xref: /freebsd/sys/net/iflib.c (revision 7f9dff23d3092aa33ad45b2b63e52469b3c13a6e)
1 /*-
2  * Copyright (c) 2014-2016, Matthew Macy <mmacy@nextbsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of iflib's software descriptor references
98  */
99 #define MEMORY_LOGGING 0
100 /*
101  * Enable mbuf vectors for compressing long mbuf chains
102  */
103 
104 /*
105  * NB:
106  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
107  *   we prefetch needs to be determined by the time spent in m_free vis a vis
108  *   the cost of a prefetch. This will of course vary based on the workload:
109  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
110  *        is quite expensive, thus suggesting very little prefetch.
111  *      - small packet forwarding which is just returning a single mbuf to
112  *        UMA will typically be very fast vis a vis the cost of a memory
113  *        access.
114  */
115 
116 
117 /*
118  * File organization:
119  *  - private structures
120  *  - iflib private utility functions
121  *  - ifnet functions
122  *  - vlan registry and other exported functions
123  *  - iflib public core functions
124  *
125  *
126  */
127 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
128 
129 struct iflib_txq;
130 typedef struct iflib_txq *iflib_txq_t;
131 struct iflib_rxq;
132 typedef struct iflib_rxq *iflib_rxq_t;
133 struct iflib_fl;
134 typedef struct iflib_fl *iflib_fl_t;
135 
136 typedef struct iflib_filter_info {
137 	driver_filter_t *ifi_filter;
138 	void *ifi_filter_arg;
139 	struct grouptask *ifi_task;
140 } *iflib_filter_info_t;
141 
142 struct iflib_ctx {
143 	KOBJ_FIELDS;
144    /*
145    * Pointer to hardware driver's softc
146    */
147 	void *ifc_softc;
148 	device_t ifc_dev;
149 	if_t ifc_ifp;
150 
151 	cpuset_t ifc_cpus;
152 	if_shared_ctx_t ifc_sctx;
153 	struct if_softc_ctx ifc_softc_ctx;
154 
155 	struct mtx ifc_mtx;
156 
157 	uint16_t ifc_nhwtxqs;
158 	uint16_t ifc_nhwrxqs;
159 
160 	iflib_txq_t ifc_txqs;
161 	iflib_rxq_t ifc_rxqs;
162 	uint32_t ifc_if_flags;
163 	uint32_t ifc_flags;
164 	uint32_t ifc_max_fl_buf_size;
165 	int ifc_in_detach;
166 
167 	int ifc_link_state;
168 	int ifc_link_irq;
169 	int ifc_pause_frames;
170 	int ifc_watchdog_events;
171 	struct cdev *ifc_led_dev;
172 	struct resource *ifc_msix_mem;
173 
174 	struct if_irq ifc_legacy_irq;
175 	struct grouptask ifc_admin_task;
176 	struct grouptask ifc_vflr_task;
177 	struct iflib_filter_info ifc_filter_info;
178 	struct ifmedia	ifc_media;
179 
180 	struct sysctl_oid *ifc_sysctl_node;
181 	uint16_t ifc_sysctl_ntxqs;
182 	uint16_t ifc_sysctl_nrxqs;
183 	uint16_t ifc_sysctl_qs_eq_override;
184 
185 	uint16_t ifc_sysctl_ntxds[8];
186 	uint16_t ifc_sysctl_nrxds[8];
187 	struct if_txrx ifc_txrx;
188 #define isc_txd_encap  ifc_txrx.ift_txd_encap
189 #define isc_txd_flush  ifc_txrx.ift_txd_flush
190 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
191 #define isc_rxd_available ifc_txrx.ift_rxd_available
192 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
193 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
194 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
195 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
197 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
198 	eventhandler_tag ifc_vlan_attach_event;
199 	eventhandler_tag ifc_vlan_detach_event;
200 	uint8_t ifc_mac[ETHER_ADDR_LEN];
201 	char ifc_mtx_name[16];
202 };
203 
204 
205 void *
206 iflib_get_softc(if_ctx_t ctx)
207 {
208 
209 	return (ctx->ifc_softc);
210 }
211 
212 device_t
213 iflib_get_dev(if_ctx_t ctx)
214 {
215 
216 	return (ctx->ifc_dev);
217 }
218 
219 if_t
220 iflib_get_ifp(if_ctx_t ctx)
221 {
222 
223 	return (ctx->ifc_ifp);
224 }
225 
226 struct ifmedia *
227 iflib_get_media(if_ctx_t ctx)
228 {
229 
230 	return (&ctx->ifc_media);
231 }
232 
233 void
234 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
235 {
236 
237 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
238 }
239 
240 if_softc_ctx_t
241 iflib_get_softc_ctx(if_ctx_t ctx)
242 {
243 
244 	return (&ctx->ifc_softc_ctx);
245 }
246 
247 if_shared_ctx_t
248 iflib_get_sctx(if_ctx_t ctx)
249 {
250 
251 	return (ctx->ifc_sctx);
252 }
253 
254 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
255 
256 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
257 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
258 
259 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
260 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
261 #define RX_SW_DESC_INUSE        (1 << 3)
262 #define TX_SW_DESC_MAPPED       (1 << 4)
263 
264 typedef struct iflib_sw_rx_desc {
265 	bus_dmamap_t    ifsd_map;         /* bus_dma map for packet */
266 	struct mbuf    *ifsd_m;           /* rx: uninitialized mbuf */
267 	caddr_t         ifsd_cl;          /* direct cluster pointer for rx */
268 	uint16_t	ifsd_flags;
269 } *iflib_rxsd_t;
270 
271 typedef struct iflib_sw_tx_desc_val {
272 	bus_dmamap_t    ifsd_map;         /* bus_dma map for packet */
273 	struct mbuf    *ifsd_m;           /* pkthdr mbuf */
274 	uint8_t		ifsd_flags;
275 } *iflib_txsd_val_t;
276 
277 typedef struct iflib_sw_tx_desc_array {
278 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
280 	uint8_t		*ifsd_flags;
281 } iflib_txsd_array_t;
282 
283 
284 /* magic number that should be high enough for any hardware */
285 #define IFLIB_MAX_TX_SEGS		128
286 #define IFLIB_MAX_RX_SEGS		32
287 #define IFLIB_RX_COPY_THRESH		128
288 #define IFLIB_MAX_RX_REFRESH		32
289 #define IFLIB_QUEUE_IDLE		0
290 #define IFLIB_QUEUE_HUNG		1
291 #define IFLIB_QUEUE_WORKING		2
292 
293 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */
294 #define TX_BATCH_SIZE			16
295 
296 #define IFLIB_RESTART_BUDGET		8
297 
298 #define	IFC_LEGACY		0x01
299 #define	IFC_QFLUSH		0x02
300 #define	IFC_MULTISEG		0x04
301 #define	IFC_DMAR		0x08
302 #define	IFC_SC_ALLOCATED	0x10
303 
304 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
305 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
306 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
307 struct iflib_txq {
308 	uint16_t	ift_in_use;
309 	uint16_t	ift_cidx;
310 	uint16_t	ift_cidx_processed;
311 	uint16_t	ift_pidx;
312 	uint8_t		ift_gen;
313 	uint8_t		ift_db_pending;
314 	uint8_t		ift_db_pending_queued;
315 	uint8_t		ift_npending;
316 	uint8_t		ift_br_offset;
317 	/* implicit pad */
318 	uint64_t	ift_processed;
319 	uint64_t	ift_cleaned;
320 #if MEMORY_LOGGING
321 	uint64_t	ift_enqueued;
322 	uint64_t	ift_dequeued;
323 #endif
324 	uint64_t	ift_no_tx_dma_setup;
325 	uint64_t	ift_no_desc_avail;
326 	uint64_t	ift_mbuf_defrag_failed;
327 	uint64_t	ift_mbuf_defrag;
328 	uint64_t	ift_map_failed;
329 	uint64_t	ift_txd_encap_efbig;
330 	uint64_t	ift_pullups;
331 
332 	struct mtx	ift_mtx;
333 	struct mtx	ift_db_mtx;
334 
335 	/* constant values */
336 	if_ctx_t	ift_ctx;
337 	struct ifmp_ring        **ift_br;
338 	struct grouptask	ift_task;
339 	uint16_t	ift_size;
340 	uint16_t	ift_id;
341 	struct callout	ift_timer;
342 	struct callout	ift_db_check;
343 
344 	iflib_txsd_array_t	ift_sds;
345 	uint8_t			ift_nbr;
346 	uint8_t			ift_qstatus;
347 	uint8_t			ift_active;
348 	uint8_t			ift_closed;
349 	int			ift_watchdog_time;
350 	struct iflib_filter_info ift_filter_info;
351 	bus_dma_tag_t		ift_desc_tag;
352 	bus_dma_tag_t		ift_tso_desc_tag;
353 	iflib_dma_info_t	ift_ifdi;
354 #define MTX_NAME_LEN 16
355 	char                    ift_mtx_name[MTX_NAME_LEN];
356 	char                    ift_db_mtx_name[MTX_NAME_LEN];
357 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
358 } __aligned(CACHE_LINE_SIZE);
359 
360 struct iflib_fl {
361 	uint16_t	ifl_cidx;
362 	uint16_t	ifl_pidx;
363 	uint16_t	ifl_credits;
364 	uint8_t		ifl_gen;
365 #if MEMORY_LOGGING
366 	uint64_t	ifl_m_enqueued;
367 	uint64_t	ifl_m_dequeued;
368 	uint64_t	ifl_cl_enqueued;
369 	uint64_t	ifl_cl_dequeued;
370 #endif
371 	/* implicit pad */
372 
373 	/* constant */
374 	uint16_t	ifl_size;
375 	uint16_t	ifl_buf_size;
376 	uint16_t	ifl_cltype;
377 	uma_zone_t	ifl_zone;
378 	iflib_rxsd_t	ifl_sds;
379 	iflib_rxq_t	ifl_rxq;
380 	uint8_t		ifl_id;
381 	bus_dma_tag_t           ifl_desc_tag;
382 	iflib_dma_info_t	ifl_ifdi;
383 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
384 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
385 }  __aligned(CACHE_LINE_SIZE);
386 
387 static inline int
388 get_inuse(int size, int cidx, int pidx, int gen)
389 {
390 	int used;
391 
392 	if (pidx > cidx)
393 		used = pidx - cidx;
394 	else if (pidx < cidx)
395 		used = size - cidx + pidx;
396 	else if (gen == 0 && pidx == cidx)
397 		used = 0;
398 	else if (gen == 1 && pidx == cidx)
399 		used = size;
400 	else
401 		panic("bad state");
402 
403 	return (used);
404 }
405 
406 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
407 
408 #define IDXDIFF(head, tail, wrap) \
409 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
410 
411 struct iflib_rxq {
412 	/* If there is a separate completion queue -
413 	 * these are the cq cidx and pidx. Otherwise
414 	 * these are unused.
415 	 */
416 	uint16_t	ifr_size;
417 	uint16_t	ifr_cq_cidx;
418 	uint16_t	ifr_cq_pidx;
419 	uint8_t		ifr_cq_gen;
420 	uint8_t		ifr_fl_offset;
421 
422 	if_ctx_t	ifr_ctx;
423 	iflib_fl_t	ifr_fl;
424 	uint64_t	ifr_rx_irq;
425 	uint16_t	ifr_id;
426 	uint8_t		ifr_lro_enabled;
427 	uint8_t		ifr_nfl;
428 	struct lro_ctrl			ifr_lc;
429 	struct grouptask        ifr_task;
430 	struct iflib_filter_info ifr_filter_info;
431 	iflib_dma_info_t		ifr_ifdi;
432 	/* dynamically allocate if any drivers need a value substantially larger than this */
433 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
434 }  __aligned(CACHE_LINE_SIZE);
435 
436 /*
437  * Only allow a single packet to take up most 1/nth of the tx ring
438  */
439 #define MAX_SINGLE_PACKET_FRACTION 12
440 #define IF_BAD_DMA (bus_addr_t)-1
441 
442 static int enable_msix = 1;
443 
444 #define mtx_held(m)	(((m)->mtx_lock & ~MTX_FLAGMASK) != (uintptr_t)0)
445 
446 
447 
448 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
449 
450 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
451 
452 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
453 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
454 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
455 
456 
457 #define TXDB_LOCK_INIT(txq)  mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF)
458 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx)
459 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx)
460 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx)
461 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx)
462 
463 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
464 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
465 
466 
467 /* Our boot-time initialization hook */
468 static int	iflib_module_event_handler(module_t, int, void *);
469 
470 static moduledata_t iflib_moduledata = {
471 	"iflib",
472 	iflib_module_event_handler,
473 	NULL
474 };
475 
476 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
477 MODULE_VERSION(iflib, 1);
478 
479 MODULE_DEPEND(iflib, pci, 1, 1, 1);
480 MODULE_DEPEND(iflib, ether, 1, 1, 1);
481 
482 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
483 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
484 
485 #ifndef IFLIB_DEBUG_COUNTERS
486 #ifdef INVARIANTS
487 #define IFLIB_DEBUG_COUNTERS 1
488 #else
489 #define IFLIB_DEBUG_COUNTERS 0
490 #endif /* !INVARIANTS */
491 #endif
492 
493 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
494                    "iflib driver parameters");
495 
496 /*
497  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
498  */
499 static int iflib_min_tx_latency = 0;
500 
501 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
502 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
503 
504 
505 #if IFLIB_DEBUG_COUNTERS
506 
507 static int iflib_tx_seen;
508 static int iflib_tx_sent;
509 static int iflib_tx_encap;
510 static int iflib_rx_allocs;
511 static int iflib_fl_refills;
512 static int iflib_fl_refills_large;
513 static int iflib_tx_frees;
514 
515 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
516 		   &iflib_tx_seen, 0, "# tx mbufs seen");
517 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
518 		   &iflib_tx_sent, 0, "# tx mbufs sent");
519 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
520 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
521 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
522 		   &iflib_tx_frees, 0, "# tx frees");
523 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
524 		   &iflib_rx_allocs, 0, "# rx allocations");
525 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
526 		   &iflib_fl_refills, 0, "# refills");
527 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
528 		   &iflib_fl_refills_large, 0, "# large refills");
529 
530 
531 static int iflib_txq_drain_flushing;
532 static int iflib_txq_drain_oactive;
533 static int iflib_txq_drain_notready;
534 static int iflib_txq_drain_encapfail;
535 
536 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
537 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
538 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
539 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
540 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
541 		   &iflib_txq_drain_notready, 0, "# drain notready");
542 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
543 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
544 
545 
546 static int iflib_encap_load_mbuf_fail;
547 static int iflib_encap_txq_avail_fail;
548 static int iflib_encap_txd_encap_fail;
549 
550 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
551 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
552 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
553 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
554 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
555 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
556 
557 static int iflib_task_fn_rxs;
558 static int iflib_rx_intr_enables;
559 static int iflib_fast_intrs;
560 static int iflib_intr_link;
561 static int iflib_intr_msix;
562 static int iflib_rx_unavail;
563 static int iflib_rx_ctx_inactive;
564 static int iflib_rx_zero_len;
565 static int iflib_rx_if_input;
566 static int iflib_rx_mbuf_null;
567 static int iflib_rxd_flush;
568 
569 static int iflib_verbose_debug;
570 
571 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
572 		   &iflib_intr_link, 0, "# intr link calls");
573 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
574 		   &iflib_intr_msix, 0, "# intr msix calls");
575 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
576 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
577 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
578 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
579 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
580 		   &iflib_fast_intrs, 0, "# fast_intr calls");
581 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
582 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
583 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
584 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
585 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
586 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
587 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
588 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
589 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
590 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
591 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
592 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
593 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
594 		   &iflib_verbose_debug, 0, "enable verbose debugging");
595 
596 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
597 static void
598 iflib_debug_reset(void)
599 {
600 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
601 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
602 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
603 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
604 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
605 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
606 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
607 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
608 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
609 }
610 
611 #else
612 #define DBG_COUNTER_INC(name)
613 static void iflib_debug_reset(void) {}
614 #endif
615 
616 
617 
618 #define IFLIB_DEBUG 0
619 
620 static void iflib_tx_structures_free(if_ctx_t ctx);
621 static void iflib_rx_structures_free(if_ctx_t ctx);
622 static int iflib_queues_alloc(if_ctx_t ctx);
623 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
624 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget);
625 static int iflib_qset_structures_setup(if_ctx_t ctx);
626 static int iflib_msix_init(if_ctx_t ctx);
627 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
628 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
629 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
630 static int iflib_register(if_ctx_t);
631 static void iflib_init_locked(if_ctx_t ctx);
632 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
633 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
634 static void iflib_ifmp_purge(iflib_txq_t txq);
635 
636 
637 #ifdef DEV_NETMAP
638 #include <sys/selinfo.h>
639 #include <net/netmap.h>
640 #include <dev/netmap/netmap_kern.h>
641 
642 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
643 
644 /*
645  * device-specific sysctl variables:
646  *
647  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
648  *	During regular operations the CRC is stripped, but on some
649  *	hardware reception of frames not multiple of 64 is slower,
650  *	so using crcstrip=0 helps in benchmarks.
651  *
652  * iflib_rx_miss, iflib_rx_miss_bufs:
653  *	count packets that might be missed due to lost interrupts.
654  */
655 SYSCTL_DECL(_dev_netmap);
656 /*
657  * The xl driver by default strips CRCs and we do not override it.
658  */
659 
660 int iflib_crcstrip = 1;
661 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
662     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
663 
664 int iflib_rx_miss, iflib_rx_miss_bufs;
665 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
666     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
667 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
668     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
669 
670 /*
671  * Register/unregister. We are already under netmap lock.
672  * Only called on the first register or the last unregister.
673  */
674 static int
675 iflib_netmap_register(struct netmap_adapter *na, int onoff)
676 {
677 	struct ifnet *ifp = na->ifp;
678 	if_ctx_t ctx = ifp->if_softc;
679 
680 	CTX_LOCK(ctx);
681 	IFDI_INTR_DISABLE(ctx);
682 
683 	/* Tell the stack that the interface is no longer active */
684 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
685 
686 	if (!CTX_IS_VF(ctx))
687 		IFDI_CRCSTRIP_SET(ctx, onoff);
688 
689 	/* enable or disable flags and callbacks in na and ifp */
690 	if (onoff) {
691 		nm_set_native_flags(na);
692 	} else {
693 		nm_clear_native_flags(na);
694 	}
695 	IFDI_INIT(ctx);
696 	IFDI_CRCSTRIP_SET(ctx, onoff); // XXX why twice ?
697 	CTX_UNLOCK(ctx);
698 	return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
699 }
700 
701 /*
702  * Reconcile kernel and user view of the transmit ring.
703  *
704  * All information is in the kring.
705  * Userspace wants to send packets up to the one before kring->rhead,
706  * kernel knows kring->nr_hwcur is the first unsent packet.
707  *
708  * Here we push packets out (as many as possible), and possibly
709  * reclaim buffers from previously completed transmission.
710  *
711  * The caller (netmap) guarantees that there is only one instance
712  * running at any time. Any interference with other driver
713  * methods should be handled by the individual drivers.
714  */
715 static int
716 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
717 {
718 	struct netmap_adapter *na = kring->na;
719 	struct ifnet *ifp = na->ifp;
720 	struct netmap_ring *ring = kring->ring;
721 	u_int nm_i;	/* index into the netmap ring */
722 	u_int nic_i;	/* index into the NIC ring */
723 	u_int n;
724 	u_int const lim = kring->nkr_num_slots - 1;
725 	u_int const head = kring->rhead;
726 	struct if_pkt_info pi;
727 
728 	/*
729 	 * interrupts on every tx packet are expensive so request
730 	 * them every half ring, or where NS_REPORT is set
731 	 */
732 	u_int report_frequency = kring->nkr_num_slots >> 1;
733 	/* device-specific */
734 	if_ctx_t ctx = ifp->if_softc;
735 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
736 
737 	pi.ipi_segs = txq->ift_segs;
738 	pi.ipi_qsidx = kring->ring_id;
739 	pi.ipi_ndescs = 0;
740 
741 	bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
742 					BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
743 
744 
745 	/*
746 	 * First part: process new packets to send.
747 	 * nm_i is the current index in the netmap ring,
748 	 * nic_i is the corresponding index in the NIC ring.
749 	 *
750 	 * If we have packets to send (nm_i != head)
751 	 * iterate over the netmap ring, fetch length and update
752 	 * the corresponding slot in the NIC ring. Some drivers also
753 	 * need to update the buffer's physical address in the NIC slot
754 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
755 	 *
756 	 * The netmap_reload_map() calls is especially expensive,
757 	 * even when (as in this case) the tag is 0, so do only
758 	 * when the buffer has actually changed.
759 	 *
760 	 * If possible do not set the report/intr bit on all slots,
761 	 * but only a few times per ring or when NS_REPORT is set.
762 	 *
763 	 * Finally, on 10G and faster drivers, it might be useful
764 	 * to prefetch the next slot and txr entry.
765 	 */
766 
767 	nm_i = kring->nr_hwcur;
768 	if (nm_i != head) {	/* we have new packets to send */
769 		nic_i = netmap_idx_k2n(kring, nm_i);
770 
771 		__builtin_prefetch(&ring->slot[nm_i]);
772 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
773 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
774 
775 		for (n = 0; nm_i != head; n++) {
776 			struct netmap_slot *slot = &ring->slot[nm_i];
777 			u_int len = slot->len;
778 			uint64_t paddr;
779 			void *addr = PNMB(na, slot, &paddr);
780 			int flags = (slot->flags & NS_REPORT ||
781 				nic_i == 0 || nic_i == report_frequency) ?
782 				IPI_TX_INTR : 0;
783 
784 			/* device-specific */
785 			pi.ipi_pidx = nic_i;
786 			pi.ipi_flags = flags;
787 
788 			/* Fill the slot in the NIC ring. */
789 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
790 
791 			/* prefetch for next round */
792 			__builtin_prefetch(&ring->slot[nm_i + 1]);
793 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
794 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
795 
796 			NM_CHECK_ADDR_LEN(na, addr, len);
797 
798 			if (slot->flags & NS_BUF_CHANGED) {
799 				/* buffer has changed, reload map */
800 				netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
801 			}
802 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
803 
804 			/* make sure changes to the buffer are synced */
805 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
806 							BUS_DMASYNC_PREWRITE);
807 
808 			nm_i = nm_next(nm_i, lim);
809 			nic_i = nm_next(nic_i, lim);
810 		}
811 		kring->nr_hwcur = head;
812 
813 		/* synchronize the NIC ring */
814 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
815 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
816 
817 		/* (re)start the tx unit up to slot nic_i (excluded) */
818 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
819 	}
820 
821 	/*
822 	 * Second part: reclaim buffers for completed transmissions.
823 	 */
824 	if (iflib_tx_credits_update(ctx, txq)) {
825 		/* some tx completed, increment avail */
826 		nic_i = txq->ift_cidx_processed;
827 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
828 	}
829 	return (0);
830 }
831 
832 /*
833  * Reconcile kernel and user view of the receive ring.
834  * Same as for the txsync, this routine must be efficient.
835  * The caller guarantees a single invocations, but races against
836  * the rest of the driver should be handled here.
837  *
838  * On call, kring->rhead is the first packet that userspace wants
839  * to keep, and kring->rcur is the wakeup point.
840  * The kernel has previously reported packets up to kring->rtail.
841  *
842  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
843  * of whether or not we received an interrupt.
844  */
845 static int
846 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
847 {
848 	struct netmap_adapter *na = kring->na;
849 	struct ifnet *ifp = na->ifp;
850 	struct netmap_ring *ring = kring->ring;
851 	u_int nm_i;	/* index into the netmap ring */
852 	u_int nic_i;	/* index into the NIC ring */
853 	u_int i, n;
854 	u_int const lim = kring->nkr_num_slots - 1;
855 	u_int const head = kring->rhead;
856 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
857 	struct if_rxd_info ri;
858 	/* device-specific */
859 	if_ctx_t ctx = ifp->if_softc;
860 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
861 	iflib_fl_t fl = rxq->ifr_fl;
862 	if (head > lim)
863 		return netmap_ring_reinit(kring);
864 
865 	bzero(&ri, sizeof(ri));
866 	ri.iri_qsidx = kring->ring_id;
867 	ri.iri_ifp = ctx->ifc_ifp;
868 	/* XXX check sync modes */
869 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++)
870 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
871 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
872 
873 	/*
874 	 * First part: import newly received packets.
875 	 *
876 	 * nm_i is the index of the next free slot in the netmap ring,
877 	 * nic_i is the index of the next received packet in the NIC ring,
878 	 * and they may differ in case if_init() has been called while
879 	 * in netmap mode. For the receive ring we have
880 	 *
881 	 *	nic_i = rxr->next_check;
882 	 *	nm_i = kring->nr_hwtail (previous)
883 	 * and
884 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
885 	 *
886 	 * rxr->next_check is set to 0 on a ring reinit
887 	 */
888 	if (netmap_no_pendintr || force_update) {
889 		int crclen = iflib_crcstrip ? 0 : 4;
890 		int error, avail;
891 		uint16_t slot_flags = kring->nkr_slot_flags;
892 
893 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
894 			nic_i = fl->ifl_cidx;
895 			nm_i = netmap_idx_n2k(kring, nic_i);
896 			avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX);
897 			for (n = 0; avail > 0; n++, avail--) {
898 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
899 				if (error)
900 					ring->slot[nm_i].len = 0;
901 				else
902 					ring->slot[nm_i].len = ri.iri_len - crclen;
903 				ring->slot[nm_i].flags = slot_flags;
904 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
905 								fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_POSTREAD);
906 				nm_i = nm_next(nm_i, lim);
907 				nic_i = nm_next(nic_i, lim);
908 			}
909 			if (n) { /* update the state variables */
910 				if (netmap_no_pendintr && !force_update) {
911 					/* diagnostics */
912 					iflib_rx_miss ++;
913 					iflib_rx_miss_bufs += n;
914 				}
915 				fl->ifl_cidx = nic_i;
916 				kring->nr_hwtail = nm_i;
917 			}
918 			kring->nr_kflags &= ~NKR_PENDINTR;
919 		}
920 	}
921 	/*
922 	 * Second part: skip past packets that userspace has released.
923 	 * (kring->nr_hwcur to head excluded),
924 	 * and make the buffers available for reception.
925 	 * As usual nm_i is the index in the netmap ring,
926 	 * nic_i is the index in the NIC ring, and
927 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
928 	 */
929 	/* XXX not sure how this will work with multiple free lists */
930 	nm_i = kring->nr_hwcur;
931 	if (nm_i != head) {
932 		nic_i = netmap_idx_k2n(kring, nm_i);
933 		for (n = 0; nm_i != head; n++) {
934 			struct netmap_slot *slot = &ring->slot[nm_i];
935 			uint64_t paddr;
936 			caddr_t vaddr;
937 			void *addr = PNMB(na, slot, &paddr);
938 
939 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
940 				goto ring_reset;
941 
942 			vaddr = addr;
943 			if (slot->flags & NS_BUF_CHANGED) {
944 				/* buffer has changed, reload map */
945 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, addr);
946 				slot->flags &= ~NS_BUF_CHANGED;
947 			}
948 			/*
949 			 * XXX we should be batching this operation - TODO
950 			 */
951 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size);
952 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map,
953 			    BUS_DMASYNC_PREREAD);
954 			nm_i = nm_next(nm_i, lim);
955 			nic_i = nm_next(nic_i, lim);
956 		}
957 		kring->nr_hwcur = head;
958 
959 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
960 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
961 		/*
962 		 * IMPORTANT: we must leave one free slot in the ring,
963 		 * so move nic_i back by one unit
964 		 */
965 		nic_i = nm_prev(nic_i, lim);
966 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
967 	}
968 
969 	return 0;
970 
971 ring_reset:
972 	return netmap_ring_reinit(kring);
973 }
974 
975 static int
976 iflib_netmap_attach(if_ctx_t ctx)
977 {
978 	struct netmap_adapter na;
979 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
980 
981 	bzero(&na, sizeof(na));
982 
983 	na.ifp = ctx->ifc_ifp;
984 	na.na_flags = NAF_BDG_MAYSLEEP;
985 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
986 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
987 
988 	na.num_tx_desc = scctx->isc_ntxd[0];
989 	na.num_rx_desc = scctx->isc_nrxd[0];
990 	na.nm_txsync = iflib_netmap_txsync;
991 	na.nm_rxsync = iflib_netmap_rxsync;
992 	na.nm_register = iflib_netmap_register;
993 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
994 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
995 	return (netmap_attach(&na));
996 }
997 
998 static void
999 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1000 {
1001 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1002 	struct netmap_slot *slot;
1003 
1004 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1005 	if (slot == 0)
1006 		return;
1007 
1008 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1009 
1010 		/*
1011 		 * In netmap mode, set the map for the packet buffer.
1012 		 * NOTE: Some drivers (not this one) also need to set
1013 		 * the physical buffer address in the NIC ring.
1014 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1015 		 * netmap slot index, si
1016 		 */
1017 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1018 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1019 	}
1020 }
1021 static void
1022 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1023 {
1024 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1025 	struct netmap_slot *slot;
1026 	iflib_rxsd_t sd;
1027 	int nrxd;
1028 
1029 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1030 	if (slot == 0)
1031 		return;
1032 	sd = rxq->ifr_fl[0].ifl_sds;
1033 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1034 	for (int i = 0; i < nrxd; i++, sd++) {
1035 			int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1036 			uint64_t paddr;
1037 			void *addr;
1038 			caddr_t vaddr;
1039 
1040 			vaddr = addr = PNMB(na, slot + sj, &paddr);
1041 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, sd->ifsd_map, addr);
1042 			/* Update descriptor and the cached value */
1043 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size);
1044 	}
1045 	/* preserve queue */
1046 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1047 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1048 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1049 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1050 	} else
1051 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1052 }
1053 
1054 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1055 
1056 #else
1057 #define iflib_netmap_txq_init(ctx, txq)
1058 #define iflib_netmap_rxq_init(ctx, rxq)
1059 #define iflib_netmap_detach(ifp)
1060 
1061 #define iflib_netmap_attach(ctx) (0)
1062 #define netmap_rx_irq(ifp, qid, budget) (0)
1063 
1064 #endif
1065 
1066 #if defined(__i386__) || defined(__amd64__)
1067 static __inline void
1068 prefetch(void *x)
1069 {
1070 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1071 }
1072 #else
1073 #define prefetch(x)
1074 #endif
1075 
1076 static void
1077 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1078 {
1079 	if (err)
1080 		return;
1081 	*(bus_addr_t *) arg = segs[0].ds_addr;
1082 }
1083 
1084 int
1085 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1086 {
1087 	int err;
1088 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1089 	device_t dev = ctx->ifc_dev;
1090 
1091 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1092 
1093 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1094 				sctx->isc_q_align, 0,	/* alignment, bounds */
1095 				BUS_SPACE_MAXADDR,	/* lowaddr */
1096 				BUS_SPACE_MAXADDR,	/* highaddr */
1097 				NULL, NULL,		/* filter, filterarg */
1098 				size,			/* maxsize */
1099 				1,			/* nsegments */
1100 				size,			/* maxsegsize */
1101 				BUS_DMA_ALLOCNOW,	/* flags */
1102 				NULL,			/* lockfunc */
1103 				NULL,			/* lockarg */
1104 				&dma->idi_tag);
1105 	if (err) {
1106 		device_printf(dev,
1107 		    "%s: bus_dma_tag_create failed: %d\n",
1108 		    __func__, err);
1109 		goto fail_0;
1110 	}
1111 
1112 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1113 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1114 	if (err) {
1115 		device_printf(dev,
1116 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1117 		    __func__, (uintmax_t)size, err);
1118 		goto fail_1;
1119 	}
1120 
1121 	dma->idi_paddr = IF_BAD_DMA;
1122 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1123 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1124 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1125 		device_printf(dev,
1126 		    "%s: bus_dmamap_load failed: %d\n",
1127 		    __func__, err);
1128 		goto fail_2;
1129 	}
1130 
1131 	dma->idi_size = size;
1132 	return (0);
1133 
1134 fail_2:
1135 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1136 fail_1:
1137 	bus_dma_tag_destroy(dma->idi_tag);
1138 fail_0:
1139 	dma->idi_tag = NULL;
1140 
1141 	return (err);
1142 }
1143 
1144 int
1145 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1146 {
1147 	int i, err;
1148 	iflib_dma_info_t *dmaiter;
1149 
1150 	dmaiter = dmalist;
1151 	for (i = 0; i < count; i++, dmaiter++) {
1152 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1153 			break;
1154 	}
1155 	if (err)
1156 		iflib_dma_free_multi(dmalist, i);
1157 	return (err);
1158 }
1159 
1160 void
1161 iflib_dma_free(iflib_dma_info_t dma)
1162 {
1163 	if (dma->idi_tag == NULL)
1164 		return;
1165 	if (dma->idi_paddr != IF_BAD_DMA) {
1166 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1167 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1168 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1169 		dma->idi_paddr = IF_BAD_DMA;
1170 	}
1171 	if (dma->idi_vaddr != NULL) {
1172 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1173 		dma->idi_vaddr = NULL;
1174 	}
1175 	bus_dma_tag_destroy(dma->idi_tag);
1176 	dma->idi_tag = NULL;
1177 }
1178 
1179 void
1180 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1181 {
1182 	int i;
1183 	iflib_dma_info_t *dmaiter = dmalist;
1184 
1185 	for (i = 0; i < count; i++, dmaiter++)
1186 		iflib_dma_free(*dmaiter);
1187 }
1188 
1189 static int
1190 iflib_fast_intr(void *arg)
1191 {
1192 	iflib_filter_info_t info = arg;
1193 	struct grouptask *gtask = info->ifi_task;
1194 
1195 	DBG_COUNTER_INC(fast_intrs);
1196 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1197 		return (FILTER_HANDLED);
1198 
1199 	GROUPTASK_ENQUEUE(gtask);
1200 	return (FILTER_HANDLED);
1201 }
1202 
1203 static int
1204 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1205 	driver_filter_t filter, driver_intr_t handler, void *arg,
1206 				 char *name)
1207 {
1208 	int rc;
1209 	struct resource *res;
1210 	void *tag;
1211 	device_t dev = ctx->ifc_dev;
1212 
1213 	MPASS(rid < 512);
1214 	irq->ii_rid = rid;
1215 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1216 				     RF_SHAREABLE | RF_ACTIVE);
1217 	if (res == NULL) {
1218 		device_printf(dev,
1219 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1220 		return (ENOMEM);
1221 	}
1222 	irq->ii_res = res;
1223 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1224 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1225 						filter, handler, arg, &tag);
1226 	if (rc != 0) {
1227 		device_printf(dev,
1228 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1229 					  rid, name ? name : "unknown", rc);
1230 		return (rc);
1231 	} else if (name)
1232 		bus_describe_intr(dev, res, tag, "%s", name);
1233 
1234 	irq->ii_tag = tag;
1235 	return (0);
1236 }
1237 
1238 
1239 /*********************************************************************
1240  *
1241  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1242  *  the information needed to transmit a packet on the wire. This is
1243  *  called only once at attach, setup is done every reset.
1244  *
1245  **********************************************************************/
1246 
1247 static int
1248 iflib_txsd_alloc(iflib_txq_t txq)
1249 {
1250 	if_ctx_t ctx = txq->ift_ctx;
1251 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1252 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1253 	device_t dev = ctx->ifc_dev;
1254 	int err, nsegments, ntsosegments;
1255 
1256 	nsegments = scctx->isc_tx_nsegments;
1257 	ntsosegments = scctx->isc_tx_tso_segments_max;
1258 	MPASS(scctx->isc_ntxd[0] > 0);
1259 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1260 	MPASS(nsegments > 0);
1261 	MPASS(ntsosegments > 0);
1262 	/*
1263 	 * Setup DMA descriptor areas.
1264 	 */
1265 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1266 			       1, 0,			/* alignment, bounds */
1267 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1268 			       BUS_SPACE_MAXADDR,	/* highaddr */
1269 			       NULL, NULL,		/* filter, filterarg */
1270 			       sctx->isc_tx_maxsize,		/* maxsize */
1271 			       nsegments,	/* nsegments */
1272 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1273 			       0,			/* flags */
1274 			       NULL,			/* lockfunc */
1275 			       NULL,			/* lockfuncarg */
1276 			       &txq->ift_desc_tag))) {
1277 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1278 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1279 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1280 		goto fail;
1281 	}
1282 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1283 			       1, 0,			/* alignment, bounds */
1284 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1285 			       BUS_SPACE_MAXADDR,	/* highaddr */
1286 			       NULL, NULL,		/* filter, filterarg */
1287 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1288 			       ntsosegments,	/* nsegments */
1289 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1290 			       0,			/* flags */
1291 			       NULL,			/* lockfunc */
1292 			       NULL,			/* lockfuncarg */
1293 			       &txq->ift_tso_desc_tag))) {
1294 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1295 
1296 		goto fail;
1297 	}
1298 	if (!(txq->ift_sds.ifsd_flags =
1299 	    (uint8_t *) malloc(sizeof(uint8_t) *
1300 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1301 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1302 		err = ENOMEM;
1303 		goto fail;
1304 	}
1305 	if (!(txq->ift_sds.ifsd_m =
1306 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1307 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1308 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1309 		err = ENOMEM;
1310 		goto fail;
1311 	}
1312 
1313         /* Create the descriptor buffer dma maps */
1314 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1315 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1316 		return (0);
1317 
1318 	if (!(txq->ift_sds.ifsd_map =
1319 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1320 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1321 		err = ENOMEM;
1322 		goto fail;
1323 	}
1324 
1325 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1326 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1327 		if (err != 0) {
1328 			device_printf(dev, "Unable to create TX DMA map\n");
1329 			goto fail;
1330 		}
1331 	}
1332 #endif
1333 	return (0);
1334 fail:
1335 	/* We free all, it handles case where we are in the middle */
1336 	iflib_tx_structures_free(ctx);
1337 	return (err);
1338 }
1339 
1340 static void
1341 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1342 {
1343 	bus_dmamap_t map;
1344 
1345 	map = NULL;
1346 	if (txq->ift_sds.ifsd_map != NULL)
1347 		map = txq->ift_sds.ifsd_map[i];
1348 	if (map != NULL) {
1349 		bus_dmamap_unload(txq->ift_desc_tag, map);
1350 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1351 		txq->ift_sds.ifsd_map[i] = NULL;
1352 	}
1353 }
1354 
1355 static void
1356 iflib_txq_destroy(iflib_txq_t txq)
1357 {
1358 	if_ctx_t ctx = txq->ift_ctx;
1359 
1360 	for (int i = 0; i < txq->ift_size; i++)
1361 		iflib_txsd_destroy(ctx, txq, i);
1362 	if (txq->ift_sds.ifsd_map != NULL) {
1363 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1364 		txq->ift_sds.ifsd_map = NULL;
1365 	}
1366 	if (txq->ift_sds.ifsd_m != NULL) {
1367 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1368 		txq->ift_sds.ifsd_m = NULL;
1369 	}
1370 	if (txq->ift_sds.ifsd_flags != NULL) {
1371 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1372 		txq->ift_sds.ifsd_flags = NULL;
1373 	}
1374 	if (txq->ift_desc_tag != NULL) {
1375 		bus_dma_tag_destroy(txq->ift_desc_tag);
1376 		txq->ift_desc_tag = NULL;
1377 	}
1378 	if (txq->ift_tso_desc_tag != NULL) {
1379 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1380 		txq->ift_tso_desc_tag = NULL;
1381 	}
1382 }
1383 
1384 static void
1385 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1386 {
1387 	struct mbuf **mp;
1388 
1389 	mp = &txq->ift_sds.ifsd_m[i];
1390 	if (*mp == NULL)
1391 		return;
1392 
1393 	if (txq->ift_sds.ifsd_map != NULL) {
1394 		bus_dmamap_sync(txq->ift_desc_tag,
1395 				txq->ift_sds.ifsd_map[i],
1396 				BUS_DMASYNC_POSTWRITE);
1397 		bus_dmamap_unload(txq->ift_desc_tag,
1398 				  txq->ift_sds.ifsd_map[i]);
1399 	}
1400 	m_free(*mp);
1401 	DBG_COUNTER_INC(tx_frees);
1402 	*mp = NULL;
1403 }
1404 
1405 static int
1406 iflib_txq_setup(iflib_txq_t txq)
1407 {
1408 	if_ctx_t ctx = txq->ift_ctx;
1409 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1410 	iflib_dma_info_t di;
1411 	int i;
1412 
1413     /* Set number of descriptors available */
1414 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1415 
1416 	/* Reset indices */
1417 	txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1418 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1419 
1420 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1421 		bzero((void *)di->idi_vaddr, di->idi_size);
1422 
1423 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1424 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1425 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1426 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1427 	return (0);
1428 }
1429 
1430 /*********************************************************************
1431  *
1432  *  Allocate memory for rx_buffer structures. Since we use one
1433  *  rx_buffer per received packet, the maximum number of rx_buffer's
1434  *  that we'll need is equal to the number of receive descriptors
1435  *  that we've allocated.
1436  *
1437  **********************************************************************/
1438 static int
1439 iflib_rxsd_alloc(iflib_rxq_t rxq)
1440 {
1441 	if_ctx_t ctx = rxq->ifr_ctx;
1442 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1443 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1444 	device_t dev = ctx->ifc_dev;
1445 	iflib_fl_t fl;
1446 	iflib_rxsd_t	rxsd;
1447 	int			err;
1448 
1449 	MPASS(scctx->isc_nrxd[0] > 0);
1450 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1451 
1452 	fl = rxq->ifr_fl;
1453 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1454 		fl->ifl_sds = malloc(sizeof(struct iflib_sw_rx_desc) *
1455 		    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB,
1456 		    M_WAITOK | M_ZERO);
1457 		if (fl->ifl_sds == NULL) {
1458 			device_printf(dev, "Unable to allocate rx sw desc memory\n");
1459 			return (ENOMEM);
1460 		}
1461 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1462 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1463 					 1, 0,			/* alignment, bounds */
1464 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1465 					 BUS_SPACE_MAXADDR,	/* highaddr */
1466 					 NULL, NULL,		/* filter, filterarg */
1467 					 sctx->isc_rx_maxsize,	/* maxsize */
1468 					 sctx->isc_rx_nsegments,	/* nsegments */
1469 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1470 					 0,			/* flags */
1471 					 NULL,			/* lockfunc */
1472 					 NULL,			/* lockarg */
1473 					 &fl->ifl_desc_tag);
1474 		if (err) {
1475 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1476 				__func__, err);
1477 			goto fail;
1478 		}
1479 
1480 		rxsd = fl->ifl_sds;
1481 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++, rxsd++) {
1482 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &rxsd->ifsd_map);
1483 			if (err) {
1484 				device_printf(dev, "%s: bus_dmamap_create failed: %d\n",
1485 					__func__, err);
1486 				goto fail;
1487 			}
1488 		}
1489 	}
1490 	return (0);
1491 
1492 fail:
1493 	iflib_rx_structures_free(ctx);
1494 	return (err);
1495 }
1496 
1497 
1498 /*
1499  * Internal service routines
1500  */
1501 
1502 struct rxq_refill_cb_arg {
1503 	int               error;
1504 	bus_dma_segment_t seg;
1505 	int               nseg;
1506 };
1507 
1508 static void
1509 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1510 {
1511 	struct rxq_refill_cb_arg *cb_arg = arg;
1512 
1513 	cb_arg->error = error;
1514 	cb_arg->seg = segs[0];
1515 	cb_arg->nseg = nseg;
1516 }
1517 
1518 
1519 #ifdef ACPI_DMAR
1520 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1521 #else
1522 #define IS_DMAR(ctx) (0)
1523 #endif
1524 
1525 /**
1526  *	rxq_refill - refill an rxq  free-buffer list
1527  *	@ctx: the iflib context
1528  *	@rxq: the free-list to refill
1529  *	@n: the number of new buffers to allocate
1530  *
1531  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1532  *	The caller must assure that @n does not exceed the queue's capacity.
1533  */
1534 static void
1535 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1536 {
1537 	struct mbuf *m;
1538 	int pidx = fl->ifl_pidx;
1539 	iflib_rxsd_t rxsd = &fl->ifl_sds[pidx];
1540 	caddr_t cl;
1541 	int n, i = 0;
1542 	uint64_t bus_addr;
1543 	int err;
1544 
1545 	n  = count;
1546 	MPASS(n > 0);
1547 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1548 
1549 	if (pidx < fl->ifl_cidx)
1550 		MPASS(pidx + n <= fl->ifl_cidx);
1551 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1552 		MPASS(fl->ifl_gen == 0);
1553 	if (pidx > fl->ifl_cidx)
1554 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1555 
1556 	DBG_COUNTER_INC(fl_refills);
1557 	if (n > 8)
1558 		DBG_COUNTER_INC(fl_refills_large);
1559 
1560 	while (n--) {
1561 		/*
1562 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1563 		 * initialized after rx.
1564 		 *
1565 		 * If the cluster is still set then we know a minimum sized packet was received
1566 		 */
1567 		if ((cl = rxsd->ifsd_cl) == NULL) {
1568 			if ((cl = rxsd->ifsd_cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1569 				break;
1570 #if MEMORY_LOGGING
1571 			fl->ifl_cl_enqueued++;
1572 #endif
1573 		}
1574 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1575 			break;
1576 		}
1577 #if MEMORY_LOGGING
1578 		fl->ifl_m_enqueued++;
1579 #endif
1580 
1581 		DBG_COUNTER_INC(rx_allocs);
1582 #ifdef notyet
1583 		if ((rxsd->ifsd_flags & RX_SW_DESC_MAP_CREATED) == 0) {
1584 			int err;
1585 
1586 			if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &rxsd->ifsd_map))) {
1587 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
1588 				uma_zfree(fl->ifl_zone, cl);
1589 				n = 0;
1590 				goto done;
1591 			}
1592 			rxsd->ifsd_flags |= RX_SW_DESC_MAP_CREATED;
1593 		}
1594 #endif
1595 #if defined(__i386__) || defined(__amd64__)
1596 		if (!IS_DMAR(ctx)) {
1597 			bus_addr = pmap_kextract((vm_offset_t)cl);
1598 		} else
1599 #endif
1600 		{
1601 			struct rxq_refill_cb_arg cb_arg;
1602 			iflib_rxq_t q;
1603 
1604 			cb_arg.error = 0;
1605 			q = fl->ifl_rxq;
1606 			err = bus_dmamap_load(fl->ifl_desc_tag, rxsd->ifsd_map,
1607 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1608 
1609 			if (err != 0 || cb_arg.error) {
1610 				/*
1611 				 * !zone_pack ?
1612 				 */
1613 				if (fl->ifl_zone == zone_pack)
1614 					uma_zfree(fl->ifl_zone, cl);
1615 				m_free(m);
1616 				n = 0;
1617 				goto done;
1618 			}
1619 			bus_addr = cb_arg.seg.ds_addr;
1620 		}
1621 		rxsd->ifsd_flags |= RX_SW_DESC_INUSE;
1622 
1623 		MPASS(rxsd->ifsd_m == NULL);
1624 		rxsd->ifsd_cl = cl;
1625 		rxsd->ifsd_m = m;
1626 		fl->ifl_bus_addrs[i] = bus_addr;
1627 		fl->ifl_vm_addrs[i] = cl;
1628 		rxsd++;
1629 		fl->ifl_credits++;
1630 		i++;
1631 		MPASS(fl->ifl_credits <= fl->ifl_size);
1632 		if (++fl->ifl_pidx == fl->ifl_size) {
1633 			fl->ifl_pidx = 0;
1634 			fl->ifl_gen = 1;
1635 			rxsd = fl->ifl_sds;
1636 		}
1637 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1638 			ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx,
1639 								 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size);
1640 			i = 0;
1641 			pidx = fl->ifl_pidx;
1642 		}
1643 	}
1644 done:
1645 	DBG_COUNTER_INC(rxd_flush);
1646 	if (fl->ifl_pidx == 0)
1647 		pidx = fl->ifl_size - 1;
1648 	else
1649 		pidx = fl->ifl_pidx - 1;
1650 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1651 }
1652 
1653 static __inline void
1654 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1655 {
1656 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1657 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1658 #ifdef INVARIANTS
1659 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1660 #endif
1661 
1662 	MPASS(fl->ifl_credits <= fl->ifl_size);
1663 	MPASS(reclaimable == delta);
1664 
1665 	if (reclaimable > 0)
1666 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1667 }
1668 
1669 static void
1670 iflib_fl_bufs_free(iflib_fl_t fl)
1671 {
1672 	iflib_dma_info_t idi = fl->ifl_ifdi;
1673 	uint32_t i;
1674 
1675 	for (i = 0; i < fl->ifl_size; i++) {
1676 		iflib_rxsd_t d = &fl->ifl_sds[i];
1677 
1678 		if (d->ifsd_flags & RX_SW_DESC_INUSE) {
1679 			bus_dmamap_unload(fl->ifl_desc_tag, d->ifsd_map);
1680 			bus_dmamap_destroy(fl->ifl_desc_tag, d->ifsd_map);
1681 			if (d->ifsd_m != NULL) {
1682 				m_init(d->ifsd_m, M_NOWAIT, MT_DATA, 0);
1683 				uma_zfree(zone_mbuf, d->ifsd_m);
1684 			}
1685 			if (d->ifsd_cl != NULL)
1686 				uma_zfree(fl->ifl_zone, d->ifsd_cl);
1687 			d->ifsd_flags = 0;
1688 		} else {
1689 			MPASS(d->ifsd_cl == NULL);
1690 			MPASS(d->ifsd_m == NULL);
1691 		}
1692 #if MEMORY_LOGGING
1693 		fl->ifl_m_dequeued++;
1694 		fl->ifl_cl_dequeued++;
1695 #endif
1696 		d->ifsd_cl = NULL;
1697 		d->ifsd_m = NULL;
1698 	}
1699 	/*
1700 	 * Reset free list values
1701 	 */
1702 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1703 	bzero(idi->idi_vaddr, idi->idi_size);
1704 }
1705 
1706 /*********************************************************************
1707  *
1708  *  Initialize a receive ring and its buffers.
1709  *
1710  **********************************************************************/
1711 static int
1712 iflib_fl_setup(iflib_fl_t fl)
1713 {
1714 	iflib_rxq_t rxq = fl->ifl_rxq;
1715 	if_ctx_t ctx = rxq->ifr_ctx;
1716 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1717 
1718 	/*
1719 	** Free current RX buffer structs and their mbufs
1720 	*/
1721 	iflib_fl_bufs_free(fl);
1722 	/* Now replenish the mbufs */
1723 	MPASS(fl->ifl_credits == 0);
1724 	/*
1725 	 * XXX don't set the max_frame_size to larger
1726 	 * than the hardware can handle
1727 	 */
1728 	if (sctx->isc_max_frame_size <= 2048)
1729 		fl->ifl_buf_size = MCLBYTES;
1730 	else if (sctx->isc_max_frame_size <= 4096)
1731 		fl->ifl_buf_size = MJUMPAGESIZE;
1732 	else if (sctx->isc_max_frame_size <= 9216)
1733 		fl->ifl_buf_size = MJUM9BYTES;
1734 	else
1735 		fl->ifl_buf_size = MJUM16BYTES;
1736 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
1737 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
1738 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
1739 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
1740 
1741 
1742 	/* avoid pre-allocating zillions of clusters to an idle card
1743 	 * potentially speeding up attach
1744 	 */
1745 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
1746 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
1747 	if (min(128, fl->ifl_size) != fl->ifl_credits)
1748 		return (ENOBUFS);
1749 	/*
1750 	 * handle failure
1751 	 */
1752 	MPASS(rxq != NULL);
1753 	MPASS(fl->ifl_ifdi != NULL);
1754 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1755 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1756 	return (0);
1757 }
1758 
1759 /*********************************************************************
1760  *
1761  *  Free receive ring data structures
1762  *
1763  **********************************************************************/
1764 static void
1765 iflib_rx_sds_free(iflib_rxq_t rxq)
1766 {
1767 	iflib_fl_t fl;
1768 	int i;
1769 
1770 	if (rxq->ifr_fl != NULL) {
1771 		for (i = 0; i < rxq->ifr_nfl; i++) {
1772 			fl = &rxq->ifr_fl[i];
1773 			if (fl->ifl_desc_tag != NULL) {
1774 				bus_dma_tag_destroy(fl->ifl_desc_tag);
1775 				fl->ifl_desc_tag = NULL;
1776 			}
1777 		}
1778 		if (rxq->ifr_fl->ifl_sds != NULL)
1779 			free(rxq->ifr_fl->ifl_sds, M_IFLIB);
1780 
1781 		free(rxq->ifr_fl, M_IFLIB);
1782 		rxq->ifr_fl = NULL;
1783 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
1784 	}
1785 }
1786 
1787 /*
1788  * MI independent logic
1789  *
1790  */
1791 static void
1792 iflib_timer(void *arg)
1793 {
1794 	iflib_txq_t txq = arg;
1795 	if_ctx_t ctx = txq->ift_ctx;
1796 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1797 
1798 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
1799 		return;
1800 	/*
1801 	** Check on the state of the TX queue(s), this
1802 	** can be done without the lock because its RO
1803 	** and the HUNG state will be static if set.
1804 	*/
1805 	IFDI_TIMER(ctx, txq->ift_id);
1806 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
1807 		(ctx->ifc_pause_frames == 0))
1808 		goto hung;
1809 
1810 	if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments ||
1811 	    ifmp_ring_is_stalled(txq->ift_br[0]))
1812 		GROUPTASK_ENQUEUE(&txq->ift_task);
1813 
1814 	ctx->ifc_pause_frames = 0;
1815 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
1816 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
1817 	return;
1818 hung:
1819 	CTX_LOCK(ctx);
1820 	if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
1821 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
1822 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
1823 
1824 	IFDI_WATCHDOG_RESET(ctx);
1825 	ctx->ifc_watchdog_events++;
1826 	ctx->ifc_pause_frames = 0;
1827 
1828 	iflib_init_locked(ctx);
1829 	CTX_UNLOCK(ctx);
1830 }
1831 
1832 static void
1833 iflib_init_locked(if_ctx_t ctx)
1834 {
1835 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1836 	if_t ifp = ctx->ifc_ifp;
1837 	iflib_fl_t fl;
1838 	iflib_txq_t txq;
1839 	iflib_rxq_t rxq;
1840 	int i, j;
1841 
1842 
1843 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1844 	IFDI_INTR_DISABLE(ctx);
1845 
1846 	/* Set hardware offload abilities */
1847 	if_clearhwassist(ifp);
1848 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
1849 		if_sethwassistbits(ifp, CSUM_IP | CSUM_TCP | CSUM_UDP, 0);
1850 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
1851 		if_sethwassistbits(ifp,  (CSUM_TCP_IPV6 | CSUM_UDP_IPV6), 0);
1852 	if (if_getcapenable(ifp) & IFCAP_TSO4)
1853 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
1854 	if (if_getcapenable(ifp) & IFCAP_TSO6)
1855 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
1856 
1857 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
1858 		CALLOUT_LOCK(txq);
1859 		callout_stop(&txq->ift_timer);
1860 		callout_stop(&txq->ift_db_check);
1861 		CALLOUT_UNLOCK(txq);
1862 		iflib_netmap_txq_init(ctx, txq);
1863 	}
1864 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1865 		iflib_netmap_rxq_init(ctx, rxq);
1866 	}
1867 #ifdef INVARIANTS
1868 	i = if_getdrvflags(ifp);
1869 #endif
1870 	IFDI_INIT(ctx);
1871 	MPASS(if_getdrvflags(ifp) == i);
1872 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1873 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
1874 			if (iflib_fl_setup(fl)) {
1875 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
1876 				goto done;
1877 			}
1878 		}
1879 	}
1880 	done:
1881 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
1882 	IFDI_INTR_ENABLE(ctx);
1883 	txq = ctx->ifc_txqs;
1884 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
1885 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
1886 			txq->ift_timer.c_cpu);
1887 }
1888 
1889 static int
1890 iflib_media_change(if_t ifp)
1891 {
1892 	if_ctx_t ctx = if_getsoftc(ifp);
1893 	int err;
1894 
1895 	CTX_LOCK(ctx);
1896 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
1897 		iflib_init_locked(ctx);
1898 	CTX_UNLOCK(ctx);
1899 	return (err);
1900 }
1901 
1902 static void
1903 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
1904 {
1905 	if_ctx_t ctx = if_getsoftc(ifp);
1906 
1907 	CTX_LOCK(ctx);
1908 	IFDI_UPDATE_ADMIN_STATUS(ctx);
1909 	IFDI_MEDIA_STATUS(ctx, ifmr);
1910 	CTX_UNLOCK(ctx);
1911 }
1912 
1913 static void
1914 iflib_stop(if_ctx_t ctx)
1915 {
1916 	iflib_txq_t txq = ctx->ifc_txqs;
1917 	iflib_rxq_t rxq = ctx->ifc_rxqs;
1918 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1919 	iflib_dma_info_t di;
1920 	iflib_fl_t fl;
1921 	int i, j;
1922 
1923 	/* Tell the stack that the interface is no longer active */
1924 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1925 
1926 	IFDI_INTR_DISABLE(ctx);
1927 	DELAY(100000);
1928 	IFDI_STOP(ctx);
1929 	DELAY(100000);
1930 
1931 	iflib_debug_reset();
1932 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
1933 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
1934 		/* make sure all transmitters have completed before proceeding XXX */
1935 
1936 		/* clean any enqueued buffers */
1937 		iflib_ifmp_purge(txq);
1938 		/* Free any existing tx buffers. */
1939 		for (j = 0; j < txq->ift_size; j++) {
1940 			iflib_txsd_free(ctx, txq, j);
1941 		}
1942 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
1943 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
1944 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
1945 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
1946 		txq->ift_pullups = 0;
1947 		ifmp_ring_reset_stats(txq->ift_br[0]);
1948 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
1949 			bzero((void *)di->idi_vaddr, di->idi_size);
1950 	}
1951 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
1952 		/* make sure all transmitters have completed before proceeding XXX */
1953 
1954 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
1955 			bzero((void *)di->idi_vaddr, di->idi_size);
1956 		/* also resets the free lists pidx/cidx */
1957 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
1958 			iflib_fl_bufs_free(fl);
1959 	}
1960 }
1961 
1962 static iflib_rxsd_t
1963 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload)
1964 {
1965 	int flid, cidx;
1966 	iflib_rxsd_t sd;
1967 	iflib_fl_t fl;
1968 	iflib_dma_info_t di;
1969 
1970 	flid = irf->irf_flid;
1971 	cidx = irf->irf_idx;
1972 	fl = &rxq->ifr_fl[flid];
1973 	fl->ifl_credits--;
1974 #if MEMORY_LOGGING
1975 	fl->ifl_m_dequeued++;
1976 	if (cltype)
1977 		fl->ifl_cl_dequeued++;
1978 #endif
1979 	sd = &fl->ifl_sds[cidx];
1980 	di = fl->ifl_ifdi;
1981 	bus_dmamap_sync(di->idi_tag, di->idi_map,
1982 			BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1983 
1984 	/* not valid assert if bxe really does SGE from non-contiguous elements */
1985 	MPASS(fl->ifl_cidx == cidx);
1986 	if (unload)
1987 		bus_dmamap_unload(fl->ifl_desc_tag, sd->ifsd_map);
1988 
1989 	if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) {
1990 		fl->ifl_cidx = 0;
1991 		fl->ifl_gen = 0;
1992 	}
1993 	/* YES ick */
1994 	if (cltype)
1995 		*cltype = fl->ifl_cltype;
1996 	return (sd);
1997 }
1998 
1999 static struct mbuf *
2000 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri)
2001 {
2002 	int i, padlen , flags, cltype;
2003 	struct mbuf *m, *mh, *mt;
2004 	iflib_rxsd_t sd;
2005 	caddr_t cl;
2006 
2007 	i = 0;
2008 	mh = NULL;
2009 	do {
2010 		sd = rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE);
2011 
2012 		MPASS(sd->ifsd_cl != NULL);
2013 		MPASS(sd->ifsd_m != NULL);
2014 
2015 		/* Don't include zero-length frags */
2016 		if (ri->iri_frags[i].irf_len == 0) {
2017 			/* XXX we can save the cluster here, but not the mbuf */
2018 			m_init(sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2019 			m_free(sd->ifsd_m);
2020 			sd->ifsd_m = NULL;
2021 			continue;
2022 		}
2023 
2024 		m = sd->ifsd_m;
2025 		if (mh == NULL) {
2026 			flags = M_PKTHDR|M_EXT;
2027 			mh = mt = m;
2028 			padlen = ri->iri_pad;
2029 		} else {
2030 			flags = M_EXT;
2031 			mt->m_next = m;
2032 			mt = m;
2033 			/* assuming padding is only on the first fragment */
2034 			padlen = 0;
2035 		}
2036 		sd->ifsd_m = NULL;
2037 		cl = sd->ifsd_cl;
2038 		sd->ifsd_cl = NULL;
2039 
2040 		/* Can these two be made one ? */
2041 		m_init(m, M_NOWAIT, MT_DATA, flags);
2042 		m_cljset(m, cl, cltype);
2043 		/*
2044 		 * These must follow m_init and m_cljset
2045 		 */
2046 		m->m_data += padlen;
2047 		ri->iri_len -= padlen;
2048 		m->m_len = ri->iri_frags[i].irf_len;
2049 	} while (++i < ri->iri_nfrags);
2050 
2051 	return (mh);
2052 }
2053 
2054 /*
2055  * Process one software descriptor
2056  */
2057 static struct mbuf *
2058 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2059 {
2060 	struct mbuf *m;
2061 	iflib_rxsd_t sd;
2062 
2063 	/* should I merge this back in now that the two paths are basically duplicated? */
2064 	if (ri->iri_nfrags == 1 &&
2065 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2066 		sd = rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE);
2067 		m = sd->ifsd_m;
2068 		sd->ifsd_m = NULL;
2069 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2070 		memcpy(m->m_data, sd->ifsd_cl, ri->iri_len);
2071 		m->m_len = ri->iri_frags[0].irf_len;
2072        } else {
2073 		m = assemble_segments(rxq, ri);
2074 	}
2075 	m->m_pkthdr.len = ri->iri_len;
2076 	m->m_pkthdr.rcvif = ri->iri_ifp;
2077 	m->m_flags |= ri->iri_flags;
2078 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2079 	m->m_pkthdr.flowid = ri->iri_flowid;
2080 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2081 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2082 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2083 	return (m);
2084 }
2085 
2086 static bool
2087 iflib_rxeof(iflib_rxq_t rxq, int budget)
2088 {
2089 	if_ctx_t ctx = rxq->ifr_ctx;
2090 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2091 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2092 	int avail, i;
2093 	uint16_t *cidxp;
2094 	struct if_rxd_info ri;
2095 	int err, budget_left, rx_bytes, rx_pkts;
2096 	iflib_fl_t fl;
2097 	struct ifnet *ifp;
2098 	int lro_enabled;
2099 	/*
2100 	 * XXX early demux data packets so that if_input processing only handles
2101 	 * acks in interrupt context
2102 	 */
2103 	struct mbuf *m, *mh, *mt;
2104 
2105 	if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) {
2106 		return (FALSE);
2107 	}
2108 
2109 	mh = mt = NULL;
2110 	MPASS(budget > 0);
2111 	rx_pkts	= rx_bytes = 0;
2112 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2113 		cidxp = &rxq->ifr_cq_cidx;
2114 	else
2115 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2116 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2117 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2118 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2119 		DBG_COUNTER_INC(rx_unavail);
2120 		return (false);
2121 	}
2122 
2123 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2124 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2125 			DBG_COUNTER_INC(rx_ctx_inactive);
2126 			break;
2127 		}
2128 		/*
2129 		 * Reset client set fields to their default values
2130 		 */
2131 		bzero(&ri, sizeof(ri));
2132 		ri.iri_qsidx = rxq->ifr_id;
2133 		ri.iri_cidx = *cidxp;
2134 		ri.iri_ifp = ctx->ifc_ifp;
2135 		ri.iri_frags = rxq->ifr_frags;
2136 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2137 
2138 		/* in lieu of handling correctly - make sure it isn't being unhandled */
2139 		MPASS(err == 0);
2140 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2141 			*cidxp = ri.iri_cidx;
2142 			/* Update our consumer index */
2143 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2144 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2145 				rxq->ifr_cq_gen = 0;
2146 			}
2147 			/* was this only a completion queue message? */
2148 			if (__predict_false(ri.iri_nfrags == 0))
2149 				continue;
2150 		}
2151 		MPASS(ri.iri_nfrags != 0);
2152 		MPASS(ri.iri_len != 0);
2153 
2154 		/* will advance the cidx on the corresponding free lists */
2155 		m = iflib_rxd_pkt_get(rxq, &ri);
2156 		if (avail == 0 && budget_left)
2157 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2158 
2159 		if (__predict_false(m == NULL)) {
2160 			DBG_COUNTER_INC(rx_mbuf_null);
2161 			continue;
2162 		}
2163 		/* imm_pkt: -- cxgb */
2164 		if (mh == NULL)
2165 			mh = mt = m;
2166 		else {
2167 			mt->m_nextpkt = m;
2168 			mt = m;
2169 		}
2170 	}
2171 	/* make sure that we can refill faster than drain */
2172 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2173 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2174 
2175 	ifp = ctx->ifc_ifp;
2176 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2177 	while (mh != NULL) {
2178 		m = mh;
2179 		mh = mh->m_nextpkt;
2180 		m->m_nextpkt = NULL;
2181 		rx_bytes += m->m_pkthdr.len;
2182 		rx_pkts++;
2183 #if defined(INET6) || defined(INET)
2184 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2185 			continue;
2186 #endif
2187 		DBG_COUNTER_INC(rx_if_input);
2188 		ifp->if_input(ifp, m);
2189 	}
2190 
2191 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2192 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2193 
2194 	/*
2195 	 * Flush any outstanding LRO work
2196 	 */
2197 #if defined(INET6) || defined(INET)
2198 	tcp_lro_flush_all(&rxq->ifr_lc);
2199 #endif
2200 	if (avail)
2201 		return true;
2202 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2203 }
2204 
2205 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2206 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2207 #define TXQ_MAX_DB_DEFERRED(size) (size >> 5)
2208 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2209 
2210 static __inline void
2211 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring)
2212 {
2213 	uint32_t dbval;
2214 
2215 	if (ring || txq->ift_db_pending >=
2216 	    TXQ_MAX_DB_DEFERRED(txq->ift_size)) {
2217 
2218 		/* the lock will only ever be contended in the !min_latency case */
2219 		if (!TXDB_TRYLOCK(txq))
2220 			return;
2221 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2222 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2223 		txq->ift_db_pending = txq->ift_npending = 0;
2224 		TXDB_UNLOCK(txq);
2225 	}
2226 }
2227 
2228 static void
2229 iflib_txd_deferred_db_check(void * arg)
2230 {
2231 	iflib_txq_t txq = arg;
2232 
2233 	/* simple non-zero boolean so use bitwise OR */
2234 	if ((txq->ift_db_pending | txq->ift_npending) &&
2235 	    txq->ift_db_pending >= txq->ift_db_pending_queued)
2236 		iflib_txd_db_check(txq->ift_ctx, txq, TRUE);
2237 	txq->ift_db_pending_queued = 0;
2238 	if (ifmp_ring_is_stalled(txq->ift_br[0]))
2239 		iflib_txq_check_drain(txq, 4);
2240 }
2241 
2242 #ifdef PKT_DEBUG
2243 static void
2244 print_pkt(if_pkt_info_t pi)
2245 {
2246 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2247 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2248 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2249 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2250 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2251 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2252 }
2253 #endif
2254 
2255 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2256 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2257 
2258 static int
2259 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2260 {
2261 	struct ether_vlan_header *eh;
2262 	struct mbuf *m, *n;
2263 
2264 	n = m = *mp;
2265 	/*
2266 	 * Determine where frame payload starts.
2267 	 * Jump over vlan headers if already present,
2268 	 * helpful for QinQ too.
2269 	 */
2270 	if (__predict_false(m->m_len < sizeof(*eh))) {
2271 		txq->ift_pullups++;
2272 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2273 			return (ENOMEM);
2274 	}
2275 	eh = mtod(m, struct ether_vlan_header *);
2276 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2277 		pi->ipi_etype = ntohs(eh->evl_proto);
2278 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2279 	} else {
2280 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2281 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2282 	}
2283 
2284 	switch (pi->ipi_etype) {
2285 #ifdef INET
2286 	case ETHERTYPE_IP:
2287 	{
2288 		struct ip *ip = NULL;
2289 		struct tcphdr *th = NULL;
2290 		int minthlen;
2291 
2292 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2293 		if (__predict_false(m->m_len < minthlen)) {
2294 			/*
2295 			 * if this code bloat is causing too much of a hit
2296 			 * move it to a separate function and mark it noinline
2297 			 */
2298 			if (m->m_len == pi->ipi_ehdrlen) {
2299 				n = m->m_next;
2300 				MPASS(n);
2301 				if (n->m_len >= sizeof(*ip))  {
2302 					ip = (struct ip *)n->m_data;
2303 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2304 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2305 				} else {
2306 					txq->ift_pullups++;
2307 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2308 						return (ENOMEM);
2309 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2310 				}
2311 			} else {
2312 				txq->ift_pullups++;
2313 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2314 					return (ENOMEM);
2315 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2316 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2317 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2318 			}
2319 		} else {
2320 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2321 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2322 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2323 		}
2324 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2325 		pi->ipi_ipproto = ip->ip_p;
2326 		pi->ipi_flags |= IPI_TX_IPV4;
2327 
2328 		if (pi->ipi_csum_flags & CSUM_IP)
2329                        ip->ip_sum = 0;
2330 
2331 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2332 			if (__predict_false(th == NULL)) {
2333 				txq->ift_pullups++;
2334 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2335 					return (ENOMEM);
2336 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2337 			}
2338 			pi->ipi_tcp_hflags = th->th_flags;
2339 			pi->ipi_tcp_hlen = th->th_off << 2;
2340 			pi->ipi_tcp_seq = th->th_seq;
2341 		}
2342 		if (IS_TSO4(pi)) {
2343 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2344 				return (ENXIO);
2345 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2346 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2347 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2348 		}
2349 		break;
2350 	}
2351 #endif
2352 #ifdef INET6
2353 	case ETHERTYPE_IPV6:
2354 	{
2355 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2356 		struct tcphdr *th;
2357 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2358 
2359 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2360 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2361 				return (ENOMEM);
2362 		}
2363 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2364 
2365 		/* XXX-BZ this will go badly in case of ext hdrs. */
2366 		pi->ipi_ipproto = ip6->ip6_nxt;
2367 		pi->ipi_flags |= IPI_TX_IPV6;
2368 
2369 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2370 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2371 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2372 					return (ENOMEM);
2373 			}
2374 			pi->ipi_tcp_hflags = th->th_flags;
2375 			pi->ipi_tcp_hlen = th->th_off << 2;
2376 		}
2377 		if (IS_TSO6(pi)) {
2378 
2379 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2380 				return (ENXIO);
2381 			/*
2382 			 * The corresponding flag is set by the stack in the IPv4
2383 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2384 			 * So, set it here because the rest of the flow requires it.
2385 			 */
2386 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2387 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2388 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2389 		}
2390 		break;
2391 	}
2392 #endif
2393 	default:
2394 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2395 		pi->ipi_ip_hlen = 0;
2396 		break;
2397 	}
2398 	*mp = m;
2399 	return (0);
2400 }
2401 
2402 
2403 static  __noinline  struct mbuf *
2404 collapse_pkthdr(struct mbuf *m0)
2405 {
2406 	struct mbuf *m, *m_next, *tmp;
2407 
2408 	m = m0;
2409 	m_next = m->m_next;
2410 	while (m_next != NULL && m_next->m_len == 0) {
2411 		m = m_next;
2412 		m->m_next = NULL;
2413 		m_free(m);
2414 		m_next = m_next->m_next;
2415 	}
2416 	m = m0;
2417 	m->m_next = m_next;
2418 	if ((m_next->m_flags & M_EXT) == 0) {
2419 		m = m_defrag(m, M_NOWAIT);
2420 	} else {
2421 		tmp = m_next->m_next;
2422 		memcpy(m_next, m, MPKTHSIZE);
2423 		m = m_next;
2424 		m->m_next = tmp;
2425 	}
2426 	return (m);
2427 }
2428 
2429 /*
2430  * If dodgy hardware rejects the scatter gather chain we've handed it
2431  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2432  * m_defrag'd mbufs
2433  */
2434 static __noinline struct mbuf *
2435 iflib_remove_mbuf(iflib_txq_t txq)
2436 {
2437 	int ntxd, i, pidx;
2438 	struct mbuf *m, *mh, **ifsd_m;
2439 
2440 	pidx = txq->ift_pidx;
2441 	ifsd_m = txq->ift_sds.ifsd_m;
2442 	ntxd = txq->ift_size;
2443 	mh = m = ifsd_m[pidx];
2444 	ifsd_m[pidx] = NULL;
2445 #if MEMORY_LOGGING
2446 	txq->ift_dequeued++;
2447 #endif
2448 	i = 1;
2449 
2450 	while (m) {
2451 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2452 #if MEMORY_LOGGING
2453 		txq->ift_dequeued++;
2454 #endif
2455 		m = m->m_next;
2456 		i++;
2457 	}
2458 	return (mh);
2459 }
2460 
2461 static int
2462 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2463 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2464 			  int max_segs, int flags)
2465 {
2466 	if_ctx_t ctx;
2467 	if_shared_ctx_t		sctx;
2468 	if_softc_ctx_t		scctx;
2469 	int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2470 	struct mbuf *m, *tmp, **ifsd_m, **mp;
2471 
2472 	m = *m0;
2473 
2474 	/*
2475 	 * Please don't ever do this
2476 	 */
2477 	if (__predict_false(m->m_len == 0))
2478 		*m0 = m = collapse_pkthdr(m);
2479 
2480 	ctx = txq->ift_ctx;
2481 	sctx = ctx->ifc_sctx;
2482 	scctx = &ctx->ifc_softc_ctx;
2483 	ifsd_m = txq->ift_sds.ifsd_m;
2484 	ntxd = txq->ift_size;
2485 	pidx = txq->ift_pidx;
2486 	if (map != NULL) {
2487 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2488 
2489 		err = bus_dmamap_load_mbuf_sg(tag, map,
2490 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2491 		if (err)
2492 			return (err);
2493 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2494 		i = 0;
2495 		next = pidx;
2496 		mask = (txq->ift_size-1);
2497 		m = *m0;
2498 		do {
2499 			mp = &ifsd_m[next];
2500 			*mp = m;
2501 			m = m->m_next;
2502 			if (__predict_false((*mp)->m_len == 0)) {
2503 				m_free(*mp);
2504 				*mp = NULL;
2505 			} else
2506 				next = (pidx + i) & (ntxd-1);
2507 		} while (m != NULL);
2508 	} else {
2509 		int buflen, sgsize, max_sgsize;
2510 		vm_offset_t vaddr;
2511 		vm_paddr_t curaddr;
2512 
2513 		count = i = 0;
2514 		maxsegsz = sctx->isc_tx_maxsize;
2515 		m = *m0;
2516 		do {
2517 			if (__predict_false(m->m_len <= 0)) {
2518 				tmp = m;
2519 				m = m->m_next;
2520 				tmp->m_next = NULL;
2521 				m_free(tmp);
2522 				continue;
2523 			}
2524 			buflen = m->m_len;
2525 			vaddr = (vm_offset_t)m->m_data;
2526 			/*
2527 			 * see if we can't be smarter about physically
2528 			 * contiguous mappings
2529 			 */
2530 			next = (pidx + count) & (ntxd-1);
2531 			MPASS(ifsd_m[next] == NULL);
2532 #if MEMORY_LOGGING
2533 			txq->ift_enqueued++;
2534 #endif
2535 			ifsd_m[next] = m;
2536 			while (buflen > 0) {
2537 				max_sgsize = MIN(buflen, maxsegsz);
2538 				curaddr = pmap_kextract(vaddr);
2539 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2540 				sgsize = MIN(sgsize, max_sgsize);
2541 				segs[i].ds_addr = curaddr;
2542 				segs[i].ds_len = sgsize;
2543 				vaddr += sgsize;
2544 				buflen -= sgsize;
2545 				i++;
2546 				if (i >= max_segs)
2547 					goto err;
2548 			}
2549 			count++;
2550 			tmp = m;
2551 			m = m->m_next;
2552 		} while (m != NULL);
2553 		*nsegs = i;
2554 	}
2555 	return (0);
2556 err:
2557 	*m0 = iflib_remove_mbuf(txq);
2558 	return (EFBIG);
2559 }
2560 
2561 static int
2562 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2563 {
2564 	if_ctx_t		ctx;
2565 	if_shared_ctx_t		sctx;
2566 	if_softc_ctx_t		scctx;
2567 	bus_dma_segment_t	*segs;
2568 	struct mbuf		*m_head;
2569 	bus_dmamap_t		map;
2570 	struct if_pkt_info	pi;
2571 	int remap = 0;
2572 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
2573 	bus_dma_tag_t desc_tag;
2574 
2575 	segs = txq->ift_segs;
2576 	ctx = txq->ift_ctx;
2577 	sctx = ctx->ifc_sctx;
2578 	scctx = &ctx->ifc_softc_ctx;
2579 	segs = txq->ift_segs;
2580 	ntxd = txq->ift_size;
2581 	m_head = *m_headp;
2582 	map = NULL;
2583 
2584 	/*
2585 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
2586 	 */
2587 	cidx = txq->ift_cidx;
2588 	pidx = txq->ift_pidx;
2589 	next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
2590 
2591 	/* prefetch the next cache line of mbuf pointers and flags */
2592 	prefetch(&txq->ift_sds.ifsd_m[next]);
2593 	if (txq->ift_sds.ifsd_map != NULL) {
2594 		prefetch(&txq->ift_sds.ifsd_map[next]);
2595 		map = txq->ift_sds.ifsd_map[pidx];
2596 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
2597 		prefetch(&txq->ift_sds.ifsd_flags[next]);
2598 	}
2599 
2600 
2601 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
2602 		desc_tag = txq->ift_tso_desc_tag;
2603 		max_segs = scctx->isc_tx_tso_segments_max;
2604 	} else {
2605 		desc_tag = txq->ift_desc_tag;
2606 		max_segs = scctx->isc_tx_nsegments;
2607 	}
2608 	m_head = *m_headp;
2609 	bzero(&pi, sizeof(pi));
2610 	pi.ipi_len = m_head->m_pkthdr.len;
2611 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
2612 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
2613 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
2614 	pi.ipi_pidx = pidx;
2615 	pi.ipi_qsidx = txq->ift_id;
2616 
2617 	/* deliberate bitwise OR to make one condition */
2618 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
2619 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
2620 			return (err);
2621 		m_head = *m_headp;
2622 	}
2623 
2624 retry:
2625 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
2626 defrag:
2627 	if (__predict_false(err)) {
2628 		switch (err) {
2629 		case EFBIG:
2630 			/* try collapse once and defrag once */
2631 			if (remap == 0)
2632 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
2633 			if (remap == 1)
2634 				m_head = m_defrag(*m_headp, M_NOWAIT);
2635 			remap++;
2636 			if (__predict_false(m_head == NULL))
2637 				goto defrag_failed;
2638 			txq->ift_mbuf_defrag++;
2639 			*m_headp = m_head;
2640 			goto retry;
2641 			break;
2642 		case ENOMEM:
2643 			txq->ift_no_tx_dma_setup++;
2644 			break;
2645 		default:
2646 			txq->ift_no_tx_dma_setup++;
2647 			m_freem(*m_headp);
2648 			DBG_COUNTER_INC(tx_frees);
2649 			*m_headp = NULL;
2650 			break;
2651 		}
2652 		txq->ift_map_failed++;
2653 		DBG_COUNTER_INC(encap_load_mbuf_fail);
2654 		return (err);
2655 	}
2656 
2657 	/*
2658 	 * XXX assumes a 1 to 1 relationship between segments and
2659 	 *        descriptors - this does not hold true on all drivers, e.g.
2660 	 *        cxgb
2661 	 */
2662 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
2663 		txq->ift_no_desc_avail++;
2664 		if (map != NULL)
2665 			bus_dmamap_unload(desc_tag, map);
2666 		DBG_COUNTER_INC(encap_txq_avail_fail);
2667 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
2668 			GROUPTASK_ENQUEUE(&txq->ift_task);
2669 		return (ENOBUFS);
2670 	}
2671 	pi.ipi_segs = segs;
2672 	pi.ipi_nsegs = nsegs;
2673 
2674 	MPASS(pidx >= 0 && pidx < txq->ift_size);
2675 #ifdef PKT_DEBUG
2676 	print_pkt(&pi);
2677 #endif
2678 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
2679 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2680 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2681 
2682 		DBG_COUNTER_INC(tx_encap);
2683 		MPASS(pi.ipi_new_pidx >= 0 &&
2684 		    pi.ipi_new_pidx < txq->ift_size);
2685 
2686 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
2687 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
2688 			ndesc += txq->ift_size;
2689 			txq->ift_gen = 1;
2690 		}
2691 		MPASS(pi.ipi_new_pidx != pidx);
2692 		MPASS(ndesc > 0);
2693 		txq->ift_in_use += ndesc;
2694 		/*
2695 		 * We update the last software descriptor again here because there may
2696 		 * be a sentinel and/or there may be more mbufs than segments
2697 		 */
2698 		txq->ift_pidx = pi.ipi_new_pidx;
2699 		txq->ift_npending += pi.ipi_ndescs;
2700 	} else if (__predict_false(err == EFBIG && remap < 2)) {
2701 		*m_headp = m_head = iflib_remove_mbuf(txq);
2702 		remap = 1;
2703 		txq->ift_txd_encap_efbig++;
2704 		goto defrag;
2705 	} else
2706 		DBG_COUNTER_INC(encap_txd_encap_fail);
2707 	return (err);
2708 
2709 defrag_failed:
2710 	txq->ift_mbuf_defrag_failed++;
2711 	txq->ift_map_failed++;
2712 	m_freem(*m_headp);
2713 	DBG_COUNTER_INC(tx_frees);
2714 	*m_headp = NULL;
2715 	return (ENOMEM);
2716 }
2717 
2718 /* forward compatibility for cxgb */
2719 #define FIRST_QSET(ctx) 0
2720 
2721 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2722 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2723 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2724 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2725 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2726 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2727 
2728 
2729 
2730 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring
2731  * doorbell writes
2732  *
2733  * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic
2734  */
2735 #define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2)
2736 
2737 static inline int
2738 iflib_txq_min_occupancy(iflib_txq_t txq)
2739 {
2740 	if_ctx_t ctx;
2741 
2742 	ctx = txq->ift_ctx;
2743 	return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx,
2744 	    txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) +
2745 	    MAX_TX_DESC(ctx));
2746 }
2747 
2748 static void
2749 iflib_tx_desc_free(iflib_txq_t txq, int n)
2750 {
2751 	int hasmap;
2752 	uint32_t qsize, cidx, mask, gen;
2753 	struct mbuf *m, **ifsd_m;
2754 	uint8_t *ifsd_flags;
2755 	bus_dmamap_t *ifsd_map;
2756 
2757 	cidx = txq->ift_cidx;
2758 	gen = txq->ift_gen;
2759 	qsize = txq->ift_size;
2760 	mask = qsize-1;
2761 	hasmap = txq->ift_sds.ifsd_map != NULL;
2762 	ifsd_flags = txq->ift_sds.ifsd_flags;
2763 	ifsd_m = txq->ift_sds.ifsd_m;
2764 	ifsd_map = txq->ift_sds.ifsd_map;
2765 
2766 	while (n--) {
2767 		prefetch(ifsd_m[(cidx + 3) & mask]);
2768 		prefetch(ifsd_m[(cidx + 4) & mask]);
2769 
2770 		if (ifsd_m[cidx] != NULL) {
2771 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
2772 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
2773 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
2774 				/*
2775 				 * does it matter if it's not the TSO tag? If so we'll
2776 				 * have to add the type to flags
2777 				 */
2778 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
2779 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
2780 			}
2781 			if ((m = ifsd_m[cidx]) != NULL) {
2782 				/* XXX we don't support any drivers that batch packets yet */
2783 				MPASS(m->m_nextpkt == NULL);
2784 
2785 				m_free(m);
2786 				ifsd_m[cidx] = NULL;
2787 #if MEMORY_LOGGING
2788 				txq->ift_dequeued++;
2789 #endif
2790 				DBG_COUNTER_INC(tx_frees);
2791 			}
2792 		}
2793 		if (__predict_false(++cidx == qsize)) {
2794 			cidx = 0;
2795 			gen = 0;
2796 		}
2797 	}
2798 	txq->ift_cidx = cidx;
2799 	txq->ift_gen = gen;
2800 }
2801 
2802 static __inline int
2803 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
2804 {
2805 	int reclaim;
2806 	if_ctx_t ctx = txq->ift_ctx;
2807 
2808 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
2809 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
2810 
2811 	/*
2812 	 * Need a rate-limiting check so that this isn't called every time
2813 	 */
2814 	iflib_tx_credits_update(ctx, txq);
2815 	reclaim = DESC_RECLAIMABLE(txq);
2816 
2817 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
2818 #ifdef INVARIANTS
2819 		if (iflib_verbose_debug) {
2820 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
2821 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
2822 			       reclaim, thresh);
2823 
2824 		}
2825 #endif
2826 		return (0);
2827 	}
2828 	iflib_tx_desc_free(txq, reclaim);
2829 	txq->ift_cleaned += reclaim;
2830 	txq->ift_in_use -= reclaim;
2831 
2832 	if (txq->ift_active == FALSE)
2833 		txq->ift_active = TRUE;
2834 
2835 	return (reclaim);
2836 }
2837 
2838 static struct mbuf **
2839 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset)
2840 {
2841 
2842 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)]));
2843 }
2844 
2845 static void
2846 iflib_txq_check_drain(iflib_txq_t txq, int budget)
2847 {
2848 
2849 	ifmp_ring_check_drainage(txq->ift_br[0], budget);
2850 }
2851 
2852 static uint32_t
2853 iflib_txq_can_drain(struct ifmp_ring *r)
2854 {
2855 	iflib_txq_t txq = r->cookie;
2856 	if_ctx_t ctx = txq->ift_ctx;
2857 
2858 	return ((TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx)) ||
2859 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false));
2860 }
2861 
2862 static uint32_t
2863 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2864 {
2865 	iflib_txq_t txq = r->cookie;
2866 	if_ctx_t ctx = txq->ift_ctx;
2867 	if_t ifp = ctx->ifc_ifp;
2868 	struct mbuf **mp, *m;
2869 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used;
2870 
2871 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
2872 			    !LINK_ACTIVE(ctx))) {
2873 		DBG_COUNTER_INC(txq_drain_notready);
2874 		return (0);
2875 	}
2876 
2877 	avail = IDXDIFF(pidx, cidx, r->size);
2878 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
2879 		DBG_COUNTER_INC(txq_drain_flushing);
2880 		for (i = 0; i < avail; i++) {
2881 			m_free(r->items[(cidx + i) & (r->size-1)]);
2882 			r->items[(cidx + i) & (r->size-1)] = NULL;
2883 		}
2884 		return (avail);
2885 	}
2886 	iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
2887 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
2888 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
2889 		CALLOUT_LOCK(txq);
2890 		callout_stop(&txq->ift_timer);
2891 		callout_stop(&txq->ift_db_check);
2892 		CALLOUT_UNLOCK(txq);
2893 		DBG_COUNTER_INC(txq_drain_oactive);
2894 		return (0);
2895 	}
2896 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
2897 	count = MIN(avail, TX_BATCH_SIZE);
2898 #ifdef INVARIANTS
2899 	if (iflib_verbose_debug)
2900 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
2901 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
2902 #endif
2903 
2904 	for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) {
2905 		mp = _ring_peek_one(r, cidx, i);
2906 		MPASS(mp != NULL && *mp != NULL);
2907 		in_use_prev = txq->ift_in_use;
2908 		if ((err = iflib_encap(txq, mp)) == ENOBUFS) {
2909 			DBG_COUNTER_INC(txq_drain_encapfail);
2910 			/* no room - bail out */
2911 			break;
2912 		}
2913 		consumed++;
2914 		if (err) {
2915 			DBG_COUNTER_INC(txq_drain_encapfail);
2916 			/* we can't send this packet - skip it */
2917 			continue;
2918 		}
2919 		pkt_sent++;
2920 		m = *mp;
2921 		DBG_COUNTER_INC(tx_sent);
2922 		bytes_sent += m->m_pkthdr.len;
2923 		if (m->m_flags & M_MCAST)
2924 			mcast_sent++;
2925 
2926 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
2927 		desc_used += (txq->ift_in_use - in_use_prev);
2928 		iflib_txd_db_check(ctx, txq, FALSE);
2929 		ETHER_BPF_MTAP(ifp, m);
2930 		if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
2931 			break;
2932 
2933 		if (desc_used > TXQ_MAX_DB_CONSUMED(txq->ift_size))
2934 			break;
2935 	}
2936 
2937 	if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending)
2938 		iflib_txd_db_check(ctx, txq, TRUE);
2939 	else if ((txq->ift_db_pending || TXQ_AVAIL(txq) < MAX_TX_DESC(ctx)) &&
2940 		 (callout_pending(&txq->ift_db_check) == 0)) {
2941 		txq->ift_db_pending_queued = txq->ift_db_pending;
2942 		callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check,
2943 				 txq, txq->ift_db_check.c_cpu);
2944 	}
2945 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
2946 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
2947 	if (mcast_sent)
2948 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
2949 #ifdef INVARIANTS
2950 	if (iflib_verbose_debug)
2951 		printf("consumed=%d\n", consumed);
2952 #endif
2953 	return (consumed);
2954 }
2955 
2956 static uint32_t
2957 iflib_txq_drain_always(struct ifmp_ring *r)
2958 {
2959 	return (1);
2960 }
2961 
2962 static uint32_t
2963 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2964 {
2965 	int i, avail;
2966 	struct mbuf **mp;
2967 	iflib_txq_t txq;
2968 
2969 	txq = r->cookie;
2970 
2971 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
2972 	CALLOUT_LOCK(txq);
2973 	callout_stop(&txq->ift_timer);
2974 	callout_stop(&txq->ift_db_check);
2975 	CALLOUT_UNLOCK(txq);
2976 
2977 	avail = IDXDIFF(pidx, cidx, r->size);
2978 	for (i = 0; i < avail; i++) {
2979 		mp = _ring_peek_one(r, cidx, i);
2980 		m_freem(*mp);
2981 	}
2982 	MPASS(ifmp_ring_is_stalled(r) == 0);
2983 	return (avail);
2984 }
2985 
2986 static void
2987 iflib_ifmp_purge(iflib_txq_t txq)
2988 {
2989 	struct ifmp_ring *r;
2990 
2991 	r = txq->ift_br[0];
2992 	r->drain = iflib_txq_drain_free;
2993 	r->can_drain = iflib_txq_drain_always;
2994 
2995 	ifmp_ring_check_drainage(r, r->size);
2996 
2997 	r->drain = iflib_txq_drain;
2998 	r->can_drain = iflib_txq_can_drain;
2999 }
3000 
3001 static void
3002 _task_fn_tx(void *context)
3003 {
3004 	iflib_txq_t txq = context;
3005 	if_ctx_t ctx = txq->ift_ctx;
3006 
3007 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3008 		return;
3009 	ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3010 }
3011 
3012 static void
3013 _task_fn_rx(void *context)
3014 {
3015 	iflib_rxq_t rxq = context;
3016 	if_ctx_t ctx = rxq->ifr_ctx;
3017 	bool more;
3018 	int rc;
3019 
3020 	DBG_COUNTER_INC(task_fn_rxs);
3021 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3022 		return;
3023 
3024 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3025 		if (ctx->ifc_flags & IFC_LEGACY)
3026 			IFDI_INTR_ENABLE(ctx);
3027 		else {
3028 			DBG_COUNTER_INC(rx_intr_enables);
3029 			rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3030 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3031 		}
3032 	}
3033 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3034 		return;
3035 	if (more)
3036 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3037 }
3038 
3039 static void
3040 _task_fn_admin(void *context)
3041 {
3042 	if_ctx_t ctx = context;
3043 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3044 	iflib_txq_t txq;
3045 	int i;
3046 
3047 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3048 		return;
3049 
3050 	CTX_LOCK(ctx);
3051 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3052 		CALLOUT_LOCK(txq);
3053 		callout_stop(&txq->ift_timer);
3054 		CALLOUT_UNLOCK(txq);
3055 	}
3056 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3057 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3058 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3059 	IFDI_LINK_INTR_ENABLE(ctx);
3060 	CTX_UNLOCK(ctx);
3061 
3062 	if (LINK_ACTIVE(ctx) == 0)
3063 		return;
3064 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3065 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3066 }
3067 
3068 
3069 static void
3070 _task_fn_iov(void *context)
3071 {
3072 	if_ctx_t ctx = context;
3073 
3074 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3075 		return;
3076 
3077 	CTX_LOCK(ctx);
3078 	IFDI_VFLR_HANDLE(ctx);
3079 	CTX_UNLOCK(ctx);
3080 }
3081 
3082 static int
3083 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3084 {
3085 	int err;
3086 	if_int_delay_info_t info;
3087 	if_ctx_t ctx;
3088 
3089 	info = (if_int_delay_info_t)arg1;
3090 	ctx = info->iidi_ctx;
3091 	info->iidi_req = req;
3092 	info->iidi_oidp = oidp;
3093 	CTX_LOCK(ctx);
3094 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3095 	CTX_UNLOCK(ctx);
3096 	return (err);
3097 }
3098 
3099 /*********************************************************************
3100  *
3101  *  IFNET FUNCTIONS
3102  *
3103  **********************************************************************/
3104 
3105 static void
3106 iflib_if_init_locked(if_ctx_t ctx)
3107 {
3108 	iflib_stop(ctx);
3109 	iflib_init_locked(ctx);
3110 }
3111 
3112 
3113 static void
3114 iflib_if_init(void *arg)
3115 {
3116 	if_ctx_t ctx = arg;
3117 
3118 	CTX_LOCK(ctx);
3119 	iflib_if_init_locked(ctx);
3120 	CTX_UNLOCK(ctx);
3121 }
3122 
3123 static int
3124 iflib_if_transmit(if_t ifp, struct mbuf *m)
3125 {
3126 	if_ctx_t	ctx = if_getsoftc(ifp);
3127 
3128 	iflib_txq_t txq;
3129 	int err, qidx;
3130 
3131 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3132 		DBG_COUNTER_INC(tx_frees);
3133 		m_freem(m);
3134 		return (ENOBUFS);
3135 	}
3136 
3137 	MPASS(m->m_nextpkt == NULL);
3138 	qidx = 0;
3139 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3140 		qidx = QIDX(ctx, m);
3141 	/*
3142 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3143 	 */
3144 	txq = &ctx->ifc_txqs[qidx];
3145 
3146 #ifdef DRIVER_BACKPRESSURE
3147 	if (txq->ift_closed) {
3148 		while (m != NULL) {
3149 			next = m->m_nextpkt;
3150 			m->m_nextpkt = NULL;
3151 			m_freem(m);
3152 			m = next;
3153 		}
3154 		return (ENOBUFS);
3155 	}
3156 #endif
3157 #ifdef notyet
3158 	qidx = count = 0;
3159 	mp = marr;
3160 	next = m;
3161 	do {
3162 		count++;
3163 		next = next->m_nextpkt;
3164 	} while (next != NULL);
3165 
3166 	if (count > nitems(marr))
3167 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3168 			/* XXX check nextpkt */
3169 			m_freem(m);
3170 			/* XXX simplify for now */
3171 			DBG_COUNTER_INC(tx_frees);
3172 			return (ENOBUFS);
3173 		}
3174 	for (next = m, i = 0; next != NULL; i++) {
3175 		mp[i] = next;
3176 		next = next->m_nextpkt;
3177 		mp[i]->m_nextpkt = NULL;
3178 	}
3179 #endif
3180 	DBG_COUNTER_INC(tx_seen);
3181 	err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE);
3182 
3183 	if (err) {
3184 		GROUPTASK_ENQUEUE(&txq->ift_task);
3185 		/* support forthcoming later */
3186 #ifdef DRIVER_BACKPRESSURE
3187 		txq->ift_closed = TRUE;
3188 #endif
3189 		ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3190 		m_freem(m);
3191 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3192 		GROUPTASK_ENQUEUE(&txq->ift_task);
3193 	}
3194 
3195 	return (err);
3196 }
3197 
3198 static void
3199 iflib_if_qflush(if_t ifp)
3200 {
3201 	if_ctx_t ctx = if_getsoftc(ifp);
3202 	iflib_txq_t txq = ctx->ifc_txqs;
3203 	int i;
3204 
3205 	CTX_LOCK(ctx);
3206 	ctx->ifc_flags |= IFC_QFLUSH;
3207 	CTX_UNLOCK(ctx);
3208 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3209 		while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0])))
3210 			iflib_txq_check_drain(txq, 0);
3211 	CTX_LOCK(ctx);
3212 	ctx->ifc_flags &= ~IFC_QFLUSH;
3213 	CTX_UNLOCK(ctx);
3214 
3215 	if_qflush(ifp);
3216 }
3217 
3218 
3219 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3220 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3221 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3222 
3223 static int
3224 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3225 {
3226 	if_ctx_t ctx = if_getsoftc(ifp);
3227 	struct ifreq	*ifr = (struct ifreq *)data;
3228 #if defined(INET) || defined(INET6)
3229 	struct ifaddr	*ifa = (struct ifaddr *)data;
3230 #endif
3231 	bool		avoid_reset = FALSE;
3232 	int		err = 0, reinit = 0, bits;
3233 
3234 	switch (command) {
3235 	case SIOCSIFADDR:
3236 #ifdef INET
3237 		if (ifa->ifa_addr->sa_family == AF_INET)
3238 			avoid_reset = TRUE;
3239 #endif
3240 #ifdef INET6
3241 		if (ifa->ifa_addr->sa_family == AF_INET6)
3242 			avoid_reset = TRUE;
3243 #endif
3244 		/*
3245 		** Calling init results in link renegotiation,
3246 		** so we avoid doing it when possible.
3247 		*/
3248 		if (avoid_reset) {
3249 			if_setflagbits(ifp, IFF_UP,0);
3250 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3251 				reinit = 1;
3252 #ifdef INET
3253 			if (!(if_getflags(ifp) & IFF_NOARP))
3254 				arp_ifinit(ifp, ifa);
3255 #endif
3256 		} else
3257 			err = ether_ioctl(ifp, command, data);
3258 		break;
3259 	case SIOCSIFMTU:
3260 		CTX_LOCK(ctx);
3261 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3262 			CTX_UNLOCK(ctx);
3263 			break;
3264 		}
3265 		bits = if_getdrvflags(ifp);
3266 		/* stop the driver and free any clusters before proceeding */
3267 		iflib_stop(ctx);
3268 
3269 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3270 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3271 				ctx->ifc_flags |= IFC_MULTISEG;
3272 			else
3273 				ctx->ifc_flags &= ~IFC_MULTISEG;
3274 			err = if_setmtu(ifp, ifr->ifr_mtu);
3275 		}
3276 		iflib_init_locked(ctx);
3277 		if_setdrvflags(ifp, bits);
3278 		CTX_UNLOCK(ctx);
3279 		break;
3280 	case SIOCSIFFLAGS:
3281 		CTX_LOCK(ctx);
3282 		if (if_getflags(ifp) & IFF_UP) {
3283 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3284 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3285 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3286 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3287 				}
3288 			} else
3289 				reinit = 1;
3290 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3291 			iflib_stop(ctx);
3292 		}
3293 		ctx->ifc_if_flags = if_getflags(ifp);
3294 		CTX_UNLOCK(ctx);
3295 		break;
3296 
3297 		break;
3298 	case SIOCADDMULTI:
3299 	case SIOCDELMULTI:
3300 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3301 			CTX_LOCK(ctx);
3302 			IFDI_INTR_DISABLE(ctx);
3303 			IFDI_MULTI_SET(ctx);
3304 			IFDI_INTR_ENABLE(ctx);
3305 			CTX_UNLOCK(ctx);
3306 		}
3307 		break;
3308 	case SIOCSIFMEDIA:
3309 		CTX_LOCK(ctx);
3310 		IFDI_MEDIA_SET(ctx);
3311 		CTX_UNLOCK(ctx);
3312 		/* falls thru */
3313 	case SIOCGIFMEDIA:
3314 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3315 		break;
3316 	case SIOCGI2C:
3317 	{
3318 		struct ifi2creq i2c;
3319 
3320 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3321 		if (err != 0)
3322 			break;
3323 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3324 			err = EINVAL;
3325 			break;
3326 		}
3327 		if (i2c.len > sizeof(i2c.data)) {
3328 			err = EINVAL;
3329 			break;
3330 		}
3331 
3332 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3333 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3334 		break;
3335 	}
3336 	case SIOCSIFCAP:
3337 	{
3338 		int mask, setmask;
3339 
3340 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3341 		setmask = 0;
3342 #ifdef TCP_OFFLOAD
3343 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3344 #endif
3345 		setmask |= (mask & IFCAP_FLAGS);
3346 
3347 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3348 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3349 		if ((mask & IFCAP_WOL) &&
3350 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3351 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3352 		if_vlancap(ifp);
3353 		/*
3354 		 * want to ensure that traffic has stopped before we change any of the flags
3355 		 */
3356 		if (setmask) {
3357 			CTX_LOCK(ctx);
3358 			bits = if_getdrvflags(ifp);
3359 			if (bits & IFF_DRV_RUNNING)
3360 				iflib_stop(ctx);
3361 			if_togglecapenable(ifp, setmask);
3362 			if (bits & IFF_DRV_RUNNING)
3363 				iflib_init_locked(ctx);
3364 			if_setdrvflags(ifp, bits);
3365 			CTX_UNLOCK(ctx);
3366 		}
3367 		break;
3368 	    }
3369 	case SIOCGPRIVATE_0:
3370 	case SIOCSDRVSPEC:
3371 	case SIOCGDRVSPEC:
3372 		CTX_LOCK(ctx);
3373 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3374 		CTX_UNLOCK(ctx);
3375 		break;
3376 	default:
3377 		err = ether_ioctl(ifp, command, data);
3378 		break;
3379 	}
3380 	if (reinit)
3381 		iflib_if_init(ctx);
3382 	return (err);
3383 }
3384 
3385 static uint64_t
3386 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3387 {
3388 	if_ctx_t ctx = if_getsoftc(ifp);
3389 
3390 	return (IFDI_GET_COUNTER(ctx, cnt));
3391 }
3392 
3393 /*********************************************************************
3394  *
3395  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3396  *
3397  **********************************************************************/
3398 
3399 static void
3400 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3401 {
3402 	if_ctx_t ctx = if_getsoftc(ifp);
3403 
3404 	if ((void *)ctx != arg)
3405 		return;
3406 
3407 	if ((vtag == 0) || (vtag > 4095))
3408 		return;
3409 
3410 	CTX_LOCK(ctx);
3411 	IFDI_VLAN_REGISTER(ctx, vtag);
3412 	/* Re-init to load the changes */
3413 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3414 		iflib_init_locked(ctx);
3415 	CTX_UNLOCK(ctx);
3416 }
3417 
3418 static void
3419 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3420 {
3421 	if_ctx_t ctx = if_getsoftc(ifp);
3422 
3423 	if ((void *)ctx != arg)
3424 		return;
3425 
3426 	if ((vtag == 0) || (vtag > 4095))
3427 		return;
3428 
3429 	CTX_LOCK(ctx);
3430 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3431 	/* Re-init to load the changes */
3432 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3433 		iflib_init_locked(ctx);
3434 	CTX_UNLOCK(ctx);
3435 }
3436 
3437 static void
3438 iflib_led_func(void *arg, int onoff)
3439 {
3440 	if_ctx_t ctx = arg;
3441 
3442 	CTX_LOCK(ctx);
3443 	IFDI_LED_FUNC(ctx, onoff);
3444 	CTX_UNLOCK(ctx);
3445 }
3446 
3447 /*********************************************************************
3448  *
3449  *  BUS FUNCTION DEFINITIONS
3450  *
3451  **********************************************************************/
3452 
3453 int
3454 iflib_device_probe(device_t dev)
3455 {
3456 	pci_vendor_info_t *ent;
3457 
3458 	uint16_t	pci_vendor_id, pci_device_id;
3459 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3460 	uint16_t	pci_rev_id;
3461 	if_shared_ctx_t sctx;
3462 
3463 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3464 		return (ENOTSUP);
3465 
3466 	pci_vendor_id = pci_get_vendor(dev);
3467 	pci_device_id = pci_get_device(dev);
3468 	pci_subvendor_id = pci_get_subvendor(dev);
3469 	pci_subdevice_id = pci_get_subdevice(dev);
3470 	pci_rev_id = pci_get_revid(dev);
3471 	if (sctx->isc_parse_devinfo != NULL)
3472 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3473 
3474 	ent = sctx->isc_vendor_info;
3475 	while (ent->pvi_vendor_id != 0) {
3476 		if (pci_vendor_id != ent->pvi_vendor_id) {
3477 			ent++;
3478 			continue;
3479 		}
3480 		if ((pci_device_id == ent->pvi_device_id) &&
3481 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3482 		     (ent->pvi_subvendor_id == 0)) &&
3483 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3484 		     (ent->pvi_subdevice_id == 0)) &&
3485 		    ((pci_rev_id == ent->pvi_rev_id) ||
3486 		     (ent->pvi_rev_id == 0))) {
3487 
3488 			device_set_desc_copy(dev, ent->pvi_name);
3489 			/* this needs to be changed to zero if the bus probing code
3490 			 * ever stops re-probing on best match because the sctx
3491 			 * may have its values over written by register calls
3492 			 * in subsequent probes
3493 			 */
3494 			return (BUS_PROBE_DEFAULT);
3495 		}
3496 		ent++;
3497 	}
3498 	return (ENXIO);
3499 }
3500 
3501 int
3502 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3503 {
3504 	int err, rid, msix, msix_bar;
3505 	if_ctx_t ctx;
3506 	if_t ifp;
3507 	if_softc_ctx_t scctx;
3508 	int i;
3509 	uint16_t main_txq;
3510 	uint16_t main_rxq;
3511 
3512 
3513 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3514 
3515 	if (sc == NULL) {
3516 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3517 		device_set_softc(dev, ctx);
3518 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3519 	}
3520 
3521 	ctx->ifc_sctx = sctx;
3522 	ctx->ifc_dev = dev;
3523 	ctx->ifc_txrx = *sctx->isc_txrx;
3524 	ctx->ifc_softc = sc;
3525 
3526 	if ((err = iflib_register(ctx)) != 0) {
3527 		device_printf(dev, "iflib_register failed %d\n", err);
3528 		return (err);
3529 	}
3530 	iflib_add_device_sysctl_pre(ctx);
3531 
3532 	scctx = &ctx->ifc_softc_ctx;
3533 	/*
3534 	 * XXX sanity check that ntxd & nrxd are a power of 2
3535 	 */
3536 	if (ctx->ifc_sysctl_ntxqs != 0)
3537 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
3538 	if (ctx->ifc_sysctl_nrxqs != 0)
3539 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
3540 
3541 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3542 		if (ctx->ifc_sysctl_ntxds[i] != 0)
3543 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
3544 		else
3545 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
3546 	}
3547 
3548 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3549 		if (ctx->ifc_sysctl_nrxds[i] != 0)
3550 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
3551 		else
3552 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
3553 	}
3554 
3555 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3556 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
3557 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
3558 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
3559 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
3560 		}
3561 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
3562 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
3563 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
3564 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
3565 		}
3566 	}
3567 
3568 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3569 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
3570 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
3571 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
3572 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
3573 		}
3574 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
3575 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
3576 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
3577 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
3578 		}
3579 	}
3580 
3581 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
3582 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
3583 		return (err);
3584 	}
3585 	if (scctx->isc_ntxqsets_max)
3586 		scctx->isc_ntxqsets = min(scctx->isc_ntxqsets, scctx->isc_ntxqsets_max);
3587 	if (scctx->isc_nrxqsets_max)
3588 		scctx->isc_nrxqsets = min(scctx->isc_nrxqsets, scctx->isc_nrxqsets_max);
3589 
3590 #ifdef ACPI_DMAR
3591 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
3592 		ctx->ifc_flags |= IFC_DMAR;
3593 #endif
3594 
3595 	msix_bar = scctx->isc_msix_bar;
3596 
3597 	ifp = ctx->ifc_ifp;
3598 
3599 	if(sctx->isc_flags & IFLIB_HAS_TXCQ)
3600 		main_txq = 1;
3601 	else
3602 		main_txq = 0;
3603 
3604 	if(sctx->isc_flags & IFLIB_HAS_RXCQ)
3605 		main_rxq = 1;
3606 	else
3607 		main_rxq = 0;
3608 
3609 	/* XXX change for per-queue sizes */
3610 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
3611 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
3612 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3613 		if (!powerof2(scctx->isc_nrxd[i])) {
3614 			/* round down instead? */
3615 			device_printf(dev, "# rx descriptors must be a power of 2\n");
3616 			err = EINVAL;
3617 			goto fail;
3618 		}
3619 	}
3620 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3621 		if (!powerof2(scctx->isc_ntxd[i])) {
3622 			device_printf(dev,
3623 			    "# tx descriptors must be a power of 2");
3624 			err = EINVAL;
3625 			goto fail;
3626 		}
3627 	}
3628 
3629 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
3630 	    MAX_SINGLE_PACKET_FRACTION)
3631 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
3632 		    MAX_SINGLE_PACKET_FRACTION);
3633 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
3634 	    MAX_SINGLE_PACKET_FRACTION)
3635 		scctx->isc_tx_tso_segments_max = max(1,
3636 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
3637 
3638 	/*
3639 	 * Protect the stack against modern hardware
3640 	 */
3641 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
3642 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
3643 
3644 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
3645 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
3646 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
3647 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
3648 	if (scctx->isc_rss_table_size == 0)
3649 		scctx->isc_rss_table_size = 64;
3650 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
3651 
3652 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
3653 	/* XXX format name */
3654 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
3655 	/*
3656 	** Now setup MSI or MSI/X, should
3657 	** return us the number of supported
3658 	** vectors. (Will be 1 for MSI)
3659 	*/
3660 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
3661 		msix = scctx->isc_vectors;
3662 	} else if (scctx->isc_msix_bar != 0)
3663 		msix = iflib_msix_init(ctx);
3664 	else {
3665 		scctx->isc_vectors = 1;
3666 		scctx->isc_ntxqsets = 1;
3667 		scctx->isc_nrxqsets = 1;
3668 		scctx->isc_intr = IFLIB_INTR_LEGACY;
3669 		msix = 0;
3670 	}
3671 	/* Get memory for the station queues */
3672 	if ((err = iflib_queues_alloc(ctx))) {
3673 		device_printf(dev, "Unable to allocate queue memory\n");
3674 		goto fail;
3675 	}
3676 
3677 	if ((err = iflib_qset_structures_setup(ctx))) {
3678 		device_printf(dev, "qset structure setup failed %d\n", err);
3679 		goto fail_queues;
3680 	}
3681 
3682 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
3683 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
3684 		goto fail_intr_free;
3685 	}
3686 	if (msix <= 1) {
3687 		rid = 0;
3688 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
3689 			MPASS(msix == 1);
3690 			rid = 1;
3691 		}
3692 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
3693 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
3694 			goto fail_intr_free;
3695 		}
3696 	}
3697 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
3698 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
3699 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
3700 		goto fail_detach;
3701 	}
3702 	if ((err = iflib_netmap_attach(ctx))) {
3703 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
3704 		goto fail_detach;
3705 	}
3706 	*ctxp = ctx;
3707 
3708 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
3709 	iflib_add_device_sysctl_post(ctx);
3710 	return (0);
3711 fail_detach:
3712 	ether_ifdetach(ctx->ifc_ifp);
3713 fail_intr_free:
3714 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
3715 		pci_release_msi(ctx->ifc_dev);
3716 fail_queues:
3717 	/* XXX free queues */
3718 fail:
3719 	IFDI_DETACH(ctx);
3720 	return (err);
3721 }
3722 
3723 int
3724 iflib_device_attach(device_t dev)
3725 {
3726 	if_ctx_t ctx;
3727 	if_shared_ctx_t sctx;
3728 
3729 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3730 		return (ENOTSUP);
3731 
3732 	pci_enable_busmaster(dev);
3733 
3734 	return (iflib_device_register(dev, NULL, sctx, &ctx));
3735 }
3736 
3737 int
3738 iflib_device_deregister(if_ctx_t ctx)
3739 {
3740 	if_t ifp = ctx->ifc_ifp;
3741 	iflib_txq_t txq;
3742 	iflib_rxq_t rxq;
3743 	device_t dev = ctx->ifc_dev;
3744 	int i;
3745 	struct taskqgroup *tqg;
3746 
3747 	/* Make sure VLANS are not using driver */
3748 	if (if_vlantrunkinuse(ifp)) {
3749 		device_printf(dev,"Vlan in use, detach first\n");
3750 		return (EBUSY);
3751 	}
3752 
3753 	CTX_LOCK(ctx);
3754 	ctx->ifc_in_detach = 1;
3755 	iflib_stop(ctx);
3756 	CTX_UNLOCK(ctx);
3757 
3758 	/* Unregister VLAN events */
3759 	if (ctx->ifc_vlan_attach_event != NULL)
3760 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
3761 	if (ctx->ifc_vlan_detach_event != NULL)
3762 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
3763 
3764 	iflib_netmap_detach(ifp);
3765 	ether_ifdetach(ifp);
3766 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
3767 	CTX_LOCK_DESTROY(ctx);
3768 	if (ctx->ifc_led_dev != NULL)
3769 		led_destroy(ctx->ifc_led_dev);
3770 	/* XXX drain any dependent tasks */
3771 	tqg = qgroup_if_io_tqg;
3772 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
3773 		callout_drain(&txq->ift_timer);
3774 		callout_drain(&txq->ift_db_check);
3775 		if (txq->ift_task.gt_uniq != NULL)
3776 			taskqgroup_detach(tqg, &txq->ift_task);
3777 	}
3778 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
3779 		if (rxq->ifr_task.gt_uniq != NULL)
3780 			taskqgroup_detach(tqg, &rxq->ifr_task);
3781 	}
3782 	tqg = qgroup_if_config_tqg;
3783 	if (ctx->ifc_admin_task.gt_uniq != NULL)
3784 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
3785 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
3786 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
3787 
3788 	IFDI_DETACH(ctx);
3789 	device_set_softc(ctx->ifc_dev, NULL);
3790 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
3791 		pci_release_msi(dev);
3792 	}
3793 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
3794 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
3795 	}
3796 	if (ctx->ifc_msix_mem != NULL) {
3797 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
3798 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
3799 		ctx->ifc_msix_mem = NULL;
3800 	}
3801 
3802 	bus_generic_detach(dev);
3803 	if_free(ifp);
3804 
3805 	iflib_tx_structures_free(ctx);
3806 	iflib_rx_structures_free(ctx);
3807 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
3808 		free(ctx->ifc_softc, M_IFLIB);
3809 	free(ctx, M_IFLIB);
3810 	return (0);
3811 }
3812 
3813 
3814 int
3815 iflib_device_detach(device_t dev)
3816 {
3817 	if_ctx_t ctx = device_get_softc(dev);
3818 
3819 	return (iflib_device_deregister(ctx));
3820 }
3821 
3822 int
3823 iflib_device_suspend(device_t dev)
3824 {
3825 	if_ctx_t ctx = device_get_softc(dev);
3826 
3827 	CTX_LOCK(ctx);
3828 	IFDI_SUSPEND(ctx);
3829 	CTX_UNLOCK(ctx);
3830 
3831 	return bus_generic_suspend(dev);
3832 }
3833 int
3834 iflib_device_shutdown(device_t dev)
3835 {
3836 	if_ctx_t ctx = device_get_softc(dev);
3837 
3838 	CTX_LOCK(ctx);
3839 	IFDI_SHUTDOWN(ctx);
3840 	CTX_UNLOCK(ctx);
3841 
3842 	return bus_generic_suspend(dev);
3843 }
3844 
3845 
3846 int
3847 iflib_device_resume(device_t dev)
3848 {
3849 	if_ctx_t ctx = device_get_softc(dev);
3850 	iflib_txq_t txq = ctx->ifc_txqs;
3851 
3852 	CTX_LOCK(ctx);
3853 	IFDI_RESUME(ctx);
3854 	iflib_init_locked(ctx);
3855 	CTX_UNLOCK(ctx);
3856 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
3857 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3858 
3859 	return (bus_generic_resume(dev));
3860 }
3861 
3862 int
3863 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
3864 {
3865 	int error;
3866 	if_ctx_t ctx = device_get_softc(dev);
3867 
3868 	CTX_LOCK(ctx);
3869 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
3870 	CTX_UNLOCK(ctx);
3871 
3872 	return (error);
3873 }
3874 
3875 void
3876 iflib_device_iov_uninit(device_t dev)
3877 {
3878 	if_ctx_t ctx = device_get_softc(dev);
3879 
3880 	CTX_LOCK(ctx);
3881 	IFDI_IOV_UNINIT(ctx);
3882 	CTX_UNLOCK(ctx);
3883 }
3884 
3885 int
3886 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
3887 {
3888 	int error;
3889 	if_ctx_t ctx = device_get_softc(dev);
3890 
3891 	CTX_LOCK(ctx);
3892 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
3893 	CTX_UNLOCK(ctx);
3894 
3895 	return (error);
3896 }
3897 
3898 /*********************************************************************
3899  *
3900  *  MODULE FUNCTION DEFINITIONS
3901  *
3902  **********************************************************************/
3903 
3904 /*
3905  * - Start a fast taskqueue thread for each core
3906  * - Start a taskqueue for control operations
3907  */
3908 static int
3909 iflib_module_init(void)
3910 {
3911 	return (0);
3912 }
3913 
3914 static int
3915 iflib_module_event_handler(module_t mod, int what, void *arg)
3916 {
3917 	int err;
3918 
3919 	switch (what) {
3920 	case MOD_LOAD:
3921 		if ((err = iflib_module_init()) != 0)
3922 			return (err);
3923 		break;
3924 	case MOD_UNLOAD:
3925 		return (EBUSY);
3926 	default:
3927 		return (EOPNOTSUPP);
3928 	}
3929 
3930 	return (0);
3931 }
3932 
3933 /*********************************************************************
3934  *
3935  *  PUBLIC FUNCTION DEFINITIONS
3936  *     ordered as in iflib.h
3937  *
3938  **********************************************************************/
3939 
3940 
3941 static void
3942 _iflib_assert(if_shared_ctx_t sctx)
3943 {
3944 	MPASS(sctx->isc_tx_maxsize);
3945 	MPASS(sctx->isc_tx_maxsegsize);
3946 
3947 	MPASS(sctx->isc_rx_maxsize);
3948 	MPASS(sctx->isc_rx_nsegments);
3949 	MPASS(sctx->isc_rx_maxsegsize);
3950 
3951 
3952 	MPASS(sctx->isc_txrx->ift_txd_encap);
3953 	MPASS(sctx->isc_txrx->ift_txd_flush);
3954 	MPASS(sctx->isc_txrx->ift_txd_credits_update);
3955 	MPASS(sctx->isc_txrx->ift_rxd_available);
3956 	MPASS(sctx->isc_txrx->ift_rxd_pkt_get);
3957 	MPASS(sctx->isc_txrx->ift_rxd_refill);
3958 	MPASS(sctx->isc_txrx->ift_rxd_flush);
3959 
3960 	MPASS(sctx->isc_nrxd_min[0]);
3961 	MPASS(sctx->isc_nrxd_max[0]);
3962 	MPASS(sctx->isc_nrxd_default[0]);
3963 	MPASS(sctx->isc_ntxd_min[0]);
3964 	MPASS(sctx->isc_ntxd_max[0]);
3965 	MPASS(sctx->isc_ntxd_default[0]);
3966 }
3967 
3968 #define DEFAULT_CAPS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3969 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3970 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO | IFCAP_HWSTATS)
3971 
3972 static int
3973 iflib_register(if_ctx_t ctx)
3974 {
3975 	if_shared_ctx_t sctx = ctx->ifc_sctx;
3976 	driver_t *driver = sctx->isc_driver;
3977 	device_t dev = ctx->ifc_dev;
3978 	if_t ifp;
3979 
3980 	_iflib_assert(sctx);
3981 
3982 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
3983 
3984 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
3985 	if (ifp == NULL) {
3986 		device_printf(dev, "can not allocate ifnet structure\n");
3987 		return (ENOMEM);
3988 	}
3989 
3990 	/*
3991 	 * Initialize our context's device specific methods
3992 	 */
3993 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
3994 	kobj_class_compile((kobj_class_t) driver);
3995 	driver->refs++;
3996 
3997 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
3998 	if_setsoftc(ifp, ctx);
3999 	if_setdev(ifp, dev);
4000 	if_setinitfn(ifp, iflib_if_init);
4001 	if_setioctlfn(ifp, iflib_if_ioctl);
4002 	if_settransmitfn(ifp, iflib_if_transmit);
4003 	if_setqflushfn(ifp, iflib_if_qflush);
4004 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4005 
4006 	/* XXX - move this in to the driver for non-default settings */
4007 	if_setcapabilities(ifp, DEFAULT_CAPS);
4008 	if_setcapenable(ifp, DEFAULT_CAPS);
4009 
4010 	ctx->ifc_vlan_attach_event =
4011 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4012 							  EVENTHANDLER_PRI_FIRST);
4013 	ctx->ifc_vlan_detach_event =
4014 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4015 							  EVENTHANDLER_PRI_FIRST);
4016 
4017 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4018 					 iflib_media_change, iflib_media_status);
4019 
4020 	return (0);
4021 }
4022 
4023 
4024 static int
4025 iflib_queues_alloc(if_ctx_t ctx)
4026 {
4027 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4028 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4029 	device_t dev = ctx->ifc_dev;
4030 	int nrxqsets = scctx->isc_nrxqsets;
4031 	int ntxqsets = scctx->isc_ntxqsets;
4032 	iflib_txq_t txq;
4033 	iflib_rxq_t rxq;
4034 	iflib_fl_t fl = NULL;
4035 	int i, j, cpu, err, txconf, rxconf;
4036 	iflib_dma_info_t ifdip;
4037 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4038 	uint32_t *txqsizes = scctx->isc_txqsizes;
4039 	uint8_t nrxqs = sctx->isc_nrxqs;
4040 	uint8_t ntxqs = sctx->isc_ntxqs;
4041 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4042 	caddr_t *vaddrs;
4043 	uint64_t *paddrs;
4044 	struct ifmp_ring **brscp;
4045 	int nbuf_rings = 1; /* XXX determine dynamically */
4046 
4047 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4048 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4049 
4050 	brscp = NULL;
4051 	txq = NULL;
4052 	rxq = NULL;
4053 
4054 /* Allocate the TX ring struct memory */
4055 	if (!(txq =
4056 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4057 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4058 		device_printf(dev, "Unable to allocate TX ring memory\n");
4059 		err = ENOMEM;
4060 		goto fail;
4061 	}
4062 
4063 	/* Now allocate the RX */
4064 	if (!(rxq =
4065 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4066 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4067 		device_printf(dev, "Unable to allocate RX ring memory\n");
4068 		err = ENOMEM;
4069 		goto rx_fail;
4070 	}
4071 	if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4072 		device_printf(dev, "Unable to buf_ring_sc * memory\n");
4073 		err = ENOMEM;
4074 		goto rx_fail;
4075 	}
4076 
4077 	ctx->ifc_txqs = txq;
4078 	ctx->ifc_rxqs = rxq;
4079 
4080 	/*
4081 	 * XXX handle allocation failure
4082 	 */
4083 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4084 		/* Set up some basics */
4085 
4086 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4087 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4088 			err = ENOMEM;
4089 			goto err_tx_desc;
4090 		}
4091 		txq->ift_ifdi = ifdip;
4092 		for (j = 0; j < ntxqs; j++, ifdip++) {
4093 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4094 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4095 				err = ENOMEM;
4096 				goto err_tx_desc;
4097 			}
4098 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4099 		}
4100 		txq->ift_ctx = ctx;
4101 		txq->ift_id = i;
4102 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4103 			txq->ift_br_offset = 1;
4104 		} else {
4105 			txq->ift_br_offset = 0;
4106 		}
4107 		/* XXX fix this */
4108 		txq->ift_timer.c_cpu = cpu;
4109 		txq->ift_db_check.c_cpu = cpu;
4110 		txq->ift_nbr = nbuf_rings;
4111 
4112 		if (iflib_txsd_alloc(txq)) {
4113 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4114 			err = ENOMEM;
4115 			goto err_tx_desc;
4116 		}
4117 
4118 		/* Initialize the TX lock */
4119 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4120 		    device_get_nameunit(dev), txq->ift_id);
4121 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4122 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4123 		callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0);
4124 
4125 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4126 			 device_get_nameunit(dev), txq->ift_id);
4127 		TXDB_LOCK_INIT(txq);
4128 
4129 		txq->ift_br = brscp + i*nbuf_rings;
4130 		for (j = 0; j < nbuf_rings; j++) {
4131 			err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain,
4132 					      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4133 			if (err) {
4134 				/* XXX free any allocated rings */
4135 				device_printf(dev, "Unable to allocate buf_ring\n");
4136 				goto err_tx_desc;
4137 			}
4138 		}
4139 	}
4140 
4141 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4142 		/* Set up some basics */
4143 
4144 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4145 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4146 			err = ENOMEM;
4147 			goto err_tx_desc;
4148 		}
4149 
4150 		rxq->ifr_ifdi = ifdip;
4151 		for (j = 0; j < nrxqs; j++, ifdip++) {
4152 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4153 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4154 				err = ENOMEM;
4155 				goto err_tx_desc;
4156 			}
4157 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4158 		}
4159 		rxq->ifr_ctx = ctx;
4160 		rxq->ifr_id = i;
4161 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4162 			rxq->ifr_fl_offset = 1;
4163 		} else {
4164 			rxq->ifr_fl_offset = 0;
4165 		}
4166 		rxq->ifr_nfl = nfree_lists;
4167 		if (!(fl =
4168 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4169 			device_printf(dev, "Unable to allocate free list memory\n");
4170 			err = ENOMEM;
4171 			goto err_tx_desc;
4172 		}
4173 		rxq->ifr_fl = fl;
4174 		for (j = 0; j < nfree_lists; j++) {
4175 			rxq->ifr_fl[j].ifl_rxq = rxq;
4176 			rxq->ifr_fl[j].ifl_id = j;
4177 			rxq->ifr_fl[j].ifl_ifdi =
4178 			    &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4179 		}
4180         /* Allocate receive buffers for the ring*/
4181 		if (iflib_rxsd_alloc(rxq)) {
4182 			device_printf(dev,
4183 			    "Critical Failure setting up receive buffers\n");
4184 			err = ENOMEM;
4185 			goto err_rx_desc;
4186 		}
4187 	}
4188 
4189 	/* TXQs */
4190 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4191 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4192 	for (i = 0; i < ntxqsets; i++) {
4193 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4194 
4195 		for (j = 0; j < ntxqs; j++, di++) {
4196 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4197 			paddrs[i*ntxqs + j] = di->idi_paddr;
4198 		}
4199 	}
4200 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4201 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4202 		iflib_tx_structures_free(ctx);
4203 		free(vaddrs, M_IFLIB);
4204 		free(paddrs, M_IFLIB);
4205 		goto err_rx_desc;
4206 	}
4207 	free(vaddrs, M_IFLIB);
4208 	free(paddrs, M_IFLIB);
4209 
4210 	/* RXQs */
4211 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4212 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4213 	for (i = 0; i < nrxqsets; i++) {
4214 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4215 
4216 		for (j = 0; j < nrxqs; j++, di++) {
4217 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4218 			paddrs[i*nrxqs + j] = di->idi_paddr;
4219 		}
4220 	}
4221 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4222 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4223 		iflib_tx_structures_free(ctx);
4224 		free(vaddrs, M_IFLIB);
4225 		free(paddrs, M_IFLIB);
4226 		goto err_rx_desc;
4227 	}
4228 	free(vaddrs, M_IFLIB);
4229 	free(paddrs, M_IFLIB);
4230 
4231 	return (0);
4232 
4233 /* XXX handle allocation failure changes */
4234 err_rx_desc:
4235 err_tx_desc:
4236 	if (ctx->ifc_rxqs != NULL)
4237 		free(ctx->ifc_rxqs, M_IFLIB);
4238 	ctx->ifc_rxqs = NULL;
4239 	if (ctx->ifc_txqs != NULL)
4240 		free(ctx->ifc_txqs, M_IFLIB);
4241 	ctx->ifc_txqs = NULL;
4242 rx_fail:
4243 	if (brscp != NULL)
4244 		free(brscp, M_IFLIB);
4245 	if (rxq != NULL)
4246 		free(rxq, M_IFLIB);
4247 	if (txq != NULL)
4248 		free(txq, M_IFLIB);
4249 fail:
4250 	return (err);
4251 }
4252 
4253 static int
4254 iflib_tx_structures_setup(if_ctx_t ctx)
4255 {
4256 	iflib_txq_t txq = ctx->ifc_txqs;
4257 	int i;
4258 
4259 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4260 		iflib_txq_setup(txq);
4261 
4262 	return (0);
4263 }
4264 
4265 static void
4266 iflib_tx_structures_free(if_ctx_t ctx)
4267 {
4268 	iflib_txq_t txq = ctx->ifc_txqs;
4269 	int i, j;
4270 
4271 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4272 		iflib_txq_destroy(txq);
4273 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4274 			iflib_dma_free(&txq->ift_ifdi[j]);
4275 	}
4276 	free(ctx->ifc_txqs, M_IFLIB);
4277 	ctx->ifc_txqs = NULL;
4278 	IFDI_QUEUES_FREE(ctx);
4279 }
4280 
4281 /*********************************************************************
4282  *
4283  *  Initialize all receive rings.
4284  *
4285  **********************************************************************/
4286 static int
4287 iflib_rx_structures_setup(if_ctx_t ctx)
4288 {
4289 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4290 	int q;
4291 #if defined(INET6) || defined(INET)
4292 	int i, err;
4293 #endif
4294 
4295 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4296 #if defined(INET6) || defined(INET)
4297 		tcp_lro_free(&rxq->ifr_lc);
4298 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4299 		    TCP_LRO_ENTRIES, min(1024,
4300 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4301 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4302 			goto fail;
4303 		}
4304 		rxq->ifr_lro_enabled = TRUE;
4305 #endif
4306 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4307 	}
4308 	return (0);
4309 #if defined(INET6) || defined(INET)
4310 fail:
4311 	/*
4312 	 * Free RX software descriptors allocated so far, we will only handle
4313 	 * the rings that completed, the failing case will have
4314 	 * cleaned up for itself. 'q' failed, so its the terminus.
4315 	 */
4316 	rxq = ctx->ifc_rxqs;
4317 	for (i = 0; i < q; ++i, rxq++) {
4318 		iflib_rx_sds_free(rxq);
4319 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4320 	}
4321 	return (err);
4322 #endif
4323 }
4324 
4325 /*********************************************************************
4326  *
4327  *  Free all receive rings.
4328  *
4329  **********************************************************************/
4330 static void
4331 iflib_rx_structures_free(if_ctx_t ctx)
4332 {
4333 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4334 
4335 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4336 		iflib_rx_sds_free(rxq);
4337 	}
4338 }
4339 
4340 static int
4341 iflib_qset_structures_setup(if_ctx_t ctx)
4342 {
4343 	int err;
4344 
4345 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4346 		return (err);
4347 
4348 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4349 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4350 		iflib_tx_structures_free(ctx);
4351 		iflib_rx_structures_free(ctx);
4352 	}
4353 	return (err);
4354 }
4355 
4356 int
4357 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4358 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4359 {
4360 
4361 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4362 }
4363 
4364 static int
4365 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4366 {
4367 	int i, cpuid, eqid, count;
4368 
4369 	CPU_COPY(&ctx->ifc_cpus, cpus);
4370 	count = CPU_COUNT(&ctx->ifc_cpus);
4371 	eqid = qid % count;
4372 	/* clear up to the qid'th bit */
4373 	for (i = 0; i < eqid; i++) {
4374 		cpuid = CPU_FFS(cpus);
4375 		MPASS(cpuid != 0);
4376 		CPU_CLR(cpuid-1, cpus);
4377 	}
4378 	cpuid = CPU_FFS(cpus);
4379 	MPASS(cpuid != 0);
4380 	return (cpuid-1);
4381 }
4382 
4383 int
4384 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4385 						iflib_intr_type_t type, driver_filter_t *filter,
4386 						void *filter_arg, int qid, char *name)
4387 {
4388 	struct grouptask *gtask;
4389 	struct taskqgroup *tqg;
4390 	iflib_filter_info_t info;
4391 	cpuset_t cpus;
4392 	gtask_fn_t *fn;
4393 	int tqrid, err, cpuid;
4394 	void *q;
4395 
4396 	info = &ctx->ifc_filter_info;
4397 	tqrid = rid;
4398 
4399 	switch (type) {
4400 	/* XXX merge tx/rx for netmap? */
4401 	case IFLIB_INTR_TX:
4402 		q = &ctx->ifc_txqs[qid];
4403 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4404 		gtask = &ctx->ifc_txqs[qid].ift_task;
4405 		tqg = qgroup_if_io_tqg;
4406 		fn = _task_fn_tx;
4407 		GROUPTASK_INIT(gtask, 0, fn, q);
4408 		break;
4409 	case IFLIB_INTR_RX:
4410 		q = &ctx->ifc_rxqs[qid];
4411 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4412 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4413 		tqg = qgroup_if_io_tqg;
4414 		fn = _task_fn_rx;
4415 		GROUPTASK_INIT(gtask, 0, fn, q);
4416 		break;
4417 	case IFLIB_INTR_ADMIN:
4418 		q = ctx;
4419 		tqrid = -1;
4420 		info = &ctx->ifc_filter_info;
4421 		gtask = &ctx->ifc_admin_task;
4422 		tqg = qgroup_if_config_tqg;
4423 		fn = _task_fn_admin;
4424 		break;
4425 	default:
4426 		panic("unknown net intr type");
4427 	}
4428 
4429 	info->ifi_filter = filter;
4430 	info->ifi_filter_arg = filter_arg;
4431 	info->ifi_task = gtask;
4432 
4433 	err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info,  name);
4434 	if (err != 0) {
4435 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4436 		return (err);
4437 	}
4438 	if (type == IFLIB_INTR_ADMIN)
4439 		return (0);
4440 
4441 	if (tqrid != -1) {
4442 		cpuid = find_nth(ctx, &cpus, qid);
4443 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4444 	} else {
4445 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4446 	}
4447 
4448 	return (0);
4449 }
4450 
4451 void
4452 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4453 {
4454 	struct grouptask *gtask;
4455 	struct taskqgroup *tqg;
4456 	gtask_fn_t *fn;
4457 	void *q;
4458 
4459 	switch (type) {
4460 	case IFLIB_INTR_TX:
4461 		q = &ctx->ifc_txqs[qid];
4462 		gtask = &ctx->ifc_txqs[qid].ift_task;
4463 		tqg = qgroup_if_io_tqg;
4464 		fn = _task_fn_tx;
4465 		break;
4466 	case IFLIB_INTR_RX:
4467 		q = &ctx->ifc_rxqs[qid];
4468 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4469 		tqg = qgroup_if_io_tqg;
4470 		fn = _task_fn_rx;
4471 		break;
4472 	case IFLIB_INTR_IOV:
4473 		q = ctx;
4474 		gtask = &ctx->ifc_vflr_task;
4475 		tqg = qgroup_if_config_tqg;
4476 		rid = -1;
4477 		fn = _task_fn_iov;
4478 		break;
4479 	default:
4480 		panic("unknown net intr type");
4481 	}
4482 	GROUPTASK_INIT(gtask, 0, fn, q);
4483 	taskqgroup_attach(tqg, gtask, q, rid, name);
4484 }
4485 
4486 void
4487 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4488 {
4489 	if (irq->ii_tag)
4490 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4491 
4492 	if (irq->ii_res)
4493 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4494 }
4495 
4496 static int
4497 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4498 {
4499 	iflib_txq_t txq = ctx->ifc_txqs;
4500 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4501 	if_irq_t irq = &ctx->ifc_legacy_irq;
4502 	iflib_filter_info_t info;
4503 	struct grouptask *gtask;
4504 	struct taskqgroup *tqg;
4505 	gtask_fn_t *fn;
4506 	int tqrid;
4507 	void *q;
4508 	int err;
4509 
4510 	q = &ctx->ifc_rxqs[0];
4511 	info = &rxq[0].ifr_filter_info;
4512 	gtask = &rxq[0].ifr_task;
4513 	tqg = qgroup_if_io_tqg;
4514 	tqrid = irq->ii_rid = *rid;
4515 	fn = _task_fn_rx;
4516 
4517 	ctx->ifc_flags |= IFC_LEGACY;
4518 	info->ifi_filter = filter;
4519 	info->ifi_filter_arg = filter_arg;
4520 	info->ifi_task = gtask;
4521 
4522 	/* We allocate a single interrupt resource */
4523 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0)
4524 		return (err);
4525 	GROUPTASK_INIT(gtask, 0, fn, q);
4526 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
4527 
4528 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
4529 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
4530 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4531 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin/link");
4532 
4533 	return (0);
4534 }
4535 
4536 void
4537 iflib_led_create(if_ctx_t ctx)
4538 {
4539 
4540 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
4541 								  device_get_nameunit(ctx->ifc_dev));
4542 }
4543 
4544 void
4545 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
4546 {
4547 
4548 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
4549 }
4550 
4551 void
4552 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
4553 {
4554 
4555 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
4556 }
4557 
4558 void
4559 iflib_admin_intr_deferred(if_ctx_t ctx)
4560 {
4561 
4562 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
4563 }
4564 
4565 void
4566 iflib_iov_intr_deferred(if_ctx_t ctx)
4567 {
4568 
4569 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
4570 }
4571 
4572 void
4573 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
4574 {
4575 
4576 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
4577 }
4578 
4579 void
4580 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
4581 	char *name)
4582 {
4583 
4584 	GROUPTASK_INIT(gtask, 0, fn, ctx);
4585 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
4586 }
4587 
4588 void
4589 iflib_config_gtask_deinit(struct grouptask *gtask)
4590 {
4591 
4592 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
4593 }
4594 
4595 void
4596 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
4597 {
4598 	if_t ifp = ctx->ifc_ifp;
4599 	iflib_txq_t txq = ctx->ifc_txqs;
4600 
4601 
4602 	if_setbaudrate(ifp, baudrate);
4603 
4604 	/* If link down, disable watchdog */
4605 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
4606 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
4607 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4608 	}
4609 	ctx->ifc_link_state = link_state;
4610 	if_link_state_change(ifp, link_state);
4611 }
4612 
4613 static int
4614 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
4615 {
4616 	int credits;
4617 
4618 	if (ctx->isc_txd_credits_update == NULL)
4619 		return (0);
4620 
4621 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0)
4622 		return (0);
4623 
4624 	txq->ift_processed += credits;
4625 	txq->ift_cidx_processed += credits;
4626 
4627 	if (txq->ift_cidx_processed >= txq->ift_size)
4628 		txq->ift_cidx_processed -= txq->ift_size;
4629 	return (credits);
4630 }
4631 
4632 static int
4633 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget)
4634 {
4635 
4636 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
4637 	    budget));
4638 }
4639 
4640 void
4641 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
4642 	const char *description, if_int_delay_info_t info,
4643 	int offset, int value)
4644 {
4645 	info->iidi_ctx = ctx;
4646 	info->iidi_offset = offset;
4647 	info->iidi_value = value;
4648 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
4649 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
4650 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
4651 	    info, 0, iflib_sysctl_int_delay, "I", description);
4652 }
4653 
4654 struct mtx *
4655 iflib_ctx_lock_get(if_ctx_t ctx)
4656 {
4657 
4658 	return (&ctx->ifc_mtx);
4659 }
4660 
4661 static int
4662 iflib_msix_init(if_ctx_t ctx)
4663 {
4664 	device_t dev = ctx->ifc_dev;
4665 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4666 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4667 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
4668 	int iflib_num_tx_queues, iflib_num_rx_queues;
4669 	int err, admincnt, bar;
4670 
4671 	iflib_num_tx_queues = scctx->isc_ntxqsets;
4672 	iflib_num_rx_queues = scctx->isc_nrxqsets;
4673 
4674 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
4675 	admincnt = sctx->isc_admin_intrcnt;
4676 	/* Override by tuneable */
4677 	if (enable_msix == 0)
4678 		goto msi;
4679 
4680 	/*
4681 	** When used in a virtualized environment
4682 	** PCI BUSMASTER capability may not be set
4683 	** so explicity set it here and rewrite
4684 	** the ENABLE in the MSIX control register
4685 	** at this point to cause the host to
4686 	** successfully initialize us.
4687 	*/
4688 	{
4689 		uint16_t pci_cmd_word;
4690 		int msix_ctrl, rid;
4691 
4692 		rid = 0;
4693 		pci_cmd_word = pci_read_config(dev, PCIR_COMMAND, 2);
4694 		pci_cmd_word |= PCIM_CMD_BUSMASTEREN;
4695 		pci_write_config(dev, PCIR_COMMAND, pci_cmd_word, 2);
4696 		pci_find_cap(dev, PCIY_MSIX, &rid);
4697 		rid += PCIR_MSIX_CTRL;
4698 		msix_ctrl = pci_read_config(dev, rid, 2);
4699 		msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
4700 		pci_write_config(dev, rid, msix_ctrl, 2);
4701 	}
4702 
4703 	/*
4704 	 * bar == -1 => "trust me I know what I'm doing"
4705 	 * https://www.youtube.com/watch?v=nnwWKkNau4I
4706 	 * Some drivers are for hardware that is so shoddily
4707 	 * documented that no one knows which bars are which
4708 	 * so the developer has to map all bars. This hack
4709 	 * allows shoddy garbage to use msix in this framework.
4710 	 */
4711 	if (bar != -1) {
4712 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
4713 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
4714 		if (ctx->ifc_msix_mem == NULL) {
4715 			/* May not be enabled */
4716 			device_printf(dev, "Unable to map MSIX table \n");
4717 			goto msi;
4718 		}
4719 	}
4720 	/* First try MSI/X */
4721 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
4722 		device_printf(dev, "System has MSIX disabled \n");
4723 		bus_release_resource(dev, SYS_RES_MEMORY,
4724 		    bar, ctx->ifc_msix_mem);
4725 		ctx->ifc_msix_mem = NULL;
4726 		goto msi;
4727 	}
4728 #if IFLIB_DEBUG
4729 	/* use only 1 qset in debug mode */
4730 	queuemsgs = min(msgs - admincnt, 1);
4731 #else
4732 	queuemsgs = msgs - admincnt;
4733 #endif
4734 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
4735 #ifdef RSS
4736 		queues = imin(queuemsgs, rss_getnumbuckets());
4737 #else
4738 		queues = queuemsgs;
4739 #endif
4740 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
4741 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
4742 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
4743 	} else {
4744 		device_printf(dev, "Unable to fetch CPU list\n");
4745 		/* Figure out a reasonable auto config value */
4746 		queues = min(queuemsgs, mp_ncpus);
4747 	}
4748 #ifdef  RSS
4749 	/* If we're doing RSS, clamp at the number of RSS buckets */
4750 	if (queues > rss_getnumbuckets())
4751 		queues = rss_getnumbuckets();
4752 #endif
4753 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
4754 		rx_queues = iflib_num_rx_queues;
4755 	else
4756 		rx_queues = queues;
4757 	/*
4758 	 * We want this to be all logical CPUs by default
4759 	 */
4760 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
4761 		tx_queues = iflib_num_tx_queues;
4762 	else
4763 		tx_queues = mp_ncpus;
4764 
4765 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
4766 #ifdef INVARIANTS
4767 		if (tx_queues != rx_queues)
4768 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
4769 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
4770 #endif
4771 		tx_queues = min(rx_queues, tx_queues);
4772 		rx_queues = min(rx_queues, tx_queues);
4773 	}
4774 
4775 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
4776 
4777 	vectors = rx_queues + admincnt;
4778 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
4779 		device_printf(dev,
4780 					  "Using MSIX interrupts with %d vectors\n", vectors);
4781 		scctx->isc_vectors = vectors;
4782 		scctx->isc_nrxqsets = rx_queues;
4783 		scctx->isc_ntxqsets = tx_queues;
4784 		scctx->isc_intr = IFLIB_INTR_MSIX;
4785 
4786 		return (vectors);
4787 	} else {
4788 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
4789 	}
4790 msi:
4791 	vectors = pci_msi_count(dev);
4792 	scctx->isc_nrxqsets = 1;
4793 	scctx->isc_ntxqsets = 1;
4794 	scctx->isc_vectors = vectors;
4795 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
4796 		device_printf(dev,"Using an MSI interrupt\n");
4797 		scctx->isc_intr = IFLIB_INTR_MSI;
4798 	} else {
4799 		device_printf(dev,"Using a Legacy interrupt\n");
4800 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4801 	}
4802 
4803 	return (vectors);
4804 }
4805 
4806 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
4807 
4808 static int
4809 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
4810 {
4811 	int rc;
4812 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
4813 	struct sbuf *sb;
4814 	char *ring_state = "UNKNOWN";
4815 
4816 	/* XXX needed ? */
4817 	rc = sysctl_wire_old_buffer(req, 0);
4818 	MPASS(rc == 0);
4819 	if (rc != 0)
4820 		return (rc);
4821 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
4822 	MPASS(sb != NULL);
4823 	if (sb == NULL)
4824 		return (ENOMEM);
4825 	if (state[3] <= 3)
4826 		ring_state = ring_states[state[3]];
4827 
4828 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
4829 		    state[0], state[1], state[2], ring_state);
4830 	rc = sbuf_finish(sb);
4831 	sbuf_delete(sb);
4832         return(rc);
4833 }
4834 
4835 enum iflib_ndesc_handler {
4836 	IFLIB_NTXD_HANDLER,
4837 	IFLIB_NRXD_HANDLER,
4838 };
4839 
4840 static int
4841 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
4842 {
4843 	if_ctx_t ctx = (void *)arg1;
4844 	enum iflib_ndesc_handler type = arg2;
4845 	char buf[256] = {0};
4846 	uint16_t *ndesc;
4847 	char *p, *next;
4848 	int nqs, rc, i;
4849 
4850 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
4851 
4852 	nqs = 8;
4853 	switch(type) {
4854 	case IFLIB_NTXD_HANDLER:
4855 		ndesc = ctx->ifc_sysctl_ntxds;
4856 		if (ctx->ifc_sctx)
4857 			nqs = ctx->ifc_sctx->isc_ntxqs;
4858 		break;
4859 	case IFLIB_NRXD_HANDLER:
4860 		ndesc = ctx->ifc_sysctl_nrxds;
4861 		if (ctx->ifc_sctx)
4862 			nqs = ctx->ifc_sctx->isc_nrxqs;
4863 		break;
4864 	}
4865 	if (nqs == 0)
4866 		nqs = 8;
4867 
4868 	for (i=0; i<8; i++) {
4869 		if (i >= nqs)
4870 			break;
4871 		if (i)
4872 			strcat(buf, ",");
4873 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
4874 	}
4875 
4876 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
4877 	if (rc || req->newptr == NULL)
4878 		return rc;
4879 
4880 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
4881 	    i++, p = strsep(&next, " ,")) {
4882 		ndesc[i] = strtoul(p, NULL, 10);
4883 	}
4884 
4885 	return(rc);
4886 }
4887 
4888 #define NAME_BUFLEN 32
4889 static void
4890 iflib_add_device_sysctl_pre(if_ctx_t ctx)
4891 {
4892         device_t dev = iflib_get_dev(ctx);
4893 	struct sysctl_oid_list *child, *oid_list;
4894 	struct sysctl_ctx_list *ctx_list;
4895 	struct sysctl_oid *node;
4896 
4897 	ctx_list = device_get_sysctl_ctx(dev);
4898 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
4899 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
4900 						      CTLFLAG_RD, NULL, "IFLIB fields");
4901 	oid_list = SYSCTL_CHILDREN(node);
4902 
4903 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
4904 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
4905 		       "driver version");
4906 
4907 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
4908 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
4909 			"# of txqs to use, 0 => use default #");
4910 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
4911 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
4912 			"# of rxqs to use, 0 => use default #");
4913 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
4914 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
4915                        "permit #txq != #rxq");
4916 
4917 	/* XXX change for per-queue sizes */
4918 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
4919 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
4920                        mp_ndesc_handler, "A",
4921                        "list of # of tx descriptors to use, 0 = use default #");
4922 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
4923 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
4924                        mp_ndesc_handler, "A",
4925                        "list of # of rx descriptors to use, 0 = use default #");
4926 }
4927 
4928 static void
4929 iflib_add_device_sysctl_post(if_ctx_t ctx)
4930 {
4931 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4932 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4933         device_t dev = iflib_get_dev(ctx);
4934 	struct sysctl_oid_list *child;
4935 	struct sysctl_ctx_list *ctx_list;
4936 	iflib_fl_t fl;
4937 	iflib_txq_t txq;
4938 	iflib_rxq_t rxq;
4939 	int i, j;
4940 	char namebuf[NAME_BUFLEN];
4941 	char *qfmt;
4942 	struct sysctl_oid *queue_node, *fl_node, *node;
4943 	struct sysctl_oid_list *queue_list, *fl_list;
4944 	ctx_list = device_get_sysctl_ctx(dev);
4945 
4946 	node = ctx->ifc_sysctl_node;
4947 	child = SYSCTL_CHILDREN(node);
4948 
4949 	if (scctx->isc_ntxqsets > 100)
4950 		qfmt = "txq%03d";
4951 	else if (scctx->isc_ntxqsets > 10)
4952 		qfmt = "txq%02d";
4953 	else
4954 		qfmt = "txq%d";
4955 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
4956 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
4957 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
4958 					     CTLFLAG_RD, NULL, "Queue Name");
4959 		queue_list = SYSCTL_CHILDREN(queue_node);
4960 #if MEMORY_LOGGING
4961 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
4962 				CTLFLAG_RD,
4963 				&txq->ift_dequeued, "total mbufs freed");
4964 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
4965 				CTLFLAG_RD,
4966 				&txq->ift_enqueued, "total mbufs enqueued");
4967 #endif
4968 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
4969 				   CTLFLAG_RD,
4970 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
4971 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
4972 				   CTLFLAG_RD,
4973 				   &txq->ift_pullups, "# of times m_pullup was called");
4974 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
4975 				   CTLFLAG_RD,
4976 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
4977 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
4978 				   CTLFLAG_RD,
4979 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
4980 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
4981 				   CTLFLAG_RD,
4982 				   &txq->ift_map_failed, "# of times dma map failed");
4983 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
4984 				   CTLFLAG_RD,
4985 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
4986 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
4987 				   CTLFLAG_RD,
4988 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
4989 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
4990 				   CTLFLAG_RD,
4991 				   &txq->ift_pidx, 1, "Producer Index");
4992 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
4993 				   CTLFLAG_RD,
4994 				   &txq->ift_cidx, 1, "Consumer Index");
4995 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
4996 				   CTLFLAG_RD,
4997 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
4998 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
4999 				   CTLFLAG_RD,
5000 				   &txq->ift_in_use, 1, "descriptors in use");
5001 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5002 				   CTLFLAG_RD,
5003 				   &txq->ift_processed, "descriptors procesed for clean");
5004 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5005 				   CTLFLAG_RD,
5006 				   &txq->ift_cleaned, "total cleaned");
5007 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5008 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state),
5009 				0, mp_ring_state_handler, "A", "soft ring state");
5010 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5011 				       CTLFLAG_RD, &txq->ift_br[0]->enqueues,
5012 				       "# of enqueues to the mp_ring for this queue");
5013 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5014 				       CTLFLAG_RD, &txq->ift_br[0]->drops,
5015 				       "# of drops in the mp_ring for this queue");
5016 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5017 				       CTLFLAG_RD, &txq->ift_br[0]->starts,
5018 				       "# of normal consumer starts in the mp_ring for this queue");
5019 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5020 				       CTLFLAG_RD, &txq->ift_br[0]->stalls,
5021 					       "# of consumer stalls in the mp_ring for this queue");
5022 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5023 			       CTLFLAG_RD, &txq->ift_br[0]->restarts,
5024 				       "# of consumer restarts in the mp_ring for this queue");
5025 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5026 				       CTLFLAG_RD, &txq->ift_br[0]->abdications,
5027 				       "# of consumer abdications in the mp_ring for this queue");
5028 	}
5029 
5030 	if (scctx->isc_nrxqsets > 100)
5031 		qfmt = "rxq%03d";
5032 	else if (scctx->isc_nrxqsets > 10)
5033 		qfmt = "rxq%02d";
5034 	else
5035 		qfmt = "rxq%d";
5036 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5037 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5038 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5039 					     CTLFLAG_RD, NULL, "Queue Name");
5040 		queue_list = SYSCTL_CHILDREN(queue_node);
5041 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5042 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5043 				       CTLFLAG_RD,
5044 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5045 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5046 				       CTLFLAG_RD,
5047 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5048 		}
5049 
5050 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5051 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5052 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5053 						     CTLFLAG_RD, NULL, "freelist Name");
5054 			fl_list = SYSCTL_CHILDREN(fl_node);
5055 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5056 				       CTLFLAG_RD,
5057 				       &fl->ifl_pidx, 1, "Producer Index");
5058 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5059 				       CTLFLAG_RD,
5060 				       &fl->ifl_cidx, 1, "Consumer Index");
5061 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5062 				       CTLFLAG_RD,
5063 				       &fl->ifl_credits, 1, "credits available");
5064 #if MEMORY_LOGGING
5065 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5066 					CTLFLAG_RD,
5067 					&fl->ifl_m_enqueued, "mbufs allocated");
5068 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5069 					CTLFLAG_RD,
5070 					&fl->ifl_m_dequeued, "mbufs freed");
5071 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5072 					CTLFLAG_RD,
5073 					&fl->ifl_cl_enqueued, "clusters allocated");
5074 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5075 					CTLFLAG_RD,
5076 					&fl->ifl_cl_dequeued, "clusters freed");
5077 #endif
5078 
5079 		}
5080 	}
5081 
5082 }
5083