xref: /freebsd/sys/net/iflib.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_acpi.h"
32 #include "opt_sched.h"
33 
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/bus.h>
37 #include <sys/eventhandler.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/module.h>
42 #include <sys/kobj.h>
43 #include <sys/rman.h>
44 #include <sys/sbuf.h>
45 #include <sys/smp.h>
46 #include <sys/socket.h>
47 #include <sys/sockio.h>
48 #include <sys/sysctl.h>
49 #include <sys/syslog.h>
50 #include <sys/taskqueue.h>
51 #include <sys/limits.h>
52 
53 #include <net/if.h>
54 #include <net/if_var.h>
55 #include <net/if_private.h>
56 #include <net/if_types.h>
57 #include <net/if_media.h>
58 #include <net/bpf.h>
59 #include <net/ethernet.h>
60 #include <net/mp_ring.h>
61 #include <net/debugnet.h>
62 #include <net/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #ifdef PCI_IOV
92 #include <dev/pci/pci_iov.h>
93 #endif
94 
95 #include <sys/bitstring.h>
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of
98  * iflib's software descriptor references
99  */
100 #define MEMORY_LOGGING 0
101 /*
102  * Enable mbuf vectors for compressing long mbuf chains
103  */
104 
105 /*
106  * NB:
107  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108  *   we prefetch needs to be determined by the time spent in m_free vis a vis
109  *   the cost of a prefetch. This will of course vary based on the workload:
110  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111  *        is quite expensive, thus suggesting very little prefetch.
112  *      - small packet forwarding which is just returning a single mbuf to
113  *        UMA will typically be very fast vis a vis the cost of a memory
114  *        access.
115  */
116 
117 /*
118  * 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 #define	IFLIB_RXEOF_MORE (1U << 0)
130 #define	IFLIB_RXEOF_EMPTY (2U << 0)
131 
132 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 static void iflib_tqg_detach(if_ctx_t ctx);
144 
145 typedef struct iflib_filter_info {
146 	driver_filter_t *ifi_filter;
147 	void *ifi_filter_arg;
148 	struct grouptask *ifi_task;
149 	void *ifi_ctx;
150 } *iflib_filter_info_t;
151 
152 struct iflib_ctx {
153 	KOBJ_FIELDS;
154 	/*
155 	 * Pointer to hardware driver's softc
156 	 */
157 	void *ifc_softc;
158 	device_t ifc_dev;
159 	if_t ifc_ifp;
160 
161 	cpuset_t ifc_cpus;
162 	if_shared_ctx_t ifc_sctx;
163 	struct if_softc_ctx ifc_softc_ctx;
164 
165 	struct sx ifc_ctx_sx;
166 	struct mtx ifc_state_mtx;
167 
168 	iflib_txq_t ifc_txqs;
169 	iflib_rxq_t ifc_rxqs;
170 	uint32_t ifc_if_flags;
171 	uint32_t ifc_flags;
172 	uint32_t ifc_max_fl_buf_size;
173 	uint32_t ifc_rx_mbuf_sz;
174 
175 	int ifc_link_state;
176 	int ifc_watchdog_events;
177 	struct cdev *ifc_led_dev;
178 	struct resource *ifc_msix_mem;
179 
180 	struct if_irq ifc_legacy_irq;
181 	struct grouptask ifc_admin_task;
182 	struct grouptask ifc_vflr_task;
183 	struct iflib_filter_info ifc_filter_info;
184 	struct ifmedia	ifc_media;
185 	struct ifmedia	*ifc_mediap;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 	uint16_t ifc_sysctl_core_offset;
194 #define	CORE_OFFSET_UNSPECIFIED	0xffff
195 	uint8_t  ifc_sysctl_separate_txrx;
196 	uint8_t  ifc_sysctl_use_logical_cores;
197 	bool	 ifc_cpus_are_physical_cores;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 #define isc_txq_select ifc_txrx.ift_txq_select
211 #define isc_txq_select_v2 ifc_txrx.ift_txq_select_v2
212 
213 	eventhandler_tag ifc_vlan_attach_event;
214 	eventhandler_tag ifc_vlan_detach_event;
215 	struct ether_addr ifc_mac;
216 };
217 
218 void *
219 iflib_get_softc(if_ctx_t ctx)
220 {
221 
222 	return (ctx->ifc_softc);
223 }
224 
225 device_t
226 iflib_get_dev(if_ctx_t ctx)
227 {
228 
229 	return (ctx->ifc_dev);
230 }
231 
232 if_t
233 iflib_get_ifp(if_ctx_t ctx)
234 {
235 
236 	return (ctx->ifc_ifp);
237 }
238 
239 struct ifmedia *
240 iflib_get_media(if_ctx_t ctx)
241 {
242 
243 	return (ctx->ifc_mediap);
244 }
245 
246 void
247 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
248 {
249 
250 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
251 }
252 
253 if_softc_ctx_t
254 iflib_get_softc_ctx(if_ctx_t ctx)
255 {
256 
257 	return (&ctx->ifc_softc_ctx);
258 }
259 
260 if_shared_ctx_t
261 iflib_get_sctx(if_ctx_t ctx)
262 {
263 
264 	return (ctx->ifc_sctx);
265 }
266 
267 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
268 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
269 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
270 
271 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
272 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
273 
274 typedef struct iflib_sw_rx_desc_array {
275 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
276 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
277 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
278 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
279 } iflib_rxsd_array_t;
280 
281 typedef struct iflib_sw_tx_desc_array {
282 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
283 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
284 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
285 } if_txsd_vec_t;
286 
287 /* magic number that should be high enough for any hardware */
288 #define IFLIB_MAX_TX_SEGS		128
289 #define IFLIB_RX_COPY_THRESH		128
290 #define IFLIB_MAX_RX_REFRESH		32
291 /* The minimum descriptors per second before we start coalescing */
292 #define IFLIB_MIN_DESC_SEC		16384
293 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
294 #define IFLIB_QUEUE_IDLE		0
295 #define IFLIB_QUEUE_HUNG		1
296 #define IFLIB_QUEUE_WORKING		2
297 /* maximum number of txqs that can share an rx interrupt */
298 #define IFLIB_MAX_TX_SHARED_INTR	4
299 
300 /* this should really scale with ring size - this is a fairly arbitrary value */
301 #define TX_BATCH_SIZE			32
302 
303 #define IFLIB_RESTART_BUDGET		8
304 
305 #define	IFC_LEGACY		0x001
306 #define	IFC_QFLUSH		0x002
307 #define	IFC_MULTISEG		0x004
308 #define	IFC_SPARE1		0x008
309 #define	IFC_SC_ALLOCATED	0x010
310 #define	IFC_INIT_DONE		0x020
311 #define	IFC_PREFETCH		0x040
312 #define	IFC_DO_RESET		0x080
313 #define	IFC_DO_WATCHDOG		0x100
314 #define	IFC_SPARE0		0x200
315 #define	IFC_SPARE2		0x400
316 #define	IFC_IN_DETACH		0x800
317 
318 #define	IFC_NETMAP_TX_IRQ	0x80000000
319 
320 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
321 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
322 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
323 
324 struct iflib_txq {
325 	qidx_t		ift_in_use;
326 	qidx_t		ift_cidx;
327 	qidx_t		ift_cidx_processed;
328 	qidx_t		ift_pidx;
329 	uint8_t		ift_gen;
330 	uint8_t		ift_br_offset;
331 	uint16_t	ift_npending;
332 	uint16_t	ift_db_pending;
333 	uint16_t	ift_rs_pending;
334 	/* implicit pad */
335 	uint8_t		ift_txd_size[8];
336 	uint64_t	ift_processed;
337 	uint64_t	ift_cleaned;
338 	uint64_t	ift_cleaned_prev;
339 #if MEMORY_LOGGING
340 	uint64_t	ift_enqueued;
341 	uint64_t	ift_dequeued;
342 #endif
343 	uint64_t	ift_no_tx_dma_setup;
344 	uint64_t	ift_no_desc_avail;
345 	uint64_t	ift_mbuf_defrag_failed;
346 	uint64_t	ift_mbuf_defrag;
347 	uint64_t	ift_map_failed;
348 	uint64_t	ift_txd_encap_efbig;
349 	uint64_t	ift_pullups;
350 	uint64_t	ift_last_timer_tick;
351 
352 	struct mtx	ift_mtx;
353 	struct mtx	ift_db_mtx;
354 
355 	/* constant values */
356 	if_ctx_t	ift_ctx;
357 	struct ifmp_ring        *ift_br;
358 	struct grouptask	ift_task;
359 	qidx_t		ift_size;
360 	uint16_t	ift_id;
361 	struct callout	ift_timer;
362 #ifdef DEV_NETMAP
363 	struct callout	ift_netmap_timer;
364 #endif /* DEV_NETMAP */
365 
366 	if_txsd_vec_t	ift_sds;
367 	uint8_t		ift_qstatus;
368 	uint8_t		ift_closed;
369 	uint8_t		ift_update_freq;
370 	struct iflib_filter_info ift_filter_info;
371 	bus_dma_tag_t	ift_buf_tag;
372 	bus_dma_tag_t	ift_tso_buf_tag;
373 	iflib_dma_info_t	ift_ifdi;
374 #define	MTX_NAME_LEN	32
375 	char                    ift_mtx_name[MTX_NAME_LEN];
376 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
377 #ifdef IFLIB_DIAGNOSTICS
378 	uint64_t ift_cpu_exec_count[256];
379 #endif
380 } __aligned(CACHE_LINE_SIZE);
381 
382 struct iflib_fl {
383 	qidx_t		ifl_cidx;
384 	qidx_t		ifl_pidx;
385 	qidx_t		ifl_credits;
386 	uint8_t		ifl_gen;
387 	uint8_t		ifl_rxd_size;
388 #if MEMORY_LOGGING
389 	uint64_t	ifl_m_enqueued;
390 	uint64_t	ifl_m_dequeued;
391 	uint64_t	ifl_cl_enqueued;
392 	uint64_t	ifl_cl_dequeued;
393 #endif
394 	/* implicit pad */
395 	bitstr_t 	*ifl_rx_bitmap;
396 	qidx_t		ifl_fragidx;
397 	/* constant */
398 	qidx_t		ifl_size;
399 	uint16_t	ifl_buf_size;
400 	uint16_t	ifl_cltype;
401 	uma_zone_t	ifl_zone;
402 	iflib_rxsd_array_t	ifl_sds;
403 	iflib_rxq_t	ifl_rxq;
404 	uint8_t		ifl_id;
405 	bus_dma_tag_t	ifl_buf_tag;
406 	iflib_dma_info_t	ifl_ifdi;
407 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
408 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
409 }  __aligned(CACHE_LINE_SIZE);
410 
411 static inline qidx_t
412 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
413 {
414 	qidx_t used;
415 
416 	if (pidx > cidx)
417 		used = pidx - cidx;
418 	else if (pidx < cidx)
419 		used = size - cidx + pidx;
420 	else if (gen == 0 && pidx == cidx)
421 		used = 0;
422 	else if (gen == 1 && pidx == cidx)
423 		used = size;
424 	else
425 		panic("bad state");
426 
427 	return (used);
428 }
429 
430 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
431 
432 #define IDXDIFF(head, tail, wrap) \
433 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
434 
435 struct iflib_rxq {
436 	if_ctx_t	ifr_ctx;
437 	iflib_fl_t	ifr_fl;
438 	uint64_t	ifr_rx_irq;
439 	struct pfil_head	*pfil;
440 	/*
441 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
442 	 * the completion queue consumer index.  Otherwise it's unused.
443 	 */
444 	qidx_t		ifr_cq_cidx;
445 	uint16_t	ifr_id;
446 	uint8_t		ifr_nfl;
447 	uint8_t		ifr_ntxqirq;
448 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
449 	uint8_t		ifr_fl_offset;
450 	struct lro_ctrl			ifr_lc;
451 	struct grouptask        ifr_task;
452 	struct callout		ifr_watchdog;
453 	struct iflib_filter_info ifr_filter_info;
454 	iflib_dma_info_t		ifr_ifdi;
455 
456 	/* dynamically allocate if any drivers need a value substantially larger than this */
457 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
458 #ifdef IFLIB_DIAGNOSTICS
459 	uint64_t ifr_cpu_exec_count[256];
460 #endif
461 }  __aligned(CACHE_LINE_SIZE);
462 
463 typedef struct if_rxsd {
464 	caddr_t *ifsd_cl;
465 	iflib_fl_t ifsd_fl;
466 } *if_rxsd_t;
467 
468 /* multiple of word size */
469 #ifdef __LP64__
470 #define PKT_INFO_SIZE	6
471 #define RXD_INFO_SIZE	5
472 #define PKT_TYPE uint64_t
473 #else
474 #define PKT_INFO_SIZE	11
475 #define RXD_INFO_SIZE	8
476 #define PKT_TYPE uint32_t
477 #endif
478 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
479 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
480 
481 typedef struct if_pkt_info_pad {
482 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
483 } *if_pkt_info_pad_t;
484 typedef struct if_rxd_info_pad {
485 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
486 } *if_rxd_info_pad_t;
487 
488 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
489 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
490 
491 static inline void
492 pkt_info_zero(if_pkt_info_t pi)
493 {
494 	if_pkt_info_pad_t pi_pad;
495 
496 	pi_pad = (if_pkt_info_pad_t)pi;
497 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
498 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
499 #ifndef __LP64__
500 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
501 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
502 #endif
503 }
504 
505 static inline void
506 rxd_info_zero(if_rxd_info_t ri)
507 {
508 	if_rxd_info_pad_t ri_pad;
509 	int i;
510 
511 	ri_pad = (if_rxd_info_pad_t)ri;
512 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
513 		ri_pad->rxd_val[i] = 0;
514 		ri_pad->rxd_val[i+1] = 0;
515 		ri_pad->rxd_val[i+2] = 0;
516 		ri_pad->rxd_val[i+3] = 0;
517 	}
518 #ifdef __LP64__
519 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
520 #endif
521 }
522 
523 /*
524  * Only allow a single packet to take up most 1/nth of the tx ring
525  */
526 #define MAX_SINGLE_PACKET_FRACTION 12
527 #define IF_BAD_DMA (bus_addr_t)-1
528 
529 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
530 
531 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
532 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
535 
536 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
537 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
538 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
539 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
540 
541 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
542 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
543 
544 /* Our boot-time initialization hook */
545 static int	iflib_module_event_handler(module_t, int, void *);
546 
547 static moduledata_t iflib_moduledata = {
548 	"iflib",
549 	iflib_module_event_handler,
550 	NULL
551 };
552 
553 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
554 MODULE_VERSION(iflib, 1);
555 
556 MODULE_DEPEND(iflib, pci, 1, 1, 1);
557 MODULE_DEPEND(iflib, ether, 1, 1, 1);
558 
559 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
560 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
561 
562 #ifndef IFLIB_DEBUG_COUNTERS
563 #ifdef INVARIANTS
564 #define IFLIB_DEBUG_COUNTERS 1
565 #else
566 #define IFLIB_DEBUG_COUNTERS 0
567 #endif /* !INVARIANTS */
568 #endif
569 
570 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
571     "iflib driver parameters");
572 
573 /*
574  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
575  */
576 static int iflib_min_tx_latency = 0;
577 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
578 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
579 static int iflib_no_tx_batch = 0;
580 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
581 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
582 static int iflib_timer_default = 1000;
583 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
584 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
585 
586 
587 #if IFLIB_DEBUG_COUNTERS
588 
589 static int iflib_tx_seen;
590 static int iflib_tx_sent;
591 static int iflib_tx_encap;
592 static int iflib_rx_allocs;
593 static int iflib_fl_refills;
594 static int iflib_fl_refills_large;
595 static int iflib_tx_frees;
596 
597 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
598 		   &iflib_tx_seen, 0, "# TX mbufs seen");
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
600 		   &iflib_tx_sent, 0, "# TX mbufs sent");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
602 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
604 		   &iflib_tx_frees, 0, "# TX frees");
605 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
606 		   &iflib_rx_allocs, 0, "# RX allocations");
607 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
608 		   &iflib_fl_refills, 0, "# refills");
609 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
610 		   &iflib_fl_refills_large, 0, "# large refills");
611 
612 static int iflib_txq_drain_flushing;
613 static int iflib_txq_drain_oactive;
614 static int iflib_txq_drain_notready;
615 
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
617 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
619 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
621 		   &iflib_txq_drain_notready, 0, "# drain notready");
622 
623 static int iflib_encap_load_mbuf_fail;
624 static int iflib_encap_pad_mbuf_fail;
625 static int iflib_encap_txq_avail_fail;
626 static int iflib_encap_txd_encap_fail;
627 
628 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
629 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
630 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
631 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
632 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
633 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
635 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
636 
637 static int iflib_task_fn_rxs;
638 static int iflib_rx_intr_enables;
639 static int iflib_fast_intrs;
640 static int iflib_rx_unavail;
641 static int iflib_rx_ctx_inactive;
642 static int iflib_rx_if_input;
643 static int iflib_rxd_flush;
644 
645 static int iflib_verbose_debug;
646 
647 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
648 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
649 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
650 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
651 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
652 		   &iflib_fast_intrs, 0, "# fast_intr calls");
653 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
654 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
656 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
658 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
660 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
661 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
662 		   &iflib_verbose_debug, 0, "enable verbose debugging");
663 
664 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
665 static void
666 iflib_debug_reset(void)
667 {
668 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
669 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
670 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
671 		iflib_txq_drain_notready =
672 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
673 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
674 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
675 		iflib_rx_unavail =
676 		iflib_rx_ctx_inactive = iflib_rx_if_input =
677 		iflib_rxd_flush = 0;
678 }
679 
680 #else
681 #define DBG_COUNTER_INC(name)
682 static void iflib_debug_reset(void) {}
683 #endif
684 
685 #define IFLIB_DEBUG 0
686 
687 static void iflib_tx_structures_free(if_ctx_t ctx);
688 static void iflib_rx_structures_free(if_ctx_t ctx);
689 static int iflib_queues_alloc(if_ctx_t ctx);
690 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
691 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
692 static int iflib_qset_structures_setup(if_ctx_t ctx);
693 static int iflib_msix_init(if_ctx_t ctx);
694 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
695 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
696 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
697 #ifdef ALTQ
698 static void iflib_altq_if_start(if_t ifp);
699 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
700 #endif
701 static int iflib_register(if_ctx_t);
702 static void iflib_deregister(if_ctx_t);
703 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
704 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
705 static void iflib_init_locked(if_ctx_t ctx);
706 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
707 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
708 static void iflib_ifmp_purge(iflib_txq_t txq);
709 static void _iflib_pre_assert(if_softc_ctx_t scctx);
710 static void iflib_stop(if_ctx_t ctx);
711 static void iflib_if_init_locked(if_ctx_t ctx);
712 static void iflib_free_intr_mem(if_ctx_t ctx);
713 #ifndef __NO_STRICT_ALIGNMENT
714 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
715 #endif
716 
717 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
718     SLIST_HEAD_INITIALIZER(cpu_offsets);
719 struct cpu_offset {
720 	SLIST_ENTRY(cpu_offset) entries;
721 	cpuset_t	set;
722 	unsigned int	refcount;
723 	uint16_t	next_cpuid;
724 };
725 static struct mtx cpu_offset_mtx;
726 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
727     MTX_DEF);
728 
729 DEBUGNET_DEFINE(iflib);
730 
731 static int
732 iflib_num_rx_descs(if_ctx_t ctx)
733 {
734 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
735 	if_shared_ctx_t sctx = ctx->ifc_sctx;
736 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
737 
738 	return scctx->isc_nrxd[first_rxq];
739 }
740 
741 static int
742 iflib_num_tx_descs(if_ctx_t ctx)
743 {
744 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
745 	if_shared_ctx_t sctx = ctx->ifc_sctx;
746 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
747 
748 	return scctx->isc_ntxd[first_txq];
749 }
750 
751 #ifdef DEV_NETMAP
752 #include <sys/selinfo.h>
753 #include <net/netmap.h>
754 #include <dev/netmap/netmap_kern.h>
755 
756 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
757 
758 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
759 static void iflib_netmap_timer(void *arg);
760 
761 /*
762  * device-specific sysctl variables:
763  *
764  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
765  *	During regular operations the CRC is stripped, but on some
766  *	hardware reception of frames not multiple of 64 is slower,
767  *	so using crcstrip=0 helps in benchmarks.
768  *
769  * iflib_rx_miss, iflib_rx_miss_bufs:
770  *	count packets that might be missed due to lost interrupts.
771  */
772 SYSCTL_DECL(_dev_netmap);
773 /*
774  * The xl driver by default strips CRCs and we do not override it.
775  */
776 
777 int iflib_crcstrip = 1;
778 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
779     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
780 
781 int iflib_rx_miss, iflib_rx_miss_bufs;
782 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
783     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
785     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
786 
787 /*
788  * Register/unregister. We are already under netmap lock.
789  * Only called on the first register or the last unregister.
790  */
791 static int
792 iflib_netmap_register(struct netmap_adapter *na, int onoff)
793 {
794 	if_t ifp = na->ifp;
795 	if_ctx_t ctx = if_getsoftc(ifp);
796 	int status;
797 
798 	CTX_LOCK(ctx);
799 	if (!CTX_IS_VF(ctx))
800 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
801 
802 	iflib_stop(ctx);
803 
804 	/*
805 	 * Enable (or disable) netmap flags, and intercept (or restore)
806 	 * ifp->if_transmit. This is done once the device has been stopped
807 	 * to prevent race conditions. Also, this must be done after
808 	 * calling netmap_disable_all_rings() and before calling
809 	 * netmap_enable_all_rings(), so that these two functions see the
810 	 * updated state of the NAF_NETMAP_ON bit.
811 	 */
812 	if (onoff) {
813 		nm_set_native_flags(na);
814 	} else {
815 		nm_clear_native_flags(na);
816 	}
817 
818 	iflib_init_locked(ctx);
819 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
820 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
821 	if (status)
822 		nm_clear_native_flags(na);
823 	CTX_UNLOCK(ctx);
824 	return (status);
825 }
826 
827 static int
828 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
829 {
830 	if_t ifp = na->ifp;
831 	if_ctx_t ctx = if_getsoftc(ifp);
832 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
833 	iflib_fl_t fl = &rxq->ifr_fl[0];
834 
835 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
836 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
837 	info->num_tx_descs = iflib_num_tx_descs(ctx);
838 	info->num_rx_descs = iflib_num_rx_descs(ctx);
839 	info->rx_buf_maxsize = fl->ifl_buf_size;
840 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
841 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
842 		info->num_rx_descs, info->rx_buf_maxsize);
843 
844 	return 0;
845 }
846 
847 static int
848 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
849 {
850 	struct netmap_adapter *na = kring->na;
851 	u_int const lim = kring->nkr_num_slots - 1;
852 	struct netmap_ring *ring = kring->ring;
853 	bus_dmamap_t *map;
854 	struct if_rxd_update iru;
855 	if_ctx_t ctx = rxq->ifr_ctx;
856 	iflib_fl_t fl = &rxq->ifr_fl[0];
857 	u_int nic_i_first, nic_i;
858 	u_int nm_i;
859 	int i, n;
860 #if IFLIB_DEBUG_COUNTERS
861 	int rf_count = 0;
862 #endif
863 
864 	/*
865 	 * This function is used both at initialization and in rxsync.
866 	 * At initialization we need to prepare (with isc_rxd_refill())
867 	 * all the netmap buffers currently owned by the kernel, in
868 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
869 	 * (except for kring->nkr_hwofs). These may be less than
870 	 * kring->nkr_num_slots if netmap_reset() was called while
871 	 * an application using the kring that still owned some
872 	 * buffers.
873 	 * At rxsync time, both indexes point to the next buffer to be
874 	 * refilled.
875 	 * In any case we publish (with isc_rxd_flush()) up to
876 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
877 	 * pointer to overrun the head/cons pointer, although this is
878 	 * not necessary for some NICs (e.g. vmx).
879 	 */
880 	if (__predict_false(init)) {
881 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
882 	} else {
883 		n = kring->rhead - kring->nr_hwcur;
884 		if (n == 0)
885 			return (0); /* Nothing to do. */
886 		if (n < 0)
887 			n += kring->nkr_num_slots;
888 	}
889 
890 	iru_init(&iru, rxq, 0 /* flid */);
891 	map = fl->ifl_sds.ifsd_map;
892 	nic_i = fl->ifl_pidx;
893 	nm_i = netmap_idx_n2k(kring, nic_i);
894 	if (__predict_false(init)) {
895 		/*
896 		 * On init/reset, nic_i must be 0, and we must
897 		 * start to refill from hwtail (see netmap_reset()).
898 		 */
899 		MPASS(nic_i == 0);
900 		MPASS(nm_i == kring->nr_hwtail);
901 	} else
902 		MPASS(nm_i == kring->nr_hwcur);
903 	DBG_COUNTER_INC(fl_refills);
904 	while (n > 0) {
905 #if IFLIB_DEBUG_COUNTERS
906 		if (++rf_count == 9)
907 			DBG_COUNTER_INC(fl_refills_large);
908 #endif
909 		nic_i_first = nic_i;
910 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
911 			struct netmap_slot *slot = &ring->slot[nm_i];
912 			uint64_t paddr;
913 			void *addr = PNMB(na, slot, &paddr);
914 
915 			MPASS(i < IFLIB_MAX_RX_REFRESH);
916 
917 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
918 			        return netmap_ring_reinit(kring);
919 
920 			fl->ifl_bus_addrs[i] = paddr +
921 			    nm_get_offset(kring, slot);
922 			fl->ifl_rxd_idxs[i] = nic_i;
923 
924 			if (__predict_false(init)) {
925 				netmap_load_map(na, fl->ifl_buf_tag,
926 				    map[nic_i], addr);
927 			} else if (slot->flags & NS_BUF_CHANGED) {
928 				/* buffer has changed, reload map */
929 				netmap_reload_map(na, fl->ifl_buf_tag,
930 				    map[nic_i], addr);
931 			}
932 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
933 			    BUS_DMASYNC_PREREAD);
934 			slot->flags &= ~NS_BUF_CHANGED;
935 
936 			nm_i = nm_next(nm_i, lim);
937 			nic_i = nm_next(nic_i, lim);
938 		}
939 
940 		iru.iru_pidx = nic_i_first;
941 		iru.iru_count = i;
942 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
943 	}
944 	fl->ifl_pidx = nic_i;
945 	/*
946 	 * At the end of the loop we must have refilled everything
947 	 * we could possibly refill.
948 	 */
949 	MPASS(nm_i == kring->rhead);
950 	kring->nr_hwcur = nm_i;
951 
952 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
953 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
954 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
955 	    nm_prev(nic_i, lim));
956 	DBG_COUNTER_INC(rxd_flush);
957 
958 	return (0);
959 }
960 
961 #define NETMAP_TX_TIMER_US	90
962 
963 /*
964  * Reconcile kernel and user view of the transmit ring.
965  *
966  * All information is in the kring.
967  * Userspace wants to send packets up to the one before kring->rhead,
968  * kernel knows kring->nr_hwcur is the first unsent packet.
969  *
970  * Here we push packets out (as many as possible), and possibly
971  * reclaim buffers from previously completed transmission.
972  *
973  * The caller (netmap) guarantees that there is only one instance
974  * running at any time. Any interference with other driver
975  * methods should be handled by the individual drivers.
976  */
977 static int
978 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
979 {
980 	struct netmap_adapter *na = kring->na;
981 	if_t ifp = na->ifp;
982 	struct netmap_ring *ring = kring->ring;
983 	u_int nm_i;	/* index into the netmap kring */
984 	u_int nic_i;	/* index into the NIC ring */
985 	u_int const lim = kring->nkr_num_slots - 1;
986 	u_int const head = kring->rhead;
987 	struct if_pkt_info pi;
988 	int tx_pkts = 0, tx_bytes = 0;
989 
990 	/*
991 	 * interrupts on every tx packet are expensive so request
992 	 * them every half ring, or where NS_REPORT is set
993 	 */
994 	u_int report_frequency = kring->nkr_num_slots >> 1;
995 	/* device-specific */
996 	if_ctx_t ctx = if_getsoftc(ifp);
997 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
998 
999 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1000 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1001 
1002 	/*
1003 	 * First part: process new packets to send.
1004 	 * nm_i is the current index in the netmap kring,
1005 	 * nic_i is the corresponding index in the NIC ring.
1006 	 *
1007 	 * If we have packets to send (nm_i != head)
1008 	 * iterate over the netmap ring, fetch length and update
1009 	 * the corresponding slot in the NIC ring. Some drivers also
1010 	 * need to update the buffer's physical address in the NIC slot
1011 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1012 	 *
1013 	 * The netmap_reload_map() calls is especially expensive,
1014 	 * even when (as in this case) the tag is 0, so do only
1015 	 * when the buffer has actually changed.
1016 	 *
1017 	 * If possible do not set the report/intr bit on all slots,
1018 	 * but only a few times per ring or when NS_REPORT is set.
1019 	 *
1020 	 * Finally, on 10G and faster drivers, it might be useful
1021 	 * to prefetch the next slot and txr entry.
1022 	 */
1023 
1024 	nm_i = kring->nr_hwcur;
1025 	if (nm_i != head) {	/* we have new packets to send */
1026 		uint32_t pkt_len = 0, seg_idx = 0;
1027 		int nic_i_start = -1, flags = 0;
1028 		pkt_info_zero(&pi);
1029 		pi.ipi_segs = txq->ift_segs;
1030 		pi.ipi_qsidx = kring->ring_id;
1031 		nic_i = netmap_idx_k2n(kring, nm_i);
1032 
1033 		__builtin_prefetch(&ring->slot[nm_i]);
1034 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1035 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1036 
1037 		while (nm_i != head) {
1038 			struct netmap_slot *slot = &ring->slot[nm_i];
1039 			uint64_t offset = nm_get_offset(kring, slot);
1040 			u_int len = slot->len;
1041 			uint64_t paddr;
1042 			void *addr = PNMB(na, slot, &paddr);
1043 
1044 			flags |= (slot->flags & NS_REPORT ||
1045 				nic_i == 0 || nic_i == report_frequency) ?
1046 				IPI_TX_INTR : 0;
1047 
1048 			/*
1049 			 * If this is the first packet fragment, save the
1050 			 * index of the first NIC slot for later.
1051 			 */
1052 			if (nic_i_start < 0)
1053 				nic_i_start = nic_i;
1054 
1055 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1056 			pi.ipi_segs[seg_idx].ds_len = len;
1057 			if (len) {
1058 				pkt_len += len;
1059 				seg_idx++;
1060 			}
1061 
1062 			if (!(slot->flags & NS_MOREFRAG)) {
1063 				pi.ipi_len = pkt_len;
1064 				pi.ipi_nsegs = seg_idx;
1065 				pi.ipi_pidx = nic_i_start;
1066 				pi.ipi_ndescs = 0;
1067 				pi.ipi_flags = flags;
1068 
1069 				/* Prepare the NIC TX ring. */
1070 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1071 				DBG_COUNTER_INC(tx_encap);
1072 
1073 				/* Update transmit counters */
1074 				tx_bytes += pi.ipi_len;
1075 				tx_pkts++;
1076 
1077 				/* Reinit per-packet info for the next one. */
1078 				flags = seg_idx = pkt_len = 0;
1079 				nic_i_start = -1;
1080 			}
1081 
1082 			/* prefetch for next round */
1083 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1084 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1085 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1086 
1087 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1088 
1089 			if (slot->flags & NS_BUF_CHANGED) {
1090 				/* buffer has changed, reload map */
1091 				netmap_reload_map(na, txq->ift_buf_tag,
1092 				    txq->ift_sds.ifsd_map[nic_i], addr);
1093 			}
1094 			/* make sure changes to the buffer are synced */
1095 			bus_dmamap_sync(txq->ift_buf_tag,
1096 			    txq->ift_sds.ifsd_map[nic_i],
1097 			    BUS_DMASYNC_PREWRITE);
1098 
1099 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1100 			nm_i = nm_next(nm_i, lim);
1101 			nic_i = nm_next(nic_i, lim);
1102 		}
1103 		kring->nr_hwcur = nm_i;
1104 
1105 		/* synchronize the NIC ring */
1106 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1107 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1108 
1109 		/* (re)start the tx unit up to slot nic_i (excluded) */
1110 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1111 	}
1112 
1113 	/*
1114 	 * Second part: reclaim buffers for completed transmissions.
1115 	 *
1116 	 * If there are unclaimed buffers, attempt to reclaim them.
1117 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1118 	 * trigger a per-tx-queue timer to try again later.
1119 	 */
1120 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1121 		if (iflib_tx_credits_update(ctx, txq)) {
1122 			/* some tx completed, increment avail */
1123 			nic_i = txq->ift_cidx_processed;
1124 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1125 		}
1126 	}
1127 
1128 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1129 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1130 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1131 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1132 			    iflib_netmap_timer, txq,
1133 			    txq->ift_netmap_timer.c_cpu, 0);
1134 		}
1135 
1136 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1137 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1138 
1139 	return (0);
1140 }
1141 
1142 /*
1143  * Reconcile kernel and user view of the receive ring.
1144  * Same as for the txsync, this routine must be efficient.
1145  * The caller guarantees a single invocations, but races against
1146  * the rest of the driver should be handled here.
1147  *
1148  * On call, kring->rhead is the first packet that userspace wants
1149  * to keep, and kring->rcur is the wakeup point.
1150  * The kernel has previously reported packets up to kring->rtail.
1151  *
1152  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1153  * of whether or not we received an interrupt.
1154  */
1155 static int
1156 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1157 {
1158 	struct netmap_adapter *na = kring->na;
1159 	struct netmap_ring *ring = kring->ring;
1160 	if_t ifp = na->ifp;
1161 	uint32_t nm_i;	/* index into the netmap ring */
1162 	uint32_t nic_i;	/* index into the NIC ring */
1163 	u_int n;
1164 	u_int const lim = kring->nkr_num_slots - 1;
1165 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1166 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1167 
1168 	if_ctx_t ctx = if_getsoftc(ifp);
1169 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1170 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1171 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1172 	iflib_fl_t fl = &rxq->ifr_fl[0];
1173 	struct if_rxd_info ri;
1174 	qidx_t *cidxp;
1175 
1176 	/*
1177 	 * netmap only uses free list 0, to avoid out of order consumption
1178 	 * of receive buffers
1179 	 */
1180 
1181 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1182 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1183 
1184 	/*
1185 	 * First part: import newly received packets.
1186 	 *
1187 	 * nm_i is the index of the next free slot in the netmap ring,
1188 	 * nic_i is the index of the next received packet in the NIC ring
1189 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1190 	 * differ in case if_init() has been called while
1191 	 * in netmap mode. For the receive ring we have
1192 	 *
1193 	 *	nic_i = fl->ifl_cidx;
1194 	 *	nm_i = kring->nr_hwtail (previous)
1195 	 * and
1196 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1197 	 *
1198 	 * fl->ifl_cidx is set to 0 on a ring reinit
1199 	 */
1200 	if (netmap_no_pendintr || force_update) {
1201 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1202 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1203 		int crclen = iflib_crcstrip ? 0 : 4;
1204 		int error, avail;
1205 
1206 		/*
1207 		 * For the free list consumer index, we use the same
1208 		 * logic as in iflib_rxeof().
1209 		 */
1210 		if (have_rxcq)
1211 			cidxp = &rxq->ifr_cq_cidx;
1212 		else
1213 			cidxp = &fl->ifl_cidx;
1214 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1215 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1216 
1217 		nic_i = fl->ifl_cidx;
1218 		nm_i = netmap_idx_n2k(kring, nic_i);
1219 		MPASS(nm_i == kring->nr_hwtail);
1220 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1221 			rxd_info_zero(&ri);
1222 			ri.iri_frags = rxq->ifr_frags;
1223 			ri.iri_qsidx = kring->ring_id;
1224 			ri.iri_ifp = ctx->ifc_ifp;
1225 			ri.iri_cidx = *cidxp;
1226 
1227 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1228 			for (i = 0; i < ri.iri_nfrags; i++) {
1229 				if (error) {
1230 					ring->slot[nm_i].len = 0;
1231 					ring->slot[nm_i].flags = 0;
1232 				} else {
1233 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1234 					if (i == (ri.iri_nfrags - 1)) {
1235 						ring->slot[nm_i].len -= crclen;
1236 						ring->slot[nm_i].flags = 0;
1237 
1238 						/* Update receive counters */
1239 						rx_bytes += ri.iri_len;
1240 						rx_pkts++;
1241 					} else
1242 						ring->slot[nm_i].flags = NS_MOREFRAG;
1243 				}
1244 
1245 				bus_dmamap_sync(fl->ifl_buf_tag,
1246 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1247 				nm_i = nm_next(nm_i, lim);
1248 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1249 			}
1250 
1251 			if (have_rxcq) {
1252 				*cidxp = ri.iri_cidx;
1253 				while (*cidxp >= scctx->isc_nrxd[0])
1254 					*cidxp -= scctx->isc_nrxd[0];
1255 			}
1256 
1257 		}
1258 		if (n) { /* update the state variables */
1259 			if (netmap_no_pendintr && !force_update) {
1260 				/* diagnostics */
1261 				iflib_rx_miss ++;
1262 				iflib_rx_miss_bufs += n;
1263 			}
1264 			kring->nr_hwtail = nm_i;
1265 		}
1266 		kring->nr_kflags &= ~NKR_PENDINTR;
1267 	}
1268 	/*
1269 	 * Second part: skip past packets that userspace has released.
1270 	 * (kring->nr_hwcur to head excluded),
1271 	 * and make the buffers available for reception.
1272 	 * As usual nm_i is the index in the netmap ring,
1273 	 * nic_i is the index in the NIC ring, and
1274 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1275 	 */
1276 	netmap_fl_refill(rxq, kring, false);
1277 
1278 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1279 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1280 
1281 	return (0);
1282 }
1283 
1284 static void
1285 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1286 {
1287 	if_ctx_t ctx = if_getsoftc(na->ifp);
1288 
1289 	CTX_LOCK(ctx);
1290 	if (onoff) {
1291 		IFDI_INTR_ENABLE(ctx);
1292 	} else {
1293 		IFDI_INTR_DISABLE(ctx);
1294 	}
1295 	CTX_UNLOCK(ctx);
1296 }
1297 
1298 static int
1299 iflib_netmap_attach(if_ctx_t ctx)
1300 {
1301 	struct netmap_adapter na;
1302 
1303 	bzero(&na, sizeof(na));
1304 
1305 	na.ifp = ctx->ifc_ifp;
1306 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1307 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1308 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1309 
1310 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1311 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1312 	na.nm_txsync = iflib_netmap_txsync;
1313 	na.nm_rxsync = iflib_netmap_rxsync;
1314 	na.nm_register = iflib_netmap_register;
1315 	na.nm_intr = iflib_netmap_intr;
1316 	na.nm_config = iflib_netmap_config;
1317 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1318 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1319 	return (netmap_attach(&na));
1320 }
1321 
1322 static int
1323 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1324 {
1325 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1326 	struct netmap_slot *slot;
1327 
1328 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1329 	if (slot == NULL)
1330 		return (0);
1331 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1332 		/*
1333 		 * In netmap mode, set the map for the packet buffer.
1334 		 * NOTE: Some drivers (not this one) also need to set
1335 		 * the physical buffer address in the NIC ring.
1336 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1337 		 * netmap slot index, si
1338 		 */
1339 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1340 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1341 		    NMB(na, slot + si));
1342 	}
1343 	return (1);
1344 }
1345 
1346 static int
1347 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1348 {
1349 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1350 	struct netmap_kring *kring;
1351 	struct netmap_slot *slot;
1352 
1353 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1354 	if (slot == NULL)
1355 		return (0);
1356 	kring = na->rx_rings[rxq->ifr_id];
1357 	netmap_fl_refill(rxq, kring, true);
1358 	return (1);
1359 }
1360 
1361 static void
1362 iflib_netmap_timer(void *arg)
1363 {
1364 	iflib_txq_t txq = arg;
1365 	if_ctx_t ctx = txq->ift_ctx;
1366 
1367 	/*
1368 	 * Wake up the netmap application, to give it a chance to
1369 	 * call txsync and reclaim more completed TX buffers.
1370 	 */
1371 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1372 }
1373 
1374 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1375 
1376 #else
1377 #define iflib_netmap_txq_init(ctx, txq) (0)
1378 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1379 #define iflib_netmap_detach(ifp)
1380 #define netmap_enable_all_rings(ifp)
1381 #define netmap_disable_all_rings(ifp)
1382 
1383 #define iflib_netmap_attach(ctx) (0)
1384 #define netmap_rx_irq(ifp, qid, budget) (0)
1385 #endif
1386 
1387 #if defined(__i386__) || defined(__amd64__)
1388 static __inline void
1389 prefetch(void *x)
1390 {
1391 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1392 }
1393 
1394 static __inline void
1395 prefetch2cachelines(void *x)
1396 {
1397 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1398 #if (CACHE_LINE_SIZE < 128)
1399 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1400 #endif
1401 }
1402 #else
1403 static __inline void
1404 prefetch(void *x)
1405 {
1406 }
1407 
1408 static __inline void
1409 prefetch2cachelines(void *x)
1410 {
1411 }
1412 #endif
1413 
1414 static void
1415 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1416 {
1417 	iflib_fl_t fl;
1418 
1419 	fl = &rxq->ifr_fl[flid];
1420 	iru->iru_paddrs = fl->ifl_bus_addrs;
1421 	iru->iru_idxs = fl->ifl_rxd_idxs;
1422 	iru->iru_qsidx = rxq->ifr_id;
1423 	iru->iru_buf_size = fl->ifl_buf_size;
1424 	iru->iru_flidx = fl->ifl_id;
1425 }
1426 
1427 static void
1428 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1429 {
1430 	if (err)
1431 		return;
1432 	*(bus_addr_t *) arg = segs[0].ds_addr;
1433 }
1434 
1435 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1436 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1437 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1438 
1439 int
1440 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1441 {
1442 	int err;
1443 	device_t dev = ctx->ifc_dev;
1444 	bus_addr_t lowaddr;
1445 
1446 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1447 
1448 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1449 				align, 0,		/* alignment, bounds */
1450 				lowaddr,		/* lowaddr */
1451 				BUS_SPACE_MAXADDR,	/* highaddr */
1452 				NULL, NULL,		/* filter, filterarg */
1453 				size,			/* maxsize */
1454 				1,			/* nsegments */
1455 				size,			/* maxsegsize */
1456 				BUS_DMA_ALLOCNOW,	/* flags */
1457 				NULL,			/* lockfunc */
1458 				NULL,			/* lockarg */
1459 				&dma->idi_tag);
1460 	if (err) {
1461 		device_printf(dev,
1462 		    "%s: bus_dma_tag_create failed: %d\n",
1463 		    __func__, err);
1464 		goto fail_0;
1465 	}
1466 
1467 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1468 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1469 	if (err) {
1470 		device_printf(dev,
1471 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1472 		    __func__, (uintmax_t)size, err);
1473 		goto fail_1;
1474 	}
1475 
1476 	dma->idi_paddr = IF_BAD_DMA;
1477 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1478 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1479 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1480 		device_printf(dev,
1481 		    "%s: bus_dmamap_load failed: %d\n",
1482 		    __func__, err);
1483 		goto fail_2;
1484 	}
1485 
1486 	dma->idi_size = size;
1487 	return (0);
1488 
1489 fail_2:
1490 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1491 fail_1:
1492 	bus_dma_tag_destroy(dma->idi_tag);
1493 fail_0:
1494 	dma->idi_tag = NULL;
1495 
1496 	return (err);
1497 }
1498 
1499 int
1500 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1501 {
1502 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1503 
1504 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1505 
1506 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1507 }
1508 
1509 int
1510 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1511 {
1512 	int i, err;
1513 	iflib_dma_info_t *dmaiter;
1514 
1515 	dmaiter = dmalist;
1516 	for (i = 0; i < count; i++, dmaiter++) {
1517 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1518 			break;
1519 	}
1520 	if (err)
1521 		iflib_dma_free_multi(dmalist, i);
1522 	return (err);
1523 }
1524 
1525 void
1526 iflib_dma_free(iflib_dma_info_t dma)
1527 {
1528 	if (dma->idi_tag == NULL)
1529 		return;
1530 	if (dma->idi_paddr != IF_BAD_DMA) {
1531 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1532 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1533 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1534 		dma->idi_paddr = IF_BAD_DMA;
1535 	}
1536 	if (dma->idi_vaddr != NULL) {
1537 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1538 		dma->idi_vaddr = NULL;
1539 	}
1540 	bus_dma_tag_destroy(dma->idi_tag);
1541 	dma->idi_tag = NULL;
1542 }
1543 
1544 void
1545 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1546 {
1547 	int i;
1548 	iflib_dma_info_t *dmaiter = dmalist;
1549 
1550 	for (i = 0; i < count; i++, dmaiter++)
1551 		iflib_dma_free(*dmaiter);
1552 }
1553 
1554 static int
1555 iflib_fast_intr(void *arg)
1556 {
1557 	iflib_filter_info_t info = arg;
1558 	struct grouptask *gtask = info->ifi_task;
1559 	int result;
1560 
1561 	DBG_COUNTER_INC(fast_intrs);
1562 	if (info->ifi_filter != NULL) {
1563 		result = info->ifi_filter(info->ifi_filter_arg);
1564 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1565 			return (result);
1566 	}
1567 
1568 	GROUPTASK_ENQUEUE(gtask);
1569 	return (FILTER_HANDLED);
1570 }
1571 
1572 static int
1573 iflib_fast_intr_rxtx(void *arg)
1574 {
1575 	iflib_filter_info_t info = arg;
1576 	struct grouptask *gtask = info->ifi_task;
1577 	if_ctx_t ctx;
1578 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1579 	iflib_txq_t txq;
1580 	void *sc;
1581 	int i, cidx, result;
1582 	qidx_t txqid;
1583 	bool intr_enable, intr_legacy;
1584 
1585 	DBG_COUNTER_INC(fast_intrs);
1586 	if (info->ifi_filter != NULL) {
1587 		result = info->ifi_filter(info->ifi_filter_arg);
1588 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1589 			return (result);
1590 	}
1591 
1592 	ctx = rxq->ifr_ctx;
1593 	sc = ctx->ifc_softc;
1594 	intr_enable = false;
1595 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1596 	MPASS(rxq->ifr_ntxqirq);
1597 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1598 		txqid = rxq->ifr_txqid[i];
1599 		txq = &ctx->ifc_txqs[txqid];
1600 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1601 		    BUS_DMASYNC_POSTREAD);
1602 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1603 			if (intr_legacy)
1604 				intr_enable = true;
1605 			else
1606 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1607 			continue;
1608 		}
1609 		GROUPTASK_ENQUEUE(&txq->ift_task);
1610 	}
1611 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1612 		cidx = rxq->ifr_cq_cidx;
1613 	else
1614 		cidx = rxq->ifr_fl[0].ifl_cidx;
1615 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1616 		GROUPTASK_ENQUEUE(gtask);
1617 	else {
1618 		if (intr_legacy)
1619 			intr_enable = true;
1620 		else
1621 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1622 		DBG_COUNTER_INC(rx_intr_enables);
1623 	}
1624 	if (intr_enable)
1625 		IFDI_INTR_ENABLE(ctx);
1626 	return (FILTER_HANDLED);
1627 }
1628 
1629 static int
1630 iflib_fast_intr_ctx(void *arg)
1631 {
1632 	iflib_filter_info_t info = arg;
1633 	struct grouptask *gtask = info->ifi_task;
1634 	int result;
1635 
1636 	DBG_COUNTER_INC(fast_intrs);
1637 	if (info->ifi_filter != NULL) {
1638 		result = info->ifi_filter(info->ifi_filter_arg);
1639 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1640 			return (result);
1641 	}
1642 
1643 	if (gtask->gt_taskqueue != NULL)
1644 		GROUPTASK_ENQUEUE(gtask);
1645 	return (FILTER_HANDLED);
1646 }
1647 
1648 static int
1649 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1650 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1651 		 const char *name)
1652 {
1653 	struct resource *res;
1654 	void *tag = NULL;
1655 	device_t dev = ctx->ifc_dev;
1656 	int flags, i, rc;
1657 
1658 	flags = RF_ACTIVE;
1659 	if (ctx->ifc_flags & IFC_LEGACY)
1660 		flags |= RF_SHAREABLE;
1661 	MPASS(rid < 512);
1662 	i = rid;
1663 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1664 	if (res == NULL) {
1665 		device_printf(dev,
1666 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1667 		return (ENOMEM);
1668 	}
1669 	irq->ii_res = res;
1670 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1671 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1672 						filter, handler, arg, &tag);
1673 	if (rc != 0) {
1674 		device_printf(dev,
1675 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1676 					  rid, name ? name : "unknown", rc);
1677 		return (rc);
1678 	} else if (name)
1679 		bus_describe_intr(dev, res, tag, "%s", name);
1680 
1681 	irq->ii_tag = tag;
1682 	return (0);
1683 }
1684 
1685 /*********************************************************************
1686  *
1687  *  Allocate DMA resources for TX buffers as well as memory for the TX
1688  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1689  *  iflib_sw_tx_desc_array structure, storing all the information that
1690  *  is needed to transmit a packet on the wire.  This is called only
1691  *  once at attach, setup is done every reset.
1692  *
1693  **********************************************************************/
1694 static int
1695 iflib_txsd_alloc(iflib_txq_t txq)
1696 {
1697 	if_ctx_t ctx = txq->ift_ctx;
1698 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1699 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1700 	device_t dev = ctx->ifc_dev;
1701 	bus_size_t tsomaxsize;
1702 	bus_addr_t lowaddr;
1703 	int err, nsegments, ntsosegments;
1704 	bool tso;
1705 
1706 	nsegments = scctx->isc_tx_nsegments;
1707 	ntsosegments = scctx->isc_tx_tso_segments_max;
1708 	tsomaxsize = scctx->isc_tx_tso_size_max;
1709 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1710 		tsomaxsize += sizeof(struct ether_vlan_header);
1711 	MPASS(scctx->isc_ntxd[0] > 0);
1712 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1713 	MPASS(nsegments > 0);
1714 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1715 		MPASS(ntsosegments > 0);
1716 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1717 	}
1718 
1719 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1720 
1721 	/*
1722 	 * Set up DMA tags for TX buffers.
1723 	 */
1724 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1725 			       1, 0,			/* alignment, bounds */
1726 			       lowaddr,			/* lowaddr */
1727 			       BUS_SPACE_MAXADDR,	/* highaddr */
1728 			       NULL, NULL,		/* filter, filterarg */
1729 			       sctx->isc_tx_maxsize,		/* maxsize */
1730 			       nsegments,	/* nsegments */
1731 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1732 			       0,			/* flags */
1733 			       NULL,			/* lockfunc */
1734 			       NULL,			/* lockfuncarg */
1735 			       &txq->ift_buf_tag))) {
1736 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1737 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1738 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1739 		goto fail;
1740 	}
1741 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1742 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1743 			       1, 0,			/* alignment, bounds */
1744 			       lowaddr,			/* lowaddr */
1745 			       BUS_SPACE_MAXADDR,	/* highaddr */
1746 			       NULL, NULL,		/* filter, filterarg */
1747 			       tsomaxsize,		/* maxsize */
1748 			       ntsosegments,	/* nsegments */
1749 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1750 			       0,			/* flags */
1751 			       NULL,			/* lockfunc */
1752 			       NULL,			/* lockfuncarg */
1753 			       &txq->ift_tso_buf_tag))) {
1754 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1755 		    err);
1756 		goto fail;
1757 	}
1758 
1759 	/* Allocate memory for the TX mbuf map. */
1760 	if (!(txq->ift_sds.ifsd_m =
1761 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1762 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1763 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1764 		err = ENOMEM;
1765 		goto fail;
1766 	}
1767 
1768 	/*
1769 	 * Create the DMA maps for TX buffers.
1770 	 */
1771 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1772 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1773 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1774 		device_printf(dev,
1775 		    "Unable to allocate TX buffer DMA map memory\n");
1776 		err = ENOMEM;
1777 		goto fail;
1778 	}
1779 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1780 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1781 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1782 		device_printf(dev,
1783 		    "Unable to allocate TSO TX buffer map memory\n");
1784 		err = ENOMEM;
1785 		goto fail;
1786 	}
1787 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1788 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1789 		    &txq->ift_sds.ifsd_map[i]);
1790 		if (err != 0) {
1791 			device_printf(dev, "Unable to create TX DMA map\n");
1792 			goto fail;
1793 		}
1794 		if (!tso)
1795 			continue;
1796 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1797 		    &txq->ift_sds.ifsd_tso_map[i]);
1798 		if (err != 0) {
1799 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1800 			goto fail;
1801 		}
1802 	}
1803 	return (0);
1804 fail:
1805 	/* We free all, it handles case where we are in the middle */
1806 	iflib_tx_structures_free(ctx);
1807 	return (err);
1808 }
1809 
1810 static void
1811 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1812 {
1813 	bus_dmamap_t map;
1814 
1815 	if (txq->ift_sds.ifsd_map != NULL) {
1816 		map = txq->ift_sds.ifsd_map[i];
1817 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1818 		bus_dmamap_unload(txq->ift_buf_tag, map);
1819 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1820 		txq->ift_sds.ifsd_map[i] = NULL;
1821 	}
1822 
1823 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1824 		map = txq->ift_sds.ifsd_tso_map[i];
1825 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1826 		    BUS_DMASYNC_POSTWRITE);
1827 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1828 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1829 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1830 	}
1831 }
1832 
1833 static void
1834 iflib_txq_destroy(iflib_txq_t txq)
1835 {
1836 	if_ctx_t ctx = txq->ift_ctx;
1837 
1838 	for (int i = 0; i < txq->ift_size; i++)
1839 		iflib_txsd_destroy(ctx, txq, i);
1840 
1841 	if (txq->ift_br != NULL) {
1842 		ifmp_ring_free(txq->ift_br);
1843 		txq->ift_br = NULL;
1844 	}
1845 
1846 	mtx_destroy(&txq->ift_mtx);
1847 
1848 	if (txq->ift_sds.ifsd_map != NULL) {
1849 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1850 		txq->ift_sds.ifsd_map = NULL;
1851 	}
1852 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1853 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1854 		txq->ift_sds.ifsd_tso_map = NULL;
1855 	}
1856 	if (txq->ift_sds.ifsd_m != NULL) {
1857 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1858 		txq->ift_sds.ifsd_m = NULL;
1859 	}
1860 	if (txq->ift_buf_tag != NULL) {
1861 		bus_dma_tag_destroy(txq->ift_buf_tag);
1862 		txq->ift_buf_tag = NULL;
1863 	}
1864 	if (txq->ift_tso_buf_tag != NULL) {
1865 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1866 		txq->ift_tso_buf_tag = NULL;
1867 	}
1868 	if (txq->ift_ifdi != NULL) {
1869 		free(txq->ift_ifdi, M_IFLIB);
1870 	}
1871 }
1872 
1873 static void
1874 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1875 {
1876 	struct mbuf **mp;
1877 
1878 	mp = &txq->ift_sds.ifsd_m[i];
1879 	if (*mp == NULL)
1880 		return;
1881 
1882 	if (txq->ift_sds.ifsd_map != NULL) {
1883 		bus_dmamap_sync(txq->ift_buf_tag,
1884 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1885 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1886 	}
1887 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1888 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1889 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1890 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1891 		    txq->ift_sds.ifsd_tso_map[i]);
1892 	}
1893 	m_freem(*mp);
1894 	DBG_COUNTER_INC(tx_frees);
1895 	*mp = NULL;
1896 }
1897 
1898 static int
1899 iflib_txq_setup(iflib_txq_t txq)
1900 {
1901 	if_ctx_t ctx = txq->ift_ctx;
1902 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1903 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1904 	iflib_dma_info_t di;
1905 	int i;
1906 
1907 	/* Set number of descriptors available */
1908 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1909 	/* XXX make configurable */
1910 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1911 
1912 	/* Reset indices */
1913 	txq->ift_cidx_processed = 0;
1914 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1915 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1916 
1917 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1918 		bzero((void *)di->idi_vaddr, di->idi_size);
1919 
1920 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1921 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1922 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1923 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1924 	return (0);
1925 }
1926 
1927 /*********************************************************************
1928  *
1929  *  Allocate DMA resources for RX buffers as well as memory for the RX
1930  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1931  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1932  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1933  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1934  *  packet, the maximum number of entries we'll need is equal to the
1935  *  number of hardware receive descriptors that we've allocated.
1936  *
1937  **********************************************************************/
1938 static int
1939 iflib_rxsd_alloc(iflib_rxq_t rxq)
1940 {
1941 	if_ctx_t ctx = rxq->ifr_ctx;
1942 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1943 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1944 	device_t dev = ctx->ifc_dev;
1945 	iflib_fl_t fl;
1946 	bus_addr_t lowaddr;
1947 	int			err;
1948 
1949 	MPASS(scctx->isc_nrxd[0] > 0);
1950 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1951 
1952 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1953 
1954 	fl = rxq->ifr_fl;
1955 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1956 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1957 		/* Set up DMA tag for RX buffers. */
1958 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1959 					 1, 0,			/* alignment, bounds */
1960 					 lowaddr,		/* lowaddr */
1961 					 BUS_SPACE_MAXADDR,	/* highaddr */
1962 					 NULL, NULL,		/* filter, filterarg */
1963 					 sctx->isc_rx_maxsize,	/* maxsize */
1964 					 sctx->isc_rx_nsegments,	/* nsegments */
1965 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1966 					 0,			/* flags */
1967 					 NULL,			/* lockfunc */
1968 					 NULL,			/* lockarg */
1969 					 &fl->ifl_buf_tag);
1970 		if (err) {
1971 			device_printf(dev,
1972 			    "Unable to allocate RX DMA tag: %d\n", err);
1973 			goto fail;
1974 		}
1975 
1976 		/* Allocate memory for the RX mbuf map. */
1977 		if (!(fl->ifl_sds.ifsd_m =
1978 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1979 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1980 			device_printf(dev,
1981 			    "Unable to allocate RX mbuf map memory\n");
1982 			err = ENOMEM;
1983 			goto fail;
1984 		}
1985 
1986 		/* Allocate memory for the direct RX cluster pointer map. */
1987 		if (!(fl->ifl_sds.ifsd_cl =
1988 		      (caddr_t *) malloc(sizeof(caddr_t) *
1989 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1990 			device_printf(dev,
1991 			    "Unable to allocate RX cluster map memory\n");
1992 			err = ENOMEM;
1993 			goto fail;
1994 		}
1995 
1996 		/* Allocate memory for the RX cluster bus address map. */
1997 		if (!(fl->ifl_sds.ifsd_ba =
1998 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1999 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2000 			device_printf(dev,
2001 			    "Unable to allocate RX bus address map memory\n");
2002 			err = ENOMEM;
2003 			goto fail;
2004 		}
2005 
2006 		/*
2007 		 * Create the DMA maps for RX buffers.
2008 		 */
2009 		if (!(fl->ifl_sds.ifsd_map =
2010 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2011 			device_printf(dev,
2012 			    "Unable to allocate RX buffer DMA map memory\n");
2013 			err = ENOMEM;
2014 			goto fail;
2015 		}
2016 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2017 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2018 			    &fl->ifl_sds.ifsd_map[i]);
2019 			if (err != 0) {
2020 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2021 				goto fail;
2022 			}
2023 		}
2024 	}
2025 	return (0);
2026 
2027 fail:
2028 	iflib_rx_structures_free(ctx);
2029 	return (err);
2030 }
2031 
2032 /*
2033  * Internal service routines
2034  */
2035 
2036 struct rxq_refill_cb_arg {
2037 	int               error;
2038 	bus_dma_segment_t seg;
2039 	int               nseg;
2040 };
2041 
2042 static void
2043 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2044 {
2045 	struct rxq_refill_cb_arg *cb_arg = arg;
2046 
2047 	cb_arg->error = error;
2048 	cb_arg->seg = segs[0];
2049 	cb_arg->nseg = nseg;
2050 }
2051 
2052 /**
2053  * iflib_fl_refill - refill an rxq free-buffer list
2054  * @ctx: the iflib context
2055  * @fl: the free list to refill
2056  * @count: the number of new buffers to allocate
2057  *
2058  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2059  * The caller must assure that @count does not exceed the queue's capacity
2060  * minus one (since we always leave a descriptor unavailable).
2061  */
2062 static uint8_t
2063 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2064 {
2065 	struct if_rxd_update iru;
2066 	struct rxq_refill_cb_arg cb_arg;
2067 	struct mbuf *m;
2068 	caddr_t cl, *sd_cl;
2069 	struct mbuf **sd_m;
2070 	bus_dmamap_t *sd_map;
2071 	bus_addr_t bus_addr, *sd_ba;
2072 	int err, frag_idx, i, idx, n, pidx;
2073 	qidx_t credits;
2074 
2075 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2076 
2077 	sd_m = fl->ifl_sds.ifsd_m;
2078 	sd_map = fl->ifl_sds.ifsd_map;
2079 	sd_cl = fl->ifl_sds.ifsd_cl;
2080 	sd_ba = fl->ifl_sds.ifsd_ba;
2081 	pidx = fl->ifl_pidx;
2082 	idx = pidx;
2083 	frag_idx = fl->ifl_fragidx;
2084 	credits = fl->ifl_credits;
2085 
2086 	i = 0;
2087 	n = count;
2088 	MPASS(n > 0);
2089 	MPASS(credits + n <= fl->ifl_size);
2090 
2091 	if (pidx < fl->ifl_cidx)
2092 		MPASS(pidx + n <= fl->ifl_cidx);
2093 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2094 		MPASS(fl->ifl_gen == 0);
2095 	if (pidx > fl->ifl_cidx)
2096 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2097 
2098 	DBG_COUNTER_INC(fl_refills);
2099 	if (n > 8)
2100 		DBG_COUNTER_INC(fl_refills_large);
2101 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2102 	while (n-- > 0) {
2103 		/*
2104 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2105 		 * initialized after rx.
2106 		 *
2107 		 * If the cluster is still set then we know a minimum sized
2108 		 * packet was received
2109 		 */
2110 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2111 		    &frag_idx);
2112 		if (frag_idx < 0)
2113 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2114 		MPASS(frag_idx >= 0);
2115 		if ((cl = sd_cl[frag_idx]) == NULL) {
2116 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2117 			if (__predict_false(cl == NULL))
2118 				break;
2119 
2120 			cb_arg.error = 0;
2121 			MPASS(sd_map != NULL);
2122 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2123 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2124 			    BUS_DMA_NOWAIT);
2125 			if (__predict_false(err != 0 || cb_arg.error)) {
2126 				uma_zfree(fl->ifl_zone, cl);
2127 				break;
2128 			}
2129 
2130 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2131 			sd_cl[frag_idx] = cl;
2132 #if MEMORY_LOGGING
2133 			fl->ifl_cl_enqueued++;
2134 #endif
2135 		} else {
2136 			bus_addr = sd_ba[frag_idx];
2137 		}
2138 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2139 		    BUS_DMASYNC_PREREAD);
2140 
2141 		if (sd_m[frag_idx] == NULL) {
2142 			m = m_gethdr_raw(M_NOWAIT, 0);
2143 			if (__predict_false(m == NULL))
2144 				break;
2145 			sd_m[frag_idx] = m;
2146 		}
2147 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2148 #if MEMORY_LOGGING
2149 		fl->ifl_m_enqueued++;
2150 #endif
2151 
2152 		DBG_COUNTER_INC(rx_allocs);
2153 		fl->ifl_rxd_idxs[i] = frag_idx;
2154 		fl->ifl_bus_addrs[i] = bus_addr;
2155 		credits++;
2156 		i++;
2157 		MPASS(credits <= fl->ifl_size);
2158 		if (++idx == fl->ifl_size) {
2159 #ifdef INVARIANTS
2160 			fl->ifl_gen = 1;
2161 #endif
2162 			idx = 0;
2163 		}
2164 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2165 			iru.iru_pidx = pidx;
2166 			iru.iru_count = i;
2167 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2168 			fl->ifl_pidx = idx;
2169 			fl->ifl_credits = credits;
2170 			pidx = idx;
2171 			i = 0;
2172 		}
2173 	}
2174 
2175 	if (n < count - 1) {
2176 		if (i != 0) {
2177 			iru.iru_pidx = pidx;
2178 			iru.iru_count = i;
2179 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2180 			fl->ifl_pidx = idx;
2181 			fl->ifl_credits = credits;
2182 		}
2183 		DBG_COUNTER_INC(rxd_flush);
2184 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2185 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2186 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2187 		    fl->ifl_id, fl->ifl_pidx);
2188 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2189 			fl->ifl_fragidx = frag_idx + 1;
2190 			if (fl->ifl_fragidx == fl->ifl_size)
2191 				fl->ifl_fragidx = 0;
2192 		} else {
2193 			fl->ifl_fragidx = frag_idx;
2194 		}
2195 	}
2196 
2197 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2198 }
2199 
2200 static inline uint8_t
2201 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2202 {
2203 	/*
2204 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2205 	 * This is important as it confuses most NICs. For instance,
2206 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2207 	 * RDH points to the next receive descriptor to be used by the NIC,
2208 	 * and RDT for the next receive descriptor to be published by the
2209 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2210 	 * The condition RDH == RDT means no descriptors are available to
2211 	 * the NIC, and thus it would be ambiguous if it also meant that
2212 	 * all the descriptors are available to the NIC.
2213 	 */
2214 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2215 #ifdef INVARIANTS
2216 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2217 #endif
2218 
2219 	MPASS(fl->ifl_credits <= fl->ifl_size);
2220 	MPASS(reclaimable == delta);
2221 
2222 	if (reclaimable > 0)
2223 		return (iflib_fl_refill(ctx, fl, reclaimable));
2224 	return (0);
2225 }
2226 
2227 uint8_t
2228 iflib_in_detach(if_ctx_t ctx)
2229 {
2230 	bool in_detach;
2231 
2232 	STATE_LOCK(ctx);
2233 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2234 	STATE_UNLOCK(ctx);
2235 	return (in_detach);
2236 }
2237 
2238 static void
2239 iflib_fl_bufs_free(iflib_fl_t fl)
2240 {
2241 	iflib_dma_info_t idi = fl->ifl_ifdi;
2242 	bus_dmamap_t sd_map;
2243 	uint32_t i;
2244 
2245 	for (i = 0; i < fl->ifl_size; i++) {
2246 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2247 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2248 
2249 		if (*sd_cl != NULL) {
2250 			sd_map = fl->ifl_sds.ifsd_map[i];
2251 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2252 			    BUS_DMASYNC_POSTREAD);
2253 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2254 			uma_zfree(fl->ifl_zone, *sd_cl);
2255 			*sd_cl = NULL;
2256 			if (*sd_m != NULL) {
2257 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2258 				m_free_raw(*sd_m);
2259 				*sd_m = NULL;
2260 			}
2261 		} else {
2262 			MPASS(*sd_m == NULL);
2263 		}
2264 #if MEMORY_LOGGING
2265 		fl->ifl_m_dequeued++;
2266 		fl->ifl_cl_dequeued++;
2267 #endif
2268 	}
2269 #ifdef INVARIANTS
2270 	for (i = 0; i < fl->ifl_size; i++) {
2271 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2272 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2273 	}
2274 #endif
2275 	/*
2276 	 * Reset free list values
2277 	 */
2278 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2279 	bzero(idi->idi_vaddr, idi->idi_size);
2280 }
2281 
2282 /*********************************************************************
2283  *
2284  *  Initialize a free list and its buffers.
2285  *
2286  **********************************************************************/
2287 static int
2288 iflib_fl_setup(iflib_fl_t fl)
2289 {
2290 	iflib_rxq_t rxq = fl->ifl_rxq;
2291 	if_ctx_t ctx = rxq->ifr_ctx;
2292 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2293 	int qidx;
2294 
2295 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2296 	/*
2297 	** Free current RX buffer structs and their mbufs
2298 	*/
2299 	iflib_fl_bufs_free(fl);
2300 	/* Now replenish the mbufs */
2301 	MPASS(fl->ifl_credits == 0);
2302 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2303 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2304 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2305 	else
2306 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2307 	/*
2308 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2309 	 * to the selected mbuf size.
2310 	 */
2311 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2312 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2313 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2314 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2315 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2316 
2317 	/*
2318 	 * Avoid pre-allocating zillions of clusters to an idle card
2319 	 * potentially speeding up attach. In any case make sure
2320 	 * to leave a descriptor unavailable. See the comment in
2321 	 * iflib_fl_refill_all().
2322 	 */
2323 	MPASS(fl->ifl_size > 0);
2324 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2325 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2326 		return (ENOBUFS);
2327 	/*
2328 	 * handle failure
2329 	 */
2330 	MPASS(rxq != NULL);
2331 	MPASS(fl->ifl_ifdi != NULL);
2332 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2333 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2334 	return (0);
2335 }
2336 
2337 /*********************************************************************
2338  *
2339  *  Free receive ring data structures
2340  *
2341  **********************************************************************/
2342 static void
2343 iflib_rx_sds_free(iflib_rxq_t rxq)
2344 {
2345 	iflib_fl_t fl;
2346 	int i, j;
2347 
2348 	if (rxq->ifr_fl != NULL) {
2349 		for (i = 0; i < rxq->ifr_nfl; i++) {
2350 			fl = &rxq->ifr_fl[i];
2351 			if (fl->ifl_buf_tag != NULL) {
2352 				if (fl->ifl_sds.ifsd_map != NULL) {
2353 					for (j = 0; j < fl->ifl_size; j++) {
2354 						bus_dmamap_sync(
2355 						    fl->ifl_buf_tag,
2356 						    fl->ifl_sds.ifsd_map[j],
2357 						    BUS_DMASYNC_POSTREAD);
2358 						bus_dmamap_unload(
2359 						    fl->ifl_buf_tag,
2360 						    fl->ifl_sds.ifsd_map[j]);
2361 						bus_dmamap_destroy(
2362 						    fl->ifl_buf_tag,
2363 						    fl->ifl_sds.ifsd_map[j]);
2364 					}
2365 				}
2366 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2367 				fl->ifl_buf_tag = NULL;
2368 			}
2369 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2370 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2371 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2372 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2373 			free(fl->ifl_rx_bitmap, M_IFLIB);
2374 			fl->ifl_sds.ifsd_m = NULL;
2375 			fl->ifl_sds.ifsd_cl = NULL;
2376 			fl->ifl_sds.ifsd_ba = NULL;
2377 			fl->ifl_sds.ifsd_map = NULL;
2378 			fl->ifl_rx_bitmap = NULL;
2379 		}
2380 		free(rxq->ifr_fl, M_IFLIB);
2381 		rxq->ifr_fl = NULL;
2382 		free(rxq->ifr_ifdi, M_IFLIB);
2383 		rxq->ifr_ifdi = NULL;
2384 		rxq->ifr_cq_cidx = 0;
2385 	}
2386 }
2387 
2388 /*
2389  * Timer routine
2390  */
2391 static void
2392 iflib_timer(void *arg)
2393 {
2394 	iflib_txq_t txq = arg;
2395 	if_ctx_t ctx = txq->ift_ctx;
2396 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2397 	uint64_t this_tick = ticks;
2398 
2399 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2400 		return;
2401 
2402 	/*
2403 	** Check on the state of the TX queue(s), this
2404 	** can be done without the lock because its RO
2405 	** and the HUNG state will be static if set.
2406 	*/
2407 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2408 		txq->ift_last_timer_tick = this_tick;
2409 		IFDI_TIMER(ctx, txq->ift_id);
2410 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2411 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2412 		     (sctx->isc_pause_frames == 0)))
2413 			goto hung;
2414 
2415 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2416 		    ifmp_ring_is_stalled(txq->ift_br)) {
2417 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2418 			    ("queue can't be marked as hung if interface is down"));
2419 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2420 		}
2421 		txq->ift_cleaned_prev = txq->ift_cleaned;
2422 	}
2423 	/* handle any laggards */
2424 	if (txq->ift_db_pending)
2425 		GROUPTASK_ENQUEUE(&txq->ift_task);
2426 
2427 	sctx->isc_pause_frames = 0;
2428 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2429 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2430 		    txq, txq->ift_timer.c_cpu);
2431 	return;
2432 
2433  hung:
2434 	device_printf(ctx->ifc_dev,
2435 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2436 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2437 	STATE_LOCK(ctx);
2438 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2439 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2440 	iflib_admin_intr_deferred(ctx);
2441 	STATE_UNLOCK(ctx);
2442 }
2443 
2444 static uint16_t
2445 iflib_get_mbuf_size_for(unsigned int size)
2446 {
2447 
2448 	if (size <= MCLBYTES)
2449 		return (MCLBYTES);
2450 	else
2451 		return (MJUMPAGESIZE);
2452 }
2453 
2454 static void
2455 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2456 {
2457 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2458 
2459 	/*
2460 	 * XXX don't set the max_frame_size to larger
2461 	 * than the hardware can handle
2462 	 */
2463 	ctx->ifc_rx_mbuf_sz =
2464 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2465 }
2466 
2467 uint32_t
2468 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2469 {
2470 
2471 	return (ctx->ifc_rx_mbuf_sz);
2472 }
2473 
2474 static void
2475 iflib_init_locked(if_ctx_t ctx)
2476 {
2477 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2478 	if_t ifp = ctx->ifc_ifp;
2479 	iflib_fl_t fl;
2480 	iflib_txq_t txq;
2481 	iflib_rxq_t rxq;
2482 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2483 
2484 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2485 	IFDI_INTR_DISABLE(ctx);
2486 
2487 	/*
2488 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2489 	 * called without first calling iflib_stop().
2490 	 */
2491 	netmap_disable_all_rings(ifp);
2492 
2493 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2494 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2495 	/* Set hardware offload abilities */
2496 	if_clearhwassist(ifp);
2497 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2498 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2499 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2500 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2501 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2502 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2503 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2504 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2505 
2506 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2507 		CALLOUT_LOCK(txq);
2508 		callout_stop(&txq->ift_timer);
2509 #ifdef DEV_NETMAP
2510 		callout_stop(&txq->ift_netmap_timer);
2511 #endif /* DEV_NETMAP */
2512 		CALLOUT_UNLOCK(txq);
2513 		(void)iflib_netmap_txq_init(ctx, txq);
2514 	}
2515 
2516 	/*
2517 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2518 	 * that drivers can use the value when setting up the hardware receive
2519 	 * buffers.
2520 	 */
2521 	iflib_calc_rx_mbuf_sz(ctx);
2522 
2523 #ifdef INVARIANTS
2524 	i = if_getdrvflags(ifp);
2525 #endif
2526 	IFDI_INIT(ctx);
2527 	MPASS(if_getdrvflags(ifp) == i);
2528 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2529 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2530 			/* This rxq is in netmap mode. Skip normal init. */
2531 			continue;
2532 		}
2533 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2534 			if (iflib_fl_setup(fl)) {
2535 				device_printf(ctx->ifc_dev,
2536 				    "setting up free list %d failed - "
2537 				    "check cluster settings\n", j);
2538 				goto done;
2539 			}
2540 		}
2541 	}
2542 done:
2543 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2544 	IFDI_INTR_ENABLE(ctx);
2545 	txq = ctx->ifc_txqs;
2546 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2547 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2548 			txq->ift_timer.c_cpu);
2549 
2550         /* Re-enable txsync/rxsync. */
2551 	netmap_enable_all_rings(ifp);
2552 }
2553 
2554 static int
2555 iflib_media_change(if_t ifp)
2556 {
2557 	if_ctx_t ctx = if_getsoftc(ifp);
2558 	int err;
2559 
2560 	CTX_LOCK(ctx);
2561 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2562 		iflib_if_init_locked(ctx);
2563 	CTX_UNLOCK(ctx);
2564 	return (err);
2565 }
2566 
2567 static void
2568 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2569 {
2570 	if_ctx_t ctx = if_getsoftc(ifp);
2571 
2572 	CTX_LOCK(ctx);
2573 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2574 	IFDI_MEDIA_STATUS(ctx, ifmr);
2575 	CTX_UNLOCK(ctx);
2576 }
2577 
2578 void
2579 iflib_stop(if_ctx_t ctx)
2580 {
2581 	iflib_txq_t txq = ctx->ifc_txqs;
2582 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2583 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2584 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2585 	iflib_dma_info_t di;
2586 	iflib_fl_t fl;
2587 	int i, j;
2588 
2589 	/* Tell the stack that the interface is no longer active */
2590 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2591 
2592 	IFDI_INTR_DISABLE(ctx);
2593 	DELAY(1000);
2594 	IFDI_STOP(ctx);
2595 	DELAY(1000);
2596 
2597 	/*
2598 	 * Stop any pending txsync/rxsync and prevent new ones
2599 	 * form starting. Processes blocked in poll() will get
2600 	 * POLLERR.
2601 	 */
2602 	netmap_disable_all_rings(ctx->ifc_ifp);
2603 
2604 	iflib_debug_reset();
2605 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2606 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2607 		/* make sure all transmitters have completed before proceeding XXX */
2608 
2609 		CALLOUT_LOCK(txq);
2610 		callout_stop(&txq->ift_timer);
2611 #ifdef DEV_NETMAP
2612 		callout_stop(&txq->ift_netmap_timer);
2613 #endif /* DEV_NETMAP */
2614 		CALLOUT_UNLOCK(txq);
2615 
2616 		/* clean any enqueued buffers */
2617 		iflib_ifmp_purge(txq);
2618 		/* Free any existing tx buffers. */
2619 		for (j = 0; j < txq->ift_size; j++) {
2620 			iflib_txsd_free(ctx, txq, j);
2621 		}
2622 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2623 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2624 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2625 			txq->ift_cidx = txq->ift_pidx;
2626 		else
2627 			txq->ift_cidx = txq->ift_pidx = 0;
2628 
2629 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2630 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2631 		txq->ift_pullups = 0;
2632 		ifmp_ring_reset_stats(txq->ift_br);
2633 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2634 			bzero((void *)di->idi_vaddr, di->idi_size);
2635 	}
2636 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2637 		if (rxq->ifr_task.gt_taskqueue != NULL)
2638 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2639 				 &rxq->ifr_task.gt_task);
2640 
2641 		rxq->ifr_cq_cidx = 0;
2642 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2643 			bzero((void *)di->idi_vaddr, di->idi_size);
2644 		/* also resets the free lists pidx/cidx */
2645 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2646 			iflib_fl_bufs_free(fl);
2647 	}
2648 }
2649 
2650 static inline caddr_t
2651 calc_next_rxd(iflib_fl_t fl, int cidx)
2652 {
2653 	qidx_t size;
2654 	int nrxd;
2655 	caddr_t start, end, cur, next;
2656 
2657 	nrxd = fl->ifl_size;
2658 	size = fl->ifl_rxd_size;
2659 	start = fl->ifl_ifdi->idi_vaddr;
2660 
2661 	if (__predict_false(size == 0))
2662 		return (start);
2663 	cur = start + size*cidx;
2664 	end = start + size*nrxd;
2665 	next = CACHE_PTR_NEXT(cur);
2666 	return (next < end ? next : start);
2667 }
2668 
2669 static inline void
2670 prefetch_pkts(iflib_fl_t fl, int cidx)
2671 {
2672 	int nextptr;
2673 	int nrxd = fl->ifl_size;
2674 	caddr_t next_rxd;
2675 
2676 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2677 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2678 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2679 	next_rxd = calc_next_rxd(fl, cidx);
2680 	prefetch(next_rxd);
2681 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2682 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2683 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2684 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2685 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2686 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2687 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2688 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2689 }
2690 
2691 static struct mbuf *
2692 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2693     int *pf_rv, if_rxd_info_t ri)
2694 {
2695 	bus_dmamap_t map;
2696 	iflib_fl_t fl;
2697 	caddr_t payload;
2698 	struct mbuf *m;
2699 	int flid, cidx, len, next;
2700 
2701 	map = NULL;
2702 	flid = irf->irf_flid;
2703 	cidx = irf->irf_idx;
2704 	fl = &rxq->ifr_fl[flid];
2705 	sd->ifsd_fl = fl;
2706 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2707 	fl->ifl_credits--;
2708 #if MEMORY_LOGGING
2709 	fl->ifl_m_dequeued++;
2710 #endif
2711 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2712 		prefetch_pkts(fl, cidx);
2713 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2714 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2715 	map = fl->ifl_sds.ifsd_map[cidx];
2716 
2717 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2718 
2719 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2720 	    irf->irf_len != 0) {
2721 		payload  = *sd->ifsd_cl;
2722 		payload +=  ri->iri_pad;
2723 		len = ri->iri_len - ri->iri_pad;
2724 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2725 		switch (*pf_rv) {
2726 		case PFIL_DROPPED:
2727 		case PFIL_CONSUMED:
2728 			/*
2729 			 * The filter ate it.  Everything is recycled.
2730 			 */
2731 			m = NULL;
2732 			unload = 0;
2733 			break;
2734 		case PFIL_REALLOCED:
2735 			/*
2736 			 * The filter copied it.  Everything is recycled.
2737 			 * 'm' points at new mbuf.
2738 			 */
2739 			unload = 0;
2740 			break;
2741 		case PFIL_PASS:
2742 			/*
2743 			 * Filter said it was OK, so receive like
2744 			 * normal
2745 			 */
2746 			m = fl->ifl_sds.ifsd_m[cidx];
2747 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2748 			break;
2749 		default:
2750 			MPASS(0);
2751 		}
2752 	} else {
2753 		m = fl->ifl_sds.ifsd_m[cidx];
2754 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2755 		if (pf_rv != NULL)
2756 			*pf_rv = PFIL_PASS;
2757 	}
2758 
2759 	if (unload && irf->irf_len != 0)
2760 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2761 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2762 	if (__predict_false(fl->ifl_cidx == 0))
2763 		fl->ifl_gen = 0;
2764 	bit_clear(fl->ifl_rx_bitmap, cidx);
2765 	return (m);
2766 }
2767 
2768 static struct mbuf *
2769 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2770 {
2771 	struct mbuf *m, *mh, *mt;
2772 	caddr_t cl;
2773 	int  *pf_rv_ptr, flags, i, padlen;
2774 	bool consumed;
2775 
2776 	i = 0;
2777 	mh = NULL;
2778 	consumed = false;
2779 	*pf_rv = PFIL_PASS;
2780 	pf_rv_ptr = pf_rv;
2781 	do {
2782 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2783 		    pf_rv_ptr, ri);
2784 
2785 		MPASS(*sd->ifsd_cl != NULL);
2786 
2787 		/*
2788 		 * Exclude zero-length frags & frags from
2789 		 * packets the filter has consumed or dropped
2790 		 */
2791 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2792 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2793 			if (mh == NULL) {
2794 				/* everything saved here */
2795 				consumed = true;
2796 				pf_rv_ptr = NULL;
2797 				continue;
2798 			}
2799 			/* XXX we can save the cluster here, but not the mbuf */
2800 			m_init(m, M_NOWAIT, MT_DATA, 0);
2801 			m_free(m);
2802 			continue;
2803 		}
2804 		if (mh == NULL) {
2805 			flags = M_PKTHDR|M_EXT;
2806 			mh = mt = m;
2807 			padlen = ri->iri_pad;
2808 		} else {
2809 			flags = M_EXT;
2810 			mt->m_next = m;
2811 			mt = m;
2812 			/* assuming padding is only on the first fragment */
2813 			padlen = 0;
2814 		}
2815 		cl = *sd->ifsd_cl;
2816 		*sd->ifsd_cl = NULL;
2817 
2818 		/* Can these two be made one ? */
2819 		m_init(m, M_NOWAIT, MT_DATA, flags);
2820 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2821 		/*
2822 		 * These must follow m_init and m_cljset
2823 		 */
2824 		m->m_data += padlen;
2825 		ri->iri_len -= padlen;
2826 		m->m_len = ri->iri_frags[i].irf_len;
2827 	} while (++i < ri->iri_nfrags);
2828 
2829 	return (mh);
2830 }
2831 
2832 /*
2833  * Process one software descriptor
2834  */
2835 static struct mbuf *
2836 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2837 {
2838 	struct if_rxsd sd;
2839 	struct mbuf *m;
2840 	int pf_rv;
2841 
2842 	/* should I merge this back in now that the two paths are basically duplicated? */
2843 	if (ri->iri_nfrags == 1 &&
2844 	    ri->iri_frags[0].irf_len != 0 &&
2845 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2846 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2847 		    &pf_rv, ri);
2848 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2849 			return (m);
2850 		if (pf_rv == PFIL_PASS) {
2851 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2852 #ifndef __NO_STRICT_ALIGNMENT
2853 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2854 				m->m_data += 2;
2855 #endif
2856 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2857 			m->m_len = ri->iri_frags[0].irf_len;
2858 			m->m_data += ri->iri_pad;
2859 			ri->iri_len -= ri->iri_pad;
2860 		}
2861 	} else {
2862 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2863 		if (m == NULL)
2864 			return (NULL);
2865 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2866 			return (m);
2867 	}
2868 	m->m_pkthdr.len = ri->iri_len;
2869 	m->m_pkthdr.rcvif = ri->iri_ifp;
2870 	m->m_flags |= ri->iri_flags;
2871 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2872 	m->m_pkthdr.flowid = ri->iri_flowid;
2873 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2874 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2875 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2876 	return (m);
2877 }
2878 
2879 #if defined(INET6) || defined(INET)
2880 static void
2881 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2882 {
2883 	CURVNET_SET(if_getvnet(lc->ifp));
2884 #if defined(INET6)
2885 	*v6 = V_ip6_forwarding;
2886 #endif
2887 #if defined(INET)
2888 	*v4 = V_ipforwarding;
2889 #endif
2890 	CURVNET_RESTORE();
2891 }
2892 
2893 /*
2894  * Returns true if it's possible this packet could be LROed.
2895  * if it returns false, it is guaranteed that tcp_lro_rx()
2896  * would not return zero.
2897  */
2898 static bool
2899 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2900 {
2901 	struct ether_header *eh;
2902 
2903 	eh = mtod(m, struct ether_header *);
2904 	switch (eh->ether_type) {
2905 #if defined(INET6)
2906 		case htons(ETHERTYPE_IPV6):
2907 			return (!v6_forwarding);
2908 #endif
2909 #if defined (INET)
2910 		case htons(ETHERTYPE_IP):
2911 			return (!v4_forwarding);
2912 #endif
2913 	}
2914 
2915 	return false;
2916 }
2917 #else
2918 static void
2919 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2920 {
2921 }
2922 #endif
2923 
2924 static void
2925 _task_fn_rx_watchdog(void *context)
2926 {
2927 	iflib_rxq_t rxq = context;
2928 
2929 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2930 }
2931 
2932 static uint8_t
2933 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2934 {
2935 	if_t ifp;
2936 	if_ctx_t ctx = rxq->ifr_ctx;
2937 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2938 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2939 	int avail, i;
2940 	qidx_t *cidxp;
2941 	struct if_rxd_info ri;
2942 	int err, budget_left, rx_bytes, rx_pkts;
2943 	iflib_fl_t fl;
2944 	int lro_enabled;
2945 	bool v4_forwarding, v6_forwarding, lro_possible;
2946 	uint8_t retval = 0;
2947 
2948 	/*
2949 	 * XXX early demux data packets so that if_input processing only handles
2950 	 * acks in interrupt context
2951 	 */
2952 	struct mbuf *m, *mh, *mt, *mf;
2953 
2954 	NET_EPOCH_ASSERT();
2955 
2956 	lro_possible = v4_forwarding = v6_forwarding = false;
2957 	ifp = ctx->ifc_ifp;
2958 	mh = mt = NULL;
2959 	MPASS(budget > 0);
2960 	rx_pkts	= rx_bytes = 0;
2961 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2962 		cidxp = &rxq->ifr_cq_cidx;
2963 	else
2964 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2965 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2966 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2967 			retval |= iflib_fl_refill_all(ctx, fl);
2968 		DBG_COUNTER_INC(rx_unavail);
2969 		return (retval);
2970 	}
2971 
2972 	/* pfil needs the vnet to be set */
2973 	CURVNET_SET_QUIET(if_getvnet(ifp));
2974 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2975 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2976 			DBG_COUNTER_INC(rx_ctx_inactive);
2977 			break;
2978 		}
2979 		/*
2980 		 * Reset client set fields to their default values
2981 		 */
2982 		rxd_info_zero(&ri);
2983 		ri.iri_qsidx = rxq->ifr_id;
2984 		ri.iri_cidx = *cidxp;
2985 		ri.iri_ifp = ifp;
2986 		ri.iri_frags = rxq->ifr_frags;
2987 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2988 
2989 		if (err)
2990 			goto err;
2991 		rx_pkts += 1;
2992 		rx_bytes += ri.iri_len;
2993 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2994 			*cidxp = ri.iri_cidx;
2995 			/* Update our consumer index */
2996 			/* XXX NB: shurd - check if this is still safe */
2997 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2998 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2999 			/* was this only a completion queue message? */
3000 			if (__predict_false(ri.iri_nfrags == 0))
3001 				continue;
3002 		}
3003 		MPASS(ri.iri_nfrags != 0);
3004 		MPASS(ri.iri_len != 0);
3005 
3006 		/* will advance the cidx on the corresponding free lists */
3007 		m = iflib_rxd_pkt_get(rxq, &ri);
3008 		avail--;
3009 		budget_left--;
3010 		if (avail == 0 && budget_left)
3011 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3012 
3013 		if (__predict_false(m == NULL))
3014 			continue;
3015 
3016 		/* imm_pkt: -- cxgb */
3017 		if (mh == NULL)
3018 			mh = mt = m;
3019 		else {
3020 			mt->m_nextpkt = m;
3021 			mt = m;
3022 		}
3023 	}
3024 	CURVNET_RESTORE();
3025 	/* make sure that we can refill faster than drain */
3026 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3027 		retval |= iflib_fl_refill_all(ctx, fl);
3028 
3029 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3030 	if (lro_enabled)
3031 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3032 	mt = mf = NULL;
3033 	while (mh != NULL) {
3034 		m = mh;
3035 		mh = mh->m_nextpkt;
3036 		m->m_nextpkt = NULL;
3037 #ifndef __NO_STRICT_ALIGNMENT
3038 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3039 			continue;
3040 #endif
3041 #if defined(INET6) || defined(INET)
3042 		if (lro_enabled) {
3043 			if (!lro_possible) {
3044 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3045 				if (lro_possible && mf != NULL) {
3046 					if_input(ifp, mf);
3047 					DBG_COUNTER_INC(rx_if_input);
3048 					mt = mf = NULL;
3049 				}
3050 			}
3051 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3052 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3053 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3054 					continue;
3055 			}
3056 		}
3057 #endif
3058 		if (lro_possible) {
3059 			if_input(ifp, m);
3060 			DBG_COUNTER_INC(rx_if_input);
3061 			continue;
3062 		}
3063 
3064 		if (mf == NULL)
3065 			mf = m;
3066 		if (mt != NULL)
3067 			mt->m_nextpkt = m;
3068 		mt = m;
3069 	}
3070 	if (mf != NULL) {
3071 		if_input(ifp, mf);
3072 		DBG_COUNTER_INC(rx_if_input);
3073 	}
3074 
3075 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3076 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3077 
3078 	/*
3079 	 * Flush any outstanding LRO work
3080 	 */
3081 #if defined(INET6) || defined(INET)
3082 	tcp_lro_flush_all(&rxq->ifr_lc);
3083 #endif
3084 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3085 		retval |= IFLIB_RXEOF_MORE;
3086 	return (retval);
3087 err:
3088 	STATE_LOCK(ctx);
3089 	ctx->ifc_flags |= IFC_DO_RESET;
3090 	iflib_admin_intr_deferred(ctx);
3091 	STATE_UNLOCK(ctx);
3092 	return (0);
3093 }
3094 
3095 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3096 static inline qidx_t
3097 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3098 {
3099 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3100 	qidx_t minthresh = txq->ift_size / 8;
3101 	if (in_use > 4*minthresh)
3102 		return (notify_count);
3103 	if (in_use > 2*minthresh)
3104 		return (notify_count >> 1);
3105 	if (in_use > minthresh)
3106 		return (notify_count >> 3);
3107 	return (0);
3108 }
3109 
3110 static inline qidx_t
3111 txq_max_rs_deferred(iflib_txq_t txq)
3112 {
3113 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3114 	qidx_t minthresh = txq->ift_size / 8;
3115 	if (txq->ift_in_use > 4*minthresh)
3116 		return (notify_count);
3117 	if (txq->ift_in_use > 2*minthresh)
3118 		return (notify_count >> 1);
3119 	if (txq->ift_in_use > minthresh)
3120 		return (notify_count >> 2);
3121 	return (2);
3122 }
3123 
3124 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3125 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3126 
3127 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3128 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3129 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3130 
3131 /* forward compatibility for cxgb */
3132 #define FIRST_QSET(ctx) 0
3133 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3134 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3135 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3136 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3137 
3138 /* XXX we should be setting this to something other than zero */
3139 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3140 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3141     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3142 
3143 static inline bool
3144 iflib_txd_db_check(iflib_txq_t txq, int ring)
3145 {
3146 	if_ctx_t ctx = txq->ift_ctx;
3147 	qidx_t dbval, max;
3148 
3149 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3150 
3151 	/* force || threshold exceeded || at the edge of the ring */
3152 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3153 
3154 		/*
3155 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3156 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3157 		 * producer index explicitly (INTC).
3158 		 */
3159 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3160 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3161 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3162 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3163 
3164 		/*
3165 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3166 		 */
3167 		txq->ift_db_pending = txq->ift_npending = 0;
3168 		return (true);
3169 	}
3170 	return (false);
3171 }
3172 
3173 #ifdef PKT_DEBUG
3174 static void
3175 print_pkt(if_pkt_info_t pi)
3176 {
3177 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3178 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3179 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3180 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3181 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3182 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3183 }
3184 #endif
3185 
3186 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3187 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3188 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3189 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3190 
3191 /**
3192  * Parses out ethernet header information in the given mbuf.
3193  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3194  *
3195  * This will account for the VLAN header if present.
3196  *
3197  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3198  * types of packets.
3199  */
3200 static int
3201 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3202 {
3203 	struct ether_vlan_header *eh;
3204 	struct mbuf *m;
3205 
3206 	m = *mp;
3207 	if (__predict_false(m->m_len < sizeof(*eh))) {
3208 		(*pullups)++;
3209 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3210 			return (ENOMEM);
3211 	}
3212 	eh = mtod(m, struct ether_vlan_header *);
3213 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3214 		pi->ipi_etype = ntohs(eh->evl_proto);
3215 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3216 	} else {
3217 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3218 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3219 	}
3220 	*mp = m;
3221 
3222 	return (0);
3223 }
3224 
3225 /**
3226  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3227  * Currently this information includes: IP ToS value, IP header version/presence
3228  *
3229  * This is missing some checks and doesn't edit the packet content as it goes,
3230  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3231  */
3232 static int
3233 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3234 {
3235 	struct mbuf *m;
3236 	int err;
3237 
3238 	*pullups = 0;
3239 	m = *mp;
3240 	if (!M_WRITABLE(m)) {
3241 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3242 			return (ENOMEM);
3243 		} else {
3244 			m_freem(*mp);
3245 			DBG_COUNTER_INC(tx_frees);
3246 			*mp = m;
3247 		}
3248 	}
3249 
3250 	/* Fills out pi->ipi_etype */
3251 	err = iflib_parse_ether_header(pi, mp, pullups);
3252 	if (err)
3253 		return (err);
3254 	m = *mp;
3255 
3256 	switch (pi->ipi_etype) {
3257 #ifdef INET
3258 	case ETHERTYPE_IP:
3259 	{
3260 		struct mbuf *n;
3261 		struct ip *ip = NULL;
3262 		int miniplen;
3263 
3264 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3265 		if (__predict_false(m->m_len < miniplen)) {
3266 			/*
3267 			 * Check for common case where the first mbuf only contains
3268 			 * the Ethernet header
3269 			 */
3270 			if (m->m_len == pi->ipi_ehdrlen) {
3271 				n = m->m_next;
3272 				MPASS(n);
3273 				/* If next mbuf contains at least the minimal IP header, then stop */
3274 				if (n->m_len >= sizeof(*ip)) {
3275 					ip = (struct ip *)n->m_data;
3276 				} else {
3277 					(*pullups)++;
3278 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3279 						return (ENOMEM);
3280 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3281 				}
3282 			} else {
3283 				(*pullups)++;
3284 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3285 					return (ENOMEM);
3286 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3287 			}
3288 		} else {
3289 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3290 		}
3291 
3292 		/* Have the IPv4 header w/ no options here */
3293 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3294 		pi->ipi_ipproto = ip->ip_p;
3295 		pi->ipi_ip_tos = ip->ip_tos;
3296 		pi->ipi_flags |= IPI_TX_IPV4;
3297 
3298 		break;
3299 	}
3300 #endif
3301 #ifdef INET6
3302 	case ETHERTYPE_IPV6:
3303 	{
3304 		struct ip6_hdr *ip6;
3305 
3306 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3307 			(*pullups)++;
3308 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3309 				return (ENOMEM);
3310 		}
3311 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3312 
3313 		/* Have the IPv6 fixed header here */
3314 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3315 		pi->ipi_ipproto = ip6->ip6_nxt;
3316 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3317 		pi->ipi_flags |= IPI_TX_IPV6;
3318 
3319 		break;
3320 	}
3321 #endif
3322 	default:
3323 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3324 		pi->ipi_ip_hlen = 0;
3325 		break;
3326 	}
3327 	*mp = m;
3328 
3329 	return (0);
3330 
3331 }
3332 
3333 static int
3334 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3335 {
3336 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3337 	struct mbuf *m;
3338 	int err;
3339 
3340 	m = *mp;
3341 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3342 	    M_WRITABLE(m) == 0) {
3343 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3344 			return (ENOMEM);
3345 		} else {
3346 			m_freem(*mp);
3347 			DBG_COUNTER_INC(tx_frees);
3348 			*mp = m;
3349 		}
3350 	}
3351 
3352 	/* Fills out pi->ipi_etype */
3353 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3354 	if (__predict_false(err))
3355 		return (err);
3356 	m = *mp;
3357 
3358 	switch (pi->ipi_etype) {
3359 #ifdef INET
3360 	case ETHERTYPE_IP:
3361 	{
3362 		struct mbuf *n;
3363 		struct ip *ip = NULL;
3364 		struct tcphdr *th = NULL;
3365 		int minthlen;
3366 
3367 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3368 		if (__predict_false(m->m_len < minthlen)) {
3369 			/*
3370 			 * if this code bloat is causing too much of a hit
3371 			 * move it to a separate function and mark it noinline
3372 			 */
3373 			if (m->m_len == pi->ipi_ehdrlen) {
3374 				n = m->m_next;
3375 				MPASS(n);
3376 				if (n->m_len >= sizeof(*ip))  {
3377 					ip = (struct ip *)n->m_data;
3378 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3379 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3380 				} else {
3381 					txq->ift_pullups++;
3382 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3383 						return (ENOMEM);
3384 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3385 				}
3386 			} else {
3387 				txq->ift_pullups++;
3388 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3389 					return (ENOMEM);
3390 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3391 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3392 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3393 			}
3394 		} else {
3395 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3396 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3397 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3398 		}
3399 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3400 		pi->ipi_ipproto = ip->ip_p;
3401 		pi->ipi_ip_tos = ip->ip_tos;
3402 		pi->ipi_flags |= IPI_TX_IPV4;
3403 
3404 		/* TCP checksum offload may require TCP header length */
3405 		if (IS_TX_OFFLOAD4(pi)) {
3406 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3407 				if (__predict_false(th == NULL)) {
3408 					txq->ift_pullups++;
3409 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3410 						return (ENOMEM);
3411 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3412 				}
3413 				pi->ipi_tcp_hflags = th->th_flags;
3414 				pi->ipi_tcp_hlen = th->th_off << 2;
3415 				pi->ipi_tcp_seq = th->th_seq;
3416 			}
3417 			if (IS_TSO4(pi)) {
3418 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3419 					return (ENXIO);
3420 				/*
3421 				 * TSO always requires hardware checksum offload.
3422 				 */
3423 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3424 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3425 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3426 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3427 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3428 					ip->ip_sum = 0;
3429 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3430 				}
3431 			}
3432 		}
3433 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3434                        ip->ip_sum = 0;
3435 
3436 		break;
3437 	}
3438 #endif
3439 #ifdef INET6
3440 	case ETHERTYPE_IPV6:
3441 	{
3442 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3443 		struct tcphdr *th;
3444 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3445 
3446 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3447 			txq->ift_pullups++;
3448 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3449 				return (ENOMEM);
3450 		}
3451 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3452 
3453 		/* XXX-BZ this will go badly in case of ext hdrs. */
3454 		pi->ipi_ipproto = ip6->ip6_nxt;
3455 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3456 		pi->ipi_flags |= IPI_TX_IPV6;
3457 
3458 		/* TCP checksum offload may require TCP header length */
3459 		if (IS_TX_OFFLOAD6(pi)) {
3460 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3461 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3462 					txq->ift_pullups++;
3463 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3464 						return (ENOMEM);
3465 				}
3466 				pi->ipi_tcp_hflags = th->th_flags;
3467 				pi->ipi_tcp_hlen = th->th_off << 2;
3468 				pi->ipi_tcp_seq = th->th_seq;
3469 			}
3470 			if (IS_TSO6(pi)) {
3471 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3472 					return (ENXIO);
3473 				/*
3474 				 * TSO always requires hardware checksum offload.
3475 				 */
3476 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3477 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3478 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3479 			}
3480 		}
3481 		break;
3482 	}
3483 #endif
3484 	default:
3485 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3486 		pi->ipi_ip_hlen = 0;
3487 		break;
3488 	}
3489 	*mp = m;
3490 
3491 	return (0);
3492 }
3493 
3494 /*
3495  * If dodgy hardware rejects the scatter gather chain we've handed it
3496  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3497  * m_defrag'd mbufs
3498  */
3499 static __noinline struct mbuf *
3500 iflib_remove_mbuf(iflib_txq_t txq)
3501 {
3502 	int ntxd, pidx;
3503 	struct mbuf *m, **ifsd_m;
3504 
3505 	ifsd_m = txq->ift_sds.ifsd_m;
3506 	ntxd = txq->ift_size;
3507 	pidx = txq->ift_pidx & (ntxd - 1);
3508 	ifsd_m = txq->ift_sds.ifsd_m;
3509 	m = ifsd_m[pidx];
3510 	ifsd_m[pidx] = NULL;
3511 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3512 	if (txq->ift_sds.ifsd_tso_map != NULL)
3513 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3514 		    txq->ift_sds.ifsd_tso_map[pidx]);
3515 #if MEMORY_LOGGING
3516 	txq->ift_dequeued++;
3517 #endif
3518 	return (m);
3519 }
3520 
3521 static inline caddr_t
3522 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3523 {
3524 	qidx_t size;
3525 	int ntxd;
3526 	caddr_t start, end, cur, next;
3527 
3528 	ntxd = txq->ift_size;
3529 	size = txq->ift_txd_size[qid];
3530 	start = txq->ift_ifdi[qid].idi_vaddr;
3531 
3532 	if (__predict_false(size == 0))
3533 		return (start);
3534 	cur = start + size*cidx;
3535 	end = start + size*ntxd;
3536 	next = CACHE_PTR_NEXT(cur);
3537 	return (next < end ? next : start);
3538 }
3539 
3540 /*
3541  * Pad an mbuf to ensure a minimum ethernet frame size.
3542  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3543  */
3544 static __noinline int
3545 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3546 {
3547 	/*
3548 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3549 	 * and ARP message is the smallest common payload I can think of
3550 	 */
3551 	static char pad[18];	/* just zeros */
3552 	int n;
3553 	struct mbuf *new_head;
3554 
3555 	if (!M_WRITABLE(*m_head)) {
3556 		new_head = m_dup(*m_head, M_NOWAIT);
3557 		if (new_head == NULL) {
3558 			m_freem(*m_head);
3559 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3560 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3561 			DBG_COUNTER_INC(tx_frees);
3562 			return ENOMEM;
3563 		}
3564 		m_freem(*m_head);
3565 		*m_head = new_head;
3566 	}
3567 
3568 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3569 	     n > 0; n -= sizeof(pad))
3570 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3571 			break;
3572 
3573 	if (n > 0) {
3574 		m_freem(*m_head);
3575 		device_printf(dev, "cannot pad short frame\n");
3576 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3577 		DBG_COUNTER_INC(tx_frees);
3578 		return (ENOBUFS);
3579 	}
3580 
3581 	return 0;
3582 }
3583 
3584 static int
3585 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3586 {
3587 	if_ctx_t		ctx;
3588 	if_shared_ctx_t		sctx;
3589 	if_softc_ctx_t		scctx;
3590 	bus_dma_tag_t		buf_tag;
3591 	bus_dma_segment_t	*segs;
3592 	struct mbuf		*m_head, **ifsd_m;
3593 	void			*next_txd;
3594 	bus_dmamap_t		map;
3595 	struct if_pkt_info	pi;
3596 	int remap = 0;
3597 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3598 
3599 	ctx = txq->ift_ctx;
3600 	sctx = ctx->ifc_sctx;
3601 	scctx = &ctx->ifc_softc_ctx;
3602 	segs = txq->ift_segs;
3603 	ntxd = txq->ift_size;
3604 	m_head = *m_headp;
3605 	map = NULL;
3606 
3607 	/*
3608 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3609 	 */
3610 	cidx = txq->ift_cidx;
3611 	pidx = txq->ift_pidx;
3612 	if (ctx->ifc_flags & IFC_PREFETCH) {
3613 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3614 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3615 			next_txd = calc_next_txd(txq, cidx, 0);
3616 			prefetch(next_txd);
3617 		}
3618 
3619 		/* prefetch the next cache line of mbuf pointers and flags */
3620 		prefetch(&txq->ift_sds.ifsd_m[next]);
3621 		prefetch(&txq->ift_sds.ifsd_map[next]);
3622 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3623 	}
3624 	map = txq->ift_sds.ifsd_map[pidx];
3625 	ifsd_m = txq->ift_sds.ifsd_m;
3626 
3627 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3628 		buf_tag = txq->ift_tso_buf_tag;
3629 		max_segs = scctx->isc_tx_tso_segments_max;
3630 		map = txq->ift_sds.ifsd_tso_map[pidx];
3631 		MPASS(buf_tag != NULL);
3632 		MPASS(max_segs > 0);
3633 	} else {
3634 		buf_tag = txq->ift_buf_tag;
3635 		max_segs = scctx->isc_tx_nsegments;
3636 		map = txq->ift_sds.ifsd_map[pidx];
3637 	}
3638 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3639 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3640 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3641 		if (err) {
3642 			DBG_COUNTER_INC(encap_txd_encap_fail);
3643 			return err;
3644 		}
3645 	}
3646 	m_head = *m_headp;
3647 
3648 	pkt_info_zero(&pi);
3649 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3650 	pi.ipi_pidx = pidx;
3651 	pi.ipi_qsidx = txq->ift_id;
3652 	pi.ipi_len = m_head->m_pkthdr.len;
3653 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3654 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3655 
3656 	/* deliberate bitwise OR to make one condition */
3657 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3658 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3659 			DBG_COUNTER_INC(encap_txd_encap_fail);
3660 			return (err);
3661 		}
3662 		m_head = *m_headp;
3663 	}
3664 
3665 retry:
3666 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3667 	    BUS_DMA_NOWAIT);
3668 defrag:
3669 	if (__predict_false(err)) {
3670 		switch (err) {
3671 		case EFBIG:
3672 			/* try collapse once and defrag once */
3673 			if (remap == 0) {
3674 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3675 				/* try defrag if collapsing fails */
3676 				if (m_head == NULL)
3677 					remap++;
3678 			}
3679 			if (remap == 1) {
3680 				txq->ift_mbuf_defrag++;
3681 				m_head = m_defrag(*m_headp, M_NOWAIT);
3682 			}
3683 			/*
3684 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3685 			 * failed to map an mbuf that was run through m_defrag
3686 			 */
3687 			MPASS(remap <= 1);
3688 			if (__predict_false(m_head == NULL || remap > 1))
3689 				goto defrag_failed;
3690 			remap++;
3691 			*m_headp = m_head;
3692 			goto retry;
3693 			break;
3694 		case ENOMEM:
3695 			txq->ift_no_tx_dma_setup++;
3696 			break;
3697 		default:
3698 			txq->ift_no_tx_dma_setup++;
3699 			m_freem(*m_headp);
3700 			DBG_COUNTER_INC(tx_frees);
3701 			*m_headp = NULL;
3702 			break;
3703 		}
3704 		txq->ift_map_failed++;
3705 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3706 		DBG_COUNTER_INC(encap_txd_encap_fail);
3707 		return (err);
3708 	}
3709 	ifsd_m[pidx] = m_head;
3710 	/*
3711 	 * XXX assumes a 1 to 1 relationship between segments and
3712 	 *        descriptors - this does not hold true on all drivers, e.g.
3713 	 *        cxgb
3714 	 */
3715 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3716 		txq->ift_no_desc_avail++;
3717 		bus_dmamap_unload(buf_tag, map);
3718 		DBG_COUNTER_INC(encap_txq_avail_fail);
3719 		DBG_COUNTER_INC(encap_txd_encap_fail);
3720 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3721 			GROUPTASK_ENQUEUE(&txq->ift_task);
3722 		return (ENOBUFS);
3723 	}
3724 	/*
3725 	 * On Intel cards we can greatly reduce the number of TX interrupts
3726 	 * we see by only setting report status on every Nth descriptor.
3727 	 * However, this also means that the driver will need to keep track
3728 	 * of the descriptors that RS was set on to check them for the DD bit.
3729 	 */
3730 	txq->ift_rs_pending += nsegs + 1;
3731 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3732 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3733 		pi.ipi_flags |= IPI_TX_INTR;
3734 		txq->ift_rs_pending = 0;
3735 	}
3736 
3737 	pi.ipi_segs = segs;
3738 	pi.ipi_nsegs = nsegs;
3739 
3740 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3741 #ifdef PKT_DEBUG
3742 	print_pkt(&pi);
3743 #endif
3744 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3745 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3746 		DBG_COUNTER_INC(tx_encap);
3747 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3748 
3749 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3750 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3751 			ndesc += txq->ift_size;
3752 			txq->ift_gen = 1;
3753 		}
3754 		/*
3755 		 * drivers can need as many as
3756 		 * two sentinels
3757 		 */
3758 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3759 		MPASS(pi.ipi_new_pidx != pidx);
3760 		MPASS(ndesc > 0);
3761 		txq->ift_in_use += ndesc;
3762 		txq->ift_db_pending += ndesc;
3763 
3764 		/*
3765 		 * We update the last software descriptor again here because there may
3766 		 * be a sentinel and/or there may be more mbufs than segments
3767 		 */
3768 		txq->ift_pidx = pi.ipi_new_pidx;
3769 		txq->ift_npending += pi.ipi_ndescs;
3770 	} else {
3771 		*m_headp = m_head = iflib_remove_mbuf(txq);
3772 		if (err == EFBIG) {
3773 			txq->ift_txd_encap_efbig++;
3774 			if (remap < 2) {
3775 				remap = 1;
3776 				goto defrag;
3777 			}
3778 		}
3779 		goto defrag_failed;
3780 	}
3781 	/*
3782 	 * err can't possibly be non-zero here, so we don't neet to test it
3783 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3784 	 */
3785 	return (err);
3786 
3787 defrag_failed:
3788 	txq->ift_mbuf_defrag_failed++;
3789 	txq->ift_map_failed++;
3790 	m_freem(*m_headp);
3791 	DBG_COUNTER_INC(tx_frees);
3792 	*m_headp = NULL;
3793 	DBG_COUNTER_INC(encap_txd_encap_fail);
3794 	return (ENOMEM);
3795 }
3796 
3797 static void
3798 iflib_tx_desc_free(iflib_txq_t txq, int n)
3799 {
3800 	uint32_t qsize, cidx, mask, gen;
3801 	struct mbuf *m, **ifsd_m;
3802 	bool do_prefetch;
3803 
3804 	cidx = txq->ift_cidx;
3805 	gen = txq->ift_gen;
3806 	qsize = txq->ift_size;
3807 	mask = qsize-1;
3808 	ifsd_m = txq->ift_sds.ifsd_m;
3809 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3810 
3811 	while (n-- > 0) {
3812 		if (do_prefetch) {
3813 			prefetch(ifsd_m[(cidx + 3) & mask]);
3814 			prefetch(ifsd_m[(cidx + 4) & mask]);
3815 		}
3816 		if ((m = ifsd_m[cidx]) != NULL) {
3817 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3818 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3819 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3820 				    txq->ift_sds.ifsd_tso_map[cidx],
3821 				    BUS_DMASYNC_POSTWRITE);
3822 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3823 				    txq->ift_sds.ifsd_tso_map[cidx]);
3824 			} else {
3825 				bus_dmamap_sync(txq->ift_buf_tag,
3826 				    txq->ift_sds.ifsd_map[cidx],
3827 				    BUS_DMASYNC_POSTWRITE);
3828 				bus_dmamap_unload(txq->ift_buf_tag,
3829 				    txq->ift_sds.ifsd_map[cidx]);
3830 			}
3831 			/* XXX we don't support any drivers that batch packets yet */
3832 			MPASS(m->m_nextpkt == NULL);
3833 			m_freem(m);
3834 			ifsd_m[cidx] = NULL;
3835 #if MEMORY_LOGGING
3836 			txq->ift_dequeued++;
3837 #endif
3838 			DBG_COUNTER_INC(tx_frees);
3839 		}
3840 		if (__predict_false(++cidx == qsize)) {
3841 			cidx = 0;
3842 			gen = 0;
3843 		}
3844 	}
3845 	txq->ift_cidx = cidx;
3846 	txq->ift_gen = gen;
3847 }
3848 
3849 static __inline int
3850 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3851 {
3852 	int reclaim;
3853 	if_ctx_t ctx = txq->ift_ctx;
3854 
3855 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3856 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3857 
3858 	/*
3859 	 * Need a rate-limiting check so that this isn't called every time
3860 	 */
3861 	iflib_tx_credits_update(ctx, txq);
3862 	reclaim = DESC_RECLAIMABLE(txq);
3863 
3864 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3865 #ifdef INVARIANTS
3866 		if (iflib_verbose_debug) {
3867 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3868 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3869 			       reclaim, thresh);
3870 		}
3871 #endif
3872 		return (0);
3873 	}
3874 	iflib_tx_desc_free(txq, reclaim);
3875 	txq->ift_cleaned += reclaim;
3876 	txq->ift_in_use -= reclaim;
3877 
3878 	return (reclaim);
3879 }
3880 
3881 static struct mbuf **
3882 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3883 {
3884 	int next, size;
3885 	struct mbuf **items;
3886 
3887 	size = r->size;
3888 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3889 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3890 
3891 	prefetch(items[(cidx + offset) & (size-1)]);
3892 	if (remaining > 1) {
3893 		prefetch2cachelines(&items[next]);
3894 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3895 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3896 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3897 	}
3898 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3899 }
3900 
3901 static void
3902 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3903 {
3904 
3905 	ifmp_ring_check_drainage(txq->ift_br, budget);
3906 }
3907 
3908 static uint32_t
3909 iflib_txq_can_drain(struct ifmp_ring *r)
3910 {
3911 	iflib_txq_t txq = r->cookie;
3912 	if_ctx_t ctx = txq->ift_ctx;
3913 
3914 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3915 		return (1);
3916 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3917 	    BUS_DMASYNC_POSTREAD);
3918 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3919 	    false));
3920 }
3921 
3922 static uint32_t
3923 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3924 {
3925 	iflib_txq_t txq = r->cookie;
3926 	if_ctx_t ctx = txq->ift_ctx;
3927 	if_t ifp = ctx->ifc_ifp;
3928 	struct mbuf *m, **mp;
3929 	int avail, bytes_sent, skipped, count, err, i;
3930 	int mcast_sent, pkt_sent, reclaimed;
3931 	bool do_prefetch, rang, ring;
3932 
3933 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3934 			    !LINK_ACTIVE(ctx))) {
3935 		DBG_COUNTER_INC(txq_drain_notready);
3936 		return (0);
3937 	}
3938 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3939 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3940 	avail = IDXDIFF(pidx, cidx, r->size);
3941 
3942 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3943 		/*
3944 		 * The driver is unloading so we need to free all pending packets.
3945 		 */
3946 		DBG_COUNTER_INC(txq_drain_flushing);
3947 		for (i = 0; i < avail; i++) {
3948 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3949 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3950 			r->items[(cidx + i) & (r->size-1)] = NULL;
3951 		}
3952 		return (avail);
3953 	}
3954 
3955 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3956 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3957 		CALLOUT_LOCK(txq);
3958 		callout_stop(&txq->ift_timer);
3959 		CALLOUT_UNLOCK(txq);
3960 		DBG_COUNTER_INC(txq_drain_oactive);
3961 		return (0);
3962 	}
3963 
3964 	/*
3965 	 * If we've reclaimed any packets this queue cannot be hung.
3966 	 */
3967 	if (reclaimed)
3968 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3969 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3970 	count = MIN(avail, TX_BATCH_SIZE);
3971 #ifdef INVARIANTS
3972 	if (iflib_verbose_debug)
3973 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3974 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3975 #endif
3976 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3977 	err = 0;
3978 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3979 		int rem = do_prefetch ? count - i : 0;
3980 
3981 		mp = _ring_peek_one(r, cidx, i, rem);
3982 		MPASS(mp != NULL && *mp != NULL);
3983 
3984 		/*
3985 		 * Completion interrupts will use the address of the txq
3986 		 * as a sentinel to enqueue _something_ in order to acquire
3987 		 * the lock on the mp_ring (there's no direct lock call).
3988 		 * We obviously whave to check for these sentinel cases
3989 		 * and skip them.
3990 		 */
3991 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3992 			skipped++;
3993 			continue;
3994 		}
3995 		err = iflib_encap(txq, mp);
3996 		if (__predict_false(err)) {
3997 			/* no room - bail out */
3998 			if (err == ENOBUFS)
3999 				break;
4000 			skipped++;
4001 			/* we can't send this packet - skip it */
4002 			continue;
4003 		}
4004 		pkt_sent++;
4005 		m = *mp;
4006 		DBG_COUNTER_INC(tx_sent);
4007 		bytes_sent += m->m_pkthdr.len;
4008 		mcast_sent += !!(m->m_flags & M_MCAST);
4009 
4010 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4011 			break;
4012 		ETHER_BPF_MTAP(ifp, m);
4013 		rang = iflib_txd_db_check(txq, false);
4014 	}
4015 
4016 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4017 	ring = rang ? false  : (iflib_min_tx_latency | err);
4018 	iflib_txd_db_check(txq, ring);
4019 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4020 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4021 	if (mcast_sent)
4022 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4023 #ifdef INVARIANTS
4024 	if (iflib_verbose_debug)
4025 		printf("consumed=%d\n", skipped + pkt_sent);
4026 #endif
4027 	return (skipped + pkt_sent);
4028 }
4029 
4030 static uint32_t
4031 iflib_txq_drain_always(struct ifmp_ring *r)
4032 {
4033 	return (1);
4034 }
4035 
4036 static uint32_t
4037 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4038 {
4039 	int i, avail;
4040 	struct mbuf **mp;
4041 	iflib_txq_t txq;
4042 
4043 	txq = r->cookie;
4044 
4045 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4046 	CALLOUT_LOCK(txq);
4047 	callout_stop(&txq->ift_timer);
4048 	CALLOUT_UNLOCK(txq);
4049 
4050 	avail = IDXDIFF(pidx, cidx, r->size);
4051 	for (i = 0; i < avail; i++) {
4052 		mp = _ring_peek_one(r, cidx, i, avail - i);
4053 		if (__predict_false(*mp == (struct mbuf *)txq))
4054 			continue;
4055 		m_freem(*mp);
4056 		DBG_COUNTER_INC(tx_frees);
4057 	}
4058 	MPASS(ifmp_ring_is_stalled(r) == 0);
4059 	return (avail);
4060 }
4061 
4062 static void
4063 iflib_ifmp_purge(iflib_txq_t txq)
4064 {
4065 	struct ifmp_ring *r;
4066 
4067 	r = txq->ift_br;
4068 	r->drain = iflib_txq_drain_free;
4069 	r->can_drain = iflib_txq_drain_always;
4070 
4071 	ifmp_ring_check_drainage(r, r->size);
4072 
4073 	r->drain = iflib_txq_drain;
4074 	r->can_drain = iflib_txq_can_drain;
4075 }
4076 
4077 static void
4078 _task_fn_tx(void *context)
4079 {
4080 	iflib_txq_t txq = context;
4081 	if_ctx_t ctx = txq->ift_ctx;
4082 	if_t ifp = ctx->ifc_ifp;
4083 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4084 
4085 #ifdef IFLIB_DIAGNOSTICS
4086 	txq->ift_cpu_exec_count[curcpu]++;
4087 #endif
4088 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4089 		return;
4090 #ifdef DEV_NETMAP
4091 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4092 	    netmap_tx_irq(ifp, txq->ift_id))
4093 		goto skip_ifmp;
4094 #endif
4095 #ifdef ALTQ
4096 	if (if_altq_is_enabled(ifp))
4097 		iflib_altq_if_start(ifp);
4098 #endif
4099 	if (txq->ift_db_pending)
4100 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4101 	else if (!abdicate)
4102 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4103 	/*
4104 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4105 	 */
4106 	if (abdicate)
4107 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4108 #ifdef DEV_NETMAP
4109 skip_ifmp:
4110 #endif
4111 	if (ctx->ifc_flags & IFC_LEGACY)
4112 		IFDI_INTR_ENABLE(ctx);
4113 	else
4114 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4115 }
4116 
4117 static void
4118 _task_fn_rx(void *context)
4119 {
4120 	iflib_rxq_t rxq = context;
4121 	if_ctx_t ctx = rxq->ifr_ctx;
4122 	uint8_t more;
4123 	uint16_t budget;
4124 #ifdef DEV_NETMAP
4125 	u_int work = 0;
4126 	int nmirq;
4127 #endif
4128 
4129 #ifdef IFLIB_DIAGNOSTICS
4130 	rxq->ifr_cpu_exec_count[curcpu]++;
4131 #endif
4132 	DBG_COUNTER_INC(task_fn_rxs);
4133 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4134 		return;
4135 #ifdef DEV_NETMAP
4136 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4137 	if (nmirq != NM_IRQ_PASS) {
4138 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4139 		goto skip_rxeof;
4140 	}
4141 #endif
4142 	budget = ctx->ifc_sysctl_rx_budget;
4143 	if (budget == 0)
4144 		budget = 16;	/* XXX */
4145 	more = iflib_rxeof(rxq, budget);
4146 #ifdef DEV_NETMAP
4147 skip_rxeof:
4148 #endif
4149 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4150 		if (ctx->ifc_flags & IFC_LEGACY)
4151 			IFDI_INTR_ENABLE(ctx);
4152 		else
4153 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4154 		DBG_COUNTER_INC(rx_intr_enables);
4155 	}
4156 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4157 		return;
4158 
4159 	if (more & IFLIB_RXEOF_MORE)
4160 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4161 	else if (more & IFLIB_RXEOF_EMPTY)
4162 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4163 }
4164 
4165 static void
4166 _task_fn_admin(void *context)
4167 {
4168 	if_ctx_t ctx = context;
4169 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4170 	iflib_txq_t txq;
4171 	int i;
4172 	bool oactive, running, do_reset, do_watchdog, in_detach;
4173 
4174 	STATE_LOCK(ctx);
4175 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4176 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4177 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4178 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4179 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4180 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4181 	STATE_UNLOCK(ctx);
4182 
4183 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4184 		return;
4185 	if (in_detach)
4186 		return;
4187 
4188 	CTX_LOCK(ctx);
4189 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4190 		CALLOUT_LOCK(txq);
4191 		callout_stop(&txq->ift_timer);
4192 		CALLOUT_UNLOCK(txq);
4193 	}
4194 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4195 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4196 	if (do_watchdog) {
4197 		ctx->ifc_watchdog_events++;
4198 		IFDI_WATCHDOG_RESET(ctx);
4199 	}
4200 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4201 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4202 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4203 		    txq->ift_timer.c_cpu);
4204 	}
4205 	IFDI_LINK_INTR_ENABLE(ctx);
4206 	if (do_reset)
4207 		iflib_if_init_locked(ctx);
4208 	CTX_UNLOCK(ctx);
4209 
4210 	if (LINK_ACTIVE(ctx) == 0)
4211 		return;
4212 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4213 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4214 }
4215 
4216 static void
4217 _task_fn_iov(void *context)
4218 {
4219 	if_ctx_t ctx = context;
4220 
4221 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4222 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4223 		return;
4224 
4225 	CTX_LOCK(ctx);
4226 	IFDI_VFLR_HANDLE(ctx);
4227 	CTX_UNLOCK(ctx);
4228 }
4229 
4230 static int
4231 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4232 {
4233 	int err;
4234 	if_int_delay_info_t info;
4235 	if_ctx_t ctx;
4236 
4237 	info = (if_int_delay_info_t)arg1;
4238 	ctx = info->iidi_ctx;
4239 	info->iidi_req = req;
4240 	info->iidi_oidp = oidp;
4241 	CTX_LOCK(ctx);
4242 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4243 	CTX_UNLOCK(ctx);
4244 	return (err);
4245 }
4246 
4247 /*********************************************************************
4248  *
4249  *  IFNET FUNCTIONS
4250  *
4251  **********************************************************************/
4252 
4253 static void
4254 iflib_if_init_locked(if_ctx_t ctx)
4255 {
4256 	iflib_stop(ctx);
4257 	iflib_init_locked(ctx);
4258 }
4259 
4260 static void
4261 iflib_if_init(void *arg)
4262 {
4263 	if_ctx_t ctx = arg;
4264 
4265 	CTX_LOCK(ctx);
4266 	iflib_if_init_locked(ctx);
4267 	CTX_UNLOCK(ctx);
4268 }
4269 
4270 static int
4271 iflib_if_transmit(if_t ifp, struct mbuf *m)
4272 {
4273 	if_ctx_t ctx = if_getsoftc(ifp);
4274 	iflib_txq_t txq;
4275 	int err, qidx;
4276 	int abdicate;
4277 
4278 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4279 		DBG_COUNTER_INC(tx_frees);
4280 		m_freem(m);
4281 		return (ENETDOWN);
4282 	}
4283 
4284 	MPASS(m->m_nextpkt == NULL);
4285 	/* ALTQ-enabled interfaces always use queue 0. */
4286 	qidx = 0;
4287 	/* Use driver-supplied queue selection method if it exists */
4288 	if (ctx->isc_txq_select_v2) {
4289 		struct if_pkt_info pi;
4290 		uint64_t early_pullups = 0;
4291 		pkt_info_zero(&pi);
4292 
4293 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4294 		if (__predict_false(err != 0)) {
4295 			/* Assign pullups for bad pkts to default queue */
4296 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4297 			DBG_COUNTER_INC(encap_txd_encap_fail);
4298 			return (err);
4299 		}
4300 		/* Let driver make queueing decision */
4301 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4302 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4303 	}
4304 	/* Backwards compatibility w/ simpler queue select */
4305 	else if (ctx->isc_txq_select)
4306 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4307 	/* If not, use iflib's standard method */
4308 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4309 		qidx = QIDX(ctx, m);
4310 
4311 	/* Set TX queue */
4312 	txq = &ctx->ifc_txqs[qidx];
4313 
4314 #ifdef DRIVER_BACKPRESSURE
4315 	if (txq->ift_closed) {
4316 		while (m != NULL) {
4317 			next = m->m_nextpkt;
4318 			m->m_nextpkt = NULL;
4319 			m_freem(m);
4320 			DBG_COUNTER_INC(tx_frees);
4321 			m = next;
4322 		}
4323 		return (ENOBUFS);
4324 	}
4325 #endif
4326 #ifdef notyet
4327 	qidx = count = 0;
4328 	mp = marr;
4329 	next = m;
4330 	do {
4331 		count++;
4332 		next = next->m_nextpkt;
4333 	} while (next != NULL);
4334 
4335 	if (count > nitems(marr))
4336 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4337 			/* XXX check nextpkt */
4338 			m_freem(m);
4339 			/* XXX simplify for now */
4340 			DBG_COUNTER_INC(tx_frees);
4341 			return (ENOBUFS);
4342 		}
4343 	for (next = m, i = 0; next != NULL; i++) {
4344 		mp[i] = next;
4345 		next = next->m_nextpkt;
4346 		mp[i]->m_nextpkt = NULL;
4347 	}
4348 #endif
4349 	DBG_COUNTER_INC(tx_seen);
4350 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4351 
4352 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4353 
4354 	if (abdicate)
4355 		GROUPTASK_ENQUEUE(&txq->ift_task);
4356  	if (err) {
4357 		if (!abdicate)
4358 			GROUPTASK_ENQUEUE(&txq->ift_task);
4359 		/* support forthcoming later */
4360 #ifdef DRIVER_BACKPRESSURE
4361 		txq->ift_closed = TRUE;
4362 #endif
4363 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4364 		m_freem(m);
4365 		DBG_COUNTER_INC(tx_frees);
4366 	}
4367 
4368 	return (err);
4369 }
4370 
4371 #ifdef ALTQ
4372 /*
4373  * The overall approach to integrating iflib with ALTQ is to continue to use
4374  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4375  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4376  * is redundant/unnecessary, but doing so minimizes the amount of
4377  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4378  * redundantly queueing to an intermediate mp_ring is swamped by the
4379  * performance limitations inherent in using ALTQ.
4380  *
4381  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4382  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4383  * given interface.  If ALTQ is enabled for an interface, then all
4384  * transmitted packets for that interface will be submitted to the ALTQ
4385  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4386  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4387  * update stats that the iflib machinery handles, and which is sensitve to
4388  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4389  * will be installed as the start routine for use by ALTQ facilities that
4390  * need to trigger queue drains on a scheduled basis.
4391  *
4392  */
4393 static void
4394 iflib_altq_if_start(if_t ifp)
4395 {
4396 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4397 	struct mbuf *m;
4398 
4399 	IFQ_LOCK(ifq);
4400 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4401 	while (m != NULL) {
4402 		iflib_if_transmit(ifp, m);
4403 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4404 	}
4405 	IFQ_UNLOCK(ifq);
4406 }
4407 
4408 static int
4409 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4410 {
4411 	int err;
4412 
4413 	if (if_altq_is_enabled(ifp)) {
4414 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4415 		if (err == 0)
4416 			iflib_altq_if_start(ifp);
4417 	} else
4418 		err = iflib_if_transmit(ifp, m);
4419 
4420 	return (err);
4421 }
4422 #endif /* ALTQ */
4423 
4424 static void
4425 iflib_if_qflush(if_t ifp)
4426 {
4427 	if_ctx_t ctx = if_getsoftc(ifp);
4428 	iflib_txq_t txq = ctx->ifc_txqs;
4429 	int i;
4430 
4431 	STATE_LOCK(ctx);
4432 	ctx->ifc_flags |= IFC_QFLUSH;
4433 	STATE_UNLOCK(ctx);
4434 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4435 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4436 			iflib_txq_check_drain(txq, 0);
4437 	STATE_LOCK(ctx);
4438 	ctx->ifc_flags &= ~IFC_QFLUSH;
4439 	STATE_UNLOCK(ctx);
4440 
4441 	/*
4442 	 * When ALTQ is enabled, this will also take care of purging the
4443 	 * ALTQ queue(s).
4444 	 */
4445 	if_qflush(ifp);
4446 }
4447 
4448 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4449 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4450 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4451 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4452 
4453 static int
4454 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4455 {
4456 	if_ctx_t ctx = if_getsoftc(ifp);
4457 	struct ifreq	*ifr = (struct ifreq *)data;
4458 #if defined(INET) || defined(INET6)
4459 	struct ifaddr	*ifa = (struct ifaddr *)data;
4460 #endif
4461 	bool		avoid_reset = false;
4462 	int		err = 0, reinit = 0, bits;
4463 
4464 	switch (command) {
4465 	case SIOCSIFADDR:
4466 #ifdef INET
4467 		if (ifa->ifa_addr->sa_family == AF_INET)
4468 			avoid_reset = true;
4469 #endif
4470 #ifdef INET6
4471 		if (ifa->ifa_addr->sa_family == AF_INET6)
4472 			avoid_reset = true;
4473 #endif
4474 		/*
4475 		** Calling init results in link renegotiation,
4476 		** so we avoid doing it when possible.
4477 		*/
4478 		if (avoid_reset) {
4479 			if_setflagbits(ifp, IFF_UP,0);
4480 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4481 				reinit = 1;
4482 #ifdef INET
4483 			if (!(if_getflags(ifp) & IFF_NOARP))
4484 				arp_ifinit(ifp, ifa);
4485 #endif
4486 		} else
4487 			err = ether_ioctl(ifp, command, data);
4488 		break;
4489 	case SIOCSIFMTU:
4490 		CTX_LOCK(ctx);
4491 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4492 			CTX_UNLOCK(ctx);
4493 			break;
4494 		}
4495 		bits = if_getdrvflags(ifp);
4496 		/* stop the driver and free any clusters before proceeding */
4497 		iflib_stop(ctx);
4498 
4499 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4500 			STATE_LOCK(ctx);
4501 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4502 				ctx->ifc_flags |= IFC_MULTISEG;
4503 			else
4504 				ctx->ifc_flags &= ~IFC_MULTISEG;
4505 			STATE_UNLOCK(ctx);
4506 			err = if_setmtu(ifp, ifr->ifr_mtu);
4507 		}
4508 		iflib_init_locked(ctx);
4509 		STATE_LOCK(ctx);
4510 		if_setdrvflags(ifp, bits);
4511 		STATE_UNLOCK(ctx);
4512 		CTX_UNLOCK(ctx);
4513 		break;
4514 	case SIOCSIFFLAGS:
4515 		CTX_LOCK(ctx);
4516 		if (if_getflags(ifp) & IFF_UP) {
4517 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4518 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4519 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4520 					CTX_UNLOCK(ctx);
4521 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4522 					CTX_LOCK(ctx);
4523 				}
4524 			} else
4525 				reinit = 1;
4526 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4527 			iflib_stop(ctx);
4528 		}
4529 		ctx->ifc_if_flags = if_getflags(ifp);
4530 		CTX_UNLOCK(ctx);
4531 		break;
4532 	case SIOCADDMULTI:
4533 	case SIOCDELMULTI:
4534 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4535 			CTX_LOCK(ctx);
4536 			IFDI_INTR_DISABLE(ctx);
4537 			IFDI_MULTI_SET(ctx);
4538 			IFDI_INTR_ENABLE(ctx);
4539 			CTX_UNLOCK(ctx);
4540 		}
4541 		break;
4542 	case SIOCSIFMEDIA:
4543 		CTX_LOCK(ctx);
4544 		IFDI_MEDIA_SET(ctx);
4545 		CTX_UNLOCK(ctx);
4546 		/* FALLTHROUGH */
4547 	case SIOCGIFMEDIA:
4548 	case SIOCGIFXMEDIA:
4549 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4550 		break;
4551 	case SIOCGI2C:
4552 	{
4553 		struct ifi2creq i2c;
4554 
4555 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4556 		if (err != 0)
4557 			break;
4558 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4559 			err = EINVAL;
4560 			break;
4561 		}
4562 		if (i2c.len > sizeof(i2c.data)) {
4563 			err = EINVAL;
4564 			break;
4565 		}
4566 
4567 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4568 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4569 			    sizeof(i2c));
4570 		break;
4571 	}
4572 	case SIOCSIFCAP:
4573 	{
4574 		int mask, setmask, oldmask;
4575 
4576 		oldmask = if_getcapenable(ifp);
4577 		mask = ifr->ifr_reqcap ^ oldmask;
4578 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4579 		setmask = 0;
4580 #ifdef TCP_OFFLOAD
4581 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4582 #endif
4583 		setmask |= (mask & IFCAP_FLAGS);
4584 		setmask |= (mask & IFCAP_WOL);
4585 
4586 		/*
4587 		 * If any RX csum has changed, change all the ones that
4588 		 * are supported by the driver.
4589 		 */
4590 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4591 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4592 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4593 		}
4594 
4595 		/*
4596 		 * want to ensure that traffic has stopped before we change any of the flags
4597 		 */
4598 		if (setmask) {
4599 			CTX_LOCK(ctx);
4600 			bits = if_getdrvflags(ifp);
4601 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4602 				iflib_stop(ctx);
4603 			STATE_LOCK(ctx);
4604 			if_togglecapenable(ifp, setmask);
4605 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4606 			STATE_UNLOCK(ctx);
4607 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4608 				iflib_init_locked(ctx);
4609 			STATE_LOCK(ctx);
4610 			if_setdrvflags(ifp, bits);
4611 			STATE_UNLOCK(ctx);
4612 			CTX_UNLOCK(ctx);
4613 		}
4614 		if_vlancap(ifp);
4615 		break;
4616 	}
4617 	case SIOCGPRIVATE_0:
4618 	case SIOCSDRVSPEC:
4619 	case SIOCGDRVSPEC:
4620 		CTX_LOCK(ctx);
4621 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4622 		CTX_UNLOCK(ctx);
4623 		break;
4624 	default:
4625 		err = ether_ioctl(ifp, command, data);
4626 		break;
4627 	}
4628 	if (reinit)
4629 		iflib_if_init(ctx);
4630 	return (err);
4631 }
4632 
4633 static uint64_t
4634 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4635 {
4636 	if_ctx_t ctx = if_getsoftc(ifp);
4637 
4638 	return (IFDI_GET_COUNTER(ctx, cnt));
4639 }
4640 
4641 /*********************************************************************
4642  *
4643  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4644  *
4645  **********************************************************************/
4646 
4647 static void
4648 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4649 {
4650 	if_ctx_t ctx = if_getsoftc(ifp);
4651 
4652 	if ((void *)ctx != arg)
4653 		return;
4654 
4655 	if ((vtag == 0) || (vtag > 4095))
4656 		return;
4657 
4658 	if (iflib_in_detach(ctx))
4659 		return;
4660 
4661 	CTX_LOCK(ctx);
4662 	/* Driver may need all untagged packets to be flushed */
4663 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4664 		iflib_stop(ctx);
4665 	IFDI_VLAN_REGISTER(ctx, vtag);
4666 	/* Re-init to load the changes, if required */
4667 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4668 		iflib_init_locked(ctx);
4669 	CTX_UNLOCK(ctx);
4670 }
4671 
4672 static void
4673 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4674 {
4675 	if_ctx_t ctx = if_getsoftc(ifp);
4676 
4677 	if ((void *)ctx != arg)
4678 		return;
4679 
4680 	if ((vtag == 0) || (vtag > 4095))
4681 		return;
4682 
4683 	CTX_LOCK(ctx);
4684 	/* Driver may need all tagged packets to be flushed */
4685 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4686 		iflib_stop(ctx);
4687 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4688 	/* Re-init to load the changes, if required */
4689 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4690 		iflib_init_locked(ctx);
4691 	CTX_UNLOCK(ctx);
4692 }
4693 
4694 static void
4695 iflib_led_func(void *arg, int onoff)
4696 {
4697 	if_ctx_t ctx = arg;
4698 
4699 	CTX_LOCK(ctx);
4700 	IFDI_LED_FUNC(ctx, onoff);
4701 	CTX_UNLOCK(ctx);
4702 }
4703 
4704 /*********************************************************************
4705  *
4706  *  BUS FUNCTION DEFINITIONS
4707  *
4708  **********************************************************************/
4709 
4710 int
4711 iflib_device_probe(device_t dev)
4712 {
4713 	const pci_vendor_info_t *ent;
4714 	if_shared_ctx_t sctx;
4715 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4716 	uint16_t pci_vendor_id;
4717 
4718 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4719 		return (ENOTSUP);
4720 
4721 	pci_vendor_id = pci_get_vendor(dev);
4722 	pci_device_id = pci_get_device(dev);
4723 	pci_subvendor_id = pci_get_subvendor(dev);
4724 	pci_subdevice_id = pci_get_subdevice(dev);
4725 	pci_rev_id = pci_get_revid(dev);
4726 	if (sctx->isc_parse_devinfo != NULL)
4727 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4728 
4729 	ent = sctx->isc_vendor_info;
4730 	while (ent->pvi_vendor_id != 0) {
4731 		if (pci_vendor_id != ent->pvi_vendor_id) {
4732 			ent++;
4733 			continue;
4734 		}
4735 		if ((pci_device_id == ent->pvi_device_id) &&
4736 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4737 		     (ent->pvi_subvendor_id == 0)) &&
4738 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4739 		     (ent->pvi_subdevice_id == 0)) &&
4740 		    ((pci_rev_id == ent->pvi_rev_id) ||
4741 		     (ent->pvi_rev_id == 0))) {
4742 			device_set_desc_copy(dev, ent->pvi_name);
4743 			/* this needs to be changed to zero if the bus probing code
4744 			 * ever stops re-probing on best match because the sctx
4745 			 * may have its values over written by register calls
4746 			 * in subsequent probes
4747 			 */
4748 			return (BUS_PROBE_DEFAULT);
4749 		}
4750 		ent++;
4751 	}
4752 	return (ENXIO);
4753 }
4754 
4755 int
4756 iflib_device_probe_vendor(device_t dev)
4757 {
4758 	int probe;
4759 
4760 	probe = iflib_device_probe(dev);
4761 	if (probe == BUS_PROBE_DEFAULT)
4762 		return (BUS_PROBE_VENDOR);
4763 	else
4764 		return (probe);
4765 }
4766 
4767 static void
4768 iflib_reset_qvalues(if_ctx_t ctx)
4769 {
4770 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4771 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4772 	device_t dev = ctx->ifc_dev;
4773 	int i;
4774 
4775 	if (ctx->ifc_sysctl_ntxqs != 0)
4776 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4777 	if (ctx->ifc_sysctl_nrxqs != 0)
4778 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4779 
4780 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4781 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4782 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4783 		else
4784 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4785 	}
4786 
4787 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4788 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4789 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4790 		else
4791 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4792 	}
4793 
4794 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4795 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4796 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4797 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4798 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4799 		}
4800 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4801 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4802 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4803 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4804 		}
4805 		if (!powerof2(scctx->isc_nrxd[i])) {
4806 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4807 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4808 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4809 		}
4810 	}
4811 
4812 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4813 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4814 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4815 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4816 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4817 		}
4818 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4819 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4820 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4821 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4822 		}
4823 		if (!powerof2(scctx->isc_ntxd[i])) {
4824 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4825 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4826 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4827 		}
4828 	}
4829 }
4830 
4831 static void
4832 iflib_add_pfil(if_ctx_t ctx)
4833 {
4834 	struct pfil_head *pfil;
4835 	struct pfil_head_args pa;
4836 	iflib_rxq_t rxq;
4837 	int i;
4838 
4839 	pa.pa_version = PFIL_VERSION;
4840 	pa.pa_flags = PFIL_IN;
4841 	pa.pa_type = PFIL_TYPE_ETHERNET;
4842 	pa.pa_headname = if_name(ctx->ifc_ifp);
4843 	pfil = pfil_head_register(&pa);
4844 
4845 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4846 		rxq->pfil = pfil;
4847 	}
4848 }
4849 
4850 static void
4851 iflib_rem_pfil(if_ctx_t ctx)
4852 {
4853 	struct pfil_head *pfil;
4854 	iflib_rxq_t rxq;
4855 	int i;
4856 
4857 	rxq = ctx->ifc_rxqs;
4858 	pfil = rxq->pfil;
4859 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4860 		rxq->pfil = NULL;
4861 	}
4862 	pfil_head_unregister(pfil);
4863 }
4864 
4865 
4866 /*
4867  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4868  * cpuid and wrapping as necessary.
4869  */
4870 static unsigned int
4871 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4872 {
4873 	unsigned int first_valid;
4874 	unsigned int last_valid;
4875 
4876 	/* cpuid should always be in the valid set */
4877 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4878 
4879 	/* valid set should never be empty */
4880 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4881 
4882 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4883 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4884 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4885 	while (n > 0) {
4886 		do {
4887 			cpuid++;
4888 			if (cpuid > last_valid)
4889 				cpuid = first_valid;
4890 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4891 		n--;
4892 	}
4893 
4894 	return (cpuid);
4895 }
4896 
4897 #if defined(SMP) && defined(SCHED_ULE)
4898 extern struct cpu_group *cpu_top;              /* CPU topology */
4899 
4900 static int
4901 find_child_with_core(int cpu, struct cpu_group *grp)
4902 {
4903 	int i;
4904 
4905 	if (grp->cg_children == 0)
4906 		return -1;
4907 
4908 	MPASS(grp->cg_child);
4909 	for (i = 0; i < grp->cg_children; i++) {
4910 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4911 			return i;
4912 	}
4913 
4914 	return -1;
4915 }
4916 
4917 
4918 /*
4919  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4920  * does not distinguish among multiple L2 neighbors if the given CPU has
4921  * more than one (it will always return the same result in that case).
4922  */
4923 static int
4924 find_l2_neighbor(int cpu)
4925 {
4926 	struct cpu_group *grp;
4927 	int i;
4928 
4929 	grp = cpu_top;
4930 	if (grp == NULL)
4931 		return -1;
4932 
4933 	/*
4934 	 * Find the smallest CPU group that contains the given core.
4935 	 */
4936 	i = 0;
4937 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4938 		/*
4939 		 * If the smallest group containing the given CPU has less
4940 		 * than two members, we conclude the given CPU has no
4941 		 * L2 neighbor.
4942 		 */
4943 		if (grp->cg_child[i].cg_count <= 1)
4944 			return (-1);
4945 		grp = &grp->cg_child[i];
4946 	}
4947 
4948 	/* Must share L2. */
4949 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4950 		return -1;
4951 
4952 	/*
4953 	 * Select the first member of the set that isn't the reference
4954 	 * CPU, which at this point is guaranteed to exist.
4955 	 */
4956 	for (i = 0; i < CPU_SETSIZE; i++) {
4957 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4958 			return (i);
4959 	}
4960 
4961 	/* Should never be reached */
4962 	return (-1);
4963 }
4964 
4965 #else
4966 static int
4967 find_l2_neighbor(int cpu)
4968 {
4969 
4970 	return (-1);
4971 }
4972 #endif
4973 
4974 /*
4975  * CPU mapping behaviors
4976  * ---------------------
4977  * 'separate txrx' refers to the separate_txrx sysctl
4978  * 'use logical' refers to the use_logical_cores sysctl
4979  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4980  *
4981  *  separate     use     INTR
4982  *    txrx     logical   CPUS   result
4983  * ---------- --------- ------ ------------------------------------------------
4984  *     -          -       X     RX and TX queues mapped to consecutive physical
4985  *                              cores with RX/TX pairs on same core and excess
4986  *                              of either following
4987  *     -          X       X     RX and TX queues mapped to consecutive cores
4988  *                              of any type with RX/TX pairs on same core and
4989  *                              excess of either following
4990  *     X          -       X     RX and TX queues mapped to consecutive physical
4991  *                              cores; all RX then all TX
4992  *     X          X       X     RX queues mapped to consecutive physical cores
4993  *                              first, then TX queues mapped to L2 neighbor of
4994  *                              the corresponding RX queue if one exists,
4995  *                              otherwise to consecutive physical cores
4996  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4997  *                              any type with RX/TX pairs on same core and excess
4998  *                              of either following
4999  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5000  *                              any type; all RX then all TX
5001  */
5002 static unsigned int
5003 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5004     bool is_tx)
5005 {
5006 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5007 	unsigned int core_index;
5008 
5009 	if (ctx->ifc_sysctl_separate_txrx) {
5010 		/*
5011 		 * When using separate CPUs for TX and RX, the assignment
5012 		 * will always be of a consecutive CPU out of the set of
5013 		 * context CPUs, except for the specific case where the
5014 		 * context CPUs are phsyical cores, the use of logical cores
5015 		 * has been enabled, the assignment is for TX, the TX qid
5016 		 * corresponds to an RX qid, and the CPU assigned to the
5017 		 * corresponding RX queue has an L2 neighbor.
5018 		 */
5019 		if (ctx->ifc_sysctl_use_logical_cores &&
5020 		    ctx->ifc_cpus_are_physical_cores &&
5021 		    is_tx && qid < scctx->isc_nrxqsets) {
5022 			int l2_neighbor;
5023 			unsigned int rx_cpuid;
5024 
5025 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5026 			l2_neighbor = find_l2_neighbor(rx_cpuid);
5027 			if (l2_neighbor != -1) {
5028 				return (l2_neighbor);
5029 			}
5030 			/*
5031 			 * ... else fall through to the normal
5032 			 * consecutive-after-RX assignment scheme.
5033 			 *
5034 			 * Note that we are assuming that all RX queue CPUs
5035 			 * have an L2 neighbor, or all do not.  If a mixed
5036 			 * scenario is possible, we will have to keep track
5037 			 * separately of how many queues prior to this one
5038 			 * were not able to be assigned to an L2 neighbor.
5039 			 */
5040 		}
5041 		if (is_tx)
5042 			core_index = scctx->isc_nrxqsets + qid;
5043 		else
5044 			core_index = qid;
5045 	} else {
5046 		core_index = qid;
5047 	}
5048 
5049 	return (cpuid_advance(ctx, base_cpuid, core_index));
5050 }
5051 
5052 static uint16_t
5053 get_ctx_core_offset(if_ctx_t ctx)
5054 {
5055 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5056 	struct cpu_offset *op;
5057 	cpuset_t assigned_cpus;
5058 	unsigned int cores_consumed;
5059 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5060 	unsigned int first_valid;
5061 	unsigned int last_valid;
5062 	unsigned int i;
5063 
5064 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5065 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5066 
5067 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5068 		/*
5069 		 * Align the user-chosen base CPU ID to the next valid CPU
5070 		 * for this device.  If the chosen base CPU ID is smaller
5071 		 * than the first valid CPU or larger than the last valid
5072 		 * CPU, we assume the user does not know what the valid
5073 		 * range is for this device and is thinking in terms of a
5074 		 * zero-based reference frame, and so we shift the given
5075 		 * value into the valid range (and wrap accordingly) so the
5076 		 * intent is translated to the proper frame of reference.
5077 		 * If the base CPU ID is within the valid first/last, but
5078 		 * does not correspond to a valid CPU, it is advanced to the
5079 		 * next valid CPU (wrapping if necessary).
5080 		 */
5081 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5082 			/* shift from zero-based to first_valid-based */
5083 			base_cpuid += first_valid;
5084 			/* wrap to range [first_valid, last_valid] */
5085 			base_cpuid = (base_cpuid - first_valid) %
5086 			    (last_valid - first_valid + 1);
5087 		}
5088 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5089 			/*
5090 			 * base_cpuid is in [first_valid, last_valid], but
5091 			 * not a member of the valid set.  In this case,
5092 			 * there will always be a member of the valid set
5093 			 * with a CPU ID that is greater than base_cpuid,
5094 			 * and we simply advance to it.
5095 			 */
5096 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5097 				base_cpuid++;
5098 		}
5099 		return (base_cpuid);
5100 	}
5101 
5102 	/*
5103 	 * Determine how many cores will be consumed by performing the CPU
5104 	 * assignments and counting how many of the assigned CPUs correspond
5105 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5106 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5107 	 * the set of context CPUs.
5108 	 *
5109 	 * Note not all assigned CPUs will be in the set of context CPUs
5110 	 * when separate CPUs are being allocated to TX and RX queues,
5111 	 * assignment to logical cores has been enabled, the set of context
5112 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5113 	 * neighbors of CPUs that RX queues have been mapped to - in this
5114 	 * case we do only want to count how many CPUs in the set of context
5115 	 * CPUs have been consumed, as that determines the next CPU in that
5116 	 * set to start allocating at for the next device for which
5117 	 * core_offset is not set.
5118 	 */
5119 	CPU_ZERO(&assigned_cpus);
5120 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5121 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5122 		    &assigned_cpus);
5123 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5124 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5125 		    &assigned_cpus);
5126 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5127 	cores_consumed = CPU_COUNT(&assigned_cpus);
5128 
5129 	mtx_lock(&cpu_offset_mtx);
5130 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5131 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5132 			base_cpuid = op->next_cpuid;
5133 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5134 			    cores_consumed);
5135 			MPASS(op->refcount < UINT_MAX);
5136 			op->refcount++;
5137 			break;
5138 		}
5139 	}
5140 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5141 		base_cpuid = first_valid;
5142 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5143 		    M_NOWAIT | M_ZERO);
5144 		if (op == NULL) {
5145 			device_printf(ctx->ifc_dev,
5146 			    "allocation for cpu offset failed.\n");
5147 		} else {
5148 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5149 			    cores_consumed);
5150 			op->refcount = 1;
5151 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5152 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5153 		}
5154 	}
5155 	mtx_unlock(&cpu_offset_mtx);
5156 
5157 	return (base_cpuid);
5158 }
5159 
5160 static void
5161 unref_ctx_core_offset(if_ctx_t ctx)
5162 {
5163 	struct cpu_offset *op, *top;
5164 
5165 	mtx_lock(&cpu_offset_mtx);
5166 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5167 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5168 			MPASS(op->refcount > 0);
5169 			op->refcount--;
5170 			if (op->refcount == 0) {
5171 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5172 				free(op, M_IFLIB);
5173 			}
5174 			break;
5175 		}
5176 	}
5177 	mtx_unlock(&cpu_offset_mtx);
5178 }
5179 
5180 int
5181 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5182 {
5183 	if_ctx_t ctx;
5184 	if_t ifp;
5185 	if_softc_ctx_t scctx;
5186 	kobjop_desc_t kobj_desc;
5187 	kobj_method_t *kobj_method;
5188 	int err, msix, rid;
5189 	int num_txd, num_rxd;
5190 
5191 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5192 
5193 	if (sc == NULL) {
5194 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5195 		device_set_softc(dev, ctx);
5196 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5197 	}
5198 
5199 	ctx->ifc_sctx = sctx;
5200 	ctx->ifc_dev = dev;
5201 	ctx->ifc_softc = sc;
5202 
5203 	if ((err = iflib_register(ctx)) != 0) {
5204 		device_printf(dev, "iflib_register failed %d\n", err);
5205 		goto fail_ctx_free;
5206 	}
5207 	iflib_add_device_sysctl_pre(ctx);
5208 
5209 	scctx = &ctx->ifc_softc_ctx;
5210 	ifp = ctx->ifc_ifp;
5211 
5212 	iflib_reset_qvalues(ctx);
5213 	IFNET_WLOCK();
5214 	CTX_LOCK(ctx);
5215 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5216 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5217 		goto fail_unlock;
5218 	}
5219 	_iflib_pre_assert(scctx);
5220 	ctx->ifc_txrx = *scctx->isc_txrx;
5221 
5222 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5223 
5224 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5225 		ctx->ifc_mediap = scctx->isc_media;
5226 
5227 #ifdef INVARIANTS
5228 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5229 		MPASS(scctx->isc_tx_csum_flags);
5230 #endif
5231 
5232 	if_setcapabilities(ifp,
5233 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5234 	if_setcapenable(ifp,
5235 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5236 
5237 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5238 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5239 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5240 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5241 
5242 	num_txd = iflib_num_tx_descs(ctx);
5243 	num_rxd = iflib_num_rx_descs(ctx);
5244 
5245 	/* XXX change for per-queue sizes */
5246 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5247 	    num_txd, num_rxd);
5248 
5249 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5250 		scctx->isc_tx_nsegments = max(1, num_txd /
5251 		    MAX_SINGLE_PACKET_FRACTION);
5252 	if (scctx->isc_tx_tso_segments_max > num_txd /
5253 	    MAX_SINGLE_PACKET_FRACTION)
5254 		scctx->isc_tx_tso_segments_max = max(1,
5255 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5256 
5257 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5258 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5259 		/*
5260 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5261 		 * but some MACs do.
5262 		 */
5263 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5264 		    IP_MAXPACKET));
5265 		/*
5266 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5267 		 * into account.  In the worst case, each of these calls will
5268 		 * add another mbuf and, thus, the requirement for another DMA
5269 		 * segment.  So for best performance, it doesn't make sense to
5270 		 * advertize a maximum of TSO segments that typically will
5271 		 * require defragmentation in iflib_encap().
5272 		 */
5273 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5274 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5275 	}
5276 	if (scctx->isc_rss_table_size == 0)
5277 		scctx->isc_rss_table_size = 64;
5278 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5279 
5280 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5281 	/* XXX format name */
5282 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5283 	    NULL, NULL, "admin");
5284 
5285 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5286 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5287 		device_printf(dev, "Unable to fetch CPU list\n");
5288 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5289 		ctx->ifc_cpus_are_physical_cores = false;
5290 	} else
5291 		ctx->ifc_cpus_are_physical_cores = true;
5292 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5293 
5294 	/*
5295 	** Now set up MSI or MSI-X, should return us the number of supported
5296 	** vectors (will be 1 for a legacy interrupt and MSI).
5297 	*/
5298 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5299 		msix = scctx->isc_vectors;
5300 	} else if (scctx->isc_msix_bar != 0)
5301 	       /*
5302 		* The simple fact that isc_msix_bar is not 0 does not mean we
5303 		* we have a good value there that is known to work.
5304 		*/
5305 		msix = iflib_msix_init(ctx);
5306 	else {
5307 		scctx->isc_vectors = 1;
5308 		scctx->isc_ntxqsets = 1;
5309 		scctx->isc_nrxqsets = 1;
5310 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5311 		msix = 0;
5312 	}
5313 	/* Get memory for the station queues */
5314 	if ((err = iflib_queues_alloc(ctx))) {
5315 		device_printf(dev, "Unable to allocate queue memory\n");
5316 		goto fail_intr_free;
5317 	}
5318 
5319 	if ((err = iflib_qset_structures_setup(ctx)))
5320 		goto fail_queues;
5321 
5322 	/*
5323 	 * Now that we know how many queues there are, get the core offset.
5324 	 */
5325 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5326 
5327 	if (msix > 1) {
5328 		/*
5329 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5330 		 * aren't the default NULL implementation.
5331 		 */
5332 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5333 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5334 		    kobj_desc);
5335 		if (kobj_method == &kobj_desc->deflt) {
5336 			device_printf(dev,
5337 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5338 			err = EOPNOTSUPP;
5339 			goto fail_queues;
5340 		}
5341 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5342 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5343 		    kobj_desc);
5344 		if (kobj_method == &kobj_desc->deflt) {
5345 			device_printf(dev,
5346 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5347 			err = EOPNOTSUPP;
5348 			goto fail_queues;
5349 		}
5350 
5351 		/*
5352 		 * Assign the MSI-X vectors.
5353 		 * Note that the default NULL ifdi_msix_intr_assign method will
5354 		 * fail here, too.
5355 		 */
5356 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5357 		if (err != 0) {
5358 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5359 			    err);
5360 			goto fail_queues;
5361 		}
5362 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5363 		rid = 0;
5364 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5365 			MPASS(msix == 1);
5366 			rid = 1;
5367 		}
5368 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5369 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5370 			goto fail_queues;
5371 		}
5372 	} else {
5373 		device_printf(dev,
5374 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5375 		err = ENODEV;
5376 		goto fail_queues;
5377 	}
5378 
5379 	/*
5380 	 * It prevents a double-locking panic with iflib_media_status when
5381 	 * the driver loads.
5382 	 */
5383 	CTX_UNLOCK(ctx);
5384 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5385 	CTX_LOCK(ctx);
5386 
5387 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5388 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5389 		goto fail_detach;
5390 	}
5391 
5392 	/*
5393 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5394 	 * This must appear after the call to ether_ifattach() because
5395 	 * ether_ifattach() sets if_hdrlen to the default value.
5396 	 */
5397 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5398 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5399 
5400 	if ((err = iflib_netmap_attach(ctx))) {
5401 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5402 		goto fail_detach;
5403 	}
5404 	*ctxp = ctx;
5405 
5406 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5407 
5408 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5409 	iflib_add_device_sysctl_post(ctx);
5410 	iflib_add_pfil(ctx);
5411 	ctx->ifc_flags |= IFC_INIT_DONE;
5412 	CTX_UNLOCK(ctx);
5413 	IFNET_WUNLOCK();
5414 
5415 	return (0);
5416 
5417 fail_detach:
5418 	ether_ifdetach(ctx->ifc_ifp);
5419 fail_queues:
5420 	iflib_tqg_detach(ctx);
5421 	iflib_tx_structures_free(ctx);
5422 	iflib_rx_structures_free(ctx);
5423 	IFDI_DETACH(ctx);
5424 	IFDI_QUEUES_FREE(ctx);
5425 fail_intr_free:
5426 	iflib_free_intr_mem(ctx);
5427 fail_unlock:
5428 	CTX_UNLOCK(ctx);
5429 	IFNET_WUNLOCK();
5430 	iflib_deregister(ctx);
5431 fail_ctx_free:
5432 	device_set_softc(ctx->ifc_dev, NULL);
5433         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5434                 free(ctx->ifc_softc, M_IFLIB);
5435         free(ctx, M_IFLIB);
5436 	return (err);
5437 }
5438 
5439 int
5440 iflib_device_attach(device_t dev)
5441 {
5442 	if_ctx_t ctx;
5443 	if_shared_ctx_t sctx;
5444 
5445 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5446 		return (ENOTSUP);
5447 
5448 	pci_enable_busmaster(dev);
5449 
5450 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5451 }
5452 
5453 int
5454 iflib_device_deregister(if_ctx_t ctx)
5455 {
5456 	if_t ifp = ctx->ifc_ifp;
5457 	device_t dev = ctx->ifc_dev;
5458 
5459 	/* Make sure VLANS are not using driver */
5460 	if (if_vlantrunkinuse(ifp)) {
5461 		device_printf(dev, "Vlan in use, detach first\n");
5462 		return (EBUSY);
5463 	}
5464 #ifdef PCI_IOV
5465 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5466 		device_printf(dev, "SR-IOV in use; detach first.\n");
5467 		return (EBUSY);
5468 	}
5469 #endif
5470 
5471 	STATE_LOCK(ctx);
5472 	ctx->ifc_flags |= IFC_IN_DETACH;
5473 	STATE_UNLOCK(ctx);
5474 
5475 	/* Unregister VLAN handlers before calling iflib_stop() */
5476 	iflib_unregister_vlan_handlers(ctx);
5477 
5478 	iflib_netmap_detach(ifp);
5479 	ether_ifdetach(ifp);
5480 
5481 	CTX_LOCK(ctx);
5482 	iflib_stop(ctx);
5483 	CTX_UNLOCK(ctx);
5484 
5485 	iflib_rem_pfil(ctx);
5486 	if (ctx->ifc_led_dev != NULL)
5487 		led_destroy(ctx->ifc_led_dev);
5488 
5489 	iflib_tqg_detach(ctx);
5490 	iflib_tx_structures_free(ctx);
5491 	iflib_rx_structures_free(ctx);
5492 
5493 	CTX_LOCK(ctx);
5494 	IFDI_DETACH(ctx);
5495 	IFDI_QUEUES_FREE(ctx);
5496 	CTX_UNLOCK(ctx);
5497 
5498 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5499 	iflib_free_intr_mem(ctx);
5500 
5501 	bus_generic_detach(dev);
5502 
5503 	iflib_deregister(ctx);
5504 
5505 	device_set_softc(ctx->ifc_dev, NULL);
5506 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5507 		free(ctx->ifc_softc, M_IFLIB);
5508 	unref_ctx_core_offset(ctx);
5509 	free(ctx, M_IFLIB);
5510 	return (0);
5511 }
5512 
5513 static void
5514 iflib_tqg_detach(if_ctx_t ctx)
5515 {
5516 	iflib_txq_t txq;
5517 	iflib_rxq_t rxq;
5518 	int i;
5519 	struct taskqgroup *tqg;
5520 
5521 	/* XXX drain any dependent tasks */
5522 	tqg = qgroup_if_io_tqg;
5523 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5524 		callout_drain(&txq->ift_timer);
5525 #ifdef DEV_NETMAP
5526 		callout_drain(&txq->ift_netmap_timer);
5527 #endif /* DEV_NETMAP */
5528 		if (txq->ift_task.gt_uniq != NULL)
5529 			taskqgroup_detach(tqg, &txq->ift_task);
5530 	}
5531 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5532 		if (rxq->ifr_task.gt_uniq != NULL)
5533 			taskqgroup_detach(tqg, &rxq->ifr_task);
5534 	}
5535 	tqg = qgroup_if_config_tqg;
5536 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5537 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5538 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5539 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5540 }
5541 
5542 static void
5543 iflib_free_intr_mem(if_ctx_t ctx)
5544 {
5545 
5546 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5547 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5548 	}
5549 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5550 		pci_release_msi(ctx->ifc_dev);
5551 	}
5552 	if (ctx->ifc_msix_mem != NULL) {
5553 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5554 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5555 		ctx->ifc_msix_mem = NULL;
5556 	}
5557 }
5558 
5559 int
5560 iflib_device_detach(device_t dev)
5561 {
5562 	if_ctx_t ctx = device_get_softc(dev);
5563 
5564 	return (iflib_device_deregister(ctx));
5565 }
5566 
5567 int
5568 iflib_device_suspend(device_t dev)
5569 {
5570 	if_ctx_t ctx = device_get_softc(dev);
5571 
5572 	CTX_LOCK(ctx);
5573 	IFDI_SUSPEND(ctx);
5574 	CTX_UNLOCK(ctx);
5575 
5576 	return bus_generic_suspend(dev);
5577 }
5578 int
5579 iflib_device_shutdown(device_t dev)
5580 {
5581 	if_ctx_t ctx = device_get_softc(dev);
5582 
5583 	CTX_LOCK(ctx);
5584 	IFDI_SHUTDOWN(ctx);
5585 	CTX_UNLOCK(ctx);
5586 
5587 	return bus_generic_suspend(dev);
5588 }
5589 
5590 int
5591 iflib_device_resume(device_t dev)
5592 {
5593 	if_ctx_t ctx = device_get_softc(dev);
5594 	iflib_txq_t txq = ctx->ifc_txqs;
5595 
5596 	CTX_LOCK(ctx);
5597 	IFDI_RESUME(ctx);
5598 	iflib_if_init_locked(ctx);
5599 	CTX_UNLOCK(ctx);
5600 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5601 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5602 
5603 	return (bus_generic_resume(dev));
5604 }
5605 
5606 int
5607 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5608 {
5609 	int error;
5610 	if_ctx_t ctx = device_get_softc(dev);
5611 
5612 	CTX_LOCK(ctx);
5613 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5614 	CTX_UNLOCK(ctx);
5615 
5616 	return (error);
5617 }
5618 
5619 void
5620 iflib_device_iov_uninit(device_t dev)
5621 {
5622 	if_ctx_t ctx = device_get_softc(dev);
5623 
5624 	CTX_LOCK(ctx);
5625 	IFDI_IOV_UNINIT(ctx);
5626 	CTX_UNLOCK(ctx);
5627 }
5628 
5629 int
5630 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5631 {
5632 	int error;
5633 	if_ctx_t ctx = device_get_softc(dev);
5634 
5635 	CTX_LOCK(ctx);
5636 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5637 	CTX_UNLOCK(ctx);
5638 
5639 	return (error);
5640 }
5641 
5642 /*********************************************************************
5643  *
5644  *  MODULE FUNCTION DEFINITIONS
5645  *
5646  **********************************************************************/
5647 
5648 /*
5649  * - Start a fast taskqueue thread for each core
5650  * - Start a taskqueue for control operations
5651  */
5652 static int
5653 iflib_module_init(void)
5654 {
5655 	iflib_timer_default = hz / 2;
5656 	return (0);
5657 }
5658 
5659 static int
5660 iflib_module_event_handler(module_t mod, int what, void *arg)
5661 {
5662 	int err;
5663 
5664 	switch (what) {
5665 	case MOD_LOAD:
5666 		if ((err = iflib_module_init()) != 0)
5667 			return (err);
5668 		break;
5669 	case MOD_UNLOAD:
5670 		return (EBUSY);
5671 	default:
5672 		return (EOPNOTSUPP);
5673 	}
5674 
5675 	return (0);
5676 }
5677 
5678 /*********************************************************************
5679  *
5680  *  PUBLIC FUNCTION DEFINITIONS
5681  *     ordered as in iflib.h
5682  *
5683  **********************************************************************/
5684 
5685 static void
5686 _iflib_assert(if_shared_ctx_t sctx)
5687 {
5688 	int i;
5689 
5690 	MPASS(sctx->isc_tx_maxsize);
5691 	MPASS(sctx->isc_tx_maxsegsize);
5692 
5693 	MPASS(sctx->isc_rx_maxsize);
5694 	MPASS(sctx->isc_rx_nsegments);
5695 	MPASS(sctx->isc_rx_maxsegsize);
5696 
5697 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5698 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5699 		MPASS(sctx->isc_nrxd_min[i]);
5700 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5701 		MPASS(sctx->isc_nrxd_max[i]);
5702 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5703 		MPASS(sctx->isc_nrxd_default[i]);
5704 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5705 	}
5706 
5707 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5708 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5709 		MPASS(sctx->isc_ntxd_min[i]);
5710 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5711 		MPASS(sctx->isc_ntxd_max[i]);
5712 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5713 		MPASS(sctx->isc_ntxd_default[i]);
5714 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5715 	}
5716 }
5717 
5718 static void
5719 _iflib_pre_assert(if_softc_ctx_t scctx)
5720 {
5721 
5722 	MPASS(scctx->isc_txrx->ift_txd_encap);
5723 	MPASS(scctx->isc_txrx->ift_txd_flush);
5724 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5725 	MPASS(scctx->isc_txrx->ift_rxd_available);
5726 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5727 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5728 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5729 }
5730 
5731 static int
5732 iflib_register(if_ctx_t ctx)
5733 {
5734 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5735 	driver_t *driver = sctx->isc_driver;
5736 	device_t dev = ctx->ifc_dev;
5737 	if_t ifp;
5738 
5739 	_iflib_assert(sctx);
5740 
5741 	CTX_LOCK_INIT(ctx);
5742 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5743 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5744 	if (ifp == NULL) {
5745 		device_printf(dev, "can not allocate ifnet structure\n");
5746 		return (ENOMEM);
5747 	}
5748 
5749 	/*
5750 	 * Initialize our context's device specific methods
5751 	 */
5752 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5753 	kobj_class_compile((kobj_class_t) driver);
5754 
5755 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5756 	if_setsoftc(ifp, ctx);
5757 	if_setdev(ifp, dev);
5758 	if_setinitfn(ifp, iflib_if_init);
5759 	if_setioctlfn(ifp, iflib_if_ioctl);
5760 #ifdef ALTQ
5761 	if_setstartfn(ifp, iflib_altq_if_start);
5762 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5763 	if_setsendqready(ifp);
5764 #else
5765 	if_settransmitfn(ifp, iflib_if_transmit);
5766 #endif
5767 	if_setqflushfn(ifp, iflib_if_qflush);
5768 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5769 	ctx->ifc_vlan_attach_event =
5770 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5771 							  EVENTHANDLER_PRI_FIRST);
5772 	ctx->ifc_vlan_detach_event =
5773 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5774 							  EVENTHANDLER_PRI_FIRST);
5775 
5776 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5777 		ctx->ifc_mediap = &ctx->ifc_media;
5778 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5779 		    iflib_media_change, iflib_media_status);
5780 	}
5781 	return (0);
5782 }
5783 
5784 static void
5785 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5786 {
5787 	/* Unregister VLAN events */
5788 	if (ctx->ifc_vlan_attach_event != NULL) {
5789 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5790 		ctx->ifc_vlan_attach_event = NULL;
5791 	}
5792 	if (ctx->ifc_vlan_detach_event != NULL) {
5793 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5794 		ctx->ifc_vlan_detach_event = NULL;
5795 	}
5796 
5797 }
5798 
5799 static void
5800 iflib_deregister(if_ctx_t ctx)
5801 {
5802 	if_t ifp = ctx->ifc_ifp;
5803 
5804 	/* Remove all media */
5805 	ifmedia_removeall(&ctx->ifc_media);
5806 
5807 	/* Ensure that VLAN event handlers are unregistered */
5808 	iflib_unregister_vlan_handlers(ctx);
5809 
5810 	/* Release kobject reference */
5811 	kobj_delete((kobj_t) ctx, NULL);
5812 
5813 	/* Free the ifnet structure */
5814 	if_free(ifp);
5815 
5816 	STATE_LOCK_DESTROY(ctx);
5817 
5818 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5819 	CTX_LOCK_DESTROY(ctx);
5820 }
5821 
5822 static int
5823 iflib_queues_alloc(if_ctx_t ctx)
5824 {
5825 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5826 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5827 	device_t dev = ctx->ifc_dev;
5828 	int nrxqsets = scctx->isc_nrxqsets;
5829 	int ntxqsets = scctx->isc_ntxqsets;
5830 	iflib_txq_t txq;
5831 	iflib_rxq_t rxq;
5832 	iflib_fl_t fl = NULL;
5833 	int i, j, cpu, err, txconf, rxconf;
5834 	iflib_dma_info_t ifdip;
5835 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5836 	uint32_t *txqsizes = scctx->isc_txqsizes;
5837 	uint8_t nrxqs = sctx->isc_nrxqs;
5838 	uint8_t ntxqs = sctx->isc_ntxqs;
5839 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5840 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5841 	caddr_t *vaddrs;
5842 	uint64_t *paddrs;
5843 
5844 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5845 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5846 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5847            ("there must be at least a rxq for each free list"));
5848 
5849 	/* Allocate the TX ring struct memory */
5850 	if (!(ctx->ifc_txqs =
5851 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5852 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5853 		device_printf(dev, "Unable to allocate TX ring memory\n");
5854 		err = ENOMEM;
5855 		goto fail;
5856 	}
5857 
5858 	/* Now allocate the RX */
5859 	if (!(ctx->ifc_rxqs =
5860 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5861 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5862 		device_printf(dev, "Unable to allocate RX ring memory\n");
5863 		err = ENOMEM;
5864 		goto rx_fail;
5865 	}
5866 
5867 	txq = ctx->ifc_txqs;
5868 	rxq = ctx->ifc_rxqs;
5869 
5870 	/*
5871 	 * XXX handle allocation failure
5872 	 */
5873 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5874 		/* Set up some basics */
5875 
5876 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5877 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5878 			device_printf(dev,
5879 			    "Unable to allocate TX DMA info memory\n");
5880 			err = ENOMEM;
5881 			goto err_tx_desc;
5882 		}
5883 		txq->ift_ifdi = ifdip;
5884 		for (j = 0; j < ntxqs; j++, ifdip++) {
5885 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5886 				device_printf(dev,
5887 				    "Unable to allocate TX descriptors\n");
5888 				err = ENOMEM;
5889 				goto err_tx_desc;
5890 			}
5891 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5892 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5893 		}
5894 		txq->ift_ctx = ctx;
5895 		txq->ift_id = i;
5896 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5897 			txq->ift_br_offset = 1;
5898 		} else {
5899 			txq->ift_br_offset = 0;
5900 		}
5901 
5902 		if (iflib_txsd_alloc(txq)) {
5903 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5904 			err = ENOMEM;
5905 			goto err_tx_desc;
5906 		}
5907 
5908 		/* Initialize the TX lock */
5909 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5910 		    device_get_nameunit(dev), txq->ift_id);
5911 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5912 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5913 		txq->ift_timer.c_cpu = cpu;
5914 #ifdef DEV_NETMAP
5915 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5916 		txq->ift_netmap_timer.c_cpu = cpu;
5917 #endif /* DEV_NETMAP */
5918 
5919 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5920 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5921 		if (err) {
5922 			/* XXX free any allocated rings */
5923 			device_printf(dev, "Unable to allocate buf_ring\n");
5924 			goto err_tx_desc;
5925 		}
5926 	}
5927 
5928 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5929 		/* Set up some basics */
5930 		callout_init(&rxq->ifr_watchdog, 1);
5931 
5932 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5933 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5934 			device_printf(dev,
5935 			    "Unable to allocate RX DMA info memory\n");
5936 			err = ENOMEM;
5937 			goto err_tx_desc;
5938 		}
5939 
5940 		rxq->ifr_ifdi = ifdip;
5941 		/* XXX this needs to be changed if #rx queues != #tx queues */
5942 		rxq->ifr_ntxqirq = 1;
5943 		rxq->ifr_txqid[0] = i;
5944 		for (j = 0; j < nrxqs; j++, ifdip++) {
5945 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5946 				device_printf(dev,
5947 				    "Unable to allocate RX descriptors\n");
5948 				err = ENOMEM;
5949 				goto err_tx_desc;
5950 			}
5951 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5952 		}
5953 		rxq->ifr_ctx = ctx;
5954 		rxq->ifr_id = i;
5955 		rxq->ifr_fl_offset = fl_offset;
5956 		rxq->ifr_nfl = nfree_lists;
5957 		if (!(fl =
5958 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5959 			device_printf(dev, "Unable to allocate free list memory\n");
5960 			err = ENOMEM;
5961 			goto err_tx_desc;
5962 		}
5963 		rxq->ifr_fl = fl;
5964 		for (j = 0; j < nfree_lists; j++) {
5965 			fl[j].ifl_rxq = rxq;
5966 			fl[j].ifl_id = j;
5967 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5968 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5969 		}
5970 		/* Allocate receive buffers for the ring */
5971 		if (iflib_rxsd_alloc(rxq)) {
5972 			device_printf(dev,
5973 			    "Critical Failure setting up receive buffers\n");
5974 			err = ENOMEM;
5975 			goto err_rx_desc;
5976 		}
5977 
5978 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5979 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5980 			    M_WAITOK);
5981 	}
5982 
5983 	/* TXQs */
5984 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5985 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5986 	for (i = 0; i < ntxqsets; i++) {
5987 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5988 
5989 		for (j = 0; j < ntxqs; j++, di++) {
5990 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5991 			paddrs[i*ntxqs + j] = di->idi_paddr;
5992 		}
5993 	}
5994 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5995 		device_printf(ctx->ifc_dev,
5996 		    "Unable to allocate device TX queue\n");
5997 		iflib_tx_structures_free(ctx);
5998 		free(vaddrs, M_IFLIB);
5999 		free(paddrs, M_IFLIB);
6000 		goto err_rx_desc;
6001 	}
6002 	free(vaddrs, M_IFLIB);
6003 	free(paddrs, M_IFLIB);
6004 
6005 	/* RXQs */
6006 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6007 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6008 	for (i = 0; i < nrxqsets; i++) {
6009 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6010 
6011 		for (j = 0; j < nrxqs; j++, di++) {
6012 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6013 			paddrs[i*nrxqs + j] = di->idi_paddr;
6014 		}
6015 	}
6016 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6017 		device_printf(ctx->ifc_dev,
6018 		    "Unable to allocate device RX queue\n");
6019 		iflib_tx_structures_free(ctx);
6020 		free(vaddrs, M_IFLIB);
6021 		free(paddrs, M_IFLIB);
6022 		goto err_rx_desc;
6023 	}
6024 	free(vaddrs, M_IFLIB);
6025 	free(paddrs, M_IFLIB);
6026 
6027 	return (0);
6028 
6029 /* XXX handle allocation failure changes */
6030 err_rx_desc:
6031 err_tx_desc:
6032 rx_fail:
6033 	if (ctx->ifc_rxqs != NULL)
6034 		free(ctx->ifc_rxqs, M_IFLIB);
6035 	ctx->ifc_rxqs = NULL;
6036 	if (ctx->ifc_txqs != NULL)
6037 		free(ctx->ifc_txqs, M_IFLIB);
6038 	ctx->ifc_txqs = NULL;
6039 fail:
6040 	return (err);
6041 }
6042 
6043 static int
6044 iflib_tx_structures_setup(if_ctx_t ctx)
6045 {
6046 	iflib_txq_t txq = ctx->ifc_txqs;
6047 	int i;
6048 
6049 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6050 		iflib_txq_setup(txq);
6051 
6052 	return (0);
6053 }
6054 
6055 static void
6056 iflib_tx_structures_free(if_ctx_t ctx)
6057 {
6058 	iflib_txq_t txq = ctx->ifc_txqs;
6059 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6060 	int i, j;
6061 
6062 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6063 		for (j = 0; j < sctx->isc_ntxqs; j++)
6064 			iflib_dma_free(&txq->ift_ifdi[j]);
6065 		iflib_txq_destroy(txq);
6066 	}
6067 	free(ctx->ifc_txqs, M_IFLIB);
6068 	ctx->ifc_txqs = NULL;
6069 }
6070 
6071 /*********************************************************************
6072  *
6073  *  Initialize all receive rings.
6074  *
6075  **********************************************************************/
6076 static int
6077 iflib_rx_structures_setup(if_ctx_t ctx)
6078 {
6079 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6080 	int q;
6081 #if defined(INET6) || defined(INET)
6082 	int err, i;
6083 #endif
6084 
6085 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6086 #if defined(INET6) || defined(INET)
6087 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6088 		    TCP_LRO_ENTRIES, min(1024,
6089 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6090 		if (err != 0) {
6091 			device_printf(ctx->ifc_dev,
6092 			    "LRO Initialization failed!\n");
6093 			goto fail;
6094 		}
6095 #endif
6096 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6097 	}
6098 	return (0);
6099 #if defined(INET6) || defined(INET)
6100 fail:
6101 	/*
6102 	 * Free LRO resources allocated so far, we will only handle
6103 	 * the rings that completed, the failing case will have
6104 	 * cleaned up for itself.  'q' failed, so its the terminus.
6105 	 */
6106 	rxq = ctx->ifc_rxqs;
6107 	for (i = 0; i < q; ++i, rxq++) {
6108 		tcp_lro_free(&rxq->ifr_lc);
6109 	}
6110 	return (err);
6111 #endif
6112 }
6113 
6114 /*********************************************************************
6115  *
6116  *  Free all receive rings.
6117  *
6118  **********************************************************************/
6119 static void
6120 iflib_rx_structures_free(if_ctx_t ctx)
6121 {
6122 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6123 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6124 	int i, j;
6125 
6126 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6127 		for (j = 0; j < sctx->isc_nrxqs; j++)
6128 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6129 		iflib_rx_sds_free(rxq);
6130 #if defined(INET6) || defined(INET)
6131 		tcp_lro_free(&rxq->ifr_lc);
6132 #endif
6133 	}
6134 	free(ctx->ifc_rxqs, M_IFLIB);
6135 	ctx->ifc_rxqs = NULL;
6136 }
6137 
6138 static int
6139 iflib_qset_structures_setup(if_ctx_t ctx)
6140 {
6141 	int err;
6142 
6143 	/*
6144 	 * It is expected that the caller takes care of freeing queues if this
6145 	 * fails.
6146 	 */
6147 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6148 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6149 		return (err);
6150 	}
6151 
6152 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6153 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6154 
6155 	return (err);
6156 }
6157 
6158 int
6159 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6160 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6161 {
6162 
6163 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6164 }
6165 
6166 /* Just to avoid copy/paste */
6167 static inline int
6168 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6169     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6170     const char *name)
6171 {
6172 	device_t dev;
6173 	unsigned int base_cpuid, cpuid;
6174 	int err;
6175 
6176 	dev = ctx->ifc_dev;
6177 	base_cpuid = ctx->ifc_sysctl_core_offset;
6178 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6179 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6180 	    irq ? irq->ii_res : NULL, name);
6181 	if (err) {
6182 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6183 		return (err);
6184 	}
6185 #ifdef notyet
6186 	if (cpuid > ctx->ifc_cpuid_highest)
6187 		ctx->ifc_cpuid_highest = cpuid;
6188 #endif
6189 	return (0);
6190 }
6191 
6192 int
6193 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6194 			iflib_intr_type_t type, driver_filter_t *filter,
6195 			void *filter_arg, int qid, const char *name)
6196 {
6197 	device_t dev;
6198 	struct grouptask *gtask;
6199 	struct taskqgroup *tqg;
6200 	iflib_filter_info_t info;
6201 	gtask_fn_t *fn;
6202 	int tqrid, err;
6203 	driver_filter_t *intr_fast;
6204 	void *q;
6205 
6206 	info = &ctx->ifc_filter_info;
6207 	tqrid = rid;
6208 
6209 	switch (type) {
6210 	/* XXX merge tx/rx for netmap? */
6211 	case IFLIB_INTR_TX:
6212 		q = &ctx->ifc_txqs[qid];
6213 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6214 		gtask = &ctx->ifc_txqs[qid].ift_task;
6215 		tqg = qgroup_if_io_tqg;
6216 		fn = _task_fn_tx;
6217 		intr_fast = iflib_fast_intr;
6218 		GROUPTASK_INIT(gtask, 0, fn, q);
6219 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6220 		break;
6221 	case IFLIB_INTR_RX:
6222 		q = &ctx->ifc_rxqs[qid];
6223 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6224 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6225 		tqg = qgroup_if_io_tqg;
6226 		fn = _task_fn_rx;
6227 		intr_fast = iflib_fast_intr;
6228 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6229 		break;
6230 	case IFLIB_INTR_RXTX:
6231 		q = &ctx->ifc_rxqs[qid];
6232 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6233 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6234 		tqg = qgroup_if_io_tqg;
6235 		fn = _task_fn_rx;
6236 		intr_fast = iflib_fast_intr_rxtx;
6237 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6238 		break;
6239 	case IFLIB_INTR_ADMIN:
6240 		q = ctx;
6241 		tqrid = -1;
6242 		info = &ctx->ifc_filter_info;
6243 		gtask = &ctx->ifc_admin_task;
6244 		tqg = qgroup_if_config_tqg;
6245 		fn = _task_fn_admin;
6246 		intr_fast = iflib_fast_intr_ctx;
6247 		break;
6248 	default:
6249 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6250 		    __func__);
6251 		return (EINVAL);
6252 	}
6253 
6254 	info->ifi_filter = filter;
6255 	info->ifi_filter_arg = filter_arg;
6256 	info->ifi_task = gtask;
6257 	info->ifi_ctx = q;
6258 
6259 	dev = ctx->ifc_dev;
6260 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6261 	if (err != 0) {
6262 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6263 		return (err);
6264 	}
6265 	if (type == IFLIB_INTR_ADMIN)
6266 		return (0);
6267 
6268 	if (tqrid != -1) {
6269 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6270 		    name);
6271 		if (err)
6272 			return (err);
6273 	} else {
6274 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6275 	}
6276 
6277 	return (0);
6278 }
6279 
6280 void
6281 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6282 			    void *arg, int qid, const char *name)
6283 {
6284 	device_t dev;
6285 	struct grouptask *gtask;
6286 	struct taskqgroup *tqg;
6287 	gtask_fn_t *fn;
6288 	void *q;
6289 	int err;
6290 
6291 	switch (type) {
6292 	case IFLIB_INTR_TX:
6293 		q = &ctx->ifc_txqs[qid];
6294 		gtask = &ctx->ifc_txqs[qid].ift_task;
6295 		tqg = qgroup_if_io_tqg;
6296 		fn = _task_fn_tx;
6297 		GROUPTASK_INIT(gtask, 0, fn, q);
6298 		break;
6299 	case IFLIB_INTR_RX:
6300 		q = &ctx->ifc_rxqs[qid];
6301 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6302 		tqg = qgroup_if_io_tqg;
6303 		fn = _task_fn_rx;
6304 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6305 		break;
6306 	case IFLIB_INTR_IOV:
6307 		q = ctx;
6308 		gtask = &ctx->ifc_vflr_task;
6309 		tqg = qgroup_if_config_tqg;
6310 		fn = _task_fn_iov;
6311 		GROUPTASK_INIT(gtask, 0, fn, q);
6312 		break;
6313 	default:
6314 		panic("unknown net intr type");
6315 	}
6316 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6317 	if (err) {
6318 		dev = ctx->ifc_dev;
6319 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6320 		    name);
6321 	}
6322 }
6323 
6324 void
6325 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6326 {
6327 
6328 	if (irq->ii_tag)
6329 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6330 
6331 	if (irq->ii_res)
6332 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6333 		    rman_get_rid(irq->ii_res), irq->ii_res);
6334 }
6335 
6336 static int
6337 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6338 {
6339 	iflib_txq_t txq = ctx->ifc_txqs;
6340 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6341 	if_irq_t irq = &ctx->ifc_legacy_irq;
6342 	iflib_filter_info_t info;
6343 	device_t dev;
6344 	struct grouptask *gtask;
6345 	struct resource *res;
6346 	struct taskqgroup *tqg;
6347 	void *q;
6348 	int err, tqrid;
6349 	bool rx_only;
6350 
6351 	q = &ctx->ifc_rxqs[0];
6352 	info = &rxq[0].ifr_filter_info;
6353 	gtask = &rxq[0].ifr_task;
6354 	tqg = qgroup_if_io_tqg;
6355 	tqrid = *rid;
6356 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6357 
6358 	ctx->ifc_flags |= IFC_LEGACY;
6359 	info->ifi_filter = filter;
6360 	info->ifi_filter_arg = filter_arg;
6361 	info->ifi_task = gtask;
6362 	info->ifi_ctx = rx_only ? ctx : q;
6363 
6364 	dev = ctx->ifc_dev;
6365 	/* We allocate a single interrupt resource */
6366 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6367 	    iflib_fast_intr_rxtx, NULL, info, name);
6368 	if (err != 0)
6369 		return (err);
6370 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6371 	res = irq->ii_res;
6372 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6373 
6374 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6375 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6376 	    "tx");
6377 	return (0);
6378 }
6379 
6380 void
6381 iflib_led_create(if_ctx_t ctx)
6382 {
6383 
6384 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6385 	    device_get_nameunit(ctx->ifc_dev));
6386 }
6387 
6388 void
6389 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6390 {
6391 
6392 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6393 }
6394 
6395 void
6396 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6397 {
6398 
6399 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6400 }
6401 
6402 void
6403 iflib_admin_intr_deferred(if_ctx_t ctx)
6404 {
6405 
6406 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6407 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6408 }
6409 
6410 void
6411 iflib_iov_intr_deferred(if_ctx_t ctx)
6412 {
6413 
6414 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6415 }
6416 
6417 void
6418 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6419 {
6420 
6421 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6422 	    name);
6423 }
6424 
6425 void
6426 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6427 	const char *name)
6428 {
6429 
6430 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6431 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6432 	    name);
6433 }
6434 
6435 void
6436 iflib_config_gtask_deinit(struct grouptask *gtask)
6437 {
6438 
6439 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6440 }
6441 
6442 void
6443 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6444 {
6445 	if_t ifp = ctx->ifc_ifp;
6446 	iflib_txq_t txq = ctx->ifc_txqs;
6447 
6448 	if_setbaudrate(ifp, baudrate);
6449 	if (baudrate >= IF_Gbps(10)) {
6450 		STATE_LOCK(ctx);
6451 		ctx->ifc_flags |= IFC_PREFETCH;
6452 		STATE_UNLOCK(ctx);
6453 	}
6454 	/* If link down, disable watchdog */
6455 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6456 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6457 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6458 	}
6459 	ctx->ifc_link_state = link_state;
6460 	if_link_state_change(ifp, link_state);
6461 }
6462 
6463 static int
6464 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6465 {
6466 	int credits;
6467 #ifdef INVARIANTS
6468 	int credits_pre = txq->ift_cidx_processed;
6469 #endif
6470 
6471 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6472 	    BUS_DMASYNC_POSTREAD);
6473 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6474 		return (0);
6475 
6476 	txq->ift_processed += credits;
6477 	txq->ift_cidx_processed += credits;
6478 
6479 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6480 	if (txq->ift_cidx_processed >= txq->ift_size)
6481 		txq->ift_cidx_processed -= txq->ift_size;
6482 	return (credits);
6483 }
6484 
6485 static int
6486 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6487 {
6488 	iflib_fl_t fl;
6489 	u_int i;
6490 
6491 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6492 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6493 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6494 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6495 	    budget));
6496 }
6497 
6498 void
6499 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6500 	const char *description, if_int_delay_info_t info,
6501 	int offset, int value)
6502 {
6503 	info->iidi_ctx = ctx;
6504 	info->iidi_offset = offset;
6505 	info->iidi_value = value;
6506 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6507 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6508 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6509 	    info, 0, iflib_sysctl_int_delay, "I", description);
6510 }
6511 
6512 struct sx *
6513 iflib_ctx_lock_get(if_ctx_t ctx)
6514 {
6515 
6516 	return (&ctx->ifc_ctx_sx);
6517 }
6518 
6519 static int
6520 iflib_msix_init(if_ctx_t ctx)
6521 {
6522 	device_t dev = ctx->ifc_dev;
6523 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6524 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6525 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6526 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6527 
6528 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6529 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6530 
6531 	if (bootverbose)
6532 		device_printf(dev, "msix_init qsets capped at %d\n",
6533 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6534 
6535 	/* Override by tuneable */
6536 	if (scctx->isc_disable_msix)
6537 		goto msi;
6538 
6539 	/* First try MSI-X */
6540 	if ((msgs = pci_msix_count(dev)) == 0) {
6541 		if (bootverbose)
6542 			device_printf(dev, "MSI-X not supported or disabled\n");
6543 		goto msi;
6544 	}
6545 
6546 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6547 	/*
6548 	 * bar == -1 => "trust me I know what I'm doing"
6549 	 * Some drivers are for hardware that is so shoddily
6550 	 * documented that no one knows which bars are which
6551 	 * so the developer has to map all bars. This hack
6552 	 * allows shoddy garbage to use MSI-X in this framework.
6553 	 */
6554 	if (bar != -1) {
6555 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6556 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6557 		if (ctx->ifc_msix_mem == NULL) {
6558 			device_printf(dev, "Unable to map MSI-X table\n");
6559 			goto msi;
6560 		}
6561 	}
6562 
6563 	admincnt = sctx->isc_admin_intrcnt;
6564 #if IFLIB_DEBUG
6565 	/* use only 1 qset in debug mode */
6566 	queuemsgs = min(msgs - admincnt, 1);
6567 #else
6568 	queuemsgs = msgs - admincnt;
6569 #endif
6570 #ifdef RSS
6571 	queues = imin(queuemsgs, rss_getnumbuckets());
6572 #else
6573 	queues = queuemsgs;
6574 #endif
6575 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6576 	if (bootverbose)
6577 		device_printf(dev,
6578 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6579 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6580 #ifdef  RSS
6581 	/* If we're doing RSS, clamp at the number of RSS buckets */
6582 	if (queues > rss_getnumbuckets())
6583 		queues = rss_getnumbuckets();
6584 #endif
6585 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6586 		rx_queues = iflib_num_rx_queues;
6587 	else
6588 		rx_queues = queues;
6589 
6590 	if (rx_queues > scctx->isc_nrxqsets)
6591 		rx_queues = scctx->isc_nrxqsets;
6592 
6593 	/*
6594 	 * We want this to be all logical CPUs by default
6595 	 */
6596 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6597 		tx_queues = iflib_num_tx_queues;
6598 	else
6599 		tx_queues = mp_ncpus;
6600 
6601 	if (tx_queues > scctx->isc_ntxqsets)
6602 		tx_queues = scctx->isc_ntxqsets;
6603 
6604 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6605 #ifdef INVARIANTS
6606 		if (tx_queues != rx_queues)
6607 			device_printf(dev,
6608 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6609 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6610 #endif
6611 		tx_queues = min(rx_queues, tx_queues);
6612 		rx_queues = min(rx_queues, tx_queues);
6613 	}
6614 
6615 	vectors = rx_queues + admincnt;
6616 	if (msgs < vectors) {
6617 		device_printf(dev,
6618 		    "insufficient number of MSI-X vectors "
6619 		    "(supported %d, need %d)\n", msgs, vectors);
6620 		goto msi;
6621 	}
6622 
6623 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6624 	    tx_queues);
6625 	msgs = vectors;
6626 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6627 		if (vectors != msgs) {
6628 			device_printf(dev,
6629 			    "Unable to allocate sufficient MSI-X vectors "
6630 			    "(got %d, need %d)\n", vectors, msgs);
6631 			pci_release_msi(dev);
6632 			if (bar != -1) {
6633 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6634 				    ctx->ifc_msix_mem);
6635 				ctx->ifc_msix_mem = NULL;
6636 			}
6637 			goto msi;
6638 		}
6639 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6640 		    vectors);
6641 		scctx->isc_vectors = vectors;
6642 		scctx->isc_nrxqsets = rx_queues;
6643 		scctx->isc_ntxqsets = tx_queues;
6644 		scctx->isc_intr = IFLIB_INTR_MSIX;
6645 
6646 		return (vectors);
6647 	} else {
6648 		device_printf(dev,
6649 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6650 		    err);
6651 		if (bar != -1) {
6652 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6653 			    ctx->ifc_msix_mem);
6654 			ctx->ifc_msix_mem = NULL;
6655 		}
6656 	}
6657 
6658 msi:
6659 	vectors = pci_msi_count(dev);
6660 	scctx->isc_nrxqsets = 1;
6661 	scctx->isc_ntxqsets = 1;
6662 	scctx->isc_vectors = vectors;
6663 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6664 		device_printf(dev,"Using an MSI interrupt\n");
6665 		scctx->isc_intr = IFLIB_INTR_MSI;
6666 	} else {
6667 		scctx->isc_vectors = 1;
6668 		device_printf(dev,"Using a Legacy interrupt\n");
6669 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6670 	}
6671 
6672 	return (vectors);
6673 }
6674 
6675 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6676 
6677 static int
6678 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6679 {
6680 	int rc;
6681 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6682 	struct sbuf *sb;
6683 	const char *ring_state = "UNKNOWN";
6684 
6685 	/* XXX needed ? */
6686 	rc = sysctl_wire_old_buffer(req, 0);
6687 	MPASS(rc == 0);
6688 	if (rc != 0)
6689 		return (rc);
6690 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6691 	MPASS(sb != NULL);
6692 	if (sb == NULL)
6693 		return (ENOMEM);
6694 	if (state[3] <= 3)
6695 		ring_state = ring_states[state[3]];
6696 
6697 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6698 		    state[0], state[1], state[2], ring_state);
6699 	rc = sbuf_finish(sb);
6700 	sbuf_delete(sb);
6701         return(rc);
6702 }
6703 
6704 enum iflib_ndesc_handler {
6705 	IFLIB_NTXD_HANDLER,
6706 	IFLIB_NRXD_HANDLER,
6707 };
6708 
6709 static int
6710 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6711 {
6712 	if_ctx_t ctx = (void *)arg1;
6713 	enum iflib_ndesc_handler type = arg2;
6714 	char buf[256] = {0};
6715 	qidx_t *ndesc;
6716 	char *p, *next;
6717 	int nqs, rc, i;
6718 
6719 	nqs = 8;
6720 	switch(type) {
6721 	case IFLIB_NTXD_HANDLER:
6722 		ndesc = ctx->ifc_sysctl_ntxds;
6723 		if (ctx->ifc_sctx)
6724 			nqs = ctx->ifc_sctx->isc_ntxqs;
6725 		break;
6726 	case IFLIB_NRXD_HANDLER:
6727 		ndesc = ctx->ifc_sysctl_nrxds;
6728 		if (ctx->ifc_sctx)
6729 			nqs = ctx->ifc_sctx->isc_nrxqs;
6730 		break;
6731 	default:
6732 		printf("%s: unhandled type\n", __func__);
6733 		return (EINVAL);
6734 	}
6735 	if (nqs == 0)
6736 		nqs = 8;
6737 
6738 	for (i=0; i<8; i++) {
6739 		if (i >= nqs)
6740 			break;
6741 		if (i)
6742 			strcat(buf, ",");
6743 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6744 	}
6745 
6746 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6747 	if (rc || req->newptr == NULL)
6748 		return rc;
6749 
6750 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6751 	    i++, p = strsep(&next, " ,")) {
6752 		ndesc[i] = strtoul(p, NULL, 10);
6753 	}
6754 
6755 	return(rc);
6756 }
6757 
6758 #define NAME_BUFLEN 32
6759 static void
6760 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6761 {
6762         device_t dev = iflib_get_dev(ctx);
6763 	struct sysctl_oid_list *child, *oid_list;
6764 	struct sysctl_ctx_list *ctx_list;
6765 	struct sysctl_oid *node;
6766 
6767 	ctx_list = device_get_sysctl_ctx(dev);
6768 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6769 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6770 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6771 	oid_list = SYSCTL_CHILDREN(node);
6772 
6773 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6774 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6775 		       "driver version");
6776 
6777 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6778 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6779 		       "# of txqs to use, 0 => use default #");
6780 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6781 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6782 		       "# of rxqs to use, 0 => use default #");
6783 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6784 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6785                        "permit #txq != #rxq");
6786 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6787 		       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6788 		       "disable MSI-X (default 0)");
6789 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6790 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6791 		       "set the RX budget");
6792 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6793 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6794 		       "cause TX to abdicate instead of running to completion");
6795 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6796 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6797 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6798 		       "offset to start using cores at");
6799 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6800 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6801 		      "use separate cores for TX and RX");
6802 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6803 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6804 		      "try to make use of logical cores for TX and RX");
6805 
6806 	/* XXX change for per-queue sizes */
6807 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6808 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6809 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6810 	    "list of # of TX descriptors to use, 0 = use default #");
6811 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6812 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6813 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6814 	    "list of # of RX descriptors to use, 0 = use default #");
6815 }
6816 
6817 static void
6818 iflib_add_device_sysctl_post(if_ctx_t ctx)
6819 {
6820 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6821 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6822         device_t dev = iflib_get_dev(ctx);
6823 	struct sysctl_oid_list *child;
6824 	struct sysctl_ctx_list *ctx_list;
6825 	iflib_fl_t fl;
6826 	iflib_txq_t txq;
6827 	iflib_rxq_t rxq;
6828 	int i, j;
6829 	char namebuf[NAME_BUFLEN];
6830 	char *qfmt;
6831 	struct sysctl_oid *queue_node, *fl_node, *node;
6832 	struct sysctl_oid_list *queue_list, *fl_list;
6833 	ctx_list = device_get_sysctl_ctx(dev);
6834 
6835 	node = ctx->ifc_sysctl_node;
6836 	child = SYSCTL_CHILDREN(node);
6837 
6838 	if (scctx->isc_ntxqsets > 100)
6839 		qfmt = "txq%03d";
6840 	else if (scctx->isc_ntxqsets > 10)
6841 		qfmt = "txq%02d";
6842 	else
6843 		qfmt = "txq%d";
6844 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6845 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6846 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6847 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6848 		queue_list = SYSCTL_CHILDREN(queue_node);
6849 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6850 			       CTLFLAG_RD,
6851 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
6852 #if MEMORY_LOGGING
6853 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6854 				CTLFLAG_RD,
6855 				&txq->ift_dequeued, "total mbufs freed");
6856 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6857 				CTLFLAG_RD,
6858 				&txq->ift_enqueued, "total mbufs enqueued");
6859 #endif
6860 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6861 				   CTLFLAG_RD,
6862 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6863 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6864 				   CTLFLAG_RD,
6865 				   &txq->ift_pullups, "# of times m_pullup was called");
6866 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6867 				   CTLFLAG_RD,
6868 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6869 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6870 				   CTLFLAG_RD,
6871 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6872 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6873 				   CTLFLAG_RD,
6874 				   &txq->ift_map_failed, "# of times DMA map failed");
6875 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6876 				   CTLFLAG_RD,
6877 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6878 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6879 				   CTLFLAG_RD,
6880 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6881 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6882 				   CTLFLAG_RD,
6883 				   &txq->ift_pidx, 1, "Producer Index");
6884 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6885 				   CTLFLAG_RD,
6886 				   &txq->ift_cidx, 1, "Consumer Index");
6887 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6888 				   CTLFLAG_RD,
6889 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6890 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6891 				   CTLFLAG_RD,
6892 				   &txq->ift_in_use, 1, "descriptors in use");
6893 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6894 				   CTLFLAG_RD,
6895 				   &txq->ift_processed, "descriptors procesed for clean");
6896 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6897 				   CTLFLAG_RD,
6898 				   &txq->ift_cleaned, "total cleaned");
6899 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6900 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6901 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6902 		    mp_ring_state_handler, "A", "soft ring state");
6903 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6904 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6905 				       "# of enqueues to the mp_ring for this queue");
6906 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6907 				       CTLFLAG_RD, &txq->ift_br->drops,
6908 				       "# of drops in the mp_ring for this queue");
6909 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6910 				       CTLFLAG_RD, &txq->ift_br->starts,
6911 				       "# of normal consumer starts in the mp_ring for this queue");
6912 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6913 				       CTLFLAG_RD, &txq->ift_br->stalls,
6914 					       "# of consumer stalls in the mp_ring for this queue");
6915 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6916 			       CTLFLAG_RD, &txq->ift_br->restarts,
6917 				       "# of consumer restarts in the mp_ring for this queue");
6918 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6919 				       CTLFLAG_RD, &txq->ift_br->abdications,
6920 				       "# of consumer abdications in the mp_ring for this queue");
6921 	}
6922 
6923 	if (scctx->isc_nrxqsets > 100)
6924 		qfmt = "rxq%03d";
6925 	else if (scctx->isc_nrxqsets > 10)
6926 		qfmt = "rxq%02d";
6927 	else
6928 		qfmt = "rxq%d";
6929 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6930 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6931 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6932 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6933 		queue_list = SYSCTL_CHILDREN(queue_node);
6934 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6935 			       CTLFLAG_RD,
6936 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
6937 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6938 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6939 				       CTLFLAG_RD,
6940 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6941 		}
6942 
6943 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6944 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6945 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6946 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6947 			fl_list = SYSCTL_CHILDREN(fl_node);
6948 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6949 				       CTLFLAG_RD,
6950 				       &fl->ifl_pidx, 1, "Producer Index");
6951 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6952 				       CTLFLAG_RD,
6953 				       &fl->ifl_cidx, 1, "Consumer Index");
6954 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6955 				       CTLFLAG_RD,
6956 				       &fl->ifl_credits, 1, "credits available");
6957 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6958 				       CTLFLAG_RD,
6959 				       &fl->ifl_buf_size, 1, "buffer size");
6960 #if MEMORY_LOGGING
6961 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6962 					CTLFLAG_RD,
6963 					&fl->ifl_m_enqueued, "mbufs allocated");
6964 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6965 					CTLFLAG_RD,
6966 					&fl->ifl_m_dequeued, "mbufs freed");
6967 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6968 					CTLFLAG_RD,
6969 					&fl->ifl_cl_enqueued, "clusters allocated");
6970 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6971 					CTLFLAG_RD,
6972 					&fl->ifl_cl_dequeued, "clusters freed");
6973 #endif
6974 		}
6975 	}
6976 
6977 }
6978 
6979 void
6980 iflib_request_reset(if_ctx_t ctx)
6981 {
6982 
6983 	STATE_LOCK(ctx);
6984 	ctx->ifc_flags |= IFC_DO_RESET;
6985 	STATE_UNLOCK(ctx);
6986 }
6987 
6988 #ifndef __NO_STRICT_ALIGNMENT
6989 static struct mbuf *
6990 iflib_fixup_rx(struct mbuf *m)
6991 {
6992 	struct mbuf *n;
6993 
6994 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6995 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6996 		m->m_data += ETHER_HDR_LEN;
6997 		n = m;
6998 	} else {
6999 		MGETHDR(n, M_NOWAIT, MT_DATA);
7000 		if (n == NULL) {
7001 			m_freem(m);
7002 			return (NULL);
7003 		}
7004 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7005 		m->m_data += ETHER_HDR_LEN;
7006 		m->m_len -= ETHER_HDR_LEN;
7007 		n->m_len = ETHER_HDR_LEN;
7008 		M_MOVE_PKTHDR(n, m);
7009 		n->m_next = m;
7010 	}
7011 	return (n);
7012 }
7013 #endif
7014 
7015 #ifdef DEBUGNET
7016 static void
7017 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7018 {
7019 	if_ctx_t ctx;
7020 
7021 	ctx = if_getsoftc(ifp);
7022 	CTX_LOCK(ctx);
7023 	*nrxr = NRXQSETS(ctx);
7024 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7025 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7026 	CTX_UNLOCK(ctx);
7027 }
7028 
7029 static void
7030 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7031 {
7032 	if_ctx_t ctx;
7033 	if_softc_ctx_t scctx;
7034 	iflib_fl_t fl;
7035 	iflib_rxq_t rxq;
7036 	int i, j;
7037 
7038 	ctx = if_getsoftc(ifp);
7039 	scctx = &ctx->ifc_softc_ctx;
7040 
7041 	switch (event) {
7042 	case DEBUGNET_START:
7043 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7044 			rxq = &ctx->ifc_rxqs[i];
7045 			for (j = 0; j < rxq->ifr_nfl; j++) {
7046 				fl = rxq->ifr_fl;
7047 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7048 			}
7049 		}
7050 		iflib_no_tx_batch = 1;
7051 		break;
7052 	default:
7053 		break;
7054 	}
7055 }
7056 
7057 static int
7058 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7059 {
7060 	if_ctx_t ctx;
7061 	iflib_txq_t txq;
7062 	int error;
7063 
7064 	ctx = if_getsoftc(ifp);
7065 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7066 	    IFF_DRV_RUNNING)
7067 		return (EBUSY);
7068 
7069 	txq = &ctx->ifc_txqs[0];
7070 	error = iflib_encap(txq, &m);
7071 	if (error == 0)
7072 		(void)iflib_txd_db_check(txq, true);
7073 	return (error);
7074 }
7075 
7076 static int
7077 iflib_debugnet_poll(if_t ifp, int count)
7078 {
7079 	struct epoch_tracker et;
7080 	if_ctx_t ctx;
7081 	if_softc_ctx_t scctx;
7082 	iflib_txq_t txq;
7083 	int i;
7084 
7085 	ctx = if_getsoftc(ifp);
7086 	scctx = &ctx->ifc_softc_ctx;
7087 
7088 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7089 	    IFF_DRV_RUNNING)
7090 		return (EBUSY);
7091 
7092 	txq = &ctx->ifc_txqs[0];
7093 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7094 
7095 	NET_EPOCH_ENTER(et);
7096 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7097 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7098 	NET_EPOCH_EXIT(et);
7099 	return (0);
7100 }
7101 #endif /* DEBUGNET */
7102