xref: /freebsd/sys/net/iflib.c (revision e4a0c92e7aea50654290e3082668932cea16b64f)
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,
579     "minimize transmit latency at the possible expense of throughput");
580 static int iflib_no_tx_batch = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
582     &iflib_no_tx_batch, 0,
583     "minimize transmit latency at the possible expense of throughput");
584 static int iflib_timer_default = 1000;
585 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
586     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
587 
588 
589 #if IFLIB_DEBUG_COUNTERS
590 
591 static int iflib_tx_seen;
592 static int iflib_tx_sent;
593 static int iflib_tx_encap;
594 static int iflib_rx_allocs;
595 static int iflib_fl_refills;
596 static int iflib_fl_refills_large;
597 static int iflib_tx_frees;
598 
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
600     "# TX mbufs seen");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
602     "# TX mbufs sent");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
604     "# TX mbufs encapped");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
606     "# TX frees");
607 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
608     "# RX allocations");
609 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
610     "# refills");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
612     &iflib_fl_refills_large, 0, "# large refills");
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
619     &iflib_txq_drain_flushing, 0, "# drain flushes");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
621     &iflib_txq_drain_oactive, 0, "# drain oactives");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
623     &iflib_txq_drain_notready, 0, "# drain notready");
624 
625 static int iflib_encap_load_mbuf_fail;
626 static int iflib_encap_pad_mbuf_fail;
627 static int iflib_encap_txq_avail_fail;
628 static int iflib_encap_txd_encap_fail;
629 
630 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
631     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
632 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
633     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
635     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
637     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
638 
639 static int iflib_task_fn_rxs;
640 static int iflib_rx_intr_enables;
641 static int iflib_fast_intrs;
642 static int iflib_rx_unavail;
643 static int iflib_rx_ctx_inactive;
644 static int iflib_rx_if_input;
645 static int iflib_rxd_flush;
646 
647 static int iflib_verbose_debug;
648 
649 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
650     "# task_fn_rx calls");
651 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
652     &iflib_rx_intr_enables, 0, "# RX intr enables");
653 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
654     "# fast_intr calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
656     "# times rxeof called with no available data");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
658     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
660     0, "# times rxeof called if_input");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
662     "# times rxd_flush called");
663 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
664     &iflib_verbose_debug, 0, "enable verbose debugging");
665 
666 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
667 static void
668 iflib_debug_reset(void)
669 {
670 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
671 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
672 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
673 		iflib_txq_drain_notready =
674 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
675 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
676 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
677 		iflib_rx_unavail =
678 		iflib_rx_ctx_inactive = iflib_rx_if_input =
679 		iflib_rxd_flush = 0;
680 }
681 
682 #else
683 #define DBG_COUNTER_INC(name)
684 static void iflib_debug_reset(void) {}
685 #endif
686 
687 #define IFLIB_DEBUG 0
688 
689 static void iflib_tx_structures_free(if_ctx_t ctx);
690 static void iflib_rx_structures_free(if_ctx_t ctx);
691 static int iflib_queues_alloc(if_ctx_t ctx);
692 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
693 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
694 static int iflib_qset_structures_setup(if_ctx_t ctx);
695 static int iflib_msix_init(if_ctx_t ctx);
696 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
697 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
698 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
699 #ifdef ALTQ
700 static void iflib_altq_if_start(if_t ifp);
701 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
702 #endif
703 static int iflib_register(if_ctx_t);
704 static void iflib_deregister(if_ctx_t);
705 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
706 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
707 static void iflib_init_locked(if_ctx_t ctx);
708 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
709 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
710 static void iflib_ifmp_purge(iflib_txq_t txq);
711 static void _iflib_pre_assert(if_softc_ctx_t scctx);
712 static void iflib_stop(if_ctx_t ctx);
713 static void iflib_if_init_locked(if_ctx_t ctx);
714 static void iflib_free_intr_mem(if_ctx_t ctx);
715 #ifndef __NO_STRICT_ALIGNMENT
716 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
717 #endif
718 
719 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
720     SLIST_HEAD_INITIALIZER(cpu_offsets);
721 struct cpu_offset {
722 	SLIST_ENTRY(cpu_offset) entries;
723 	cpuset_t	set;
724 	unsigned int	refcount;
725 	uint16_t	next_cpuid;
726 };
727 static struct mtx cpu_offset_mtx;
728 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
729     MTX_DEF);
730 
731 DEBUGNET_DEFINE(iflib);
732 
733 static int
734 iflib_num_rx_descs(if_ctx_t ctx)
735 {
736 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
737 	if_shared_ctx_t sctx = ctx->ifc_sctx;
738 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
739 
740 	return scctx->isc_nrxd[first_rxq];
741 }
742 
743 static int
744 iflib_num_tx_descs(if_ctx_t ctx)
745 {
746 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
747 	if_shared_ctx_t sctx = ctx->ifc_sctx;
748 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
749 
750 	return scctx->isc_ntxd[first_txq];
751 }
752 
753 #ifdef DEV_NETMAP
754 #include <sys/selinfo.h>
755 #include <net/netmap.h>
756 #include <dev/netmap/netmap_kern.h>
757 
758 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
759 
760 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
761 static void iflib_netmap_timer(void *arg);
762 
763 /*
764  * device-specific sysctl variables:
765  *
766  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
767  *	During regular operations the CRC is stripped, but on some
768  *	hardware reception of frames not multiple of 64 is slower,
769  *	so using crcstrip=0 helps in benchmarks.
770  *
771  * iflib_rx_miss, iflib_rx_miss_bufs:
772  *	count packets that might be missed due to lost interrupts.
773  */
774 SYSCTL_DECL(_dev_netmap);
775 /*
776  * The xl driver by default strips CRCs and we do not override it.
777  */
778 
779 int iflib_crcstrip = 1;
780 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
781     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
782 
783 int iflib_rx_miss, iflib_rx_miss_bufs;
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
785     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
786 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
787     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
788 
789 /*
790  * Register/unregister. We are already under netmap lock.
791  * Only called on the first register or the last unregister.
792  */
793 static int
794 iflib_netmap_register(struct netmap_adapter *na, int onoff)
795 {
796 	if_t ifp = na->ifp;
797 	if_ctx_t ctx = if_getsoftc(ifp);
798 	int status;
799 
800 	CTX_LOCK(ctx);
801 	if (!CTX_IS_VF(ctx))
802 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
803 
804 	iflib_stop(ctx);
805 
806 	/*
807 	 * Enable (or disable) netmap flags, and intercept (or restore)
808 	 * ifp->if_transmit. This is done once the device has been stopped
809 	 * to prevent race conditions. Also, this must be done after
810 	 * calling netmap_disable_all_rings() and before calling
811 	 * netmap_enable_all_rings(), so that these two functions see the
812 	 * updated state of the NAF_NETMAP_ON bit.
813 	 */
814 	if (onoff) {
815 		nm_set_native_flags(na);
816 	} else {
817 		nm_clear_native_flags(na);
818 	}
819 
820 	iflib_init_locked(ctx);
821 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
822 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
823 	if (status)
824 		nm_clear_native_flags(na);
825 	CTX_UNLOCK(ctx);
826 	return (status);
827 }
828 
829 static int
830 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
831 {
832 	if_t ifp = na->ifp;
833 	if_ctx_t ctx = if_getsoftc(ifp);
834 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
835 	iflib_fl_t fl = &rxq->ifr_fl[0];
836 
837 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
838 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
839 	info->num_tx_descs = iflib_num_tx_descs(ctx);
840 	info->num_rx_descs = iflib_num_rx_descs(ctx);
841 	info->rx_buf_maxsize = fl->ifl_buf_size;
842 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
843 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
844 		info->num_rx_descs, info->rx_buf_maxsize);
845 
846 	return 0;
847 }
848 
849 static int
850 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
851 {
852 	struct netmap_adapter *na = kring->na;
853 	u_int const lim = kring->nkr_num_slots - 1;
854 	struct netmap_ring *ring = kring->ring;
855 	bus_dmamap_t *map;
856 	struct if_rxd_update iru;
857 	if_ctx_t ctx = rxq->ifr_ctx;
858 	iflib_fl_t fl = &rxq->ifr_fl[0];
859 	u_int nic_i_first, nic_i;
860 	u_int nm_i;
861 	int i, n;
862 #if IFLIB_DEBUG_COUNTERS
863 	int rf_count = 0;
864 #endif
865 
866 	/*
867 	 * This function is used both at initialization and in rxsync.
868 	 * At initialization we need to prepare (with isc_rxd_refill())
869 	 * all the netmap buffers currently owned by the kernel, in
870 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
871 	 * (except for kring->nkr_hwofs). These may be less than
872 	 * kring->nkr_num_slots if netmap_reset() was called while
873 	 * an application using the kring that still owned some
874 	 * buffers.
875 	 * At rxsync time, both indexes point to the next buffer to be
876 	 * refilled.
877 	 * In any case we publish (with isc_rxd_flush()) up to
878 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
879 	 * pointer to overrun the head/cons pointer, although this is
880 	 * not necessary for some NICs (e.g. vmx).
881 	 */
882 	if (__predict_false(init)) {
883 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
884 	} else {
885 		n = kring->rhead - kring->nr_hwcur;
886 		if (n == 0)
887 			return (0); /* Nothing to do. */
888 		if (n < 0)
889 			n += kring->nkr_num_slots;
890 	}
891 
892 	iru_init(&iru, rxq, 0 /* flid */);
893 	map = fl->ifl_sds.ifsd_map;
894 	nic_i = fl->ifl_pidx;
895 	nm_i = netmap_idx_n2k(kring, nic_i);
896 	if (__predict_false(init)) {
897 		/*
898 		 * On init/reset, nic_i must be 0, and we must
899 		 * start to refill from hwtail (see netmap_reset()).
900 		 */
901 		MPASS(nic_i == 0);
902 		MPASS(nm_i == kring->nr_hwtail);
903 	} else
904 		MPASS(nm_i == kring->nr_hwcur);
905 	DBG_COUNTER_INC(fl_refills);
906 	while (n > 0) {
907 #if IFLIB_DEBUG_COUNTERS
908 		if (++rf_count == 9)
909 			DBG_COUNTER_INC(fl_refills_large);
910 #endif
911 		nic_i_first = nic_i;
912 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
913 			struct netmap_slot *slot = &ring->slot[nm_i];
914 			uint64_t paddr;
915 			void *addr = PNMB(na, slot, &paddr);
916 
917 			MPASS(i < IFLIB_MAX_RX_REFRESH);
918 
919 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
920 			        return netmap_ring_reinit(kring);
921 
922 			fl->ifl_bus_addrs[i] = paddr +
923 			    nm_get_offset(kring, slot);
924 			fl->ifl_rxd_idxs[i] = nic_i;
925 
926 			if (__predict_false(init)) {
927 				netmap_load_map(na, fl->ifl_buf_tag,
928 				    map[nic_i], addr);
929 			} else if (slot->flags & NS_BUF_CHANGED) {
930 				/* buffer has changed, reload map */
931 				netmap_reload_map(na, fl->ifl_buf_tag,
932 				    map[nic_i], addr);
933 			}
934 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
935 			    BUS_DMASYNC_PREREAD);
936 			slot->flags &= ~NS_BUF_CHANGED;
937 
938 			nm_i = nm_next(nm_i, lim);
939 			nic_i = nm_next(nic_i, lim);
940 		}
941 
942 		iru.iru_pidx = nic_i_first;
943 		iru.iru_count = i;
944 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
945 	}
946 	fl->ifl_pidx = nic_i;
947 	/*
948 	 * At the end of the loop we must have refilled everything
949 	 * we could possibly refill.
950 	 */
951 	MPASS(nm_i == kring->rhead);
952 	kring->nr_hwcur = nm_i;
953 
954 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
955 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
956 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
957 	    nm_prev(nic_i, lim));
958 	DBG_COUNTER_INC(rxd_flush);
959 
960 	return (0);
961 }
962 
963 #define NETMAP_TX_TIMER_US	90
964 
965 /*
966  * Reconcile kernel and user view of the transmit ring.
967  *
968  * All information is in the kring.
969  * Userspace wants to send packets up to the one before kring->rhead,
970  * kernel knows kring->nr_hwcur is the first unsent packet.
971  *
972  * Here we push packets out (as many as possible), and possibly
973  * reclaim buffers from previously completed transmission.
974  *
975  * The caller (netmap) guarantees that there is only one instance
976  * running at any time. Any interference with other driver
977  * methods should be handled by the individual drivers.
978  */
979 static int
980 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
981 {
982 	struct netmap_adapter *na = kring->na;
983 	if_t ifp = na->ifp;
984 	struct netmap_ring *ring = kring->ring;
985 	u_int nm_i;	/* index into the netmap kring */
986 	u_int nic_i;	/* index into the NIC ring */
987 	u_int const lim = kring->nkr_num_slots - 1;
988 	u_int const head = kring->rhead;
989 	struct if_pkt_info pi;
990 	int tx_pkts = 0, tx_bytes = 0;
991 
992 	/*
993 	 * interrupts on every tx packet are expensive so request
994 	 * them every half ring, or where NS_REPORT is set
995 	 */
996 	u_int report_frequency = kring->nkr_num_slots >> 1;
997 	/* device-specific */
998 	if_ctx_t ctx = if_getsoftc(ifp);
999 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1000 
1001 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1002 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1003 
1004 	/*
1005 	 * First part: process new packets to send.
1006 	 * nm_i is the current index in the netmap kring,
1007 	 * nic_i is the corresponding index in the NIC ring.
1008 	 *
1009 	 * If we have packets to send (nm_i != head)
1010 	 * iterate over the netmap ring, fetch length and update
1011 	 * the corresponding slot in the NIC ring. Some drivers also
1012 	 * need to update the buffer's physical address in the NIC slot
1013 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1014 	 *
1015 	 * The netmap_reload_map() calls is especially expensive,
1016 	 * even when (as in this case) the tag is 0, so do only
1017 	 * when the buffer has actually changed.
1018 	 *
1019 	 * If possible do not set the report/intr bit on all slots,
1020 	 * but only a few times per ring or when NS_REPORT is set.
1021 	 *
1022 	 * Finally, on 10G and faster drivers, it might be useful
1023 	 * to prefetch the next slot and txr entry.
1024 	 */
1025 
1026 	nm_i = kring->nr_hwcur;
1027 	if (nm_i != head) {	/* we have new packets to send */
1028 		uint32_t pkt_len = 0, seg_idx = 0;
1029 		int nic_i_start = -1, flags = 0;
1030 		pkt_info_zero(&pi);
1031 		pi.ipi_segs = txq->ift_segs;
1032 		pi.ipi_qsidx = kring->ring_id;
1033 		nic_i = netmap_idx_k2n(kring, nm_i);
1034 
1035 		__builtin_prefetch(&ring->slot[nm_i]);
1036 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1037 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1038 
1039 		while (nm_i != head) {
1040 			struct netmap_slot *slot = &ring->slot[nm_i];
1041 			uint64_t offset = nm_get_offset(kring, slot);
1042 			u_int len = slot->len;
1043 			uint64_t paddr;
1044 			void *addr = PNMB(na, slot, &paddr);
1045 
1046 			flags |= (slot->flags & NS_REPORT ||
1047 				nic_i == 0 || nic_i == report_frequency) ?
1048 				IPI_TX_INTR : 0;
1049 
1050 			/*
1051 			 * If this is the first packet fragment, save the
1052 			 * index of the first NIC slot for later.
1053 			 */
1054 			if (nic_i_start < 0)
1055 				nic_i_start = nic_i;
1056 
1057 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1058 			pi.ipi_segs[seg_idx].ds_len = len;
1059 			if (len) {
1060 				pkt_len += len;
1061 				seg_idx++;
1062 			}
1063 
1064 			if (!(slot->flags & NS_MOREFRAG)) {
1065 				pi.ipi_len = pkt_len;
1066 				pi.ipi_nsegs = seg_idx;
1067 				pi.ipi_pidx = nic_i_start;
1068 				pi.ipi_ndescs = 0;
1069 				pi.ipi_flags = flags;
1070 
1071 				/* Prepare the NIC TX ring. */
1072 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1073 				DBG_COUNTER_INC(tx_encap);
1074 
1075 				/* Update transmit counters */
1076 				tx_bytes += pi.ipi_len;
1077 				tx_pkts++;
1078 
1079 				/* Reinit per-packet info for the next one. */
1080 				flags = seg_idx = pkt_len = 0;
1081 				nic_i_start = -1;
1082 			}
1083 
1084 			/* prefetch for next round */
1085 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1086 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1087 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1088 
1089 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1090 
1091 			if (slot->flags & NS_BUF_CHANGED) {
1092 				/* buffer has changed, reload map */
1093 				netmap_reload_map(na, txq->ift_buf_tag,
1094 				    txq->ift_sds.ifsd_map[nic_i], addr);
1095 			}
1096 			/* make sure changes to the buffer are synced */
1097 			bus_dmamap_sync(txq->ift_buf_tag,
1098 			    txq->ift_sds.ifsd_map[nic_i],
1099 			    BUS_DMASYNC_PREWRITE);
1100 
1101 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1102 			nm_i = nm_next(nm_i, lim);
1103 			nic_i = nm_next(nic_i, lim);
1104 		}
1105 		kring->nr_hwcur = nm_i;
1106 
1107 		/* synchronize the NIC ring */
1108 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1109 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1110 
1111 		/* (re)start the tx unit up to slot nic_i (excluded) */
1112 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1113 	}
1114 
1115 	/*
1116 	 * Second part: reclaim buffers for completed transmissions.
1117 	 *
1118 	 * If there are unclaimed buffers, attempt to reclaim them.
1119 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1120 	 * trigger a per-tx-queue timer to try again later.
1121 	 */
1122 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1123 		if (iflib_tx_credits_update(ctx, txq)) {
1124 			/* some tx completed, increment avail */
1125 			nic_i = txq->ift_cidx_processed;
1126 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1127 		}
1128 	}
1129 
1130 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1131 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1132 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1133 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1134 			    iflib_netmap_timer, txq,
1135 			    txq->ift_netmap_timer.c_cpu, 0);
1136 		}
1137 
1138 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1139 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1140 
1141 	return (0);
1142 }
1143 
1144 /*
1145  * Reconcile kernel and user view of the receive ring.
1146  * Same as for the txsync, this routine must be efficient.
1147  * The caller guarantees a single invocations, but races against
1148  * the rest of the driver should be handled here.
1149  *
1150  * On call, kring->rhead is the first packet that userspace wants
1151  * to keep, and kring->rcur is the wakeup point.
1152  * The kernel has previously reported packets up to kring->rtail.
1153  *
1154  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1155  * of whether or not we received an interrupt.
1156  */
1157 static int
1158 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1159 {
1160 	struct netmap_adapter *na = kring->na;
1161 	struct netmap_ring *ring = kring->ring;
1162 	if_t ifp = na->ifp;
1163 	uint32_t nm_i;	/* index into the netmap ring */
1164 	uint32_t nic_i;	/* index into the NIC ring */
1165 	u_int n;
1166 	u_int const lim = kring->nkr_num_slots - 1;
1167 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1168 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1169 
1170 	if_ctx_t ctx = if_getsoftc(ifp);
1171 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1172 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1173 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1174 	iflib_fl_t fl = &rxq->ifr_fl[0];
1175 	struct if_rxd_info ri;
1176 	qidx_t *cidxp;
1177 
1178 	/*
1179 	 * netmap only uses free list 0, to avoid out of order consumption
1180 	 * of receive buffers
1181 	 */
1182 
1183 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1184 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1185 
1186 	/*
1187 	 * First part: import newly received packets.
1188 	 *
1189 	 * nm_i is the index of the next free slot in the netmap ring,
1190 	 * nic_i is the index of the next received packet in the NIC ring
1191 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1192 	 * differ in case if_init() has been called while
1193 	 * in netmap mode. For the receive ring we have
1194 	 *
1195 	 *	nic_i = fl->ifl_cidx;
1196 	 *	nm_i = kring->nr_hwtail (previous)
1197 	 * and
1198 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1199 	 *
1200 	 * fl->ifl_cidx is set to 0 on a ring reinit
1201 	 */
1202 	if (netmap_no_pendintr || force_update) {
1203 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1204 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1205 		int crclen = iflib_crcstrip ? 0 : 4;
1206 		int error, avail;
1207 
1208 		/*
1209 		 * For the free list consumer index, we use the same
1210 		 * logic as in iflib_rxeof().
1211 		 */
1212 		if (have_rxcq)
1213 			cidxp = &rxq->ifr_cq_cidx;
1214 		else
1215 			cidxp = &fl->ifl_cidx;
1216 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1217 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1218 
1219 		nic_i = fl->ifl_cidx;
1220 		nm_i = netmap_idx_n2k(kring, nic_i);
1221 		MPASS(nm_i == kring->nr_hwtail);
1222 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1223 			rxd_info_zero(&ri);
1224 			ri.iri_frags = rxq->ifr_frags;
1225 			ri.iri_qsidx = kring->ring_id;
1226 			ri.iri_ifp = ctx->ifc_ifp;
1227 			ri.iri_cidx = *cidxp;
1228 
1229 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1230 			for (i = 0; i < ri.iri_nfrags; i++) {
1231 				if (error) {
1232 					ring->slot[nm_i].len = 0;
1233 					ring->slot[nm_i].flags = 0;
1234 				} else {
1235 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1236 					if (i == (ri.iri_nfrags - 1)) {
1237 						ring->slot[nm_i].len -= crclen;
1238 						ring->slot[nm_i].flags = 0;
1239 
1240 						/* Update receive counters */
1241 						rx_bytes += ri.iri_len;
1242 						rx_pkts++;
1243 					} else
1244 						ring->slot[nm_i].flags = NS_MOREFRAG;
1245 				}
1246 
1247 				bus_dmamap_sync(fl->ifl_buf_tag,
1248 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1249 				nm_i = nm_next(nm_i, lim);
1250 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1251 			}
1252 
1253 			if (have_rxcq) {
1254 				*cidxp = ri.iri_cidx;
1255 				while (*cidxp >= scctx->isc_nrxd[0])
1256 					*cidxp -= scctx->isc_nrxd[0];
1257 			}
1258 
1259 		}
1260 		if (n) { /* update the state variables */
1261 			if (netmap_no_pendintr && !force_update) {
1262 				/* diagnostics */
1263 				iflib_rx_miss ++;
1264 				iflib_rx_miss_bufs += n;
1265 			}
1266 			kring->nr_hwtail = nm_i;
1267 		}
1268 		kring->nr_kflags &= ~NKR_PENDINTR;
1269 	}
1270 	/*
1271 	 * Second part: skip past packets that userspace has released.
1272 	 * (kring->nr_hwcur to head excluded),
1273 	 * and make the buffers available for reception.
1274 	 * As usual nm_i is the index in the netmap ring,
1275 	 * nic_i is the index in the NIC ring, and
1276 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1277 	 */
1278 	netmap_fl_refill(rxq, kring, false);
1279 
1280 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1281 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1282 
1283 	return (0);
1284 }
1285 
1286 static void
1287 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1288 {
1289 	if_ctx_t ctx = if_getsoftc(na->ifp);
1290 
1291 	CTX_LOCK(ctx);
1292 	if (onoff) {
1293 		IFDI_INTR_ENABLE(ctx);
1294 	} else {
1295 		IFDI_INTR_DISABLE(ctx);
1296 	}
1297 	CTX_UNLOCK(ctx);
1298 }
1299 
1300 static int
1301 iflib_netmap_attach(if_ctx_t ctx)
1302 {
1303 	struct netmap_adapter na;
1304 
1305 	bzero(&na, sizeof(na));
1306 
1307 	na.ifp = ctx->ifc_ifp;
1308 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1309 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1310 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1311 
1312 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1313 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1314 	na.nm_txsync = iflib_netmap_txsync;
1315 	na.nm_rxsync = iflib_netmap_rxsync;
1316 	na.nm_register = iflib_netmap_register;
1317 	na.nm_intr = iflib_netmap_intr;
1318 	na.nm_config = iflib_netmap_config;
1319 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1320 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1321 	return (netmap_attach(&na));
1322 }
1323 
1324 static int
1325 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1326 {
1327 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1328 	struct netmap_slot *slot;
1329 
1330 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1331 	if (slot == NULL)
1332 		return (0);
1333 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1334 		/*
1335 		 * In netmap mode, set the map for the packet buffer.
1336 		 * NOTE: Some drivers (not this one) also need to set
1337 		 * the physical buffer address in the NIC ring.
1338 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1339 		 * netmap slot index, si
1340 		 */
1341 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1342 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1343 		    NMB(na, slot + si));
1344 	}
1345 	return (1);
1346 }
1347 
1348 static int
1349 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1350 {
1351 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1352 	struct netmap_kring *kring;
1353 	struct netmap_slot *slot;
1354 
1355 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1356 	if (slot == NULL)
1357 		return (0);
1358 	kring = na->rx_rings[rxq->ifr_id];
1359 	netmap_fl_refill(rxq, kring, true);
1360 	return (1);
1361 }
1362 
1363 static void
1364 iflib_netmap_timer(void *arg)
1365 {
1366 	iflib_txq_t txq = arg;
1367 	if_ctx_t ctx = txq->ift_ctx;
1368 
1369 	/*
1370 	 * Wake up the netmap application, to give it a chance to
1371 	 * call txsync and reclaim more completed TX buffers.
1372 	 */
1373 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1374 }
1375 
1376 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1377 
1378 #else
1379 #define iflib_netmap_txq_init(ctx, txq) (0)
1380 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1381 #define iflib_netmap_detach(ifp)
1382 #define netmap_enable_all_rings(ifp)
1383 #define netmap_disable_all_rings(ifp)
1384 
1385 #define iflib_netmap_attach(ctx) (0)
1386 #define netmap_rx_irq(ifp, qid, budget) (0)
1387 #endif
1388 
1389 #if defined(__i386__) || defined(__amd64__)
1390 static __inline void
1391 prefetch(void *x)
1392 {
1393 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1394 }
1395 
1396 static __inline void
1397 prefetch2cachelines(void *x)
1398 {
1399 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1400 #if (CACHE_LINE_SIZE < 128)
1401 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1402 #endif
1403 }
1404 #else
1405 static __inline void
1406 prefetch(void *x)
1407 {
1408 }
1409 
1410 static __inline void
1411 prefetch2cachelines(void *x)
1412 {
1413 }
1414 #endif
1415 
1416 static void
1417 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1418 {
1419 	iflib_fl_t fl;
1420 
1421 	fl = &rxq->ifr_fl[flid];
1422 	iru->iru_paddrs = fl->ifl_bus_addrs;
1423 	iru->iru_idxs = fl->ifl_rxd_idxs;
1424 	iru->iru_qsidx = rxq->ifr_id;
1425 	iru->iru_buf_size = fl->ifl_buf_size;
1426 	iru->iru_flidx = fl->ifl_id;
1427 }
1428 
1429 static void
1430 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1431 {
1432 	if (err)
1433 		return;
1434 	*(bus_addr_t *) arg = segs[0].ds_addr;
1435 }
1436 
1437 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1438 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1439 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1440 
1441 int
1442 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1443 {
1444 	int err;
1445 	device_t dev = ctx->ifc_dev;
1446 	bus_addr_t lowaddr;
1447 
1448 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1449 
1450 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1451 				align, 0,		/* alignment, bounds */
1452 				lowaddr,		/* lowaddr */
1453 				BUS_SPACE_MAXADDR,	/* highaddr */
1454 				NULL, NULL,		/* filter, filterarg */
1455 				size,			/* maxsize */
1456 				1,			/* nsegments */
1457 				size,			/* maxsegsize */
1458 				BUS_DMA_ALLOCNOW,	/* flags */
1459 				NULL,			/* lockfunc */
1460 				NULL,			/* lockarg */
1461 				&dma->idi_tag);
1462 	if (err) {
1463 		device_printf(dev,
1464 		    "%s: bus_dma_tag_create failed: %d\n",
1465 		    __func__, err);
1466 		goto fail_0;
1467 	}
1468 
1469 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1470 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1471 	if (err) {
1472 		device_printf(dev,
1473 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1474 		    __func__, (uintmax_t)size, err);
1475 		goto fail_1;
1476 	}
1477 
1478 	dma->idi_paddr = IF_BAD_DMA;
1479 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1480 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1481 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1482 		device_printf(dev,
1483 		    "%s: bus_dmamap_load failed: %d\n",
1484 		    __func__, err);
1485 		goto fail_2;
1486 	}
1487 
1488 	dma->idi_size = size;
1489 	return (0);
1490 
1491 fail_2:
1492 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1493 fail_1:
1494 	bus_dma_tag_destroy(dma->idi_tag);
1495 fail_0:
1496 	dma->idi_tag = NULL;
1497 
1498 	return (err);
1499 }
1500 
1501 int
1502 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1503 {
1504 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1505 
1506 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1507 
1508 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1509 }
1510 
1511 int
1512 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1513 {
1514 	int i, err;
1515 	iflib_dma_info_t *dmaiter;
1516 
1517 	dmaiter = dmalist;
1518 	for (i = 0; i < count; i++, dmaiter++) {
1519 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1520 			break;
1521 	}
1522 	if (err)
1523 		iflib_dma_free_multi(dmalist, i);
1524 	return (err);
1525 }
1526 
1527 void
1528 iflib_dma_free(iflib_dma_info_t dma)
1529 {
1530 	if (dma->idi_tag == NULL)
1531 		return;
1532 	if (dma->idi_paddr != IF_BAD_DMA) {
1533 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1534 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1535 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1536 		dma->idi_paddr = IF_BAD_DMA;
1537 	}
1538 	if (dma->idi_vaddr != NULL) {
1539 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1540 		dma->idi_vaddr = NULL;
1541 	}
1542 	bus_dma_tag_destroy(dma->idi_tag);
1543 	dma->idi_tag = NULL;
1544 }
1545 
1546 void
1547 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1548 {
1549 	int i;
1550 	iflib_dma_info_t *dmaiter = dmalist;
1551 
1552 	for (i = 0; i < count; i++, dmaiter++)
1553 		iflib_dma_free(*dmaiter);
1554 }
1555 
1556 static int
1557 iflib_fast_intr(void *arg)
1558 {
1559 	iflib_filter_info_t info = arg;
1560 	struct grouptask *gtask = info->ifi_task;
1561 	int result;
1562 
1563 	DBG_COUNTER_INC(fast_intrs);
1564 	if (info->ifi_filter != NULL) {
1565 		result = info->ifi_filter(info->ifi_filter_arg);
1566 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1567 			return (result);
1568 	}
1569 
1570 	GROUPTASK_ENQUEUE(gtask);
1571 	return (FILTER_HANDLED);
1572 }
1573 
1574 static int
1575 iflib_fast_intr_rxtx(void *arg)
1576 {
1577 	iflib_filter_info_t info = arg;
1578 	struct grouptask *gtask = info->ifi_task;
1579 	if_ctx_t ctx;
1580 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1581 	iflib_txq_t txq;
1582 	void *sc;
1583 	int i, cidx, result;
1584 	qidx_t txqid;
1585 	bool intr_enable, intr_legacy;
1586 
1587 	DBG_COUNTER_INC(fast_intrs);
1588 	if (info->ifi_filter != NULL) {
1589 		result = info->ifi_filter(info->ifi_filter_arg);
1590 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1591 			return (result);
1592 	}
1593 
1594 	ctx = rxq->ifr_ctx;
1595 	sc = ctx->ifc_softc;
1596 	intr_enable = false;
1597 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1598 	MPASS(rxq->ifr_ntxqirq);
1599 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1600 		txqid = rxq->ifr_txqid[i];
1601 		txq = &ctx->ifc_txqs[txqid];
1602 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1603 		    BUS_DMASYNC_POSTREAD);
1604 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1605 			if (intr_legacy)
1606 				intr_enable = true;
1607 			else
1608 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1609 			continue;
1610 		}
1611 		GROUPTASK_ENQUEUE(&txq->ift_task);
1612 	}
1613 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1614 		cidx = rxq->ifr_cq_cidx;
1615 	else
1616 		cidx = rxq->ifr_fl[0].ifl_cidx;
1617 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1618 		GROUPTASK_ENQUEUE(gtask);
1619 	else {
1620 		if (intr_legacy)
1621 			intr_enable = true;
1622 		else
1623 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1624 		DBG_COUNTER_INC(rx_intr_enables);
1625 	}
1626 	if (intr_enable)
1627 		IFDI_INTR_ENABLE(ctx);
1628 	return (FILTER_HANDLED);
1629 }
1630 
1631 static int
1632 iflib_fast_intr_ctx(void *arg)
1633 {
1634 	iflib_filter_info_t info = arg;
1635 	struct grouptask *gtask = info->ifi_task;
1636 	int result;
1637 
1638 	DBG_COUNTER_INC(fast_intrs);
1639 	if (info->ifi_filter != NULL) {
1640 		result = info->ifi_filter(info->ifi_filter_arg);
1641 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1642 			return (result);
1643 	}
1644 
1645 	if (gtask->gt_taskqueue != NULL)
1646 		GROUPTASK_ENQUEUE(gtask);
1647 	return (FILTER_HANDLED);
1648 }
1649 
1650 static int
1651 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1652 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1653 		 const char *name)
1654 {
1655 	struct resource *res;
1656 	void *tag = NULL;
1657 	device_t dev = ctx->ifc_dev;
1658 	int flags, i, rc;
1659 
1660 	flags = RF_ACTIVE;
1661 	if (ctx->ifc_flags & IFC_LEGACY)
1662 		flags |= RF_SHAREABLE;
1663 	MPASS(rid < 512);
1664 	i = rid;
1665 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1666 	if (res == NULL) {
1667 		device_printf(dev,
1668 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1669 		return (ENOMEM);
1670 	}
1671 	irq->ii_res = res;
1672 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1673 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1674 						filter, handler, arg, &tag);
1675 	if (rc != 0) {
1676 		device_printf(dev,
1677 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1678 					  rid, name ? name : "unknown", rc);
1679 		return (rc);
1680 	} else if (name)
1681 		bus_describe_intr(dev, res, tag, "%s", name);
1682 
1683 	irq->ii_tag = tag;
1684 	return (0);
1685 }
1686 
1687 /*********************************************************************
1688  *
1689  *  Allocate DMA resources for TX buffers as well as memory for the TX
1690  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1691  *  iflib_sw_tx_desc_array structure, storing all the information that
1692  *  is needed to transmit a packet on the wire.  This is called only
1693  *  once at attach, setup is done every reset.
1694  *
1695  **********************************************************************/
1696 static int
1697 iflib_txsd_alloc(iflib_txq_t txq)
1698 {
1699 	if_ctx_t ctx = txq->ift_ctx;
1700 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1701 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1702 	device_t dev = ctx->ifc_dev;
1703 	bus_size_t tsomaxsize;
1704 	bus_addr_t lowaddr;
1705 	int err, nsegments, ntsosegments;
1706 	bool tso;
1707 
1708 	nsegments = scctx->isc_tx_nsegments;
1709 	ntsosegments = scctx->isc_tx_tso_segments_max;
1710 	tsomaxsize = scctx->isc_tx_tso_size_max;
1711 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1712 		tsomaxsize += sizeof(struct ether_vlan_header);
1713 	MPASS(scctx->isc_ntxd[0] > 0);
1714 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1715 	MPASS(nsegments > 0);
1716 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1717 		MPASS(ntsosegments > 0);
1718 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1719 	}
1720 
1721 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1722 
1723 	/*
1724 	 * Set up DMA tags for TX buffers.
1725 	 */
1726 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1727 			       1, 0,			/* alignment, bounds */
1728 			       lowaddr,			/* lowaddr */
1729 			       BUS_SPACE_MAXADDR,	/* highaddr */
1730 			       NULL, NULL,		/* filter, filterarg */
1731 			       sctx->isc_tx_maxsize,		/* maxsize */
1732 			       nsegments,	/* nsegments */
1733 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1734 			       0,			/* flags */
1735 			       NULL,			/* lockfunc */
1736 			       NULL,			/* lockfuncarg */
1737 			       &txq->ift_buf_tag))) {
1738 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1739 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1740 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1741 		goto fail;
1742 	}
1743 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1744 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1745 			       1, 0,			/* alignment, bounds */
1746 			       lowaddr,			/* lowaddr */
1747 			       BUS_SPACE_MAXADDR,	/* highaddr */
1748 			       NULL, NULL,		/* filter, filterarg */
1749 			       tsomaxsize,		/* maxsize */
1750 			       ntsosegments,	/* nsegments */
1751 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1752 			       0,			/* flags */
1753 			       NULL,			/* lockfunc */
1754 			       NULL,			/* lockfuncarg */
1755 			       &txq->ift_tso_buf_tag))) {
1756 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1757 		    err);
1758 		goto fail;
1759 	}
1760 
1761 	/* Allocate memory for the TX mbuf map. */
1762 	if (!(txq->ift_sds.ifsd_m =
1763 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1764 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1765 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1766 		err = ENOMEM;
1767 		goto fail;
1768 	}
1769 
1770 	/*
1771 	 * Create the DMA maps for TX buffers.
1772 	 */
1773 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1774 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1775 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1776 		device_printf(dev,
1777 		    "Unable to allocate TX buffer DMA map memory\n");
1778 		err = ENOMEM;
1779 		goto fail;
1780 	}
1781 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1782 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1783 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1784 		device_printf(dev,
1785 		    "Unable to allocate TSO TX buffer map memory\n");
1786 		err = ENOMEM;
1787 		goto fail;
1788 	}
1789 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1790 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1791 		    &txq->ift_sds.ifsd_map[i]);
1792 		if (err != 0) {
1793 			device_printf(dev, "Unable to create TX DMA map\n");
1794 			goto fail;
1795 		}
1796 		if (!tso)
1797 			continue;
1798 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1799 		    &txq->ift_sds.ifsd_tso_map[i]);
1800 		if (err != 0) {
1801 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1802 			goto fail;
1803 		}
1804 	}
1805 	return (0);
1806 fail:
1807 	/* We free all, it handles case where we are in the middle */
1808 	iflib_tx_structures_free(ctx);
1809 	return (err);
1810 }
1811 
1812 static void
1813 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1814 {
1815 	bus_dmamap_t map;
1816 
1817 	if (txq->ift_sds.ifsd_map != NULL) {
1818 		map = txq->ift_sds.ifsd_map[i];
1819 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1820 		bus_dmamap_unload(txq->ift_buf_tag, map);
1821 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1822 		txq->ift_sds.ifsd_map[i] = NULL;
1823 	}
1824 
1825 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1826 		map = txq->ift_sds.ifsd_tso_map[i];
1827 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1828 		    BUS_DMASYNC_POSTWRITE);
1829 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1830 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1831 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1832 	}
1833 }
1834 
1835 static void
1836 iflib_txq_destroy(iflib_txq_t txq)
1837 {
1838 	if_ctx_t ctx = txq->ift_ctx;
1839 
1840 	for (int i = 0; i < txq->ift_size; i++)
1841 		iflib_txsd_destroy(ctx, txq, i);
1842 
1843 	if (txq->ift_br != NULL) {
1844 		ifmp_ring_free(txq->ift_br);
1845 		txq->ift_br = NULL;
1846 	}
1847 
1848 	mtx_destroy(&txq->ift_mtx);
1849 
1850 	if (txq->ift_sds.ifsd_map != NULL) {
1851 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1852 		txq->ift_sds.ifsd_map = NULL;
1853 	}
1854 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1855 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1856 		txq->ift_sds.ifsd_tso_map = NULL;
1857 	}
1858 	if (txq->ift_sds.ifsd_m != NULL) {
1859 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1860 		txq->ift_sds.ifsd_m = NULL;
1861 	}
1862 	if (txq->ift_buf_tag != NULL) {
1863 		bus_dma_tag_destroy(txq->ift_buf_tag);
1864 		txq->ift_buf_tag = NULL;
1865 	}
1866 	if (txq->ift_tso_buf_tag != NULL) {
1867 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1868 		txq->ift_tso_buf_tag = NULL;
1869 	}
1870 	if (txq->ift_ifdi != NULL) {
1871 		free(txq->ift_ifdi, M_IFLIB);
1872 	}
1873 }
1874 
1875 static void
1876 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1877 {
1878 	struct mbuf **mp;
1879 
1880 	mp = &txq->ift_sds.ifsd_m[i];
1881 	if (*mp == NULL)
1882 		return;
1883 
1884 	if (txq->ift_sds.ifsd_map != NULL) {
1885 		bus_dmamap_sync(txq->ift_buf_tag,
1886 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1887 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1888 	}
1889 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1890 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1891 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1892 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1893 		    txq->ift_sds.ifsd_tso_map[i]);
1894 	}
1895 	m_freem(*mp);
1896 	DBG_COUNTER_INC(tx_frees);
1897 	*mp = NULL;
1898 }
1899 
1900 static int
1901 iflib_txq_setup(iflib_txq_t txq)
1902 {
1903 	if_ctx_t ctx = txq->ift_ctx;
1904 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1905 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1906 	iflib_dma_info_t di;
1907 	int i;
1908 
1909 	/* Set number of descriptors available */
1910 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1911 	/* XXX make configurable */
1912 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1913 
1914 	/* Reset indices */
1915 	txq->ift_cidx_processed = 0;
1916 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1917 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1918 
1919 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1920 		bzero((void *)di->idi_vaddr, di->idi_size);
1921 
1922 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1923 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1924 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1925 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1926 	return (0);
1927 }
1928 
1929 /*********************************************************************
1930  *
1931  *  Allocate DMA resources for RX buffers as well as memory for the RX
1932  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1933  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1934  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1935  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1936  *  packet, the maximum number of entries we'll need is equal to the
1937  *  number of hardware receive descriptors that we've allocated.
1938  *
1939  **********************************************************************/
1940 static int
1941 iflib_rxsd_alloc(iflib_rxq_t rxq)
1942 {
1943 	if_ctx_t ctx = rxq->ifr_ctx;
1944 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1945 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1946 	device_t dev = ctx->ifc_dev;
1947 	iflib_fl_t fl;
1948 	bus_addr_t lowaddr;
1949 	int			err;
1950 
1951 	MPASS(scctx->isc_nrxd[0] > 0);
1952 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1953 
1954 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1955 
1956 	fl = rxq->ifr_fl;
1957 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1958 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1959 		/* Set up DMA tag for RX buffers. */
1960 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1961 					 1, 0,			/* alignment, bounds */
1962 					 lowaddr,		/* lowaddr */
1963 					 BUS_SPACE_MAXADDR,	/* highaddr */
1964 					 NULL, NULL,		/* filter, filterarg */
1965 					 sctx->isc_rx_maxsize,	/* maxsize */
1966 					 sctx->isc_rx_nsegments,	/* nsegments */
1967 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1968 					 0,			/* flags */
1969 					 NULL,			/* lockfunc */
1970 					 NULL,			/* lockarg */
1971 					 &fl->ifl_buf_tag);
1972 		if (err) {
1973 			device_printf(dev,
1974 			    "Unable to allocate RX DMA tag: %d\n", err);
1975 			goto fail;
1976 		}
1977 
1978 		/* Allocate memory for the RX mbuf map. */
1979 		if (!(fl->ifl_sds.ifsd_m =
1980 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1981 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1982 			device_printf(dev,
1983 			    "Unable to allocate RX mbuf map memory\n");
1984 			err = ENOMEM;
1985 			goto fail;
1986 		}
1987 
1988 		/* Allocate memory for the direct RX cluster pointer map. */
1989 		if (!(fl->ifl_sds.ifsd_cl =
1990 		      (caddr_t *) malloc(sizeof(caddr_t) *
1991 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1992 			device_printf(dev,
1993 			    "Unable to allocate RX cluster map memory\n");
1994 			err = ENOMEM;
1995 			goto fail;
1996 		}
1997 
1998 		/* Allocate memory for the RX cluster bus address map. */
1999 		if (!(fl->ifl_sds.ifsd_ba =
2000 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2001 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2002 			device_printf(dev,
2003 			    "Unable to allocate RX bus address map memory\n");
2004 			err = ENOMEM;
2005 			goto fail;
2006 		}
2007 
2008 		/*
2009 		 * Create the DMA maps for RX buffers.
2010 		 */
2011 		if (!(fl->ifl_sds.ifsd_map =
2012 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2013 			device_printf(dev,
2014 			    "Unable to allocate RX buffer DMA map memory\n");
2015 			err = ENOMEM;
2016 			goto fail;
2017 		}
2018 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2019 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2020 			    &fl->ifl_sds.ifsd_map[i]);
2021 			if (err != 0) {
2022 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2023 				goto fail;
2024 			}
2025 		}
2026 	}
2027 	return (0);
2028 
2029 fail:
2030 	iflib_rx_structures_free(ctx);
2031 	return (err);
2032 }
2033 
2034 /*
2035  * Internal service routines
2036  */
2037 
2038 struct rxq_refill_cb_arg {
2039 	int               error;
2040 	bus_dma_segment_t seg;
2041 	int               nseg;
2042 };
2043 
2044 static void
2045 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2046 {
2047 	struct rxq_refill_cb_arg *cb_arg = arg;
2048 
2049 	cb_arg->error = error;
2050 	cb_arg->seg = segs[0];
2051 	cb_arg->nseg = nseg;
2052 }
2053 
2054 /**
2055  * iflib_fl_refill - refill an rxq free-buffer list
2056  * @ctx: the iflib context
2057  * @fl: the free list to refill
2058  * @count: the number of new buffers to allocate
2059  *
2060  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2061  * The caller must assure that @count does not exceed the queue's capacity
2062  * minus one (since we always leave a descriptor unavailable).
2063  */
2064 static uint8_t
2065 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2066 {
2067 	struct if_rxd_update iru;
2068 	struct rxq_refill_cb_arg cb_arg;
2069 	struct mbuf *m;
2070 	caddr_t cl, *sd_cl;
2071 	struct mbuf **sd_m;
2072 	bus_dmamap_t *sd_map;
2073 	bus_addr_t bus_addr, *sd_ba;
2074 	int err, frag_idx, i, idx, n, pidx;
2075 	qidx_t credits;
2076 
2077 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2078 
2079 	sd_m = fl->ifl_sds.ifsd_m;
2080 	sd_map = fl->ifl_sds.ifsd_map;
2081 	sd_cl = fl->ifl_sds.ifsd_cl;
2082 	sd_ba = fl->ifl_sds.ifsd_ba;
2083 	pidx = fl->ifl_pidx;
2084 	idx = pidx;
2085 	frag_idx = fl->ifl_fragidx;
2086 	credits = fl->ifl_credits;
2087 
2088 	i = 0;
2089 	n = count;
2090 	MPASS(n > 0);
2091 	MPASS(credits + n <= fl->ifl_size);
2092 
2093 	if (pidx < fl->ifl_cidx)
2094 		MPASS(pidx + n <= fl->ifl_cidx);
2095 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2096 		MPASS(fl->ifl_gen == 0);
2097 	if (pidx > fl->ifl_cidx)
2098 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2099 
2100 	DBG_COUNTER_INC(fl_refills);
2101 	if (n > 8)
2102 		DBG_COUNTER_INC(fl_refills_large);
2103 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2104 	while (n-- > 0) {
2105 		/*
2106 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2107 		 * initialized after rx.
2108 		 *
2109 		 * If the cluster is still set then we know a minimum sized
2110 		 * packet was received
2111 		 */
2112 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2113 		    &frag_idx);
2114 		if (frag_idx < 0)
2115 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2116 		MPASS(frag_idx >= 0);
2117 		if ((cl = sd_cl[frag_idx]) == NULL) {
2118 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2119 			if (__predict_false(cl == NULL))
2120 				break;
2121 
2122 			cb_arg.error = 0;
2123 			MPASS(sd_map != NULL);
2124 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2125 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2126 			    BUS_DMA_NOWAIT);
2127 			if (__predict_false(err != 0 || cb_arg.error)) {
2128 				uma_zfree(fl->ifl_zone, cl);
2129 				break;
2130 			}
2131 
2132 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2133 			sd_cl[frag_idx] = cl;
2134 #if MEMORY_LOGGING
2135 			fl->ifl_cl_enqueued++;
2136 #endif
2137 		} else {
2138 			bus_addr = sd_ba[frag_idx];
2139 		}
2140 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2141 		    BUS_DMASYNC_PREREAD);
2142 
2143 		if (sd_m[frag_idx] == NULL) {
2144 			m = m_gethdr_raw(M_NOWAIT, 0);
2145 			if (__predict_false(m == NULL))
2146 				break;
2147 			sd_m[frag_idx] = m;
2148 		}
2149 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2150 #if MEMORY_LOGGING
2151 		fl->ifl_m_enqueued++;
2152 #endif
2153 
2154 		DBG_COUNTER_INC(rx_allocs);
2155 		fl->ifl_rxd_idxs[i] = frag_idx;
2156 		fl->ifl_bus_addrs[i] = bus_addr;
2157 		credits++;
2158 		i++;
2159 		MPASS(credits <= fl->ifl_size);
2160 		if (++idx == fl->ifl_size) {
2161 #ifdef INVARIANTS
2162 			fl->ifl_gen = 1;
2163 #endif
2164 			idx = 0;
2165 		}
2166 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2167 			iru.iru_pidx = pidx;
2168 			iru.iru_count = i;
2169 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2170 			fl->ifl_pidx = idx;
2171 			fl->ifl_credits = credits;
2172 			pidx = idx;
2173 			i = 0;
2174 		}
2175 	}
2176 
2177 	if (n < count - 1) {
2178 		if (i != 0) {
2179 			iru.iru_pidx = pidx;
2180 			iru.iru_count = i;
2181 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2182 			fl->ifl_pidx = idx;
2183 			fl->ifl_credits = credits;
2184 		}
2185 		DBG_COUNTER_INC(rxd_flush);
2186 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2187 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2188 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2189 		    fl->ifl_id, fl->ifl_pidx);
2190 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2191 			fl->ifl_fragidx = frag_idx + 1;
2192 			if (fl->ifl_fragidx == fl->ifl_size)
2193 				fl->ifl_fragidx = 0;
2194 		} else {
2195 			fl->ifl_fragidx = frag_idx;
2196 		}
2197 	}
2198 
2199 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2200 }
2201 
2202 static inline uint8_t
2203 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2204 {
2205 	/*
2206 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2207 	 * This is important as it confuses most NICs. For instance,
2208 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2209 	 * RDH points to the next receive descriptor to be used by the NIC,
2210 	 * and RDT for the next receive descriptor to be published by the
2211 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2212 	 * The condition RDH == RDT means no descriptors are available to
2213 	 * the NIC, and thus it would be ambiguous if it also meant that
2214 	 * all the descriptors are available to the NIC.
2215 	 */
2216 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2217 #ifdef INVARIANTS
2218 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2219 #endif
2220 
2221 	MPASS(fl->ifl_credits <= fl->ifl_size);
2222 	MPASS(reclaimable == delta);
2223 
2224 	if (reclaimable > 0)
2225 		return (iflib_fl_refill(ctx, fl, reclaimable));
2226 	return (0);
2227 }
2228 
2229 uint8_t
2230 iflib_in_detach(if_ctx_t ctx)
2231 {
2232 	bool in_detach;
2233 
2234 	STATE_LOCK(ctx);
2235 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2236 	STATE_UNLOCK(ctx);
2237 	return (in_detach);
2238 }
2239 
2240 static void
2241 iflib_fl_bufs_free(iflib_fl_t fl)
2242 {
2243 	iflib_dma_info_t idi = fl->ifl_ifdi;
2244 	bus_dmamap_t sd_map;
2245 	uint32_t i;
2246 
2247 	for (i = 0; i < fl->ifl_size; i++) {
2248 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2249 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2250 
2251 		if (*sd_cl != NULL) {
2252 			sd_map = fl->ifl_sds.ifsd_map[i];
2253 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2254 			    BUS_DMASYNC_POSTREAD);
2255 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2256 			uma_zfree(fl->ifl_zone, *sd_cl);
2257 			*sd_cl = NULL;
2258 			if (*sd_m != NULL) {
2259 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2260 				m_free_raw(*sd_m);
2261 				*sd_m = NULL;
2262 			}
2263 		} else {
2264 			MPASS(*sd_m == NULL);
2265 		}
2266 #if MEMORY_LOGGING
2267 		fl->ifl_m_dequeued++;
2268 		fl->ifl_cl_dequeued++;
2269 #endif
2270 	}
2271 #ifdef INVARIANTS
2272 	for (i = 0; i < fl->ifl_size; i++) {
2273 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2274 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2275 	}
2276 #endif
2277 	/*
2278 	 * Reset free list values
2279 	 */
2280 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2281 	bzero(idi->idi_vaddr, idi->idi_size);
2282 }
2283 
2284 /*********************************************************************
2285  *
2286  *  Initialize a free list and its buffers.
2287  *
2288  **********************************************************************/
2289 static int
2290 iflib_fl_setup(iflib_fl_t fl)
2291 {
2292 	iflib_rxq_t rxq = fl->ifl_rxq;
2293 	if_ctx_t ctx = rxq->ifr_ctx;
2294 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2295 	int qidx;
2296 
2297 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2298 	/*
2299 	** Free current RX buffer structs and their mbufs
2300 	*/
2301 	iflib_fl_bufs_free(fl);
2302 	/* Now replenish the mbufs */
2303 	MPASS(fl->ifl_credits == 0);
2304 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2305 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2306 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2307 	else
2308 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2309 	/*
2310 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2311 	 * to the selected mbuf size.
2312 	 */
2313 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2314 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2315 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2316 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2317 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2318 
2319 	/*
2320 	 * Avoid pre-allocating zillions of clusters to an idle card
2321 	 * potentially speeding up attach. In any case make sure
2322 	 * to leave a descriptor unavailable. See the comment in
2323 	 * iflib_fl_refill_all().
2324 	 */
2325 	MPASS(fl->ifl_size > 0);
2326 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2327 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2328 		return (ENOBUFS);
2329 	/*
2330 	 * handle failure
2331 	 */
2332 	MPASS(rxq != NULL);
2333 	MPASS(fl->ifl_ifdi != NULL);
2334 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2335 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2336 	return (0);
2337 }
2338 
2339 /*********************************************************************
2340  *
2341  *  Free receive ring data structures
2342  *
2343  **********************************************************************/
2344 static void
2345 iflib_rx_sds_free(iflib_rxq_t rxq)
2346 {
2347 	iflib_fl_t fl;
2348 	int i, j;
2349 
2350 	if (rxq->ifr_fl != NULL) {
2351 		for (i = 0; i < rxq->ifr_nfl; i++) {
2352 			fl = &rxq->ifr_fl[i];
2353 			if (fl->ifl_buf_tag != NULL) {
2354 				if (fl->ifl_sds.ifsd_map != NULL) {
2355 					for (j = 0; j < fl->ifl_size; j++) {
2356 						bus_dmamap_sync(
2357 						    fl->ifl_buf_tag,
2358 						    fl->ifl_sds.ifsd_map[j],
2359 						    BUS_DMASYNC_POSTREAD);
2360 						bus_dmamap_unload(
2361 						    fl->ifl_buf_tag,
2362 						    fl->ifl_sds.ifsd_map[j]);
2363 						bus_dmamap_destroy(
2364 						    fl->ifl_buf_tag,
2365 						    fl->ifl_sds.ifsd_map[j]);
2366 					}
2367 				}
2368 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2369 				fl->ifl_buf_tag = NULL;
2370 			}
2371 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2372 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2373 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2374 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2375 			free(fl->ifl_rx_bitmap, M_IFLIB);
2376 			fl->ifl_sds.ifsd_m = NULL;
2377 			fl->ifl_sds.ifsd_cl = NULL;
2378 			fl->ifl_sds.ifsd_ba = NULL;
2379 			fl->ifl_sds.ifsd_map = NULL;
2380 			fl->ifl_rx_bitmap = NULL;
2381 		}
2382 		free(rxq->ifr_fl, M_IFLIB);
2383 		rxq->ifr_fl = NULL;
2384 		free(rxq->ifr_ifdi, M_IFLIB);
2385 		rxq->ifr_ifdi = NULL;
2386 		rxq->ifr_cq_cidx = 0;
2387 	}
2388 }
2389 
2390 /*
2391  * Timer routine
2392  */
2393 static void
2394 iflib_timer(void *arg)
2395 {
2396 	iflib_txq_t txq = arg;
2397 	if_ctx_t ctx = txq->ift_ctx;
2398 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2399 	uint64_t this_tick = ticks;
2400 
2401 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2402 		return;
2403 
2404 	/*
2405 	** Check on the state of the TX queue(s), this
2406 	** can be done without the lock because its RO
2407 	** and the HUNG state will be static if set.
2408 	*/
2409 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2410 		txq->ift_last_timer_tick = this_tick;
2411 		IFDI_TIMER(ctx, txq->ift_id);
2412 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2413 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2414 		     (sctx->isc_pause_frames == 0)))
2415 			goto hung;
2416 
2417 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2418 		    ifmp_ring_is_stalled(txq->ift_br)) {
2419 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2420 			    ("queue can't be marked as hung if interface is down"));
2421 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2422 		}
2423 		txq->ift_cleaned_prev = txq->ift_cleaned;
2424 	}
2425 	/* handle any laggards */
2426 	if (txq->ift_db_pending)
2427 		GROUPTASK_ENQUEUE(&txq->ift_task);
2428 
2429 	sctx->isc_pause_frames = 0;
2430 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2431 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2432 		    txq, txq->ift_timer.c_cpu);
2433 	return;
2434 
2435  hung:
2436 	device_printf(ctx->ifc_dev,
2437 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2438 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2439 	STATE_LOCK(ctx);
2440 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2441 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2442 	iflib_admin_intr_deferred(ctx);
2443 	STATE_UNLOCK(ctx);
2444 }
2445 
2446 static uint16_t
2447 iflib_get_mbuf_size_for(unsigned int size)
2448 {
2449 
2450 	if (size <= MCLBYTES)
2451 		return (MCLBYTES);
2452 	else
2453 		return (MJUMPAGESIZE);
2454 }
2455 
2456 static void
2457 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2458 {
2459 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2460 
2461 	/*
2462 	 * XXX don't set the max_frame_size to larger
2463 	 * than the hardware can handle
2464 	 */
2465 	ctx->ifc_rx_mbuf_sz =
2466 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2467 }
2468 
2469 uint32_t
2470 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2471 {
2472 
2473 	return (ctx->ifc_rx_mbuf_sz);
2474 }
2475 
2476 static void
2477 iflib_init_locked(if_ctx_t ctx)
2478 {
2479 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2480 	if_t ifp = ctx->ifc_ifp;
2481 	iflib_fl_t fl;
2482 	iflib_txq_t txq;
2483 	iflib_rxq_t rxq;
2484 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2485 
2486 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2487 	IFDI_INTR_DISABLE(ctx);
2488 
2489 	/*
2490 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2491 	 * called without first calling iflib_stop().
2492 	 */
2493 	netmap_disable_all_rings(ifp);
2494 
2495 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2496 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2497 	/* Set hardware offload abilities */
2498 	if_clearhwassist(ifp);
2499 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2500 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2501 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2502 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2503 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2504 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2505 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2506 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2507 
2508 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2509 		CALLOUT_LOCK(txq);
2510 		callout_stop(&txq->ift_timer);
2511 #ifdef DEV_NETMAP
2512 		callout_stop(&txq->ift_netmap_timer);
2513 #endif /* DEV_NETMAP */
2514 		CALLOUT_UNLOCK(txq);
2515 		(void)iflib_netmap_txq_init(ctx, txq);
2516 	}
2517 
2518 	/*
2519 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2520 	 * that drivers can use the value when setting up the hardware receive
2521 	 * buffers.
2522 	 */
2523 	iflib_calc_rx_mbuf_sz(ctx);
2524 
2525 #ifdef INVARIANTS
2526 	i = if_getdrvflags(ifp);
2527 #endif
2528 	IFDI_INIT(ctx);
2529 	MPASS(if_getdrvflags(ifp) == i);
2530 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2531 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2532 			/* This rxq is in netmap mode. Skip normal init. */
2533 			continue;
2534 		}
2535 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2536 			if (iflib_fl_setup(fl)) {
2537 				device_printf(ctx->ifc_dev,
2538 				    "setting up free list %d failed - "
2539 				    "check cluster settings\n", j);
2540 				goto done;
2541 			}
2542 		}
2543 	}
2544 done:
2545 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2546 	IFDI_INTR_ENABLE(ctx);
2547 	txq = ctx->ifc_txqs;
2548 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2549 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2550 			txq->ift_timer.c_cpu);
2551 
2552         /* Re-enable txsync/rxsync. */
2553 	netmap_enable_all_rings(ifp);
2554 }
2555 
2556 static int
2557 iflib_media_change(if_t ifp)
2558 {
2559 	if_ctx_t ctx = if_getsoftc(ifp);
2560 	int err;
2561 
2562 	CTX_LOCK(ctx);
2563 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2564 		iflib_if_init_locked(ctx);
2565 	CTX_UNLOCK(ctx);
2566 	return (err);
2567 }
2568 
2569 static void
2570 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2571 {
2572 	if_ctx_t ctx = if_getsoftc(ifp);
2573 
2574 	CTX_LOCK(ctx);
2575 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2576 	IFDI_MEDIA_STATUS(ctx, ifmr);
2577 	CTX_UNLOCK(ctx);
2578 }
2579 
2580 void
2581 iflib_stop(if_ctx_t ctx)
2582 {
2583 	iflib_txq_t txq = ctx->ifc_txqs;
2584 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2585 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2586 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2587 	iflib_dma_info_t di;
2588 	iflib_fl_t fl;
2589 	int i, j;
2590 
2591 	/* Tell the stack that the interface is no longer active */
2592 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2593 
2594 	IFDI_INTR_DISABLE(ctx);
2595 	DELAY(1000);
2596 	IFDI_STOP(ctx);
2597 	DELAY(1000);
2598 
2599 	/*
2600 	 * Stop any pending txsync/rxsync and prevent new ones
2601 	 * form starting. Processes blocked in poll() will get
2602 	 * POLLERR.
2603 	 */
2604 	netmap_disable_all_rings(ctx->ifc_ifp);
2605 
2606 	iflib_debug_reset();
2607 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2608 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2609 		/* make sure all transmitters have completed before proceeding XXX */
2610 
2611 		CALLOUT_LOCK(txq);
2612 		callout_stop(&txq->ift_timer);
2613 #ifdef DEV_NETMAP
2614 		callout_stop(&txq->ift_netmap_timer);
2615 #endif /* DEV_NETMAP */
2616 		CALLOUT_UNLOCK(txq);
2617 
2618 		/* clean any enqueued buffers */
2619 		iflib_ifmp_purge(txq);
2620 		/* Free any existing tx buffers. */
2621 		for (j = 0; j < txq->ift_size; j++) {
2622 			iflib_txsd_free(ctx, txq, j);
2623 		}
2624 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2625 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2626 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2627 			txq->ift_cidx = txq->ift_pidx;
2628 		else
2629 			txq->ift_cidx = txq->ift_pidx = 0;
2630 
2631 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2632 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2633 		txq->ift_pullups = 0;
2634 		ifmp_ring_reset_stats(txq->ift_br);
2635 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2636 			bzero((void *)di->idi_vaddr, di->idi_size);
2637 	}
2638 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2639 		if (rxq->ifr_task.gt_taskqueue != NULL)
2640 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2641 				 &rxq->ifr_task.gt_task);
2642 
2643 		rxq->ifr_cq_cidx = 0;
2644 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2645 			bzero((void *)di->idi_vaddr, di->idi_size);
2646 		/* also resets the free lists pidx/cidx */
2647 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2648 			iflib_fl_bufs_free(fl);
2649 	}
2650 }
2651 
2652 static inline caddr_t
2653 calc_next_rxd(iflib_fl_t fl, int cidx)
2654 {
2655 	qidx_t size;
2656 	int nrxd;
2657 	caddr_t start, end, cur, next;
2658 
2659 	nrxd = fl->ifl_size;
2660 	size = fl->ifl_rxd_size;
2661 	start = fl->ifl_ifdi->idi_vaddr;
2662 
2663 	if (__predict_false(size == 0))
2664 		return (start);
2665 	cur = start + size*cidx;
2666 	end = start + size*nrxd;
2667 	next = CACHE_PTR_NEXT(cur);
2668 	return (next < end ? next : start);
2669 }
2670 
2671 static inline void
2672 prefetch_pkts(iflib_fl_t fl, int cidx)
2673 {
2674 	int nextptr;
2675 	int nrxd = fl->ifl_size;
2676 	caddr_t next_rxd;
2677 
2678 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2679 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2680 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2681 	next_rxd = calc_next_rxd(fl, cidx);
2682 	prefetch(next_rxd);
2683 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2684 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2685 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2686 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2687 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2688 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2689 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2690 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2691 }
2692 
2693 static struct mbuf *
2694 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2695     int *pf_rv, if_rxd_info_t ri)
2696 {
2697 	bus_dmamap_t map;
2698 	iflib_fl_t fl;
2699 	caddr_t payload;
2700 	struct mbuf *m;
2701 	int flid, cidx, len, next;
2702 
2703 	map = NULL;
2704 	flid = irf->irf_flid;
2705 	cidx = irf->irf_idx;
2706 	fl = &rxq->ifr_fl[flid];
2707 	sd->ifsd_fl = fl;
2708 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2709 	fl->ifl_credits--;
2710 #if MEMORY_LOGGING
2711 	fl->ifl_m_dequeued++;
2712 #endif
2713 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2714 		prefetch_pkts(fl, cidx);
2715 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2716 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2717 	map = fl->ifl_sds.ifsd_map[cidx];
2718 
2719 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2720 
2721 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2722 	    irf->irf_len != 0) {
2723 		payload  = *sd->ifsd_cl;
2724 		payload +=  ri->iri_pad;
2725 		len = ri->iri_len - ri->iri_pad;
2726 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2727 		switch (*pf_rv) {
2728 		case PFIL_DROPPED:
2729 		case PFIL_CONSUMED:
2730 			/*
2731 			 * The filter ate it.  Everything is recycled.
2732 			 */
2733 			m = NULL;
2734 			unload = 0;
2735 			break;
2736 		case PFIL_REALLOCED:
2737 			/*
2738 			 * The filter copied it.  Everything is recycled.
2739 			 * 'm' points at new mbuf.
2740 			 */
2741 			unload = 0;
2742 			break;
2743 		case PFIL_PASS:
2744 			/*
2745 			 * Filter said it was OK, so receive like
2746 			 * normal
2747 			 */
2748 			m = fl->ifl_sds.ifsd_m[cidx];
2749 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2750 			break;
2751 		default:
2752 			MPASS(0);
2753 		}
2754 	} else {
2755 		m = fl->ifl_sds.ifsd_m[cidx];
2756 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2757 		if (pf_rv != NULL)
2758 			*pf_rv = PFIL_PASS;
2759 	}
2760 
2761 	if (unload && irf->irf_len != 0)
2762 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2763 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2764 	if (__predict_false(fl->ifl_cidx == 0))
2765 		fl->ifl_gen = 0;
2766 	bit_clear(fl->ifl_rx_bitmap, cidx);
2767 	return (m);
2768 }
2769 
2770 static struct mbuf *
2771 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2772 {
2773 	struct mbuf *m, *mh, *mt;
2774 	caddr_t cl;
2775 	int  *pf_rv_ptr, flags, i, padlen;
2776 	bool consumed;
2777 
2778 	i = 0;
2779 	mh = NULL;
2780 	consumed = false;
2781 	*pf_rv = PFIL_PASS;
2782 	pf_rv_ptr = pf_rv;
2783 	do {
2784 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2785 		    pf_rv_ptr, ri);
2786 
2787 		MPASS(*sd->ifsd_cl != NULL);
2788 
2789 		/*
2790 		 * Exclude zero-length frags & frags from
2791 		 * packets the filter has consumed or dropped
2792 		 */
2793 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2794 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2795 			if (mh == NULL) {
2796 				/* everything saved here */
2797 				consumed = true;
2798 				pf_rv_ptr = NULL;
2799 				continue;
2800 			}
2801 			/* XXX we can save the cluster here, but not the mbuf */
2802 			m_init(m, M_NOWAIT, MT_DATA, 0);
2803 			m_free(m);
2804 			continue;
2805 		}
2806 		if (mh == NULL) {
2807 			flags = M_PKTHDR|M_EXT;
2808 			mh = mt = m;
2809 			padlen = ri->iri_pad;
2810 		} else {
2811 			flags = M_EXT;
2812 			mt->m_next = m;
2813 			mt = m;
2814 			/* assuming padding is only on the first fragment */
2815 			padlen = 0;
2816 		}
2817 		cl = *sd->ifsd_cl;
2818 		*sd->ifsd_cl = NULL;
2819 
2820 		/* Can these two be made one ? */
2821 		m_init(m, M_NOWAIT, MT_DATA, flags);
2822 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2823 		/*
2824 		 * These must follow m_init and m_cljset
2825 		 */
2826 		m->m_data += padlen;
2827 		ri->iri_len -= padlen;
2828 		m->m_len = ri->iri_frags[i].irf_len;
2829 	} while (++i < ri->iri_nfrags);
2830 
2831 	return (mh);
2832 }
2833 
2834 /*
2835  * Process one software descriptor
2836  */
2837 static struct mbuf *
2838 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2839 {
2840 	struct if_rxsd sd;
2841 	struct mbuf *m;
2842 	int pf_rv;
2843 
2844 	/* should I merge this back in now that the two paths are basically duplicated? */
2845 	if (ri->iri_nfrags == 1 &&
2846 	    ri->iri_frags[0].irf_len != 0 &&
2847 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2848 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2849 		    &pf_rv, ri);
2850 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2851 			return (m);
2852 		if (pf_rv == PFIL_PASS) {
2853 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2854 #ifndef __NO_STRICT_ALIGNMENT
2855 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2856 				m->m_data += 2;
2857 #endif
2858 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2859 			m->m_len = ri->iri_frags[0].irf_len;
2860 			m->m_data += ri->iri_pad;
2861 			ri->iri_len -= ri->iri_pad;
2862 		}
2863 	} else {
2864 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2865 		if (m == NULL)
2866 			return (NULL);
2867 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2868 			return (m);
2869 	}
2870 	m->m_pkthdr.len = ri->iri_len;
2871 	m->m_pkthdr.rcvif = ri->iri_ifp;
2872 	m->m_flags |= ri->iri_flags;
2873 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2874 	m->m_pkthdr.flowid = ri->iri_flowid;
2875 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2876 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2877 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2878 	return (m);
2879 }
2880 
2881 #if defined(INET6) || defined(INET)
2882 static void
2883 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2884 {
2885 	CURVNET_SET(if_getvnet(lc->ifp));
2886 #if defined(INET6)
2887 	*v6 = V_ip6_forwarding;
2888 #endif
2889 #if defined(INET)
2890 	*v4 = V_ipforwarding;
2891 #endif
2892 	CURVNET_RESTORE();
2893 }
2894 
2895 /*
2896  * Returns true if it's possible this packet could be LROed.
2897  * if it returns false, it is guaranteed that tcp_lro_rx()
2898  * would not return zero.
2899  */
2900 static bool
2901 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2902 {
2903 	struct ether_header *eh;
2904 
2905 	eh = mtod(m, struct ether_header *);
2906 	switch (eh->ether_type) {
2907 #if defined(INET6)
2908 		case htons(ETHERTYPE_IPV6):
2909 			return (!v6_forwarding);
2910 #endif
2911 #if defined (INET)
2912 		case htons(ETHERTYPE_IP):
2913 			return (!v4_forwarding);
2914 #endif
2915 	}
2916 
2917 	return false;
2918 }
2919 #else
2920 static void
2921 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2922 {
2923 }
2924 #endif
2925 
2926 static void
2927 _task_fn_rx_watchdog(void *context)
2928 {
2929 	iflib_rxq_t rxq = context;
2930 
2931 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2932 }
2933 
2934 static uint8_t
2935 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2936 {
2937 	if_t ifp;
2938 	if_ctx_t ctx = rxq->ifr_ctx;
2939 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2940 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2941 	int avail, i;
2942 	qidx_t *cidxp;
2943 	struct if_rxd_info ri;
2944 	int err, budget_left, rx_bytes, rx_pkts;
2945 	iflib_fl_t fl;
2946 	int lro_enabled;
2947 	bool v4_forwarding, v6_forwarding, lro_possible;
2948 	uint8_t retval = 0;
2949 
2950 	/*
2951 	 * XXX early demux data packets so that if_input processing only handles
2952 	 * acks in interrupt context
2953 	 */
2954 	struct mbuf *m, *mh, *mt, *mf;
2955 
2956 	NET_EPOCH_ASSERT();
2957 
2958 	lro_possible = v4_forwarding = v6_forwarding = false;
2959 	ifp = ctx->ifc_ifp;
2960 	mh = mt = NULL;
2961 	MPASS(budget > 0);
2962 	rx_pkts	= rx_bytes = 0;
2963 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2964 		cidxp = &rxq->ifr_cq_cidx;
2965 	else
2966 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2967 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2968 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2969 			retval |= iflib_fl_refill_all(ctx, fl);
2970 		DBG_COUNTER_INC(rx_unavail);
2971 		return (retval);
2972 	}
2973 
2974 	/* pfil needs the vnet to be set */
2975 	CURVNET_SET_QUIET(if_getvnet(ifp));
2976 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2977 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2978 			DBG_COUNTER_INC(rx_ctx_inactive);
2979 			break;
2980 		}
2981 		/*
2982 		 * Reset client set fields to their default values
2983 		 */
2984 		rxd_info_zero(&ri);
2985 		ri.iri_qsidx = rxq->ifr_id;
2986 		ri.iri_cidx = *cidxp;
2987 		ri.iri_ifp = ifp;
2988 		ri.iri_frags = rxq->ifr_frags;
2989 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2990 
2991 		if (err)
2992 			goto err;
2993 		rx_pkts += 1;
2994 		rx_bytes += ri.iri_len;
2995 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2996 			*cidxp = ri.iri_cidx;
2997 			/* Update our consumer index */
2998 			/* XXX NB: shurd - check if this is still safe */
2999 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3000 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3001 			/* was this only a completion queue message? */
3002 			if (__predict_false(ri.iri_nfrags == 0))
3003 				continue;
3004 		}
3005 		MPASS(ri.iri_nfrags != 0);
3006 		MPASS(ri.iri_len != 0);
3007 
3008 		/* will advance the cidx on the corresponding free lists */
3009 		m = iflib_rxd_pkt_get(rxq, &ri);
3010 		avail--;
3011 		budget_left--;
3012 		if (avail == 0 && budget_left)
3013 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3014 
3015 		if (__predict_false(m == NULL))
3016 			continue;
3017 
3018 		/* imm_pkt: -- cxgb */
3019 		if (mh == NULL)
3020 			mh = mt = m;
3021 		else {
3022 			mt->m_nextpkt = m;
3023 			mt = m;
3024 		}
3025 	}
3026 	CURVNET_RESTORE();
3027 	/* make sure that we can refill faster than drain */
3028 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3029 		retval |= iflib_fl_refill_all(ctx, fl);
3030 
3031 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3032 	if (lro_enabled)
3033 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3034 	mt = mf = NULL;
3035 	while (mh != NULL) {
3036 		m = mh;
3037 		mh = mh->m_nextpkt;
3038 		m->m_nextpkt = NULL;
3039 #ifndef __NO_STRICT_ALIGNMENT
3040 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3041 			continue;
3042 #endif
3043 #if defined(INET6) || defined(INET)
3044 		if (lro_enabled) {
3045 			if (!lro_possible) {
3046 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3047 				if (lro_possible && mf != NULL) {
3048 					if_input(ifp, mf);
3049 					DBG_COUNTER_INC(rx_if_input);
3050 					mt = mf = NULL;
3051 				}
3052 			}
3053 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3054 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3055 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3056 					continue;
3057 			}
3058 		}
3059 #endif
3060 		if (lro_possible) {
3061 			if_input(ifp, m);
3062 			DBG_COUNTER_INC(rx_if_input);
3063 			continue;
3064 		}
3065 
3066 		if (mf == NULL)
3067 			mf = m;
3068 		if (mt != NULL)
3069 			mt->m_nextpkt = m;
3070 		mt = m;
3071 	}
3072 	if (mf != NULL) {
3073 		if_input(ifp, mf);
3074 		DBG_COUNTER_INC(rx_if_input);
3075 	}
3076 
3077 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3078 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3079 
3080 	/*
3081 	 * Flush any outstanding LRO work
3082 	 */
3083 #if defined(INET6) || defined(INET)
3084 	tcp_lro_flush_all(&rxq->ifr_lc);
3085 #endif
3086 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3087 		retval |= IFLIB_RXEOF_MORE;
3088 	return (retval);
3089 err:
3090 	STATE_LOCK(ctx);
3091 	ctx->ifc_flags |= IFC_DO_RESET;
3092 	iflib_admin_intr_deferred(ctx);
3093 	STATE_UNLOCK(ctx);
3094 	return (0);
3095 }
3096 
3097 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3098 static inline qidx_t
3099 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3100 {
3101 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3102 	qidx_t minthresh = txq->ift_size / 8;
3103 	if (in_use > 4*minthresh)
3104 		return (notify_count);
3105 	if (in_use > 2*minthresh)
3106 		return (notify_count >> 1);
3107 	if (in_use > minthresh)
3108 		return (notify_count >> 3);
3109 	return (0);
3110 }
3111 
3112 static inline qidx_t
3113 txq_max_rs_deferred(iflib_txq_t txq)
3114 {
3115 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3116 	qidx_t minthresh = txq->ift_size / 8;
3117 	if (txq->ift_in_use > 4*minthresh)
3118 		return (notify_count);
3119 	if (txq->ift_in_use > 2*minthresh)
3120 		return (notify_count >> 1);
3121 	if (txq->ift_in_use > minthresh)
3122 		return (notify_count >> 2);
3123 	return (2);
3124 }
3125 
3126 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3127 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3128 
3129 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3130 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3131 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3132 
3133 /* forward compatibility for cxgb */
3134 #define FIRST_QSET(ctx) 0
3135 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3136 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3137 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3138 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3139 
3140 /* XXX we should be setting this to something other than zero */
3141 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3142 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3143     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3144 
3145 static inline bool
3146 iflib_txd_db_check(iflib_txq_t txq, int ring)
3147 {
3148 	if_ctx_t ctx = txq->ift_ctx;
3149 	qidx_t dbval, max;
3150 
3151 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3152 
3153 	/* force || threshold exceeded || at the edge of the ring */
3154 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3155 
3156 		/*
3157 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3158 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3159 		 * producer index explicitly (INTC).
3160 		 */
3161 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3162 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3163 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3164 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3165 
3166 		/*
3167 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3168 		 */
3169 		txq->ift_db_pending = txq->ift_npending = 0;
3170 		return (true);
3171 	}
3172 	return (false);
3173 }
3174 
3175 #ifdef PKT_DEBUG
3176 static void
3177 print_pkt(if_pkt_info_t pi)
3178 {
3179 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3180 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3181 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3182 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3183 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3184 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3185 }
3186 #endif
3187 
3188 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3189 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3190 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3191 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3192 
3193 /**
3194  * Parses out ethernet header information in the given mbuf.
3195  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3196  *
3197  * This will account for the VLAN header if present.
3198  *
3199  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3200  * types of packets.
3201  */
3202 static int
3203 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3204 {
3205 	struct ether_vlan_header *eh;
3206 	struct mbuf *m;
3207 
3208 	m = *mp;
3209 	if (__predict_false(m->m_len < sizeof(*eh))) {
3210 		(*pullups)++;
3211 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3212 			return (ENOMEM);
3213 	}
3214 	eh = mtod(m, struct ether_vlan_header *);
3215 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3216 		pi->ipi_etype = ntohs(eh->evl_proto);
3217 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3218 	} else {
3219 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3220 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3221 	}
3222 	*mp = m;
3223 
3224 	return (0);
3225 }
3226 
3227 /**
3228  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3229  * Currently this information includes: IP ToS value, IP header version/presence
3230  *
3231  * This is missing some checks and doesn't edit the packet content as it goes,
3232  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3233  */
3234 static int
3235 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3236 {
3237 	struct mbuf *m;
3238 	int err;
3239 
3240 	*pullups = 0;
3241 	m = *mp;
3242 	if (!M_WRITABLE(m)) {
3243 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3244 			return (ENOMEM);
3245 		} else {
3246 			m_freem(*mp);
3247 			DBG_COUNTER_INC(tx_frees);
3248 			*mp = m;
3249 		}
3250 	}
3251 
3252 	/* Fills out pi->ipi_etype */
3253 	err = iflib_parse_ether_header(pi, mp, pullups);
3254 	if (err)
3255 		return (err);
3256 	m = *mp;
3257 
3258 	switch (pi->ipi_etype) {
3259 #ifdef INET
3260 	case ETHERTYPE_IP:
3261 	{
3262 		struct mbuf *n;
3263 		struct ip *ip = NULL;
3264 		int miniplen;
3265 
3266 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3267 		if (__predict_false(m->m_len < miniplen)) {
3268 			/*
3269 			 * Check for common case where the first mbuf only contains
3270 			 * the Ethernet header
3271 			 */
3272 			if (m->m_len == pi->ipi_ehdrlen) {
3273 				n = m->m_next;
3274 				MPASS(n);
3275 				/* If next mbuf contains at least the minimal IP header, then stop */
3276 				if (n->m_len >= sizeof(*ip)) {
3277 					ip = (struct ip *)n->m_data;
3278 				} else {
3279 					(*pullups)++;
3280 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3281 						return (ENOMEM);
3282 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3283 				}
3284 			} else {
3285 				(*pullups)++;
3286 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3287 					return (ENOMEM);
3288 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3289 			}
3290 		} else {
3291 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3292 		}
3293 
3294 		/* Have the IPv4 header w/ no options here */
3295 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3296 		pi->ipi_ipproto = ip->ip_p;
3297 		pi->ipi_ip_tos = ip->ip_tos;
3298 		pi->ipi_flags |= IPI_TX_IPV4;
3299 
3300 		break;
3301 	}
3302 #endif
3303 #ifdef INET6
3304 	case ETHERTYPE_IPV6:
3305 	{
3306 		struct ip6_hdr *ip6;
3307 
3308 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3309 			(*pullups)++;
3310 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3311 				return (ENOMEM);
3312 		}
3313 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3314 
3315 		/* Have the IPv6 fixed header here */
3316 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3317 		pi->ipi_ipproto = ip6->ip6_nxt;
3318 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3319 		pi->ipi_flags |= IPI_TX_IPV6;
3320 
3321 		break;
3322 	}
3323 #endif
3324 	default:
3325 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3326 		pi->ipi_ip_hlen = 0;
3327 		break;
3328 	}
3329 	*mp = m;
3330 
3331 	return (0);
3332 
3333 }
3334 
3335 static int
3336 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3337 {
3338 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3339 	struct mbuf *m;
3340 	int err;
3341 
3342 	m = *mp;
3343 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3344 	    M_WRITABLE(m) == 0) {
3345 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3346 			return (ENOMEM);
3347 		} else {
3348 			m_freem(*mp);
3349 			DBG_COUNTER_INC(tx_frees);
3350 			*mp = m;
3351 		}
3352 	}
3353 
3354 	/* Fills out pi->ipi_etype */
3355 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3356 	if (__predict_false(err))
3357 		return (err);
3358 	m = *mp;
3359 
3360 	switch (pi->ipi_etype) {
3361 #ifdef INET
3362 	case ETHERTYPE_IP:
3363 	{
3364 		struct mbuf *n;
3365 		struct ip *ip = NULL;
3366 		struct tcphdr *th = NULL;
3367 		int minthlen;
3368 
3369 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3370 		if (__predict_false(m->m_len < minthlen)) {
3371 			/*
3372 			 * if this code bloat is causing too much of a hit
3373 			 * move it to a separate function and mark it noinline
3374 			 */
3375 			if (m->m_len == pi->ipi_ehdrlen) {
3376 				n = m->m_next;
3377 				MPASS(n);
3378 				if (n->m_len >= sizeof(*ip))  {
3379 					ip = (struct ip *)n->m_data;
3380 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3381 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3382 				} else {
3383 					txq->ift_pullups++;
3384 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3385 						return (ENOMEM);
3386 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3387 				}
3388 			} else {
3389 				txq->ift_pullups++;
3390 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3391 					return (ENOMEM);
3392 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3393 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3394 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3395 			}
3396 		} else {
3397 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3398 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3399 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3400 		}
3401 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3402 		pi->ipi_ipproto = ip->ip_p;
3403 		pi->ipi_ip_tos = ip->ip_tos;
3404 		pi->ipi_flags |= IPI_TX_IPV4;
3405 
3406 		/* TCP checksum offload may require TCP header length */
3407 		if (IS_TX_OFFLOAD4(pi)) {
3408 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3409 				if (__predict_false(th == NULL)) {
3410 					txq->ift_pullups++;
3411 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3412 						return (ENOMEM);
3413 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3414 				}
3415 				pi->ipi_tcp_hflags = th->th_flags;
3416 				pi->ipi_tcp_hlen = th->th_off << 2;
3417 				pi->ipi_tcp_seq = th->th_seq;
3418 			}
3419 			if (IS_TSO4(pi)) {
3420 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3421 					return (ENXIO);
3422 				/*
3423 				 * TSO always requires hardware checksum offload.
3424 				 */
3425 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3426 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3427 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3428 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3429 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3430 					ip->ip_sum = 0;
3431 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3432 				}
3433 			}
3434 		}
3435 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3436                        ip->ip_sum = 0;
3437 
3438 		break;
3439 	}
3440 #endif
3441 #ifdef INET6
3442 	case ETHERTYPE_IPV6:
3443 	{
3444 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3445 		struct tcphdr *th;
3446 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3447 
3448 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3449 			txq->ift_pullups++;
3450 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3451 				return (ENOMEM);
3452 		}
3453 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3454 
3455 		/* XXX-BZ this will go badly in case of ext hdrs. */
3456 		pi->ipi_ipproto = ip6->ip6_nxt;
3457 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3458 		pi->ipi_flags |= IPI_TX_IPV6;
3459 
3460 		/* TCP checksum offload may require TCP header length */
3461 		if (IS_TX_OFFLOAD6(pi)) {
3462 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3463 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3464 					txq->ift_pullups++;
3465 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3466 						return (ENOMEM);
3467 				}
3468 				pi->ipi_tcp_hflags = th->th_flags;
3469 				pi->ipi_tcp_hlen = th->th_off << 2;
3470 				pi->ipi_tcp_seq = th->th_seq;
3471 			}
3472 			if (IS_TSO6(pi)) {
3473 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3474 					return (ENXIO);
3475 				/*
3476 				 * TSO always requires hardware checksum offload.
3477 				 */
3478 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3479 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3480 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3481 			}
3482 		}
3483 		break;
3484 	}
3485 #endif
3486 	default:
3487 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3488 		pi->ipi_ip_hlen = 0;
3489 		break;
3490 	}
3491 	*mp = m;
3492 
3493 	return (0);
3494 }
3495 
3496 /*
3497  * If dodgy hardware rejects the scatter gather chain we've handed it
3498  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3499  * m_defrag'd mbufs
3500  */
3501 static __noinline struct mbuf *
3502 iflib_remove_mbuf(iflib_txq_t txq)
3503 {
3504 	int ntxd, pidx;
3505 	struct mbuf *m, **ifsd_m;
3506 
3507 	ifsd_m = txq->ift_sds.ifsd_m;
3508 	ntxd = txq->ift_size;
3509 	pidx = txq->ift_pidx & (ntxd - 1);
3510 	ifsd_m = txq->ift_sds.ifsd_m;
3511 	m = ifsd_m[pidx];
3512 	ifsd_m[pidx] = NULL;
3513 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3514 	if (txq->ift_sds.ifsd_tso_map != NULL)
3515 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3516 		    txq->ift_sds.ifsd_tso_map[pidx]);
3517 #if MEMORY_LOGGING
3518 	txq->ift_dequeued++;
3519 #endif
3520 	return (m);
3521 }
3522 
3523 static inline caddr_t
3524 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3525 {
3526 	qidx_t size;
3527 	int ntxd;
3528 	caddr_t start, end, cur, next;
3529 
3530 	ntxd = txq->ift_size;
3531 	size = txq->ift_txd_size[qid];
3532 	start = txq->ift_ifdi[qid].idi_vaddr;
3533 
3534 	if (__predict_false(size == 0))
3535 		return (start);
3536 	cur = start + size*cidx;
3537 	end = start + size*ntxd;
3538 	next = CACHE_PTR_NEXT(cur);
3539 	return (next < end ? next : start);
3540 }
3541 
3542 /*
3543  * Pad an mbuf to ensure a minimum ethernet frame size.
3544  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3545  */
3546 static __noinline int
3547 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3548 {
3549 	/*
3550 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3551 	 * and ARP message is the smallest common payload I can think of
3552 	 */
3553 	static char pad[18];	/* just zeros */
3554 	int n;
3555 	struct mbuf *new_head;
3556 
3557 	if (!M_WRITABLE(*m_head)) {
3558 		new_head = m_dup(*m_head, M_NOWAIT);
3559 		if (new_head == NULL) {
3560 			m_freem(*m_head);
3561 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3562 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3563 			DBG_COUNTER_INC(tx_frees);
3564 			return ENOMEM;
3565 		}
3566 		m_freem(*m_head);
3567 		*m_head = new_head;
3568 	}
3569 
3570 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3571 	     n > 0; n -= sizeof(pad))
3572 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3573 			break;
3574 
3575 	if (n > 0) {
3576 		m_freem(*m_head);
3577 		device_printf(dev, "cannot pad short frame\n");
3578 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3579 		DBG_COUNTER_INC(tx_frees);
3580 		return (ENOBUFS);
3581 	}
3582 
3583 	return 0;
3584 }
3585 
3586 static int
3587 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3588 {
3589 	if_ctx_t		ctx;
3590 	if_shared_ctx_t		sctx;
3591 	if_softc_ctx_t		scctx;
3592 	bus_dma_tag_t		buf_tag;
3593 	bus_dma_segment_t	*segs;
3594 	struct mbuf		*m_head, **ifsd_m;
3595 	void			*next_txd;
3596 	bus_dmamap_t		map;
3597 	struct if_pkt_info	pi;
3598 	int remap = 0;
3599 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3600 
3601 	ctx = txq->ift_ctx;
3602 	sctx = ctx->ifc_sctx;
3603 	scctx = &ctx->ifc_softc_ctx;
3604 	segs = txq->ift_segs;
3605 	ntxd = txq->ift_size;
3606 	m_head = *m_headp;
3607 	map = NULL;
3608 
3609 	/*
3610 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3611 	 */
3612 	cidx = txq->ift_cidx;
3613 	pidx = txq->ift_pidx;
3614 	if (ctx->ifc_flags & IFC_PREFETCH) {
3615 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3616 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3617 			next_txd = calc_next_txd(txq, cidx, 0);
3618 			prefetch(next_txd);
3619 		}
3620 
3621 		/* prefetch the next cache line of mbuf pointers and flags */
3622 		prefetch(&txq->ift_sds.ifsd_m[next]);
3623 		prefetch(&txq->ift_sds.ifsd_map[next]);
3624 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3625 	}
3626 	map = txq->ift_sds.ifsd_map[pidx];
3627 	ifsd_m = txq->ift_sds.ifsd_m;
3628 
3629 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3630 		buf_tag = txq->ift_tso_buf_tag;
3631 		max_segs = scctx->isc_tx_tso_segments_max;
3632 		map = txq->ift_sds.ifsd_tso_map[pidx];
3633 		MPASS(buf_tag != NULL);
3634 		MPASS(max_segs > 0);
3635 	} else {
3636 		buf_tag = txq->ift_buf_tag;
3637 		max_segs = scctx->isc_tx_nsegments;
3638 		map = txq->ift_sds.ifsd_map[pidx];
3639 	}
3640 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3641 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3642 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3643 		if (err) {
3644 			DBG_COUNTER_INC(encap_txd_encap_fail);
3645 			return err;
3646 		}
3647 	}
3648 	m_head = *m_headp;
3649 
3650 	pkt_info_zero(&pi);
3651 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3652 	pi.ipi_pidx = pidx;
3653 	pi.ipi_qsidx = txq->ift_id;
3654 	pi.ipi_len = m_head->m_pkthdr.len;
3655 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3656 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3657 
3658 	/* deliberate bitwise OR to make one condition */
3659 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3660 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3661 			DBG_COUNTER_INC(encap_txd_encap_fail);
3662 			return (err);
3663 		}
3664 		m_head = *m_headp;
3665 	}
3666 
3667 retry:
3668 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3669 	    BUS_DMA_NOWAIT);
3670 defrag:
3671 	if (__predict_false(err)) {
3672 		switch (err) {
3673 		case EFBIG:
3674 			/* try collapse once and defrag once */
3675 			if (remap == 0) {
3676 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3677 				/* try defrag if collapsing fails */
3678 				if (m_head == NULL)
3679 					remap++;
3680 			}
3681 			if (remap == 1) {
3682 				txq->ift_mbuf_defrag++;
3683 				m_head = m_defrag(*m_headp, M_NOWAIT);
3684 			}
3685 			/*
3686 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3687 			 * failed to map an mbuf that was run through m_defrag
3688 			 */
3689 			MPASS(remap <= 1);
3690 			if (__predict_false(m_head == NULL || remap > 1))
3691 				goto defrag_failed;
3692 			remap++;
3693 			*m_headp = m_head;
3694 			goto retry;
3695 			break;
3696 		case ENOMEM:
3697 			txq->ift_no_tx_dma_setup++;
3698 			break;
3699 		default:
3700 			txq->ift_no_tx_dma_setup++;
3701 			m_freem(*m_headp);
3702 			DBG_COUNTER_INC(tx_frees);
3703 			*m_headp = NULL;
3704 			break;
3705 		}
3706 		txq->ift_map_failed++;
3707 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3708 		DBG_COUNTER_INC(encap_txd_encap_fail);
3709 		return (err);
3710 	}
3711 	ifsd_m[pidx] = m_head;
3712 	/*
3713 	 * XXX assumes a 1 to 1 relationship between segments and
3714 	 *        descriptors - this does not hold true on all drivers, e.g.
3715 	 *        cxgb
3716 	 */
3717 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3718 		txq->ift_no_desc_avail++;
3719 		bus_dmamap_unload(buf_tag, map);
3720 		DBG_COUNTER_INC(encap_txq_avail_fail);
3721 		DBG_COUNTER_INC(encap_txd_encap_fail);
3722 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3723 			GROUPTASK_ENQUEUE(&txq->ift_task);
3724 		return (ENOBUFS);
3725 	}
3726 	/*
3727 	 * On Intel cards we can greatly reduce the number of TX interrupts
3728 	 * we see by only setting report status on every Nth descriptor.
3729 	 * However, this also means that the driver will need to keep track
3730 	 * of the descriptors that RS was set on to check them for the DD bit.
3731 	 */
3732 	txq->ift_rs_pending += nsegs + 1;
3733 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3734 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3735 		pi.ipi_flags |= IPI_TX_INTR;
3736 		txq->ift_rs_pending = 0;
3737 	}
3738 
3739 	pi.ipi_segs = segs;
3740 	pi.ipi_nsegs = nsegs;
3741 
3742 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3743 #ifdef PKT_DEBUG
3744 	print_pkt(&pi);
3745 #endif
3746 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3747 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3748 		DBG_COUNTER_INC(tx_encap);
3749 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3750 
3751 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3752 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3753 			ndesc += txq->ift_size;
3754 			txq->ift_gen = 1;
3755 		}
3756 		/*
3757 		 * drivers can need as many as
3758 		 * two sentinels
3759 		 */
3760 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3761 		MPASS(pi.ipi_new_pidx != pidx);
3762 		MPASS(ndesc > 0);
3763 		txq->ift_in_use += ndesc;
3764 		txq->ift_db_pending += ndesc;
3765 
3766 		/*
3767 		 * We update the last software descriptor again here because there may
3768 		 * be a sentinel and/or there may be more mbufs than segments
3769 		 */
3770 		txq->ift_pidx = pi.ipi_new_pidx;
3771 		txq->ift_npending += pi.ipi_ndescs;
3772 	} else {
3773 		*m_headp = m_head = iflib_remove_mbuf(txq);
3774 		if (err == EFBIG) {
3775 			txq->ift_txd_encap_efbig++;
3776 			if (remap < 2) {
3777 				remap = 1;
3778 				goto defrag;
3779 			}
3780 		}
3781 		goto defrag_failed;
3782 	}
3783 	/*
3784 	 * err can't possibly be non-zero here, so we don't neet to test it
3785 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3786 	 */
3787 	return (err);
3788 
3789 defrag_failed:
3790 	txq->ift_mbuf_defrag_failed++;
3791 	txq->ift_map_failed++;
3792 	m_freem(*m_headp);
3793 	DBG_COUNTER_INC(tx_frees);
3794 	*m_headp = NULL;
3795 	DBG_COUNTER_INC(encap_txd_encap_fail);
3796 	return (ENOMEM);
3797 }
3798 
3799 static void
3800 iflib_tx_desc_free(iflib_txq_t txq, int n)
3801 {
3802 	uint32_t qsize, cidx, mask, gen;
3803 	struct mbuf *m, **ifsd_m;
3804 	bool do_prefetch;
3805 
3806 	cidx = txq->ift_cidx;
3807 	gen = txq->ift_gen;
3808 	qsize = txq->ift_size;
3809 	mask = qsize-1;
3810 	ifsd_m = txq->ift_sds.ifsd_m;
3811 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3812 
3813 	while (n-- > 0) {
3814 		if (do_prefetch) {
3815 			prefetch(ifsd_m[(cidx + 3) & mask]);
3816 			prefetch(ifsd_m[(cidx + 4) & mask]);
3817 		}
3818 		if ((m = ifsd_m[cidx]) != NULL) {
3819 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3820 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3821 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3822 				    txq->ift_sds.ifsd_tso_map[cidx],
3823 				    BUS_DMASYNC_POSTWRITE);
3824 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3825 				    txq->ift_sds.ifsd_tso_map[cidx]);
3826 			} else {
3827 				bus_dmamap_sync(txq->ift_buf_tag,
3828 				    txq->ift_sds.ifsd_map[cidx],
3829 				    BUS_DMASYNC_POSTWRITE);
3830 				bus_dmamap_unload(txq->ift_buf_tag,
3831 				    txq->ift_sds.ifsd_map[cidx]);
3832 			}
3833 			/* XXX we don't support any drivers that batch packets yet */
3834 			MPASS(m->m_nextpkt == NULL);
3835 			m_freem(m);
3836 			ifsd_m[cidx] = NULL;
3837 #if MEMORY_LOGGING
3838 			txq->ift_dequeued++;
3839 #endif
3840 			DBG_COUNTER_INC(tx_frees);
3841 		}
3842 		if (__predict_false(++cidx == qsize)) {
3843 			cidx = 0;
3844 			gen = 0;
3845 		}
3846 	}
3847 	txq->ift_cidx = cidx;
3848 	txq->ift_gen = gen;
3849 }
3850 
3851 static __inline int
3852 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3853 {
3854 	int reclaim;
3855 	if_ctx_t ctx = txq->ift_ctx;
3856 
3857 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3858 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3859 
3860 	/*
3861 	 * Need a rate-limiting check so that this isn't called every time
3862 	 */
3863 	iflib_tx_credits_update(ctx, txq);
3864 	reclaim = DESC_RECLAIMABLE(txq);
3865 
3866 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3867 #ifdef INVARIANTS
3868 		if (iflib_verbose_debug) {
3869 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3870 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3871 			       reclaim, thresh);
3872 		}
3873 #endif
3874 		return (0);
3875 	}
3876 	iflib_tx_desc_free(txq, reclaim);
3877 	txq->ift_cleaned += reclaim;
3878 	txq->ift_in_use -= reclaim;
3879 
3880 	return (reclaim);
3881 }
3882 
3883 static struct mbuf **
3884 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3885 {
3886 	int next, size;
3887 	struct mbuf **items;
3888 
3889 	size = r->size;
3890 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3891 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3892 
3893 	prefetch(items[(cidx + offset) & (size-1)]);
3894 	if (remaining > 1) {
3895 		prefetch2cachelines(&items[next]);
3896 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3897 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3898 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3899 	}
3900 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3901 }
3902 
3903 static void
3904 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3905 {
3906 
3907 	ifmp_ring_check_drainage(txq->ift_br, budget);
3908 }
3909 
3910 static uint32_t
3911 iflib_txq_can_drain(struct ifmp_ring *r)
3912 {
3913 	iflib_txq_t txq = r->cookie;
3914 	if_ctx_t ctx = txq->ift_ctx;
3915 
3916 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3917 		return (1);
3918 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3919 	    BUS_DMASYNC_POSTREAD);
3920 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3921 	    false));
3922 }
3923 
3924 static uint32_t
3925 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3926 {
3927 	iflib_txq_t txq = r->cookie;
3928 	if_ctx_t ctx = txq->ift_ctx;
3929 	if_t ifp = ctx->ifc_ifp;
3930 	struct mbuf *m, **mp;
3931 	int avail, bytes_sent, skipped, count, err, i;
3932 	int mcast_sent, pkt_sent, reclaimed;
3933 	bool do_prefetch, rang, ring;
3934 
3935 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3936 			    !LINK_ACTIVE(ctx))) {
3937 		DBG_COUNTER_INC(txq_drain_notready);
3938 		return (0);
3939 	}
3940 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3941 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3942 	avail = IDXDIFF(pidx, cidx, r->size);
3943 
3944 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3945 		/*
3946 		 * The driver is unloading so we need to free all pending packets.
3947 		 */
3948 		DBG_COUNTER_INC(txq_drain_flushing);
3949 		for (i = 0; i < avail; i++) {
3950 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3951 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3952 			r->items[(cidx + i) & (r->size-1)] = NULL;
3953 		}
3954 		return (avail);
3955 	}
3956 
3957 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3958 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3959 		CALLOUT_LOCK(txq);
3960 		callout_stop(&txq->ift_timer);
3961 		CALLOUT_UNLOCK(txq);
3962 		DBG_COUNTER_INC(txq_drain_oactive);
3963 		return (0);
3964 	}
3965 
3966 	/*
3967 	 * If we've reclaimed any packets this queue cannot be hung.
3968 	 */
3969 	if (reclaimed)
3970 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3971 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3972 	count = MIN(avail, TX_BATCH_SIZE);
3973 #ifdef INVARIANTS
3974 	if (iflib_verbose_debug)
3975 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3976 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3977 #endif
3978 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3979 	err = 0;
3980 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3981 		int rem = do_prefetch ? count - i : 0;
3982 
3983 		mp = _ring_peek_one(r, cidx, i, rem);
3984 		MPASS(mp != NULL && *mp != NULL);
3985 
3986 		/*
3987 		 * Completion interrupts will use the address of the txq
3988 		 * as a sentinel to enqueue _something_ in order to acquire
3989 		 * the lock on the mp_ring (there's no direct lock call).
3990 		 * We obviously whave to check for these sentinel cases
3991 		 * and skip them.
3992 		 */
3993 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3994 			skipped++;
3995 			continue;
3996 		}
3997 		err = iflib_encap(txq, mp);
3998 		if (__predict_false(err)) {
3999 			/* no room - bail out */
4000 			if (err == ENOBUFS)
4001 				break;
4002 			skipped++;
4003 			/* we can't send this packet - skip it */
4004 			continue;
4005 		}
4006 		pkt_sent++;
4007 		m = *mp;
4008 		DBG_COUNTER_INC(tx_sent);
4009 		bytes_sent += m->m_pkthdr.len;
4010 		mcast_sent += !!(m->m_flags & M_MCAST);
4011 
4012 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4013 			break;
4014 		ETHER_BPF_MTAP(ifp, m);
4015 		rang = iflib_txd_db_check(txq, false);
4016 	}
4017 
4018 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4019 	ring = rang ? false  : (iflib_min_tx_latency | err);
4020 	iflib_txd_db_check(txq, ring);
4021 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4022 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4023 	if (mcast_sent)
4024 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4025 #ifdef INVARIANTS
4026 	if (iflib_verbose_debug)
4027 		printf("consumed=%d\n", skipped + pkt_sent);
4028 #endif
4029 	return (skipped + pkt_sent);
4030 }
4031 
4032 static uint32_t
4033 iflib_txq_drain_always(struct ifmp_ring *r)
4034 {
4035 	return (1);
4036 }
4037 
4038 static uint32_t
4039 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4040 {
4041 	int i, avail;
4042 	struct mbuf **mp;
4043 	iflib_txq_t txq;
4044 
4045 	txq = r->cookie;
4046 
4047 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4048 	CALLOUT_LOCK(txq);
4049 	callout_stop(&txq->ift_timer);
4050 	CALLOUT_UNLOCK(txq);
4051 
4052 	avail = IDXDIFF(pidx, cidx, r->size);
4053 	for (i = 0; i < avail; i++) {
4054 		mp = _ring_peek_one(r, cidx, i, avail - i);
4055 		if (__predict_false(*mp == (struct mbuf *)txq))
4056 			continue;
4057 		m_freem(*mp);
4058 		DBG_COUNTER_INC(tx_frees);
4059 	}
4060 	MPASS(ifmp_ring_is_stalled(r) == 0);
4061 	return (avail);
4062 }
4063 
4064 static void
4065 iflib_ifmp_purge(iflib_txq_t txq)
4066 {
4067 	struct ifmp_ring *r;
4068 
4069 	r = txq->ift_br;
4070 	r->drain = iflib_txq_drain_free;
4071 	r->can_drain = iflib_txq_drain_always;
4072 
4073 	ifmp_ring_check_drainage(r, r->size);
4074 
4075 	r->drain = iflib_txq_drain;
4076 	r->can_drain = iflib_txq_can_drain;
4077 }
4078 
4079 static void
4080 _task_fn_tx(void *context)
4081 {
4082 	iflib_txq_t txq = context;
4083 	if_ctx_t ctx = txq->ift_ctx;
4084 	if_t ifp = ctx->ifc_ifp;
4085 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4086 
4087 #ifdef IFLIB_DIAGNOSTICS
4088 	txq->ift_cpu_exec_count[curcpu]++;
4089 #endif
4090 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4091 		return;
4092 #ifdef DEV_NETMAP
4093 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4094 	    netmap_tx_irq(ifp, txq->ift_id))
4095 		goto skip_ifmp;
4096 #endif
4097 #ifdef ALTQ
4098 	if (if_altq_is_enabled(ifp))
4099 		iflib_altq_if_start(ifp);
4100 #endif
4101 	if (txq->ift_db_pending)
4102 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4103 	else if (!abdicate)
4104 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4105 	/*
4106 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4107 	 */
4108 	if (abdicate)
4109 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4110 #ifdef DEV_NETMAP
4111 skip_ifmp:
4112 #endif
4113 	if (ctx->ifc_flags & IFC_LEGACY)
4114 		IFDI_INTR_ENABLE(ctx);
4115 	else
4116 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4117 }
4118 
4119 static void
4120 _task_fn_rx(void *context)
4121 {
4122 	iflib_rxq_t rxq = context;
4123 	if_ctx_t ctx = rxq->ifr_ctx;
4124 	uint8_t more;
4125 	uint16_t budget;
4126 #ifdef DEV_NETMAP
4127 	u_int work = 0;
4128 	int nmirq;
4129 #endif
4130 
4131 #ifdef IFLIB_DIAGNOSTICS
4132 	rxq->ifr_cpu_exec_count[curcpu]++;
4133 #endif
4134 	DBG_COUNTER_INC(task_fn_rxs);
4135 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4136 		return;
4137 #ifdef DEV_NETMAP
4138 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4139 	if (nmirq != NM_IRQ_PASS) {
4140 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4141 		goto skip_rxeof;
4142 	}
4143 #endif
4144 	budget = ctx->ifc_sysctl_rx_budget;
4145 	if (budget == 0)
4146 		budget = 16;	/* XXX */
4147 	more = iflib_rxeof(rxq, budget);
4148 #ifdef DEV_NETMAP
4149 skip_rxeof:
4150 #endif
4151 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4152 		if (ctx->ifc_flags & IFC_LEGACY)
4153 			IFDI_INTR_ENABLE(ctx);
4154 		else
4155 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4156 		DBG_COUNTER_INC(rx_intr_enables);
4157 	}
4158 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4159 		return;
4160 
4161 	if (more & IFLIB_RXEOF_MORE)
4162 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4163 	else if (more & IFLIB_RXEOF_EMPTY)
4164 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4165 }
4166 
4167 static void
4168 _task_fn_admin(void *context)
4169 {
4170 	if_ctx_t ctx = context;
4171 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4172 	iflib_txq_t txq;
4173 	int i;
4174 	bool oactive, running, do_reset, do_watchdog, in_detach;
4175 
4176 	STATE_LOCK(ctx);
4177 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4178 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4179 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4180 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4181 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4182 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4183 	STATE_UNLOCK(ctx);
4184 
4185 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4186 		return;
4187 	if (in_detach)
4188 		return;
4189 
4190 	CTX_LOCK(ctx);
4191 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4192 		CALLOUT_LOCK(txq);
4193 		callout_stop(&txq->ift_timer);
4194 		CALLOUT_UNLOCK(txq);
4195 	}
4196 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4197 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4198 	if (do_watchdog) {
4199 		ctx->ifc_watchdog_events++;
4200 		IFDI_WATCHDOG_RESET(ctx);
4201 	}
4202 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4203 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4204 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4205 		    txq->ift_timer.c_cpu);
4206 	}
4207 	IFDI_LINK_INTR_ENABLE(ctx);
4208 	if (do_reset)
4209 		iflib_if_init_locked(ctx);
4210 	CTX_UNLOCK(ctx);
4211 
4212 	if (LINK_ACTIVE(ctx) == 0)
4213 		return;
4214 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4215 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4216 }
4217 
4218 static void
4219 _task_fn_iov(void *context)
4220 {
4221 	if_ctx_t ctx = context;
4222 
4223 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4224 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4225 		return;
4226 
4227 	CTX_LOCK(ctx);
4228 	IFDI_VFLR_HANDLE(ctx);
4229 	CTX_UNLOCK(ctx);
4230 }
4231 
4232 static int
4233 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4234 {
4235 	int err;
4236 	if_int_delay_info_t info;
4237 	if_ctx_t ctx;
4238 
4239 	info = (if_int_delay_info_t)arg1;
4240 	ctx = info->iidi_ctx;
4241 	info->iidi_req = req;
4242 	info->iidi_oidp = oidp;
4243 	CTX_LOCK(ctx);
4244 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4245 	CTX_UNLOCK(ctx);
4246 	return (err);
4247 }
4248 
4249 /*********************************************************************
4250  *
4251  *  IFNET FUNCTIONS
4252  *
4253  **********************************************************************/
4254 
4255 static void
4256 iflib_if_init_locked(if_ctx_t ctx)
4257 {
4258 	iflib_stop(ctx);
4259 	iflib_init_locked(ctx);
4260 }
4261 
4262 static void
4263 iflib_if_init(void *arg)
4264 {
4265 	if_ctx_t ctx = arg;
4266 
4267 	CTX_LOCK(ctx);
4268 	iflib_if_init_locked(ctx);
4269 	CTX_UNLOCK(ctx);
4270 }
4271 
4272 static int
4273 iflib_if_transmit(if_t ifp, struct mbuf *m)
4274 {
4275 	if_ctx_t ctx = if_getsoftc(ifp);
4276 	iflib_txq_t txq;
4277 	int err, qidx;
4278 	int abdicate;
4279 
4280 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4281 		DBG_COUNTER_INC(tx_frees);
4282 		m_freem(m);
4283 		return (ENETDOWN);
4284 	}
4285 
4286 	MPASS(m->m_nextpkt == NULL);
4287 	/* ALTQ-enabled interfaces always use queue 0. */
4288 	qidx = 0;
4289 	/* Use driver-supplied queue selection method if it exists */
4290 	if (ctx->isc_txq_select_v2) {
4291 		struct if_pkt_info pi;
4292 		uint64_t early_pullups = 0;
4293 		pkt_info_zero(&pi);
4294 
4295 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4296 		if (__predict_false(err != 0)) {
4297 			/* Assign pullups for bad pkts to default queue */
4298 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4299 			DBG_COUNTER_INC(encap_txd_encap_fail);
4300 			return (err);
4301 		}
4302 		/* Let driver make queueing decision */
4303 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4304 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4305 	}
4306 	/* Backwards compatibility w/ simpler queue select */
4307 	else if (ctx->isc_txq_select)
4308 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4309 	/* If not, use iflib's standard method */
4310 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4311 		qidx = QIDX(ctx, m);
4312 
4313 	/* Set TX queue */
4314 	txq = &ctx->ifc_txqs[qidx];
4315 
4316 #ifdef DRIVER_BACKPRESSURE
4317 	if (txq->ift_closed) {
4318 		while (m != NULL) {
4319 			next = m->m_nextpkt;
4320 			m->m_nextpkt = NULL;
4321 			m_freem(m);
4322 			DBG_COUNTER_INC(tx_frees);
4323 			m = next;
4324 		}
4325 		return (ENOBUFS);
4326 	}
4327 #endif
4328 #ifdef notyet
4329 	qidx = count = 0;
4330 	mp = marr;
4331 	next = m;
4332 	do {
4333 		count++;
4334 		next = next->m_nextpkt;
4335 	} while (next != NULL);
4336 
4337 	if (count > nitems(marr))
4338 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4339 			/* XXX check nextpkt */
4340 			m_freem(m);
4341 			/* XXX simplify for now */
4342 			DBG_COUNTER_INC(tx_frees);
4343 			return (ENOBUFS);
4344 		}
4345 	for (next = m, i = 0; next != NULL; i++) {
4346 		mp[i] = next;
4347 		next = next->m_nextpkt;
4348 		mp[i]->m_nextpkt = NULL;
4349 	}
4350 #endif
4351 	DBG_COUNTER_INC(tx_seen);
4352 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4353 
4354 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4355 
4356 	if (abdicate)
4357 		GROUPTASK_ENQUEUE(&txq->ift_task);
4358  	if (err) {
4359 		if (!abdicate)
4360 			GROUPTASK_ENQUEUE(&txq->ift_task);
4361 		/* support forthcoming later */
4362 #ifdef DRIVER_BACKPRESSURE
4363 		txq->ift_closed = TRUE;
4364 #endif
4365 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4366 		m_freem(m);
4367 		DBG_COUNTER_INC(tx_frees);
4368 	}
4369 
4370 	return (err);
4371 }
4372 
4373 #ifdef ALTQ
4374 /*
4375  * The overall approach to integrating iflib with ALTQ is to continue to use
4376  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4377  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4378  * is redundant/unnecessary, but doing so minimizes the amount of
4379  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4380  * redundantly queueing to an intermediate mp_ring is swamped by the
4381  * performance limitations inherent in using ALTQ.
4382  *
4383  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4384  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4385  * given interface.  If ALTQ is enabled for an interface, then all
4386  * transmitted packets for that interface will be submitted to the ALTQ
4387  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4388  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4389  * update stats that the iflib machinery handles, and which is sensitve to
4390  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4391  * will be installed as the start routine for use by ALTQ facilities that
4392  * need to trigger queue drains on a scheduled basis.
4393  *
4394  */
4395 static void
4396 iflib_altq_if_start(if_t ifp)
4397 {
4398 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4399 	struct mbuf *m;
4400 
4401 	IFQ_LOCK(ifq);
4402 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4403 	while (m != NULL) {
4404 		iflib_if_transmit(ifp, m);
4405 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4406 	}
4407 	IFQ_UNLOCK(ifq);
4408 }
4409 
4410 static int
4411 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4412 {
4413 	int err;
4414 
4415 	if (if_altq_is_enabled(ifp)) {
4416 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4417 		if (err == 0)
4418 			iflib_altq_if_start(ifp);
4419 	} else
4420 		err = iflib_if_transmit(ifp, m);
4421 
4422 	return (err);
4423 }
4424 #endif /* ALTQ */
4425 
4426 static void
4427 iflib_if_qflush(if_t ifp)
4428 {
4429 	if_ctx_t ctx = if_getsoftc(ifp);
4430 	iflib_txq_t txq = ctx->ifc_txqs;
4431 	int i;
4432 
4433 	STATE_LOCK(ctx);
4434 	ctx->ifc_flags |= IFC_QFLUSH;
4435 	STATE_UNLOCK(ctx);
4436 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4437 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4438 			iflib_txq_check_drain(txq, 0);
4439 	STATE_LOCK(ctx);
4440 	ctx->ifc_flags &= ~IFC_QFLUSH;
4441 	STATE_UNLOCK(ctx);
4442 
4443 	/*
4444 	 * When ALTQ is enabled, this will also take care of purging the
4445 	 * ALTQ queue(s).
4446 	 */
4447 	if_qflush(ifp);
4448 }
4449 
4450 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4451 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4452 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4453 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4454 
4455 static int
4456 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4457 {
4458 	if_ctx_t ctx = if_getsoftc(ifp);
4459 	struct ifreq	*ifr = (struct ifreq *)data;
4460 #if defined(INET) || defined(INET6)
4461 	struct ifaddr	*ifa = (struct ifaddr *)data;
4462 #endif
4463 	bool		avoid_reset = false;
4464 	int		err = 0, reinit = 0, bits;
4465 
4466 	switch (command) {
4467 	case SIOCSIFADDR:
4468 #ifdef INET
4469 		if (ifa->ifa_addr->sa_family == AF_INET)
4470 			avoid_reset = true;
4471 #endif
4472 #ifdef INET6
4473 		if (ifa->ifa_addr->sa_family == AF_INET6)
4474 			avoid_reset = true;
4475 #endif
4476 		/*
4477 		** Calling init results in link renegotiation,
4478 		** so we avoid doing it when possible.
4479 		*/
4480 		if (avoid_reset) {
4481 			if_setflagbits(ifp, IFF_UP,0);
4482 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4483 				reinit = 1;
4484 #ifdef INET
4485 			if (!(if_getflags(ifp) & IFF_NOARP))
4486 				arp_ifinit(ifp, ifa);
4487 #endif
4488 		} else
4489 			err = ether_ioctl(ifp, command, data);
4490 		break;
4491 	case SIOCSIFMTU:
4492 		CTX_LOCK(ctx);
4493 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4494 			CTX_UNLOCK(ctx);
4495 			break;
4496 		}
4497 		bits = if_getdrvflags(ifp);
4498 		/* stop the driver and free any clusters before proceeding */
4499 		iflib_stop(ctx);
4500 
4501 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4502 			STATE_LOCK(ctx);
4503 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4504 				ctx->ifc_flags |= IFC_MULTISEG;
4505 			else
4506 				ctx->ifc_flags &= ~IFC_MULTISEG;
4507 			STATE_UNLOCK(ctx);
4508 			err = if_setmtu(ifp, ifr->ifr_mtu);
4509 		}
4510 		iflib_init_locked(ctx);
4511 		STATE_LOCK(ctx);
4512 		if_setdrvflags(ifp, bits);
4513 		STATE_UNLOCK(ctx);
4514 		CTX_UNLOCK(ctx);
4515 		break;
4516 	case SIOCSIFFLAGS:
4517 		CTX_LOCK(ctx);
4518 		if (if_getflags(ifp) & IFF_UP) {
4519 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4520 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4521 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4522 					CTX_UNLOCK(ctx);
4523 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4524 					CTX_LOCK(ctx);
4525 				}
4526 			} else
4527 				reinit = 1;
4528 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4529 			iflib_stop(ctx);
4530 		}
4531 		ctx->ifc_if_flags = if_getflags(ifp);
4532 		CTX_UNLOCK(ctx);
4533 		break;
4534 	case SIOCADDMULTI:
4535 	case SIOCDELMULTI:
4536 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4537 			CTX_LOCK(ctx);
4538 			IFDI_INTR_DISABLE(ctx);
4539 			IFDI_MULTI_SET(ctx);
4540 			IFDI_INTR_ENABLE(ctx);
4541 			CTX_UNLOCK(ctx);
4542 		}
4543 		break;
4544 	case SIOCSIFMEDIA:
4545 		CTX_LOCK(ctx);
4546 		IFDI_MEDIA_SET(ctx);
4547 		CTX_UNLOCK(ctx);
4548 		/* FALLTHROUGH */
4549 	case SIOCGIFMEDIA:
4550 	case SIOCGIFXMEDIA:
4551 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4552 		break;
4553 	case SIOCGI2C:
4554 	{
4555 		struct ifi2creq i2c;
4556 
4557 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4558 		if (err != 0)
4559 			break;
4560 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4561 			err = EINVAL;
4562 			break;
4563 		}
4564 		if (i2c.len > sizeof(i2c.data)) {
4565 			err = EINVAL;
4566 			break;
4567 		}
4568 
4569 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4570 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4571 			    sizeof(i2c));
4572 		break;
4573 	}
4574 	case SIOCSIFCAP:
4575 	{
4576 		int mask, setmask, oldmask;
4577 
4578 		oldmask = if_getcapenable(ifp);
4579 		mask = ifr->ifr_reqcap ^ oldmask;
4580 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4581 		setmask = 0;
4582 #ifdef TCP_OFFLOAD
4583 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4584 #endif
4585 		setmask |= (mask & IFCAP_FLAGS);
4586 		setmask |= (mask & IFCAP_WOL);
4587 
4588 		/*
4589 		 * If any RX csum has changed, change all the ones that
4590 		 * are supported by the driver.
4591 		 */
4592 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4593 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4594 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4595 		}
4596 
4597 		/*
4598 		 * want to ensure that traffic has stopped before we change any of the flags
4599 		 */
4600 		if (setmask) {
4601 			CTX_LOCK(ctx);
4602 			bits = if_getdrvflags(ifp);
4603 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4604 				iflib_stop(ctx);
4605 			STATE_LOCK(ctx);
4606 			if_togglecapenable(ifp, setmask);
4607 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4608 			STATE_UNLOCK(ctx);
4609 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4610 				iflib_init_locked(ctx);
4611 			STATE_LOCK(ctx);
4612 			if_setdrvflags(ifp, bits);
4613 			STATE_UNLOCK(ctx);
4614 			CTX_UNLOCK(ctx);
4615 		}
4616 		if_vlancap(ifp);
4617 		break;
4618 	}
4619 	case SIOCGPRIVATE_0:
4620 	case SIOCSDRVSPEC:
4621 	case SIOCGDRVSPEC:
4622 		CTX_LOCK(ctx);
4623 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4624 		CTX_UNLOCK(ctx);
4625 		break;
4626 	default:
4627 		err = ether_ioctl(ifp, command, data);
4628 		break;
4629 	}
4630 	if (reinit)
4631 		iflib_if_init(ctx);
4632 	return (err);
4633 }
4634 
4635 static uint64_t
4636 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4637 {
4638 	if_ctx_t ctx = if_getsoftc(ifp);
4639 
4640 	return (IFDI_GET_COUNTER(ctx, cnt));
4641 }
4642 
4643 /*********************************************************************
4644  *
4645  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4646  *
4647  **********************************************************************/
4648 
4649 static void
4650 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4651 {
4652 	if_ctx_t ctx = if_getsoftc(ifp);
4653 
4654 	if ((void *)ctx != arg)
4655 		return;
4656 
4657 	if ((vtag == 0) || (vtag > 4095))
4658 		return;
4659 
4660 	if (iflib_in_detach(ctx))
4661 		return;
4662 
4663 	CTX_LOCK(ctx);
4664 	/* Driver may need all untagged packets to be flushed */
4665 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4666 		iflib_stop(ctx);
4667 	IFDI_VLAN_REGISTER(ctx, vtag);
4668 	/* Re-init to load the changes, if required */
4669 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4670 		iflib_init_locked(ctx);
4671 	CTX_UNLOCK(ctx);
4672 }
4673 
4674 static void
4675 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4676 {
4677 	if_ctx_t ctx = if_getsoftc(ifp);
4678 
4679 	if ((void *)ctx != arg)
4680 		return;
4681 
4682 	if ((vtag == 0) || (vtag > 4095))
4683 		return;
4684 
4685 	CTX_LOCK(ctx);
4686 	/* Driver may need all tagged packets to be flushed */
4687 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4688 		iflib_stop(ctx);
4689 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4690 	/* Re-init to load the changes, if required */
4691 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4692 		iflib_init_locked(ctx);
4693 	CTX_UNLOCK(ctx);
4694 }
4695 
4696 static void
4697 iflib_led_func(void *arg, int onoff)
4698 {
4699 	if_ctx_t ctx = arg;
4700 
4701 	CTX_LOCK(ctx);
4702 	IFDI_LED_FUNC(ctx, onoff);
4703 	CTX_UNLOCK(ctx);
4704 }
4705 
4706 /*********************************************************************
4707  *
4708  *  BUS FUNCTION DEFINITIONS
4709  *
4710  **********************************************************************/
4711 
4712 int
4713 iflib_device_probe(device_t dev)
4714 {
4715 	const pci_vendor_info_t *ent;
4716 	if_shared_ctx_t sctx;
4717 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4718 	uint16_t pci_vendor_id;
4719 
4720 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4721 		return (ENOTSUP);
4722 
4723 	pci_vendor_id = pci_get_vendor(dev);
4724 	pci_device_id = pci_get_device(dev);
4725 	pci_subvendor_id = pci_get_subvendor(dev);
4726 	pci_subdevice_id = pci_get_subdevice(dev);
4727 	pci_rev_id = pci_get_revid(dev);
4728 	if (sctx->isc_parse_devinfo != NULL)
4729 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4730 
4731 	ent = sctx->isc_vendor_info;
4732 	while (ent->pvi_vendor_id != 0) {
4733 		if (pci_vendor_id != ent->pvi_vendor_id) {
4734 			ent++;
4735 			continue;
4736 		}
4737 		if ((pci_device_id == ent->pvi_device_id) &&
4738 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4739 		     (ent->pvi_subvendor_id == 0)) &&
4740 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4741 		     (ent->pvi_subdevice_id == 0)) &&
4742 		    ((pci_rev_id == ent->pvi_rev_id) ||
4743 		     (ent->pvi_rev_id == 0))) {
4744 			device_set_desc_copy(dev, ent->pvi_name);
4745 			/* this needs to be changed to zero if the bus probing code
4746 			 * ever stops re-probing on best match because the sctx
4747 			 * may have its values over written by register calls
4748 			 * in subsequent probes
4749 			 */
4750 			return (BUS_PROBE_DEFAULT);
4751 		}
4752 		ent++;
4753 	}
4754 	return (ENXIO);
4755 }
4756 
4757 int
4758 iflib_device_probe_vendor(device_t dev)
4759 {
4760 	int probe;
4761 
4762 	probe = iflib_device_probe(dev);
4763 	if (probe == BUS_PROBE_DEFAULT)
4764 		return (BUS_PROBE_VENDOR);
4765 	else
4766 		return (probe);
4767 }
4768 
4769 static void
4770 iflib_reset_qvalues(if_ctx_t ctx)
4771 {
4772 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4773 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4774 	device_t dev = ctx->ifc_dev;
4775 	int i;
4776 
4777 	if (ctx->ifc_sysctl_ntxqs != 0)
4778 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4779 	if (ctx->ifc_sysctl_nrxqs != 0)
4780 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4781 
4782 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4783 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4784 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4785 		else
4786 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4787 	}
4788 
4789 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4790 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4791 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4792 		else
4793 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4794 	}
4795 
4796 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4797 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4798 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4799 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4800 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4801 		}
4802 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4803 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4804 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4805 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4806 		}
4807 		if (!powerof2(scctx->isc_nrxd[i])) {
4808 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4809 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4810 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4811 		}
4812 	}
4813 
4814 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4815 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4816 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4817 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4818 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4819 		}
4820 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4821 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4822 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4823 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4824 		}
4825 		if (!powerof2(scctx->isc_ntxd[i])) {
4826 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4827 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4828 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4829 		}
4830 	}
4831 }
4832 
4833 static void
4834 iflib_add_pfil(if_ctx_t ctx)
4835 {
4836 	struct pfil_head *pfil;
4837 	struct pfil_head_args pa;
4838 	iflib_rxq_t rxq;
4839 	int i;
4840 
4841 	pa.pa_version = PFIL_VERSION;
4842 	pa.pa_flags = PFIL_IN;
4843 	pa.pa_type = PFIL_TYPE_ETHERNET;
4844 	pa.pa_headname = if_name(ctx->ifc_ifp);
4845 	pfil = pfil_head_register(&pa);
4846 
4847 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4848 		rxq->pfil = pfil;
4849 	}
4850 }
4851 
4852 static void
4853 iflib_rem_pfil(if_ctx_t ctx)
4854 {
4855 	struct pfil_head *pfil;
4856 	iflib_rxq_t rxq;
4857 	int i;
4858 
4859 	rxq = ctx->ifc_rxqs;
4860 	pfil = rxq->pfil;
4861 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4862 		rxq->pfil = NULL;
4863 	}
4864 	pfil_head_unregister(pfil);
4865 }
4866 
4867 
4868 /*
4869  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4870  * cpuid and wrapping as necessary.
4871  */
4872 static unsigned int
4873 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4874 {
4875 	unsigned int first_valid;
4876 	unsigned int last_valid;
4877 
4878 	/* cpuid should always be in the valid set */
4879 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4880 
4881 	/* valid set should never be empty */
4882 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4883 
4884 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4885 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4886 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4887 	while (n > 0) {
4888 		do {
4889 			cpuid++;
4890 			if (cpuid > last_valid)
4891 				cpuid = first_valid;
4892 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4893 		n--;
4894 	}
4895 
4896 	return (cpuid);
4897 }
4898 
4899 #if defined(SMP) && defined(SCHED_ULE)
4900 extern struct cpu_group *cpu_top;              /* CPU topology */
4901 
4902 static int
4903 find_child_with_core(int cpu, struct cpu_group *grp)
4904 {
4905 	int i;
4906 
4907 	if (grp->cg_children == 0)
4908 		return -1;
4909 
4910 	MPASS(grp->cg_child);
4911 	for (i = 0; i < grp->cg_children; i++) {
4912 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4913 			return i;
4914 	}
4915 
4916 	return -1;
4917 }
4918 
4919 
4920 /*
4921  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4922  * does not distinguish among multiple L2 neighbors if the given CPU has
4923  * more than one (it will always return the same result in that case).
4924  */
4925 static int
4926 find_l2_neighbor(int cpu)
4927 {
4928 	struct cpu_group *grp;
4929 	int i;
4930 
4931 	grp = cpu_top;
4932 	if (grp == NULL)
4933 		return -1;
4934 
4935 	/*
4936 	 * Find the smallest CPU group that contains the given core.
4937 	 */
4938 	i = 0;
4939 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4940 		/*
4941 		 * If the smallest group containing the given CPU has less
4942 		 * than two members, we conclude the given CPU has no
4943 		 * L2 neighbor.
4944 		 */
4945 		if (grp->cg_child[i].cg_count <= 1)
4946 			return (-1);
4947 		grp = &grp->cg_child[i];
4948 	}
4949 
4950 	/* Must share L2. */
4951 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4952 		return -1;
4953 
4954 	/*
4955 	 * Select the first member of the set that isn't the reference
4956 	 * CPU, which at this point is guaranteed to exist.
4957 	 */
4958 	for (i = 0; i < CPU_SETSIZE; i++) {
4959 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4960 			return (i);
4961 	}
4962 
4963 	/* Should never be reached */
4964 	return (-1);
4965 }
4966 
4967 #else
4968 static int
4969 find_l2_neighbor(int cpu)
4970 {
4971 
4972 	return (-1);
4973 }
4974 #endif
4975 
4976 /*
4977  * CPU mapping behaviors
4978  * ---------------------
4979  * 'separate txrx' refers to the separate_txrx sysctl
4980  * 'use logical' refers to the use_logical_cores sysctl
4981  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4982  *
4983  *  separate     use     INTR
4984  *    txrx     logical   CPUS   result
4985  * ---------- --------- ------ ------------------------------------------------
4986  *     -          -       X     RX and TX queues mapped to consecutive physical
4987  *                              cores with RX/TX pairs on same core and excess
4988  *                              of either following
4989  *     -          X       X     RX and TX queues mapped to consecutive cores
4990  *                              of any type with RX/TX pairs on same core and
4991  *                              excess of either following
4992  *     X          -       X     RX and TX queues mapped to consecutive physical
4993  *                              cores; all RX then all TX
4994  *     X          X       X     RX queues mapped to consecutive physical cores
4995  *                              first, then TX queues mapped to L2 neighbor of
4996  *                              the corresponding RX queue if one exists,
4997  *                              otherwise to consecutive physical cores
4998  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4999  *                              any type with RX/TX pairs on same core and excess
5000  *                              of either following
5001  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5002  *                              any type; all RX then all TX
5003  */
5004 static unsigned int
5005 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5006     bool is_tx)
5007 {
5008 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5009 	unsigned int core_index;
5010 
5011 	if (ctx->ifc_sysctl_separate_txrx) {
5012 		/*
5013 		 * When using separate CPUs for TX and RX, the assignment
5014 		 * will always be of a consecutive CPU out of the set of
5015 		 * context CPUs, except for the specific case where the
5016 		 * context CPUs are phsyical cores, the use of logical cores
5017 		 * has been enabled, the assignment is for TX, the TX qid
5018 		 * corresponds to an RX qid, and the CPU assigned to the
5019 		 * corresponding RX queue has an L2 neighbor.
5020 		 */
5021 		if (ctx->ifc_sysctl_use_logical_cores &&
5022 		    ctx->ifc_cpus_are_physical_cores &&
5023 		    is_tx && qid < scctx->isc_nrxqsets) {
5024 			int l2_neighbor;
5025 			unsigned int rx_cpuid;
5026 
5027 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5028 			l2_neighbor = find_l2_neighbor(rx_cpuid);
5029 			if (l2_neighbor != -1) {
5030 				return (l2_neighbor);
5031 			}
5032 			/*
5033 			 * ... else fall through to the normal
5034 			 * consecutive-after-RX assignment scheme.
5035 			 *
5036 			 * Note that we are assuming that all RX queue CPUs
5037 			 * have an L2 neighbor, or all do not.  If a mixed
5038 			 * scenario is possible, we will have to keep track
5039 			 * separately of how many queues prior to this one
5040 			 * were not able to be assigned to an L2 neighbor.
5041 			 */
5042 		}
5043 		if (is_tx)
5044 			core_index = scctx->isc_nrxqsets + qid;
5045 		else
5046 			core_index = qid;
5047 	} else {
5048 		core_index = qid;
5049 	}
5050 
5051 	return (cpuid_advance(ctx, base_cpuid, core_index));
5052 }
5053 
5054 static uint16_t
5055 get_ctx_core_offset(if_ctx_t ctx)
5056 {
5057 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5058 	struct cpu_offset *op;
5059 	cpuset_t assigned_cpus;
5060 	unsigned int cores_consumed;
5061 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5062 	unsigned int first_valid;
5063 	unsigned int last_valid;
5064 	unsigned int i;
5065 
5066 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5067 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5068 
5069 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5070 		/*
5071 		 * Align the user-chosen base CPU ID to the next valid CPU
5072 		 * for this device.  If the chosen base CPU ID is smaller
5073 		 * than the first valid CPU or larger than the last valid
5074 		 * CPU, we assume the user does not know what the valid
5075 		 * range is for this device and is thinking in terms of a
5076 		 * zero-based reference frame, and so we shift the given
5077 		 * value into the valid range (and wrap accordingly) so the
5078 		 * intent is translated to the proper frame of reference.
5079 		 * If the base CPU ID is within the valid first/last, but
5080 		 * does not correspond to a valid CPU, it is advanced to the
5081 		 * next valid CPU (wrapping if necessary).
5082 		 */
5083 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5084 			/* shift from zero-based to first_valid-based */
5085 			base_cpuid += first_valid;
5086 			/* wrap to range [first_valid, last_valid] */
5087 			base_cpuid = (base_cpuid - first_valid) %
5088 			    (last_valid - first_valid + 1);
5089 		}
5090 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5091 			/*
5092 			 * base_cpuid is in [first_valid, last_valid], but
5093 			 * not a member of the valid set.  In this case,
5094 			 * there will always be a member of the valid set
5095 			 * with a CPU ID that is greater than base_cpuid,
5096 			 * and we simply advance to it.
5097 			 */
5098 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5099 				base_cpuid++;
5100 		}
5101 		return (base_cpuid);
5102 	}
5103 
5104 	/*
5105 	 * Determine how many cores will be consumed by performing the CPU
5106 	 * assignments and counting how many of the assigned CPUs correspond
5107 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5108 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5109 	 * the set of context CPUs.
5110 	 *
5111 	 * Note not all assigned CPUs will be in the set of context CPUs
5112 	 * when separate CPUs are being allocated to TX and RX queues,
5113 	 * assignment to logical cores has been enabled, the set of context
5114 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5115 	 * neighbors of CPUs that RX queues have been mapped to - in this
5116 	 * case we do only want to count how many CPUs in the set of context
5117 	 * CPUs have been consumed, as that determines the next CPU in that
5118 	 * set to start allocating at for the next device for which
5119 	 * core_offset is not set.
5120 	 */
5121 	CPU_ZERO(&assigned_cpus);
5122 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5123 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5124 		    &assigned_cpus);
5125 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5126 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5127 		    &assigned_cpus);
5128 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5129 	cores_consumed = CPU_COUNT(&assigned_cpus);
5130 
5131 	mtx_lock(&cpu_offset_mtx);
5132 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5133 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5134 			base_cpuid = op->next_cpuid;
5135 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5136 			    cores_consumed);
5137 			MPASS(op->refcount < UINT_MAX);
5138 			op->refcount++;
5139 			break;
5140 		}
5141 	}
5142 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5143 		base_cpuid = first_valid;
5144 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5145 		    M_NOWAIT | M_ZERO);
5146 		if (op == NULL) {
5147 			device_printf(ctx->ifc_dev,
5148 			    "allocation for cpu offset failed.\n");
5149 		} else {
5150 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5151 			    cores_consumed);
5152 			op->refcount = 1;
5153 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5154 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5155 		}
5156 	}
5157 	mtx_unlock(&cpu_offset_mtx);
5158 
5159 	return (base_cpuid);
5160 }
5161 
5162 static void
5163 unref_ctx_core_offset(if_ctx_t ctx)
5164 {
5165 	struct cpu_offset *op, *top;
5166 
5167 	mtx_lock(&cpu_offset_mtx);
5168 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5169 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5170 			MPASS(op->refcount > 0);
5171 			op->refcount--;
5172 			if (op->refcount == 0) {
5173 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5174 				free(op, M_IFLIB);
5175 			}
5176 			break;
5177 		}
5178 	}
5179 	mtx_unlock(&cpu_offset_mtx);
5180 }
5181 
5182 int
5183 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5184 {
5185 	if_ctx_t ctx;
5186 	if_t ifp;
5187 	if_softc_ctx_t scctx;
5188 	kobjop_desc_t kobj_desc;
5189 	kobj_method_t *kobj_method;
5190 	int err, msix, rid;
5191 	int num_txd, num_rxd;
5192 
5193 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5194 
5195 	if (sc == NULL) {
5196 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5197 		device_set_softc(dev, ctx);
5198 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5199 	}
5200 
5201 	ctx->ifc_sctx = sctx;
5202 	ctx->ifc_dev = dev;
5203 	ctx->ifc_softc = sc;
5204 
5205 	if ((err = iflib_register(ctx)) != 0) {
5206 		device_printf(dev, "iflib_register failed %d\n", err);
5207 		goto fail_ctx_free;
5208 	}
5209 	iflib_add_device_sysctl_pre(ctx);
5210 
5211 	scctx = &ctx->ifc_softc_ctx;
5212 	ifp = ctx->ifc_ifp;
5213 
5214 	iflib_reset_qvalues(ctx);
5215 	IFNET_WLOCK();
5216 	CTX_LOCK(ctx);
5217 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5218 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5219 		goto fail_unlock;
5220 	}
5221 	_iflib_pre_assert(scctx);
5222 	ctx->ifc_txrx = *scctx->isc_txrx;
5223 
5224 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5225 
5226 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5227 		ctx->ifc_mediap = scctx->isc_media;
5228 
5229 #ifdef INVARIANTS
5230 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5231 		MPASS(scctx->isc_tx_csum_flags);
5232 #endif
5233 
5234 	if_setcapabilities(ifp,
5235 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5236 	if_setcapenable(ifp,
5237 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5238 
5239 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5240 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5241 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5242 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5243 
5244 	num_txd = iflib_num_tx_descs(ctx);
5245 	num_rxd = iflib_num_rx_descs(ctx);
5246 
5247 	/* XXX change for per-queue sizes */
5248 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5249 	    num_txd, num_rxd);
5250 
5251 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5252 		scctx->isc_tx_nsegments = max(1, num_txd /
5253 		    MAX_SINGLE_PACKET_FRACTION);
5254 	if (scctx->isc_tx_tso_segments_max > num_txd /
5255 	    MAX_SINGLE_PACKET_FRACTION)
5256 		scctx->isc_tx_tso_segments_max = max(1,
5257 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5258 
5259 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5260 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5261 		/*
5262 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5263 		 * but some MACs do.
5264 		 */
5265 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5266 		    IP_MAXPACKET));
5267 		/*
5268 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5269 		 * into account.  In the worst case, each of these calls will
5270 		 * add another mbuf and, thus, the requirement for another DMA
5271 		 * segment.  So for best performance, it doesn't make sense to
5272 		 * advertize a maximum of TSO segments that typically will
5273 		 * require defragmentation in iflib_encap().
5274 		 */
5275 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5276 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5277 	}
5278 	if (scctx->isc_rss_table_size == 0)
5279 		scctx->isc_rss_table_size = 64;
5280 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5281 
5282 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5283 	/* XXX format name */
5284 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5285 	    NULL, NULL, "admin");
5286 
5287 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5288 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5289 		device_printf(dev, "Unable to fetch CPU list\n");
5290 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5291 		ctx->ifc_cpus_are_physical_cores = false;
5292 	} else
5293 		ctx->ifc_cpus_are_physical_cores = true;
5294 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5295 
5296 	/*
5297 	** Now set up MSI or MSI-X, should return us the number of supported
5298 	** vectors (will be 1 for a legacy interrupt and MSI).
5299 	*/
5300 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5301 		msix = scctx->isc_vectors;
5302 	} else if (scctx->isc_msix_bar != 0)
5303 	       /*
5304 		* The simple fact that isc_msix_bar is not 0 does not mean we
5305 		* we have a good value there that is known to work.
5306 		*/
5307 		msix = iflib_msix_init(ctx);
5308 	else {
5309 		scctx->isc_vectors = 1;
5310 		scctx->isc_ntxqsets = 1;
5311 		scctx->isc_nrxqsets = 1;
5312 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5313 		msix = 0;
5314 	}
5315 	/* Get memory for the station queues */
5316 	if ((err = iflib_queues_alloc(ctx))) {
5317 		device_printf(dev, "Unable to allocate queue memory\n");
5318 		goto fail_intr_free;
5319 	}
5320 
5321 	if ((err = iflib_qset_structures_setup(ctx)))
5322 		goto fail_queues;
5323 
5324 	/*
5325 	 * Now that we know how many queues there are, get the core offset.
5326 	 */
5327 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5328 
5329 	if (msix > 1) {
5330 		/*
5331 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5332 		 * aren't the default NULL implementation.
5333 		 */
5334 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5335 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5336 		    kobj_desc);
5337 		if (kobj_method == &kobj_desc->deflt) {
5338 			device_printf(dev,
5339 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5340 			err = EOPNOTSUPP;
5341 			goto fail_queues;
5342 		}
5343 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5344 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5345 		    kobj_desc);
5346 		if (kobj_method == &kobj_desc->deflt) {
5347 			device_printf(dev,
5348 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5349 			err = EOPNOTSUPP;
5350 			goto fail_queues;
5351 		}
5352 
5353 		/*
5354 		 * Assign the MSI-X vectors.
5355 		 * Note that the default NULL ifdi_msix_intr_assign method will
5356 		 * fail here, too.
5357 		 */
5358 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5359 		if (err != 0) {
5360 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5361 			    err);
5362 			goto fail_queues;
5363 		}
5364 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5365 		rid = 0;
5366 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5367 			MPASS(msix == 1);
5368 			rid = 1;
5369 		}
5370 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5371 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5372 			goto fail_queues;
5373 		}
5374 	} else {
5375 		device_printf(dev,
5376 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5377 		err = ENODEV;
5378 		goto fail_queues;
5379 	}
5380 
5381 	/*
5382 	 * It prevents a double-locking panic with iflib_media_status when
5383 	 * the driver loads.
5384 	 */
5385 	CTX_UNLOCK(ctx);
5386 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5387 	CTX_LOCK(ctx);
5388 
5389 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5390 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5391 		goto fail_detach;
5392 	}
5393 
5394 	/*
5395 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5396 	 * This must appear after the call to ether_ifattach() because
5397 	 * ether_ifattach() sets if_hdrlen to the default value.
5398 	 */
5399 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5400 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5401 
5402 	if ((err = iflib_netmap_attach(ctx))) {
5403 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5404 		goto fail_detach;
5405 	}
5406 	*ctxp = ctx;
5407 
5408 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5409 
5410 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5411 	iflib_add_device_sysctl_post(ctx);
5412 	iflib_add_pfil(ctx);
5413 	ctx->ifc_flags |= IFC_INIT_DONE;
5414 	CTX_UNLOCK(ctx);
5415 	IFNET_WUNLOCK();
5416 
5417 	return (0);
5418 
5419 fail_detach:
5420 	ether_ifdetach(ctx->ifc_ifp);
5421 fail_queues:
5422 	iflib_tqg_detach(ctx);
5423 	iflib_tx_structures_free(ctx);
5424 	iflib_rx_structures_free(ctx);
5425 	IFDI_DETACH(ctx);
5426 	IFDI_QUEUES_FREE(ctx);
5427 fail_intr_free:
5428 	iflib_free_intr_mem(ctx);
5429 fail_unlock:
5430 	CTX_UNLOCK(ctx);
5431 	IFNET_WUNLOCK();
5432 	iflib_deregister(ctx);
5433 fail_ctx_free:
5434 	device_set_softc(ctx->ifc_dev, NULL);
5435         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5436                 free(ctx->ifc_softc, M_IFLIB);
5437         free(ctx, M_IFLIB);
5438 	return (err);
5439 }
5440 
5441 int
5442 iflib_device_attach(device_t dev)
5443 {
5444 	if_ctx_t ctx;
5445 	if_shared_ctx_t sctx;
5446 
5447 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5448 		return (ENOTSUP);
5449 
5450 	pci_enable_busmaster(dev);
5451 
5452 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5453 }
5454 
5455 int
5456 iflib_device_deregister(if_ctx_t ctx)
5457 {
5458 	if_t ifp = ctx->ifc_ifp;
5459 	device_t dev = ctx->ifc_dev;
5460 
5461 	/* Make sure VLANS are not using driver */
5462 	if (if_vlantrunkinuse(ifp)) {
5463 		device_printf(dev, "Vlan in use, detach first\n");
5464 		return (EBUSY);
5465 	}
5466 #ifdef PCI_IOV
5467 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5468 		device_printf(dev, "SR-IOV in use; detach first.\n");
5469 		return (EBUSY);
5470 	}
5471 #endif
5472 
5473 	STATE_LOCK(ctx);
5474 	ctx->ifc_flags |= IFC_IN_DETACH;
5475 	STATE_UNLOCK(ctx);
5476 
5477 	/* Unregister VLAN handlers before calling iflib_stop() */
5478 	iflib_unregister_vlan_handlers(ctx);
5479 
5480 	iflib_netmap_detach(ifp);
5481 	ether_ifdetach(ifp);
5482 
5483 	CTX_LOCK(ctx);
5484 	iflib_stop(ctx);
5485 	CTX_UNLOCK(ctx);
5486 
5487 	iflib_rem_pfil(ctx);
5488 	if (ctx->ifc_led_dev != NULL)
5489 		led_destroy(ctx->ifc_led_dev);
5490 
5491 	iflib_tqg_detach(ctx);
5492 	iflib_tx_structures_free(ctx);
5493 	iflib_rx_structures_free(ctx);
5494 
5495 	CTX_LOCK(ctx);
5496 	IFDI_DETACH(ctx);
5497 	IFDI_QUEUES_FREE(ctx);
5498 	CTX_UNLOCK(ctx);
5499 
5500 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5501 	iflib_free_intr_mem(ctx);
5502 
5503 	bus_generic_detach(dev);
5504 
5505 	iflib_deregister(ctx);
5506 
5507 	device_set_softc(ctx->ifc_dev, NULL);
5508 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5509 		free(ctx->ifc_softc, M_IFLIB);
5510 	unref_ctx_core_offset(ctx);
5511 	free(ctx, M_IFLIB);
5512 	return (0);
5513 }
5514 
5515 static void
5516 iflib_tqg_detach(if_ctx_t ctx)
5517 {
5518 	iflib_txq_t txq;
5519 	iflib_rxq_t rxq;
5520 	int i;
5521 	struct taskqgroup *tqg;
5522 
5523 	/* XXX drain any dependent tasks */
5524 	tqg = qgroup_if_io_tqg;
5525 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5526 		callout_drain(&txq->ift_timer);
5527 #ifdef DEV_NETMAP
5528 		callout_drain(&txq->ift_netmap_timer);
5529 #endif /* DEV_NETMAP */
5530 		if (txq->ift_task.gt_uniq != NULL)
5531 			taskqgroup_detach(tqg, &txq->ift_task);
5532 	}
5533 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5534 		if (rxq->ifr_task.gt_uniq != NULL)
5535 			taskqgroup_detach(tqg, &rxq->ifr_task);
5536 	}
5537 	tqg = qgroup_if_config_tqg;
5538 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5539 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5540 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5541 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5542 }
5543 
5544 static void
5545 iflib_free_intr_mem(if_ctx_t ctx)
5546 {
5547 
5548 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5549 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5550 	}
5551 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5552 		pci_release_msi(ctx->ifc_dev);
5553 	}
5554 	if (ctx->ifc_msix_mem != NULL) {
5555 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5556 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5557 		ctx->ifc_msix_mem = NULL;
5558 	}
5559 }
5560 
5561 int
5562 iflib_device_detach(device_t dev)
5563 {
5564 	if_ctx_t ctx = device_get_softc(dev);
5565 
5566 	return (iflib_device_deregister(ctx));
5567 }
5568 
5569 int
5570 iflib_device_suspend(device_t dev)
5571 {
5572 	if_ctx_t ctx = device_get_softc(dev);
5573 
5574 	CTX_LOCK(ctx);
5575 	IFDI_SUSPEND(ctx);
5576 	CTX_UNLOCK(ctx);
5577 
5578 	return bus_generic_suspend(dev);
5579 }
5580 int
5581 iflib_device_shutdown(device_t dev)
5582 {
5583 	if_ctx_t ctx = device_get_softc(dev);
5584 
5585 	CTX_LOCK(ctx);
5586 	IFDI_SHUTDOWN(ctx);
5587 	CTX_UNLOCK(ctx);
5588 
5589 	return bus_generic_suspend(dev);
5590 }
5591 
5592 int
5593 iflib_device_resume(device_t dev)
5594 {
5595 	if_ctx_t ctx = device_get_softc(dev);
5596 	iflib_txq_t txq = ctx->ifc_txqs;
5597 
5598 	CTX_LOCK(ctx);
5599 	IFDI_RESUME(ctx);
5600 	iflib_if_init_locked(ctx);
5601 	CTX_UNLOCK(ctx);
5602 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5603 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5604 
5605 	return (bus_generic_resume(dev));
5606 }
5607 
5608 int
5609 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5610 {
5611 	int error;
5612 	if_ctx_t ctx = device_get_softc(dev);
5613 
5614 	CTX_LOCK(ctx);
5615 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5616 	CTX_UNLOCK(ctx);
5617 
5618 	return (error);
5619 }
5620 
5621 void
5622 iflib_device_iov_uninit(device_t dev)
5623 {
5624 	if_ctx_t ctx = device_get_softc(dev);
5625 
5626 	CTX_LOCK(ctx);
5627 	IFDI_IOV_UNINIT(ctx);
5628 	CTX_UNLOCK(ctx);
5629 }
5630 
5631 int
5632 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5633 {
5634 	int error;
5635 	if_ctx_t ctx = device_get_softc(dev);
5636 
5637 	CTX_LOCK(ctx);
5638 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5639 	CTX_UNLOCK(ctx);
5640 
5641 	return (error);
5642 }
5643 
5644 /*********************************************************************
5645  *
5646  *  MODULE FUNCTION DEFINITIONS
5647  *
5648  **********************************************************************/
5649 
5650 /*
5651  * - Start a fast taskqueue thread for each core
5652  * - Start a taskqueue for control operations
5653  */
5654 static int
5655 iflib_module_init(void)
5656 {
5657 	iflib_timer_default = hz / 2;
5658 	return (0);
5659 }
5660 
5661 static int
5662 iflib_module_event_handler(module_t mod, int what, void *arg)
5663 {
5664 	int err;
5665 
5666 	switch (what) {
5667 	case MOD_LOAD:
5668 		if ((err = iflib_module_init()) != 0)
5669 			return (err);
5670 		break;
5671 	case MOD_UNLOAD:
5672 		return (EBUSY);
5673 	default:
5674 		return (EOPNOTSUPP);
5675 	}
5676 
5677 	return (0);
5678 }
5679 
5680 /*********************************************************************
5681  *
5682  *  PUBLIC FUNCTION DEFINITIONS
5683  *     ordered as in iflib.h
5684  *
5685  **********************************************************************/
5686 
5687 static void
5688 _iflib_assert(if_shared_ctx_t sctx)
5689 {
5690 	int i;
5691 
5692 	MPASS(sctx->isc_tx_maxsize);
5693 	MPASS(sctx->isc_tx_maxsegsize);
5694 
5695 	MPASS(sctx->isc_rx_maxsize);
5696 	MPASS(sctx->isc_rx_nsegments);
5697 	MPASS(sctx->isc_rx_maxsegsize);
5698 
5699 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5700 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5701 		MPASS(sctx->isc_nrxd_min[i]);
5702 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5703 		MPASS(sctx->isc_nrxd_max[i]);
5704 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5705 		MPASS(sctx->isc_nrxd_default[i]);
5706 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5707 	}
5708 
5709 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5710 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5711 		MPASS(sctx->isc_ntxd_min[i]);
5712 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5713 		MPASS(sctx->isc_ntxd_max[i]);
5714 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5715 		MPASS(sctx->isc_ntxd_default[i]);
5716 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5717 	}
5718 }
5719 
5720 static void
5721 _iflib_pre_assert(if_softc_ctx_t scctx)
5722 {
5723 
5724 	MPASS(scctx->isc_txrx->ift_txd_encap);
5725 	MPASS(scctx->isc_txrx->ift_txd_flush);
5726 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5727 	MPASS(scctx->isc_txrx->ift_rxd_available);
5728 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5729 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5730 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5731 }
5732 
5733 static int
5734 iflib_register(if_ctx_t ctx)
5735 {
5736 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5737 	driver_t *driver = sctx->isc_driver;
5738 	device_t dev = ctx->ifc_dev;
5739 	if_t ifp;
5740 
5741 	_iflib_assert(sctx);
5742 
5743 	CTX_LOCK_INIT(ctx);
5744 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5745 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5746 	if (ifp == NULL) {
5747 		device_printf(dev, "can not allocate ifnet structure\n");
5748 		return (ENOMEM);
5749 	}
5750 
5751 	/*
5752 	 * Initialize our context's device specific methods
5753 	 */
5754 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5755 	kobj_class_compile((kobj_class_t) driver);
5756 
5757 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5758 	if_setsoftc(ifp, ctx);
5759 	if_setdev(ifp, dev);
5760 	if_setinitfn(ifp, iflib_if_init);
5761 	if_setioctlfn(ifp, iflib_if_ioctl);
5762 #ifdef ALTQ
5763 	if_setstartfn(ifp, iflib_altq_if_start);
5764 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5765 	if_setsendqready(ifp);
5766 #else
5767 	if_settransmitfn(ifp, iflib_if_transmit);
5768 #endif
5769 	if_setqflushfn(ifp, iflib_if_qflush);
5770 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5771 	ctx->ifc_vlan_attach_event =
5772 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5773 							  EVENTHANDLER_PRI_FIRST);
5774 	ctx->ifc_vlan_detach_event =
5775 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5776 							  EVENTHANDLER_PRI_FIRST);
5777 
5778 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5779 		ctx->ifc_mediap = &ctx->ifc_media;
5780 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5781 		    iflib_media_change, iflib_media_status);
5782 	}
5783 	return (0);
5784 }
5785 
5786 static void
5787 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5788 {
5789 	/* Unregister VLAN events */
5790 	if (ctx->ifc_vlan_attach_event != NULL) {
5791 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5792 		ctx->ifc_vlan_attach_event = NULL;
5793 	}
5794 	if (ctx->ifc_vlan_detach_event != NULL) {
5795 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5796 		ctx->ifc_vlan_detach_event = NULL;
5797 	}
5798 
5799 }
5800 
5801 static void
5802 iflib_deregister(if_ctx_t ctx)
5803 {
5804 	if_t ifp = ctx->ifc_ifp;
5805 
5806 	/* Remove all media */
5807 	ifmedia_removeall(&ctx->ifc_media);
5808 
5809 	/* Ensure that VLAN event handlers are unregistered */
5810 	iflib_unregister_vlan_handlers(ctx);
5811 
5812 	/* Release kobject reference */
5813 	kobj_delete((kobj_t) ctx, NULL);
5814 
5815 	/* Free the ifnet structure */
5816 	if_free(ifp);
5817 
5818 	STATE_LOCK_DESTROY(ctx);
5819 
5820 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5821 	CTX_LOCK_DESTROY(ctx);
5822 }
5823 
5824 static int
5825 iflib_queues_alloc(if_ctx_t ctx)
5826 {
5827 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5828 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5829 	device_t dev = ctx->ifc_dev;
5830 	int nrxqsets = scctx->isc_nrxqsets;
5831 	int ntxqsets = scctx->isc_ntxqsets;
5832 	iflib_txq_t txq;
5833 	iflib_rxq_t rxq;
5834 	iflib_fl_t fl = NULL;
5835 	int i, j, cpu, err, txconf, rxconf;
5836 	iflib_dma_info_t ifdip;
5837 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5838 	uint32_t *txqsizes = scctx->isc_txqsizes;
5839 	uint8_t nrxqs = sctx->isc_nrxqs;
5840 	uint8_t ntxqs = sctx->isc_ntxqs;
5841 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5842 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5843 	caddr_t *vaddrs;
5844 	uint64_t *paddrs;
5845 
5846 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5847 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5848 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5849            ("there must be at least a rxq for each free list"));
5850 
5851 	/* Allocate the TX ring struct memory */
5852 	if (!(ctx->ifc_txqs =
5853 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5854 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5855 		device_printf(dev, "Unable to allocate TX ring memory\n");
5856 		err = ENOMEM;
5857 		goto fail;
5858 	}
5859 
5860 	/* Now allocate the RX */
5861 	if (!(ctx->ifc_rxqs =
5862 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5863 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5864 		device_printf(dev, "Unable to allocate RX ring memory\n");
5865 		err = ENOMEM;
5866 		goto rx_fail;
5867 	}
5868 
5869 	txq = ctx->ifc_txqs;
5870 	rxq = ctx->ifc_rxqs;
5871 
5872 	/*
5873 	 * XXX handle allocation failure
5874 	 */
5875 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5876 		/* Set up some basics */
5877 
5878 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5879 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5880 			device_printf(dev,
5881 			    "Unable to allocate TX DMA info memory\n");
5882 			err = ENOMEM;
5883 			goto err_tx_desc;
5884 		}
5885 		txq->ift_ifdi = ifdip;
5886 		for (j = 0; j < ntxqs; j++, ifdip++) {
5887 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5888 				device_printf(dev,
5889 				    "Unable to allocate TX descriptors\n");
5890 				err = ENOMEM;
5891 				goto err_tx_desc;
5892 			}
5893 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5894 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5895 		}
5896 		txq->ift_ctx = ctx;
5897 		txq->ift_id = i;
5898 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5899 			txq->ift_br_offset = 1;
5900 		} else {
5901 			txq->ift_br_offset = 0;
5902 		}
5903 
5904 		if (iflib_txsd_alloc(txq)) {
5905 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5906 			err = ENOMEM;
5907 			goto err_tx_desc;
5908 		}
5909 
5910 		/* Initialize the TX lock */
5911 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5912 		    device_get_nameunit(dev), txq->ift_id);
5913 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5914 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5915 		txq->ift_timer.c_cpu = cpu;
5916 #ifdef DEV_NETMAP
5917 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5918 		txq->ift_netmap_timer.c_cpu = cpu;
5919 #endif /* DEV_NETMAP */
5920 
5921 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5922 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5923 		if (err) {
5924 			/* XXX free any allocated rings */
5925 			device_printf(dev, "Unable to allocate buf_ring\n");
5926 			goto err_tx_desc;
5927 		}
5928 	}
5929 
5930 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5931 		/* Set up some basics */
5932 		callout_init(&rxq->ifr_watchdog, 1);
5933 
5934 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5935 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5936 			device_printf(dev,
5937 			    "Unable to allocate RX DMA info memory\n");
5938 			err = ENOMEM;
5939 			goto err_tx_desc;
5940 		}
5941 
5942 		rxq->ifr_ifdi = ifdip;
5943 		/* XXX this needs to be changed if #rx queues != #tx queues */
5944 		rxq->ifr_ntxqirq = 1;
5945 		rxq->ifr_txqid[0] = i;
5946 		for (j = 0; j < nrxqs; j++, ifdip++) {
5947 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5948 				device_printf(dev,
5949 				    "Unable to allocate RX descriptors\n");
5950 				err = ENOMEM;
5951 				goto err_tx_desc;
5952 			}
5953 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5954 		}
5955 		rxq->ifr_ctx = ctx;
5956 		rxq->ifr_id = i;
5957 		rxq->ifr_fl_offset = fl_offset;
5958 		rxq->ifr_nfl = nfree_lists;
5959 		if (!(fl =
5960 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5961 			device_printf(dev, "Unable to allocate free list memory\n");
5962 			err = ENOMEM;
5963 			goto err_tx_desc;
5964 		}
5965 		rxq->ifr_fl = fl;
5966 		for (j = 0; j < nfree_lists; j++) {
5967 			fl[j].ifl_rxq = rxq;
5968 			fl[j].ifl_id = j;
5969 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5970 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5971 		}
5972 		/* Allocate receive buffers for the ring */
5973 		if (iflib_rxsd_alloc(rxq)) {
5974 			device_printf(dev,
5975 			    "Critical Failure setting up receive buffers\n");
5976 			err = ENOMEM;
5977 			goto err_rx_desc;
5978 		}
5979 
5980 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5981 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5982 			    M_WAITOK);
5983 	}
5984 
5985 	/* TXQs */
5986 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5987 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5988 	for (i = 0; i < ntxqsets; i++) {
5989 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5990 
5991 		for (j = 0; j < ntxqs; j++, di++) {
5992 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5993 			paddrs[i*ntxqs + j] = di->idi_paddr;
5994 		}
5995 	}
5996 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5997 		device_printf(ctx->ifc_dev,
5998 		    "Unable to allocate device TX queue\n");
5999 		iflib_tx_structures_free(ctx);
6000 		free(vaddrs, M_IFLIB);
6001 		free(paddrs, M_IFLIB);
6002 		goto err_rx_desc;
6003 	}
6004 	free(vaddrs, M_IFLIB);
6005 	free(paddrs, M_IFLIB);
6006 
6007 	/* RXQs */
6008 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6009 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6010 	for (i = 0; i < nrxqsets; i++) {
6011 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6012 
6013 		for (j = 0; j < nrxqs; j++, di++) {
6014 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6015 			paddrs[i*nrxqs + j] = di->idi_paddr;
6016 		}
6017 	}
6018 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6019 		device_printf(ctx->ifc_dev,
6020 		    "Unable to allocate device RX queue\n");
6021 		iflib_tx_structures_free(ctx);
6022 		free(vaddrs, M_IFLIB);
6023 		free(paddrs, M_IFLIB);
6024 		goto err_rx_desc;
6025 	}
6026 	free(vaddrs, M_IFLIB);
6027 	free(paddrs, M_IFLIB);
6028 
6029 	return (0);
6030 
6031 /* XXX handle allocation failure changes */
6032 err_rx_desc:
6033 err_tx_desc:
6034 rx_fail:
6035 	if (ctx->ifc_rxqs != NULL)
6036 		free(ctx->ifc_rxqs, M_IFLIB);
6037 	ctx->ifc_rxqs = NULL;
6038 	if (ctx->ifc_txqs != NULL)
6039 		free(ctx->ifc_txqs, M_IFLIB);
6040 	ctx->ifc_txqs = NULL;
6041 fail:
6042 	return (err);
6043 }
6044 
6045 static int
6046 iflib_tx_structures_setup(if_ctx_t ctx)
6047 {
6048 	iflib_txq_t txq = ctx->ifc_txqs;
6049 	int i;
6050 
6051 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6052 		iflib_txq_setup(txq);
6053 
6054 	return (0);
6055 }
6056 
6057 static void
6058 iflib_tx_structures_free(if_ctx_t ctx)
6059 {
6060 	iflib_txq_t txq = ctx->ifc_txqs;
6061 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6062 	int i, j;
6063 
6064 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6065 		for (j = 0; j < sctx->isc_ntxqs; j++)
6066 			iflib_dma_free(&txq->ift_ifdi[j]);
6067 		iflib_txq_destroy(txq);
6068 	}
6069 	free(ctx->ifc_txqs, M_IFLIB);
6070 	ctx->ifc_txqs = NULL;
6071 }
6072 
6073 /*********************************************************************
6074  *
6075  *  Initialize all receive rings.
6076  *
6077  **********************************************************************/
6078 static int
6079 iflib_rx_structures_setup(if_ctx_t ctx)
6080 {
6081 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6082 	int q;
6083 #if defined(INET6) || defined(INET)
6084 	int err, i;
6085 #endif
6086 
6087 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6088 #if defined(INET6) || defined(INET)
6089 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6090 		    TCP_LRO_ENTRIES, min(1024,
6091 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6092 		if (err != 0) {
6093 			device_printf(ctx->ifc_dev,
6094 			    "LRO Initialization failed!\n");
6095 			goto fail;
6096 		}
6097 #endif
6098 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6099 	}
6100 	return (0);
6101 #if defined(INET6) || defined(INET)
6102 fail:
6103 	/*
6104 	 * Free LRO resources allocated so far, we will only handle
6105 	 * the rings that completed, the failing case will have
6106 	 * cleaned up for itself.  'q' failed, so its the terminus.
6107 	 */
6108 	rxq = ctx->ifc_rxqs;
6109 	for (i = 0; i < q; ++i, rxq++) {
6110 		tcp_lro_free(&rxq->ifr_lc);
6111 	}
6112 	return (err);
6113 #endif
6114 }
6115 
6116 /*********************************************************************
6117  *
6118  *  Free all receive rings.
6119  *
6120  **********************************************************************/
6121 static void
6122 iflib_rx_structures_free(if_ctx_t ctx)
6123 {
6124 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6125 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6126 	int i, j;
6127 
6128 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6129 		for (j = 0; j < sctx->isc_nrxqs; j++)
6130 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6131 		iflib_rx_sds_free(rxq);
6132 #if defined(INET6) || defined(INET)
6133 		tcp_lro_free(&rxq->ifr_lc);
6134 #endif
6135 	}
6136 	free(ctx->ifc_rxqs, M_IFLIB);
6137 	ctx->ifc_rxqs = NULL;
6138 }
6139 
6140 static int
6141 iflib_qset_structures_setup(if_ctx_t ctx)
6142 {
6143 	int err;
6144 
6145 	/*
6146 	 * It is expected that the caller takes care of freeing queues if this
6147 	 * fails.
6148 	 */
6149 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6150 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6151 		return (err);
6152 	}
6153 
6154 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6155 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6156 
6157 	return (err);
6158 }
6159 
6160 int
6161 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6162 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6163 {
6164 
6165 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6166 }
6167 
6168 /* Just to avoid copy/paste */
6169 static inline int
6170 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6171     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6172     const char *name)
6173 {
6174 	device_t dev;
6175 	unsigned int base_cpuid, cpuid;
6176 	int err;
6177 
6178 	dev = ctx->ifc_dev;
6179 	base_cpuid = ctx->ifc_sysctl_core_offset;
6180 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6181 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6182 	    irq ? irq->ii_res : NULL, name);
6183 	if (err) {
6184 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6185 		return (err);
6186 	}
6187 #ifdef notyet
6188 	if (cpuid > ctx->ifc_cpuid_highest)
6189 		ctx->ifc_cpuid_highest = cpuid;
6190 #endif
6191 	return (0);
6192 }
6193 
6194 int
6195 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6196 			iflib_intr_type_t type, driver_filter_t *filter,
6197 			void *filter_arg, int qid, const char *name)
6198 {
6199 	device_t dev;
6200 	struct grouptask *gtask;
6201 	struct taskqgroup *tqg;
6202 	iflib_filter_info_t info;
6203 	gtask_fn_t *fn;
6204 	int tqrid, err;
6205 	driver_filter_t *intr_fast;
6206 	void *q;
6207 
6208 	info = &ctx->ifc_filter_info;
6209 	tqrid = rid;
6210 
6211 	switch (type) {
6212 	/* XXX merge tx/rx for netmap? */
6213 	case IFLIB_INTR_TX:
6214 		q = &ctx->ifc_txqs[qid];
6215 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6216 		gtask = &ctx->ifc_txqs[qid].ift_task;
6217 		tqg = qgroup_if_io_tqg;
6218 		fn = _task_fn_tx;
6219 		intr_fast = iflib_fast_intr;
6220 		GROUPTASK_INIT(gtask, 0, fn, q);
6221 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6222 		break;
6223 	case IFLIB_INTR_RX:
6224 		q = &ctx->ifc_rxqs[qid];
6225 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6226 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6227 		tqg = qgroup_if_io_tqg;
6228 		fn = _task_fn_rx;
6229 		intr_fast = iflib_fast_intr;
6230 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6231 		break;
6232 	case IFLIB_INTR_RXTX:
6233 		q = &ctx->ifc_rxqs[qid];
6234 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6235 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6236 		tqg = qgroup_if_io_tqg;
6237 		fn = _task_fn_rx;
6238 		intr_fast = iflib_fast_intr_rxtx;
6239 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6240 		break;
6241 	case IFLIB_INTR_ADMIN:
6242 		q = ctx;
6243 		tqrid = -1;
6244 		info = &ctx->ifc_filter_info;
6245 		gtask = &ctx->ifc_admin_task;
6246 		tqg = qgroup_if_config_tqg;
6247 		fn = _task_fn_admin;
6248 		intr_fast = iflib_fast_intr_ctx;
6249 		break;
6250 	default:
6251 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6252 		    __func__);
6253 		return (EINVAL);
6254 	}
6255 
6256 	info->ifi_filter = filter;
6257 	info->ifi_filter_arg = filter_arg;
6258 	info->ifi_task = gtask;
6259 	info->ifi_ctx = q;
6260 
6261 	dev = ctx->ifc_dev;
6262 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6263 	if (err != 0) {
6264 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6265 		return (err);
6266 	}
6267 	if (type == IFLIB_INTR_ADMIN)
6268 		return (0);
6269 
6270 	if (tqrid != -1) {
6271 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6272 		    name);
6273 		if (err)
6274 			return (err);
6275 	} else {
6276 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6277 	}
6278 
6279 	return (0);
6280 }
6281 
6282 void
6283 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6284 			    void *arg, int qid, const char *name)
6285 {
6286 	device_t dev;
6287 	struct grouptask *gtask;
6288 	struct taskqgroup *tqg;
6289 	gtask_fn_t *fn;
6290 	void *q;
6291 	int err;
6292 
6293 	switch (type) {
6294 	case IFLIB_INTR_TX:
6295 		q = &ctx->ifc_txqs[qid];
6296 		gtask = &ctx->ifc_txqs[qid].ift_task;
6297 		tqg = qgroup_if_io_tqg;
6298 		fn = _task_fn_tx;
6299 		GROUPTASK_INIT(gtask, 0, fn, q);
6300 		break;
6301 	case IFLIB_INTR_RX:
6302 		q = &ctx->ifc_rxqs[qid];
6303 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6304 		tqg = qgroup_if_io_tqg;
6305 		fn = _task_fn_rx;
6306 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6307 		break;
6308 	case IFLIB_INTR_IOV:
6309 		q = ctx;
6310 		gtask = &ctx->ifc_vflr_task;
6311 		tqg = qgroup_if_config_tqg;
6312 		fn = _task_fn_iov;
6313 		GROUPTASK_INIT(gtask, 0, fn, q);
6314 		break;
6315 	default:
6316 		panic("unknown net intr type");
6317 	}
6318 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6319 	if (err) {
6320 		dev = ctx->ifc_dev;
6321 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6322 		    name);
6323 	}
6324 }
6325 
6326 void
6327 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6328 {
6329 
6330 	if (irq->ii_tag)
6331 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6332 
6333 	if (irq->ii_res)
6334 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6335 		    rman_get_rid(irq->ii_res), irq->ii_res);
6336 }
6337 
6338 static int
6339 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6340 {
6341 	iflib_txq_t txq = ctx->ifc_txqs;
6342 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6343 	if_irq_t irq = &ctx->ifc_legacy_irq;
6344 	iflib_filter_info_t info;
6345 	device_t dev;
6346 	struct grouptask *gtask;
6347 	struct resource *res;
6348 	struct taskqgroup *tqg;
6349 	void *q;
6350 	int err, tqrid;
6351 	bool rx_only;
6352 
6353 	q = &ctx->ifc_rxqs[0];
6354 	info = &rxq[0].ifr_filter_info;
6355 	gtask = &rxq[0].ifr_task;
6356 	tqg = qgroup_if_io_tqg;
6357 	tqrid = *rid;
6358 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6359 
6360 	ctx->ifc_flags |= IFC_LEGACY;
6361 	info->ifi_filter = filter;
6362 	info->ifi_filter_arg = filter_arg;
6363 	info->ifi_task = gtask;
6364 	info->ifi_ctx = rx_only ? ctx : q;
6365 
6366 	dev = ctx->ifc_dev;
6367 	/* We allocate a single interrupt resource */
6368 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6369 	    iflib_fast_intr_rxtx, NULL, info, name);
6370 	if (err != 0)
6371 		return (err);
6372 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6373 	res = irq->ii_res;
6374 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6375 
6376 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6377 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6378 	    "tx");
6379 	return (0);
6380 }
6381 
6382 void
6383 iflib_led_create(if_ctx_t ctx)
6384 {
6385 
6386 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6387 	    device_get_nameunit(ctx->ifc_dev));
6388 }
6389 
6390 void
6391 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6392 {
6393 
6394 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6395 }
6396 
6397 void
6398 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6399 {
6400 
6401 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6402 }
6403 
6404 void
6405 iflib_admin_intr_deferred(if_ctx_t ctx)
6406 {
6407 
6408 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6409 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6410 }
6411 
6412 void
6413 iflib_iov_intr_deferred(if_ctx_t ctx)
6414 {
6415 
6416 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6417 }
6418 
6419 void
6420 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6421 {
6422 
6423 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6424 	    name);
6425 }
6426 
6427 void
6428 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6429 	const char *name)
6430 {
6431 
6432 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6433 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6434 	    name);
6435 }
6436 
6437 void
6438 iflib_config_gtask_deinit(struct grouptask *gtask)
6439 {
6440 
6441 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6442 }
6443 
6444 void
6445 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6446 {
6447 	if_t ifp = ctx->ifc_ifp;
6448 	iflib_txq_t txq = ctx->ifc_txqs;
6449 
6450 	if_setbaudrate(ifp, baudrate);
6451 	if (baudrate >= IF_Gbps(10)) {
6452 		STATE_LOCK(ctx);
6453 		ctx->ifc_flags |= IFC_PREFETCH;
6454 		STATE_UNLOCK(ctx);
6455 	}
6456 	/* If link down, disable watchdog */
6457 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6458 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6459 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6460 	}
6461 	ctx->ifc_link_state = link_state;
6462 	if_link_state_change(ifp, link_state);
6463 }
6464 
6465 static int
6466 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6467 {
6468 	int credits;
6469 #ifdef INVARIANTS
6470 	int credits_pre = txq->ift_cidx_processed;
6471 #endif
6472 
6473 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6474 	    BUS_DMASYNC_POSTREAD);
6475 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6476 		return (0);
6477 
6478 	txq->ift_processed += credits;
6479 	txq->ift_cidx_processed += credits;
6480 
6481 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6482 	if (txq->ift_cidx_processed >= txq->ift_size)
6483 		txq->ift_cidx_processed -= txq->ift_size;
6484 	return (credits);
6485 }
6486 
6487 static int
6488 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6489 {
6490 	iflib_fl_t fl;
6491 	u_int i;
6492 
6493 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6494 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6495 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6496 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6497 	    budget));
6498 }
6499 
6500 void
6501 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6502 	const char *description, if_int_delay_info_t info,
6503 	int offset, int value)
6504 {
6505 	info->iidi_ctx = ctx;
6506 	info->iidi_offset = offset;
6507 	info->iidi_value = value;
6508 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6509 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6510 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6511 	    info, 0, iflib_sysctl_int_delay, "I", description);
6512 }
6513 
6514 struct sx *
6515 iflib_ctx_lock_get(if_ctx_t ctx)
6516 {
6517 
6518 	return (&ctx->ifc_ctx_sx);
6519 }
6520 
6521 static int
6522 iflib_msix_init(if_ctx_t ctx)
6523 {
6524 	device_t dev = ctx->ifc_dev;
6525 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6526 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6527 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6528 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6529 
6530 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6531 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6532 
6533 	if (bootverbose)
6534 		device_printf(dev, "msix_init qsets capped at %d\n",
6535 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6536 
6537 	/* Override by tuneable */
6538 	if (scctx->isc_disable_msix)
6539 		goto msi;
6540 
6541 	/* First try MSI-X */
6542 	if ((msgs = pci_msix_count(dev)) == 0) {
6543 		if (bootverbose)
6544 			device_printf(dev, "MSI-X not supported or disabled\n");
6545 		goto msi;
6546 	}
6547 
6548 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6549 	/*
6550 	 * bar == -1 => "trust me I know what I'm doing"
6551 	 * Some drivers are for hardware that is so shoddily
6552 	 * documented that no one knows which bars are which
6553 	 * so the developer has to map all bars. This hack
6554 	 * allows shoddy garbage to use MSI-X in this framework.
6555 	 */
6556 	if (bar != -1) {
6557 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6558 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6559 		if (ctx->ifc_msix_mem == NULL) {
6560 			device_printf(dev, "Unable to map MSI-X table\n");
6561 			goto msi;
6562 		}
6563 	}
6564 
6565 	admincnt = sctx->isc_admin_intrcnt;
6566 #if IFLIB_DEBUG
6567 	/* use only 1 qset in debug mode */
6568 	queuemsgs = min(msgs - admincnt, 1);
6569 #else
6570 	queuemsgs = msgs - admincnt;
6571 #endif
6572 #ifdef RSS
6573 	queues = imin(queuemsgs, rss_getnumbuckets());
6574 #else
6575 	queues = queuemsgs;
6576 #endif
6577 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6578 	if (bootverbose)
6579 		device_printf(dev,
6580 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6581 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6582 #ifdef  RSS
6583 	/* If we're doing RSS, clamp at the number of RSS buckets */
6584 	if (queues > rss_getnumbuckets())
6585 		queues = rss_getnumbuckets();
6586 #endif
6587 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6588 		rx_queues = iflib_num_rx_queues;
6589 	else
6590 		rx_queues = queues;
6591 
6592 	if (rx_queues > scctx->isc_nrxqsets)
6593 		rx_queues = scctx->isc_nrxqsets;
6594 
6595 	/*
6596 	 * We want this to be all logical CPUs by default
6597 	 */
6598 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6599 		tx_queues = iflib_num_tx_queues;
6600 	else
6601 		tx_queues = mp_ncpus;
6602 
6603 	if (tx_queues > scctx->isc_ntxqsets)
6604 		tx_queues = scctx->isc_ntxqsets;
6605 
6606 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6607 #ifdef INVARIANTS
6608 		if (tx_queues != rx_queues)
6609 			device_printf(dev,
6610 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6611 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6612 #endif
6613 		tx_queues = min(rx_queues, tx_queues);
6614 		rx_queues = min(rx_queues, tx_queues);
6615 	}
6616 
6617 	vectors = rx_queues + admincnt;
6618 	if (msgs < vectors) {
6619 		device_printf(dev,
6620 		    "insufficient number of MSI-X vectors "
6621 		    "(supported %d, need %d)\n", msgs, vectors);
6622 		goto msi;
6623 	}
6624 
6625 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6626 	    tx_queues);
6627 	msgs = vectors;
6628 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6629 		if (vectors != msgs) {
6630 			device_printf(dev,
6631 			    "Unable to allocate sufficient MSI-X vectors "
6632 			    "(got %d, need %d)\n", vectors, msgs);
6633 			pci_release_msi(dev);
6634 			if (bar != -1) {
6635 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6636 				    ctx->ifc_msix_mem);
6637 				ctx->ifc_msix_mem = NULL;
6638 			}
6639 			goto msi;
6640 		}
6641 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6642 		    vectors);
6643 		scctx->isc_vectors = vectors;
6644 		scctx->isc_nrxqsets = rx_queues;
6645 		scctx->isc_ntxqsets = tx_queues;
6646 		scctx->isc_intr = IFLIB_INTR_MSIX;
6647 
6648 		return (vectors);
6649 	} else {
6650 		device_printf(dev,
6651 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6652 		    err);
6653 		if (bar != -1) {
6654 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6655 			    ctx->ifc_msix_mem);
6656 			ctx->ifc_msix_mem = NULL;
6657 		}
6658 	}
6659 
6660 msi:
6661 	vectors = pci_msi_count(dev);
6662 	scctx->isc_nrxqsets = 1;
6663 	scctx->isc_ntxqsets = 1;
6664 	scctx->isc_vectors = vectors;
6665 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6666 		device_printf(dev,"Using an MSI interrupt\n");
6667 		scctx->isc_intr = IFLIB_INTR_MSI;
6668 	} else {
6669 		scctx->isc_vectors = 1;
6670 		device_printf(dev,"Using a Legacy interrupt\n");
6671 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6672 	}
6673 
6674 	return (vectors);
6675 }
6676 
6677 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6678 
6679 static int
6680 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6681 {
6682 	int rc;
6683 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6684 	struct sbuf *sb;
6685 	const char *ring_state = "UNKNOWN";
6686 
6687 	/* XXX needed ? */
6688 	rc = sysctl_wire_old_buffer(req, 0);
6689 	MPASS(rc == 0);
6690 	if (rc != 0)
6691 		return (rc);
6692 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6693 	MPASS(sb != NULL);
6694 	if (sb == NULL)
6695 		return (ENOMEM);
6696 	if (state[3] <= 3)
6697 		ring_state = ring_states[state[3]];
6698 
6699 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6700 		    state[0], state[1], state[2], ring_state);
6701 	rc = sbuf_finish(sb);
6702 	sbuf_delete(sb);
6703         return(rc);
6704 }
6705 
6706 enum iflib_ndesc_handler {
6707 	IFLIB_NTXD_HANDLER,
6708 	IFLIB_NRXD_HANDLER,
6709 };
6710 
6711 static int
6712 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6713 {
6714 	if_ctx_t ctx = (void *)arg1;
6715 	enum iflib_ndesc_handler type = arg2;
6716 	char buf[256] = {0};
6717 	qidx_t *ndesc;
6718 	char *p, *next;
6719 	int nqs, rc, i;
6720 
6721 	nqs = 8;
6722 	switch(type) {
6723 	case IFLIB_NTXD_HANDLER:
6724 		ndesc = ctx->ifc_sysctl_ntxds;
6725 		if (ctx->ifc_sctx)
6726 			nqs = ctx->ifc_sctx->isc_ntxqs;
6727 		break;
6728 	case IFLIB_NRXD_HANDLER:
6729 		ndesc = ctx->ifc_sysctl_nrxds;
6730 		if (ctx->ifc_sctx)
6731 			nqs = ctx->ifc_sctx->isc_nrxqs;
6732 		break;
6733 	default:
6734 		printf("%s: unhandled type\n", __func__);
6735 		return (EINVAL);
6736 	}
6737 	if (nqs == 0)
6738 		nqs = 8;
6739 
6740 	for (i=0; i<8; i++) {
6741 		if (i >= nqs)
6742 			break;
6743 		if (i)
6744 			strcat(buf, ",");
6745 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6746 	}
6747 
6748 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6749 	if (rc || req->newptr == NULL)
6750 		return rc;
6751 
6752 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6753 	    i++, p = strsep(&next, " ,")) {
6754 		ndesc[i] = strtoul(p, NULL, 10);
6755 	}
6756 
6757 	return(rc);
6758 }
6759 
6760 #define NAME_BUFLEN 32
6761 static void
6762 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6763 {
6764         device_t dev = iflib_get_dev(ctx);
6765 	struct sysctl_oid_list *child, *oid_list;
6766 	struct sysctl_ctx_list *ctx_list;
6767 	struct sysctl_oid *node;
6768 
6769 	ctx_list = device_get_sysctl_ctx(dev);
6770 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6771 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child,
6772 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
6773 	    "IFLIB fields");
6774 	oid_list = SYSCTL_CHILDREN(node);
6775 
6776 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6777 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
6778 
6779 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6780 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6781 	    "# of txqs to use, 0 => use default #");
6782 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6783 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6784 	    "# of rxqs to use, 0 => use default #");
6785 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6786 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6787 	    "permit #txq != #rxq");
6788 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6789 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6790 	    "disable MSI-X (default 0)");
6791 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6792 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
6793 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6794 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6795 	    "cause TX to abdicate instead of running to completion");
6796 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6797 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6798 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6799 	    "offset to start using cores at");
6800 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6801 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6802 	    "use separate cores for TX and RX");
6803 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6804 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6805 	    "try to make use of logical cores for TX and RX");
6806 
6807 	/* XXX change for per-queue sizes */
6808 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6809 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6810 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6811 	    "list of # of TX descriptors to use, 0 = use default #");
6812 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6813 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6814 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6815 	    "list of # of RX descriptors to use, 0 = use default #");
6816 }
6817 
6818 static void
6819 iflib_add_device_sysctl_post(if_ctx_t ctx)
6820 {
6821 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6822 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6823         device_t dev = iflib_get_dev(ctx);
6824 	struct sysctl_oid_list *child;
6825 	struct sysctl_ctx_list *ctx_list;
6826 	iflib_fl_t fl;
6827 	iflib_txq_t txq;
6828 	iflib_rxq_t rxq;
6829 	int i, j;
6830 	char namebuf[NAME_BUFLEN];
6831 	char *qfmt;
6832 	struct sysctl_oid *queue_node, *fl_node, *node;
6833 	struct sysctl_oid_list *queue_list, *fl_list;
6834 	ctx_list = device_get_sysctl_ctx(dev);
6835 
6836 	node = ctx->ifc_sysctl_node;
6837 	child = SYSCTL_CHILDREN(node);
6838 
6839 	if (scctx->isc_ntxqsets > 100)
6840 		qfmt = "txq%03d";
6841 	else if (scctx->isc_ntxqsets > 10)
6842 		qfmt = "txq%02d";
6843 	else
6844 		qfmt = "txq%d";
6845 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6846 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6847 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6848 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6849 		queue_list = SYSCTL_CHILDREN(queue_node);
6850 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6851 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
6852 		    "cpu this queue is bound to");
6853 #if MEMORY_LOGGING
6854 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6855 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
6856 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6857 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
6858 #endif
6859 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6860 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
6861 		    "# of times m_defrag was called");
6862 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6863 		    CTLFLAG_RD, &txq->ift_pullups,
6864 		    "# of times m_pullup was called");
6865 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6866 		    "mbuf_defrag_failed", CTLFLAG_RD,
6867 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6868 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6869 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
6870 		    "# of times no descriptors were available");
6871 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6872 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
6873 		    "# of times DMA map failed");
6874 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6875 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
6876 		    "# of times txd_encap returned EFBIG");
6877 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6878 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
6879 		    "# of times map failed for other than EFBIG");
6880 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6881 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
6882 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6883 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
6884 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
6885 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
6886 		    1, "Consumer Index seen by credit update");
6887 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6888 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
6889 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6890 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
6891 		    "descriptors procesed for clean");
6892 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6893 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
6894 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6895 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6896 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6897 		    mp_ring_state_handler, "A", "soft ring state");
6898 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6899 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
6900 		    "# of enqueues to the mp_ring for this queue");
6901 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6902 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
6903 		    "# of drops in the mp_ring for this queue");
6904 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6905 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
6906 		    "# of normal consumer starts in mp_ring for this queue");
6907 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6908 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
6909 		    "# of consumer stalls in the mp_ring for this queue");
6910 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6911 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
6912 		    "# of consumer restarts in the mp_ring for this queue");
6913 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6914 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
6915 		    "# of consumer abdications in the mp_ring for this queue");
6916 	}
6917 
6918 	if (scctx->isc_nrxqsets > 100)
6919 		qfmt = "rxq%03d";
6920 	else if (scctx->isc_nrxqsets > 10)
6921 		qfmt = "rxq%02d";
6922 	else
6923 		qfmt = "rxq%d";
6924 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6925 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6926 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6927 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6928 		queue_list = SYSCTL_CHILDREN(queue_node);
6929 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6930 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
6931 		    "cpu this queue is bound to");
6932 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6933 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
6934 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
6935 			    "Consumer Index");
6936 		}
6937 
6938 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6939 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6940 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
6941 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
6942 			    NULL, "freelist Name");
6943 			fl_list = SYSCTL_CHILDREN(fl_node);
6944 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6945 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
6946 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6947 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
6948 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6949 			    CTLFLAG_RD, &fl->ifl_credits, 1,
6950 			    "credits available");
6951 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6952 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
6953 #if MEMORY_LOGGING
6954 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6955 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
6956 			    "mbufs allocated");
6957 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6958 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
6959 			    "mbufs freed");
6960 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6961 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
6962 			    "clusters allocated");
6963 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6964 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
6965 			    "clusters freed");
6966 #endif
6967 		}
6968 	}
6969 
6970 }
6971 
6972 void
6973 iflib_request_reset(if_ctx_t ctx)
6974 {
6975 
6976 	STATE_LOCK(ctx);
6977 	ctx->ifc_flags |= IFC_DO_RESET;
6978 	STATE_UNLOCK(ctx);
6979 }
6980 
6981 #ifndef __NO_STRICT_ALIGNMENT
6982 static struct mbuf *
6983 iflib_fixup_rx(struct mbuf *m)
6984 {
6985 	struct mbuf *n;
6986 
6987 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6988 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6989 		m->m_data += ETHER_HDR_LEN;
6990 		n = m;
6991 	} else {
6992 		MGETHDR(n, M_NOWAIT, MT_DATA);
6993 		if (n == NULL) {
6994 			m_freem(m);
6995 			return (NULL);
6996 		}
6997 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6998 		m->m_data += ETHER_HDR_LEN;
6999 		m->m_len -= ETHER_HDR_LEN;
7000 		n->m_len = ETHER_HDR_LEN;
7001 		M_MOVE_PKTHDR(n, m);
7002 		n->m_next = m;
7003 	}
7004 	return (n);
7005 }
7006 #endif
7007 
7008 #ifdef DEBUGNET
7009 static void
7010 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7011 {
7012 	if_ctx_t ctx;
7013 
7014 	ctx = if_getsoftc(ifp);
7015 	CTX_LOCK(ctx);
7016 	*nrxr = NRXQSETS(ctx);
7017 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7018 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7019 	CTX_UNLOCK(ctx);
7020 }
7021 
7022 static void
7023 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7024 {
7025 	if_ctx_t ctx;
7026 	if_softc_ctx_t scctx;
7027 	iflib_fl_t fl;
7028 	iflib_rxq_t rxq;
7029 	int i, j;
7030 
7031 	ctx = if_getsoftc(ifp);
7032 	scctx = &ctx->ifc_softc_ctx;
7033 
7034 	switch (event) {
7035 	case DEBUGNET_START:
7036 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7037 			rxq = &ctx->ifc_rxqs[i];
7038 			for (j = 0; j < rxq->ifr_nfl; j++) {
7039 				fl = rxq->ifr_fl;
7040 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7041 			}
7042 		}
7043 		iflib_no_tx_batch = 1;
7044 		break;
7045 	default:
7046 		break;
7047 	}
7048 }
7049 
7050 static int
7051 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7052 {
7053 	if_ctx_t ctx;
7054 	iflib_txq_t txq;
7055 	int error;
7056 
7057 	ctx = if_getsoftc(ifp);
7058 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7059 	    IFF_DRV_RUNNING)
7060 		return (EBUSY);
7061 
7062 	txq = &ctx->ifc_txqs[0];
7063 	error = iflib_encap(txq, &m);
7064 	if (error == 0)
7065 		(void)iflib_txd_db_check(txq, true);
7066 	return (error);
7067 }
7068 
7069 static int
7070 iflib_debugnet_poll(if_t ifp, int count)
7071 {
7072 	struct epoch_tracker et;
7073 	if_ctx_t ctx;
7074 	if_softc_ctx_t scctx;
7075 	iflib_txq_t txq;
7076 	int i;
7077 
7078 	ctx = if_getsoftc(ifp);
7079 	scctx = &ctx->ifc_softc_ctx;
7080 
7081 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7082 	    IFF_DRV_RUNNING)
7083 		return (EBUSY);
7084 
7085 	txq = &ctx->ifc_txqs[0];
7086 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7087 
7088 	NET_EPOCH_ENTER(et);
7089 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7090 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7091 	NET_EPOCH_EXIT(et);
7092 	return (0);
7093 }
7094 #endif /* DEBUGNET */
7095