endif
2328 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2329 		prefetch_pkts(fl, cidx);
2330 	if (fl->ifl_sds.ifsd_map != NULL) {
2331 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2332 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2333 		map = fl->ifl_sds.ifsd_map[cidx];
2334 		di = fl->ifl_ifdi;
2335 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2336 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2337 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2338 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2339 
2340 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2341 		MPASS(fl->ifl_cidx == cidx);
2342 		if (unload)
2343 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2344 	}
2345 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2346 	if (__predict_false(fl->ifl_cidx == 0))
2347 		fl->ifl_gen = 0;
2348 	if (map != NULL)
2349 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2350 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2351 }
2352 
2353 static struct mbuf *
2354 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2355 {
2356 	int i, padlen , flags;
2357 	struct mbuf *m, *mh, *mt;
2358 	caddr_t cl;
2359 
2360 	i = 0;
2361 	mh = NULL;
2362 	do {
2363 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2364 
2365 		MPASS(*sd->ifsd_cl != NULL);
2366 		MPASS(*sd->ifsd_m != NULL);
2367 
2368 		/* Don't include zero-length frags */
2369 		if (ri->iri_frags[i].irf_len == 0) {
2370 			/* XXX we can save the cluster here, but not the mbuf */
2371 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2372 			m_free(*sd->ifsd_m);
2373 			*sd->ifsd_m = NULL;
2374 			continue;
2375 		}
2376 		m = *sd->ifsd_m;
2377 		*sd->ifsd_m = NULL;
2378 		if (mh == NULL) {
2379 			flags = M_PKTHDR|M_EXT;
2380 			mh = mt = m;
2381 			padlen = ri->iri_pad;
2382 		} else {
2383 			flags = M_EXT;
2384 			mt->m_next = m;
2385 			mt = m;
2386 			/* assuming padding is only on the first fragment */
2387 			padlen = 0;
2388 		}
2389 		cl = *sd->ifsd_cl;
2390 		*sd->ifsd_cl = NULL;
2391 
2392 		/* Can these two be made one ? */
2393 		m_init(m, M_NOWAIT, MT_DATA, flags);
2394 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2395 		/*
2396 		 * These must follow m_init and m_cljset
2397 		 */
2398 		m->m_data += padlen;
2399 		ri->iri_len -= padlen;
2400 		m->m_len = ri->iri_frags[i].irf_len;
2401 	} while (++i < ri->iri_nfrags);
2402 
2403 	return (mh);
2404 }
2405 
2406 /*
2407  * Process one software descriptor
2408  */
2409 static struct mbuf *
2410 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2411 {
2412 	struct if_rxsd sd;
2413 	struct mbuf *m;
2414 
2415 	/* should I merge this back in now that the two paths are basically duplicated? */
2416 	if (ri->iri_nfrags == 1 &&
2417 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2418 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2419 		m = *sd.ifsd_m;
2420 		*sd.ifsd_m = NULL;
2421 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2422 #ifndef __NO_STRICT_ALIGNMENT
2423 		if (!IP_ALIGNED(m))
2424 			m->m_data += 2;
2425 #endif
2426 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2427 		m->m_len = ri->iri_frags[0].irf_len;
2428        } else {
2429 		m = assemble_segments(rxq, ri, &sd);
2430 	}
2431 	m->m_pkthdr.len = ri->iri_len;
2432 	m->m_pkthdr.rcvif = ri->iri_ifp;
2433 	m->m_flags |= ri->iri_flags;
2434 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2435 	m->m_pkthdr.flowid = ri->iri_flowid;
2436 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2437 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2438 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2439 	return (m);
2440 }
2441 
2442 static bool
2443 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2444 {
2445 	if_ctx_t ctx = rxq->ifr_ctx;
2446 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2447 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2448 	int avail, i;
2449 	qidx_t *cidxp;
2450 	struct if_rxd_info ri;
2451 	int err, budget_left, rx_bytes, rx_pkts;
2452 	iflib_fl_t fl;
2453 	struct ifnet *ifp;
2454 	int lro_enabled;
2455 
2456 	/*
2457 	 * XXX early demux data packets so that if_input processing only handles
2458 	 * acks in interrupt context
2459 	 */
2460 	struct mbuf *m, *mh, *mt;
2461 
2462 	ifp = ctx->ifc_ifp;
2463 #ifdef DEV_NETMAP
2464 	if (ifp->if_capenable & IFCAP_NETMAP) {
2465 		u_int work = 0;
2466 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2467 			return (FALSE);
2468 	}
2469 #endif
2470 
2471 	mh = mt = NULL;
2472 	MPASS(budget > 0);
2473 	rx_pkts	= rx_bytes = 0;
2474 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2475 		cidxp = &rxq->ifr_cq_cidx;
2476 	else
2477 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2478 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2479 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2480 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2481 		DBG_COUNTER_INC(rx_unavail);
2482 		return (false);
2483 	}
2484 
2485 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2486 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2487 			DBG_COUNTER_INC(rx_ctx_inactive);
2488 			break;
2489 		}
2490 		/*
2491 		 * Reset client set fields to their default values
2492 		 */
2493 		rxd_info_zero(&ri);
2494 		ri.iri_qsidx = rxq->ifr_id;
2495 		ri.iri_cidx = *cidxp;
2496 		ri.iri_ifp = ifp;
2497 		ri.iri_frags = rxq->ifr_frags;
2498 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2499 
2500 		if (err)
2501 			goto err;
2502 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2503 			*cidxp = ri.iri_cidx;
2504 			/* Update our consumer index */
2505 			/* XXX NB: shurd - check if this is still safe */
2506 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2507 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2508 				rxq->ifr_cq_gen = 0;
2509 			}
2510 			/* was this only a completion queue message? */
2511 			if (__predict_false(ri.iri_nfrags == 0))
2512 				continue;
2513 		}
2514 		MPASS(ri.iri_nfrags != 0);
2515 		MPASS(ri.iri_len != 0);
2516 
2517 		/* will advance the cidx on the corresponding free lists */
2518 		m = iflib_rxd_pkt_get(rxq, &ri);
2519 		if (avail == 0 && budget_left)
2520 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2521 
2522 		if (__predict_false(m == NULL)) {
2523 			DBG_COUNTER_INC(rx_mbuf_null);
2524 			continue;
2525 		}
2526 		/* imm_pkt: -- cxgb */
2527 		if (mh == NULL)
2528 			mh = mt = m;
2529 		else {
2530 			mt->m_nextpkt = m;
2531 			mt = m;
2532 		}
2533 	}
2534 	/* make sure that we can refill faster than drain */
2535 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2536 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2537 
2538 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2539 	while (mh != NULL) {
2540 		m = mh;
2541 		mh = mh->m_nextpkt;
2542 		m->m_nextpkt = NULL;
2543 #ifndef __NO_STRICT_ALIGNMENT
2544 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2545 			continue;
2546 #endif
2547 		rx_bytes += m->m_pkthdr.len;
2548 		rx_pkts++;
2549 #if defined(INET6) || defined(INET)
2550 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2551 			continue;
2552 #endif
2553 		DBG_COUNTER_INC(rx_if_input);
2554 		ifp->if_input(ifp, m);
2555 	}
2556 
2557 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2558 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2559 
2560 	/*
2561 	 * Flush any outstanding LRO work
2562 	 */
2563 #if defined(INET6) || defined(INET)
2564 	tcp_lro_flush_all(&rxq->ifr_lc);
2565 #endif
2566 	if (avail)
2567 		return true;
2568 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2569 err:
2570 	CTX_LOCK(ctx);
2571 	ctx->ifc_flags |= IFC_DO_RESET;
2572 	iflib_admin_intr_deferred(ctx);
2573 	CTX_UNLOCK(ctx);
2574 	return (false);
2575 }
2576 
2577 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2578 static inline qidx_t
2579 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2580 {
2581 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2582 	qidx_t minthresh = txq->ift_size / 8;
2583 	if (in_use > 4*minthresh)
2584 		return (notify_count);
2585 	if (in_use > 2*minthresh)
2586 		return (notify_count >> 1);
2587 	if (in_use > minthresh)
2588 		return (notify_count >> 3);
2589 	return (0);
2590 }
2591 
2592 static inline qidx_t
2593 txq_max_rs_deferred(iflib_txq_t txq)
2594 {
2595 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2596 	qidx_t minthresh = txq->ift_size / 8;
2597 	if (txq->ift_in_use > 4*minthresh)
2598 		return (notify_count);
2599 	if (txq->ift_in_use > 2*minthresh)
2600 		return (notify_count >> 1);
2601 	if (txq->ift_in_use > minthresh)
2602 		return (notify_count >> 2);
2603 	return (2);
2604 }
2605 
2606 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2607 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2608 
2609 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2610 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2611 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2612 
2613 /* forward compatibility for cxgb */
2614 #define FIRST_QSET(ctx) 0
2615 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2616 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2617 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2618 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2619 
2620 /* XXX we should be setting this to something other than zero */
2621 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2622 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2623 
2624 static inline bool
2625 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2626 {
2627 	qidx_t dbval, max;
2628 	bool rang;
2629 
2630 	rang = false;
2631 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2632 	if (ring || txq->ift_db_pending >= max) {
2633 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2634 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2635 		txq->ift_db_pending = txq->ift_npending = 0;
2636 		rang = true;
2637 	}
2638 	return (rang);
2639 }
2640 
2641 #ifdef PKT_DEBUG
2642 static void
2643 print_pkt(if_pkt_info_t pi)
2644 {
2645 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2646 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2647 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2648 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2649 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2650 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2651 }
2652 #endif
2653 
2654 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2655 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2656 
2657 static int
2658 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2659 {
2660 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2661 	struct ether_vlan_header *eh;
2662 	struct mbuf *m, *n;
2663 
2664 	n = m = *mp;
2665 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2666 	    M_WRITABLE(m) == 0) {
2667 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2668 			return (ENOMEM);
2669 		} else {
2670 			m_freem(*mp);
2671 			n = *mp = m;
2672 		}
2673 	}
2674 
2675 	/*
2676 	 * Determine where frame payload starts.
2677 	 * Jump over vlan headers if already present,
2678 	 * helpful for QinQ too.
2679 	 */
2680 	if (__predict_false(m->m_len < sizeof(*eh))) {
2681 		txq->ift_pullups++;
2682 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2683 			return (ENOMEM);
2684 	}
2685 	eh = mtod(m, struct ether_vlan_header *);
2686 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2687 		pi->ipi_etype = ntohs(eh->evl_proto);
2688 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2689 	} else {
2690 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2691 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2692 	}
2693 
2694 	switch (pi->ipi_etype) {
2695 #ifdef INET
2696 	case ETHERTYPE_IP:
2697 	{
2698 		struct ip *ip = NULL;
2699 		struct tcphdr *th = NULL;
2700 		int minthlen;
2701 
2702 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2703 		if (__predict_false(m->m_len < minthlen)) {
2704 			/*
2705 			 * if this code bloat is causing too much of a hit
2706 			 * move it to a separate function and mark it noinline
2707 			 */
2708 			if (m->m_len == pi->ipi_ehdrlen) {
2709 				n = m->m_next;
2710 				MPASS(n);
2711 				if (n->m_len >= sizeof(*ip))  {
2712 					ip = (struct ip *)n->m_data;
2713 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2714 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2715 				} else {
2716 					txq->ift_pullups++;
2717 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2718 						return (ENOMEM);
2719 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2720 				}
2721 			} else {
2722 				txq->ift_pullups++;
2723 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2724 					return (ENOMEM);
2725 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2726 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2727 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2728 			}
2729 		} else {
2730 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2731 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2732 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2733 		}
2734 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2735 		pi->ipi_ipproto = ip->ip_p;
2736 		pi->ipi_flags |= IPI_TX_IPV4;
2737 
2738 		if (pi->ipi_csum_flags & CSUM_IP)
2739                        ip->ip_sum = 0;
2740 
2741 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2742 			if (__predict_false(th == NULL)) {
2743 				txq->ift_pullups++;
2744 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2745 					return (ENOMEM);
2746 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2747 			}
2748 			pi->ipi_tcp_hflags = th->th_flags;
2749 			pi->ipi_tcp_hlen = th->th_off << 2;
2750 			pi->ipi_tcp_seq = th->th_seq;
2751 		}
2752 		if (IS_TSO4(pi)) {
2753 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2754 				return (ENXIO);
2755 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2756 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2757 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2758 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2759 				ip->ip_sum = 0;
2760 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2761 			}
2762 		}
2763 		break;
2764 	}
2765 #endif
2766 #ifdef INET6
2767 	case ETHERTYPE_IPV6:
2768 	{
2769 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2770 		struct tcphdr *th;
2771 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2772 
2773 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2774 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2775 				return (ENOMEM);
2776 		}
2777 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2778 
2779 		/* XXX-BZ this will go badly in case of ext hdrs. */
2780 		pi->ipi_ipproto = ip6->ip6_nxt;
2781 		pi->ipi_flags |= IPI_TX_IPV6;
2782 
2783 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2784 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2785 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2786 					return (ENOMEM);
2787 			}
2788 			pi->ipi_tcp_hflags = th->th_flags;
2789 			pi->ipi_tcp_hlen = th->th_off << 2;
2790 		}
2791 		if (IS_TSO6(pi)) {
2792 
2793 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2794 				return (ENXIO);
2795 			/*
2796 			 * The corresponding flag is set by the stack in the IPv4
2797 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2798 			 * So, set it here because the rest of the flow requires it.
2799 			 */
2800 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2801 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2802 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2803 		}
2804 		break;
2805 	}
2806 #endif
2807 	default:
2808 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2809 		pi->ipi_ip_hlen = 0;
2810 		break;
2811 	}
2812 	*mp = m;
2813 
2814 	return (0);
2815 }
2816 
2817 static  __noinline  struct mbuf *
2818 collapse_pkthdr(struct mbuf *m0)
2819 {
2820 	struct mbuf *m, *m_next, *tmp;
2821 
2822 	m = m0;
2823 	m_next = m->m_next;
2824 	while (m_next != NULL && m_next->m_len == 0) {
2825 		m = m_next;
2826 		m->m_next = NULL;
2827 		m_free(m);
2828 		m_next = m_next->m_next;
2829 	}
2830 	m = m0;
2831 	m->m_next = m_next;
2832 	if ((m_next->m_flags & M_EXT) == 0) {
2833 		m = m_defrag(m, M_NOWAIT);
2834 	} else {
2835 		tmp = m_next->m_next;
2836 		memcpy(m_next, m, MPKTHSIZE);
2837 		m = m_next;
2838 		m->m_next = tmp;
2839 	}
2840 	return (m);
2841 }
2842 
2843 /*
2844  * If dodgy hardware rejects the scatter gather chain we've handed it
2845  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2846  * m_defrag'd mbufs
2847  */
2848 static __noinline struct mbuf *
2849 iflib_remove_mbuf(iflib_txq_t txq)
2850 {
2851 	int ntxd, i, pidx;
2852 	struct mbuf *m, *mh, **ifsd_m;
2853 
2854 	pidx = txq->ift_pidx;
2855 	ifsd_m = txq->ift_sds.ifsd_m;
2856 	ntxd = txq->ift_size;
2857 	mh = m = ifsd_m[pidx];
2858 	ifsd_m[pidx] = NULL;
2859 #if MEMORY_LOGGING
2860 	txq->ift_dequeued++;
2861 #endif
2862 	i = 1;
2863 
2864 	while (m) {
2865 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2866 #if MEMORY_LOGGING
2867 		txq->ift_dequeued++;
2868 #endif
2869 		m = m->m_next;
2870 		i++;
2871 	}
2872 	return (mh);
2873 }
2874 
2875 static int
2876 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2877 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2878 			  int max_segs, int flags)
2879 {
2880 	if_ctx_t ctx;
2881 	if_shared_ctx_t		sctx;
2882 	if_softc_ctx_t		scctx;
2883 	int i, next, pidx, err, maxsegsz, ntxd, count;
2884 	struct mbuf *m, *tmp, **ifsd_m;
2885 
2886 	m = *m0;
2887 
2888 	/*
2889 	 * Please don't ever do this
2890 	 */
2891 	if (__predict_false(m->m_len == 0))
2892 		*m0 = m = collapse_pkthdr(m);
2893 
2894 	ctx = txq->ift_ctx;
2895 	sctx = ctx->ifc_sctx;
2896 	scctx = &ctx->ifc_softc_ctx;
2897 	ifsd_m = txq->ift_sds.ifsd_m;
2898 	ntxd = txq->ift_size;
2899 	pidx = txq->ift_pidx;
2900 	if (map != NULL) {
2901 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2902 
2903 		err = bus_dmamap_load_mbuf_sg(tag, map,
2904 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2905 		if (err)
2906 			return (err);
2907 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2908 		count = 0;
2909 		m = *m0;
2910 		do {
2911 			if (__predict_false(m->m_len <= 0)) {
2912 				tmp = m;
2913 				m = m->m_next;
2914 				tmp->m_next = NULL;
2915 				m_free(tmp);
2916 				continue;
2917 			}
2918 			next = (pidx + count) & (ntxd-1);
2919 			MPASS(ifsd_m[next] == NULL);
2920 			ifsd_m[next] = m;
2921 			count++;
2922 			tmp = m;
2923 			m = m->m_next;
2924 		} while (m != NULL);
2925 	} else {
2926 		int buflen, sgsize, max_sgsize;
2927 		vm_offset_t vaddr;
2928 		vm_paddr_t curaddr;
2929 
2930 		count = i = 0;
2931 		maxsegsz = sctx->isc_tx_maxsize;
2932 		m = *m0;
2933 		do {
2934 			if (__predict_false(m->m_len <= 0)) {
2935 				tmp = m;
2936 				m = m->m_next;
2937 				tmp->m_next = NULL;
2938 				m_free(tmp);
2939 				continue;
2940 			}
2941 			buflen = m->m_len;
2942 			vaddr = (vm_offset_t)m->m_data;
2943 			/*
2944 			 * see if we can't be smarter about physically
2945 			 * contiguous mappings
2946 			 */
2947 			next = (pidx + count) & (ntxd-1);
2948 			MPASS(ifsd_m[next] == NULL);
2949 #if MEMORY_LOGGING
2950 			txq->ift_enqueued++;
2951 #endif
2952 			ifsd_m[next] = m;
2953 			while (buflen > 0) {
2954 				max_sgsize = MIN(buflen, maxsegsz);
2955 				curaddr = pmap_kextract(vaddr);
2956 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2957 				sgsize = MIN(sgsize, max_sgsize);
2958 				segs[i].ds_addr = curaddr;
2959 				segs[i].ds_len = sgsize;
2960 				vaddr += sgsize;
2961 				buflen -= sgsize;
2962 				i++;
2963 				if (i >= max_segs)
2964 					goto err;
2965 			}
2966 			count++;
2967 			tmp = m;
2968 			m = m->m_next;
2969 		} while (m != NULL);
2970 		*nsegs = i;
2971 	}
2972 	return (0);
2973 err:
2974 	*m0 = iflib_remove_mbuf(txq);
2975 	return (EFBIG);
2976 }
2977 
2978 static inline caddr_t
2979 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2980 {
2981 	qidx_t size;
2982 	int ntxd;
2983 	caddr_t start, end, cur, next;
2984 
2985 	ntxd = txq->ift_size;
2986 	size = txq->ift_txd_size[qid];
2987 	start = txq->ift_ifdi[qid].idi_vaddr;
2988 
2989 	if (__predict_false(size == 0))
2990 		return (start);
2991 	cur = start + size*cidx;
2992 	end = start + size*ntxd;
2993 	next = CACHE_PTR_NEXT(cur);
2994 	return (next < end ? next : start);
2995 }
2996 
2997 static int
2998 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2999 {
3000 	if_ctx_t		ctx;
3001 	if_shared_ctx_t		sctx;
3002 	if_softc_ctx_t		scctx;
3003 	bus_dma_segment_t	*segs;
3004 	struct mbuf		*m_head;
3005 	void			*next_txd;
3006 	bus_dmamap_t		map;
3007 	struct if_pkt_info	pi;
3008 	int remap = 0;
3009 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3010 	bus_dma_tag_t desc_tag;
3011 
3012 	segs = txq->ift_segs;
3013 	ctx = txq->ift_ctx;
3014 	sctx = ctx->ifc_sctx;
3015 	scctx = &ctx->ifc_softc_ctx;
3016 	segs = txq->ift_segs;
3017 	ntxd = txq->ift_size;
3018 	m_head = *m_headp;
3019 	map = NULL;
3020 
3021 	/*
3022 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3023 	 */
3024 	cidx = txq->ift_cidx;
3025 	pidx = txq->ift_pidx;
3026 	if (ctx->ifc_flags & IFC_PREFETCH) {
3027 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3028 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3029 			next_txd = calc_next_txd(txq, cidx, 0);
3030 			prefetch(next_txd);
3031 		}
3032 
3033 		/* prefetch the next cache line of mbuf pointers and flags */
3034 		prefetch(&txq->ift_sds.ifsd_m[next]);
3035 		if (txq->ift_sds.ifsd_map != NULL) {
3036 			prefetch(&txq->ift_sds.ifsd_map[next]);
3037 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3038 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3039 		}
3040 	} else if (txq->ift_sds.ifsd_map != NULL)
3041 		map = txq->ift_sds.ifsd_map[pidx];
3042 
3043 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3044 		desc_tag = txq->ift_tso_desc_tag;
3045 		max_segs = scctx->isc_tx_tso_segments_max;
3046 	} else {
3047 		desc_tag = txq->ift_desc_tag;
3048 		max_segs = scctx->isc_tx_nsegments;
3049 	}
3050 	m_head = *m_headp;
3051 
3052 	pkt_info_zero(&pi);
3053 	pi.ipi_len = m_head->m_pkthdr.len;
3054 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3055 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3056 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3057 	pi.ipi_pidx = pidx;
3058 	pi.ipi_qsidx = txq->ift_id;
3059 
3060 	/* deliberate bitwise OR to make one condition */
3061 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3062 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3063 			return (err);
3064 		m_head = *m_headp;
3065 	}
3066 
3067 retry:
3068 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3069 defrag:
3070 	if (__predict_false(err)) {
3071 		switch (err) {
3072 		case EFBIG:
3073 			/* try collapse once and defrag once */
3074 			if (remap == 0)
3075 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3076 			if (remap == 1)
3077 				m_head = m_defrag(*m_headp, M_NOWAIT);
3078 			remap++;
3079 			if (__predict_false(m_head == NULL))
3080 				goto defrag_failed;
3081 			txq->ift_mbuf_defrag++;
3082 			*m_headp = m_head;
3083 			goto retry;
3084 			break;
3085 		case ENOMEM:
3086 			txq->ift_no_tx_dma_setup++;
3087 			break;
3088 		default:
3089 			txq->ift_no_tx_dma_setup++;
3090 			m_freem(*m_headp);
3091 			DBG_COUNTER_INC(tx_frees);
3092 			*m_headp = NULL;
3093 			break;
3094 		}
3095 		txq->ift_map_failed++;
3096 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3097 		return (err);
3098 	}
3099 
3100 	/*
3101 	 * XXX assumes a 1 to 1 relationship between segments and
3102 	 *        descriptors - this does not hold true on all drivers, e.g.
3103 	 *        cxgb
3104 	 */
3105 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3106 		txq->ift_no_desc_avail++;
3107 		if (map != NULL)
3108 			bus_dmamap_unload(desc_tag, map);
3109 		DBG_COUNTER_INC(encap_txq_avail_fail);
3110 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3111 			GROUPTASK_ENQUEUE(&txq->ift_task);
3112 		return (ENOBUFS);
3113 	}
3114 	/*
3115 	 * On Intel cards we can greatly reduce the number of TX interrupts
3116 	 * we see by only setting report status on every Nth descriptor.
3117 	 * However, this also means that the driver will need to keep track
3118 	 * of the descriptors that RS was set on to check them for the DD bit.
3119 	 */
3120 	txq->ift_rs_pending += nsegs + 1;
3121 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3122 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3123 		pi.ipi_flags |= IPI_TX_INTR;
3124 		txq->ift_rs_pending = 0;
3125 	}
3126 
3127 	pi.ipi_segs = segs;
3128 	pi.ipi_nsegs = nsegs;
3129 
3130 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3131 #ifdef PKT_DEBUG
3132 	print_pkt(&pi);
3133 #endif
3134 	if (map != NULL)
3135 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3136 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3137 		if (map != NULL)
3138 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3139 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3140 		DBG_COUNTER_INC(tx_encap);
3141 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3142 
3143 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3144 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3145 			ndesc += txq->ift_size;
3146 			txq->ift_gen = 1;
3147 		}
3148 		/*
3149 		 * drivers can need as many as
3150 		 * two sentinels
3151 		 */
3152 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3153 		MPASS(pi.ipi_new_pidx != pidx);
3154 		MPASS(ndesc > 0);
3155 		txq->ift_in_use += ndesc;
3156 
3157 		/*
3158 		 * We update the last software descriptor again here because there may
3159 		 * be a sentinel and/or there may be more mbufs than segments
3160 		 */
3161 		txq->ift_pidx = pi.ipi_new_pidx;
3162 		txq->ift_npending += pi.ipi_ndescs;
3163 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3164 		*m_headp = m_head = iflib_remove_mbuf(txq);
3165 		remap = 1;
3166 		txq->ift_txd_encap_efbig++;
3167 		goto defrag;
3168 	} else
3169 		DBG_COUNTER_INC(encap_txd_encap_fail);
3170 	return (err);
3171 
3172 defrag_failed:
3173 	txq->ift_mbuf_defrag_failed++;
3174 	txq->ift_map_failed++;
3175 	m_freem(*m_headp);
3176 	DBG_COUNTER_INC(tx_frees);
3177 	*m_headp = NULL;
3178 	return (ENOMEM);
3179 }
3180 
3181 static void
3182 iflib_tx_desc_free(iflib_txq_t txq, int n)
3183 {
3184 	int hasmap;
3185 	uint32_t qsize, cidx, mask, gen;
3186 	struct mbuf *m, **ifsd_m;
3187 	uint8_t *ifsd_flags;
3188 	bus_dmamap_t *ifsd_map;
3189 	bool do_prefetch;
3190 
3191 	cidx = txq->ift_cidx;
3192 	gen = txq->ift_gen;
3193 	qsize = txq->ift_size;
3194 	mask = qsize-1;
3195 	hasmap = txq->ift_sds.ifsd_map != NULL;
3196 	ifsd_flags = txq->ift_sds.ifsd_flags;
3197 	ifsd_m = txq->ift_sds.ifsd_m;
3198 	ifsd_map = txq->ift_sds.ifsd_map;
3199 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3200 
3201 	while (n--) {
3202 		if (do_prefetch) {
3203 			prefetch(ifsd_m[(cidx + 3) & mask]);
3204 			prefetch(ifsd_m[(cidx + 4) & mask]);
3205 		}
3206 		if (ifsd_m[cidx] != NULL) {
3207 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3208 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3209 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3210 				/*
3211 				 * does it matter if it's not the TSO tag? If so we'll
3212 				 * have to add the type to flags
3213 				 */
3214 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3215 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3216 			}
3217 			if ((m = ifsd_m[cidx]) != NULL) {
3218 				/* XXX we don't support any drivers that batch packets yet */
3219 				MPASS(m->m_nextpkt == NULL);
3220 
3221 				m_free(m);
3222 				ifsd_m[cidx] = NULL;
3223 #if MEMORY_LOGGING
3224 				txq->ift_dequeued++;
3225 #endif
3226 				DBG_COUNTER_INC(tx_frees);
3227 			}
3228 		}
3229 		if (__predict_false(++cidx == qsize)) {
3230 			cidx = 0;
3231 			gen = 0;
3232 		}
3233 	}
3234 	txq->ift_cidx = cidx;
3235 	txq->ift_gen = gen;
3236 }
3237 
3238 static __inline int
3239 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3240 {
3241 	int reclaim;
3242 	if_ctx_t ctx = txq->ift_ctx;
3243 
3244 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3245 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3246 
3247 	/*
3248 	 * Need a rate-limiting check so that this isn't called every time
3249 	 */
3250 	iflib_tx_credits_update(ctx, txq);
3251 	reclaim = DESC_RECLAIMABLE(txq);
3252 
3253 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3254 #ifdef INVARIANTS
3255 		if (iflib_verbose_debug) {
3256 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3257 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3258 			       reclaim, thresh);
3259 
3260 		}
3261 #endif
3262 		return (0);
3263 	}
3264 	iflib_tx_desc_free(txq, reclaim);
3265 	txq->ift_cleaned += reclaim;
3266 	txq->ift_in_use -= reclaim;
3267 
3268 	return (reclaim);
3269 }
3270 
3271 static struct mbuf **
3272 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3273 {
3274 	int next, size;
3275 	struct mbuf **items;
3276 
3277 	size = r->size;
3278 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3279 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3280 
3281 	prefetch(items[(cidx + offset) & (size-1)]);
3282 	if (remaining > 1) {
3283 		prefetch(&items[next]);
3284 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3285 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3286 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3287 	}
3288 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3289 }
3290 
3291 static void
3292 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3293 {
3294 
3295 	ifmp_ring_check_drainage(txq->ift_br, budget);
3296 }
3297 
3298 static uint32_t
3299 iflib_txq_can_drain(struct ifmp_ring *r)
3300 {
3301 	iflib_txq_t txq = r->cookie;
3302 	if_ctx_t ctx = txq->ift_ctx;
3303 
3304 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3305 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3306 }
3307 
3308 static uint32_t
3309 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3310 {
3311 	iflib_txq_t txq = r->cookie;
3312 	if_ctx_t ctx = txq->ift_ctx;
3313 	struct ifnet *ifp = ctx->ifc_ifp;
3314 	struct mbuf **mp, *m;
3315 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3316 	int reclaimed, err, in_use_prev, desc_used;
3317 	bool do_prefetch, ring, rang;
3318 
3319 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3320 			    !LINK_ACTIVE(ctx))) {
3321 		DBG_COUNTER_INC(txq_drain_notready);
3322 		return (0);
3323 	}
3324 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3325 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3326 	avail = IDXDIFF(pidx, cidx, r->size);
3327 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3328 		DBG_COUNTER_INC(txq_drain_flushing);
3329 		for (i = 0; i < avail; i++) {
3330 			m_free(r->items[(cidx + i) & (r->size-1)]);
3331 			r->items[(cidx + i) & (r->size-1)] = NULL;
3332 		}
3333 		return (avail);
3334 	}
3335 
3336 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3337 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3338 		CALLOUT_LOCK(txq);
3339 		callout_stop(&txq->ift_timer);
3340 		CALLOUT_UNLOCK(txq);
3341 		DBG_COUNTER_INC(txq_drain_oactive);
3342 		return (0);
3343 	}
3344 	if (reclaimed)
3345 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3346 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3347 	count = MIN(avail, TX_BATCH_SIZE);
3348 #ifdef INVARIANTS
3349 	if (iflib_verbose_debug)
3350 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3351 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3352 #endif
3353 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3354 	avail = TXQ_AVAIL(txq);
3355 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3356 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3357 
3358 		mp = _ring_peek_one(r, cidx, i, rem);
3359 		MPASS(mp != NULL && *mp != NULL);
3360 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3361 			consumed++;
3362 			reclaimed++;
3363 			continue;
3364 		}
3365 		in_use_prev = txq->ift_in_use;
3366 		pidx_prev = txq->ift_pidx;
3367 		err = iflib_encap(txq, mp);
3368 		if (__predict_false(err)) {
3369 			DBG_COUNTER_INC(txq_drain_encapfail);
3370 			/* no room - bail out */
3371 			if (err == ENOBUFS)
3372 				break;
3373 			consumed++;
3374 			DBG_COUNTER_INC(txq_drain_encapfail);
3375 			/* we can't send this packet - skip it */
3376 			continue;
3377 		}
3378 		consumed++;
3379 		pkt_sent++;
3380 		m = *mp;
3381 		DBG_COUNTER_INC(tx_sent);
3382 		bytes_sent += m->m_pkthdr.len;
3383 		mcast_sent += !!(m->m_flags & M_MCAST);
3384 		avail = TXQ_AVAIL(txq);
3385 
3386 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3387 		desc_used += (txq->ift_in_use - in_use_prev);
3388 		ETHER_BPF_MTAP(ifp, m);
3389 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3390 			break;
3391 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3392 	}
3393 
3394 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3395 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3396 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3397 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3398 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3399 	if (mcast_sent)
3400 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3401 #ifdef INVARIANTS
3402 	if (iflib_verbose_debug)
3403 		printf("consumed=%d\n", consumed);
3404 #endif
3405 	return (consumed);
3406 }
3407 
3408 static uint32_t
3409 iflib_txq_drain_always(struct ifmp_ring *r)
3410 {
3411 	return (1);
3412 }
3413 
3414 static uint32_t
3415 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3416 {
3417 	int i, avail;
3418 	struct mbuf **mp;
3419 	iflib_txq_t txq;
3420 
3421 	txq = r->cookie;
3422 
3423 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3424 	CALLOUT_LOCK(txq);
3425 	callout_stop(&txq->ift_timer);
3426 	CALLOUT_UNLOCK(txq);
3427 
3428 	avail = IDXDIFF(pidx, cidx, r->size);
3429 	for (i = 0; i < avail; i++) {
3430 		mp = _ring_peek_one(r, cidx, i, avail - i);
3431 		if (__predict_false(*mp == (struct mbuf *)txq))
3432 			continue;
3433 		m_freem(*mp);
3434 	}
3435 	MPASS(ifmp_ring_is_stalled(r) == 0);
3436 	return (avail);
3437 }
3438 
3439 static void
3440 iflib_ifmp_purge(iflib_txq_t txq)
3441 {
3442 	struct ifmp_ring *r;
3443 
3444 	r = txq->ift_br;
3445 	r->drain = iflib_txq_drain_free;
3446 	r->can_drain = iflib_txq_drain_always;
3447 
3448 	ifmp_ring_check_drainage(r, r->size);
3449 
3450 	r->drain = iflib_txq_drain;
3451 	r->can_drain = iflib_txq_can_drain;
3452 }
3453 
3454 static void
3455 _task_fn_tx(void *context)
3456 {
3457 	iflib_txq_t txq = context;
3458 	if_ctx_t ctx = txq->ift_ctx;
3459 	struct ifnet *ifp = ctx->ifc_ifp;
3460 	int rc;
3461 
3462 #ifdef IFLIB_DIAGNOSTICS
3463 	txq->ift_cpu_exec_count[curcpu]++;
3464 #endif
3465 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3466 		return;
3467 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3468 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3469 			netmap_tx_irq(ifp, txq->ift_id);
3470 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3471 		return;
3472 	}
3473 	if (txq->ift_db_pending)
3474 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3475 	else
3476 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3477 	if (ctx->ifc_flags & IFC_LEGACY)
3478 		IFDI_INTR_ENABLE(ctx);
3479 	else {
3480 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3481 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3482 	}
3483 }
3484 
3485 static void
3486 _task_fn_rx(void *context)
3487 {
3488 	iflib_rxq_t rxq = context;
3489 	if_ctx_t ctx = rxq->ifr_ctx;
3490 	bool more;
3491 	int rc;
3492 
3493 #ifdef IFLIB_DIAGNOSTICS
3494 	rxq->ifr_cpu_exec_count[curcpu]++;
3495 #endif
3496 	DBG_COUNTER_INC(task_fn_rxs);
3497 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3498 		return;
3499 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3500 		if (ctx->ifc_flags & IFC_LEGACY)
3501 			IFDI_INTR_ENABLE(ctx);
3502 		else {
3503 			DBG_COUNTER_INC(rx_intr_enables);
3504 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3505 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3506 		}
3507 	}
3508 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3509 		return;
3510 	if (more)
3511 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3512 }
3513 
3514 static void
3515 _task_fn_admin(void *context)
3516 {
3517 	if_ctx_t ctx = context;
3518 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3519 	iflib_txq_t txq;
3520 	int i;
3521 
3522 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3523 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3524 			return;
3525 		}
3526 	}
3527 
3528 	CTX_LOCK(ctx);
3529 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3530 		CALLOUT_LOCK(txq);
3531 		callout_stop(&txq->ift_timer);
3532 		CALLOUT_UNLOCK(txq);
3533 	}
3534 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3535 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3536 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3537 	IFDI_LINK_INTR_ENABLE(ctx);
3538 	if (ctx->ifc_flags & IFC_DO_RESET) {
3539 		ctx->ifc_flags &= ~IFC_DO_RESET;
3540 		iflib_if_init_locked(ctx);
3541 	}
3542 	CTX_UNLOCK(ctx);
3543 
3544 	if (LINK_ACTIVE(ctx) == 0)
3545 		return;
3546 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3547 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3548 }
3549 
3550 
3551 static void
3552 _task_fn_iov(void *context)
3553 {
3554 	if_ctx_t ctx = context;
3555 
3556 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3557 		return;
3558 
3559 	CTX_LOCK(ctx);
3560 	IFDI_VFLR_HANDLE(ctx);
3561 	CTX_UNLOCK(ctx);
3562 }
3563 
3564 static int
3565 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3566 {
3567 	int err;
3568 	if_int_delay_info_t info;
3569 	if_ctx_t ctx;
3570 
3571 	info = (if_int_delay_info_t)arg1;
3572 	ctx = info->iidi_ctx;
3573 	info->iidi_req = req;
3574 	info->iidi_oidp = oidp;
3575 	CTX_LOCK(ctx);
3576 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3577 	CTX_UNLOCK(ctx);
3578 	return (err);
3579 }
3580 
3581 /*********************************************************************
3582  *
3583  *  IFNET FUNCTIONS
3584  *
3585  **********************************************************************/
3586 
3587 static void
3588 iflib_if_init_locked(if_ctx_t ctx)
3589 {
3590 	iflib_stop(ctx);
3591 	iflib_init_locked(ctx);
3592 }
3593 
3594 
3595 static void
3596 iflib_if_init(void *arg)
3597 {
3598 	if_ctx_t ctx = arg;
3599 
3600 	CTX_LOCK(ctx);
3601 	iflib_if_init_locked(ctx);
3602 	CTX_UNLOCK(ctx);
3603 }
3604 
3605 static int
3606 iflib_if_transmit(if_t ifp, struct mbuf *m)
3607 {
3608 	if_ctx_t	ctx = if_getsoftc(ifp);
3609 
3610 	iflib_txq_t txq;
3611 	int err, qidx;
3612 
3613 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3614 		DBG_COUNTER_INC(tx_frees);
3615 		m_freem(m);
3616 		return (ENOBUFS);
3617 	}
3618 
3619 	MPASS(m->m_nextpkt == NULL);
3620 	qidx = 0;
3621 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3622 		qidx = QIDX(ctx, m);
3623 	/*
3624 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3625 	 */
3626 	txq = &ctx->ifc_txqs[qidx];
3627 
3628 #ifdef DRIVER_BACKPRESSURE
3629 	if (txq->ift_closed) {
3630 		while (m != NULL) {
3631 			next = m->m_nextpkt;
3632 			m->m_nextpkt = NULL;
3633 			m_freem(m);
3634 			m = next;
3635 		}
3636 		return (ENOBUFS);
3637 	}
3638 #endif
3639 #ifdef notyet
3640 	qidx = count = 0;
3641 	mp = marr;
3642 	next = m;
3643 	do {
3644 		count++;
3645 		next = next->m_nextpkt;
3646 	} while (next != NULL);
3647 
3648 	if (count > nitems(marr))
3649 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3650 			/* XXX check nextpkt */
3651 			m_freem(m);
3652 			/* XXX simplify for now */
3653 			DBG_COUNTER_INC(tx_frees);
3654 			return (ENOBUFS);
3655 		}
3656 	for (next = m, i = 0; next != NULL; i++) {
3657 		mp[i] = next;
3658 		next = next->m_nextpkt;
3659 		mp[i]->m_nextpkt = NULL;
3660 	}
3661 #endif
3662 	DBG_COUNTER_INC(tx_seen);
3663 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3664 
3665 	if (err) {
3666 		GROUPTASK_ENQUEUE(&txq->ift_task);
3667 		/* support forthcoming later */
3668 #ifdef DRIVER_BACKPRESSURE
3669 		txq->ift_closed = TRUE;
3670 #endif
3671 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3672 		m_freem(m);
3673 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3674 		GROUPTASK_ENQUEUE(&txq->ift_task);
3675 	}
3676 
3677 	return (err);
3678 }
3679 
3680 static void
3681 iflib_if_qflush(if_t ifp)
3682 {
3683 	if_ctx_t ctx = if_getsoftc(ifp);
3684 	iflib_txq_t txq = ctx->ifc_txqs;
3685 	int i;
3686 
3687 	CTX_LOCK(ctx);
3688 	ctx->ifc_flags |= IFC_QFLUSH;
3689 	CTX_UNLOCK(ctx);
3690 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3691 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3692 			iflib_txq_check_drain(txq, 0);
3693 	CTX_LOCK(ctx);
3694 	ctx->ifc_flags &= ~IFC_QFLUSH;
3695 	CTX_UNLOCK(ctx);
3696 
3697 	if_qflush(ifp);
3698 }
3699 
3700 
3701 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3702 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3703 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3704 
3705 static int
3706 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3707 {
3708 	if_ctx_t ctx = if_getsoftc(ifp);
3709 	struct ifreq	*ifr = (struct ifreq *)data;
3710 #if defined(INET) || defined(INET6)
3711 	struct ifaddr	*ifa = (struct ifaddr *)data;
3712 #endif
3713 	bool		avoid_reset = FALSE;
3714 	int		err = 0, reinit = 0, bits;
3715 
3716 	switch (command) {
3717 	case SIOCSIFADDR:
3718 #ifdef INET
3719 		if (ifa->ifa_addr->sa_family == AF_INET)
3720 			avoid_reset = TRUE;
3721 #endif
3722 #ifdef INET6
3723 		if (ifa->ifa_addr->sa_family == AF_INET6)
3724 			avoid_reset = TRUE;
3725 #endif
3726 		/*
3727 		** Calling init results in link renegotiation,
3728 		** so we avoid doing it when possible.
3729 		*/
3730 		if (avoid_reset) {
3731 			if_setflagbits(ifp, IFF_UP,0);
3732 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3733 				reinit = 1;
3734 #ifdef INET
3735 			if (!(if_getflags(ifp) & IFF_NOARP))
3736 				arp_ifinit(ifp, ifa);
3737 #endif
3738 		} else
3739 			err = ether_ioctl(ifp, command, data);
3740 		break;
3741 	case SIOCSIFMTU:
3742 		CTX_LOCK(ctx);
3743 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3744 			CTX_UNLOCK(ctx);
3745 			break;
3746 		}
3747 		bits = if_getdrvflags(ifp);
3748 		/* stop the driver and free any clusters before proceeding */
3749 		iflib_stop(ctx);
3750 
3751 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3752 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3753 				ctx->ifc_flags |= IFC_MULTISEG;
3754 			else
3755 				ctx->ifc_flags &= ~IFC_MULTISEG;
3756 			err = if_setmtu(ifp, ifr->ifr_mtu);
3757 		}
3758 		iflib_init_locked(ctx);
3759 		if_setdrvflags(ifp, bits);
3760 		CTX_UNLOCK(ctx);
3761 		break;
3762 	case SIOCSIFFLAGS:
3763 		CTX_LOCK(ctx);
3764 		if (if_getflags(ifp) & IFF_UP) {
3765 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3766 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3767 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3768 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3769 				}
3770 			} else
3771 				reinit = 1;
3772 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3773 			iflib_stop(ctx);
3774 		}
3775 		ctx->ifc_if_flags = if_getflags(ifp);
3776 		CTX_UNLOCK(ctx);
3777 		break;
3778 	case SIOCADDMULTI:
3779 	case SIOCDELMULTI:
3780 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3781 			CTX_LOCK(ctx);
3782 			IFDI_INTR_DISABLE(ctx);
3783 			IFDI_MULTI_SET(ctx);
3784 			IFDI_INTR_ENABLE(ctx);
3785 			CTX_UNLOCK(ctx);
3786 		}
3787 		break;
3788 	case SIOCSIFMEDIA:
3789 		CTX_LOCK(ctx);
3790 		IFDI_MEDIA_SET(ctx);
3791 		CTX_UNLOCK(ctx);
3792 		/* falls thru */
3793 	case SIOCGIFMEDIA:
3794 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3795 		break;
3796 	case SIOCGI2C:
3797 	{
3798 		struct ifi2creq i2c;
3799 
3800 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3801 		if (err != 0)
3802 			break;
3803 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3804 			err = EINVAL;
3805 			break;
3806 		}
3807 		if (i2c.len > sizeof(i2c.data)) {
3808 			err = EINVAL;
3809 			break;
3810 		}
3811 
3812 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3813 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3814 		break;
3815 	}
3816 	case SIOCSIFCAP:
3817 	{
3818 		int mask, setmask;
3819 
3820 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3821 		setmask = 0;
3822 #ifdef TCP_OFFLOAD
3823 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3824 #endif
3825 		setmask |= (mask & IFCAP_FLAGS);
3826 
3827 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3828 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3829 		if ((mask & IFCAP_WOL) &&
3830 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3831 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3832 		if_vlancap(ifp);
3833 		/*
3834 		 * want to ensure that traffic has stopped before we change any of the flags
3835 		 */
3836 		if (setmask) {
3837 			CTX_LOCK(ctx);
3838 			bits = if_getdrvflags(ifp);
3839 			if (bits & IFF_DRV_RUNNING)
3840 				iflib_stop(ctx);
3841 			if_togglecapenable(ifp, setmask);
3842 			if (bits & IFF_DRV_RUNNING)
3843 				iflib_init_locked(ctx);
3844 			if_setdrvflags(ifp, bits);
3845 			CTX_UNLOCK(ctx);
3846 		}
3847 		break;
3848 	    }
3849 	case SIOCGPRIVATE_0:
3850 	case SIOCSDRVSPEC:
3851 	case SIOCGDRVSPEC:
3852 		CTX_LOCK(ctx);
3853 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3854 		CTX_UNLOCK(ctx);
3855 		break;
3856 	default:
3857 		err = ether_ioctl(ifp, command, data);
3858 		break;
3859 	}
3860 	if (reinit)
3861 		iflib_if_init(ctx);
3862 	return (err);
3863 }
3864 
3865 static uint64_t
3866 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3867 {
3868 	if_ctx_t ctx = if_getsoftc(ifp);
3869 
3870 	return (IFDI_GET_COUNTER(ctx, cnt));
3871 }
3872 
3873 /*********************************************************************
3874  *
3875  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3876  *
3877  **********************************************************************/
3878 
3879 static void
3880 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3881 {
3882 	if_ctx_t ctx = if_getsoftc(ifp);
3883 
3884 	if ((void *)ctx != arg)
3885 		return;
3886 
3887 	if ((vtag == 0) || (vtag > 4095))
3888 		return;
3889 
3890 	CTX_LOCK(ctx);
3891 	IFDI_VLAN_REGISTER(ctx, vtag);
3892 	/* Re-init to load the changes */
3893 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3894 		iflib_init_locked(ctx);
3895 	CTX_UNLOCK(ctx);
3896 }
3897 
3898 static void
3899 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3900 {
3901 	if_ctx_t ctx = if_getsoftc(ifp);
3902 
3903 	if ((void *)ctx != arg)
3904 		return;
3905 
3906 	if ((vtag == 0) || (vtag > 4095))
3907 		return;
3908 
3909 	CTX_LOCK(ctx);
3910 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3911 	/* Re-init to load the changes */
3912 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3913 		iflib_init_locked(ctx);
3914 	CTX_UNLOCK(ctx);
3915 }
3916 
3917 static void
3918 iflib_led_func(void *arg, int onoff)
3919 {
3920 	if_ctx_t ctx = arg;
3921 
3922 	CTX_LOCK(ctx);
3923 	IFDI_LED_FUNC(ctx, onoff);
3924 	CTX_UNLOCK(ctx);
3925 }
3926 
3927 /*********************************************************************
3928  *
3929  *  BUS FUNCTION DEFINITIONS
3930  *
3931  **********************************************************************/
3932 
3933 int
3934 iflib_device_probe(device_t dev)
3935 {
3936 	pci_vendor_info_t *ent;
3937 
3938 	uint16_t	pci_vendor_id, pci_device_id;
3939 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3940 	uint16_t	pci_rev_id;
3941 	if_shared_ctx_t sctx;
3942 
3943 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3944 		return (ENOTSUP);
3945 
3946 	pci_vendor_id = pci_get_vendor(dev);
3947 	pci_device_id = pci_get_device(dev);
3948 	pci_subvendor_id = pci_get_subvendor(dev);
3949 	pci_subdevice_id = pci_get_subdevice(dev);
3950 	pci_rev_id = pci_get_revid(dev);
3951 	if (sctx->isc_parse_devinfo != NULL)
3952 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3953 
3954 	ent = sctx->isc_vendor_info;
3955 	while (ent->pvi_vendor_id != 0) {
3956 		if (pci_vendor_id != ent->pvi_vendor_id) {
3957 			ent++;
3958 			continue;
3959 		}
3960 		if ((pci_device_id == ent->pvi_device_id) &&
3961 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3962 		     (ent->pvi_subvendor_id == 0)) &&
3963 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3964 		     (ent->pvi_subdevice_id == 0)) &&
3965 		    ((pci_rev_id == ent->pvi_rev_id) ||
3966 		     (ent->pvi_rev_id == 0))) {
3967 
3968 			device_set_desc_copy(dev, ent->pvi_name);
3969 			/* this needs to be changed to zero if the bus probing code
3970 			 * ever stops re-probing on best match because the sctx
3971 			 * may have its values over written by register calls
3972 			 * in subsequent probes
3973 			 */
3974 			return (BUS_PROBE_DEFAULT);
3975 		}
3976 		ent++;
3977 	}
3978 	return (ENXIO);
3979 }
3980 
3981 int
3982 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3983 {
3984 	int err, rid, msix, msix_bar;
3985 	if_ctx_t ctx;
3986 	if_t ifp;
3987 	if_softc_ctx_t scctx;
3988 	int i;
3989 	uint16_t main_txq;
3990 	uint16_t main_rxq;
3991 
3992 
3993 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3994 
3995 	if (sc == NULL) {
3996 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3997 		device_set_softc(dev, ctx);
3998 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3999 	}
4000 
4001 	ctx->ifc_sctx = sctx;
4002 	ctx->ifc_dev = dev;
4003 	ctx->ifc_softc = sc;
4004 
4005 	if ((err = iflib_register(ctx)) != 0) {
4006 		device_printf(dev, "iflib_register failed %d\n", err);
4007 		return (err);
4008 	}
4009 	iflib_add_device_sysctl_pre(ctx);
4010 
4011 	scctx = &ctx->ifc_softc_ctx;
4012 	ifp = ctx->ifc_ifp;
4013 
4014 	/*
4015 	 * XXX sanity check that ntxd & nrxd are a power of 2
4016 	 */
4017 	if (ctx->ifc_sysctl_ntxqs != 0)
4018 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4019 	if (ctx->ifc_sysctl_nrxqs != 0)
4020 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4021 
4022 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4023 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4024 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4025 		else
4026 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4027 	}
4028 
4029 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4030 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4031 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4032 		else
4033 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4034 	}
4035 
4036 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4037 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4038 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4039 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4040 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4041 		}
4042 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4043 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4044 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4045 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4046 		}
4047 	}
4048 
4049 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4050 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4051 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4052 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4053 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4054 		}
4055 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4056 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4057 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4058 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4059 		}
4060 	}
4061 
4062 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4063 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4064 		return (err);
4065 	}
4066 	_iflib_pre_assert(scctx);
4067 	ctx->ifc_txrx = *scctx->isc_txrx;
4068 
4069 #ifdef INVARIANTS
4070 	MPASS(scctx->isc_capenable);
4071 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4072 		MPASS(scctx->isc_tx_csum_flags);
4073 #endif
4074 
4075 	if_setcapabilities(ifp, scctx->isc_capenable);
4076 	if_setcapenable(ifp, scctx->isc_capenable);
4077 
4078 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4079 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4080 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4081 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4082 
4083 #ifdef ACPI_DMAR
4084 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4085 		ctx->ifc_flags |= IFC_DMAR;
4086 #elif !(defined(__i386__) || defined(__amd64__))
4087 	/* set unconditionally for !x86 */
4088 	ctx->ifc_flags |= IFC_DMAR;
4089 #endif
4090 
4091 	msix_bar = scctx->isc_msix_bar;
4092 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4093 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4094 
4095 	/* XXX change for per-queue sizes */
4096 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4097 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4098 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4099 		if (!powerof2(scctx->isc_nrxd[i])) {
4100 			/* round down instead? */
4101 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4102 			err = EINVAL;
4103 			goto fail;
4104 		}
4105 	}
4106 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4107 		if (!powerof2(scctx->isc_ntxd[i])) {
4108 			device_printf(dev,
4109 			    "# tx descriptors must be a power of 2");
4110 			err = EINVAL;
4111 			goto fail;
4112 		}
4113 	}
4114 
4115 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4116 	    MAX_SINGLE_PACKET_FRACTION)
4117 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4118 		    MAX_SINGLE_PACKET_FRACTION);
4119 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4120 	    MAX_SINGLE_PACKET_FRACTION)
4121 		scctx->isc_tx_tso_segments_max = max(1,
4122 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4123 
4124 	/*
4125 	 * Protect the stack against modern hardware
4126 	 */
4127 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4128 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4129 
4130 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4131 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4132 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4133 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4134 	if (scctx->isc_rss_table_size == 0)
4135 		scctx->isc_rss_table_size = 64;
4136 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4137 
4138 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4139 	/* XXX format name */
4140 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4141 	/*
4142 	** Now setup MSI or MSI/X, should
4143 	** return us the number of supported
4144 	** vectors. (Will be 1 for MSI)
4145 	*/
4146 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4147 		msix = scctx->isc_vectors;
4148 	} else if (scctx->isc_msix_bar != 0)
4149 	       /*
4150 		* The simple fact that isc_msix_bar is not 0 does not mean we
4151 		* we have a good value there that is known to work.
4152 		*/
4153 		msix = iflib_msix_init(ctx);
4154 	else {
4155 		scctx->isc_vectors = 1;
4156 		scctx->isc_ntxqsets = 1;
4157 		scctx->isc_nrxqsets = 1;
4158 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4159 		msix = 0;
4160 	}
4161 	/* Get memory for the station queues */
4162 	if ((err = iflib_queues_alloc(ctx))) {
4163 		device_printf(dev, "Unable to allocate queue memory\n");
4164 		goto fail;
4165 	}
4166 
4167 	if ((err = iflib_qset_structures_setup(ctx))) {
4168 		device_printf(dev, "qset structure setup failed %d\n", err);
4169 		goto fail_queues;
4170 	}
4171 
4172 	/*
4173 	 * Group taskqueues aren't properly set up until SMP is started,
4174 	 * so we disable interrupts until we can handle them post
4175 	 * SI_SUB_SMP.
4176 	 *
4177 	 * XXX: disabling interrupts doesn't actually work, at least for
4178 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4179 	 * we do null handling and depend on this not causing too large an
4180 	 * interrupt storm.
4181 	 */
4182 	IFDI_INTR_DISABLE(ctx);
4183 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4184 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4185 		goto fail_intr_free;
4186 	}
4187 	if (msix <= 1) {
4188 		rid = 0;
4189 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4190 			MPASS(msix == 1);
4191 			rid = 1;
4192 		}
4193 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4194 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4195 			goto fail_intr_free;
4196 		}
4197 	}
4198 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4199 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4200 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4201 		goto fail_detach;
4202 	}
4203 	if ((err = iflib_netmap_attach(ctx))) {
4204 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4205 		goto fail_detach;
4206 	}
4207 	*ctxp = ctx;
4208 
4209 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4210 	iflib_add_device_sysctl_post(ctx);
4211 	ctx->ifc_flags |= IFC_INIT_DONE;
4212 	return (0);
4213 fail_detach:
4214 	ether_ifdetach(ctx->ifc_ifp);
4215 fail_intr_free:
4216 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4217 		pci_release_msi(ctx->ifc_dev);
4218 fail_queues:
4219 	/* XXX free queues */
4220 fail:
4221 	IFDI_DETACH(ctx);
4222 	return (err);
4223 }
4224 
4225 int
4226 iflib_device_attach(device_t dev)
4227 {
4228 	if_ctx_t ctx;
4229 	if_shared_ctx_t sctx;
4230 
4231 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4232 		return (ENOTSUP);
4233 
4234 	pci_enable_busmaster(dev);
4235 
4236 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4237 }
4238 
4239 int
4240 iflib_device_deregister(if_ctx_t ctx)
4241 {
4242 	if_t ifp = ctx->ifc_ifp;
4243 	iflib_txq_t txq;
4244 	iflib_rxq_t rxq;
4245 	device_t dev = ctx->ifc_dev;
4246 	int i;
4247 	struct taskqgroup *tqg;
4248 
4249 	/* Make sure VLANS are not using driver */
4250 	if (if_vlantrunkinuse(ifp)) {
4251 		device_printf(dev,"Vlan in use, detach first\n");
4252 		return (EBUSY);
4253 	}
4254 
4255 	CTX_LOCK(ctx);
4256 	ctx->ifc_in_detach = 1;
4257 	iflib_stop(ctx);
4258 	CTX_UNLOCK(ctx);
4259 
4260 	/* Unregister VLAN events */
4261 	if (ctx->ifc_vlan_attach_event != NULL)
4262 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4263 	if (ctx->ifc_vlan_detach_event != NULL)
4264 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4265 
4266 	iflib_netmap_detach(ifp);
4267 	ether_ifdetach(ifp);
4268 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4269 	CTX_LOCK_DESTROY(ctx);
4270 	if (ctx->ifc_led_dev != NULL)
4271 		led_destroy(ctx->ifc_led_dev);
4272 	/* XXX drain any dependent tasks */
4273 	tqg = qgroup_if_io_tqg;
4274 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4275 		callout_drain(&txq->ift_timer);
4276 		if (txq->ift_task.gt_uniq != NULL)
4277 			taskqgroup_detach(tqg, &txq->ift_task);
4278 	}
4279 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4280 		if (rxq->ifr_task.gt_uniq != NULL)
4281 			taskqgroup_detach(tqg, &rxq->ifr_task);
4282 	}
4283 	tqg = qgroup_if_config_tqg;
4284 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4285 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4286 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4287 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4288 
4289 	IFDI_DETACH(ctx);
4290 	device_set_softc(ctx->ifc_dev, NULL);
4291 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4292 		pci_release_msi(dev);
4293 	}
4294 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4295 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4296 	}
4297 	if (ctx->ifc_msix_mem != NULL) {
4298 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4299 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4300 		ctx->ifc_msix_mem = NULL;
4301 	}
4302 
4303 	bus_generic_detach(dev);
4304 	if_free(ifp);
4305 
4306 	iflib_tx_structures_free(ctx);
4307 	iflib_rx_structures_free(ctx);
4308 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4309 		free(ctx->ifc_softc, M_IFLIB);
4310 	free(ctx, M_IFLIB);
4311 	return (0);
4312 }
4313 
4314 
4315 int
4316 iflib_device_detach(device_t dev)
4317 {
4318 	if_ctx_t ctx = device_get_softc(dev);
4319 
4320 	return (iflib_device_deregister(ctx));
4321 }
4322 
4323 int
4324 iflib_device_suspend(device_t dev)
4325 {
4326 	if_ctx_t ctx = device_get_softc(dev);
4327 
4328 	CTX_LOCK(ctx);
4329 	IFDI_SUSPEND(ctx);
4330 	CTX_UNLOCK(ctx);
4331 
4332 	return bus_generic_suspend(dev);
4333 }
4334 int
4335 iflib_device_shutdown(device_t dev)
4336 {
4337 	if_ctx_t ctx = device_get_softc(dev);
4338 
4339 	CTX_LOCK(ctx);
4340 	IFDI_SHUTDOWN(ctx);
4341 	CTX_UNLOCK(ctx);
4342 
4343 	return bus_generic_suspend(dev);
4344 }
4345 
4346 
4347 int
4348 iflib_device_resume(device_t dev)
4349 {
4350 	if_ctx_t ctx = device_get_softc(dev);
4351 	iflib_txq_t txq = ctx->ifc_txqs;
4352 
4353 	CTX_LOCK(ctx);
4354 	IFDI_RESUME(ctx);
4355 	iflib_init_locked(ctx);
4356 	CTX_UNLOCK(ctx);
4357 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4358 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4359 
4360 	return (bus_generic_resume(dev));
4361 }
4362 
4363 int
4364 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4365 {
4366 	int error;
4367 	if_ctx_t ctx = device_get_softc(dev);
4368 
4369 	CTX_LOCK(ctx);
4370 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4371 	CTX_UNLOCK(ctx);
4372 
4373 	return (error);
4374 }
4375 
4376 void
4377 iflib_device_iov_uninit(device_t dev)
4378 {
4379 	if_ctx_t ctx = device_get_softc(dev);
4380 
4381 	CTX_LOCK(ctx);
4382 	IFDI_IOV_UNINIT(ctx);
4383 	CTX_UNLOCK(ctx);
4384 }
4385 
4386 int
4387 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4388 {
4389 	int error;
4390 	if_ctx_t ctx = device_get_softc(dev);
4391 
4392 	CTX_LOCK(ctx);
4393 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4394 	CTX_UNLOCK(ctx);
4395 
4396 	return (error);
4397 }
4398 
4399 /*********************************************************************
4400  *
4401  *  MODULE FUNCTION DEFINITIONS
4402  *
4403  **********************************************************************/
4404 
4405 /*
4406  * - Start a fast taskqueue thread for each core
4407  * - Start a taskqueue for control operations
4408  */
4409 static int
4410 iflib_module_init(void)
4411 {
4412 	return (0);
4413 }
4414 
4415 static int
4416 iflib_module_event_handler(module_t mod, int what, void *arg)
4417 {
4418 	int err;
4419 
4420 	switch (what) {
4421 	case MOD_LOAD:
4422 		if ((err = iflib_module_init()) != 0)
4423 			return (err);
4424 		break;
4425 	case MOD_UNLOAD:
4426 		return (EBUSY);
4427 	default:
4428 		return (EOPNOTSUPP);
4429 	}
4430 
4431 	return (0);
4432 }
4433 
4434 /*********************************************************************
4435  *
4436  *  PUBLIC FUNCTION DEFINITIONS
4437  *     ordered as in iflib.h
4438  *
4439  **********************************************************************/
4440 
4441 
4442 static void
4443 _iflib_assert(if_shared_ctx_t sctx)
4444 {
4445 	MPASS(sctx->isc_tx_maxsize);
4446 	MPASS(sctx->isc_tx_maxsegsize);
4447 
4448 	MPASS(sctx->isc_rx_maxsize);
4449 	MPASS(sctx->isc_rx_nsegments);
4450 	MPASS(sctx->isc_rx_maxsegsize);
4451 
4452 	MPASS(sctx->isc_nrxd_min[0]);
4453 	MPASS(sctx->isc_nrxd_max[0]);
4454 	MPASS(sctx->isc_nrxd_default[0]);
4455 	MPASS(sctx->isc_ntxd_min[0]);
4456 	MPASS(sctx->isc_ntxd_max[0]);
4457 	MPASS(sctx->isc_ntxd_default[0]);
4458 }
4459 
4460 static void
4461 _iflib_pre_assert(if_softc_ctx_t scctx)
4462 {
4463 
4464 	MPASS(scctx->isc_txrx->ift_txd_encap);
4465 	MPASS(scctx->isc_txrx->ift_txd_flush);
4466 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4467 	MPASS(scctx->isc_txrx->ift_rxd_available);
4468 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4469 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4470 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4471 }
4472 
4473 static int
4474 iflib_register(if_ctx_t ctx)
4475 {
4476 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4477 	driver_t *driver = sctx->isc_driver;
4478 	device_t dev = ctx->ifc_dev;
4479 	if_t ifp;
4480 
4481 	_iflib_assert(sctx);
4482 
4483 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4484 
4485 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4486 	if (ifp == NULL) {
4487 		device_printf(dev, "can not allocate ifnet structure\n");
4488 		return (ENOMEM);
4489 	}
4490 
4491 	/*
4492 	 * Initialize our context's device specific methods
4493 	 */
4494 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4495 	kobj_class_compile((kobj_class_t) driver);
4496 	driver->refs++;
4497 
4498 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4499 	if_setsoftc(ifp, ctx);
4500 	if_setdev(ifp, dev);
4501 	if_setinitfn(ifp, iflib_if_init);
4502 	if_setioctlfn(ifp, iflib_if_ioctl);
4503 	if_settransmitfn(ifp, iflib_if_transmit);
4504 	if_setqflushfn(ifp, iflib_if_qflush);
4505 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4506 
4507 	ctx->ifc_vlan_attach_event =
4508 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4509 							  EVENTHANDLER_PRI_FIRST);
4510 	ctx->ifc_vlan_detach_event =
4511 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4512 							  EVENTHANDLER_PRI_FIRST);
4513 
4514 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4515 					 iflib_media_change, iflib_media_status);
4516 
4517 	return (0);
4518 }
4519 
4520 
4521 static int
4522 iflib_queues_alloc(if_ctx_t ctx)
4523 {
4524 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4525 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4526 	device_t dev = ctx->ifc_dev;
4527 	int nrxqsets = scctx->isc_nrxqsets;
4528 	int ntxqsets = scctx->isc_ntxqsets;
4529 	iflib_txq_t txq;
4530 	iflib_rxq_t rxq;
4531 	iflib_fl_t fl = NULL;
4532 	int i, j, cpu, err, txconf, rxconf;
4533 	iflib_dma_info_t ifdip;
4534 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4535 	uint32_t *txqsizes = scctx->isc_txqsizes;
4536 	uint8_t nrxqs = sctx->isc_nrxqs;
4537 	uint8_t ntxqs = sctx->isc_ntxqs;
4538 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4539 	caddr_t *vaddrs;
4540 	uint64_t *paddrs;
4541 	struct ifmp_ring **brscp;
4542 
4543 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4544 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4545 
4546 	brscp = NULL;
4547 	txq = NULL;
4548 	rxq = NULL;
4549 
4550 /* Allocate the TX ring struct memory */
4551 	if (!(txq =
4552 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4553 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4554 		device_printf(dev, "Unable to allocate TX ring memory\n");
4555 		err = ENOMEM;
4556 		goto fail;
4557 	}
4558 
4559 	/* Now allocate the RX */
4560 	if (!(rxq =
4561 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4562 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4563 		device_printf(dev, "Unable to allocate RX ring memory\n");
4564 		err = ENOMEM;
4565 		goto rx_fail;
4566 	}
4567 
4568 	ctx->ifc_txqs = txq;
4569 	ctx->ifc_rxqs = rxq;
4570 
4571 	/*
4572 	 * XXX handle allocation failure
4573 	 */
4574 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4575 		/* Set up some basics */
4576 
4577 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4578 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4579 			err = ENOMEM;
4580 			goto err_tx_desc;
4581 		}
4582 		txq->ift_ifdi = ifdip;
4583 		for (j = 0; j < ntxqs; j++, ifdip++) {
4584 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4585 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4586 				err = ENOMEM;
4587 				goto err_tx_desc;
4588 			}
4589 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4590 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4591 		}
4592 		txq->ift_ctx = ctx;
4593 		txq->ift_id = i;
4594 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4595 			txq->ift_br_offset = 1;
4596 		} else {
4597 			txq->ift_br_offset = 0;
4598 		}
4599 		/* XXX fix this */
4600 		txq->ift_timer.c_cpu = cpu;
4601 
4602 		if (iflib_txsd_alloc(txq)) {
4603 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4604 			err = ENOMEM;
4605 			goto err_tx_desc;
4606 		}
4607 
4608 		/* Initialize the TX lock */
4609 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4610 		    device_get_nameunit(dev), txq->ift_id);
4611 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4612 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4613 
4614 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4615 			 device_get_nameunit(dev), txq->ift_id);
4616 
4617 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4618 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4619 		if (err) {
4620 			/* XXX free any allocated rings */
4621 			device_printf(dev, "Unable to allocate buf_ring\n");
4622 			goto err_tx_desc;
4623 		}
4624 	}
4625 
4626 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4627 		/* Set up some basics */
4628 
4629 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4630 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4631 			err = ENOMEM;
4632 			goto err_tx_desc;
4633 		}
4634 
4635 		rxq->ifr_ifdi = ifdip;
4636 		/* XXX this needs to be changed if #rx queues != #tx queues */
4637 		rxq->ifr_ntxqirq = 1;
4638 		rxq->ifr_txqid[0] = i;
4639 		for (j = 0; j < nrxqs; j++, ifdip++) {
4640 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4641 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4642 				err = ENOMEM;
4643 				goto err_tx_desc;
4644 			}
4645 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4646 		}
4647 		rxq->ifr_ctx = ctx;
4648 		rxq->ifr_id = i;
4649 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4650 			rxq->ifr_fl_offset = 1;
4651 		} else {
4652 			rxq->ifr_fl_offset = 0;
4653 		}
4654 		rxq->ifr_nfl = nfree_lists;
4655 		if (!(fl =
4656 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4657 			device_printf(dev, "Unable to allocate free list memory\n");
4658 			err = ENOMEM;
4659 			goto err_tx_desc;
4660 		}
4661 		rxq->ifr_fl = fl;
4662 		for (j = 0; j < nfree_lists; j++) {
4663 			fl[j].ifl_rxq = rxq;
4664 			fl[j].ifl_id = j;
4665 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4666 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4667 		}
4668         /* Allocate receive buffers for the ring*/
4669 		if (iflib_rxsd_alloc(rxq)) {
4670 			device_printf(dev,
4671 			    "Critical Failure setting up receive buffers\n");
4672 			err = ENOMEM;
4673 			goto err_rx_desc;
4674 		}
4675 	}
4676 
4677 	/* TXQs */
4678 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4679 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4680 	for (i = 0; i < ntxqsets; i++) {
4681 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4682 
4683 		for (j = 0; j < ntxqs; j++, di++) {
4684 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4685 			paddrs[i*ntxqs + j] = di->idi_paddr;
4686 		}
4687 	}
4688 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4689 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4690 		iflib_tx_structures_free(ctx);
4691 		free(vaddrs, M_IFLIB);
4692 		free(paddrs, M_IFLIB);
4693 		goto err_rx_desc;
4694 	}
4695 	free(vaddrs, M_IFLIB);
4696 	free(paddrs, M_IFLIB);
4697 
4698 	/* RXQs */
4699 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4700 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4701 	for (i = 0; i < nrxqsets; i++) {
4702 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4703 
4704 		for (j = 0; j < nrxqs; j++, di++) {
4705 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4706 			paddrs[i*nrxqs + j] = di->idi_paddr;
4707 		}
4708 	}
4709 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4710 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4711 		iflib_tx_structures_free(ctx);
4712 		free(vaddrs, M_IFLIB);
4713 		free(paddrs, M_IFLIB);
4714 		goto err_rx_desc;
4715 	}
4716 	free(vaddrs, M_IFLIB);
4717 	free(paddrs, M_IFLIB);
4718 
4719 	return (0);
4720 
4721 /* XXX handle allocation failure changes */
4722 err_rx_desc:
4723 err_tx_desc:
4724 	if (ctx->ifc_rxqs != NULL)
4725 		free(ctx->ifc_rxqs, M_IFLIB);
4726 	ctx->ifc_rxqs = NULL;
4727 	if (ctx->ifc_txqs != NULL)
4728 		free(ctx->ifc_txqs, M_IFLIB);
4729 	ctx->ifc_txqs = NULL;
4730 rx_fail:
4731 	if (brscp != NULL)
4732 		free(brscp, M_IFLIB);
4733 	if (rxq != NULL)
4734 		free(rxq, M_IFLIB);
4735 	if (txq != NULL)
4736 		free(txq, M_IFLIB);
4737 fail:
4738 	return (err);
4739 }
4740 
4741 static int
4742 iflib_tx_structures_setup(if_ctx_t ctx)
4743 {
4744 	iflib_txq_t txq = ctx->ifc_txqs;
4745 	int i;
4746 
4747 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4748 		iflib_txq_setup(txq);
4749 
4750 	return (0);
4751 }
4752 
4753 static void
4754 iflib_tx_structures_free(if_ctx_t ctx)
4755 {
4756 	iflib_txq_t txq = ctx->ifc_txqs;
4757 	int i, j;
4758 
4759 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4760 		iflib_txq_destroy(txq);
4761 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4762 			iflib_dma_free(&txq->ift_ifdi[j]);
4763 	}
4764 	free(ctx->ifc_txqs, M_IFLIB);
4765 	ctx->ifc_txqs = NULL;
4766 	IFDI_QUEUES_FREE(ctx);
4767 }
4768 
4769 /*********************************************************************
4770  *
4771  *  Initialize all receive rings.
4772  *
4773  **********************************************************************/
4774 static int
4775 iflib_rx_structures_setup(if_ctx_t ctx)
4776 {
4777 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4778 	int q;
4779 #if defined(INET6) || defined(INET)
4780 	int i, err;
4781 #endif
4782 
4783 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4784 #if defined(INET6) || defined(INET)
4785 		tcp_lro_free(&rxq->ifr_lc);
4786 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4787 		    TCP_LRO_ENTRIES, min(1024,
4788 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4789 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4790 			goto fail;
4791 		}
4792 		rxq->ifr_lro_enabled = TRUE;
4793 #endif
4794 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4795 	}
4796 	return (0);
4797 #if defined(INET6) || defined(INET)
4798 fail:
4799 	/*
4800 	 * Free RX software descriptors allocated so far, we will only handle
4801 	 * the rings that completed, the failing case will have
4802 	 * cleaned up for itself. 'q' failed, so its the terminus.
4803 	 */
4804 	rxq = ctx->ifc_rxqs;
4805 	for (i = 0; i < q; ++i, rxq++) {
4806 		iflib_rx_sds_free(rxq);
4807 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4808 	}
4809 	return (err);
4810 #endif
4811 }
4812 
4813 /*********************************************************************
4814  *
4815  *  Free all receive rings.
4816  *
4817  **********************************************************************/
4818 static void
4819 iflib_rx_structures_free(if_ctx_t ctx)
4820 {
4821 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4822 
4823 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4824 		iflib_rx_sds_free(rxq);
4825 	}
4826 }
4827 
4828 static int
4829 iflib_qset_structures_setup(if_ctx_t ctx)
4830 {
4831 	int err;
4832 
4833 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4834 		return (err);
4835 
4836 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4837 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4838 		iflib_tx_structures_free(ctx);
4839 		iflib_rx_structures_free(ctx);
4840 	}
4841 	return (err);
4842 }
4843 
4844 int
4845 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4846 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4847 {
4848 
4849 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4850 }
4851 
4852 static int
4853 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4854 {
4855 	int i, cpuid, eqid, count;
4856 
4857 	CPU_COPY(&ctx->ifc_cpus, cpus);
4858 	count = CPU_COUNT(&ctx->ifc_cpus);
4859 	eqid = qid % count;
4860 	/* clear up to the qid'th bit */
4861 	for (i = 0; i < eqid; i++) {
4862 		cpuid = CPU_FFS(cpus);
4863 		MPASS(cpuid != 0);
4864 		CPU_CLR(cpuid-1, cpus);
4865 	}
4866 	cpuid = CPU_FFS(cpus);
4867 	MPASS(cpuid != 0);
4868 	return (cpuid-1);
4869 }
4870 
4871 int
4872 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4873 						iflib_intr_type_t type, driver_filter_t *filter,
4874 						void *filter_arg, int qid, char *name)
4875 {
4876 	struct grouptask *gtask;
4877 	struct taskqgroup *tqg;
4878 	iflib_filter_info_t info;
4879 	cpuset_t cpus;
4880 	gtask_fn_t *fn;
4881 	int tqrid, err, cpuid;
4882 	driver_filter_t *intr_fast;
4883 	void *q;
4884 
4885 	info = &ctx->ifc_filter_info;
4886 	tqrid = rid;
4887 
4888 	switch (type) {
4889 	/* XXX merge tx/rx for netmap? */
4890 	case IFLIB_INTR_TX:
4891 		q = &ctx->ifc_txqs[qid];
4892 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4893 		gtask = &ctx->ifc_txqs[qid].ift_task;
4894 		tqg = qgroup_if_io_tqg;
4895 		fn = _task_fn_tx;
4896 		intr_fast = iflib_fast_intr;
4897 		GROUPTASK_INIT(gtask, 0, fn, q);
4898 		break;
4899 	case IFLIB_INTR_RX:
4900 		q = &ctx->ifc_rxqs[qid];
4901 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4902 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4903 		tqg = qgroup_if_io_tqg;
4904 		fn = _task_fn_rx;
4905 		intr_fast = iflib_fast_intr;
4906 		GROUPTASK_INIT(gtask, 0, fn, q);
4907 		break;
4908 	case IFLIB_INTR_RXTX:
4909 		q = &ctx->ifc_rxqs[qid];
4910 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4911 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4912 		tqg = qgroup_if_io_tqg;
4913 		fn = _task_fn_rx;
4914 		intr_fast = iflib_fast_intr_rxtx;
4915 		GROUPTASK_INIT(gtask, 0, fn, q);
4916 		break;
4917 	case IFLIB_INTR_ADMIN:
4918 		q = ctx;
4919 		tqrid = -1;
4920 		info = &ctx->ifc_filter_info;
4921 		gtask = &ctx->ifc_admin_task;
4922 		tqg = qgroup_if_config_tqg;
4923 		fn = _task_fn_admin;
4924 		intr_fast = iflib_fast_intr_ctx;
4925 		break;
4926 	default:
4927 		panic("unknown net intr type");
4928 	}
4929 
4930 	info->ifi_filter = filter;
4931 	info->ifi_filter_arg = filter_arg;
4932 	info->ifi_task = gtask;
4933 	info->ifi_ctx = q;
4934 
4935 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4936 	if (err != 0) {
4937 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4938 		return (err);
4939 	}
4940 	if (type == IFLIB_INTR_ADMIN)
4941 		return (0);
4942 
4943 	if (tqrid != -1) {
4944 		cpuid = find_nth(ctx, &cpus, qid);
4945 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4946 	} else {
4947 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4948 	}
4949 
4950 	return (0);
4951 }
4952 
4953 void
4954 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4955 {
4956 	struct grouptask *gtask;
4957 	struct taskqgroup *tqg;
4958 	gtask_fn_t *fn;
4959 	void *q;
4960 
4961 	switch (type) {
4962 	case IFLIB_INTR_TX:
4963 		q = &ctx->ifc_txqs[qid];
4964 		gtask = &ctx->ifc_txqs[qid].ift_task;
4965 		tqg = qgroup_if_io_tqg;
4966 		fn = _task_fn_tx;
4967 		break;
4968 	case IFLIB_INTR_RX:
4969 		q = &ctx->ifc_rxqs[qid];
4970 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4971 		tqg = qgroup_if_io_tqg;
4972 		fn = _task_fn_rx;
4973 		break;
4974 	case IFLIB_INTR_IOV:
4975 		q = ctx;
4976 		gtask = &ctx->ifc_vflr_task;
4977 		tqg = qgroup_if_config_tqg;
4978 		rid = -1;
4979 		fn = _task_fn_iov;
4980 		break;
4981 	default:
4982 		panic("unknown net intr type");
4983 	}
4984 	GROUPTASK_INIT(gtask, 0, fn, q);
4985 	taskqgroup_attach(tqg, gtask, q, rid, name);
4986 }
4987 
4988 void
4989 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4990 {
4991 	if (irq->ii_tag)
4992 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4993 
4994 	if (irq->ii_res)
4995 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4996 }
4997 
4998 static int
4999 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5000 {
5001 	iflib_txq_t txq = ctx->ifc_txqs;
5002 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5003 	if_irq_t irq = &ctx->ifc_legacy_irq;
5004 	iflib_filter_info_t info;
5005 	struct grouptask *gtask;
5006 	struct taskqgroup *tqg;
5007 	gtask_fn_t *fn;
5008 	int tqrid;
5009 	void *q;
5010 	int err;
5011 
5012 	q = &ctx->ifc_rxqs[0];
5013 	info = &rxq[0].ifr_filter_info;
5014 	gtask = &rxq[0].ifr_task;
5015 	tqg = qgroup_if_io_tqg;
5016 	tqrid = irq->ii_rid = *rid;
5017 	fn = _task_fn_rx;
5018 
5019 	ctx->ifc_flags |= IFC_LEGACY;
5020 	info->ifi_filter = filter;
5021 	info->ifi_filter_arg = filter_arg;
5022 	info->ifi_task = gtask;
5023 	info->ifi_ctx = ctx;
5024 
5025 	/* We allocate a single interrupt resource */
5026 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5027 		return (err);
5028 	GROUPTASK_INIT(gtask, 0, fn, q);
5029 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5030 
5031 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5032 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5033 	return (0);
5034 }
5035 
5036 void
5037 iflib_led_create(if_ctx_t ctx)
5038 {
5039 
5040 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5041 	    device_get_nameunit(ctx->ifc_dev));
5042 }
5043 
5044 void
5045 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5046 {
5047 
5048 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5049 }
5050 
5051 void
5052 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5053 {
5054 
5055 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5056 }
5057 
5058 void
5059 iflib_admin_intr_deferred(if_ctx_t ctx)
5060 {
5061 #ifdef INVARIANTS
5062 	struct grouptask *gtask;
5063 
5064 	gtask = &ctx->ifc_admin_task;
5065 	MPASS(gtask->gt_taskqueue != NULL);
5066 #endif
5067 
5068 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5069 }
5070 
5071 void
5072 iflib_iov_intr_deferred(if_ctx_t ctx)
5073 {
5074 
5075 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5076 }
5077 
5078 void
5079 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5080 {
5081 
5082 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5083 }
5084 
5085 void
5086 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5087 	char *name)
5088 {
5089 
5090 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5091 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5092 }
5093 
5094 void
5095 iflib_config_gtask_deinit(struct grouptask *gtask)
5096 {
5097 
5098 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5099 }
5100 
5101 void
5102 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5103 {
5104 	if_t ifp = ctx->ifc_ifp;
5105 	iflib_txq_t txq = ctx->ifc_txqs;
5106 
5107 	if_setbaudrate(ifp, baudrate);
5108 	if (baudrate >= IF_Gbps(10))
5109 		ctx->ifc_flags |= IFC_PREFETCH;
5110 
5111 	/* If link down, disable watchdog */
5112 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5113 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5114 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5115 	}
5116 	ctx->ifc_link_state = link_state;
5117 	if_link_state_change(ifp, link_state);
5118 }
5119 
5120 static int
5121 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5122 {
5123 	int credits;
5124 #ifdef INVARIANTS
5125 	int credits_pre = txq->ift_cidx_processed;
5126 #endif
5127 
5128 	if (ctx->isc_txd_credits_update == NULL)
5129 		return (0);
5130 
5131 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5132 		return (0);
5133 
5134 	txq->ift_processed += credits;
5135 	txq->ift_cidx_processed += credits;
5136 
5137 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5138 	if (txq->ift_cidx_processed >= txq->ift_size)
5139 		txq->ift_cidx_processed -= txq->ift_size;
5140 	return (credits);
5141 }
5142 
5143 static int
5144 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5145 {
5146 
5147 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5148 	    budget));
5149 }
5150 
5151 void
5152 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5153 	const char *description, if_int_delay_info_t info,
5154 	int offset, int value)
5155 {
5156 	info->iidi_ctx = ctx;
5157 	info->iidi_offset = offset;
5158 	info->iidi_value = value;
5159 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5160 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5161 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5162 	    info, 0, iflib_sysctl_int_delay, "I", description);
5163 }
5164 
5165 struct mtx *
5166 iflib_ctx_lock_get(if_ctx_t ctx)
5167 {
5168 
5169 	return (&ctx->ifc_mtx);
5170 }
5171 
5172 static int
5173 iflib_msix_init(if_ctx_t ctx)
5174 {
5175 	device_t dev = ctx->ifc_dev;
5176 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5177 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5178 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5179 	int iflib_num_tx_queues, iflib_num_rx_queues;
5180 	int err, admincnt, bar;
5181 
5182 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5183 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5184 
5185 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5186 
5187 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5188 	admincnt = sctx->isc_admin_intrcnt;
5189 	/* Override by tuneable */
5190 	if (scctx->isc_disable_msix)
5191 		goto msi;
5192 
5193 	/*
5194 	** When used in a virtualized environment
5195 	** PCI BUSMASTER capability may not be set
5196 	** so explicity set it here and rewrite
5197 	** the ENABLE in the MSIX control register
5198 	** at this point to cause the host to
5199 	** successfully initialize us.
5200 	*/
5201 	{
5202 		int msix_ctrl, rid;
5203 
5204  		pci_enable_busmaster(dev);
5205 		rid = 0;
5206 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5207 			rid += PCIR_MSIX_CTRL;
5208 			msix_ctrl = pci_read_config(dev, rid, 2);
5209 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5210 			pci_write_config(dev, rid, msix_ctrl, 2);
5211 		} else {
5212 			device_printf(dev, "PCIY_MSIX capability not found; "
5213 			                   "or rid %d == 0.\n", rid);
5214 			goto msi;
5215 		}
5216 	}
5217 
5218 	/*
5219 	 * bar == -1 => "trust me I know what I'm doing"
5220 	 * Some drivers are for hardware that is so shoddily
5221 	 * documented that no one knows which bars are which
5222 	 * so the developer has to map all bars. This hack
5223 	 * allows shoddy garbage to use msix in this framework.
5224 	 */
5225 	if (bar != -1) {
5226 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5227 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5228 		if (ctx->ifc_msix_mem == NULL) {
5229 			/* May not be enabled */
5230 			device_printf(dev, "Unable to map MSIX table \n");
5231 			goto msi;
5232 		}
5233 	}
5234 	/* First try MSI/X */
5235 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5236 		device_printf(dev, "System has MSIX disabled \n");
5237 		bus_release_resource(dev, SYS_RES_MEMORY,
5238 		    bar, ctx->ifc_msix_mem);
5239 		ctx->ifc_msix_mem = NULL;
5240 		goto msi;
5241 	}
5242 #if IFLIB_DEBUG
5243 	/* use only 1 qset in debug mode */
5244 	queuemsgs = min(msgs - admincnt, 1);
5245 #else
5246 	queuemsgs = msgs - admincnt;
5247 #endif
5248 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5249 #ifdef RSS
5250 		queues = imin(queuemsgs, rss_getnumbuckets());
5251 #else
5252 		queues = queuemsgs;
5253 #endif
5254 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5255 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5256 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5257 	} else {
5258 		device_printf(dev, "Unable to fetch CPU list\n");
5259 		/* Figure out a reasonable auto config value */
5260 		queues = min(queuemsgs, mp_ncpus);
5261 	}
5262 #ifdef  RSS
5263 	/* If we're doing RSS, clamp at the number of RSS buckets */
5264 	if (queues > rss_getnumbuckets())
5265 		queues = rss_getnumbuckets();
5266 #endif
5267 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5268 		rx_queues = iflib_num_rx_queues;
5269 	else
5270 		rx_queues = queues;
5271 	/*
5272 	 * We want this to be all logical CPUs by default
5273 	 */
5274 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5275 		tx_queues = iflib_num_tx_queues;
5276 	else
5277 		tx_queues = mp_ncpus;
5278 
5279 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5280 #ifdef INVARIANTS
5281 		if (tx_queues != rx_queues)
5282 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5283 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5284 #endif
5285 		tx_queues = min(rx_queues, tx_queues);
5286 		rx_queues = min(rx_queues, tx_queues);
5287 	}
5288 
5289 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5290 
5291 	vectors = rx_queues + admincnt;
5292 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5293 		device_printf(dev,
5294 					  "Using MSIX interrupts with %d vectors\n", vectors);
5295 		scctx->isc_vectors = vectors;
5296 		scctx->isc_nrxqsets = rx_queues;
5297 		scctx->isc_ntxqsets = tx_queues;
5298 		scctx->isc_intr = IFLIB_INTR_MSIX;
5299 
5300 		return (vectors);
5301 	} else {
5302 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5303 	}
5304 msi:
5305 	vectors = pci_msi_count(dev);
5306 	scctx->isc_nrxqsets = 1;
5307 	scctx->isc_ntxqsets = 1;
5308 	scctx->isc_vectors = vectors;
5309 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5310 		device_printf(dev,"Using an MSI interrupt\n");
5311 		scctx->isc_intr = IFLIB_INTR_MSI;
5312 	} else {
5313 		device_printf(dev,"Using a Legacy interrupt\n");
5314 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5315 	}
5316 
5317 	return (vectors);
5318 }
5319 
5320 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5321 
5322 static int
5323 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5324 {
5325 	int rc;
5326 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5327 	struct sbuf *sb;
5328 	char *ring_state = "UNKNOWN";
5329 
5330 	/* XXX needed ? */
5331 	rc = sysctl_wire_old_buffer(req, 0);
5332 	MPASS(rc == 0);
5333 	if (rc != 0)
5334 		return (rc);
5335 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5336 	MPASS(sb != NULL);
5337 	if (sb == NULL)
5338 		return (ENOMEM);
5339 	if (state[3] <= 3)
5340 		ring_state = ring_states[state[3]];
5341 
5342 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5343 		    state[0], state[1], state[2], ring_state);
5344 	rc = sbuf_finish(sb);
5345 	sbuf_delete(sb);
5346         return(rc);
5347 }
5348 
5349 enum iflib_ndesc_handler {
5350 	IFLIB_NTXD_HANDLER,
5351 	IFLIB_NRXD_HANDLER,
5352 };
5353 
5354 static int
5355 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5356 {
5357 	if_ctx_t ctx = (void *)arg1;
5358 	enum iflib_ndesc_handler type = arg2;
5359 	char buf[256] = {0};
5360 	qidx_t *ndesc;
5361 	char *p, *next;
5362 	int nqs, rc, i;
5363 
5364 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5365 
5366 	nqs = 8;
5367 	switch(type) {
5368 	case IFLIB_NTXD_HANDLER:
5369 		ndesc = ctx->ifc_sysctl_ntxds;
5370 		if (ctx->ifc_sctx)
5371 			nqs = ctx->ifc_sctx->isc_ntxqs;
5372 		break;
5373 	case IFLIB_NRXD_HANDLER:
5374 		ndesc = ctx->ifc_sysctl_nrxds;
5375 		if (ctx->ifc_sctx)
5376 			nqs = ctx->ifc_sctx->isc_nrxqs;
5377 		break;
5378 	}
5379 	if (nqs == 0)
5380 		nqs = 8;
5381 
5382 	for (i=0; i<8; i++) {
5383 		if (i >= nqs)
5384 			break;
5385 		if (i)
5386 			strcat(buf, ",");
5387 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5388 	}
5389 
5390 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5391 	if (rc || req->newptr == NULL)
5392 		return rc;
5393 
5394 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5395 	    i++, p = strsep(&next, " ,")) {
5396 		ndesc[i] = strtoul(p, NULL, 10);
5397 	}
5398 
5399 	return(rc);
5400 }
5401 
5402 #define NAME_BUFLEN 32
5403 static void
5404 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5405 {
5406         device_t dev = iflib_get_dev(ctx);
5407 	struct sysctl_oid_list *child, *oid_list;
5408 	struct sysctl_ctx_list *ctx_list;
5409 	struct sysctl_oid *node;
5410 
5411 	ctx_list = device_get_sysctl_ctx(dev);
5412 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5413 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5414 						      CTLFLAG_RD, NULL, "IFLIB fields");
5415 	oid_list = SYSCTL_CHILDREN(node);
5416 
5417 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5418 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5419 		       "driver version");
5420 
5421 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5422 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5423 			"# of txqs to use, 0 => use default #");
5424 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5425 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5426 			"# of rxqs to use, 0 => use default #");
5427 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5428 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5429                        "permit #txq != #rxq");
5430        SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5431                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5432                       "disable MSIX (default 0)");
5433 
5434 	/* XXX change for per-queue sizes */
5435 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5436 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5437                        mp_ndesc_handler, "A",
5438                        "list of # of tx descriptors to use, 0 = use default #");
5439 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5440 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5441                        mp_ndesc_handler, "A",
5442                        "list of # of rx descriptors to use, 0 = use default #");
5443 }
5444 
5445 static void
5446 iflib_add_device_sysctl_post(if_ctx_t ctx)
5447 {
5448 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5449 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5450         device_t dev = iflib_get_dev(ctx);
5451 	struct sysctl_oid_list *child;
5452 	struct sysctl_ctx_list *ctx_list;
5453 	iflib_fl_t fl;
5454 	iflib_txq_t txq;
5455 	iflib_rxq_t rxq;
5456 	int i, j;
5457 	char namebuf[NAME_BUFLEN];
5458 	char *qfmt;
5459 	struct sysctl_oid *queue_node, *fl_node, *node;
5460 	struct sysctl_oid_list *queue_list, *fl_list;
5461 	ctx_list = device_get_sysctl_ctx(dev);
5462 
5463 	node = ctx->ifc_sysctl_node;
5464 	child = SYSCTL_CHILDREN(node);
5465 
5466 	if (scctx->isc_ntxqsets > 100)
5467 		qfmt = "txq%03d";
5468 	else if (scctx->isc_ntxqsets > 10)
5469 		qfmt = "txq%02d";
5470 	else
5471 		qfmt = "txq%d";
5472 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5473 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5474 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5475 					     CTLFLAG_RD, NULL, "Queue Name");
5476 		queue_list = SYSCTL_CHILDREN(queue_node);
5477 #if MEMORY_LOGGING
5478 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5479 				CTLFLAG_RD,
5480 				&txq->ift_dequeued, "total mbufs freed");
5481 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5482 				CTLFLAG_RD,
5483 				&txq->ift_enqueued, "total mbufs enqueued");
5484 #endif
5485 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5486 				   CTLFLAG_RD,
5487 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5488 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5489 				   CTLFLAG_RD,
5490 				   &txq->ift_pullups, "# of times m_pullup was called");
5491 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5492 				   CTLFLAG_RD,
5493 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5494 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5495 				   CTLFLAG_RD,
5496 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5497 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5498 				   CTLFLAG_RD,
5499 				   &txq->ift_map_failed, "# of times dma map failed");
5500 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5501 				   CTLFLAG_RD,
5502 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5503 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5504 				   CTLFLAG_RD,
5505 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5506 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5507 				   CTLFLAG_RD,
5508 				   &txq->ift_pidx, 1, "Producer Index");
5509 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5510 				   CTLFLAG_RD,
5511 				   &txq->ift_cidx, 1, "Consumer Index");
5512 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5513 				   CTLFLAG_RD,
5514 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5515 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5516 				   CTLFLAG_RD,
5517 				   &txq->ift_in_use, 1, "descriptors in use");
5518 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5519 				   CTLFLAG_RD,
5520 				   &txq->ift_processed, "descriptors procesed for clean");
5521 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5522 				   CTLFLAG_RD,
5523 				   &txq->ift_cleaned, "total cleaned");
5524 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5525 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5526 				0, mp_ring_state_handler, "A", "soft ring state");
5527 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5528 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5529 				       "# of enqueues to the mp_ring for this queue");
5530 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5531 				       CTLFLAG_RD, &txq->ift_br->drops,
5532 				       "# of drops in the mp_ring for this queue");
5533 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5534 				       CTLFLAG_RD, &txq->ift_br->starts,
5535 				       "# of normal consumer starts in the mp_ring for this queue");
5536 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5537 				       CTLFLAG_RD, &txq->ift_br->stalls,
5538 					       "# of consumer stalls in the mp_ring for this queue");
5539 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5540 			       CTLFLAG_RD, &txq->ift_br->restarts,
5541 				       "# of consumer restarts in the mp_ring for this queue");
5542 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5543 				       CTLFLAG_RD, &txq->ift_br->abdications,
5544 				       "# of consumer abdications in the mp_ring for this queue");
5545 	}
5546 
5547 	if (scctx->isc_nrxqsets > 100)
5548 		qfmt = "rxq%03d";
5549 	else if (scctx->isc_nrxqsets > 10)
5550 		qfmt = "rxq%02d";
5551 	else
5552 		qfmt = "rxq%d";
5553 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5554 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5555 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5556 					     CTLFLAG_RD, NULL, "Queue Name");
5557 		queue_list = SYSCTL_CHILDREN(queue_node);
5558 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5559 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5560 				       CTLFLAG_RD,
5561 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5562 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5563 				       CTLFLAG_RD,
5564 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5565 		}
5566 
5567 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5568 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5569 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5570 						     CTLFLAG_RD, NULL, "freelist Name");
5571 			fl_list = SYSCTL_CHILDREN(fl_node);
5572 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5573 				       CTLFLAG_RD,
5574 				       &fl->ifl_pidx, 1, "Producer Index");
5575 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5576 				       CTLFLAG_RD,
5577 				       &fl->ifl_cidx, 1, "Consumer Index");
5578 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5579 				       CTLFLAG_RD,
5580 				       &fl->ifl_credits, 1, "credits available");
5581 #if MEMORY_LOGGING
5582 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5583 					CTLFLAG_RD,
5584 					&fl->ifl_m_enqueued, "mbufs allocated");
5585 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5586 					CTLFLAG_RD,
5587 					&fl->ifl_m_dequeued, "mbufs freed");
5588 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5589 					CTLFLAG_RD,
5590 					&fl->ifl_cl_enqueued, "clusters allocated");
5591 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5592 					CTLFLAG_RD,
5593 					&fl->ifl_cl_dequeued, "clusters freed");
5594 #endif
5595 
5596 		}
5597 	}
5598 
5599 }
5600 
5601 #ifndef __NO_STRICT_ALIGNMENT
5602 static struct mbuf *
5603 iflib_fixup_rx(struct mbuf *m)
5604 {
5605 	struct mbuf *n;
5606 
5607 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5608 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5609 		m->m_data += ETHER_HDR_LEN;
5610 		n = m;
5611 	} else {
5612 		MGETHDR(n, M_NOWAIT, MT_DATA);
5613 		if (n == NULL) {
5614 			m_freem(m);
5615 			return (NULL);
5616 		}
5617 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5618 		m->m_data += ETHER_HDR_LEN;
5619 		m->m_len -= ETHER_HDR_LEN;
5620 		n->m_len = ETHER_HDR_LEN;
5621 		M_MOVE_PKTHDR(n, m);
5622 		n->m_next = m;
5623 	}
5624 	return (n);
5625 }
5626 #endif
5627