xref: /freebsd/sys/net/iflib.c (revision 1ae4848cd0177755e88302ccb9afdd3ad69e68c1)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/sockio.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/kobj.h>
46 #include <sys/rman.h>
47 #include <sys/sbuf.h>
48 #include <sys/smp.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #if defined(__i386__) || defined(__amd64__)
92 #include <sys/memdesc.h>
93 #include <machine/bus.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
96 #include <x86/include/busdma_impl.h>
97 #include <x86/iommu/busdma_dmar.h>
98 #endif
99 
100 #include <sys/bitstring.h>
101 /*
102  * enable accounting of every mbuf as it comes in to and goes out of
103  * iflib's software descriptor references
104  */
105 #define MEMORY_LOGGING 0
106 /*
107  * Enable mbuf vectors for compressing long mbuf chains
108  */
109 
110 /*
111  * NB:
112  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
113  *   we prefetch needs to be determined by the time spent in m_free vis a vis
114  *   the cost of a prefetch. This will of course vary based on the workload:
115  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
116  *        is quite expensive, thus suggesting very little prefetch.
117  *      - small packet forwarding which is just returning a single mbuf to
118  *        UMA will typically be very fast vis a vis the cost of a memory
119  *        access.
120  */
121 
122 
123 /*
124  * File organization:
125  *  - private structures
126  *  - iflib private utility functions
127  *  - ifnet functions
128  *  - vlan registry and other exported functions
129  *  - iflib public core functions
130  *
131  *
132  */
133 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155    /*
156    * Pointer to hardware driver's softc
157    */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	uint16_t ifc_nhwtxqs;
170 
171 	iflib_txq_t ifc_txqs;
172 	iflib_rxq_t ifc_rxqs;
173 	uint32_t ifc_if_flags;
174 	uint32_t ifc_flags;
175 	uint32_t ifc_max_fl_buf_size;
176 	int ifc_in_detach;
177 
178 	int ifc_link_state;
179 	int ifc_link_irq;
180 	int ifc_watchdog_events;
181 	struct cdev *ifc_led_dev;
182 	struct resource *ifc_msix_mem;
183 
184 	struct if_irq ifc_legacy_irq;
185 	struct grouptask ifc_admin_task;
186 	struct grouptask ifc_vflr_task;
187 	struct iflib_filter_info ifc_filter_info;
188 	struct ifmedia	ifc_media;
189 
190 	struct sysctl_oid *ifc_sysctl_node;
191 	uint16_t ifc_sysctl_ntxqs;
192 	uint16_t ifc_sysctl_nrxqs;
193 	uint16_t ifc_sysctl_qs_eq_override;
194 	uint16_t ifc_sysctl_rx_budget;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	uint8_t ifc_mac[ETHER_ADDR_LEN];
212 	char ifc_mtx_name[16];
213 };
214 
215 
216 void *
217 iflib_get_softc(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_softc);
221 }
222 
223 device_t
224 iflib_get_dev(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_dev);
228 }
229 
230 if_t
231 iflib_get_ifp(if_ctx_t ctx)
232 {
233 
234 	return (ctx->ifc_ifp);
235 }
236 
237 struct ifmedia *
238 iflib_get_media(if_ctx_t ctx)
239 {
240 
241 	return (&ctx->ifc_media);
242 }
243 
244 void
245 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
246 {
247 
248 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
249 }
250 
251 if_softc_ctx_t
252 iflib_get_softc_ctx(if_ctx_t ctx)
253 {
254 
255 	return (&ctx->ifc_softc_ctx);
256 }
257 
258 if_shared_ctx_t
259 iflib_get_sctx(if_ctx_t ctx)
260 {
261 
262 	return (ctx->ifc_sctx);
263 }
264 
265 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
266 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
267 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
268 
269 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
270 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
271 
272 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
273 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
274 #define RX_SW_DESC_INUSE        (1 << 3)
275 #define TX_SW_DESC_MAPPED       (1 << 4)
276 
277 #define	M_TOOBIG		M_PROTO1
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	uint8_t		*ifsd_flags;
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
289 	uint8_t		*ifsd_flags;
290 } if_txsd_vec_t;
291 
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 /* bnxt supports 64 with hardware LRO enabled */
296 #define IFLIB_MAX_RX_SEGS		64
297 #define IFLIB_RX_COPY_THRESH		128
298 #define IFLIB_MAX_RX_REFRESH		32
299 /* The minimum descriptors per second before we start coalescing */
300 #define IFLIB_MIN_DESC_SEC		16384
301 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
302 #define IFLIB_QUEUE_IDLE		0
303 #define IFLIB_QUEUE_HUNG		1
304 #define IFLIB_QUEUE_WORKING		2
305 /* maximum number of txqs that can share an rx interrupt */
306 #define IFLIB_MAX_TX_SHARED_INTR	4
307 
308 /* this should really scale with ring size - this is a fairly arbitrary value */
309 #define TX_BATCH_SIZE			32
310 
311 #define IFLIB_RESTART_BUDGET		8
312 
313 #define	IFC_LEGACY		0x001
314 #define	IFC_QFLUSH		0x002
315 #define	IFC_MULTISEG		0x004
316 #define	IFC_DMAR		0x008
317 #define	IFC_SC_ALLOCATED	0x010
318 #define	IFC_INIT_DONE		0x020
319 #define	IFC_PREFETCH		0x040
320 #define	IFC_DO_RESET		0x080
321 #define	IFC_DO_WATCHDOG		0x100
322 #define	IFC_CHECK_HUNG		0x200
323 
324 
325 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
326 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
327 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
328 struct iflib_txq {
329 	qidx_t		ift_in_use;
330 	qidx_t		ift_cidx;
331 	qidx_t		ift_cidx_processed;
332 	qidx_t		ift_pidx;
333 	uint8_t		ift_gen;
334 	uint8_t		ift_br_offset;
335 	uint16_t	ift_npending;
336 	uint16_t	ift_db_pending;
337 	uint16_t	ift_rs_pending;
338 	/* implicit pad */
339 	uint8_t		ift_txd_size[8];
340 	uint64_t	ift_processed;
341 	uint64_t	ift_cleaned;
342 	uint64_t	ift_cleaned_prev;
343 #if MEMORY_LOGGING
344 	uint64_t	ift_enqueued;
345 	uint64_t	ift_dequeued;
346 #endif
347 	uint64_t	ift_no_tx_dma_setup;
348 	uint64_t	ift_no_desc_avail;
349 	uint64_t	ift_mbuf_defrag_failed;
350 	uint64_t	ift_mbuf_defrag;
351 	uint64_t	ift_map_failed;
352 	uint64_t	ift_txd_encap_efbig;
353 	uint64_t	ift_pullups;
354 
355 	struct mtx	ift_mtx;
356 	struct mtx	ift_db_mtx;
357 
358 	/* constant values */
359 	if_ctx_t	ift_ctx;
360 	struct ifmp_ring        *ift_br;
361 	struct grouptask	ift_task;
362 	qidx_t		ift_size;
363 	uint16_t	ift_id;
364 	struct callout	ift_timer;
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_desc_tag;
372 	bus_dma_tag_t		ift_tso_desc_tag;
373 	iflib_dma_info_t	ift_ifdi;
374 #define MTX_NAME_LEN 16
375 	char                    ift_mtx_name[MTX_NAME_LEN];
376 	char                    ift_db_mtx_name[MTX_NAME_LEN];
377 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
378 #ifdef IFLIB_DIAGNOSTICS
379 	uint64_t ift_cpu_exec_count[256];
380 #endif
381 } __aligned(CACHE_LINE_SIZE);
382 
383 struct iflib_fl {
384 	qidx_t		ifl_cidx;
385 	qidx_t		ifl_pidx;
386 	qidx_t		ifl_credits;
387 	uint8_t		ifl_gen;
388 	uint8_t		ifl_rxd_size;
389 #if MEMORY_LOGGING
390 	uint64_t	ifl_m_enqueued;
391 	uint64_t	ifl_m_dequeued;
392 	uint64_t	ifl_cl_enqueued;
393 	uint64_t	ifl_cl_dequeued;
394 #endif
395 	/* implicit pad */
396 
397 	bitstr_t 	*ifl_rx_bitmap;
398 	qidx_t		ifl_fragidx;
399 	/* constant */
400 	qidx_t		ifl_size;
401 	uint16_t	ifl_buf_size;
402 	uint16_t	ifl_cltype;
403 	uma_zone_t	ifl_zone;
404 	iflib_rxsd_array_t	ifl_sds;
405 	iflib_rxq_t	ifl_rxq;
406 	uint8_t		ifl_id;
407 	bus_dma_tag_t           ifl_desc_tag;
408 	iflib_dma_info_t	ifl_ifdi;
409 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
410 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
411 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
412 }  __aligned(CACHE_LINE_SIZE);
413 
414 static inline qidx_t
415 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
416 {
417 	qidx_t used;
418 
419 	if (pidx > cidx)
420 		used = pidx - cidx;
421 	else if (pidx < cidx)
422 		used = size - cidx + pidx;
423 	else if (gen == 0 && pidx == cidx)
424 		used = 0;
425 	else if (gen == 1 && pidx == cidx)
426 		used = size;
427 	else
428 		panic("bad state");
429 
430 	return (used);
431 }
432 
433 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
434 
435 #define IDXDIFF(head, tail, wrap) \
436 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
437 
438 struct iflib_rxq {
439 	/* If there is a separate completion queue -
440 	 * these are the cq cidx and pidx. Otherwise
441 	 * these are unused.
442 	 */
443 	qidx_t		ifr_size;
444 	qidx_t		ifr_cq_cidx;
445 	qidx_t		ifr_cq_pidx;
446 	uint8_t		ifr_cq_gen;
447 	uint8_t		ifr_fl_offset;
448 
449 	if_ctx_t	ifr_ctx;
450 	iflib_fl_t	ifr_fl;
451 	uint64_t	ifr_rx_irq;
452 	uint16_t	ifr_id;
453 	uint8_t		ifr_lro_enabled;
454 	uint8_t		ifr_nfl;
455 	uint8_t		ifr_ntxqirq;
456 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
457 	struct lro_ctrl			ifr_lc;
458 	struct grouptask        ifr_task;
459 	struct iflib_filter_info ifr_filter_info;
460 	iflib_dma_info_t		ifr_ifdi;
461 
462 	/* dynamically allocate if any drivers need a value substantially larger than this */
463 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
464 #ifdef IFLIB_DIAGNOSTICS
465 	uint64_t ifr_cpu_exec_count[256];
466 #endif
467 }  __aligned(CACHE_LINE_SIZE);
468 
469 typedef struct if_rxsd {
470 	caddr_t *ifsd_cl;
471 	struct mbuf **ifsd_m;
472 	iflib_fl_t ifsd_fl;
473 	qidx_t ifsd_cidx;
474 } *if_rxsd_t;
475 
476 /* multiple of word size */
477 #ifdef __LP64__
478 #define PKT_INFO_SIZE	6
479 #define RXD_INFO_SIZE	5
480 #define PKT_TYPE uint64_t
481 #else
482 #define PKT_INFO_SIZE	11
483 #define RXD_INFO_SIZE	8
484 #define PKT_TYPE uint32_t
485 #endif
486 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
487 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
488 
489 typedef struct if_pkt_info_pad {
490 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
491 } *if_pkt_info_pad_t;
492 typedef struct if_rxd_info_pad {
493 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
494 } *if_rxd_info_pad_t;
495 
496 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
497 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
498 
499 
500 static inline void
501 pkt_info_zero(if_pkt_info_t pi)
502 {
503 	if_pkt_info_pad_t pi_pad;
504 
505 	pi_pad = (if_pkt_info_pad_t)pi;
506 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
507 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
508 #ifndef __LP64__
509 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
510 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
511 #endif
512 }
513 
514 static inline void
515 rxd_info_zero(if_rxd_info_t ri)
516 {
517 	if_rxd_info_pad_t ri_pad;
518 	int i;
519 
520 	ri_pad = (if_rxd_info_pad_t)ri;
521 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
522 		ri_pad->rxd_val[i] = 0;
523 		ri_pad->rxd_val[i+1] = 0;
524 		ri_pad->rxd_val[i+2] = 0;
525 		ri_pad->rxd_val[i+3] = 0;
526 	}
527 #ifdef __LP64__
528 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
529 #endif
530 }
531 
532 /*
533  * Only allow a single packet to take up most 1/nth of the tx ring
534  */
535 #define MAX_SINGLE_PACKET_FRACTION 12
536 #define IF_BAD_DMA (bus_addr_t)-1
537 
538 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
539 
540 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
541 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
542 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
543 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
544 
545 
546 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
547 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
548 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
549 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
550 
551 
552 
553 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
554 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
555 
556 
557 /* Our boot-time initialization hook */
558 static int	iflib_module_event_handler(module_t, int, void *);
559 
560 static moduledata_t iflib_moduledata = {
561 	"iflib",
562 	iflib_module_event_handler,
563 	NULL
564 };
565 
566 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
567 MODULE_VERSION(iflib, 1);
568 
569 MODULE_DEPEND(iflib, pci, 1, 1, 1);
570 MODULE_DEPEND(iflib, ether, 1, 1, 1);
571 
572 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
573 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
574 
575 #ifndef IFLIB_DEBUG_COUNTERS
576 #ifdef INVARIANTS
577 #define IFLIB_DEBUG_COUNTERS 1
578 #else
579 #define IFLIB_DEBUG_COUNTERS 0
580 #endif /* !INVARIANTS */
581 #endif
582 
583 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
584                    "iflib driver parameters");
585 
586 /*
587  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
588  */
589 static int iflib_min_tx_latency = 0;
590 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
591 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
592 static int iflib_no_tx_batch = 0;
593 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
594 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
595 
596 
597 #if IFLIB_DEBUG_COUNTERS
598 
599 static int iflib_tx_seen;
600 static int iflib_tx_sent;
601 static int iflib_tx_encap;
602 static int iflib_rx_allocs;
603 static int iflib_fl_refills;
604 static int iflib_fl_refills_large;
605 static int iflib_tx_frees;
606 
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
608 		   &iflib_tx_seen, 0, "# tx mbufs seen");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
610 		   &iflib_tx_sent, 0, "# tx mbufs sent");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
612 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
614 		   &iflib_tx_frees, 0, "# tx frees");
615 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
616 		   &iflib_rx_allocs, 0, "# rx allocations");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
618 		   &iflib_fl_refills, 0, "# refills");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
620 		   &iflib_fl_refills_large, 0, "# large refills");
621 
622 
623 static int iflib_txq_drain_flushing;
624 static int iflib_txq_drain_oactive;
625 static int iflib_txq_drain_notready;
626 static int iflib_txq_drain_encapfail;
627 
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
629 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
631 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
633 		   &iflib_txq_drain_notready, 0, "# drain notready");
634 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
635 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
636 
637 
638 static int iflib_encap_load_mbuf_fail;
639 static int iflib_encap_pad_mbuf_fail;
640 static int iflib_encap_txq_avail_fail;
641 static int iflib_encap_txd_encap_fail;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
644 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
646 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
647 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
648 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
649 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
650 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
651 
652 static int iflib_task_fn_rxs;
653 static int iflib_rx_intr_enables;
654 static int iflib_fast_intrs;
655 static int iflib_intr_link;
656 static int iflib_intr_msix;
657 static int iflib_rx_unavail;
658 static int iflib_rx_ctx_inactive;
659 static int iflib_rx_zero_len;
660 static int iflib_rx_if_input;
661 static int iflib_rx_mbuf_null;
662 static int iflib_rxd_flush;
663 
664 static int iflib_verbose_debug;
665 
666 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
667 		   &iflib_intr_link, 0, "# intr link calls");
668 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
669 		   &iflib_intr_msix, 0, "# intr msix calls");
670 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
671 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
673 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
674 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
675 		   &iflib_fast_intrs, 0, "# fast_intr calls");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
677 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
678 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
679 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
680 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
681 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
682 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
683 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
684 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
685 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
686 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
687 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
688 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
689 		   &iflib_verbose_debug, 0, "enable verbose debugging");
690 
691 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
692 static void
693 iflib_debug_reset(void)
694 {
695 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
696 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
697 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
698 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
699 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
700 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
701 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
702 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
703 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
704 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
705 }
706 
707 #else
708 #define DBG_COUNTER_INC(name)
709 static void iflib_debug_reset(void) {}
710 #endif
711 
712 
713 
714 #define IFLIB_DEBUG 0
715 
716 static void iflib_tx_structures_free(if_ctx_t ctx);
717 static void iflib_rx_structures_free(if_ctx_t ctx);
718 static int iflib_queues_alloc(if_ctx_t ctx);
719 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
720 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
721 static int iflib_qset_structures_setup(if_ctx_t ctx);
722 static int iflib_msix_init(if_ctx_t ctx);
723 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
724 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
725 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
726 static int iflib_register(if_ctx_t);
727 static void iflib_init_locked(if_ctx_t ctx);
728 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
729 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
730 static void iflib_ifmp_purge(iflib_txq_t txq);
731 static void _iflib_pre_assert(if_softc_ctx_t scctx);
732 static void iflib_stop(if_ctx_t ctx);
733 static void iflib_if_init_locked(if_ctx_t ctx);
734 #ifndef __NO_STRICT_ALIGNMENT
735 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
736 #endif
737 
738 #ifdef DEV_NETMAP
739 #include <sys/selinfo.h>
740 #include <net/netmap.h>
741 #include <dev/netmap/netmap_kern.h>
742 
743 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
744 
745 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
746 
747 /*
748  * device-specific sysctl variables:
749  *
750  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
751  *	During regular operations the CRC is stripped, but on some
752  *	hardware reception of frames not multiple of 64 is slower,
753  *	so using crcstrip=0 helps in benchmarks.
754  *
755  * iflib_rx_miss, iflib_rx_miss_bufs:
756  *	count packets that might be missed due to lost interrupts.
757  */
758 SYSCTL_DECL(_dev_netmap);
759 /*
760  * The xl driver by default strips CRCs and we do not override it.
761  */
762 
763 int iflib_crcstrip = 1;
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
765     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
766 
767 int iflib_rx_miss, iflib_rx_miss_bufs;
768 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
769     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
770 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
771     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
772 
773 /*
774  * Register/unregister. We are already under netmap lock.
775  * Only called on the first register or the last unregister.
776  */
777 static int
778 iflib_netmap_register(struct netmap_adapter *na, int onoff)
779 {
780 	struct ifnet *ifp = na->ifp;
781 	if_ctx_t ctx = ifp->if_softc;
782 	int status;
783 
784 	CTX_LOCK(ctx);
785 	IFDI_INTR_DISABLE(ctx);
786 
787 	/* Tell the stack that the interface is no longer active */
788 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
789 
790 	if (!CTX_IS_VF(ctx))
791 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
792 
793 	/* enable or disable flags and callbacks in na and ifp */
794 	if (onoff) {
795 		nm_set_native_flags(na);
796 	} else {
797 		nm_clear_native_flags(na);
798 	}
799 	iflib_stop(ctx);
800 	iflib_init_locked(ctx);
801 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
802 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
803 	if (status)
804 		nm_clear_native_flags(na);
805 	CTX_UNLOCK(ctx);
806 	return (status);
807 }
808 
809 static int
810 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
811 {
812 	struct netmap_adapter *na = kring->na;
813 	u_int const lim = kring->nkr_num_slots - 1;
814 	u_int head = kring->rhead;
815 	struct netmap_ring *ring = kring->ring;
816 	bus_dmamap_t *map;
817 	struct if_rxd_update iru;
818 	if_ctx_t ctx = rxq->ifr_ctx;
819 	iflib_fl_t fl = &rxq->ifr_fl[0];
820 	uint32_t refill_pidx, nic_i;
821 
822 	if (nm_i == head && __predict_true(!init))
823 		return 0;
824 	iru_init(&iru, rxq, 0 /* flid */);
825 	map = fl->ifl_sds.ifsd_map;
826 	refill_pidx = netmap_idx_k2n(kring, nm_i);
827 	/*
828 	 * IMPORTANT: we must leave one free slot in the ring,
829 	 * so move head back by one unit
830 	 */
831 	head = nm_prev(head, lim);
832 	nic_i = UINT_MAX;
833 	while (nm_i != head) {
834 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
835 			struct netmap_slot *slot = &ring->slot[nm_i];
836 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
837 			uint32_t nic_i_dma = refill_pidx;
838 			nic_i = netmap_idx_k2n(kring, nm_i);
839 
840 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
841 
842 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
843 			        return netmap_ring_reinit(kring);
844 
845 			fl->ifl_vm_addrs[tmp_pidx] = addr;
846 			if (__predict_false(init) && map) {
847 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
848 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
849 				/* buffer has changed, reload map */
850 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
851 			}
852 			slot->flags &= ~NS_BUF_CHANGED;
853 
854 			nm_i = nm_next(nm_i, lim);
855 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
856 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
857 				continue;
858 
859 			iru.iru_pidx = refill_pidx;
860 			iru.iru_count = tmp_pidx+1;
861 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
862 
863 			refill_pidx = nic_i;
864 			if (map == NULL)
865 				continue;
866 
867 			for (int n = 0; n < iru.iru_count; n++) {
868 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
869 						BUS_DMASYNC_PREREAD);
870 				/* XXX - change this to not use the netmap func*/
871 				nic_i_dma = nm_next(nic_i_dma, lim);
872 			}
873 		}
874 	}
875 	kring->nr_hwcur = head;
876 
877 	if (map)
878 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
879 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
880 	if (__predict_true(nic_i != UINT_MAX))
881 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
882 	return (0);
883 }
884 
885 /*
886  * Reconcile kernel and user view of the transmit ring.
887  *
888  * All information is in the kring.
889  * Userspace wants to send packets up to the one before kring->rhead,
890  * kernel knows kring->nr_hwcur is the first unsent packet.
891  *
892  * Here we push packets out (as many as possible), and possibly
893  * reclaim buffers from previously completed transmission.
894  *
895  * The caller (netmap) guarantees that there is only one instance
896  * running at any time. Any interference with other driver
897  * methods should be handled by the individual drivers.
898  */
899 static int
900 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
901 {
902 	struct netmap_adapter *na = kring->na;
903 	struct ifnet *ifp = na->ifp;
904 	struct netmap_ring *ring = kring->ring;
905 	u_int nm_i;	/* index into the netmap ring */
906 	u_int nic_i;	/* index into the NIC ring */
907 	u_int n;
908 	u_int const lim = kring->nkr_num_slots - 1;
909 	u_int const head = kring->rhead;
910 	struct if_pkt_info pi;
911 
912 	/*
913 	 * interrupts on every tx packet are expensive so request
914 	 * them every half ring, or where NS_REPORT is set
915 	 */
916 	u_int report_frequency = kring->nkr_num_slots >> 1;
917 	/* device-specific */
918 	if_ctx_t ctx = ifp->if_softc;
919 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
920 
921 	if (txq->ift_sds.ifsd_map)
922 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
923 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
924 
925 
926 	/*
927 	 * First part: process new packets to send.
928 	 * nm_i is the current index in the netmap ring,
929 	 * nic_i is the corresponding index in the NIC ring.
930 	 *
931 	 * If we have packets to send (nm_i != head)
932 	 * iterate over the netmap ring, fetch length and update
933 	 * the corresponding slot in the NIC ring. Some drivers also
934 	 * need to update the buffer's physical address in the NIC slot
935 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
936 	 *
937 	 * The netmap_reload_map() calls is especially expensive,
938 	 * even when (as in this case) the tag is 0, so do only
939 	 * when the buffer has actually changed.
940 	 *
941 	 * If possible do not set the report/intr bit on all slots,
942 	 * but only a few times per ring or when NS_REPORT is set.
943 	 *
944 	 * Finally, on 10G and faster drivers, it might be useful
945 	 * to prefetch the next slot and txr entry.
946 	 */
947 
948 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
949 	pkt_info_zero(&pi);
950 	pi.ipi_segs = txq->ift_segs;
951 	pi.ipi_qsidx = kring->ring_id;
952 	if (nm_i != head) {	/* we have new packets to send */
953 		nic_i = netmap_idx_k2n(kring, nm_i);
954 
955 		__builtin_prefetch(&ring->slot[nm_i]);
956 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
957 		if (txq->ift_sds.ifsd_map)
958 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
959 
960 		for (n = 0; nm_i != head; n++) {
961 			struct netmap_slot *slot = &ring->slot[nm_i];
962 			u_int len = slot->len;
963 			uint64_t paddr;
964 			void *addr = PNMB(na, slot, &paddr);
965 			int flags = (slot->flags & NS_REPORT ||
966 				nic_i == 0 || nic_i == report_frequency) ?
967 				IPI_TX_INTR : 0;
968 
969 			/* device-specific */
970 			pi.ipi_len = len;
971 			pi.ipi_segs[0].ds_addr = paddr;
972 			pi.ipi_segs[0].ds_len = len;
973 			pi.ipi_nsegs = 1;
974 			pi.ipi_ndescs = 0;
975 			pi.ipi_pidx = nic_i;
976 			pi.ipi_flags = flags;
977 
978 			/* Fill the slot in the NIC ring. */
979 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
980 
981 			/* prefetch for next round */
982 			__builtin_prefetch(&ring->slot[nm_i + 1]);
983 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
984 			if (txq->ift_sds.ifsd_map) {
985 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
986 
987 				NM_CHECK_ADDR_LEN(na, addr, len);
988 
989 				if (slot->flags & NS_BUF_CHANGED) {
990 					/* buffer has changed, reload map */
991 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
992 				}
993 				/* make sure changes to the buffer are synced */
994 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
995 						BUS_DMASYNC_PREWRITE);
996 			}
997 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
998 			nm_i = nm_next(nm_i, lim);
999 			nic_i = nm_next(nic_i, lim);
1000 		}
1001 		kring->nr_hwcur = head;
1002 
1003 		/* synchronize the NIC ring */
1004 		if (txq->ift_sds.ifsd_map)
1005 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1006 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1007 
1008 		/* (re)start the tx unit up to slot nic_i (excluded) */
1009 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1010 	}
1011 
1012 	/*
1013 	 * Second part: reclaim buffers for completed transmissions.
1014 	 */
1015 	if (iflib_tx_credits_update(ctx, txq)) {
1016 		/* some tx completed, increment avail */
1017 		nic_i = txq->ift_cidx_processed;
1018 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1019 	}
1020 	return (0);
1021 }
1022 
1023 /*
1024  * Reconcile kernel and user view of the receive ring.
1025  * Same as for the txsync, this routine must be efficient.
1026  * The caller guarantees a single invocations, but races against
1027  * the rest of the driver should be handled here.
1028  *
1029  * On call, kring->rhead is the first packet that userspace wants
1030  * to keep, and kring->rcur is the wakeup point.
1031  * The kernel has previously reported packets up to kring->rtail.
1032  *
1033  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1034  * of whether or not we received an interrupt.
1035  */
1036 static int
1037 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1038 {
1039 	struct netmap_adapter *na = kring->na;
1040 	struct netmap_ring *ring = kring->ring;
1041 	uint32_t nm_i;	/* index into the netmap ring */
1042 	uint32_t nic_i;	/* index into the NIC ring */
1043 	u_int i, n;
1044 	u_int const lim = kring->nkr_num_slots - 1;
1045 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1046 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1047 	struct if_rxd_info ri;
1048 
1049 	struct ifnet *ifp = na->ifp;
1050 	if_ctx_t ctx = ifp->if_softc;
1051 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1052 	iflib_fl_t fl = rxq->ifr_fl;
1053 	if (head > lim)
1054 		return netmap_ring_reinit(kring);
1055 
1056 	/* XXX check sync modes */
1057 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1058 		if (fl->ifl_sds.ifsd_map == NULL)
1059 			continue;
1060 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1061 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1062 	}
1063 	/*
1064 	 * First part: import newly received packets.
1065 	 *
1066 	 * nm_i is the index of the next free slot in the netmap ring,
1067 	 * nic_i is the index of the next received packet in the NIC ring,
1068 	 * and they may differ in case if_init() has been called while
1069 	 * in netmap mode. For the receive ring we have
1070 	 *
1071 	 *	nic_i = rxr->next_check;
1072 	 *	nm_i = kring->nr_hwtail (previous)
1073 	 * and
1074 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1075 	 *
1076 	 * rxr->next_check is set to 0 on a ring reinit
1077 	 */
1078 	if (netmap_no_pendintr || force_update) {
1079 		int crclen = iflib_crcstrip ? 0 : 4;
1080 		int error, avail;
1081 
1082 		for (i = 0; i < rxq->ifr_nfl; i++) {
1083 			fl = &rxq->ifr_fl[i];
1084 			nic_i = fl->ifl_cidx;
1085 			nm_i = netmap_idx_n2k(kring, nic_i);
1086 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1087 			for (n = 0; avail > 0; n++, avail--) {
1088 				rxd_info_zero(&ri);
1089 				ri.iri_frags = rxq->ifr_frags;
1090 				ri.iri_qsidx = kring->ring_id;
1091 				ri.iri_ifp = ctx->ifc_ifp;
1092 				ri.iri_cidx = nic_i;
1093 
1094 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1095 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1096 				ring->slot[nm_i].flags = 0;
1097 				if (fl->ifl_sds.ifsd_map)
1098 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1099 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1100 				nm_i = nm_next(nm_i, lim);
1101 				nic_i = nm_next(nic_i, lim);
1102 			}
1103 			if (n) { /* update the state variables */
1104 				if (netmap_no_pendintr && !force_update) {
1105 					/* diagnostics */
1106 					iflib_rx_miss ++;
1107 					iflib_rx_miss_bufs += n;
1108 				}
1109 				fl->ifl_cidx = nic_i;
1110 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1111 			}
1112 			kring->nr_kflags &= ~NKR_PENDINTR;
1113 		}
1114 	}
1115 	/*
1116 	 * Second part: skip past packets that userspace has released.
1117 	 * (kring->nr_hwcur to head excluded),
1118 	 * and make the buffers available for reception.
1119 	 * As usual nm_i is the index in the netmap ring,
1120 	 * nic_i is the index in the NIC ring, and
1121 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1122 	 */
1123 	/* XXX not sure how this will work with multiple free lists */
1124 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1125 
1126 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1127 }
1128 
1129 static void
1130 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1131 {
1132 	struct ifnet *ifp = na->ifp;
1133 	if_ctx_t ctx = ifp->if_softc;
1134 
1135 	CTX_LOCK(ctx);
1136 	if (onoff) {
1137 		IFDI_INTR_ENABLE(ctx);
1138 	} else {
1139 		IFDI_INTR_DISABLE(ctx);
1140 	}
1141 	CTX_UNLOCK(ctx);
1142 }
1143 
1144 
1145 static int
1146 iflib_netmap_attach(if_ctx_t ctx)
1147 {
1148 	struct netmap_adapter na;
1149 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1150 
1151 	bzero(&na, sizeof(na));
1152 
1153 	na.ifp = ctx->ifc_ifp;
1154 	na.na_flags = NAF_BDG_MAYSLEEP;
1155 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1156 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1157 
1158 	na.num_tx_desc = scctx->isc_ntxd[0];
1159 	na.num_rx_desc = scctx->isc_nrxd[0];
1160 	na.nm_txsync = iflib_netmap_txsync;
1161 	na.nm_rxsync = iflib_netmap_rxsync;
1162 	na.nm_register = iflib_netmap_register;
1163 	na.nm_intr = iflib_netmap_intr;
1164 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1165 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1166 	return (netmap_attach(&na));
1167 }
1168 
1169 static void
1170 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1171 {
1172 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1173 	struct netmap_slot *slot;
1174 
1175 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1176 	if (slot == NULL)
1177 		return;
1178 	if (txq->ift_sds.ifsd_map == NULL)
1179 		return;
1180 
1181 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1182 
1183 		/*
1184 		 * In netmap mode, set the map for the packet buffer.
1185 		 * NOTE: Some drivers (not this one) also need to set
1186 		 * the physical buffer address in the NIC ring.
1187 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1188 		 * netmap slot index, si
1189 		 */
1190 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1191 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1192 	}
1193 }
1194 
1195 static void
1196 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1197 {
1198 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1199 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1200 	struct netmap_slot *slot;
1201 	uint32_t nm_i;
1202 
1203 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1204 	if (slot == NULL)
1205 		return;
1206 	nm_i = netmap_idx_n2k(kring, 0);
1207 	netmap_fl_refill(rxq, kring, nm_i, true);
1208 }
1209 
1210 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1211 
1212 #else
1213 #define iflib_netmap_txq_init(ctx, txq)
1214 #define iflib_netmap_rxq_init(ctx, rxq)
1215 #define iflib_netmap_detach(ifp)
1216 
1217 #define iflib_netmap_attach(ctx) (0)
1218 #define netmap_rx_irq(ifp, qid, budget) (0)
1219 #define netmap_tx_irq(ifp, qid) do {} while (0)
1220 
1221 #endif
1222 
1223 #if defined(__i386__) || defined(__amd64__)
1224 static __inline void
1225 prefetch(void *x)
1226 {
1227 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1228 }
1229 static __inline void
1230 prefetch2cachelines(void *x)
1231 {
1232 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1233 #if (CACHE_LINE_SIZE < 128)
1234 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1235 #endif
1236 }
1237 #else
1238 #define prefetch(x)
1239 #define prefetch2cachelines(x)
1240 #endif
1241 
1242 static void
1243 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1244 {
1245 	iflib_fl_t fl;
1246 
1247 	fl = &rxq->ifr_fl[flid];
1248 	iru->iru_paddrs = fl->ifl_bus_addrs;
1249 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1250 	iru->iru_idxs = fl->ifl_rxd_idxs;
1251 	iru->iru_qsidx = rxq->ifr_id;
1252 	iru->iru_buf_size = fl->ifl_buf_size;
1253 	iru->iru_flidx = fl->ifl_id;
1254 }
1255 
1256 static void
1257 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1258 {
1259 	if (err)
1260 		return;
1261 	*(bus_addr_t *) arg = segs[0].ds_addr;
1262 }
1263 
1264 int
1265 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1266 {
1267 	int err;
1268 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1269 	device_t dev = ctx->ifc_dev;
1270 
1271 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1272 
1273 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1274 				sctx->isc_q_align, 0,	/* alignment, bounds */
1275 				BUS_SPACE_MAXADDR,	/* lowaddr */
1276 				BUS_SPACE_MAXADDR,	/* highaddr */
1277 				NULL, NULL,		/* filter, filterarg */
1278 				size,			/* maxsize */
1279 				1,			/* nsegments */
1280 				size,			/* maxsegsize */
1281 				BUS_DMA_ALLOCNOW,	/* flags */
1282 				NULL,			/* lockfunc */
1283 				NULL,			/* lockarg */
1284 				&dma->idi_tag);
1285 	if (err) {
1286 		device_printf(dev,
1287 		    "%s: bus_dma_tag_create failed: %d\n",
1288 		    __func__, err);
1289 		goto fail_0;
1290 	}
1291 
1292 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1293 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1294 	if (err) {
1295 		device_printf(dev,
1296 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1297 		    __func__, (uintmax_t)size, err);
1298 		goto fail_1;
1299 	}
1300 
1301 	dma->idi_paddr = IF_BAD_DMA;
1302 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1303 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1304 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1305 		device_printf(dev,
1306 		    "%s: bus_dmamap_load failed: %d\n",
1307 		    __func__, err);
1308 		goto fail_2;
1309 	}
1310 
1311 	dma->idi_size = size;
1312 	return (0);
1313 
1314 fail_2:
1315 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1316 fail_1:
1317 	bus_dma_tag_destroy(dma->idi_tag);
1318 fail_0:
1319 	dma->idi_tag = NULL;
1320 
1321 	return (err);
1322 }
1323 
1324 int
1325 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1326 {
1327 	int i, err;
1328 	iflib_dma_info_t *dmaiter;
1329 
1330 	dmaiter = dmalist;
1331 	for (i = 0; i < count; i++, dmaiter++) {
1332 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1333 			break;
1334 	}
1335 	if (err)
1336 		iflib_dma_free_multi(dmalist, i);
1337 	return (err);
1338 }
1339 
1340 void
1341 iflib_dma_free(iflib_dma_info_t dma)
1342 {
1343 	if (dma->idi_tag == NULL)
1344 		return;
1345 	if (dma->idi_paddr != IF_BAD_DMA) {
1346 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1347 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1348 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1349 		dma->idi_paddr = IF_BAD_DMA;
1350 	}
1351 	if (dma->idi_vaddr != NULL) {
1352 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1353 		dma->idi_vaddr = NULL;
1354 	}
1355 	bus_dma_tag_destroy(dma->idi_tag);
1356 	dma->idi_tag = NULL;
1357 }
1358 
1359 void
1360 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1361 {
1362 	int i;
1363 	iflib_dma_info_t *dmaiter = dmalist;
1364 
1365 	for (i = 0; i < count; i++, dmaiter++)
1366 		iflib_dma_free(*dmaiter);
1367 }
1368 
1369 #ifdef EARLY_AP_STARTUP
1370 static const int iflib_started = 1;
1371 #else
1372 /*
1373  * We used to abuse the smp_started flag to decide if the queues have been
1374  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1375  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1376  * is set.  Run a SYSINIT() strictly after that to just set a usable
1377  * completion flag.
1378  */
1379 
1380 static int iflib_started;
1381 
1382 static void
1383 iflib_record_started(void *arg)
1384 {
1385 	iflib_started = 1;
1386 }
1387 
1388 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1389 	iflib_record_started, NULL);
1390 #endif
1391 
1392 static int
1393 iflib_fast_intr(void *arg)
1394 {
1395 	iflib_filter_info_t info = arg;
1396 	struct grouptask *gtask = info->ifi_task;
1397 	if (!iflib_started)
1398 		return (FILTER_HANDLED);
1399 
1400 	DBG_COUNTER_INC(fast_intrs);
1401 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1402 		return (FILTER_HANDLED);
1403 
1404 	GROUPTASK_ENQUEUE(gtask);
1405 	return (FILTER_HANDLED);
1406 }
1407 
1408 static int
1409 iflib_fast_intr_rxtx(void *arg)
1410 {
1411 	iflib_filter_info_t info = arg;
1412 	struct grouptask *gtask = info->ifi_task;
1413 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1414 	if_ctx_t ctx = NULL;;
1415 	int i, cidx;
1416 
1417 	if (!iflib_started)
1418 		return (FILTER_HANDLED);
1419 
1420 	DBG_COUNTER_INC(fast_intrs);
1421 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1422 		return (FILTER_HANDLED);
1423 
1424 	MPASS(rxq->ifr_ntxqirq);
1425 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1426 		qidx_t txqid = rxq->ifr_txqid[i];
1427 
1428 		ctx = rxq->ifr_ctx;
1429 
1430 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1431 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1432 			continue;
1433 		}
1434 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1435 	}
1436 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1437 		cidx = rxq->ifr_cq_cidx;
1438 	else
1439 		cidx = rxq->ifr_fl[0].ifl_cidx;
1440 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1441 		GROUPTASK_ENQUEUE(gtask);
1442 	else
1443 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1444 	return (FILTER_HANDLED);
1445 }
1446 
1447 
1448 static int
1449 iflib_fast_intr_ctx(void *arg)
1450 {
1451 	iflib_filter_info_t info = arg;
1452 	struct grouptask *gtask = info->ifi_task;
1453 
1454 	if (!iflib_started)
1455 		return (FILTER_HANDLED);
1456 
1457 	DBG_COUNTER_INC(fast_intrs);
1458 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1459 		return (FILTER_HANDLED);
1460 
1461 	GROUPTASK_ENQUEUE(gtask);
1462 	return (FILTER_HANDLED);
1463 }
1464 
1465 static int
1466 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1467 	driver_filter_t filter, driver_intr_t handler, void *arg,
1468 				 char *name)
1469 {
1470 	int rc, flags;
1471 	struct resource *res;
1472 	void *tag = NULL;
1473 	device_t dev = ctx->ifc_dev;
1474 
1475 	flags = RF_ACTIVE;
1476 	if (ctx->ifc_flags & IFC_LEGACY)
1477 		flags |= RF_SHAREABLE;
1478 	MPASS(rid < 512);
1479 	irq->ii_rid = rid;
1480 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1481 	if (res == NULL) {
1482 		device_printf(dev,
1483 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1484 		return (ENOMEM);
1485 	}
1486 	irq->ii_res = res;
1487 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1488 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1489 						filter, handler, arg, &tag);
1490 	if (rc != 0) {
1491 		device_printf(dev,
1492 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1493 					  rid, name ? name : "unknown", rc);
1494 		return (rc);
1495 	} else if (name)
1496 		bus_describe_intr(dev, res, tag, "%s", name);
1497 
1498 	irq->ii_tag = tag;
1499 	return (0);
1500 }
1501 
1502 
1503 /*********************************************************************
1504  *
1505  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1506  *  the information needed to transmit a packet on the wire. This is
1507  *  called only once at attach, setup is done every reset.
1508  *
1509  **********************************************************************/
1510 
1511 static int
1512 iflib_txsd_alloc(iflib_txq_t txq)
1513 {
1514 	if_ctx_t ctx = txq->ift_ctx;
1515 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1516 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1517 	device_t dev = ctx->ifc_dev;
1518 	int err, nsegments, ntsosegments;
1519 
1520 	nsegments = scctx->isc_tx_nsegments;
1521 	ntsosegments = scctx->isc_tx_tso_segments_max;
1522 	MPASS(scctx->isc_ntxd[0] > 0);
1523 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1524 	MPASS(nsegments > 0);
1525 	MPASS(ntsosegments > 0);
1526 	/*
1527 	 * Setup DMA descriptor areas.
1528 	 */
1529 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1530 			       1, 0,			/* alignment, bounds */
1531 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1532 			       BUS_SPACE_MAXADDR,	/* highaddr */
1533 			       NULL, NULL,		/* filter, filterarg */
1534 			       sctx->isc_tx_maxsize,		/* maxsize */
1535 			       nsegments,	/* nsegments */
1536 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1537 			       0,			/* flags */
1538 			       NULL,			/* lockfunc */
1539 			       NULL,			/* lockfuncarg */
1540 			       &txq->ift_desc_tag))) {
1541 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1542 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1543 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1544 		goto fail;
1545 	}
1546 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1547 			       1, 0,			/* alignment, bounds */
1548 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1549 			       BUS_SPACE_MAXADDR,	/* highaddr */
1550 			       NULL, NULL,		/* filter, filterarg */
1551 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1552 			       ntsosegments,	/* nsegments */
1553 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1554 			       0,			/* flags */
1555 			       NULL,			/* lockfunc */
1556 			       NULL,			/* lockfuncarg */
1557 			       &txq->ift_tso_desc_tag))) {
1558 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1559 
1560 		goto fail;
1561 	}
1562 	if (!(txq->ift_sds.ifsd_flags =
1563 	    (uint8_t *) malloc(sizeof(uint8_t) *
1564 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1565 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1566 		err = ENOMEM;
1567 		goto fail;
1568 	}
1569 	if (!(txq->ift_sds.ifsd_m =
1570 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1571 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1572 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1573 		err = ENOMEM;
1574 		goto fail;
1575 	}
1576 
1577         /* Create the descriptor buffer dma maps */
1578 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1579 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1580 		return (0);
1581 
1582 	if (!(txq->ift_sds.ifsd_map =
1583 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1584 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1585 		err = ENOMEM;
1586 		goto fail;
1587 	}
1588 
1589 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1590 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1591 		if (err != 0) {
1592 			device_printf(dev, "Unable to create TX DMA map\n");
1593 			goto fail;
1594 		}
1595 	}
1596 #endif
1597 	return (0);
1598 fail:
1599 	/* We free all, it handles case where we are in the middle */
1600 	iflib_tx_structures_free(ctx);
1601 	return (err);
1602 }
1603 
1604 static void
1605 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1606 {
1607 	bus_dmamap_t map;
1608 
1609 	map = NULL;
1610 	if (txq->ift_sds.ifsd_map != NULL)
1611 		map = txq->ift_sds.ifsd_map[i];
1612 	if (map != NULL) {
1613 		bus_dmamap_unload(txq->ift_desc_tag, map);
1614 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1615 		txq->ift_sds.ifsd_map[i] = NULL;
1616 	}
1617 }
1618 
1619 static void
1620 iflib_txq_destroy(iflib_txq_t txq)
1621 {
1622 	if_ctx_t ctx = txq->ift_ctx;
1623 
1624 	for (int i = 0; i < txq->ift_size; i++)
1625 		iflib_txsd_destroy(ctx, txq, i);
1626 	if (txq->ift_sds.ifsd_map != NULL) {
1627 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1628 		txq->ift_sds.ifsd_map = NULL;
1629 	}
1630 	if (txq->ift_sds.ifsd_m != NULL) {
1631 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1632 		txq->ift_sds.ifsd_m = NULL;
1633 	}
1634 	if (txq->ift_sds.ifsd_flags != NULL) {
1635 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1636 		txq->ift_sds.ifsd_flags = NULL;
1637 	}
1638 	if (txq->ift_desc_tag != NULL) {
1639 		bus_dma_tag_destroy(txq->ift_desc_tag);
1640 		txq->ift_desc_tag = NULL;
1641 	}
1642 	if (txq->ift_tso_desc_tag != NULL) {
1643 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1644 		txq->ift_tso_desc_tag = NULL;
1645 	}
1646 }
1647 
1648 static void
1649 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1650 {
1651 	struct mbuf **mp;
1652 
1653 	mp = &txq->ift_sds.ifsd_m[i];
1654 	if (*mp == NULL)
1655 		return;
1656 
1657 	if (txq->ift_sds.ifsd_map != NULL) {
1658 		bus_dmamap_sync(txq->ift_desc_tag,
1659 				txq->ift_sds.ifsd_map[i],
1660 				BUS_DMASYNC_POSTWRITE);
1661 		bus_dmamap_unload(txq->ift_desc_tag,
1662 				  txq->ift_sds.ifsd_map[i]);
1663 	}
1664 	m_free(*mp);
1665 	DBG_COUNTER_INC(tx_frees);
1666 	*mp = NULL;
1667 }
1668 
1669 static int
1670 iflib_txq_setup(iflib_txq_t txq)
1671 {
1672 	if_ctx_t ctx = txq->ift_ctx;
1673 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1674 	iflib_dma_info_t di;
1675 	int i;
1676 
1677 	/* Set number of descriptors available */
1678 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1679 	/* XXX make configurable */
1680 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1681 
1682 	/* Reset indices */
1683 	txq->ift_cidx_processed = 0;
1684 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1685 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1686 
1687 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1688 		bzero((void *)di->idi_vaddr, di->idi_size);
1689 
1690 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1691 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1692 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1693 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1694 	return (0);
1695 }
1696 
1697 /*********************************************************************
1698  *
1699  *  Allocate memory for rx_buffer structures. Since we use one
1700  *  rx_buffer per received packet, the maximum number of rx_buffer's
1701  *  that we'll need is equal to the number of receive descriptors
1702  *  that we've allocated.
1703  *
1704  **********************************************************************/
1705 static int
1706 iflib_rxsd_alloc(iflib_rxq_t rxq)
1707 {
1708 	if_ctx_t ctx = rxq->ifr_ctx;
1709 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1710 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1711 	device_t dev = ctx->ifc_dev;
1712 	iflib_fl_t fl;
1713 	int			err;
1714 
1715 	MPASS(scctx->isc_nrxd[0] > 0);
1716 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1717 
1718 	fl = rxq->ifr_fl;
1719 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1720 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1721 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1722 					 1, 0,			/* alignment, bounds */
1723 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1724 					 BUS_SPACE_MAXADDR,	/* highaddr */
1725 					 NULL, NULL,		/* filter, filterarg */
1726 					 sctx->isc_rx_maxsize,	/* maxsize */
1727 					 sctx->isc_rx_nsegments,	/* nsegments */
1728 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1729 					 0,			/* flags */
1730 					 NULL,			/* lockfunc */
1731 					 NULL,			/* lockarg */
1732 					 &fl->ifl_desc_tag);
1733 		if (err) {
1734 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1735 				__func__, err);
1736 			goto fail;
1737 		}
1738 		if (!(fl->ifl_sds.ifsd_flags =
1739 		      (uint8_t *) malloc(sizeof(uint8_t) *
1740 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1741 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1742 			err = ENOMEM;
1743 			goto fail;
1744 		}
1745 		if (!(fl->ifl_sds.ifsd_m =
1746 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1747 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1748 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1749 			err = ENOMEM;
1750 			goto fail;
1751 		}
1752 		if (!(fl->ifl_sds.ifsd_cl =
1753 		      (caddr_t *) malloc(sizeof(caddr_t) *
1754 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1755 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1756 			err = ENOMEM;
1757 			goto fail;
1758 		}
1759 
1760 		/* Create the descriptor buffer dma maps */
1761 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1762 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1763 			continue;
1764 
1765 		if (!(fl->ifl_sds.ifsd_map =
1766 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1767 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1768 			err = ENOMEM;
1769 			goto fail;
1770 		}
1771 
1772 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1773 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1774 			if (err != 0) {
1775 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1776 				goto fail;
1777 			}
1778 		}
1779 #endif
1780 	}
1781 	return (0);
1782 
1783 fail:
1784 	iflib_rx_structures_free(ctx);
1785 	return (err);
1786 }
1787 
1788 
1789 /*
1790  * Internal service routines
1791  */
1792 
1793 struct rxq_refill_cb_arg {
1794 	int               error;
1795 	bus_dma_segment_t seg;
1796 	int               nseg;
1797 };
1798 
1799 static void
1800 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1801 {
1802 	struct rxq_refill_cb_arg *cb_arg = arg;
1803 
1804 	cb_arg->error = error;
1805 	cb_arg->seg = segs[0];
1806 	cb_arg->nseg = nseg;
1807 }
1808 
1809 
1810 #ifdef ACPI_DMAR
1811 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1812 #else
1813 #define IS_DMAR(ctx) (0)
1814 #endif
1815 
1816 /**
1817  *	rxq_refill - refill an rxq  free-buffer list
1818  *	@ctx: the iflib context
1819  *	@rxq: the free-list to refill
1820  *	@n: the number of new buffers to allocate
1821  *
1822  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1823  *	The caller must assure that @n does not exceed the queue's capacity.
1824  */
1825 static void
1826 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1827 {
1828 	struct mbuf *m;
1829 	int idx, frag_idx = fl->ifl_fragidx;
1830         int pidx = fl->ifl_pidx;
1831 	caddr_t cl, *sd_cl;
1832 	struct mbuf **sd_m;
1833 	uint8_t *sd_flags;
1834 	struct if_rxd_update iru;
1835 	bus_dmamap_t *sd_map;
1836 	int n, i = 0;
1837 	uint64_t bus_addr;
1838 	int err;
1839 	qidx_t credits;
1840 
1841 	sd_m = fl->ifl_sds.ifsd_m;
1842 	sd_map = fl->ifl_sds.ifsd_map;
1843 	sd_cl = fl->ifl_sds.ifsd_cl;
1844 	sd_flags = fl->ifl_sds.ifsd_flags;
1845 	idx = pidx;
1846 	credits = fl->ifl_credits;
1847 
1848 	n  = count;
1849 	MPASS(n > 0);
1850 	MPASS(credits + n <= fl->ifl_size);
1851 
1852 	if (pidx < fl->ifl_cidx)
1853 		MPASS(pidx + n <= fl->ifl_cidx);
1854 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1855 		MPASS(fl->ifl_gen == 0);
1856 	if (pidx > fl->ifl_cidx)
1857 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1858 
1859 	DBG_COUNTER_INC(fl_refills);
1860 	if (n > 8)
1861 		DBG_COUNTER_INC(fl_refills_large);
1862 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1863 	while (n--) {
1864 		/*
1865 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1866 		 * initialized after rx.
1867 		 *
1868 		 * If the cluster is still set then we know a minimum sized packet was received
1869 		 */
1870 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1871 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1872                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1873 		if ((cl = sd_cl[frag_idx]) == NULL) {
1874                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1875 				break;
1876 #if MEMORY_LOGGING
1877 			fl->ifl_cl_enqueued++;
1878 #endif
1879 		}
1880 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1881 			break;
1882 		}
1883 #if MEMORY_LOGGING
1884 		fl->ifl_m_enqueued++;
1885 #endif
1886 
1887 		DBG_COUNTER_INC(rx_allocs);
1888 #if defined(__i386__) || defined(__amd64__)
1889 		if (!IS_DMAR(ctx)) {
1890 			bus_addr = pmap_kextract((vm_offset_t)cl);
1891 		} else
1892 #endif
1893 		{
1894 			struct rxq_refill_cb_arg cb_arg;
1895 
1896 			cb_arg.error = 0;
1897 			MPASS(sd_map != NULL);
1898 			MPASS(sd_map[frag_idx] != NULL);
1899 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1900 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1901 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1902 					BUS_DMASYNC_PREREAD);
1903 
1904 			if (err != 0 || cb_arg.error) {
1905 				/*
1906 				 * !zone_pack ?
1907 				 */
1908 				if (fl->ifl_zone == zone_pack)
1909 					uma_zfree(fl->ifl_zone, cl);
1910 				m_free(m);
1911 				n = 0;
1912 				goto done;
1913 			}
1914 			bus_addr = cb_arg.seg.ds_addr;
1915 		}
1916                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1917 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1918 
1919 		MPASS(sd_m[frag_idx] == NULL);
1920 		sd_cl[frag_idx] = cl;
1921 		sd_m[frag_idx] = m;
1922 		fl->ifl_rxd_idxs[i] = frag_idx;
1923 		fl->ifl_bus_addrs[i] = bus_addr;
1924 		fl->ifl_vm_addrs[i] = cl;
1925 		credits++;
1926 		i++;
1927 		MPASS(credits <= fl->ifl_size);
1928 		if (++idx == fl->ifl_size) {
1929 			fl->ifl_gen = 1;
1930 			idx = 0;
1931 		}
1932 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1933 			iru.iru_pidx = pidx;
1934 			iru.iru_count = i;
1935 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1936 			i = 0;
1937 			pidx = idx;
1938 			fl->ifl_pidx = idx;
1939 			fl->ifl_credits = credits;
1940 		}
1941 
1942 	}
1943 done:
1944 	if (i) {
1945 		iru.iru_pidx = pidx;
1946 		iru.iru_count = i;
1947 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1948 		fl->ifl_pidx = idx;
1949 		fl->ifl_credits = credits;
1950 	}
1951 	DBG_COUNTER_INC(rxd_flush);
1952 	if (fl->ifl_pidx == 0)
1953 		pidx = fl->ifl_size - 1;
1954 	else
1955 		pidx = fl->ifl_pidx - 1;
1956 
1957 	if (sd_map)
1958 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1959 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1960 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1961 	fl->ifl_fragidx = frag_idx;
1962 }
1963 
1964 static __inline void
1965 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1966 {
1967 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1968 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1969 #ifdef INVARIANTS
1970 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1971 #endif
1972 
1973 	MPASS(fl->ifl_credits <= fl->ifl_size);
1974 	MPASS(reclaimable == delta);
1975 
1976 	if (reclaimable > 0)
1977 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1978 }
1979 
1980 static void
1981 iflib_fl_bufs_free(iflib_fl_t fl)
1982 {
1983 	iflib_dma_info_t idi = fl->ifl_ifdi;
1984 	uint32_t i;
1985 
1986 	for (i = 0; i < fl->ifl_size; i++) {
1987 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1988 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1989 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1990 
1991 		if (*sd_flags & RX_SW_DESC_INUSE) {
1992 			if (fl->ifl_sds.ifsd_map != NULL) {
1993 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1994 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1995 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
1996 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1997 			}
1998 			if (*sd_m != NULL) {
1999 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2000 				uma_zfree(zone_mbuf, *sd_m);
2001 			}
2002 			if (*sd_cl != NULL)
2003 				uma_zfree(fl->ifl_zone, *sd_cl);
2004 			*sd_flags = 0;
2005 		} else {
2006 			MPASS(*sd_cl == NULL);
2007 			MPASS(*sd_m == NULL);
2008 		}
2009 #if MEMORY_LOGGING
2010 		fl->ifl_m_dequeued++;
2011 		fl->ifl_cl_dequeued++;
2012 #endif
2013 		*sd_cl = NULL;
2014 		*sd_m = NULL;
2015 	}
2016 #ifdef INVARIANTS
2017 	for (i = 0; i < fl->ifl_size; i++) {
2018 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2019 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2020 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2021 	}
2022 #endif
2023 	/*
2024 	 * Reset free list values
2025 	 */
2026 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2027 	bzero(idi->idi_vaddr, idi->idi_size);
2028 }
2029 
2030 /*********************************************************************
2031  *
2032  *  Initialize a receive ring and its buffers.
2033  *
2034  **********************************************************************/
2035 static int
2036 iflib_fl_setup(iflib_fl_t fl)
2037 {
2038 	iflib_rxq_t rxq = fl->ifl_rxq;
2039 	if_ctx_t ctx = rxq->ifr_ctx;
2040 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2041 
2042 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2043 	/*
2044 	** Free current RX buffer structs and their mbufs
2045 	*/
2046 	iflib_fl_bufs_free(fl);
2047 	/* Now replenish the mbufs */
2048 	MPASS(fl->ifl_credits == 0);
2049 	/*
2050 	 * XXX don't set the max_frame_size to larger
2051 	 * than the hardware can handle
2052 	 */
2053 	if (sctx->isc_max_frame_size <= 2048)
2054 		fl->ifl_buf_size = MCLBYTES;
2055 #ifndef CONTIGMALLOC_WORKS
2056 	else
2057 		fl->ifl_buf_size = MJUMPAGESIZE;
2058 #else
2059 	else if (sctx->isc_max_frame_size <= 4096)
2060 		fl->ifl_buf_size = MJUMPAGESIZE;
2061 	else if (sctx->isc_max_frame_size <= 9216)
2062 		fl->ifl_buf_size = MJUM9BYTES;
2063 	else
2064 		fl->ifl_buf_size = MJUM16BYTES;
2065 #endif
2066 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2067 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2068 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2069 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2070 
2071 
2072 	/* avoid pre-allocating zillions of clusters to an idle card
2073 	 * potentially speeding up attach
2074 	 */
2075 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2076 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2077 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2078 		return (ENOBUFS);
2079 	/*
2080 	 * handle failure
2081 	 */
2082 	MPASS(rxq != NULL);
2083 	MPASS(fl->ifl_ifdi != NULL);
2084 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2085 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2086 	return (0);
2087 }
2088 
2089 /*********************************************************************
2090  *
2091  *  Free receive ring data structures
2092  *
2093  **********************************************************************/
2094 static void
2095 iflib_rx_sds_free(iflib_rxq_t rxq)
2096 {
2097 	iflib_fl_t fl;
2098 	int i;
2099 
2100 	if (rxq->ifr_fl != NULL) {
2101 		for (i = 0; i < rxq->ifr_nfl; i++) {
2102 			fl = &rxq->ifr_fl[i];
2103 			if (fl->ifl_desc_tag != NULL) {
2104 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2105 				fl->ifl_desc_tag = NULL;
2106 			}
2107 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2108 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2109 			/* XXX destroy maps first */
2110 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2111 			fl->ifl_sds.ifsd_m = NULL;
2112 			fl->ifl_sds.ifsd_cl = NULL;
2113 			fl->ifl_sds.ifsd_map = NULL;
2114 		}
2115 		free(rxq->ifr_fl, M_IFLIB);
2116 		rxq->ifr_fl = NULL;
2117 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2118 	}
2119 }
2120 
2121 /*
2122  * MI independent logic
2123  *
2124  */
2125 static void
2126 iflib_timer(void *arg)
2127 {
2128 	iflib_txq_t txq = arg;
2129 	if_ctx_t ctx = txq->ift_ctx;
2130 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2131 
2132 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2133 		return;
2134 	/*
2135 	** Check on the state of the TX queue(s), this
2136 	** can be done without the lock because its RO
2137 	** and the HUNG state will be static if set.
2138 	*/
2139 	IFDI_TIMER(ctx, txq->ift_id);
2140 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2141 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2142 	     (sctx->isc_pause_frames == 0)))
2143 		goto hung;
2144 
2145 	if (ifmp_ring_is_stalled(txq->ift_br))
2146 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2147 	txq->ift_cleaned_prev = txq->ift_cleaned;
2148 	/* handle any laggards */
2149 	if (txq->ift_db_pending)
2150 		GROUPTASK_ENQUEUE(&txq->ift_task);
2151 
2152 	sctx->isc_pause_frames = 0;
2153 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2154 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2155 	return;
2156  hung:
2157 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2158 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2159 	STATE_LOCK(ctx);
2160 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2161 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2162 	iflib_admin_intr_deferred(ctx);
2163 	STATE_UNLOCK(ctx);
2164 }
2165 
2166 static void
2167 iflib_init_locked(if_ctx_t ctx)
2168 {
2169 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2170 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2171 	if_t ifp = ctx->ifc_ifp;
2172 	iflib_fl_t fl;
2173 	iflib_txq_t txq;
2174 	iflib_rxq_t rxq;
2175 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2176 
2177 
2178 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2179 	IFDI_INTR_DISABLE(ctx);
2180 
2181 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2182 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2183 	/* Set hardware offload abilities */
2184 	if_clearhwassist(ifp);
2185 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2186 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2187 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2188 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2189 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2190 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2191 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2192 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2193 
2194 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2195 		CALLOUT_LOCK(txq);
2196 		callout_stop(&txq->ift_timer);
2197 		CALLOUT_UNLOCK(txq);
2198 		iflib_netmap_txq_init(ctx, txq);
2199 	}
2200 #ifdef INVARIANTS
2201 	i = if_getdrvflags(ifp);
2202 #endif
2203 	IFDI_INIT(ctx);
2204 	MPASS(if_getdrvflags(ifp) == i);
2205 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2206 		/* XXX this should really be done on a per-queue basis */
2207 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2208 			MPASS(rxq->ifr_id == i);
2209 			iflib_netmap_rxq_init(ctx, rxq);
2210 			continue;
2211 		}
2212 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2213 			if (iflib_fl_setup(fl)) {
2214 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2215 				goto done;
2216 			}
2217 		}
2218 	}
2219 	done:
2220 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2221 	IFDI_INTR_ENABLE(ctx);
2222 	txq = ctx->ifc_txqs;
2223 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2224 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2225 			txq->ift_timer.c_cpu);
2226 }
2227 
2228 static int
2229 iflib_media_change(if_t ifp)
2230 {
2231 	if_ctx_t ctx = if_getsoftc(ifp);
2232 	int err;
2233 
2234 	CTX_LOCK(ctx);
2235 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2236 		iflib_init_locked(ctx);
2237 	CTX_UNLOCK(ctx);
2238 	return (err);
2239 }
2240 
2241 static void
2242 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2243 {
2244 	if_ctx_t ctx = if_getsoftc(ifp);
2245 
2246 	CTX_LOCK(ctx);
2247 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2248 	IFDI_MEDIA_STATUS(ctx, ifmr);
2249 	CTX_UNLOCK(ctx);
2250 }
2251 
2252 static void
2253 iflib_stop(if_ctx_t ctx)
2254 {
2255 	iflib_txq_t txq = ctx->ifc_txqs;
2256 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2257 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2258 	iflib_dma_info_t di;
2259 	iflib_fl_t fl;
2260 	int i, j;
2261 
2262 	/* Tell the stack that the interface is no longer active */
2263 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2264 
2265 	IFDI_INTR_DISABLE(ctx);
2266 	DELAY(1000);
2267 	IFDI_STOP(ctx);
2268 	DELAY(1000);
2269 
2270 	iflib_debug_reset();
2271 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2272 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2273 		/* make sure all transmitters have completed before proceeding XXX */
2274 
2275 		CALLOUT_LOCK(txq);
2276 		callout_stop(&txq->ift_timer);
2277 		CALLOUT_UNLOCK(txq);
2278 
2279 		/* clean any enqueued buffers */
2280 		iflib_ifmp_purge(txq);
2281 		/* Free any existing tx buffers. */
2282 		for (j = 0; j < txq->ift_size; j++) {
2283 			iflib_txsd_free(ctx, txq, j);
2284 		}
2285 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2286 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2287 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2288 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2289 		txq->ift_pullups = 0;
2290 		ifmp_ring_reset_stats(txq->ift_br);
2291 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2292 			bzero((void *)di->idi_vaddr, di->idi_size);
2293 	}
2294 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2295 		/* make sure all transmitters have completed before proceeding XXX */
2296 
2297 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2298 			bzero((void *)di->idi_vaddr, di->idi_size);
2299 		/* also resets the free lists pidx/cidx */
2300 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2301 			iflib_fl_bufs_free(fl);
2302 	}
2303 }
2304 
2305 static inline caddr_t
2306 calc_next_rxd(iflib_fl_t fl, int cidx)
2307 {
2308 	qidx_t size;
2309 	int nrxd;
2310 	caddr_t start, end, cur, next;
2311 
2312 	nrxd = fl->ifl_size;
2313 	size = fl->ifl_rxd_size;
2314 	start = fl->ifl_ifdi->idi_vaddr;
2315 
2316 	if (__predict_false(size == 0))
2317 		return (start);
2318 	cur = start + size*cidx;
2319 	end = start + size*nrxd;
2320 	next = CACHE_PTR_NEXT(cur);
2321 	return (next < end ? next : start);
2322 }
2323 
2324 static inline void
2325 prefetch_pkts(iflib_fl_t fl, int cidx)
2326 {
2327 	int nextptr;
2328 	int nrxd = fl->ifl_size;
2329 	caddr_t next_rxd;
2330 
2331 
2332 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2333 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2334 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2335 	next_rxd = calc_next_rxd(fl, cidx);
2336 	prefetch(next_rxd);
2337 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2338 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2339 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2340 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2341 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2342 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2343 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2344 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2345 }
2346 
2347 static void
2348 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2349 {
2350 	int flid, cidx;
2351 	bus_dmamap_t map;
2352 	iflib_fl_t fl;
2353 	iflib_dma_info_t di;
2354 	int next;
2355 
2356 	map = NULL;
2357 	flid = irf->irf_flid;
2358 	cidx = irf->irf_idx;
2359 	fl = &rxq->ifr_fl[flid];
2360 	sd->ifsd_fl = fl;
2361 	sd->ifsd_cidx = cidx;
2362 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2363 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2364 	fl->ifl_credits--;
2365 #if MEMORY_LOGGING
2366 	fl->ifl_m_dequeued++;
2367 #endif
2368 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2369 		prefetch_pkts(fl, cidx);
2370 	if (fl->ifl_sds.ifsd_map != NULL) {
2371 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2372 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2373 		map = fl->ifl_sds.ifsd_map[cidx];
2374 		di = fl->ifl_ifdi;
2375 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2376 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2377 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2378 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2379 
2380 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2381 		MPASS(fl->ifl_cidx == cidx);
2382 		if (unload)
2383 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2384 	}
2385 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2386 	if (__predict_false(fl->ifl_cidx == 0))
2387 		fl->ifl_gen = 0;
2388 	if (map != NULL)
2389 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2390 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2391         bit_clear(fl->ifl_rx_bitmap, cidx);
2392 }
2393 
2394 static struct mbuf *
2395 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2396 {
2397 	int i, padlen , flags;
2398 	struct mbuf *m, *mh, *mt;
2399 	caddr_t cl;
2400 
2401 	i = 0;
2402 	mh = NULL;
2403 	do {
2404 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2405 
2406 		MPASS(*sd->ifsd_cl != NULL);
2407 		MPASS(*sd->ifsd_m != NULL);
2408 
2409 		/* Don't include zero-length frags */
2410 		if (ri->iri_frags[i].irf_len == 0) {
2411 			/* XXX we can save the cluster here, but not the mbuf */
2412 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2413 			m_free(*sd->ifsd_m);
2414 			*sd->ifsd_m = NULL;
2415 			continue;
2416 		}
2417 		m = *sd->ifsd_m;
2418 		*sd->ifsd_m = NULL;
2419 		if (mh == NULL) {
2420 			flags = M_PKTHDR|M_EXT;
2421 			mh = mt = m;
2422 			padlen = ri->iri_pad;
2423 		} else {
2424 			flags = M_EXT;
2425 			mt->m_next = m;
2426 			mt = m;
2427 			/* assuming padding is only on the first fragment */
2428 			padlen = 0;
2429 		}
2430 		cl = *sd->ifsd_cl;
2431 		*sd->ifsd_cl = NULL;
2432 
2433 		/* Can these two be made one ? */
2434 		m_init(m, M_NOWAIT, MT_DATA, flags);
2435 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2436 		/*
2437 		 * These must follow m_init and m_cljset
2438 		 */
2439 		m->m_data += padlen;
2440 		ri->iri_len -= padlen;
2441 		m->m_len = ri->iri_frags[i].irf_len;
2442 	} while (++i < ri->iri_nfrags);
2443 
2444 	return (mh);
2445 }
2446 
2447 /*
2448  * Process one software descriptor
2449  */
2450 static struct mbuf *
2451 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2452 {
2453 	struct if_rxsd sd;
2454 	struct mbuf *m;
2455 
2456 	/* should I merge this back in now that the two paths are basically duplicated? */
2457 	if (ri->iri_nfrags == 1 &&
2458 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2459 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2460 		m = *sd.ifsd_m;
2461 		*sd.ifsd_m = NULL;
2462 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2463 #ifndef __NO_STRICT_ALIGNMENT
2464 		if (!IP_ALIGNED(m))
2465 			m->m_data += 2;
2466 #endif
2467 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2468 		m->m_len = ri->iri_frags[0].irf_len;
2469        } else {
2470 		m = assemble_segments(rxq, ri, &sd);
2471 	}
2472 	m->m_pkthdr.len = ri->iri_len;
2473 	m->m_pkthdr.rcvif = ri->iri_ifp;
2474 	m->m_flags |= ri->iri_flags;
2475 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2476 	m->m_pkthdr.flowid = ri->iri_flowid;
2477 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2478 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2479 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2480 	return (m);
2481 }
2482 
2483 #if defined(INET6) || defined(INET)
2484 static void
2485 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2486 {
2487 	CURVNET_SET(lc->ifp->if_vnet);
2488 #if defined(INET6)
2489 	*v6 = VNET(ip6_forwarding);
2490 #endif
2491 #if defined(INET)
2492 	*v4 = VNET(ipforwarding);
2493 #endif
2494 	CURVNET_RESTORE();
2495 }
2496 
2497 /*
2498  * Returns true if it's possible this packet could be LROed.
2499  * if it returns false, it is guaranteed that tcp_lro_rx()
2500  * would not return zero.
2501  */
2502 static bool
2503 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2504 {
2505 	struct ether_header *eh;
2506 	uint16_t eh_type;
2507 
2508 	eh = mtod(m, struct ether_header *);
2509 	eh_type = ntohs(eh->ether_type);
2510 	switch (eh_type) {
2511 #if defined(INET6)
2512 		case ETHERTYPE_IPV6:
2513 			return !v6_forwarding;
2514 #endif
2515 #if defined (INET)
2516 		case ETHERTYPE_IP:
2517 			return !v4_forwarding;
2518 #endif
2519 	}
2520 
2521 	return false;
2522 }
2523 #else
2524 static void
2525 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2526 {
2527 }
2528 #endif
2529 
2530 static bool
2531 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2532 {
2533 	if_ctx_t ctx = rxq->ifr_ctx;
2534 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2535 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2536 	int avail, i;
2537 	qidx_t *cidxp;
2538 	struct if_rxd_info ri;
2539 	int err, budget_left, rx_bytes, rx_pkts;
2540 	iflib_fl_t fl;
2541 	struct ifnet *ifp;
2542 	int lro_enabled;
2543 	bool lro_possible = false;
2544 	bool v4_forwarding, v6_forwarding;
2545 
2546 	/*
2547 	 * XXX early demux data packets so that if_input processing only handles
2548 	 * acks in interrupt context
2549 	 */
2550 	struct mbuf *m, *mh, *mt, *mf;
2551 
2552 	ifp = ctx->ifc_ifp;
2553 	mh = mt = NULL;
2554 	MPASS(budget > 0);
2555 	rx_pkts	= rx_bytes = 0;
2556 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2557 		cidxp = &rxq->ifr_cq_cidx;
2558 	else
2559 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2560 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2561 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2562 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2563 		DBG_COUNTER_INC(rx_unavail);
2564 		return (false);
2565 	}
2566 
2567 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2568 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2569 			DBG_COUNTER_INC(rx_ctx_inactive);
2570 			break;
2571 		}
2572 		/*
2573 		 * Reset client set fields to their default values
2574 		 */
2575 		rxd_info_zero(&ri);
2576 		ri.iri_qsidx = rxq->ifr_id;
2577 		ri.iri_cidx = *cidxp;
2578 		ri.iri_ifp = ifp;
2579 		ri.iri_frags = rxq->ifr_frags;
2580 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2581 
2582 		if (err)
2583 			goto err;
2584 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2585 			*cidxp = ri.iri_cidx;
2586 			/* Update our consumer index */
2587 			/* XXX NB: shurd - check if this is still safe */
2588 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2589 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2590 				rxq->ifr_cq_gen = 0;
2591 			}
2592 			/* was this only a completion queue message? */
2593 			if (__predict_false(ri.iri_nfrags == 0))
2594 				continue;
2595 		}
2596 		MPASS(ri.iri_nfrags != 0);
2597 		MPASS(ri.iri_len != 0);
2598 
2599 		/* will advance the cidx on the corresponding free lists */
2600 		m = iflib_rxd_pkt_get(rxq, &ri);
2601 		if (avail == 0 && budget_left)
2602 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2603 
2604 		if (__predict_false(m == NULL)) {
2605 			DBG_COUNTER_INC(rx_mbuf_null);
2606 			continue;
2607 		}
2608 		/* imm_pkt: -- cxgb */
2609 		if (mh == NULL)
2610 			mh = mt = m;
2611 		else {
2612 			mt->m_nextpkt = m;
2613 			mt = m;
2614 		}
2615 	}
2616 	/* make sure that we can refill faster than drain */
2617 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2618 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2619 
2620 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2621 	if (lro_enabled)
2622 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2623 	mt = mf = NULL;
2624 	while (mh != NULL) {
2625 		m = mh;
2626 		mh = mh->m_nextpkt;
2627 		m->m_nextpkt = NULL;
2628 #ifndef __NO_STRICT_ALIGNMENT
2629 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2630 			continue;
2631 #endif
2632 		rx_bytes += m->m_pkthdr.len;
2633 		rx_pkts++;
2634 #if defined(INET6) || defined(INET)
2635 		if (lro_enabled) {
2636 			if (!lro_possible) {
2637 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2638 				if (lro_possible && mf != NULL) {
2639 					ifp->if_input(ifp, mf);
2640 					DBG_COUNTER_INC(rx_if_input);
2641 					mt = mf = NULL;
2642 				}
2643 			}
2644 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2645 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2646 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2647 					continue;
2648 			}
2649 		}
2650 #endif
2651 		if (lro_possible) {
2652 			ifp->if_input(ifp, m);
2653 			DBG_COUNTER_INC(rx_if_input);
2654 			continue;
2655 		}
2656 
2657 		if (mf == NULL)
2658 			mf = m;
2659 		if (mt != NULL)
2660 			mt->m_nextpkt = m;
2661 		mt = m;
2662 	}
2663 	if (mf != NULL) {
2664 		ifp->if_input(ifp, mf);
2665 		DBG_COUNTER_INC(rx_if_input);
2666 	}
2667 
2668 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2669 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2670 
2671 	/*
2672 	 * Flush any outstanding LRO work
2673 	 */
2674 #if defined(INET6) || defined(INET)
2675 	tcp_lro_flush_all(&rxq->ifr_lc);
2676 #endif
2677 	if (avail)
2678 		return true;
2679 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2680 err:
2681 	STATE_LOCK(ctx);
2682 	ctx->ifc_flags |= IFC_DO_RESET;
2683 	iflib_admin_intr_deferred(ctx);
2684 	STATE_UNLOCK(ctx);
2685 	return (false);
2686 }
2687 
2688 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2689 static inline qidx_t
2690 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2691 {
2692 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2693 	qidx_t minthresh = txq->ift_size / 8;
2694 	if (in_use > 4*minthresh)
2695 		return (notify_count);
2696 	if (in_use > 2*minthresh)
2697 		return (notify_count >> 1);
2698 	if (in_use > minthresh)
2699 		return (notify_count >> 3);
2700 	return (0);
2701 }
2702 
2703 static inline qidx_t
2704 txq_max_rs_deferred(iflib_txq_t txq)
2705 {
2706 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2707 	qidx_t minthresh = txq->ift_size / 8;
2708 	if (txq->ift_in_use > 4*minthresh)
2709 		return (notify_count);
2710 	if (txq->ift_in_use > 2*minthresh)
2711 		return (notify_count >> 1);
2712 	if (txq->ift_in_use > minthresh)
2713 		return (notify_count >> 2);
2714 	return (2);
2715 }
2716 
2717 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2718 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2719 
2720 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2721 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2722 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2723 
2724 /* forward compatibility for cxgb */
2725 #define FIRST_QSET(ctx) 0
2726 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2727 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2728 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2729 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2730 
2731 /* XXX we should be setting this to something other than zero */
2732 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2733 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2734 
2735 static inline bool
2736 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2737 {
2738 	qidx_t dbval, max;
2739 	bool rang;
2740 
2741 	rang = false;
2742 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2743 	if (ring || txq->ift_db_pending >= max) {
2744 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2745 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2746 		txq->ift_db_pending = txq->ift_npending = 0;
2747 		rang = true;
2748 	}
2749 	return (rang);
2750 }
2751 
2752 #ifdef PKT_DEBUG
2753 static void
2754 print_pkt(if_pkt_info_t pi)
2755 {
2756 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2757 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2758 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2759 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2760 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2761 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2762 }
2763 #endif
2764 
2765 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2766 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2767 
2768 static int
2769 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2770 {
2771 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2772 	struct ether_vlan_header *eh;
2773 	struct mbuf *m, *n;
2774 
2775 	n = m = *mp;
2776 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2777 	    M_WRITABLE(m) == 0) {
2778 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2779 			return (ENOMEM);
2780 		} else {
2781 			m_freem(*mp);
2782 			n = *mp = m;
2783 		}
2784 	}
2785 
2786 	/*
2787 	 * Determine where frame payload starts.
2788 	 * Jump over vlan headers if already present,
2789 	 * helpful for QinQ too.
2790 	 */
2791 	if (__predict_false(m->m_len < sizeof(*eh))) {
2792 		txq->ift_pullups++;
2793 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2794 			return (ENOMEM);
2795 	}
2796 	eh = mtod(m, struct ether_vlan_header *);
2797 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2798 		pi->ipi_etype = ntohs(eh->evl_proto);
2799 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2800 	} else {
2801 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2802 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2803 	}
2804 
2805 	switch (pi->ipi_etype) {
2806 #ifdef INET
2807 	case ETHERTYPE_IP:
2808 	{
2809 		struct ip *ip = NULL;
2810 		struct tcphdr *th = NULL;
2811 		int minthlen;
2812 
2813 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2814 		if (__predict_false(m->m_len < minthlen)) {
2815 			/*
2816 			 * if this code bloat is causing too much of a hit
2817 			 * move it to a separate function and mark it noinline
2818 			 */
2819 			if (m->m_len == pi->ipi_ehdrlen) {
2820 				n = m->m_next;
2821 				MPASS(n);
2822 				if (n->m_len >= sizeof(*ip))  {
2823 					ip = (struct ip *)n->m_data;
2824 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2825 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2826 				} else {
2827 					txq->ift_pullups++;
2828 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2829 						return (ENOMEM);
2830 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2831 				}
2832 			} else {
2833 				txq->ift_pullups++;
2834 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2835 					return (ENOMEM);
2836 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2837 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2838 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2839 			}
2840 		} else {
2841 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2842 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2843 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2844 		}
2845 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2846 		pi->ipi_ipproto = ip->ip_p;
2847 		pi->ipi_flags |= IPI_TX_IPV4;
2848 
2849 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2850                        ip->ip_sum = 0;
2851 
2852 		if (IS_TSO4(pi)) {
2853 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2854 				if (__predict_false(th == NULL)) {
2855 					txq->ift_pullups++;
2856 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2857 						return (ENOMEM);
2858 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2859 				}
2860 				pi->ipi_tcp_hflags = th->th_flags;
2861 				pi->ipi_tcp_hlen = th->th_off << 2;
2862 				pi->ipi_tcp_seq = th->th_seq;
2863 			}
2864 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2865 				return (ENXIO);
2866 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2867 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2868 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2869 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2870 				ip->ip_sum = 0;
2871 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2872 			}
2873 		}
2874 		break;
2875 	}
2876 #endif
2877 #ifdef INET6
2878 	case ETHERTYPE_IPV6:
2879 	{
2880 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2881 		struct tcphdr *th;
2882 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2883 
2884 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2885 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2886 				return (ENOMEM);
2887 		}
2888 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2889 
2890 		/* XXX-BZ this will go badly in case of ext hdrs. */
2891 		pi->ipi_ipproto = ip6->ip6_nxt;
2892 		pi->ipi_flags |= IPI_TX_IPV6;
2893 
2894 		if (IS_TSO6(pi)) {
2895 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2896 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2897 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2898 						return (ENOMEM);
2899 				}
2900 				pi->ipi_tcp_hflags = th->th_flags;
2901 				pi->ipi_tcp_hlen = th->th_off << 2;
2902 			}
2903 
2904 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2905 				return (ENXIO);
2906 			/*
2907 			 * The corresponding flag is set by the stack in the IPv4
2908 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2909 			 * So, set it here because the rest of the flow requires it.
2910 			 */
2911 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2912 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2913 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2914 		}
2915 		break;
2916 	}
2917 #endif
2918 	default:
2919 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2920 		pi->ipi_ip_hlen = 0;
2921 		break;
2922 	}
2923 	*mp = m;
2924 
2925 	return (0);
2926 }
2927 
2928 static  __noinline  struct mbuf *
2929 collapse_pkthdr(struct mbuf *m0)
2930 {
2931 	struct mbuf *m, *m_next, *tmp;
2932 
2933 	m = m0;
2934 	m_next = m->m_next;
2935 	while (m_next != NULL && m_next->m_len == 0) {
2936 		m = m_next;
2937 		m->m_next = NULL;
2938 		m_free(m);
2939 		m_next = m_next->m_next;
2940 	}
2941 	m = m0;
2942 	m->m_next = m_next;
2943 	if ((m_next->m_flags & M_EXT) == 0) {
2944 		m = m_defrag(m, M_NOWAIT);
2945 	} else {
2946 		tmp = m_next->m_next;
2947 		memcpy(m_next, m, MPKTHSIZE);
2948 		m = m_next;
2949 		m->m_next = tmp;
2950 	}
2951 	return (m);
2952 }
2953 
2954 /*
2955  * If dodgy hardware rejects the scatter gather chain we've handed it
2956  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2957  * m_defrag'd mbufs
2958  */
2959 static __noinline struct mbuf *
2960 iflib_remove_mbuf(iflib_txq_t txq)
2961 {
2962 	int ntxd, i, pidx;
2963 	struct mbuf *m, *mh, **ifsd_m;
2964 
2965 	pidx = txq->ift_pidx;
2966 	ifsd_m = txq->ift_sds.ifsd_m;
2967 	ntxd = txq->ift_size;
2968 	mh = m = ifsd_m[pidx];
2969 	ifsd_m[pidx] = NULL;
2970 #if MEMORY_LOGGING
2971 	txq->ift_dequeued++;
2972 #endif
2973 	i = 1;
2974 
2975 	while (m) {
2976 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2977 #if MEMORY_LOGGING
2978 		txq->ift_dequeued++;
2979 #endif
2980 		m = m->m_next;
2981 		i++;
2982 	}
2983 	return (mh);
2984 }
2985 
2986 static int
2987 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2988 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2989 			  int max_segs, int flags)
2990 {
2991 	if_ctx_t ctx;
2992 	if_shared_ctx_t		sctx;
2993 	if_softc_ctx_t		scctx;
2994 	int i, next, pidx, err, ntxd, count;
2995 	struct mbuf *m, *tmp, **ifsd_m;
2996 
2997 	m = *m0;
2998 
2999 	/*
3000 	 * Please don't ever do this
3001 	 */
3002 	if (__predict_false(m->m_len == 0))
3003 		*m0 = m = collapse_pkthdr(m);
3004 
3005 	ctx = txq->ift_ctx;
3006 	sctx = ctx->ifc_sctx;
3007 	scctx = &ctx->ifc_softc_ctx;
3008 	ifsd_m = txq->ift_sds.ifsd_m;
3009 	ntxd = txq->ift_size;
3010 	pidx = txq->ift_pidx;
3011 	if (map != NULL) {
3012 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3013 
3014 		err = bus_dmamap_load_mbuf_sg(tag, map,
3015 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3016 		if (err)
3017 			return (err);
3018 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3019 		count = 0;
3020 		m = *m0;
3021 		do {
3022 			if (__predict_false(m->m_len <= 0)) {
3023 				tmp = m;
3024 				m = m->m_next;
3025 				tmp->m_next = NULL;
3026 				m_free(tmp);
3027 				continue;
3028 			}
3029 			m = m->m_next;
3030 			count++;
3031 		} while (m != NULL);
3032 		if (count > *nsegs) {
3033 			ifsd_m[pidx] = *m0;
3034 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3035 			return (0);
3036 		}
3037 		m = *m0;
3038 		count = 0;
3039 		do {
3040 			next = (pidx + count) & (ntxd-1);
3041 			MPASS(ifsd_m[next] == NULL);
3042 			ifsd_m[next] = m;
3043 			count++;
3044 			tmp = m;
3045 			m = m->m_next;
3046 		} while (m != NULL);
3047 	} else {
3048 		int buflen, sgsize, maxsegsz, max_sgsize;
3049 		vm_offset_t vaddr;
3050 		vm_paddr_t curaddr;
3051 
3052 		count = i = 0;
3053 		m = *m0;
3054 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3055 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3056 		else
3057 			maxsegsz = sctx->isc_tx_maxsegsize;
3058 
3059 		do {
3060 			if (__predict_false(m->m_len <= 0)) {
3061 				tmp = m;
3062 				m = m->m_next;
3063 				tmp->m_next = NULL;
3064 				m_free(tmp);
3065 				continue;
3066 			}
3067 			buflen = m->m_len;
3068 			vaddr = (vm_offset_t)m->m_data;
3069 			/*
3070 			 * see if we can't be smarter about physically
3071 			 * contiguous mappings
3072 			 */
3073 			next = (pidx + count) & (ntxd-1);
3074 			MPASS(ifsd_m[next] == NULL);
3075 #if MEMORY_LOGGING
3076 			txq->ift_enqueued++;
3077 #endif
3078 			ifsd_m[next] = m;
3079 			while (buflen > 0) {
3080 				if (i >= max_segs)
3081 					goto err;
3082 				max_sgsize = MIN(buflen, maxsegsz);
3083 				curaddr = pmap_kextract(vaddr);
3084 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3085 				sgsize = MIN(sgsize, max_sgsize);
3086 				segs[i].ds_addr = curaddr;
3087 				segs[i].ds_len = sgsize;
3088 				vaddr += sgsize;
3089 				buflen -= sgsize;
3090 				i++;
3091 			}
3092 			count++;
3093 			tmp = m;
3094 			m = m->m_next;
3095 		} while (m != NULL);
3096 		*nsegs = i;
3097 	}
3098 	return (0);
3099 err:
3100 	*m0 = iflib_remove_mbuf(txq);
3101 	return (EFBIG);
3102 }
3103 
3104 static inline caddr_t
3105 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3106 {
3107 	qidx_t size;
3108 	int ntxd;
3109 	caddr_t start, end, cur, next;
3110 
3111 	ntxd = txq->ift_size;
3112 	size = txq->ift_txd_size[qid];
3113 	start = txq->ift_ifdi[qid].idi_vaddr;
3114 
3115 	if (__predict_false(size == 0))
3116 		return (start);
3117 	cur = start + size*cidx;
3118 	end = start + size*ntxd;
3119 	next = CACHE_PTR_NEXT(cur);
3120 	return (next < end ? next : start);
3121 }
3122 
3123 /*
3124  * Pad an mbuf to ensure a minimum ethernet frame size.
3125  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3126  */
3127 static __noinline int
3128 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3129 {
3130 	/*
3131 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3132 	 * and ARP message is the smallest common payload I can think of
3133 	 */
3134 	static char pad[18];	/* just zeros */
3135 	int n;
3136 	struct mbuf *new_head;
3137 
3138 	if (!M_WRITABLE(*m_head)) {
3139 		new_head = m_dup(*m_head, M_NOWAIT);
3140 		if (new_head == NULL) {
3141 			m_freem(*m_head);
3142 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3143 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3144 			return ENOMEM;
3145 		}
3146 		m_freem(*m_head);
3147 		*m_head = new_head;
3148 	}
3149 
3150 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3151 	     n > 0; n -= sizeof(pad))
3152 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3153 			break;
3154 
3155 	if (n > 0) {
3156 		m_freem(*m_head);
3157 		device_printf(dev, "cannot pad short frame\n");
3158 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3159 		return (ENOBUFS);
3160 	}
3161 
3162 	return 0;
3163 }
3164 
3165 static int
3166 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3167 {
3168 	if_ctx_t		ctx;
3169 	if_shared_ctx_t		sctx;
3170 	if_softc_ctx_t		scctx;
3171 	bus_dma_segment_t	*segs;
3172 	struct mbuf		*m_head;
3173 	void			*next_txd;
3174 	bus_dmamap_t		map;
3175 	struct if_pkt_info	pi;
3176 	int remap = 0;
3177 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3178 	bus_dma_tag_t desc_tag;
3179 
3180 	segs = txq->ift_segs;
3181 	ctx = txq->ift_ctx;
3182 	sctx = ctx->ifc_sctx;
3183 	scctx = &ctx->ifc_softc_ctx;
3184 	segs = txq->ift_segs;
3185 	ntxd = txq->ift_size;
3186 	m_head = *m_headp;
3187 	map = NULL;
3188 
3189 	/*
3190 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3191 	 */
3192 	cidx = txq->ift_cidx;
3193 	pidx = txq->ift_pidx;
3194 	if (ctx->ifc_flags & IFC_PREFETCH) {
3195 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3196 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3197 			next_txd = calc_next_txd(txq, cidx, 0);
3198 			prefetch(next_txd);
3199 		}
3200 
3201 		/* prefetch the next cache line of mbuf pointers and flags */
3202 		prefetch(&txq->ift_sds.ifsd_m[next]);
3203 		if (txq->ift_sds.ifsd_map != NULL) {
3204 			prefetch(&txq->ift_sds.ifsd_map[next]);
3205 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3206 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3207 		}
3208 	} else if (txq->ift_sds.ifsd_map != NULL)
3209 		map = txq->ift_sds.ifsd_map[pidx];
3210 
3211 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3212 		desc_tag = txq->ift_tso_desc_tag;
3213 		max_segs = scctx->isc_tx_tso_segments_max;
3214 	} else {
3215 		desc_tag = txq->ift_desc_tag;
3216 		max_segs = scctx->isc_tx_nsegments;
3217 	}
3218 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3219 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3220 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3221 		if (err)
3222 			return err;
3223 	}
3224 	m_head = *m_headp;
3225 
3226 	pkt_info_zero(&pi);
3227 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3228 	pi.ipi_pidx = pidx;
3229 	pi.ipi_qsidx = txq->ift_id;
3230 	pi.ipi_len = m_head->m_pkthdr.len;
3231 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3232 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3233 
3234 	/* deliberate bitwise OR to make one condition */
3235 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3236 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3237 			return (err);
3238 		m_head = *m_headp;
3239 	}
3240 
3241 retry:
3242 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3243 defrag:
3244 	if (__predict_false(err)) {
3245 		switch (err) {
3246 		case EFBIG:
3247 			/* try collapse once and defrag once */
3248 			if (remap == 0) {
3249 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3250 				/* try defrag if collapsing fails */
3251 				if (m_head == NULL)
3252 					remap++;
3253 			}
3254 			if (remap == 1)
3255 				m_head = m_defrag(*m_headp, M_NOWAIT);
3256 			remap++;
3257 			if (__predict_false(m_head == NULL))
3258 				goto defrag_failed;
3259 			txq->ift_mbuf_defrag++;
3260 			*m_headp = m_head;
3261 			goto retry;
3262 			break;
3263 		case ENOMEM:
3264 			txq->ift_no_tx_dma_setup++;
3265 			break;
3266 		default:
3267 			txq->ift_no_tx_dma_setup++;
3268 			m_freem(*m_headp);
3269 			DBG_COUNTER_INC(tx_frees);
3270 			*m_headp = NULL;
3271 			break;
3272 		}
3273 		txq->ift_map_failed++;
3274 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3275 		return (err);
3276 	}
3277 
3278 	/*
3279 	 * XXX assumes a 1 to 1 relationship between segments and
3280 	 *        descriptors - this does not hold true on all drivers, e.g.
3281 	 *        cxgb
3282 	 */
3283 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3284 		txq->ift_no_desc_avail++;
3285 		if (map != NULL)
3286 			bus_dmamap_unload(desc_tag, map);
3287 		DBG_COUNTER_INC(encap_txq_avail_fail);
3288 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3289 			GROUPTASK_ENQUEUE(&txq->ift_task);
3290 		return (ENOBUFS);
3291 	}
3292 	/*
3293 	 * On Intel cards we can greatly reduce the number of TX interrupts
3294 	 * we see by only setting report status on every Nth descriptor.
3295 	 * However, this also means that the driver will need to keep track
3296 	 * of the descriptors that RS was set on to check them for the DD bit.
3297 	 */
3298 	txq->ift_rs_pending += nsegs + 1;
3299 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3300 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3301 		pi.ipi_flags |= IPI_TX_INTR;
3302 		txq->ift_rs_pending = 0;
3303 	}
3304 
3305 	pi.ipi_segs = segs;
3306 	pi.ipi_nsegs = nsegs;
3307 
3308 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3309 #ifdef PKT_DEBUG
3310 	print_pkt(&pi);
3311 #endif
3312 	if (map != NULL)
3313 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3314 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3315 		if (map != NULL)
3316 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3317 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3318 		DBG_COUNTER_INC(tx_encap);
3319 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3320 
3321 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3322 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3323 			ndesc += txq->ift_size;
3324 			txq->ift_gen = 1;
3325 		}
3326 		/*
3327 		 * drivers can need as many as
3328 		 * two sentinels
3329 		 */
3330 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3331 		MPASS(pi.ipi_new_pidx != pidx);
3332 		MPASS(ndesc > 0);
3333 		txq->ift_in_use += ndesc;
3334 
3335 		/*
3336 		 * We update the last software descriptor again here because there may
3337 		 * be a sentinel and/or there may be more mbufs than segments
3338 		 */
3339 		txq->ift_pidx = pi.ipi_new_pidx;
3340 		txq->ift_npending += pi.ipi_ndescs;
3341 	} else {
3342 		*m_headp = m_head = iflib_remove_mbuf(txq);
3343 		if (err == EFBIG) {
3344 			txq->ift_txd_encap_efbig++;
3345 			if (remap < 2) {
3346 				remap = 1;
3347 				goto defrag;
3348 			}
3349 		}
3350 		DBG_COUNTER_INC(encap_txd_encap_fail);
3351 		goto defrag_failed;
3352 	}
3353 	return (err);
3354 
3355 defrag_failed:
3356 	txq->ift_mbuf_defrag_failed++;
3357 	txq->ift_map_failed++;
3358 	m_freem(*m_headp);
3359 	DBG_COUNTER_INC(tx_frees);
3360 	*m_headp = NULL;
3361 	return (ENOMEM);
3362 }
3363 
3364 static void
3365 iflib_tx_desc_free(iflib_txq_t txq, int n)
3366 {
3367 	int hasmap;
3368 	uint32_t qsize, cidx, mask, gen;
3369 	struct mbuf *m, **ifsd_m;
3370 	uint8_t *ifsd_flags;
3371 	bus_dmamap_t *ifsd_map;
3372 	bool do_prefetch;
3373 
3374 	cidx = txq->ift_cidx;
3375 	gen = txq->ift_gen;
3376 	qsize = txq->ift_size;
3377 	mask = qsize-1;
3378 	hasmap = txq->ift_sds.ifsd_map != NULL;
3379 	ifsd_flags = txq->ift_sds.ifsd_flags;
3380 	ifsd_m = txq->ift_sds.ifsd_m;
3381 	ifsd_map = txq->ift_sds.ifsd_map;
3382 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3383 
3384 	while (n--) {
3385 		if (do_prefetch) {
3386 			prefetch(ifsd_m[(cidx + 3) & mask]);
3387 			prefetch(ifsd_m[(cidx + 4) & mask]);
3388 		}
3389 		if (ifsd_m[cidx] != NULL) {
3390 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3391 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3392 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3393 				/*
3394 				 * does it matter if it's not the TSO tag? If so we'll
3395 				 * have to add the type to flags
3396 				 */
3397 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3398 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3399 			}
3400 			if ((m = ifsd_m[cidx]) != NULL) {
3401 				/* XXX we don't support any drivers that batch packets yet */
3402 				MPASS(m->m_nextpkt == NULL);
3403 				/* if the number of clusters exceeds the number of segments
3404 				 * there won't be space on the ring to save a pointer to each
3405 				 * cluster so we simply free the list here
3406 				 */
3407 				if (m->m_flags & M_TOOBIG) {
3408 					m_freem(m);
3409 				} else {
3410 					m_free(m);
3411 				}
3412 				ifsd_m[cidx] = NULL;
3413 #if MEMORY_LOGGING
3414 				txq->ift_dequeued++;
3415 #endif
3416 				DBG_COUNTER_INC(tx_frees);
3417 			}
3418 		}
3419 		if (__predict_false(++cidx == qsize)) {
3420 			cidx = 0;
3421 			gen = 0;
3422 		}
3423 	}
3424 	txq->ift_cidx = cidx;
3425 	txq->ift_gen = gen;
3426 }
3427 
3428 static __inline int
3429 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3430 {
3431 	int reclaim;
3432 	if_ctx_t ctx = txq->ift_ctx;
3433 
3434 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3435 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3436 
3437 	/*
3438 	 * Need a rate-limiting check so that this isn't called every time
3439 	 */
3440 	iflib_tx_credits_update(ctx, txq);
3441 	reclaim = DESC_RECLAIMABLE(txq);
3442 
3443 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3444 #ifdef INVARIANTS
3445 		if (iflib_verbose_debug) {
3446 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3447 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3448 			       reclaim, thresh);
3449 
3450 		}
3451 #endif
3452 		return (0);
3453 	}
3454 	iflib_tx_desc_free(txq, reclaim);
3455 	txq->ift_cleaned += reclaim;
3456 	txq->ift_in_use -= reclaim;
3457 
3458 	return (reclaim);
3459 }
3460 
3461 static struct mbuf **
3462 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3463 {
3464 	int next, size;
3465 	struct mbuf **items;
3466 
3467 	size = r->size;
3468 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3469 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3470 
3471 	prefetch(items[(cidx + offset) & (size-1)]);
3472 	if (remaining > 1) {
3473 		prefetch2cachelines(&items[next]);
3474 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3475 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3476 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3477 	}
3478 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3479 }
3480 
3481 static void
3482 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3483 {
3484 
3485 	ifmp_ring_check_drainage(txq->ift_br, budget);
3486 }
3487 
3488 static uint32_t
3489 iflib_txq_can_drain(struct ifmp_ring *r)
3490 {
3491 	iflib_txq_t txq = r->cookie;
3492 	if_ctx_t ctx = txq->ift_ctx;
3493 
3494 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3495 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3496 }
3497 
3498 static uint32_t
3499 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3500 {
3501 	iflib_txq_t txq = r->cookie;
3502 	if_ctx_t ctx = txq->ift_ctx;
3503 	struct ifnet *ifp = ctx->ifc_ifp;
3504 	struct mbuf **mp, *m;
3505 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3506 	int reclaimed, err, in_use_prev, desc_used;
3507 	bool do_prefetch, ring, rang;
3508 
3509 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3510 			    !LINK_ACTIVE(ctx))) {
3511 		DBG_COUNTER_INC(txq_drain_notready);
3512 		return (0);
3513 	}
3514 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3515 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3516 	avail = IDXDIFF(pidx, cidx, r->size);
3517 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3518 		DBG_COUNTER_INC(txq_drain_flushing);
3519 		for (i = 0; i < avail; i++) {
3520 			m_free(r->items[(cidx + i) & (r->size-1)]);
3521 			r->items[(cidx + i) & (r->size-1)] = NULL;
3522 		}
3523 		return (avail);
3524 	}
3525 
3526 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3527 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3528 		CALLOUT_LOCK(txq);
3529 		callout_stop(&txq->ift_timer);
3530 		CALLOUT_UNLOCK(txq);
3531 		DBG_COUNTER_INC(txq_drain_oactive);
3532 		return (0);
3533 	}
3534 	if (reclaimed)
3535 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3536 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3537 	count = MIN(avail, TX_BATCH_SIZE);
3538 #ifdef INVARIANTS
3539 	if (iflib_verbose_debug)
3540 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3541 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3542 #endif
3543 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3544 	avail = TXQ_AVAIL(txq);
3545 	err = 0;
3546 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3547 		int rem = do_prefetch ? count - i : 0;
3548 
3549 		mp = _ring_peek_one(r, cidx, i, rem);
3550 		MPASS(mp != NULL && *mp != NULL);
3551 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3552 			consumed++;
3553 			reclaimed++;
3554 			continue;
3555 		}
3556 		in_use_prev = txq->ift_in_use;
3557 		err = iflib_encap(txq, mp);
3558 		if (__predict_false(err)) {
3559 			DBG_COUNTER_INC(txq_drain_encapfail);
3560 			/* no room - bail out */
3561 			if (err == ENOBUFS)
3562 				break;
3563 			consumed++;
3564 			DBG_COUNTER_INC(txq_drain_encapfail);
3565 			/* we can't send this packet - skip it */
3566 			continue;
3567 		}
3568 		consumed++;
3569 		pkt_sent++;
3570 		m = *mp;
3571 		DBG_COUNTER_INC(tx_sent);
3572 		bytes_sent += m->m_pkthdr.len;
3573 		mcast_sent += !!(m->m_flags & M_MCAST);
3574 		avail = TXQ_AVAIL(txq);
3575 
3576 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3577 		desc_used += (txq->ift_in_use - in_use_prev);
3578 		ETHER_BPF_MTAP(ifp, m);
3579 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3580 			break;
3581 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3582 	}
3583 
3584 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3585 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3586 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3587 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3588 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3589 	if (mcast_sent)
3590 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3591 #ifdef INVARIANTS
3592 	if (iflib_verbose_debug)
3593 		printf("consumed=%d\n", consumed);
3594 #endif
3595 	return (consumed);
3596 }
3597 
3598 static uint32_t
3599 iflib_txq_drain_always(struct ifmp_ring *r)
3600 {
3601 	return (1);
3602 }
3603 
3604 static uint32_t
3605 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3606 {
3607 	int i, avail;
3608 	struct mbuf **mp;
3609 	iflib_txq_t txq;
3610 
3611 	txq = r->cookie;
3612 
3613 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3614 	CALLOUT_LOCK(txq);
3615 	callout_stop(&txq->ift_timer);
3616 	CALLOUT_UNLOCK(txq);
3617 
3618 	avail = IDXDIFF(pidx, cidx, r->size);
3619 	for (i = 0; i < avail; i++) {
3620 		mp = _ring_peek_one(r, cidx, i, avail - i);
3621 		if (__predict_false(*mp == (struct mbuf *)txq))
3622 			continue;
3623 		m_freem(*mp);
3624 	}
3625 	MPASS(ifmp_ring_is_stalled(r) == 0);
3626 	return (avail);
3627 }
3628 
3629 static void
3630 iflib_ifmp_purge(iflib_txq_t txq)
3631 {
3632 	struct ifmp_ring *r;
3633 
3634 	r = txq->ift_br;
3635 	r->drain = iflib_txq_drain_free;
3636 	r->can_drain = iflib_txq_drain_always;
3637 
3638 	ifmp_ring_check_drainage(r, r->size);
3639 
3640 	r->drain = iflib_txq_drain;
3641 	r->can_drain = iflib_txq_can_drain;
3642 }
3643 
3644 static void
3645 _task_fn_tx(void *context)
3646 {
3647 	iflib_txq_t txq = context;
3648 	if_ctx_t ctx = txq->ift_ctx;
3649 	struct ifnet *ifp = ctx->ifc_ifp;
3650 
3651 #ifdef IFLIB_DIAGNOSTICS
3652 	txq->ift_cpu_exec_count[curcpu]++;
3653 #endif
3654 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3655 		return;
3656 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3657 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3658 			netmap_tx_irq(ifp, txq->ift_id);
3659 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3660 		return;
3661 	}
3662 	if (txq->ift_db_pending)
3663 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3664 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3665 	if (ctx->ifc_flags & IFC_LEGACY)
3666 		IFDI_INTR_ENABLE(ctx);
3667 	else {
3668 #ifdef INVARIANTS
3669 		int rc =
3670 #endif
3671 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3672 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3673 	}
3674 }
3675 
3676 static void
3677 _task_fn_rx(void *context)
3678 {
3679 	iflib_rxq_t rxq = context;
3680 	if_ctx_t ctx = rxq->ifr_ctx;
3681 	bool more;
3682 	uint16_t budget;
3683 
3684 #ifdef IFLIB_DIAGNOSTICS
3685 	rxq->ifr_cpu_exec_count[curcpu]++;
3686 #endif
3687 	DBG_COUNTER_INC(task_fn_rxs);
3688 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3689 		return;
3690 	more = true;
3691 #ifdef DEV_NETMAP
3692 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3693 		u_int work = 0;
3694 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3695 			more = false;
3696 		}
3697 	}
3698 #endif
3699 	budget = ctx->ifc_sysctl_rx_budget;
3700 	if (budget == 0)
3701 		budget = 16;	/* XXX */
3702 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3703 		if (ctx->ifc_flags & IFC_LEGACY)
3704 			IFDI_INTR_ENABLE(ctx);
3705 		else {
3706 #ifdef INVARIANTS
3707 			int rc =
3708 #endif
3709 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3710 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3711 			DBG_COUNTER_INC(rx_intr_enables);
3712 		}
3713 	}
3714 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3715 		return;
3716 	if (more)
3717 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3718 }
3719 
3720 static void
3721 _task_fn_admin(void *context)
3722 {
3723 	if_ctx_t ctx = context;
3724 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3725 	iflib_txq_t txq;
3726 	int i;
3727 	bool oactive, running, do_reset, do_watchdog;
3728 
3729 	STATE_LOCK(ctx);
3730 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3731 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3732 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3733 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3734 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3735 	STATE_UNLOCK(ctx);
3736 
3737 	if (!running & !oactive)
3738 		return;
3739 
3740 	CTX_LOCK(ctx);
3741 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3742 		CALLOUT_LOCK(txq);
3743 		callout_stop(&txq->ift_timer);
3744 		CALLOUT_UNLOCK(txq);
3745 	}
3746 	if (do_watchdog) {
3747 		ctx->ifc_watchdog_events++;
3748 		IFDI_WATCHDOG_RESET(ctx);
3749 	}
3750 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3751 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3752 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3753 	IFDI_LINK_INTR_ENABLE(ctx);
3754 	if (do_reset)
3755 		iflib_if_init_locked(ctx);
3756 	CTX_UNLOCK(ctx);
3757 
3758 	if (LINK_ACTIVE(ctx) == 0)
3759 		return;
3760 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3761 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3762 }
3763 
3764 
3765 static void
3766 _task_fn_iov(void *context)
3767 {
3768 	if_ctx_t ctx = context;
3769 
3770 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3771 		return;
3772 
3773 	CTX_LOCK(ctx);
3774 	IFDI_VFLR_HANDLE(ctx);
3775 	CTX_UNLOCK(ctx);
3776 }
3777 
3778 static int
3779 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3780 {
3781 	int err;
3782 	if_int_delay_info_t info;
3783 	if_ctx_t ctx;
3784 
3785 	info = (if_int_delay_info_t)arg1;
3786 	ctx = info->iidi_ctx;
3787 	info->iidi_req = req;
3788 	info->iidi_oidp = oidp;
3789 	CTX_LOCK(ctx);
3790 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3791 	CTX_UNLOCK(ctx);
3792 	return (err);
3793 }
3794 
3795 /*********************************************************************
3796  *
3797  *  IFNET FUNCTIONS
3798  *
3799  **********************************************************************/
3800 
3801 static void
3802 iflib_if_init_locked(if_ctx_t ctx)
3803 {
3804 	iflib_stop(ctx);
3805 	iflib_init_locked(ctx);
3806 }
3807 
3808 
3809 static void
3810 iflib_if_init(void *arg)
3811 {
3812 	if_ctx_t ctx = arg;
3813 
3814 	CTX_LOCK(ctx);
3815 	iflib_if_init_locked(ctx);
3816 	CTX_UNLOCK(ctx);
3817 }
3818 
3819 static int
3820 iflib_if_transmit(if_t ifp, struct mbuf *m)
3821 {
3822 	if_ctx_t	ctx = if_getsoftc(ifp);
3823 
3824 	iflib_txq_t txq;
3825 	int err, qidx;
3826 
3827 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3828 		DBG_COUNTER_INC(tx_frees);
3829 		m_freem(m);
3830 		return (ENOBUFS);
3831 	}
3832 
3833 	MPASS(m->m_nextpkt == NULL);
3834 	qidx = 0;
3835 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3836 		qidx = QIDX(ctx, m);
3837 	/*
3838 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3839 	 */
3840 	txq = &ctx->ifc_txqs[qidx];
3841 
3842 #ifdef DRIVER_BACKPRESSURE
3843 	if (txq->ift_closed) {
3844 		while (m != NULL) {
3845 			next = m->m_nextpkt;
3846 			m->m_nextpkt = NULL;
3847 			m_freem(m);
3848 			m = next;
3849 		}
3850 		return (ENOBUFS);
3851 	}
3852 #endif
3853 #ifdef notyet
3854 	qidx = count = 0;
3855 	mp = marr;
3856 	next = m;
3857 	do {
3858 		count++;
3859 		next = next->m_nextpkt;
3860 	} while (next != NULL);
3861 
3862 	if (count > nitems(marr))
3863 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3864 			/* XXX check nextpkt */
3865 			m_freem(m);
3866 			/* XXX simplify for now */
3867 			DBG_COUNTER_INC(tx_frees);
3868 			return (ENOBUFS);
3869 		}
3870 	for (next = m, i = 0; next != NULL; i++) {
3871 		mp[i] = next;
3872 		next = next->m_nextpkt;
3873 		mp[i]->m_nextpkt = NULL;
3874 	}
3875 #endif
3876 	DBG_COUNTER_INC(tx_seen);
3877 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3878 
3879 	GROUPTASK_ENQUEUE(&txq->ift_task);
3880 	if (err) {
3881 		/* support forthcoming later */
3882 #ifdef DRIVER_BACKPRESSURE
3883 		txq->ift_closed = TRUE;
3884 #endif
3885 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3886 		m_freem(m);
3887 	}
3888 
3889 	return (err);
3890 }
3891 
3892 static void
3893 iflib_if_qflush(if_t ifp)
3894 {
3895 	if_ctx_t ctx = if_getsoftc(ifp);
3896 	iflib_txq_t txq = ctx->ifc_txqs;
3897 	int i;
3898 
3899 	STATE_LOCK(ctx);
3900 	ctx->ifc_flags |= IFC_QFLUSH;
3901 	STATE_UNLOCK(ctx);
3902 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3903 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3904 			iflib_txq_check_drain(txq, 0);
3905 	STATE_LOCK(ctx);
3906 	ctx->ifc_flags &= ~IFC_QFLUSH;
3907 	STATE_UNLOCK(ctx);
3908 
3909 	if_qflush(ifp);
3910 }
3911 
3912 
3913 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3914 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3915 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3916 
3917 static int
3918 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3919 {
3920 	if_ctx_t ctx = if_getsoftc(ifp);
3921 	struct ifreq	*ifr = (struct ifreq *)data;
3922 #if defined(INET) || defined(INET6)
3923 	struct ifaddr	*ifa = (struct ifaddr *)data;
3924 #endif
3925 	bool		avoid_reset = FALSE;
3926 	int		err = 0, reinit = 0, bits;
3927 
3928 	switch (command) {
3929 	case SIOCSIFADDR:
3930 #ifdef INET
3931 		if (ifa->ifa_addr->sa_family == AF_INET)
3932 			avoid_reset = TRUE;
3933 #endif
3934 #ifdef INET6
3935 		if (ifa->ifa_addr->sa_family == AF_INET6)
3936 			avoid_reset = TRUE;
3937 #endif
3938 		/*
3939 		** Calling init results in link renegotiation,
3940 		** so we avoid doing it when possible.
3941 		*/
3942 		if (avoid_reset) {
3943 			if_setflagbits(ifp, IFF_UP,0);
3944 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3945 				reinit = 1;
3946 #ifdef INET
3947 			if (!(if_getflags(ifp) & IFF_NOARP))
3948 				arp_ifinit(ifp, ifa);
3949 #endif
3950 		} else
3951 			err = ether_ioctl(ifp, command, data);
3952 		break;
3953 	case SIOCSIFMTU:
3954 		CTX_LOCK(ctx);
3955 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3956 			CTX_UNLOCK(ctx);
3957 			break;
3958 		}
3959 		bits = if_getdrvflags(ifp);
3960 		/* stop the driver and free any clusters before proceeding */
3961 		iflib_stop(ctx);
3962 
3963 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3964 			STATE_LOCK(ctx);
3965 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3966 				ctx->ifc_flags |= IFC_MULTISEG;
3967 			else
3968 				ctx->ifc_flags &= ~IFC_MULTISEG;
3969 			STATE_UNLOCK(ctx);
3970 			err = if_setmtu(ifp, ifr->ifr_mtu);
3971 		}
3972 		iflib_init_locked(ctx);
3973 		STATE_LOCK(ctx);
3974 		if_setdrvflags(ifp, bits);
3975 		STATE_UNLOCK(ctx);
3976 		CTX_UNLOCK(ctx);
3977 		break;
3978 	case SIOCSIFFLAGS:
3979 		CTX_LOCK(ctx);
3980 		if (if_getflags(ifp) & IFF_UP) {
3981 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3982 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3983 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3984 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3985 				}
3986 			} else
3987 				reinit = 1;
3988 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3989 			iflib_stop(ctx);
3990 		}
3991 		ctx->ifc_if_flags = if_getflags(ifp);
3992 		CTX_UNLOCK(ctx);
3993 		break;
3994 	case SIOCADDMULTI:
3995 	case SIOCDELMULTI:
3996 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3997 			CTX_LOCK(ctx);
3998 			IFDI_INTR_DISABLE(ctx);
3999 			IFDI_MULTI_SET(ctx);
4000 			IFDI_INTR_ENABLE(ctx);
4001 			CTX_UNLOCK(ctx);
4002 		}
4003 		break;
4004 	case SIOCSIFMEDIA:
4005 		CTX_LOCK(ctx);
4006 		IFDI_MEDIA_SET(ctx);
4007 		CTX_UNLOCK(ctx);
4008 		/* falls thru */
4009 	case SIOCGIFMEDIA:
4010 	case SIOCGIFXMEDIA:
4011 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4012 		break;
4013 	case SIOCGI2C:
4014 	{
4015 		struct ifi2creq i2c;
4016 
4017 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4018 		if (err != 0)
4019 			break;
4020 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4021 			err = EINVAL;
4022 			break;
4023 		}
4024 		if (i2c.len > sizeof(i2c.data)) {
4025 			err = EINVAL;
4026 			break;
4027 		}
4028 
4029 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4030 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4031 			    sizeof(i2c));
4032 		break;
4033 	}
4034 	case SIOCSIFCAP:
4035 	{
4036 		int mask, setmask;
4037 
4038 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4039 		setmask = 0;
4040 #ifdef TCP_OFFLOAD
4041 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4042 #endif
4043 		setmask |= (mask & IFCAP_FLAGS);
4044 
4045 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4046 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4047 		if ((mask & IFCAP_WOL) &&
4048 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4049 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4050 		if_vlancap(ifp);
4051 		/*
4052 		 * want to ensure that traffic has stopped before we change any of the flags
4053 		 */
4054 		if (setmask) {
4055 			CTX_LOCK(ctx);
4056 			bits = if_getdrvflags(ifp);
4057 			if (bits & IFF_DRV_RUNNING)
4058 				iflib_stop(ctx);
4059 			STATE_LOCK(ctx);
4060 			if_togglecapenable(ifp, setmask);
4061 			STATE_UNLOCK(ctx);
4062 			if (bits & IFF_DRV_RUNNING)
4063 				iflib_init_locked(ctx);
4064 			STATE_LOCK(ctx);
4065 			if_setdrvflags(ifp, bits);
4066 			STATE_UNLOCK(ctx);
4067 			CTX_UNLOCK(ctx);
4068 		}
4069 		break;
4070 	    }
4071 	case SIOCGPRIVATE_0:
4072 	case SIOCSDRVSPEC:
4073 	case SIOCGDRVSPEC:
4074 		CTX_LOCK(ctx);
4075 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4076 		CTX_UNLOCK(ctx);
4077 		break;
4078 	default:
4079 		err = ether_ioctl(ifp, command, data);
4080 		break;
4081 	}
4082 	if (reinit)
4083 		iflib_if_init(ctx);
4084 	return (err);
4085 }
4086 
4087 static uint64_t
4088 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4089 {
4090 	if_ctx_t ctx = if_getsoftc(ifp);
4091 
4092 	return (IFDI_GET_COUNTER(ctx, cnt));
4093 }
4094 
4095 /*********************************************************************
4096  *
4097  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4098  *
4099  **********************************************************************/
4100 
4101 static void
4102 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4103 {
4104 	if_ctx_t ctx = if_getsoftc(ifp);
4105 
4106 	if ((void *)ctx != arg)
4107 		return;
4108 
4109 	if ((vtag == 0) || (vtag > 4095))
4110 		return;
4111 
4112 	CTX_LOCK(ctx);
4113 	IFDI_VLAN_REGISTER(ctx, vtag);
4114 	/* Re-init to load the changes */
4115 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4116 		iflib_if_init_locked(ctx);
4117 	CTX_UNLOCK(ctx);
4118 }
4119 
4120 static void
4121 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4122 {
4123 	if_ctx_t ctx = if_getsoftc(ifp);
4124 
4125 	if ((void *)ctx != arg)
4126 		return;
4127 
4128 	if ((vtag == 0) || (vtag > 4095))
4129 		return;
4130 
4131 	CTX_LOCK(ctx);
4132 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4133 	/* Re-init to load the changes */
4134 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4135 		iflib_if_init_locked(ctx);
4136 	CTX_UNLOCK(ctx);
4137 }
4138 
4139 static void
4140 iflib_led_func(void *arg, int onoff)
4141 {
4142 	if_ctx_t ctx = arg;
4143 
4144 	CTX_LOCK(ctx);
4145 	IFDI_LED_FUNC(ctx, onoff);
4146 	CTX_UNLOCK(ctx);
4147 }
4148 
4149 /*********************************************************************
4150  *
4151  *  BUS FUNCTION DEFINITIONS
4152  *
4153  **********************************************************************/
4154 
4155 int
4156 iflib_device_probe(device_t dev)
4157 {
4158 	pci_vendor_info_t *ent;
4159 
4160 	uint16_t	pci_vendor_id, pci_device_id;
4161 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4162 	uint16_t	pci_rev_id;
4163 	if_shared_ctx_t sctx;
4164 
4165 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4166 		return (ENOTSUP);
4167 
4168 	pci_vendor_id = pci_get_vendor(dev);
4169 	pci_device_id = pci_get_device(dev);
4170 	pci_subvendor_id = pci_get_subvendor(dev);
4171 	pci_subdevice_id = pci_get_subdevice(dev);
4172 	pci_rev_id = pci_get_revid(dev);
4173 	if (sctx->isc_parse_devinfo != NULL)
4174 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4175 
4176 	ent = sctx->isc_vendor_info;
4177 	while (ent->pvi_vendor_id != 0) {
4178 		if (pci_vendor_id != ent->pvi_vendor_id) {
4179 			ent++;
4180 			continue;
4181 		}
4182 		if ((pci_device_id == ent->pvi_device_id) &&
4183 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4184 		     (ent->pvi_subvendor_id == 0)) &&
4185 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4186 		     (ent->pvi_subdevice_id == 0)) &&
4187 		    ((pci_rev_id == ent->pvi_rev_id) ||
4188 		     (ent->pvi_rev_id == 0))) {
4189 
4190 			device_set_desc_copy(dev, ent->pvi_name);
4191 			/* this needs to be changed to zero if the bus probing code
4192 			 * ever stops re-probing on best match because the sctx
4193 			 * may have its values over written by register calls
4194 			 * in subsequent probes
4195 			 */
4196 			return (BUS_PROBE_DEFAULT);
4197 		}
4198 		ent++;
4199 	}
4200 	return (ENXIO);
4201 }
4202 
4203 int
4204 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4205 {
4206 	int err, rid, msix;
4207 	if_ctx_t ctx;
4208 	if_t ifp;
4209 	if_softc_ctx_t scctx;
4210 	int i;
4211 	uint16_t main_txq;
4212 	uint16_t main_rxq;
4213 
4214 
4215 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4216 
4217 	if (sc == NULL) {
4218 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4219 		device_set_softc(dev, ctx);
4220 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4221 	}
4222 
4223 	ctx->ifc_sctx = sctx;
4224 	ctx->ifc_dev = dev;
4225 	ctx->ifc_softc = sc;
4226 
4227 	if ((err = iflib_register(ctx)) != 0) {
4228 		device_printf(dev, "iflib_register failed %d\n", err);
4229 		return (err);
4230 	}
4231 	iflib_add_device_sysctl_pre(ctx);
4232 
4233 	scctx = &ctx->ifc_softc_ctx;
4234 	ifp = ctx->ifc_ifp;
4235 	ctx->ifc_nhwtxqs = sctx->isc_ntxqs;
4236 
4237 	/*
4238 	 * XXX sanity check that ntxd & nrxd are a power of 2
4239 	 */
4240 	if (ctx->ifc_sysctl_ntxqs != 0)
4241 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4242 	if (ctx->ifc_sysctl_nrxqs != 0)
4243 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4244 
4245 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4246 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4247 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4248 		else
4249 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4250 	}
4251 
4252 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4253 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4254 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4255 		else
4256 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4257 	}
4258 
4259 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4260 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4261 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4262 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4263 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4264 		}
4265 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4266 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4267 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4268 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4269 		}
4270 	}
4271 
4272 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4273 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4274 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4275 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4276 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4277 		}
4278 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4279 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4280 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4281 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4282 		}
4283 	}
4284 
4285 	CTX_LOCK(ctx);
4286 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4287 		CTX_UNLOCK(ctx);
4288 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4289 		return (err);
4290 	}
4291 	_iflib_pre_assert(scctx);
4292 	ctx->ifc_txrx = *scctx->isc_txrx;
4293 
4294 #ifdef INVARIANTS
4295 	MPASS(scctx->isc_capenable);
4296 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4297 		MPASS(scctx->isc_tx_csum_flags);
4298 #endif
4299 
4300 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4301 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4302 
4303 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4304 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4305 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4306 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4307 
4308 #ifdef ACPI_DMAR
4309 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4310 		ctx->ifc_flags |= IFC_DMAR;
4311 #elif !(defined(__i386__) || defined(__amd64__))
4312 	/* set unconditionally for !x86 */
4313 	ctx->ifc_flags |= IFC_DMAR;
4314 #endif
4315 
4316 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4317 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4318 
4319 	/* XXX change for per-queue sizes */
4320 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4321 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4322 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4323 		if (!powerof2(scctx->isc_nrxd[i])) {
4324 			/* round down instead? */
4325 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4326 			err = EINVAL;
4327 			goto fail;
4328 		}
4329 	}
4330 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4331 		if (!powerof2(scctx->isc_ntxd[i])) {
4332 			device_printf(dev,
4333 			    "# tx descriptors must be a power of 2");
4334 			err = EINVAL;
4335 			goto fail;
4336 		}
4337 	}
4338 
4339 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4340 	    MAX_SINGLE_PACKET_FRACTION)
4341 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4342 		    MAX_SINGLE_PACKET_FRACTION);
4343 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4344 	    MAX_SINGLE_PACKET_FRACTION)
4345 		scctx->isc_tx_tso_segments_max = max(1,
4346 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4347 
4348 	/*
4349 	 * Protect the stack against modern hardware
4350 	 */
4351 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4352 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4353 
4354 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4355 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4356 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4357 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4358 	if (scctx->isc_rss_table_size == 0)
4359 		scctx->isc_rss_table_size = 64;
4360 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4361 
4362 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4363 	/* XXX format name */
4364 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4365 
4366 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4367 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4368 		device_printf(dev, "Unable to fetch CPU list\n");
4369 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4370 	}
4371 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4372 
4373 	/*
4374 	** Now setup MSI or MSI/X, should
4375 	** return us the number of supported
4376 	** vectors. (Will be 1 for MSI)
4377 	*/
4378 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4379 		msix = scctx->isc_vectors;
4380 	} else if (scctx->isc_msix_bar != 0)
4381 	       /*
4382 		* The simple fact that isc_msix_bar is not 0 does not mean we
4383 		* we have a good value there that is known to work.
4384 		*/
4385 		msix = iflib_msix_init(ctx);
4386 	else {
4387 		scctx->isc_vectors = 1;
4388 		scctx->isc_ntxqsets = 1;
4389 		scctx->isc_nrxqsets = 1;
4390 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4391 		msix = 0;
4392 	}
4393 	/* Get memory for the station queues */
4394 	if ((err = iflib_queues_alloc(ctx))) {
4395 		device_printf(dev, "Unable to allocate queue memory\n");
4396 		goto fail;
4397 	}
4398 
4399 	if ((err = iflib_qset_structures_setup(ctx))) {
4400 		device_printf(dev, "qset structure setup failed %d\n", err);
4401 		goto fail_queues;
4402 	}
4403 
4404 	/*
4405 	 * Group taskqueues aren't properly set up until SMP is started,
4406 	 * so we disable interrupts until we can handle them post
4407 	 * SI_SUB_SMP.
4408 	 *
4409 	 * XXX: disabling interrupts doesn't actually work, at least for
4410 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4411 	 * we do null handling and depend on this not causing too large an
4412 	 * interrupt storm.
4413 	 */
4414 	IFDI_INTR_DISABLE(ctx);
4415 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4416 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4417 		goto fail_intr_free;
4418 	}
4419 	if (msix <= 1) {
4420 		rid = 0;
4421 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4422 			MPASS(msix == 1);
4423 			rid = 1;
4424 		}
4425 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4426 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4427 			goto fail_intr_free;
4428 		}
4429 	}
4430 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4431 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4432 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4433 		goto fail_detach;
4434 	}
4435 	if ((err = iflib_netmap_attach(ctx))) {
4436 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4437 		goto fail_detach;
4438 	}
4439 	*ctxp = ctx;
4440 
4441 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4442 	iflib_add_device_sysctl_post(ctx);
4443 	ctx->ifc_flags |= IFC_INIT_DONE;
4444 	CTX_UNLOCK(ctx);
4445 	return (0);
4446 fail_detach:
4447 	ether_ifdetach(ctx->ifc_ifp);
4448 fail_intr_free:
4449 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4450 		pci_release_msi(ctx->ifc_dev);
4451 fail_queues:
4452 	/* XXX free queues */
4453 fail:
4454 	IFDI_DETACH(ctx);
4455 	CTX_UNLOCK(ctx);
4456 	return (err);
4457 }
4458 
4459 int
4460 iflib_device_attach(device_t dev)
4461 {
4462 	if_ctx_t ctx;
4463 	if_shared_ctx_t sctx;
4464 
4465 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4466 		return (ENOTSUP);
4467 
4468 	pci_enable_busmaster(dev);
4469 
4470 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4471 }
4472 
4473 int
4474 iflib_device_deregister(if_ctx_t ctx)
4475 {
4476 	if_t ifp = ctx->ifc_ifp;
4477 	iflib_txq_t txq;
4478 	iflib_rxq_t rxq;
4479 	device_t dev = ctx->ifc_dev;
4480 	int i, j;
4481 	struct taskqgroup *tqg;
4482 	iflib_fl_t fl;
4483 
4484 	/* Make sure VLANS are not using driver */
4485 	if (if_vlantrunkinuse(ifp)) {
4486 		device_printf(dev,"Vlan in use, detach first\n");
4487 		return (EBUSY);
4488 	}
4489 
4490 	CTX_LOCK(ctx);
4491 	ctx->ifc_in_detach = 1;
4492 	iflib_stop(ctx);
4493 	CTX_UNLOCK(ctx);
4494 
4495 	/* Unregister VLAN events */
4496 	if (ctx->ifc_vlan_attach_event != NULL)
4497 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4498 	if (ctx->ifc_vlan_detach_event != NULL)
4499 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4500 
4501 	iflib_netmap_detach(ifp);
4502 	ether_ifdetach(ifp);
4503 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4504 	CTX_LOCK_DESTROY(ctx);
4505 	if (ctx->ifc_led_dev != NULL)
4506 		led_destroy(ctx->ifc_led_dev);
4507 	/* XXX drain any dependent tasks */
4508 	tqg = qgroup_if_io_tqg;
4509 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4510 		callout_drain(&txq->ift_timer);
4511 		if (txq->ift_task.gt_uniq != NULL)
4512 			taskqgroup_detach(tqg, &txq->ift_task);
4513 	}
4514 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4515 		if (rxq->ifr_task.gt_uniq != NULL)
4516 			taskqgroup_detach(tqg, &rxq->ifr_task);
4517 
4518 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4519 			free(fl->ifl_rx_bitmap, M_IFLIB);
4520 
4521 	}
4522 	tqg = qgroup_if_config_tqg;
4523 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4524 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4525 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4526 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4527 
4528 	IFDI_DETACH(ctx);
4529 	device_set_softc(ctx->ifc_dev, NULL);
4530 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4531 		pci_release_msi(dev);
4532 	}
4533 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4534 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4535 	}
4536 	if (ctx->ifc_msix_mem != NULL) {
4537 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4538 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4539 		ctx->ifc_msix_mem = NULL;
4540 	}
4541 
4542 	bus_generic_detach(dev);
4543 	if_free(ifp);
4544 
4545 	iflib_tx_structures_free(ctx);
4546 	iflib_rx_structures_free(ctx);
4547 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4548 		free(ctx->ifc_softc, M_IFLIB);
4549 	free(ctx, M_IFLIB);
4550 	return (0);
4551 }
4552 
4553 
4554 int
4555 iflib_device_detach(device_t dev)
4556 {
4557 	if_ctx_t ctx = device_get_softc(dev);
4558 
4559 	return (iflib_device_deregister(ctx));
4560 }
4561 
4562 int
4563 iflib_device_suspend(device_t dev)
4564 {
4565 	if_ctx_t ctx = device_get_softc(dev);
4566 
4567 	CTX_LOCK(ctx);
4568 	IFDI_SUSPEND(ctx);
4569 	CTX_UNLOCK(ctx);
4570 
4571 	return bus_generic_suspend(dev);
4572 }
4573 int
4574 iflib_device_shutdown(device_t dev)
4575 {
4576 	if_ctx_t ctx = device_get_softc(dev);
4577 
4578 	CTX_LOCK(ctx);
4579 	IFDI_SHUTDOWN(ctx);
4580 	CTX_UNLOCK(ctx);
4581 
4582 	return bus_generic_suspend(dev);
4583 }
4584 
4585 
4586 int
4587 iflib_device_resume(device_t dev)
4588 {
4589 	if_ctx_t ctx = device_get_softc(dev);
4590 	iflib_txq_t txq = ctx->ifc_txqs;
4591 
4592 	CTX_LOCK(ctx);
4593 	IFDI_RESUME(ctx);
4594 	iflib_init_locked(ctx);
4595 	CTX_UNLOCK(ctx);
4596 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4597 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4598 
4599 	return (bus_generic_resume(dev));
4600 }
4601 
4602 int
4603 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4604 {
4605 	int error;
4606 	if_ctx_t ctx = device_get_softc(dev);
4607 
4608 	CTX_LOCK(ctx);
4609 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4610 	CTX_UNLOCK(ctx);
4611 
4612 	return (error);
4613 }
4614 
4615 void
4616 iflib_device_iov_uninit(device_t dev)
4617 {
4618 	if_ctx_t ctx = device_get_softc(dev);
4619 
4620 	CTX_LOCK(ctx);
4621 	IFDI_IOV_UNINIT(ctx);
4622 	CTX_UNLOCK(ctx);
4623 }
4624 
4625 int
4626 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4627 {
4628 	int error;
4629 	if_ctx_t ctx = device_get_softc(dev);
4630 
4631 	CTX_LOCK(ctx);
4632 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4633 	CTX_UNLOCK(ctx);
4634 
4635 	return (error);
4636 }
4637 
4638 /*********************************************************************
4639  *
4640  *  MODULE FUNCTION DEFINITIONS
4641  *
4642  **********************************************************************/
4643 
4644 /*
4645  * - Start a fast taskqueue thread for each core
4646  * - Start a taskqueue for control operations
4647  */
4648 static int
4649 iflib_module_init(void)
4650 {
4651 	return (0);
4652 }
4653 
4654 static int
4655 iflib_module_event_handler(module_t mod, int what, void *arg)
4656 {
4657 	int err;
4658 
4659 	switch (what) {
4660 	case MOD_LOAD:
4661 		if ((err = iflib_module_init()) != 0)
4662 			return (err);
4663 		break;
4664 	case MOD_UNLOAD:
4665 		return (EBUSY);
4666 	default:
4667 		return (EOPNOTSUPP);
4668 	}
4669 
4670 	return (0);
4671 }
4672 
4673 /*********************************************************************
4674  *
4675  *  PUBLIC FUNCTION DEFINITIONS
4676  *     ordered as in iflib.h
4677  *
4678  **********************************************************************/
4679 
4680 
4681 static void
4682 _iflib_assert(if_shared_ctx_t sctx)
4683 {
4684 	MPASS(sctx->isc_tx_maxsize);
4685 	MPASS(sctx->isc_tx_maxsegsize);
4686 
4687 	MPASS(sctx->isc_rx_maxsize);
4688 	MPASS(sctx->isc_rx_nsegments);
4689 	MPASS(sctx->isc_rx_maxsegsize);
4690 
4691 	MPASS(sctx->isc_nrxd_min[0]);
4692 	MPASS(sctx->isc_nrxd_max[0]);
4693 	MPASS(sctx->isc_nrxd_default[0]);
4694 	MPASS(sctx->isc_ntxd_min[0]);
4695 	MPASS(sctx->isc_ntxd_max[0]);
4696 	MPASS(sctx->isc_ntxd_default[0]);
4697 }
4698 
4699 static void
4700 _iflib_pre_assert(if_softc_ctx_t scctx)
4701 {
4702 
4703 	MPASS(scctx->isc_txrx->ift_txd_encap);
4704 	MPASS(scctx->isc_txrx->ift_txd_flush);
4705 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4706 	MPASS(scctx->isc_txrx->ift_rxd_available);
4707 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4708 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4709 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4710 }
4711 
4712 static int
4713 iflib_register(if_ctx_t ctx)
4714 {
4715 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4716 	driver_t *driver = sctx->isc_driver;
4717 	device_t dev = ctx->ifc_dev;
4718 	if_t ifp;
4719 
4720 	_iflib_assert(sctx);
4721 
4722 	CTX_LOCK_INIT(ctx);
4723 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4724 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4725 	if (ifp == NULL) {
4726 		device_printf(dev, "can not allocate ifnet structure\n");
4727 		return (ENOMEM);
4728 	}
4729 
4730 	/*
4731 	 * Initialize our context's device specific methods
4732 	 */
4733 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4734 	kobj_class_compile((kobj_class_t) driver);
4735 	driver->refs++;
4736 
4737 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4738 	if_setsoftc(ifp, ctx);
4739 	if_setdev(ifp, dev);
4740 	if_setinitfn(ifp, iflib_if_init);
4741 	if_setioctlfn(ifp, iflib_if_ioctl);
4742 	if_settransmitfn(ifp, iflib_if_transmit);
4743 	if_setqflushfn(ifp, iflib_if_qflush);
4744 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4745 
4746 	ctx->ifc_vlan_attach_event =
4747 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4748 							  EVENTHANDLER_PRI_FIRST);
4749 	ctx->ifc_vlan_detach_event =
4750 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4751 							  EVENTHANDLER_PRI_FIRST);
4752 
4753 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4754 					 iflib_media_change, iflib_media_status);
4755 
4756 	return (0);
4757 }
4758 
4759 
4760 static int
4761 iflib_queues_alloc(if_ctx_t ctx)
4762 {
4763 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4764 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4765 	device_t dev = ctx->ifc_dev;
4766 	int nrxqsets = scctx->isc_nrxqsets;
4767 	int ntxqsets = scctx->isc_ntxqsets;
4768 	iflib_txq_t txq;
4769 	iflib_rxq_t rxq;
4770 	iflib_fl_t fl = NULL;
4771 	int i, j, cpu, err, txconf, rxconf;
4772 	iflib_dma_info_t ifdip;
4773 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4774 	uint32_t *txqsizes = scctx->isc_txqsizes;
4775 	uint8_t nrxqs = sctx->isc_nrxqs;
4776 	uint8_t ntxqs = sctx->isc_ntxqs;
4777 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4778 	caddr_t *vaddrs;
4779 	uint64_t *paddrs;
4780 
4781 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4782 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4783 
4784 /* Allocate the TX ring struct memory */
4785 	if (!(ctx->ifc_txqs =
4786 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4787 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4788 		device_printf(dev, "Unable to allocate TX ring memory\n");
4789 		err = ENOMEM;
4790 		goto fail;
4791 	}
4792 
4793 	/* Now allocate the RX */
4794 	if (!(ctx->ifc_rxqs =
4795 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4796 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4797 		device_printf(dev, "Unable to allocate RX ring memory\n");
4798 		err = ENOMEM;
4799 		goto rx_fail;
4800 	}
4801 
4802 	txq = ctx->ifc_txqs;
4803 	rxq = ctx->ifc_rxqs;
4804 
4805 	/*
4806 	 * XXX handle allocation failure
4807 	 */
4808 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4809 		/* Set up some basics */
4810 
4811 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4812 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4813 			err = ENOMEM;
4814 			goto err_tx_desc;
4815 		}
4816 		txq->ift_ifdi = ifdip;
4817 		for (j = 0; j < ntxqs; j++, ifdip++) {
4818 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4819 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4820 				err = ENOMEM;
4821 				goto err_tx_desc;
4822 			}
4823 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4824 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4825 		}
4826 		txq->ift_ctx = ctx;
4827 		txq->ift_id = i;
4828 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4829 			txq->ift_br_offset = 1;
4830 		} else {
4831 			txq->ift_br_offset = 0;
4832 		}
4833 		/* XXX fix this */
4834 		txq->ift_timer.c_cpu = cpu;
4835 
4836 		if (iflib_txsd_alloc(txq)) {
4837 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4838 			err = ENOMEM;
4839 			goto err_tx_desc;
4840 		}
4841 
4842 		/* Initialize the TX lock */
4843 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4844 		    device_get_nameunit(dev), txq->ift_id);
4845 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4846 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4847 
4848 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4849 			 device_get_nameunit(dev), txq->ift_id);
4850 
4851 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4852 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4853 		if (err) {
4854 			/* XXX free any allocated rings */
4855 			device_printf(dev, "Unable to allocate buf_ring\n");
4856 			goto err_tx_desc;
4857 		}
4858 	}
4859 
4860 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4861 		/* Set up some basics */
4862 
4863 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4864 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4865 			err = ENOMEM;
4866 			goto err_tx_desc;
4867 		}
4868 
4869 		rxq->ifr_ifdi = ifdip;
4870 		/* XXX this needs to be changed if #rx queues != #tx queues */
4871 		rxq->ifr_ntxqirq = 1;
4872 		rxq->ifr_txqid[0] = i;
4873 		for (j = 0; j < nrxqs; j++, ifdip++) {
4874 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4875 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4876 				err = ENOMEM;
4877 				goto err_tx_desc;
4878 			}
4879 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4880 		}
4881 		rxq->ifr_ctx = ctx;
4882 		rxq->ifr_id = i;
4883 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4884 			rxq->ifr_fl_offset = 1;
4885 		} else {
4886 			rxq->ifr_fl_offset = 0;
4887 		}
4888 		rxq->ifr_nfl = nfree_lists;
4889 		if (!(fl =
4890 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4891 			device_printf(dev, "Unable to allocate free list memory\n");
4892 			err = ENOMEM;
4893 			goto err_tx_desc;
4894 		}
4895 		rxq->ifr_fl = fl;
4896 		for (j = 0; j < nfree_lists; j++) {
4897 			fl[j].ifl_rxq = rxq;
4898 			fl[j].ifl_id = j;
4899 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4900 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4901 		}
4902         /* Allocate receive buffers for the ring*/
4903 		if (iflib_rxsd_alloc(rxq)) {
4904 			device_printf(dev,
4905 			    "Critical Failure setting up receive buffers\n");
4906 			err = ENOMEM;
4907 			goto err_rx_desc;
4908 		}
4909 
4910 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4911 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4912 	}
4913 
4914 	/* TXQs */
4915 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4916 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4917 	for (i = 0; i < ntxqsets; i++) {
4918 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4919 
4920 		for (j = 0; j < ntxqs; j++, di++) {
4921 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4922 			paddrs[i*ntxqs + j] = di->idi_paddr;
4923 		}
4924 	}
4925 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4926 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4927 		iflib_tx_structures_free(ctx);
4928 		free(vaddrs, M_IFLIB);
4929 		free(paddrs, M_IFLIB);
4930 		goto err_rx_desc;
4931 	}
4932 	free(vaddrs, M_IFLIB);
4933 	free(paddrs, M_IFLIB);
4934 
4935 	/* RXQs */
4936 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4937 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4938 	for (i = 0; i < nrxqsets; i++) {
4939 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4940 
4941 		for (j = 0; j < nrxqs; j++, di++) {
4942 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4943 			paddrs[i*nrxqs + j] = di->idi_paddr;
4944 		}
4945 	}
4946 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4947 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4948 		iflib_tx_structures_free(ctx);
4949 		free(vaddrs, M_IFLIB);
4950 		free(paddrs, M_IFLIB);
4951 		goto err_rx_desc;
4952 	}
4953 	free(vaddrs, M_IFLIB);
4954 	free(paddrs, M_IFLIB);
4955 
4956 	return (0);
4957 
4958 /* XXX handle allocation failure changes */
4959 err_rx_desc:
4960 err_tx_desc:
4961 rx_fail:
4962 	if (ctx->ifc_rxqs != NULL)
4963 		free(ctx->ifc_rxqs, M_IFLIB);
4964 	ctx->ifc_rxqs = NULL;
4965 	if (ctx->ifc_txqs != NULL)
4966 		free(ctx->ifc_txqs, M_IFLIB);
4967 	ctx->ifc_txqs = NULL;
4968 fail:
4969 	return (err);
4970 }
4971 
4972 static int
4973 iflib_tx_structures_setup(if_ctx_t ctx)
4974 {
4975 	iflib_txq_t txq = ctx->ifc_txqs;
4976 	int i;
4977 
4978 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4979 		iflib_txq_setup(txq);
4980 
4981 	return (0);
4982 }
4983 
4984 static void
4985 iflib_tx_structures_free(if_ctx_t ctx)
4986 {
4987 	iflib_txq_t txq = ctx->ifc_txqs;
4988 	int i, j;
4989 
4990 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4991 		iflib_txq_destroy(txq);
4992 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4993 			iflib_dma_free(&txq->ift_ifdi[j]);
4994 	}
4995 	free(ctx->ifc_txqs, M_IFLIB);
4996 	ctx->ifc_txqs = NULL;
4997 	IFDI_QUEUES_FREE(ctx);
4998 }
4999 
5000 /*********************************************************************
5001  *
5002  *  Initialize all receive rings.
5003  *
5004  **********************************************************************/
5005 static int
5006 iflib_rx_structures_setup(if_ctx_t ctx)
5007 {
5008 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5009 	int q;
5010 #if defined(INET6) || defined(INET)
5011 	int i, err;
5012 #endif
5013 
5014 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5015 #if defined(INET6) || defined(INET)
5016 		tcp_lro_free(&rxq->ifr_lc);
5017 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5018 		    TCP_LRO_ENTRIES, min(1024,
5019 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5020 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5021 			goto fail;
5022 		}
5023 		rxq->ifr_lro_enabled = TRUE;
5024 #endif
5025 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5026 	}
5027 	return (0);
5028 #if defined(INET6) || defined(INET)
5029 fail:
5030 	/*
5031 	 * Free RX software descriptors allocated so far, we will only handle
5032 	 * the rings that completed, the failing case will have
5033 	 * cleaned up for itself. 'q' failed, so its the terminus.
5034 	 */
5035 	rxq = ctx->ifc_rxqs;
5036 	for (i = 0; i < q; ++i, rxq++) {
5037 		iflib_rx_sds_free(rxq);
5038 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5039 	}
5040 	return (err);
5041 #endif
5042 }
5043 
5044 /*********************************************************************
5045  *
5046  *  Free all receive rings.
5047  *
5048  **********************************************************************/
5049 static void
5050 iflib_rx_structures_free(if_ctx_t ctx)
5051 {
5052 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5053 
5054 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5055 		iflib_rx_sds_free(rxq);
5056 	}
5057 }
5058 
5059 static int
5060 iflib_qset_structures_setup(if_ctx_t ctx)
5061 {
5062 	int err;
5063 
5064 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5065 		return (err);
5066 
5067 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5068 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5069 		iflib_tx_structures_free(ctx);
5070 		iflib_rx_structures_free(ctx);
5071 	}
5072 	return (err);
5073 }
5074 
5075 int
5076 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5077 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5078 {
5079 
5080 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5081 }
5082 
5083 #ifdef SMP
5084 static int
5085 find_nth(if_ctx_t ctx, int qid)
5086 {
5087 	cpuset_t cpus;
5088 	int i, cpuid, eqid, count;
5089 
5090 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5091 	count = CPU_COUNT(&cpus);
5092 	eqid = qid % count;
5093 	/* clear up to the qid'th bit */
5094 	for (i = 0; i < eqid; i++) {
5095 		cpuid = CPU_FFS(&cpus);
5096 		MPASS(cpuid != 0);
5097 		CPU_CLR(cpuid-1, &cpus);
5098 	}
5099 	cpuid = CPU_FFS(&cpus);
5100 	MPASS(cpuid != 0);
5101 	return (cpuid-1);
5102 }
5103 
5104 #ifdef SCHED_ULE
5105 extern struct cpu_group *cpu_top;              /* CPU topology */
5106 
5107 static int
5108 find_child_with_core(int cpu, struct cpu_group *grp)
5109 {
5110 	int i;
5111 
5112 	if (grp->cg_children == 0)
5113 		return -1;
5114 
5115 	MPASS(grp->cg_child);
5116 	for (i = 0; i < grp->cg_children; i++) {
5117 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5118 			return i;
5119 	}
5120 
5121 	return -1;
5122 }
5123 
5124 /*
5125  * Find the nth "close" core to the specified core
5126  * "close" is defined as the deepest level that shares
5127  * at least an L2 cache.  With threads, this will be
5128  * threads on the same core.  If the sahred cache is L3
5129  * or higher, simply returns the same core.
5130  */
5131 static int
5132 find_close_core(int cpu, int core_offset)
5133 {
5134 	struct cpu_group *grp;
5135 	int i;
5136 	int fcpu;
5137 	cpuset_t cs;
5138 
5139 	grp = cpu_top;
5140 	if (grp == NULL)
5141 		return cpu;
5142 	i = 0;
5143 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5144 		/* If the child only has one cpu, don't descend */
5145 		if (grp->cg_child[i].cg_count <= 1)
5146 			break;
5147 		grp = &grp->cg_child[i];
5148 	}
5149 
5150 	/* If they don't share at least an L2 cache, use the same CPU */
5151 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5152 		return cpu;
5153 
5154 	/* Now pick one */
5155 	CPU_COPY(&grp->cg_mask, &cs);
5156 
5157 	/* Add the selected CPU offset to core offset. */
5158 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5159 		if (fcpu - 1 == cpu)
5160 			break;
5161 		CPU_CLR(fcpu - 1, &cs);
5162 	}
5163 	MPASS(fcpu);
5164 
5165 	core_offset += i;
5166 
5167 	CPU_COPY(&grp->cg_mask, &cs);
5168 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5169 		MPASS(CPU_FFS(&cs));
5170 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5171 	}
5172 	MPASS(CPU_FFS(&cs));
5173 	return CPU_FFS(&cs) - 1;
5174 }
5175 #else
5176 static int
5177 find_close_core(int cpu, int core_offset __unused)
5178 {
5179 	return cpu;
5180 }
5181 #endif
5182 
5183 static int
5184 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5185 {
5186 	switch (type) {
5187 	case IFLIB_INTR_TX:
5188 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5189 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5190 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5191 	case IFLIB_INTR_RX:
5192 	case IFLIB_INTR_RXTX:
5193 		/* RX queues get the specified core */
5194 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5195 	default:
5196 		return -1;
5197 	}
5198 }
5199 #else
5200 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5201 #define find_close_core(cpuid, tid)	CPU_FIRST()
5202 #define find_nth(ctx, gid)		CPU_FIRST()
5203 #endif
5204 
5205 /* Just to avoid copy/paste */
5206 static inline int
5207 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5208     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5209 {
5210 	int cpuid;
5211 	int err, tid;
5212 
5213 	cpuid = find_nth(ctx, qid);
5214 	tid = get_core_offset(ctx, type, qid);
5215 	MPASS(tid >= 0);
5216 	cpuid = find_close_core(cpuid, tid);
5217 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5218 	if (err) {
5219 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5220 		return (err);
5221 	}
5222 #ifdef notyet
5223 	if (cpuid > ctx->ifc_cpuid_highest)
5224 		ctx->ifc_cpuid_highest = cpuid;
5225 #endif
5226 	return 0;
5227 }
5228 
5229 int
5230 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5231 						iflib_intr_type_t type, driver_filter_t *filter,
5232 						void *filter_arg, int qid, char *name)
5233 {
5234 	struct grouptask *gtask;
5235 	struct taskqgroup *tqg;
5236 	iflib_filter_info_t info;
5237 	gtask_fn_t *fn;
5238 	int tqrid, err;
5239 	driver_filter_t *intr_fast;
5240 	void *q;
5241 
5242 	info = &ctx->ifc_filter_info;
5243 	tqrid = rid;
5244 
5245 	switch (type) {
5246 	/* XXX merge tx/rx for netmap? */
5247 	case IFLIB_INTR_TX:
5248 		q = &ctx->ifc_txqs[qid];
5249 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5250 		gtask = &ctx->ifc_txqs[qid].ift_task;
5251 		tqg = qgroup_if_io_tqg;
5252 		fn = _task_fn_tx;
5253 		intr_fast = iflib_fast_intr;
5254 		GROUPTASK_INIT(gtask, 0, fn, q);
5255 		break;
5256 	case IFLIB_INTR_RX:
5257 		q = &ctx->ifc_rxqs[qid];
5258 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5259 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5260 		tqg = qgroup_if_io_tqg;
5261 		fn = _task_fn_rx;
5262 		intr_fast = iflib_fast_intr;
5263 		GROUPTASK_INIT(gtask, 0, fn, q);
5264 		break;
5265 	case IFLIB_INTR_RXTX:
5266 		q = &ctx->ifc_rxqs[qid];
5267 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5268 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5269 		tqg = qgroup_if_io_tqg;
5270 		fn = _task_fn_rx;
5271 		intr_fast = iflib_fast_intr_rxtx;
5272 		GROUPTASK_INIT(gtask, 0, fn, q);
5273 		break;
5274 	case IFLIB_INTR_ADMIN:
5275 		q = ctx;
5276 		tqrid = -1;
5277 		info = &ctx->ifc_filter_info;
5278 		gtask = &ctx->ifc_admin_task;
5279 		tqg = qgroup_if_config_tqg;
5280 		fn = _task_fn_admin;
5281 		intr_fast = iflib_fast_intr_ctx;
5282 		break;
5283 	default:
5284 		panic("unknown net intr type");
5285 	}
5286 
5287 	info->ifi_filter = filter;
5288 	info->ifi_filter_arg = filter_arg;
5289 	info->ifi_task = gtask;
5290 	info->ifi_ctx = q;
5291 
5292 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5293 	if (err != 0) {
5294 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5295 		return (err);
5296 	}
5297 	if (type == IFLIB_INTR_ADMIN)
5298 		return (0);
5299 
5300 	if (tqrid != -1) {
5301 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5302 		if (err)
5303 			return (err);
5304 	} else {
5305 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5306 	}
5307 
5308 	return (0);
5309 }
5310 
5311 void
5312 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5313 {
5314 	struct grouptask *gtask;
5315 	struct taskqgroup *tqg;
5316 	gtask_fn_t *fn;
5317 	void *q;
5318 	int irq_num = -1;
5319 	int err;
5320 
5321 	switch (type) {
5322 	case IFLIB_INTR_TX:
5323 		q = &ctx->ifc_txqs[qid];
5324 		gtask = &ctx->ifc_txqs[qid].ift_task;
5325 		tqg = qgroup_if_io_tqg;
5326 		fn = _task_fn_tx;
5327 		if (irq != NULL)
5328 			irq_num = rman_get_start(irq->ii_res);
5329 		break;
5330 	case IFLIB_INTR_RX:
5331 		q = &ctx->ifc_rxqs[qid];
5332 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5333 		tqg = qgroup_if_io_tqg;
5334 		fn = _task_fn_rx;
5335 		if (irq != NULL)
5336 			irq_num = rman_get_start(irq->ii_res);
5337 		break;
5338 	case IFLIB_INTR_IOV:
5339 		q = ctx;
5340 		gtask = &ctx->ifc_vflr_task;
5341 		tqg = qgroup_if_config_tqg;
5342 		fn = _task_fn_iov;
5343 		break;
5344 	default:
5345 		panic("unknown net intr type");
5346 	}
5347 	GROUPTASK_INIT(gtask, 0, fn, q);
5348 	if (irq_num != -1) {
5349 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5350 		if (err)
5351 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5352 	}
5353 	else {
5354 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5355 	}
5356 }
5357 
5358 void
5359 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5360 {
5361 	if (irq->ii_tag)
5362 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5363 
5364 	if (irq->ii_res)
5365 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5366 }
5367 
5368 static int
5369 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5370 {
5371 	iflib_txq_t txq = ctx->ifc_txqs;
5372 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5373 	if_irq_t irq = &ctx->ifc_legacy_irq;
5374 	iflib_filter_info_t info;
5375 	struct grouptask *gtask;
5376 	struct taskqgroup *tqg;
5377 	gtask_fn_t *fn;
5378 	int tqrid;
5379 	void *q;
5380 	int err;
5381 
5382 	q = &ctx->ifc_rxqs[0];
5383 	info = &rxq[0].ifr_filter_info;
5384 	gtask = &rxq[0].ifr_task;
5385 	tqg = qgroup_if_io_tqg;
5386 	tqrid = irq->ii_rid = *rid;
5387 	fn = _task_fn_rx;
5388 
5389 	ctx->ifc_flags |= IFC_LEGACY;
5390 	info->ifi_filter = filter;
5391 	info->ifi_filter_arg = filter_arg;
5392 	info->ifi_task = gtask;
5393 	info->ifi_ctx = ctx;
5394 
5395 	/* We allocate a single interrupt resource */
5396 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5397 		return (err);
5398 	GROUPTASK_INIT(gtask, 0, fn, q);
5399 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5400 
5401 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5402 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5403 	return (0);
5404 }
5405 
5406 void
5407 iflib_led_create(if_ctx_t ctx)
5408 {
5409 
5410 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5411 	    device_get_nameunit(ctx->ifc_dev));
5412 }
5413 
5414 void
5415 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5416 {
5417 
5418 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5419 }
5420 
5421 void
5422 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5423 {
5424 
5425 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5426 }
5427 
5428 void
5429 iflib_admin_intr_deferred(if_ctx_t ctx)
5430 {
5431 #ifdef INVARIANTS
5432 	struct grouptask *gtask;
5433 
5434 	gtask = &ctx->ifc_admin_task;
5435 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5436 #endif
5437 
5438 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5439 }
5440 
5441 void
5442 iflib_iov_intr_deferred(if_ctx_t ctx)
5443 {
5444 
5445 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5446 }
5447 
5448 void
5449 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5450 {
5451 
5452 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5453 }
5454 
5455 void
5456 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5457 	const char *name)
5458 {
5459 
5460 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5461 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5462 }
5463 
5464 void
5465 iflib_config_gtask_deinit(struct grouptask *gtask)
5466 {
5467 
5468 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5469 }
5470 
5471 void
5472 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5473 {
5474 	if_t ifp = ctx->ifc_ifp;
5475 	iflib_txq_t txq = ctx->ifc_txqs;
5476 
5477 	if_setbaudrate(ifp, baudrate);
5478 	if (baudrate >= IF_Gbps(10)) {
5479 		STATE_LOCK(ctx);
5480 		ctx->ifc_flags |= IFC_PREFETCH;
5481 		STATE_UNLOCK(ctx);
5482 	}
5483 	/* If link down, disable watchdog */
5484 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5485 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5486 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5487 	}
5488 	ctx->ifc_link_state = link_state;
5489 	if_link_state_change(ifp, link_state);
5490 }
5491 
5492 static int
5493 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5494 {
5495 	int credits;
5496 #ifdef INVARIANTS
5497 	int credits_pre = txq->ift_cidx_processed;
5498 #endif
5499 
5500 	if (ctx->isc_txd_credits_update == NULL)
5501 		return (0);
5502 
5503 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5504 		return (0);
5505 
5506 	txq->ift_processed += credits;
5507 	txq->ift_cidx_processed += credits;
5508 
5509 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5510 	if (txq->ift_cidx_processed >= txq->ift_size)
5511 		txq->ift_cidx_processed -= txq->ift_size;
5512 	return (credits);
5513 }
5514 
5515 static int
5516 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5517 {
5518 
5519 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5520 	    budget));
5521 }
5522 
5523 void
5524 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5525 	const char *description, if_int_delay_info_t info,
5526 	int offset, int value)
5527 {
5528 	info->iidi_ctx = ctx;
5529 	info->iidi_offset = offset;
5530 	info->iidi_value = value;
5531 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5532 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5533 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5534 	    info, 0, iflib_sysctl_int_delay, "I", description);
5535 }
5536 
5537 struct sx *
5538 iflib_ctx_lock_get(if_ctx_t ctx)
5539 {
5540 
5541 	return (&ctx->ifc_ctx_sx);
5542 }
5543 
5544 static int
5545 iflib_msix_init(if_ctx_t ctx)
5546 {
5547 	device_t dev = ctx->ifc_dev;
5548 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5549 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5550 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5551 	int iflib_num_tx_queues, iflib_num_rx_queues;
5552 	int err, admincnt, bar;
5553 
5554 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5555 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5556 
5557 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5558 
5559 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5560 	admincnt = sctx->isc_admin_intrcnt;
5561 	/* Override by global tuneable */
5562 	{
5563 		int i;
5564 		size_t len = sizeof(i);
5565 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5566 		if (err == 0) {
5567 			if (i == 0)
5568 				goto msi;
5569 		}
5570 		else {
5571 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5572 		}
5573 	}
5574 	/* Override by tuneable */
5575 	if (scctx->isc_disable_msix)
5576 		goto msi;
5577 
5578 	/*
5579 	** When used in a virtualized environment
5580 	** PCI BUSMASTER capability may not be set
5581 	** so explicity set it here and rewrite
5582 	** the ENABLE in the MSIX control register
5583 	** at this point to cause the host to
5584 	** successfully initialize us.
5585 	*/
5586 	{
5587 		int msix_ctrl, rid;
5588 
5589  		pci_enable_busmaster(dev);
5590 		rid = 0;
5591 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5592 			rid += PCIR_MSIX_CTRL;
5593 			msix_ctrl = pci_read_config(dev, rid, 2);
5594 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5595 			pci_write_config(dev, rid, msix_ctrl, 2);
5596 		} else {
5597 			device_printf(dev, "PCIY_MSIX capability not found; "
5598 			                   "or rid %d == 0.\n", rid);
5599 			goto msi;
5600 		}
5601 	}
5602 
5603 	/*
5604 	 * bar == -1 => "trust me I know what I'm doing"
5605 	 * Some drivers are for hardware that is so shoddily
5606 	 * documented that no one knows which bars are which
5607 	 * so the developer has to map all bars. This hack
5608 	 * allows shoddy garbage to use msix in this framework.
5609 	 */
5610 	if (bar != -1) {
5611 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5612 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5613 		if (ctx->ifc_msix_mem == NULL) {
5614 			/* May not be enabled */
5615 			device_printf(dev, "Unable to map MSIX table \n");
5616 			goto msi;
5617 		}
5618 	}
5619 	/* First try MSI/X */
5620 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5621 		device_printf(dev, "System has MSIX disabled \n");
5622 		bus_release_resource(dev, SYS_RES_MEMORY,
5623 		    bar, ctx->ifc_msix_mem);
5624 		ctx->ifc_msix_mem = NULL;
5625 		goto msi;
5626 	}
5627 #if IFLIB_DEBUG
5628 	/* use only 1 qset in debug mode */
5629 	queuemsgs = min(msgs - admincnt, 1);
5630 #else
5631 	queuemsgs = msgs - admincnt;
5632 #endif
5633 #ifdef RSS
5634 	queues = imin(queuemsgs, rss_getnumbuckets());
5635 #else
5636 	queues = queuemsgs;
5637 #endif
5638 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5639 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5640 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5641 #ifdef  RSS
5642 	/* If we're doing RSS, clamp at the number of RSS buckets */
5643 	if (queues > rss_getnumbuckets())
5644 		queues = rss_getnumbuckets();
5645 #endif
5646 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5647 		rx_queues = iflib_num_rx_queues;
5648 	else
5649 		rx_queues = queues;
5650 
5651 	if (rx_queues > scctx->isc_nrxqsets)
5652 		rx_queues = scctx->isc_nrxqsets;
5653 
5654 	/*
5655 	 * We want this to be all logical CPUs by default
5656 	 */
5657 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5658 		tx_queues = iflib_num_tx_queues;
5659 	else
5660 		tx_queues = mp_ncpus;
5661 
5662 	if (tx_queues > scctx->isc_ntxqsets)
5663 		tx_queues = scctx->isc_ntxqsets;
5664 
5665 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5666 #ifdef INVARIANTS
5667 		if (tx_queues != rx_queues)
5668 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5669 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5670 #endif
5671 		tx_queues = min(rx_queues, tx_queues);
5672 		rx_queues = min(rx_queues, tx_queues);
5673 	}
5674 
5675 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5676 
5677 	vectors = rx_queues + admincnt;
5678 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5679 		device_printf(dev,
5680 					  "Using MSIX interrupts with %d vectors\n", vectors);
5681 		scctx->isc_vectors = vectors;
5682 		scctx->isc_nrxqsets = rx_queues;
5683 		scctx->isc_ntxqsets = tx_queues;
5684 		scctx->isc_intr = IFLIB_INTR_MSIX;
5685 
5686 		return (vectors);
5687 	} else {
5688 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5689 	}
5690 msi:
5691 	vectors = pci_msi_count(dev);
5692 	scctx->isc_nrxqsets = 1;
5693 	scctx->isc_ntxqsets = 1;
5694 	scctx->isc_vectors = vectors;
5695 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5696 		device_printf(dev,"Using an MSI interrupt\n");
5697 		scctx->isc_intr = IFLIB_INTR_MSI;
5698 	} else {
5699 		device_printf(dev,"Using a Legacy interrupt\n");
5700 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5701 	}
5702 
5703 	return (vectors);
5704 }
5705 
5706 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5707 
5708 static int
5709 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5710 {
5711 	int rc;
5712 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5713 	struct sbuf *sb;
5714 	char *ring_state = "UNKNOWN";
5715 
5716 	/* XXX needed ? */
5717 	rc = sysctl_wire_old_buffer(req, 0);
5718 	MPASS(rc == 0);
5719 	if (rc != 0)
5720 		return (rc);
5721 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5722 	MPASS(sb != NULL);
5723 	if (sb == NULL)
5724 		return (ENOMEM);
5725 	if (state[3] <= 3)
5726 		ring_state = ring_states[state[3]];
5727 
5728 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5729 		    state[0], state[1], state[2], ring_state);
5730 	rc = sbuf_finish(sb);
5731 	sbuf_delete(sb);
5732         return(rc);
5733 }
5734 
5735 enum iflib_ndesc_handler {
5736 	IFLIB_NTXD_HANDLER,
5737 	IFLIB_NRXD_HANDLER,
5738 };
5739 
5740 static int
5741 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5742 {
5743 	if_ctx_t ctx = (void *)arg1;
5744 	enum iflib_ndesc_handler type = arg2;
5745 	char buf[256] = {0};
5746 	qidx_t *ndesc;
5747 	char *p, *next;
5748 	int nqs, rc, i;
5749 
5750 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5751 
5752 	nqs = 8;
5753 	switch(type) {
5754 	case IFLIB_NTXD_HANDLER:
5755 		ndesc = ctx->ifc_sysctl_ntxds;
5756 		if (ctx->ifc_sctx)
5757 			nqs = ctx->ifc_sctx->isc_ntxqs;
5758 		break;
5759 	case IFLIB_NRXD_HANDLER:
5760 		ndesc = ctx->ifc_sysctl_nrxds;
5761 		if (ctx->ifc_sctx)
5762 			nqs = ctx->ifc_sctx->isc_nrxqs;
5763 		break;
5764 	default:
5765 			panic("unhandled type");
5766 	}
5767 	if (nqs == 0)
5768 		nqs = 8;
5769 
5770 	for (i=0; i<8; i++) {
5771 		if (i >= nqs)
5772 			break;
5773 		if (i)
5774 			strcat(buf, ",");
5775 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5776 	}
5777 
5778 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5779 	if (rc || req->newptr == NULL)
5780 		return rc;
5781 
5782 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5783 	    i++, p = strsep(&next, " ,")) {
5784 		ndesc[i] = strtoul(p, NULL, 10);
5785 	}
5786 
5787 	return(rc);
5788 }
5789 
5790 #define NAME_BUFLEN 32
5791 static void
5792 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5793 {
5794         device_t dev = iflib_get_dev(ctx);
5795 	struct sysctl_oid_list *child, *oid_list;
5796 	struct sysctl_ctx_list *ctx_list;
5797 	struct sysctl_oid *node;
5798 
5799 	ctx_list = device_get_sysctl_ctx(dev);
5800 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5801 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5802 						      CTLFLAG_RD, NULL, "IFLIB fields");
5803 	oid_list = SYSCTL_CHILDREN(node);
5804 
5805 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5806 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5807 		       "driver version");
5808 
5809 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5810 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5811 			"# of txqs to use, 0 => use default #");
5812 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5813 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5814 			"# of rxqs to use, 0 => use default #");
5815 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5816 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5817                        "permit #txq != #rxq");
5818 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5819                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5820                       "disable MSIX (default 0)");
5821 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5822 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5823                        "set the rx budget");
5824 
5825 	/* XXX change for per-queue sizes */
5826 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5827 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5828                        mp_ndesc_handler, "A",
5829                        "list of # of tx descriptors to use, 0 = use default #");
5830 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5831 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5832                        mp_ndesc_handler, "A",
5833                        "list of # of rx descriptors to use, 0 = use default #");
5834 }
5835 
5836 static void
5837 iflib_add_device_sysctl_post(if_ctx_t ctx)
5838 {
5839 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5840 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5841         device_t dev = iflib_get_dev(ctx);
5842 	struct sysctl_oid_list *child;
5843 	struct sysctl_ctx_list *ctx_list;
5844 	iflib_fl_t fl;
5845 	iflib_txq_t txq;
5846 	iflib_rxq_t rxq;
5847 	int i, j;
5848 	char namebuf[NAME_BUFLEN];
5849 	char *qfmt;
5850 	struct sysctl_oid *queue_node, *fl_node, *node;
5851 	struct sysctl_oid_list *queue_list, *fl_list;
5852 	ctx_list = device_get_sysctl_ctx(dev);
5853 
5854 	node = ctx->ifc_sysctl_node;
5855 	child = SYSCTL_CHILDREN(node);
5856 
5857 	if (scctx->isc_ntxqsets > 100)
5858 		qfmt = "txq%03d";
5859 	else if (scctx->isc_ntxqsets > 10)
5860 		qfmt = "txq%02d";
5861 	else
5862 		qfmt = "txq%d";
5863 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5864 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5865 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5866 					     CTLFLAG_RD, NULL, "Queue Name");
5867 		queue_list = SYSCTL_CHILDREN(queue_node);
5868 #if MEMORY_LOGGING
5869 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5870 				CTLFLAG_RD,
5871 				&txq->ift_dequeued, "total mbufs freed");
5872 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5873 				CTLFLAG_RD,
5874 				&txq->ift_enqueued, "total mbufs enqueued");
5875 #endif
5876 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5877 				   CTLFLAG_RD,
5878 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5879 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5880 				   CTLFLAG_RD,
5881 				   &txq->ift_pullups, "# of times m_pullup was called");
5882 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5883 				   CTLFLAG_RD,
5884 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5885 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5886 				   CTLFLAG_RD,
5887 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5888 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5889 				   CTLFLAG_RD,
5890 				   &txq->ift_map_failed, "# of times dma map failed");
5891 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5892 				   CTLFLAG_RD,
5893 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5894 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5895 				   CTLFLAG_RD,
5896 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5897 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5898 				   CTLFLAG_RD,
5899 				   &txq->ift_pidx, 1, "Producer Index");
5900 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5901 				   CTLFLAG_RD,
5902 				   &txq->ift_cidx, 1, "Consumer Index");
5903 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5904 				   CTLFLAG_RD,
5905 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5906 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5907 				   CTLFLAG_RD,
5908 				   &txq->ift_in_use, 1, "descriptors in use");
5909 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5910 				   CTLFLAG_RD,
5911 				   &txq->ift_processed, "descriptors procesed for clean");
5912 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5913 				   CTLFLAG_RD,
5914 				   &txq->ift_cleaned, "total cleaned");
5915 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5916 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5917 				0, mp_ring_state_handler, "A", "soft ring state");
5918 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5919 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5920 				       "# of enqueues to the mp_ring for this queue");
5921 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5922 				       CTLFLAG_RD, &txq->ift_br->drops,
5923 				       "# of drops in the mp_ring for this queue");
5924 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5925 				       CTLFLAG_RD, &txq->ift_br->starts,
5926 				       "# of normal consumer starts in the mp_ring for this queue");
5927 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5928 				       CTLFLAG_RD, &txq->ift_br->stalls,
5929 					       "# of consumer stalls in the mp_ring for this queue");
5930 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5931 			       CTLFLAG_RD, &txq->ift_br->restarts,
5932 				       "# of consumer restarts in the mp_ring for this queue");
5933 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5934 				       CTLFLAG_RD, &txq->ift_br->abdications,
5935 				       "# of consumer abdications in the mp_ring for this queue");
5936 	}
5937 
5938 	if (scctx->isc_nrxqsets > 100)
5939 		qfmt = "rxq%03d";
5940 	else if (scctx->isc_nrxqsets > 10)
5941 		qfmt = "rxq%02d";
5942 	else
5943 		qfmt = "rxq%d";
5944 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5945 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5946 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5947 					     CTLFLAG_RD, NULL, "Queue Name");
5948 		queue_list = SYSCTL_CHILDREN(queue_node);
5949 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5950 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5951 				       CTLFLAG_RD,
5952 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5953 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5954 				       CTLFLAG_RD,
5955 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5956 		}
5957 
5958 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5959 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5960 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5961 						     CTLFLAG_RD, NULL, "freelist Name");
5962 			fl_list = SYSCTL_CHILDREN(fl_node);
5963 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5964 				       CTLFLAG_RD,
5965 				       &fl->ifl_pidx, 1, "Producer Index");
5966 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5967 				       CTLFLAG_RD,
5968 				       &fl->ifl_cidx, 1, "Consumer Index");
5969 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5970 				       CTLFLAG_RD,
5971 				       &fl->ifl_credits, 1, "credits available");
5972 #if MEMORY_LOGGING
5973 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5974 					CTLFLAG_RD,
5975 					&fl->ifl_m_enqueued, "mbufs allocated");
5976 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5977 					CTLFLAG_RD,
5978 					&fl->ifl_m_dequeued, "mbufs freed");
5979 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5980 					CTLFLAG_RD,
5981 					&fl->ifl_cl_enqueued, "clusters allocated");
5982 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5983 					CTLFLAG_RD,
5984 					&fl->ifl_cl_dequeued, "clusters freed");
5985 #endif
5986 
5987 		}
5988 	}
5989 
5990 }
5991 
5992 #ifndef __NO_STRICT_ALIGNMENT
5993 static struct mbuf *
5994 iflib_fixup_rx(struct mbuf *m)
5995 {
5996 	struct mbuf *n;
5997 
5998 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5999 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6000 		m->m_data += ETHER_HDR_LEN;
6001 		n = m;
6002 	} else {
6003 		MGETHDR(n, M_NOWAIT, MT_DATA);
6004 		if (n == NULL) {
6005 			m_freem(m);
6006 			return (NULL);
6007 		}
6008 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6009 		m->m_data += ETHER_HDR_LEN;
6010 		m->m_len -= ETHER_HDR_LEN;
6011 		n->m_len = ETHER_HDR_LEN;
6012 		M_MOVE_PKTHDR(n, m);
6013 		n->m_next = m;
6014 	}
6015 	return (n);
6016 }
6017 #endif
6018