xref: /freebsd/sys/net/iflib.c (revision 5c1d8c4b73afad759f45def8f38f6895b404863b)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
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 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/sockio.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/kobj.h>
46 #include <sys/rman.h>
47 #include <sys/sbuf.h>
48 #include <sys/smp.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #if defined(__i386__) || defined(__amd64__)
92 #include <sys/memdesc.h>
93 #include <machine/bus.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
96 #include <x86/include/busdma_impl.h>
97 #include <x86/iommu/busdma_dmar.h>
98 #endif
99 
100 #include <sys/bitstring.h>
101 /*
102  * enable accounting of every mbuf as it comes in to and goes out of
103  * iflib's software descriptor references
104  */
105 #define MEMORY_LOGGING 0
106 /*
107  * Enable mbuf vectors for compressing long mbuf chains
108  */
109 
110 /*
111  * NB:
112  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
113  *   we prefetch needs to be determined by the time spent in m_free vis a vis
114  *   the cost of a prefetch. This will of course vary based on the workload:
115  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
116  *        is quite expensive, thus suggesting very little prefetch.
117  *      - small packet forwarding which is just returning a single mbuf to
118  *        UMA will typically be very fast vis a vis the cost of a memory
119  *        access.
120  */
121 
122 
123 /*
124  * File organization:
125  *  - private structures
126  *  - iflib private utility functions
127  *  - ifnet functions
128  *  - vlan registry and other exported functions
129  *  - iflib public core functions
130  *
131  *
132  */
133 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155    /*
156    * Pointer to hardware driver's softc
157    */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct mtx ifc_ctx_mtx;
167 	struct mtx ifc_state_mtx;
168 
169 	uint16_t ifc_nhwtxqs;
170 	uint16_t ifc_nhwrxqs;
171 
172 	iflib_txq_t ifc_txqs;
173 	iflib_rxq_t ifc_rxqs;
174 	uint32_t ifc_if_flags;
175 	uint32_t ifc_flags;
176 	uint32_t ifc_max_fl_buf_size;
177 	int ifc_in_detach;
178 
179 	int ifc_link_state;
180 	int ifc_link_irq;
181 	int ifc_watchdog_events;
182 	struct cdev *ifc_led_dev;
183 	struct resource *ifc_msix_mem;
184 
185 	struct if_irq ifc_legacy_irq;
186 	struct grouptask ifc_admin_task;
187 	struct grouptask ifc_vflr_task;
188 	struct iflib_filter_info ifc_filter_info;
189 	struct ifmedia	ifc_media;
190 
191 	struct sysctl_oid *ifc_sysctl_node;
192 	uint16_t ifc_sysctl_ntxqs;
193 	uint16_t ifc_sysctl_nrxqs;
194 	uint16_t ifc_sysctl_qs_eq_override;
195 	uint16_t ifc_sysctl_rx_budget;
196 
197 	qidx_t ifc_sysctl_ntxds[8];
198 	qidx_t ifc_sysctl_nrxds[8];
199 	struct if_txrx ifc_txrx;
200 #define isc_txd_encap  ifc_txrx.ift_txd_encap
201 #define isc_txd_flush  ifc_txrx.ift_txd_flush
202 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
203 #define isc_rxd_available ifc_txrx.ift_rxd_available
204 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	uint8_t ifc_mac[ETHER_ADDR_LEN];
213 	char ifc_mtx_name[16];
214 };
215 
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (&ctx->ifc_media);
243 }
244 
245 void
246 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
247 {
248 
249 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
250 }
251 
252 if_softc_ctx_t
253 iflib_get_softc_ctx(if_ctx_t ctx)
254 {
255 
256 	return (&ctx->ifc_softc_ctx);
257 }
258 
259 if_shared_ctx_t
260 iflib_get_sctx(if_ctx_t ctx)
261 {
262 
263 	return (ctx->ifc_sctx);
264 }
265 
266 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
267 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
268 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
269 
270 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
271 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
272 
273 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
274 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
275 #define RX_SW_DESC_INUSE        (1 << 3)
276 #define TX_SW_DESC_MAPPED       (1 << 4)
277 
278 #define	M_TOOBIG		M_PROTO1
279 
280 typedef struct iflib_sw_rx_desc_array {
281 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
282 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
283 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
284 	uint8_t		*ifsd_flags;
285 } iflib_rxsd_array_t;
286 
287 typedef struct iflib_sw_tx_desc_array {
288 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 	uint8_t		*ifsd_flags;
291 } if_txsd_vec_t;
292 
293 
294 /* magic number that should be high enough for any hardware */
295 #define IFLIB_MAX_TX_SEGS		128
296 /* bnxt supports 64 with hardware LRO enabled */
297 #define IFLIB_MAX_RX_SEGS		64
298 #define IFLIB_RX_COPY_THRESH		128
299 #define IFLIB_MAX_RX_REFRESH		32
300 /* The minimum descriptors per second before we start coalescing */
301 #define IFLIB_MIN_DESC_SEC		16384
302 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
303 #define IFLIB_QUEUE_IDLE		0
304 #define IFLIB_QUEUE_HUNG		1
305 #define IFLIB_QUEUE_WORKING		2
306 /* maximum number of txqs that can share an rx interrupt */
307 #define IFLIB_MAX_TX_SHARED_INTR	4
308 
309 /* this should really scale with ring size - this is a fairly arbitrary value */
310 #define TX_BATCH_SIZE			32
311 
312 #define IFLIB_RESTART_BUDGET		8
313 
314 #define	IFC_LEGACY		0x001
315 #define	IFC_QFLUSH		0x002
316 #define	IFC_MULTISEG		0x004
317 #define	IFC_DMAR		0x008
318 #define	IFC_SC_ALLOCATED	0x010
319 #define	IFC_INIT_DONE		0x020
320 #define	IFC_PREFETCH		0x040
321 #define	IFC_DO_RESET		0x080
322 #define	IFC_DO_WATCHDOG		0x100
323 #define	IFC_CHECK_HUNG		0x200
324 
325 
326 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
327 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
328 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
329 struct iflib_txq {
330 	qidx_t		ift_in_use;
331 	qidx_t		ift_cidx;
332 	qidx_t		ift_cidx_processed;
333 	qidx_t		ift_pidx;
334 	uint8_t		ift_gen;
335 	uint8_t		ift_br_offset;
336 	uint16_t	ift_npending;
337 	uint16_t	ift_db_pending;
338 	uint16_t	ift_rs_pending;
339 	/* implicit pad */
340 	uint8_t		ift_txd_size[8];
341 	uint64_t	ift_processed;
342 	uint64_t	ift_cleaned;
343 	uint64_t	ift_cleaned_prev;
344 #if MEMORY_LOGGING
345 	uint64_t	ift_enqueued;
346 	uint64_t	ift_dequeued;
347 #endif
348 	uint64_t	ift_no_tx_dma_setup;
349 	uint64_t	ift_no_desc_avail;
350 	uint64_t	ift_mbuf_defrag_failed;
351 	uint64_t	ift_mbuf_defrag;
352 	uint64_t	ift_map_failed;
353 	uint64_t	ift_txd_encap_efbig;
354 	uint64_t	ift_pullups;
355 
356 	struct mtx	ift_mtx;
357 	struct mtx	ift_db_mtx;
358 
359 	/* constant values */
360 	if_ctx_t	ift_ctx;
361 	struct ifmp_ring        *ift_br;
362 	struct grouptask	ift_task;
363 	qidx_t		ift_size;
364 	uint16_t	ift_id;
365 	struct callout	ift_timer;
366 
367 	if_txsd_vec_t	ift_sds;
368 	uint8_t		ift_qstatus;
369 	uint8_t		ift_closed;
370 	uint8_t		ift_update_freq;
371 	struct iflib_filter_info ift_filter_info;
372 	bus_dma_tag_t		ift_desc_tag;
373 	bus_dma_tag_t		ift_tso_desc_tag;
374 	iflib_dma_info_t	ift_ifdi;
375 #define MTX_NAME_LEN 16
376 	char                    ift_mtx_name[MTX_NAME_LEN];
377 	char                    ift_db_mtx_name[MTX_NAME_LEN];
378 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
379 #ifdef IFLIB_DIAGNOSTICS
380 	uint64_t ift_cpu_exec_count[256];
381 #endif
382 } __aligned(CACHE_LINE_SIZE);
383 
384 struct iflib_fl {
385 	qidx_t		ifl_cidx;
386 	qidx_t		ifl_pidx;
387 	qidx_t		ifl_credits;
388 	uint8_t		ifl_gen;
389 	uint8_t		ifl_rxd_size;
390 #if MEMORY_LOGGING
391 	uint64_t	ifl_m_enqueued;
392 	uint64_t	ifl_m_dequeued;
393 	uint64_t	ifl_cl_enqueued;
394 	uint64_t	ifl_cl_dequeued;
395 #endif
396 	/* implicit pad */
397 
398 	bitstr_t 	*ifl_rx_bitmap;
399 	qidx_t		ifl_fragidx;
400 	/* constant */
401 	qidx_t		ifl_size;
402 	uint16_t	ifl_buf_size;
403 	uint16_t	ifl_cltype;
404 	uma_zone_t	ifl_zone;
405 	iflib_rxsd_array_t	ifl_sds;
406 	iflib_rxq_t	ifl_rxq;
407 	uint8_t		ifl_id;
408 	bus_dma_tag_t           ifl_desc_tag;
409 	iflib_dma_info_t	ifl_ifdi;
410 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
411 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
412 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
413 }  __aligned(CACHE_LINE_SIZE);
414 
415 static inline qidx_t
416 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
417 {
418 	qidx_t used;
419 
420 	if (pidx > cidx)
421 		used = pidx - cidx;
422 	else if (pidx < cidx)
423 		used = size - cidx + pidx;
424 	else if (gen == 0 && pidx == cidx)
425 		used = 0;
426 	else if (gen == 1 && pidx == cidx)
427 		used = size;
428 	else
429 		panic("bad state");
430 
431 	return (used);
432 }
433 
434 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
435 
436 #define IDXDIFF(head, tail, wrap) \
437 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
438 
439 struct iflib_rxq {
440 	/* If there is a separate completion queue -
441 	 * these are the cq cidx and pidx. Otherwise
442 	 * these are unused.
443 	 */
444 	qidx_t		ifr_size;
445 	qidx_t		ifr_cq_cidx;
446 	qidx_t		ifr_cq_pidx;
447 	uint8_t		ifr_cq_gen;
448 	uint8_t		ifr_fl_offset;
449 
450 	if_ctx_t	ifr_ctx;
451 	iflib_fl_t	ifr_fl;
452 	uint64_t	ifr_rx_irq;
453 	uint16_t	ifr_id;
454 	uint8_t		ifr_lro_enabled;
455 	uint8_t		ifr_nfl;
456 	uint8_t		ifr_ntxqirq;
457 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
458 	struct lro_ctrl			ifr_lc;
459 	struct grouptask        ifr_task;
460 	struct iflib_filter_info ifr_filter_info;
461 	iflib_dma_info_t		ifr_ifdi;
462 
463 	/* dynamically allocate if any drivers need a value substantially larger than this */
464 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
465 #ifdef IFLIB_DIAGNOSTICS
466 	uint64_t ifr_cpu_exec_count[256];
467 #endif
468 }  __aligned(CACHE_LINE_SIZE);
469 
470 typedef struct if_rxsd {
471 	caddr_t *ifsd_cl;
472 	struct mbuf **ifsd_m;
473 	iflib_fl_t ifsd_fl;
474 	qidx_t ifsd_cidx;
475 } *if_rxsd_t;
476 
477 /* multiple of word size */
478 #ifdef __LP64__
479 #define PKT_INFO_SIZE	6
480 #define RXD_INFO_SIZE	5
481 #define PKT_TYPE uint64_t
482 #else
483 #define PKT_INFO_SIZE	11
484 #define RXD_INFO_SIZE	8
485 #define PKT_TYPE uint32_t
486 #endif
487 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
488 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
489 
490 typedef struct if_pkt_info_pad {
491 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
492 } *if_pkt_info_pad_t;
493 typedef struct if_rxd_info_pad {
494 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
495 } *if_rxd_info_pad_t;
496 
497 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
498 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
499 
500 
501 static inline void
502 pkt_info_zero(if_pkt_info_t pi)
503 {
504 	if_pkt_info_pad_t pi_pad;
505 
506 	pi_pad = (if_pkt_info_pad_t)pi;
507 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
508 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
509 #ifndef __LP64__
510 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
511 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
512 #endif
513 }
514 
515 static inline void
516 rxd_info_zero(if_rxd_info_t ri)
517 {
518 	if_rxd_info_pad_t ri_pad;
519 	int i;
520 
521 	ri_pad = (if_rxd_info_pad_t)ri;
522 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
523 		ri_pad->rxd_val[i] = 0;
524 		ri_pad->rxd_val[i+1] = 0;
525 		ri_pad->rxd_val[i+2] = 0;
526 		ri_pad->rxd_val[i+3] = 0;
527 	}
528 #ifdef __LP64__
529 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
530 #endif
531 }
532 
533 /*
534  * Only allow a single packet to take up most 1/nth of the tx ring
535  */
536 #define MAX_SINGLE_PACKET_FRACTION 12
537 #define IF_BAD_DMA (bus_addr_t)-1
538 
539 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
540 
541 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_ctx_mtx, _name, "iflib ctx lock", MTX_DEF)
542 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_ctx_mtx)
543 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_ctx_mtx)
544 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_ctx_mtx)
545 
546 
547 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
548 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
549 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
550 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
551 
552 
553 
554 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
555 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
556 
557 
558 /* Our boot-time initialization hook */
559 static int	iflib_module_event_handler(module_t, int, void *);
560 
561 static moduledata_t iflib_moduledata = {
562 	"iflib",
563 	iflib_module_event_handler,
564 	NULL
565 };
566 
567 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
568 MODULE_VERSION(iflib, 1);
569 
570 MODULE_DEPEND(iflib, pci, 1, 1, 1);
571 MODULE_DEPEND(iflib, ether, 1, 1, 1);
572 
573 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
574 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
575 
576 #ifndef IFLIB_DEBUG_COUNTERS
577 #ifdef INVARIANTS
578 #define IFLIB_DEBUG_COUNTERS 1
579 #else
580 #define IFLIB_DEBUG_COUNTERS 0
581 #endif /* !INVARIANTS */
582 #endif
583 
584 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
585                    "iflib driver parameters");
586 
587 /*
588  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
589  */
590 static int iflib_min_tx_latency = 0;
591 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
592 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
593 static int iflib_no_tx_batch = 0;
594 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
595 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
596 
597 
598 #if IFLIB_DEBUG_COUNTERS
599 
600 static int iflib_tx_seen;
601 static int iflib_tx_sent;
602 static int iflib_tx_encap;
603 static int iflib_rx_allocs;
604 static int iflib_fl_refills;
605 static int iflib_fl_refills_large;
606 static int iflib_tx_frees;
607 
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
609 		   &iflib_tx_seen, 0, "# tx mbufs seen");
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
611 		   &iflib_tx_sent, 0, "# tx mbufs sent");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
613 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
615 		   &iflib_tx_frees, 0, "# tx frees");
616 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
617 		   &iflib_rx_allocs, 0, "# rx allocations");
618 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
619 		   &iflib_fl_refills, 0, "# refills");
620 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
621 		   &iflib_fl_refills_large, 0, "# large refills");
622 
623 
624 static int iflib_txq_drain_flushing;
625 static int iflib_txq_drain_oactive;
626 static int iflib_txq_drain_notready;
627 static int iflib_txq_drain_encapfail;
628 
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
630 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
632 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
634 		   &iflib_txq_drain_notready, 0, "# drain notready");
635 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
636 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
637 
638 
639 static int iflib_encap_load_mbuf_fail;
640 static int iflib_encap_pad_mbuf_fail;
641 static int iflib_encap_txq_avail_fail;
642 static int iflib_encap_txd_encap_fail;
643 
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
645 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
647 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
649 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
650 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
651 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
652 
653 static int iflib_task_fn_rxs;
654 static int iflib_rx_intr_enables;
655 static int iflib_fast_intrs;
656 static int iflib_intr_link;
657 static int iflib_intr_msix;
658 static int iflib_rx_unavail;
659 static int iflib_rx_ctx_inactive;
660 static int iflib_rx_zero_len;
661 static int iflib_rx_if_input;
662 static int iflib_rx_mbuf_null;
663 static int iflib_rxd_flush;
664 
665 static int iflib_verbose_debug;
666 
667 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
668 		   &iflib_intr_link, 0, "# intr link calls");
669 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
670 		   &iflib_intr_msix, 0, "# intr msix calls");
671 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
672 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
673 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
674 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
675 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
676 		   &iflib_fast_intrs, 0, "# fast_intr calls");
677 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
678 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
679 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
680 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
681 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
682 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
683 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
684 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
685 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
686 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
687 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
688 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
689 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
690 		   &iflib_verbose_debug, 0, "enable verbose debugging");
691 
692 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
693 static void
694 iflib_debug_reset(void)
695 {
696 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
697 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
698 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
699 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
700 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
701 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
702 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
703 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
704 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
705 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
706 }
707 
708 #else
709 #define DBG_COUNTER_INC(name)
710 static void iflib_debug_reset(void) {}
711 #endif
712 
713 
714 
715 #define IFLIB_DEBUG 0
716 
717 static void iflib_tx_structures_free(if_ctx_t ctx);
718 static void iflib_rx_structures_free(if_ctx_t ctx);
719 static int iflib_queues_alloc(if_ctx_t ctx);
720 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
721 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
722 static int iflib_qset_structures_setup(if_ctx_t ctx);
723 static int iflib_msix_init(if_ctx_t ctx);
724 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
725 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
726 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
727 static int iflib_register(if_ctx_t);
728 static void iflib_init_locked(if_ctx_t ctx);
729 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
730 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
731 static void iflib_ifmp_purge(iflib_txq_t txq);
732 static void _iflib_pre_assert(if_softc_ctx_t scctx);
733 static void iflib_stop(if_ctx_t ctx);
734 static void iflib_if_init_locked(if_ctx_t ctx);
735 #ifndef __NO_STRICT_ALIGNMENT
736 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
737 #endif
738 
739 #ifdef DEV_NETMAP
740 #include <sys/selinfo.h>
741 #include <net/netmap.h>
742 #include <dev/netmap/netmap_kern.h>
743 
744 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
745 
746 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
747 
748 /*
749  * device-specific sysctl variables:
750  *
751  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
752  *	During regular operations the CRC is stripped, but on some
753  *	hardware reception of frames not multiple of 64 is slower,
754  *	so using crcstrip=0 helps in benchmarks.
755  *
756  * iflib_rx_miss, iflib_rx_miss_bufs:
757  *	count packets that might be missed due to lost interrupts.
758  */
759 SYSCTL_DECL(_dev_netmap);
760 /*
761  * The xl driver by default strips CRCs and we do not override it.
762  */
763 
764 int iflib_crcstrip = 1;
765 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
766     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
767 
768 int iflib_rx_miss, iflib_rx_miss_bufs;
769 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
770     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
771 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
772     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
773 
774 /*
775  * Register/unregister. We are already under netmap lock.
776  * Only called on the first register or the last unregister.
777  */
778 static int
779 iflib_netmap_register(struct netmap_adapter *na, int onoff)
780 {
781 	struct ifnet *ifp = na->ifp;
782 	if_ctx_t ctx = ifp->if_softc;
783 	int status;
784 
785 	CTX_LOCK(ctx);
786 	IFDI_INTR_DISABLE(ctx);
787 
788 	/* Tell the stack that the interface is no longer active */
789 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
790 
791 	if (!CTX_IS_VF(ctx))
792 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
793 
794 	/* enable or disable flags and callbacks in na and ifp */
795 	if (onoff) {
796 		nm_set_native_flags(na);
797 	} else {
798 		nm_clear_native_flags(na);
799 	}
800 	iflib_stop(ctx);
801 	iflib_init_locked(ctx);
802 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
803 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
804 	if (status)
805 		nm_clear_native_flags(na);
806 	CTX_UNLOCK(ctx);
807 	return (status);
808 }
809 
810 static int
811 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
812 {
813 	struct netmap_adapter *na = kring->na;
814 	u_int const lim = kring->nkr_num_slots - 1;
815 	u_int head = kring->rhead;
816 	struct netmap_ring *ring = kring->ring;
817 	bus_dmamap_t *map;
818 	struct if_rxd_update iru;
819 	if_ctx_t ctx = rxq->ifr_ctx;
820 	iflib_fl_t fl = &rxq->ifr_fl[0];
821 	uint32_t refill_pidx, nic_i;
822 
823 	if (nm_i == head && __predict_true(!init))
824 		return 0;
825 	iru_init(&iru, rxq, 0 /* flid */);
826 	map = fl->ifl_sds.ifsd_map;
827 	refill_pidx = netmap_idx_k2n(kring, nm_i);
828 	/*
829 	 * IMPORTANT: we must leave one free slot in the ring,
830 	 * so move head back by one unit
831 	 */
832 	head = nm_prev(head, lim);
833 	while (nm_i != head) {
834 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
835 			struct netmap_slot *slot = &ring->slot[nm_i];
836 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
837 			uint32_t nic_i_dma = refill_pidx;
838 			nic_i = netmap_idx_k2n(kring, nm_i);
839 
840 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
841 
842 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
843 			        return netmap_ring_reinit(kring);
844 
845 			fl->ifl_vm_addrs[tmp_pidx] = addr;
846 			if (__predict_false(init) && map) {
847 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
848 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
849 				/* buffer has changed, reload map */
850 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
851 			}
852 			slot->flags &= ~NS_BUF_CHANGED;
853 
854 			nm_i = nm_next(nm_i, lim);
855 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
856 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
857 				continue;
858 
859 			iru.iru_pidx = refill_pidx;
860 			iru.iru_count = tmp_pidx+1;
861 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
862 
863 			refill_pidx = nic_i;
864 			if (map == NULL)
865 				continue;
866 
867 			for (int n = 0; n < iru.iru_count; n++) {
868 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
869 						BUS_DMASYNC_PREREAD);
870 				/* XXX - change this to not use the netmap func*/
871 				nic_i_dma = nm_next(nic_i_dma, lim);
872 			}
873 		}
874 	}
875 	kring->nr_hwcur = head;
876 
877 	if (map)
878 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
879 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
880 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
881 	return (0);
882 }
883 
884 /*
885  * Reconcile kernel and user view of the transmit ring.
886  *
887  * All information is in the kring.
888  * Userspace wants to send packets up to the one before kring->rhead,
889  * kernel knows kring->nr_hwcur is the first unsent packet.
890  *
891  * Here we push packets out (as many as possible), and possibly
892  * reclaim buffers from previously completed transmission.
893  *
894  * The caller (netmap) guarantees that there is only one instance
895  * running at any time. Any interference with other driver
896  * methods should be handled by the individual drivers.
897  */
898 static int
899 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
900 {
901 	struct netmap_adapter *na = kring->na;
902 	struct ifnet *ifp = na->ifp;
903 	struct netmap_ring *ring = kring->ring;
904 	u_int nm_i;	/* index into the netmap ring */
905 	u_int nic_i;	/* index into the NIC ring */
906 	u_int n;
907 	u_int const lim = kring->nkr_num_slots - 1;
908 	u_int const head = kring->rhead;
909 	struct if_pkt_info pi;
910 
911 	/*
912 	 * interrupts on every tx packet are expensive so request
913 	 * them every half ring, or where NS_REPORT is set
914 	 */
915 	u_int report_frequency = kring->nkr_num_slots >> 1;
916 	/* device-specific */
917 	if_ctx_t ctx = ifp->if_softc;
918 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
919 
920 	if (txq->ift_sds.ifsd_map)
921 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
922 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
923 
924 
925 	/*
926 	 * First part: process new packets to send.
927 	 * nm_i is the current index in the netmap ring,
928 	 * nic_i is the corresponding index in the NIC ring.
929 	 *
930 	 * If we have packets to send (nm_i != head)
931 	 * iterate over the netmap ring, fetch length and update
932 	 * the corresponding slot in the NIC ring. Some drivers also
933 	 * need to update the buffer's physical address in the NIC slot
934 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
935 	 *
936 	 * The netmap_reload_map() calls is especially expensive,
937 	 * even when (as in this case) the tag is 0, so do only
938 	 * when the buffer has actually changed.
939 	 *
940 	 * If possible do not set the report/intr bit on all slots,
941 	 * but only a few times per ring or when NS_REPORT is set.
942 	 *
943 	 * Finally, on 10G and faster drivers, it might be useful
944 	 * to prefetch the next slot and txr entry.
945 	 */
946 
947 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
948 	pkt_info_zero(&pi);
949 	pi.ipi_segs = txq->ift_segs;
950 	pi.ipi_qsidx = kring->ring_id;
951 	if (nm_i != head) {	/* we have new packets to send */
952 		nic_i = netmap_idx_k2n(kring, nm_i);
953 
954 		__builtin_prefetch(&ring->slot[nm_i]);
955 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
956 		if (txq->ift_sds.ifsd_map)
957 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
958 
959 		for (n = 0; nm_i != head; n++) {
960 			struct netmap_slot *slot = &ring->slot[nm_i];
961 			u_int len = slot->len;
962 			uint64_t paddr;
963 			void *addr = PNMB(na, slot, &paddr);
964 			int flags = (slot->flags & NS_REPORT ||
965 				nic_i == 0 || nic_i == report_frequency) ?
966 				IPI_TX_INTR : 0;
967 
968 			/* device-specific */
969 			pi.ipi_len = len;
970 			pi.ipi_segs[0].ds_addr = paddr;
971 			pi.ipi_segs[0].ds_len = len;
972 			pi.ipi_nsegs = 1;
973 			pi.ipi_ndescs = 0;
974 			pi.ipi_pidx = nic_i;
975 			pi.ipi_flags = flags;
976 
977 			/* Fill the slot in the NIC ring. */
978 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
979 
980 			/* prefetch for next round */
981 			__builtin_prefetch(&ring->slot[nm_i + 1]);
982 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
983 			if (txq->ift_sds.ifsd_map) {
984 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
985 
986 				NM_CHECK_ADDR_LEN(na, addr, len);
987 
988 				if (slot->flags & NS_BUF_CHANGED) {
989 					/* buffer has changed, reload map */
990 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
991 				}
992 				/* make sure changes to the buffer are synced */
993 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
994 						BUS_DMASYNC_PREWRITE);
995 			}
996 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
997 			nm_i = nm_next(nm_i, lim);
998 			nic_i = nm_next(nic_i, lim);
999 		}
1000 		kring->nr_hwcur = head;
1001 
1002 		/* synchronize the NIC ring */
1003 		if (txq->ift_sds.ifsd_map)
1004 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1005 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1006 
1007 		/* (re)start the tx unit up to slot nic_i (excluded) */
1008 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1009 	}
1010 
1011 	/*
1012 	 * Second part: reclaim buffers for completed transmissions.
1013 	 */
1014 	if (iflib_tx_credits_update(ctx, txq)) {
1015 		/* some tx completed, increment avail */
1016 		nic_i = txq->ift_cidx_processed;
1017 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1018 	}
1019 	return (0);
1020 }
1021 
1022 /*
1023  * Reconcile kernel and user view of the receive ring.
1024  * Same as for the txsync, this routine must be efficient.
1025  * The caller guarantees a single invocations, but races against
1026  * the rest of the driver should be handled here.
1027  *
1028  * On call, kring->rhead is the first packet that userspace wants
1029  * to keep, and kring->rcur is the wakeup point.
1030  * The kernel has previously reported packets up to kring->rtail.
1031  *
1032  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1033  * of whether or not we received an interrupt.
1034  */
1035 static int
1036 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1037 {
1038 	struct netmap_adapter *na = kring->na;
1039 	struct netmap_ring *ring = kring->ring;
1040 	uint32_t nm_i;	/* index into the netmap ring */
1041 	uint32_t nic_i;	/* index into the NIC ring */
1042 	u_int i, n;
1043 	u_int const lim = kring->nkr_num_slots - 1;
1044 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1045 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1046 	struct if_rxd_info ri;
1047 
1048 	struct ifnet *ifp = na->ifp;
1049 	if_ctx_t ctx = ifp->if_softc;
1050 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1051 	iflib_fl_t fl = rxq->ifr_fl;
1052 	if (head > lim)
1053 		return netmap_ring_reinit(kring);
1054 
1055 	/* XXX check sync modes */
1056 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1057 		if (fl->ifl_sds.ifsd_map == NULL)
1058 			continue;
1059 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1060 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1061 	}
1062 	/*
1063 	 * First part: import newly received packets.
1064 	 *
1065 	 * nm_i is the index of the next free slot in the netmap ring,
1066 	 * nic_i is the index of the next received packet in the NIC ring,
1067 	 * and they may differ in case if_init() has been called while
1068 	 * in netmap mode. For the receive ring we have
1069 	 *
1070 	 *	nic_i = rxr->next_check;
1071 	 *	nm_i = kring->nr_hwtail (previous)
1072 	 * and
1073 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1074 	 *
1075 	 * rxr->next_check is set to 0 on a ring reinit
1076 	 */
1077 	if (netmap_no_pendintr || force_update) {
1078 		int crclen = iflib_crcstrip ? 0 : 4;
1079 		int error, avail;
1080 
1081 		for (i = 0; i < rxq->ifr_nfl; i++) {
1082 			fl = &rxq->ifr_fl[i];
1083 			nic_i = fl->ifl_cidx;
1084 			nm_i = netmap_idx_n2k(kring, nic_i);
1085 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1086 			for (n = 0; avail > 0; n++, avail--) {
1087 				rxd_info_zero(&ri);
1088 				ri.iri_frags = rxq->ifr_frags;
1089 				ri.iri_qsidx = kring->ring_id;
1090 				ri.iri_ifp = ctx->ifc_ifp;
1091 				ri.iri_cidx = nic_i;
1092 
1093 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1094 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1095 				ring->slot[nm_i].flags = 0;
1096 				if (fl->ifl_sds.ifsd_map)
1097 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1098 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1099 				nm_i = nm_next(nm_i, lim);
1100 				nic_i = nm_next(nic_i, lim);
1101 			}
1102 			if (n) { /* update the state variables */
1103 				if (netmap_no_pendintr && !force_update) {
1104 					/* diagnostics */
1105 					iflib_rx_miss ++;
1106 					iflib_rx_miss_bufs += n;
1107 				}
1108 				fl->ifl_cidx = nic_i;
1109 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1110 			}
1111 			kring->nr_kflags &= ~NKR_PENDINTR;
1112 		}
1113 	}
1114 	/*
1115 	 * Second part: skip past packets that userspace has released.
1116 	 * (kring->nr_hwcur to head excluded),
1117 	 * and make the buffers available for reception.
1118 	 * As usual nm_i is the index in the netmap ring,
1119 	 * nic_i is the index in the NIC ring, and
1120 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1121 	 */
1122 	/* XXX not sure how this will work with multiple free lists */
1123 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1124 
1125 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1126 }
1127 
1128 static void
1129 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1130 {
1131 	struct ifnet *ifp = na->ifp;
1132 	if_ctx_t ctx = ifp->if_softc;
1133 
1134 	CTX_LOCK(ctx);
1135 	if (onoff) {
1136 		IFDI_INTR_ENABLE(ctx);
1137 	} else {
1138 		IFDI_INTR_DISABLE(ctx);
1139 	}
1140 	CTX_UNLOCK(ctx);
1141 }
1142 
1143 
1144 static int
1145 iflib_netmap_attach(if_ctx_t ctx)
1146 {
1147 	struct netmap_adapter na;
1148 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1149 
1150 	bzero(&na, sizeof(na));
1151 
1152 	na.ifp = ctx->ifc_ifp;
1153 	na.na_flags = NAF_BDG_MAYSLEEP;
1154 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1155 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1156 
1157 	na.num_tx_desc = scctx->isc_ntxd[0];
1158 	na.num_rx_desc = scctx->isc_nrxd[0];
1159 	na.nm_txsync = iflib_netmap_txsync;
1160 	na.nm_rxsync = iflib_netmap_rxsync;
1161 	na.nm_register = iflib_netmap_register;
1162 	na.nm_intr = iflib_netmap_intr;
1163 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1164 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1165 	return (netmap_attach(&na));
1166 }
1167 
1168 static void
1169 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1170 {
1171 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1172 	struct netmap_slot *slot;
1173 
1174 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1175 	if (slot == NULL)
1176 		return;
1177 	if (txq->ift_sds.ifsd_map == NULL)
1178 		return;
1179 
1180 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1181 
1182 		/*
1183 		 * In netmap mode, set the map for the packet buffer.
1184 		 * NOTE: Some drivers (not this one) also need to set
1185 		 * the physical buffer address in the NIC ring.
1186 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1187 		 * netmap slot index, si
1188 		 */
1189 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1190 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1191 	}
1192 }
1193 
1194 static void
1195 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1196 {
1197 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1198 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1199 	struct netmap_slot *slot;
1200 	uint32_t nm_i;
1201 
1202 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1203 	if (slot == NULL)
1204 		return;
1205 	nm_i = netmap_idx_n2k(kring, 0);
1206 	netmap_fl_refill(rxq, kring, nm_i, true);
1207 }
1208 
1209 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1210 
1211 #else
1212 #define iflib_netmap_txq_init(ctx, txq)
1213 #define iflib_netmap_rxq_init(ctx, rxq)
1214 #define iflib_netmap_detach(ifp)
1215 
1216 #define iflib_netmap_attach(ctx) (0)
1217 #define netmap_rx_irq(ifp, qid, budget) (0)
1218 #define netmap_tx_irq(ifp, qid) do {} while (0)
1219 
1220 #endif
1221 
1222 #if defined(__i386__) || defined(__amd64__)
1223 static __inline void
1224 prefetch(void *x)
1225 {
1226 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1227 }
1228 static __inline void
1229 prefetch2cachelines(void *x)
1230 {
1231 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1232 #if (CACHE_LINE_SIZE < 128)
1233 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1234 #endif
1235 }
1236 #else
1237 #define prefetch(x)
1238 #define prefetch2cachelines(x)
1239 #endif
1240 
1241 static void
1242 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1243 {
1244 	iflib_fl_t fl;
1245 
1246 	fl = &rxq->ifr_fl[flid];
1247 	iru->iru_paddrs = fl->ifl_bus_addrs;
1248 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1249 	iru->iru_idxs = fl->ifl_rxd_idxs;
1250 	iru->iru_qsidx = rxq->ifr_id;
1251 	iru->iru_buf_size = fl->ifl_buf_size;
1252 	iru->iru_flidx = fl->ifl_id;
1253 }
1254 
1255 static void
1256 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1257 {
1258 	if (err)
1259 		return;
1260 	*(bus_addr_t *) arg = segs[0].ds_addr;
1261 }
1262 
1263 int
1264 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1265 {
1266 	int err;
1267 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1268 	device_t dev = ctx->ifc_dev;
1269 
1270 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1271 
1272 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1273 				sctx->isc_q_align, 0,	/* alignment, bounds */
1274 				BUS_SPACE_MAXADDR,	/* lowaddr */
1275 				BUS_SPACE_MAXADDR,	/* highaddr */
1276 				NULL, NULL,		/* filter, filterarg */
1277 				size,			/* maxsize */
1278 				1,			/* nsegments */
1279 				size,			/* maxsegsize */
1280 				BUS_DMA_ALLOCNOW,	/* flags */
1281 				NULL,			/* lockfunc */
1282 				NULL,			/* lockarg */
1283 				&dma->idi_tag);
1284 	if (err) {
1285 		device_printf(dev,
1286 		    "%s: bus_dma_tag_create failed: %d\n",
1287 		    __func__, err);
1288 		goto fail_0;
1289 	}
1290 
1291 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1292 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1293 	if (err) {
1294 		device_printf(dev,
1295 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1296 		    __func__, (uintmax_t)size, err);
1297 		goto fail_1;
1298 	}
1299 
1300 	dma->idi_paddr = IF_BAD_DMA;
1301 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1302 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1303 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1304 		device_printf(dev,
1305 		    "%s: bus_dmamap_load failed: %d\n",
1306 		    __func__, err);
1307 		goto fail_2;
1308 	}
1309 
1310 	dma->idi_size = size;
1311 	return (0);
1312 
1313 fail_2:
1314 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1315 fail_1:
1316 	bus_dma_tag_destroy(dma->idi_tag);
1317 fail_0:
1318 	dma->idi_tag = NULL;
1319 
1320 	return (err);
1321 }
1322 
1323 int
1324 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1325 {
1326 	int i, err;
1327 	iflib_dma_info_t *dmaiter;
1328 
1329 	dmaiter = dmalist;
1330 	for (i = 0; i < count; i++, dmaiter++) {
1331 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1332 			break;
1333 	}
1334 	if (err)
1335 		iflib_dma_free_multi(dmalist, i);
1336 	return (err);
1337 }
1338 
1339 void
1340 iflib_dma_free(iflib_dma_info_t dma)
1341 {
1342 	if (dma->idi_tag == NULL)
1343 		return;
1344 	if (dma->idi_paddr != IF_BAD_DMA) {
1345 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1346 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1347 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1348 		dma->idi_paddr = IF_BAD_DMA;
1349 	}
1350 	if (dma->idi_vaddr != NULL) {
1351 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1352 		dma->idi_vaddr = NULL;
1353 	}
1354 	bus_dma_tag_destroy(dma->idi_tag);
1355 	dma->idi_tag = NULL;
1356 }
1357 
1358 void
1359 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1360 {
1361 	int i;
1362 	iflib_dma_info_t *dmaiter = dmalist;
1363 
1364 	for (i = 0; i < count; i++, dmaiter++)
1365 		iflib_dma_free(*dmaiter);
1366 }
1367 
1368 #ifdef EARLY_AP_STARTUP
1369 static const int iflib_started = 1;
1370 #else
1371 /*
1372  * We used to abuse the smp_started flag to decide if the queues have been
1373  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1374  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1375  * is set.  Run a SYSINIT() strictly after that to just set a usable
1376  * completion flag.
1377  */
1378 
1379 static int iflib_started;
1380 
1381 static void
1382 iflib_record_started(void *arg)
1383 {
1384 	iflib_started = 1;
1385 }
1386 
1387 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1388 	iflib_record_started, NULL);
1389 #endif
1390 
1391 static int
1392 iflib_fast_intr(void *arg)
1393 {
1394 	iflib_filter_info_t info = arg;
1395 	struct grouptask *gtask = info->ifi_task;
1396 	if (!iflib_started)
1397 		return (FILTER_HANDLED);
1398 
1399 	DBG_COUNTER_INC(fast_intrs);
1400 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1401 		return (FILTER_HANDLED);
1402 
1403 	GROUPTASK_ENQUEUE(gtask);
1404 	return (FILTER_HANDLED);
1405 }
1406 
1407 static int
1408 iflib_fast_intr_rxtx(void *arg)
1409 {
1410 	iflib_filter_info_t info = arg;
1411 	struct grouptask *gtask = info->ifi_task;
1412 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1413 	if_ctx_t ctx;
1414 	int i, cidx;
1415 
1416 	if (!iflib_started)
1417 		return (FILTER_HANDLED);
1418 
1419 	DBG_COUNTER_INC(fast_intrs);
1420 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1421 		return (FILTER_HANDLED);
1422 
1423 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1424 		qidx_t txqid = rxq->ifr_txqid[i];
1425 
1426 		ctx = rxq->ifr_ctx;
1427 
1428 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1429 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1430 			continue;
1431 		}
1432 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1433 	}
1434 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1435 		cidx = rxq->ifr_cq_cidx;
1436 	else
1437 		cidx = rxq->ifr_fl[0].ifl_cidx;
1438 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1439 		GROUPTASK_ENQUEUE(gtask);
1440 	else
1441 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1442 	return (FILTER_HANDLED);
1443 }
1444 
1445 
1446 static int
1447 iflib_fast_intr_ctx(void *arg)
1448 {
1449 	iflib_filter_info_t info = arg;
1450 	struct grouptask *gtask = info->ifi_task;
1451 
1452 	if (!iflib_started)
1453 		return (FILTER_HANDLED);
1454 
1455 	DBG_COUNTER_INC(fast_intrs);
1456 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1457 		return (FILTER_HANDLED);
1458 
1459 	GROUPTASK_ENQUEUE(gtask);
1460 	return (FILTER_HANDLED);
1461 }
1462 
1463 static int
1464 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1465 	driver_filter_t filter, driver_intr_t handler, void *arg,
1466 				 char *name)
1467 {
1468 	int rc, flags;
1469 	struct resource *res;
1470 	void *tag = NULL;
1471 	device_t dev = ctx->ifc_dev;
1472 
1473 	flags = RF_ACTIVE;
1474 	if (ctx->ifc_flags & IFC_LEGACY)
1475 		flags |= RF_SHAREABLE;
1476 	MPASS(rid < 512);
1477 	irq->ii_rid = rid;
1478 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1479 	if (res == NULL) {
1480 		device_printf(dev,
1481 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1482 		return (ENOMEM);
1483 	}
1484 	irq->ii_res = res;
1485 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1486 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1487 						filter, handler, arg, &tag);
1488 	if (rc != 0) {
1489 		device_printf(dev,
1490 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1491 					  rid, name ? name : "unknown", rc);
1492 		return (rc);
1493 	} else if (name)
1494 		bus_describe_intr(dev, res, tag, "%s", name);
1495 
1496 	irq->ii_tag = tag;
1497 	return (0);
1498 }
1499 
1500 
1501 /*********************************************************************
1502  *
1503  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1504  *  the information needed to transmit a packet on the wire. This is
1505  *  called only once at attach, setup is done every reset.
1506  *
1507  **********************************************************************/
1508 
1509 static int
1510 iflib_txsd_alloc(iflib_txq_t txq)
1511 {
1512 	if_ctx_t ctx = txq->ift_ctx;
1513 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1514 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1515 	device_t dev = ctx->ifc_dev;
1516 	int err, nsegments, ntsosegments;
1517 
1518 	nsegments = scctx->isc_tx_nsegments;
1519 	ntsosegments = scctx->isc_tx_tso_segments_max;
1520 	MPASS(scctx->isc_ntxd[0] > 0);
1521 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1522 	MPASS(nsegments > 0);
1523 	MPASS(ntsosegments > 0);
1524 	/*
1525 	 * Setup DMA descriptor areas.
1526 	 */
1527 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1528 			       1, 0,			/* alignment, bounds */
1529 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1530 			       BUS_SPACE_MAXADDR,	/* highaddr */
1531 			       NULL, NULL,		/* filter, filterarg */
1532 			       sctx->isc_tx_maxsize,		/* maxsize */
1533 			       nsegments,	/* nsegments */
1534 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1535 			       0,			/* flags */
1536 			       NULL,			/* lockfunc */
1537 			       NULL,			/* lockfuncarg */
1538 			       &txq->ift_desc_tag))) {
1539 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1540 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1541 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1542 		goto fail;
1543 	}
1544 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1545 			       1, 0,			/* alignment, bounds */
1546 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1547 			       BUS_SPACE_MAXADDR,	/* highaddr */
1548 			       NULL, NULL,		/* filter, filterarg */
1549 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1550 			       ntsosegments,	/* nsegments */
1551 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1552 			       0,			/* flags */
1553 			       NULL,			/* lockfunc */
1554 			       NULL,			/* lockfuncarg */
1555 			       &txq->ift_tso_desc_tag))) {
1556 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1557 
1558 		goto fail;
1559 	}
1560 	if (!(txq->ift_sds.ifsd_flags =
1561 	    (uint8_t *) malloc(sizeof(uint8_t) *
1562 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1563 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1564 		err = ENOMEM;
1565 		goto fail;
1566 	}
1567 	if (!(txq->ift_sds.ifsd_m =
1568 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1569 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1570 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1571 		err = ENOMEM;
1572 		goto fail;
1573 	}
1574 
1575         /* Create the descriptor buffer dma maps */
1576 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1577 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1578 		return (0);
1579 
1580 	if (!(txq->ift_sds.ifsd_map =
1581 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1582 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1583 		err = ENOMEM;
1584 		goto fail;
1585 	}
1586 
1587 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1588 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1589 		if (err != 0) {
1590 			device_printf(dev, "Unable to create TX DMA map\n");
1591 			goto fail;
1592 		}
1593 	}
1594 #endif
1595 	return (0);
1596 fail:
1597 	/* We free all, it handles case where we are in the middle */
1598 	iflib_tx_structures_free(ctx);
1599 	return (err);
1600 }
1601 
1602 static void
1603 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1604 {
1605 	bus_dmamap_t map;
1606 
1607 	map = NULL;
1608 	if (txq->ift_sds.ifsd_map != NULL)
1609 		map = txq->ift_sds.ifsd_map[i];
1610 	if (map != NULL) {
1611 		bus_dmamap_unload(txq->ift_desc_tag, map);
1612 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1613 		txq->ift_sds.ifsd_map[i] = NULL;
1614 	}
1615 }
1616 
1617 static void
1618 iflib_txq_destroy(iflib_txq_t txq)
1619 {
1620 	if_ctx_t ctx = txq->ift_ctx;
1621 
1622 	for (int i = 0; i < txq->ift_size; i++)
1623 		iflib_txsd_destroy(ctx, txq, i);
1624 	if (txq->ift_sds.ifsd_map != NULL) {
1625 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1626 		txq->ift_sds.ifsd_map = NULL;
1627 	}
1628 	if (txq->ift_sds.ifsd_m != NULL) {
1629 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1630 		txq->ift_sds.ifsd_m = NULL;
1631 	}
1632 	if (txq->ift_sds.ifsd_flags != NULL) {
1633 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1634 		txq->ift_sds.ifsd_flags = NULL;
1635 	}
1636 	if (txq->ift_desc_tag != NULL) {
1637 		bus_dma_tag_destroy(txq->ift_desc_tag);
1638 		txq->ift_desc_tag = NULL;
1639 	}
1640 	if (txq->ift_tso_desc_tag != NULL) {
1641 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1642 		txq->ift_tso_desc_tag = NULL;
1643 	}
1644 }
1645 
1646 static void
1647 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1648 {
1649 	struct mbuf **mp;
1650 
1651 	mp = &txq->ift_sds.ifsd_m[i];
1652 	if (*mp == NULL)
1653 		return;
1654 
1655 	if (txq->ift_sds.ifsd_map != NULL) {
1656 		bus_dmamap_sync(txq->ift_desc_tag,
1657 				txq->ift_sds.ifsd_map[i],
1658 				BUS_DMASYNC_POSTWRITE);
1659 		bus_dmamap_unload(txq->ift_desc_tag,
1660 				  txq->ift_sds.ifsd_map[i]);
1661 	}
1662 	m_free(*mp);
1663 	DBG_COUNTER_INC(tx_frees);
1664 	*mp = NULL;
1665 }
1666 
1667 static int
1668 iflib_txq_setup(iflib_txq_t txq)
1669 {
1670 	if_ctx_t ctx = txq->ift_ctx;
1671 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1672 	iflib_dma_info_t di;
1673 	int i;
1674 
1675 	/* Set number of descriptors available */
1676 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1677 	/* XXX make configurable */
1678 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1679 
1680 	/* Reset indices */
1681 	txq->ift_cidx_processed = 0;
1682 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1683 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1684 
1685 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1686 		bzero((void *)di->idi_vaddr, di->idi_size);
1687 
1688 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1689 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1690 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1691 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1692 	return (0);
1693 }
1694 
1695 /*********************************************************************
1696  *
1697  *  Allocate memory for rx_buffer structures. Since we use one
1698  *  rx_buffer per received packet, the maximum number of rx_buffer's
1699  *  that we'll need is equal to the number of receive descriptors
1700  *  that we've allocated.
1701  *
1702  **********************************************************************/
1703 static int
1704 iflib_rxsd_alloc(iflib_rxq_t rxq)
1705 {
1706 	if_ctx_t ctx = rxq->ifr_ctx;
1707 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1708 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1709 	device_t dev = ctx->ifc_dev;
1710 	iflib_fl_t fl;
1711 	int			err;
1712 
1713 	MPASS(scctx->isc_nrxd[0] > 0);
1714 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1715 
1716 	fl = rxq->ifr_fl;
1717 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1718 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1719 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1720 					 1, 0,			/* alignment, bounds */
1721 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1722 					 BUS_SPACE_MAXADDR,	/* highaddr */
1723 					 NULL, NULL,		/* filter, filterarg */
1724 					 sctx->isc_rx_maxsize,	/* maxsize */
1725 					 sctx->isc_rx_nsegments,	/* nsegments */
1726 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1727 					 0,			/* flags */
1728 					 NULL,			/* lockfunc */
1729 					 NULL,			/* lockarg */
1730 					 &fl->ifl_desc_tag);
1731 		if (err) {
1732 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1733 				__func__, err);
1734 			goto fail;
1735 		}
1736 		if (!(fl->ifl_sds.ifsd_flags =
1737 		      (uint8_t *) malloc(sizeof(uint8_t) *
1738 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1739 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1740 			err = ENOMEM;
1741 			goto fail;
1742 		}
1743 		if (!(fl->ifl_sds.ifsd_m =
1744 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1745 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1746 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1747 			err = ENOMEM;
1748 			goto fail;
1749 		}
1750 		if (!(fl->ifl_sds.ifsd_cl =
1751 		      (caddr_t *) malloc(sizeof(caddr_t) *
1752 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1753 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1754 			err = ENOMEM;
1755 			goto fail;
1756 		}
1757 
1758 		/* Create the descriptor buffer dma maps */
1759 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1760 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1761 			continue;
1762 
1763 		if (!(fl->ifl_sds.ifsd_map =
1764 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1765 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1766 			err = ENOMEM;
1767 			goto fail;
1768 		}
1769 
1770 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1771 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1772 			if (err != 0) {
1773 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1774 				goto fail;
1775 			}
1776 		}
1777 #endif
1778 	}
1779 	return (0);
1780 
1781 fail:
1782 	iflib_rx_structures_free(ctx);
1783 	return (err);
1784 }
1785 
1786 
1787 /*
1788  * Internal service routines
1789  */
1790 
1791 struct rxq_refill_cb_arg {
1792 	int               error;
1793 	bus_dma_segment_t seg;
1794 	int               nseg;
1795 };
1796 
1797 static void
1798 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1799 {
1800 	struct rxq_refill_cb_arg *cb_arg = arg;
1801 
1802 	cb_arg->error = error;
1803 	cb_arg->seg = segs[0];
1804 	cb_arg->nseg = nseg;
1805 }
1806 
1807 
1808 #ifdef ACPI_DMAR
1809 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1810 #else
1811 #define IS_DMAR(ctx) (0)
1812 #endif
1813 
1814 /**
1815  *	rxq_refill - refill an rxq  free-buffer list
1816  *	@ctx: the iflib context
1817  *	@rxq: the free-list to refill
1818  *	@n: the number of new buffers to allocate
1819  *
1820  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1821  *	The caller must assure that @n does not exceed the queue's capacity.
1822  */
1823 static void
1824 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1825 {
1826 	struct mbuf *m;
1827 	int idx, frag_idx = fl->ifl_fragidx;
1828         int pidx = fl->ifl_pidx;
1829 	caddr_t cl, *sd_cl;
1830 	struct mbuf **sd_m;
1831 	uint8_t *sd_flags;
1832 	struct if_rxd_update iru;
1833 	bus_dmamap_t *sd_map;
1834 	int n, i = 0;
1835 	uint64_t bus_addr;
1836 	int err;
1837 	qidx_t credits;
1838 
1839 	sd_m = fl->ifl_sds.ifsd_m;
1840 	sd_map = fl->ifl_sds.ifsd_map;
1841 	sd_cl = fl->ifl_sds.ifsd_cl;
1842 	sd_flags = fl->ifl_sds.ifsd_flags;
1843 	idx = pidx;
1844 	credits = fl->ifl_credits;
1845 
1846 	n  = count;
1847 	MPASS(n > 0);
1848 	MPASS(credits + n <= fl->ifl_size);
1849 
1850 	if (pidx < fl->ifl_cidx)
1851 		MPASS(pidx + n <= fl->ifl_cidx);
1852 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1853 		MPASS(fl->ifl_gen == 0);
1854 	if (pidx > fl->ifl_cidx)
1855 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1856 
1857 	DBG_COUNTER_INC(fl_refills);
1858 	if (n > 8)
1859 		DBG_COUNTER_INC(fl_refills_large);
1860 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1861 	while (n--) {
1862 		/*
1863 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1864 		 * initialized after rx.
1865 		 *
1866 		 * If the cluster is still set then we know a minimum sized packet was received
1867 		 */
1868 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1869 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1870                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1871 		if ((cl = sd_cl[frag_idx]) == NULL) {
1872                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1873 				break;
1874 #if MEMORY_LOGGING
1875 			fl->ifl_cl_enqueued++;
1876 #endif
1877 		}
1878 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1879 			break;
1880 		}
1881 #if MEMORY_LOGGING
1882 		fl->ifl_m_enqueued++;
1883 #endif
1884 
1885 		DBG_COUNTER_INC(rx_allocs);
1886 #if defined(__i386__) || defined(__amd64__)
1887 		if (!IS_DMAR(ctx)) {
1888 			bus_addr = pmap_kextract((vm_offset_t)cl);
1889 		} else
1890 #endif
1891 		{
1892 			struct rxq_refill_cb_arg cb_arg;
1893 			iflib_rxq_t q;
1894 
1895 			cb_arg.error = 0;
1896 			q = fl->ifl_rxq;
1897 			MPASS(sd_map != NULL);
1898 			MPASS(sd_map[frag_idx] != NULL);
1899 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1900 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1901 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1902 					BUS_DMASYNC_PREREAD);
1903 
1904 			if (err != 0 || cb_arg.error) {
1905 				/*
1906 				 * !zone_pack ?
1907 				 */
1908 				if (fl->ifl_zone == zone_pack)
1909 					uma_zfree(fl->ifl_zone, cl);
1910 				m_free(m);
1911 				n = 0;
1912 				goto done;
1913 			}
1914 			bus_addr = cb_arg.seg.ds_addr;
1915 		}
1916                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1917 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1918 
1919 		MPASS(sd_m[frag_idx] == NULL);
1920 		sd_cl[frag_idx] = cl;
1921 		sd_m[frag_idx] = m;
1922 		fl->ifl_rxd_idxs[i] = frag_idx;
1923 		fl->ifl_bus_addrs[i] = bus_addr;
1924 		fl->ifl_vm_addrs[i] = cl;
1925 		credits++;
1926 		i++;
1927 		MPASS(credits <= fl->ifl_size);
1928 		if (++idx == fl->ifl_size) {
1929 			fl->ifl_gen = 1;
1930 			idx = 0;
1931 		}
1932 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1933 			iru.iru_pidx = pidx;
1934 			iru.iru_count = i;
1935 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1936 			i = 0;
1937 			pidx = idx;
1938 			fl->ifl_pidx = idx;
1939 			fl->ifl_credits = credits;
1940 		}
1941 
1942 	}
1943 done:
1944 	if (i) {
1945 		iru.iru_pidx = pidx;
1946 		iru.iru_count = i;
1947 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1948 		fl->ifl_pidx = idx;
1949 		fl->ifl_credits = credits;
1950 	}
1951 	DBG_COUNTER_INC(rxd_flush);
1952 	if (fl->ifl_pidx == 0)
1953 		pidx = fl->ifl_size - 1;
1954 	else
1955 		pidx = fl->ifl_pidx - 1;
1956 
1957 	if (sd_map)
1958 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1959 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1960 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1961 	fl->ifl_fragidx = frag_idx;
1962 }
1963 
1964 static __inline void
1965 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1966 {
1967 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1968 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1969 #ifdef INVARIANTS
1970 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1971 #endif
1972 
1973 	MPASS(fl->ifl_credits <= fl->ifl_size);
1974 	MPASS(reclaimable == delta);
1975 
1976 	if (reclaimable > 0)
1977 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1978 }
1979 
1980 static void
1981 iflib_fl_bufs_free(iflib_fl_t fl)
1982 {
1983 	iflib_dma_info_t idi = fl->ifl_ifdi;
1984 	uint32_t i;
1985 
1986 	for (i = 0; i < fl->ifl_size; i++) {
1987 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1988 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1989 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1990 
1991 		if (*sd_flags & RX_SW_DESC_INUSE) {
1992 			if (fl->ifl_sds.ifsd_map != NULL) {
1993 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1994 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1995 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
1996 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1997 			}
1998 			if (*sd_m != NULL) {
1999 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2000 				uma_zfree(zone_mbuf, *sd_m);
2001 			}
2002 			if (*sd_cl != NULL)
2003 				uma_zfree(fl->ifl_zone, *sd_cl);
2004 			*sd_flags = 0;
2005 		} else {
2006 			MPASS(*sd_cl == NULL);
2007 			MPASS(*sd_m == NULL);
2008 		}
2009 #if MEMORY_LOGGING
2010 		fl->ifl_m_dequeued++;
2011 		fl->ifl_cl_dequeued++;
2012 #endif
2013 		*sd_cl = NULL;
2014 		*sd_m = NULL;
2015 	}
2016 #ifdef INVARIANTS
2017 	for (i = 0; i < fl->ifl_size; i++) {
2018 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2019 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2020 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2021 	}
2022 #endif
2023 	/*
2024 	 * Reset free list values
2025 	 */
2026 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2027 	bzero(idi->idi_vaddr, idi->idi_size);
2028 }
2029 
2030 /*********************************************************************
2031  *
2032  *  Initialize a receive ring and its buffers.
2033  *
2034  **********************************************************************/
2035 static int
2036 iflib_fl_setup(iflib_fl_t fl)
2037 {
2038 	iflib_rxq_t rxq = fl->ifl_rxq;
2039 	if_ctx_t ctx = rxq->ifr_ctx;
2040 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2041 
2042 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2043 	/*
2044 	** Free current RX buffer structs and their mbufs
2045 	*/
2046 	iflib_fl_bufs_free(fl);
2047 	/* Now replenish the mbufs */
2048 	MPASS(fl->ifl_credits == 0);
2049 	/*
2050 	 * XXX don't set the max_frame_size to larger
2051 	 * than the hardware can handle
2052 	 */
2053 	if (sctx->isc_max_frame_size <= 2048)
2054 		fl->ifl_buf_size = MCLBYTES;
2055 #ifndef CONTIGMALLOC_WORKS
2056 	else
2057 		fl->ifl_buf_size = MJUMPAGESIZE;
2058 #else
2059 	else if (sctx->isc_max_frame_size <= 4096)
2060 		fl->ifl_buf_size = MJUMPAGESIZE;
2061 	else if (sctx->isc_max_frame_size <= 9216)
2062 		fl->ifl_buf_size = MJUM9BYTES;
2063 	else
2064 		fl->ifl_buf_size = MJUM16BYTES;
2065 #endif
2066 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2067 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2068 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2069 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2070 
2071 
2072 	/* avoid pre-allocating zillions of clusters to an idle card
2073 	 * potentially speeding up attach
2074 	 */
2075 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2076 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2077 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2078 		return (ENOBUFS);
2079 	/*
2080 	 * handle failure
2081 	 */
2082 	MPASS(rxq != NULL);
2083 	MPASS(fl->ifl_ifdi != NULL);
2084 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2085 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2086 	return (0);
2087 }
2088 
2089 /*********************************************************************
2090  *
2091  *  Free receive ring data structures
2092  *
2093  **********************************************************************/
2094 static void
2095 iflib_rx_sds_free(iflib_rxq_t rxq)
2096 {
2097 	iflib_fl_t fl;
2098 	int i;
2099 
2100 	if (rxq->ifr_fl != NULL) {
2101 		for (i = 0; i < rxq->ifr_nfl; i++) {
2102 			fl = &rxq->ifr_fl[i];
2103 			if (fl->ifl_desc_tag != NULL) {
2104 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2105 				fl->ifl_desc_tag = NULL;
2106 			}
2107 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2108 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2109 			/* XXX destroy maps first */
2110 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2111 			fl->ifl_sds.ifsd_m = NULL;
2112 			fl->ifl_sds.ifsd_cl = NULL;
2113 			fl->ifl_sds.ifsd_map = NULL;
2114 		}
2115 		free(rxq->ifr_fl, M_IFLIB);
2116 		rxq->ifr_fl = NULL;
2117 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2118 	}
2119 }
2120 
2121 /*
2122  * MI independent logic
2123  *
2124  */
2125 static void
2126 iflib_timer(void *arg)
2127 {
2128 	iflib_txq_t txq = arg;
2129 	if_ctx_t ctx = txq->ift_ctx;
2130 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2131 
2132 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2133 		return;
2134 	/*
2135 	** Check on the state of the TX queue(s), this
2136 	** can be done without the lock because its RO
2137 	** and the HUNG state will be static if set.
2138 	*/
2139 	IFDI_TIMER(ctx, txq->ift_id);
2140 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2141 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2142 	     (sctx->isc_pause_frames == 0)))
2143 		goto hung;
2144 
2145 	if (ifmp_ring_is_stalled(txq->ift_br))
2146 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2147 	txq->ift_cleaned_prev = txq->ift_cleaned;
2148 	/* handle any laggards */
2149 	if (txq->ift_db_pending)
2150 		GROUPTASK_ENQUEUE(&txq->ift_task);
2151 
2152 	sctx->isc_pause_frames = 0;
2153 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2154 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2155 	return;
2156  hung:
2157 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2158 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2159 	STATE_LOCK(ctx);
2160 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2161 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2162 	iflib_admin_intr_deferred(ctx);
2163 	STATE_UNLOCK(ctx);
2164 }
2165 
2166 static void
2167 iflib_init_locked(if_ctx_t ctx)
2168 {
2169 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2170 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2171 	if_t ifp = ctx->ifc_ifp;
2172 	iflib_fl_t fl;
2173 	iflib_txq_t txq;
2174 	iflib_rxq_t rxq;
2175 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2176 
2177 
2178 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2179 	IFDI_INTR_DISABLE(ctx);
2180 
2181 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2182 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2183 	/* Set hardware offload abilities */
2184 	if_clearhwassist(ifp);
2185 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2186 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2187 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2188 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2189 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2190 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2191 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2192 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2193 
2194 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2195 		CALLOUT_LOCK(txq);
2196 		callout_stop(&txq->ift_timer);
2197 		CALLOUT_UNLOCK(txq);
2198 		iflib_netmap_txq_init(ctx, txq);
2199 	}
2200 #ifdef INVARIANTS
2201 	i = if_getdrvflags(ifp);
2202 #endif
2203 	IFDI_INIT(ctx);
2204 	MPASS(if_getdrvflags(ifp) == i);
2205 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2206 		/* XXX this should really be done on a per-queue basis */
2207 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2208 			MPASS(rxq->ifr_id == i);
2209 			iflib_netmap_rxq_init(ctx, rxq);
2210 			continue;
2211 		}
2212 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2213 			if (iflib_fl_setup(fl)) {
2214 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2215 				goto done;
2216 			}
2217 		}
2218 	}
2219 	done:
2220 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2221 	IFDI_INTR_ENABLE(ctx);
2222 	txq = ctx->ifc_txqs;
2223 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2224 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2225 			txq->ift_timer.c_cpu);
2226 }
2227 
2228 static int
2229 iflib_media_change(if_t ifp)
2230 {
2231 	if_ctx_t ctx = if_getsoftc(ifp);
2232 	int err;
2233 
2234 	CTX_LOCK(ctx);
2235 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2236 		iflib_init_locked(ctx);
2237 	CTX_UNLOCK(ctx);
2238 	return (err);
2239 }
2240 
2241 static void
2242 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2243 {
2244 	if_ctx_t ctx = if_getsoftc(ifp);
2245 
2246 	CTX_LOCK(ctx);
2247 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2248 	IFDI_MEDIA_STATUS(ctx, ifmr);
2249 	CTX_UNLOCK(ctx);
2250 }
2251 
2252 static void
2253 iflib_stop(if_ctx_t ctx)
2254 {
2255 	iflib_txq_t txq = ctx->ifc_txqs;
2256 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2257 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2258 	iflib_dma_info_t di;
2259 	iflib_fl_t fl;
2260 	int i, j;
2261 
2262 	/* Tell the stack that the interface is no longer active */
2263 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2264 
2265 	IFDI_INTR_DISABLE(ctx);
2266 	DELAY(1000);
2267 	IFDI_STOP(ctx);
2268 	DELAY(1000);
2269 
2270 	iflib_debug_reset();
2271 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2272 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2273 		/* make sure all transmitters have completed before proceeding XXX */
2274 
2275 		CALLOUT_LOCK(txq);
2276 		callout_stop(&txq->ift_timer);
2277 		CALLOUT_UNLOCK(txq);
2278 
2279 		/* clean any enqueued buffers */
2280 		iflib_ifmp_purge(txq);
2281 		/* Free any existing tx buffers. */
2282 		for (j = 0; j < txq->ift_size; j++) {
2283 			iflib_txsd_free(ctx, txq, j);
2284 		}
2285 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2286 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2287 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2288 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2289 		txq->ift_pullups = 0;
2290 		ifmp_ring_reset_stats(txq->ift_br);
2291 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2292 			bzero((void *)di->idi_vaddr, di->idi_size);
2293 	}
2294 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2295 		/* make sure all transmitters have completed before proceeding XXX */
2296 
2297 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2298 			bzero((void *)di->idi_vaddr, di->idi_size);
2299 		/* also resets the free lists pidx/cidx */
2300 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2301 			iflib_fl_bufs_free(fl);
2302 	}
2303 }
2304 
2305 static inline caddr_t
2306 calc_next_rxd(iflib_fl_t fl, int cidx)
2307 {
2308 	qidx_t size;
2309 	int nrxd;
2310 	caddr_t start, end, cur, next;
2311 
2312 	nrxd = fl->ifl_size;
2313 	size = fl->ifl_rxd_size;
2314 	start = fl->ifl_ifdi->idi_vaddr;
2315 
2316 	if (__predict_false(size == 0))
2317 		return (start);
2318 	cur = start + size*cidx;
2319 	end = start + size*nrxd;
2320 	next = CACHE_PTR_NEXT(cur);
2321 	return (next < end ? next : start);
2322 }
2323 
2324 static inline void
2325 prefetch_pkts(iflib_fl_t fl, int cidx)
2326 {
2327 	int nextptr;
2328 	int nrxd = fl->ifl_size;
2329 	caddr_t next_rxd;
2330 
2331 
2332 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2333 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2334 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2335 	next_rxd = calc_next_rxd(fl, cidx);
2336 	prefetch(next_rxd);
2337 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2338 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2339 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2340 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2341 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2342 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2343 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2344 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2345 }
2346 
2347 static void
2348 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2349 {
2350 	int flid, cidx;
2351 	bus_dmamap_t map;
2352 	iflib_fl_t fl;
2353 	iflib_dma_info_t di;
2354 	int next;
2355 
2356 	map = NULL;
2357 	flid = irf->irf_flid;
2358 	cidx = irf->irf_idx;
2359 	fl = &rxq->ifr_fl[flid];
2360 	sd->ifsd_fl = fl;
2361 	sd->ifsd_cidx = cidx;
2362 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2363 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2364 	fl->ifl_credits--;
2365 #if MEMORY_LOGGING
2366 	fl->ifl_m_dequeued++;
2367 #endif
2368 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2369 		prefetch_pkts(fl, cidx);
2370 	if (fl->ifl_sds.ifsd_map != NULL) {
2371 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2372 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2373 		map = fl->ifl_sds.ifsd_map[cidx];
2374 		di = fl->ifl_ifdi;
2375 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2376 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2377 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2378 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2379 
2380 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2381 		MPASS(fl->ifl_cidx == cidx);
2382 		if (unload)
2383 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2384 	}
2385 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2386 	if (__predict_false(fl->ifl_cidx == 0))
2387 		fl->ifl_gen = 0;
2388 	if (map != NULL)
2389 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2390 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2391         bit_clear(fl->ifl_rx_bitmap, cidx);
2392 }
2393 
2394 static struct mbuf *
2395 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2396 {
2397 	int i, padlen , flags;
2398 	struct mbuf *m, *mh, *mt;
2399 	caddr_t cl;
2400 
2401 	i = 0;
2402 	mh = NULL;
2403 	do {
2404 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2405 
2406 		MPASS(*sd->ifsd_cl != NULL);
2407 		MPASS(*sd->ifsd_m != NULL);
2408 
2409 		/* Don't include zero-length frags */
2410 		if (ri->iri_frags[i].irf_len == 0) {
2411 			/* XXX we can save the cluster here, but not the mbuf */
2412 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2413 			m_free(*sd->ifsd_m);
2414 			*sd->ifsd_m = NULL;
2415 			continue;
2416 		}
2417 		m = *sd->ifsd_m;
2418 		*sd->ifsd_m = NULL;
2419 		if (mh == NULL) {
2420 			flags = M_PKTHDR|M_EXT;
2421 			mh = mt = m;
2422 			padlen = ri->iri_pad;
2423 		} else {
2424 			flags = M_EXT;
2425 			mt->m_next = m;
2426 			mt = m;
2427 			/* assuming padding is only on the first fragment */
2428 			padlen = 0;
2429 		}
2430 		cl = *sd->ifsd_cl;
2431 		*sd->ifsd_cl = NULL;
2432 
2433 		/* Can these two be made one ? */
2434 		m_init(m, M_NOWAIT, MT_DATA, flags);
2435 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2436 		/*
2437 		 * These must follow m_init and m_cljset
2438 		 */
2439 		m->m_data += padlen;
2440 		ri->iri_len -= padlen;
2441 		m->m_len = ri->iri_frags[i].irf_len;
2442 	} while (++i < ri->iri_nfrags);
2443 
2444 	return (mh);
2445 }
2446 
2447 /*
2448  * Process one software descriptor
2449  */
2450 static struct mbuf *
2451 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2452 {
2453 	struct if_rxsd sd;
2454 	struct mbuf *m;
2455 
2456 	/* should I merge this back in now that the two paths are basically duplicated? */
2457 	if (ri->iri_nfrags == 1 &&
2458 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2459 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2460 		m = *sd.ifsd_m;
2461 		*sd.ifsd_m = NULL;
2462 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2463 #ifndef __NO_STRICT_ALIGNMENT
2464 		if (!IP_ALIGNED(m))
2465 			m->m_data += 2;
2466 #endif
2467 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2468 		m->m_len = ri->iri_frags[0].irf_len;
2469        } else {
2470 		m = assemble_segments(rxq, ri, &sd);
2471 	}
2472 	m->m_pkthdr.len = ri->iri_len;
2473 	m->m_pkthdr.rcvif = ri->iri_ifp;
2474 	m->m_flags |= ri->iri_flags;
2475 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2476 	m->m_pkthdr.flowid = ri->iri_flowid;
2477 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2478 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2479 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2480 	return (m);
2481 }
2482 
2483 #if defined(INET6) || defined(INET)
2484 static void
2485 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2486 {
2487 	CURVNET_SET(lc->ifp->if_vnet);
2488 #if defined(INET6)
2489 	*v6 = VNET(ip6_forwarding);
2490 #endif
2491 #if defined(INET)
2492 	*v4 = VNET(ipforwarding);
2493 #endif
2494 	CURVNET_RESTORE();
2495 }
2496 
2497 /*
2498  * Returns true if it's possible this packet could be LROed.
2499  * if it returns false, it is guaranteed that tcp_lro_rx()
2500  * would not return zero.
2501  */
2502 static bool
2503 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2504 {
2505 	struct ether_header *eh;
2506 	uint16_t eh_type;
2507 
2508 	eh = mtod(m, struct ether_header *);
2509 	eh_type = ntohs(eh->ether_type);
2510 	switch (eh_type) {
2511 #if defined(INET6)
2512 		case ETHERTYPE_IPV6:
2513 			return !v6_forwarding;
2514 #endif
2515 #if defined (INET)
2516 		case ETHERTYPE_IP:
2517 			return !v4_forwarding;
2518 #endif
2519 	}
2520 
2521 	return false;
2522 }
2523 #else
2524 static void
2525 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2526 {
2527 }
2528 #endif
2529 
2530 static bool
2531 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2532 {
2533 	if_ctx_t ctx = rxq->ifr_ctx;
2534 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2535 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2536 	int avail, i;
2537 	qidx_t *cidxp;
2538 	struct if_rxd_info ri;
2539 	int err, budget_left, rx_bytes, rx_pkts;
2540 	iflib_fl_t fl;
2541 	struct ifnet *ifp;
2542 	int lro_enabled;
2543 	bool lro_possible = false;
2544 	bool v4_forwarding, v6_forwarding;
2545 
2546 	/*
2547 	 * XXX early demux data packets so that if_input processing only handles
2548 	 * acks in interrupt context
2549 	 */
2550 	struct mbuf *m, *mh, *mt, *mf;
2551 
2552 	ifp = ctx->ifc_ifp;
2553 	mh = mt = NULL;
2554 	MPASS(budget > 0);
2555 	rx_pkts	= rx_bytes = 0;
2556 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2557 		cidxp = &rxq->ifr_cq_cidx;
2558 	else
2559 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2560 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2561 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2562 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2563 		DBG_COUNTER_INC(rx_unavail);
2564 		return (false);
2565 	}
2566 
2567 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2568 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2569 			DBG_COUNTER_INC(rx_ctx_inactive);
2570 			break;
2571 		}
2572 		/*
2573 		 * Reset client set fields to their default values
2574 		 */
2575 		rxd_info_zero(&ri);
2576 		ri.iri_qsidx = rxq->ifr_id;
2577 		ri.iri_cidx = *cidxp;
2578 		ri.iri_ifp = ifp;
2579 		ri.iri_frags = rxq->ifr_frags;
2580 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2581 
2582 		if (err)
2583 			goto err;
2584 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2585 			*cidxp = ri.iri_cidx;
2586 			/* Update our consumer index */
2587 			/* XXX NB: shurd - check if this is still safe */
2588 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2589 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2590 				rxq->ifr_cq_gen = 0;
2591 			}
2592 			/* was this only a completion queue message? */
2593 			if (__predict_false(ri.iri_nfrags == 0))
2594 				continue;
2595 		}
2596 		MPASS(ri.iri_nfrags != 0);
2597 		MPASS(ri.iri_len != 0);
2598 
2599 		/* will advance the cidx on the corresponding free lists */
2600 		m = iflib_rxd_pkt_get(rxq, &ri);
2601 		if (avail == 0 && budget_left)
2602 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2603 
2604 		if (__predict_false(m == NULL)) {
2605 			DBG_COUNTER_INC(rx_mbuf_null);
2606 			continue;
2607 		}
2608 		/* imm_pkt: -- cxgb */
2609 		if (mh == NULL)
2610 			mh = mt = m;
2611 		else {
2612 			mt->m_nextpkt = m;
2613 			mt = m;
2614 		}
2615 	}
2616 	/* make sure that we can refill faster than drain */
2617 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2618 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2619 
2620 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2621 	if (lro_enabled)
2622 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2623 	mt = mf = NULL;
2624 	while (mh != NULL) {
2625 		m = mh;
2626 		mh = mh->m_nextpkt;
2627 		m->m_nextpkt = NULL;
2628 #ifndef __NO_STRICT_ALIGNMENT
2629 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2630 			continue;
2631 #endif
2632 		rx_bytes += m->m_pkthdr.len;
2633 		rx_pkts++;
2634 #if defined(INET6) || defined(INET)
2635 		if (lro_enabled) {
2636 			if (!lro_possible) {
2637 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2638 				if (lro_possible && mf != NULL) {
2639 					ifp->if_input(ifp, mf);
2640 					DBG_COUNTER_INC(rx_if_input);
2641 					mt = mf = NULL;
2642 				}
2643 			}
2644 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2645 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2646 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2647 					continue;
2648 			}
2649 		}
2650 #endif
2651 		if (lro_possible) {
2652 			ifp->if_input(ifp, m);
2653 			DBG_COUNTER_INC(rx_if_input);
2654 			continue;
2655 		}
2656 
2657 		if (mf == NULL)
2658 			mf = m;
2659 		if (mt != NULL)
2660 			mt->m_nextpkt = m;
2661 		mt = m;
2662 	}
2663 	if (mf != NULL) {
2664 		ifp->if_input(ifp, mf);
2665 		DBG_COUNTER_INC(rx_if_input);
2666 	}
2667 
2668 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2669 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2670 
2671 	/*
2672 	 * Flush any outstanding LRO work
2673 	 */
2674 #if defined(INET6) || defined(INET)
2675 	tcp_lro_flush_all(&rxq->ifr_lc);
2676 #endif
2677 	if (avail)
2678 		return true;
2679 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2680 err:
2681 	STATE_LOCK(ctx);
2682 	ctx->ifc_flags |= IFC_DO_RESET;
2683 	iflib_admin_intr_deferred(ctx);
2684 	STATE_UNLOCK(ctx);
2685 	return (false);
2686 }
2687 
2688 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2689 static inline qidx_t
2690 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2691 {
2692 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2693 	qidx_t minthresh = txq->ift_size / 8;
2694 	if (in_use > 4*minthresh)
2695 		return (notify_count);
2696 	if (in_use > 2*minthresh)
2697 		return (notify_count >> 1);
2698 	if (in_use > minthresh)
2699 		return (notify_count >> 3);
2700 	return (0);
2701 }
2702 
2703 static inline qidx_t
2704 txq_max_rs_deferred(iflib_txq_t txq)
2705 {
2706 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2707 	qidx_t minthresh = txq->ift_size / 8;
2708 	if (txq->ift_in_use > 4*minthresh)
2709 		return (notify_count);
2710 	if (txq->ift_in_use > 2*minthresh)
2711 		return (notify_count >> 1);
2712 	if (txq->ift_in_use > minthresh)
2713 		return (notify_count >> 2);
2714 	return (2);
2715 }
2716 
2717 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2718 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2719 
2720 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2721 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2722 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2723 
2724 /* forward compatibility for cxgb */
2725 #define FIRST_QSET(ctx) 0
2726 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2727 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2728 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2729 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2730 
2731 /* XXX we should be setting this to something other than zero */
2732 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2733 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2734 
2735 static inline bool
2736 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2737 {
2738 	qidx_t dbval, max;
2739 	bool rang;
2740 
2741 	rang = false;
2742 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2743 	if (ring || txq->ift_db_pending >= max) {
2744 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2745 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2746 		txq->ift_db_pending = txq->ift_npending = 0;
2747 		rang = true;
2748 	}
2749 	return (rang);
2750 }
2751 
2752 #ifdef PKT_DEBUG
2753 static void
2754 print_pkt(if_pkt_info_t pi)
2755 {
2756 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2757 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2758 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2759 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2760 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2761 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2762 }
2763 #endif
2764 
2765 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2766 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2767 
2768 static int
2769 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2770 {
2771 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2772 	struct ether_vlan_header *eh;
2773 	struct mbuf *m, *n;
2774 
2775 	n = m = *mp;
2776 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2777 	    M_WRITABLE(m) == 0) {
2778 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2779 			return (ENOMEM);
2780 		} else {
2781 			m_freem(*mp);
2782 			n = *mp = m;
2783 		}
2784 	}
2785 
2786 	/*
2787 	 * Determine where frame payload starts.
2788 	 * Jump over vlan headers if already present,
2789 	 * helpful for QinQ too.
2790 	 */
2791 	if (__predict_false(m->m_len < sizeof(*eh))) {
2792 		txq->ift_pullups++;
2793 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2794 			return (ENOMEM);
2795 	}
2796 	eh = mtod(m, struct ether_vlan_header *);
2797 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2798 		pi->ipi_etype = ntohs(eh->evl_proto);
2799 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2800 	} else {
2801 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2802 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2803 	}
2804 
2805 	switch (pi->ipi_etype) {
2806 #ifdef INET
2807 	case ETHERTYPE_IP:
2808 	{
2809 		struct ip *ip = NULL;
2810 		struct tcphdr *th = NULL;
2811 		int minthlen;
2812 
2813 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2814 		if (__predict_false(m->m_len < minthlen)) {
2815 			/*
2816 			 * if this code bloat is causing too much of a hit
2817 			 * move it to a separate function and mark it noinline
2818 			 */
2819 			if (m->m_len == pi->ipi_ehdrlen) {
2820 				n = m->m_next;
2821 				MPASS(n);
2822 				if (n->m_len >= sizeof(*ip))  {
2823 					ip = (struct ip *)n->m_data;
2824 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2825 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2826 				} else {
2827 					txq->ift_pullups++;
2828 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2829 						return (ENOMEM);
2830 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2831 				}
2832 			} else {
2833 				txq->ift_pullups++;
2834 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2835 					return (ENOMEM);
2836 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2837 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2838 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2839 			}
2840 		} else {
2841 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2842 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2843 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2844 		}
2845 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2846 		pi->ipi_ipproto = ip->ip_p;
2847 		pi->ipi_flags |= IPI_TX_IPV4;
2848 
2849 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2850                        ip->ip_sum = 0;
2851 
2852 		if (IS_TSO4(pi)) {
2853 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2854 				if (__predict_false(th == NULL)) {
2855 					txq->ift_pullups++;
2856 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2857 						return (ENOMEM);
2858 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2859 				}
2860 				pi->ipi_tcp_hflags = th->th_flags;
2861 				pi->ipi_tcp_hlen = th->th_off << 2;
2862 				pi->ipi_tcp_seq = th->th_seq;
2863 			}
2864 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2865 				return (ENXIO);
2866 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2867 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2868 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2869 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2870 				ip->ip_sum = 0;
2871 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2872 			}
2873 		}
2874 		break;
2875 	}
2876 #endif
2877 #ifdef INET6
2878 	case ETHERTYPE_IPV6:
2879 	{
2880 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2881 		struct tcphdr *th;
2882 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2883 
2884 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2885 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2886 				return (ENOMEM);
2887 		}
2888 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2889 
2890 		/* XXX-BZ this will go badly in case of ext hdrs. */
2891 		pi->ipi_ipproto = ip6->ip6_nxt;
2892 		pi->ipi_flags |= IPI_TX_IPV6;
2893 
2894 		if (IS_TSO6(pi)) {
2895 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2896 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2897 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2898 						return (ENOMEM);
2899 				}
2900 				pi->ipi_tcp_hflags = th->th_flags;
2901 				pi->ipi_tcp_hlen = th->th_off << 2;
2902 			}
2903 
2904 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2905 				return (ENXIO);
2906 			/*
2907 			 * The corresponding flag is set by the stack in the IPv4
2908 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2909 			 * So, set it here because the rest of the flow requires it.
2910 			 */
2911 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2912 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2913 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2914 		}
2915 		break;
2916 	}
2917 #endif
2918 	default:
2919 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2920 		pi->ipi_ip_hlen = 0;
2921 		break;
2922 	}
2923 	*mp = m;
2924 
2925 	return (0);
2926 }
2927 
2928 static  __noinline  struct mbuf *
2929 collapse_pkthdr(struct mbuf *m0)
2930 {
2931 	struct mbuf *m, *m_next, *tmp;
2932 
2933 	m = m0;
2934 	m_next = m->m_next;
2935 	while (m_next != NULL && m_next->m_len == 0) {
2936 		m = m_next;
2937 		m->m_next = NULL;
2938 		m_free(m);
2939 		m_next = m_next->m_next;
2940 	}
2941 	m = m0;
2942 	m->m_next = m_next;
2943 	if ((m_next->m_flags & M_EXT) == 0) {
2944 		m = m_defrag(m, M_NOWAIT);
2945 	} else {
2946 		tmp = m_next->m_next;
2947 		memcpy(m_next, m, MPKTHSIZE);
2948 		m = m_next;
2949 		m->m_next = tmp;
2950 	}
2951 	return (m);
2952 }
2953 
2954 /*
2955  * If dodgy hardware rejects the scatter gather chain we've handed it
2956  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2957  * m_defrag'd mbufs
2958  */
2959 static __noinline struct mbuf *
2960 iflib_remove_mbuf(iflib_txq_t txq)
2961 {
2962 	int ntxd, i, pidx;
2963 	struct mbuf *m, *mh, **ifsd_m;
2964 
2965 	pidx = txq->ift_pidx;
2966 	ifsd_m = txq->ift_sds.ifsd_m;
2967 	ntxd = txq->ift_size;
2968 	mh = m = ifsd_m[pidx];
2969 	ifsd_m[pidx] = NULL;
2970 #if MEMORY_LOGGING
2971 	txq->ift_dequeued++;
2972 #endif
2973 	i = 1;
2974 
2975 	while (m) {
2976 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2977 #if MEMORY_LOGGING
2978 		txq->ift_dequeued++;
2979 #endif
2980 		m = m->m_next;
2981 		i++;
2982 	}
2983 	return (mh);
2984 }
2985 
2986 static int
2987 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2988 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2989 			  int max_segs, int flags)
2990 {
2991 	if_ctx_t ctx;
2992 	if_shared_ctx_t		sctx;
2993 	if_softc_ctx_t		scctx;
2994 	int i, next, pidx, err, ntxd, count;
2995 	struct mbuf *m, *tmp, **ifsd_m;
2996 
2997 	m = *m0;
2998 
2999 	/*
3000 	 * Please don't ever do this
3001 	 */
3002 	if (__predict_false(m->m_len == 0))
3003 		*m0 = m = collapse_pkthdr(m);
3004 
3005 	ctx = txq->ift_ctx;
3006 	sctx = ctx->ifc_sctx;
3007 	scctx = &ctx->ifc_softc_ctx;
3008 	ifsd_m = txq->ift_sds.ifsd_m;
3009 	ntxd = txq->ift_size;
3010 	pidx = txq->ift_pidx;
3011 	if (map != NULL) {
3012 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3013 
3014 		err = bus_dmamap_load_mbuf_sg(tag, map,
3015 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3016 		if (err)
3017 			return (err);
3018 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3019 		count = 0;
3020 		m = *m0;
3021 		do {
3022 			if (__predict_false(m->m_len <= 0)) {
3023 				tmp = m;
3024 				m = m->m_next;
3025 				tmp->m_next = NULL;
3026 				m_free(tmp);
3027 				continue;
3028 			}
3029 			m = m->m_next;
3030 			count++;
3031 		} while (m != NULL);
3032 		if (count > *nsegs) {
3033 			ifsd_m[pidx] = *m0;
3034 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3035 			return (0);
3036 		}
3037 		m = *m0;
3038 		count = 0;
3039 		do {
3040 			next = (pidx + count) & (ntxd-1);
3041 			MPASS(ifsd_m[next] == NULL);
3042 			ifsd_m[next] = m;
3043 			count++;
3044 			tmp = m;
3045 			m = m->m_next;
3046 		} while (m != NULL);
3047 	} else {
3048 		int buflen, sgsize, maxsegsz, max_sgsize;
3049 		vm_offset_t vaddr;
3050 		vm_paddr_t curaddr;
3051 
3052 		count = i = 0;
3053 		m = *m0;
3054 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3055 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3056 		else
3057 			maxsegsz = sctx->isc_tx_maxsegsize;
3058 
3059 		do {
3060 			if (__predict_false(m->m_len <= 0)) {
3061 				tmp = m;
3062 				m = m->m_next;
3063 				tmp->m_next = NULL;
3064 				m_free(tmp);
3065 				continue;
3066 			}
3067 			buflen = m->m_len;
3068 			vaddr = (vm_offset_t)m->m_data;
3069 			/*
3070 			 * see if we can't be smarter about physically
3071 			 * contiguous mappings
3072 			 */
3073 			next = (pidx + count) & (ntxd-1);
3074 			MPASS(ifsd_m[next] == NULL);
3075 #if MEMORY_LOGGING
3076 			txq->ift_enqueued++;
3077 #endif
3078 			ifsd_m[next] = m;
3079 			while (buflen > 0) {
3080 				if (i >= max_segs)
3081 					goto err;
3082 				max_sgsize = MIN(buflen, maxsegsz);
3083 				curaddr = pmap_kextract(vaddr);
3084 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3085 				sgsize = MIN(sgsize, max_sgsize);
3086 				segs[i].ds_addr = curaddr;
3087 				segs[i].ds_len = sgsize;
3088 				vaddr += sgsize;
3089 				buflen -= sgsize;
3090 				i++;
3091 			}
3092 			count++;
3093 			tmp = m;
3094 			m = m->m_next;
3095 		} while (m != NULL);
3096 		*nsegs = i;
3097 	}
3098 	return (0);
3099 err:
3100 	*m0 = iflib_remove_mbuf(txq);
3101 	return (EFBIG);
3102 }
3103 
3104 static inline caddr_t
3105 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3106 {
3107 	qidx_t size;
3108 	int ntxd;
3109 	caddr_t start, end, cur, next;
3110 
3111 	ntxd = txq->ift_size;
3112 	size = txq->ift_txd_size[qid];
3113 	start = txq->ift_ifdi[qid].idi_vaddr;
3114 
3115 	if (__predict_false(size == 0))
3116 		return (start);
3117 	cur = start + size*cidx;
3118 	end = start + size*ntxd;
3119 	next = CACHE_PTR_NEXT(cur);
3120 	return (next < end ? next : start);
3121 }
3122 
3123 /*
3124  * Pad an mbuf to ensure a minimum ethernet frame size.
3125  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3126  */
3127 static __noinline int
3128 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3129 {
3130 	/*
3131 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3132 	 * and ARP message is the smallest common payload I can think of
3133 	 */
3134 	static char pad[18];	/* just zeros */
3135 	int n;
3136 	struct mbuf *new_head;
3137 
3138 	if (!M_WRITABLE(*m_head)) {
3139 		new_head = m_dup(*m_head, M_NOWAIT);
3140 		if (new_head == NULL) {
3141 			m_freem(*m_head);
3142 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3143 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3144 			return ENOMEM;
3145 		}
3146 		m_freem(*m_head);
3147 		*m_head = new_head;
3148 	}
3149 
3150 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3151 	     n > 0; n -= sizeof(pad))
3152 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3153 			break;
3154 
3155 	if (n > 0) {
3156 		m_freem(*m_head);
3157 		device_printf(dev, "cannot pad short frame\n");
3158 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3159 		return (ENOBUFS);
3160 	}
3161 
3162 	return 0;
3163 }
3164 
3165 static int
3166 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3167 {
3168 	if_ctx_t		ctx;
3169 	if_shared_ctx_t		sctx;
3170 	if_softc_ctx_t		scctx;
3171 	bus_dma_segment_t	*segs;
3172 	struct mbuf		*m_head;
3173 	void			*next_txd;
3174 	bus_dmamap_t		map;
3175 	struct if_pkt_info	pi;
3176 	int remap = 0;
3177 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3178 	bus_dma_tag_t desc_tag;
3179 
3180 	segs = txq->ift_segs;
3181 	ctx = txq->ift_ctx;
3182 	sctx = ctx->ifc_sctx;
3183 	scctx = &ctx->ifc_softc_ctx;
3184 	segs = txq->ift_segs;
3185 	ntxd = txq->ift_size;
3186 	m_head = *m_headp;
3187 	map = NULL;
3188 
3189 	/*
3190 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3191 	 */
3192 	cidx = txq->ift_cidx;
3193 	pidx = txq->ift_pidx;
3194 	if (ctx->ifc_flags & IFC_PREFETCH) {
3195 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3196 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3197 			next_txd = calc_next_txd(txq, cidx, 0);
3198 			prefetch(next_txd);
3199 		}
3200 
3201 		/* prefetch the next cache line of mbuf pointers and flags */
3202 		prefetch(&txq->ift_sds.ifsd_m[next]);
3203 		if (txq->ift_sds.ifsd_map != NULL) {
3204 			prefetch(&txq->ift_sds.ifsd_map[next]);
3205 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3206 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3207 		}
3208 	} else if (txq->ift_sds.ifsd_map != NULL)
3209 		map = txq->ift_sds.ifsd_map[pidx];
3210 
3211 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3212 		desc_tag = txq->ift_tso_desc_tag;
3213 		max_segs = scctx->isc_tx_tso_segments_max;
3214 	} else {
3215 		desc_tag = txq->ift_desc_tag;
3216 		max_segs = scctx->isc_tx_nsegments;
3217 	}
3218 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3219 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3220 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3221 		if (err)
3222 			return err;
3223 	}
3224 	m_head = *m_headp;
3225 
3226 	pkt_info_zero(&pi);
3227 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3228 	pi.ipi_pidx = pidx;
3229 	pi.ipi_qsidx = txq->ift_id;
3230 	pi.ipi_len = m_head->m_pkthdr.len;
3231 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3232 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3233 
3234 	/* deliberate bitwise OR to make one condition */
3235 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3236 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3237 			return (err);
3238 		m_head = *m_headp;
3239 	}
3240 
3241 retry:
3242 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3243 defrag:
3244 	if (__predict_false(err)) {
3245 		switch (err) {
3246 		case EFBIG:
3247 			/* try collapse once and defrag once */
3248 			if (remap == 0)
3249 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3250 			if (remap == 1)
3251 				m_head = m_defrag(*m_headp, M_NOWAIT);
3252 			remap++;
3253 			if (__predict_false(m_head == NULL))
3254 				goto defrag_failed;
3255 			txq->ift_mbuf_defrag++;
3256 			*m_headp = m_head;
3257 			goto retry;
3258 			break;
3259 		case ENOMEM:
3260 			txq->ift_no_tx_dma_setup++;
3261 			break;
3262 		default:
3263 			txq->ift_no_tx_dma_setup++;
3264 			m_freem(*m_headp);
3265 			DBG_COUNTER_INC(tx_frees);
3266 			*m_headp = NULL;
3267 			break;
3268 		}
3269 		txq->ift_map_failed++;
3270 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3271 		return (err);
3272 	}
3273 
3274 	/*
3275 	 * XXX assumes a 1 to 1 relationship between segments and
3276 	 *        descriptors - this does not hold true on all drivers, e.g.
3277 	 *        cxgb
3278 	 */
3279 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3280 		txq->ift_no_desc_avail++;
3281 		if (map != NULL)
3282 			bus_dmamap_unload(desc_tag, map);
3283 		DBG_COUNTER_INC(encap_txq_avail_fail);
3284 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3285 			GROUPTASK_ENQUEUE(&txq->ift_task);
3286 		return (ENOBUFS);
3287 	}
3288 	/*
3289 	 * On Intel cards we can greatly reduce the number of TX interrupts
3290 	 * we see by only setting report status on every Nth descriptor.
3291 	 * However, this also means that the driver will need to keep track
3292 	 * of the descriptors that RS was set on to check them for the DD bit.
3293 	 */
3294 	txq->ift_rs_pending += nsegs + 1;
3295 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3296 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3297 		pi.ipi_flags |= IPI_TX_INTR;
3298 		txq->ift_rs_pending = 0;
3299 	}
3300 
3301 	pi.ipi_segs = segs;
3302 	pi.ipi_nsegs = nsegs;
3303 
3304 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3305 #ifdef PKT_DEBUG
3306 	print_pkt(&pi);
3307 #endif
3308 	if (map != NULL)
3309 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3310 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3311 		if (map != NULL)
3312 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3313 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3314 		DBG_COUNTER_INC(tx_encap);
3315 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3316 
3317 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3318 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3319 			ndesc += txq->ift_size;
3320 			txq->ift_gen = 1;
3321 		}
3322 		/*
3323 		 * drivers can need as many as
3324 		 * two sentinels
3325 		 */
3326 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3327 		MPASS(pi.ipi_new_pidx != pidx);
3328 		MPASS(ndesc > 0);
3329 		txq->ift_in_use += ndesc;
3330 
3331 		/*
3332 		 * We update the last software descriptor again here because there may
3333 		 * be a sentinel and/or there may be more mbufs than segments
3334 		 */
3335 		txq->ift_pidx = pi.ipi_new_pidx;
3336 		txq->ift_npending += pi.ipi_ndescs;
3337 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3338 		*m_headp = m_head = iflib_remove_mbuf(txq);
3339 		remap = 1;
3340 		txq->ift_txd_encap_efbig++;
3341 		goto defrag;
3342 	} else
3343 		DBG_COUNTER_INC(encap_txd_encap_fail);
3344 	return (err);
3345 
3346 defrag_failed:
3347 	txq->ift_mbuf_defrag_failed++;
3348 	txq->ift_map_failed++;
3349 	m_freem(*m_headp);
3350 	DBG_COUNTER_INC(tx_frees);
3351 	*m_headp = NULL;
3352 	return (ENOMEM);
3353 }
3354 
3355 static void
3356 iflib_tx_desc_free(iflib_txq_t txq, int n)
3357 {
3358 	int hasmap;
3359 	uint32_t qsize, cidx, mask, gen;
3360 	struct mbuf *m, **ifsd_m;
3361 	uint8_t *ifsd_flags;
3362 	bus_dmamap_t *ifsd_map;
3363 	bool do_prefetch;
3364 
3365 	cidx = txq->ift_cidx;
3366 	gen = txq->ift_gen;
3367 	qsize = txq->ift_size;
3368 	mask = qsize-1;
3369 	hasmap = txq->ift_sds.ifsd_map != NULL;
3370 	ifsd_flags = txq->ift_sds.ifsd_flags;
3371 	ifsd_m = txq->ift_sds.ifsd_m;
3372 	ifsd_map = txq->ift_sds.ifsd_map;
3373 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3374 
3375 	while (n--) {
3376 		if (do_prefetch) {
3377 			prefetch(ifsd_m[(cidx + 3) & mask]);
3378 			prefetch(ifsd_m[(cidx + 4) & mask]);
3379 		}
3380 		if (ifsd_m[cidx] != NULL) {
3381 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3382 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3383 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3384 				/*
3385 				 * does it matter if it's not the TSO tag? If so we'll
3386 				 * have to add the type to flags
3387 				 */
3388 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3389 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3390 			}
3391 			if ((m = ifsd_m[cidx]) != NULL) {
3392 				/* XXX we don't support any drivers that batch packets yet */
3393 				MPASS(m->m_nextpkt == NULL);
3394 				/* if the number of clusters exceeds the number of segments
3395 				 * there won't be space on the ring to save a pointer to each
3396 				 * cluster so we simply free the list here
3397 				 */
3398 				if (m->m_flags & M_TOOBIG) {
3399 					m_freem(m);
3400 				} else {
3401 					m_free(m);
3402 				}
3403 				ifsd_m[cidx] = NULL;
3404 #if MEMORY_LOGGING
3405 				txq->ift_dequeued++;
3406 #endif
3407 				DBG_COUNTER_INC(tx_frees);
3408 			}
3409 		}
3410 		if (__predict_false(++cidx == qsize)) {
3411 			cidx = 0;
3412 			gen = 0;
3413 		}
3414 	}
3415 	txq->ift_cidx = cidx;
3416 	txq->ift_gen = gen;
3417 }
3418 
3419 static __inline int
3420 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3421 {
3422 	int reclaim;
3423 	if_ctx_t ctx = txq->ift_ctx;
3424 
3425 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3426 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3427 
3428 	/*
3429 	 * Need a rate-limiting check so that this isn't called every time
3430 	 */
3431 	iflib_tx_credits_update(ctx, txq);
3432 	reclaim = DESC_RECLAIMABLE(txq);
3433 
3434 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3435 #ifdef INVARIANTS
3436 		if (iflib_verbose_debug) {
3437 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3438 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3439 			       reclaim, thresh);
3440 
3441 		}
3442 #endif
3443 		return (0);
3444 	}
3445 	iflib_tx_desc_free(txq, reclaim);
3446 	txq->ift_cleaned += reclaim;
3447 	txq->ift_in_use -= reclaim;
3448 
3449 	return (reclaim);
3450 }
3451 
3452 static struct mbuf **
3453 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3454 {
3455 	int next, size;
3456 	struct mbuf **items;
3457 
3458 	size = r->size;
3459 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3460 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3461 
3462 	prefetch(items[(cidx + offset) & (size-1)]);
3463 	if (remaining > 1) {
3464 		prefetch2cachelines(&items[next]);
3465 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3466 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3467 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3468 	}
3469 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3470 }
3471 
3472 static void
3473 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3474 {
3475 
3476 	ifmp_ring_check_drainage(txq->ift_br, budget);
3477 }
3478 
3479 static uint32_t
3480 iflib_txq_can_drain(struct ifmp_ring *r)
3481 {
3482 	iflib_txq_t txq = r->cookie;
3483 	if_ctx_t ctx = txq->ift_ctx;
3484 
3485 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3486 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3487 }
3488 
3489 static uint32_t
3490 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3491 {
3492 	iflib_txq_t txq = r->cookie;
3493 	if_ctx_t ctx = txq->ift_ctx;
3494 	struct ifnet *ifp = ctx->ifc_ifp;
3495 	struct mbuf **mp, *m;
3496 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3497 	int reclaimed, err, in_use_prev, desc_used;
3498 	bool do_prefetch, ring, rang;
3499 
3500 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3501 			    !LINK_ACTIVE(ctx))) {
3502 		DBG_COUNTER_INC(txq_drain_notready);
3503 		return (0);
3504 	}
3505 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3506 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3507 	avail = IDXDIFF(pidx, cidx, r->size);
3508 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3509 		DBG_COUNTER_INC(txq_drain_flushing);
3510 		for (i = 0; i < avail; i++) {
3511 			m_free(r->items[(cidx + i) & (r->size-1)]);
3512 			r->items[(cidx + i) & (r->size-1)] = NULL;
3513 		}
3514 		return (avail);
3515 	}
3516 
3517 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3518 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3519 		CALLOUT_LOCK(txq);
3520 		callout_stop(&txq->ift_timer);
3521 		CALLOUT_UNLOCK(txq);
3522 		DBG_COUNTER_INC(txq_drain_oactive);
3523 		return (0);
3524 	}
3525 	if (reclaimed)
3526 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3527 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3528 	count = MIN(avail, TX_BATCH_SIZE);
3529 #ifdef INVARIANTS
3530 	if (iflib_verbose_debug)
3531 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3532 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3533 #endif
3534 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3535 	avail = TXQ_AVAIL(txq);
3536 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3537 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3538 
3539 		mp = _ring_peek_one(r, cidx, i, rem);
3540 		MPASS(mp != NULL && *mp != NULL);
3541 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3542 			consumed++;
3543 			reclaimed++;
3544 			continue;
3545 		}
3546 		in_use_prev = txq->ift_in_use;
3547 		pidx_prev = txq->ift_pidx;
3548 		err = iflib_encap(txq, mp);
3549 		if (__predict_false(err)) {
3550 			DBG_COUNTER_INC(txq_drain_encapfail);
3551 			/* no room - bail out */
3552 			if (err == ENOBUFS)
3553 				break;
3554 			consumed++;
3555 			DBG_COUNTER_INC(txq_drain_encapfail);
3556 			/* we can't send this packet - skip it */
3557 			continue;
3558 		}
3559 		consumed++;
3560 		pkt_sent++;
3561 		m = *mp;
3562 		DBG_COUNTER_INC(tx_sent);
3563 		bytes_sent += m->m_pkthdr.len;
3564 		mcast_sent += !!(m->m_flags & M_MCAST);
3565 		avail = TXQ_AVAIL(txq);
3566 
3567 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3568 		desc_used += (txq->ift_in_use - in_use_prev);
3569 		ETHER_BPF_MTAP(ifp, m);
3570 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3571 			break;
3572 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3573 	}
3574 
3575 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3576 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3577 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3578 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3579 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3580 	if (mcast_sent)
3581 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3582 #ifdef INVARIANTS
3583 	if (iflib_verbose_debug)
3584 		printf("consumed=%d\n", consumed);
3585 #endif
3586 	return (consumed);
3587 }
3588 
3589 static uint32_t
3590 iflib_txq_drain_always(struct ifmp_ring *r)
3591 {
3592 	return (1);
3593 }
3594 
3595 static uint32_t
3596 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3597 {
3598 	int i, avail;
3599 	struct mbuf **mp;
3600 	iflib_txq_t txq;
3601 
3602 	txq = r->cookie;
3603 
3604 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3605 	CALLOUT_LOCK(txq);
3606 	callout_stop(&txq->ift_timer);
3607 	CALLOUT_UNLOCK(txq);
3608 
3609 	avail = IDXDIFF(pidx, cidx, r->size);
3610 	for (i = 0; i < avail; i++) {
3611 		mp = _ring_peek_one(r, cidx, i, avail - i);
3612 		if (__predict_false(*mp == (struct mbuf *)txq))
3613 			continue;
3614 		m_freem(*mp);
3615 	}
3616 	MPASS(ifmp_ring_is_stalled(r) == 0);
3617 	return (avail);
3618 }
3619 
3620 static void
3621 iflib_ifmp_purge(iflib_txq_t txq)
3622 {
3623 	struct ifmp_ring *r;
3624 
3625 	r = txq->ift_br;
3626 	r->drain = iflib_txq_drain_free;
3627 	r->can_drain = iflib_txq_drain_always;
3628 
3629 	ifmp_ring_check_drainage(r, r->size);
3630 
3631 	r->drain = iflib_txq_drain;
3632 	r->can_drain = iflib_txq_can_drain;
3633 }
3634 
3635 static void
3636 _task_fn_tx(void *context)
3637 {
3638 	iflib_txq_t txq = context;
3639 	if_ctx_t ctx = txq->ift_ctx;
3640 	struct ifnet *ifp = ctx->ifc_ifp;
3641 	int rc;
3642 
3643 #ifdef IFLIB_DIAGNOSTICS
3644 	txq->ift_cpu_exec_count[curcpu]++;
3645 #endif
3646 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3647 		return;
3648 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3649 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3650 			netmap_tx_irq(ifp, txq->ift_id);
3651 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3652 		return;
3653 	}
3654 	if (txq->ift_db_pending)
3655 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3656 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3657 	if (ctx->ifc_flags & IFC_LEGACY)
3658 		IFDI_INTR_ENABLE(ctx);
3659 	else {
3660 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3661 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3662 	}
3663 }
3664 
3665 static void
3666 _task_fn_rx(void *context)
3667 {
3668 	iflib_rxq_t rxq = context;
3669 	if_ctx_t ctx = rxq->ifr_ctx;
3670 	bool more;
3671 	int rc;
3672 	uint16_t budget;
3673 
3674 #ifdef IFLIB_DIAGNOSTICS
3675 	rxq->ifr_cpu_exec_count[curcpu]++;
3676 #endif
3677 	DBG_COUNTER_INC(task_fn_rxs);
3678 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3679 		return;
3680 	more = true;
3681 #ifdef DEV_NETMAP
3682 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3683 		u_int work = 0;
3684 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3685 			more = false;
3686 		}
3687 	}
3688 #endif
3689 	budget = ctx->ifc_sysctl_rx_budget;
3690 	if (budget == 0)
3691 		budget = 16;	/* XXX */
3692 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3693 		if (ctx->ifc_flags & IFC_LEGACY)
3694 			IFDI_INTR_ENABLE(ctx);
3695 		else {
3696 			DBG_COUNTER_INC(rx_intr_enables);
3697 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3698 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3699 		}
3700 	}
3701 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3702 		return;
3703 	if (more)
3704 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3705 }
3706 
3707 static void
3708 _task_fn_admin(void *context)
3709 {
3710 	if_ctx_t ctx = context;
3711 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3712 	iflib_txq_t txq;
3713 	int i;
3714 	bool oactive, running, do_reset, do_watchdog;
3715 
3716 	STATE_LOCK(ctx);
3717 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3718 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3719 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3720 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3721 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3722 	STATE_UNLOCK(ctx);
3723 
3724 	if (!running & !oactive)
3725 		return;
3726 
3727 	CTX_LOCK(ctx);
3728 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3729 		CALLOUT_LOCK(txq);
3730 		callout_stop(&txq->ift_timer);
3731 		CALLOUT_UNLOCK(txq);
3732 	}
3733 	if (do_watchdog) {
3734 		ctx->ifc_watchdog_events++;
3735 		IFDI_WATCHDOG_RESET(ctx);
3736 	}
3737 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3738 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3739 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3740 	IFDI_LINK_INTR_ENABLE(ctx);
3741 	if (do_reset)
3742 		iflib_if_init_locked(ctx);
3743 	CTX_UNLOCK(ctx);
3744 
3745 	if (LINK_ACTIVE(ctx) == 0)
3746 		return;
3747 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3748 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3749 }
3750 
3751 
3752 static void
3753 _task_fn_iov(void *context)
3754 {
3755 	if_ctx_t ctx = context;
3756 
3757 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3758 		return;
3759 
3760 	CTX_LOCK(ctx);
3761 	IFDI_VFLR_HANDLE(ctx);
3762 	CTX_UNLOCK(ctx);
3763 }
3764 
3765 static int
3766 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3767 {
3768 	int err;
3769 	if_int_delay_info_t info;
3770 	if_ctx_t ctx;
3771 
3772 	info = (if_int_delay_info_t)arg1;
3773 	ctx = info->iidi_ctx;
3774 	info->iidi_req = req;
3775 	info->iidi_oidp = oidp;
3776 	CTX_LOCK(ctx);
3777 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3778 	CTX_UNLOCK(ctx);
3779 	return (err);
3780 }
3781 
3782 /*********************************************************************
3783  *
3784  *  IFNET FUNCTIONS
3785  *
3786  **********************************************************************/
3787 
3788 static void
3789 iflib_if_init_locked(if_ctx_t ctx)
3790 {
3791 	iflib_stop(ctx);
3792 	iflib_init_locked(ctx);
3793 }
3794 
3795 
3796 static void
3797 iflib_if_init(void *arg)
3798 {
3799 	if_ctx_t ctx = arg;
3800 
3801 	CTX_LOCK(ctx);
3802 	iflib_if_init_locked(ctx);
3803 	CTX_UNLOCK(ctx);
3804 }
3805 
3806 static int
3807 iflib_if_transmit(if_t ifp, struct mbuf *m)
3808 {
3809 	if_ctx_t	ctx = if_getsoftc(ifp);
3810 
3811 	iflib_txq_t txq;
3812 	int err, qidx;
3813 
3814 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3815 		DBG_COUNTER_INC(tx_frees);
3816 		m_freem(m);
3817 		return (ENOBUFS);
3818 	}
3819 
3820 	MPASS(m->m_nextpkt == NULL);
3821 	qidx = 0;
3822 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3823 		qidx = QIDX(ctx, m);
3824 	/*
3825 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3826 	 */
3827 	txq = &ctx->ifc_txqs[qidx];
3828 
3829 #ifdef DRIVER_BACKPRESSURE
3830 	if (txq->ift_closed) {
3831 		while (m != NULL) {
3832 			next = m->m_nextpkt;
3833 			m->m_nextpkt = NULL;
3834 			m_freem(m);
3835 			m = next;
3836 		}
3837 		return (ENOBUFS);
3838 	}
3839 #endif
3840 #ifdef notyet
3841 	qidx = count = 0;
3842 	mp = marr;
3843 	next = m;
3844 	do {
3845 		count++;
3846 		next = next->m_nextpkt;
3847 	} while (next != NULL);
3848 
3849 	if (count > nitems(marr))
3850 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3851 			/* XXX check nextpkt */
3852 			m_freem(m);
3853 			/* XXX simplify for now */
3854 			DBG_COUNTER_INC(tx_frees);
3855 			return (ENOBUFS);
3856 		}
3857 	for (next = m, i = 0; next != NULL; i++) {
3858 		mp[i] = next;
3859 		next = next->m_nextpkt;
3860 		mp[i]->m_nextpkt = NULL;
3861 	}
3862 #endif
3863 	DBG_COUNTER_INC(tx_seen);
3864 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3865 
3866 	GROUPTASK_ENQUEUE(&txq->ift_task);
3867 	if (err) {
3868 		/* support forthcoming later */
3869 #ifdef DRIVER_BACKPRESSURE
3870 		txq->ift_closed = TRUE;
3871 #endif
3872 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3873 		m_freem(m);
3874 	}
3875 
3876 	return (err);
3877 }
3878 
3879 static void
3880 iflib_if_qflush(if_t ifp)
3881 {
3882 	if_ctx_t ctx = if_getsoftc(ifp);
3883 	iflib_txq_t txq = ctx->ifc_txqs;
3884 	int i;
3885 
3886 	STATE_LOCK(ctx);
3887 	ctx->ifc_flags |= IFC_QFLUSH;
3888 	STATE_UNLOCK(ctx);
3889 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3890 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3891 			iflib_txq_check_drain(txq, 0);
3892 	STATE_LOCK(ctx);
3893 	ctx->ifc_flags &= ~IFC_QFLUSH;
3894 	STATE_UNLOCK(ctx);
3895 
3896 	if_qflush(ifp);
3897 }
3898 
3899 
3900 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3901 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3902 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3903 
3904 static int
3905 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3906 {
3907 	if_ctx_t ctx = if_getsoftc(ifp);
3908 	struct ifreq	*ifr = (struct ifreq *)data;
3909 #if defined(INET) || defined(INET6)
3910 	struct ifaddr	*ifa = (struct ifaddr *)data;
3911 #endif
3912 	bool		avoid_reset = FALSE;
3913 	int		err = 0, reinit = 0, bits;
3914 
3915 	switch (command) {
3916 	case SIOCSIFADDR:
3917 #ifdef INET
3918 		if (ifa->ifa_addr->sa_family == AF_INET)
3919 			avoid_reset = TRUE;
3920 #endif
3921 #ifdef INET6
3922 		if (ifa->ifa_addr->sa_family == AF_INET6)
3923 			avoid_reset = TRUE;
3924 #endif
3925 		/*
3926 		** Calling init results in link renegotiation,
3927 		** so we avoid doing it when possible.
3928 		*/
3929 		if (avoid_reset) {
3930 			if_setflagbits(ifp, IFF_UP,0);
3931 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3932 				reinit = 1;
3933 #ifdef INET
3934 			if (!(if_getflags(ifp) & IFF_NOARP))
3935 				arp_ifinit(ifp, ifa);
3936 #endif
3937 		} else
3938 			err = ether_ioctl(ifp, command, data);
3939 		break;
3940 	case SIOCSIFMTU:
3941 		CTX_LOCK(ctx);
3942 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3943 			CTX_UNLOCK(ctx);
3944 			break;
3945 		}
3946 		bits = if_getdrvflags(ifp);
3947 		/* stop the driver and free any clusters before proceeding */
3948 		iflib_stop(ctx);
3949 
3950 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3951 			STATE_LOCK(ctx);
3952 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3953 				ctx->ifc_flags |= IFC_MULTISEG;
3954 			else
3955 				ctx->ifc_flags &= ~IFC_MULTISEG;
3956 			STATE_UNLOCK(ctx);
3957 			err = if_setmtu(ifp, ifr->ifr_mtu);
3958 		}
3959 		iflib_init_locked(ctx);
3960 		STATE_LOCK(ctx);
3961 		if_setdrvflags(ifp, bits);
3962 		STATE_UNLOCK(ctx);
3963 		CTX_UNLOCK(ctx);
3964 		break;
3965 	case SIOCSIFFLAGS:
3966 		CTX_LOCK(ctx);
3967 		if (if_getflags(ifp) & IFF_UP) {
3968 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3969 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3970 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3971 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3972 				}
3973 			} else
3974 				reinit = 1;
3975 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3976 			iflib_stop(ctx);
3977 		}
3978 		ctx->ifc_if_flags = if_getflags(ifp);
3979 		CTX_UNLOCK(ctx);
3980 		break;
3981 	case SIOCADDMULTI:
3982 	case SIOCDELMULTI:
3983 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3984 			CTX_LOCK(ctx);
3985 			IFDI_INTR_DISABLE(ctx);
3986 			IFDI_MULTI_SET(ctx);
3987 			IFDI_INTR_ENABLE(ctx);
3988 			CTX_UNLOCK(ctx);
3989 		}
3990 		break;
3991 	case SIOCSIFMEDIA:
3992 		CTX_LOCK(ctx);
3993 		IFDI_MEDIA_SET(ctx);
3994 		CTX_UNLOCK(ctx);
3995 		/* falls thru */
3996 	case SIOCGIFMEDIA:
3997 	case SIOCGIFXMEDIA:
3998 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3999 		break;
4000 	case SIOCGI2C:
4001 	{
4002 		struct ifi2creq i2c;
4003 
4004 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4005 		if (err != 0)
4006 			break;
4007 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4008 			err = EINVAL;
4009 			break;
4010 		}
4011 		if (i2c.len > sizeof(i2c.data)) {
4012 			err = EINVAL;
4013 			break;
4014 		}
4015 
4016 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4017 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4018 			    sizeof(i2c));
4019 		break;
4020 	}
4021 	case SIOCSIFCAP:
4022 	{
4023 		int mask, setmask;
4024 
4025 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4026 		setmask = 0;
4027 #ifdef TCP_OFFLOAD
4028 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4029 #endif
4030 		setmask |= (mask & IFCAP_FLAGS);
4031 
4032 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4033 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4034 		if ((mask & IFCAP_WOL) &&
4035 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4036 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4037 		if_vlancap(ifp);
4038 		/*
4039 		 * want to ensure that traffic has stopped before we change any of the flags
4040 		 */
4041 		if (setmask) {
4042 			CTX_LOCK(ctx);
4043 			bits = if_getdrvflags(ifp);
4044 			if (bits & IFF_DRV_RUNNING)
4045 				iflib_stop(ctx);
4046 			STATE_LOCK(ctx);
4047 			if_togglecapenable(ifp, setmask);
4048 			STATE_UNLOCK(ctx);
4049 			if (bits & IFF_DRV_RUNNING)
4050 				iflib_init_locked(ctx);
4051 			STATE_LOCK(ctx);
4052 			if_setdrvflags(ifp, bits);
4053 			STATE_UNLOCK(ctx);
4054 			CTX_UNLOCK(ctx);
4055 		}
4056 		break;
4057 	    }
4058 	case SIOCGPRIVATE_0:
4059 	case SIOCSDRVSPEC:
4060 	case SIOCGDRVSPEC:
4061 		CTX_LOCK(ctx);
4062 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4063 		CTX_UNLOCK(ctx);
4064 		break;
4065 	default:
4066 		err = ether_ioctl(ifp, command, data);
4067 		break;
4068 	}
4069 	if (reinit)
4070 		iflib_if_init(ctx);
4071 	return (err);
4072 }
4073 
4074 static uint64_t
4075 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4076 {
4077 	if_ctx_t ctx = if_getsoftc(ifp);
4078 
4079 	return (IFDI_GET_COUNTER(ctx, cnt));
4080 }
4081 
4082 /*********************************************************************
4083  *
4084  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4085  *
4086  **********************************************************************/
4087 
4088 static void
4089 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4090 {
4091 	if_ctx_t ctx = if_getsoftc(ifp);
4092 
4093 	if ((void *)ctx != arg)
4094 		return;
4095 
4096 	if ((vtag == 0) || (vtag > 4095))
4097 		return;
4098 
4099 	CTX_LOCK(ctx);
4100 	IFDI_VLAN_REGISTER(ctx, vtag);
4101 	/* Re-init to load the changes */
4102 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4103 		iflib_if_init_locked(ctx);
4104 	CTX_UNLOCK(ctx);
4105 }
4106 
4107 static void
4108 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4109 {
4110 	if_ctx_t ctx = if_getsoftc(ifp);
4111 
4112 	if ((void *)ctx != arg)
4113 		return;
4114 
4115 	if ((vtag == 0) || (vtag > 4095))
4116 		return;
4117 
4118 	CTX_LOCK(ctx);
4119 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4120 	/* Re-init to load the changes */
4121 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4122 		iflib_if_init_locked(ctx);
4123 	CTX_UNLOCK(ctx);
4124 }
4125 
4126 static void
4127 iflib_led_func(void *arg, int onoff)
4128 {
4129 	if_ctx_t ctx = arg;
4130 
4131 	CTX_LOCK(ctx);
4132 	IFDI_LED_FUNC(ctx, onoff);
4133 	CTX_UNLOCK(ctx);
4134 }
4135 
4136 /*********************************************************************
4137  *
4138  *  BUS FUNCTION DEFINITIONS
4139  *
4140  **********************************************************************/
4141 
4142 int
4143 iflib_device_probe(device_t dev)
4144 {
4145 	pci_vendor_info_t *ent;
4146 
4147 	uint16_t	pci_vendor_id, pci_device_id;
4148 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4149 	uint16_t	pci_rev_id;
4150 	if_shared_ctx_t sctx;
4151 
4152 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4153 		return (ENOTSUP);
4154 
4155 	pci_vendor_id = pci_get_vendor(dev);
4156 	pci_device_id = pci_get_device(dev);
4157 	pci_subvendor_id = pci_get_subvendor(dev);
4158 	pci_subdevice_id = pci_get_subdevice(dev);
4159 	pci_rev_id = pci_get_revid(dev);
4160 	if (sctx->isc_parse_devinfo != NULL)
4161 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4162 
4163 	ent = sctx->isc_vendor_info;
4164 	while (ent->pvi_vendor_id != 0) {
4165 		if (pci_vendor_id != ent->pvi_vendor_id) {
4166 			ent++;
4167 			continue;
4168 		}
4169 		if ((pci_device_id == ent->pvi_device_id) &&
4170 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4171 		     (ent->pvi_subvendor_id == 0)) &&
4172 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4173 		     (ent->pvi_subdevice_id == 0)) &&
4174 		    ((pci_rev_id == ent->pvi_rev_id) ||
4175 		     (ent->pvi_rev_id == 0))) {
4176 
4177 			device_set_desc_copy(dev, ent->pvi_name);
4178 			/* this needs to be changed to zero if the bus probing code
4179 			 * ever stops re-probing on best match because the sctx
4180 			 * may have its values over written by register calls
4181 			 * in subsequent probes
4182 			 */
4183 			return (BUS_PROBE_DEFAULT);
4184 		}
4185 		ent++;
4186 	}
4187 	return (ENXIO);
4188 }
4189 
4190 int
4191 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4192 {
4193 	int err, rid, msix, msix_bar;
4194 	if_ctx_t ctx;
4195 	if_t ifp;
4196 	if_softc_ctx_t scctx;
4197 	int i;
4198 	uint16_t main_txq;
4199 	uint16_t main_rxq;
4200 
4201 
4202 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4203 
4204 	if (sc == NULL) {
4205 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4206 		device_set_softc(dev, ctx);
4207 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4208 	}
4209 
4210 	ctx->ifc_sctx = sctx;
4211 	ctx->ifc_dev = dev;
4212 	ctx->ifc_softc = sc;
4213 
4214 	if ((err = iflib_register(ctx)) != 0) {
4215 		device_printf(dev, "iflib_register failed %d\n", err);
4216 		return (err);
4217 	}
4218 	iflib_add_device_sysctl_pre(ctx);
4219 
4220 	scctx = &ctx->ifc_softc_ctx;
4221 	ifp = ctx->ifc_ifp;
4222 
4223 	/*
4224 	 * XXX sanity check that ntxd & nrxd are a power of 2
4225 	 */
4226 	if (ctx->ifc_sysctl_ntxqs != 0)
4227 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4228 	if (ctx->ifc_sysctl_nrxqs != 0)
4229 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4230 
4231 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4232 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4233 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4234 		else
4235 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4236 	}
4237 
4238 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4239 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4240 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4241 		else
4242 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4243 	}
4244 
4245 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4246 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4247 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4248 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4249 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4250 		}
4251 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4252 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4253 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4254 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4255 		}
4256 	}
4257 
4258 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4259 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4260 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4261 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4262 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4263 		}
4264 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4265 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4266 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4267 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4268 		}
4269 	}
4270 
4271 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4272 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4273 		return (err);
4274 	}
4275 	_iflib_pre_assert(scctx);
4276 	ctx->ifc_txrx = *scctx->isc_txrx;
4277 
4278 #ifdef INVARIANTS
4279 	MPASS(scctx->isc_capenable);
4280 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4281 		MPASS(scctx->isc_tx_csum_flags);
4282 #endif
4283 
4284 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4285 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4286 
4287 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4288 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4289 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4290 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4291 
4292 #ifdef ACPI_DMAR
4293 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4294 		ctx->ifc_flags |= IFC_DMAR;
4295 #elif !(defined(__i386__) || defined(__amd64__))
4296 	/* set unconditionally for !x86 */
4297 	ctx->ifc_flags |= IFC_DMAR;
4298 #endif
4299 
4300 	msix_bar = scctx->isc_msix_bar;
4301 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4302 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4303 
4304 	/* XXX change for per-queue sizes */
4305 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4306 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4307 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4308 		if (!powerof2(scctx->isc_nrxd[i])) {
4309 			/* round down instead? */
4310 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4311 			err = EINVAL;
4312 			goto fail;
4313 		}
4314 	}
4315 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4316 		if (!powerof2(scctx->isc_ntxd[i])) {
4317 			device_printf(dev,
4318 			    "# tx descriptors must be a power of 2");
4319 			err = EINVAL;
4320 			goto fail;
4321 		}
4322 	}
4323 
4324 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4325 	    MAX_SINGLE_PACKET_FRACTION)
4326 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4327 		    MAX_SINGLE_PACKET_FRACTION);
4328 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4329 	    MAX_SINGLE_PACKET_FRACTION)
4330 		scctx->isc_tx_tso_segments_max = max(1,
4331 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4332 
4333 	/*
4334 	 * Protect the stack against modern hardware
4335 	 */
4336 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4337 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4338 
4339 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4340 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4341 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4342 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4343 	if (scctx->isc_rss_table_size == 0)
4344 		scctx->isc_rss_table_size = 64;
4345 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4346 
4347 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4348 	/* XXX format name */
4349 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4350 
4351 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4352 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4353 		device_printf(dev, "Unable to fetch CPU list\n");
4354 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4355 	}
4356 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4357 
4358 	/*
4359 	** Now setup MSI or MSI/X, should
4360 	** return us the number of supported
4361 	** vectors. (Will be 1 for MSI)
4362 	*/
4363 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4364 		msix = scctx->isc_vectors;
4365 	} else if (scctx->isc_msix_bar != 0)
4366 	       /*
4367 		* The simple fact that isc_msix_bar is not 0 does not mean we
4368 		* we have a good value there that is known to work.
4369 		*/
4370 		msix = iflib_msix_init(ctx);
4371 	else {
4372 		scctx->isc_vectors = 1;
4373 		scctx->isc_ntxqsets = 1;
4374 		scctx->isc_nrxqsets = 1;
4375 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4376 		msix = 0;
4377 	}
4378 	/* Get memory for the station queues */
4379 	if ((err = iflib_queues_alloc(ctx))) {
4380 		device_printf(dev, "Unable to allocate queue memory\n");
4381 		goto fail;
4382 	}
4383 
4384 	if ((err = iflib_qset_structures_setup(ctx))) {
4385 		device_printf(dev, "qset structure setup failed %d\n", err);
4386 		goto fail_queues;
4387 	}
4388 
4389 	/*
4390 	 * Group taskqueues aren't properly set up until SMP is started,
4391 	 * so we disable interrupts until we can handle them post
4392 	 * SI_SUB_SMP.
4393 	 *
4394 	 * XXX: disabling interrupts doesn't actually work, at least for
4395 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4396 	 * we do null handling and depend on this not causing too large an
4397 	 * interrupt storm.
4398 	 */
4399 	IFDI_INTR_DISABLE(ctx);
4400 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4401 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4402 		goto fail_intr_free;
4403 	}
4404 	if (msix <= 1) {
4405 		rid = 0;
4406 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4407 			MPASS(msix == 1);
4408 			rid = 1;
4409 		}
4410 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4411 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4412 			goto fail_intr_free;
4413 		}
4414 	}
4415 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4416 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4417 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4418 		goto fail_detach;
4419 	}
4420 	if ((err = iflib_netmap_attach(ctx))) {
4421 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4422 		goto fail_detach;
4423 	}
4424 	*ctxp = ctx;
4425 
4426 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4427 	iflib_add_device_sysctl_post(ctx);
4428 	ctx->ifc_flags |= IFC_INIT_DONE;
4429 	return (0);
4430 fail_detach:
4431 	ether_ifdetach(ctx->ifc_ifp);
4432 fail_intr_free:
4433 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4434 		pci_release_msi(ctx->ifc_dev);
4435 fail_queues:
4436 	/* XXX free queues */
4437 fail:
4438 	IFDI_DETACH(ctx);
4439 	return (err);
4440 }
4441 
4442 int
4443 iflib_device_attach(device_t dev)
4444 {
4445 	if_ctx_t ctx;
4446 	if_shared_ctx_t sctx;
4447 
4448 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4449 		return (ENOTSUP);
4450 
4451 	pci_enable_busmaster(dev);
4452 
4453 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4454 }
4455 
4456 int
4457 iflib_device_deregister(if_ctx_t ctx)
4458 {
4459 	if_t ifp = ctx->ifc_ifp;
4460 	iflib_txq_t txq;
4461 	iflib_rxq_t rxq;
4462 	device_t dev = ctx->ifc_dev;
4463 	int i, j;
4464 	struct taskqgroup *tqg;
4465 	iflib_fl_t fl;
4466 
4467 	/* Make sure VLANS are not using driver */
4468 	if (if_vlantrunkinuse(ifp)) {
4469 		device_printf(dev,"Vlan in use, detach first\n");
4470 		return (EBUSY);
4471 	}
4472 
4473 	CTX_LOCK(ctx);
4474 	ctx->ifc_in_detach = 1;
4475 	iflib_stop(ctx);
4476 	CTX_UNLOCK(ctx);
4477 
4478 	/* Unregister VLAN events */
4479 	if (ctx->ifc_vlan_attach_event != NULL)
4480 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4481 	if (ctx->ifc_vlan_detach_event != NULL)
4482 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4483 
4484 	iflib_netmap_detach(ifp);
4485 	ether_ifdetach(ifp);
4486 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4487 	CTX_LOCK_DESTROY(ctx);
4488 	if (ctx->ifc_led_dev != NULL)
4489 		led_destroy(ctx->ifc_led_dev);
4490 	/* XXX drain any dependent tasks */
4491 	tqg = qgroup_if_io_tqg;
4492 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4493 		callout_drain(&txq->ift_timer);
4494 		if (txq->ift_task.gt_uniq != NULL)
4495 			taskqgroup_detach(tqg, &txq->ift_task);
4496 	}
4497 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4498 		if (rxq->ifr_task.gt_uniq != NULL)
4499 			taskqgroup_detach(tqg, &rxq->ifr_task);
4500 
4501 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4502 			free(fl->ifl_rx_bitmap, M_IFLIB);
4503 
4504 	}
4505 	tqg = qgroup_if_config_tqg;
4506 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4507 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4508 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4509 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4510 
4511 	IFDI_DETACH(ctx);
4512 	device_set_softc(ctx->ifc_dev, NULL);
4513 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4514 		pci_release_msi(dev);
4515 	}
4516 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4517 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4518 	}
4519 	if (ctx->ifc_msix_mem != NULL) {
4520 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4521 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4522 		ctx->ifc_msix_mem = NULL;
4523 	}
4524 
4525 	bus_generic_detach(dev);
4526 	if_free(ifp);
4527 
4528 	iflib_tx_structures_free(ctx);
4529 	iflib_rx_structures_free(ctx);
4530 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4531 		free(ctx->ifc_softc, M_IFLIB);
4532 	free(ctx, M_IFLIB);
4533 	return (0);
4534 }
4535 
4536 
4537 int
4538 iflib_device_detach(device_t dev)
4539 {
4540 	if_ctx_t ctx = device_get_softc(dev);
4541 
4542 	return (iflib_device_deregister(ctx));
4543 }
4544 
4545 int
4546 iflib_device_suspend(device_t dev)
4547 {
4548 	if_ctx_t ctx = device_get_softc(dev);
4549 
4550 	CTX_LOCK(ctx);
4551 	IFDI_SUSPEND(ctx);
4552 	CTX_UNLOCK(ctx);
4553 
4554 	return bus_generic_suspend(dev);
4555 }
4556 int
4557 iflib_device_shutdown(device_t dev)
4558 {
4559 	if_ctx_t ctx = device_get_softc(dev);
4560 
4561 	CTX_LOCK(ctx);
4562 	IFDI_SHUTDOWN(ctx);
4563 	CTX_UNLOCK(ctx);
4564 
4565 	return bus_generic_suspend(dev);
4566 }
4567 
4568 
4569 int
4570 iflib_device_resume(device_t dev)
4571 {
4572 	if_ctx_t ctx = device_get_softc(dev);
4573 	iflib_txq_t txq = ctx->ifc_txqs;
4574 
4575 	CTX_LOCK(ctx);
4576 	IFDI_RESUME(ctx);
4577 	iflib_init_locked(ctx);
4578 	CTX_UNLOCK(ctx);
4579 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4580 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4581 
4582 	return (bus_generic_resume(dev));
4583 }
4584 
4585 int
4586 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4587 {
4588 	int error;
4589 	if_ctx_t ctx = device_get_softc(dev);
4590 
4591 	CTX_LOCK(ctx);
4592 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4593 	CTX_UNLOCK(ctx);
4594 
4595 	return (error);
4596 }
4597 
4598 void
4599 iflib_device_iov_uninit(device_t dev)
4600 {
4601 	if_ctx_t ctx = device_get_softc(dev);
4602 
4603 	CTX_LOCK(ctx);
4604 	IFDI_IOV_UNINIT(ctx);
4605 	CTX_UNLOCK(ctx);
4606 }
4607 
4608 int
4609 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4610 {
4611 	int error;
4612 	if_ctx_t ctx = device_get_softc(dev);
4613 
4614 	CTX_LOCK(ctx);
4615 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4616 	CTX_UNLOCK(ctx);
4617 
4618 	return (error);
4619 }
4620 
4621 /*********************************************************************
4622  *
4623  *  MODULE FUNCTION DEFINITIONS
4624  *
4625  **********************************************************************/
4626 
4627 /*
4628  * - Start a fast taskqueue thread for each core
4629  * - Start a taskqueue for control operations
4630  */
4631 static int
4632 iflib_module_init(void)
4633 {
4634 	return (0);
4635 }
4636 
4637 static int
4638 iflib_module_event_handler(module_t mod, int what, void *arg)
4639 {
4640 	int err;
4641 
4642 	switch (what) {
4643 	case MOD_LOAD:
4644 		if ((err = iflib_module_init()) != 0)
4645 			return (err);
4646 		break;
4647 	case MOD_UNLOAD:
4648 		return (EBUSY);
4649 	default:
4650 		return (EOPNOTSUPP);
4651 	}
4652 
4653 	return (0);
4654 }
4655 
4656 /*********************************************************************
4657  *
4658  *  PUBLIC FUNCTION DEFINITIONS
4659  *     ordered as in iflib.h
4660  *
4661  **********************************************************************/
4662 
4663 
4664 static void
4665 _iflib_assert(if_shared_ctx_t sctx)
4666 {
4667 	MPASS(sctx->isc_tx_maxsize);
4668 	MPASS(sctx->isc_tx_maxsegsize);
4669 
4670 	MPASS(sctx->isc_rx_maxsize);
4671 	MPASS(sctx->isc_rx_nsegments);
4672 	MPASS(sctx->isc_rx_maxsegsize);
4673 
4674 	MPASS(sctx->isc_nrxd_min[0]);
4675 	MPASS(sctx->isc_nrxd_max[0]);
4676 	MPASS(sctx->isc_nrxd_default[0]);
4677 	MPASS(sctx->isc_ntxd_min[0]);
4678 	MPASS(sctx->isc_ntxd_max[0]);
4679 	MPASS(sctx->isc_ntxd_default[0]);
4680 }
4681 
4682 static void
4683 _iflib_pre_assert(if_softc_ctx_t scctx)
4684 {
4685 
4686 	MPASS(scctx->isc_txrx->ift_txd_encap);
4687 	MPASS(scctx->isc_txrx->ift_txd_flush);
4688 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4689 	MPASS(scctx->isc_txrx->ift_rxd_available);
4690 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4691 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4692 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4693 }
4694 
4695 static int
4696 iflib_register(if_ctx_t ctx)
4697 {
4698 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4699 	driver_t *driver = sctx->isc_driver;
4700 	device_t dev = ctx->ifc_dev;
4701 	if_t ifp;
4702 
4703 	_iflib_assert(sctx);
4704 
4705 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4706 
4707 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4708 	if (ifp == NULL) {
4709 		device_printf(dev, "can not allocate ifnet structure\n");
4710 		return (ENOMEM);
4711 	}
4712 
4713 	/*
4714 	 * Initialize our context's device specific methods
4715 	 */
4716 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4717 	kobj_class_compile((kobj_class_t) driver);
4718 	driver->refs++;
4719 
4720 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4721 	if_setsoftc(ifp, ctx);
4722 	if_setdev(ifp, dev);
4723 	if_setinitfn(ifp, iflib_if_init);
4724 	if_setioctlfn(ifp, iflib_if_ioctl);
4725 	if_settransmitfn(ifp, iflib_if_transmit);
4726 	if_setqflushfn(ifp, iflib_if_qflush);
4727 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4728 
4729 	ctx->ifc_vlan_attach_event =
4730 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4731 							  EVENTHANDLER_PRI_FIRST);
4732 	ctx->ifc_vlan_detach_event =
4733 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4734 							  EVENTHANDLER_PRI_FIRST);
4735 
4736 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4737 					 iflib_media_change, iflib_media_status);
4738 
4739 	return (0);
4740 }
4741 
4742 
4743 static int
4744 iflib_queues_alloc(if_ctx_t ctx)
4745 {
4746 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4747 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4748 	device_t dev = ctx->ifc_dev;
4749 	int nrxqsets = scctx->isc_nrxqsets;
4750 	int ntxqsets = scctx->isc_ntxqsets;
4751 	iflib_txq_t txq;
4752 	iflib_rxq_t rxq;
4753 	iflib_fl_t fl = NULL;
4754 	int i, j, cpu, err, txconf, rxconf;
4755 	iflib_dma_info_t ifdip;
4756 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4757 	uint32_t *txqsizes = scctx->isc_txqsizes;
4758 	uint8_t nrxqs = sctx->isc_nrxqs;
4759 	uint8_t ntxqs = sctx->isc_ntxqs;
4760 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4761 	caddr_t *vaddrs;
4762 	uint64_t *paddrs;
4763 	struct ifmp_ring **brscp;
4764 
4765 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4766 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4767 
4768 	brscp = NULL;
4769 	txq = NULL;
4770 	rxq = NULL;
4771 
4772 /* Allocate the TX ring struct memory */
4773 	if (!(txq =
4774 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4775 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4776 		device_printf(dev, "Unable to allocate TX ring memory\n");
4777 		err = ENOMEM;
4778 		goto fail;
4779 	}
4780 
4781 	/* Now allocate the RX */
4782 	if (!(rxq =
4783 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4784 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4785 		device_printf(dev, "Unable to allocate RX ring memory\n");
4786 		err = ENOMEM;
4787 		goto rx_fail;
4788 	}
4789 
4790 	ctx->ifc_txqs = txq;
4791 	ctx->ifc_rxqs = rxq;
4792 
4793 	/*
4794 	 * XXX handle allocation failure
4795 	 */
4796 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4797 		/* Set up some basics */
4798 
4799 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4800 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4801 			err = ENOMEM;
4802 			goto err_tx_desc;
4803 		}
4804 		txq->ift_ifdi = ifdip;
4805 		for (j = 0; j < ntxqs; j++, ifdip++) {
4806 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4807 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4808 				err = ENOMEM;
4809 				goto err_tx_desc;
4810 			}
4811 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4812 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4813 		}
4814 		txq->ift_ctx = ctx;
4815 		txq->ift_id = i;
4816 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4817 			txq->ift_br_offset = 1;
4818 		} else {
4819 			txq->ift_br_offset = 0;
4820 		}
4821 		/* XXX fix this */
4822 		txq->ift_timer.c_cpu = cpu;
4823 
4824 		if (iflib_txsd_alloc(txq)) {
4825 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4826 			err = ENOMEM;
4827 			goto err_tx_desc;
4828 		}
4829 
4830 		/* Initialize the TX lock */
4831 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4832 		    device_get_nameunit(dev), txq->ift_id);
4833 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4834 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4835 
4836 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4837 			 device_get_nameunit(dev), txq->ift_id);
4838 
4839 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4840 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4841 		if (err) {
4842 			/* XXX free any allocated rings */
4843 			device_printf(dev, "Unable to allocate buf_ring\n");
4844 			goto err_tx_desc;
4845 		}
4846 	}
4847 
4848 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4849 		/* Set up some basics */
4850 
4851 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4852 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4853 			err = ENOMEM;
4854 			goto err_tx_desc;
4855 		}
4856 
4857 		rxq->ifr_ifdi = ifdip;
4858 		/* XXX this needs to be changed if #rx queues != #tx queues */
4859 		rxq->ifr_ntxqirq = 1;
4860 		rxq->ifr_txqid[0] = i;
4861 		for (j = 0; j < nrxqs; j++, ifdip++) {
4862 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4863 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4864 				err = ENOMEM;
4865 				goto err_tx_desc;
4866 			}
4867 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4868 		}
4869 		rxq->ifr_ctx = ctx;
4870 		rxq->ifr_id = i;
4871 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4872 			rxq->ifr_fl_offset = 1;
4873 		} else {
4874 			rxq->ifr_fl_offset = 0;
4875 		}
4876 		rxq->ifr_nfl = nfree_lists;
4877 		if (!(fl =
4878 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4879 			device_printf(dev, "Unable to allocate free list memory\n");
4880 			err = ENOMEM;
4881 			goto err_tx_desc;
4882 		}
4883 		rxq->ifr_fl = fl;
4884 		for (j = 0; j < nfree_lists; j++) {
4885 			fl[j].ifl_rxq = rxq;
4886 			fl[j].ifl_id = j;
4887 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4888 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4889 		}
4890         /* Allocate receive buffers for the ring*/
4891 		if (iflib_rxsd_alloc(rxq)) {
4892 			device_printf(dev,
4893 			    "Critical Failure setting up receive buffers\n");
4894 			err = ENOMEM;
4895 			goto err_rx_desc;
4896 		}
4897 
4898 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4899 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4900 	}
4901 
4902 	/* TXQs */
4903 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4904 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4905 	for (i = 0; i < ntxqsets; i++) {
4906 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4907 
4908 		for (j = 0; j < ntxqs; j++, di++) {
4909 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4910 			paddrs[i*ntxqs + j] = di->idi_paddr;
4911 		}
4912 	}
4913 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4914 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4915 		iflib_tx_structures_free(ctx);
4916 		free(vaddrs, M_IFLIB);
4917 		free(paddrs, M_IFLIB);
4918 		goto err_rx_desc;
4919 	}
4920 	free(vaddrs, M_IFLIB);
4921 	free(paddrs, M_IFLIB);
4922 
4923 	/* RXQs */
4924 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4925 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4926 	for (i = 0; i < nrxqsets; i++) {
4927 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4928 
4929 		for (j = 0; j < nrxqs; j++, di++) {
4930 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4931 			paddrs[i*nrxqs + j] = di->idi_paddr;
4932 		}
4933 	}
4934 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4935 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4936 		iflib_tx_structures_free(ctx);
4937 		free(vaddrs, M_IFLIB);
4938 		free(paddrs, M_IFLIB);
4939 		goto err_rx_desc;
4940 	}
4941 	free(vaddrs, M_IFLIB);
4942 	free(paddrs, M_IFLIB);
4943 
4944 	return (0);
4945 
4946 /* XXX handle allocation failure changes */
4947 err_rx_desc:
4948 err_tx_desc:
4949 	if (ctx->ifc_rxqs != NULL)
4950 		free(ctx->ifc_rxqs, M_IFLIB);
4951 	ctx->ifc_rxqs = NULL;
4952 	if (ctx->ifc_txqs != NULL)
4953 		free(ctx->ifc_txqs, M_IFLIB);
4954 	ctx->ifc_txqs = NULL;
4955 rx_fail:
4956 	if (brscp != NULL)
4957 		free(brscp, M_IFLIB);
4958 	if (rxq != NULL)
4959 		free(rxq, M_IFLIB);
4960 	if (txq != NULL)
4961 		free(txq, M_IFLIB);
4962 fail:
4963 	return (err);
4964 }
4965 
4966 static int
4967 iflib_tx_structures_setup(if_ctx_t ctx)
4968 {
4969 	iflib_txq_t txq = ctx->ifc_txqs;
4970 	int i;
4971 
4972 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4973 		iflib_txq_setup(txq);
4974 
4975 	return (0);
4976 }
4977 
4978 static void
4979 iflib_tx_structures_free(if_ctx_t ctx)
4980 {
4981 	iflib_txq_t txq = ctx->ifc_txqs;
4982 	int i, j;
4983 
4984 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4985 		iflib_txq_destroy(txq);
4986 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4987 			iflib_dma_free(&txq->ift_ifdi[j]);
4988 	}
4989 	free(ctx->ifc_txqs, M_IFLIB);
4990 	ctx->ifc_txqs = NULL;
4991 	IFDI_QUEUES_FREE(ctx);
4992 }
4993 
4994 /*********************************************************************
4995  *
4996  *  Initialize all receive rings.
4997  *
4998  **********************************************************************/
4999 static int
5000 iflib_rx_structures_setup(if_ctx_t ctx)
5001 {
5002 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5003 	int q;
5004 #if defined(INET6) || defined(INET)
5005 	int i, err;
5006 #endif
5007 
5008 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5009 #if defined(INET6) || defined(INET)
5010 		tcp_lro_free(&rxq->ifr_lc);
5011 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5012 		    TCP_LRO_ENTRIES, min(1024,
5013 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5014 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5015 			goto fail;
5016 		}
5017 		rxq->ifr_lro_enabled = TRUE;
5018 #endif
5019 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5020 	}
5021 	return (0);
5022 #if defined(INET6) || defined(INET)
5023 fail:
5024 	/*
5025 	 * Free RX software descriptors allocated so far, we will only handle
5026 	 * the rings that completed, the failing case will have
5027 	 * cleaned up for itself. 'q' failed, so its the terminus.
5028 	 */
5029 	rxq = ctx->ifc_rxqs;
5030 	for (i = 0; i < q; ++i, rxq++) {
5031 		iflib_rx_sds_free(rxq);
5032 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5033 	}
5034 	return (err);
5035 #endif
5036 }
5037 
5038 /*********************************************************************
5039  *
5040  *  Free all receive rings.
5041  *
5042  **********************************************************************/
5043 static void
5044 iflib_rx_structures_free(if_ctx_t ctx)
5045 {
5046 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5047 
5048 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5049 		iflib_rx_sds_free(rxq);
5050 	}
5051 }
5052 
5053 static int
5054 iflib_qset_structures_setup(if_ctx_t ctx)
5055 {
5056 	int err;
5057 
5058 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5059 		return (err);
5060 
5061 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5062 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5063 		iflib_tx_structures_free(ctx);
5064 		iflib_rx_structures_free(ctx);
5065 	}
5066 	return (err);
5067 }
5068 
5069 int
5070 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5071 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5072 {
5073 
5074 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5075 }
5076 
5077 #ifdef SMP
5078 static int
5079 find_nth(if_ctx_t ctx, int qid)
5080 {
5081 	cpuset_t cpus;
5082 	int i, cpuid, eqid, count;
5083 
5084 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5085 	count = CPU_COUNT(&cpus);
5086 	eqid = qid % count;
5087 	/* clear up to the qid'th bit */
5088 	for (i = 0; i < eqid; i++) {
5089 		cpuid = CPU_FFS(&cpus);
5090 		MPASS(cpuid != 0);
5091 		CPU_CLR(cpuid-1, &cpus);
5092 	}
5093 	cpuid = CPU_FFS(&cpus);
5094 	MPASS(cpuid != 0);
5095 	return (cpuid-1);
5096 }
5097 
5098 #ifdef SCHED_ULE
5099 extern struct cpu_group *cpu_top;              /* CPU topology */
5100 
5101 static int
5102 find_child_with_core(int cpu, struct cpu_group *grp)
5103 {
5104 	int i;
5105 
5106 	if (grp->cg_children == 0)
5107 		return -1;
5108 
5109 	MPASS(grp->cg_child);
5110 	for (i = 0; i < grp->cg_children; i++) {
5111 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5112 			return i;
5113 	}
5114 
5115 	return -1;
5116 }
5117 
5118 /*
5119  * Find the nth thread on the specified core
5120  */
5121 static int
5122 find_thread(int cpu, int thread_num)
5123 {
5124 	struct cpu_group *grp;
5125 	int i;
5126 	cpuset_t cs;
5127 
5128 	grp = cpu_top;
5129 	if (grp == NULL)
5130 		return cpu;
5131 	i = 0;
5132 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5133 		/* If the child only has one cpu, don't descend */
5134 		if (grp->cg_child[i].cg_count <= 1)
5135 			break;
5136 		grp = &grp->cg_child[i];
5137 	}
5138 
5139 	/* If they don't share at least an L2 cache, use the same CPU */
5140 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5141 		return cpu;
5142 
5143 	/* Now pick one */
5144 	CPU_COPY(&grp->cg_mask, &cs);
5145 	for (i = thread_num % grp->cg_count; i > 0; i--) {
5146 		MPASS(CPU_FFS(&cs));
5147 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5148 	}
5149 	MPASS(CPU_FFS(&cs));
5150 	return CPU_FFS(&cs) - 1;
5151 }
5152 #else
5153 static int
5154 find_thread(int cpu, int thread_num __unused)
5155 {
5156 	return cpu;
5157 }
5158 #endif
5159 
5160 static int
5161 get_thread_num(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5162 {
5163 	switch (type) {
5164 	case IFLIB_INTR_TX:
5165 		/* TX queues get threads on the same core as the corresponding RX queue */
5166 		/* XXX handle multiple RX threads per core and more than two threads per core */
5167 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5168 	case IFLIB_INTR_RX:
5169 	case IFLIB_INTR_RXTX:
5170 		/* RX queues get the first thread on their core */
5171 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5172 	default:
5173 		return -1;
5174 	}
5175 }
5176 #else
5177 #define get_thread_num(ctx, type, qid)	CPU_FIRST()
5178 #define find_thread(cpuid, tid)		CPU_FIRST()
5179 #define find_nth(ctx, gid)		CPU_FIRST()
5180 #endif
5181 
5182 /* Just to avoid copy/paste */
5183 static inline int
5184 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5185     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5186 {
5187 	int cpuid;
5188 	int err, tid;
5189 
5190 	cpuid = find_nth(ctx, qid);
5191 	tid = get_thread_num(ctx, type, qid);
5192 	MPASS(tid >= 0);
5193 	cpuid = find_thread(cpuid, tid);
5194 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5195 	if (err) {
5196 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5197 		return (err);
5198 	}
5199 #ifdef notyet
5200 	if (cpuid > ctx->ifc_cpuid_highest)
5201 		ctx->ifc_cpuid_highest = cpuid;
5202 #endif
5203 	return 0;
5204 }
5205 
5206 int
5207 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5208 						iflib_intr_type_t type, driver_filter_t *filter,
5209 						void *filter_arg, int qid, char *name)
5210 {
5211 	struct grouptask *gtask;
5212 	struct taskqgroup *tqg;
5213 	iflib_filter_info_t info;
5214 	gtask_fn_t *fn;
5215 	int tqrid, err;
5216 	driver_filter_t *intr_fast;
5217 	void *q;
5218 
5219 	info = &ctx->ifc_filter_info;
5220 	tqrid = rid;
5221 
5222 	switch (type) {
5223 	/* XXX merge tx/rx for netmap? */
5224 	case IFLIB_INTR_TX:
5225 		q = &ctx->ifc_txqs[qid];
5226 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5227 		gtask = &ctx->ifc_txqs[qid].ift_task;
5228 		tqg = qgroup_if_io_tqg;
5229 		fn = _task_fn_tx;
5230 		intr_fast = iflib_fast_intr;
5231 		GROUPTASK_INIT(gtask, 0, fn, q);
5232 		break;
5233 	case IFLIB_INTR_RX:
5234 		q = &ctx->ifc_rxqs[qid];
5235 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5236 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5237 		tqg = qgroup_if_io_tqg;
5238 		fn = _task_fn_rx;
5239 		intr_fast = iflib_fast_intr;
5240 		GROUPTASK_INIT(gtask, 0, fn, q);
5241 		break;
5242 	case IFLIB_INTR_RXTX:
5243 		q = &ctx->ifc_rxqs[qid];
5244 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5245 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5246 		tqg = qgroup_if_io_tqg;
5247 		fn = _task_fn_rx;
5248 		intr_fast = iflib_fast_intr_rxtx;
5249 		GROUPTASK_INIT(gtask, 0, fn, q);
5250 		break;
5251 	case IFLIB_INTR_ADMIN:
5252 		q = ctx;
5253 		tqrid = -1;
5254 		info = &ctx->ifc_filter_info;
5255 		gtask = &ctx->ifc_admin_task;
5256 		tqg = qgroup_if_config_tqg;
5257 		fn = _task_fn_admin;
5258 		intr_fast = iflib_fast_intr_ctx;
5259 		break;
5260 	default:
5261 		panic("unknown net intr type");
5262 	}
5263 
5264 	info->ifi_filter = filter;
5265 	info->ifi_filter_arg = filter_arg;
5266 	info->ifi_task = gtask;
5267 	info->ifi_ctx = q;
5268 
5269 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5270 	if (err != 0) {
5271 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5272 		return (err);
5273 	}
5274 	if (type == IFLIB_INTR_ADMIN)
5275 		return (0);
5276 
5277 	if (tqrid != -1) {
5278 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5279 		if (err)
5280 			return (err);
5281 	} else {
5282 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5283 	}
5284 
5285 	return (0);
5286 }
5287 
5288 void
5289 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5290 {
5291 	struct grouptask *gtask;
5292 	struct taskqgroup *tqg;
5293 	gtask_fn_t *fn;
5294 	void *q;
5295 	int irq_num = -1;
5296 	int err;
5297 
5298 	switch (type) {
5299 	case IFLIB_INTR_TX:
5300 		q = &ctx->ifc_txqs[qid];
5301 		gtask = &ctx->ifc_txqs[qid].ift_task;
5302 		tqg = qgroup_if_io_tqg;
5303 		fn = _task_fn_tx;
5304 		if (irq != NULL)
5305 			irq_num = rman_get_start(irq->ii_res);
5306 		break;
5307 	case IFLIB_INTR_RX:
5308 		q = &ctx->ifc_rxqs[qid];
5309 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5310 		tqg = qgroup_if_io_tqg;
5311 		fn = _task_fn_rx;
5312 		if (irq != NULL)
5313 			irq_num = rman_get_start(irq->ii_res);
5314 		break;
5315 	case IFLIB_INTR_IOV:
5316 		q = ctx;
5317 		gtask = &ctx->ifc_vflr_task;
5318 		tqg = qgroup_if_config_tqg;
5319 		fn = _task_fn_iov;
5320 		break;
5321 	default:
5322 		panic("unknown net intr type");
5323 	}
5324 	GROUPTASK_INIT(gtask, 0, fn, q);
5325 	if (irq_num != -1) {
5326 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5327 		if (err)
5328 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5329 	}
5330 	else {
5331 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5332 	}
5333 }
5334 
5335 void
5336 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5337 {
5338 	if (irq->ii_tag)
5339 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5340 
5341 	if (irq->ii_res)
5342 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5343 }
5344 
5345 static int
5346 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5347 {
5348 	iflib_txq_t txq = ctx->ifc_txqs;
5349 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5350 	if_irq_t irq = &ctx->ifc_legacy_irq;
5351 	iflib_filter_info_t info;
5352 	struct grouptask *gtask;
5353 	struct taskqgroup *tqg;
5354 	gtask_fn_t *fn;
5355 	int tqrid;
5356 	void *q;
5357 	int err;
5358 
5359 	q = &ctx->ifc_rxqs[0];
5360 	info = &rxq[0].ifr_filter_info;
5361 	gtask = &rxq[0].ifr_task;
5362 	tqg = qgroup_if_io_tqg;
5363 	tqrid = irq->ii_rid = *rid;
5364 	fn = _task_fn_rx;
5365 
5366 	ctx->ifc_flags |= IFC_LEGACY;
5367 	info->ifi_filter = filter;
5368 	info->ifi_filter_arg = filter_arg;
5369 	info->ifi_task = gtask;
5370 	info->ifi_ctx = ctx;
5371 
5372 	/* We allocate a single interrupt resource */
5373 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5374 		return (err);
5375 	GROUPTASK_INIT(gtask, 0, fn, q);
5376 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5377 
5378 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5379 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5380 	return (0);
5381 }
5382 
5383 void
5384 iflib_led_create(if_ctx_t ctx)
5385 {
5386 
5387 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5388 	    device_get_nameunit(ctx->ifc_dev));
5389 }
5390 
5391 void
5392 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5393 {
5394 
5395 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5396 }
5397 
5398 void
5399 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5400 {
5401 
5402 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5403 }
5404 
5405 void
5406 iflib_admin_intr_deferred(if_ctx_t ctx)
5407 {
5408 #ifdef INVARIANTS
5409 	struct grouptask *gtask;
5410 
5411 	gtask = &ctx->ifc_admin_task;
5412 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5413 #endif
5414 
5415 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5416 }
5417 
5418 void
5419 iflib_iov_intr_deferred(if_ctx_t ctx)
5420 {
5421 
5422 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5423 }
5424 
5425 void
5426 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5427 {
5428 
5429 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5430 }
5431 
5432 void
5433 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5434 	char *name)
5435 {
5436 
5437 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5438 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5439 }
5440 
5441 void
5442 iflib_config_gtask_deinit(struct grouptask *gtask)
5443 {
5444 
5445 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5446 }
5447 
5448 void
5449 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5450 {
5451 	if_t ifp = ctx->ifc_ifp;
5452 	iflib_txq_t txq = ctx->ifc_txqs;
5453 
5454 	if_setbaudrate(ifp, baudrate);
5455 	if (baudrate >= IF_Gbps(10)) {
5456 		STATE_LOCK(ctx);
5457 		ctx->ifc_flags |= IFC_PREFETCH;
5458 		STATE_UNLOCK(ctx);
5459 	}
5460 	/* If link down, disable watchdog */
5461 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5462 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5463 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5464 	}
5465 	ctx->ifc_link_state = link_state;
5466 	if_link_state_change(ifp, link_state);
5467 }
5468 
5469 static int
5470 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5471 {
5472 	int credits;
5473 #ifdef INVARIANTS
5474 	int credits_pre = txq->ift_cidx_processed;
5475 #endif
5476 
5477 	if (ctx->isc_txd_credits_update == NULL)
5478 		return (0);
5479 
5480 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5481 		return (0);
5482 
5483 	txq->ift_processed += credits;
5484 	txq->ift_cidx_processed += credits;
5485 
5486 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5487 	if (txq->ift_cidx_processed >= txq->ift_size)
5488 		txq->ift_cidx_processed -= txq->ift_size;
5489 	return (credits);
5490 }
5491 
5492 static int
5493 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5494 {
5495 
5496 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5497 	    budget));
5498 }
5499 
5500 void
5501 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5502 	const char *description, if_int_delay_info_t info,
5503 	int offset, int value)
5504 {
5505 	info->iidi_ctx = ctx;
5506 	info->iidi_offset = offset;
5507 	info->iidi_value = value;
5508 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5509 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5510 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5511 	    info, 0, iflib_sysctl_int_delay, "I", description);
5512 }
5513 
5514 struct mtx *
5515 iflib_ctx_lock_get(if_ctx_t ctx)
5516 {
5517 
5518 	return (&ctx->ifc_ctx_mtx);
5519 }
5520 
5521 static int
5522 iflib_msix_init(if_ctx_t ctx)
5523 {
5524 	device_t dev = ctx->ifc_dev;
5525 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5526 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5527 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5528 	int iflib_num_tx_queues, iflib_num_rx_queues;
5529 	int err, admincnt, bar;
5530 
5531 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5532 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5533 
5534 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5535 
5536 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5537 	admincnt = sctx->isc_admin_intrcnt;
5538 	/* Override by global tuneable */
5539 	{
5540 		int i;
5541 		size_t len = sizeof(i);
5542 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5543 		if (err == 0) {
5544 			if (i == 0)
5545 				goto msi;
5546 		}
5547 		else {
5548 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5549 		}
5550 	}
5551 	/* Override by tuneable */
5552 	if (scctx->isc_disable_msix)
5553 		goto msi;
5554 
5555 	/*
5556 	** When used in a virtualized environment
5557 	** PCI BUSMASTER capability may not be set
5558 	** so explicity set it here and rewrite
5559 	** the ENABLE in the MSIX control register
5560 	** at this point to cause the host to
5561 	** successfully initialize us.
5562 	*/
5563 	{
5564 		int msix_ctrl, rid;
5565 
5566  		pci_enable_busmaster(dev);
5567 		rid = 0;
5568 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5569 			rid += PCIR_MSIX_CTRL;
5570 			msix_ctrl = pci_read_config(dev, rid, 2);
5571 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5572 			pci_write_config(dev, rid, msix_ctrl, 2);
5573 		} else {
5574 			device_printf(dev, "PCIY_MSIX capability not found; "
5575 			                   "or rid %d == 0.\n", rid);
5576 			goto msi;
5577 		}
5578 	}
5579 
5580 	/*
5581 	 * bar == -1 => "trust me I know what I'm doing"
5582 	 * Some drivers are for hardware that is so shoddily
5583 	 * documented that no one knows which bars are which
5584 	 * so the developer has to map all bars. This hack
5585 	 * allows shoddy garbage to use msix in this framework.
5586 	 */
5587 	if (bar != -1) {
5588 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5589 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5590 		if (ctx->ifc_msix_mem == NULL) {
5591 			/* May not be enabled */
5592 			device_printf(dev, "Unable to map MSIX table \n");
5593 			goto msi;
5594 		}
5595 	}
5596 	/* First try MSI/X */
5597 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5598 		device_printf(dev, "System has MSIX disabled \n");
5599 		bus_release_resource(dev, SYS_RES_MEMORY,
5600 		    bar, ctx->ifc_msix_mem);
5601 		ctx->ifc_msix_mem = NULL;
5602 		goto msi;
5603 	}
5604 #if IFLIB_DEBUG
5605 	/* use only 1 qset in debug mode */
5606 	queuemsgs = min(msgs - admincnt, 1);
5607 #else
5608 	queuemsgs = msgs - admincnt;
5609 #endif
5610 #ifdef RSS
5611 	queues = imin(queuemsgs, rss_getnumbuckets());
5612 #else
5613 	queues = queuemsgs;
5614 #endif
5615 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5616 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5617 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5618 #ifdef  RSS
5619 	/* If we're doing RSS, clamp at the number of RSS buckets */
5620 	if (queues > rss_getnumbuckets())
5621 		queues = rss_getnumbuckets();
5622 #endif
5623 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5624 		rx_queues = iflib_num_rx_queues;
5625 	else
5626 		rx_queues = queues;
5627 
5628 	if (rx_queues > scctx->isc_nrxqsets)
5629 		rx_queues = scctx->isc_nrxqsets;
5630 
5631 	/*
5632 	 * We want this to be all logical CPUs by default
5633 	 */
5634 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5635 		tx_queues = iflib_num_tx_queues;
5636 	else
5637 		tx_queues = mp_ncpus;
5638 
5639 	if (tx_queues > scctx->isc_ntxqsets)
5640 		tx_queues = scctx->isc_ntxqsets;
5641 
5642 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5643 #ifdef INVARIANTS
5644 		if (tx_queues != rx_queues)
5645 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5646 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5647 #endif
5648 		tx_queues = min(rx_queues, tx_queues);
5649 		rx_queues = min(rx_queues, tx_queues);
5650 	}
5651 
5652 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5653 
5654 	vectors = rx_queues + admincnt;
5655 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5656 		device_printf(dev,
5657 					  "Using MSIX interrupts with %d vectors\n", vectors);
5658 		scctx->isc_vectors = vectors;
5659 		scctx->isc_nrxqsets = rx_queues;
5660 		scctx->isc_ntxqsets = tx_queues;
5661 		scctx->isc_intr = IFLIB_INTR_MSIX;
5662 
5663 		return (vectors);
5664 	} else {
5665 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5666 	}
5667 msi:
5668 	vectors = pci_msi_count(dev);
5669 	scctx->isc_nrxqsets = 1;
5670 	scctx->isc_ntxqsets = 1;
5671 	scctx->isc_vectors = vectors;
5672 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5673 		device_printf(dev,"Using an MSI interrupt\n");
5674 		scctx->isc_intr = IFLIB_INTR_MSI;
5675 	} else {
5676 		device_printf(dev,"Using a Legacy interrupt\n");
5677 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5678 	}
5679 
5680 	return (vectors);
5681 }
5682 
5683 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5684 
5685 static int
5686 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5687 {
5688 	int rc;
5689 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5690 	struct sbuf *sb;
5691 	char *ring_state = "UNKNOWN";
5692 
5693 	/* XXX needed ? */
5694 	rc = sysctl_wire_old_buffer(req, 0);
5695 	MPASS(rc == 0);
5696 	if (rc != 0)
5697 		return (rc);
5698 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5699 	MPASS(sb != NULL);
5700 	if (sb == NULL)
5701 		return (ENOMEM);
5702 	if (state[3] <= 3)
5703 		ring_state = ring_states[state[3]];
5704 
5705 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5706 		    state[0], state[1], state[2], ring_state);
5707 	rc = sbuf_finish(sb);
5708 	sbuf_delete(sb);
5709         return(rc);
5710 }
5711 
5712 enum iflib_ndesc_handler {
5713 	IFLIB_NTXD_HANDLER,
5714 	IFLIB_NRXD_HANDLER,
5715 };
5716 
5717 static int
5718 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5719 {
5720 	if_ctx_t ctx = (void *)arg1;
5721 	enum iflib_ndesc_handler type = arg2;
5722 	char buf[256] = {0};
5723 	qidx_t *ndesc;
5724 	char *p, *next;
5725 	int nqs, rc, i;
5726 
5727 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5728 
5729 	nqs = 8;
5730 	switch(type) {
5731 	case IFLIB_NTXD_HANDLER:
5732 		ndesc = ctx->ifc_sysctl_ntxds;
5733 		if (ctx->ifc_sctx)
5734 			nqs = ctx->ifc_sctx->isc_ntxqs;
5735 		break;
5736 	case IFLIB_NRXD_HANDLER:
5737 		ndesc = ctx->ifc_sysctl_nrxds;
5738 		if (ctx->ifc_sctx)
5739 			nqs = ctx->ifc_sctx->isc_nrxqs;
5740 		break;
5741 	}
5742 	if (nqs == 0)
5743 		nqs = 8;
5744 
5745 	for (i=0; i<8; i++) {
5746 		if (i >= nqs)
5747 			break;
5748 		if (i)
5749 			strcat(buf, ",");
5750 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5751 	}
5752 
5753 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5754 	if (rc || req->newptr == NULL)
5755 		return rc;
5756 
5757 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5758 	    i++, p = strsep(&next, " ,")) {
5759 		ndesc[i] = strtoul(p, NULL, 10);
5760 	}
5761 
5762 	return(rc);
5763 }
5764 
5765 #define NAME_BUFLEN 32
5766 static void
5767 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5768 {
5769         device_t dev = iflib_get_dev(ctx);
5770 	struct sysctl_oid_list *child, *oid_list;
5771 	struct sysctl_ctx_list *ctx_list;
5772 	struct sysctl_oid *node;
5773 
5774 	ctx_list = device_get_sysctl_ctx(dev);
5775 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5776 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5777 						      CTLFLAG_RD, NULL, "IFLIB fields");
5778 	oid_list = SYSCTL_CHILDREN(node);
5779 
5780 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5781 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5782 		       "driver version");
5783 
5784 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5785 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5786 			"# of txqs to use, 0 => use default #");
5787 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5788 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5789 			"# of rxqs to use, 0 => use default #");
5790 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5791 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5792                        "permit #txq != #rxq");
5793 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5794                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5795                       "disable MSIX (default 0)");
5796 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5797 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5798                        "set the rx budget");
5799 
5800 	/* XXX change for per-queue sizes */
5801 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5802 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5803                        mp_ndesc_handler, "A",
5804                        "list of # of tx descriptors to use, 0 = use default #");
5805 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5806 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5807                        mp_ndesc_handler, "A",
5808                        "list of # of rx descriptors to use, 0 = use default #");
5809 }
5810 
5811 static void
5812 iflib_add_device_sysctl_post(if_ctx_t ctx)
5813 {
5814 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5815 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5816         device_t dev = iflib_get_dev(ctx);
5817 	struct sysctl_oid_list *child;
5818 	struct sysctl_ctx_list *ctx_list;
5819 	iflib_fl_t fl;
5820 	iflib_txq_t txq;
5821 	iflib_rxq_t rxq;
5822 	int i, j;
5823 	char namebuf[NAME_BUFLEN];
5824 	char *qfmt;
5825 	struct sysctl_oid *queue_node, *fl_node, *node;
5826 	struct sysctl_oid_list *queue_list, *fl_list;
5827 	ctx_list = device_get_sysctl_ctx(dev);
5828 
5829 	node = ctx->ifc_sysctl_node;
5830 	child = SYSCTL_CHILDREN(node);
5831 
5832 	if (scctx->isc_ntxqsets > 100)
5833 		qfmt = "txq%03d";
5834 	else if (scctx->isc_ntxqsets > 10)
5835 		qfmt = "txq%02d";
5836 	else
5837 		qfmt = "txq%d";
5838 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5839 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5840 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5841 					     CTLFLAG_RD, NULL, "Queue Name");
5842 		queue_list = SYSCTL_CHILDREN(queue_node);
5843 #if MEMORY_LOGGING
5844 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5845 				CTLFLAG_RD,
5846 				&txq->ift_dequeued, "total mbufs freed");
5847 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5848 				CTLFLAG_RD,
5849 				&txq->ift_enqueued, "total mbufs enqueued");
5850 #endif
5851 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5852 				   CTLFLAG_RD,
5853 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5854 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5855 				   CTLFLAG_RD,
5856 				   &txq->ift_pullups, "# of times m_pullup was called");
5857 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5858 				   CTLFLAG_RD,
5859 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5860 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5861 				   CTLFLAG_RD,
5862 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5863 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5864 				   CTLFLAG_RD,
5865 				   &txq->ift_map_failed, "# of times dma map failed");
5866 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5867 				   CTLFLAG_RD,
5868 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5869 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5870 				   CTLFLAG_RD,
5871 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5872 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5873 				   CTLFLAG_RD,
5874 				   &txq->ift_pidx, 1, "Producer Index");
5875 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5876 				   CTLFLAG_RD,
5877 				   &txq->ift_cidx, 1, "Consumer Index");
5878 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5879 				   CTLFLAG_RD,
5880 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5881 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5882 				   CTLFLAG_RD,
5883 				   &txq->ift_in_use, 1, "descriptors in use");
5884 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5885 				   CTLFLAG_RD,
5886 				   &txq->ift_processed, "descriptors procesed for clean");
5887 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5888 				   CTLFLAG_RD,
5889 				   &txq->ift_cleaned, "total cleaned");
5890 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5891 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5892 				0, mp_ring_state_handler, "A", "soft ring state");
5893 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5894 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5895 				       "# of enqueues to the mp_ring for this queue");
5896 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5897 				       CTLFLAG_RD, &txq->ift_br->drops,
5898 				       "# of drops in the mp_ring for this queue");
5899 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5900 				       CTLFLAG_RD, &txq->ift_br->starts,
5901 				       "# of normal consumer starts in the mp_ring for this queue");
5902 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5903 				       CTLFLAG_RD, &txq->ift_br->stalls,
5904 					       "# of consumer stalls in the mp_ring for this queue");
5905 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5906 			       CTLFLAG_RD, &txq->ift_br->restarts,
5907 				       "# of consumer restarts in the mp_ring for this queue");
5908 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5909 				       CTLFLAG_RD, &txq->ift_br->abdications,
5910 				       "# of consumer abdications in the mp_ring for this queue");
5911 	}
5912 
5913 	if (scctx->isc_nrxqsets > 100)
5914 		qfmt = "rxq%03d";
5915 	else if (scctx->isc_nrxqsets > 10)
5916 		qfmt = "rxq%02d";
5917 	else
5918 		qfmt = "rxq%d";
5919 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5920 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5921 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5922 					     CTLFLAG_RD, NULL, "Queue Name");
5923 		queue_list = SYSCTL_CHILDREN(queue_node);
5924 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5925 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5926 				       CTLFLAG_RD,
5927 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5928 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5929 				       CTLFLAG_RD,
5930 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5931 		}
5932 
5933 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5934 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5935 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5936 						     CTLFLAG_RD, NULL, "freelist Name");
5937 			fl_list = SYSCTL_CHILDREN(fl_node);
5938 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5939 				       CTLFLAG_RD,
5940 				       &fl->ifl_pidx, 1, "Producer Index");
5941 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5942 				       CTLFLAG_RD,
5943 				       &fl->ifl_cidx, 1, "Consumer Index");
5944 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5945 				       CTLFLAG_RD,
5946 				       &fl->ifl_credits, 1, "credits available");
5947 #if MEMORY_LOGGING
5948 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5949 					CTLFLAG_RD,
5950 					&fl->ifl_m_enqueued, "mbufs allocated");
5951 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5952 					CTLFLAG_RD,
5953 					&fl->ifl_m_dequeued, "mbufs freed");
5954 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5955 					CTLFLAG_RD,
5956 					&fl->ifl_cl_enqueued, "clusters allocated");
5957 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5958 					CTLFLAG_RD,
5959 					&fl->ifl_cl_dequeued, "clusters freed");
5960 #endif
5961 
5962 		}
5963 	}
5964 
5965 }
5966 
5967 #ifndef __NO_STRICT_ALIGNMENT
5968 static struct mbuf *
5969 iflib_fixup_rx(struct mbuf *m)
5970 {
5971 	struct mbuf *n;
5972 
5973 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5974 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5975 		m->m_data += ETHER_HDR_LEN;
5976 		n = m;
5977 	} else {
5978 		MGETHDR(n, M_NOWAIT, MT_DATA);
5979 		if (n == NULL) {
5980 			m_freem(m);
5981 			return (NULL);
5982 		}
5983 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5984 		m->m_data += ETHER_HDR_LEN;
5985 		m->m_len -= ETHER_HDR_LEN;
5986 		n->m_len = ETHER_HDR_LEN;
5987 		M_MOVE_PKTHDR(n, m);
5988 		n->m_next = m;
5989 	}
5990 	return (n);
5991 }
5992 #endif
5993