xref: /freebsd/sys/net/iflib.c (revision 63d1fd5970ec814904aa0f4580b10a0d302d08b2)
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 #ifdef IFLIB_DIAGNOSTICS
359 	uint64_t ift_cpu_exec_count[256];
360 #endif
361 } __aligned(CACHE_LINE_SIZE);
362 
363 struct iflib_fl {
364 	uint16_t	ifl_cidx;
365 	uint16_t	ifl_pidx;
366 	uint16_t	ifl_credits;
367 	uint8_t		ifl_gen;
368 #if MEMORY_LOGGING
369 	uint64_t	ifl_m_enqueued;
370 	uint64_t	ifl_m_dequeued;
371 	uint64_t	ifl_cl_enqueued;
372 	uint64_t	ifl_cl_dequeued;
373 #endif
374 	/* implicit pad */
375 
376 	/* constant */
377 	uint16_t	ifl_size;
378 	uint16_t	ifl_buf_size;
379 	uint16_t	ifl_cltype;
380 	uma_zone_t	ifl_zone;
381 	iflib_rxsd_t	ifl_sds;
382 	iflib_rxq_t	ifl_rxq;
383 	uint8_t		ifl_id;
384 	bus_dma_tag_t           ifl_desc_tag;
385 	iflib_dma_info_t	ifl_ifdi;
386 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
387 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
388 }  __aligned(CACHE_LINE_SIZE);
389 
390 static inline int
391 get_inuse(int size, int cidx, int pidx, int gen)
392 {
393 	int used;
394 
395 	if (pidx > cidx)
396 		used = pidx - cidx;
397 	else if (pidx < cidx)
398 		used = size - cidx + pidx;
399 	else if (gen == 0 && pidx == cidx)
400 		used = 0;
401 	else if (gen == 1 && pidx == cidx)
402 		used = size;
403 	else
404 		panic("bad state");
405 
406 	return (used);
407 }
408 
409 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
410 
411 #define IDXDIFF(head, tail, wrap) \
412 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
413 
414 struct iflib_rxq {
415 	/* If there is a separate completion queue -
416 	 * these are the cq cidx and pidx. Otherwise
417 	 * these are unused.
418 	 */
419 	uint16_t	ifr_size;
420 	uint16_t	ifr_cq_cidx;
421 	uint16_t	ifr_cq_pidx;
422 	uint8_t		ifr_cq_gen;
423 	uint8_t		ifr_fl_offset;
424 
425 	if_ctx_t	ifr_ctx;
426 	iflib_fl_t	ifr_fl;
427 	uint64_t	ifr_rx_irq;
428 	uint16_t	ifr_id;
429 	uint8_t		ifr_lro_enabled;
430 	uint8_t		ifr_nfl;
431 	struct lro_ctrl			ifr_lc;
432 	struct grouptask        ifr_task;
433 	struct iflib_filter_info ifr_filter_info;
434 	iflib_dma_info_t		ifr_ifdi;
435 	/* dynamically allocate if any drivers need a value substantially larger than this */
436 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
437 #ifdef IFLIB_DIAGNOSTICS
438 	uint64_t ifr_cpu_exec_count[256];
439 #endif
440 }  __aligned(CACHE_LINE_SIZE);
441 
442 /*
443  * Only allow a single packet to take up most 1/nth of the tx ring
444  */
445 #define MAX_SINGLE_PACKET_FRACTION 12
446 #define IF_BAD_DMA (bus_addr_t)-1
447 
448 static int enable_msix = 1;
449 
450 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
451 
452 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
453 
454 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
455 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
456 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
457 
458 
459 #define TXDB_LOCK_INIT(txq)  mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF)
460 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx)
461 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx)
462 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx)
463 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx)
464 
465 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
466 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
467 
468 
469 /* Our boot-time initialization hook */
470 static int	iflib_module_event_handler(module_t, int, void *);
471 
472 static moduledata_t iflib_moduledata = {
473 	"iflib",
474 	iflib_module_event_handler,
475 	NULL
476 };
477 
478 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
479 MODULE_VERSION(iflib, 1);
480 
481 MODULE_DEPEND(iflib, pci, 1, 1, 1);
482 MODULE_DEPEND(iflib, ether, 1, 1, 1);
483 
484 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
485 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
486 
487 #ifndef IFLIB_DEBUG_COUNTERS
488 #ifdef INVARIANTS
489 #define IFLIB_DEBUG_COUNTERS 1
490 #else
491 #define IFLIB_DEBUG_COUNTERS 0
492 #endif /* !INVARIANTS */
493 #endif
494 
495 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
496                    "iflib driver parameters");
497 
498 /*
499  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
500  */
501 static int iflib_min_tx_latency = 0;
502 
503 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
504 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
505 
506 
507 #if IFLIB_DEBUG_COUNTERS
508 
509 static int iflib_tx_seen;
510 static int iflib_tx_sent;
511 static int iflib_tx_encap;
512 static int iflib_rx_allocs;
513 static int iflib_fl_refills;
514 static int iflib_fl_refills_large;
515 static int iflib_tx_frees;
516 
517 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
518 		   &iflib_tx_seen, 0, "# tx mbufs seen");
519 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
520 		   &iflib_tx_sent, 0, "# tx mbufs sent");
521 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
522 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
523 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
524 		   &iflib_tx_frees, 0, "# tx frees");
525 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
526 		   &iflib_rx_allocs, 0, "# rx allocations");
527 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
528 		   &iflib_fl_refills, 0, "# refills");
529 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
530 		   &iflib_fl_refills_large, 0, "# large refills");
531 
532 
533 static int iflib_txq_drain_flushing;
534 static int iflib_txq_drain_oactive;
535 static int iflib_txq_drain_notready;
536 static int iflib_txq_drain_encapfail;
537 
538 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
539 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
540 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
541 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
542 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
543 		   &iflib_txq_drain_notready, 0, "# drain notready");
544 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
545 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
546 
547 
548 static int iflib_encap_load_mbuf_fail;
549 static int iflib_encap_txq_avail_fail;
550 static int iflib_encap_txd_encap_fail;
551 
552 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
553 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
554 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
555 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
556 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
557 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
558 
559 static int iflib_task_fn_rxs;
560 static int iflib_rx_intr_enables;
561 static int iflib_fast_intrs;
562 static int iflib_intr_link;
563 static int iflib_intr_msix;
564 static int iflib_rx_unavail;
565 static int iflib_rx_ctx_inactive;
566 static int iflib_rx_zero_len;
567 static int iflib_rx_if_input;
568 static int iflib_rx_mbuf_null;
569 static int iflib_rxd_flush;
570 
571 static int iflib_verbose_debug;
572 
573 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
574 		   &iflib_intr_link, 0, "# intr link calls");
575 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
576 		   &iflib_intr_msix, 0, "# intr msix calls");
577 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
578 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
579 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
580 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
581 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
582 		   &iflib_fast_intrs, 0, "# fast_intr calls");
583 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
584 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
585 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
586 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
587 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
588 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
589 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
590 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
591 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
592 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
593 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
594 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
595 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
596 		   &iflib_verbose_debug, 0, "enable verbose debugging");
597 
598 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
599 static void
600 iflib_debug_reset(void)
601 {
602 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
603 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
604 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
605 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
606 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
607 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
608 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
609 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
610 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
611 }
612 
613 #else
614 #define DBG_COUNTER_INC(name)
615 static void iflib_debug_reset(void) {}
616 #endif
617 
618 
619 
620 #define IFLIB_DEBUG 0
621 
622 static void iflib_tx_structures_free(if_ctx_t ctx);
623 static void iflib_rx_structures_free(if_ctx_t ctx);
624 static int iflib_queues_alloc(if_ctx_t ctx);
625 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
626 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget);
627 static int iflib_qset_structures_setup(if_ctx_t ctx);
628 static int iflib_msix_init(if_ctx_t ctx);
629 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
630 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
631 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
632 static int iflib_register(if_ctx_t);
633 static void iflib_init_locked(if_ctx_t ctx);
634 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
635 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
636 static void iflib_ifmp_purge(iflib_txq_t txq);
637 static void _iflib_pre_assert(if_softc_ctx_t scctx);
638 
639 #ifdef DEV_NETMAP
640 #include <sys/selinfo.h>
641 #include <net/netmap.h>
642 #include <dev/netmap/netmap_kern.h>
643 
644 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
645 
646 /*
647  * device-specific sysctl variables:
648  *
649  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
650  *	During regular operations the CRC is stripped, but on some
651  *	hardware reception of frames not multiple of 64 is slower,
652  *	so using crcstrip=0 helps in benchmarks.
653  *
654  * iflib_rx_miss, iflib_rx_miss_bufs:
655  *	count packets that might be missed due to lost interrupts.
656  */
657 SYSCTL_DECL(_dev_netmap);
658 /*
659  * The xl driver by default strips CRCs and we do not override it.
660  */
661 
662 int iflib_crcstrip = 1;
663 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
664     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
665 
666 int iflib_rx_miss, iflib_rx_miss_bufs;
667 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
668     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
669 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
670     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
671 
672 /*
673  * Register/unregister. We are already under netmap lock.
674  * Only called on the first register or the last unregister.
675  */
676 static int
677 iflib_netmap_register(struct netmap_adapter *na, int onoff)
678 {
679 	struct ifnet *ifp = na->ifp;
680 	if_ctx_t ctx = ifp->if_softc;
681 
682 	CTX_LOCK(ctx);
683 	IFDI_INTR_DISABLE(ctx);
684 
685 	/* Tell the stack that the interface is no longer active */
686 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
687 
688 	if (!CTX_IS_VF(ctx))
689 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
690 
691 	/* enable or disable flags and callbacks in na and ifp */
692 	if (onoff) {
693 		nm_set_native_flags(na);
694 	} else {
695 		nm_clear_native_flags(na);
696 	}
697 	IFDI_INIT(ctx);
698 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
699 	CTX_UNLOCK(ctx);
700 	return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
701 }
702 
703 /*
704  * Reconcile kernel and user view of the transmit ring.
705  *
706  * All information is in the kring.
707  * Userspace wants to send packets up to the one before kring->rhead,
708  * kernel knows kring->nr_hwcur is the first unsent packet.
709  *
710  * Here we push packets out (as many as possible), and possibly
711  * reclaim buffers from previously completed transmission.
712  *
713  * The caller (netmap) guarantees that there is only one instance
714  * running at any time. Any interference with other driver
715  * methods should be handled by the individual drivers.
716  */
717 static int
718 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
719 {
720 	struct netmap_adapter *na = kring->na;
721 	struct ifnet *ifp = na->ifp;
722 	struct netmap_ring *ring = kring->ring;
723 	u_int nm_i;	/* index into the netmap ring */
724 	u_int nic_i;	/* index into the NIC ring */
725 	u_int n;
726 	u_int const lim = kring->nkr_num_slots - 1;
727 	u_int const head = kring->rhead;
728 	struct if_pkt_info pi;
729 
730 	/*
731 	 * interrupts on every tx packet are expensive so request
732 	 * them every half ring, or where NS_REPORT is set
733 	 */
734 	u_int report_frequency = kring->nkr_num_slots >> 1;
735 	/* device-specific */
736 	if_ctx_t ctx = ifp->if_softc;
737 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
738 
739 	pi.ipi_segs = txq->ift_segs;
740 	pi.ipi_qsidx = kring->ring_id;
741 	pi.ipi_ndescs = 0;
742 
743 	bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
744 					BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
745 
746 
747 	/*
748 	 * First part: process new packets to send.
749 	 * nm_i is the current index in the netmap ring,
750 	 * nic_i is the corresponding index in the NIC ring.
751 	 *
752 	 * If we have packets to send (nm_i != head)
753 	 * iterate over the netmap ring, fetch length and update
754 	 * the corresponding slot in the NIC ring. Some drivers also
755 	 * need to update the buffer's physical address in the NIC slot
756 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
757 	 *
758 	 * The netmap_reload_map() calls is especially expensive,
759 	 * even when (as in this case) the tag is 0, so do only
760 	 * when the buffer has actually changed.
761 	 *
762 	 * If possible do not set the report/intr bit on all slots,
763 	 * but only a few times per ring or when NS_REPORT is set.
764 	 *
765 	 * Finally, on 10G and faster drivers, it might be useful
766 	 * to prefetch the next slot and txr entry.
767 	 */
768 
769 	nm_i = kring->nr_hwcur;
770 	if (nm_i != head) {	/* we have new packets to send */
771 		nic_i = netmap_idx_k2n(kring, nm_i);
772 
773 		__builtin_prefetch(&ring->slot[nm_i]);
774 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
775 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
776 
777 		for (n = 0; nm_i != head; n++) {
778 			struct netmap_slot *slot = &ring->slot[nm_i];
779 			u_int len = slot->len;
780 			uint64_t paddr;
781 			void *addr = PNMB(na, slot, &paddr);
782 			int flags = (slot->flags & NS_REPORT ||
783 				nic_i == 0 || nic_i == report_frequency) ?
784 				IPI_TX_INTR : 0;
785 
786 			/* device-specific */
787 			pi.ipi_pidx = nic_i;
788 			pi.ipi_flags = flags;
789 
790 			/* Fill the slot in the NIC ring. */
791 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
792 
793 			/* prefetch for next round */
794 			__builtin_prefetch(&ring->slot[nm_i + 1]);
795 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
796 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
797 
798 			NM_CHECK_ADDR_LEN(na, addr, len);
799 
800 			if (slot->flags & NS_BUF_CHANGED) {
801 				/* buffer has changed, reload map */
802 				netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
803 			}
804 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
805 
806 			/* make sure changes to the buffer are synced */
807 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
808 							BUS_DMASYNC_PREWRITE);
809 
810 			nm_i = nm_next(nm_i, lim);
811 			nic_i = nm_next(nic_i, lim);
812 		}
813 		kring->nr_hwcur = head;
814 
815 		/* synchronize the NIC ring */
816 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
817 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
818 
819 		/* (re)start the tx unit up to slot nic_i (excluded) */
820 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
821 	}
822 
823 	/*
824 	 * Second part: reclaim buffers for completed transmissions.
825 	 */
826 	if (iflib_tx_credits_update(ctx, txq)) {
827 		/* some tx completed, increment avail */
828 		nic_i = txq->ift_cidx_processed;
829 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
830 	}
831 	return (0);
832 }
833 
834 /*
835  * Reconcile kernel and user view of the receive ring.
836  * Same as for the txsync, this routine must be efficient.
837  * The caller guarantees a single invocations, but races against
838  * the rest of the driver should be handled here.
839  *
840  * On call, kring->rhead is the first packet that userspace wants
841  * to keep, and kring->rcur is the wakeup point.
842  * The kernel has previously reported packets up to kring->rtail.
843  *
844  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
845  * of whether or not we received an interrupt.
846  */
847 static int
848 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
849 {
850 	struct netmap_adapter *na = kring->na;
851 	struct ifnet *ifp = na->ifp;
852 	struct netmap_ring *ring = kring->ring;
853 	u_int nm_i;	/* index into the netmap ring */
854 	u_int nic_i;	/* index into the NIC ring */
855 	u_int i, n;
856 	u_int const lim = kring->nkr_num_slots - 1;
857 	u_int const head = kring->rhead;
858 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
859 	struct if_rxd_info ri;
860 	/* device-specific */
861 	if_ctx_t ctx = ifp->if_softc;
862 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
863 	iflib_fl_t fl = rxq->ifr_fl;
864 	if (head > lim)
865 		return netmap_ring_reinit(kring);
866 
867 	bzero(&ri, sizeof(ri));
868 	ri.iri_qsidx = kring->ring_id;
869 	ri.iri_ifp = ctx->ifc_ifp;
870 	/* XXX check sync modes */
871 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++)
872 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
873 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
874 
875 	/*
876 	 * First part: import newly received packets.
877 	 *
878 	 * nm_i is the index of the next free slot in the netmap ring,
879 	 * nic_i is the index of the next received packet in the NIC ring,
880 	 * and they may differ in case if_init() has been called while
881 	 * in netmap mode. For the receive ring we have
882 	 *
883 	 *	nic_i = rxr->next_check;
884 	 *	nm_i = kring->nr_hwtail (previous)
885 	 * and
886 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
887 	 *
888 	 * rxr->next_check is set to 0 on a ring reinit
889 	 */
890 	if (netmap_no_pendintr || force_update) {
891 		int crclen = iflib_crcstrip ? 0 : 4;
892 		int error, avail;
893 		uint16_t slot_flags = kring->nkr_slot_flags;
894 
895 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
896 			nic_i = fl->ifl_cidx;
897 			nm_i = netmap_idx_n2k(kring, nic_i);
898 			avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX);
899 			for (n = 0; avail > 0; n++, avail--) {
900 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
901 				if (error)
902 					ring->slot[nm_i].len = 0;
903 				else
904 					ring->slot[nm_i].len = ri.iri_len - crclen;
905 				ring->slot[nm_i].flags = slot_flags;
906 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
907 								fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_POSTREAD);
908 				nm_i = nm_next(nm_i, lim);
909 				nic_i = nm_next(nic_i, lim);
910 			}
911 			if (n) { /* update the state variables */
912 				if (netmap_no_pendintr && !force_update) {
913 					/* diagnostics */
914 					iflib_rx_miss ++;
915 					iflib_rx_miss_bufs += n;
916 				}
917 				fl->ifl_cidx = nic_i;
918 				kring->nr_hwtail = nm_i;
919 			}
920 			kring->nr_kflags &= ~NKR_PENDINTR;
921 		}
922 	}
923 	/*
924 	 * Second part: skip past packets that userspace has released.
925 	 * (kring->nr_hwcur to head excluded),
926 	 * and make the buffers available for reception.
927 	 * As usual nm_i is the index in the netmap ring,
928 	 * nic_i is the index in the NIC ring, and
929 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
930 	 */
931 	/* XXX not sure how this will work with multiple free lists */
932 	nm_i = kring->nr_hwcur;
933 	if (nm_i != head) {
934 		nic_i = netmap_idx_k2n(kring, nm_i);
935 		for (n = 0; nm_i != head; n++) {
936 			struct netmap_slot *slot = &ring->slot[nm_i];
937 			uint64_t paddr;
938 			caddr_t vaddr;
939 			void *addr = PNMB(na, slot, &paddr);
940 
941 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
942 				goto ring_reset;
943 
944 			vaddr = addr;
945 			if (slot->flags & NS_BUF_CHANGED) {
946 				/* buffer has changed, reload map */
947 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, addr);
948 				slot->flags &= ~NS_BUF_CHANGED;
949 			}
950 			/*
951 			 * XXX we should be batching this operation - TODO
952 			 */
953 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size);
954 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map,
955 			    BUS_DMASYNC_PREREAD);
956 			nm_i = nm_next(nm_i, lim);
957 			nic_i = nm_next(nic_i, lim);
958 		}
959 		kring->nr_hwcur = head;
960 
961 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
962 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
963 		/*
964 		 * IMPORTANT: we must leave one free slot in the ring,
965 		 * so move nic_i back by one unit
966 		 */
967 		nic_i = nm_prev(nic_i, lim);
968 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
969 	}
970 
971 	return 0;
972 
973 ring_reset:
974 	return netmap_ring_reinit(kring);
975 }
976 
977 static int
978 iflib_netmap_attach(if_ctx_t ctx)
979 {
980 	struct netmap_adapter na;
981 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
982 
983 	bzero(&na, sizeof(na));
984 
985 	na.ifp = ctx->ifc_ifp;
986 	na.na_flags = NAF_BDG_MAYSLEEP;
987 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
988 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
989 
990 	na.num_tx_desc = scctx->isc_ntxd[0];
991 	na.num_rx_desc = scctx->isc_nrxd[0];
992 	na.nm_txsync = iflib_netmap_txsync;
993 	na.nm_rxsync = iflib_netmap_rxsync;
994 	na.nm_register = iflib_netmap_register;
995 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
996 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
997 	return (netmap_attach(&na));
998 }
999 
1000 static void
1001 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1002 {
1003 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1004 	struct netmap_slot *slot;
1005 
1006 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1007 	if (slot == 0)
1008 		return;
1009 
1010 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1011 
1012 		/*
1013 		 * In netmap mode, set the map for the packet buffer.
1014 		 * NOTE: Some drivers (not this one) also need to set
1015 		 * the physical buffer address in the NIC ring.
1016 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1017 		 * netmap slot index, si
1018 		 */
1019 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1020 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1021 	}
1022 }
1023 static void
1024 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1025 {
1026 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1027 	struct netmap_slot *slot;
1028 	iflib_rxsd_t sd;
1029 	int nrxd;
1030 
1031 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1032 	if (slot == 0)
1033 		return;
1034 	sd = rxq->ifr_fl[0].ifl_sds;
1035 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1036 	for (int i = 0; i < nrxd; i++, sd++) {
1037 			int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1038 			uint64_t paddr;
1039 			void *addr;
1040 			caddr_t vaddr;
1041 
1042 			vaddr = addr = PNMB(na, slot + sj, &paddr);
1043 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, sd->ifsd_map, addr);
1044 			/* Update descriptor and the cached value */
1045 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size);
1046 	}
1047 	/* preserve queue */
1048 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1049 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1050 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1051 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1052 	} else
1053 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1054 }
1055 
1056 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1057 
1058 #else
1059 #define iflib_netmap_txq_init(ctx, txq)
1060 #define iflib_netmap_rxq_init(ctx, rxq)
1061 #define iflib_netmap_detach(ifp)
1062 
1063 #define iflib_netmap_attach(ctx) (0)
1064 #define netmap_rx_irq(ifp, qid, budget) (0)
1065 
1066 #endif
1067 
1068 #if defined(__i386__) || defined(__amd64__)
1069 static __inline void
1070 prefetch(void *x)
1071 {
1072 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1073 }
1074 #else
1075 #define prefetch(x)
1076 #endif
1077 
1078 static void
1079 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1080 {
1081 	if (err)
1082 		return;
1083 	*(bus_addr_t *) arg = segs[0].ds_addr;
1084 }
1085 
1086 int
1087 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1088 {
1089 	int err;
1090 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1091 	device_t dev = ctx->ifc_dev;
1092 
1093 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1094 
1095 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1096 				sctx->isc_q_align, 0,	/* alignment, bounds */
1097 				BUS_SPACE_MAXADDR,	/* lowaddr */
1098 				BUS_SPACE_MAXADDR,	/* highaddr */
1099 				NULL, NULL,		/* filter, filterarg */
1100 				size,			/* maxsize */
1101 				1,			/* nsegments */
1102 				size,			/* maxsegsize */
1103 				BUS_DMA_ALLOCNOW,	/* flags */
1104 				NULL,			/* lockfunc */
1105 				NULL,			/* lockarg */
1106 				&dma->idi_tag);
1107 	if (err) {
1108 		device_printf(dev,
1109 		    "%s: bus_dma_tag_create failed: %d\n",
1110 		    __func__, err);
1111 		goto fail_0;
1112 	}
1113 
1114 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1115 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1116 	if (err) {
1117 		device_printf(dev,
1118 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1119 		    __func__, (uintmax_t)size, err);
1120 		goto fail_1;
1121 	}
1122 
1123 	dma->idi_paddr = IF_BAD_DMA;
1124 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1125 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1126 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1127 		device_printf(dev,
1128 		    "%s: bus_dmamap_load failed: %d\n",
1129 		    __func__, err);
1130 		goto fail_2;
1131 	}
1132 
1133 	dma->idi_size = size;
1134 	return (0);
1135 
1136 fail_2:
1137 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1138 fail_1:
1139 	bus_dma_tag_destroy(dma->idi_tag);
1140 fail_0:
1141 	dma->idi_tag = NULL;
1142 
1143 	return (err);
1144 }
1145 
1146 int
1147 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1148 {
1149 	int i, err;
1150 	iflib_dma_info_t *dmaiter;
1151 
1152 	dmaiter = dmalist;
1153 	for (i = 0; i < count; i++, dmaiter++) {
1154 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1155 			break;
1156 	}
1157 	if (err)
1158 		iflib_dma_free_multi(dmalist, i);
1159 	return (err);
1160 }
1161 
1162 void
1163 iflib_dma_free(iflib_dma_info_t dma)
1164 {
1165 	if (dma->idi_tag == NULL)
1166 		return;
1167 	if (dma->idi_paddr != IF_BAD_DMA) {
1168 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1169 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1170 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1171 		dma->idi_paddr = IF_BAD_DMA;
1172 	}
1173 	if (dma->idi_vaddr != NULL) {
1174 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1175 		dma->idi_vaddr = NULL;
1176 	}
1177 	bus_dma_tag_destroy(dma->idi_tag);
1178 	dma->idi_tag = NULL;
1179 }
1180 
1181 void
1182 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1183 {
1184 	int i;
1185 	iflib_dma_info_t *dmaiter = dmalist;
1186 
1187 	for (i = 0; i < count; i++, dmaiter++)
1188 		iflib_dma_free(*dmaiter);
1189 }
1190 
1191 static int
1192 iflib_fast_intr(void *arg)
1193 {
1194 	iflib_filter_info_t info = arg;
1195 	struct grouptask *gtask = info->ifi_task;
1196 
1197 	if (!smp_started)
1198 		return (FILTER_HANDLED);
1199 
1200 	DBG_COUNTER_INC(fast_intrs);
1201 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1202 		return (FILTER_HANDLED);
1203 
1204 	GROUPTASK_ENQUEUE(gtask);
1205 	return (FILTER_HANDLED);
1206 }
1207 
1208 static int
1209 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1210 	driver_filter_t filter, driver_intr_t handler, void *arg,
1211 				 char *name)
1212 {
1213 	int rc;
1214 	struct resource *res;
1215 	void *tag;
1216 	device_t dev = ctx->ifc_dev;
1217 
1218 	MPASS(rid < 512);
1219 	irq->ii_rid = rid;
1220 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1221 				     RF_SHAREABLE | RF_ACTIVE);
1222 	if (res == NULL) {
1223 		device_printf(dev,
1224 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1225 		return (ENOMEM);
1226 	}
1227 	irq->ii_res = res;
1228 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1229 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1230 						filter, handler, arg, &tag);
1231 	if (rc != 0) {
1232 		device_printf(dev,
1233 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1234 					  rid, name ? name : "unknown", rc);
1235 		return (rc);
1236 	} else if (name)
1237 		bus_describe_intr(dev, res, tag, "%s", name);
1238 
1239 	irq->ii_tag = tag;
1240 	return (0);
1241 }
1242 
1243 
1244 /*********************************************************************
1245  *
1246  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1247  *  the information needed to transmit a packet on the wire. This is
1248  *  called only once at attach, setup is done every reset.
1249  *
1250  **********************************************************************/
1251 
1252 static int
1253 iflib_txsd_alloc(iflib_txq_t txq)
1254 {
1255 	if_ctx_t ctx = txq->ift_ctx;
1256 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1257 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1258 	device_t dev = ctx->ifc_dev;
1259 	int err, nsegments, ntsosegments;
1260 
1261 	nsegments = scctx->isc_tx_nsegments;
1262 	ntsosegments = scctx->isc_tx_tso_segments_max;
1263 	MPASS(scctx->isc_ntxd[0] > 0);
1264 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1265 	MPASS(nsegments > 0);
1266 	MPASS(ntsosegments > 0);
1267 	/*
1268 	 * Setup DMA descriptor areas.
1269 	 */
1270 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1271 			       1, 0,			/* alignment, bounds */
1272 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1273 			       BUS_SPACE_MAXADDR,	/* highaddr */
1274 			       NULL, NULL,		/* filter, filterarg */
1275 			       sctx->isc_tx_maxsize,		/* maxsize */
1276 			       nsegments,	/* nsegments */
1277 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1278 			       0,			/* flags */
1279 			       NULL,			/* lockfunc */
1280 			       NULL,			/* lockfuncarg */
1281 			       &txq->ift_desc_tag))) {
1282 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1283 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1284 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1285 		goto fail;
1286 	}
1287 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1288 			       1, 0,			/* alignment, bounds */
1289 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1290 			       BUS_SPACE_MAXADDR,	/* highaddr */
1291 			       NULL, NULL,		/* filter, filterarg */
1292 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1293 			       ntsosegments,	/* nsegments */
1294 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1295 			       0,			/* flags */
1296 			       NULL,			/* lockfunc */
1297 			       NULL,			/* lockfuncarg */
1298 			       &txq->ift_tso_desc_tag))) {
1299 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1300 
1301 		goto fail;
1302 	}
1303 	if (!(txq->ift_sds.ifsd_flags =
1304 	    (uint8_t *) malloc(sizeof(uint8_t) *
1305 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1306 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1307 		err = ENOMEM;
1308 		goto fail;
1309 	}
1310 	if (!(txq->ift_sds.ifsd_m =
1311 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1312 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1313 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1314 		err = ENOMEM;
1315 		goto fail;
1316 	}
1317 
1318         /* Create the descriptor buffer dma maps */
1319 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1320 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1321 		return (0);
1322 
1323 	if (!(txq->ift_sds.ifsd_map =
1324 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1325 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1326 		err = ENOMEM;
1327 		goto fail;
1328 	}
1329 
1330 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1331 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1332 		if (err != 0) {
1333 			device_printf(dev, "Unable to create TX DMA map\n");
1334 			goto fail;
1335 		}
1336 	}
1337 #endif
1338 	return (0);
1339 fail:
1340 	/* We free all, it handles case where we are in the middle */
1341 	iflib_tx_structures_free(ctx);
1342 	return (err);
1343 }
1344 
1345 static void
1346 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1347 {
1348 	bus_dmamap_t map;
1349 
1350 	map = NULL;
1351 	if (txq->ift_sds.ifsd_map != NULL)
1352 		map = txq->ift_sds.ifsd_map[i];
1353 	if (map != NULL) {
1354 		bus_dmamap_unload(txq->ift_desc_tag, map);
1355 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1356 		txq->ift_sds.ifsd_map[i] = NULL;
1357 	}
1358 }
1359 
1360 static void
1361 iflib_txq_destroy(iflib_txq_t txq)
1362 {
1363 	if_ctx_t ctx = txq->ift_ctx;
1364 
1365 	for (int i = 0; i < txq->ift_size; i++)
1366 		iflib_txsd_destroy(ctx, txq, i);
1367 	if (txq->ift_sds.ifsd_map != NULL) {
1368 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1369 		txq->ift_sds.ifsd_map = NULL;
1370 	}
1371 	if (txq->ift_sds.ifsd_m != NULL) {
1372 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1373 		txq->ift_sds.ifsd_m = NULL;
1374 	}
1375 	if (txq->ift_sds.ifsd_flags != NULL) {
1376 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1377 		txq->ift_sds.ifsd_flags = NULL;
1378 	}
1379 	if (txq->ift_desc_tag != NULL) {
1380 		bus_dma_tag_destroy(txq->ift_desc_tag);
1381 		txq->ift_desc_tag = NULL;
1382 	}
1383 	if (txq->ift_tso_desc_tag != NULL) {
1384 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1385 		txq->ift_tso_desc_tag = NULL;
1386 	}
1387 }
1388 
1389 static void
1390 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1391 {
1392 	struct mbuf **mp;
1393 
1394 	mp = &txq->ift_sds.ifsd_m[i];
1395 	if (*mp == NULL)
1396 		return;
1397 
1398 	if (txq->ift_sds.ifsd_map != NULL) {
1399 		bus_dmamap_sync(txq->ift_desc_tag,
1400 				txq->ift_sds.ifsd_map[i],
1401 				BUS_DMASYNC_POSTWRITE);
1402 		bus_dmamap_unload(txq->ift_desc_tag,
1403 				  txq->ift_sds.ifsd_map[i]);
1404 	}
1405 	m_free(*mp);
1406 	DBG_COUNTER_INC(tx_frees);
1407 	*mp = NULL;
1408 }
1409 
1410 static int
1411 iflib_txq_setup(iflib_txq_t txq)
1412 {
1413 	if_ctx_t ctx = txq->ift_ctx;
1414 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1415 	iflib_dma_info_t di;
1416 	int i;
1417 
1418 	/* Set number of descriptors available */
1419 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1420 
1421 	/* Reset indices */
1422 	txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1423 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1424 
1425 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1426 		bzero((void *)di->idi_vaddr, di->idi_size);
1427 
1428 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1429 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1430 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1431 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1432 	return (0);
1433 }
1434 
1435 /*********************************************************************
1436  *
1437  *  Allocate memory for rx_buffer structures. Since we use one
1438  *  rx_buffer per received packet, the maximum number of rx_buffer's
1439  *  that we'll need is equal to the number of receive descriptors
1440  *  that we've allocated.
1441  *
1442  **********************************************************************/
1443 static int
1444 iflib_rxsd_alloc(iflib_rxq_t rxq)
1445 {
1446 	if_ctx_t ctx = rxq->ifr_ctx;
1447 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1448 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1449 	device_t dev = ctx->ifc_dev;
1450 	iflib_fl_t fl;
1451 	iflib_rxsd_t	rxsd;
1452 	int			err;
1453 
1454 	MPASS(scctx->isc_nrxd[0] > 0);
1455 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1456 
1457 	fl = rxq->ifr_fl;
1458 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1459 		fl->ifl_sds = malloc(sizeof(struct iflib_sw_rx_desc) *
1460 		    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB,
1461 		    M_WAITOK | M_ZERO);
1462 		if (fl->ifl_sds == NULL) {
1463 			device_printf(dev, "Unable to allocate rx sw desc memory\n");
1464 			return (ENOMEM);
1465 		}
1466 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1467 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1468 					 1, 0,			/* alignment, bounds */
1469 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1470 					 BUS_SPACE_MAXADDR,	/* highaddr */
1471 					 NULL, NULL,		/* filter, filterarg */
1472 					 sctx->isc_rx_maxsize,	/* maxsize */
1473 					 sctx->isc_rx_nsegments,	/* nsegments */
1474 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1475 					 0,			/* flags */
1476 					 NULL,			/* lockfunc */
1477 					 NULL,			/* lockarg */
1478 					 &fl->ifl_desc_tag);
1479 		if (err) {
1480 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1481 				__func__, err);
1482 			goto fail;
1483 		}
1484 
1485 		rxsd = fl->ifl_sds;
1486 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++, rxsd++) {
1487 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &rxsd->ifsd_map);
1488 			if (err) {
1489 				device_printf(dev, "%s: bus_dmamap_create failed: %d\n",
1490 					__func__, err);
1491 				goto fail;
1492 			}
1493 		}
1494 	}
1495 	return (0);
1496 
1497 fail:
1498 	iflib_rx_structures_free(ctx);
1499 	return (err);
1500 }
1501 
1502 
1503 /*
1504  * Internal service routines
1505  */
1506 
1507 struct rxq_refill_cb_arg {
1508 	int               error;
1509 	bus_dma_segment_t seg;
1510 	int               nseg;
1511 };
1512 
1513 static void
1514 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1515 {
1516 	struct rxq_refill_cb_arg *cb_arg = arg;
1517 
1518 	cb_arg->error = error;
1519 	cb_arg->seg = segs[0];
1520 	cb_arg->nseg = nseg;
1521 }
1522 
1523 
1524 #ifdef ACPI_DMAR
1525 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1526 #else
1527 #define IS_DMAR(ctx) (0)
1528 #endif
1529 
1530 /**
1531  *	rxq_refill - refill an rxq  free-buffer list
1532  *	@ctx: the iflib context
1533  *	@rxq: the free-list to refill
1534  *	@n: the number of new buffers to allocate
1535  *
1536  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1537  *	The caller must assure that @n does not exceed the queue's capacity.
1538  */
1539 static void
1540 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1541 {
1542 	struct mbuf *m;
1543 	int pidx = fl->ifl_pidx;
1544 	iflib_rxsd_t rxsd = &fl->ifl_sds[pidx];
1545 	caddr_t cl;
1546 	int n, i = 0;
1547 	uint64_t bus_addr;
1548 	int err;
1549 
1550 	n  = count;
1551 	MPASS(n > 0);
1552 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1553 
1554 	if (pidx < fl->ifl_cidx)
1555 		MPASS(pidx + n <= fl->ifl_cidx);
1556 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1557 		MPASS(fl->ifl_gen == 0);
1558 	if (pidx > fl->ifl_cidx)
1559 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1560 
1561 	DBG_COUNTER_INC(fl_refills);
1562 	if (n > 8)
1563 		DBG_COUNTER_INC(fl_refills_large);
1564 
1565 	while (n--) {
1566 		/*
1567 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1568 		 * initialized after rx.
1569 		 *
1570 		 * If the cluster is still set then we know a minimum sized packet was received
1571 		 */
1572 		if ((cl = rxsd->ifsd_cl) == NULL) {
1573 			if ((cl = rxsd->ifsd_cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1574 				break;
1575 #if MEMORY_LOGGING
1576 			fl->ifl_cl_enqueued++;
1577 #endif
1578 		}
1579 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1580 			break;
1581 		}
1582 #if MEMORY_LOGGING
1583 		fl->ifl_m_enqueued++;
1584 #endif
1585 
1586 		DBG_COUNTER_INC(rx_allocs);
1587 #ifdef notyet
1588 		if ((rxsd->ifsd_flags & RX_SW_DESC_MAP_CREATED) == 0) {
1589 			int err;
1590 
1591 			if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &rxsd->ifsd_map))) {
1592 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
1593 				uma_zfree(fl->ifl_zone, cl);
1594 				n = 0;
1595 				goto done;
1596 			}
1597 			rxsd->ifsd_flags |= RX_SW_DESC_MAP_CREATED;
1598 		}
1599 #endif
1600 #if defined(__i386__) || defined(__amd64__)
1601 		if (!IS_DMAR(ctx)) {
1602 			bus_addr = pmap_kextract((vm_offset_t)cl);
1603 		} else
1604 #endif
1605 		{
1606 			struct rxq_refill_cb_arg cb_arg;
1607 			iflib_rxq_t q;
1608 
1609 			cb_arg.error = 0;
1610 			q = fl->ifl_rxq;
1611 			err = bus_dmamap_load(fl->ifl_desc_tag, rxsd->ifsd_map,
1612 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1613 
1614 			if (err != 0 || cb_arg.error) {
1615 				/*
1616 				 * !zone_pack ?
1617 				 */
1618 				if (fl->ifl_zone == zone_pack)
1619 					uma_zfree(fl->ifl_zone, cl);
1620 				m_free(m);
1621 				n = 0;
1622 				goto done;
1623 			}
1624 			bus_addr = cb_arg.seg.ds_addr;
1625 		}
1626 		rxsd->ifsd_flags |= RX_SW_DESC_INUSE;
1627 
1628 		MPASS(rxsd->ifsd_m == NULL);
1629 		rxsd->ifsd_cl = cl;
1630 		rxsd->ifsd_m = m;
1631 		fl->ifl_bus_addrs[i] = bus_addr;
1632 		fl->ifl_vm_addrs[i] = cl;
1633 		rxsd++;
1634 		fl->ifl_credits++;
1635 		i++;
1636 		MPASS(fl->ifl_credits <= fl->ifl_size);
1637 		if (++fl->ifl_pidx == fl->ifl_size) {
1638 			fl->ifl_pidx = 0;
1639 			fl->ifl_gen = 1;
1640 			rxsd = fl->ifl_sds;
1641 		}
1642 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1643 			ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx,
1644 								 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size);
1645 			i = 0;
1646 			pidx = fl->ifl_pidx;
1647 		}
1648 	}
1649 done:
1650 	DBG_COUNTER_INC(rxd_flush);
1651 	if (fl->ifl_pidx == 0)
1652 		pidx = fl->ifl_size - 1;
1653 	else
1654 		pidx = fl->ifl_pidx - 1;
1655 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1656 }
1657 
1658 static __inline void
1659 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1660 {
1661 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1662 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1663 #ifdef INVARIANTS
1664 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1665 #endif
1666 
1667 	MPASS(fl->ifl_credits <= fl->ifl_size);
1668 	MPASS(reclaimable == delta);
1669 
1670 	if (reclaimable > 0)
1671 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1672 }
1673 
1674 static void
1675 iflib_fl_bufs_free(iflib_fl_t fl)
1676 {
1677 	iflib_dma_info_t idi = fl->ifl_ifdi;
1678 	uint32_t i;
1679 
1680 	for (i = 0; i < fl->ifl_size; i++) {
1681 		iflib_rxsd_t d = &fl->ifl_sds[i];
1682 
1683 		if (d->ifsd_flags & RX_SW_DESC_INUSE) {
1684 			bus_dmamap_unload(fl->ifl_desc_tag, d->ifsd_map);
1685 			bus_dmamap_destroy(fl->ifl_desc_tag, d->ifsd_map);
1686 			if (d->ifsd_m != NULL) {
1687 				m_init(d->ifsd_m, M_NOWAIT, MT_DATA, 0);
1688 				uma_zfree(zone_mbuf, d->ifsd_m);
1689 			}
1690 			if (d->ifsd_cl != NULL)
1691 				uma_zfree(fl->ifl_zone, d->ifsd_cl);
1692 			d->ifsd_flags = 0;
1693 		} else {
1694 			MPASS(d->ifsd_cl == NULL);
1695 			MPASS(d->ifsd_m == NULL);
1696 		}
1697 #if MEMORY_LOGGING
1698 		fl->ifl_m_dequeued++;
1699 		fl->ifl_cl_dequeued++;
1700 #endif
1701 		d->ifsd_cl = NULL;
1702 		d->ifsd_m = NULL;
1703 	}
1704 	/*
1705 	 * Reset free list values
1706 	 */
1707 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1708 	bzero(idi->idi_vaddr, idi->idi_size);
1709 }
1710 
1711 /*********************************************************************
1712  *
1713  *  Initialize a receive ring and its buffers.
1714  *
1715  **********************************************************************/
1716 static int
1717 iflib_fl_setup(iflib_fl_t fl)
1718 {
1719 	iflib_rxq_t rxq = fl->ifl_rxq;
1720 	if_ctx_t ctx = rxq->ifr_ctx;
1721 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1722 
1723 	/*
1724 	** Free current RX buffer structs and their mbufs
1725 	*/
1726 	iflib_fl_bufs_free(fl);
1727 	/* Now replenish the mbufs */
1728 	MPASS(fl->ifl_credits == 0);
1729 	/*
1730 	 * XXX don't set the max_frame_size to larger
1731 	 * than the hardware can handle
1732 	 */
1733 	if (sctx->isc_max_frame_size <= 2048)
1734 		fl->ifl_buf_size = MCLBYTES;
1735 	else if (sctx->isc_max_frame_size <= 4096)
1736 		fl->ifl_buf_size = MJUMPAGESIZE;
1737 	else if (sctx->isc_max_frame_size <= 9216)
1738 		fl->ifl_buf_size = MJUM9BYTES;
1739 	else
1740 		fl->ifl_buf_size = MJUM16BYTES;
1741 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
1742 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
1743 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
1744 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
1745 
1746 
1747 	/* avoid pre-allocating zillions of clusters to an idle card
1748 	 * potentially speeding up attach
1749 	 */
1750 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
1751 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
1752 	if (min(128, fl->ifl_size) != fl->ifl_credits)
1753 		return (ENOBUFS);
1754 	/*
1755 	 * handle failure
1756 	 */
1757 	MPASS(rxq != NULL);
1758 	MPASS(fl->ifl_ifdi != NULL);
1759 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1760 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1761 	return (0);
1762 }
1763 
1764 /*********************************************************************
1765  *
1766  *  Free receive ring data structures
1767  *
1768  **********************************************************************/
1769 static void
1770 iflib_rx_sds_free(iflib_rxq_t rxq)
1771 {
1772 	iflib_fl_t fl;
1773 	int i;
1774 
1775 	if (rxq->ifr_fl != NULL) {
1776 		for (i = 0; i < rxq->ifr_nfl; i++) {
1777 			fl = &rxq->ifr_fl[i];
1778 			if (fl->ifl_desc_tag != NULL) {
1779 				bus_dma_tag_destroy(fl->ifl_desc_tag);
1780 				fl->ifl_desc_tag = NULL;
1781 			}
1782 		}
1783 		if (rxq->ifr_fl->ifl_sds != NULL)
1784 			free(rxq->ifr_fl->ifl_sds, M_IFLIB);
1785 
1786 		free(rxq->ifr_fl, M_IFLIB);
1787 		rxq->ifr_fl = NULL;
1788 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
1789 	}
1790 }
1791 
1792 /*
1793  * MI independent logic
1794  *
1795  */
1796 static void
1797 iflib_timer(void *arg)
1798 {
1799 	iflib_txq_t txq = arg;
1800 	if_ctx_t ctx = txq->ift_ctx;
1801 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1802 
1803 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
1804 		return;
1805 	/*
1806 	** Check on the state of the TX queue(s), this
1807 	** can be done without the lock because its RO
1808 	** and the HUNG state will be static if set.
1809 	*/
1810 	IFDI_TIMER(ctx, txq->ift_id);
1811 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
1812 		(ctx->ifc_pause_frames == 0))
1813 		goto hung;
1814 
1815 	if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments ||
1816 	    ifmp_ring_is_stalled(txq->ift_br[0]))
1817 		GROUPTASK_ENQUEUE(&txq->ift_task);
1818 
1819 	ctx->ifc_pause_frames = 0;
1820 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
1821 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
1822 	return;
1823 hung:
1824 	CTX_LOCK(ctx);
1825 	if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
1826 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
1827 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
1828 
1829 	IFDI_WATCHDOG_RESET(ctx);
1830 	ctx->ifc_watchdog_events++;
1831 	ctx->ifc_pause_frames = 0;
1832 
1833 	iflib_init_locked(ctx);
1834 	CTX_UNLOCK(ctx);
1835 }
1836 
1837 static void
1838 iflib_init_locked(if_ctx_t ctx)
1839 {
1840 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1841 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1842 	if_t ifp = ctx->ifc_ifp;
1843 	iflib_fl_t fl;
1844 	iflib_txq_t txq;
1845 	iflib_rxq_t rxq;
1846 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
1847 
1848 
1849 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1850 	IFDI_INTR_DISABLE(ctx);
1851 
1852 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
1853 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
1854 	/* Set hardware offload abilities */
1855 	if_clearhwassist(ifp);
1856 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
1857 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
1858 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
1859 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
1860 	if (if_getcapenable(ifp) & IFCAP_TSO4)
1861 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
1862 	if (if_getcapenable(ifp) & IFCAP_TSO6)
1863 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
1864 
1865 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
1866 		CALLOUT_LOCK(txq);
1867 		callout_stop(&txq->ift_timer);
1868 		callout_stop(&txq->ift_db_check);
1869 		CALLOUT_UNLOCK(txq);
1870 		iflib_netmap_txq_init(ctx, txq);
1871 	}
1872 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1873 		iflib_netmap_rxq_init(ctx, rxq);
1874 	}
1875 #ifdef INVARIANTS
1876 	i = if_getdrvflags(ifp);
1877 #endif
1878 	IFDI_INIT(ctx);
1879 	MPASS(if_getdrvflags(ifp) == i);
1880 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1881 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
1882 			if (iflib_fl_setup(fl)) {
1883 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
1884 				goto done;
1885 			}
1886 		}
1887 	}
1888 	done:
1889 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
1890 	IFDI_INTR_ENABLE(ctx);
1891 	txq = ctx->ifc_txqs;
1892 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
1893 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
1894 			txq->ift_timer.c_cpu);
1895 }
1896 
1897 static int
1898 iflib_media_change(if_t ifp)
1899 {
1900 	if_ctx_t ctx = if_getsoftc(ifp);
1901 	int err;
1902 
1903 	CTX_LOCK(ctx);
1904 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
1905 		iflib_init_locked(ctx);
1906 	CTX_UNLOCK(ctx);
1907 	return (err);
1908 }
1909 
1910 static void
1911 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
1912 {
1913 	if_ctx_t ctx = if_getsoftc(ifp);
1914 
1915 	CTX_LOCK(ctx);
1916 	IFDI_UPDATE_ADMIN_STATUS(ctx);
1917 	IFDI_MEDIA_STATUS(ctx, ifmr);
1918 	CTX_UNLOCK(ctx);
1919 }
1920 
1921 static void
1922 iflib_stop(if_ctx_t ctx)
1923 {
1924 	iflib_txq_t txq = ctx->ifc_txqs;
1925 	iflib_rxq_t rxq = ctx->ifc_rxqs;
1926 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1927 	iflib_dma_info_t di;
1928 	iflib_fl_t fl;
1929 	int i, j;
1930 
1931 	/* Tell the stack that the interface is no longer active */
1932 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1933 
1934 	IFDI_INTR_DISABLE(ctx);
1935 	DELAY(100000);
1936 	IFDI_STOP(ctx);
1937 	DELAY(100000);
1938 
1939 	iflib_debug_reset();
1940 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
1941 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
1942 		/* make sure all transmitters have completed before proceeding XXX */
1943 
1944 		/* clean any enqueued buffers */
1945 		iflib_ifmp_purge(txq);
1946 		/* Free any existing tx buffers. */
1947 		for (j = 0; j < txq->ift_size; j++) {
1948 			iflib_txsd_free(ctx, txq, j);
1949 		}
1950 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
1951 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
1952 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
1953 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
1954 		txq->ift_pullups = 0;
1955 		ifmp_ring_reset_stats(txq->ift_br[0]);
1956 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
1957 			bzero((void *)di->idi_vaddr, di->idi_size);
1958 	}
1959 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
1960 		/* make sure all transmitters have completed before proceeding XXX */
1961 
1962 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
1963 			bzero((void *)di->idi_vaddr, di->idi_size);
1964 		/* also resets the free lists pidx/cidx */
1965 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
1966 			iflib_fl_bufs_free(fl);
1967 	}
1968 }
1969 
1970 static iflib_rxsd_t
1971 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload)
1972 {
1973 	int flid, cidx;
1974 	iflib_rxsd_t sd;
1975 	iflib_fl_t fl;
1976 	iflib_dma_info_t di;
1977 
1978 	flid = irf->irf_flid;
1979 	cidx = irf->irf_idx;
1980 	fl = &rxq->ifr_fl[flid];
1981 	fl->ifl_credits--;
1982 #if MEMORY_LOGGING
1983 	fl->ifl_m_dequeued++;
1984 	if (cltype)
1985 		fl->ifl_cl_dequeued++;
1986 #endif
1987 	sd = &fl->ifl_sds[cidx];
1988 	di = fl->ifl_ifdi;
1989 	bus_dmamap_sync(di->idi_tag, di->idi_map,
1990 			BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1991 
1992 	/* not valid assert if bxe really does SGE from non-contiguous elements */
1993 	MPASS(fl->ifl_cidx == cidx);
1994 	if (unload)
1995 		bus_dmamap_unload(fl->ifl_desc_tag, sd->ifsd_map);
1996 
1997 	if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) {
1998 		fl->ifl_cidx = 0;
1999 		fl->ifl_gen = 0;
2000 	}
2001 	/* YES ick */
2002 	if (cltype)
2003 		*cltype = fl->ifl_cltype;
2004 	return (sd);
2005 }
2006 
2007 static struct mbuf *
2008 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri)
2009 {
2010 	int i, padlen , flags, cltype;
2011 	struct mbuf *m, *mh, *mt;
2012 	iflib_rxsd_t sd;
2013 	caddr_t cl;
2014 
2015 	i = 0;
2016 	mh = NULL;
2017 	do {
2018 		sd = rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE);
2019 
2020 		MPASS(sd->ifsd_cl != NULL);
2021 		MPASS(sd->ifsd_m != NULL);
2022 
2023 		/* Don't include zero-length frags */
2024 		if (ri->iri_frags[i].irf_len == 0) {
2025 			/* XXX we can save the cluster here, but not the mbuf */
2026 			m_init(sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2027 			m_free(sd->ifsd_m);
2028 			sd->ifsd_m = NULL;
2029 			continue;
2030 		}
2031 
2032 		m = sd->ifsd_m;
2033 		if (mh == NULL) {
2034 			flags = M_PKTHDR|M_EXT;
2035 			mh = mt = m;
2036 			padlen = ri->iri_pad;
2037 		} else {
2038 			flags = M_EXT;
2039 			mt->m_next = m;
2040 			mt = m;
2041 			/* assuming padding is only on the first fragment */
2042 			padlen = 0;
2043 		}
2044 		sd->ifsd_m = NULL;
2045 		cl = sd->ifsd_cl;
2046 		sd->ifsd_cl = NULL;
2047 
2048 		/* Can these two be made one ? */
2049 		m_init(m, M_NOWAIT, MT_DATA, flags);
2050 		m_cljset(m, cl, cltype);
2051 		/*
2052 		 * These must follow m_init and m_cljset
2053 		 */
2054 		m->m_data += padlen;
2055 		ri->iri_len -= padlen;
2056 		m->m_len = ri->iri_frags[i].irf_len;
2057 	} while (++i < ri->iri_nfrags);
2058 
2059 	return (mh);
2060 }
2061 
2062 /*
2063  * Process one software descriptor
2064  */
2065 static struct mbuf *
2066 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2067 {
2068 	struct mbuf *m;
2069 	iflib_rxsd_t sd;
2070 
2071 	/* should I merge this back in now that the two paths are basically duplicated? */
2072 	if (ri->iri_nfrags == 1 &&
2073 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2074 		sd = rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE);
2075 		m = sd->ifsd_m;
2076 		sd->ifsd_m = NULL;
2077 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2078 		memcpy(m->m_data, sd->ifsd_cl, ri->iri_len);
2079 		m->m_len = ri->iri_frags[0].irf_len;
2080        } else {
2081 		m = assemble_segments(rxq, ri);
2082 	}
2083 	m->m_pkthdr.len = ri->iri_len;
2084 	m->m_pkthdr.rcvif = ri->iri_ifp;
2085 	m->m_flags |= ri->iri_flags;
2086 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2087 	m->m_pkthdr.flowid = ri->iri_flowid;
2088 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2089 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2090 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2091 	return (m);
2092 }
2093 
2094 static bool
2095 iflib_rxeof(iflib_rxq_t rxq, int budget)
2096 {
2097 	if_ctx_t ctx = rxq->ifr_ctx;
2098 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2099 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2100 	int avail, i;
2101 	uint16_t *cidxp;
2102 	struct if_rxd_info ri;
2103 	int err, budget_left, rx_bytes, rx_pkts;
2104 	iflib_fl_t fl;
2105 	struct ifnet *ifp;
2106 	int lro_enabled;
2107 	/*
2108 	 * XXX early demux data packets so that if_input processing only handles
2109 	 * acks in interrupt context
2110 	 */
2111 	struct mbuf *m, *mh, *mt;
2112 
2113 	if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) {
2114 		return (FALSE);
2115 	}
2116 
2117 	mh = mt = NULL;
2118 	MPASS(budget > 0);
2119 	rx_pkts	= rx_bytes = 0;
2120 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2121 		cidxp = &rxq->ifr_cq_cidx;
2122 	else
2123 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2124 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2125 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2126 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2127 		DBG_COUNTER_INC(rx_unavail);
2128 		return (false);
2129 	}
2130 
2131 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2132 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2133 			DBG_COUNTER_INC(rx_ctx_inactive);
2134 			break;
2135 		}
2136 		/*
2137 		 * Reset client set fields to their default values
2138 		 */
2139 		bzero(&ri, sizeof(ri));
2140 		ri.iri_qsidx = rxq->ifr_id;
2141 		ri.iri_cidx = *cidxp;
2142 		ri.iri_ifp = ctx->ifc_ifp;
2143 		ri.iri_frags = rxq->ifr_frags;
2144 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2145 
2146 		/* in lieu of handling correctly - make sure it isn't being unhandled */
2147 		MPASS(err == 0);
2148 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2149 			*cidxp = ri.iri_cidx;
2150 			/* Update our consumer index */
2151 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2152 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2153 				rxq->ifr_cq_gen = 0;
2154 			}
2155 			/* was this only a completion queue message? */
2156 			if (__predict_false(ri.iri_nfrags == 0))
2157 				continue;
2158 		}
2159 		MPASS(ri.iri_nfrags != 0);
2160 		MPASS(ri.iri_len != 0);
2161 
2162 		/* will advance the cidx on the corresponding free lists */
2163 		m = iflib_rxd_pkt_get(rxq, &ri);
2164 		if (avail == 0 && budget_left)
2165 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2166 
2167 		if (__predict_false(m == NULL)) {
2168 			DBG_COUNTER_INC(rx_mbuf_null);
2169 			continue;
2170 		}
2171 		/* imm_pkt: -- cxgb */
2172 		if (mh == NULL)
2173 			mh = mt = m;
2174 		else {
2175 			mt->m_nextpkt = m;
2176 			mt = m;
2177 		}
2178 	}
2179 	/* make sure that we can refill faster than drain */
2180 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2181 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2182 
2183 	ifp = ctx->ifc_ifp;
2184 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2185 	while (mh != NULL) {
2186 		m = mh;
2187 		mh = mh->m_nextpkt;
2188 		m->m_nextpkt = NULL;
2189 		rx_bytes += m->m_pkthdr.len;
2190 		rx_pkts++;
2191 #if defined(INET6) || defined(INET)
2192 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2193 			continue;
2194 #endif
2195 		DBG_COUNTER_INC(rx_if_input);
2196 		ifp->if_input(ifp, m);
2197 	}
2198 
2199 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2200 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2201 
2202 	/*
2203 	 * Flush any outstanding LRO work
2204 	 */
2205 #if defined(INET6) || defined(INET)
2206 	tcp_lro_flush_all(&rxq->ifr_lc);
2207 #endif
2208 	if (avail)
2209 		return true;
2210 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2211 }
2212 
2213 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2214 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2215 #define TXQ_MAX_DB_DEFERRED(size) (size >> 5)
2216 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2217 
2218 static __inline void
2219 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring)
2220 {
2221 	uint32_t dbval;
2222 
2223 	if (ring || txq->ift_db_pending >=
2224 	    TXQ_MAX_DB_DEFERRED(txq->ift_size)) {
2225 
2226 		/* the lock will only ever be contended in the !min_latency case */
2227 		if (!TXDB_TRYLOCK(txq))
2228 			return;
2229 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2230 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2231 		txq->ift_db_pending = txq->ift_npending = 0;
2232 		TXDB_UNLOCK(txq);
2233 	}
2234 }
2235 
2236 static void
2237 iflib_txd_deferred_db_check(void * arg)
2238 {
2239 	iflib_txq_t txq = arg;
2240 
2241 	/* simple non-zero boolean so use bitwise OR */
2242 	if ((txq->ift_db_pending | txq->ift_npending) &&
2243 	    txq->ift_db_pending >= txq->ift_db_pending_queued)
2244 		iflib_txd_db_check(txq->ift_ctx, txq, TRUE);
2245 	txq->ift_db_pending_queued = 0;
2246 	if (ifmp_ring_is_stalled(txq->ift_br[0]))
2247 		iflib_txq_check_drain(txq, 4);
2248 }
2249 
2250 #ifdef PKT_DEBUG
2251 static void
2252 print_pkt(if_pkt_info_t pi)
2253 {
2254 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2255 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2256 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2257 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2258 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2259 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2260 }
2261 #endif
2262 
2263 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2264 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2265 
2266 static int
2267 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2268 {
2269 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2270 	struct ether_vlan_header *eh;
2271 	struct mbuf *m, *n;
2272 
2273 	n = m = *mp;
2274 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2275 	    M_WRITABLE(m) == 0) {
2276 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2277 			return (ENOMEM);
2278 		} else {
2279 			m_freem(*mp);
2280 			n = *mp = m;
2281 		}
2282 	}
2283 
2284 	/*
2285 	 * Determine where frame payload starts.
2286 	 * Jump over vlan headers if already present,
2287 	 * helpful for QinQ too.
2288 	 */
2289 	if (__predict_false(m->m_len < sizeof(*eh))) {
2290 		txq->ift_pullups++;
2291 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2292 			return (ENOMEM);
2293 	}
2294 	eh = mtod(m, struct ether_vlan_header *);
2295 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2296 		pi->ipi_etype = ntohs(eh->evl_proto);
2297 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2298 	} else {
2299 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2300 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2301 	}
2302 
2303 	switch (pi->ipi_etype) {
2304 #ifdef INET
2305 	case ETHERTYPE_IP:
2306 	{
2307 		struct ip *ip = NULL;
2308 		struct tcphdr *th = NULL;
2309 		int minthlen;
2310 
2311 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2312 		if (__predict_false(m->m_len < minthlen)) {
2313 			/*
2314 			 * if this code bloat is causing too much of a hit
2315 			 * move it to a separate function and mark it noinline
2316 			 */
2317 			if (m->m_len == pi->ipi_ehdrlen) {
2318 				n = m->m_next;
2319 				MPASS(n);
2320 				if (n->m_len >= sizeof(*ip))  {
2321 					ip = (struct ip *)n->m_data;
2322 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2323 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2324 				} else {
2325 					txq->ift_pullups++;
2326 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2327 						return (ENOMEM);
2328 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2329 				}
2330 			} else {
2331 				txq->ift_pullups++;
2332 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2333 					return (ENOMEM);
2334 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2335 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2336 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2337 			}
2338 		} else {
2339 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2340 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2341 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2342 		}
2343 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2344 		pi->ipi_ipproto = ip->ip_p;
2345 		pi->ipi_flags |= IPI_TX_IPV4;
2346 
2347 		if (pi->ipi_csum_flags & CSUM_IP)
2348                        ip->ip_sum = 0;
2349 
2350 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2351 			if (__predict_false(th == NULL)) {
2352 				txq->ift_pullups++;
2353 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2354 					return (ENOMEM);
2355 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2356 			}
2357 			pi->ipi_tcp_hflags = th->th_flags;
2358 			pi->ipi_tcp_hlen = th->th_off << 2;
2359 			pi->ipi_tcp_seq = th->th_seq;
2360 		}
2361 		if (IS_TSO4(pi)) {
2362 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2363 				return (ENXIO);
2364 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2365 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2366 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2367 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2368 				ip->ip_sum = 0;
2369 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2370 			}
2371 		}
2372 		break;
2373 	}
2374 #endif
2375 #ifdef INET6
2376 	case ETHERTYPE_IPV6:
2377 	{
2378 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2379 		struct tcphdr *th;
2380 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2381 
2382 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2383 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2384 				return (ENOMEM);
2385 		}
2386 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2387 
2388 		/* XXX-BZ this will go badly in case of ext hdrs. */
2389 		pi->ipi_ipproto = ip6->ip6_nxt;
2390 		pi->ipi_flags |= IPI_TX_IPV6;
2391 
2392 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2393 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2394 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2395 					return (ENOMEM);
2396 			}
2397 			pi->ipi_tcp_hflags = th->th_flags;
2398 			pi->ipi_tcp_hlen = th->th_off << 2;
2399 		}
2400 		if (IS_TSO6(pi)) {
2401 
2402 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2403 				return (ENXIO);
2404 			/*
2405 			 * The corresponding flag is set by the stack in the IPv4
2406 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2407 			 * So, set it here because the rest of the flow requires it.
2408 			 */
2409 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2410 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2411 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2412 		}
2413 		break;
2414 	}
2415 #endif
2416 	default:
2417 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2418 		pi->ipi_ip_hlen = 0;
2419 		break;
2420 	}
2421 	*mp = m;
2422 
2423 	return (0);
2424 }
2425 
2426 static  __noinline  struct mbuf *
2427 collapse_pkthdr(struct mbuf *m0)
2428 {
2429 	struct mbuf *m, *m_next, *tmp;
2430 
2431 	m = m0;
2432 	m_next = m->m_next;
2433 	while (m_next != NULL && m_next->m_len == 0) {
2434 		m = m_next;
2435 		m->m_next = NULL;
2436 		m_free(m);
2437 		m_next = m_next->m_next;
2438 	}
2439 	m = m0;
2440 	m->m_next = m_next;
2441 	if ((m_next->m_flags & M_EXT) == 0) {
2442 		m = m_defrag(m, M_NOWAIT);
2443 	} else {
2444 		tmp = m_next->m_next;
2445 		memcpy(m_next, m, MPKTHSIZE);
2446 		m = m_next;
2447 		m->m_next = tmp;
2448 	}
2449 	return (m);
2450 }
2451 
2452 /*
2453  * If dodgy hardware rejects the scatter gather chain we've handed it
2454  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2455  * m_defrag'd mbufs
2456  */
2457 static __noinline struct mbuf *
2458 iflib_remove_mbuf(iflib_txq_t txq)
2459 {
2460 	int ntxd, i, pidx;
2461 	struct mbuf *m, *mh, **ifsd_m;
2462 
2463 	pidx = txq->ift_pidx;
2464 	ifsd_m = txq->ift_sds.ifsd_m;
2465 	ntxd = txq->ift_size;
2466 	mh = m = ifsd_m[pidx];
2467 	ifsd_m[pidx] = NULL;
2468 #if MEMORY_LOGGING
2469 	txq->ift_dequeued++;
2470 #endif
2471 	i = 1;
2472 
2473 	while (m) {
2474 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2475 #if MEMORY_LOGGING
2476 		txq->ift_dequeued++;
2477 #endif
2478 		m = m->m_next;
2479 		i++;
2480 	}
2481 	return (mh);
2482 }
2483 
2484 static int
2485 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2486 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2487 			  int max_segs, int flags)
2488 {
2489 	if_ctx_t ctx;
2490 	if_shared_ctx_t		sctx;
2491 	if_softc_ctx_t		scctx;
2492 	int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2493 	struct mbuf *m, *tmp, **ifsd_m, **mp;
2494 
2495 	m = *m0;
2496 
2497 	/*
2498 	 * Please don't ever do this
2499 	 */
2500 	if (__predict_false(m->m_len == 0))
2501 		*m0 = m = collapse_pkthdr(m);
2502 
2503 	ctx = txq->ift_ctx;
2504 	sctx = ctx->ifc_sctx;
2505 	scctx = &ctx->ifc_softc_ctx;
2506 	ifsd_m = txq->ift_sds.ifsd_m;
2507 	ntxd = txq->ift_size;
2508 	pidx = txq->ift_pidx;
2509 	if (map != NULL) {
2510 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2511 
2512 		err = bus_dmamap_load_mbuf_sg(tag, map,
2513 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2514 		if (err)
2515 			return (err);
2516 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2517 		i = 0;
2518 		next = pidx;
2519 		mask = (txq->ift_size-1);
2520 		m = *m0;
2521 		do {
2522 			mp = &ifsd_m[next];
2523 			*mp = m;
2524 			m = m->m_next;
2525 			if (__predict_false((*mp)->m_len == 0)) {
2526 				m_free(*mp);
2527 				*mp = NULL;
2528 			} else
2529 				next = (pidx + i) & (ntxd-1);
2530 		} while (m != NULL);
2531 	} else {
2532 		int buflen, sgsize, max_sgsize;
2533 		vm_offset_t vaddr;
2534 		vm_paddr_t curaddr;
2535 
2536 		count = i = 0;
2537 		maxsegsz = sctx->isc_tx_maxsize;
2538 		m = *m0;
2539 		do {
2540 			if (__predict_false(m->m_len <= 0)) {
2541 				tmp = m;
2542 				m = m->m_next;
2543 				tmp->m_next = NULL;
2544 				m_free(tmp);
2545 				continue;
2546 			}
2547 			buflen = m->m_len;
2548 			vaddr = (vm_offset_t)m->m_data;
2549 			/*
2550 			 * see if we can't be smarter about physically
2551 			 * contiguous mappings
2552 			 */
2553 			next = (pidx + count) & (ntxd-1);
2554 			MPASS(ifsd_m[next] == NULL);
2555 #if MEMORY_LOGGING
2556 			txq->ift_enqueued++;
2557 #endif
2558 			ifsd_m[next] = m;
2559 			while (buflen > 0) {
2560 				max_sgsize = MIN(buflen, maxsegsz);
2561 				curaddr = pmap_kextract(vaddr);
2562 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2563 				sgsize = MIN(sgsize, max_sgsize);
2564 				segs[i].ds_addr = curaddr;
2565 				segs[i].ds_len = sgsize;
2566 				vaddr += sgsize;
2567 				buflen -= sgsize;
2568 				i++;
2569 				if (i >= max_segs)
2570 					goto err;
2571 			}
2572 			count++;
2573 			tmp = m;
2574 			m = m->m_next;
2575 		} while (m != NULL);
2576 		*nsegs = i;
2577 	}
2578 	return (0);
2579 err:
2580 	*m0 = iflib_remove_mbuf(txq);
2581 	return (EFBIG);
2582 }
2583 
2584 static int
2585 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2586 {
2587 	if_ctx_t		ctx;
2588 	if_shared_ctx_t		sctx;
2589 	if_softc_ctx_t		scctx;
2590 	bus_dma_segment_t	*segs;
2591 	struct mbuf		*m_head;
2592 	bus_dmamap_t		map;
2593 	struct if_pkt_info	pi;
2594 	int remap = 0;
2595 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
2596 	bus_dma_tag_t desc_tag;
2597 
2598 	segs = txq->ift_segs;
2599 	ctx = txq->ift_ctx;
2600 	sctx = ctx->ifc_sctx;
2601 	scctx = &ctx->ifc_softc_ctx;
2602 	segs = txq->ift_segs;
2603 	ntxd = txq->ift_size;
2604 	m_head = *m_headp;
2605 	map = NULL;
2606 
2607 	/*
2608 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
2609 	 */
2610 	cidx = txq->ift_cidx;
2611 	pidx = txq->ift_pidx;
2612 	next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
2613 
2614 	/* prefetch the next cache line of mbuf pointers and flags */
2615 	prefetch(&txq->ift_sds.ifsd_m[next]);
2616 	if (txq->ift_sds.ifsd_map != NULL) {
2617 		prefetch(&txq->ift_sds.ifsd_map[next]);
2618 		map = txq->ift_sds.ifsd_map[pidx];
2619 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
2620 		prefetch(&txq->ift_sds.ifsd_flags[next]);
2621 	}
2622 
2623 
2624 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
2625 		desc_tag = txq->ift_tso_desc_tag;
2626 		max_segs = scctx->isc_tx_tso_segments_max;
2627 	} else {
2628 		desc_tag = txq->ift_desc_tag;
2629 		max_segs = scctx->isc_tx_nsegments;
2630 	}
2631 	m_head = *m_headp;
2632 	bzero(&pi, sizeof(pi));
2633 	pi.ipi_len = m_head->m_pkthdr.len;
2634 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
2635 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
2636 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
2637 	pi.ipi_pidx = pidx;
2638 	pi.ipi_qsidx = txq->ift_id;
2639 
2640 	/* deliberate bitwise OR to make one condition */
2641 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
2642 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
2643 			return (err);
2644 		m_head = *m_headp;
2645 	}
2646 
2647 retry:
2648 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
2649 defrag:
2650 	if (__predict_false(err)) {
2651 		switch (err) {
2652 		case EFBIG:
2653 			/* try collapse once and defrag once */
2654 			if (remap == 0)
2655 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
2656 			if (remap == 1)
2657 				m_head = m_defrag(*m_headp, M_NOWAIT);
2658 			remap++;
2659 			if (__predict_false(m_head == NULL))
2660 				goto defrag_failed;
2661 			txq->ift_mbuf_defrag++;
2662 			*m_headp = m_head;
2663 			goto retry;
2664 			break;
2665 		case ENOMEM:
2666 			txq->ift_no_tx_dma_setup++;
2667 			break;
2668 		default:
2669 			txq->ift_no_tx_dma_setup++;
2670 			m_freem(*m_headp);
2671 			DBG_COUNTER_INC(tx_frees);
2672 			*m_headp = NULL;
2673 			break;
2674 		}
2675 		txq->ift_map_failed++;
2676 		DBG_COUNTER_INC(encap_load_mbuf_fail);
2677 		return (err);
2678 	}
2679 
2680 	/*
2681 	 * XXX assumes a 1 to 1 relationship between segments and
2682 	 *        descriptors - this does not hold true on all drivers, e.g.
2683 	 *        cxgb
2684 	 */
2685 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
2686 		txq->ift_no_desc_avail++;
2687 		if (map != NULL)
2688 			bus_dmamap_unload(desc_tag, map);
2689 		DBG_COUNTER_INC(encap_txq_avail_fail);
2690 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
2691 			GROUPTASK_ENQUEUE(&txq->ift_task);
2692 		return (ENOBUFS);
2693 	}
2694 	pi.ipi_segs = segs;
2695 	pi.ipi_nsegs = nsegs;
2696 
2697 	MPASS(pidx >= 0 && pidx < txq->ift_size);
2698 #ifdef PKT_DEBUG
2699 	print_pkt(&pi);
2700 #endif
2701 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
2702 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2703 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2704 
2705 		DBG_COUNTER_INC(tx_encap);
2706 		MPASS(pi.ipi_new_pidx >= 0 &&
2707 		    pi.ipi_new_pidx < txq->ift_size);
2708 
2709 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
2710 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
2711 			ndesc += txq->ift_size;
2712 			txq->ift_gen = 1;
2713 		}
2714 		/*
2715 		 * drivers can need as many as
2716 		 * two sentinels
2717 		 */
2718 		MPASS(ndesc <= pi.ipi_nsegs + 2);
2719 		MPASS(pi.ipi_new_pidx != pidx);
2720 		MPASS(ndesc > 0);
2721 		txq->ift_in_use += ndesc;
2722 		/*
2723 		 * We update the last software descriptor again here because there may
2724 		 * be a sentinel and/or there may be more mbufs than segments
2725 		 */
2726 		txq->ift_pidx = pi.ipi_new_pidx;
2727 		txq->ift_npending += pi.ipi_ndescs;
2728 	} else if (__predict_false(err == EFBIG && remap < 2)) {
2729 		*m_headp = m_head = iflib_remove_mbuf(txq);
2730 		remap = 1;
2731 		txq->ift_txd_encap_efbig++;
2732 		goto defrag;
2733 	} else
2734 		DBG_COUNTER_INC(encap_txd_encap_fail);
2735 	return (err);
2736 
2737 defrag_failed:
2738 	txq->ift_mbuf_defrag_failed++;
2739 	txq->ift_map_failed++;
2740 	m_freem(*m_headp);
2741 	DBG_COUNTER_INC(tx_frees);
2742 	*m_headp = NULL;
2743 	return (ENOMEM);
2744 }
2745 
2746 /* forward compatibility for cxgb */
2747 #define FIRST_QSET(ctx) 0
2748 
2749 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2750 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2751 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2752 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2753 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2754 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2755 
2756 
2757 
2758 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring
2759  * doorbell writes
2760  *
2761  * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic
2762  */
2763 #define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2)
2764 
2765 static inline int
2766 iflib_txq_min_occupancy(iflib_txq_t txq)
2767 {
2768 	if_ctx_t ctx;
2769 
2770 	ctx = txq->ift_ctx;
2771 	return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx,
2772 	    txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) +
2773 	    MAX_TX_DESC(ctx));
2774 }
2775 
2776 static void
2777 iflib_tx_desc_free(iflib_txq_t txq, int n)
2778 {
2779 	int hasmap;
2780 	uint32_t qsize, cidx, mask, gen;
2781 	struct mbuf *m, **ifsd_m;
2782 	uint8_t *ifsd_flags;
2783 	bus_dmamap_t *ifsd_map;
2784 
2785 	cidx = txq->ift_cidx;
2786 	gen = txq->ift_gen;
2787 	qsize = txq->ift_size;
2788 	mask = qsize-1;
2789 	hasmap = txq->ift_sds.ifsd_map != NULL;
2790 	ifsd_flags = txq->ift_sds.ifsd_flags;
2791 	ifsd_m = txq->ift_sds.ifsd_m;
2792 	ifsd_map = txq->ift_sds.ifsd_map;
2793 
2794 	while (n--) {
2795 		prefetch(ifsd_m[(cidx + 3) & mask]);
2796 		prefetch(ifsd_m[(cidx + 4) & mask]);
2797 
2798 		if (ifsd_m[cidx] != NULL) {
2799 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
2800 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
2801 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
2802 				/*
2803 				 * does it matter if it's not the TSO tag? If so we'll
2804 				 * have to add the type to flags
2805 				 */
2806 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
2807 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
2808 			}
2809 			if ((m = ifsd_m[cidx]) != NULL) {
2810 				/* XXX we don't support any drivers that batch packets yet */
2811 				MPASS(m->m_nextpkt == NULL);
2812 
2813 				m_free(m);
2814 				ifsd_m[cidx] = NULL;
2815 #if MEMORY_LOGGING
2816 				txq->ift_dequeued++;
2817 #endif
2818 				DBG_COUNTER_INC(tx_frees);
2819 			}
2820 		}
2821 		if (__predict_false(++cidx == qsize)) {
2822 			cidx = 0;
2823 			gen = 0;
2824 		}
2825 	}
2826 	txq->ift_cidx = cidx;
2827 	txq->ift_gen = gen;
2828 }
2829 
2830 static __inline int
2831 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
2832 {
2833 	int reclaim;
2834 	if_ctx_t ctx = txq->ift_ctx;
2835 
2836 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
2837 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
2838 
2839 	/*
2840 	 * Need a rate-limiting check so that this isn't called every time
2841 	 */
2842 	iflib_tx_credits_update(ctx, txq);
2843 	reclaim = DESC_RECLAIMABLE(txq);
2844 
2845 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
2846 #ifdef INVARIANTS
2847 		if (iflib_verbose_debug) {
2848 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
2849 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
2850 			       reclaim, thresh);
2851 
2852 		}
2853 #endif
2854 		return (0);
2855 	}
2856 	iflib_tx_desc_free(txq, reclaim);
2857 	txq->ift_cleaned += reclaim;
2858 	txq->ift_in_use -= reclaim;
2859 
2860 	if (txq->ift_active == FALSE)
2861 		txq->ift_active = TRUE;
2862 
2863 	return (reclaim);
2864 }
2865 
2866 static struct mbuf **
2867 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset)
2868 {
2869 
2870 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)]));
2871 }
2872 
2873 static void
2874 iflib_txq_check_drain(iflib_txq_t txq, int budget)
2875 {
2876 
2877 	ifmp_ring_check_drainage(txq->ift_br[0], budget);
2878 }
2879 
2880 static uint32_t
2881 iflib_txq_can_drain(struct ifmp_ring *r)
2882 {
2883 	iflib_txq_t txq = r->cookie;
2884 	if_ctx_t ctx = txq->ift_ctx;
2885 
2886 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
2887 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false));
2888 }
2889 
2890 static uint32_t
2891 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2892 {
2893 	iflib_txq_t txq = r->cookie;
2894 	if_ctx_t ctx = txq->ift_ctx;
2895 	if_t ifp = ctx->ifc_ifp;
2896 	struct mbuf **mp, *m;
2897 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used;
2898 
2899 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
2900 			    !LINK_ACTIVE(ctx))) {
2901 		DBG_COUNTER_INC(txq_drain_notready);
2902 		return (0);
2903 	}
2904 
2905 	avail = IDXDIFF(pidx, cidx, r->size);
2906 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
2907 		DBG_COUNTER_INC(txq_drain_flushing);
2908 		for (i = 0; i < avail; i++) {
2909 			m_free(r->items[(cidx + i) & (r->size-1)]);
2910 			r->items[(cidx + i) & (r->size-1)] = NULL;
2911 		}
2912 		return (avail);
2913 	}
2914 	iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
2915 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
2916 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
2917 		CALLOUT_LOCK(txq);
2918 		callout_stop(&txq->ift_timer);
2919 		callout_stop(&txq->ift_db_check);
2920 		CALLOUT_UNLOCK(txq);
2921 		DBG_COUNTER_INC(txq_drain_oactive);
2922 		return (0);
2923 	}
2924 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
2925 	count = MIN(avail, TX_BATCH_SIZE);
2926 #ifdef INVARIANTS
2927 	if (iflib_verbose_debug)
2928 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
2929 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
2930 #endif
2931 
2932 	for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) {
2933 		mp = _ring_peek_one(r, cidx, i);
2934 		MPASS(mp != NULL && *mp != NULL);
2935 		in_use_prev = txq->ift_in_use;
2936 		if ((err = iflib_encap(txq, mp)) == ENOBUFS) {
2937 			DBG_COUNTER_INC(txq_drain_encapfail);
2938 			/* no room - bail out */
2939 			break;
2940 		}
2941 		consumed++;
2942 		if (err) {
2943 			DBG_COUNTER_INC(txq_drain_encapfail);
2944 			/* we can't send this packet - skip it */
2945 			continue;
2946 		}
2947 		pkt_sent++;
2948 		m = *mp;
2949 		DBG_COUNTER_INC(tx_sent);
2950 		bytes_sent += m->m_pkthdr.len;
2951 		if (m->m_flags & M_MCAST)
2952 			mcast_sent++;
2953 
2954 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
2955 		desc_used += (txq->ift_in_use - in_use_prev);
2956 		iflib_txd_db_check(ctx, txq, FALSE);
2957 		ETHER_BPF_MTAP(ifp, m);
2958 		if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
2959 			break;
2960 
2961 		if (desc_used >= TXQ_MAX_DB_CONSUMED(txq->ift_size))
2962 			break;
2963 	}
2964 
2965 	if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending)
2966 		iflib_txd_db_check(ctx, txq, TRUE);
2967 	else if ((txq->ift_db_pending || TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2) &&
2968 		 (callout_pending(&txq->ift_db_check) == 0)) {
2969 		txq->ift_db_pending_queued = txq->ift_db_pending;
2970 		callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check,
2971 				 txq, txq->ift_db_check.c_cpu);
2972 	}
2973 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
2974 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
2975 	if (mcast_sent)
2976 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
2977 #ifdef INVARIANTS
2978 	if (iflib_verbose_debug)
2979 		printf("consumed=%d\n", consumed);
2980 #endif
2981 	return (consumed);
2982 }
2983 
2984 static uint32_t
2985 iflib_txq_drain_always(struct ifmp_ring *r)
2986 {
2987 	return (1);
2988 }
2989 
2990 static uint32_t
2991 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2992 {
2993 	int i, avail;
2994 	struct mbuf **mp;
2995 	iflib_txq_t txq;
2996 
2997 	txq = r->cookie;
2998 
2999 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3000 	CALLOUT_LOCK(txq);
3001 	callout_stop(&txq->ift_timer);
3002 	callout_stop(&txq->ift_db_check);
3003 	CALLOUT_UNLOCK(txq);
3004 
3005 	avail = IDXDIFF(pidx, cidx, r->size);
3006 	for (i = 0; i < avail; i++) {
3007 		mp = _ring_peek_one(r, cidx, i);
3008 		m_freem(*mp);
3009 	}
3010 	MPASS(ifmp_ring_is_stalled(r) == 0);
3011 	return (avail);
3012 }
3013 
3014 static void
3015 iflib_ifmp_purge(iflib_txq_t txq)
3016 {
3017 	struct ifmp_ring *r;
3018 
3019 	r = txq->ift_br[0];
3020 	r->drain = iflib_txq_drain_free;
3021 	r->can_drain = iflib_txq_drain_always;
3022 
3023 	ifmp_ring_check_drainage(r, r->size);
3024 
3025 	r->drain = iflib_txq_drain;
3026 	r->can_drain = iflib_txq_can_drain;
3027 }
3028 
3029 static void
3030 _task_fn_tx(void *context)
3031 {
3032 	iflib_txq_t txq = context;
3033 	if_ctx_t ctx = txq->ift_ctx;
3034 
3035 #ifdef IFLIB_DIAGNOSTICS
3036 	txq->ift_cpu_exec_count[curcpu]++;
3037 #endif
3038 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3039 		return;
3040 	ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3041 }
3042 
3043 static void
3044 _task_fn_rx(void *context)
3045 {
3046 	iflib_rxq_t rxq = context;
3047 	if_ctx_t ctx = rxq->ifr_ctx;
3048 	bool more;
3049 	int rc;
3050 
3051 #ifdef IFLIB_DIAGNOSTICS
3052 	rxq->ifr_cpu_exec_count[curcpu]++;
3053 #endif
3054 	DBG_COUNTER_INC(task_fn_rxs);
3055 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3056 		return;
3057 
3058 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3059 		if (ctx->ifc_flags & IFC_LEGACY)
3060 			IFDI_INTR_ENABLE(ctx);
3061 		else {
3062 			DBG_COUNTER_INC(rx_intr_enables);
3063 			rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3064 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3065 		}
3066 	}
3067 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3068 		return;
3069 	if (more)
3070 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3071 }
3072 
3073 static void
3074 _task_fn_admin(void *context)
3075 {
3076 	if_ctx_t ctx = context;
3077 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3078 	iflib_txq_t txq;
3079 	int i;
3080 
3081 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3082 		return;
3083 
3084 	CTX_LOCK(ctx);
3085 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3086 		CALLOUT_LOCK(txq);
3087 		callout_stop(&txq->ift_timer);
3088 		CALLOUT_UNLOCK(txq);
3089 	}
3090 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3091 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3092 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3093 	IFDI_LINK_INTR_ENABLE(ctx);
3094 	CTX_UNLOCK(ctx);
3095 
3096 	if (LINK_ACTIVE(ctx) == 0)
3097 		return;
3098 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3099 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3100 }
3101 
3102 
3103 static void
3104 _task_fn_iov(void *context)
3105 {
3106 	if_ctx_t ctx = context;
3107 
3108 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3109 		return;
3110 
3111 	CTX_LOCK(ctx);
3112 	IFDI_VFLR_HANDLE(ctx);
3113 	CTX_UNLOCK(ctx);
3114 }
3115 
3116 static int
3117 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3118 {
3119 	int err;
3120 	if_int_delay_info_t info;
3121 	if_ctx_t ctx;
3122 
3123 	info = (if_int_delay_info_t)arg1;
3124 	ctx = info->iidi_ctx;
3125 	info->iidi_req = req;
3126 	info->iidi_oidp = oidp;
3127 	CTX_LOCK(ctx);
3128 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3129 	CTX_UNLOCK(ctx);
3130 	return (err);
3131 }
3132 
3133 /*********************************************************************
3134  *
3135  *  IFNET FUNCTIONS
3136  *
3137  **********************************************************************/
3138 
3139 static void
3140 iflib_if_init_locked(if_ctx_t ctx)
3141 {
3142 	iflib_stop(ctx);
3143 	iflib_init_locked(ctx);
3144 }
3145 
3146 
3147 static void
3148 iflib_if_init(void *arg)
3149 {
3150 	if_ctx_t ctx = arg;
3151 
3152 	CTX_LOCK(ctx);
3153 	iflib_if_init_locked(ctx);
3154 	CTX_UNLOCK(ctx);
3155 }
3156 
3157 static int
3158 iflib_if_transmit(if_t ifp, struct mbuf *m)
3159 {
3160 	if_ctx_t	ctx = if_getsoftc(ifp);
3161 
3162 	iflib_txq_t txq;
3163 	int err, qidx;
3164 
3165 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3166 		DBG_COUNTER_INC(tx_frees);
3167 		m_freem(m);
3168 		return (ENOBUFS);
3169 	}
3170 
3171 	MPASS(m->m_nextpkt == NULL);
3172 	qidx = 0;
3173 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3174 		qidx = QIDX(ctx, m);
3175 	/*
3176 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3177 	 */
3178 	txq = &ctx->ifc_txqs[qidx];
3179 
3180 #ifdef DRIVER_BACKPRESSURE
3181 	if (txq->ift_closed) {
3182 		while (m != NULL) {
3183 			next = m->m_nextpkt;
3184 			m->m_nextpkt = NULL;
3185 			m_freem(m);
3186 			m = next;
3187 		}
3188 		return (ENOBUFS);
3189 	}
3190 #endif
3191 #ifdef notyet
3192 	qidx = count = 0;
3193 	mp = marr;
3194 	next = m;
3195 	do {
3196 		count++;
3197 		next = next->m_nextpkt;
3198 	} while (next != NULL);
3199 
3200 	if (count > nitems(marr))
3201 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3202 			/* XXX check nextpkt */
3203 			m_freem(m);
3204 			/* XXX simplify for now */
3205 			DBG_COUNTER_INC(tx_frees);
3206 			return (ENOBUFS);
3207 		}
3208 	for (next = m, i = 0; next != NULL; i++) {
3209 		mp[i] = next;
3210 		next = next->m_nextpkt;
3211 		mp[i]->m_nextpkt = NULL;
3212 	}
3213 #endif
3214 	DBG_COUNTER_INC(tx_seen);
3215 	err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE);
3216 
3217 	if (err) {
3218 		GROUPTASK_ENQUEUE(&txq->ift_task);
3219 		/* support forthcoming later */
3220 #ifdef DRIVER_BACKPRESSURE
3221 		txq->ift_closed = TRUE;
3222 #endif
3223 		ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3224 		m_freem(m);
3225 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3226 		GROUPTASK_ENQUEUE(&txq->ift_task);
3227 	}
3228 
3229 	return (err);
3230 }
3231 
3232 static void
3233 iflib_if_qflush(if_t ifp)
3234 {
3235 	if_ctx_t ctx = if_getsoftc(ifp);
3236 	iflib_txq_t txq = ctx->ifc_txqs;
3237 	int i;
3238 
3239 	CTX_LOCK(ctx);
3240 	ctx->ifc_flags |= IFC_QFLUSH;
3241 	CTX_UNLOCK(ctx);
3242 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3243 		while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0])))
3244 			iflib_txq_check_drain(txq, 0);
3245 	CTX_LOCK(ctx);
3246 	ctx->ifc_flags &= ~IFC_QFLUSH;
3247 	CTX_UNLOCK(ctx);
3248 
3249 	if_qflush(ifp);
3250 }
3251 
3252 
3253 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3254 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3255 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3256 
3257 static int
3258 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3259 {
3260 	if_ctx_t ctx = if_getsoftc(ifp);
3261 	struct ifreq	*ifr = (struct ifreq *)data;
3262 #if defined(INET) || defined(INET6)
3263 	struct ifaddr	*ifa = (struct ifaddr *)data;
3264 #endif
3265 	bool		avoid_reset = FALSE;
3266 	int		err = 0, reinit = 0, bits;
3267 
3268 	switch (command) {
3269 	case SIOCSIFADDR:
3270 #ifdef INET
3271 		if (ifa->ifa_addr->sa_family == AF_INET)
3272 			avoid_reset = TRUE;
3273 #endif
3274 #ifdef INET6
3275 		if (ifa->ifa_addr->sa_family == AF_INET6)
3276 			avoid_reset = TRUE;
3277 #endif
3278 		/*
3279 		** Calling init results in link renegotiation,
3280 		** so we avoid doing it when possible.
3281 		*/
3282 		if (avoid_reset) {
3283 			if_setflagbits(ifp, IFF_UP,0);
3284 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3285 				reinit = 1;
3286 #ifdef INET
3287 			if (!(if_getflags(ifp) & IFF_NOARP))
3288 				arp_ifinit(ifp, ifa);
3289 #endif
3290 		} else
3291 			err = ether_ioctl(ifp, command, data);
3292 		break;
3293 	case SIOCSIFMTU:
3294 		CTX_LOCK(ctx);
3295 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3296 			CTX_UNLOCK(ctx);
3297 			break;
3298 		}
3299 		bits = if_getdrvflags(ifp);
3300 		/* stop the driver and free any clusters before proceeding */
3301 		iflib_stop(ctx);
3302 
3303 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3304 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3305 				ctx->ifc_flags |= IFC_MULTISEG;
3306 			else
3307 				ctx->ifc_flags &= ~IFC_MULTISEG;
3308 			err = if_setmtu(ifp, ifr->ifr_mtu);
3309 		}
3310 		iflib_init_locked(ctx);
3311 		if_setdrvflags(ifp, bits);
3312 		CTX_UNLOCK(ctx);
3313 		break;
3314 	case SIOCSIFFLAGS:
3315 		CTX_LOCK(ctx);
3316 		if (if_getflags(ifp) & IFF_UP) {
3317 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3318 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3319 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3320 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3321 				}
3322 			} else
3323 				reinit = 1;
3324 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3325 			iflib_stop(ctx);
3326 		}
3327 		ctx->ifc_if_flags = if_getflags(ifp);
3328 		CTX_UNLOCK(ctx);
3329 		break;
3330 
3331 		break;
3332 	case SIOCADDMULTI:
3333 	case SIOCDELMULTI:
3334 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3335 			CTX_LOCK(ctx);
3336 			IFDI_INTR_DISABLE(ctx);
3337 			IFDI_MULTI_SET(ctx);
3338 			IFDI_INTR_ENABLE(ctx);
3339 			CTX_UNLOCK(ctx);
3340 		}
3341 		break;
3342 	case SIOCSIFMEDIA:
3343 		CTX_LOCK(ctx);
3344 		IFDI_MEDIA_SET(ctx);
3345 		CTX_UNLOCK(ctx);
3346 		/* falls thru */
3347 	case SIOCGIFMEDIA:
3348 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3349 		break;
3350 	case SIOCGI2C:
3351 	{
3352 		struct ifi2creq i2c;
3353 
3354 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3355 		if (err != 0)
3356 			break;
3357 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3358 			err = EINVAL;
3359 			break;
3360 		}
3361 		if (i2c.len > sizeof(i2c.data)) {
3362 			err = EINVAL;
3363 			break;
3364 		}
3365 
3366 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3367 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3368 		break;
3369 	}
3370 	case SIOCSIFCAP:
3371 	{
3372 		int mask, setmask;
3373 
3374 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3375 		setmask = 0;
3376 #ifdef TCP_OFFLOAD
3377 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3378 #endif
3379 		setmask |= (mask & IFCAP_FLAGS);
3380 
3381 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3382 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3383 		if ((mask & IFCAP_WOL) &&
3384 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3385 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3386 		if_vlancap(ifp);
3387 		/*
3388 		 * want to ensure that traffic has stopped before we change any of the flags
3389 		 */
3390 		if (setmask) {
3391 			CTX_LOCK(ctx);
3392 			bits = if_getdrvflags(ifp);
3393 			if (bits & IFF_DRV_RUNNING)
3394 				iflib_stop(ctx);
3395 			if_togglecapenable(ifp, setmask);
3396 			if (bits & IFF_DRV_RUNNING)
3397 				iflib_init_locked(ctx);
3398 			if_setdrvflags(ifp, bits);
3399 			CTX_UNLOCK(ctx);
3400 		}
3401 		break;
3402 	    }
3403 	case SIOCGPRIVATE_0:
3404 	case SIOCSDRVSPEC:
3405 	case SIOCGDRVSPEC:
3406 		CTX_LOCK(ctx);
3407 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3408 		CTX_UNLOCK(ctx);
3409 		break;
3410 	default:
3411 		err = ether_ioctl(ifp, command, data);
3412 		break;
3413 	}
3414 	if (reinit)
3415 		iflib_if_init(ctx);
3416 	return (err);
3417 }
3418 
3419 static uint64_t
3420 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3421 {
3422 	if_ctx_t ctx = if_getsoftc(ifp);
3423 
3424 	return (IFDI_GET_COUNTER(ctx, cnt));
3425 }
3426 
3427 /*********************************************************************
3428  *
3429  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3430  *
3431  **********************************************************************/
3432 
3433 static void
3434 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3435 {
3436 	if_ctx_t ctx = if_getsoftc(ifp);
3437 
3438 	if ((void *)ctx != arg)
3439 		return;
3440 
3441 	if ((vtag == 0) || (vtag > 4095))
3442 		return;
3443 
3444 	CTX_LOCK(ctx);
3445 	IFDI_VLAN_REGISTER(ctx, vtag);
3446 	/* Re-init to load the changes */
3447 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3448 		iflib_init_locked(ctx);
3449 	CTX_UNLOCK(ctx);
3450 }
3451 
3452 static void
3453 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3454 {
3455 	if_ctx_t ctx = if_getsoftc(ifp);
3456 
3457 	if ((void *)ctx != arg)
3458 		return;
3459 
3460 	if ((vtag == 0) || (vtag > 4095))
3461 		return;
3462 
3463 	CTX_LOCK(ctx);
3464 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3465 	/* Re-init to load the changes */
3466 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3467 		iflib_init_locked(ctx);
3468 	CTX_UNLOCK(ctx);
3469 }
3470 
3471 static void
3472 iflib_led_func(void *arg, int onoff)
3473 {
3474 	if_ctx_t ctx = arg;
3475 
3476 	CTX_LOCK(ctx);
3477 	IFDI_LED_FUNC(ctx, onoff);
3478 	CTX_UNLOCK(ctx);
3479 }
3480 
3481 /*********************************************************************
3482  *
3483  *  BUS FUNCTION DEFINITIONS
3484  *
3485  **********************************************************************/
3486 
3487 int
3488 iflib_device_probe(device_t dev)
3489 {
3490 	pci_vendor_info_t *ent;
3491 
3492 	uint16_t	pci_vendor_id, pci_device_id;
3493 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3494 	uint16_t	pci_rev_id;
3495 	if_shared_ctx_t sctx;
3496 
3497 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3498 		return (ENOTSUP);
3499 
3500 	pci_vendor_id = pci_get_vendor(dev);
3501 	pci_device_id = pci_get_device(dev);
3502 	pci_subvendor_id = pci_get_subvendor(dev);
3503 	pci_subdevice_id = pci_get_subdevice(dev);
3504 	pci_rev_id = pci_get_revid(dev);
3505 	if (sctx->isc_parse_devinfo != NULL)
3506 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3507 
3508 	ent = sctx->isc_vendor_info;
3509 	while (ent->pvi_vendor_id != 0) {
3510 		if (pci_vendor_id != ent->pvi_vendor_id) {
3511 			ent++;
3512 			continue;
3513 		}
3514 		if ((pci_device_id == ent->pvi_device_id) &&
3515 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3516 		     (ent->pvi_subvendor_id == 0)) &&
3517 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3518 		     (ent->pvi_subdevice_id == 0)) &&
3519 		    ((pci_rev_id == ent->pvi_rev_id) ||
3520 		     (ent->pvi_rev_id == 0))) {
3521 
3522 			device_set_desc_copy(dev, ent->pvi_name);
3523 			/* this needs to be changed to zero if the bus probing code
3524 			 * ever stops re-probing on best match because the sctx
3525 			 * may have its values over written by register calls
3526 			 * in subsequent probes
3527 			 */
3528 			return (BUS_PROBE_DEFAULT);
3529 		}
3530 		ent++;
3531 	}
3532 	return (ENXIO);
3533 }
3534 
3535 int
3536 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3537 {
3538 	int err, rid, msix, msix_bar;
3539 	if_ctx_t ctx;
3540 	if_t ifp;
3541 	if_softc_ctx_t scctx;
3542 	int i;
3543 	uint16_t main_txq;
3544 	uint16_t main_rxq;
3545 
3546 
3547 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3548 
3549 	if (sc == NULL) {
3550 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3551 		device_set_softc(dev, ctx);
3552 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3553 	}
3554 
3555 	ctx->ifc_sctx = sctx;
3556 	ctx->ifc_dev = dev;
3557 	ctx->ifc_softc = sc;
3558 
3559 	if ((err = iflib_register(ctx)) != 0) {
3560 		device_printf(dev, "iflib_register failed %d\n", err);
3561 		return (err);
3562 	}
3563 	iflib_add_device_sysctl_pre(ctx);
3564 
3565 	scctx = &ctx->ifc_softc_ctx;
3566 	ifp = ctx->ifc_ifp;
3567 
3568 	/*
3569 	 * XXX sanity check that ntxd & nrxd are a power of 2
3570 	 */
3571 	if (ctx->ifc_sysctl_ntxqs != 0)
3572 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
3573 	if (ctx->ifc_sysctl_nrxqs != 0)
3574 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
3575 
3576 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3577 		if (ctx->ifc_sysctl_ntxds[i] != 0)
3578 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
3579 		else
3580 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
3581 	}
3582 
3583 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3584 		if (ctx->ifc_sysctl_nrxds[i] != 0)
3585 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
3586 		else
3587 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
3588 	}
3589 
3590 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3591 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
3592 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
3593 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
3594 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
3595 		}
3596 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
3597 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
3598 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
3599 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
3600 		}
3601 	}
3602 
3603 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3604 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
3605 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
3606 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
3607 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
3608 		}
3609 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
3610 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
3611 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
3612 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
3613 		}
3614 	}
3615 
3616 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
3617 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
3618 		return (err);
3619 	}
3620 	_iflib_pre_assert(scctx);
3621 	ctx->ifc_txrx = *scctx->isc_txrx;
3622 
3623 #ifdef INVARIANTS
3624 	MPASS(scctx->isc_capenable);
3625 	if (scctx->isc_capenable & IFCAP_TXCSUM)
3626 		MPASS(scctx->isc_tx_csum_flags);
3627 #endif
3628 
3629 	if_setcapabilities(ifp, scctx->isc_capenable);
3630 	if_setcapenable(ifp, scctx->isc_capenable);
3631 
3632 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
3633 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
3634 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
3635 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
3636 
3637 #ifdef ACPI_DMAR
3638 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
3639 		ctx->ifc_flags |= IFC_DMAR;
3640 #endif
3641 
3642 	msix_bar = scctx->isc_msix_bar;
3643 
3644 	if(sctx->isc_flags & IFLIB_HAS_TXCQ)
3645 		main_txq = 1;
3646 	else
3647 		main_txq = 0;
3648 
3649 	if(sctx->isc_flags & IFLIB_HAS_RXCQ)
3650 		main_rxq = 1;
3651 	else
3652 		main_rxq = 0;
3653 
3654 	/* XXX change for per-queue sizes */
3655 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
3656 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
3657 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3658 		if (!powerof2(scctx->isc_nrxd[i])) {
3659 			/* round down instead? */
3660 			device_printf(dev, "# rx descriptors must be a power of 2\n");
3661 			err = EINVAL;
3662 			goto fail;
3663 		}
3664 	}
3665 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3666 		if (!powerof2(scctx->isc_ntxd[i])) {
3667 			device_printf(dev,
3668 			    "# tx descriptors must be a power of 2");
3669 			err = EINVAL;
3670 			goto fail;
3671 		}
3672 	}
3673 
3674 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
3675 	    MAX_SINGLE_PACKET_FRACTION)
3676 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
3677 		    MAX_SINGLE_PACKET_FRACTION);
3678 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
3679 	    MAX_SINGLE_PACKET_FRACTION)
3680 		scctx->isc_tx_tso_segments_max = max(1,
3681 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
3682 
3683 	/*
3684 	 * Protect the stack against modern hardware
3685 	 */
3686 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
3687 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
3688 
3689 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
3690 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
3691 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
3692 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
3693 	if (scctx->isc_rss_table_size == 0)
3694 		scctx->isc_rss_table_size = 64;
3695 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
3696 
3697 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
3698 	/* XXX format name */
3699 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
3700 	/*
3701 	** Now setup MSI or MSI/X, should
3702 	** return us the number of supported
3703 	** vectors. (Will be 1 for MSI)
3704 	*/
3705 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
3706 		msix = scctx->isc_vectors;
3707 	} else if (scctx->isc_msix_bar != 0)
3708 		msix = iflib_msix_init(ctx);
3709 	else {
3710 		scctx->isc_vectors = 1;
3711 		scctx->isc_ntxqsets = 1;
3712 		scctx->isc_nrxqsets = 1;
3713 		scctx->isc_intr = IFLIB_INTR_LEGACY;
3714 		msix = 0;
3715 	}
3716 	/* Get memory for the station queues */
3717 	if ((err = iflib_queues_alloc(ctx))) {
3718 		device_printf(dev, "Unable to allocate queue memory\n");
3719 		goto fail;
3720 	}
3721 
3722 	if ((err = iflib_qset_structures_setup(ctx))) {
3723 		device_printf(dev, "qset structure setup failed %d\n", err);
3724 		goto fail_queues;
3725 	}
3726 
3727 	IFDI_INTR_DISABLE(ctx);
3728 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
3729 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
3730 		goto fail_intr_free;
3731 	}
3732 	if (msix <= 1) {
3733 		rid = 0;
3734 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
3735 			MPASS(msix == 1);
3736 			rid = 1;
3737 		}
3738 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
3739 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
3740 			goto fail_intr_free;
3741 		}
3742 	}
3743 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
3744 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
3745 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
3746 		goto fail_detach;
3747 	}
3748 	if ((err = iflib_netmap_attach(ctx))) {
3749 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
3750 		goto fail_detach;
3751 	}
3752 	*ctxp = ctx;
3753 
3754 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
3755 	iflib_add_device_sysctl_post(ctx);
3756 	return (0);
3757 fail_detach:
3758 	ether_ifdetach(ctx->ifc_ifp);
3759 fail_intr_free:
3760 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
3761 		pci_release_msi(ctx->ifc_dev);
3762 fail_queues:
3763 	/* XXX free queues */
3764 fail:
3765 	IFDI_DETACH(ctx);
3766 	return (err);
3767 }
3768 
3769 int
3770 iflib_device_attach(device_t dev)
3771 {
3772 	if_ctx_t ctx;
3773 	if_shared_ctx_t sctx;
3774 
3775 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3776 		return (ENOTSUP);
3777 
3778 	pci_enable_busmaster(dev);
3779 
3780 	return (iflib_device_register(dev, NULL, sctx, &ctx));
3781 }
3782 
3783 int
3784 iflib_device_deregister(if_ctx_t ctx)
3785 {
3786 	if_t ifp = ctx->ifc_ifp;
3787 	iflib_txq_t txq;
3788 	iflib_rxq_t rxq;
3789 	device_t dev = ctx->ifc_dev;
3790 	int i;
3791 	struct taskqgroup *tqg;
3792 
3793 	/* Make sure VLANS are not using driver */
3794 	if (if_vlantrunkinuse(ifp)) {
3795 		device_printf(dev,"Vlan in use, detach first\n");
3796 		return (EBUSY);
3797 	}
3798 
3799 	CTX_LOCK(ctx);
3800 	ctx->ifc_in_detach = 1;
3801 	iflib_stop(ctx);
3802 	CTX_UNLOCK(ctx);
3803 
3804 	/* Unregister VLAN events */
3805 	if (ctx->ifc_vlan_attach_event != NULL)
3806 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
3807 	if (ctx->ifc_vlan_detach_event != NULL)
3808 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
3809 
3810 	iflib_netmap_detach(ifp);
3811 	ether_ifdetach(ifp);
3812 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
3813 	CTX_LOCK_DESTROY(ctx);
3814 	if (ctx->ifc_led_dev != NULL)
3815 		led_destroy(ctx->ifc_led_dev);
3816 	/* XXX drain any dependent tasks */
3817 	tqg = qgroup_if_io_tqg;
3818 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
3819 		callout_drain(&txq->ift_timer);
3820 		callout_drain(&txq->ift_db_check);
3821 		if (txq->ift_task.gt_uniq != NULL)
3822 			taskqgroup_detach(tqg, &txq->ift_task);
3823 	}
3824 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
3825 		if (rxq->ifr_task.gt_uniq != NULL)
3826 			taskqgroup_detach(tqg, &rxq->ifr_task);
3827 	}
3828 	tqg = qgroup_if_config_tqg;
3829 	if (ctx->ifc_admin_task.gt_uniq != NULL)
3830 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
3831 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
3832 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
3833 
3834 	IFDI_DETACH(ctx);
3835 	device_set_softc(ctx->ifc_dev, NULL);
3836 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
3837 		pci_release_msi(dev);
3838 	}
3839 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
3840 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
3841 	}
3842 	if (ctx->ifc_msix_mem != NULL) {
3843 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
3844 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
3845 		ctx->ifc_msix_mem = NULL;
3846 	}
3847 
3848 	bus_generic_detach(dev);
3849 	if_free(ifp);
3850 
3851 	iflib_tx_structures_free(ctx);
3852 	iflib_rx_structures_free(ctx);
3853 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
3854 		free(ctx->ifc_softc, M_IFLIB);
3855 	free(ctx, M_IFLIB);
3856 	return (0);
3857 }
3858 
3859 
3860 int
3861 iflib_device_detach(device_t dev)
3862 {
3863 	if_ctx_t ctx = device_get_softc(dev);
3864 
3865 	return (iflib_device_deregister(ctx));
3866 }
3867 
3868 int
3869 iflib_device_suspend(device_t dev)
3870 {
3871 	if_ctx_t ctx = device_get_softc(dev);
3872 
3873 	CTX_LOCK(ctx);
3874 	IFDI_SUSPEND(ctx);
3875 	CTX_UNLOCK(ctx);
3876 
3877 	return bus_generic_suspend(dev);
3878 }
3879 int
3880 iflib_device_shutdown(device_t dev)
3881 {
3882 	if_ctx_t ctx = device_get_softc(dev);
3883 
3884 	CTX_LOCK(ctx);
3885 	IFDI_SHUTDOWN(ctx);
3886 	CTX_UNLOCK(ctx);
3887 
3888 	return bus_generic_suspend(dev);
3889 }
3890 
3891 
3892 int
3893 iflib_device_resume(device_t dev)
3894 {
3895 	if_ctx_t ctx = device_get_softc(dev);
3896 	iflib_txq_t txq = ctx->ifc_txqs;
3897 
3898 	CTX_LOCK(ctx);
3899 	IFDI_RESUME(ctx);
3900 	iflib_init_locked(ctx);
3901 	CTX_UNLOCK(ctx);
3902 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
3903 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3904 
3905 	return (bus_generic_resume(dev));
3906 }
3907 
3908 int
3909 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
3910 {
3911 	int error;
3912 	if_ctx_t ctx = device_get_softc(dev);
3913 
3914 	CTX_LOCK(ctx);
3915 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
3916 	CTX_UNLOCK(ctx);
3917 
3918 	return (error);
3919 }
3920 
3921 void
3922 iflib_device_iov_uninit(device_t dev)
3923 {
3924 	if_ctx_t ctx = device_get_softc(dev);
3925 
3926 	CTX_LOCK(ctx);
3927 	IFDI_IOV_UNINIT(ctx);
3928 	CTX_UNLOCK(ctx);
3929 }
3930 
3931 int
3932 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
3933 {
3934 	int error;
3935 	if_ctx_t ctx = device_get_softc(dev);
3936 
3937 	CTX_LOCK(ctx);
3938 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
3939 	CTX_UNLOCK(ctx);
3940 
3941 	return (error);
3942 }
3943 
3944 /*********************************************************************
3945  *
3946  *  MODULE FUNCTION DEFINITIONS
3947  *
3948  **********************************************************************/
3949 
3950 /*
3951  * - Start a fast taskqueue thread for each core
3952  * - Start a taskqueue for control operations
3953  */
3954 static int
3955 iflib_module_init(void)
3956 {
3957 	return (0);
3958 }
3959 
3960 static int
3961 iflib_module_event_handler(module_t mod, int what, void *arg)
3962 {
3963 	int err;
3964 
3965 	switch (what) {
3966 	case MOD_LOAD:
3967 		if ((err = iflib_module_init()) != 0)
3968 			return (err);
3969 		break;
3970 	case MOD_UNLOAD:
3971 		return (EBUSY);
3972 	default:
3973 		return (EOPNOTSUPP);
3974 	}
3975 
3976 	return (0);
3977 }
3978 
3979 /*********************************************************************
3980  *
3981  *  PUBLIC FUNCTION DEFINITIONS
3982  *     ordered as in iflib.h
3983  *
3984  **********************************************************************/
3985 
3986 
3987 static void
3988 _iflib_assert(if_shared_ctx_t sctx)
3989 {
3990 	MPASS(sctx->isc_tx_maxsize);
3991 	MPASS(sctx->isc_tx_maxsegsize);
3992 
3993 	MPASS(sctx->isc_rx_maxsize);
3994 	MPASS(sctx->isc_rx_nsegments);
3995 	MPASS(sctx->isc_rx_maxsegsize);
3996 
3997 	MPASS(sctx->isc_nrxd_min[0]);
3998 	MPASS(sctx->isc_nrxd_max[0]);
3999 	MPASS(sctx->isc_nrxd_default[0]);
4000 	MPASS(sctx->isc_ntxd_min[0]);
4001 	MPASS(sctx->isc_ntxd_max[0]);
4002 	MPASS(sctx->isc_ntxd_default[0]);
4003 }
4004 
4005 static void
4006 _iflib_pre_assert(if_softc_ctx_t scctx)
4007 {
4008 
4009 	MPASS(scctx->isc_txrx->ift_txd_encap);
4010 	MPASS(scctx->isc_txrx->ift_txd_flush);
4011 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4012 	MPASS(scctx->isc_txrx->ift_rxd_available);
4013 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4014 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4015 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4016 }
4017 
4018 static int
4019 iflib_register(if_ctx_t ctx)
4020 {
4021 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4022 	driver_t *driver = sctx->isc_driver;
4023 	device_t dev = ctx->ifc_dev;
4024 	if_t ifp;
4025 
4026 	_iflib_assert(sctx);
4027 
4028 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4029 
4030 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4031 	if (ifp == NULL) {
4032 		device_printf(dev, "can not allocate ifnet structure\n");
4033 		return (ENOMEM);
4034 	}
4035 
4036 	/*
4037 	 * Initialize our context's device specific methods
4038 	 */
4039 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4040 	kobj_class_compile((kobj_class_t) driver);
4041 	driver->refs++;
4042 
4043 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4044 	if_setsoftc(ifp, ctx);
4045 	if_setdev(ifp, dev);
4046 	if_setinitfn(ifp, iflib_if_init);
4047 	if_setioctlfn(ifp, iflib_if_ioctl);
4048 	if_settransmitfn(ifp, iflib_if_transmit);
4049 	if_setqflushfn(ifp, iflib_if_qflush);
4050 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4051 
4052 	ctx->ifc_vlan_attach_event =
4053 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4054 							  EVENTHANDLER_PRI_FIRST);
4055 	ctx->ifc_vlan_detach_event =
4056 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4057 							  EVENTHANDLER_PRI_FIRST);
4058 
4059 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4060 					 iflib_media_change, iflib_media_status);
4061 
4062 	return (0);
4063 }
4064 
4065 
4066 static int
4067 iflib_queues_alloc(if_ctx_t ctx)
4068 {
4069 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4070 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4071 	device_t dev = ctx->ifc_dev;
4072 	int nrxqsets = scctx->isc_nrxqsets;
4073 	int ntxqsets = scctx->isc_ntxqsets;
4074 	iflib_txq_t txq;
4075 	iflib_rxq_t rxq;
4076 	iflib_fl_t fl = NULL;
4077 	int i, j, cpu, err, txconf, rxconf;
4078 	iflib_dma_info_t ifdip;
4079 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4080 	uint32_t *txqsizes = scctx->isc_txqsizes;
4081 	uint8_t nrxqs = sctx->isc_nrxqs;
4082 	uint8_t ntxqs = sctx->isc_ntxqs;
4083 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4084 	caddr_t *vaddrs;
4085 	uint64_t *paddrs;
4086 	struct ifmp_ring **brscp;
4087 	int nbuf_rings = 1; /* XXX determine dynamically */
4088 
4089 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4090 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4091 
4092 	brscp = NULL;
4093 	txq = NULL;
4094 	rxq = NULL;
4095 
4096 /* Allocate the TX ring struct memory */
4097 	if (!(txq =
4098 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4099 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4100 		device_printf(dev, "Unable to allocate TX ring memory\n");
4101 		err = ENOMEM;
4102 		goto fail;
4103 	}
4104 
4105 	/* Now allocate the RX */
4106 	if (!(rxq =
4107 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4108 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4109 		device_printf(dev, "Unable to allocate RX ring memory\n");
4110 		err = ENOMEM;
4111 		goto rx_fail;
4112 	}
4113 	if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4114 		device_printf(dev, "Unable to buf_ring_sc * memory\n");
4115 		err = ENOMEM;
4116 		goto rx_fail;
4117 	}
4118 
4119 	ctx->ifc_txqs = txq;
4120 	ctx->ifc_rxqs = rxq;
4121 
4122 	/*
4123 	 * XXX handle allocation failure
4124 	 */
4125 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4126 		/* Set up some basics */
4127 
4128 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4129 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4130 			err = ENOMEM;
4131 			goto err_tx_desc;
4132 		}
4133 		txq->ift_ifdi = ifdip;
4134 		for (j = 0; j < ntxqs; j++, ifdip++) {
4135 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4136 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4137 				err = ENOMEM;
4138 				goto err_tx_desc;
4139 			}
4140 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4141 		}
4142 		txq->ift_ctx = ctx;
4143 		txq->ift_id = i;
4144 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4145 			txq->ift_br_offset = 1;
4146 		} else {
4147 			txq->ift_br_offset = 0;
4148 		}
4149 		/* XXX fix this */
4150 		txq->ift_timer.c_cpu = cpu;
4151 		txq->ift_db_check.c_cpu = cpu;
4152 		txq->ift_nbr = nbuf_rings;
4153 
4154 		if (iflib_txsd_alloc(txq)) {
4155 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4156 			err = ENOMEM;
4157 			goto err_tx_desc;
4158 		}
4159 
4160 		/* Initialize the TX lock */
4161 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4162 		    device_get_nameunit(dev), txq->ift_id);
4163 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4164 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4165 		callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0);
4166 
4167 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4168 			 device_get_nameunit(dev), txq->ift_id);
4169 		TXDB_LOCK_INIT(txq);
4170 
4171 		txq->ift_br = brscp + i*nbuf_rings;
4172 		for (j = 0; j < nbuf_rings; j++) {
4173 			err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain,
4174 					      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4175 			if (err) {
4176 				/* XXX free any allocated rings */
4177 				device_printf(dev, "Unable to allocate buf_ring\n");
4178 				goto err_tx_desc;
4179 			}
4180 		}
4181 	}
4182 
4183 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4184 		/* Set up some basics */
4185 
4186 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4187 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4188 			err = ENOMEM;
4189 			goto err_tx_desc;
4190 		}
4191 
4192 		rxq->ifr_ifdi = ifdip;
4193 		for (j = 0; j < nrxqs; j++, ifdip++) {
4194 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4195 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4196 				err = ENOMEM;
4197 				goto err_tx_desc;
4198 			}
4199 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4200 		}
4201 		rxq->ifr_ctx = ctx;
4202 		rxq->ifr_id = i;
4203 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4204 			rxq->ifr_fl_offset = 1;
4205 		} else {
4206 			rxq->ifr_fl_offset = 0;
4207 		}
4208 		rxq->ifr_nfl = nfree_lists;
4209 		if (!(fl =
4210 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4211 			device_printf(dev, "Unable to allocate free list memory\n");
4212 			err = ENOMEM;
4213 			goto err_tx_desc;
4214 		}
4215 		rxq->ifr_fl = fl;
4216 		for (j = 0; j < nfree_lists; j++) {
4217 			rxq->ifr_fl[j].ifl_rxq = rxq;
4218 			rxq->ifr_fl[j].ifl_id = j;
4219 			rxq->ifr_fl[j].ifl_ifdi =
4220 			    &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4221 		}
4222         /* Allocate receive buffers for the ring*/
4223 		if (iflib_rxsd_alloc(rxq)) {
4224 			device_printf(dev,
4225 			    "Critical Failure setting up receive buffers\n");
4226 			err = ENOMEM;
4227 			goto err_rx_desc;
4228 		}
4229 	}
4230 
4231 	/* TXQs */
4232 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4233 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4234 	for (i = 0; i < ntxqsets; i++) {
4235 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4236 
4237 		for (j = 0; j < ntxqs; j++, di++) {
4238 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4239 			paddrs[i*ntxqs + j] = di->idi_paddr;
4240 		}
4241 	}
4242 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4243 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4244 		iflib_tx_structures_free(ctx);
4245 		free(vaddrs, M_IFLIB);
4246 		free(paddrs, M_IFLIB);
4247 		goto err_rx_desc;
4248 	}
4249 	free(vaddrs, M_IFLIB);
4250 	free(paddrs, M_IFLIB);
4251 
4252 	/* RXQs */
4253 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4254 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4255 	for (i = 0; i < nrxqsets; i++) {
4256 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4257 
4258 		for (j = 0; j < nrxqs; j++, di++) {
4259 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4260 			paddrs[i*nrxqs + j] = di->idi_paddr;
4261 		}
4262 	}
4263 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4264 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4265 		iflib_tx_structures_free(ctx);
4266 		free(vaddrs, M_IFLIB);
4267 		free(paddrs, M_IFLIB);
4268 		goto err_rx_desc;
4269 	}
4270 	free(vaddrs, M_IFLIB);
4271 	free(paddrs, M_IFLIB);
4272 
4273 	return (0);
4274 
4275 /* XXX handle allocation failure changes */
4276 err_rx_desc:
4277 err_tx_desc:
4278 	if (ctx->ifc_rxqs != NULL)
4279 		free(ctx->ifc_rxqs, M_IFLIB);
4280 	ctx->ifc_rxqs = NULL;
4281 	if (ctx->ifc_txqs != NULL)
4282 		free(ctx->ifc_txqs, M_IFLIB);
4283 	ctx->ifc_txqs = NULL;
4284 rx_fail:
4285 	if (brscp != NULL)
4286 		free(brscp, M_IFLIB);
4287 	if (rxq != NULL)
4288 		free(rxq, M_IFLIB);
4289 	if (txq != NULL)
4290 		free(txq, M_IFLIB);
4291 fail:
4292 	return (err);
4293 }
4294 
4295 static int
4296 iflib_tx_structures_setup(if_ctx_t ctx)
4297 {
4298 	iflib_txq_t txq = ctx->ifc_txqs;
4299 	int i;
4300 
4301 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4302 		iflib_txq_setup(txq);
4303 
4304 	return (0);
4305 }
4306 
4307 static void
4308 iflib_tx_structures_free(if_ctx_t ctx)
4309 {
4310 	iflib_txq_t txq = ctx->ifc_txqs;
4311 	int i, j;
4312 
4313 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4314 		iflib_txq_destroy(txq);
4315 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4316 			iflib_dma_free(&txq->ift_ifdi[j]);
4317 	}
4318 	free(ctx->ifc_txqs, M_IFLIB);
4319 	ctx->ifc_txqs = NULL;
4320 	IFDI_QUEUES_FREE(ctx);
4321 }
4322 
4323 /*********************************************************************
4324  *
4325  *  Initialize all receive rings.
4326  *
4327  **********************************************************************/
4328 static int
4329 iflib_rx_structures_setup(if_ctx_t ctx)
4330 {
4331 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4332 	int q;
4333 #if defined(INET6) || defined(INET)
4334 	int i, err;
4335 #endif
4336 
4337 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4338 #if defined(INET6) || defined(INET)
4339 		tcp_lro_free(&rxq->ifr_lc);
4340 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4341 		    TCP_LRO_ENTRIES, min(1024,
4342 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4343 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4344 			goto fail;
4345 		}
4346 		rxq->ifr_lro_enabled = TRUE;
4347 #endif
4348 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4349 	}
4350 	return (0);
4351 #if defined(INET6) || defined(INET)
4352 fail:
4353 	/*
4354 	 * Free RX software descriptors allocated so far, we will only handle
4355 	 * the rings that completed, the failing case will have
4356 	 * cleaned up for itself. 'q' failed, so its the terminus.
4357 	 */
4358 	rxq = ctx->ifc_rxqs;
4359 	for (i = 0; i < q; ++i, rxq++) {
4360 		iflib_rx_sds_free(rxq);
4361 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4362 	}
4363 	return (err);
4364 #endif
4365 }
4366 
4367 /*********************************************************************
4368  *
4369  *  Free all receive rings.
4370  *
4371  **********************************************************************/
4372 static void
4373 iflib_rx_structures_free(if_ctx_t ctx)
4374 {
4375 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4376 
4377 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4378 		iflib_rx_sds_free(rxq);
4379 	}
4380 }
4381 
4382 static int
4383 iflib_qset_structures_setup(if_ctx_t ctx)
4384 {
4385 	int err;
4386 
4387 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4388 		return (err);
4389 
4390 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4391 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4392 		iflib_tx_structures_free(ctx);
4393 		iflib_rx_structures_free(ctx);
4394 	}
4395 	return (err);
4396 }
4397 
4398 int
4399 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4400 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4401 {
4402 
4403 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4404 }
4405 
4406 static int
4407 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4408 {
4409 	int i, cpuid, eqid, count;
4410 
4411 	CPU_COPY(&ctx->ifc_cpus, cpus);
4412 	count = CPU_COUNT(&ctx->ifc_cpus);
4413 	eqid = qid % count;
4414 	/* clear up to the qid'th bit */
4415 	for (i = 0; i < eqid; i++) {
4416 		cpuid = CPU_FFS(cpus);
4417 		MPASS(cpuid != 0);
4418 		CPU_CLR(cpuid-1, cpus);
4419 	}
4420 	cpuid = CPU_FFS(cpus);
4421 	MPASS(cpuid != 0);
4422 	return (cpuid-1);
4423 }
4424 
4425 int
4426 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4427 						iflib_intr_type_t type, driver_filter_t *filter,
4428 						void *filter_arg, int qid, char *name)
4429 {
4430 	struct grouptask *gtask;
4431 	struct taskqgroup *tqg;
4432 	iflib_filter_info_t info;
4433 	cpuset_t cpus;
4434 	gtask_fn_t *fn;
4435 	int tqrid, err, cpuid;
4436 	void *q;
4437 
4438 	info = &ctx->ifc_filter_info;
4439 	tqrid = rid;
4440 
4441 	switch (type) {
4442 	/* XXX merge tx/rx for netmap? */
4443 	case IFLIB_INTR_TX:
4444 		q = &ctx->ifc_txqs[qid];
4445 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4446 		gtask = &ctx->ifc_txqs[qid].ift_task;
4447 		tqg = qgroup_if_io_tqg;
4448 		fn = _task_fn_tx;
4449 		GROUPTASK_INIT(gtask, 0, fn, q);
4450 		break;
4451 	case IFLIB_INTR_RX:
4452 		q = &ctx->ifc_rxqs[qid];
4453 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4454 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4455 		tqg = qgroup_if_io_tqg;
4456 		fn = _task_fn_rx;
4457 		GROUPTASK_INIT(gtask, 0, fn, q);
4458 		break;
4459 	case IFLIB_INTR_ADMIN:
4460 		q = ctx;
4461 		tqrid = -1;
4462 		info = &ctx->ifc_filter_info;
4463 		gtask = &ctx->ifc_admin_task;
4464 		tqg = qgroup_if_config_tqg;
4465 		fn = _task_fn_admin;
4466 		break;
4467 	default:
4468 		panic("unknown net intr type");
4469 	}
4470 
4471 	info->ifi_filter = filter;
4472 	info->ifi_filter_arg = filter_arg;
4473 	info->ifi_task = gtask;
4474 
4475 	err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info,  name);
4476 	if (err != 0) {
4477 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4478 		return (err);
4479 	}
4480 	if (type == IFLIB_INTR_ADMIN)
4481 		return (0);
4482 
4483 	if (tqrid != -1) {
4484 		cpuid = find_nth(ctx, &cpus, qid);
4485 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4486 	} else {
4487 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4488 	}
4489 
4490 	return (0);
4491 }
4492 
4493 void
4494 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4495 {
4496 	struct grouptask *gtask;
4497 	struct taskqgroup *tqg;
4498 	gtask_fn_t *fn;
4499 	void *q;
4500 
4501 	switch (type) {
4502 	case IFLIB_INTR_TX:
4503 		q = &ctx->ifc_txqs[qid];
4504 		gtask = &ctx->ifc_txqs[qid].ift_task;
4505 		tqg = qgroup_if_io_tqg;
4506 		fn = _task_fn_tx;
4507 		break;
4508 	case IFLIB_INTR_RX:
4509 		q = &ctx->ifc_rxqs[qid];
4510 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4511 		tqg = qgroup_if_io_tqg;
4512 		fn = _task_fn_rx;
4513 		break;
4514 	case IFLIB_INTR_IOV:
4515 		q = ctx;
4516 		gtask = &ctx->ifc_vflr_task;
4517 		tqg = qgroup_if_config_tqg;
4518 		rid = -1;
4519 		fn = _task_fn_iov;
4520 		break;
4521 	default:
4522 		panic("unknown net intr type");
4523 	}
4524 	GROUPTASK_INIT(gtask, 0, fn, q);
4525 	taskqgroup_attach(tqg, gtask, q, rid, name);
4526 }
4527 
4528 void
4529 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4530 {
4531 	if (irq->ii_tag)
4532 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4533 
4534 	if (irq->ii_res)
4535 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4536 }
4537 
4538 static int
4539 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4540 {
4541 	iflib_txq_t txq = ctx->ifc_txqs;
4542 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4543 	if_irq_t irq = &ctx->ifc_legacy_irq;
4544 	iflib_filter_info_t info;
4545 	struct grouptask *gtask;
4546 	struct taskqgroup *tqg;
4547 	gtask_fn_t *fn;
4548 	int tqrid;
4549 	void *q;
4550 	int err;
4551 
4552 	/*
4553 	 * group taskqueues aren't properly set up until SMP is started
4554 	 * so we disable interrupts until we can handle them post
4555 	 * SI_SUB_SMP
4556 	 */
4557 	IFDI_INTR_DISABLE(ctx);
4558 
4559 	q = &ctx->ifc_rxqs[0];
4560 	info = &rxq[0].ifr_filter_info;
4561 	gtask = &rxq[0].ifr_task;
4562 	tqg = qgroup_if_io_tqg;
4563 	tqrid = irq->ii_rid = *rid;
4564 	fn = _task_fn_rx;
4565 
4566 	ctx->ifc_flags |= IFC_LEGACY;
4567 	info->ifi_filter = filter;
4568 	info->ifi_filter_arg = filter_arg;
4569 	info->ifi_task = gtask;
4570 
4571 	/* We allocate a single interrupt resource */
4572 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0)
4573 		return (err);
4574 	GROUPTASK_INIT(gtask, 0, fn, q);
4575 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
4576 
4577 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
4578 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
4579 	return (0);
4580 }
4581 
4582 void
4583 iflib_led_create(if_ctx_t ctx)
4584 {
4585 
4586 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
4587 								  device_get_nameunit(ctx->ifc_dev));
4588 }
4589 
4590 void
4591 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
4592 {
4593 
4594 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
4595 }
4596 
4597 void
4598 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
4599 {
4600 
4601 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
4602 }
4603 
4604 void
4605 iflib_admin_intr_deferred(if_ctx_t ctx)
4606 {
4607 #ifdef INVARIANTS
4608 	struct grouptask *gtask;
4609 
4610 	gtask = &ctx->ifc_admin_task;
4611 	MPASS(gtask->gt_taskqueue != NULL);
4612 #endif
4613 
4614 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
4615 }
4616 
4617 void
4618 iflib_iov_intr_deferred(if_ctx_t ctx)
4619 {
4620 
4621 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
4622 }
4623 
4624 void
4625 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
4626 {
4627 
4628 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
4629 }
4630 
4631 void
4632 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
4633 	char *name)
4634 {
4635 
4636 	GROUPTASK_INIT(gtask, 0, fn, ctx);
4637 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
4638 }
4639 
4640 void
4641 iflib_config_gtask_deinit(struct grouptask *gtask)
4642 {
4643 
4644 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
4645 }
4646 
4647 void
4648 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
4649 {
4650 	if_t ifp = ctx->ifc_ifp;
4651 	iflib_txq_t txq = ctx->ifc_txqs;
4652 
4653 	if_setbaudrate(ifp, baudrate);
4654 
4655 	/* If link down, disable watchdog */
4656 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
4657 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
4658 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4659 	}
4660 	ctx->ifc_link_state = link_state;
4661 	if_link_state_change(ifp, link_state);
4662 }
4663 
4664 static int
4665 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
4666 {
4667 	int credits;
4668 #ifdef INVARIANTS
4669 	int credits_pre = txq->ift_cidx_processed;
4670 #endif
4671 
4672 	if (ctx->isc_txd_credits_update == NULL)
4673 		return (0);
4674 
4675 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0)
4676 		return (0);
4677 
4678 	txq->ift_processed += credits;
4679 	txq->ift_cidx_processed += credits;
4680 
4681 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
4682 	if (txq->ift_cidx_processed >= txq->ift_size)
4683 		txq->ift_cidx_processed -= txq->ift_size;
4684 	return (credits);
4685 }
4686 
4687 static int
4688 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget)
4689 {
4690 
4691 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
4692 	    budget));
4693 }
4694 
4695 void
4696 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
4697 	const char *description, if_int_delay_info_t info,
4698 	int offset, int value)
4699 {
4700 	info->iidi_ctx = ctx;
4701 	info->iidi_offset = offset;
4702 	info->iidi_value = value;
4703 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
4704 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
4705 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
4706 	    info, 0, iflib_sysctl_int_delay, "I", description);
4707 }
4708 
4709 struct mtx *
4710 iflib_ctx_lock_get(if_ctx_t ctx)
4711 {
4712 
4713 	return (&ctx->ifc_mtx);
4714 }
4715 
4716 static int
4717 iflib_msix_init(if_ctx_t ctx)
4718 {
4719 	device_t dev = ctx->ifc_dev;
4720 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4721 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4722 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
4723 	int iflib_num_tx_queues, iflib_num_rx_queues;
4724 	int err, admincnt, bar;
4725 
4726 	iflib_num_tx_queues = scctx->isc_ntxqsets;
4727 	iflib_num_rx_queues = scctx->isc_nrxqsets;
4728 
4729 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
4730 
4731 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
4732 	admincnt = sctx->isc_admin_intrcnt;
4733 	/* Override by tuneable */
4734 	if (enable_msix == 0)
4735 		goto msi;
4736 
4737 	/*
4738 	** When used in a virtualized environment
4739 	** PCI BUSMASTER capability may not be set
4740 	** so explicity set it here and rewrite
4741 	** the ENABLE in the MSIX control register
4742 	** at this point to cause the host to
4743 	** successfully initialize us.
4744 	*/
4745 	{
4746 		uint16_t pci_cmd_word;
4747 		int msix_ctrl, rid;
4748 
4749 		rid = 0;
4750 		pci_cmd_word = pci_read_config(dev, PCIR_COMMAND, 2);
4751 		pci_cmd_word |= PCIM_CMD_BUSMASTEREN;
4752 		pci_write_config(dev, PCIR_COMMAND, pci_cmd_word, 2);
4753 		pci_find_cap(dev, PCIY_MSIX, &rid);
4754 		rid += PCIR_MSIX_CTRL;
4755 		msix_ctrl = pci_read_config(dev, rid, 2);
4756 		msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
4757 		pci_write_config(dev, rid, msix_ctrl, 2);
4758 	}
4759 
4760 	/*
4761 	 * bar == -1 => "trust me I know what I'm doing"
4762 	 * https://www.youtube.com/watch?v=nnwWKkNau4I
4763 	 * Some drivers are for hardware that is so shoddily
4764 	 * documented that no one knows which bars are which
4765 	 * so the developer has to map all bars. This hack
4766 	 * allows shoddy garbage to use msix in this framework.
4767 	 */
4768 	if (bar != -1) {
4769 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
4770 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
4771 		if (ctx->ifc_msix_mem == NULL) {
4772 			/* May not be enabled */
4773 			device_printf(dev, "Unable to map MSIX table \n");
4774 			goto msi;
4775 		}
4776 	}
4777 	/* First try MSI/X */
4778 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
4779 		device_printf(dev, "System has MSIX disabled \n");
4780 		bus_release_resource(dev, SYS_RES_MEMORY,
4781 		    bar, ctx->ifc_msix_mem);
4782 		ctx->ifc_msix_mem = NULL;
4783 		goto msi;
4784 	}
4785 #if IFLIB_DEBUG
4786 	/* use only 1 qset in debug mode */
4787 	queuemsgs = min(msgs - admincnt, 1);
4788 #else
4789 	queuemsgs = msgs - admincnt;
4790 #endif
4791 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
4792 #ifdef RSS
4793 		queues = imin(queuemsgs, rss_getnumbuckets());
4794 #else
4795 		queues = queuemsgs;
4796 #endif
4797 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
4798 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
4799 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
4800 	} else {
4801 		device_printf(dev, "Unable to fetch CPU list\n");
4802 		/* Figure out a reasonable auto config value */
4803 		queues = min(queuemsgs, mp_ncpus);
4804 	}
4805 #ifdef  RSS
4806 	/* If we're doing RSS, clamp at the number of RSS buckets */
4807 	if (queues > rss_getnumbuckets())
4808 		queues = rss_getnumbuckets();
4809 #endif
4810 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
4811 		rx_queues = iflib_num_rx_queues;
4812 	else
4813 		rx_queues = queues;
4814 	/*
4815 	 * We want this to be all logical CPUs by default
4816 	 */
4817 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
4818 		tx_queues = iflib_num_tx_queues;
4819 	else
4820 		tx_queues = mp_ncpus;
4821 
4822 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
4823 #ifdef INVARIANTS
4824 		if (tx_queues != rx_queues)
4825 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
4826 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
4827 #endif
4828 		tx_queues = min(rx_queues, tx_queues);
4829 		rx_queues = min(rx_queues, tx_queues);
4830 	}
4831 
4832 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
4833 
4834 	vectors = rx_queues + admincnt;
4835 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
4836 		device_printf(dev,
4837 					  "Using MSIX interrupts with %d vectors\n", vectors);
4838 		scctx->isc_vectors = vectors;
4839 		scctx->isc_nrxqsets = rx_queues;
4840 		scctx->isc_ntxqsets = tx_queues;
4841 		scctx->isc_intr = IFLIB_INTR_MSIX;
4842 
4843 		return (vectors);
4844 	} else {
4845 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
4846 	}
4847 msi:
4848 	vectors = pci_msi_count(dev);
4849 	scctx->isc_nrxqsets = 1;
4850 	scctx->isc_ntxqsets = 1;
4851 	scctx->isc_vectors = vectors;
4852 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
4853 		device_printf(dev,"Using an MSI interrupt\n");
4854 		scctx->isc_intr = IFLIB_INTR_MSI;
4855 	} else {
4856 		device_printf(dev,"Using a Legacy interrupt\n");
4857 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4858 	}
4859 
4860 	return (vectors);
4861 }
4862 
4863 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
4864 
4865 static int
4866 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
4867 {
4868 	int rc;
4869 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
4870 	struct sbuf *sb;
4871 	char *ring_state = "UNKNOWN";
4872 
4873 	/* XXX needed ? */
4874 	rc = sysctl_wire_old_buffer(req, 0);
4875 	MPASS(rc == 0);
4876 	if (rc != 0)
4877 		return (rc);
4878 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
4879 	MPASS(sb != NULL);
4880 	if (sb == NULL)
4881 		return (ENOMEM);
4882 	if (state[3] <= 3)
4883 		ring_state = ring_states[state[3]];
4884 
4885 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
4886 		    state[0], state[1], state[2], ring_state);
4887 	rc = sbuf_finish(sb);
4888 	sbuf_delete(sb);
4889         return(rc);
4890 }
4891 
4892 enum iflib_ndesc_handler {
4893 	IFLIB_NTXD_HANDLER,
4894 	IFLIB_NRXD_HANDLER,
4895 };
4896 
4897 static int
4898 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
4899 {
4900 	if_ctx_t ctx = (void *)arg1;
4901 	enum iflib_ndesc_handler type = arg2;
4902 	char buf[256] = {0};
4903 	uint16_t *ndesc;
4904 	char *p, *next;
4905 	int nqs, rc, i;
4906 
4907 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
4908 
4909 	nqs = 8;
4910 	switch(type) {
4911 	case IFLIB_NTXD_HANDLER:
4912 		ndesc = ctx->ifc_sysctl_ntxds;
4913 		if (ctx->ifc_sctx)
4914 			nqs = ctx->ifc_sctx->isc_ntxqs;
4915 		break;
4916 	case IFLIB_NRXD_HANDLER:
4917 		ndesc = ctx->ifc_sysctl_nrxds;
4918 		if (ctx->ifc_sctx)
4919 			nqs = ctx->ifc_sctx->isc_nrxqs;
4920 		break;
4921 	}
4922 	if (nqs == 0)
4923 		nqs = 8;
4924 
4925 	for (i=0; i<8; i++) {
4926 		if (i >= nqs)
4927 			break;
4928 		if (i)
4929 			strcat(buf, ",");
4930 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
4931 	}
4932 
4933 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
4934 	if (rc || req->newptr == NULL)
4935 		return rc;
4936 
4937 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
4938 	    i++, p = strsep(&next, " ,")) {
4939 		ndesc[i] = strtoul(p, NULL, 10);
4940 	}
4941 
4942 	return(rc);
4943 }
4944 
4945 #define NAME_BUFLEN 32
4946 static void
4947 iflib_add_device_sysctl_pre(if_ctx_t ctx)
4948 {
4949         device_t dev = iflib_get_dev(ctx);
4950 	struct sysctl_oid_list *child, *oid_list;
4951 	struct sysctl_ctx_list *ctx_list;
4952 	struct sysctl_oid *node;
4953 
4954 	ctx_list = device_get_sysctl_ctx(dev);
4955 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
4956 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
4957 						      CTLFLAG_RD, NULL, "IFLIB fields");
4958 	oid_list = SYSCTL_CHILDREN(node);
4959 
4960 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
4961 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
4962 		       "driver version");
4963 
4964 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
4965 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
4966 			"# of txqs to use, 0 => use default #");
4967 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
4968 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
4969 			"# of rxqs to use, 0 => use default #");
4970 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
4971 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
4972                        "permit #txq != #rxq");
4973 
4974 	/* XXX change for per-queue sizes */
4975 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
4976 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
4977                        mp_ndesc_handler, "A",
4978                        "list of # of tx descriptors to use, 0 = use default #");
4979 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
4980 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
4981                        mp_ndesc_handler, "A",
4982                        "list of # of rx descriptors to use, 0 = use default #");
4983 }
4984 
4985 static void
4986 iflib_add_device_sysctl_post(if_ctx_t ctx)
4987 {
4988 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4989 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4990         device_t dev = iflib_get_dev(ctx);
4991 	struct sysctl_oid_list *child;
4992 	struct sysctl_ctx_list *ctx_list;
4993 	iflib_fl_t fl;
4994 	iflib_txq_t txq;
4995 	iflib_rxq_t rxq;
4996 	int i, j;
4997 	char namebuf[NAME_BUFLEN];
4998 	char *qfmt;
4999 	struct sysctl_oid *queue_node, *fl_node, *node;
5000 	struct sysctl_oid_list *queue_list, *fl_list;
5001 	ctx_list = device_get_sysctl_ctx(dev);
5002 
5003 	node = ctx->ifc_sysctl_node;
5004 	child = SYSCTL_CHILDREN(node);
5005 
5006 	if (scctx->isc_ntxqsets > 100)
5007 		qfmt = "txq%03d";
5008 	else if (scctx->isc_ntxqsets > 10)
5009 		qfmt = "txq%02d";
5010 	else
5011 		qfmt = "txq%d";
5012 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5013 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5014 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5015 					     CTLFLAG_RD, NULL, "Queue Name");
5016 		queue_list = SYSCTL_CHILDREN(queue_node);
5017 #if MEMORY_LOGGING
5018 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5019 				CTLFLAG_RD,
5020 				&txq->ift_dequeued, "total mbufs freed");
5021 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5022 				CTLFLAG_RD,
5023 				&txq->ift_enqueued, "total mbufs enqueued");
5024 #endif
5025 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5026 				   CTLFLAG_RD,
5027 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5028 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5029 				   CTLFLAG_RD,
5030 				   &txq->ift_pullups, "# of times m_pullup was called");
5031 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5032 				   CTLFLAG_RD,
5033 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5034 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5035 				   CTLFLAG_RD,
5036 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5037 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5038 				   CTLFLAG_RD,
5039 				   &txq->ift_map_failed, "# of times dma map failed");
5040 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5041 				   CTLFLAG_RD,
5042 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5043 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5044 				   CTLFLAG_RD,
5045 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5046 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5047 				   CTLFLAG_RD,
5048 				   &txq->ift_pidx, 1, "Producer Index");
5049 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5050 				   CTLFLAG_RD,
5051 				   &txq->ift_cidx, 1, "Consumer Index");
5052 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5053 				   CTLFLAG_RD,
5054 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5055 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5056 				   CTLFLAG_RD,
5057 				   &txq->ift_in_use, 1, "descriptors in use");
5058 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5059 				   CTLFLAG_RD,
5060 				   &txq->ift_processed, "descriptors procesed for clean");
5061 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5062 				   CTLFLAG_RD,
5063 				   &txq->ift_cleaned, "total cleaned");
5064 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5065 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state),
5066 				0, mp_ring_state_handler, "A", "soft ring state");
5067 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5068 				       CTLFLAG_RD, &txq->ift_br[0]->enqueues,
5069 				       "# of enqueues to the mp_ring for this queue");
5070 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5071 				       CTLFLAG_RD, &txq->ift_br[0]->drops,
5072 				       "# of drops in the mp_ring for this queue");
5073 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5074 				       CTLFLAG_RD, &txq->ift_br[0]->starts,
5075 				       "# of normal consumer starts in the mp_ring for this queue");
5076 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5077 				       CTLFLAG_RD, &txq->ift_br[0]->stalls,
5078 					       "# of consumer stalls in the mp_ring for this queue");
5079 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5080 			       CTLFLAG_RD, &txq->ift_br[0]->restarts,
5081 				       "# of consumer restarts in the mp_ring for this queue");
5082 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5083 				       CTLFLAG_RD, &txq->ift_br[0]->abdications,
5084 				       "# of consumer abdications in the mp_ring for this queue");
5085 	}
5086 
5087 	if (scctx->isc_nrxqsets > 100)
5088 		qfmt = "rxq%03d";
5089 	else if (scctx->isc_nrxqsets > 10)
5090 		qfmt = "rxq%02d";
5091 	else
5092 		qfmt = "rxq%d";
5093 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5094 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5095 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5096 					     CTLFLAG_RD, NULL, "Queue Name");
5097 		queue_list = SYSCTL_CHILDREN(queue_node);
5098 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5099 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5100 				       CTLFLAG_RD,
5101 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5102 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5103 				       CTLFLAG_RD,
5104 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5105 		}
5106 
5107 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5108 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5109 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5110 						     CTLFLAG_RD, NULL, "freelist Name");
5111 			fl_list = SYSCTL_CHILDREN(fl_node);
5112 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5113 				       CTLFLAG_RD,
5114 				       &fl->ifl_pidx, 1, "Producer Index");
5115 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5116 				       CTLFLAG_RD,
5117 				       &fl->ifl_cidx, 1, "Consumer Index");
5118 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5119 				       CTLFLAG_RD,
5120 				       &fl->ifl_credits, 1, "credits available");
5121 #if MEMORY_LOGGING
5122 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5123 					CTLFLAG_RD,
5124 					&fl->ifl_m_enqueued, "mbufs allocated");
5125 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5126 					CTLFLAG_RD,
5127 					&fl->ifl_m_dequeued, "mbufs freed");
5128 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5129 					CTLFLAG_RD,
5130 					&fl->ifl_cl_enqueued, "clusters allocated");
5131 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5132 					CTLFLAG_RD,
5133 					&fl->ifl_cl_dequeued, "clusters freed");
5134 #endif
5135 
5136 		}
5137 	}
5138 
5139 }
5140