xref: /freebsd/sys/net/iflib.c (revision cbd30a72ca196976c1c700400ecd424baa1b9c16)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of iflib's software descriptor references
98  */
99 #define MEMORY_LOGGING 0
100 /*
101  * Enable mbuf vectors for compressing long mbuf chains
102  */
103 
104 /*
105  * NB:
106  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
107  *   we prefetch needs to be determined by the time spent in m_free vis a vis
108  *   the cost of a prefetch. This will of course vary based on the workload:
109  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
110  *        is quite expensive, thus suggesting very little prefetch.
111  *      - small packet forwarding which is just returning a single mbuf to
112  *        UMA will typically be very fast vis a vis the cost of a memory
113  *        access.
114  */
115 
116 
117 /*
118  * File organization:
119  *  - private structures
120  *  - iflib private utility functions
121  *  - ifnet functions
122  *  - vlan registry and other exported functions
123  *  - iflib public core functions
124  *
125  *
126  */
127 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
128 
129 struct iflib_txq;
130 typedef struct iflib_txq *iflib_txq_t;
131 struct iflib_rxq;
132 typedef struct iflib_rxq *iflib_rxq_t;
133 struct iflib_fl;
134 typedef struct iflib_fl *iflib_fl_t;
135 
136 struct iflib_ctx;
137 
138 typedef struct iflib_filter_info {
139 	driver_filter_t *ifi_filter;
140 	void *ifi_filter_arg;
141 	struct grouptask *ifi_task;
142 	struct iflib_ctx *ifi_ctx;
143 } *iflib_filter_info_t;
144 
145 struct iflib_ctx {
146 	KOBJ_FIELDS;
147    /*
148    * Pointer to hardware driver's softc
149    */
150 	void *ifc_softc;
151 	device_t ifc_dev;
152 	if_t ifc_ifp;
153 
154 	cpuset_t ifc_cpus;
155 	if_shared_ctx_t ifc_sctx;
156 	struct if_softc_ctx ifc_softc_ctx;
157 
158 	struct mtx ifc_mtx;
159 
160 	uint16_t ifc_nhwtxqs;
161 	uint16_t ifc_nhwrxqs;
162 
163 	iflib_txq_t ifc_txqs;
164 	iflib_rxq_t ifc_rxqs;
165 	uint32_t ifc_if_flags;
166 	uint32_t ifc_flags;
167 	uint32_t ifc_max_fl_buf_size;
168 	int ifc_in_detach;
169 
170 	int ifc_link_state;
171 	int ifc_link_irq;
172 	int ifc_pause_frames;
173 	int ifc_watchdog_events;
174 	struct cdev *ifc_led_dev;
175 	struct resource *ifc_msix_mem;
176 
177 	struct if_irq ifc_legacy_irq;
178 	struct grouptask ifc_admin_task;
179 	struct grouptask ifc_vflr_task;
180 	struct iflib_filter_info ifc_filter_info;
181 	struct ifmedia	ifc_media;
182 
183 	struct sysctl_oid *ifc_sysctl_node;
184 	uint16_t ifc_sysctl_ntxqs;
185 	uint16_t ifc_sysctl_nrxqs;
186 	uint16_t ifc_sysctl_qs_eq_override;
187 
188 	uint16_t ifc_sysctl_ntxds[8];
189 	uint16_t ifc_sysctl_nrxds[8];
190 	struct if_txrx ifc_txrx;
191 #define isc_txd_encap  ifc_txrx.ift_txd_encap
192 #define isc_txd_flush  ifc_txrx.ift_txd_flush
193 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
194 #define isc_rxd_available ifc_txrx.ift_rxd_available
195 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
197 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
198 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
201 	eventhandler_tag ifc_vlan_attach_event;
202 	eventhandler_tag ifc_vlan_detach_event;
203 	uint8_t ifc_mac[ETHER_ADDR_LEN];
204 	char ifc_mtx_name[16];
205 };
206 
207 
208 void *
209 iflib_get_softc(if_ctx_t ctx)
210 {
211 
212 	return (ctx->ifc_softc);
213 }
214 
215 device_t
216 iflib_get_dev(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_dev);
220 }
221 
222 if_t
223 iflib_get_ifp(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_ifp);
227 }
228 
229 struct ifmedia *
230 iflib_get_media(if_ctx_t ctx)
231 {
232 
233 	return (&ctx->ifc_media);
234 }
235 
236 void
237 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
238 {
239 
240 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
241 }
242 
243 if_softc_ctx_t
244 iflib_get_softc_ctx(if_ctx_t ctx)
245 {
246 
247 	return (&ctx->ifc_softc_ctx);
248 }
249 
250 if_shared_ctx_t
251 iflib_get_sctx(if_ctx_t ctx)
252 {
253 
254 	return (ctx->ifc_sctx);
255 }
256 
257 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
258 
259 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
260 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
261 
262 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
263 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
264 #define RX_SW_DESC_INUSE        (1 << 3)
265 #define TX_SW_DESC_MAPPED       (1 << 4)
266 
267 typedef struct iflib_sw_rx_desc_array {
268 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
269 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
270 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
271 	uint8_t		*ifsd_flags;
272 } iflib_rxsd_array_t;
273 
274 typedef struct iflib_sw_tx_desc_array {
275 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
276 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
277 	uint8_t		*ifsd_flags;
278 } iflib_txsd_array_t;
279 
280 
281 /* magic number that should be high enough for any hardware */
282 #define IFLIB_MAX_TX_SEGS		128
283 #define IFLIB_MAX_RX_SEGS		32
284 #define IFLIB_RX_COPY_THRESH		63
285 #define IFLIB_MAX_RX_REFRESH		32
286 #define IFLIB_QUEUE_IDLE		0
287 #define IFLIB_QUEUE_HUNG		1
288 #define IFLIB_QUEUE_WORKING		2
289 
290 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */
291 #define TX_BATCH_SIZE			16
292 
293 #define IFLIB_RESTART_BUDGET		8
294 
295 #define	IFC_LEGACY		0x01
296 #define	IFC_QFLUSH		0x02
297 #define	IFC_MULTISEG		0x04
298 #define	IFC_DMAR		0x08
299 #define	IFC_SC_ALLOCATED	0x10
300 #define	IFC_INIT_DONE		0x20
301 
302 
303 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
304 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
305 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
306 struct iflib_txq {
307 	uint16_t	ift_in_use;
308 	uint16_t	ift_cidx;
309 	uint16_t	ift_cidx_processed;
310 	uint16_t	ift_pidx;
311 	uint8_t		ift_gen;
312 	uint8_t		ift_db_pending;
313 	uint8_t		ift_db_pending_queued;
314 	uint8_t		ift_npending;
315 	uint8_t		ift_br_offset;
316 	/* implicit pad */
317 	uint64_t	ift_processed;
318 	uint64_t	ift_cleaned;
319 #if MEMORY_LOGGING
320 	uint64_t	ift_enqueued;
321 	uint64_t	ift_dequeued;
322 #endif
323 	uint64_t	ift_no_tx_dma_setup;
324 	uint64_t	ift_no_desc_avail;
325 	uint64_t	ift_mbuf_defrag_failed;
326 	uint64_t	ift_mbuf_defrag;
327 	uint64_t	ift_map_failed;
328 	uint64_t	ift_txd_encap_efbig;
329 	uint64_t	ift_pullups;
330 
331 	struct mtx	ift_mtx;
332 	struct mtx	ift_db_mtx;
333 
334 	/* constant values */
335 	if_ctx_t	ift_ctx;
336 	struct ifmp_ring        **ift_br;
337 	struct grouptask	ift_task;
338 	uint16_t	ift_size;
339 	uint16_t	ift_id;
340 	struct callout	ift_timer;
341 	struct callout	ift_db_check;
342 
343 	iflib_txsd_array_t	ift_sds;
344 	uint8_t			ift_nbr;
345 	uint8_t			ift_qstatus;
346 	uint8_t			ift_active;
347 	uint8_t			ift_closed;
348 	int			ift_watchdog_time;
349 	struct iflib_filter_info ift_filter_info;
350 	bus_dma_tag_t		ift_desc_tag;
351 	bus_dma_tag_t		ift_tso_desc_tag;
352 	iflib_dma_info_t	ift_ifdi;
353 #define MTX_NAME_LEN 16
354 	char                    ift_mtx_name[MTX_NAME_LEN];
355 	char                    ift_db_mtx_name[MTX_NAME_LEN];
356 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
357 #ifdef IFLIB_DIAGNOSTICS
358 	uint64_t ift_cpu_exec_count[256];
359 #endif
360 } __aligned(CACHE_LINE_SIZE);
361 
362 struct iflib_fl {
363 	uint16_t	ifl_cidx;
364 	uint16_t	ifl_pidx;
365 	uint16_t	ifl_credits;
366 	uint8_t		ifl_gen;
367 #if MEMORY_LOGGING
368 	uint64_t	ifl_m_enqueued;
369 	uint64_t	ifl_m_dequeued;
370 	uint64_t	ifl_cl_enqueued;
371 	uint64_t	ifl_cl_dequeued;
372 #endif
373 	/* implicit pad */
374 
375 	/* constant */
376 	uint16_t	ifl_size;
377 	uint16_t	ifl_buf_size;
378 	uint16_t	ifl_cltype;
379 	uma_zone_t	ifl_zone;
380 	iflib_rxsd_array_t	ifl_sds;
381 	iflib_rxq_t	ifl_rxq;
382 	uint8_t		ifl_id;
383 	bus_dma_tag_t           ifl_desc_tag;
384 	iflib_dma_info_t	ifl_ifdi;
385 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
386 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
387 }  __aligned(CACHE_LINE_SIZE);
388 
389 static inline int
390 get_inuse(int size, int cidx, int pidx, int gen)
391 {
392 	int used;
393 
394 	if (pidx > cidx)
395 		used = pidx - cidx;
396 	else if (pidx < cidx)
397 		used = size - cidx + pidx;
398 	else if (gen == 0 && pidx == cidx)
399 		used = 0;
400 	else if (gen == 1 && pidx == cidx)
401 		used = size;
402 	else
403 		panic("bad state");
404 
405 	return (used);
406 }
407 
408 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
409 
410 #define IDXDIFF(head, tail, wrap) \
411 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
412 
413 struct iflib_rxq {
414 	/* If there is a separate completion queue -
415 	 * these are the cq cidx and pidx. Otherwise
416 	 * these are unused.
417 	 */
418 	uint16_t	ifr_size;
419 	uint16_t	ifr_cq_cidx;
420 	uint16_t	ifr_cq_pidx;
421 	uint8_t		ifr_cq_gen;
422 	uint8_t		ifr_fl_offset;
423 
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	uint16_t	ifr_id;
428 	uint8_t		ifr_lro_enabled;
429 	uint8_t		ifr_nfl;
430 	struct lro_ctrl			ifr_lc;
431 	struct grouptask        ifr_task;
432 	struct iflib_filter_info ifr_filter_info;
433 	iflib_dma_info_t		ifr_ifdi;
434 	/* dynamically allocate if any drivers need a value substantially larger than this */
435 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
436 #ifdef IFLIB_DIAGNOSTICS
437 	uint64_t ifr_cpu_exec_count[256];
438 #endif
439 }  __aligned(CACHE_LINE_SIZE);
440 
441 /*
442  * Only allow a single packet to take up most 1/nth of the tx ring
443  */
444 #define MAX_SINGLE_PACKET_FRACTION 12
445 #define IF_BAD_DMA (bus_addr_t)-1
446 
447 static int enable_msix = 1;
448 
449 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
450 
451 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
452 
453 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
454 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
455 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
456 
457 
458 #define TXDB_LOCK_INIT(txq)  mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF)
459 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx)
460 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx)
461 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx)
462 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx)
463 
464 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
465 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
466 
467 
468 /* Our boot-time initialization hook */
469 static int	iflib_module_event_handler(module_t, int, void *);
470 
471 static moduledata_t iflib_moduledata = {
472 	"iflib",
473 	iflib_module_event_handler,
474 	NULL
475 };
476 
477 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
478 MODULE_VERSION(iflib, 1);
479 
480 MODULE_DEPEND(iflib, pci, 1, 1, 1);
481 MODULE_DEPEND(iflib, ether, 1, 1, 1);
482 
483 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
484 
485 #ifndef IFLIB_DEBUG_COUNTERS
486 #ifdef INVARIANTS
487 #define IFLIB_DEBUG_COUNTERS 1
488 #else
489 #define IFLIB_DEBUG_COUNTERS 0
490 #endif /* !INVARIANTS */
491 #endif
492 
493 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
494                    "iflib driver parameters");
495 
496 /*
497  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
498  */
499 static int iflib_min_tx_latency = 0;
500 
501 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
502 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
503 
504 
505 #if IFLIB_DEBUG_COUNTERS
506 
507 static int iflib_tx_seen;
508 static int iflib_tx_sent;
509 static int iflib_tx_encap;
510 static int iflib_rx_allocs;
511 static int iflib_fl_refills;
512 static int iflib_fl_refills_large;
513 static int iflib_tx_frees;
514 
515 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
516 		   &iflib_tx_seen, 0, "# tx mbufs seen");
517 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
518 		   &iflib_tx_sent, 0, "# tx mbufs sent");
519 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
520 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
521 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
522 		   &iflib_tx_frees, 0, "# tx frees");
523 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
524 		   &iflib_rx_allocs, 0, "# rx allocations");
525 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
526 		   &iflib_fl_refills, 0, "# refills");
527 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
528 		   &iflib_fl_refills_large, 0, "# large refills");
529 
530 
531 static int iflib_txq_drain_flushing;
532 static int iflib_txq_drain_oactive;
533 static int iflib_txq_drain_notready;
534 static int iflib_txq_drain_encapfail;
535 
536 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
537 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
538 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
539 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
540 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
541 		   &iflib_txq_drain_notready, 0, "# drain notready");
542 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
543 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
544 
545 
546 static int iflib_encap_load_mbuf_fail;
547 static int iflib_encap_txq_avail_fail;
548 static int iflib_encap_txd_encap_fail;
549 
550 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
551 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
552 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
553 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
554 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
555 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
556 
557 static int iflib_task_fn_rxs;
558 static int iflib_rx_intr_enables;
559 static int iflib_fast_intrs;
560 static int iflib_intr_link;
561 static int iflib_intr_msix;
562 static int iflib_rx_unavail;
563 static int iflib_rx_ctx_inactive;
564 static int iflib_rx_zero_len;
565 static int iflib_rx_if_input;
566 static int iflib_rx_mbuf_null;
567 static int iflib_rxd_flush;
568 
569 static int iflib_verbose_debug;
570 
571 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
572 		   &iflib_intr_link, 0, "# intr link calls");
573 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
574 		   &iflib_intr_msix, 0, "# intr msix calls");
575 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
576 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
577 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
578 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
579 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
580 		   &iflib_fast_intrs, 0, "# fast_intr calls");
581 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
582 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
583 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
584 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
585 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
586 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
587 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
588 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
589 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
590 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
591 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
592 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
593 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
594 		   &iflib_verbose_debug, 0, "enable verbose debugging");
595 
596 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
597 static void
598 iflib_debug_reset(void)
599 {
600 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
601 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
602 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
603 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
604 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
605 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
606 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
607 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
608 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
609 }
610 
611 #else
612 #define DBG_COUNTER_INC(name)
613 static void iflib_debug_reset(void) {}
614 #endif
615 
616 
617 
618 #define IFLIB_DEBUG 0
619 
620 static void iflib_tx_structures_free(if_ctx_t ctx);
621 static void iflib_rx_structures_free(if_ctx_t ctx);
622 static int iflib_queues_alloc(if_ctx_t ctx);
623 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
624 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget);
625 static int iflib_qset_structures_setup(if_ctx_t ctx);
626 static int iflib_msix_init(if_ctx_t ctx);
627 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
628 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
629 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
630 static int iflib_register(if_ctx_t);
631 static void iflib_init_locked(if_ctx_t ctx);
632 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
633 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
634 static void iflib_ifmp_purge(iflib_txq_t txq);
635 static void _iflib_pre_assert(if_softc_ctx_t scctx);
636 
637 #ifdef DEV_NETMAP
638 #include <sys/selinfo.h>
639 #include <net/netmap.h>
640 #include <dev/netmap/netmap_kern.h>
641 
642 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
643 
644 /*
645  * device-specific sysctl variables:
646  *
647  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
648  *	During regular operations the CRC is stripped, but on some
649  *	hardware reception of frames not multiple of 64 is slower,
650  *	so using crcstrip=0 helps in benchmarks.
651  *
652  * iflib_rx_miss, iflib_rx_miss_bufs:
653  *	count packets that might be missed due to lost interrupts.
654  */
655 SYSCTL_DECL(_dev_netmap);
656 /*
657  * The xl driver by default strips CRCs and we do not override it.
658  */
659 
660 int iflib_crcstrip = 1;
661 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
662     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
663 
664 int iflib_rx_miss, iflib_rx_miss_bufs;
665 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
666     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
667 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
668     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
669 
670 /*
671  * Register/unregister. We are already under netmap lock.
672  * Only called on the first register or the last unregister.
673  */
674 static int
675 iflib_netmap_register(struct netmap_adapter *na, int onoff)
676 {
677 	struct ifnet *ifp = na->ifp;
678 	if_ctx_t ctx = ifp->if_softc;
679 
680 	CTX_LOCK(ctx);
681 	IFDI_INTR_DISABLE(ctx);
682 
683 	/* Tell the stack that the interface is no longer active */
684 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
685 
686 	if (!CTX_IS_VF(ctx))
687 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
688 
689 	/* enable or disable flags and callbacks in na and ifp */
690 	if (onoff) {
691 		nm_set_native_flags(na);
692 	} else {
693 		nm_clear_native_flags(na);
694 	}
695 	IFDI_INIT(ctx);
696 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
697 	CTX_UNLOCK(ctx);
698 	return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
699 }
700 
701 /*
702  * Reconcile kernel and user view of the transmit ring.
703  *
704  * All information is in the kring.
705  * Userspace wants to send packets up to the one before kring->rhead,
706  * kernel knows kring->nr_hwcur is the first unsent packet.
707  *
708  * Here we push packets out (as many as possible), and possibly
709  * reclaim buffers from previously completed transmission.
710  *
711  * The caller (netmap) guarantees that there is only one instance
712  * running at any time. Any interference with other driver
713  * methods should be handled by the individual drivers.
714  */
715 static int
716 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
717 {
718 	struct netmap_adapter *na = kring->na;
719 	struct ifnet *ifp = na->ifp;
720 	struct netmap_ring *ring = kring->ring;
721 	u_int nm_i;	/* index into the netmap ring */
722 	u_int nic_i;	/* index into the NIC ring */
723 	u_int n;
724 	u_int const lim = kring->nkr_num_slots - 1;
725 	u_int const head = kring->rhead;
726 	struct if_pkt_info pi;
727 
728 	/*
729 	 * interrupts on every tx packet are expensive so request
730 	 * them every half ring, or where NS_REPORT is set
731 	 */
732 	u_int report_frequency = kring->nkr_num_slots >> 1;
733 	/* device-specific */
734 	if_ctx_t ctx = ifp->if_softc;
735 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
736 
737 	pi.ipi_segs = txq->ift_segs;
738 	pi.ipi_qsidx = kring->ring_id;
739 	pi.ipi_ndescs = 0;
740 
741 	bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
742 					BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
743 
744 
745 	/*
746 	 * First part: process new packets to send.
747 	 * nm_i is the current index in the netmap ring,
748 	 * nic_i is the corresponding index in the NIC ring.
749 	 *
750 	 * If we have packets to send (nm_i != head)
751 	 * iterate over the netmap ring, fetch length and update
752 	 * the corresponding slot in the NIC ring. Some drivers also
753 	 * need to update the buffer's physical address in the NIC slot
754 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
755 	 *
756 	 * The netmap_reload_map() calls is especially expensive,
757 	 * even when (as in this case) the tag is 0, so do only
758 	 * when the buffer has actually changed.
759 	 *
760 	 * If possible do not set the report/intr bit on all slots,
761 	 * but only a few times per ring or when NS_REPORT is set.
762 	 *
763 	 * Finally, on 10G and faster drivers, it might be useful
764 	 * to prefetch the next slot and txr entry.
765 	 */
766 
767 	nm_i = kring->nr_hwcur;
768 	if (nm_i != head) {	/* we have new packets to send */
769 		nic_i = netmap_idx_k2n(kring, nm_i);
770 
771 		__builtin_prefetch(&ring->slot[nm_i]);
772 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
773 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
774 
775 		for (n = 0; nm_i != head; n++) {
776 			struct netmap_slot *slot = &ring->slot[nm_i];
777 			u_int len = slot->len;
778 			uint64_t paddr;
779 			void *addr = PNMB(na, slot, &paddr);
780 			int flags = (slot->flags & NS_REPORT ||
781 				nic_i == 0 || nic_i == report_frequency) ?
782 				IPI_TX_INTR : 0;
783 
784 			/* device-specific */
785 			pi.ipi_pidx = nic_i;
786 			pi.ipi_flags = flags;
787 
788 			/* Fill the slot in the NIC ring. */
789 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
790 
791 			/* prefetch for next round */
792 			__builtin_prefetch(&ring->slot[nm_i + 1]);
793 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
794 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
795 
796 			NM_CHECK_ADDR_LEN(na, addr, len);
797 
798 			if (slot->flags & NS_BUF_CHANGED) {
799 				/* buffer has changed, reload map */
800 				netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
801 			}
802 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
803 
804 			/* make sure changes to the buffer are synced */
805 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
806 							BUS_DMASYNC_PREWRITE);
807 
808 			nm_i = nm_next(nm_i, lim);
809 			nic_i = nm_next(nic_i, lim);
810 		}
811 		kring->nr_hwcur = head;
812 
813 		/* synchronize the NIC ring */
814 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
815 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
816 
817 		/* (re)start the tx unit up to slot nic_i (excluded) */
818 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
819 	}
820 
821 	/*
822 	 * Second part: reclaim buffers for completed transmissions.
823 	 */
824 	if (iflib_tx_credits_update(ctx, txq)) {
825 		/* some tx completed, increment avail */
826 		nic_i = txq->ift_cidx_processed;
827 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
828 	}
829 	return (0);
830 }
831 
832 /*
833  * Reconcile kernel and user view of the receive ring.
834  * Same as for the txsync, this routine must be efficient.
835  * The caller guarantees a single invocations, but races against
836  * the rest of the driver should be handled here.
837  *
838  * On call, kring->rhead is the first packet that userspace wants
839  * to keep, and kring->rcur is the wakeup point.
840  * The kernel has previously reported packets up to kring->rtail.
841  *
842  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
843  * of whether or not we received an interrupt.
844  */
845 static int
846 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
847 {
848 	struct netmap_adapter *na = kring->na;
849 	struct ifnet *ifp = na->ifp;
850 	struct netmap_ring *ring = kring->ring;
851 	u_int nm_i;	/* index into the netmap ring */
852 	u_int nic_i;	/* index into the NIC ring */
853 	u_int i, n;
854 	u_int const lim = kring->nkr_num_slots - 1;
855 	u_int const head = kring->rhead;
856 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
857 	struct if_rxd_info ri;
858 	/* device-specific */
859 	if_ctx_t ctx = ifp->if_softc;
860 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
861 	iflib_fl_t fl = rxq->ifr_fl;
862 	if (head > lim)
863 		return netmap_ring_reinit(kring);
864 
865 	bzero(&ri, sizeof(ri));
866 	ri.iri_qsidx = kring->ring_id;
867 	ri.iri_ifp = ctx->ifc_ifp;
868 	/* XXX check sync modes */
869 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++)
870 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
871 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
872 
873 	/*
874 	 * First part: import newly received packets.
875 	 *
876 	 * nm_i is the index of the next free slot in the netmap ring,
877 	 * nic_i is the index of the next received packet in the NIC ring,
878 	 * and they may differ in case if_init() has been called while
879 	 * in netmap mode. For the receive ring we have
880 	 *
881 	 *	nic_i = rxr->next_check;
882 	 *	nm_i = kring->nr_hwtail (previous)
883 	 * and
884 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
885 	 *
886 	 * rxr->next_check is set to 0 on a ring reinit
887 	 */
888 	if (netmap_no_pendintr || force_update) {
889 		int crclen = iflib_crcstrip ? 0 : 4;
890 		int error, avail;
891 		uint16_t slot_flags = kring->nkr_slot_flags;
892 
893 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
894 			nic_i = fl->ifl_cidx;
895 			nm_i = netmap_idx_n2k(kring, nic_i);
896 			avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX);
897 			for (n = 0; avail > 0; n++, avail--) {
898 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
899 				if (error)
900 					ring->slot[nm_i].len = 0;
901 				else
902 					ring->slot[nm_i].len = ri.iri_len - crclen;
903 				ring->slot[nm_i].flags = slot_flags;
904 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
905 								fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
906 				nm_i = nm_next(nm_i, lim);
907 				nic_i = nm_next(nic_i, lim);
908 			}
909 			if (n) { /* update the state variables */
910 				if (netmap_no_pendintr && !force_update) {
911 					/* diagnostics */
912 					iflib_rx_miss ++;
913 					iflib_rx_miss_bufs += n;
914 				}
915 				fl->ifl_cidx = nic_i;
916 				kring->nr_hwtail = nm_i;
917 			}
918 			kring->nr_kflags &= ~NKR_PENDINTR;
919 		}
920 	}
921 	/*
922 	 * Second part: skip past packets that userspace has released.
923 	 * (kring->nr_hwcur to head excluded),
924 	 * and make the buffers available for reception.
925 	 * As usual nm_i is the index in the netmap ring,
926 	 * nic_i is the index in the NIC ring, and
927 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
928 	 */
929 	/* XXX not sure how this will work with multiple free lists */
930 	nm_i = kring->nr_hwcur;
931 	if (nm_i != head) {
932 		nic_i = netmap_idx_k2n(kring, nm_i);
933 		for (n = 0; nm_i != head; n++) {
934 			struct netmap_slot *slot = &ring->slot[nm_i];
935 			uint64_t paddr;
936 			caddr_t vaddr;
937 			void *addr = PNMB(na, slot, &paddr);
938 
939 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
940 				goto ring_reset;
941 
942 			vaddr = addr;
943 			if (slot->flags & NS_BUF_CHANGED) {
944 				/* buffer has changed, reload map */
945 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
946 				slot->flags &= ~NS_BUF_CHANGED;
947 			}
948 			/*
949 			 * XXX we should be batching this operation - TODO
950 			 */
951 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size);
952 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
953 			    BUS_DMASYNC_PREREAD);
954 			nm_i = nm_next(nm_i, lim);
955 			nic_i = nm_next(nic_i, lim);
956 		}
957 		kring->nr_hwcur = head;
958 
959 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
960 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
961 		/*
962 		 * IMPORTANT: we must leave one free slot in the ring,
963 		 * so move nic_i back by one unit
964 		 */
965 		nic_i = nm_prev(nic_i, lim);
966 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
967 	}
968 
969 	return 0;
970 
971 ring_reset:
972 	return netmap_ring_reinit(kring);
973 }
974 
975 static int
976 iflib_netmap_attach(if_ctx_t ctx)
977 {
978 	struct netmap_adapter na;
979 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
980 
981 	bzero(&na, sizeof(na));
982 
983 	na.ifp = ctx->ifc_ifp;
984 	na.na_flags = NAF_BDG_MAYSLEEP;
985 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
986 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
987 
988 	na.num_tx_desc = scctx->isc_ntxd[0];
989 	na.num_rx_desc = scctx->isc_nrxd[0];
990 	na.nm_txsync = iflib_netmap_txsync;
991 	na.nm_rxsync = iflib_netmap_rxsync;
992 	na.nm_register = iflib_netmap_register;
993 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
994 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
995 	return (netmap_attach(&na));
996 }
997 
998 static void
999 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1000 {
1001 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1002 	struct netmap_slot *slot;
1003 
1004 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1005 	if (slot == NULL)
1006 		return;
1007 
1008 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1009 
1010 		/*
1011 		 * In netmap mode, set the map for the packet buffer.
1012 		 * NOTE: Some drivers (not this one) also need to set
1013 		 * the physical buffer address in the NIC ring.
1014 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1015 		 * netmap slot index, si
1016 		 */
1017 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1018 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1019 	}
1020 }
1021 static void
1022 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1023 {
1024 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1025 	struct netmap_slot *slot;
1026 	bus_dmamap_t *map;
1027 	int nrxd;
1028 
1029 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1030 	if (slot == NULL)
1031 		return;
1032 	map = rxq->ifr_fl[0].ifl_sds.ifsd_map;
1033 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1034 	for (int i = 0; i < nrxd; i++, map++) {
1035 			int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1036 			uint64_t paddr;
1037 			void *addr;
1038 			caddr_t vaddr;
1039 
1040 			vaddr = addr = PNMB(na, slot + sj, &paddr);
1041 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1042 			/* Update descriptor and the cached value */
1043 			ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size);
1044 	}
1045 	/* preserve queue */
1046 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1047 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1048 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1049 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1050 	} else
1051 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1052 }
1053 
1054 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1055 
1056 #else
1057 #define iflib_netmap_txq_init(ctx, txq)
1058 #define iflib_netmap_rxq_init(ctx, rxq)
1059 #define iflib_netmap_detach(ifp)
1060 
1061 #define iflib_netmap_attach(ctx) (0)
1062 #define netmap_rx_irq(ifp, qid, budget) (0)
1063 
1064 #endif
1065 
1066 #if defined(__i386__) || defined(__amd64__)
1067 static __inline void
1068 prefetch(void *x)
1069 {
1070 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1071 }
1072 #else
1073 #define prefetch(x)
1074 #endif
1075 
1076 static void
1077 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1078 {
1079 	if (err)
1080 		return;
1081 	*(bus_addr_t *) arg = segs[0].ds_addr;
1082 }
1083 
1084 int
1085 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1086 {
1087 	int err;
1088 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1089 	device_t dev = ctx->ifc_dev;
1090 
1091 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1092 
1093 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1094 				sctx->isc_q_align, 0,	/* alignment, bounds */
1095 				BUS_SPACE_MAXADDR,	/* lowaddr */
1096 				BUS_SPACE_MAXADDR,	/* highaddr */
1097 				NULL, NULL,		/* filter, filterarg */
1098 				size,			/* maxsize */
1099 				1,			/* nsegments */
1100 				size,			/* maxsegsize */
1101 				BUS_DMA_ALLOCNOW,	/* flags */
1102 				NULL,			/* lockfunc */
1103 				NULL,			/* lockarg */
1104 				&dma->idi_tag);
1105 	if (err) {
1106 		device_printf(dev,
1107 		    "%s: bus_dma_tag_create failed: %d\n",
1108 		    __func__, err);
1109 		goto fail_0;
1110 	}
1111 
1112 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1113 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1114 	if (err) {
1115 		device_printf(dev,
1116 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1117 		    __func__, (uintmax_t)size, err);
1118 		goto fail_1;
1119 	}
1120 
1121 	dma->idi_paddr = IF_BAD_DMA;
1122 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1123 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1124 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1125 		device_printf(dev,
1126 		    "%s: bus_dmamap_load failed: %d\n",
1127 		    __func__, err);
1128 		goto fail_2;
1129 	}
1130 
1131 	dma->idi_size = size;
1132 	return (0);
1133 
1134 fail_2:
1135 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1136 fail_1:
1137 	bus_dma_tag_destroy(dma->idi_tag);
1138 fail_0:
1139 	dma->idi_tag = NULL;
1140 
1141 	return (err);
1142 }
1143 
1144 int
1145 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1146 {
1147 	int i, err;
1148 	iflib_dma_info_t *dmaiter;
1149 
1150 	dmaiter = dmalist;
1151 	for (i = 0; i < count; i++, dmaiter++) {
1152 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1153 			break;
1154 	}
1155 	if (err)
1156 		iflib_dma_free_multi(dmalist, i);
1157 	return (err);
1158 }
1159 
1160 void
1161 iflib_dma_free(iflib_dma_info_t dma)
1162 {
1163 	if (dma->idi_tag == NULL)
1164 		return;
1165 	if (dma->idi_paddr != IF_BAD_DMA) {
1166 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1167 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1168 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1169 		dma->idi_paddr = IF_BAD_DMA;
1170 	}
1171 	if (dma->idi_vaddr != NULL) {
1172 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1173 		dma->idi_vaddr = NULL;
1174 	}
1175 	bus_dma_tag_destroy(dma->idi_tag);
1176 	dma->idi_tag = NULL;
1177 }
1178 
1179 void
1180 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1181 {
1182 	int i;
1183 	iflib_dma_info_t *dmaiter = dmalist;
1184 
1185 	for (i = 0; i < count; i++, dmaiter++)
1186 		iflib_dma_free(*dmaiter);
1187 }
1188 
1189 #ifdef EARLY_AP_STARTUP
1190 static const int iflib_started = 1;
1191 #else
1192 /*
1193  * We used to abuse the smp_started flag to decide if the queues have been
1194  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1195  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1196  * is set.  Run a SYSINIT() strictly after that to just set a usable
1197  * completion flag.
1198  */
1199 
1200 static int iflib_started;
1201 
1202 static void
1203 iflib_record_started(void *arg)
1204 {
1205 	iflib_started = 1;
1206 }
1207 
1208 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1209 	iflib_record_started, NULL);
1210 #endif
1211 
1212 static int
1213 iflib_fast_intr(void *arg)
1214 {
1215 	iflib_filter_info_t info = arg;
1216 	struct grouptask *gtask = info->ifi_task;
1217 
1218 	if (!iflib_started)
1219 		return (FILTER_HANDLED);
1220 
1221 	DBG_COUNTER_INC(fast_intrs);
1222 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1223 		return (FILTER_HANDLED);
1224 
1225 	GROUPTASK_ENQUEUE(gtask);
1226 	return (FILTER_HANDLED);
1227 }
1228 
1229 static int
1230 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1231 	driver_filter_t filter, driver_intr_t handler, void *arg,
1232 				 char *name)
1233 {
1234 	int rc;
1235 	struct resource *res;
1236 	void *tag;
1237 	device_t dev = ctx->ifc_dev;
1238 
1239 	MPASS(rid < 512);
1240 	irq->ii_rid = rid;
1241 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1242 				     RF_SHAREABLE | RF_ACTIVE);
1243 	if (res == NULL) {
1244 		device_printf(dev,
1245 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1246 		return (ENOMEM);
1247 	}
1248 	irq->ii_res = res;
1249 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1250 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1251 						filter, handler, arg, &tag);
1252 	if (rc != 0) {
1253 		device_printf(dev,
1254 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1255 					  rid, name ? name : "unknown", rc);
1256 		return (rc);
1257 	} else if (name)
1258 		bus_describe_intr(dev, res, tag, "%s", name);
1259 
1260 	irq->ii_tag = tag;
1261 	return (0);
1262 }
1263 
1264 
1265 /*********************************************************************
1266  *
1267  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1268  *  the information needed to transmit a packet on the wire. This is
1269  *  called only once at attach, setup is done every reset.
1270  *
1271  **********************************************************************/
1272 
1273 static int
1274 iflib_txsd_alloc(iflib_txq_t txq)
1275 {
1276 	if_ctx_t ctx = txq->ift_ctx;
1277 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1278 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1279 	device_t dev = ctx->ifc_dev;
1280 	int err, nsegments, ntsosegments;
1281 
1282 	nsegments = scctx->isc_tx_nsegments;
1283 	ntsosegments = scctx->isc_tx_tso_segments_max;
1284 	MPASS(scctx->isc_ntxd[0] > 0);
1285 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1286 	MPASS(nsegments > 0);
1287 	MPASS(ntsosegments > 0);
1288 	/*
1289 	 * Setup DMA descriptor areas.
1290 	 */
1291 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1292 			       1, 0,			/* alignment, bounds */
1293 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1294 			       BUS_SPACE_MAXADDR,	/* highaddr */
1295 			       NULL, NULL,		/* filter, filterarg */
1296 			       sctx->isc_tx_maxsize,		/* maxsize */
1297 			       nsegments,	/* nsegments */
1298 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1299 			       0,			/* flags */
1300 			       NULL,			/* lockfunc */
1301 			       NULL,			/* lockfuncarg */
1302 			       &txq->ift_desc_tag))) {
1303 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1304 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1305 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1306 		goto fail;
1307 	}
1308 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1309 			       1, 0,			/* alignment, bounds */
1310 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1311 			       BUS_SPACE_MAXADDR,	/* highaddr */
1312 			       NULL, NULL,		/* filter, filterarg */
1313 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1314 			       ntsosegments,	/* nsegments */
1315 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1316 			       0,			/* flags */
1317 			       NULL,			/* lockfunc */
1318 			       NULL,			/* lockfuncarg */
1319 			       &txq->ift_tso_desc_tag))) {
1320 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1321 
1322 		goto fail;
1323 	}
1324 	if (!(txq->ift_sds.ifsd_flags =
1325 	    (uint8_t *) malloc(sizeof(uint8_t) *
1326 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1327 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1328 		err = ENOMEM;
1329 		goto fail;
1330 	}
1331 	if (!(txq->ift_sds.ifsd_m =
1332 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1333 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1334 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1335 		err = ENOMEM;
1336 		goto fail;
1337 	}
1338 
1339         /* Create the descriptor buffer dma maps */
1340 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1341 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1342 		return (0);
1343 
1344 	if (!(txq->ift_sds.ifsd_map =
1345 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1346 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1347 		err = ENOMEM;
1348 		goto fail;
1349 	}
1350 
1351 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1352 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1353 		if (err != 0) {
1354 			device_printf(dev, "Unable to create TX DMA map\n");
1355 			goto fail;
1356 		}
1357 	}
1358 #endif
1359 	return (0);
1360 fail:
1361 	/* We free all, it handles case where we are in the middle */
1362 	iflib_tx_structures_free(ctx);
1363 	return (err);
1364 }
1365 
1366 static void
1367 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1368 {
1369 	bus_dmamap_t map;
1370 
1371 	map = NULL;
1372 	if (txq->ift_sds.ifsd_map != NULL)
1373 		map = txq->ift_sds.ifsd_map[i];
1374 	if (map != NULL) {
1375 		bus_dmamap_unload(txq->ift_desc_tag, map);
1376 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1377 		txq->ift_sds.ifsd_map[i] = NULL;
1378 	}
1379 }
1380 
1381 static void
1382 iflib_txq_destroy(iflib_txq_t txq)
1383 {
1384 	if_ctx_t ctx = txq->ift_ctx;
1385 
1386 	for (int i = 0; i < txq->ift_size; i++)
1387 		iflib_txsd_destroy(ctx, txq, i);
1388 	if (txq->ift_sds.ifsd_map != NULL) {
1389 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1390 		txq->ift_sds.ifsd_map = NULL;
1391 	}
1392 	if (txq->ift_sds.ifsd_m != NULL) {
1393 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1394 		txq->ift_sds.ifsd_m = NULL;
1395 	}
1396 	if (txq->ift_sds.ifsd_flags != NULL) {
1397 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1398 		txq->ift_sds.ifsd_flags = NULL;
1399 	}
1400 	if (txq->ift_desc_tag != NULL) {
1401 		bus_dma_tag_destroy(txq->ift_desc_tag);
1402 		txq->ift_desc_tag = NULL;
1403 	}
1404 	if (txq->ift_tso_desc_tag != NULL) {
1405 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1406 		txq->ift_tso_desc_tag = NULL;
1407 	}
1408 }
1409 
1410 static void
1411 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1412 {
1413 	struct mbuf **mp;
1414 
1415 	mp = &txq->ift_sds.ifsd_m[i];
1416 	if (*mp == NULL)
1417 		return;
1418 
1419 	if (txq->ift_sds.ifsd_map != NULL) {
1420 		bus_dmamap_sync(txq->ift_desc_tag,
1421 				txq->ift_sds.ifsd_map[i],
1422 				BUS_DMASYNC_POSTWRITE);
1423 		bus_dmamap_unload(txq->ift_desc_tag,
1424 				  txq->ift_sds.ifsd_map[i]);
1425 	}
1426 	m_free(*mp);
1427 	DBG_COUNTER_INC(tx_frees);
1428 	*mp = NULL;
1429 }
1430 
1431 static int
1432 iflib_txq_setup(iflib_txq_t txq)
1433 {
1434 	if_ctx_t ctx = txq->ift_ctx;
1435 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1436 	iflib_dma_info_t di;
1437 	int i;
1438 
1439 	/* Set number of descriptors available */
1440 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1441 
1442 	/* Reset indices */
1443 	txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1444 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1445 
1446 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1447 		bzero((void *)di->idi_vaddr, di->idi_size);
1448 
1449 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1450 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1451 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1452 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1453 	return (0);
1454 }
1455 
1456 /*********************************************************************
1457  *
1458  *  Allocate memory for rx_buffer structures. Since we use one
1459  *  rx_buffer per received packet, the maximum number of rx_buffer's
1460  *  that we'll need is equal to the number of receive descriptors
1461  *  that we've allocated.
1462  *
1463  **********************************************************************/
1464 static int
1465 iflib_rxsd_alloc(iflib_rxq_t rxq)
1466 {
1467 	if_ctx_t ctx = rxq->ifr_ctx;
1468 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1469 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1470 	device_t dev = ctx->ifc_dev;
1471 	iflib_fl_t fl;
1472 	int			err;
1473 
1474 	MPASS(scctx->isc_nrxd[0] > 0);
1475 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1476 
1477 	fl = rxq->ifr_fl;
1478 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1479 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1480 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1481 					 1, 0,			/* alignment, bounds */
1482 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1483 					 BUS_SPACE_MAXADDR,	/* highaddr */
1484 					 NULL, NULL,		/* filter, filterarg */
1485 					 sctx->isc_rx_maxsize,	/* maxsize */
1486 					 sctx->isc_rx_nsegments,	/* nsegments */
1487 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1488 					 0,			/* flags */
1489 					 NULL,			/* lockfunc */
1490 					 NULL,			/* lockarg */
1491 					 &fl->ifl_desc_tag);
1492 		if (err) {
1493 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1494 				__func__, err);
1495 			goto fail;
1496 		}
1497 		if (!(fl->ifl_sds.ifsd_flags =
1498 		      (uint8_t *) malloc(sizeof(uint8_t) *
1499 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1500 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1501 			err = ENOMEM;
1502 			goto fail;
1503 		}
1504 		if (!(fl->ifl_sds.ifsd_m =
1505 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1506 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1507 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1508 			err = ENOMEM;
1509 			goto fail;
1510 		}
1511 		if (!(fl->ifl_sds.ifsd_cl =
1512 		      (caddr_t *) malloc(sizeof(caddr_t) *
1513 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1514 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1515 			err = ENOMEM;
1516 			goto fail;
1517 		}
1518 
1519 		/* Create the descriptor buffer dma maps */
1520 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1521 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1522 			continue;
1523 
1524 		if (!(fl->ifl_sds.ifsd_map =
1525 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1526 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1527 			err = ENOMEM;
1528 			goto fail;
1529 		}
1530 
1531 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1532 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1533 			if (err != 0) {
1534 				device_printf(dev, "Unable to create TX DMA map\n");
1535 				goto fail;
1536 			}
1537 		}
1538 #endif
1539 	}
1540 	return (0);
1541 
1542 fail:
1543 	iflib_rx_structures_free(ctx);
1544 	return (err);
1545 }
1546 
1547 
1548 /*
1549  * Internal service routines
1550  */
1551 
1552 struct rxq_refill_cb_arg {
1553 	int               error;
1554 	bus_dma_segment_t seg;
1555 	int               nseg;
1556 };
1557 
1558 static void
1559 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1560 {
1561 	struct rxq_refill_cb_arg *cb_arg = arg;
1562 
1563 	cb_arg->error = error;
1564 	cb_arg->seg = segs[0];
1565 	cb_arg->nseg = nseg;
1566 }
1567 
1568 
1569 #ifdef ACPI_DMAR
1570 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1571 #else
1572 #define IS_DMAR(ctx) (0)
1573 #endif
1574 
1575 /**
1576  *	rxq_refill - refill an rxq  free-buffer list
1577  *	@ctx: the iflib context
1578  *	@rxq: the free-list to refill
1579  *	@n: the number of new buffers to allocate
1580  *
1581  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1582  *	The caller must assure that @n does not exceed the queue's capacity.
1583  */
1584 static void
1585 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1586 {
1587 	struct mbuf *m;
1588 	int idx, pidx = fl->ifl_pidx;
1589 	caddr_t cl, *sd_cl;
1590 	struct mbuf **sd_m;
1591 	uint8_t *sd_flags;
1592 	bus_dmamap_t *sd_map;
1593 	int n, i = 0;
1594 	uint64_t bus_addr;
1595 	int err;
1596 
1597 	sd_m = fl->ifl_sds.ifsd_m;
1598 	sd_map = fl->ifl_sds.ifsd_map;
1599 	sd_cl = fl->ifl_sds.ifsd_cl;
1600 	sd_flags = fl->ifl_sds.ifsd_flags;
1601 	idx = pidx;
1602 
1603 	n  = count;
1604 	MPASS(n > 0);
1605 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1606 
1607 	if (pidx < fl->ifl_cidx)
1608 		MPASS(pidx + n <= fl->ifl_cidx);
1609 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1610 		MPASS(fl->ifl_gen == 0);
1611 	if (pidx > fl->ifl_cidx)
1612 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1613 
1614 	DBG_COUNTER_INC(fl_refills);
1615 	if (n > 8)
1616 		DBG_COUNTER_INC(fl_refills_large);
1617 
1618 	while (n--) {
1619 		/*
1620 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1621 		 * initialized after rx.
1622 		 *
1623 		 * If the cluster is still set then we know a minimum sized packet was received
1624 		 */
1625 		if ((cl = sd_cl[idx]) == NULL) {
1626 			if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1627 				break;
1628 #if MEMORY_LOGGING
1629 			fl->ifl_cl_enqueued++;
1630 #endif
1631 		}
1632 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1633 			break;
1634 		}
1635 #if MEMORY_LOGGING
1636 		fl->ifl_m_enqueued++;
1637 #endif
1638 
1639 		DBG_COUNTER_INC(rx_allocs);
1640 #ifdef notyet
1641 		if ((sd_flags[pidx] & RX_SW_DESC_MAP_CREATED) == 0) {
1642 			int err;
1643 
1644 			if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &sd_map[idx]))) {
1645 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
1646 				uma_zfree(fl->ifl_zone, cl);
1647 				n = 0;
1648 				goto done;
1649 			}
1650 			sd_flags[idx] |= RX_SW_DESC_MAP_CREATED;
1651 		}
1652 #endif
1653 #if defined(__i386__) || defined(__amd64__)
1654 		if (!IS_DMAR(ctx)) {
1655 			bus_addr = pmap_kextract((vm_offset_t)cl);
1656 		} else
1657 #endif
1658 		{
1659 			struct rxq_refill_cb_arg cb_arg;
1660 			iflib_rxq_t q;
1661 
1662 			cb_arg.error = 0;
1663 			q = fl->ifl_rxq;
1664 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1665 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1666 
1667 			if (err != 0 || cb_arg.error) {
1668 				/*
1669 				 * !zone_pack ?
1670 				 */
1671 				if (fl->ifl_zone == zone_pack)
1672 					uma_zfree(fl->ifl_zone, cl);
1673 				m_free(m);
1674 				n = 0;
1675 				goto done;
1676 			}
1677 			bus_addr = cb_arg.seg.ds_addr;
1678 		}
1679 		sd_flags[idx] |= RX_SW_DESC_INUSE;
1680 
1681 		MPASS(sd_m[idx] == NULL);
1682 		sd_cl[idx] = cl;
1683 		sd_m[idx] = m;
1684 		fl->ifl_bus_addrs[i] = bus_addr;
1685 		fl->ifl_vm_addrs[i] = cl;
1686 		fl->ifl_credits++;
1687 		i++;
1688 		MPASS(fl->ifl_credits <= fl->ifl_size);
1689 		if (++idx == fl->ifl_size) {
1690 			fl->ifl_gen = 1;
1691 			idx = 0;
1692 		}
1693 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1694 			ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx,
1695 								 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size);
1696 			i = 0;
1697 			pidx = idx;
1698 		}
1699 		fl->ifl_pidx = idx;
1700 
1701 	}
1702 done:
1703 	DBG_COUNTER_INC(rxd_flush);
1704 	if (fl->ifl_pidx == 0)
1705 		pidx = fl->ifl_size - 1;
1706 	else
1707 		pidx = fl->ifl_pidx - 1;
1708 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1709 }
1710 
1711 static __inline void
1712 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1713 {
1714 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1715 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1716 #ifdef INVARIANTS
1717 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1718 #endif
1719 
1720 	MPASS(fl->ifl_credits <= fl->ifl_size);
1721 	MPASS(reclaimable == delta);
1722 
1723 	if (reclaimable > 0)
1724 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1725 }
1726 
1727 static void
1728 iflib_fl_bufs_free(iflib_fl_t fl)
1729 {
1730 	iflib_dma_info_t idi = fl->ifl_ifdi;
1731 	uint32_t i;
1732 
1733 	for (i = 0; i < fl->ifl_size; i++) {
1734 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1735 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1736 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1737 
1738 		if (*sd_flags & RX_SW_DESC_INUSE) {
1739 			if (fl->ifl_sds.ifsd_map != NULL) {
1740 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1741 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1742 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1743 			}
1744 			if (*sd_m != NULL) {
1745 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1746 				uma_zfree(zone_mbuf, *sd_m);
1747 			}
1748 			if (*sd_cl != NULL)
1749 				uma_zfree(fl->ifl_zone, *sd_cl);
1750 			*sd_flags = 0;
1751 		} else {
1752 			MPASS(*sd_cl == NULL);
1753 			MPASS(*sd_m == NULL);
1754 		}
1755 #if MEMORY_LOGGING
1756 		fl->ifl_m_dequeued++;
1757 		fl->ifl_cl_dequeued++;
1758 #endif
1759 		*sd_cl = NULL;
1760 		*sd_m = NULL;
1761 	}
1762 	/*
1763 	 * Reset free list values
1764 	 */
1765 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1766 	bzero(idi->idi_vaddr, idi->idi_size);
1767 }
1768 
1769 /*********************************************************************
1770  *
1771  *  Initialize a receive ring and its buffers.
1772  *
1773  **********************************************************************/
1774 static int
1775 iflib_fl_setup(iflib_fl_t fl)
1776 {
1777 	iflib_rxq_t rxq = fl->ifl_rxq;
1778 	if_ctx_t ctx = rxq->ifr_ctx;
1779 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1780 
1781 	/*
1782 	** Free current RX buffer structs and their mbufs
1783 	*/
1784 	iflib_fl_bufs_free(fl);
1785 	/* Now replenish the mbufs */
1786 	MPASS(fl->ifl_credits == 0);
1787 	/*
1788 	 * XXX don't set the max_frame_size to larger
1789 	 * than the hardware can handle
1790 	 */
1791 	if (sctx->isc_max_frame_size <= 2048)
1792 		fl->ifl_buf_size = MCLBYTES;
1793 	else if (sctx->isc_max_frame_size <= 4096)
1794 		fl->ifl_buf_size = MJUMPAGESIZE;
1795 	else if (sctx->isc_max_frame_size <= 9216)
1796 		fl->ifl_buf_size = MJUM9BYTES;
1797 	else
1798 		fl->ifl_buf_size = MJUM16BYTES;
1799 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
1800 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
1801 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
1802 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
1803 
1804 
1805 	/* avoid pre-allocating zillions of clusters to an idle card
1806 	 * potentially speeding up attach
1807 	 */
1808 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
1809 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
1810 	if (min(128, fl->ifl_size) != fl->ifl_credits)
1811 		return (ENOBUFS);
1812 	/*
1813 	 * handle failure
1814 	 */
1815 	MPASS(rxq != NULL);
1816 	MPASS(fl->ifl_ifdi != NULL);
1817 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1818 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1819 	return (0);
1820 }
1821 
1822 /*********************************************************************
1823  *
1824  *  Free receive ring data structures
1825  *
1826  **********************************************************************/
1827 static void
1828 iflib_rx_sds_free(iflib_rxq_t rxq)
1829 {
1830 	iflib_fl_t fl;
1831 	int i;
1832 
1833 	if (rxq->ifr_fl != NULL) {
1834 		for (i = 0; i < rxq->ifr_nfl; i++) {
1835 			fl = &rxq->ifr_fl[i];
1836 			if (fl->ifl_desc_tag != NULL) {
1837 				bus_dma_tag_destroy(fl->ifl_desc_tag);
1838 				fl->ifl_desc_tag = NULL;
1839 			}
1840 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
1841 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
1842 			/* XXX destroy maps first */
1843 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
1844 			fl->ifl_sds.ifsd_m = NULL;
1845 			fl->ifl_sds.ifsd_cl = NULL;
1846 			fl->ifl_sds.ifsd_map = NULL;
1847 		}
1848 		free(rxq->ifr_fl, M_IFLIB);
1849 		rxq->ifr_fl = NULL;
1850 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
1851 	}
1852 }
1853 
1854 /*
1855  * MI independent logic
1856  *
1857  */
1858 static void
1859 iflib_timer(void *arg)
1860 {
1861 	iflib_txq_t txq = arg;
1862 	if_ctx_t ctx = txq->ift_ctx;
1863 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1864 
1865 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
1866 		return;
1867 	/*
1868 	** Check on the state of the TX queue(s), this
1869 	** can be done without the lock because its RO
1870 	** and the HUNG state will be static if set.
1871 	*/
1872 	IFDI_TIMER(ctx, txq->ift_id);
1873 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
1874 		(ctx->ifc_pause_frames == 0))
1875 		goto hung;
1876 
1877 	if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments ||
1878 	    ifmp_ring_is_stalled(txq->ift_br[0]))
1879 		GROUPTASK_ENQUEUE(&txq->ift_task);
1880 
1881 	ctx->ifc_pause_frames = 0;
1882 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
1883 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
1884 	return;
1885 hung:
1886 	CTX_LOCK(ctx);
1887 	if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
1888 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
1889 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
1890 
1891 	IFDI_WATCHDOG_RESET(ctx);
1892 	ctx->ifc_watchdog_events++;
1893 	ctx->ifc_pause_frames = 0;
1894 
1895 	iflib_init_locked(ctx);
1896 	CTX_UNLOCK(ctx);
1897 }
1898 
1899 static void
1900 iflib_init_locked(if_ctx_t ctx)
1901 {
1902 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1903 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1904 	if_t ifp = ctx->ifc_ifp;
1905 	iflib_fl_t fl;
1906 	iflib_txq_t txq;
1907 	iflib_rxq_t rxq;
1908 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
1909 
1910 
1911 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1912 	IFDI_INTR_DISABLE(ctx);
1913 
1914 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
1915 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
1916 	/* Set hardware offload abilities */
1917 	if_clearhwassist(ifp);
1918 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
1919 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
1920 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
1921 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
1922 	if (if_getcapenable(ifp) & IFCAP_TSO4)
1923 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
1924 	if (if_getcapenable(ifp) & IFCAP_TSO6)
1925 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
1926 
1927 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
1928 		CALLOUT_LOCK(txq);
1929 		callout_stop(&txq->ift_timer);
1930 		callout_stop(&txq->ift_db_check);
1931 		CALLOUT_UNLOCK(txq);
1932 		iflib_netmap_txq_init(ctx, txq);
1933 	}
1934 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1935 		iflib_netmap_rxq_init(ctx, rxq);
1936 	}
1937 #ifdef INVARIANTS
1938 	i = if_getdrvflags(ifp);
1939 #endif
1940 	IFDI_INIT(ctx);
1941 	MPASS(if_getdrvflags(ifp) == i);
1942 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1943 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
1944 			if (iflib_fl_setup(fl)) {
1945 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
1946 				goto done;
1947 			}
1948 		}
1949 	}
1950 	done:
1951 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
1952 	IFDI_INTR_ENABLE(ctx);
1953 	txq = ctx->ifc_txqs;
1954 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
1955 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
1956 			txq->ift_timer.c_cpu);
1957 }
1958 
1959 static int
1960 iflib_media_change(if_t ifp)
1961 {
1962 	if_ctx_t ctx = if_getsoftc(ifp);
1963 	int err;
1964 
1965 	CTX_LOCK(ctx);
1966 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
1967 		iflib_init_locked(ctx);
1968 	CTX_UNLOCK(ctx);
1969 	return (err);
1970 }
1971 
1972 static void
1973 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
1974 {
1975 	if_ctx_t ctx = if_getsoftc(ifp);
1976 
1977 	CTX_LOCK(ctx);
1978 	IFDI_UPDATE_ADMIN_STATUS(ctx);
1979 	IFDI_MEDIA_STATUS(ctx, ifmr);
1980 	CTX_UNLOCK(ctx);
1981 }
1982 
1983 static void
1984 iflib_stop(if_ctx_t ctx)
1985 {
1986 	iflib_txq_t txq = ctx->ifc_txqs;
1987 	iflib_rxq_t rxq = ctx->ifc_rxqs;
1988 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1989 	iflib_dma_info_t di;
1990 	iflib_fl_t fl;
1991 	int i, j;
1992 
1993 	/* Tell the stack that the interface is no longer active */
1994 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1995 
1996 	IFDI_INTR_DISABLE(ctx);
1997 	DELAY(100000);
1998 	IFDI_STOP(ctx);
1999 	DELAY(100000);
2000 
2001 	iflib_debug_reset();
2002 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2003 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2004 		/* make sure all transmitters have completed before proceeding XXX */
2005 
2006 		/* clean any enqueued buffers */
2007 		iflib_ifmp_purge(txq);
2008 		/* Free any existing tx buffers. */
2009 		for (j = 0; j < txq->ift_size; j++) {
2010 			iflib_txsd_free(ctx, txq, j);
2011 		}
2012 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2013 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2014 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2015 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2016 		txq->ift_pullups = 0;
2017 		ifmp_ring_reset_stats(txq->ift_br[0]);
2018 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2019 			bzero((void *)di->idi_vaddr, di->idi_size);
2020 	}
2021 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2022 		/* make sure all transmitters have completed before proceeding XXX */
2023 
2024 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2025 			bzero((void *)di->idi_vaddr, di->idi_size);
2026 		/* also resets the free lists pidx/cidx */
2027 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2028 			iflib_fl_bufs_free(fl);
2029 	}
2030 }
2031 
2032 static inline void
2033 prefetch_pkts(iflib_fl_t fl, int cidx)
2034 {
2035 	int nextptr;
2036 	int nrxd = fl->ifl_size;
2037 
2038 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2039 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2040 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2041 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2042 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2043 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2044 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2045 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2046 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2047 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2048 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2049 }
2050 
2051 static void
2052 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload, iflib_fl_t *pfl, int *pcidx)
2053 {
2054 	int flid, cidx;
2055 	bus_dmamap_t map;
2056 	iflib_fl_t fl;
2057 	iflib_dma_info_t di;
2058 	int next;
2059 
2060 	flid = irf->irf_flid;
2061 	cidx = irf->irf_idx;
2062 	fl = &rxq->ifr_fl[flid];
2063 	fl->ifl_credits--;
2064 #if MEMORY_LOGGING
2065 	fl->ifl_m_dequeued++;
2066 	if (cltype)
2067 		fl->ifl_cl_dequeued++;
2068 #endif
2069 	prefetch_pkts(fl, cidx);
2070 	if (fl->ifl_sds.ifsd_map != NULL) {
2071 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2072 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2073 		map = fl->ifl_sds.ifsd_map[cidx];
2074 		di = fl->ifl_ifdi;
2075 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2076 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2077 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2078 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2079 
2080 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2081 		MPASS(fl->ifl_cidx == cidx);
2082 		if (unload)
2083 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2084 	}
2085 	if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) {
2086 		fl->ifl_cidx = 0;
2087 		fl->ifl_gen = 0;
2088 	}
2089 	/* YES ick */
2090 	if (cltype)
2091 		*cltype = fl->ifl_cltype;
2092 	*pfl = fl;
2093 	*pcidx = cidx;
2094 }
2095 
2096 static struct mbuf *
2097 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri)
2098 {
2099 	int i, padlen , flags, cltype;
2100 	struct mbuf *m, *mh, *mt, *sd_m;
2101 	iflib_fl_t fl;
2102 	int cidx;
2103 	caddr_t cl, sd_cl;
2104 
2105 	i = 0;
2106 	mh = NULL;
2107 	do {
2108 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE, &fl, &cidx);
2109 		sd_m = fl->ifl_sds.ifsd_m[cidx];
2110 		sd_cl = fl->ifl_sds.ifsd_cl[cidx];
2111 
2112 		MPASS(sd_cl != NULL);
2113 		MPASS(sd_m != NULL);
2114 
2115 		/* Don't include zero-length frags */
2116 		if (ri->iri_frags[i].irf_len == 0) {
2117 			/* XXX we can save the cluster here, but not the mbuf */
2118 			m_init(sd_m, M_NOWAIT, MT_DATA, 0);
2119 			m_free(sd_m);
2120 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2121 			continue;
2122 		}
2123 		m = sd_m;
2124 		if (mh == NULL) {
2125 			flags = M_PKTHDR|M_EXT;
2126 			mh = mt = m;
2127 			padlen = ri->iri_pad;
2128 		} else {
2129 			flags = M_EXT;
2130 			mt->m_next = m;
2131 			mt = m;
2132 			/* assuming padding is only on the first fragment */
2133 			padlen = 0;
2134 		}
2135 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2136 		cl = fl->ifl_sds.ifsd_cl[cidx];
2137 		fl->ifl_sds.ifsd_cl[cidx] = NULL;
2138 
2139 		/* Can these two be made one ? */
2140 		m_init(m, M_NOWAIT, MT_DATA, flags);
2141 		m_cljset(m, cl, cltype);
2142 		/*
2143 		 * These must follow m_init and m_cljset
2144 		 */
2145 		m->m_data += padlen;
2146 		ri->iri_len -= padlen;
2147 		m->m_len = ri->iri_frags[i].irf_len;
2148 	} while (++i < ri->iri_nfrags);
2149 
2150 	return (mh);
2151 }
2152 
2153 /*
2154  * Process one software descriptor
2155  */
2156 static struct mbuf *
2157 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2158 {
2159 	struct mbuf *m;
2160 	iflib_fl_t fl;
2161 	caddr_t sd_cl;
2162 	int cidx;
2163 
2164 	/* should I merge this back in now that the two paths are basically duplicated? */
2165 	if (ri->iri_nfrags == 1 &&
2166 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2167 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE, &fl, &cidx);
2168 		m = fl->ifl_sds.ifsd_m[cidx];
2169 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2170 		sd_cl = fl->ifl_sds.ifsd_cl[cidx];
2171 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2172 		memcpy(m->m_data, sd_cl, ri->iri_len);
2173 		m->m_len = ri->iri_frags[0].irf_len;
2174        } else {
2175 		m = assemble_segments(rxq, ri);
2176 	}
2177 	m->m_pkthdr.len = ri->iri_len;
2178 	m->m_pkthdr.rcvif = ri->iri_ifp;
2179 	m->m_flags |= ri->iri_flags;
2180 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2181 	m->m_pkthdr.flowid = ri->iri_flowid;
2182 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2183 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2184 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2185 	return (m);
2186 }
2187 
2188 static bool
2189 iflib_rxeof(iflib_rxq_t rxq, int budget)
2190 {
2191 	if_ctx_t ctx = rxq->ifr_ctx;
2192 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2193 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2194 	int avail, i;
2195 	uint16_t *cidxp;
2196 	struct if_rxd_info ri;
2197 	int err, budget_left, rx_bytes, rx_pkts;
2198 	iflib_fl_t fl;
2199 	struct ifnet *ifp;
2200 	int lro_enabled;
2201 	/*
2202 	 * XXX early demux data packets so that if_input processing only handles
2203 	 * acks in interrupt context
2204 	 */
2205 	struct mbuf *m, *mh, *mt;
2206 
2207 	if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) {
2208 		return (FALSE);
2209 	}
2210 
2211 	mh = mt = NULL;
2212 	MPASS(budget > 0);
2213 	rx_pkts	= rx_bytes = 0;
2214 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2215 		cidxp = &rxq->ifr_cq_cidx;
2216 	else
2217 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2218 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2219 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2220 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2221 		DBG_COUNTER_INC(rx_unavail);
2222 		return (false);
2223 	}
2224 
2225 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2226 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2227 			DBG_COUNTER_INC(rx_ctx_inactive);
2228 			break;
2229 		}
2230 		/*
2231 		 * Reset client set fields to their default values
2232 		 */
2233 		bzero(&ri, sizeof(ri));
2234 		ri.iri_qsidx = rxq->ifr_id;
2235 		ri.iri_cidx = *cidxp;
2236 		ri.iri_ifp = ctx->ifc_ifp;
2237 		ri.iri_frags = rxq->ifr_frags;
2238 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2239 
2240 		/* in lieu of handling correctly - make sure it isn't being unhandled */
2241 		MPASS(err == 0);
2242 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2243 			*cidxp = ri.iri_cidx;
2244 			/* Update our consumer index */
2245 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2246 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2247 				rxq->ifr_cq_gen = 0;
2248 			}
2249 			/* was this only a completion queue message? */
2250 			if (__predict_false(ri.iri_nfrags == 0))
2251 				continue;
2252 		}
2253 		MPASS(ri.iri_nfrags != 0);
2254 		MPASS(ri.iri_len != 0);
2255 
2256 		/* will advance the cidx on the corresponding free lists */
2257 		m = iflib_rxd_pkt_get(rxq, &ri);
2258 		if (avail == 0 && budget_left)
2259 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2260 
2261 		if (__predict_false(m == NULL)) {
2262 			DBG_COUNTER_INC(rx_mbuf_null);
2263 			continue;
2264 		}
2265 		/* imm_pkt: -- cxgb */
2266 		if (mh == NULL)
2267 			mh = mt = m;
2268 		else {
2269 			mt->m_nextpkt = m;
2270 			mt = m;
2271 		}
2272 	}
2273 	/* make sure that we can refill faster than drain */
2274 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2275 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2276 
2277 	ifp = ctx->ifc_ifp;
2278 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2279 	while (mh != NULL) {
2280 		m = mh;
2281 		mh = mh->m_nextpkt;
2282 		m->m_nextpkt = NULL;
2283 		rx_bytes += m->m_pkthdr.len;
2284 		rx_pkts++;
2285 #if defined(INET6) || defined(INET)
2286 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2287 			continue;
2288 #endif
2289 		DBG_COUNTER_INC(rx_if_input);
2290 		ifp->if_input(ifp, m);
2291 	}
2292 
2293 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2294 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2295 
2296 	/*
2297 	 * Flush any outstanding LRO work
2298 	 */
2299 #if defined(INET6) || defined(INET)
2300 	tcp_lro_flush_all(&rxq->ifr_lc);
2301 #endif
2302 	if (avail)
2303 		return true;
2304 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2305 }
2306 
2307 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2308 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2309 #define TXQ_MAX_DB_DEFERRED(size) (size >> 5)
2310 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2311 
2312 static __inline void
2313 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring)
2314 {
2315 	uint32_t dbval;
2316 
2317 	if (ring || txq->ift_db_pending >=
2318 	    TXQ_MAX_DB_DEFERRED(txq->ift_size)) {
2319 
2320 		/* the lock will only ever be contended in the !min_latency case */
2321 		if (!TXDB_TRYLOCK(txq))
2322 			return;
2323 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2324 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2325 		txq->ift_db_pending = txq->ift_npending = 0;
2326 		TXDB_UNLOCK(txq);
2327 	}
2328 }
2329 
2330 static void
2331 iflib_txd_deferred_db_check(void * arg)
2332 {
2333 	iflib_txq_t txq = arg;
2334 
2335 	/* simple non-zero boolean so use bitwise OR */
2336 	if ((txq->ift_db_pending | txq->ift_npending) &&
2337 	    txq->ift_db_pending >= txq->ift_db_pending_queued)
2338 		iflib_txd_db_check(txq->ift_ctx, txq, TRUE);
2339 	txq->ift_db_pending_queued = 0;
2340 	if (ifmp_ring_is_stalled(txq->ift_br[0]))
2341 		iflib_txq_check_drain(txq, 4);
2342 }
2343 
2344 #ifdef PKT_DEBUG
2345 static void
2346 print_pkt(if_pkt_info_t pi)
2347 {
2348 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2349 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2350 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2351 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2352 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2353 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2354 }
2355 #endif
2356 
2357 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2358 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2359 
2360 static int
2361 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2362 {
2363 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2364 	struct ether_vlan_header *eh;
2365 	struct mbuf *m, *n;
2366 
2367 	n = m = *mp;
2368 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2369 	    M_WRITABLE(m) == 0) {
2370 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2371 			return (ENOMEM);
2372 		} else {
2373 			m_freem(*mp);
2374 			n = *mp = m;
2375 		}
2376 	}
2377 
2378 	/*
2379 	 * Determine where frame payload starts.
2380 	 * Jump over vlan headers if already present,
2381 	 * helpful for QinQ too.
2382 	 */
2383 	if (__predict_false(m->m_len < sizeof(*eh))) {
2384 		txq->ift_pullups++;
2385 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2386 			return (ENOMEM);
2387 	}
2388 	eh = mtod(m, struct ether_vlan_header *);
2389 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2390 		pi->ipi_etype = ntohs(eh->evl_proto);
2391 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2392 	} else {
2393 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2394 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2395 	}
2396 
2397 	switch (pi->ipi_etype) {
2398 #ifdef INET
2399 	case ETHERTYPE_IP:
2400 	{
2401 		struct ip *ip = NULL;
2402 		struct tcphdr *th = NULL;
2403 		int minthlen;
2404 
2405 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2406 		if (__predict_false(m->m_len < minthlen)) {
2407 			/*
2408 			 * if this code bloat is causing too much of a hit
2409 			 * move it to a separate function and mark it noinline
2410 			 */
2411 			if (m->m_len == pi->ipi_ehdrlen) {
2412 				n = m->m_next;
2413 				MPASS(n);
2414 				if (n->m_len >= sizeof(*ip))  {
2415 					ip = (struct ip *)n->m_data;
2416 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2417 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2418 				} else {
2419 					txq->ift_pullups++;
2420 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2421 						return (ENOMEM);
2422 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2423 				}
2424 			} else {
2425 				txq->ift_pullups++;
2426 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2427 					return (ENOMEM);
2428 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2429 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2430 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2431 			}
2432 		} else {
2433 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2434 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2435 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2436 		}
2437 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2438 		pi->ipi_ipproto = ip->ip_p;
2439 		pi->ipi_flags |= IPI_TX_IPV4;
2440 
2441 		if (pi->ipi_csum_flags & CSUM_IP)
2442                        ip->ip_sum = 0;
2443 
2444 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2445 			if (__predict_false(th == NULL)) {
2446 				txq->ift_pullups++;
2447 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2448 					return (ENOMEM);
2449 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2450 			}
2451 			pi->ipi_tcp_hflags = th->th_flags;
2452 			pi->ipi_tcp_hlen = th->th_off << 2;
2453 			pi->ipi_tcp_seq = th->th_seq;
2454 		}
2455 		if (IS_TSO4(pi)) {
2456 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2457 				return (ENXIO);
2458 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2459 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2460 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2461 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2462 				ip->ip_sum = 0;
2463 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2464 			}
2465 		}
2466 		break;
2467 	}
2468 #endif
2469 #ifdef INET6
2470 	case ETHERTYPE_IPV6:
2471 	{
2472 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2473 		struct tcphdr *th;
2474 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2475 
2476 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2477 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2478 				return (ENOMEM);
2479 		}
2480 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2481 
2482 		/* XXX-BZ this will go badly in case of ext hdrs. */
2483 		pi->ipi_ipproto = ip6->ip6_nxt;
2484 		pi->ipi_flags |= IPI_TX_IPV6;
2485 
2486 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2487 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2488 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2489 					return (ENOMEM);
2490 			}
2491 			pi->ipi_tcp_hflags = th->th_flags;
2492 			pi->ipi_tcp_hlen = th->th_off << 2;
2493 		}
2494 		if (IS_TSO6(pi)) {
2495 
2496 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2497 				return (ENXIO);
2498 			/*
2499 			 * The corresponding flag is set by the stack in the IPv4
2500 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2501 			 * So, set it here because the rest of the flow requires it.
2502 			 */
2503 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2504 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2505 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2506 		}
2507 		break;
2508 	}
2509 #endif
2510 	default:
2511 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2512 		pi->ipi_ip_hlen = 0;
2513 		break;
2514 	}
2515 	*mp = m;
2516 
2517 	return (0);
2518 }
2519 
2520 static  __noinline  struct mbuf *
2521 collapse_pkthdr(struct mbuf *m0)
2522 {
2523 	struct mbuf *m, *m_next, *tmp;
2524 
2525 	m = m0;
2526 	m_next = m->m_next;
2527 	while (m_next != NULL && m_next->m_len == 0) {
2528 		m = m_next;
2529 		m->m_next = NULL;
2530 		m_free(m);
2531 		m_next = m_next->m_next;
2532 	}
2533 	m = m0;
2534 	m->m_next = m_next;
2535 	if ((m_next->m_flags & M_EXT) == 0) {
2536 		m = m_defrag(m, M_NOWAIT);
2537 	} else {
2538 		tmp = m_next->m_next;
2539 		memcpy(m_next, m, MPKTHSIZE);
2540 		m = m_next;
2541 		m->m_next = tmp;
2542 	}
2543 	return (m);
2544 }
2545 
2546 /*
2547  * If dodgy hardware rejects the scatter gather chain we've handed it
2548  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2549  * m_defrag'd mbufs
2550  */
2551 static __noinline struct mbuf *
2552 iflib_remove_mbuf(iflib_txq_t txq)
2553 {
2554 	int ntxd, i, pidx;
2555 	struct mbuf *m, *mh, **ifsd_m;
2556 
2557 	pidx = txq->ift_pidx;
2558 	ifsd_m = txq->ift_sds.ifsd_m;
2559 	ntxd = txq->ift_size;
2560 	mh = m = ifsd_m[pidx];
2561 	ifsd_m[pidx] = NULL;
2562 #if MEMORY_LOGGING
2563 	txq->ift_dequeued++;
2564 #endif
2565 	i = 1;
2566 
2567 	while (m) {
2568 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2569 #if MEMORY_LOGGING
2570 		txq->ift_dequeued++;
2571 #endif
2572 		m = m->m_next;
2573 		i++;
2574 	}
2575 	return (mh);
2576 }
2577 
2578 static int
2579 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2580 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2581 			  int max_segs, int flags)
2582 {
2583 	if_ctx_t ctx;
2584 	if_shared_ctx_t		sctx;
2585 	if_softc_ctx_t		scctx;
2586 	int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2587 	struct mbuf *m, *tmp, **ifsd_m, **mp;
2588 
2589 	m = *m0;
2590 
2591 	/*
2592 	 * Please don't ever do this
2593 	 */
2594 	if (__predict_false(m->m_len == 0))
2595 		*m0 = m = collapse_pkthdr(m);
2596 
2597 	ctx = txq->ift_ctx;
2598 	sctx = ctx->ifc_sctx;
2599 	scctx = &ctx->ifc_softc_ctx;
2600 	ifsd_m = txq->ift_sds.ifsd_m;
2601 	ntxd = txq->ift_size;
2602 	pidx = txq->ift_pidx;
2603 	if (map != NULL) {
2604 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2605 
2606 		err = bus_dmamap_load_mbuf_sg(tag, map,
2607 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2608 		if (err)
2609 			return (err);
2610 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2611 		i = 0;
2612 		next = pidx;
2613 		mask = (txq->ift_size-1);
2614 		m = *m0;
2615 		do {
2616 			mp = &ifsd_m[next];
2617 			*mp = m;
2618 			m = m->m_next;
2619 			if (__predict_false((*mp)->m_len == 0)) {
2620 				m_free(*mp);
2621 				*mp = NULL;
2622 			} else
2623 				next = (pidx + i) & (ntxd-1);
2624 		} while (m != NULL);
2625 	} else {
2626 		int buflen, sgsize, max_sgsize;
2627 		vm_offset_t vaddr;
2628 		vm_paddr_t curaddr;
2629 
2630 		count = i = 0;
2631 		maxsegsz = sctx->isc_tx_maxsize;
2632 		m = *m0;
2633 		do {
2634 			if (__predict_false(m->m_len <= 0)) {
2635 				tmp = m;
2636 				m = m->m_next;
2637 				tmp->m_next = NULL;
2638 				m_free(tmp);
2639 				continue;
2640 			}
2641 			buflen = m->m_len;
2642 			vaddr = (vm_offset_t)m->m_data;
2643 			/*
2644 			 * see if we can't be smarter about physically
2645 			 * contiguous mappings
2646 			 */
2647 			next = (pidx + count) & (ntxd-1);
2648 			MPASS(ifsd_m[next] == NULL);
2649 #if MEMORY_LOGGING
2650 			txq->ift_enqueued++;
2651 #endif
2652 			ifsd_m[next] = m;
2653 			while (buflen > 0) {
2654 				max_sgsize = MIN(buflen, maxsegsz);
2655 				curaddr = pmap_kextract(vaddr);
2656 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2657 				sgsize = MIN(sgsize, max_sgsize);
2658 				segs[i].ds_addr = curaddr;
2659 				segs[i].ds_len = sgsize;
2660 				vaddr += sgsize;
2661 				buflen -= sgsize;
2662 				i++;
2663 				if (i >= max_segs)
2664 					goto err;
2665 			}
2666 			count++;
2667 			tmp = m;
2668 			m = m->m_next;
2669 		} while (m != NULL);
2670 		*nsegs = i;
2671 	}
2672 	return (0);
2673 err:
2674 	*m0 = iflib_remove_mbuf(txq);
2675 	return (EFBIG);
2676 }
2677 
2678 static int
2679 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2680 {
2681 	if_ctx_t		ctx;
2682 	if_shared_ctx_t		sctx;
2683 	if_softc_ctx_t		scctx;
2684 	bus_dma_segment_t	*segs;
2685 	struct mbuf		*m_head;
2686 	bus_dmamap_t		map;
2687 	struct if_pkt_info	pi;
2688 	int remap = 0;
2689 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
2690 	bus_dma_tag_t desc_tag;
2691 
2692 	segs = txq->ift_segs;
2693 	ctx = txq->ift_ctx;
2694 	sctx = ctx->ifc_sctx;
2695 	scctx = &ctx->ifc_softc_ctx;
2696 	segs = txq->ift_segs;
2697 	ntxd = txq->ift_size;
2698 	m_head = *m_headp;
2699 	map = NULL;
2700 
2701 	/*
2702 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
2703 	 */
2704 	cidx = txq->ift_cidx;
2705 	pidx = txq->ift_pidx;
2706 	next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
2707 
2708 	/* prefetch the next cache line of mbuf pointers and flags */
2709 	prefetch(&txq->ift_sds.ifsd_m[next]);
2710 	if (txq->ift_sds.ifsd_map != NULL) {
2711 		prefetch(&txq->ift_sds.ifsd_map[next]);
2712 		map = txq->ift_sds.ifsd_map[pidx];
2713 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
2714 		prefetch(&txq->ift_sds.ifsd_flags[next]);
2715 	}
2716 
2717 
2718 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
2719 		desc_tag = txq->ift_tso_desc_tag;
2720 		max_segs = scctx->isc_tx_tso_segments_max;
2721 	} else {
2722 		desc_tag = txq->ift_desc_tag;
2723 		max_segs = scctx->isc_tx_nsegments;
2724 	}
2725 	m_head = *m_headp;
2726 	bzero(&pi, sizeof(pi));
2727 	pi.ipi_len = m_head->m_pkthdr.len;
2728 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
2729 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
2730 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
2731 	pi.ipi_pidx = pidx;
2732 	pi.ipi_qsidx = txq->ift_id;
2733 
2734 	/* deliberate bitwise OR to make one condition */
2735 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
2736 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
2737 			return (err);
2738 		m_head = *m_headp;
2739 	}
2740 
2741 retry:
2742 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
2743 defrag:
2744 	if (__predict_false(err)) {
2745 		switch (err) {
2746 		case EFBIG:
2747 			/* try collapse once and defrag once */
2748 			if (remap == 0)
2749 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
2750 			if (remap == 1)
2751 				m_head = m_defrag(*m_headp, M_NOWAIT);
2752 			remap++;
2753 			if (__predict_false(m_head == NULL))
2754 				goto defrag_failed;
2755 			txq->ift_mbuf_defrag++;
2756 			*m_headp = m_head;
2757 			goto retry;
2758 			break;
2759 		case ENOMEM:
2760 			txq->ift_no_tx_dma_setup++;
2761 			break;
2762 		default:
2763 			txq->ift_no_tx_dma_setup++;
2764 			m_freem(*m_headp);
2765 			DBG_COUNTER_INC(tx_frees);
2766 			*m_headp = NULL;
2767 			break;
2768 		}
2769 		txq->ift_map_failed++;
2770 		DBG_COUNTER_INC(encap_load_mbuf_fail);
2771 		return (err);
2772 	}
2773 
2774 	/*
2775 	 * XXX assumes a 1 to 1 relationship between segments and
2776 	 *        descriptors - this does not hold true on all drivers, e.g.
2777 	 *        cxgb
2778 	 */
2779 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
2780 		txq->ift_no_desc_avail++;
2781 		if (map != NULL)
2782 			bus_dmamap_unload(desc_tag, map);
2783 		DBG_COUNTER_INC(encap_txq_avail_fail);
2784 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
2785 			GROUPTASK_ENQUEUE(&txq->ift_task);
2786 		return (ENOBUFS);
2787 	}
2788 	pi.ipi_segs = segs;
2789 	pi.ipi_nsegs = nsegs;
2790 
2791 	MPASS(pidx >= 0 && pidx < txq->ift_size);
2792 #ifdef PKT_DEBUG
2793 	print_pkt(&pi);
2794 #endif
2795 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
2796 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2797 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2798 
2799 		DBG_COUNTER_INC(tx_encap);
2800 		MPASS(pi.ipi_new_pidx >= 0 &&
2801 		    pi.ipi_new_pidx < txq->ift_size);
2802 
2803 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
2804 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
2805 			ndesc += txq->ift_size;
2806 			txq->ift_gen = 1;
2807 		}
2808 		/*
2809 		 * drivers can need as many as
2810 		 * two sentinels
2811 		 */
2812 		MPASS(ndesc <= pi.ipi_nsegs + 2);
2813 		MPASS(pi.ipi_new_pidx != pidx);
2814 		MPASS(ndesc > 0);
2815 		txq->ift_in_use += ndesc;
2816 		/*
2817 		 * We update the last software descriptor again here because there may
2818 		 * be a sentinel and/or there may be more mbufs than segments
2819 		 */
2820 		txq->ift_pidx = pi.ipi_new_pidx;
2821 		txq->ift_npending += pi.ipi_ndescs;
2822 	} else if (__predict_false(err == EFBIG && remap < 2)) {
2823 		*m_headp = m_head = iflib_remove_mbuf(txq);
2824 		remap = 1;
2825 		txq->ift_txd_encap_efbig++;
2826 		goto defrag;
2827 	} else
2828 		DBG_COUNTER_INC(encap_txd_encap_fail);
2829 	return (err);
2830 
2831 defrag_failed:
2832 	txq->ift_mbuf_defrag_failed++;
2833 	txq->ift_map_failed++;
2834 	m_freem(*m_headp);
2835 	DBG_COUNTER_INC(tx_frees);
2836 	*m_headp = NULL;
2837 	return (ENOMEM);
2838 }
2839 
2840 /* forward compatibility for cxgb */
2841 #define FIRST_QSET(ctx) 0
2842 
2843 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2844 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2845 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2846 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2847 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2848 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2849 
2850 
2851 
2852 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring
2853  * doorbell writes
2854  *
2855  * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic
2856  */
2857 #define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2)
2858 
2859 static inline int
2860 iflib_txq_min_occupancy(iflib_txq_t txq)
2861 {
2862 	if_ctx_t ctx;
2863 
2864 	ctx = txq->ift_ctx;
2865 	return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx,
2866 	    txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) +
2867 	    MAX_TX_DESC(ctx));
2868 }
2869 
2870 static void
2871 iflib_tx_desc_free(iflib_txq_t txq, int n)
2872 {
2873 	int hasmap;
2874 	uint32_t qsize, cidx, mask, gen;
2875 	struct mbuf *m, **ifsd_m;
2876 	uint8_t *ifsd_flags;
2877 	bus_dmamap_t *ifsd_map;
2878 
2879 	cidx = txq->ift_cidx;
2880 	gen = txq->ift_gen;
2881 	qsize = txq->ift_size;
2882 	mask = qsize-1;
2883 	hasmap = txq->ift_sds.ifsd_map != NULL;
2884 	ifsd_flags = txq->ift_sds.ifsd_flags;
2885 	ifsd_m = txq->ift_sds.ifsd_m;
2886 	ifsd_map = txq->ift_sds.ifsd_map;
2887 
2888 	while (n--) {
2889 		prefetch(ifsd_m[(cidx + 3) & mask]);
2890 		prefetch(ifsd_m[(cidx + 4) & mask]);
2891 
2892 		if (ifsd_m[cidx] != NULL) {
2893 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
2894 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
2895 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
2896 				/*
2897 				 * does it matter if it's not the TSO tag? If so we'll
2898 				 * have to add the type to flags
2899 				 */
2900 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
2901 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
2902 			}
2903 			if ((m = ifsd_m[cidx]) != NULL) {
2904 				/* XXX we don't support any drivers that batch packets yet */
2905 				MPASS(m->m_nextpkt == NULL);
2906 
2907 				m_free(m);
2908 				ifsd_m[cidx] = NULL;
2909 #if MEMORY_LOGGING
2910 				txq->ift_dequeued++;
2911 #endif
2912 				DBG_COUNTER_INC(tx_frees);
2913 			}
2914 		}
2915 		if (__predict_false(++cidx == qsize)) {
2916 			cidx = 0;
2917 			gen = 0;
2918 		}
2919 	}
2920 	txq->ift_cidx = cidx;
2921 	txq->ift_gen = gen;
2922 }
2923 
2924 static __inline int
2925 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
2926 {
2927 	int reclaim;
2928 	if_ctx_t ctx = txq->ift_ctx;
2929 
2930 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
2931 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
2932 
2933 	/*
2934 	 * Need a rate-limiting check so that this isn't called every time
2935 	 */
2936 	iflib_tx_credits_update(ctx, txq);
2937 	reclaim = DESC_RECLAIMABLE(txq);
2938 
2939 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
2940 #ifdef INVARIANTS
2941 		if (iflib_verbose_debug) {
2942 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
2943 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
2944 			       reclaim, thresh);
2945 
2946 		}
2947 #endif
2948 		return (0);
2949 	}
2950 	iflib_tx_desc_free(txq, reclaim);
2951 	txq->ift_cleaned += reclaim;
2952 	txq->ift_in_use -= reclaim;
2953 
2954 	if (txq->ift_active == FALSE)
2955 		txq->ift_active = TRUE;
2956 
2957 	return (reclaim);
2958 }
2959 
2960 static struct mbuf **
2961 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset)
2962 {
2963 
2964 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)]));
2965 }
2966 
2967 static void
2968 iflib_txq_check_drain(iflib_txq_t txq, int budget)
2969 {
2970 
2971 	ifmp_ring_check_drainage(txq->ift_br[0], budget);
2972 }
2973 
2974 static uint32_t
2975 iflib_txq_can_drain(struct ifmp_ring *r)
2976 {
2977 	iflib_txq_t txq = r->cookie;
2978 	if_ctx_t ctx = txq->ift_ctx;
2979 
2980 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
2981 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false));
2982 }
2983 
2984 static uint32_t
2985 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2986 {
2987 	iflib_txq_t txq = r->cookie;
2988 	if_ctx_t ctx = txq->ift_ctx;
2989 	if_t ifp = ctx->ifc_ifp;
2990 	struct mbuf **mp, *m;
2991 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used;
2992 
2993 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
2994 			    !LINK_ACTIVE(ctx))) {
2995 		DBG_COUNTER_INC(txq_drain_notready);
2996 		return (0);
2997 	}
2998 
2999 	avail = IDXDIFF(pidx, cidx, r->size);
3000 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3001 		DBG_COUNTER_INC(txq_drain_flushing);
3002 		for (i = 0; i < avail; i++) {
3003 			m_free(r->items[(cidx + i) & (r->size-1)]);
3004 			r->items[(cidx + i) & (r->size-1)] = NULL;
3005 		}
3006 		return (avail);
3007 	}
3008 	iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3009 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3010 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3011 		CALLOUT_LOCK(txq);
3012 		callout_stop(&txq->ift_timer);
3013 		callout_stop(&txq->ift_db_check);
3014 		CALLOUT_UNLOCK(txq);
3015 		DBG_COUNTER_INC(txq_drain_oactive);
3016 		return (0);
3017 	}
3018 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3019 	count = MIN(avail, TX_BATCH_SIZE);
3020 #ifdef INVARIANTS
3021 	if (iflib_verbose_debug)
3022 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3023 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3024 #endif
3025 
3026 	for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) {
3027 		mp = _ring_peek_one(r, cidx, i);
3028 		MPASS(mp != NULL && *mp != NULL);
3029 		in_use_prev = txq->ift_in_use;
3030 		if ((err = iflib_encap(txq, mp)) == ENOBUFS) {
3031 			DBG_COUNTER_INC(txq_drain_encapfail);
3032 			/* no room - bail out */
3033 			break;
3034 		}
3035 		consumed++;
3036 		if (err) {
3037 			DBG_COUNTER_INC(txq_drain_encapfail);
3038 			/* we can't send this packet - skip it */
3039 			continue;
3040 		}
3041 		pkt_sent++;
3042 		m = *mp;
3043 		DBG_COUNTER_INC(tx_sent);
3044 		bytes_sent += m->m_pkthdr.len;
3045 		if (m->m_flags & M_MCAST)
3046 			mcast_sent++;
3047 
3048 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3049 		desc_used += (txq->ift_in_use - in_use_prev);
3050 		iflib_txd_db_check(ctx, txq, FALSE);
3051 		ETHER_BPF_MTAP(ifp, m);
3052 		if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3053 			break;
3054 
3055 		if (desc_used >= TXQ_MAX_DB_CONSUMED(txq->ift_size))
3056 			break;
3057 	}
3058 
3059 	if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending)
3060 		iflib_txd_db_check(ctx, txq, TRUE);
3061 	else if ((txq->ift_db_pending || TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2) &&
3062 		 (callout_pending(&txq->ift_db_check) == 0)) {
3063 		txq->ift_db_pending_queued = txq->ift_db_pending;
3064 		callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check,
3065 				 txq, txq->ift_db_check.c_cpu);
3066 	}
3067 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3068 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3069 	if (mcast_sent)
3070 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3071 #ifdef INVARIANTS
3072 	if (iflib_verbose_debug)
3073 		printf("consumed=%d\n", consumed);
3074 #endif
3075 	return (consumed);
3076 }
3077 
3078 static uint32_t
3079 iflib_txq_drain_always(struct ifmp_ring *r)
3080 {
3081 	return (1);
3082 }
3083 
3084 static uint32_t
3085 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3086 {
3087 	int i, avail;
3088 	struct mbuf **mp;
3089 	iflib_txq_t txq;
3090 
3091 	txq = r->cookie;
3092 
3093 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3094 	CALLOUT_LOCK(txq);
3095 	callout_stop(&txq->ift_timer);
3096 	callout_stop(&txq->ift_db_check);
3097 	CALLOUT_UNLOCK(txq);
3098 
3099 	avail = IDXDIFF(pidx, cidx, r->size);
3100 	for (i = 0; i < avail; i++) {
3101 		mp = _ring_peek_one(r, cidx, i);
3102 		m_freem(*mp);
3103 	}
3104 	MPASS(ifmp_ring_is_stalled(r) == 0);
3105 	return (avail);
3106 }
3107 
3108 static void
3109 iflib_ifmp_purge(iflib_txq_t txq)
3110 {
3111 	struct ifmp_ring *r;
3112 
3113 	r = txq->ift_br[0];
3114 	r->drain = iflib_txq_drain_free;
3115 	r->can_drain = iflib_txq_drain_always;
3116 
3117 	ifmp_ring_check_drainage(r, r->size);
3118 
3119 	r->drain = iflib_txq_drain;
3120 	r->can_drain = iflib_txq_can_drain;
3121 }
3122 
3123 static void
3124 _task_fn_tx(void *context)
3125 {
3126 	iflib_txq_t txq = context;
3127 	if_ctx_t ctx = txq->ift_ctx;
3128 
3129 #ifdef IFLIB_DIAGNOSTICS
3130 	txq->ift_cpu_exec_count[curcpu]++;
3131 #endif
3132 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3133 		return;
3134 	ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3135 }
3136 
3137 static void
3138 _task_fn_rx(void *context)
3139 {
3140 	iflib_rxq_t rxq = context;
3141 	if_ctx_t ctx = rxq->ifr_ctx;
3142 	bool more;
3143 	int rc;
3144 
3145 #ifdef IFLIB_DIAGNOSTICS
3146 	rxq->ifr_cpu_exec_count[curcpu]++;
3147 #endif
3148 	DBG_COUNTER_INC(task_fn_rxs);
3149 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3150 		return;
3151 
3152 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3153 		if (ctx->ifc_flags & IFC_LEGACY)
3154 			IFDI_INTR_ENABLE(ctx);
3155 		else {
3156 			DBG_COUNTER_INC(rx_intr_enables);
3157 			rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3158 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3159 		}
3160 	}
3161 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3162 		return;
3163 	if (more)
3164 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3165 }
3166 
3167 static void
3168 _task_fn_admin(void *context)
3169 {
3170 	if_ctx_t ctx = context;
3171 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3172 	iflib_txq_t txq;
3173 	int i;
3174 
3175 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3176 		return;
3177 
3178 	CTX_LOCK(ctx);
3179 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3180 		CALLOUT_LOCK(txq);
3181 		callout_stop(&txq->ift_timer);
3182 		CALLOUT_UNLOCK(txq);
3183 	}
3184 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3185 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3186 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3187 	IFDI_LINK_INTR_ENABLE(ctx);
3188 	CTX_UNLOCK(ctx);
3189 
3190 	if (LINK_ACTIVE(ctx) == 0)
3191 		return;
3192 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3193 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3194 }
3195 
3196 
3197 static void
3198 _task_fn_iov(void *context)
3199 {
3200 	if_ctx_t ctx = context;
3201 
3202 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3203 		return;
3204 
3205 	CTX_LOCK(ctx);
3206 	IFDI_VFLR_HANDLE(ctx);
3207 	CTX_UNLOCK(ctx);
3208 }
3209 
3210 static int
3211 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3212 {
3213 	int err;
3214 	if_int_delay_info_t info;
3215 	if_ctx_t ctx;
3216 
3217 	info = (if_int_delay_info_t)arg1;
3218 	ctx = info->iidi_ctx;
3219 	info->iidi_req = req;
3220 	info->iidi_oidp = oidp;
3221 	CTX_LOCK(ctx);
3222 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3223 	CTX_UNLOCK(ctx);
3224 	return (err);
3225 }
3226 
3227 /*********************************************************************
3228  *
3229  *  IFNET FUNCTIONS
3230  *
3231  **********************************************************************/
3232 
3233 static void
3234 iflib_if_init_locked(if_ctx_t ctx)
3235 {
3236 	iflib_stop(ctx);
3237 	iflib_init_locked(ctx);
3238 }
3239 
3240 
3241 static void
3242 iflib_if_init(void *arg)
3243 {
3244 	if_ctx_t ctx = arg;
3245 
3246 	CTX_LOCK(ctx);
3247 	iflib_if_init_locked(ctx);
3248 	CTX_UNLOCK(ctx);
3249 }
3250 
3251 static int
3252 iflib_if_transmit(if_t ifp, struct mbuf *m)
3253 {
3254 	if_ctx_t	ctx = if_getsoftc(ifp);
3255 
3256 	iflib_txq_t txq;
3257 	int err, qidx;
3258 
3259 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3260 		DBG_COUNTER_INC(tx_frees);
3261 		m_freem(m);
3262 		return (ENOBUFS);
3263 	}
3264 
3265 	MPASS(m->m_nextpkt == NULL);
3266 	qidx = 0;
3267 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3268 		qidx = QIDX(ctx, m);
3269 	/*
3270 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3271 	 */
3272 	txq = &ctx->ifc_txqs[qidx];
3273 
3274 #ifdef DRIVER_BACKPRESSURE
3275 	if (txq->ift_closed) {
3276 		while (m != NULL) {
3277 			next = m->m_nextpkt;
3278 			m->m_nextpkt = NULL;
3279 			m_freem(m);
3280 			m = next;
3281 		}
3282 		return (ENOBUFS);
3283 	}
3284 #endif
3285 #ifdef notyet
3286 	qidx = count = 0;
3287 	mp = marr;
3288 	next = m;
3289 	do {
3290 		count++;
3291 		next = next->m_nextpkt;
3292 	} while (next != NULL);
3293 
3294 	if (count > nitems(marr))
3295 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3296 			/* XXX check nextpkt */
3297 			m_freem(m);
3298 			/* XXX simplify for now */
3299 			DBG_COUNTER_INC(tx_frees);
3300 			return (ENOBUFS);
3301 		}
3302 	for (next = m, i = 0; next != NULL; i++) {
3303 		mp[i] = next;
3304 		next = next->m_nextpkt;
3305 		mp[i]->m_nextpkt = NULL;
3306 	}
3307 #endif
3308 	DBG_COUNTER_INC(tx_seen);
3309 	err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE);
3310 
3311 	if (err) {
3312 		GROUPTASK_ENQUEUE(&txq->ift_task);
3313 		/* support forthcoming later */
3314 #ifdef DRIVER_BACKPRESSURE
3315 		txq->ift_closed = TRUE;
3316 #endif
3317 		ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3318 		m_freem(m);
3319 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3320 		GROUPTASK_ENQUEUE(&txq->ift_task);
3321 	}
3322 
3323 	return (err);
3324 }
3325 
3326 static void
3327 iflib_if_qflush(if_t ifp)
3328 {
3329 	if_ctx_t ctx = if_getsoftc(ifp);
3330 	iflib_txq_t txq = ctx->ifc_txqs;
3331 	int i;
3332 
3333 	CTX_LOCK(ctx);
3334 	ctx->ifc_flags |= IFC_QFLUSH;
3335 	CTX_UNLOCK(ctx);
3336 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3337 		while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0])))
3338 			iflib_txq_check_drain(txq, 0);
3339 	CTX_LOCK(ctx);
3340 	ctx->ifc_flags &= ~IFC_QFLUSH;
3341 	CTX_UNLOCK(ctx);
3342 
3343 	if_qflush(ifp);
3344 }
3345 
3346 
3347 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3348 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3349 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3350 
3351 static int
3352 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3353 {
3354 	if_ctx_t ctx = if_getsoftc(ifp);
3355 	struct ifreq	*ifr = (struct ifreq *)data;
3356 #if defined(INET) || defined(INET6)
3357 	struct ifaddr	*ifa = (struct ifaddr *)data;
3358 #endif
3359 	bool		avoid_reset = FALSE;
3360 	int		err = 0, reinit = 0, bits;
3361 
3362 	switch (command) {
3363 	case SIOCSIFADDR:
3364 #ifdef INET
3365 		if (ifa->ifa_addr->sa_family == AF_INET)
3366 			avoid_reset = TRUE;
3367 #endif
3368 #ifdef INET6
3369 		if (ifa->ifa_addr->sa_family == AF_INET6)
3370 			avoid_reset = TRUE;
3371 #endif
3372 		/*
3373 		** Calling init results in link renegotiation,
3374 		** so we avoid doing it when possible.
3375 		*/
3376 		if (avoid_reset) {
3377 			if_setflagbits(ifp, IFF_UP,0);
3378 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3379 				reinit = 1;
3380 #ifdef INET
3381 			if (!(if_getflags(ifp) & IFF_NOARP))
3382 				arp_ifinit(ifp, ifa);
3383 #endif
3384 		} else
3385 			err = ether_ioctl(ifp, command, data);
3386 		break;
3387 	case SIOCSIFMTU:
3388 		CTX_LOCK(ctx);
3389 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3390 			CTX_UNLOCK(ctx);
3391 			break;
3392 		}
3393 		bits = if_getdrvflags(ifp);
3394 		/* stop the driver and free any clusters before proceeding */
3395 		iflib_stop(ctx);
3396 
3397 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3398 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3399 				ctx->ifc_flags |= IFC_MULTISEG;
3400 			else
3401 				ctx->ifc_flags &= ~IFC_MULTISEG;
3402 			err = if_setmtu(ifp, ifr->ifr_mtu);
3403 		}
3404 		iflib_init_locked(ctx);
3405 		if_setdrvflags(ifp, bits);
3406 		CTX_UNLOCK(ctx);
3407 		break;
3408 	case SIOCSIFFLAGS:
3409 		CTX_LOCK(ctx);
3410 		if (if_getflags(ifp) & IFF_UP) {
3411 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3412 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3413 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3414 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3415 				}
3416 			} else
3417 				reinit = 1;
3418 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3419 			iflib_stop(ctx);
3420 		}
3421 		ctx->ifc_if_flags = if_getflags(ifp);
3422 		CTX_UNLOCK(ctx);
3423 		break;
3424 	case SIOCADDMULTI:
3425 	case SIOCDELMULTI:
3426 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3427 			CTX_LOCK(ctx);
3428 			IFDI_INTR_DISABLE(ctx);
3429 			IFDI_MULTI_SET(ctx);
3430 			IFDI_INTR_ENABLE(ctx);
3431 			CTX_UNLOCK(ctx);
3432 		}
3433 		break;
3434 	case SIOCSIFMEDIA:
3435 		CTX_LOCK(ctx);
3436 		IFDI_MEDIA_SET(ctx);
3437 		CTX_UNLOCK(ctx);
3438 		/* falls thru */
3439 	case SIOCGIFMEDIA:
3440 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3441 		break;
3442 	case SIOCGI2C:
3443 	{
3444 		struct ifi2creq i2c;
3445 
3446 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3447 		if (err != 0)
3448 			break;
3449 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3450 			err = EINVAL;
3451 			break;
3452 		}
3453 		if (i2c.len > sizeof(i2c.data)) {
3454 			err = EINVAL;
3455 			break;
3456 		}
3457 
3458 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3459 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3460 		break;
3461 	}
3462 	case SIOCSIFCAP:
3463 	{
3464 		int mask, setmask;
3465 
3466 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3467 		setmask = 0;
3468 #ifdef TCP_OFFLOAD
3469 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3470 #endif
3471 		setmask |= (mask & IFCAP_FLAGS);
3472 
3473 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3474 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3475 		if ((mask & IFCAP_WOL) &&
3476 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3477 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3478 		if_vlancap(ifp);
3479 		/*
3480 		 * want to ensure that traffic has stopped before we change any of the flags
3481 		 */
3482 		if (setmask) {
3483 			CTX_LOCK(ctx);
3484 			bits = if_getdrvflags(ifp);
3485 			if (bits & IFF_DRV_RUNNING)
3486 				iflib_stop(ctx);
3487 			if_togglecapenable(ifp, setmask);
3488 			if (bits & IFF_DRV_RUNNING)
3489 				iflib_init_locked(ctx);
3490 			if_setdrvflags(ifp, bits);
3491 			CTX_UNLOCK(ctx);
3492 		}
3493 		break;
3494 	    }
3495 	case SIOCGPRIVATE_0:
3496 	case SIOCSDRVSPEC:
3497 	case SIOCGDRVSPEC:
3498 		CTX_LOCK(ctx);
3499 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3500 		CTX_UNLOCK(ctx);
3501 		break;
3502 	default:
3503 		err = ether_ioctl(ifp, command, data);
3504 		break;
3505 	}
3506 	if (reinit)
3507 		iflib_if_init(ctx);
3508 	return (err);
3509 }
3510 
3511 static uint64_t
3512 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3513 {
3514 	if_ctx_t ctx = if_getsoftc(ifp);
3515 
3516 	return (IFDI_GET_COUNTER(ctx, cnt));
3517 }
3518 
3519 /*********************************************************************
3520  *
3521  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3522  *
3523  **********************************************************************/
3524 
3525 static void
3526 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3527 {
3528 	if_ctx_t ctx = if_getsoftc(ifp);
3529 
3530 	if ((void *)ctx != arg)
3531 		return;
3532 
3533 	if ((vtag == 0) || (vtag > 4095))
3534 		return;
3535 
3536 	CTX_LOCK(ctx);
3537 	IFDI_VLAN_REGISTER(ctx, vtag);
3538 	/* Re-init to load the changes */
3539 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3540 		iflib_init_locked(ctx);
3541 	CTX_UNLOCK(ctx);
3542 }
3543 
3544 static void
3545 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3546 {
3547 	if_ctx_t ctx = if_getsoftc(ifp);
3548 
3549 	if ((void *)ctx != arg)
3550 		return;
3551 
3552 	if ((vtag == 0) || (vtag > 4095))
3553 		return;
3554 
3555 	CTX_LOCK(ctx);
3556 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3557 	/* Re-init to load the changes */
3558 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3559 		iflib_init_locked(ctx);
3560 	CTX_UNLOCK(ctx);
3561 }
3562 
3563 static void
3564 iflib_led_func(void *arg, int onoff)
3565 {
3566 	if_ctx_t ctx = arg;
3567 
3568 	CTX_LOCK(ctx);
3569 	IFDI_LED_FUNC(ctx, onoff);
3570 	CTX_UNLOCK(ctx);
3571 }
3572 
3573 /*********************************************************************
3574  *
3575  *  BUS FUNCTION DEFINITIONS
3576  *
3577  **********************************************************************/
3578 
3579 int
3580 iflib_device_probe(device_t dev)
3581 {
3582 	pci_vendor_info_t *ent;
3583 
3584 	uint16_t	pci_vendor_id, pci_device_id;
3585 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3586 	uint16_t	pci_rev_id;
3587 	if_shared_ctx_t sctx;
3588 
3589 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3590 		return (ENOTSUP);
3591 
3592 	pci_vendor_id = pci_get_vendor(dev);
3593 	pci_device_id = pci_get_device(dev);
3594 	pci_subvendor_id = pci_get_subvendor(dev);
3595 	pci_subdevice_id = pci_get_subdevice(dev);
3596 	pci_rev_id = pci_get_revid(dev);
3597 	if (sctx->isc_parse_devinfo != NULL)
3598 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3599 
3600 	ent = sctx->isc_vendor_info;
3601 	while (ent->pvi_vendor_id != 0) {
3602 		if (pci_vendor_id != ent->pvi_vendor_id) {
3603 			ent++;
3604 			continue;
3605 		}
3606 		if ((pci_device_id == ent->pvi_device_id) &&
3607 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3608 		     (ent->pvi_subvendor_id == 0)) &&
3609 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3610 		     (ent->pvi_subdevice_id == 0)) &&
3611 		    ((pci_rev_id == ent->pvi_rev_id) ||
3612 		     (ent->pvi_rev_id == 0))) {
3613 
3614 			device_set_desc_copy(dev, ent->pvi_name);
3615 			/* this needs to be changed to zero if the bus probing code
3616 			 * ever stops re-probing on best match because the sctx
3617 			 * may have its values over written by register calls
3618 			 * in subsequent probes
3619 			 */
3620 			return (BUS_PROBE_DEFAULT);
3621 		}
3622 		ent++;
3623 	}
3624 	return (ENXIO);
3625 }
3626 
3627 int
3628 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3629 {
3630 	int err, rid, msix, msix_bar;
3631 	if_ctx_t ctx;
3632 	if_t ifp;
3633 	if_softc_ctx_t scctx;
3634 	int i;
3635 	uint16_t main_txq;
3636 	uint16_t main_rxq;
3637 
3638 
3639 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3640 
3641 	if (sc == NULL) {
3642 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3643 		device_set_softc(dev, ctx);
3644 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3645 	}
3646 
3647 	ctx->ifc_sctx = sctx;
3648 	ctx->ifc_dev = dev;
3649 	ctx->ifc_softc = sc;
3650 
3651 	if ((err = iflib_register(ctx)) != 0) {
3652 		device_printf(dev, "iflib_register failed %d\n", err);
3653 		return (err);
3654 	}
3655 	iflib_add_device_sysctl_pre(ctx);
3656 
3657 	scctx = &ctx->ifc_softc_ctx;
3658 	ifp = ctx->ifc_ifp;
3659 
3660 	/*
3661 	 * XXX sanity check that ntxd & nrxd are a power of 2
3662 	 */
3663 	if (ctx->ifc_sysctl_ntxqs != 0)
3664 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
3665 	if (ctx->ifc_sysctl_nrxqs != 0)
3666 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
3667 
3668 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3669 		if (ctx->ifc_sysctl_ntxds[i] != 0)
3670 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
3671 		else
3672 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
3673 	}
3674 
3675 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3676 		if (ctx->ifc_sysctl_nrxds[i] != 0)
3677 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
3678 		else
3679 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
3680 	}
3681 
3682 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3683 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
3684 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
3685 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
3686 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
3687 		}
3688 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
3689 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
3690 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
3691 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
3692 		}
3693 	}
3694 
3695 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3696 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
3697 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
3698 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
3699 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
3700 		}
3701 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
3702 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
3703 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
3704 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
3705 		}
3706 	}
3707 
3708 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
3709 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
3710 		return (err);
3711 	}
3712 	_iflib_pre_assert(scctx);
3713 	ctx->ifc_txrx = *scctx->isc_txrx;
3714 
3715 #ifdef INVARIANTS
3716 	MPASS(scctx->isc_capenable);
3717 	if (scctx->isc_capenable & IFCAP_TXCSUM)
3718 		MPASS(scctx->isc_tx_csum_flags);
3719 #endif
3720 
3721 	if_setcapabilities(ifp, scctx->isc_capenable);
3722 	if_setcapenable(ifp, scctx->isc_capenable);
3723 
3724 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
3725 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
3726 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
3727 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
3728 
3729 #ifdef ACPI_DMAR
3730 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
3731 		ctx->ifc_flags |= IFC_DMAR;
3732 #endif
3733 
3734 	msix_bar = scctx->isc_msix_bar;
3735 
3736 	if(sctx->isc_flags & IFLIB_HAS_TXCQ)
3737 		main_txq = 1;
3738 	else
3739 		main_txq = 0;
3740 
3741 	if(sctx->isc_flags & IFLIB_HAS_RXCQ)
3742 		main_rxq = 1;
3743 	else
3744 		main_rxq = 0;
3745 
3746 	/* XXX change for per-queue sizes */
3747 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
3748 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
3749 	for (i = 0; i < sctx->isc_nrxqs; i++) {
3750 		if (!powerof2(scctx->isc_nrxd[i])) {
3751 			/* round down instead? */
3752 			device_printf(dev, "# rx descriptors must be a power of 2\n");
3753 			err = EINVAL;
3754 			goto fail;
3755 		}
3756 	}
3757 	for (i = 0; i < sctx->isc_ntxqs; i++) {
3758 		if (!powerof2(scctx->isc_ntxd[i])) {
3759 			device_printf(dev,
3760 			    "# tx descriptors must be a power of 2");
3761 			err = EINVAL;
3762 			goto fail;
3763 		}
3764 	}
3765 
3766 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
3767 	    MAX_SINGLE_PACKET_FRACTION)
3768 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
3769 		    MAX_SINGLE_PACKET_FRACTION);
3770 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
3771 	    MAX_SINGLE_PACKET_FRACTION)
3772 		scctx->isc_tx_tso_segments_max = max(1,
3773 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
3774 
3775 	/*
3776 	 * Protect the stack against modern hardware
3777 	 */
3778 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
3779 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
3780 
3781 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
3782 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
3783 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
3784 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
3785 	if (scctx->isc_rss_table_size == 0)
3786 		scctx->isc_rss_table_size = 64;
3787 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
3788 
3789 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
3790 	/* XXX format name */
3791 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
3792 	/*
3793 	** Now setup MSI or MSI/X, should
3794 	** return us the number of supported
3795 	** vectors. (Will be 1 for MSI)
3796 	*/
3797 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
3798 		msix = scctx->isc_vectors;
3799 	} else if (scctx->isc_msix_bar != 0)
3800 	       /*
3801 		* The simple fact that isc_msix_bar is not 0 does not mean we
3802 		* we have a good value there that is known to work.
3803 		*/
3804 		msix = iflib_msix_init(ctx);
3805 	else {
3806 		scctx->isc_vectors = 1;
3807 		scctx->isc_ntxqsets = 1;
3808 		scctx->isc_nrxqsets = 1;
3809 		scctx->isc_intr = IFLIB_INTR_LEGACY;
3810 		msix = 0;
3811 	}
3812 	/* Get memory for the station queues */
3813 	if ((err = iflib_queues_alloc(ctx))) {
3814 		device_printf(dev, "Unable to allocate queue memory\n");
3815 		goto fail;
3816 	}
3817 
3818 	if ((err = iflib_qset_structures_setup(ctx))) {
3819 		device_printf(dev, "qset structure setup failed %d\n", err);
3820 		goto fail_queues;
3821 	}
3822 
3823 	/*
3824 	 * Group taskqueues aren't properly set up until SMP is started,
3825 	 * so we disable interrupts until we can handle them post
3826 	 * SI_SUB_SMP.
3827 	 *
3828 	 * XXX: disabling interrupts doesn't actually work, at least for
3829 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
3830 	 * we do null handling and depend on this not causing too large an
3831 	 * interrupt storm.
3832 	 */
3833 	IFDI_INTR_DISABLE(ctx);
3834 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
3835 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
3836 		goto fail_intr_free;
3837 	}
3838 	if (msix <= 1) {
3839 		rid = 0;
3840 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
3841 			MPASS(msix == 1);
3842 			rid = 1;
3843 		}
3844 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
3845 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
3846 			goto fail_intr_free;
3847 		}
3848 	}
3849 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
3850 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
3851 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
3852 		goto fail_detach;
3853 	}
3854 	if ((err = iflib_netmap_attach(ctx))) {
3855 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
3856 		goto fail_detach;
3857 	}
3858 	*ctxp = ctx;
3859 
3860 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
3861 	iflib_add_device_sysctl_post(ctx);
3862 	ctx->ifc_flags |= IFC_INIT_DONE;
3863 	return (0);
3864 fail_detach:
3865 	ether_ifdetach(ctx->ifc_ifp);
3866 fail_intr_free:
3867 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
3868 		pci_release_msi(ctx->ifc_dev);
3869 fail_queues:
3870 	/* XXX free queues */
3871 fail:
3872 	IFDI_DETACH(ctx);
3873 	return (err);
3874 }
3875 
3876 int
3877 iflib_device_attach(device_t dev)
3878 {
3879 	if_ctx_t ctx;
3880 	if_shared_ctx_t sctx;
3881 
3882 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3883 		return (ENOTSUP);
3884 
3885 	pci_enable_busmaster(dev);
3886 
3887 	return (iflib_device_register(dev, NULL, sctx, &ctx));
3888 }
3889 
3890 int
3891 iflib_device_deregister(if_ctx_t ctx)
3892 {
3893 	if_t ifp = ctx->ifc_ifp;
3894 	iflib_txq_t txq;
3895 	iflib_rxq_t rxq;
3896 	device_t dev = ctx->ifc_dev;
3897 	int i;
3898 	struct taskqgroup *tqg;
3899 
3900 	/* Make sure VLANS are not using driver */
3901 	if (if_vlantrunkinuse(ifp)) {
3902 		device_printf(dev,"Vlan in use, detach first\n");
3903 		return (EBUSY);
3904 	}
3905 
3906 	CTX_LOCK(ctx);
3907 	ctx->ifc_in_detach = 1;
3908 	iflib_stop(ctx);
3909 	CTX_UNLOCK(ctx);
3910 
3911 	/* Unregister VLAN events */
3912 	if (ctx->ifc_vlan_attach_event != NULL)
3913 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
3914 	if (ctx->ifc_vlan_detach_event != NULL)
3915 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
3916 
3917 	iflib_netmap_detach(ifp);
3918 	ether_ifdetach(ifp);
3919 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
3920 	CTX_LOCK_DESTROY(ctx);
3921 	if (ctx->ifc_led_dev != NULL)
3922 		led_destroy(ctx->ifc_led_dev);
3923 	/* XXX drain any dependent tasks */
3924 	tqg = qgroup_softirq;
3925 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
3926 		callout_drain(&txq->ift_timer);
3927 		callout_drain(&txq->ift_db_check);
3928 		if (txq->ift_task.gt_uniq != NULL)
3929 			taskqgroup_detach(tqg, &txq->ift_task);
3930 	}
3931 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
3932 		if (rxq->ifr_task.gt_uniq != NULL)
3933 			taskqgroup_detach(tqg, &rxq->ifr_task);
3934 	}
3935 	tqg = qgroup_if_config_tqg;
3936 	if (ctx->ifc_admin_task.gt_uniq != NULL)
3937 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
3938 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
3939 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
3940 
3941 	IFDI_DETACH(ctx);
3942 	device_set_softc(ctx->ifc_dev, NULL);
3943 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
3944 		pci_release_msi(dev);
3945 	}
3946 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
3947 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
3948 	}
3949 	if (ctx->ifc_msix_mem != NULL) {
3950 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
3951 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
3952 		ctx->ifc_msix_mem = NULL;
3953 	}
3954 
3955 	bus_generic_detach(dev);
3956 	if_free(ifp);
3957 
3958 	iflib_tx_structures_free(ctx);
3959 	iflib_rx_structures_free(ctx);
3960 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
3961 		free(ctx->ifc_softc, M_IFLIB);
3962 	free(ctx, M_IFLIB);
3963 	return (0);
3964 }
3965 
3966 
3967 int
3968 iflib_device_detach(device_t dev)
3969 {
3970 	if_ctx_t ctx = device_get_softc(dev);
3971 
3972 	return (iflib_device_deregister(ctx));
3973 }
3974 
3975 int
3976 iflib_device_suspend(device_t dev)
3977 {
3978 	if_ctx_t ctx = device_get_softc(dev);
3979 
3980 	CTX_LOCK(ctx);
3981 	IFDI_SUSPEND(ctx);
3982 	CTX_UNLOCK(ctx);
3983 
3984 	return bus_generic_suspend(dev);
3985 }
3986 int
3987 iflib_device_shutdown(device_t dev)
3988 {
3989 	if_ctx_t ctx = device_get_softc(dev);
3990 
3991 	CTX_LOCK(ctx);
3992 	IFDI_SHUTDOWN(ctx);
3993 	CTX_UNLOCK(ctx);
3994 
3995 	return bus_generic_suspend(dev);
3996 }
3997 
3998 
3999 int
4000 iflib_device_resume(device_t dev)
4001 {
4002 	if_ctx_t ctx = device_get_softc(dev);
4003 	iflib_txq_t txq = ctx->ifc_txqs;
4004 
4005 	CTX_LOCK(ctx);
4006 	IFDI_RESUME(ctx);
4007 	iflib_init_locked(ctx);
4008 	CTX_UNLOCK(ctx);
4009 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4010 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4011 
4012 	return (bus_generic_resume(dev));
4013 }
4014 
4015 int
4016 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4017 {
4018 	int error;
4019 	if_ctx_t ctx = device_get_softc(dev);
4020 
4021 	CTX_LOCK(ctx);
4022 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4023 	CTX_UNLOCK(ctx);
4024 
4025 	return (error);
4026 }
4027 
4028 void
4029 iflib_device_iov_uninit(device_t dev)
4030 {
4031 	if_ctx_t ctx = device_get_softc(dev);
4032 
4033 	CTX_LOCK(ctx);
4034 	IFDI_IOV_UNINIT(ctx);
4035 	CTX_UNLOCK(ctx);
4036 }
4037 
4038 int
4039 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4040 {
4041 	int error;
4042 	if_ctx_t ctx = device_get_softc(dev);
4043 
4044 	CTX_LOCK(ctx);
4045 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4046 	CTX_UNLOCK(ctx);
4047 
4048 	return (error);
4049 }
4050 
4051 /*********************************************************************
4052  *
4053  *  MODULE FUNCTION DEFINITIONS
4054  *
4055  **********************************************************************/
4056 
4057 /*
4058  * - Start a fast taskqueue thread for each core
4059  * - Start a taskqueue for control operations
4060  */
4061 static int
4062 iflib_module_init(void)
4063 {
4064 	return (0);
4065 }
4066 
4067 static int
4068 iflib_module_event_handler(module_t mod, int what, void *arg)
4069 {
4070 	int err;
4071 
4072 	switch (what) {
4073 	case MOD_LOAD:
4074 		if ((err = iflib_module_init()) != 0)
4075 			return (err);
4076 		break;
4077 	case MOD_UNLOAD:
4078 		return (EBUSY);
4079 	default:
4080 		return (EOPNOTSUPP);
4081 	}
4082 
4083 	return (0);
4084 }
4085 
4086 /*********************************************************************
4087  *
4088  *  PUBLIC FUNCTION DEFINITIONS
4089  *     ordered as in iflib.h
4090  *
4091  **********************************************************************/
4092 
4093 
4094 static void
4095 _iflib_assert(if_shared_ctx_t sctx)
4096 {
4097 	MPASS(sctx->isc_tx_maxsize);
4098 	MPASS(sctx->isc_tx_maxsegsize);
4099 
4100 	MPASS(sctx->isc_rx_maxsize);
4101 	MPASS(sctx->isc_rx_nsegments);
4102 	MPASS(sctx->isc_rx_maxsegsize);
4103 
4104 	MPASS(sctx->isc_nrxd_min[0]);
4105 	MPASS(sctx->isc_nrxd_max[0]);
4106 	MPASS(sctx->isc_nrxd_default[0]);
4107 	MPASS(sctx->isc_ntxd_min[0]);
4108 	MPASS(sctx->isc_ntxd_max[0]);
4109 	MPASS(sctx->isc_ntxd_default[0]);
4110 }
4111 
4112 static void
4113 _iflib_pre_assert(if_softc_ctx_t scctx)
4114 {
4115 
4116 	MPASS(scctx->isc_txrx->ift_txd_encap);
4117 	MPASS(scctx->isc_txrx->ift_txd_flush);
4118 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4119 	MPASS(scctx->isc_txrx->ift_rxd_available);
4120 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4121 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4122 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4123 }
4124 
4125 static int
4126 iflib_register(if_ctx_t ctx)
4127 {
4128 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4129 	driver_t *driver = sctx->isc_driver;
4130 	device_t dev = ctx->ifc_dev;
4131 	if_t ifp;
4132 
4133 	_iflib_assert(sctx);
4134 
4135 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4136 
4137 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4138 	if (ifp == NULL) {
4139 		device_printf(dev, "can not allocate ifnet structure\n");
4140 		return (ENOMEM);
4141 	}
4142 
4143 	/*
4144 	 * Initialize our context's device specific methods
4145 	 */
4146 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4147 	kobj_class_compile((kobj_class_t) driver);
4148 	driver->refs++;
4149 
4150 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4151 	if_setsoftc(ifp, ctx);
4152 	if_setdev(ifp, dev);
4153 	if_setinitfn(ifp, iflib_if_init);
4154 	if_setioctlfn(ifp, iflib_if_ioctl);
4155 	if_settransmitfn(ifp, iflib_if_transmit);
4156 	if_setqflushfn(ifp, iflib_if_qflush);
4157 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4158 
4159 	ctx->ifc_vlan_attach_event =
4160 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4161 							  EVENTHANDLER_PRI_FIRST);
4162 	ctx->ifc_vlan_detach_event =
4163 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4164 							  EVENTHANDLER_PRI_FIRST);
4165 
4166 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4167 					 iflib_media_change, iflib_media_status);
4168 
4169 	return (0);
4170 }
4171 
4172 
4173 static int
4174 iflib_queues_alloc(if_ctx_t ctx)
4175 {
4176 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4177 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4178 	device_t dev = ctx->ifc_dev;
4179 	int nrxqsets = scctx->isc_nrxqsets;
4180 	int ntxqsets = scctx->isc_ntxqsets;
4181 	iflib_txq_t txq;
4182 	iflib_rxq_t rxq;
4183 	iflib_fl_t fl = NULL;
4184 	int i, j, cpu, err, txconf, rxconf;
4185 	iflib_dma_info_t ifdip;
4186 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4187 	uint32_t *txqsizes = scctx->isc_txqsizes;
4188 	uint8_t nrxqs = sctx->isc_nrxqs;
4189 	uint8_t ntxqs = sctx->isc_ntxqs;
4190 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4191 	caddr_t *vaddrs;
4192 	uint64_t *paddrs;
4193 	struct ifmp_ring **brscp;
4194 	int nbuf_rings = 1; /* XXX determine dynamically */
4195 
4196 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4197 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4198 
4199 	brscp = NULL;
4200 	txq = NULL;
4201 	rxq = NULL;
4202 
4203 /* Allocate the TX ring struct memory */
4204 	if (!(txq =
4205 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4206 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4207 		device_printf(dev, "Unable to allocate TX ring memory\n");
4208 		err = ENOMEM;
4209 		goto fail;
4210 	}
4211 
4212 	/* Now allocate the RX */
4213 	if (!(rxq =
4214 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4215 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4216 		device_printf(dev, "Unable to allocate RX ring memory\n");
4217 		err = ENOMEM;
4218 		goto rx_fail;
4219 	}
4220 	if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4221 		device_printf(dev, "Unable to buf_ring_sc * memory\n");
4222 		err = ENOMEM;
4223 		goto rx_fail;
4224 	}
4225 
4226 	ctx->ifc_txqs = txq;
4227 	ctx->ifc_rxqs = rxq;
4228 
4229 	/*
4230 	 * XXX handle allocation failure
4231 	 */
4232 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4233 		/* Set up some basics */
4234 
4235 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4236 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4237 			err = ENOMEM;
4238 			goto err_tx_desc;
4239 		}
4240 		txq->ift_ifdi = ifdip;
4241 		for (j = 0; j < ntxqs; j++, ifdip++) {
4242 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4243 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4244 				err = ENOMEM;
4245 				goto err_tx_desc;
4246 			}
4247 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4248 		}
4249 		txq->ift_ctx = ctx;
4250 		txq->ift_id = i;
4251 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4252 			txq->ift_br_offset = 1;
4253 		} else {
4254 			txq->ift_br_offset = 0;
4255 		}
4256 		/* XXX fix this */
4257 		txq->ift_timer.c_cpu = cpu;
4258 		txq->ift_db_check.c_cpu = cpu;
4259 		txq->ift_nbr = nbuf_rings;
4260 
4261 		if (iflib_txsd_alloc(txq)) {
4262 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4263 			err = ENOMEM;
4264 			goto err_tx_desc;
4265 		}
4266 
4267 		/* Initialize the TX lock */
4268 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4269 		    device_get_nameunit(dev), txq->ift_id);
4270 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4271 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4272 		callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0);
4273 
4274 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4275 			 device_get_nameunit(dev), txq->ift_id);
4276 		TXDB_LOCK_INIT(txq);
4277 
4278 		txq->ift_br = brscp + i*nbuf_rings;
4279 		for (j = 0; j < nbuf_rings; j++) {
4280 			err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain,
4281 					      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4282 			if (err) {
4283 				/* XXX free any allocated rings */
4284 				device_printf(dev, "Unable to allocate buf_ring\n");
4285 				goto err_tx_desc;
4286 			}
4287 		}
4288 	}
4289 
4290 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4291 		/* Set up some basics */
4292 
4293 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4294 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4295 			err = ENOMEM;
4296 			goto err_tx_desc;
4297 		}
4298 
4299 		rxq->ifr_ifdi = ifdip;
4300 		for (j = 0; j < nrxqs; j++, ifdip++) {
4301 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4302 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4303 				err = ENOMEM;
4304 				goto err_tx_desc;
4305 			}
4306 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4307 		}
4308 		rxq->ifr_ctx = ctx;
4309 		rxq->ifr_id = i;
4310 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4311 			rxq->ifr_fl_offset = 1;
4312 		} else {
4313 			rxq->ifr_fl_offset = 0;
4314 		}
4315 		rxq->ifr_nfl = nfree_lists;
4316 		if (!(fl =
4317 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4318 			device_printf(dev, "Unable to allocate free list memory\n");
4319 			err = ENOMEM;
4320 			goto err_tx_desc;
4321 		}
4322 		rxq->ifr_fl = fl;
4323 		for (j = 0; j < nfree_lists; j++) {
4324 			rxq->ifr_fl[j].ifl_rxq = rxq;
4325 			rxq->ifr_fl[j].ifl_id = j;
4326 			rxq->ifr_fl[j].ifl_ifdi =
4327 			    &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4328 		}
4329         /* Allocate receive buffers for the ring*/
4330 		if (iflib_rxsd_alloc(rxq)) {
4331 			device_printf(dev,
4332 			    "Critical Failure setting up receive buffers\n");
4333 			err = ENOMEM;
4334 			goto err_rx_desc;
4335 		}
4336 	}
4337 
4338 	/* TXQs */
4339 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4340 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4341 	for (i = 0; i < ntxqsets; i++) {
4342 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4343 
4344 		for (j = 0; j < ntxqs; j++, di++) {
4345 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4346 			paddrs[i*ntxqs + j] = di->idi_paddr;
4347 		}
4348 	}
4349 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4350 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4351 		iflib_tx_structures_free(ctx);
4352 		free(vaddrs, M_IFLIB);
4353 		free(paddrs, M_IFLIB);
4354 		goto err_rx_desc;
4355 	}
4356 	free(vaddrs, M_IFLIB);
4357 	free(paddrs, M_IFLIB);
4358 
4359 	/* RXQs */
4360 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4361 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4362 	for (i = 0; i < nrxqsets; i++) {
4363 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4364 
4365 		for (j = 0; j < nrxqs; j++, di++) {
4366 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4367 			paddrs[i*nrxqs + j] = di->idi_paddr;
4368 		}
4369 	}
4370 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4371 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4372 		iflib_tx_structures_free(ctx);
4373 		free(vaddrs, M_IFLIB);
4374 		free(paddrs, M_IFLIB);
4375 		goto err_rx_desc;
4376 	}
4377 	free(vaddrs, M_IFLIB);
4378 	free(paddrs, M_IFLIB);
4379 
4380 	return (0);
4381 
4382 /* XXX handle allocation failure changes */
4383 err_rx_desc:
4384 err_tx_desc:
4385 	if (ctx->ifc_rxqs != NULL)
4386 		free(ctx->ifc_rxqs, M_IFLIB);
4387 	ctx->ifc_rxqs = NULL;
4388 	if (ctx->ifc_txqs != NULL)
4389 		free(ctx->ifc_txqs, M_IFLIB);
4390 	ctx->ifc_txqs = NULL;
4391 rx_fail:
4392 	if (brscp != NULL)
4393 		free(brscp, M_IFLIB);
4394 	if (rxq != NULL)
4395 		free(rxq, M_IFLIB);
4396 	if (txq != NULL)
4397 		free(txq, M_IFLIB);
4398 fail:
4399 	return (err);
4400 }
4401 
4402 static int
4403 iflib_tx_structures_setup(if_ctx_t ctx)
4404 {
4405 	iflib_txq_t txq = ctx->ifc_txqs;
4406 	int i;
4407 
4408 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4409 		iflib_txq_setup(txq);
4410 
4411 	return (0);
4412 }
4413 
4414 static void
4415 iflib_tx_structures_free(if_ctx_t ctx)
4416 {
4417 	iflib_txq_t txq = ctx->ifc_txqs;
4418 	int i, j;
4419 
4420 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4421 		iflib_txq_destroy(txq);
4422 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4423 			iflib_dma_free(&txq->ift_ifdi[j]);
4424 	}
4425 	free(ctx->ifc_txqs, M_IFLIB);
4426 	ctx->ifc_txqs = NULL;
4427 	IFDI_QUEUES_FREE(ctx);
4428 }
4429 
4430 /*********************************************************************
4431  *
4432  *  Initialize all receive rings.
4433  *
4434  **********************************************************************/
4435 static int
4436 iflib_rx_structures_setup(if_ctx_t ctx)
4437 {
4438 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4439 	int q;
4440 #if defined(INET6) || defined(INET)
4441 	int i, err;
4442 #endif
4443 
4444 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4445 #if defined(INET6) || defined(INET)
4446 		tcp_lro_free(&rxq->ifr_lc);
4447 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4448 		    TCP_LRO_ENTRIES, min(1024,
4449 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4450 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4451 			goto fail;
4452 		}
4453 		rxq->ifr_lro_enabled = TRUE;
4454 #endif
4455 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4456 	}
4457 	return (0);
4458 #if defined(INET6) || defined(INET)
4459 fail:
4460 	/*
4461 	 * Free RX software descriptors allocated so far, we will only handle
4462 	 * the rings that completed, the failing case will have
4463 	 * cleaned up for itself. 'q' failed, so its the terminus.
4464 	 */
4465 	rxq = ctx->ifc_rxqs;
4466 	for (i = 0; i < q; ++i, rxq++) {
4467 		iflib_rx_sds_free(rxq);
4468 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4469 	}
4470 	return (err);
4471 #endif
4472 }
4473 
4474 /*********************************************************************
4475  *
4476  *  Free all receive rings.
4477  *
4478  **********************************************************************/
4479 static void
4480 iflib_rx_structures_free(if_ctx_t ctx)
4481 {
4482 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4483 
4484 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4485 		iflib_rx_sds_free(rxq);
4486 	}
4487 }
4488 
4489 static int
4490 iflib_qset_structures_setup(if_ctx_t ctx)
4491 {
4492 	int err;
4493 
4494 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4495 		return (err);
4496 
4497 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4498 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4499 		iflib_tx_structures_free(ctx);
4500 		iflib_rx_structures_free(ctx);
4501 	}
4502 	return (err);
4503 }
4504 
4505 int
4506 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4507 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4508 {
4509 
4510 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4511 }
4512 
4513 static int
4514 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4515 {
4516 	int i, cpuid, eqid, count;
4517 
4518 	CPU_COPY(&ctx->ifc_cpus, cpus);
4519 	count = CPU_COUNT(&ctx->ifc_cpus);
4520 	eqid = qid % count;
4521 	/* clear up to the qid'th bit */
4522 	for (i = 0; i < eqid; i++) {
4523 		cpuid = CPU_FFS(cpus);
4524 		MPASS(cpuid != 0);
4525 		CPU_CLR(cpuid-1, cpus);
4526 	}
4527 	cpuid = CPU_FFS(cpus);
4528 	MPASS(cpuid != 0);
4529 	return (cpuid-1);
4530 }
4531 
4532 int
4533 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4534 						iflib_intr_type_t type, driver_filter_t *filter,
4535 						void *filter_arg, int qid, char *name)
4536 {
4537 	struct grouptask *gtask;
4538 	struct taskqgroup *tqg;
4539 	iflib_filter_info_t info;
4540 	cpuset_t cpus;
4541 	gtask_fn_t *fn;
4542 	int tqrid, err, cpuid;
4543 	void *q;
4544 
4545 	info = &ctx->ifc_filter_info;
4546 	tqrid = rid;
4547 
4548 	switch (type) {
4549 	/* XXX merge tx/rx for netmap? */
4550 	case IFLIB_INTR_TX:
4551 		q = &ctx->ifc_txqs[qid];
4552 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4553 		gtask = &ctx->ifc_txqs[qid].ift_task;
4554 		tqg = qgroup_softirq;
4555 		fn = _task_fn_tx;
4556 		GROUPTASK_INIT(gtask, 0, fn, q);
4557 		break;
4558 	case IFLIB_INTR_RX:
4559 		q = &ctx->ifc_rxqs[qid];
4560 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4561 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4562 		tqg = qgroup_softirq;
4563 		fn = _task_fn_rx;
4564 		GROUPTASK_INIT(gtask, 0, fn, q);
4565 		break;
4566 	case IFLIB_INTR_ADMIN:
4567 		q = ctx;
4568 		tqrid = -1;
4569 		info = &ctx->ifc_filter_info;
4570 		gtask = &ctx->ifc_admin_task;
4571 		tqg = qgroup_if_config_tqg;
4572 		fn = _task_fn_admin;
4573 		break;
4574 	default:
4575 		panic("unknown net intr type");
4576 	}
4577 
4578 	info->ifi_filter = filter;
4579 	info->ifi_filter_arg = filter_arg;
4580 	info->ifi_task = gtask;
4581 	info->ifi_ctx = ctx;
4582 
4583 	err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info,  name);
4584 	if (err != 0) {
4585 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4586 		return (err);
4587 	}
4588 	if (type == IFLIB_INTR_ADMIN)
4589 		return (0);
4590 
4591 	if (tqrid != -1) {
4592 		cpuid = find_nth(ctx, &cpus, qid);
4593 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4594 	} else {
4595 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4596 	}
4597 
4598 	return (0);
4599 }
4600 
4601 void
4602 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4603 {
4604 	struct grouptask *gtask;
4605 	struct taskqgroup *tqg;
4606 	gtask_fn_t *fn;
4607 	void *q;
4608 
4609 	switch (type) {
4610 	case IFLIB_INTR_TX:
4611 		q = &ctx->ifc_txqs[qid];
4612 		gtask = &ctx->ifc_txqs[qid].ift_task;
4613 		tqg = qgroup_softirq;
4614 		fn = _task_fn_tx;
4615 		break;
4616 	case IFLIB_INTR_RX:
4617 		q = &ctx->ifc_rxqs[qid];
4618 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4619 		tqg = qgroup_softirq;
4620 		fn = _task_fn_rx;
4621 		break;
4622 	case IFLIB_INTR_IOV:
4623 		q = ctx;
4624 		gtask = &ctx->ifc_vflr_task;
4625 		tqg = qgroup_if_config_tqg;
4626 		rid = -1;
4627 		fn = _task_fn_iov;
4628 		break;
4629 	default:
4630 		panic("unknown net intr type");
4631 	}
4632 	GROUPTASK_INIT(gtask, 0, fn, q);
4633 	taskqgroup_attach(tqg, gtask, q, rid, name);
4634 }
4635 
4636 void
4637 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4638 {
4639 	if (irq->ii_tag)
4640 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4641 
4642 	if (irq->ii_res)
4643 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4644 }
4645 
4646 static int
4647 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4648 {
4649 	iflib_txq_t txq = ctx->ifc_txqs;
4650 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4651 	if_irq_t irq = &ctx->ifc_legacy_irq;
4652 	iflib_filter_info_t info;
4653 	struct grouptask *gtask;
4654 	struct taskqgroup *tqg;
4655 	gtask_fn_t *fn;
4656 	int tqrid;
4657 	void *q;
4658 	int err;
4659 
4660 	q = &ctx->ifc_rxqs[0];
4661 	info = &rxq[0].ifr_filter_info;
4662 	gtask = &rxq[0].ifr_task;
4663 	tqg = qgroup_softirq;
4664 	tqrid = irq->ii_rid = *rid;
4665 	fn = _task_fn_rx;
4666 
4667 	ctx->ifc_flags |= IFC_LEGACY;
4668 	info->ifi_filter = filter;
4669 	info->ifi_filter_arg = filter_arg;
4670 	info->ifi_task = gtask;
4671 	info->ifi_ctx = ctx;
4672 
4673 	/* We allocate a single interrupt resource */
4674 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0)
4675 		return (err);
4676 	GROUPTASK_INIT(gtask, 0, fn, q);
4677 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
4678 
4679 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
4680 	taskqgroup_attach(qgroup_softirq, &txq->ift_task, txq, tqrid, "tx");
4681 	return (0);
4682 }
4683 
4684 void
4685 iflib_led_create(if_ctx_t ctx)
4686 {
4687 
4688 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
4689 								  device_get_nameunit(ctx->ifc_dev));
4690 }
4691 
4692 void
4693 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
4694 {
4695 
4696 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
4697 }
4698 
4699 void
4700 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
4701 {
4702 
4703 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
4704 }
4705 
4706 void
4707 iflib_admin_intr_deferred(if_ctx_t ctx)
4708 {
4709 #ifdef INVARIANTS
4710 	struct grouptask *gtask;
4711 
4712 	gtask = &ctx->ifc_admin_task;
4713 	MPASS(gtask->gt_taskqueue != NULL);
4714 #endif
4715 
4716 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
4717 }
4718 
4719 void
4720 iflib_iov_intr_deferred(if_ctx_t ctx)
4721 {
4722 
4723 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
4724 }
4725 
4726 void
4727 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
4728 {
4729 
4730 	taskqgroup_attach_cpu(qgroup_softirq, gt, uniq, cpu, -1, name);
4731 }
4732 
4733 void
4734 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
4735 	char *name)
4736 {
4737 
4738 	GROUPTASK_INIT(gtask, 0, fn, ctx);
4739 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
4740 }
4741 
4742 void
4743 iflib_config_gtask_deinit(struct grouptask *gtask)
4744 {
4745 
4746 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
4747 }
4748 
4749 void
4750 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
4751 {
4752 	if_t ifp = ctx->ifc_ifp;
4753 	iflib_txq_t txq = ctx->ifc_txqs;
4754 
4755 	if_setbaudrate(ifp, baudrate);
4756 
4757 	/* If link down, disable watchdog */
4758 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
4759 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
4760 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4761 	}
4762 	ctx->ifc_link_state = link_state;
4763 	if_link_state_change(ifp, link_state);
4764 }
4765 
4766 static int
4767 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
4768 {
4769 	int credits;
4770 #ifdef INVARIANTS
4771 	int credits_pre = txq->ift_cidx_processed;
4772 #endif
4773 
4774 	if (ctx->isc_txd_credits_update == NULL)
4775 		return (0);
4776 
4777 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0)
4778 		return (0);
4779 
4780 	txq->ift_processed += credits;
4781 	txq->ift_cidx_processed += credits;
4782 
4783 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
4784 	if (txq->ift_cidx_processed >= txq->ift_size)
4785 		txq->ift_cidx_processed -= txq->ift_size;
4786 	return (credits);
4787 }
4788 
4789 static int
4790 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget)
4791 {
4792 
4793 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
4794 	    budget));
4795 }
4796 
4797 void
4798 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
4799 	const char *description, if_int_delay_info_t info,
4800 	int offset, int value)
4801 {
4802 	info->iidi_ctx = ctx;
4803 	info->iidi_offset = offset;
4804 	info->iidi_value = value;
4805 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
4806 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
4807 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
4808 	    info, 0, iflib_sysctl_int_delay, "I", description);
4809 }
4810 
4811 struct mtx *
4812 iflib_ctx_lock_get(if_ctx_t ctx)
4813 {
4814 
4815 	return (&ctx->ifc_mtx);
4816 }
4817 
4818 static int
4819 iflib_msix_init(if_ctx_t ctx)
4820 {
4821 	device_t dev = ctx->ifc_dev;
4822 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4823 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4824 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
4825 	int iflib_num_tx_queues, iflib_num_rx_queues;
4826 	int err, admincnt, bar;
4827 
4828 	iflib_num_tx_queues = scctx->isc_ntxqsets;
4829 	iflib_num_rx_queues = scctx->isc_nrxqsets;
4830 
4831 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
4832 
4833 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
4834 	admincnt = sctx->isc_admin_intrcnt;
4835 	/* Override by tuneable */
4836 	if (enable_msix == 0)
4837 		goto msi;
4838 
4839 	/*
4840 	** When used in a virtualized environment
4841 	** PCI BUSMASTER capability may not be set
4842 	** so explicity set it here and rewrite
4843 	** the ENABLE in the MSIX control register
4844 	** at this point to cause the host to
4845 	** successfully initialize us.
4846 	*/
4847 	{
4848 		int msix_ctrl, rid;
4849 
4850  		pci_enable_busmaster(dev);
4851 		rid = 0;
4852 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
4853 			rid += PCIR_MSIX_CTRL;
4854 			msix_ctrl = pci_read_config(dev, rid, 2);
4855 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
4856 			pci_write_config(dev, rid, msix_ctrl, 2);
4857 		} else {
4858 			device_printf(dev, "PCIY_MSIX capability not found; "
4859 			                   "or rid %d == 0.\n", rid);
4860 			goto msi;
4861 		}
4862 	}
4863 
4864 	/*
4865 	 * bar == -1 => "trust me I know what I'm doing"
4866 	 * https://www.youtube.com/watch?v=nnwWKkNau4I
4867 	 * Some drivers are for hardware that is so shoddily
4868 	 * documented that no one knows which bars are which
4869 	 * so the developer has to map all bars. This hack
4870 	 * allows shoddy garbage to use msix in this framework.
4871 	 */
4872 	if (bar != -1) {
4873 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
4874 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
4875 		if (ctx->ifc_msix_mem == NULL) {
4876 			/* May not be enabled */
4877 			device_printf(dev, "Unable to map MSIX table \n");
4878 			goto msi;
4879 		}
4880 	}
4881 	/* First try MSI/X */
4882 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
4883 		device_printf(dev, "System has MSIX disabled \n");
4884 		bus_release_resource(dev, SYS_RES_MEMORY,
4885 		    bar, ctx->ifc_msix_mem);
4886 		ctx->ifc_msix_mem = NULL;
4887 		goto msi;
4888 	}
4889 #if IFLIB_DEBUG
4890 	/* use only 1 qset in debug mode */
4891 	queuemsgs = min(msgs - admincnt, 1);
4892 #else
4893 	queuemsgs = msgs - admincnt;
4894 #endif
4895 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
4896 #ifdef RSS
4897 		queues = imin(queuemsgs, rss_getnumbuckets());
4898 #else
4899 		queues = queuemsgs;
4900 #endif
4901 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
4902 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
4903 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
4904 	} else {
4905 		device_printf(dev, "Unable to fetch CPU list\n");
4906 		/* Figure out a reasonable auto config value */
4907 		queues = min(queuemsgs, mp_ncpus);
4908 	}
4909 #ifdef  RSS
4910 	/* If we're doing RSS, clamp at the number of RSS buckets */
4911 	if (queues > rss_getnumbuckets())
4912 		queues = rss_getnumbuckets();
4913 #endif
4914 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
4915 		rx_queues = iflib_num_rx_queues;
4916 	else
4917 		rx_queues = queues;
4918 	/*
4919 	 * We want this to be all logical CPUs by default
4920 	 */
4921 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
4922 		tx_queues = iflib_num_tx_queues;
4923 	else
4924 		tx_queues = mp_ncpus;
4925 
4926 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
4927 #ifdef INVARIANTS
4928 		if (tx_queues != rx_queues)
4929 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
4930 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
4931 #endif
4932 		tx_queues = min(rx_queues, tx_queues);
4933 		rx_queues = min(rx_queues, tx_queues);
4934 	}
4935 
4936 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
4937 
4938 	vectors = rx_queues + admincnt;
4939 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
4940 		device_printf(dev,
4941 					  "Using MSIX interrupts with %d vectors\n", vectors);
4942 		scctx->isc_vectors = vectors;
4943 		scctx->isc_nrxqsets = rx_queues;
4944 		scctx->isc_ntxqsets = tx_queues;
4945 		scctx->isc_intr = IFLIB_INTR_MSIX;
4946 
4947 		return (vectors);
4948 	} else {
4949 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
4950 	}
4951 msi:
4952 	vectors = pci_msi_count(dev);
4953 	scctx->isc_nrxqsets = 1;
4954 	scctx->isc_ntxqsets = 1;
4955 	scctx->isc_vectors = vectors;
4956 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
4957 		device_printf(dev,"Using an MSI interrupt\n");
4958 		scctx->isc_intr = IFLIB_INTR_MSI;
4959 	} else {
4960 		device_printf(dev,"Using a Legacy interrupt\n");
4961 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4962 	}
4963 
4964 	return (vectors);
4965 }
4966 
4967 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
4968 
4969 static int
4970 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
4971 {
4972 	int rc;
4973 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
4974 	struct sbuf *sb;
4975 	char *ring_state = "UNKNOWN";
4976 
4977 	/* XXX needed ? */
4978 	rc = sysctl_wire_old_buffer(req, 0);
4979 	MPASS(rc == 0);
4980 	if (rc != 0)
4981 		return (rc);
4982 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
4983 	MPASS(sb != NULL);
4984 	if (sb == NULL)
4985 		return (ENOMEM);
4986 	if (state[3] <= 3)
4987 		ring_state = ring_states[state[3]];
4988 
4989 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
4990 		    state[0], state[1], state[2], ring_state);
4991 	rc = sbuf_finish(sb);
4992 	sbuf_delete(sb);
4993         return(rc);
4994 }
4995 
4996 enum iflib_ndesc_handler {
4997 	IFLIB_NTXD_HANDLER,
4998 	IFLIB_NRXD_HANDLER,
4999 };
5000 
5001 static int
5002 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5003 {
5004 	if_ctx_t ctx = (void *)arg1;
5005 	enum iflib_ndesc_handler type = arg2;
5006 	char buf[256] = {0};
5007 	uint16_t *ndesc;
5008 	char *p, *next;
5009 	int nqs, rc, i;
5010 
5011 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5012 
5013 	nqs = 8;
5014 	switch(type) {
5015 	case IFLIB_NTXD_HANDLER:
5016 		ndesc = ctx->ifc_sysctl_ntxds;
5017 		if (ctx->ifc_sctx)
5018 			nqs = ctx->ifc_sctx->isc_ntxqs;
5019 		break;
5020 	case IFLIB_NRXD_HANDLER:
5021 		ndesc = ctx->ifc_sysctl_nrxds;
5022 		if (ctx->ifc_sctx)
5023 			nqs = ctx->ifc_sctx->isc_nrxqs;
5024 		break;
5025 	}
5026 	if (nqs == 0)
5027 		nqs = 8;
5028 
5029 	for (i=0; i<8; i++) {
5030 		if (i >= nqs)
5031 			break;
5032 		if (i)
5033 			strcat(buf, ",");
5034 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5035 	}
5036 
5037 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5038 	if (rc || req->newptr == NULL)
5039 		return rc;
5040 
5041 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5042 	    i++, p = strsep(&next, " ,")) {
5043 		ndesc[i] = strtoul(p, NULL, 10);
5044 	}
5045 
5046 	return(rc);
5047 }
5048 
5049 #define NAME_BUFLEN 32
5050 static void
5051 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5052 {
5053         device_t dev = iflib_get_dev(ctx);
5054 	struct sysctl_oid_list *child, *oid_list;
5055 	struct sysctl_ctx_list *ctx_list;
5056 	struct sysctl_oid *node;
5057 
5058 	ctx_list = device_get_sysctl_ctx(dev);
5059 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5060 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5061 						      CTLFLAG_RD, NULL, "IFLIB fields");
5062 	oid_list = SYSCTL_CHILDREN(node);
5063 
5064 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5065 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5066 		       "driver version");
5067 
5068 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5069 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5070 			"# of txqs to use, 0 => use default #");
5071 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5072 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5073 			"# of rxqs to use, 0 => use default #");
5074 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5075 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5076                        "permit #txq != #rxq");
5077 
5078 	/* XXX change for per-queue sizes */
5079 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5080 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5081                        mp_ndesc_handler, "A",
5082                        "list of # of tx descriptors to use, 0 = use default #");
5083 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5084 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5085                        mp_ndesc_handler, "A",
5086                        "list of # of rx descriptors to use, 0 = use default #");
5087 }
5088 
5089 static void
5090 iflib_add_device_sysctl_post(if_ctx_t ctx)
5091 {
5092 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5093 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5094         device_t dev = iflib_get_dev(ctx);
5095 	struct sysctl_oid_list *child;
5096 	struct sysctl_ctx_list *ctx_list;
5097 	iflib_fl_t fl;
5098 	iflib_txq_t txq;
5099 	iflib_rxq_t rxq;
5100 	int i, j;
5101 	char namebuf[NAME_BUFLEN];
5102 	char *qfmt;
5103 	struct sysctl_oid *queue_node, *fl_node, *node;
5104 	struct sysctl_oid_list *queue_list, *fl_list;
5105 	ctx_list = device_get_sysctl_ctx(dev);
5106 
5107 	node = ctx->ifc_sysctl_node;
5108 	child = SYSCTL_CHILDREN(node);
5109 
5110 	if (scctx->isc_ntxqsets > 100)
5111 		qfmt = "txq%03d";
5112 	else if (scctx->isc_ntxqsets > 10)
5113 		qfmt = "txq%02d";
5114 	else
5115 		qfmt = "txq%d";
5116 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5117 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5118 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5119 					     CTLFLAG_RD, NULL, "Queue Name");
5120 		queue_list = SYSCTL_CHILDREN(queue_node);
5121 #if MEMORY_LOGGING
5122 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5123 				CTLFLAG_RD,
5124 				&txq->ift_dequeued, "total mbufs freed");
5125 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5126 				CTLFLAG_RD,
5127 				&txq->ift_enqueued, "total mbufs enqueued");
5128 #endif
5129 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5130 				   CTLFLAG_RD,
5131 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5132 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5133 				   CTLFLAG_RD,
5134 				   &txq->ift_pullups, "# of times m_pullup was called");
5135 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5136 				   CTLFLAG_RD,
5137 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5138 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5139 				   CTLFLAG_RD,
5140 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5141 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5142 				   CTLFLAG_RD,
5143 				   &txq->ift_map_failed, "# of times dma map failed");
5144 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5145 				   CTLFLAG_RD,
5146 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5147 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5148 				   CTLFLAG_RD,
5149 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5150 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5151 				   CTLFLAG_RD,
5152 				   &txq->ift_pidx, 1, "Producer Index");
5153 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5154 				   CTLFLAG_RD,
5155 				   &txq->ift_cidx, 1, "Consumer Index");
5156 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5157 				   CTLFLAG_RD,
5158 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5159 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5160 				   CTLFLAG_RD,
5161 				   &txq->ift_in_use, 1, "descriptors in use");
5162 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5163 				   CTLFLAG_RD,
5164 				   &txq->ift_processed, "descriptors procesed for clean");
5165 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5166 				   CTLFLAG_RD,
5167 				   &txq->ift_cleaned, "total cleaned");
5168 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5169 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state),
5170 				0, mp_ring_state_handler, "A", "soft ring state");
5171 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5172 				       CTLFLAG_RD, &txq->ift_br[0]->enqueues,
5173 				       "# of enqueues to the mp_ring for this queue");
5174 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5175 				       CTLFLAG_RD, &txq->ift_br[0]->drops,
5176 				       "# of drops in the mp_ring for this queue");
5177 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5178 				       CTLFLAG_RD, &txq->ift_br[0]->starts,
5179 				       "# of normal consumer starts in the mp_ring for this queue");
5180 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5181 				       CTLFLAG_RD, &txq->ift_br[0]->stalls,
5182 					       "# of consumer stalls in the mp_ring for this queue");
5183 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5184 			       CTLFLAG_RD, &txq->ift_br[0]->restarts,
5185 				       "# of consumer restarts in the mp_ring for this queue");
5186 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5187 				       CTLFLAG_RD, &txq->ift_br[0]->abdications,
5188 				       "# of consumer abdications in the mp_ring for this queue");
5189 	}
5190 
5191 	if (scctx->isc_nrxqsets > 100)
5192 		qfmt = "rxq%03d";
5193 	else if (scctx->isc_nrxqsets > 10)
5194 		qfmt = "rxq%02d";
5195 	else
5196 		qfmt = "rxq%d";
5197 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5198 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5199 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5200 					     CTLFLAG_RD, NULL, "Queue Name");
5201 		queue_list = SYSCTL_CHILDREN(queue_node);
5202 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5203 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5204 				       CTLFLAG_RD,
5205 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5206 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5207 				       CTLFLAG_RD,
5208 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5209 		}
5210 
5211 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5212 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5213 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5214 						     CTLFLAG_RD, NULL, "freelist Name");
5215 			fl_list = SYSCTL_CHILDREN(fl_node);
5216 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5217 				       CTLFLAG_RD,
5218 				       &fl->ifl_pidx, 1, "Producer Index");
5219 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5220 				       CTLFLAG_RD,
5221 				       &fl->ifl_cidx, 1, "Consumer Index");
5222 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5223 				       CTLFLAG_RD,
5224 				       &fl->ifl_credits, 1, "credits available");
5225 #if MEMORY_LOGGING
5226 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5227 					CTLFLAG_RD,
5228 					&fl->ifl_m_enqueued, "mbufs allocated");
5229 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5230 					CTLFLAG_RD,
5231 					&fl->ifl_m_dequeued, "mbufs freed");
5232 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5233 					CTLFLAG_RD,
5234 					&fl->ifl_cl_enqueued, "clusters allocated");
5235 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5236 					CTLFLAG_RD,
5237 					&fl->ifl_cl_dequeued, "clusters freed");
5238 #endif
5239 
5240 		}
5241 	}
5242 
5243 }
5244