xref: /freebsd/sys/net/iflib.c (revision b602ba1b5fd92bb226e32f5720885f856a5cb0bb)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_acpi.h"
32 
33 #include <sys/param.h>
34 #include <sys/types.h>
35 #include <sys/bus.h>
36 #include <sys/eventhandler.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/module.h>
41 #include <sys/kobj.h>
42 #include <sys/proc.h>
43 #include <sys/rman.h>
44 #include <sys/sbuf.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/socket.h>
48 #include <sys/sockio.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/if_private.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 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/udp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet6/ip6_var.h>
77 
78 #include <machine/bus.h>
79 #include <machine/in_cksum.h>
80 
81 #include <vm/vm.h>
82 #include <vm/pmap.h>
83 
84 #include <dev/led/led.h>
85 #include <dev/pci/pcireg.h>
86 #include <dev/pci/pcivar.h>
87 #include <dev/pci/pci_private.h>
88 
89 #include <net/iflib.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE	(1U << 0)
132 #define	IFLIB_RXEOF_EMPTY	(2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 #ifndef ALTQ
147 static int  iflib_simple_transmit(if_t ifp, struct mbuf *m);
148 #endif
149 
150 typedef struct iflib_filter_info {
151 	driver_filter_t *ifi_filter;
152 	void *ifi_filter_arg;
153 	struct grouptask *ifi_task;
154 	void *ifi_ctx;
155 } *iflib_filter_info_t;
156 
157 struct iflib_ctx {
158 	KOBJ_FIELDS;
159 	/*
160 	 * Pointer to hardware driver's softc
161 	 */
162 	void *ifc_softc;
163 	device_t ifc_dev;
164 	if_t ifc_ifp;
165 
166 	cpuset_t ifc_cpus;
167 	if_shared_ctx_t ifc_sctx;
168 	struct if_softc_ctx ifc_softc_ctx;
169 
170 	struct sx ifc_ctx_sx;
171 	struct mtx ifc_state_mtx;
172 
173 	iflib_txq_t ifc_txqs;
174 	iflib_rxq_t ifc_rxqs;
175 	uint32_t ifc_if_flags;
176 	uint32_t ifc_flags;
177 	uint32_t ifc_max_fl_buf_size;
178 	uint32_t ifc_rx_mbuf_sz;
179 
180 	int ifc_link_state;
181 	int ifc_watchdog_events;
182 	struct cdev *ifc_led_dev;
183 	struct resource *ifc_msix_mem;
184 
185 	struct if_irq ifc_legacy_irq;
186 	struct task ifc_admin_task;
187 	struct task ifc_vflr_task;
188 	struct taskqueue *ifc_tq;
189 	struct iflib_filter_info ifc_filter_info;
190 	struct ifmedia	ifc_media;
191 	struct ifmedia	*ifc_mediap;
192 
193 	struct sysctl_oid *ifc_sysctl_node;
194 	uint16_t ifc_sysctl_ntxqs;
195 	uint16_t ifc_sysctl_nrxqs;
196 	uint16_t ifc_sysctl_qs_eq_override;
197 	uint16_t ifc_sysctl_rx_budget;
198 	uint16_t ifc_sysctl_tx_abdicate;
199 	uint16_t ifc_sysctl_core_offset;
200 #define	CORE_OFFSET_UNSPECIFIED	0xffff
201 	uint8_t  ifc_sysctl_separate_txrx;
202 	uint8_t  ifc_sysctl_use_logical_cores;
203 	uint16_t ifc_sysctl_extra_msix_vectors;
204 	bool     ifc_cpus_are_physical_cores;
205 	bool     ifc_sysctl_simple_tx;
206 	bool     ifc_sysctl_tx_defer_mfree;
207 	uint16_t ifc_sysctl_tx_reclaim_thresh;
208 	uint16_t ifc_sysctl_tx_reclaim_ticks;
209 
210 	qidx_t ifc_sysctl_ntxds[8];
211 	qidx_t ifc_sysctl_nrxds[8];
212 	struct if_txrx ifc_txrx;
213 #define isc_txd_encap		ifc_txrx.ift_txd_encap
214 #define isc_txd_flush		ifc_txrx.ift_txd_flush
215 #define isc_txd_credits_update	ifc_txrx.ift_txd_credits_update
216 #define isc_rxd_available	ifc_txrx.ift_rxd_available
217 #define isc_rxd_pkt_get		ifc_txrx.ift_rxd_pkt_get
218 #define isc_rxd_refill		ifc_txrx.ift_rxd_refill
219 #define isc_rxd_flush		ifc_txrx.ift_rxd_flush
220 #define isc_legacy_intr		ifc_txrx.ift_legacy_intr
221 #define isc_txq_select		ifc_txrx.ift_txq_select
222 #define isc_txq_select_v2	ifc_txrx.ift_txq_select_v2
223 
224 	eventhandler_tag ifc_vlan_attach_event;
225 	eventhandler_tag ifc_vlan_detach_event;
226 	struct ether_addr ifc_mac;
227 };
228 
229 void *
iflib_get_softc(if_ctx_t ctx)230 iflib_get_softc(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_softc);
234 }
235 
236 device_t
iflib_get_dev(if_ctx_t ctx)237 iflib_get_dev(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_dev);
241 }
242 
243 if_t
iflib_get_ifp(if_ctx_t ctx)244 iflib_get_ifp(if_ctx_t ctx)
245 {
246 
247 	return (ctx->ifc_ifp);
248 }
249 
250 struct ifmedia *
iflib_get_media(if_ctx_t ctx)251 iflib_get_media(if_ctx_t ctx)
252 {
253 
254 	return (ctx->ifc_mediap);
255 }
256 
257 void
iflib_set_mac(if_ctx_t ctx,uint8_t mac[ETHER_ADDR_LEN])258 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
259 {
260 
261 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
262 }
263 
264 if_softc_ctx_t
iflib_get_softc_ctx(if_ctx_t ctx)265 iflib_get_softc_ctx(if_ctx_t ctx)
266 {
267 
268 	return (&ctx->ifc_softc_ctx);
269 }
270 
271 if_shared_ctx_t
iflib_get_sctx(if_ctx_t ctx)272 iflib_get_sctx(if_ctx_t ctx)
273 {
274 
275 	return (ctx->ifc_sctx);
276 }
277 
278 uint16_t
iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)279 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
280 {
281 
282 	return (ctx->ifc_sysctl_extra_msix_vectors);
283 }
284 
285 #define IP_ALIGNED(m)		((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
286 #define CACHE_PTR_INCREMENT	(CACHE_LINE_SIZE / sizeof(void *))
287 #define CACHE_PTR_NEXT(ptr)	((void *)(roundup2(ptr, CACHE_LINE_SIZE)))
288 
289 #define LINK_ACTIVE(ctx)	((ctx)->ifc_link_state == LINK_STATE_UP)
290 #define CTX_IS_VF(ctx)		((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
291 
292 typedef struct iflib_sw_rx_desc_array {
293 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
294 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
295 	caddr_t		*ifsd_cl;	/* direct cluster pointer for rx */
296 	bus_addr_t	*ifsd_ba;	/* bus addr of cluster for rx */
297 } iflib_rxsd_array_t;
298 
299 typedef struct iflib_sw_tx_desc_array {
300 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
301 	bus_dmamap_t	*ifsd_tso_map;	/* bus_dma maps for TSO packet */
302 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
303 	struct mbuf	**ifsd_m_defer;	/* deferred mbuf ptr */
304 	struct mbuf	**ifsd_m_deferb;/* deferred mbuf backing ptr */
305 } if_txsd_vec_t;
306 
307 /* magic number that should be high enough for any hardware */
308 #define IFLIB_MAX_TX_SEGS		128
309 #define IFLIB_RX_COPY_THRESH		128
310 #define IFLIB_MAX_RX_REFRESH		32
311 /* The minimum descriptors per second before we start coalescing */
312 #define IFLIB_MIN_DESC_SEC		16384
313 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
314 #define IFLIB_QUEUE_IDLE		0
315 #define IFLIB_QUEUE_HUNG		1
316 #define IFLIB_QUEUE_WORKING		2
317 /* maximum number of txqs that can share an rx interrupt */
318 #define IFLIB_MAX_TX_SHARED_INTR	4
319 
320 /* this should really scale with ring size - this is a fairly arbitrary value */
321 #define TX_BATCH_SIZE			32
322 
323 #define IFLIB_RESTART_BUDGET		8
324 
325 
326 /*
327  * Encode TSO or !TSO in the low bits of the tx ifsd_m pointer so as
328  * to avoid defref'ing the mbuf to determine the correct busdma resources
329  * to release
330  */
331 #define IFLIB_TSO		(1ULL << 0)
332 #define IFLIB_NO_TSO		(2ULL << 0)
333 #define IFLIB_FLAGS_MASK	(0x3ULL)
334 #define IFLIB_SAVE_MBUF(mbuf, flags)	((void *)(((uintptr_t)mbuf) | flags))
335 #define IFLIB_GET_FLAGS(a)	((uintptr_t)a & IFLIB_FLAGS_MASK)
336 #define IFLIB_GET_MBUF(a)	((struct mbuf *)((uintptr_t)a & ~IFLIB_FLAGS_MASK))
337 
338 
339 #define	IFC_LEGACY		0x001
340 #define	IFC_QFLUSH		0x002
341 #define	IFC_MULTISEG		0x004
342 #define	IFC_SPARE1		0x008
343 #define	IFC_SC_ALLOCATED	0x010
344 #define	IFC_INIT_DONE		0x020
345 #define	IFC_PREFETCH		0x040
346 #define	IFC_DO_RESET		0x080
347 #define	IFC_DO_WATCHDOG		0x100
348 #define	IFC_SPARE0		0x200
349 #define	IFC_SPARE2		0x400
350 #define	IFC_IN_DETACH		0x800
351 
352 #define	IFC_NETMAP_TX_IRQ	0x80000000
353 
354 #define CSUM_OFFLOAD		(CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \
355 				 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \
356 				 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP)
357 
358 struct iflib_txq {
359 	qidx_t		ift_in_use;
360 	qidx_t		ift_cidx;
361 	qidx_t		ift_cidx_processed;
362 	qidx_t		ift_pidx;
363 	uint8_t		ift_gen;
364 	uint8_t		ift_br_offset:1,
365 			ift_defer_mfree:1,
366 			ift_spare_bits0:6;
367 	uint16_t	ift_npending;
368 	uint16_t	ift_db_pending;
369 	uint16_t	ift_rs_pending;
370 	uint32_t	ift_last_reclaim;
371 	uint16_t	ift_reclaim_thresh;
372 	uint16_t	ift_reclaim_ticks;
373 	uint8_t		ift_txd_size[8];
374 	uint64_t	ift_processed;
375 	uint64_t	ift_cleaned;
376 	uint64_t	ift_cleaned_prev;
377 #if MEMORY_LOGGING
378 	uint64_t	ift_enqueued;
379 	uint64_t	ift_dequeued;
380 #endif
381 	uint64_t	ift_no_tx_dma_setup;
382 	uint64_t	ift_no_desc_avail;
383 	uint64_t	ift_mbuf_defrag_failed;
384 	uint64_t	ift_mbuf_defrag;
385 	uint64_t	ift_map_failed;
386 	uint64_t	ift_txd_encap_efbig;
387 	uint64_t	ift_pullups;
388 	uint64_t	ift_last_timer_tick;
389 
390 	struct mtx	ift_mtx;
391 	struct mtx	ift_db_mtx;
392 
393 	/* constant values */
394 	if_ctx_t	ift_ctx;
395 	struct ifmp_ring	*ift_br;
396 	struct grouptask	ift_task;
397 	qidx_t		ift_size;
398 	qidx_t		ift_pad;
399 	uint16_t	ift_id;
400 	struct callout	ift_timer;
401 #ifdef DEV_NETMAP
402 	struct callout	ift_netmap_timer;
403 #endif /* DEV_NETMAP */
404 
405 	if_txsd_vec_t	ift_sds;
406 	uint8_t		ift_qstatus;
407 	uint8_t		ift_closed;
408 	uint8_t		ift_update_freq;
409 	struct iflib_filter_info ift_filter_info;
410 	bus_dma_tag_t	ift_buf_tag;
411 	bus_dma_tag_t	ift_tso_buf_tag;
412 	iflib_dma_info_t	ift_ifdi;
413 #define	MTX_NAME_LEN	32
414 	char		ift_mtx_name[MTX_NAME_LEN];
415 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
416 #ifdef IFLIB_DIAGNOSTICS
417 	uint64_t ift_cpu_exec_count[256];
418 #endif
419 } __aligned(CACHE_LINE_SIZE);
420 
421 struct iflib_fl {
422 	qidx_t		ifl_cidx;
423 	qidx_t		ifl_pidx;
424 	qidx_t		ifl_credits;
425 	uint8_t		ifl_gen;
426 	uint8_t		ifl_rxd_size;
427 #if MEMORY_LOGGING
428 	uint64_t	ifl_m_enqueued;
429 	uint64_t	ifl_m_dequeued;
430 	uint64_t	ifl_cl_enqueued;
431 	uint64_t	ifl_cl_dequeued;
432 #endif
433 	/* implicit pad */
434 	bitstr_t	*ifl_rx_bitmap;
435 	qidx_t		ifl_fragidx;
436 	/* constant */
437 	qidx_t		ifl_size;
438 	uint16_t	ifl_buf_size;
439 	uint16_t	ifl_cltype;
440 	uma_zone_t	ifl_zone;
441 	iflib_rxsd_array_t	ifl_sds;
442 	iflib_rxq_t	ifl_rxq;
443 	uint8_t		ifl_id;
444 	bus_dma_tag_t	ifl_buf_tag;
445 	iflib_dma_info_t	ifl_ifdi;
446 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
447 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
448 }  __aligned(CACHE_LINE_SIZE);
449 
450 static inline qidx_t
get_inuse(int size,qidx_t cidx,qidx_t pidx,uint8_t gen)451 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
452 {
453 	qidx_t used;
454 
455 	if (pidx > cidx)
456 		used = pidx - cidx;
457 	else if (pidx < cidx)
458 		used = size - cidx + pidx;
459 	else if (gen == 0 && pidx == cidx)
460 		used = 0;
461 	else if (gen == 1 && pidx == cidx)
462 		used = size;
463 	else
464 		panic("bad state");
465 
466 	return (used);
467 }
468 
469 #define TXQ_AVAIL(txq) ((txq->ift_size - txq->ift_pad) -\
470 	    get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
471 
472 #define IDXDIFF(head, tail, wrap) \
473 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
474 
475 struct iflib_rxq {
476 	if_ctx_t	ifr_ctx;
477 	iflib_fl_t	ifr_fl;
478 	uint64_t	ifr_rx_irq;
479 	struct pfil_head	*pfil;
480 	/*
481 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
482 	 * the completion queue consumer index.  Otherwise it's unused.
483 	 */
484 	qidx_t		ifr_cq_cidx;
485 	uint16_t	ifr_id;
486 	uint8_t		ifr_nfl;
487 	uint8_t		ifr_ntxqirq;
488 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
489 	uint8_t		ifr_fl_offset;
490 	struct lro_ctrl		ifr_lc;
491 	struct grouptask	ifr_task;
492 	struct callout		ifr_watchdog;
493 	struct iflib_filter_info ifr_filter_info;
494 	iflib_dma_info_t	ifr_ifdi;
495 
496 	/* dynamically allocate if any drivers need a value substantially larger than this */
497 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
498 #ifdef IFLIB_DIAGNOSTICS
499 	uint64_t ifr_cpu_exec_count[256];
500 #endif
501 }  __aligned(CACHE_LINE_SIZE);
502 
503 typedef struct if_rxsd {
504 	caddr_t *ifsd_cl;
505 	iflib_fl_t ifsd_fl;
506 } *if_rxsd_t;
507 
508 /*
509  * Only allow a single packet to take up most 1/nth of the tx ring
510  */
511 #define MAX_SINGLE_PACKET_FRACTION 12
512 #define IF_BAD_DMA	((bus_addr_t)-1)
513 
514 #define CTX_ACTIVE(ctx)	((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
515 
516 #define CTX_LOCK_INIT(_sc)	sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
517 #define CTX_LOCK(ctx)		sx_xlock(&(ctx)->ifc_ctx_sx)
518 #define CTX_UNLOCK(ctx)		sx_xunlock(&(ctx)->ifc_ctx_sx)
519 #define CTX_LOCK_DESTROY(ctx)	sx_destroy(&(ctx)->ifc_ctx_sx)
520 
521 #define STATE_LOCK_INIT(_sc, _name)	mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
522 #define STATE_LOCK(ctx)		mtx_lock(&(ctx)->ifc_state_mtx)
523 #define STATE_UNLOCK(ctx)	mtx_unlock(&(ctx)->ifc_state_mtx)
524 #define STATE_LOCK_DESTROY(ctx)	mtx_destroy(&(ctx)->ifc_state_mtx)
525 
526 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
527 #define CALLOUT_UNLOCK(txq)	mtx_unlock(&txq->ift_mtx)
528 
529 /* Our boot-time initialization hook */
530 static int	iflib_module_event_handler(module_t, int, void *);
531 
532 static moduledata_t iflib_moduledata = {
533 	"iflib",
534 	iflib_module_event_handler,
535 	NULL
536 };
537 
538 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
539 MODULE_VERSION(iflib, 1);
540 
541 MODULE_DEPEND(iflib, pci, 1, 1, 1);
542 MODULE_DEPEND(iflib, ether, 1, 1, 1);
543 
544 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
545 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
546 
547 #ifndef IFLIB_DEBUG_COUNTERS
548 #ifdef INVARIANTS
549 #define IFLIB_DEBUG_COUNTERS 1
550 #else
551 #define IFLIB_DEBUG_COUNTERS 0
552 #endif /* !INVARIANTS */
553 #endif
554 
555 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
556     "iflib driver parameters");
557 
558 /*
559  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
560  */
561 static int iflib_min_tx_latency = 0;
562 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
563     &iflib_min_tx_latency, 0,
564     "minimize transmit latency at the possible expense of throughput");
565 static int iflib_no_tx_batch = 0;
566 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
567     &iflib_no_tx_batch, 0,
568     "minimize transmit latency at the possible expense of throughput");
569 static int iflib_timer_default = 1000;
570 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
571     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
572 
573 
574 #if IFLIB_DEBUG_COUNTERS
575 
576 static int iflib_tx_seen;
577 static int iflib_tx_sent;
578 static int iflib_tx_encap;
579 static int iflib_rx_allocs;
580 static int iflib_fl_refills;
581 static int iflib_fl_refills_large;
582 static int iflib_tx_frees;
583 
584 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
585     "# TX mbufs seen");
586 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
587     "# TX mbufs sent");
588 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
589     "# TX mbufs encapped");
590 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
591     "# TX frees");
592 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
593     "# RX allocations");
594 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
595     "# refills");
596 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
597     &iflib_fl_refills_large, 0, "# large refills");
598 
599 static int iflib_txq_drain_flushing;
600 static int iflib_txq_drain_oactive;
601 static int iflib_txq_drain_notready;
602 
603 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
604     &iflib_txq_drain_flushing, 0, "# drain flushes");
605 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
606     &iflib_txq_drain_oactive, 0, "# drain oactives");
607 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
608     &iflib_txq_drain_notready, 0, "# drain notready");
609 
610 static int iflib_encap_load_mbuf_fail;
611 static int iflib_encap_pad_mbuf_fail;
612 static int iflib_encap_txq_avail_fail;
613 static int iflib_encap_txd_encap_fail;
614 
615 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
616     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
617 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
618     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
619 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
620     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
621 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
622     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
623 
624 static int iflib_task_fn_rxs;
625 static int iflib_rx_intr_enables;
626 static int iflib_fast_intrs;
627 static int iflib_rx_unavail;
628 static int iflib_rx_ctx_inactive;
629 static int iflib_rx_if_input;
630 static int iflib_rxd_flush;
631 
632 static int iflib_verbose_debug;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
635     "# task_fn_rx calls");
636 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
637     &iflib_rx_intr_enables, 0, "# RX intr enables");
638 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
639     "# fast_intr calls");
640 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
641     "# times rxeof called with no available data");
642 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
643     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
644 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
645     0, "# times rxeof called if_input");
646 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
647     "# times rxd_flush called");
648 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
649     &iflib_verbose_debug, 0, "enable verbose debugging");
650 
651 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
652 static void
iflib_debug_reset(void)653 iflib_debug_reset(void)
654 {
655 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
656 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
657 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
658 		iflib_txq_drain_notready =
659 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
660 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
661 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
662 		iflib_rx_unavail =
663 		iflib_rx_ctx_inactive = iflib_rx_if_input =
664 		iflib_rxd_flush = 0;
665 }
666 
667 #else
668 #define DBG_COUNTER_INC(name)
iflib_debug_reset(void)669 static void iflib_debug_reset(void) {}
670 #endif
671 
672 #define IFLIB_DEBUG 0
673 
674 static void iflib_tx_structures_free(if_ctx_t ctx);
675 static void iflib_rx_structures_free(if_ctx_t ctx);
676 static int iflib_queues_alloc(if_ctx_t ctx);
677 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
678 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
679 static int iflib_qset_structures_setup(if_ctx_t ctx);
680 static int iflib_msix_init(if_ctx_t ctx);
681 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
682 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
683 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
684 #ifdef ALTQ
685 static void iflib_altq_if_start(if_t ifp);
686 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
687 #endif
688 static void iflib_register(if_ctx_t);
689 static void iflib_deregister(if_ctx_t);
690 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
691 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
692 static void iflib_init_locked(if_ctx_t ctx);
693 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
694 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
695 static void iflib_ifmp_purge(iflib_txq_t txq);
696 static void _iflib_pre_assert(if_softc_ctx_t scctx);
697 static void iflib_stop(if_ctx_t ctx);
698 static void iflib_if_init_locked(if_ctx_t ctx);
699 static void iflib_free_intr_mem(if_ctx_t ctx);
700 #ifndef __NO_STRICT_ALIGNMENT
701 static struct mbuf *iflib_fixup_rx(struct mbuf *m);
702 #endif
703 static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq,
704     struct mbuf **m_defer);
705 
706 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
707     SLIST_HEAD_INITIALIZER(cpu_offsets);
708 struct cpu_offset {
709 	SLIST_ENTRY(cpu_offset) entries;
710 	cpuset_t	set;
711 	unsigned int	refcount;
712 	uint16_t	next_cpuid;
713 };
714 static struct mtx cpu_offset_mtx;
715 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
716     MTX_DEF);
717 
718 DEBUGNET_DEFINE(iflib);
719 
720 static int
iflib_num_rx_descs(if_ctx_t ctx)721 iflib_num_rx_descs(if_ctx_t ctx)
722 {
723 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
724 	if_shared_ctx_t sctx = ctx->ifc_sctx;
725 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
726 
727 	return (scctx->isc_nrxd[first_rxq]);
728 }
729 
730 static int
iflib_num_tx_descs(if_ctx_t ctx)731 iflib_num_tx_descs(if_ctx_t ctx)
732 {
733 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
734 	if_shared_ctx_t sctx = ctx->ifc_sctx;
735 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
736 
737 	return (scctx->isc_ntxd[first_txq]);
738 }
739 
740 #ifdef DEV_NETMAP
741 #include <sys/selinfo.h>
742 #include <net/netmap.h>
743 #include <dev/netmap/netmap_kern.h>
744 
745 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
746 
747 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
748 static void iflib_netmap_timer(void *arg);
749 
750 /*
751  * device-specific sysctl variables:
752  *
753  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
754  *	During regular operations the CRC is stripped, but on some
755  *	hardware reception of frames not multiple of 64 is slower,
756  *	so using crcstrip=0 helps in benchmarks.
757  *
758  * iflib_rx_miss, iflib_rx_miss_bufs:
759  *	count packets that might be missed due to lost interrupts.
760  */
761 SYSCTL_DECL(_dev_netmap);
762 /*
763  * The xl driver by default strips CRCs and we do not override it.
764  */
765 
766 int iflib_crcstrip = 1;
767 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
768     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
769 
770 int iflib_rx_miss, iflib_rx_miss_bufs;
771 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
772     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
773 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
774     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
775 
776 /*
777  * Register/unregister. We are already under netmap lock.
778  * Only called on the first register or the last unregister.
779  */
780 static int
iflib_netmap_register(struct netmap_adapter * na,int onoff)781 iflib_netmap_register(struct netmap_adapter *na, int onoff)
782 {
783 	if_t ifp = na->ifp;
784 	if_ctx_t ctx = if_getsoftc(ifp);
785 	int status;
786 
787 	CTX_LOCK(ctx);
788 	if (!CTX_IS_VF(ctx))
789 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
790 
791 	iflib_stop(ctx);
792 
793 	/*
794 	 * Enable (or disable) netmap flags, and intercept (or restore)
795 	 * ifp->if_transmit. This is done once the device has been stopped
796 	 * to prevent race conditions. Also, this must be done after
797 	 * calling netmap_disable_all_rings() and before calling
798 	 * netmap_enable_all_rings(), so that these two functions see the
799 	 * updated state of the NAF_NETMAP_ON bit.
800 	 */
801 	if (onoff) {
802 		nm_set_native_flags(na);
803 	} else {
804 		nm_clear_native_flags(na);
805 	}
806 
807 	iflib_init_locked(ctx);
808 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
809 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
810 	if (status)
811 		nm_clear_native_flags(na);
812 	CTX_UNLOCK(ctx);
813 	return (status);
814 }
815 
816 static int
iflib_netmap_config(struct netmap_adapter * na,struct nm_config_info * info)817 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
818 {
819 	if_t ifp = na->ifp;
820 	if_ctx_t ctx = if_getsoftc(ifp);
821 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
822 	iflib_fl_t fl = &rxq->ifr_fl[0];
823 
824 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
825 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
826 	info->num_tx_descs = iflib_num_tx_descs(ctx);
827 	info->num_rx_descs = iflib_num_rx_descs(ctx);
828 	info->rx_buf_maxsize = fl->ifl_buf_size;
829 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
830 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
831 		info->num_rx_descs, info->rx_buf_maxsize);
832 
833 	return (0);
834 }
835 
836 static int
netmap_fl_refill(iflib_rxq_t rxq,struct netmap_kring * kring,bool init)837 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
838 {
839 	struct netmap_adapter *na = kring->na;
840 	u_int const lim = kring->nkr_num_slots - 1;
841 	struct netmap_ring *ring = kring->ring;
842 	bus_dmamap_t *map;
843 	struct if_rxd_update iru;
844 	if_ctx_t ctx = rxq->ifr_ctx;
845 	iflib_fl_t fl = &rxq->ifr_fl[0];
846 	u_int nic_i_first, nic_i;
847 	u_int nm_i;
848 	int i, n;
849 #if IFLIB_DEBUG_COUNTERS
850 	int rf_count = 0;
851 #endif
852 
853 	/*
854 	 * This function is used both at initialization and in rxsync.
855 	 * At initialization we need to prepare (with isc_rxd_refill())
856 	 * all the netmap buffers currently owned by the kernel, in
857 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
858 	 * (except for kring->nkr_hwofs). These may be less than
859 	 * kring->nkr_num_slots if netmap_reset() was called while
860 	 * an application using the kring that still owned some
861 	 * buffers.
862 	 * At rxsync time, both indexes point to the next buffer to be
863 	 * refilled.
864 	 * In any case we publish (with isc_rxd_flush()) up to
865 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
866 	 * pointer to overrun the head/cons pointer, although this is
867 	 * not necessary for some NICs (e.g. vmx).
868 	 */
869 	if (__predict_false(init)) {
870 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
871 	} else {
872 		n = kring->rhead - kring->nr_hwcur;
873 		if (n == 0)
874 			return (0); /* Nothing to do. */
875 		if (n < 0)
876 			n += kring->nkr_num_slots;
877 	}
878 
879 	iru_init(&iru, rxq, 0 /* flid */);
880 	map = fl->ifl_sds.ifsd_map;
881 	nic_i = fl->ifl_pidx;
882 	nm_i = netmap_idx_n2k(kring, nic_i);
883 	if (__predict_false(init)) {
884 		/*
885 		 * On init/reset, nic_i must be 0, and we must
886 		 * start to refill from hwtail (see netmap_reset()).
887 		 */
888 		MPASS(nic_i == 0);
889 		MPASS(nm_i == kring->nr_hwtail);
890 	} else
891 		MPASS(nm_i == kring->nr_hwcur);
892 	DBG_COUNTER_INC(fl_refills);
893 	while (n > 0) {
894 #if IFLIB_DEBUG_COUNTERS
895 		if (++rf_count == 9)
896 			DBG_COUNTER_INC(fl_refills_large);
897 #endif
898 		nic_i_first = nic_i;
899 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
900 			struct netmap_slot *slot = &ring->slot[nm_i];
901 			uint64_t paddr;
902 			void *addr = PNMB(na, slot, &paddr);
903 
904 			MPASS(i < IFLIB_MAX_RX_REFRESH);
905 
906 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
907 				return (netmap_ring_reinit(kring));
908 
909 			fl->ifl_bus_addrs[i] = paddr +
910 			    nm_get_offset(kring, slot);
911 			fl->ifl_rxd_idxs[i] = nic_i;
912 
913 			if (__predict_false(init)) {
914 				netmap_load_map(na, fl->ifl_buf_tag,
915 				    map[nic_i], addr);
916 			} else if (slot->flags & NS_BUF_CHANGED) {
917 				/* buffer has changed, reload map */
918 				netmap_reload_map(na, fl->ifl_buf_tag,
919 				    map[nic_i], addr);
920 			}
921 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
922 			    BUS_DMASYNC_PREREAD);
923 			slot->flags &= ~NS_BUF_CHANGED;
924 
925 			nm_i = nm_next(nm_i, lim);
926 			nic_i = nm_next(nic_i, lim);
927 		}
928 
929 		iru.iru_pidx = nic_i_first;
930 		iru.iru_count = i;
931 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
932 	}
933 	fl->ifl_pidx = nic_i;
934 	/*
935 	 * At the end of the loop we must have refilled everything
936 	 * we could possibly refill.
937 	 */
938 	MPASS(nm_i == kring->rhead);
939 	kring->nr_hwcur = nm_i;
940 
941 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
942 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
943 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
944 	    nm_prev(nic_i, lim));
945 	DBG_COUNTER_INC(rxd_flush);
946 
947 	return (0);
948 }
949 
950 #define NETMAP_TX_TIMER_US	90
951 
952 /*
953  * Reconcile kernel and user view of the transmit ring.
954  *
955  * All information is in the kring.
956  * Userspace wants to send packets up to the one before kring->rhead,
957  * kernel knows kring->nr_hwcur is the first unsent packet.
958  *
959  * Here we push packets out (as many as possible), and possibly
960  * reclaim buffers from previously completed transmission.
961  *
962  * The caller (netmap) guarantees that there is only one instance
963  * running at any time. Any interference with other driver
964  * methods should be handled by the individual drivers.
965  */
966 static int
iflib_netmap_txsync(struct netmap_kring * kring,int flags)967 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
968 {
969 	struct netmap_adapter *na = kring->na;
970 	if_t ifp = na->ifp;
971 	struct netmap_ring *ring = kring->ring;
972 	u_int nm_i;	/* index into the netmap kring */
973 	u_int nic_i;	/* index into the NIC ring */
974 	u_int const lim = kring->nkr_num_slots - 1;
975 	u_int const head = kring->rhead;
976 	struct if_pkt_info pi;
977 	int tx_pkts = 0, tx_bytes = 0;
978 
979 	/*
980 	 * interrupts on every tx packet are expensive so request
981 	 * them every half ring, or where NS_REPORT is set
982 	 */
983 	u_int report_frequency = kring->nkr_num_slots >> 1;
984 	/* device-specific */
985 	if_ctx_t ctx = if_getsoftc(ifp);
986 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
987 
988 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
989 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
990 
991 	/*
992 	 * First part: process new packets to send.
993 	 * nm_i is the current index in the netmap kring,
994 	 * nic_i is the corresponding index in the NIC ring.
995 	 *
996 	 * If we have packets to send (nm_i != head)
997 	 * iterate over the netmap ring, fetch length and update
998 	 * the corresponding slot in the NIC ring. Some drivers also
999 	 * need to update the buffer's physical address in the NIC slot
1000 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1001 	 *
1002 	 * The netmap_reload_map() calls is especially expensive,
1003 	 * even when (as in this case) the tag is 0, so do only
1004 	 * when the buffer has actually changed.
1005 	 *
1006 	 * If possible do not set the report/intr bit on all slots,
1007 	 * but only a few times per ring or when NS_REPORT is set.
1008 	 *
1009 	 * Finally, on 10G and faster drivers, it might be useful
1010 	 * to prefetch the next slot and txr entry.
1011 	 */
1012 
1013 	nm_i = kring->nr_hwcur;
1014 	if (nm_i != head) {	/* we have new packets to send */
1015 		uint32_t pkt_len = 0, seg_idx = 0;
1016 		int nic_i_start = -1, flags = 0;
1017 		memset(&pi, 0, sizeof(pi));
1018 		pi.ipi_segs = txq->ift_segs;
1019 		pi.ipi_qsidx = kring->ring_id;
1020 		nic_i = netmap_idx_k2n(kring, nm_i);
1021 
1022 		__builtin_prefetch(&ring->slot[nm_i]);
1023 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1024 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1025 
1026 		while (nm_i != head) {
1027 			struct netmap_slot *slot = &ring->slot[nm_i];
1028 			uint64_t offset = nm_get_offset(kring, slot);
1029 			u_int len = slot->len;
1030 			uint64_t paddr;
1031 			void *addr = PNMB(na, slot, &paddr);
1032 
1033 			flags |= (slot->flags & NS_REPORT ||
1034 				nic_i == 0 || nic_i == report_frequency) ?
1035 				IPI_TX_INTR : 0;
1036 
1037 			/*
1038 			 * If this is the first packet fragment, save the
1039 			 * index of the first NIC slot for later.
1040 			 */
1041 			if (nic_i_start < 0)
1042 				nic_i_start = nic_i;
1043 
1044 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1045 			pi.ipi_segs[seg_idx].ds_len = len;
1046 			if (len) {
1047 				pkt_len += len;
1048 				seg_idx++;
1049 			}
1050 
1051 			if (!(slot->flags & NS_MOREFRAG)) {
1052 				pi.ipi_len = pkt_len;
1053 				pi.ipi_nsegs = seg_idx;
1054 				pi.ipi_pidx = nic_i_start;
1055 				pi.ipi_ndescs = 0;
1056 				pi.ipi_flags = flags;
1057 
1058 				/* Prepare the NIC TX ring. */
1059 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1060 				DBG_COUNTER_INC(tx_encap);
1061 
1062 				/* Update transmit counters */
1063 				tx_bytes += pi.ipi_len;
1064 				tx_pkts++;
1065 
1066 				/* Reinit per-packet info for the next one. */
1067 				flags = seg_idx = pkt_len = 0;
1068 				nic_i_start = -1;
1069 			}
1070 
1071 			/* prefetch for next round */
1072 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1073 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1074 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1075 
1076 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1077 
1078 			if (slot->flags & NS_BUF_CHANGED) {
1079 				/* buffer has changed, reload map */
1080 				netmap_reload_map(na, txq->ift_buf_tag,
1081 				    txq->ift_sds.ifsd_map[nic_i], addr);
1082 			}
1083 			/* make sure changes to the buffer are synced */
1084 			bus_dmamap_sync(txq->ift_buf_tag,
1085 			    txq->ift_sds.ifsd_map[nic_i],
1086 			    BUS_DMASYNC_PREWRITE);
1087 
1088 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1089 			nm_i = nm_next(nm_i, lim);
1090 			nic_i = nm_next(nic_i, lim);
1091 		}
1092 		kring->nr_hwcur = nm_i;
1093 
1094 		/* synchronize the NIC ring */
1095 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1096 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1097 
1098 		/* (re)start the tx unit up to slot nic_i (excluded) */
1099 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1100 	}
1101 
1102 	/*
1103 	 * Second part: reclaim buffers for completed transmissions.
1104 	 *
1105 	 * If there are unclaimed buffers, attempt to reclaim them.
1106 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1107 	 * trigger a per-tx-queue timer to try again later.
1108 	 */
1109 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1110 		if (iflib_tx_credits_update(ctx, txq)) {
1111 			/* some tx completed, increment avail */
1112 			nic_i = txq->ift_cidx_processed;
1113 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1114 		}
1115 	}
1116 
1117 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1118 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1119 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1120 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1121 			    iflib_netmap_timer, txq,
1122 			    txq->ift_netmap_timer.c_cpu, 0);
1123 		}
1124 
1125 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1126 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1127 
1128 	return (0);
1129 }
1130 
1131 /*
1132  * Reconcile kernel and user view of the receive ring.
1133  * Same as for the txsync, this routine must be efficient.
1134  * The caller guarantees a single invocations, but races against
1135  * the rest of the driver should be handled here.
1136  *
1137  * On call, kring->rhead is the first packet that userspace wants
1138  * to keep, and kring->rcur is the wakeup point.
1139  * The kernel has previously reported packets up to kring->rtail.
1140  *
1141  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1142  * of whether or not we received an interrupt.
1143  */
1144 static int
iflib_netmap_rxsync(struct netmap_kring * kring,int flags)1145 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1146 {
1147 	struct netmap_adapter *na = kring->na;
1148 	struct netmap_ring *ring = kring->ring;
1149 	if_t ifp = na->ifp;
1150 	uint32_t nm_i;	/* index into the netmap ring */
1151 	uint32_t nic_i;	/* index into the NIC ring */
1152 	u_int n;
1153 	u_int const lim = kring->nkr_num_slots - 1;
1154 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1155 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1156 
1157 	if_ctx_t ctx = if_getsoftc(ifp);
1158 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1159 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1160 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1161 	iflib_fl_t fl = &rxq->ifr_fl[0];
1162 	struct if_rxd_info ri;
1163 	qidx_t *cidxp;
1164 
1165 	/*
1166 	 * netmap only uses free list 0, to avoid out of order consumption
1167 	 * of receive buffers
1168 	 */
1169 
1170 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1171 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1172 
1173 	/*
1174 	 * First part: import newly received packets.
1175 	 *
1176 	 * nm_i is the index of the next free slot in the netmap ring,
1177 	 * nic_i is the index of the next received packet in the NIC ring
1178 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1179 	 * differ in case if_init() has been called while
1180 	 * in netmap mode. For the receive ring we have
1181 	 *
1182 	 *	nic_i = fl->ifl_cidx;
1183 	 *	nm_i = kring->nr_hwtail (previous)
1184 	 * and
1185 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1186 	 *
1187 	 * fl->ifl_cidx is set to 0 on a ring reinit
1188 	 */
1189 	if (netmap_no_pendintr || force_update) {
1190 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1191 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1192 		int crclen = iflib_crcstrip ? 0 : 4;
1193 		int error, avail;
1194 
1195 		/*
1196 		 * For the free list consumer index, we use the same
1197 		 * logic as in iflib_rxeof().
1198 		 */
1199 		if (have_rxcq)
1200 			cidxp = &rxq->ifr_cq_cidx;
1201 		else
1202 			cidxp = &fl->ifl_cidx;
1203 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1204 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1205 
1206 		nic_i = fl->ifl_cidx;
1207 		nm_i = netmap_idx_n2k(kring, nic_i);
1208 		MPASS(nm_i == kring->nr_hwtail);
1209 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1210 			memset(&ri, 0, sizeof(ri));
1211 			ri.iri_frags = rxq->ifr_frags;
1212 			ri.iri_qsidx = kring->ring_id;
1213 			ri.iri_ifp = ctx->ifc_ifp;
1214 			ri.iri_cidx = *cidxp;
1215 
1216 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1217 			for (i = 0; i < ri.iri_nfrags; i++) {
1218 				if (error) {
1219 					ring->slot[nm_i].len = 0;
1220 					ring->slot[nm_i].flags = 0;
1221 				} else {
1222 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1223 					if (i == (ri.iri_nfrags - 1)) {
1224 						ring->slot[nm_i].len -= crclen;
1225 						ring->slot[nm_i].flags = 0;
1226 
1227 						/* Update receive counters */
1228 						rx_bytes += ri.iri_len;
1229 						rx_pkts++;
1230 					} else
1231 						ring->slot[nm_i].flags = NS_MOREFRAG;
1232 				}
1233 
1234 				bus_dmamap_sync(fl->ifl_buf_tag,
1235 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1236 				nm_i = nm_next(nm_i, lim);
1237 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1238 			}
1239 
1240 			if (have_rxcq) {
1241 				*cidxp = ri.iri_cidx;
1242 				while (*cidxp >= scctx->isc_nrxd[0])
1243 					*cidxp -= scctx->isc_nrxd[0];
1244 			}
1245 
1246 		}
1247 		if (n) { /* update the state variables */
1248 			if (netmap_no_pendintr && !force_update) {
1249 				/* diagnostics */
1250 				iflib_rx_miss++;
1251 				iflib_rx_miss_bufs += n;
1252 			}
1253 			kring->nr_hwtail = nm_i;
1254 		}
1255 		kring->nr_kflags &= ~NKR_PENDINTR;
1256 	}
1257 	/*
1258 	 * Second part: skip past packets that userspace has released.
1259 	 * (kring->nr_hwcur to head excluded),
1260 	 * and make the buffers available for reception.
1261 	 * As usual nm_i is the index in the netmap ring,
1262 	 * nic_i is the index in the NIC ring, and
1263 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1264 	 */
1265 	netmap_fl_refill(rxq, kring, false);
1266 
1267 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1268 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1269 
1270 	return (0);
1271 }
1272 
1273 static void
iflib_netmap_intr(struct netmap_adapter * na,int onoff)1274 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1275 {
1276 	if_ctx_t ctx = if_getsoftc(na->ifp);
1277 
1278 	CTX_LOCK(ctx);
1279 	if (onoff) {
1280 		IFDI_INTR_ENABLE(ctx);
1281 	} else {
1282 		IFDI_INTR_DISABLE(ctx);
1283 	}
1284 	CTX_UNLOCK(ctx);
1285 }
1286 
1287 static int
iflib_netmap_attach(if_ctx_t ctx)1288 iflib_netmap_attach(if_ctx_t ctx)
1289 {
1290 	struct netmap_adapter na;
1291 
1292 	bzero(&na, sizeof(na));
1293 
1294 	na.ifp = ctx->ifc_ifp;
1295 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1296 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1297 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1298 
1299 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1300 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1301 	na.nm_txsync = iflib_netmap_txsync;
1302 	na.nm_rxsync = iflib_netmap_rxsync;
1303 	na.nm_register = iflib_netmap_register;
1304 	na.nm_intr = iflib_netmap_intr;
1305 	na.nm_config = iflib_netmap_config;
1306 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1307 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1308 	return (netmap_attach(&na));
1309 }
1310 
1311 static int
iflib_netmap_txq_init(if_ctx_t ctx,iflib_txq_t txq)1312 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1313 {
1314 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1315 	struct netmap_slot *slot;
1316 
1317 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1318 	if (slot == NULL)
1319 		return (0);
1320 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1321 		/*
1322 		 * In netmap mode, set the map for the packet buffer.
1323 		 * NOTE: Some drivers (not this one) also need to set
1324 		 * the physical buffer address in the NIC ring.
1325 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1326 		 * netmap slot index, si
1327 		 */
1328 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1329 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1330 		    NMB(na, slot + si));
1331 	}
1332 	return (1);
1333 }
1334 
1335 static int
iflib_netmap_rxq_init(if_ctx_t ctx,iflib_rxq_t rxq)1336 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1337 {
1338 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1339 	struct netmap_kring *kring;
1340 	struct netmap_slot *slot;
1341 
1342 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1343 	if (slot == NULL)
1344 		return (0);
1345 	kring = na->rx_rings[rxq->ifr_id];
1346 	netmap_fl_refill(rxq, kring, true);
1347 	return (1);
1348 }
1349 
1350 static void
iflib_netmap_timer(void * arg)1351 iflib_netmap_timer(void *arg)
1352 {
1353 	iflib_txq_t txq = arg;
1354 	if_ctx_t ctx = txq->ift_ctx;
1355 
1356 	/*
1357 	 * Wake up the netmap application, to give it a chance to
1358 	 * call txsync and reclaim more completed TX buffers.
1359 	 */
1360 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1361 }
1362 
1363 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1364 
1365 #else
1366 #define iflib_netmap_txq_init(ctx, txq) (0)
1367 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1368 #define iflib_netmap_detach(ifp)
1369 #define netmap_enable_all_rings(ifp)
1370 #define netmap_disable_all_rings(ifp)
1371 
1372 #define iflib_netmap_attach(ctx) (0)
1373 #define netmap_rx_irq(ifp, qid, budget) (0)
1374 #endif
1375 
1376 #if defined(__i386__) || defined(__amd64__)
1377 static __inline void
prefetch(void * x)1378 prefetch(void *x)
1379 {
1380 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1381 }
1382 
1383 static __inline void
prefetch2cachelines(void * x)1384 prefetch2cachelines(void *x)
1385 {
1386 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1387 #if (CACHE_LINE_SIZE < 128)
1388 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long)))));
1389 #endif
1390 }
1391 #else
1392 static __inline void
prefetch(void * x)1393 prefetch(void *x)
1394 {
1395 }
1396 
1397 static __inline void
prefetch2cachelines(void * x)1398 prefetch2cachelines(void *x)
1399 {
1400 }
1401 #endif
1402 
1403 static void
iru_init(if_rxd_update_t iru,iflib_rxq_t rxq,uint8_t flid)1404 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1405 {
1406 	iflib_fl_t fl;
1407 
1408 	fl = &rxq->ifr_fl[flid];
1409 	iru->iru_paddrs = fl->ifl_bus_addrs;
1410 	iru->iru_idxs = fl->ifl_rxd_idxs;
1411 	iru->iru_qsidx = rxq->ifr_id;
1412 	iru->iru_buf_size = fl->ifl_buf_size;
1413 	iru->iru_flidx = fl->ifl_id;
1414 }
1415 
1416 static void
_iflib_dmamap_cb(void * arg,bus_dma_segment_t * segs,int nseg,int err)1417 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1418 {
1419 	if (err)
1420 		return;
1421 	*(bus_addr_t *) arg = segs[0].ds_addr;
1422 }
1423 
1424 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1425 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1426 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1427 
1428 int
iflib_dma_alloc_align(if_ctx_t ctx,int size,int align,iflib_dma_info_t dma,int mapflags)1429 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1430 {
1431 	int err;
1432 	device_t dev = ctx->ifc_dev;
1433 	bus_addr_t lowaddr;
1434 
1435 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1436 
1437 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1438 		    align, 0,		/* alignment, bounds */
1439 		    lowaddr,		/* lowaddr */
1440 		    BUS_SPACE_MAXADDR,	/* highaddr */
1441 		    NULL, NULL,		/* filter, filterarg */
1442 		    size,		/* maxsize */
1443 		    1,			/* nsegments */
1444 		    size,		/* maxsegsize */
1445 		    BUS_DMA_ALLOCNOW,	/* flags */
1446 		    NULL,		/* lockfunc */
1447 		    NULL,		/* lockarg */
1448 		    &dma->idi_tag);
1449 	if (err) {
1450 		device_printf(dev,
1451 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1452 		    __func__, err, size, align);
1453 		goto fail_0;
1454 	}
1455 
1456 	err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr,
1457 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1458 	if (err) {
1459 		device_printf(dev,
1460 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1461 		    __func__, (uintmax_t)size, err);
1462 		goto fail_1;
1463 	}
1464 
1465 	dma->idi_paddr = IF_BAD_DMA;
1466 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1467 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1468 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1469 		device_printf(dev,
1470 		    "%s: bus_dmamap_load failed: %d\n",
1471 		    __func__, err);
1472 		goto fail_2;
1473 	}
1474 
1475 	dma->idi_size = size;
1476 	return (0);
1477 
1478 fail_2:
1479 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1480 fail_1:
1481 	bus_dma_tag_destroy(dma->idi_tag);
1482 fail_0:
1483 	dma->idi_tag = NULL;
1484 
1485 	return (err);
1486 }
1487 
1488 int
iflib_dma_alloc(if_ctx_t ctx,int size,iflib_dma_info_t dma,int mapflags)1489 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1490 {
1491 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1492 
1493 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1494 
1495 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1496 }
1497 
1498 int
iflib_dma_alloc_multi(if_ctx_t ctx,int * sizes,iflib_dma_info_t * dmalist,int mapflags,int count)1499 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1500 {
1501 	int i, err;
1502 	iflib_dma_info_t *dmaiter;
1503 
1504 	dmaiter = dmalist;
1505 	for (i = 0; i < count; i++, dmaiter++) {
1506 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1507 			break;
1508 	}
1509 	if (err)
1510 		iflib_dma_free_multi(dmalist, i);
1511 	return (err);
1512 }
1513 
1514 void
iflib_dma_free(iflib_dma_info_t dma)1515 iflib_dma_free(iflib_dma_info_t dma)
1516 {
1517 	if (dma->idi_tag == NULL)
1518 		return;
1519 	if (dma->idi_paddr != IF_BAD_DMA) {
1520 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1521 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1522 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1523 		dma->idi_paddr = IF_BAD_DMA;
1524 	}
1525 	if (dma->idi_vaddr != NULL) {
1526 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1527 		dma->idi_vaddr = NULL;
1528 	}
1529 	bus_dma_tag_destroy(dma->idi_tag);
1530 	dma->idi_tag = NULL;
1531 }
1532 
1533 void
iflib_dma_free_multi(iflib_dma_info_t * dmalist,int count)1534 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1535 {
1536 	int i;
1537 	iflib_dma_info_t *dmaiter = dmalist;
1538 
1539 	for (i = 0; i < count; i++, dmaiter++)
1540 		iflib_dma_free(*dmaiter);
1541 }
1542 
1543 static int
iflib_fast_intr(void * arg)1544 iflib_fast_intr(void *arg)
1545 {
1546 	iflib_filter_info_t info = arg;
1547 	struct grouptask *gtask = info->ifi_task;
1548 	int result;
1549 
1550 	DBG_COUNTER_INC(fast_intrs);
1551 	if (info->ifi_filter != NULL) {
1552 		result = info->ifi_filter(info->ifi_filter_arg);
1553 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1554 			return (result);
1555 	}
1556 
1557 	GROUPTASK_ENQUEUE(gtask);
1558 	return (FILTER_HANDLED);
1559 }
1560 
1561 static int
iflib_fast_intr_rxtx(void * arg)1562 iflib_fast_intr_rxtx(void *arg)
1563 {
1564 	iflib_filter_info_t info = arg;
1565 	struct grouptask *gtask = info->ifi_task;
1566 	if_ctx_t ctx;
1567 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1568 	iflib_txq_t txq;
1569 	void *sc;
1570 	int i, cidx, result;
1571 	qidx_t txqid;
1572 	bool intr_enable, intr_legacy;
1573 
1574 	DBG_COUNTER_INC(fast_intrs);
1575 	if (info->ifi_filter != NULL) {
1576 		result = info->ifi_filter(info->ifi_filter_arg);
1577 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1578 			return (result);
1579 	}
1580 
1581 	ctx = rxq->ifr_ctx;
1582 	sc = ctx->ifc_softc;
1583 	intr_enable = false;
1584 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1585 	MPASS(rxq->ifr_ntxqirq);
1586 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1587 		txqid = rxq->ifr_txqid[i];
1588 		txq = &ctx->ifc_txqs[txqid];
1589 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1590 		    BUS_DMASYNC_POSTREAD);
1591 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1592 			if (intr_legacy)
1593 				intr_enable = true;
1594 			else
1595 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1596 			continue;
1597 		}
1598 		GROUPTASK_ENQUEUE(&txq->ift_task);
1599 	}
1600 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1601 		cidx = rxq->ifr_cq_cidx;
1602 	else
1603 		cidx = rxq->ifr_fl[0].ifl_cidx;
1604 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1605 		GROUPTASK_ENQUEUE(gtask);
1606 	else {
1607 		if (intr_legacy)
1608 			intr_enable = true;
1609 		else
1610 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1611 		DBG_COUNTER_INC(rx_intr_enables);
1612 	}
1613 	if (intr_enable)
1614 		IFDI_INTR_ENABLE(ctx);
1615 	return (FILTER_HANDLED);
1616 }
1617 
1618 static int
iflib_fast_intr_ctx(void * arg)1619 iflib_fast_intr_ctx(void *arg)
1620 {
1621 	iflib_filter_info_t info = arg;
1622 	if_ctx_t ctx = info->ifi_ctx;
1623 	int result;
1624 
1625 	DBG_COUNTER_INC(fast_intrs);
1626 	if (info->ifi_filter != NULL) {
1627 		result = info->ifi_filter(info->ifi_filter_arg);
1628 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1629 			return (result);
1630 	}
1631 
1632 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
1633 	return (FILTER_HANDLED);
1634 }
1635 
1636 static int
_iflib_irq_alloc(if_ctx_t ctx,if_irq_t irq,int rid,driver_filter_t filter,driver_intr_t handler,void * arg,const char * name)1637 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1638 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1639 		 const char *name)
1640 {
1641 	struct resource *res;
1642 	void *tag = NULL;
1643 	device_t dev = ctx->ifc_dev;
1644 	int flags, i, rc;
1645 
1646 	flags = RF_ACTIVE;
1647 	if (ctx->ifc_flags & IFC_LEGACY)
1648 		flags |= RF_SHAREABLE;
1649 	MPASS(rid < 512);
1650 	i = rid;
1651 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1652 	if (res == NULL) {
1653 		device_printf(dev,
1654 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1655 		return (ENOMEM);
1656 	}
1657 	irq->ii_res = res;
1658 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1659 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1660 		    filter, handler, arg, &tag);
1661 	if (rc != 0) {
1662 		device_printf(dev,
1663 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1664 		    rid, name ? name : "unknown", rc);
1665 		return (rc);
1666 	} else if (name)
1667 		bus_describe_intr(dev, res, tag, "%s", name);
1668 
1669 	irq->ii_tag = tag;
1670 	return (0);
1671 }
1672 
1673 /*********************************************************************
1674  *
1675  *  Allocate DMA resources for TX buffers as well as memory for the TX
1676  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1677  *  iflib_sw_tx_desc_array structure, storing all the information that
1678  *  is needed to transmit a packet on the wire.  This is called only
1679  *  once at attach, setup is done every reset.
1680  *
1681  **********************************************************************/
1682 static int
iflib_txsd_alloc(iflib_txq_t txq)1683 iflib_txsd_alloc(iflib_txq_t txq)
1684 {
1685 	if_ctx_t ctx = txq->ift_ctx;
1686 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1687 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1688 	device_t dev = ctx->ifc_dev;
1689 	bus_size_t tsomaxsize;
1690 	bus_addr_t lowaddr;
1691 	int err, nsegments, ntsosegments;
1692 	bool tso;
1693 
1694 	nsegments = scctx->isc_tx_nsegments;
1695 	ntsosegments = scctx->isc_tx_tso_segments_max;
1696 	tsomaxsize = scctx->isc_tx_tso_size_max;
1697 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1698 		tsomaxsize += sizeof(struct ether_vlan_header);
1699 	MPASS(scctx->isc_ntxd[0] > 0);
1700 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1701 	MPASS(nsegments > 0);
1702 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1703 		MPASS(ntsosegments > 0);
1704 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1705 	}
1706 
1707 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1708 
1709 	/*
1710 	 * Set up DMA tags for TX buffers.
1711 	 */
1712 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1713 		    1, 0,			/* alignment, bounds */
1714 		    lowaddr,			/* lowaddr */
1715 		    BUS_SPACE_MAXADDR,		/* highaddr */
1716 		    NULL, NULL,			/* filter, filterarg */
1717 		    sctx->isc_tx_maxsize,	/* maxsize */
1718 		    nsegments,			/* nsegments */
1719 		    sctx->isc_tx_maxsegsize,	/* maxsegsize */
1720 		    0,				/* flags */
1721 		    NULL,			/* lockfunc */
1722 		    NULL,			/* lockfuncarg */
1723 		    &txq->ift_buf_tag))) {
1724 		device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err);
1725 		device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1726 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1727 		goto fail;
1728 	}
1729 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1730 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1731 		    1, 0,			/* alignment, bounds */
1732 		    lowaddr,			/* lowaddr */
1733 		    BUS_SPACE_MAXADDR,		/* highaddr */
1734 		    NULL, NULL,			/* filter, filterarg */
1735 		    tsomaxsize,			/* maxsize */
1736 		    ntsosegments,		/* nsegments */
1737 		    sctx->isc_tso_maxsegsize,	/* maxsegsize */
1738 		    0,				/* flags */
1739 		    NULL,			/* lockfunc */
1740 		    NULL,			/* lockfuncarg */
1741 		    &txq->ift_tso_buf_tag))) {
1742 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1743 		    err);
1744 		goto fail;
1745 	}
1746 
1747 	/* Allocate memory for the TX mbuf map. */
1748 	if (!(txq->ift_sds.ifsd_m =
1749 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1750 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1751 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1752 		err = ENOMEM;
1753 		goto fail;
1754 	}
1755 	if (ctx->ifc_sysctl_simple_tx) {
1756 		if (!(txq->ift_sds.ifsd_m_defer =
1757 			(struct mbuf **) malloc(sizeof(struct mbuf *) *
1758 			    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1759 			device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1760 			err = ENOMEM;
1761 			goto fail;
1762 		}
1763 	}
1764 	txq->ift_sds.ifsd_m_deferb = txq->ift_sds.ifsd_m_defer;
1765 	/*
1766 	 * Create the DMA maps for TX buffers.
1767 	 */
1768 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1769 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1770 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1771 		device_printf(dev,
1772 		    "Unable to allocate TX buffer DMA map memory\n");
1773 		err = ENOMEM;
1774 		goto fail;
1775 	}
1776 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1777 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1778 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1779 		device_printf(dev,
1780 		    "Unable to allocate TSO TX buffer map memory\n");
1781 		err = ENOMEM;
1782 		goto fail;
1783 	}
1784 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1785 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1786 		    &txq->ift_sds.ifsd_map[i]);
1787 		if (err != 0) {
1788 			device_printf(dev, "Unable to create TX DMA map\n");
1789 			goto fail;
1790 		}
1791 		if (!tso)
1792 			continue;
1793 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1794 		    &txq->ift_sds.ifsd_tso_map[i]);
1795 		if (err != 0) {
1796 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1797 			goto fail;
1798 		}
1799 	}
1800 	return (0);
1801 fail:
1802 	/* We free all, it handles case where we are in the middle */
1803 	iflib_tx_structures_free(ctx);
1804 	return (err);
1805 }
1806 
1807 static void
iflib_txsd_destroy(if_ctx_t ctx,iflib_txq_t txq,int i)1808 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1809 {
1810 	bus_dmamap_t map;
1811 
1812 	if (txq->ift_sds.ifsd_map != NULL) {
1813 		map = txq->ift_sds.ifsd_map[i];
1814 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1815 		bus_dmamap_unload(txq->ift_buf_tag, map);
1816 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1817 		txq->ift_sds.ifsd_map[i] = NULL;
1818 	}
1819 
1820 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1821 		map = txq->ift_sds.ifsd_tso_map[i];
1822 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1823 		    BUS_DMASYNC_POSTWRITE);
1824 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1825 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1826 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1827 	}
1828 }
1829 
1830 static void
iflib_txq_destroy(iflib_txq_t txq)1831 iflib_txq_destroy(iflib_txq_t txq)
1832 {
1833 	if_ctx_t ctx = txq->ift_ctx;
1834 
1835 	for (int i = 0; i < txq->ift_size; i++)
1836 		iflib_txsd_destroy(ctx, txq, i);
1837 
1838 	if (txq->ift_br != NULL) {
1839 		ifmp_ring_free(txq->ift_br);
1840 		txq->ift_br = NULL;
1841 	}
1842 
1843 	mtx_destroy(&txq->ift_mtx);
1844 
1845 	if (txq->ift_sds.ifsd_map != NULL) {
1846 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1847 		txq->ift_sds.ifsd_map = NULL;
1848 	}
1849 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1850 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1851 		txq->ift_sds.ifsd_tso_map = NULL;
1852 	}
1853 	if (txq->ift_sds.ifsd_m != NULL) {
1854 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1855 		txq->ift_sds.ifsd_m = NULL;
1856 	}
1857 	if (txq->ift_sds.ifsd_m_defer != NULL) {
1858 		free(txq->ift_sds.ifsd_m_defer, M_IFLIB);
1859 		txq->ift_sds.ifsd_m_defer = NULL;
1860 	}
1861 	if (txq->ift_buf_tag != NULL) {
1862 		bus_dma_tag_destroy(txq->ift_buf_tag);
1863 		txq->ift_buf_tag = NULL;
1864 	}
1865 	if (txq->ift_tso_buf_tag != NULL) {
1866 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1867 		txq->ift_tso_buf_tag = NULL;
1868 	}
1869 	if (txq->ift_ifdi != NULL) {
1870 		free(txq->ift_ifdi, M_IFLIB);
1871 	}
1872 }
1873 
1874 static void
iflib_txsd_free(if_ctx_t ctx,iflib_txq_t txq,int i)1875 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1876 {
1877 	struct mbuf *m;
1878 
1879 	m = IFLIB_GET_MBUF(txq->ift_sds.ifsd_m[i]);
1880 	if (m == NULL)
1881 		return;
1882 
1883 	if (txq->ift_sds.ifsd_map != NULL) {
1884 		bus_dmamap_sync(txq->ift_buf_tag,
1885 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1886 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1887 	}
1888 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1889 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1890 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1891 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1892 		    txq->ift_sds.ifsd_tso_map[i]);
1893 	}
1894 	txq->ift_sds.ifsd_m[i] = NULL;
1895 	m_freem(m);
1896 	DBG_COUNTER_INC(tx_frees);
1897 }
1898 
1899 static int
iflib_txq_setup(iflib_txq_t txq)1900 iflib_txq_setup(iflib_txq_t txq)
1901 {
1902 	if_ctx_t ctx = txq->ift_ctx;
1903 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1904 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1905 	iflib_dma_info_t di;
1906 	int i;
1907 
1908 	/* Set number of descriptors available */
1909 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1910 	/* XXX make configurable */
1911 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1912 
1913 	/* Reset indices */
1914 	txq->ift_cidx_processed = 0;
1915 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1916 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1917 	txq->ift_pad = scctx->isc_tx_pad;
1918 
1919 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1920 		bzero((void *)di->idi_vaddr, di->idi_size);
1921 
1922 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1923 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1924 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1925 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1926 	return (0);
1927 }
1928 
1929 /*********************************************************************
1930  *
1931  *  Allocate DMA resources for RX buffers as well as memory for the RX
1932  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1933  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1934  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1935  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1936  *  packet, the maximum number of entries we'll need is equal to the
1937  *  number of hardware receive descriptors that we've allocated.
1938  *
1939  **********************************************************************/
1940 static int
iflib_rxsd_alloc(iflib_rxq_t rxq)1941 iflib_rxsd_alloc(iflib_rxq_t rxq)
1942 {
1943 	if_ctx_t ctx = rxq->ifr_ctx;
1944 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1945 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1946 	device_t dev = ctx->ifc_dev;
1947 	iflib_fl_t fl;
1948 	bus_addr_t lowaddr;
1949 	int err;
1950 
1951 	MPASS(scctx->isc_nrxd[0] > 0);
1952 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1953 
1954 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1955 
1956 	fl = rxq->ifr_fl;
1957 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1958 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1959 		/* Set up DMA tag for RX buffers. */
1960 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1961 			    1, 0,			/* alignment, bounds */
1962 			    lowaddr,			/* lowaddr */
1963 			    BUS_SPACE_MAXADDR,		/* highaddr */
1964 			    NULL, NULL,			/* filter, filterarg */
1965 			    sctx->isc_rx_maxsize,	/* maxsize */
1966 			    sctx->isc_rx_nsegments,	/* nsegments */
1967 			    sctx->isc_rx_maxsegsize,	/* maxsegsize */
1968 			    0,				/* flags */
1969 			    NULL,			/* lockfunc */
1970 			    NULL,			/* lockarg */
1971 			    &fl->ifl_buf_tag);
1972 		if (err) {
1973 			device_printf(dev,
1974 			    "Unable to allocate RX DMA tag: %d\n", err);
1975 			goto fail;
1976 		}
1977 
1978 		/* Allocate memory for the RX mbuf map. */
1979 		if (!(fl->ifl_sds.ifsd_m =
1980 		    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1981 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1982 			device_printf(dev,
1983 			    "Unable to allocate RX mbuf map memory\n");
1984 			err = ENOMEM;
1985 			goto fail;
1986 		}
1987 
1988 		/* Allocate memory for the direct RX cluster pointer map. */
1989 		if (!(fl->ifl_sds.ifsd_cl =
1990 		    (caddr_t *) malloc(sizeof(caddr_t) *
1991 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1992 			device_printf(dev,
1993 			    "Unable to allocate RX cluster map memory\n");
1994 			err = ENOMEM;
1995 			goto fail;
1996 		}
1997 
1998 		/* Allocate memory for the RX cluster bus address map. */
1999 		if (!(fl->ifl_sds.ifsd_ba =
2000 		    (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2001 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2002 			device_printf(dev,
2003 			    "Unable to allocate RX bus address map memory\n");
2004 			err = ENOMEM;
2005 			goto fail;
2006 		}
2007 
2008 		/*
2009 		 * Create the DMA maps for RX buffers.
2010 		 */
2011 		if (!(fl->ifl_sds.ifsd_map =
2012 		    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2013 			device_printf(dev,
2014 			    "Unable to allocate RX buffer DMA map memory\n");
2015 			err = ENOMEM;
2016 			goto fail;
2017 		}
2018 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2019 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2020 			    &fl->ifl_sds.ifsd_map[i]);
2021 			if (err != 0) {
2022 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2023 				goto fail;
2024 			}
2025 		}
2026 	}
2027 	return (0);
2028 
2029 fail:
2030 	iflib_rx_structures_free(ctx);
2031 	return (err);
2032 }
2033 
2034 /*
2035  * Internal service routines
2036  */
2037 
2038 struct rxq_refill_cb_arg {
2039 	int               error;
2040 	bus_dma_segment_t seg;
2041 	int               nseg;
2042 };
2043 
2044 static void
_rxq_refill_cb(void * arg,bus_dma_segment_t * segs,int nseg,int error)2045 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2046 {
2047 	struct rxq_refill_cb_arg *cb_arg = arg;
2048 
2049 	cb_arg->error = error;
2050 	cb_arg->seg = segs[0];
2051 	cb_arg->nseg = nseg;
2052 }
2053 
2054 /**
2055  * iflib_fl_refill - refill an rxq free-buffer list
2056  * @ctx: the iflib context
2057  * @fl: the free list to refill
2058  * @count: the number of new buffers to allocate
2059  *
2060  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2061  * The caller must assure that @count does not exceed the queue's capacity
2062  * minus one (since we always leave a descriptor unavailable).
2063  */
2064 static uint8_t
iflib_fl_refill(if_ctx_t ctx,iflib_fl_t fl,int count)2065 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2066 {
2067 	struct if_rxd_update iru;
2068 	struct rxq_refill_cb_arg cb_arg;
2069 	struct mbuf *m;
2070 	caddr_t cl, *sd_cl;
2071 	struct mbuf **sd_m;
2072 	bus_dmamap_t *sd_map;
2073 	bus_addr_t bus_addr, *sd_ba;
2074 	int err, frag_idx, i, idx, n, pidx;
2075 	qidx_t credits;
2076 
2077 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2078 
2079 	sd_m = fl->ifl_sds.ifsd_m;
2080 	sd_map = fl->ifl_sds.ifsd_map;
2081 	sd_cl = fl->ifl_sds.ifsd_cl;
2082 	sd_ba = fl->ifl_sds.ifsd_ba;
2083 	pidx = fl->ifl_pidx;
2084 	idx = pidx;
2085 	frag_idx = fl->ifl_fragidx;
2086 	credits = fl->ifl_credits;
2087 
2088 	i = 0;
2089 	n = count;
2090 	MPASS(n > 0);
2091 	MPASS(credits + n <= fl->ifl_size);
2092 
2093 	if (pidx < fl->ifl_cidx)
2094 		MPASS(pidx + n <= fl->ifl_cidx);
2095 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2096 		MPASS(fl->ifl_gen == 0);
2097 	if (pidx > fl->ifl_cidx)
2098 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2099 
2100 	DBG_COUNTER_INC(fl_refills);
2101 	if (n > 8)
2102 		DBG_COUNTER_INC(fl_refills_large);
2103 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2104 	while (n-- > 0) {
2105 		/*
2106 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2107 		 * initialized after rx.
2108 		 *
2109 		 * If the cluster is still set then we know a minimum sized
2110 		 * packet was received
2111 		 */
2112 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2113 		    &frag_idx);
2114 		if (frag_idx < 0)
2115 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2116 		MPASS(frag_idx >= 0);
2117 		if ((cl = sd_cl[frag_idx]) == NULL) {
2118 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2119 			if (__predict_false(cl == NULL))
2120 				break;
2121 
2122 			cb_arg.error = 0;
2123 			MPASS(sd_map != NULL);
2124 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2125 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2126 			    BUS_DMA_NOWAIT);
2127 			if (__predict_false(err != 0 || cb_arg.error)) {
2128 				uma_zfree(fl->ifl_zone, cl);
2129 				break;
2130 			}
2131 
2132 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2133 			sd_cl[frag_idx] = cl;
2134 #if MEMORY_LOGGING
2135 			fl->ifl_cl_enqueued++;
2136 #endif
2137 		} else {
2138 			bus_addr = sd_ba[frag_idx];
2139 		}
2140 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2141 		    BUS_DMASYNC_PREREAD);
2142 
2143 		if (sd_m[frag_idx] == NULL) {
2144 			m = m_gethdr_raw(M_NOWAIT, 0);
2145 			if (__predict_false(m == NULL))
2146 				break;
2147 			sd_m[frag_idx] = m;
2148 		}
2149 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2150 #if MEMORY_LOGGING
2151 		fl->ifl_m_enqueued++;
2152 #endif
2153 
2154 		DBG_COUNTER_INC(rx_allocs);
2155 		fl->ifl_rxd_idxs[i] = frag_idx;
2156 		fl->ifl_bus_addrs[i] = bus_addr;
2157 		credits++;
2158 		i++;
2159 		MPASS(credits <= fl->ifl_size);
2160 		if (++idx == fl->ifl_size) {
2161 #ifdef INVARIANTS
2162 			fl->ifl_gen = 1;
2163 #endif
2164 			idx = 0;
2165 		}
2166 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2167 			iru.iru_pidx = pidx;
2168 			iru.iru_count = i;
2169 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2170 			fl->ifl_pidx = idx;
2171 			fl->ifl_credits = credits;
2172 			pidx = idx;
2173 			i = 0;
2174 		}
2175 	}
2176 
2177 	if (n < count - 1) {
2178 		if (i != 0) {
2179 			iru.iru_pidx = pidx;
2180 			iru.iru_count = i;
2181 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2182 			fl->ifl_pidx = idx;
2183 			fl->ifl_credits = credits;
2184 		}
2185 		DBG_COUNTER_INC(rxd_flush);
2186 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2187 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2188 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2189 		    fl->ifl_id, fl->ifl_pidx);
2190 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2191 			fl->ifl_fragidx = frag_idx + 1;
2192 			if (fl->ifl_fragidx == fl->ifl_size)
2193 				fl->ifl_fragidx = 0;
2194 		} else {
2195 			fl->ifl_fragidx = frag_idx;
2196 		}
2197 	}
2198 
2199 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2200 }
2201 
2202 static inline uint8_t
iflib_fl_refill_all(if_ctx_t ctx,iflib_fl_t fl)2203 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2204 {
2205 	/*
2206 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2207 	 * This is important as it confuses most NICs. For instance,
2208 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2209 	 * RDH points to the next receive descriptor to be used by the NIC,
2210 	 * and RDT for the next receive descriptor to be published by the
2211 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2212 	 * The condition RDH == RDT means no descriptors are available to
2213 	 * the NIC, and thus it would be ambiguous if it also meant that
2214 	 * all the descriptors are available to the NIC.
2215 	 */
2216 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2217 #ifdef INVARIANTS
2218 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2219 #endif
2220 
2221 	MPASS(fl->ifl_credits <= fl->ifl_size);
2222 	MPASS(reclaimable == delta);
2223 
2224 	if (reclaimable > 0)
2225 		return (iflib_fl_refill(ctx, fl, reclaimable));
2226 	return (0);
2227 }
2228 
2229 uint8_t
iflib_in_detach(if_ctx_t ctx)2230 iflib_in_detach(if_ctx_t ctx)
2231 {
2232 	bool in_detach;
2233 
2234 	STATE_LOCK(ctx);
2235 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2236 	STATE_UNLOCK(ctx);
2237 	return (in_detach);
2238 }
2239 
2240 static void
iflib_fl_bufs_free(iflib_fl_t fl)2241 iflib_fl_bufs_free(iflib_fl_t fl)
2242 {
2243 	iflib_dma_info_t idi = fl->ifl_ifdi;
2244 	bus_dmamap_t sd_map;
2245 	uint32_t i;
2246 
2247 	for (i = 0; i < fl->ifl_size; i++) {
2248 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2249 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2250 
2251 		if (*sd_cl != NULL) {
2252 			sd_map = fl->ifl_sds.ifsd_map[i];
2253 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2254 			    BUS_DMASYNC_POSTREAD);
2255 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2256 			uma_zfree(fl->ifl_zone, *sd_cl);
2257 			*sd_cl = NULL;
2258 			if (*sd_m != NULL) {
2259 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2260 				m_free_raw(*sd_m);
2261 				*sd_m = NULL;
2262 			}
2263 		} else {
2264 			MPASS(*sd_m == NULL);
2265 		}
2266 #if MEMORY_LOGGING
2267 		fl->ifl_m_dequeued++;
2268 		fl->ifl_cl_dequeued++;
2269 #endif
2270 	}
2271 #ifdef INVARIANTS
2272 	for (i = 0; i < fl->ifl_size; i++) {
2273 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2274 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2275 	}
2276 #endif
2277 	/*
2278 	 * Reset free list values
2279 	 */
2280 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2281 	bzero(idi->idi_vaddr, idi->idi_size);
2282 }
2283 
2284 /*********************************************************************
2285  *
2286  *  Initialize a free list and its buffers.
2287  *
2288  **********************************************************************/
2289 static int
iflib_fl_setup(iflib_fl_t fl)2290 iflib_fl_setup(iflib_fl_t fl)
2291 {
2292 	iflib_rxq_t rxq = fl->ifl_rxq;
2293 	if_ctx_t ctx = rxq->ifr_ctx;
2294 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2295 	int qidx;
2296 
2297 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2298 	/*
2299 	 * Free current RX buffer structs and their mbufs
2300 	 */
2301 	iflib_fl_bufs_free(fl);
2302 	/* Now replenish the mbufs */
2303 	MPASS(fl->ifl_credits == 0);
2304 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2305 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2306 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2307 	else
2308 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2309 	/*
2310 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2311 	 * to the selected mbuf size.
2312 	 */
2313 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2314 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2315 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2316 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2317 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2318 
2319 	/*
2320 	 * Avoid pre-allocating zillions of clusters to an idle card
2321 	 * potentially speeding up attach. In any case make sure
2322 	 * to leave a descriptor unavailable. See the comment in
2323 	 * iflib_fl_refill_all().
2324 	 */
2325 	MPASS(fl->ifl_size > 0);
2326 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2327 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2328 		return (ENOBUFS);
2329 	/*
2330 	 * handle failure
2331 	 */
2332 	MPASS(rxq != NULL);
2333 	MPASS(fl->ifl_ifdi != NULL);
2334 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2335 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2336 	return (0);
2337 }
2338 
2339 /*********************************************************************
2340  *
2341  *  Free receive ring data structures
2342  *
2343  **********************************************************************/
2344 static void
iflib_rx_sds_free(iflib_rxq_t rxq)2345 iflib_rx_sds_free(iflib_rxq_t rxq)
2346 {
2347 	iflib_fl_t fl;
2348 	int i, j;
2349 
2350 	if (rxq->ifr_fl != NULL) {
2351 		for (i = 0; i < rxq->ifr_nfl; i++) {
2352 			fl = &rxq->ifr_fl[i];
2353 			if (fl->ifl_buf_tag != NULL) {
2354 				if (fl->ifl_sds.ifsd_map != NULL) {
2355 					for (j = 0; j < fl->ifl_size; j++) {
2356 						bus_dmamap_sync(
2357 						    fl->ifl_buf_tag,
2358 						    fl->ifl_sds.ifsd_map[j],
2359 						    BUS_DMASYNC_POSTREAD);
2360 						bus_dmamap_unload(
2361 						    fl->ifl_buf_tag,
2362 						    fl->ifl_sds.ifsd_map[j]);
2363 						bus_dmamap_destroy(
2364 						    fl->ifl_buf_tag,
2365 						    fl->ifl_sds.ifsd_map[j]);
2366 					}
2367 				}
2368 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2369 				fl->ifl_buf_tag = NULL;
2370 			}
2371 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2372 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2373 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2374 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2375 			free(fl->ifl_rx_bitmap, M_IFLIB);
2376 			fl->ifl_sds.ifsd_m = NULL;
2377 			fl->ifl_sds.ifsd_cl = NULL;
2378 			fl->ifl_sds.ifsd_ba = NULL;
2379 			fl->ifl_sds.ifsd_map = NULL;
2380 			fl->ifl_rx_bitmap = NULL;
2381 		}
2382 		free(rxq->ifr_fl, M_IFLIB);
2383 		rxq->ifr_fl = NULL;
2384 		free(rxq->ifr_ifdi, M_IFLIB);
2385 		rxq->ifr_ifdi = NULL;
2386 		rxq->ifr_cq_cidx = 0;
2387 	}
2388 }
2389 
2390 /*
2391  * Timer routine
2392  */
2393 static void
iflib_timer(void * arg)2394 iflib_timer(void *arg)
2395 {
2396 	iflib_txq_t txq = arg;
2397 	if_ctx_t ctx = txq->ift_ctx;
2398 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2399 	uint64_t this_tick = ticks;
2400 
2401 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2402 		return;
2403 
2404 	/*
2405 	** Check on the state of the TX queue(s), this
2406 	** can be done without the lock because its RO
2407 	** and the HUNG state will be static if set.
2408 	*/
2409 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2410 		txq->ift_last_timer_tick = this_tick;
2411 		IFDI_TIMER(ctx, txq->ift_id);
2412 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2413 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2414 		     (sctx->isc_pause_frames == 0)))
2415 			goto hung;
2416 
2417 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2418 		    ifmp_ring_is_stalled(txq->ift_br)) {
2419 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2420 			    ("queue can't be marked as hung if interface is down"));
2421 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2422 		}
2423 		txq->ift_cleaned_prev = txq->ift_cleaned;
2424 	}
2425 	/* handle any laggards */
2426 	if (txq->ift_db_pending)
2427 		GROUPTASK_ENQUEUE(&txq->ift_task);
2428 
2429 	sctx->isc_pause_frames = 0;
2430 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2431 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2432 		    txq, txq->ift_timer.c_cpu);
2433 	return;
2434 
2435  hung:
2436 	device_printf(ctx->ifc_dev,
2437 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2438 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2439 	STATE_LOCK(ctx);
2440 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2441 	ctx->ifc_flags |= (IFC_DO_WATCHDOG | IFC_DO_RESET);
2442 	iflib_admin_intr_deferred(ctx);
2443 	STATE_UNLOCK(ctx);
2444 }
2445 
2446 static uint16_t
iflib_get_mbuf_size_for(unsigned int size)2447 iflib_get_mbuf_size_for(unsigned int size)
2448 {
2449 
2450 	if (size <= MCLBYTES)
2451 		return (MCLBYTES);
2452 	else
2453 		return (MJUMPAGESIZE);
2454 }
2455 
2456 static void
iflib_calc_rx_mbuf_sz(if_ctx_t ctx)2457 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2458 {
2459 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2460 
2461 	/*
2462 	 * XXX don't set the max_frame_size to larger
2463 	 * than the hardware can handle
2464 	 */
2465 	ctx->ifc_rx_mbuf_sz =
2466 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2467 }
2468 
2469 uint32_t
iflib_get_rx_mbuf_sz(if_ctx_t ctx)2470 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2471 {
2472 
2473 	return (ctx->ifc_rx_mbuf_sz);
2474 }
2475 
2476 static void
iflib_init_locked(if_ctx_t ctx)2477 iflib_init_locked(if_ctx_t ctx)
2478 {
2479 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2480 	if_t ifp = ctx->ifc_ifp;
2481 	iflib_fl_t fl;
2482 	iflib_txq_t txq;
2483 	iflib_rxq_t rxq;
2484 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2485 
2486 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2487 	IFDI_INTR_DISABLE(ctx);
2488 
2489 	/*
2490 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2491 	 * called without first calling iflib_stop().
2492 	 */
2493 	netmap_disable_all_rings(ifp);
2494 
2495 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2496 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2497 	/* Set hardware offload abilities */
2498 	if_clearhwassist(ifp);
2499 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2500 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2501 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2502 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2503 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2504 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2505 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2506 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2507 
2508 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2509 		CALLOUT_LOCK(txq);
2510 		callout_stop(&txq->ift_timer);
2511 #ifdef DEV_NETMAP
2512 		callout_stop(&txq->ift_netmap_timer);
2513 #endif /* DEV_NETMAP */
2514 		CALLOUT_UNLOCK(txq);
2515 		(void)iflib_netmap_txq_init(ctx, txq);
2516 	}
2517 
2518 	/*
2519 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2520 	 * that drivers can use the value when setting up the hardware receive
2521 	 * buffers.
2522 	 */
2523 	iflib_calc_rx_mbuf_sz(ctx);
2524 
2525 #ifdef INVARIANTS
2526 	i = if_getdrvflags(ifp);
2527 #endif
2528 	IFDI_INIT(ctx);
2529 	MPASS(if_getdrvflags(ifp) == i);
2530 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2531 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2532 			/* This rxq is in netmap mode. Skip normal init. */
2533 			continue;
2534 		}
2535 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2536 			if (iflib_fl_setup(fl)) {
2537 				device_printf(ctx->ifc_dev,
2538 				    "setting up free list %d failed - "
2539 				    "check cluster settings\n", j);
2540 				goto done;
2541 			}
2542 		}
2543 	}
2544 done:
2545 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2546 	IFDI_INTR_ENABLE(ctx);
2547 	txq = ctx->ifc_txqs;
2548 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2549 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2550 			txq->ift_timer.c_cpu);
2551 
2552 	/* Re-enable txsync/rxsync. */
2553 	netmap_enable_all_rings(ifp);
2554 }
2555 
2556 static int
iflib_media_change(if_t ifp)2557 iflib_media_change(if_t ifp)
2558 {
2559 	if_ctx_t ctx = if_getsoftc(ifp);
2560 	int err;
2561 
2562 	CTX_LOCK(ctx);
2563 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2564 		iflib_if_init_locked(ctx);
2565 	CTX_UNLOCK(ctx);
2566 	return (err);
2567 }
2568 
2569 static void
iflib_media_status(if_t ifp,struct ifmediareq * ifmr)2570 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2571 {
2572 	if_ctx_t ctx = if_getsoftc(ifp);
2573 
2574 	CTX_LOCK(ctx);
2575 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2576 	IFDI_MEDIA_STATUS(ctx, ifmr);
2577 	CTX_UNLOCK(ctx);
2578 }
2579 
2580 static void
iflib_stop(if_ctx_t ctx)2581 iflib_stop(if_ctx_t ctx)
2582 {
2583 	iflib_txq_t txq = ctx->ifc_txqs;
2584 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2585 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2586 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2587 	iflib_dma_info_t di;
2588 	iflib_fl_t fl;
2589 	int i, j;
2590 
2591 	/* Tell the stack that the interface is no longer active */
2592 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2593 
2594 	IFDI_INTR_DISABLE(ctx);
2595 	DELAY(1000);
2596 	IFDI_STOP(ctx);
2597 	DELAY(1000);
2598 
2599 	/*
2600 	 * Stop any pending txsync/rxsync and prevent new ones
2601 	 * form starting. Processes blocked in poll() will get
2602 	 * POLLERR.
2603 	 */
2604 	netmap_disable_all_rings(ctx->ifc_ifp);
2605 
2606 	iflib_debug_reset();
2607 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2608 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2609 		/* make sure all transmitters have completed before proceeding XXX */
2610 
2611 		CALLOUT_LOCK(txq);
2612 		callout_stop(&txq->ift_timer);
2613 #ifdef DEV_NETMAP
2614 		callout_stop(&txq->ift_netmap_timer);
2615 #endif /* DEV_NETMAP */
2616 		CALLOUT_UNLOCK(txq);
2617 
2618 		if (!ctx->ifc_sysctl_simple_tx) {
2619 			/* clean any enqueued buffers */
2620 			iflib_ifmp_purge(txq);
2621 		}
2622 		/* Free any existing tx buffers. */
2623 		for (j = 0; j < txq->ift_size; j++) {
2624 			iflib_txsd_free(ctx, txq, j);
2625 		}
2626 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2627 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2628 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2629 			txq->ift_cidx = txq->ift_pidx;
2630 		else
2631 			txq->ift_cidx = txq->ift_pidx = 0;
2632 
2633 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2634 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2635 		txq->ift_pullups = 0;
2636 		ifmp_ring_reset_stats(txq->ift_br);
2637 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2638 			bzero((void *)di->idi_vaddr, di->idi_size);
2639 	}
2640 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2641 		if (rxq->ifr_task.gt_taskqueue != NULL)
2642 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2643 				 &rxq->ifr_task.gt_task);
2644 
2645 		rxq->ifr_cq_cidx = 0;
2646 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2647 			bzero((void *)di->idi_vaddr, di->idi_size);
2648 		/* also resets the free lists pidx/cidx */
2649 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2650 			iflib_fl_bufs_free(fl);
2651 	}
2652 }
2653 
2654 static inline caddr_t
calc_next_rxd(iflib_fl_t fl,int cidx)2655 calc_next_rxd(iflib_fl_t fl, int cidx)
2656 {
2657 	qidx_t size;
2658 	int nrxd;
2659 	caddr_t start, end, cur, next;
2660 
2661 	nrxd = fl->ifl_size;
2662 	size = fl->ifl_rxd_size;
2663 	start = fl->ifl_ifdi->idi_vaddr;
2664 
2665 	if (__predict_false(size == 0))
2666 		return (start);
2667 	cur = start + size * cidx;
2668 	end = start + size * nrxd;
2669 	next = CACHE_PTR_NEXT(cur);
2670 	return (next < end ? next : start);
2671 }
2672 
2673 static inline void
prefetch_pkts(iflib_fl_t fl,int cidx)2674 prefetch_pkts(iflib_fl_t fl, int cidx)
2675 {
2676 	int nextptr;
2677 	int nrxd = fl->ifl_size;
2678 	caddr_t next_rxd;
2679 
2680 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1);
2681 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2682 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2683 	next_rxd = calc_next_rxd(fl, cidx);
2684 	prefetch(next_rxd);
2685 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]);
2686 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]);
2687 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]);
2688 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]);
2689 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]);
2690 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]);
2691 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]);
2692 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]);
2693 }
2694 
2695 static struct mbuf *
rxd_frag_to_sd(iflib_rxq_t rxq,if_rxd_frag_t irf,bool unload,if_rxsd_t sd,int * pf_rv,if_rxd_info_t ri)2696 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2697     int *pf_rv, if_rxd_info_t ri)
2698 {
2699 	bus_dmamap_t map;
2700 	iflib_fl_t fl;
2701 	caddr_t payload;
2702 	struct mbuf *m;
2703 	int flid, cidx, len, next;
2704 
2705 	map = NULL;
2706 	flid = irf->irf_flid;
2707 	cidx = irf->irf_idx;
2708 	fl = &rxq->ifr_fl[flid];
2709 	sd->ifsd_fl = fl;
2710 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2711 	fl->ifl_credits--;
2712 #if MEMORY_LOGGING
2713 	fl->ifl_m_dequeued++;
2714 #endif
2715 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2716 		prefetch_pkts(fl, cidx);
2717 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1);
2718 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2719 	map = fl->ifl_sds.ifsd_map[cidx];
2720 
2721 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2722 
2723 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2724 	    irf->irf_len != 0) {
2725 		payload  = *sd->ifsd_cl;
2726 		payload +=  ri->iri_pad;
2727 		len = ri->iri_len - ri->iri_pad;
2728 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2729 		switch (*pf_rv) {
2730 		case PFIL_DROPPED:
2731 		case PFIL_CONSUMED:
2732 			/*
2733 			 * The filter ate it.  Everything is recycled.
2734 			 */
2735 			m = NULL;
2736 			unload = 0;
2737 			break;
2738 		case PFIL_REALLOCED:
2739 			/*
2740 			 * The filter copied it.  Everything is recycled.
2741 			 * 'm' points at new mbuf.
2742 			 */
2743 			unload = 0;
2744 			break;
2745 		case PFIL_PASS:
2746 			/*
2747 			 * Filter said it was OK, so receive like
2748 			 * normal
2749 			 */
2750 			m = fl->ifl_sds.ifsd_m[cidx];
2751 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2752 			break;
2753 		default:
2754 			MPASS(0);
2755 		}
2756 	} else {
2757 		m = fl->ifl_sds.ifsd_m[cidx];
2758 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2759 		if (pf_rv != NULL)
2760 			*pf_rv = PFIL_PASS;
2761 	}
2762 
2763 	if (unload && irf->irf_len != 0)
2764 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2765 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1);
2766 	if (__predict_false(fl->ifl_cidx == 0))
2767 		fl->ifl_gen = 0;
2768 	bit_clear(fl->ifl_rx_bitmap, cidx);
2769 	return (m);
2770 }
2771 
2772 static struct mbuf *
assemble_segments(iflib_rxq_t rxq,if_rxd_info_t ri,if_rxsd_t sd,int * pf_rv)2773 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2774 {
2775 	struct mbuf *m, *mh, *mt;
2776 	caddr_t cl;
2777 	int  *pf_rv_ptr, flags, i, padlen;
2778 	bool consumed;
2779 
2780 	i = 0;
2781 	mh = NULL;
2782 	consumed = false;
2783 	*pf_rv = PFIL_PASS;
2784 	pf_rv_ptr = pf_rv;
2785 	do {
2786 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2787 		    pf_rv_ptr, ri);
2788 
2789 		MPASS(*sd->ifsd_cl != NULL);
2790 
2791 		/*
2792 		 * Exclude zero-length frags & frags from
2793 		 * packets the filter has consumed or dropped
2794 		 */
2795 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2796 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2797 			if (mh == NULL) {
2798 				/* everything saved here */
2799 				consumed = true;
2800 				pf_rv_ptr = NULL;
2801 				continue;
2802 			}
2803 			/* XXX we can save the cluster here, but not the mbuf */
2804 			m_init(m, M_NOWAIT, MT_DATA, 0);
2805 			m_free(m);
2806 			continue;
2807 		}
2808 		if (mh == NULL) {
2809 			flags = M_PKTHDR | M_EXT;
2810 			mh = mt = m;
2811 			padlen = ri->iri_pad;
2812 		} else {
2813 			flags = M_EXT;
2814 			mt->m_next = m;
2815 			mt = m;
2816 			/* assuming padding is only on the first fragment */
2817 			padlen = 0;
2818 		}
2819 		cl = *sd->ifsd_cl;
2820 		*sd->ifsd_cl = NULL;
2821 
2822 		/* Can these two be made one ? */
2823 		m_init(m, M_NOWAIT, MT_DATA, flags);
2824 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2825 		/*
2826 		 * These must follow m_init and m_cljset
2827 		 */
2828 		m->m_data += padlen;
2829 		ri->iri_len -= padlen;
2830 		m->m_len = ri->iri_frags[i].irf_len;
2831 	} while (++i < ri->iri_nfrags);
2832 
2833 	return (mh);
2834 }
2835 
2836 /*
2837  * Process one software descriptor
2838  */
2839 static struct mbuf *
iflib_rxd_pkt_get(iflib_rxq_t rxq,if_rxd_info_t ri)2840 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2841 {
2842 	struct if_rxsd sd;
2843 	struct mbuf *m;
2844 	int pf_rv;
2845 
2846 	/* should I merge this back in now that the two paths are basically duplicated? */
2847 	if (ri->iri_nfrags == 1 &&
2848 	    ri->iri_frags[0].irf_len != 0 &&
2849 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2850 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2851 		    &pf_rv, ri);
2852 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2853 			return (m);
2854 		if (pf_rv == PFIL_PASS) {
2855 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2856 #ifndef __NO_STRICT_ALIGNMENT
2857 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2858 				m->m_data += 2;
2859 #endif
2860 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2861 			m->m_len = ri->iri_frags[0].irf_len;
2862 			m->m_data += ri->iri_pad;
2863 			ri->iri_len -= ri->iri_pad;
2864 		}
2865 	} else {
2866 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2867 		if (m == NULL)
2868 			return (NULL);
2869 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2870 			return (m);
2871 	}
2872 	m->m_pkthdr.len = ri->iri_len;
2873 	m->m_pkthdr.rcvif = ri->iri_ifp;
2874 	m->m_flags |= ri->iri_flags;
2875 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2876 	m->m_pkthdr.flowid = ri->iri_flowid;
2877 #ifdef NUMA
2878 	m->m_pkthdr.numa_domain = if_getnumadomain(ri->iri_ifp);
2879 #endif
2880 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2881 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2882 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2883 	return (m);
2884 }
2885 
2886 static void
_task_fn_rx_watchdog(void * context)2887 _task_fn_rx_watchdog(void *context)
2888 {
2889 	iflib_rxq_t rxq = context;
2890 
2891 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2892 }
2893 
2894 static uint8_t
iflib_rxeof(iflib_rxq_t rxq,qidx_t budget)2895 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2896 {
2897 	if_t ifp;
2898 	if_ctx_t ctx = rxq->ifr_ctx;
2899 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2900 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2901 	int avail, i;
2902 	qidx_t *cidxp;
2903 	struct if_rxd_info ri;
2904 	int err, budget_left, rx_bytes, rx_pkts;
2905 	iflib_fl_t fl;
2906 #if defined(INET6) || defined(INET)
2907 	int lro_enabled;
2908 #endif
2909 	uint8_t retval = 0;
2910 
2911 	/*
2912 	 * XXX early demux data packets so that if_input processing only handles
2913 	 * acks in interrupt context
2914 	 */
2915 	struct mbuf *m, *mh, *mt;
2916 
2917 	NET_EPOCH_ASSERT();
2918 
2919 	ifp = ctx->ifc_ifp;
2920 	mh = mt = NULL;
2921 	MPASS(budget > 0);
2922 	rx_pkts	= rx_bytes = 0;
2923 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2924 		cidxp = &rxq->ifr_cq_cidx;
2925 	else
2926 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2927 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2928 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2929 			retval |= iflib_fl_refill_all(ctx, fl);
2930 		DBG_COUNTER_INC(rx_unavail);
2931 		return (retval);
2932 	}
2933 
2934 #if defined(INET6) || defined(INET)
2935 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2936 #endif
2937 
2938 	/* pfil needs the vnet to be set */
2939 	CURVNET_SET_QUIET(if_getvnet(ifp));
2940 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2941 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2942 			DBG_COUNTER_INC(rx_ctx_inactive);
2943 			break;
2944 		}
2945 		/*
2946 		 * Reset client set fields to their default values
2947 		 */
2948 		memset(&ri, 0, sizeof(ri));
2949 		ri.iri_qsidx = rxq->ifr_id;
2950 		ri.iri_cidx = *cidxp;
2951 		ri.iri_ifp = ifp;
2952 		ri.iri_frags = rxq->ifr_frags;
2953 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2954 
2955 		if (err)
2956 			goto err;
2957 		rx_pkts += 1;
2958 		rx_bytes += ri.iri_len;
2959 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2960 			*cidxp = ri.iri_cidx;
2961 			/* Update our consumer index */
2962 			/* XXX NB: shurd - check if this is still safe */
2963 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2964 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2965 			/* was this only a completion queue message? */
2966 			if (__predict_false(ri.iri_nfrags == 0))
2967 				continue;
2968 		}
2969 		MPASS(ri.iri_nfrags != 0);
2970 		MPASS(ri.iri_len != 0);
2971 
2972 		/* will advance the cidx on the corresponding free lists */
2973 		m = iflib_rxd_pkt_get(rxq, &ri);
2974 		avail--;
2975 		budget_left--;
2976 		if (avail == 0 && budget_left)
2977 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2978 
2979 		if (__predict_false(m == NULL))
2980 			continue;
2981 
2982 #ifndef __NO_STRICT_ALIGNMENT
2983 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2984 			continue;
2985 #endif
2986 #if defined(INET6) || defined(INET)
2987 		if (lro_enabled) {
2988 			tcp_lro_queue_mbuf(&rxq->ifr_lc, m);
2989 			continue;
2990 		}
2991 #endif
2992 
2993 		if (mh == NULL)
2994 			mh = mt = m;
2995 		else {
2996 			mt->m_nextpkt = m;
2997 			mt = m;
2998 		}
2999 	}
3000 	CURVNET_RESTORE();
3001 	/* make sure that we can refill faster than drain */
3002 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3003 		retval |= iflib_fl_refill_all(ctx, fl);
3004 
3005 	if (mh != NULL) {
3006 		if_input(ifp, mh);
3007 		DBG_COUNTER_INC(rx_if_input);
3008 	}
3009 
3010 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3011 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3012 
3013 	/*
3014 	 * Flush any outstanding LRO work
3015 	 */
3016 #if defined(INET6) || defined(INET)
3017 	tcp_lro_flush_all(&rxq->ifr_lc);
3018 #endif
3019 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3020 		retval |= IFLIB_RXEOF_MORE;
3021 	return (retval);
3022 err:
3023 	STATE_LOCK(ctx);
3024 	ctx->ifc_flags |= IFC_DO_RESET;
3025 	iflib_admin_intr_deferred(ctx);
3026 	STATE_UNLOCK(ctx);
3027 	return (0);
3028 }
3029 
3030 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1)
3031 static inline qidx_t
txq_max_db_deferred(iflib_txq_t txq,qidx_t in_use)3032 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3033 {
3034 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3035 	qidx_t minthresh = txq->ift_size / 8;
3036 	if (in_use > 4 * minthresh)
3037 		return (notify_count);
3038 	if (in_use > 2 * minthresh)
3039 		return (notify_count >> 1);
3040 	if (in_use > minthresh)
3041 		return (notify_count >> 3);
3042 	return (0);
3043 }
3044 
3045 static inline qidx_t
txq_max_rs_deferred(iflib_txq_t txq)3046 txq_max_rs_deferred(iflib_txq_t txq)
3047 {
3048 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3049 	qidx_t minthresh = txq->ift_size / 8;
3050 	if (txq->ift_in_use > 4 * minthresh)
3051 		return (notify_count);
3052 	if (txq->ift_in_use > 2 * minthresh)
3053 		return (notify_count >> 1);
3054 	if (txq->ift_in_use > minthresh)
3055 		return (notify_count >> 2);
3056 	return (2);
3057 }
3058 
3059 #define M_CSUM_FLAGS(m)		((m)->m_pkthdr.csum_flags)
3060 #define M_HAS_VLANTAG(m)	(m->m_flags & M_VLANTAG)
3061 
3062 #define TXQ_MAX_DB_DEFERRED(txq, in_use)	txq_max_db_deferred((txq), (in_use))
3063 #define TXQ_MAX_RS_DEFERRED(txq)	txq_max_rs_deferred(txq)
3064 #define TXQ_MAX_DB_CONSUMED(size)	(size >> 4)
3065 
3066 /* forward compatibility for cxgb */
3067 #define FIRST_QSET(ctx) 0
3068 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3069 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3070 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3071 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3072 
3073 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3074     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3075 
3076 static inline bool
iflib_txd_db_check(iflib_txq_t txq,int ring)3077 iflib_txd_db_check(iflib_txq_t txq, int ring)
3078 {
3079 	if_ctx_t ctx = txq->ift_ctx;
3080 	qidx_t dbval, max;
3081 
3082 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3083 
3084 	/* force || threshold exceeded || at the edge of the ring */
3085 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx))) {
3086 
3087 		/*
3088 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3089 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3090 		 * producer index explicitly (INTC).
3091 		 */
3092 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3093 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3094 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3095 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3096 
3097 		/*
3098 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3099 		 */
3100 		txq->ift_db_pending = txq->ift_npending = 0;
3101 		return (true);
3102 	}
3103 	return (false);
3104 }
3105 
3106 #ifdef PKT_DEBUG
3107 static void
print_pkt(if_pkt_info_t pi)3108 print_pkt(if_pkt_info_t pi)
3109 {
3110 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3111 	    pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3112 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3113 	    pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3114 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3115 	    pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3116 }
3117 #endif
3118 
3119 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3120 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3121 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3122 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3123 
3124 /**
3125  * Parses out ethernet header information in the given mbuf.
3126  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3127  *
3128  * This will account for the VLAN header if present.
3129  *
3130  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3131  * types of packets.
3132  */
3133 static int
iflib_parse_ether_header(if_pkt_info_t pi,struct mbuf ** mp,uint64_t * pullups)3134 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3135 {
3136 	struct ether_vlan_header *eh;
3137 	struct mbuf *m;
3138 
3139 	m = *mp;
3140 	if (__predict_false(m->m_len < sizeof(*eh))) {
3141 		(*pullups)++;
3142 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3143 			return (ENOMEM);
3144 	}
3145 	eh = mtod(m, struct ether_vlan_header *);
3146 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3147 		pi->ipi_etype = ntohs(eh->evl_proto);
3148 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3149 	} else {
3150 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3151 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3152 	}
3153 	*mp = m;
3154 
3155 	return (0);
3156 }
3157 
3158 /**
3159  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3160  * Currently this information includes: IP ToS value, IP header version/presence
3161  *
3162  * This is missing some checks and doesn't edit the packet content as it goes,
3163  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3164  */
3165 static int
iflib_parse_header_partial(if_pkt_info_t pi,struct mbuf ** mp,uint64_t * pullups)3166 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3167 {
3168 	struct mbuf *m;
3169 	int err;
3170 
3171 	*pullups = 0;
3172 	m = *mp;
3173 	if (!M_WRITABLE(m)) {
3174 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3175 			return (ENOMEM);
3176 		} else {
3177 			m_freem(*mp);
3178 			DBG_COUNTER_INC(tx_frees);
3179 			*mp = m;
3180 		}
3181 	}
3182 
3183 	/* Fills out pi->ipi_etype */
3184 	err = iflib_parse_ether_header(pi, mp, pullups);
3185 	if (err)
3186 		return (err);
3187 	m = *mp;
3188 
3189 	switch (pi->ipi_etype) {
3190 #ifdef INET
3191 	case ETHERTYPE_IP:
3192 	{
3193 		struct mbuf *n;
3194 		struct ip *ip = NULL;
3195 		int miniplen;
3196 
3197 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3198 		if (__predict_false(m->m_len < miniplen)) {
3199 			/*
3200 			 * Check for common case where the first mbuf only contains
3201 			 * the Ethernet header
3202 			 */
3203 			if (m->m_len == pi->ipi_ehdrlen) {
3204 				n = m->m_next;
3205 				MPASS(n);
3206 				/* If next mbuf contains at least the minimal IP header, then stop */
3207 				if (n->m_len >= sizeof(*ip)) {
3208 					ip = (struct ip *)n->m_data;
3209 				} else {
3210 					(*pullups)++;
3211 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3212 						return (ENOMEM);
3213 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3214 				}
3215 			} else {
3216 				(*pullups)++;
3217 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3218 					return (ENOMEM);
3219 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3220 			}
3221 		} else {
3222 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3223 		}
3224 
3225 		/* Have the IPv4 header w/ no options here */
3226 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3227 		pi->ipi_ipproto = ip->ip_p;
3228 		pi->ipi_ip_tos = ip->ip_tos;
3229 		pi->ipi_flags |= IPI_TX_IPV4;
3230 
3231 		break;
3232 	}
3233 #endif
3234 #ifdef INET6
3235 	case ETHERTYPE_IPV6:
3236 	{
3237 		struct ip6_hdr *ip6;
3238 
3239 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3240 			(*pullups)++;
3241 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3242 				return (ENOMEM);
3243 		}
3244 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3245 
3246 		/* Have the IPv6 fixed header here */
3247 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3248 		pi->ipi_ipproto = ip6->ip6_nxt;
3249 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3250 		pi->ipi_flags |= IPI_TX_IPV6;
3251 
3252 		break;
3253 	}
3254 #endif
3255 	default:
3256 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3257 		pi->ipi_ip_hlen = 0;
3258 		break;
3259 	}
3260 	*mp = m;
3261 
3262 	return (0);
3263 
3264 }
3265 
3266 static int
iflib_parse_header(iflib_txq_t txq,if_pkt_info_t pi,struct mbuf ** mp)3267 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3268 {
3269 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3270 	struct mbuf *m;
3271 	int err;
3272 
3273 	m = *mp;
3274 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3275 	    M_WRITABLE(m) == 0) {
3276 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3277 			return (ENOMEM);
3278 		} else {
3279 			m_freem(*mp);
3280 			DBG_COUNTER_INC(tx_frees);
3281 			*mp = m;
3282 		}
3283 	}
3284 
3285 	/* Fills out pi->ipi_etype */
3286 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3287 	if (__predict_false(err))
3288 		return (err);
3289 	m = *mp;
3290 
3291 	switch (pi->ipi_etype) {
3292 #ifdef INET
3293 	case ETHERTYPE_IP:
3294 	{
3295 		struct ip *ip;
3296 		struct tcphdr *th;
3297 		uint8_t hlen;
3298 
3299 		hlen = pi->ipi_ehdrlen + sizeof(*ip);
3300 		if (__predict_false(m->m_len < hlen)) {
3301 			txq->ift_pullups++;
3302 			if (__predict_false((m = m_pullup(m, hlen)) == NULL))
3303 				return (ENOMEM);
3304 		}
3305 		ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3306 		hlen = pi->ipi_ehdrlen + (ip->ip_hl << 2);
3307 		if (ip->ip_p == IPPROTO_TCP) {
3308 			hlen += sizeof(*th);
3309 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3310 		} else if (ip->ip_p == IPPROTO_UDP) {
3311 			hlen += sizeof(struct udphdr);
3312 		}
3313 		if (__predict_false(m->m_len < hlen)) {
3314 			txq->ift_pullups++;
3315 			if ((m = m_pullup(m, hlen)) == NULL)
3316 				return (ENOMEM);
3317 		}
3318 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3319 		pi->ipi_ipproto = ip->ip_p;
3320 		pi->ipi_ip_tos = ip->ip_tos;
3321 		pi->ipi_flags |= IPI_TX_IPV4;
3322 
3323 		/* TCP checksum offload may require TCP header length */
3324 		if (IS_TX_OFFLOAD4(pi)) {
3325 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3326 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3327 				pi->ipi_tcp_hlen = th->th_off << 2;
3328 				pi->ipi_tcp_seq = th->th_seq;
3329 			}
3330 			if (IS_TSO4(pi)) {
3331 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3332 					return (ENXIO);
3333 				/*
3334 				 * TSO always requires hardware checksum offload.
3335 				 */
3336 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3337 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3338 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3339 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3340 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3341 					ip->ip_sum = 0;
3342 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3343 				}
3344 			}
3345 		}
3346 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3347 			ip->ip_sum = 0;
3348 
3349 		break;
3350 	}
3351 #endif
3352 #ifdef INET6
3353 	case ETHERTYPE_IPV6:
3354 	{
3355 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3356 		struct tcphdr *th;
3357 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3358 
3359 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3360 			txq->ift_pullups++;
3361 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3362 				return (ENOMEM);
3363 		}
3364 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3365 
3366 		/* XXX-BZ this will go badly in case of ext hdrs. */
3367 		pi->ipi_ipproto = ip6->ip6_nxt;
3368 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3369 		pi->ipi_flags |= IPI_TX_IPV6;
3370 
3371 		/* TCP checksum offload may require TCP header length */
3372 		if (IS_TX_OFFLOAD6(pi)) {
3373 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3374 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3375 					txq->ift_pullups++;
3376 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3377 						return (ENOMEM);
3378 				}
3379 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3380 				pi->ipi_tcp_hlen = th->th_off << 2;
3381 				pi->ipi_tcp_seq = th->th_seq;
3382 			}
3383 			if (IS_TSO6(pi)) {
3384 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3385 					return (ENXIO);
3386 				/*
3387 				 * TSO always requires hardware checksum offload.
3388 				 */
3389 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3390 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3391 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3392 			}
3393 		}
3394 		break;
3395 	}
3396 #endif
3397 	default:
3398 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3399 		pi->ipi_ip_hlen = 0;
3400 		break;
3401 	}
3402 	*mp = m;
3403 
3404 	return (0);
3405 }
3406 
3407 /*
3408  * If dodgy hardware rejects the scatter gather chain we've handed it
3409  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3410  * m_defrag'd mbufs
3411  */
3412 static __noinline struct mbuf *
iflib_remove_mbuf(iflib_txq_t txq)3413 iflib_remove_mbuf(iflib_txq_t txq)
3414 {
3415 	int ntxd, pidx;
3416 	struct mbuf *m, **ifsd_m;
3417 
3418 	ifsd_m = txq->ift_sds.ifsd_m;
3419 	ntxd = txq->ift_size;
3420 	pidx = txq->ift_pidx & (ntxd - 1);
3421 	ifsd_m = txq->ift_sds.ifsd_m;
3422 	m = IFLIB_GET_MBUF(ifsd_m[pidx]);
3423 	ifsd_m[pidx] = NULL;
3424 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3425 	if (txq->ift_sds.ifsd_tso_map != NULL)
3426 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3427 		    txq->ift_sds.ifsd_tso_map[pidx]);
3428 #if MEMORY_LOGGING
3429 	txq->ift_dequeued++;
3430 #endif
3431 	return (m);
3432 }
3433 
3434 /*
3435  * Pad an mbuf to ensure a minimum ethernet frame size.
3436  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3437  */
3438 static __noinline int
iflib_ether_pad(device_t dev,struct mbuf ** m_head,uint16_t min_frame_size)3439 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3440 {
3441 	/*
3442 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3443 	 * and ARP message is the smallest common payload I can think of
3444 	 */
3445 	static char pad[18];	/* just zeros */
3446 	int n;
3447 	struct mbuf *new_head;
3448 
3449 	if (!M_WRITABLE(*m_head)) {
3450 		new_head = m_dup(*m_head, M_NOWAIT);
3451 		if (new_head == NULL) {
3452 			m_freem(*m_head);
3453 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3454 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3455 			DBG_COUNTER_INC(tx_frees);
3456 			return (ENOMEM);
3457 		}
3458 		m_freem(*m_head);
3459 		*m_head = new_head;
3460 	}
3461 
3462 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3463 	     n > 0; n -= sizeof(pad))
3464 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3465 			break;
3466 
3467 	if (n > 0) {
3468 		m_freem(*m_head);
3469 		device_printf(dev, "cannot pad short frame\n");
3470 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3471 		DBG_COUNTER_INC(tx_frees);
3472 		return (ENOBUFS);
3473 	}
3474 
3475 	return (0);
3476 }
3477 
3478 static int
iflib_encap(iflib_txq_t txq,struct mbuf ** m_headp)3479 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3480 {
3481 	if_ctx_t		ctx;
3482 	if_shared_ctx_t		sctx;
3483 	if_softc_ctx_t		scctx;
3484 	bus_dma_tag_t		buf_tag;
3485 	bus_dma_segment_t	*segs;
3486 	struct mbuf		*m_head, **ifsd_m;
3487 	bus_dmamap_t		map;
3488 	struct if_pkt_info	pi;
3489 	uintptr_t		flags;
3490 	int remap = 0;
3491 	int err, nsegs, ndesc, max_segs, pidx;
3492 
3493 	ctx = txq->ift_ctx;
3494 	sctx = ctx->ifc_sctx;
3495 	scctx = &ctx->ifc_softc_ctx;
3496 	segs = txq->ift_segs;
3497 	m_head = *m_headp;
3498 	map = NULL;
3499 
3500 	/*
3501 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3502 	 */
3503 	pidx = txq->ift_pidx;
3504 	map = txq->ift_sds.ifsd_map[pidx];
3505 	ifsd_m = txq->ift_sds.ifsd_m;
3506 
3507 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3508 		buf_tag = txq->ift_tso_buf_tag;
3509 		max_segs = scctx->isc_tx_tso_segments_max;
3510 		map = txq->ift_sds.ifsd_tso_map[pidx];
3511 		MPASS(buf_tag != NULL);
3512 		MPASS(max_segs > 0);
3513 		flags = IFLIB_TSO;
3514 	} else {
3515 		buf_tag = txq->ift_buf_tag;
3516 		max_segs = scctx->isc_tx_nsegments;
3517 		map = txq->ift_sds.ifsd_map[pidx];
3518 		flags = IFLIB_NO_TSO;
3519 	}
3520 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3521 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3522 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3523 		if (err) {
3524 			DBG_COUNTER_INC(encap_txd_encap_fail);
3525 			return (err);
3526 		}
3527 	}
3528 	m_head = *m_headp;
3529 
3530 	memset(&pi, 0, sizeof(pi));
3531 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST));
3532 	pi.ipi_pidx = pidx;
3533 	pi.ipi_qsidx = txq->ift_id;
3534 	pi.ipi_len = m_head->m_pkthdr.len;
3535 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3536 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3537 
3538 	/* deliberate bitwise OR to make one condition */
3539 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3540 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3541 			DBG_COUNTER_INC(encap_txd_encap_fail);
3542 			return (err);
3543 		}
3544 		m_head = *m_headp;
3545 	}
3546 
3547 retry:
3548 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3549 	    BUS_DMA_NOWAIT);
3550 defrag:
3551 	if (__predict_false(err)) {
3552 		switch (err) {
3553 		case EFBIG:
3554 			/* try collapse once and defrag once */
3555 			if (remap == 0) {
3556 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3557 				/* try defrag if collapsing fails */
3558 				if (m_head == NULL)
3559 					remap++;
3560 			}
3561 			if (remap == 1) {
3562 				txq->ift_mbuf_defrag++;
3563 				m_head = m_defrag(*m_headp, M_NOWAIT);
3564 			}
3565 			/*
3566 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3567 			 * failed to map an mbuf that was run through m_defrag
3568 			 */
3569 			MPASS(remap <= 1);
3570 			if (__predict_false(m_head == NULL || remap > 1))
3571 				goto defrag_failed;
3572 			remap++;
3573 			*m_headp = m_head;
3574 			goto retry;
3575 			break;
3576 		case ENOMEM:
3577 			txq->ift_no_tx_dma_setup++;
3578 			break;
3579 		default:
3580 			txq->ift_no_tx_dma_setup++;
3581 			m_freem(*m_headp);
3582 			DBG_COUNTER_INC(tx_frees);
3583 			*m_headp = NULL;
3584 			break;
3585 		}
3586 		txq->ift_map_failed++;
3587 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3588 		DBG_COUNTER_INC(encap_txd_encap_fail);
3589 		return (err);
3590 	}
3591 	ifsd_m[pidx] = IFLIB_SAVE_MBUF(m_head, flags);
3592 	if (m_head->m_pkthdr.csum_flags & CSUM_SND_TAG)
3593 		pi.ipi_mbuf = m_head;
3594 	else
3595 		pi.ipi_mbuf = NULL;
3596 	/*
3597 	 * XXX assumes a 1 to 1 relationship between segments and
3598 	 *        descriptors - this does not hold true on all drivers, e.g.
3599 	 *        cxgb
3600 	 */
3601 	if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3602 		(void)iflib_completed_tx_reclaim(txq, NULL);
3603 		if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3604 			txq->ift_no_desc_avail++;
3605 			bus_dmamap_unload(buf_tag, map);
3606 			DBG_COUNTER_INC(encap_txq_avail_fail);
3607 			DBG_COUNTER_INC(encap_txd_encap_fail);
3608 			if (ctx->ifc_sysctl_simple_tx) {
3609 				*m_headp = m_head = iflib_remove_mbuf(txq);
3610 				m_freem(*m_headp);
3611 				DBG_COUNTER_INC(tx_frees);
3612 				*m_headp = NULL;
3613 			}
3614 			if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3615 				GROUPTASK_ENQUEUE(&txq->ift_task);
3616 			return (ENOBUFS);
3617 		}
3618 	}
3619 	/*
3620 	 * On Intel cards we can greatly reduce the number of TX interrupts
3621 	 * we see by only setting report status on every Nth descriptor.
3622 	 * However, this also means that the driver will need to keep track
3623 	 * of the descriptors that RS was set on to check them for the DD bit.
3624 	 */
3625 	txq->ift_rs_pending += nsegs + 1;
3626 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3627 	    iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx)) {
3628 		pi.ipi_flags |= IPI_TX_INTR;
3629 		txq->ift_rs_pending = 0;
3630 	}
3631 
3632 	pi.ipi_segs = segs;
3633 	pi.ipi_nsegs = nsegs;
3634 
3635 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3636 #ifdef PKT_DEBUG
3637 	print_pkt(&pi);
3638 #endif
3639 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3640 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3641 		DBG_COUNTER_INC(tx_encap);
3642 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3643 
3644 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3645 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3646 			ndesc += txq->ift_size;
3647 			txq->ift_gen = 1;
3648 		}
3649 		/*
3650 		 * drivers can need up to ift_pad sentinels
3651 		 */
3652 		MPASS(ndesc <= pi.ipi_nsegs + txq->ift_pad);
3653 		MPASS(pi.ipi_new_pidx != pidx);
3654 		MPASS(ndesc > 0);
3655 		txq->ift_in_use += ndesc;
3656 		txq->ift_db_pending += ndesc;
3657 
3658 		/*
3659 		 * We update the last software descriptor again here because there may
3660 		 * be a sentinel and/or there may be more mbufs than segments
3661 		 */
3662 		txq->ift_pidx = pi.ipi_new_pidx;
3663 		txq->ift_npending += pi.ipi_ndescs;
3664 	} else {
3665 		*m_headp = m_head = iflib_remove_mbuf(txq);
3666 		if (err == EFBIG) {
3667 			txq->ift_txd_encap_efbig++;
3668 			if (remap < 2) {
3669 				remap = 1;
3670 				goto defrag;
3671 			}
3672 		}
3673 		goto defrag_failed;
3674 	}
3675 	/*
3676 	 * err can't possibly be non-zero here, so we don't neet to test it
3677 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3678 	 */
3679 	return (err);
3680 
3681 defrag_failed:
3682 	txq->ift_mbuf_defrag_failed++;
3683 	txq->ift_map_failed++;
3684 	m_freem(*m_headp);
3685 	DBG_COUNTER_INC(tx_frees);
3686 	*m_headp = NULL;
3687 	DBG_COUNTER_INC(encap_txd_encap_fail);
3688 	return (ENOMEM);
3689 }
3690 
3691 static void
iflib_tx_desc_free(iflib_txq_t txq,int n,struct mbuf ** m_defer)3692 iflib_tx_desc_free(iflib_txq_t txq, int n, struct mbuf **m_defer)
3693 {
3694 	uint32_t qsize, cidx, gen;
3695 	struct mbuf *m, **ifsd_m;
3696 	uintptr_t flags;
3697 
3698 	cidx = txq->ift_cidx;
3699 	gen = txq->ift_gen;
3700 	qsize = txq->ift_size;
3701 	ifsd_m =txq->ift_sds.ifsd_m;
3702 
3703 	while (n-- > 0) {
3704 		if ((m = IFLIB_GET_MBUF(ifsd_m[cidx])) != NULL) {
3705 			flags = IFLIB_GET_FLAGS(ifsd_m[cidx]);
3706 			MPASS(flags != 0);
3707 			if (flags & IFLIB_TSO) {
3708 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3709 				    txq->ift_sds.ifsd_tso_map[cidx],
3710 				    BUS_DMASYNC_POSTWRITE);
3711 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3712 				    txq->ift_sds.ifsd_tso_map[cidx]);
3713 			} else {
3714 				bus_dmamap_sync(txq->ift_buf_tag,
3715 				    txq->ift_sds.ifsd_map[cidx],
3716 				    BUS_DMASYNC_POSTWRITE);
3717 				bus_dmamap_unload(txq->ift_buf_tag,
3718 				    txq->ift_sds.ifsd_map[cidx]);
3719 			}
3720 			/* XXX we don't support any drivers that batch packets yet */
3721 			MPASS(m->m_nextpkt == NULL);
3722 			if (m_defer == NULL) {
3723 				m_freem(m);
3724 			} else if (m != NULL) {
3725 				*m_defer = m;
3726 				m_defer++;
3727 			}
3728 			ifsd_m[cidx] = NULL;
3729 #if MEMORY_LOGGING
3730 			txq->ift_dequeued++;
3731 #endif
3732 			DBG_COUNTER_INC(tx_frees);
3733 		}
3734 		if (__predict_false(++cidx == qsize)) {
3735 			cidx = 0;
3736 			gen = 0;
3737 		}
3738 	}
3739 	txq->ift_cidx = cidx;
3740 	txq->ift_gen = gen;
3741 }
3742 
3743 static __inline int
iflib_txq_can_reclaim(iflib_txq_t txq)3744 iflib_txq_can_reclaim(iflib_txq_t txq)
3745 {
3746 	int reclaim, thresh;
3747 
3748 	thresh = txq->ift_reclaim_thresh;
3749 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3750 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3751 
3752 	if (ticks <= (txq->ift_last_reclaim + txq->ift_reclaim_ticks) &&
3753 	    txq->ift_in_use < thresh)
3754 		return (false);
3755 	iflib_tx_credits_update(txq->ift_ctx, txq);
3756 	reclaim = DESC_RECLAIMABLE(txq);
3757 	if (reclaim <= thresh) {
3758 #ifdef INVARIANTS
3759 		if (iflib_verbose_debug) {
3760 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__,
3761 			    txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3762 			    reclaim, thresh);
3763 		}
3764 #endif
3765 		return (0);
3766 	}
3767 	return (reclaim);
3768 }
3769 
3770 static __inline void
_iflib_completed_tx_reclaim(iflib_txq_t txq,struct mbuf ** m_defer,int reclaim)3771 _iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer, int reclaim)
3772 {
3773 	txq->ift_last_reclaim = ticks;
3774 	iflib_tx_desc_free(txq, reclaim, m_defer);
3775 	txq->ift_cleaned += reclaim;
3776 	txq->ift_in_use -= reclaim;
3777 }
3778 
3779 static __inline int
iflib_completed_tx_reclaim(iflib_txq_t txq,struct mbuf ** m_defer)3780 iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer)
3781 {
3782 	int reclaim;
3783 
3784 	reclaim = iflib_txq_can_reclaim(txq);
3785 	if (reclaim == 0)
3786 		return (0);
3787 	_iflib_completed_tx_reclaim(txq, m_defer, reclaim);
3788 	return (reclaim);
3789 }
3790 
3791 static struct mbuf **
_ring_peek_one(struct ifmp_ring * r,int cidx,int offset,int remaining)3792 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3793 {
3794 	int next, size;
3795 	struct mbuf **items;
3796 
3797 	size = r->size;
3798 	next = (cidx + CACHE_PTR_INCREMENT) & (size - 1);
3799 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3800 
3801 	prefetch(items[(cidx + offset) & (size - 1)]);
3802 	if (remaining > 1) {
3803 		prefetch2cachelines(&items[next]);
3804 		prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]);
3805 		prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]);
3806 		prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]);
3807 	}
3808 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)]));
3809 }
3810 
3811 static void
iflib_txq_check_drain(iflib_txq_t txq,int budget)3812 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3813 {
3814 
3815 	ifmp_ring_check_drainage(txq->ift_br, budget);
3816 }
3817 
3818 static uint32_t
iflib_txq_can_drain(struct ifmp_ring * r)3819 iflib_txq_can_drain(struct ifmp_ring *r)
3820 {
3821 	iflib_txq_t txq = r->cookie;
3822 	if_ctx_t ctx = txq->ift_ctx;
3823 
3824 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx))
3825 		return (1);
3826 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3827 	    BUS_DMASYNC_POSTREAD);
3828 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3829 	    false));
3830 }
3831 
3832 static uint32_t
iflib_txq_drain(struct ifmp_ring * r,uint32_t cidx,uint32_t pidx)3833 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3834 {
3835 	iflib_txq_t txq = r->cookie;
3836 	if_ctx_t ctx = txq->ift_ctx;
3837 	if_t ifp = ctx->ifc_ifp;
3838 	struct mbuf *m, **mp;
3839 	int avail, bytes_sent, skipped, count, err, i;
3840 	int mcast_sent, pkt_sent, reclaimed;
3841 	bool do_prefetch, rang, ring;
3842 
3843 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3844 			    !LINK_ACTIVE(ctx))) {
3845 		DBG_COUNTER_INC(txq_drain_notready);
3846 		return (0);
3847 	}
3848 	reclaimed = iflib_completed_tx_reclaim(txq, NULL);
3849 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3850 	avail = IDXDIFF(pidx, cidx, r->size);
3851 
3852 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3853 		/*
3854 		 * The driver is unloading so we need to free all pending packets.
3855 		 */
3856 		DBG_COUNTER_INC(txq_drain_flushing);
3857 		for (i = 0; i < avail; i++) {
3858 			if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq))
3859 				m_freem(r->items[(cidx + i) & (r->size - 1)]);
3860 			r->items[(cidx + i) & (r->size - 1)] = NULL;
3861 		}
3862 		return (avail);
3863 	}
3864 
3865 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3866 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3867 		CALLOUT_LOCK(txq);
3868 		callout_stop(&txq->ift_timer);
3869 		CALLOUT_UNLOCK(txq);
3870 		DBG_COUNTER_INC(txq_drain_oactive);
3871 		return (0);
3872 	}
3873 
3874 	/*
3875 	 * If we've reclaimed any packets this queue cannot be hung.
3876 	 */
3877 	if (reclaimed)
3878 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3879 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3880 	count = MIN(avail, TX_BATCH_SIZE);
3881 #ifdef INVARIANTS
3882 	if (iflib_verbose_debug)
3883 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__,
3884 		    avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3885 #endif
3886 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3887 	err = 0;
3888 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx); i++) {
3889 		int rem = do_prefetch ? count - i : 0;
3890 
3891 		mp = _ring_peek_one(r, cidx, i, rem);
3892 		MPASS(mp != NULL && *mp != NULL);
3893 
3894 		/*
3895 		 * Completion interrupts will use the address of the txq
3896 		 * as a sentinel to enqueue _something_ in order to acquire
3897 		 * the lock on the mp_ring (there's no direct lock call).
3898 		 * We obviously whave to check for these sentinel cases
3899 		 * and skip them.
3900 		 */
3901 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3902 			skipped++;
3903 			continue;
3904 		}
3905 		err = iflib_encap(txq, mp);
3906 		if (__predict_false(err)) {
3907 			/* no room - bail out */
3908 			if (err == ENOBUFS)
3909 				break;
3910 			skipped++;
3911 			/* we can't send this packet - skip it */
3912 			continue;
3913 		}
3914 		pkt_sent++;
3915 		m = *mp;
3916 		DBG_COUNTER_INC(tx_sent);
3917 		bytes_sent += m->m_pkthdr.len;
3918 		mcast_sent += !!(m->m_flags & M_MCAST);
3919 
3920 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
3921 			break;
3922 		ETHER_BPF_MTAP(ifp, m);
3923 		rang = iflib_txd_db_check(txq, false);
3924 	}
3925 
3926 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3927 	ring = rang ? false  : (iflib_min_tx_latency | err | (!!txq->ift_reclaim_thresh));
3928 	iflib_txd_db_check(txq, ring);
3929 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3930 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3931 	if (mcast_sent)
3932 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3933 #ifdef INVARIANTS
3934 	if (iflib_verbose_debug)
3935 		printf("consumed=%d\n", skipped + pkt_sent);
3936 #endif
3937 	return (skipped + pkt_sent);
3938 }
3939 
3940 static uint32_t
iflib_txq_drain_always(struct ifmp_ring * r)3941 iflib_txq_drain_always(struct ifmp_ring *r)
3942 {
3943 	return (1);
3944 }
3945 
3946 static uint32_t
iflib_txq_drain_free(struct ifmp_ring * r,uint32_t cidx,uint32_t pidx)3947 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3948 {
3949 	int i, avail;
3950 	struct mbuf **mp;
3951 	iflib_txq_t txq;
3952 
3953 	txq = r->cookie;
3954 
3955 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3956 	CALLOUT_LOCK(txq);
3957 	callout_stop(&txq->ift_timer);
3958 	CALLOUT_UNLOCK(txq);
3959 
3960 	avail = IDXDIFF(pidx, cidx, r->size);
3961 	for (i = 0; i < avail; i++) {
3962 		mp = _ring_peek_one(r, cidx, i, avail - i);
3963 		if (__predict_false(*mp == (struct mbuf *)txq))
3964 			continue;
3965 		m_freem(*mp);
3966 		DBG_COUNTER_INC(tx_frees);
3967 	}
3968 	MPASS(ifmp_ring_is_stalled(r) == 0);
3969 	return (avail);
3970 }
3971 
3972 static void
iflib_ifmp_purge(iflib_txq_t txq)3973 iflib_ifmp_purge(iflib_txq_t txq)
3974 {
3975 	struct ifmp_ring *r;
3976 
3977 	r = txq->ift_br;
3978 	r->drain = iflib_txq_drain_free;
3979 	r->can_drain = iflib_txq_drain_always;
3980 
3981 	ifmp_ring_check_drainage(r, r->size);
3982 
3983 	r->drain = iflib_txq_drain;
3984 	r->can_drain = iflib_txq_can_drain;
3985 }
3986 
3987 static void
_task_fn_tx(void * context)3988 _task_fn_tx(void *context)
3989 {
3990 	iflib_txq_t txq = context;
3991 	if_ctx_t ctx = txq->ift_ctx;
3992 	if_t ifp = ctx->ifc_ifp;
3993 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3994 
3995 #ifdef IFLIB_DIAGNOSTICS
3996 	txq->ift_cpu_exec_count[curcpu]++;
3997 #endif
3998 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3999 		return;
4000 #ifdef DEV_NETMAP
4001 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4002 	    netmap_tx_irq(ifp, txq->ift_id))
4003 		goto skip_ifmp;
4004 #endif
4005         if (ctx->ifc_sysctl_simple_tx) {
4006                 mtx_lock(&txq->ift_mtx);
4007                 (void)iflib_completed_tx_reclaim(txq, NULL);
4008                 mtx_unlock(&txq->ift_mtx);
4009                 goto skip_ifmp;
4010         }
4011 #ifdef ALTQ
4012 	if (if_altq_is_enabled(ifp))
4013 		iflib_altq_if_start(ifp);
4014 #endif
4015 	if (txq->ift_db_pending)
4016 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4017 	else if (!abdicate)
4018 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4019 	/*
4020 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4021 	 */
4022 	if (abdicate)
4023 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4024 
4025 skip_ifmp:
4026 	if (ctx->ifc_flags & IFC_LEGACY)
4027 		IFDI_INTR_ENABLE(ctx);
4028 	else
4029 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4030 }
4031 
4032 static void
_task_fn_rx(void * context)4033 _task_fn_rx(void *context)
4034 {
4035 	iflib_rxq_t rxq = context;
4036 	if_ctx_t ctx = rxq->ifr_ctx;
4037 	uint8_t more;
4038 	uint16_t budget;
4039 #ifdef DEV_NETMAP
4040 	u_int work = 0;
4041 	int nmirq;
4042 #endif
4043 
4044 #ifdef IFLIB_DIAGNOSTICS
4045 	rxq->ifr_cpu_exec_count[curcpu]++;
4046 #endif
4047 	DBG_COUNTER_INC(task_fn_rxs);
4048 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4049 		return;
4050 #ifdef DEV_NETMAP
4051 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4052 	if (nmirq != NM_IRQ_PASS) {
4053 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4054 		goto skip_rxeof;
4055 	}
4056 #endif
4057 	budget = ctx->ifc_sysctl_rx_budget;
4058 	if (budget == 0)
4059 		budget = 16;	/* XXX */
4060 	more = iflib_rxeof(rxq, budget);
4061 #ifdef DEV_NETMAP
4062 skip_rxeof:
4063 #endif
4064 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4065 		if (ctx->ifc_flags & IFC_LEGACY)
4066 			IFDI_INTR_ENABLE(ctx);
4067 		else
4068 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4069 		DBG_COUNTER_INC(rx_intr_enables);
4070 	}
4071 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4072 		return;
4073 
4074 	if (more & IFLIB_RXEOF_MORE)
4075 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4076 	else if (more & IFLIB_RXEOF_EMPTY)
4077 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4078 }
4079 
4080 static void
_task_fn_admin(void * context,int pending)4081 _task_fn_admin(void *context, int pending)
4082 {
4083 	if_ctx_t ctx = context;
4084 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4085 	iflib_txq_t txq;
4086 	int i;
4087 	bool oactive, running, do_reset, do_watchdog, in_detach;
4088 
4089 	STATE_LOCK(ctx);
4090 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4091 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4092 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4093 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4094 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4095 	ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_WATCHDOG);
4096 	STATE_UNLOCK(ctx);
4097 
4098 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4099 		return;
4100 	if (in_detach)
4101 		return;
4102 
4103 	CTX_LOCK(ctx);
4104 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4105 		CALLOUT_LOCK(txq);
4106 		callout_stop(&txq->ift_timer);
4107 		CALLOUT_UNLOCK(txq);
4108 	}
4109 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4110 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4111 	if (do_watchdog) {
4112 		ctx->ifc_watchdog_events++;
4113 		IFDI_WATCHDOG_RESET(ctx);
4114 	}
4115 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4116 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4117 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4118 		    txq->ift_timer.c_cpu);
4119 	}
4120 	IFDI_LINK_INTR_ENABLE(ctx);
4121 	if (do_reset)
4122 		iflib_if_init_locked(ctx);
4123 	CTX_UNLOCK(ctx);
4124 
4125 	if (LINK_ACTIVE(ctx) == 0)
4126 		return;
4127 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4128 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4129 }
4130 
4131 static void
_task_fn_iov(void * context,int pending)4132 _task_fn_iov(void *context, int pending)
4133 {
4134 	if_ctx_t ctx = context;
4135 
4136 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4137 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4138 		return;
4139 
4140 	CTX_LOCK(ctx);
4141 	IFDI_VFLR_HANDLE(ctx);
4142 	CTX_UNLOCK(ctx);
4143 }
4144 
4145 static int
iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)4146 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4147 {
4148 	int err;
4149 	if_int_delay_info_t info;
4150 	if_ctx_t ctx;
4151 
4152 	info = (if_int_delay_info_t)arg1;
4153 	ctx = info->iidi_ctx;
4154 	info->iidi_req = req;
4155 	info->iidi_oidp = oidp;
4156 	CTX_LOCK(ctx);
4157 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4158 	CTX_UNLOCK(ctx);
4159 	return (err);
4160 }
4161 
4162 /*********************************************************************
4163  *
4164  *  IFNET FUNCTIONS
4165  *
4166  **********************************************************************/
4167 
4168 static void
iflib_if_init_locked(if_ctx_t ctx)4169 iflib_if_init_locked(if_ctx_t ctx)
4170 {
4171 	iflib_stop(ctx);
4172 	iflib_init_locked(ctx);
4173 }
4174 
4175 static void
iflib_if_init(void * arg)4176 iflib_if_init(void *arg)
4177 {
4178 	if_ctx_t ctx = arg;
4179 
4180 	CTX_LOCK(ctx);
4181 	iflib_if_init_locked(ctx);
4182 	CTX_UNLOCK(ctx);
4183 }
4184 
4185 static int
iflib_if_transmit(if_t ifp,struct mbuf * m)4186 iflib_if_transmit(if_t ifp, struct mbuf *m)
4187 {
4188 	if_ctx_t ctx = if_getsoftc(ifp);
4189 	iflib_txq_t txq;
4190 	int err, qidx;
4191 	int abdicate;
4192 
4193 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4194 		DBG_COUNTER_INC(tx_frees);
4195 		m_freem(m);
4196 		return (ENETDOWN);
4197 	}
4198 
4199 	MPASS(m->m_nextpkt == NULL);
4200 	/* ALTQ-enabled interfaces always use queue 0. */
4201 	qidx = 0;
4202 	/* Use driver-supplied queue selection method if it exists */
4203 	if (ctx->isc_txq_select_v2) {
4204 		struct if_pkt_info pi;
4205 		uint64_t early_pullups = 0;
4206 		memset(&pi, 0, sizeof(pi));
4207 
4208 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4209 		if (__predict_false(err != 0)) {
4210 			/* Assign pullups for bad pkts to default queue */
4211 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4212 			DBG_COUNTER_INC(encap_txd_encap_fail);
4213 			return (err);
4214 		}
4215 		/* Let driver make queueing decision */
4216 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4217 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4218 	}
4219 	/* Backwards compatibility w/ simpler queue select */
4220 	else if (ctx->isc_txq_select)
4221 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4222 	/* If not, use iflib's standard method */
4223 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4224 		qidx = QIDX(ctx, m);
4225 
4226 	/* Set TX queue */
4227 	txq = &ctx->ifc_txqs[qidx];
4228 
4229 #ifdef DRIVER_BACKPRESSURE
4230 	if (txq->ift_closed) {
4231 		while (m != NULL) {
4232 			next = m->m_nextpkt;
4233 			m->m_nextpkt = NULL;
4234 			m_freem(m);
4235 			DBG_COUNTER_INC(tx_frees);
4236 			m = next;
4237 		}
4238 		return (ENOBUFS);
4239 	}
4240 #endif
4241 #ifdef notyet
4242 	qidx = count = 0;
4243 	mp = marr;
4244 	next = m;
4245 	do {
4246 		count++;
4247 		next = next->m_nextpkt;
4248 	} while (next != NULL);
4249 
4250 	if (count > nitems(marr))
4251 		if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4252 			/* XXX check nextpkt */
4253 			m_freem(m);
4254 			/* XXX simplify for now */
4255 			DBG_COUNTER_INC(tx_frees);
4256 			return (ENOBUFS);
4257 		}
4258 	for (next = m, i = 0; next != NULL; i++) {
4259 		mp[i] = next;
4260 		next = next->m_nextpkt;
4261 		mp[i]->m_nextpkt = NULL;
4262 	}
4263 #endif
4264 	DBG_COUNTER_INC(tx_seen);
4265 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4266 
4267 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4268 
4269 	if (abdicate)
4270 		GROUPTASK_ENQUEUE(&txq->ift_task);
4271 	if (err) {
4272 		if (!abdicate)
4273 			GROUPTASK_ENQUEUE(&txq->ift_task);
4274 		/* support forthcoming later */
4275 #ifdef DRIVER_BACKPRESSURE
4276 		txq->ift_closed = TRUE;
4277 #endif
4278 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4279 		m_freem(m);
4280 		DBG_COUNTER_INC(tx_frees);
4281 		if (err == ENOBUFS)
4282 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
4283 		else
4284 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
4285 	}
4286 
4287 	return (err);
4288 }
4289 
4290 #ifdef ALTQ
4291 /*
4292  * The overall approach to integrating iflib with ALTQ is to continue to use
4293  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4294  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4295  * is redundant/unnecessary, but doing so minimizes the amount of
4296  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4297  * redundantly queueing to an intermediate mp_ring is swamped by the
4298  * performance limitations inherent in using ALTQ.
4299  *
4300  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4301  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4302  * given interface.  If ALTQ is enabled for an interface, then all
4303  * transmitted packets for that interface will be submitted to the ALTQ
4304  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4305  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4306  * update stats that the iflib machinery handles, and which is sensitve to
4307  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4308  * will be installed as the start routine for use by ALTQ facilities that
4309  * need to trigger queue drains on a scheduled basis.
4310  *
4311  */
4312 static void
iflib_altq_if_start(if_t ifp)4313 iflib_altq_if_start(if_t ifp)
4314 {
4315 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4316 	struct mbuf *m;
4317 
4318 	IFQ_LOCK(ifq);
4319 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4320 	while (m != NULL) {
4321 		iflib_if_transmit(ifp, m);
4322 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4323 	}
4324 	IFQ_UNLOCK(ifq);
4325 }
4326 
4327 static int
iflib_altq_if_transmit(if_t ifp,struct mbuf * m)4328 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4329 {
4330 	int err;
4331 
4332 	if (if_altq_is_enabled(ifp)) {
4333 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4334 		if (err == 0)
4335 			iflib_altq_if_start(ifp);
4336 	} else
4337 		err = iflib_if_transmit(ifp, m);
4338 
4339 	return (err);
4340 }
4341 #endif /* ALTQ */
4342 
4343 static void
iflib_if_qflush(if_t ifp)4344 iflib_if_qflush(if_t ifp)
4345 {
4346 	if_ctx_t ctx = if_getsoftc(ifp);
4347 	iflib_txq_t txq = ctx->ifc_txqs;
4348 	int i;
4349 
4350 	STATE_LOCK(ctx);
4351 	ctx->ifc_flags |= IFC_QFLUSH;
4352 	STATE_UNLOCK(ctx);
4353 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4354 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4355 			iflib_txq_check_drain(txq, 0);
4356 	STATE_LOCK(ctx);
4357 	ctx->ifc_flags &= ~IFC_QFLUSH;
4358 	STATE_UNLOCK(ctx);
4359 
4360 	/*
4361 	 * When ALTQ is enabled, this will also take care of purging the
4362 	 * ALTQ queue(s).
4363 	 */
4364 	if_qflush(ifp);
4365 }
4366 
4367 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4368 		    IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4369 		    IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4370 		    IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4371 
4372 static int
iflib_if_ioctl(if_t ifp,u_long command,caddr_t data)4373 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4374 {
4375 	if_ctx_t ctx = if_getsoftc(ifp);
4376 	struct ifreq	*ifr = (struct ifreq *)data;
4377 #if defined(INET) || defined(INET6)
4378 	struct ifaddr	*ifa = (struct ifaddr *)data;
4379 #endif
4380 	bool		avoid_reset = false;
4381 	int		err = 0, reinit = 0, bits;
4382 
4383 	switch (command) {
4384 	case SIOCSIFADDR:
4385 #ifdef INET
4386 		if (ifa->ifa_addr->sa_family == AF_INET)
4387 			avoid_reset = true;
4388 #endif
4389 #ifdef INET6
4390 		if (ifa->ifa_addr->sa_family == AF_INET6)
4391 			avoid_reset = true;
4392 #endif
4393 		/*
4394 		 * Calling init results in link renegotiation,
4395 		 * so we avoid doing it when possible.
4396 		 */
4397 		if (avoid_reset) {
4398 			if_setflagbits(ifp, IFF_UP, 0);
4399 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4400 				reinit = 1;
4401 #ifdef INET
4402 			if (!(if_getflags(ifp) & IFF_NOARP))
4403 				arp_ifinit(ifp, ifa);
4404 #endif
4405 		} else
4406 			err = ether_ioctl(ifp, command, data);
4407 		break;
4408 	case SIOCSIFMTU:
4409 		CTX_LOCK(ctx);
4410 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4411 			CTX_UNLOCK(ctx);
4412 			break;
4413 		}
4414 		bits = if_getdrvflags(ifp);
4415 		/* stop the driver and free any clusters before proceeding */
4416 		iflib_stop(ctx);
4417 
4418 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4419 			STATE_LOCK(ctx);
4420 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4421 				ctx->ifc_flags |= IFC_MULTISEG;
4422 			else
4423 				ctx->ifc_flags &= ~IFC_MULTISEG;
4424 			STATE_UNLOCK(ctx);
4425 			err = if_setmtu(ifp, ifr->ifr_mtu);
4426 		}
4427 		iflib_init_locked(ctx);
4428 		STATE_LOCK(ctx);
4429 		if_setdrvflags(ifp, bits);
4430 		STATE_UNLOCK(ctx);
4431 		CTX_UNLOCK(ctx);
4432 		break;
4433 	case SIOCSIFFLAGS:
4434 		CTX_LOCK(ctx);
4435 		if (if_getflags(ifp) & IFF_UP) {
4436 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4437 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4438 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4439 					CTX_UNLOCK(ctx);
4440 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4441 					CTX_LOCK(ctx);
4442 				}
4443 			} else
4444 				reinit = 1;
4445 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4446 			iflib_stop(ctx);
4447 		}
4448 		ctx->ifc_if_flags = if_getflags(ifp);
4449 		CTX_UNLOCK(ctx);
4450 		break;
4451 	case SIOCADDMULTI:
4452 	case SIOCDELMULTI:
4453 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4454 			CTX_LOCK(ctx);
4455 			IFDI_INTR_DISABLE(ctx);
4456 			IFDI_MULTI_SET(ctx);
4457 			IFDI_INTR_ENABLE(ctx);
4458 			CTX_UNLOCK(ctx);
4459 		}
4460 		break;
4461 	case SIOCSIFMEDIA:
4462 		CTX_LOCK(ctx);
4463 		IFDI_MEDIA_SET(ctx);
4464 		CTX_UNLOCK(ctx);
4465 		/* FALLTHROUGH */
4466 	case SIOCGIFMEDIA:
4467 	case SIOCGIFXMEDIA:
4468 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4469 		break;
4470 	case SIOCGI2C:
4471 	{
4472 		struct ifi2creq i2c;
4473 
4474 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4475 		if (err != 0)
4476 			break;
4477 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4478 			err = EINVAL;
4479 			break;
4480 		}
4481 		if (i2c.len > sizeof(i2c.data)) {
4482 			err = EINVAL;
4483 			break;
4484 		}
4485 
4486 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4487 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4488 			    sizeof(i2c));
4489 		break;
4490 	}
4491 	case SIOCSIFCAP:
4492 	{
4493 		int mask, setmask, oldmask;
4494 
4495 		oldmask = if_getcapenable(ifp);
4496 		mask = ifr->ifr_reqcap ^ oldmask;
4497 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4498 		setmask = 0;
4499 #ifdef TCP_OFFLOAD
4500 		setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6);
4501 #endif
4502 		setmask |= (mask & IFCAP_FLAGS);
4503 		setmask |= (mask & IFCAP_WOL);
4504 
4505 		/*
4506 		 * If any RX csum has changed, change all the ones that
4507 		 * are supported by the driver.
4508 		 */
4509 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4510 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4511 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4512 		}
4513 
4514 		/*
4515 		 * want to ensure that traffic has stopped before we change any of the flags
4516 		 */
4517 		if (setmask) {
4518 			CTX_LOCK(ctx);
4519 			bits = if_getdrvflags(ifp);
4520 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4521 				iflib_stop(ctx);
4522 			STATE_LOCK(ctx);
4523 			if_togglecapenable(ifp, setmask);
4524 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4525 			STATE_UNLOCK(ctx);
4526 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4527 				iflib_init_locked(ctx);
4528 			STATE_LOCK(ctx);
4529 			if_setdrvflags(ifp, bits);
4530 			STATE_UNLOCK(ctx);
4531 			CTX_UNLOCK(ctx);
4532 		}
4533 		if_vlancap(ifp);
4534 		break;
4535 	}
4536 	case SIOCGPRIVATE_0:
4537 	case SIOCSDRVSPEC:
4538 	case SIOCGDRVSPEC:
4539 		CTX_LOCK(ctx);
4540 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4541 		CTX_UNLOCK(ctx);
4542 		break;
4543 	default:
4544 		err = ether_ioctl(ifp, command, data);
4545 		break;
4546 	}
4547 	if (reinit)
4548 		iflib_if_init(ctx);
4549 	return (err);
4550 }
4551 
4552 static uint64_t
iflib_if_get_counter(if_t ifp,ift_counter cnt)4553 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4554 {
4555 	if_ctx_t ctx = if_getsoftc(ifp);
4556 
4557 	return (IFDI_GET_COUNTER(ctx, cnt));
4558 }
4559 
4560 /*********************************************************************
4561  *
4562  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4563  *
4564  **********************************************************************/
4565 
4566 static void
iflib_vlan_register(void * arg,if_t ifp,uint16_t vtag)4567 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4568 {
4569 	if_ctx_t ctx = if_getsoftc(ifp);
4570 
4571 	if ((void *)ctx != arg)
4572 		return;
4573 
4574 	if ((vtag == 0) || (vtag > 4095))
4575 		return;
4576 
4577 	if (iflib_in_detach(ctx))
4578 		return;
4579 
4580 	CTX_LOCK(ctx);
4581 	/* Driver may need all untagged packets to be flushed */
4582 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4583 		iflib_stop(ctx);
4584 	IFDI_VLAN_REGISTER(ctx, vtag);
4585 	/* Re-init to load the changes, if required */
4586 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4587 		iflib_init_locked(ctx);
4588 	CTX_UNLOCK(ctx);
4589 }
4590 
4591 static void
iflib_vlan_unregister(void * arg,if_t ifp,uint16_t vtag)4592 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4593 {
4594 	if_ctx_t ctx = if_getsoftc(ifp);
4595 
4596 	if ((void *)ctx != arg)
4597 		return;
4598 
4599 	if ((vtag == 0) || (vtag > 4095))
4600 		return;
4601 
4602 	CTX_LOCK(ctx);
4603 	/* Driver may need all tagged packets to be flushed */
4604 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4605 		iflib_stop(ctx);
4606 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4607 	/* Re-init to load the changes, if required */
4608 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4609 		iflib_init_locked(ctx);
4610 	CTX_UNLOCK(ctx);
4611 }
4612 
4613 static void
iflib_led_func(void * arg,int onoff)4614 iflib_led_func(void *arg, int onoff)
4615 {
4616 	if_ctx_t ctx = arg;
4617 
4618 	CTX_LOCK(ctx);
4619 	IFDI_LED_FUNC(ctx, onoff);
4620 	CTX_UNLOCK(ctx);
4621 }
4622 
4623 /*********************************************************************
4624  *
4625  *  BUS FUNCTION DEFINITIONS
4626  *
4627  **********************************************************************/
4628 
4629 int
iflib_device_probe(device_t dev)4630 iflib_device_probe(device_t dev)
4631 {
4632 	const pci_vendor_info_t *ent;
4633 	if_shared_ctx_t sctx;
4634 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4635 	uint16_t pci_vendor_id;
4636 
4637 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4638 		return (ENOTSUP);
4639 
4640 	pci_vendor_id = pci_get_vendor(dev);
4641 	pci_device_id = pci_get_device(dev);
4642 	pci_subvendor_id = pci_get_subvendor(dev);
4643 	pci_subdevice_id = pci_get_subdevice(dev);
4644 	pci_rev_id = pci_get_revid(dev);
4645 	if (sctx->isc_parse_devinfo != NULL)
4646 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4647 
4648 	ent = sctx->isc_vendor_info;
4649 	while (ent->pvi_vendor_id != 0) {
4650 		if (pci_vendor_id != ent->pvi_vendor_id) {
4651 			ent++;
4652 			continue;
4653 		}
4654 		if ((pci_device_id == ent->pvi_device_id) &&
4655 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4656 		     (ent->pvi_subvendor_id == 0)) &&
4657 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4658 		     (ent->pvi_subdevice_id == 0)) &&
4659 		    ((pci_rev_id == ent->pvi_rev_id) ||
4660 		     (ent->pvi_rev_id == 0))) {
4661 			device_set_desc_copy(dev, ent->pvi_name);
4662 			/* this needs to be changed to zero if the bus probing code
4663 			 * ever stops re-probing on best match because the sctx
4664 			 * may have its values over written by register calls
4665 			 * in subsequent probes
4666 			 */
4667 			return (BUS_PROBE_DEFAULT);
4668 		}
4669 		ent++;
4670 	}
4671 	return (ENXIO);
4672 }
4673 
4674 int
iflib_device_probe_vendor(device_t dev)4675 iflib_device_probe_vendor(device_t dev)
4676 {
4677 	int probe;
4678 
4679 	probe = iflib_device_probe(dev);
4680 	if (probe == BUS_PROBE_DEFAULT)
4681 		return (BUS_PROBE_VENDOR);
4682 	else
4683 		return (probe);
4684 }
4685 
4686 static void
iflib_reset_qvalues(if_ctx_t ctx)4687 iflib_reset_qvalues(if_ctx_t ctx)
4688 {
4689 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4690 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4691 	device_t dev = ctx->ifc_dev;
4692 	int i;
4693 
4694 	if (ctx->ifc_sysctl_ntxqs != 0)
4695 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4696 	if (ctx->ifc_sysctl_nrxqs != 0)
4697 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4698 
4699 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4700 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4701 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4702 		else
4703 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4704 	}
4705 
4706 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4707 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4708 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4709 		else
4710 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4711 	}
4712 
4713 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4714 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4715 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4716 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4717 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4718 		}
4719 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4720 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4721 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4722 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4723 		}
4724 		if (!powerof2(scctx->isc_nrxd[i])) {
4725 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4726 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4727 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4728 		}
4729 	}
4730 
4731 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4732 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4733 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4734 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4735 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4736 		}
4737 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4738 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4739 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4740 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4741 		}
4742 		if (!powerof2(scctx->isc_ntxd[i])) {
4743 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4744 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4745 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4746 		}
4747 	}
4748 	scctx->isc_tx_pad = 2;
4749 }
4750 
4751 static void
iflib_add_pfil(if_ctx_t ctx)4752 iflib_add_pfil(if_ctx_t ctx)
4753 {
4754 	struct pfil_head *pfil;
4755 	struct pfil_head_args pa;
4756 	iflib_rxq_t rxq;
4757 	int i;
4758 
4759 	pa.pa_version = PFIL_VERSION;
4760 	pa.pa_flags = PFIL_IN;
4761 	pa.pa_type = PFIL_TYPE_ETHERNET;
4762 	pa.pa_headname = if_name(ctx->ifc_ifp);
4763 	pfil = pfil_head_register(&pa);
4764 
4765 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4766 		rxq->pfil = pfil;
4767 	}
4768 }
4769 
4770 static void
iflib_rem_pfil(if_ctx_t ctx)4771 iflib_rem_pfil(if_ctx_t ctx)
4772 {
4773 	struct pfil_head *pfil;
4774 	iflib_rxq_t rxq;
4775 	int i;
4776 
4777 	rxq = ctx->ifc_rxqs;
4778 	pfil = rxq->pfil;
4779 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4780 		rxq->pfil = NULL;
4781 	}
4782 	pfil_head_unregister(pfil);
4783 }
4784 
4785 
4786 /*
4787  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4788  * cpuid and wrapping as necessary.
4789  */
4790 static unsigned int
cpuid_advance(if_ctx_t ctx,unsigned int cpuid,unsigned int n)4791 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4792 {
4793 	unsigned int first_valid;
4794 	unsigned int last_valid;
4795 
4796 	/* cpuid should always be in the valid set */
4797 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4798 
4799 	/* valid set should never be empty */
4800 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4801 
4802 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4803 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4804 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4805 	while (n > 0) {
4806 		do {
4807 			cpuid++;
4808 			if (cpuid > last_valid)
4809 				cpuid = first_valid;
4810 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4811 		n--;
4812 	}
4813 
4814 	return (cpuid);
4815 }
4816 
4817 /*
4818  * CPU mapping behaviors
4819  * ---------------------
4820  * 'separate txrx' refers to the separate_txrx sysctl
4821  * 'use logical' refers to the use_logical_cores sysctl
4822  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4823  *
4824  *  separate     use     INTR
4825  *    txrx     logical   CPUS   result
4826  * ---------- --------- ------ ------------------------------------------------
4827  *     -          -       X     RX and TX queues mapped to consecutive physical
4828  *                              cores with RX/TX pairs on same core and excess
4829  *                              of either following
4830  *     -          X       X     RX and TX queues mapped to consecutive cores
4831  *                              of any type with RX/TX pairs on same core and
4832  *                              excess of either following
4833  *     X          -       X     RX and TX queues mapped to consecutive physical
4834  *                              cores; all RX then all TX
4835  *     X          X       X     RX queues mapped to consecutive physical cores
4836  *                              first, then TX queues mapped to L2 neighbor of
4837  *                              the corresponding RX queue if one exists,
4838  *                              otherwise to consecutive physical cores
4839  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4840  *                              any type with RX/TX pairs on same core and excess
4841  *                              of either following
4842  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4843  *                              any type; all RX then all TX
4844  */
4845 static unsigned int
get_cpuid_for_queue(if_ctx_t ctx,unsigned int base_cpuid,unsigned int qid,bool is_tx)4846 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4847     bool is_tx)
4848 {
4849 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4850 	unsigned int core_index;
4851 
4852 	if (ctx->ifc_sysctl_separate_txrx) {
4853 		/*
4854 		 * When using separate CPUs for TX and RX, the assignment
4855 		 * will always be of a consecutive CPU out of the set of
4856 		 * context CPUs, except for the specific case where the
4857 		 * context CPUs are phsyical cores, the use of logical cores
4858 		 * has been enabled, the assignment is for TX, the TX qid
4859 		 * corresponds to an RX qid, and the CPU assigned to the
4860 		 * corresponding RX queue has an L2 neighbor.
4861 		 */
4862 		if (ctx->ifc_sysctl_use_logical_cores &&
4863 		    ctx->ifc_cpus_are_physical_cores &&
4864 		    is_tx && qid < scctx->isc_nrxqsets) {
4865 			int l2_neighbor;
4866 			unsigned int rx_cpuid;
4867 
4868 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4869 			l2_neighbor = sched_find_l2_neighbor(rx_cpuid);
4870 			if (l2_neighbor != -1) {
4871 				return (l2_neighbor);
4872 			}
4873 			/*
4874 			 * ... else fall through to the normal
4875 			 * consecutive-after-RX assignment scheme.
4876 			 *
4877 			 * Note that we are assuming that all RX queue CPUs
4878 			 * have an L2 neighbor, or all do not.  If a mixed
4879 			 * scenario is possible, we will have to keep track
4880 			 * separately of how many queues prior to this one
4881 			 * were not able to be assigned to an L2 neighbor.
4882 			 */
4883 		}
4884 		if (is_tx)
4885 			core_index = scctx->isc_nrxqsets + qid;
4886 		else
4887 			core_index = qid;
4888 	} else {
4889 		core_index = qid;
4890 	}
4891 
4892 	return (cpuid_advance(ctx, base_cpuid, core_index));
4893 }
4894 
4895 static uint16_t
get_ctx_core_offset(if_ctx_t ctx)4896 get_ctx_core_offset(if_ctx_t ctx)
4897 {
4898 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4899 	struct cpu_offset *op;
4900 	cpuset_t assigned_cpus;
4901 	unsigned int cores_consumed;
4902 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4903 	unsigned int first_valid;
4904 	unsigned int last_valid;
4905 	unsigned int i;
4906 
4907 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4908 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4909 
4910 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4911 		/*
4912 		 * Align the user-chosen base CPU ID to the next valid CPU
4913 		 * for this device.  If the chosen base CPU ID is smaller
4914 		 * than the first valid CPU or larger than the last valid
4915 		 * CPU, we assume the user does not know what the valid
4916 		 * range is for this device and is thinking in terms of a
4917 		 * zero-based reference frame, and so we shift the given
4918 		 * value into the valid range (and wrap accordingly) so the
4919 		 * intent is translated to the proper frame of reference.
4920 		 * If the base CPU ID is within the valid first/last, but
4921 		 * does not correspond to a valid CPU, it is advanced to the
4922 		 * next valid CPU (wrapping if necessary).
4923 		 */
4924 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
4925 			/* shift from zero-based to first_valid-based */
4926 			base_cpuid += first_valid;
4927 			/* wrap to range [first_valid, last_valid] */
4928 			base_cpuid = (base_cpuid - first_valid) %
4929 			    (last_valid - first_valid + 1);
4930 		}
4931 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
4932 			/*
4933 			 * base_cpuid is in [first_valid, last_valid], but
4934 			 * not a member of the valid set.  In this case,
4935 			 * there will always be a member of the valid set
4936 			 * with a CPU ID that is greater than base_cpuid,
4937 			 * and we simply advance to it.
4938 			 */
4939 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
4940 				base_cpuid++;
4941 		}
4942 		return (base_cpuid);
4943 	}
4944 
4945 	/*
4946 	 * Determine how many cores will be consumed by performing the CPU
4947 	 * assignments and counting how many of the assigned CPUs correspond
4948 	 * to CPUs in the set of context CPUs.  This is done using the CPU
4949 	 * ID first_valid as the base CPU ID, as the base CPU must be within
4950 	 * the set of context CPUs.
4951 	 *
4952 	 * Note not all assigned CPUs will be in the set of context CPUs
4953 	 * when separate CPUs are being allocated to TX and RX queues,
4954 	 * assignment to logical cores has been enabled, the set of context
4955 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
4956 	 * neighbors of CPUs that RX queues have been mapped to - in this
4957 	 * case we do only want to count how many CPUs in the set of context
4958 	 * CPUs have been consumed, as that determines the next CPU in that
4959 	 * set to start allocating at for the next device for which
4960 	 * core_offset is not set.
4961 	 */
4962 	CPU_ZERO(&assigned_cpus);
4963 	for (i = 0; i < scctx->isc_ntxqsets; i++)
4964 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
4965 		    &assigned_cpus);
4966 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4967 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
4968 		    &assigned_cpus);
4969 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
4970 	cores_consumed = CPU_COUNT(&assigned_cpus);
4971 
4972 	mtx_lock(&cpu_offset_mtx);
4973 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4974 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4975 			base_cpuid = op->next_cpuid;
4976 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
4977 			    cores_consumed);
4978 			MPASS(op->refcount < UINT_MAX);
4979 			op->refcount++;
4980 			break;
4981 		}
4982 	}
4983 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
4984 		base_cpuid = first_valid;
4985 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4986 		    M_NOWAIT | M_ZERO);
4987 		if (op == NULL) {
4988 			device_printf(ctx->ifc_dev,
4989 			    "allocation for cpu offset failed.\n");
4990 		} else {
4991 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
4992 			    cores_consumed);
4993 			op->refcount = 1;
4994 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4995 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4996 		}
4997 	}
4998 	mtx_unlock(&cpu_offset_mtx);
4999 
5000 	return (base_cpuid);
5001 }
5002 
5003 static void
unref_ctx_core_offset(if_ctx_t ctx)5004 unref_ctx_core_offset(if_ctx_t ctx)
5005 {
5006 	struct cpu_offset *op, *top;
5007 
5008 	mtx_lock(&cpu_offset_mtx);
5009 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5010 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5011 			MPASS(op->refcount > 0);
5012 			op->refcount--;
5013 			if (op->refcount == 0) {
5014 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5015 				free(op, M_IFLIB);
5016 			}
5017 			break;
5018 		}
5019 	}
5020 	mtx_unlock(&cpu_offset_mtx);
5021 }
5022 
5023 int
iflib_device_register(device_t dev,void * sc,if_shared_ctx_t sctx,if_ctx_t * ctxp)5024 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5025 {
5026 	if_ctx_t ctx;
5027 	if_t ifp;
5028 	if_softc_ctx_t scctx;
5029 	kobjop_desc_t kobj_desc;
5030 	kobj_method_t *kobj_method;
5031 	int err, msix, rid;
5032 	int num_txd, num_rxd;
5033 	char namebuf[TASKQUEUE_NAMELEN];
5034 
5035 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO);
5036 
5037 	if (sc == NULL) {
5038 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO);
5039 		device_set_softc(dev, ctx);
5040 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5041 	}
5042 
5043 	ctx->ifc_sctx = sctx;
5044 	ctx->ifc_dev = dev;
5045 	ctx->ifc_softc = sc;
5046 
5047 	iflib_register(ctx);
5048 	iflib_add_device_sysctl_pre(ctx);
5049 
5050 	scctx = &ctx->ifc_softc_ctx;
5051 	ifp = ctx->ifc_ifp;
5052 	if (ctx->ifc_sysctl_simple_tx) {
5053 #ifndef ALTQ
5054 		if_settransmitfn(ifp, iflib_simple_transmit);
5055 		device_printf(dev, "using simple if_transmit\n");
5056 #else
5057 		device_printf(dev, "ALTQ prevents using simple if_transmit\n");
5058 #endif
5059 	}
5060 	iflib_reset_qvalues(ctx);
5061 	IFNET_WLOCK();
5062 	CTX_LOCK(ctx);
5063 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5064 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5065 		goto fail_unlock;
5066 	}
5067 	_iflib_pre_assert(scctx);
5068 	ctx->ifc_txrx = *scctx->isc_txrx;
5069 
5070 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5071 
5072 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5073 		ctx->ifc_mediap = scctx->isc_media;
5074 
5075 #ifdef INVARIANTS
5076 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5077 		MPASS(scctx->isc_tx_csum_flags);
5078 #endif
5079 
5080 	if_setcapabilities(ifp,
5081 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5082 	if_setcapenable(ifp,
5083 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5084 
5085 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5086 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5087 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5088 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5089 
5090 	num_txd = iflib_num_tx_descs(ctx);
5091 	num_rxd = iflib_num_rx_descs(ctx);
5092 
5093 	/* XXX change for per-queue sizes */
5094 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5095 	    num_txd, num_rxd);
5096 
5097 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5098 		scctx->isc_tx_nsegments = max(1, num_txd /
5099 		    MAX_SINGLE_PACKET_FRACTION);
5100 	if (scctx->isc_tx_tso_segments_max > num_txd /
5101 	    MAX_SINGLE_PACKET_FRACTION)
5102 		scctx->isc_tx_tso_segments_max = max(1,
5103 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5104 
5105 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5106 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5107 		/*
5108 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5109 		 * but some MACs do.
5110 		 */
5111 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5112 		    IP_MAXPACKET));
5113 		/*
5114 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5115 		 * into account.  In the worst case, each of these calls will
5116 		 * add another mbuf and, thus, the requirement for another DMA
5117 		 * segment.  So for best performance, it doesn't make sense to
5118 		 * advertize a maximum of TSO segments that typically will
5119 		 * require defragmentation in iflib_encap().
5120 		 */
5121 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5122 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5123 	}
5124 	if (scctx->isc_rss_table_size == 0)
5125 		scctx->isc_rss_table_size = 64;
5126 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1;
5127 
5128 	/* Create and start admin taskqueue */
5129 	snprintf(namebuf, TASKQUEUE_NAMELEN, "if_%s_tq", device_get_nameunit(dev));
5130 	ctx->ifc_tq = taskqueue_create_fast(namebuf, M_NOWAIT,
5131 	    taskqueue_thread_enqueue, &ctx->ifc_tq);
5132 	if (ctx->ifc_tq == NULL) {
5133 		device_printf(dev, "Unable to create admin taskqueue\n");
5134 		return (ENOMEM);
5135 	}
5136 
5137 	err = taskqueue_start_threads(&ctx->ifc_tq, 1, PI_NET, "%s", namebuf);
5138 	if (err) {
5139 		device_printf(dev,
5140 		    "Unable to start admin taskqueue threads error: %d\n",
5141 		    err);
5142 		taskqueue_free(ctx->ifc_tq);
5143 		return (err);
5144 	}
5145 
5146 	TASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5147 
5148 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5149 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5150 		device_printf(dev, "Unable to fetch CPU list\n");
5151 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5152 		ctx->ifc_cpus_are_physical_cores = false;
5153 	} else
5154 		ctx->ifc_cpus_are_physical_cores = true;
5155 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5156 
5157 	/*
5158 	 * Now set up MSI or MSI-X, should return us the number of supported
5159 	 * vectors (will be 1 for a legacy interrupt and MSI).
5160 	 */
5161 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5162 		msix = scctx->isc_vectors;
5163 	} else if (scctx->isc_msix_bar != 0)
5164 		/*
5165 		 * The simple fact that isc_msix_bar is not 0 does not mean we
5166 		 * we have a good value there that is known to work.
5167 		 */
5168 		msix = iflib_msix_init(ctx);
5169 	else {
5170 		scctx->isc_vectors = 1;
5171 		scctx->isc_ntxqsets = 1;
5172 		scctx->isc_nrxqsets = 1;
5173 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5174 		msix = 0;
5175 	}
5176 	/* Get memory for the station queues */
5177 	if ((err = iflib_queues_alloc(ctx))) {
5178 		device_printf(dev, "Unable to allocate queue memory\n");
5179 		goto fail_intr_free;
5180 	}
5181 
5182 	if ((err = iflib_qset_structures_setup(ctx)))
5183 		goto fail_queues;
5184 
5185 	/*
5186 	 * Now that we know how many queues there are, get the core offset.
5187 	 */
5188 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5189 
5190 	if (msix > 1) {
5191 		/*
5192 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5193 		 * aren't the default NULL implementation.
5194 		 */
5195 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5196 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5197 		    kobj_desc);
5198 		if (kobj_method == &kobj_desc->deflt) {
5199 			device_printf(dev,
5200 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5201 			err = EOPNOTSUPP;
5202 			goto fail_queues;
5203 		}
5204 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5205 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5206 		    kobj_desc);
5207 		if (kobj_method == &kobj_desc->deflt) {
5208 			device_printf(dev,
5209 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5210 			err = EOPNOTSUPP;
5211 			goto fail_queues;
5212 		}
5213 
5214 		/*
5215 		 * Assign the MSI-X vectors.
5216 		 * Note that the default NULL ifdi_msix_intr_assign method will
5217 		 * fail here, too.
5218 		 */
5219 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5220 		if (err != 0) {
5221 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5222 			    err);
5223 			goto fail_queues;
5224 		}
5225 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5226 		rid = 0;
5227 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5228 			MPASS(msix == 1);
5229 			rid = 1;
5230 		}
5231 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5232 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5233 			goto fail_queues;
5234 		}
5235 	} else {
5236 		device_printf(dev,
5237 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5238 		err = ENODEV;
5239 		goto fail_queues;
5240 	}
5241 
5242 	/*
5243 	 * It prevents a double-locking panic with iflib_media_status when
5244 	 * the driver loads.
5245 	 */
5246 	CTX_UNLOCK(ctx);
5247 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5248 	CTX_LOCK(ctx);
5249 
5250 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5251 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5252 		goto fail_detach;
5253 	}
5254 
5255 	/*
5256 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5257 	 * This must appear after the call to ether_ifattach() because
5258 	 * ether_ifattach() sets if_hdrlen to the default value.
5259 	 */
5260 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5261 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5262 
5263 	if ((err = iflib_netmap_attach(ctx))) {
5264 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5265 		goto fail_detach;
5266 	}
5267 	*ctxp = ctx;
5268 
5269 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5270 
5271 	iflib_add_device_sysctl_post(ctx);
5272 	iflib_add_pfil(ctx);
5273 	ctx->ifc_flags |= IFC_INIT_DONE;
5274 	CTX_UNLOCK(ctx);
5275 	IFNET_WUNLOCK();
5276 
5277 	return (0);
5278 
5279 fail_detach:
5280 	ether_ifdetach(ctx->ifc_ifp);
5281 fail_queues:
5282 	taskqueue_free(ctx->ifc_tq);
5283 	iflib_tqg_detach(ctx);
5284 	iflib_tx_structures_free(ctx);
5285 	iflib_rx_structures_free(ctx);
5286 	IFDI_DETACH(ctx);
5287 	IFDI_QUEUES_FREE(ctx);
5288 fail_intr_free:
5289 	iflib_free_intr_mem(ctx);
5290 fail_unlock:
5291 	CTX_UNLOCK(ctx);
5292 	IFNET_WUNLOCK();
5293 	iflib_deregister(ctx);
5294 	device_set_softc(ctx->ifc_dev, NULL);
5295 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5296 		free(ctx->ifc_softc, M_IFLIB);
5297 	free(ctx, M_IFLIB);
5298 	return (err);
5299 }
5300 
5301 int
iflib_device_attach(device_t dev)5302 iflib_device_attach(device_t dev)
5303 {
5304 	if_ctx_t ctx;
5305 	if_shared_ctx_t sctx;
5306 
5307 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5308 		return (ENOTSUP);
5309 
5310 	pci_enable_busmaster(dev);
5311 
5312 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5313 }
5314 
5315 int
iflib_device_deregister(if_ctx_t ctx)5316 iflib_device_deregister(if_ctx_t ctx)
5317 {
5318 	if_t ifp = ctx->ifc_ifp;
5319 	device_t dev = ctx->ifc_dev;
5320 
5321 	/* Make sure VLANS are not using driver */
5322 	if (if_vlantrunkinuse(ifp)) {
5323 		device_printf(dev, "Vlan in use, detach first\n");
5324 		return (EBUSY);
5325 	}
5326 #ifdef PCI_IOV
5327 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5328 		device_printf(dev, "SR-IOV in use; detach first.\n");
5329 		return (EBUSY);
5330 	}
5331 #endif
5332 
5333 	STATE_LOCK(ctx);
5334 	ctx->ifc_flags |= IFC_IN_DETACH;
5335 	STATE_UNLOCK(ctx);
5336 
5337 	/* Unregister VLAN handlers before calling iflib_stop() */
5338 	iflib_unregister_vlan_handlers(ctx);
5339 
5340 	iflib_netmap_detach(ifp);
5341 	ether_ifdetach(ifp);
5342 
5343 	CTX_LOCK(ctx);
5344 	iflib_stop(ctx);
5345 	CTX_UNLOCK(ctx);
5346 
5347 	iflib_rem_pfil(ctx);
5348 	if (ctx->ifc_led_dev != NULL)
5349 		led_destroy(ctx->ifc_led_dev);
5350 
5351 	iflib_tqg_detach(ctx);
5352 	iflib_tx_structures_free(ctx);
5353 	iflib_rx_structures_free(ctx);
5354 
5355 	CTX_LOCK(ctx);
5356 	IFDI_DETACH(ctx);
5357 	IFDI_QUEUES_FREE(ctx);
5358 	CTX_UNLOCK(ctx);
5359 
5360 	taskqueue_free(ctx->ifc_tq);
5361 	ctx->ifc_tq = NULL;
5362 
5363 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5364 	iflib_free_intr_mem(ctx);
5365 
5366 	bus_generic_detach(dev);
5367 
5368 	iflib_deregister(ctx);
5369 
5370 	device_set_softc(ctx->ifc_dev, NULL);
5371 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5372 		free(ctx->ifc_softc, M_IFLIB);
5373 	unref_ctx_core_offset(ctx);
5374 	free(ctx, M_IFLIB);
5375 	return (0);
5376 }
5377 
5378 static void
iflib_tqg_detach(if_ctx_t ctx)5379 iflib_tqg_detach(if_ctx_t ctx)
5380 {
5381 	iflib_txq_t txq;
5382 	iflib_rxq_t rxq;
5383 	int i;
5384 	struct taskqgroup *tqg;
5385 
5386 	/* XXX drain any dependent tasks */
5387 	tqg = qgroup_if_io_tqg;
5388 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5389 		callout_drain(&txq->ift_timer);
5390 #ifdef DEV_NETMAP
5391 		callout_drain(&txq->ift_netmap_timer);
5392 #endif /* DEV_NETMAP */
5393 		if (txq->ift_task.gt_uniq != NULL)
5394 			taskqgroup_detach(tqg, &txq->ift_task);
5395 	}
5396 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5397 		if (rxq->ifr_task.gt_uniq != NULL)
5398 			taskqgroup_detach(tqg, &rxq->ifr_task);
5399 	}
5400 }
5401 
5402 static void
iflib_free_intr_mem(if_ctx_t ctx)5403 iflib_free_intr_mem(if_ctx_t ctx)
5404 {
5405 
5406 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5407 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5408 	}
5409 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5410 		pci_release_msi(ctx->ifc_dev);
5411 	}
5412 	if (ctx->ifc_msix_mem != NULL) {
5413 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5414 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5415 		ctx->ifc_msix_mem = NULL;
5416 	}
5417 }
5418 
5419 int
iflib_device_detach(device_t dev)5420 iflib_device_detach(device_t dev)
5421 {
5422 	if_ctx_t ctx = device_get_softc(dev);
5423 
5424 	return (iflib_device_deregister(ctx));
5425 }
5426 
5427 int
iflib_device_suspend(device_t dev)5428 iflib_device_suspend(device_t dev)
5429 {
5430 	if_ctx_t ctx = device_get_softc(dev);
5431 
5432 	CTX_LOCK(ctx);
5433 	IFDI_SUSPEND(ctx);
5434 	CTX_UNLOCK(ctx);
5435 
5436 	return (bus_generic_suspend(dev));
5437 }
5438 int
iflib_device_shutdown(device_t dev)5439 iflib_device_shutdown(device_t dev)
5440 {
5441 	if_ctx_t ctx = device_get_softc(dev);
5442 
5443 	CTX_LOCK(ctx);
5444 	IFDI_SHUTDOWN(ctx);
5445 	CTX_UNLOCK(ctx);
5446 
5447 	return (bus_generic_suspend(dev));
5448 }
5449 
5450 int
iflib_device_resume(device_t dev)5451 iflib_device_resume(device_t dev)
5452 {
5453 	if_ctx_t ctx = device_get_softc(dev);
5454 	iflib_txq_t txq = ctx->ifc_txqs;
5455 
5456 	CTX_LOCK(ctx);
5457 	IFDI_RESUME(ctx);
5458 	iflib_if_init_locked(ctx);
5459 	CTX_UNLOCK(ctx);
5460 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5461 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5462 
5463 	return (bus_generic_resume(dev));
5464 }
5465 
5466 int
iflib_device_iov_init(device_t dev,uint16_t num_vfs,const nvlist_t * params)5467 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5468 {
5469 	int error;
5470 	if_ctx_t ctx = device_get_softc(dev);
5471 
5472 	CTX_LOCK(ctx);
5473 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5474 	CTX_UNLOCK(ctx);
5475 
5476 	return (error);
5477 }
5478 
5479 void
iflib_device_iov_uninit(device_t dev)5480 iflib_device_iov_uninit(device_t dev)
5481 {
5482 	if_ctx_t ctx = device_get_softc(dev);
5483 
5484 	CTX_LOCK(ctx);
5485 	IFDI_IOV_UNINIT(ctx);
5486 	CTX_UNLOCK(ctx);
5487 }
5488 
5489 int
iflib_device_iov_add_vf(device_t dev,uint16_t vfnum,const nvlist_t * params)5490 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5491 {
5492 	int error;
5493 	if_ctx_t ctx = device_get_softc(dev);
5494 
5495 	CTX_LOCK(ctx);
5496 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5497 	CTX_UNLOCK(ctx);
5498 
5499 	return (error);
5500 }
5501 
5502 /*********************************************************************
5503  *
5504  *  MODULE FUNCTION DEFINITIONS
5505  *
5506  **********************************************************************/
5507 
5508 /*
5509  * - Start a fast taskqueue thread for each core
5510  * - Start a taskqueue for control operations
5511  */
5512 static int
iflib_module_init(void)5513 iflib_module_init(void)
5514 {
5515 	iflib_timer_default = hz / 2;
5516 	return (0);
5517 }
5518 
5519 static int
iflib_module_event_handler(module_t mod,int what,void * arg)5520 iflib_module_event_handler(module_t mod, int what, void *arg)
5521 {
5522 	int err;
5523 
5524 	switch (what) {
5525 	case MOD_LOAD:
5526 		if ((err = iflib_module_init()) != 0)
5527 			return (err);
5528 		break;
5529 	case MOD_UNLOAD:
5530 		return (EBUSY);
5531 	default:
5532 		return (EOPNOTSUPP);
5533 	}
5534 
5535 	return (0);
5536 }
5537 
5538 /*********************************************************************
5539  *
5540  *  PUBLIC FUNCTION DEFINITIONS
5541  *     ordered as in iflib.h
5542  *
5543  **********************************************************************/
5544 
5545 static void
_iflib_assert(if_shared_ctx_t sctx)5546 _iflib_assert(if_shared_ctx_t sctx)
5547 {
5548 	int i;
5549 
5550 	MPASS(sctx->isc_tx_maxsize);
5551 	MPASS(sctx->isc_tx_maxsegsize);
5552 
5553 	MPASS(sctx->isc_rx_maxsize);
5554 	MPASS(sctx->isc_rx_nsegments);
5555 	MPASS(sctx->isc_rx_maxsegsize);
5556 
5557 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5558 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5559 		MPASS(sctx->isc_nrxd_min[i]);
5560 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5561 		MPASS(sctx->isc_nrxd_max[i]);
5562 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5563 		MPASS(sctx->isc_nrxd_default[i]);
5564 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5565 	}
5566 
5567 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5568 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5569 		MPASS(sctx->isc_ntxd_min[i]);
5570 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5571 		MPASS(sctx->isc_ntxd_max[i]);
5572 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5573 		MPASS(sctx->isc_ntxd_default[i]);
5574 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5575 	}
5576 }
5577 
5578 static void
_iflib_pre_assert(if_softc_ctx_t scctx)5579 _iflib_pre_assert(if_softc_ctx_t scctx)
5580 {
5581 
5582 	MPASS(scctx->isc_txrx->ift_txd_encap);
5583 	MPASS(scctx->isc_txrx->ift_txd_flush);
5584 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5585 	MPASS(scctx->isc_txrx->ift_rxd_available);
5586 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5587 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5588 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5589 }
5590 
5591 static void
iflib_register(if_ctx_t ctx)5592 iflib_register(if_ctx_t ctx)
5593 {
5594 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5595 	driver_t *driver = sctx->isc_driver;
5596 	device_t dev = ctx->ifc_dev;
5597 	if_t ifp;
5598 
5599 	_iflib_assert(sctx);
5600 
5601 	CTX_LOCK_INIT(ctx);
5602 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5603 	ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev);
5604 
5605 	/*
5606 	 * Initialize our context's device specific methods
5607 	 */
5608 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5609 	kobj_class_compile((kobj_class_t) driver);
5610 
5611 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5612 	if_setsoftc(ifp, ctx);
5613 	if_setdev(ifp, dev);
5614 	if_setinitfn(ifp, iflib_if_init);
5615 	if_setioctlfn(ifp, iflib_if_ioctl);
5616 #ifdef ALTQ
5617 	if_setstartfn(ifp, iflib_altq_if_start);
5618 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5619 	if_setsendqready(ifp);
5620 #else
5621 	if_settransmitfn(ifp, iflib_if_transmit);
5622 #endif
5623 	if_setqflushfn(ifp, iflib_if_qflush);
5624 	if_setgetcounterfn(ifp, iflib_if_get_counter);
5625 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5626 	ctx->ifc_vlan_attach_event =
5627 	    EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5628 		    EVENTHANDLER_PRI_FIRST);
5629 	ctx->ifc_vlan_detach_event =
5630 	    EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5631 		    EVENTHANDLER_PRI_FIRST);
5632 
5633 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5634 		ctx->ifc_mediap = &ctx->ifc_media;
5635 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5636 		    iflib_media_change, iflib_media_status);
5637 	}
5638 }
5639 
5640 static void
iflib_unregister_vlan_handlers(if_ctx_t ctx)5641 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5642 {
5643 	/* Unregister VLAN events */
5644 	if (ctx->ifc_vlan_attach_event != NULL) {
5645 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5646 		ctx->ifc_vlan_attach_event = NULL;
5647 	}
5648 	if (ctx->ifc_vlan_detach_event != NULL) {
5649 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5650 		ctx->ifc_vlan_detach_event = NULL;
5651 	}
5652 
5653 }
5654 
5655 static void
iflib_deregister(if_ctx_t ctx)5656 iflib_deregister(if_ctx_t ctx)
5657 {
5658 	if_t ifp = ctx->ifc_ifp;
5659 
5660 	/* Remove all media */
5661 	ifmedia_removeall(&ctx->ifc_media);
5662 
5663 	/* Ensure that VLAN event handlers are unregistered */
5664 	iflib_unregister_vlan_handlers(ctx);
5665 
5666 	/* Release kobject reference */
5667 	kobj_delete((kobj_t) ctx, NULL);
5668 
5669 	/* Free the ifnet structure */
5670 	if_free(ifp);
5671 
5672 	STATE_LOCK_DESTROY(ctx);
5673 
5674 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5675 	CTX_LOCK_DESTROY(ctx);
5676 }
5677 
5678 static int
iflib_queues_alloc(if_ctx_t ctx)5679 iflib_queues_alloc(if_ctx_t ctx)
5680 {
5681 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5682 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5683 	device_t dev = ctx->ifc_dev;
5684 	int nrxqsets = scctx->isc_nrxqsets;
5685 	int ntxqsets = scctx->isc_ntxqsets;
5686 	iflib_txq_t txq;
5687 	iflib_rxq_t rxq;
5688 	iflib_fl_t fl = NULL;
5689 	int i, j, cpu, err, txconf, rxconf;
5690 	iflib_dma_info_t ifdip;
5691 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5692 	uint32_t *txqsizes = scctx->isc_txqsizes;
5693 	uint8_t nrxqs = sctx->isc_nrxqs;
5694 	uint8_t ntxqs = sctx->isc_ntxqs;
5695 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5696 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5697 	caddr_t *vaddrs;
5698 	uint64_t *paddrs;
5699 
5700 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5701 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5702 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5703 	    ("there must be at least a rxq for each free list"));
5704 
5705 	/* Allocate the TX ring struct memory */
5706 	if (!(ctx->ifc_txqs =
5707 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5708 		    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5709 		device_printf(dev, "Unable to allocate TX ring memory\n");
5710 		err = ENOMEM;
5711 		goto fail;
5712 	}
5713 
5714 	/* Now allocate the RX */
5715 	if (!(ctx->ifc_rxqs =
5716 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5717 		    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5718 		device_printf(dev, "Unable to allocate RX ring memory\n");
5719 		err = ENOMEM;
5720 		goto rx_fail;
5721 	}
5722 
5723 	txq = ctx->ifc_txqs;
5724 	rxq = ctx->ifc_rxqs;
5725 
5726 	/*
5727 	 * XXX handle allocation failure
5728 	 */
5729 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5730 		/* Set up some basics */
5731 
5732 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5733 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5734 			device_printf(dev,
5735 			    "Unable to allocate TX DMA info memory\n");
5736 			err = ENOMEM;
5737 			goto err_tx_desc;
5738 		}
5739 		txq->ift_ifdi = ifdip;
5740 		for (j = 0; j < ntxqs; j++, ifdip++) {
5741 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5742 				device_printf(dev,
5743 				    "Unable to allocate TX descriptors\n");
5744 				err = ENOMEM;
5745 				goto err_tx_desc;
5746 			}
5747 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5748 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5749 		}
5750 		txq->ift_ctx = ctx;
5751 		txq->ift_id = i;
5752 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5753 			txq->ift_br_offset = 1;
5754 		} else {
5755 			txq->ift_br_offset = 0;
5756 		}
5757 
5758 		if (iflib_txsd_alloc(txq)) {
5759 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5760 			err = ENOMEM;
5761 			goto err_tx_desc;
5762 		}
5763 
5764 		/* Initialize the TX lock */
5765 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5766 		    device_get_nameunit(dev), txq->ift_id);
5767 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5768 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5769 		txq->ift_timer.c_cpu = cpu;
5770 #ifdef DEV_NETMAP
5771 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5772 		txq->ift_netmap_timer.c_cpu = cpu;
5773 #endif /* DEV_NETMAP */
5774 
5775 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5776 		    iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5777 		if (err) {
5778 			/* XXX free any allocated rings */
5779 			device_printf(dev, "Unable to allocate buf_ring\n");
5780 			goto err_tx_desc;
5781 		}
5782 		txq->ift_reclaim_thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
5783 	}
5784 
5785 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5786 		/* Set up some basics */
5787 		callout_init(&rxq->ifr_watchdog, 1);
5788 
5789 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5790 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5791 			device_printf(dev,
5792 			    "Unable to allocate RX DMA info memory\n");
5793 			err = ENOMEM;
5794 			goto err_tx_desc;
5795 		}
5796 
5797 		rxq->ifr_ifdi = ifdip;
5798 		/* XXX this needs to be changed if #rx queues != #tx queues */
5799 		rxq->ifr_ntxqirq = 1;
5800 		rxq->ifr_txqid[0] = i;
5801 		for (j = 0; j < nrxqs; j++, ifdip++) {
5802 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5803 				device_printf(dev,
5804 				    "Unable to allocate RX descriptors\n");
5805 				err = ENOMEM;
5806 				goto err_tx_desc;
5807 			}
5808 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5809 		}
5810 		rxq->ifr_ctx = ctx;
5811 		rxq->ifr_id = i;
5812 		rxq->ifr_fl_offset = fl_offset;
5813 		rxq->ifr_nfl = nfree_lists;
5814 		if (!(fl =
5815 		    (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5816 			device_printf(dev, "Unable to allocate free list memory\n");
5817 			err = ENOMEM;
5818 			goto err_tx_desc;
5819 		}
5820 		rxq->ifr_fl = fl;
5821 		for (j = 0; j < nfree_lists; j++) {
5822 			fl[j].ifl_rxq = rxq;
5823 			fl[j].ifl_id = j;
5824 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5825 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5826 		}
5827 		/* Allocate receive buffers for the ring */
5828 		if (iflib_rxsd_alloc(rxq)) {
5829 			device_printf(dev,
5830 			    "Critical Failure setting up receive buffers\n");
5831 			err = ENOMEM;
5832 			goto err_rx_desc;
5833 		}
5834 
5835 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5836 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5837 			    M_WAITOK);
5838 	}
5839 
5840 	/* TXQs */
5841 	vaddrs = malloc(sizeof(caddr_t)  * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
5842 	paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
5843 	for (i = 0; i < ntxqsets; i++) {
5844 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5845 
5846 		for (j = 0; j < ntxqs; j++, di++) {
5847 			vaddrs[i * ntxqs + j] = di->idi_vaddr;
5848 			paddrs[i * ntxqs + j] = di->idi_paddr;
5849 		}
5850 	}
5851 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5852 		device_printf(ctx->ifc_dev,
5853 		    "Unable to allocate device TX queue\n");
5854 		iflib_tx_structures_free(ctx);
5855 		free(vaddrs, M_IFLIB);
5856 		free(paddrs, M_IFLIB);
5857 		goto err_rx_desc;
5858 	}
5859 	free(vaddrs, M_IFLIB);
5860 	free(paddrs, M_IFLIB);
5861 
5862 	/* RXQs */
5863 	vaddrs = malloc(sizeof(caddr_t)  * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
5864 	paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
5865 	for (i = 0; i < nrxqsets; i++) {
5866 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5867 
5868 		for (j = 0; j < nrxqs; j++, di++) {
5869 			vaddrs[i * nrxqs + j] = di->idi_vaddr;
5870 			paddrs[i * nrxqs + j] = di->idi_paddr;
5871 		}
5872 	}
5873 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5874 		device_printf(ctx->ifc_dev,
5875 		    "Unable to allocate device RX queue\n");
5876 		iflib_tx_structures_free(ctx);
5877 		free(vaddrs, M_IFLIB);
5878 		free(paddrs, M_IFLIB);
5879 		goto err_rx_desc;
5880 	}
5881 	free(vaddrs, M_IFLIB);
5882 	free(paddrs, M_IFLIB);
5883 
5884 	return (0);
5885 
5886 /* XXX handle allocation failure changes */
5887 err_rx_desc:
5888 err_tx_desc:
5889 rx_fail:
5890 	if (ctx->ifc_rxqs != NULL)
5891 		free(ctx->ifc_rxqs, M_IFLIB);
5892 	ctx->ifc_rxqs = NULL;
5893 	if (ctx->ifc_txqs != NULL)
5894 		free(ctx->ifc_txqs, M_IFLIB);
5895 	ctx->ifc_txqs = NULL;
5896 fail:
5897 	return (err);
5898 }
5899 
5900 static int
iflib_tx_structures_setup(if_ctx_t ctx)5901 iflib_tx_structures_setup(if_ctx_t ctx)
5902 {
5903 	iflib_txq_t txq = ctx->ifc_txqs;
5904 	int i;
5905 
5906 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5907 		iflib_txq_setup(txq);
5908 
5909 	return (0);
5910 }
5911 
5912 static void
iflib_tx_structures_free(if_ctx_t ctx)5913 iflib_tx_structures_free(if_ctx_t ctx)
5914 {
5915 	iflib_txq_t txq = ctx->ifc_txqs;
5916 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5917 	int i, j;
5918 
5919 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5920 		for (j = 0; j < sctx->isc_ntxqs; j++)
5921 			iflib_dma_free(&txq->ift_ifdi[j]);
5922 		iflib_txq_destroy(txq);
5923 	}
5924 	free(ctx->ifc_txqs, M_IFLIB);
5925 	ctx->ifc_txqs = NULL;
5926 }
5927 
5928 /*********************************************************************
5929  *
5930  *  Initialize all receive rings.
5931  *
5932  **********************************************************************/
5933 static int
iflib_rx_structures_setup(if_ctx_t ctx)5934 iflib_rx_structures_setup(if_ctx_t ctx)
5935 {
5936 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5937 	int q;
5938 #if defined(INET6) || defined(INET)
5939 	int err, i;
5940 #endif
5941 
5942 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5943 #if defined(INET6) || defined(INET)
5944 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5945 		    TCP_LRO_ENTRIES, min(1024,
5946 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5947 		if (err != 0) {
5948 			device_printf(ctx->ifc_dev,
5949 			    "LRO Initialization failed!\n");
5950 			goto fail;
5951 		}
5952 #endif
5953 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5954 	}
5955 	return (0);
5956 #if defined(INET6) || defined(INET)
5957 fail:
5958 	/*
5959 	 * Free LRO resources allocated so far, we will only handle
5960 	 * the rings that completed, the failing case will have
5961 	 * cleaned up for itself.  'q' failed, so its the terminus.
5962 	 */
5963 	rxq = ctx->ifc_rxqs;
5964 	for (i = 0; i < q; ++i, rxq++) {
5965 		tcp_lro_free(&rxq->ifr_lc);
5966 	}
5967 	return (err);
5968 #endif
5969 }
5970 
5971 /*********************************************************************
5972  *
5973  *  Free all receive rings.
5974  *
5975  **********************************************************************/
5976 static void
iflib_rx_structures_free(if_ctx_t ctx)5977 iflib_rx_structures_free(if_ctx_t ctx)
5978 {
5979 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5980 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5981 	int i, j;
5982 
5983 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5984 		for (j = 0; j < sctx->isc_nrxqs; j++)
5985 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5986 		iflib_rx_sds_free(rxq);
5987 #if defined(INET6) || defined(INET)
5988 		tcp_lro_free(&rxq->ifr_lc);
5989 #endif
5990 	}
5991 	free(ctx->ifc_rxqs, M_IFLIB);
5992 	ctx->ifc_rxqs = NULL;
5993 }
5994 
5995 static int
iflib_qset_structures_setup(if_ctx_t ctx)5996 iflib_qset_structures_setup(if_ctx_t ctx)
5997 {
5998 	int err;
5999 
6000 	/*
6001 	 * It is expected that the caller takes care of freeing queues if this
6002 	 * fails.
6003 	 */
6004 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6005 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6006 		return (err);
6007 	}
6008 
6009 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6010 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6011 
6012 	return (err);
6013 }
6014 
6015 int
iflib_irq_alloc(if_ctx_t ctx,if_irq_t irq,int rid,driver_filter_t filter,void * filter_arg,driver_intr_t handler,void * arg,const char * name)6016 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6017 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6018 {
6019 
6020 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6021 }
6022 
6023 /* Just to avoid copy/paste */
6024 static inline int
iflib_irq_set_affinity(if_ctx_t ctx,if_irq_t irq,iflib_intr_type_t type,int qid,struct grouptask * gtask,struct taskqgroup * tqg,void * uniq,const char * name)6025 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6026     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6027     const char *name)
6028 {
6029 	device_t dev;
6030 	unsigned int base_cpuid, cpuid;
6031 	int err;
6032 
6033 	dev = ctx->ifc_dev;
6034 	base_cpuid = ctx->ifc_sysctl_core_offset;
6035 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6036 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6037 	    irq ? irq->ii_res : NULL, name);
6038 	if (err) {
6039 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6040 		return (err);
6041 	}
6042 #ifdef notyet
6043 	if (cpuid > ctx->ifc_cpuid_highest)
6044 		ctx->ifc_cpuid_highest = cpuid;
6045 #endif
6046 	return (0);
6047 }
6048 
6049 /*
6050  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6051  * resources.
6052  *
6053  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6054  * only.
6055  *
6056  * XXX: Could be removed if subctx's dev has its intr resource allocation
6057  * methods replaced with custom ones?
6058  */
6059 int
iflib_irq_alloc_generic_subctx(if_ctx_t ctx,if_ctx_t subctx,if_irq_t irq,int rid,iflib_intr_type_t type,driver_filter_t * filter,void * filter_arg,int qid,const char * name)6060 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6061 			       int rid, iflib_intr_type_t type,
6062 			       driver_filter_t *filter, void *filter_arg,
6063 			       int qid, const char *name)
6064 {
6065 	device_t dev, subdev;
6066 	struct grouptask *gtask;
6067 	struct taskqgroup *tqg;
6068 	iflib_filter_info_t info;
6069 	gtask_fn_t *fn;
6070 	int tqrid, err;
6071 	driver_filter_t *intr_fast;
6072 	void *q;
6073 
6074 	MPASS(ctx != NULL);
6075 	MPASS(subctx != NULL);
6076 
6077 	tqrid = rid;
6078 	dev = ctx->ifc_dev;
6079 	subdev = subctx->ifc_dev;
6080 
6081 	switch (type) {
6082 	case IFLIB_INTR_RXTX:
6083 		q = &subctx->ifc_rxqs[qid];
6084 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6085 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6086 		tqg = qgroup_if_io_tqg;
6087 		fn = _task_fn_rx;
6088 		intr_fast = iflib_fast_intr_rxtx;
6089 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6090 		break;
6091 	default:
6092 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6093 		    __func__, device_get_nameunit(subdev), type);
6094 		return (EINVAL);
6095 	}
6096 
6097 	info->ifi_filter = filter;
6098 	info->ifi_filter_arg = filter_arg;
6099 	info->ifi_task = gtask;
6100 	info->ifi_ctx = q;
6101 
6102 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6103 
6104 	/* Allocate interrupts from hardware using parent context */
6105 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6106 	if (err != 0) {
6107 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6108 		    device_get_nameunit(subdev), err);
6109 		return (err);
6110 	}
6111 
6112 	if (tqrid != -1) {
6113 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6114 		    name);
6115 		if (err)
6116 			return (err);
6117 	} else {
6118 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6119 	}
6120 
6121 	return (0);
6122 }
6123 
6124 int
iflib_irq_alloc_generic(if_ctx_t ctx,if_irq_t irq,int rid,iflib_intr_type_t type,driver_filter_t * filter,void * filter_arg,int qid,const char * name)6125 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6126 			iflib_intr_type_t type, driver_filter_t *filter,
6127 			void *filter_arg, int qid, const char *name)
6128 {
6129 	device_t dev;
6130 	struct grouptask *gtask;
6131 	struct taskqgroup *tqg;
6132 	iflib_filter_info_t info;
6133 	gtask_fn_t *fn;
6134 	int tqrid, err;
6135 	driver_filter_t *intr_fast;
6136 	void *q;
6137 
6138 	info = &ctx->ifc_filter_info;
6139 	tqrid = rid;
6140 
6141 	switch (type) {
6142 	/* XXX merge tx/rx for netmap? */
6143 	case IFLIB_INTR_TX:
6144 		q = &ctx->ifc_txqs[qid];
6145 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6146 		gtask = &ctx->ifc_txqs[qid].ift_task;
6147 		tqg = qgroup_if_io_tqg;
6148 		fn = _task_fn_tx;
6149 		intr_fast = iflib_fast_intr;
6150 		GROUPTASK_INIT(gtask, 0, fn, q);
6151 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6152 		break;
6153 	case IFLIB_INTR_RX:
6154 		q = &ctx->ifc_rxqs[qid];
6155 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6156 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6157 		tqg = qgroup_if_io_tqg;
6158 		fn = _task_fn_rx;
6159 		intr_fast = iflib_fast_intr;
6160 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6161 		break;
6162 	case IFLIB_INTR_RXTX:
6163 		q = &ctx->ifc_rxqs[qid];
6164 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6165 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6166 		tqg = qgroup_if_io_tqg;
6167 		fn = _task_fn_rx;
6168 		intr_fast = iflib_fast_intr_rxtx;
6169 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6170 		break;
6171 	case IFLIB_INTR_ADMIN:
6172 		q = ctx;
6173 		tqrid = -1;
6174 		info = &ctx->ifc_filter_info;
6175 		gtask = NULL;
6176 		intr_fast = iflib_fast_intr_ctx;
6177 		break;
6178 	default:
6179 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6180 		    __func__);
6181 		return (EINVAL);
6182 	}
6183 
6184 	info->ifi_filter = filter;
6185 	info->ifi_filter_arg = filter_arg;
6186 	info->ifi_task = gtask;
6187 	info->ifi_ctx = q;
6188 
6189 	dev = ctx->ifc_dev;
6190 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6191 	if (err != 0) {
6192 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6193 		return (err);
6194 	}
6195 	if (type == IFLIB_INTR_ADMIN)
6196 		return (0);
6197 
6198 	if (tqrid != -1) {
6199 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6200 		    name);
6201 		if (err)
6202 			return (err);
6203 	} else {
6204 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6205 	}
6206 
6207 	return (0);
6208 }
6209 
6210 void
iflib_softirq_alloc_generic(if_ctx_t ctx,if_irq_t irq,iflib_intr_type_t type,void * arg,int qid,const char * name)6211 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6212 			    void *arg, int qid, const char *name)
6213 {
6214 	device_t dev;
6215 	struct grouptask *gtask;
6216 	struct taskqgroup *tqg;
6217 	gtask_fn_t *fn;
6218 	void *q;
6219 	int err;
6220 
6221 	switch (type) {
6222 	case IFLIB_INTR_TX:
6223 		q = &ctx->ifc_txqs[qid];
6224 		gtask = &ctx->ifc_txqs[qid].ift_task;
6225 		tqg = qgroup_if_io_tqg;
6226 		fn = _task_fn_tx;
6227 		GROUPTASK_INIT(gtask, 0, fn, q);
6228 		break;
6229 	case IFLIB_INTR_RX:
6230 		q = &ctx->ifc_rxqs[qid];
6231 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6232 		tqg = qgroup_if_io_tqg;
6233 		fn = _task_fn_rx;
6234 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6235 		break;
6236 	case IFLIB_INTR_IOV:
6237 		TASK_INIT(&ctx->ifc_vflr_task, 0, _task_fn_iov, ctx);
6238 		return;
6239 	default:
6240 		panic("unknown net intr type");
6241 	}
6242 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6243 	if (err) {
6244 		dev = ctx->ifc_dev;
6245 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6246 		    name);
6247 	}
6248 }
6249 
6250 void
iflib_irq_free(if_ctx_t ctx,if_irq_t irq)6251 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6252 {
6253 
6254 	if (irq->ii_tag)
6255 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6256 
6257 	if (irq->ii_res)
6258 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6259 		    rman_get_rid(irq->ii_res), irq->ii_res);
6260 }
6261 
6262 static int
iflib_legacy_setup(if_ctx_t ctx,driver_filter_t filter,void * filter_arg,int * rid,const char * name)6263 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6264 {
6265 	iflib_txq_t txq = ctx->ifc_txqs;
6266 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6267 	if_irq_t irq = &ctx->ifc_legacy_irq;
6268 	iflib_filter_info_t info;
6269 	device_t dev;
6270 	struct grouptask *gtask;
6271 	struct resource *res;
6272 	int err, tqrid;
6273 	bool rx_only;
6274 
6275 	info = &rxq->ifr_filter_info;
6276 	gtask = &rxq->ifr_task;
6277 	tqrid = *rid;
6278 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6279 
6280 	ctx->ifc_flags |= IFC_LEGACY;
6281 	info->ifi_filter = filter;
6282 	info->ifi_filter_arg = filter_arg;
6283 	info->ifi_task = gtask;
6284 	info->ifi_ctx = rxq;
6285 
6286 	dev = ctx->ifc_dev;
6287 	/* We allocate a single interrupt resource */
6288 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr :
6289 	    iflib_fast_intr_rxtx, NULL, info, name);
6290 	if (err != 0)
6291 		return (err);
6292 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq);
6293 	res = irq->ii_res;
6294 	taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name);
6295 
6296 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6297 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6298 	    "tx");
6299 	return (0);
6300 }
6301 
6302 void
iflib_led_create(if_ctx_t ctx)6303 iflib_led_create(if_ctx_t ctx)
6304 {
6305 
6306 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6307 	    device_get_nameunit(ctx->ifc_dev));
6308 }
6309 
6310 void
iflib_tx_intr_deferred(if_ctx_t ctx,int txqid)6311 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6312 {
6313 
6314 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6315 }
6316 
6317 void
iflib_rx_intr_deferred(if_ctx_t ctx,int rxqid)6318 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6319 {
6320 
6321 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6322 }
6323 
6324 void
iflib_admin_intr_deferred(if_ctx_t ctx)6325 iflib_admin_intr_deferred(if_ctx_t ctx)
6326 {
6327 
6328 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
6329 }
6330 
6331 void
iflib_iov_intr_deferred(if_ctx_t ctx)6332 iflib_iov_intr_deferred(if_ctx_t ctx)
6333 {
6334 
6335 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_vflr_task);
6336 }
6337 
6338 void
iflib_io_tqg_attach(struct grouptask * gt,void * uniq,int cpu,const char * name)6339 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6340 {
6341 
6342 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6343 	    name);
6344 }
6345 
6346 void
iflib_config_task_init(if_ctx_t ctx,struct task * config_task,task_fn_t * fn)6347 iflib_config_task_init(if_ctx_t ctx, struct task *config_task, task_fn_t *fn)
6348 {
6349 	TASK_INIT(config_task, 0, fn, ctx);
6350 }
6351 
6352 void
iflib_config_task_enqueue(if_ctx_t ctx,struct task * config_task)6353 iflib_config_task_enqueue(if_ctx_t ctx, struct task *config_task)
6354 {
6355 	taskqueue_enqueue(ctx->ifc_tq, config_task);
6356 }
6357 
6358 void
iflib_link_state_change(if_ctx_t ctx,int link_state,uint64_t baudrate)6359 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6360 {
6361 	if_t ifp = ctx->ifc_ifp;
6362 	iflib_txq_t txq = ctx->ifc_txqs;
6363 
6364 	if_setbaudrate(ifp, baudrate);
6365 	if (baudrate >= IF_Gbps(10)) {
6366 		STATE_LOCK(ctx);
6367 		ctx->ifc_flags |= IFC_PREFETCH;
6368 		STATE_UNLOCK(ctx);
6369 	}
6370 	/* If link down, disable watchdog */
6371 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6372 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6373 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6374 	}
6375 	ctx->ifc_link_state = link_state;
6376 	if_link_state_change(ifp, link_state);
6377 }
6378 
6379 static int
iflib_tx_credits_update(if_ctx_t ctx,iflib_txq_t txq)6380 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6381 {
6382 	int credits;
6383 #ifdef INVARIANTS
6384 	int credits_pre = txq->ift_cidx_processed;
6385 #endif
6386 
6387 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6388 	    BUS_DMASYNC_POSTREAD);
6389 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6390 		return (0);
6391 
6392 	txq->ift_processed += credits;
6393 	txq->ift_cidx_processed += credits;
6394 
6395 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6396 	if (txq->ift_cidx_processed >= txq->ift_size)
6397 		txq->ift_cidx_processed -= txq->ift_size;
6398 	return (credits);
6399 }
6400 
6401 static int
iflib_rxd_avail(if_ctx_t ctx,iflib_rxq_t rxq,qidx_t cidx,qidx_t budget)6402 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6403 {
6404 	iflib_fl_t fl;
6405 	u_int i;
6406 
6407 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6408 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6409 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6410 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6411 	    budget));
6412 }
6413 
6414 void
iflib_add_int_delay_sysctl(if_ctx_t ctx,const char * name,const char * description,if_int_delay_info_t info,int offset,int value)6415 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6416 	const char *description, if_int_delay_info_t info,
6417 	int offset, int value)
6418 {
6419 	info->iidi_ctx = ctx;
6420 	info->iidi_offset = offset;
6421 	info->iidi_value = value;
6422 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6423 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6424 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6425 	    info, 0, iflib_sysctl_int_delay, "I", description);
6426 }
6427 
6428 struct sx *
iflib_ctx_lock_get(if_ctx_t ctx)6429 iflib_ctx_lock_get(if_ctx_t ctx)
6430 {
6431 
6432 	return (&ctx->ifc_ctx_sx);
6433 }
6434 
6435 static int
iflib_msix_init(if_ctx_t ctx)6436 iflib_msix_init(if_ctx_t ctx)
6437 {
6438 	device_t dev = ctx->ifc_dev;
6439 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6440 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6441 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6442 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6443 
6444 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6445 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6446 
6447 	if (bootverbose)
6448 		device_printf(dev, "msix_init qsets capped at %d\n",
6449 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6450 
6451 	/* Override by tuneable */
6452 	if (scctx->isc_disable_msix)
6453 		goto msi;
6454 
6455 	/* First try MSI-X */
6456 	if ((msgs = pci_msix_count(dev)) == 0) {
6457 		if (bootverbose)
6458 			device_printf(dev, "MSI-X not supported or disabled\n");
6459 		goto msi;
6460 	}
6461 
6462 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6463 	/*
6464 	 * bar == -1 => "trust me I know what I'm doing"
6465 	 * Some drivers are for hardware that is so shoddily
6466 	 * documented that no one knows which bars are which
6467 	 * so the developer has to map all bars. This hack
6468 	 * allows shoddy garbage to use MSI-X in this framework.
6469 	 */
6470 	if (bar != -1) {
6471 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6472 		    SYS_RES_MEMORY, &bar, RF_ACTIVE);
6473 		if (ctx->ifc_msix_mem == NULL) {
6474 			device_printf(dev, "Unable to map MSI-X table\n");
6475 			goto msi;
6476 		}
6477 	}
6478 
6479 	admincnt = sctx->isc_admin_intrcnt;
6480 #if IFLIB_DEBUG
6481 	/* use only 1 qset in debug mode */
6482 	queuemsgs = min(msgs - admincnt, 1);
6483 #else
6484 	queuemsgs = msgs - admincnt;
6485 #endif
6486 #ifdef RSS
6487 	queues = imin(queuemsgs, rss_getnumbuckets());
6488 #else
6489 	queues = queuemsgs;
6490 #endif
6491 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6492 	if (bootverbose)
6493 		device_printf(dev,
6494 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6495 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6496 #ifdef  RSS
6497 	/* If we're doing RSS, clamp at the number of RSS buckets */
6498 	if (queues > rss_getnumbuckets())
6499 		queues = rss_getnumbuckets();
6500 #endif
6501 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6502 		rx_queues = iflib_num_rx_queues;
6503 	else
6504 		rx_queues = queues;
6505 
6506 	if (rx_queues > scctx->isc_nrxqsets)
6507 		rx_queues = scctx->isc_nrxqsets;
6508 
6509 	/*
6510 	 * We want this to be all logical CPUs by default
6511 	 */
6512 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6513 		tx_queues = iflib_num_tx_queues;
6514 	else
6515 		tx_queues = mp_ncpus;
6516 
6517 	if (tx_queues > scctx->isc_ntxqsets)
6518 		tx_queues = scctx->isc_ntxqsets;
6519 
6520 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6521 #ifdef INVARIANTS
6522 		if (tx_queues != rx_queues)
6523 			device_printf(dev,
6524 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6525 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6526 #endif
6527 		tx_queues = min(rx_queues, tx_queues);
6528 		rx_queues = min(rx_queues, tx_queues);
6529 	}
6530 
6531 	vectors = rx_queues + admincnt;
6532 	if (msgs < vectors) {
6533 		device_printf(dev,
6534 		    "insufficient number of MSI-X vectors "
6535 		    "(supported %d, need %d)\n", msgs, vectors);
6536 		goto msi;
6537 	}
6538 
6539 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6540 	    tx_queues);
6541 	msgs = vectors;
6542 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6543 		if (vectors != msgs) {
6544 			device_printf(dev,
6545 			    "Unable to allocate sufficient MSI-X vectors "
6546 			    "(got %d, need %d)\n", vectors, msgs);
6547 			pci_release_msi(dev);
6548 			if (bar != -1) {
6549 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6550 				    ctx->ifc_msix_mem);
6551 				ctx->ifc_msix_mem = NULL;
6552 			}
6553 			goto msi;
6554 		}
6555 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6556 		    vectors);
6557 		scctx->isc_vectors = vectors;
6558 		scctx->isc_nrxqsets = rx_queues;
6559 		scctx->isc_ntxqsets = tx_queues;
6560 		scctx->isc_intr = IFLIB_INTR_MSIX;
6561 
6562 		return (vectors);
6563 	} else {
6564 		device_printf(dev,
6565 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6566 		    err);
6567 		if (bar != -1) {
6568 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6569 			    ctx->ifc_msix_mem);
6570 			ctx->ifc_msix_mem = NULL;
6571 		}
6572 	}
6573 
6574 msi:
6575 	vectors = pci_msi_count(dev);
6576 	scctx->isc_nrxqsets = 1;
6577 	scctx->isc_ntxqsets = 1;
6578 	scctx->isc_vectors = vectors;
6579 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6580 		device_printf(dev, "Using an MSI interrupt\n");
6581 		scctx->isc_intr = IFLIB_INTR_MSI;
6582 	} else {
6583 		scctx->isc_vectors = 1;
6584 		device_printf(dev, "Using a Legacy interrupt\n");
6585 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6586 	}
6587 
6588 	return (vectors);
6589 }
6590 
6591 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6592 
6593 static int
mp_ring_state_handler(SYSCTL_HANDLER_ARGS)6594 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6595 {
6596 	int rc;
6597 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6598 	struct sbuf *sb;
6599 	const char *ring_state = "UNKNOWN";
6600 
6601 	/* XXX needed ? */
6602 	rc = sysctl_wire_old_buffer(req, 0);
6603 	MPASS(rc == 0);
6604 	if (rc != 0)
6605 		return (rc);
6606 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6607 	MPASS(sb != NULL);
6608 	if (sb == NULL)
6609 		return (ENOMEM);
6610 	if (state[3] <= 3)
6611 		ring_state = ring_states[state[3]];
6612 
6613 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6614 		    state[0], state[1], state[2], ring_state);
6615 	rc = sbuf_finish(sb);
6616 	sbuf_delete(sb);
6617 	return (rc);
6618 }
6619 
6620 enum iflib_ndesc_handler {
6621 	IFLIB_NTXD_HANDLER,
6622 	IFLIB_NRXD_HANDLER,
6623 };
6624 
6625 static int
mp_ndesc_handler(SYSCTL_HANDLER_ARGS)6626 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6627 {
6628 	if_ctx_t ctx = (void *)arg1;
6629 	enum iflib_ndesc_handler type = arg2;
6630 	char buf[256] = {0};
6631 	qidx_t *ndesc;
6632 	char *p, *next;
6633 	int nqs, rc, i;
6634 
6635 	nqs = 8;
6636 	switch (type) {
6637 	case IFLIB_NTXD_HANDLER:
6638 		ndesc = ctx->ifc_sysctl_ntxds;
6639 		if (ctx->ifc_sctx)
6640 			nqs = ctx->ifc_sctx->isc_ntxqs;
6641 		break;
6642 	case IFLIB_NRXD_HANDLER:
6643 		ndesc = ctx->ifc_sysctl_nrxds;
6644 		if (ctx->ifc_sctx)
6645 			nqs = ctx->ifc_sctx->isc_nrxqs;
6646 		break;
6647 	default:
6648 		printf("%s: unhandled type\n", __func__);
6649 		return (EINVAL);
6650 	}
6651 	if (nqs == 0)
6652 		nqs = 8;
6653 
6654 	for (i = 0; i < 8; i++) {
6655 		if (i >= nqs)
6656 			break;
6657 		if (i)
6658 			strcat(buf, ",");
6659 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6660 	}
6661 
6662 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6663 	if (rc || req->newptr == NULL)
6664 		return (rc);
6665 
6666 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6667 	    i++, p = strsep(&next, " ,")) {
6668 		ndesc[i] = strtoul(p, NULL, 10);
6669 	}
6670 
6671 	return (rc);
6672 }
6673 
6674 static int
iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)6675 iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)
6676 {
6677 	if_ctx_t ctx = (void *)arg1;
6678 	iflib_txq_t txq;
6679 	int i, err;
6680 	int thresh;
6681 
6682 	thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
6683 	err = sysctl_handle_int(oidp, &thresh, arg2, req);
6684 	if (err != 0) {
6685 		return err;
6686 	}
6687 
6688 	if (thresh == ctx->ifc_sysctl_tx_reclaim_thresh)
6689 		return 0;
6690 
6691 	if (thresh > ctx->ifc_softc_ctx.isc_ntxd[0] / 2) {
6692 		device_printf(ctx->ifc_dev, "TX Reclaim thresh must be <= %d\n",
6693 		    ctx->ifc_softc_ctx.isc_ntxd[0] / 2);
6694 		return (EINVAL);
6695 	}
6696 
6697 	ctx->ifc_sysctl_tx_reclaim_thresh = thresh;
6698 	if (ctx->ifc_txqs == NULL)
6699 		return (err);
6700 
6701 	txq = &ctx->ifc_txqs[0];
6702 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6703 		txq->ift_reclaim_thresh = thresh;
6704 	}
6705 	return (err);
6706 }
6707 
6708 static int
iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)6709 iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)
6710 {
6711 	if_ctx_t ctx = (void *)arg1;
6712 	iflib_txq_t txq;
6713 	int i, err;
6714 	int ticks;
6715 
6716 	ticks = ctx->ifc_sysctl_tx_reclaim_ticks;
6717 	err = sysctl_handle_int(oidp, &ticks, arg2, req);
6718 	if (err != 0) {
6719 		return err;
6720 	}
6721 
6722 	if (ticks == ctx->ifc_sysctl_tx_reclaim_ticks)
6723 		return 0;
6724 
6725 	if (ticks > hz) {
6726 		device_printf(ctx->ifc_dev,
6727 		    "TX Reclaim ticks must be <= hz (%d)\n", hz);
6728 		return (EINVAL);
6729 	}
6730 
6731 	ctx->ifc_sysctl_tx_reclaim_ticks = ticks;
6732 	if (ctx->ifc_txqs == NULL)
6733 		return (err);
6734 
6735 	txq = &ctx->ifc_txqs[0];
6736 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6737 		txq->ift_reclaim_ticks = ticks;
6738 	}
6739 	return (err);
6740 }
6741 
6742 static int
iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)6743 iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)
6744 {
6745 	if_ctx_t ctx = (void *)arg1;
6746 	iflib_txq_t txq;
6747 	int i, err;
6748 	int defer;
6749 
6750 	defer = ctx->ifc_sysctl_tx_defer_mfree;
6751 	err = sysctl_handle_int(oidp, &defer, arg2, req);
6752 	if (err != 0) {
6753 		return err;
6754 	}
6755 
6756 	if (defer == ctx->ifc_sysctl_tx_defer_mfree)
6757 		return 0;
6758 
6759 	ctx->ifc_sysctl_tx_defer_mfree = defer;
6760 	if (ctx->ifc_txqs == NULL)
6761 		return (err);
6762 
6763 	txq = &ctx->ifc_txqs[0];
6764 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6765 		txq->ift_defer_mfree = defer;
6766 	}
6767 	return (err);
6768 }
6769 
6770 #define NAME_BUFLEN 32
6771 static void
iflib_add_device_sysctl_pre(if_ctx_t ctx)6772 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6773 {
6774 	device_t dev = iflib_get_dev(ctx);
6775 	struct sysctl_oid_list *child, *oid_list;
6776 	struct sysctl_ctx_list *ctx_list;
6777 	struct sysctl_oid *node;
6778 
6779 	ctx_list = device_get_sysctl_ctx(dev);
6780 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6781 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child,
6782 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
6783 	    "IFLIB fields");
6784 	oid_list = SYSCTL_CHILDREN(node);
6785 
6786 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6787 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
6788 
6789 	SYSCTL_ADD_BOOL(ctx_list, oid_list, OID_AUTO, "simple_tx",
6790 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_simple_tx, 0,
6791 	    "use simple tx ring");
6792 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6793 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6794 	    "# of txqs to use, 0 => use default #");
6795 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6796 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6797 	    "# of rxqs to use, 0 => use default #");
6798 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6799 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6800 	    "permit #txq != #rxq");
6801 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6802 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6803 	    "disable MSI-X (default 0)");
6804 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6805 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
6806 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6807 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6808 	    "cause TX to abdicate instead of running to completion");
6809 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6810 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6811 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6812 	    "offset to start using cores at");
6813 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6814 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6815 	    "use separate cores for TX and RX");
6816 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6817 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6818 	    "try to make use of logical cores for TX and RX");
6819 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "use_extra_msix_vectors",
6820 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
6821 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
6822 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "allocated_msix_vectors",
6823 	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
6824 	    "total # of MSI-X vectors allocated by driver");
6825 
6826 	/* XXX change for per-queue sizes */
6827 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6828 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6829 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6830 	    "list of # of TX descriptors to use, 0 = use default #");
6831 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6832 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6833 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6834 	    "list of # of RX descriptors to use, 0 = use default #");
6835 }
6836 
6837 static void
iflib_add_device_sysctl_post(if_ctx_t ctx)6838 iflib_add_device_sysctl_post(if_ctx_t ctx)
6839 {
6840 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6841 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6842 	device_t dev = iflib_get_dev(ctx);
6843 	struct sysctl_oid_list *child;
6844 	struct sysctl_ctx_list *ctx_list;
6845 	iflib_fl_t fl;
6846 	iflib_txq_t txq;
6847 	iflib_rxq_t rxq;
6848 	int i, j;
6849 	char namebuf[NAME_BUFLEN];
6850 	char *qfmt;
6851 	struct sysctl_oid *queue_node, *fl_node, *node;
6852 	struct sysctl_oid_list *queue_list, *fl_list;
6853 	ctx_list = device_get_sysctl_ctx(dev);
6854 
6855 	node = ctx->ifc_sysctl_node;
6856 	child = SYSCTL_CHILDREN(node);
6857 
6858        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_thresh",
6859            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
6860            0, iflib_handle_tx_reclaim_thresh, "I",
6861            "Number of TX descs outstanding before reclaim is called");
6862 
6863        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_ticks",
6864            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
6865            0, iflib_handle_tx_reclaim_ticks, "I",
6866            "Number of ticks before a TX reclaim is forced");
6867 
6868        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_defer_mfree",
6869            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
6870            0, iflib_handle_tx_defer_mfree, "I",
6871            "Free completed transmits outside of TX ring lock");
6872 
6873 	if (scctx->isc_ntxqsets > 100)
6874 		qfmt = "txq%03d";
6875 	else if (scctx->isc_ntxqsets > 10)
6876 		qfmt = "txq%02d";
6877 	else
6878 		qfmt = "txq%d";
6879 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6880 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6881 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6882 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6883 		queue_list = SYSCTL_CHILDREN(queue_node);
6884 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6885 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
6886 		    "cpu this queue is bound to");
6887 #if MEMORY_LOGGING
6888 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6889 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
6890 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6891 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
6892 #endif
6893 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6894 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
6895 		    "# of times m_defrag was called");
6896 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6897 		    CTLFLAG_RD, &txq->ift_pullups,
6898 		    "# of times m_pullup was called");
6899 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6900 		    "mbuf_defrag_failed", CTLFLAG_RD,
6901 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6902 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6903 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
6904 		    "# of times no descriptors were available");
6905 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6906 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
6907 		    "# of times DMA map failed");
6908 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6909 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
6910 		    "# of times txd_encap returned EFBIG");
6911 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6912 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
6913 		    "# of times map failed for other than EFBIG");
6914 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6915 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
6916 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6917 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
6918 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
6919 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
6920 		    1, "Consumer Index seen by credit update");
6921 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6922 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
6923 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
6924 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
6925 		    "descriptors procesed for clean");
6926 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6927 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
6928 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6929 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6930 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6931 		    mp_ring_state_handler, "A", "soft ring state");
6932 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6933 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
6934 		    "# of enqueues to the mp_ring for this queue");
6935 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6936 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
6937 		    "# of drops in the mp_ring for this queue");
6938 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6939 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
6940 		    "# of normal consumer starts in mp_ring for this queue");
6941 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6942 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
6943 		    "# of consumer stalls in the mp_ring for this queue");
6944 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6945 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
6946 		    "# of consumer restarts in the mp_ring for this queue");
6947 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
6948 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
6949 		    "# of consumer abdications in the mp_ring for this queue");
6950 	}
6951 
6952 	if (scctx->isc_nrxqsets > 100)
6953 		qfmt = "rxq%03d";
6954 	else if (scctx->isc_nrxqsets > 10)
6955 		qfmt = "rxq%02d";
6956 	else
6957 		qfmt = "rxq%d";
6958 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6959 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6960 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6961 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6962 		queue_list = SYSCTL_CHILDREN(queue_node);
6963 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6964 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
6965 		    "cpu this queue is bound to");
6966 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6967 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
6968 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
6969 			    "Consumer Index");
6970 		}
6971 
6972 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6973 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6974 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
6975 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
6976 			    NULL, "freelist Name");
6977 			fl_list = SYSCTL_CHILDREN(fl_node);
6978 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6979 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
6980 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6981 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
6982 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6983 			    CTLFLAG_RD, &fl->ifl_credits, 1,
6984 			    "credits available");
6985 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6986 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
6987 #if MEMORY_LOGGING
6988 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6989 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
6990 			    "mbufs allocated");
6991 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6992 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
6993 			    "mbufs freed");
6994 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6995 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
6996 			    "clusters allocated");
6997 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
6998 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
6999 			    "clusters freed");
7000 #endif
7001 		}
7002 	}
7003 
7004 }
7005 
7006 void
iflib_request_reset(if_ctx_t ctx)7007 iflib_request_reset(if_ctx_t ctx)
7008 {
7009 
7010 	STATE_LOCK(ctx);
7011 	ctx->ifc_flags |= IFC_DO_RESET;
7012 	STATE_UNLOCK(ctx);
7013 }
7014 
7015 #ifndef __NO_STRICT_ALIGNMENT
7016 static struct mbuf *
iflib_fixup_rx(struct mbuf * m)7017 iflib_fixup_rx(struct mbuf *m)
7018 {
7019 	struct mbuf *n;
7020 
7021 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7022 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7023 		m->m_data += ETHER_HDR_LEN;
7024 		n = m;
7025 	} else {
7026 		MGETHDR(n, M_NOWAIT, MT_DATA);
7027 		if (n == NULL) {
7028 			m_freem(m);
7029 			return (NULL);
7030 		}
7031 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7032 		m->m_data += ETHER_HDR_LEN;
7033 		m->m_len -= ETHER_HDR_LEN;
7034 		n->m_len = ETHER_HDR_LEN;
7035 		M_MOVE_PKTHDR(n, m);
7036 		n->m_next = m;
7037 	}
7038 	return (n);
7039 }
7040 #endif
7041 
7042 #ifdef DEBUGNET
7043 static void
iflib_debugnet_init(if_t ifp,int * nrxr,int * ncl,int * clsize)7044 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7045 {
7046 	if_ctx_t ctx;
7047 
7048 	ctx = if_getsoftc(ifp);
7049 	CTX_LOCK(ctx);
7050 	*nrxr = NRXQSETS(ctx);
7051 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7052 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7053 	CTX_UNLOCK(ctx);
7054 }
7055 
7056 static void
iflib_debugnet_event(if_t ifp,enum debugnet_ev event)7057 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7058 {
7059 	if_ctx_t ctx;
7060 	if_softc_ctx_t scctx;
7061 	iflib_fl_t fl;
7062 	iflib_rxq_t rxq;
7063 	int i, j;
7064 
7065 	ctx = if_getsoftc(ifp);
7066 	scctx = &ctx->ifc_softc_ctx;
7067 
7068 	switch (event) {
7069 	case DEBUGNET_START:
7070 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7071 			rxq = &ctx->ifc_rxqs[i];
7072 			for (j = 0; j < rxq->ifr_nfl; j++) {
7073 				fl = rxq->ifr_fl;
7074 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7075 			}
7076 		}
7077 		iflib_no_tx_batch = 1;
7078 		break;
7079 	default:
7080 		break;
7081 	}
7082 }
7083 
7084 static int
iflib_debugnet_transmit(if_t ifp,struct mbuf * m)7085 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7086 {
7087 	if_ctx_t ctx;
7088 	iflib_txq_t txq;
7089 	int error;
7090 
7091 	ctx = if_getsoftc(ifp);
7092 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7093 	    IFF_DRV_RUNNING)
7094 		return (EBUSY);
7095 
7096 	txq = &ctx->ifc_txqs[0];
7097 	error = iflib_encap(txq, &m);
7098 	if (error == 0)
7099 		(void)iflib_txd_db_check(txq, true);
7100 	return (error);
7101 }
7102 
7103 static int
iflib_debugnet_poll(if_t ifp,int count)7104 iflib_debugnet_poll(if_t ifp, int count)
7105 {
7106 	struct epoch_tracker et;
7107 	if_ctx_t ctx;
7108 	if_softc_ctx_t scctx;
7109 	iflib_txq_t txq;
7110 	int i;
7111 
7112 	ctx = if_getsoftc(ifp);
7113 	scctx = &ctx->ifc_softc_ctx;
7114 
7115 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7116 	    IFF_DRV_RUNNING)
7117 		return (EBUSY);
7118 
7119 	txq = &ctx->ifc_txqs[0];
7120 	(void)iflib_completed_tx_reclaim(txq, NULL);
7121 
7122 	NET_EPOCH_ENTER(et);
7123 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7124 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7125 	NET_EPOCH_EXIT(et);
7126 	return (0);
7127 }
7128 #endif /* DEBUGNET */
7129 
7130 #ifndef ALTQ
7131 static inline iflib_txq_t
iflib_simple_select_queue(if_ctx_t ctx,struct mbuf * m)7132 iflib_simple_select_queue(if_ctx_t ctx, struct mbuf *m)
7133 {
7134 	int qidx;
7135 
7136 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
7137 		qidx = QIDX(ctx, m);
7138 	else
7139 		qidx = NTXQSETS(ctx) + FIRST_QSET(ctx) - 1;
7140 	return (&ctx->ifc_txqs[qidx]);
7141 }
7142 
7143 static int
iflib_simple_transmit(if_t ifp,struct mbuf * m)7144 iflib_simple_transmit(if_t ifp, struct mbuf *m)
7145 {
7146 	if_ctx_t ctx;
7147 	iflib_txq_t txq;
7148 	struct mbuf **m_defer;
7149 	int error, i, reclaimable;
7150 	int bytes_sent = 0, pkt_sent = 0, mcast_sent = 0;
7151 
7152 
7153 	ctx = if_getsoftc(ifp);
7154 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0
7155 		|| !LINK_ACTIVE(ctx))) {
7156 		DBG_COUNTER_INC(tx_frees);
7157 		m_freem(m);
7158 		return (ENETDOWN);
7159 	}
7160 
7161 	txq = iflib_simple_select_queue(ctx, m);
7162 	mtx_lock(&txq->ift_mtx);
7163 	error = iflib_encap(txq, &m);
7164 	if (error == 0) {
7165 		pkt_sent++;
7166 		bytes_sent += m->m_pkthdr.len;
7167 		mcast_sent += !!(m->m_flags & M_MCAST);
7168 		(void)iflib_txd_db_check(txq, true);
7169 	} else {
7170 		if (error == ENOBUFS)
7171 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
7172 		else
7173 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
7174 	}
7175 	m_defer = NULL;
7176 	reclaimable = iflib_txq_can_reclaim(txq);
7177 	if (reclaimable != 0) {
7178 		/*
7179 		 * Try to set m_defer to the deferred mbuf reclaim array.  If
7180 		 * we can, the frees will happen outside the tx lock.  If we
7181 		 * can't, it means another thread is still proccessing frees.
7182 		 */
7183 		if (txq->ift_defer_mfree &&
7184 		    atomic_cmpset_acq_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7185 			(uintptr_t )txq->ift_sds.ifsd_m_deferb, 0)) {
7186 			m_defer = txq->ift_sds.ifsd_m_deferb;
7187 		}
7188 		_iflib_completed_tx_reclaim(txq, m_defer, reclaimable);
7189 	}
7190 	mtx_unlock(&txq->ift_mtx);
7191 
7192 	/*
7193 	 * Process mbuf frees outside the tx lock
7194 	 */
7195 	if (m_defer != NULL) {
7196 		for (i = 0; m_defer[i] != NULL; i++) {
7197 			m_freem(m_defer[i]);
7198 			m_defer[i] = NULL;
7199 		}
7200 		atomic_store_rel_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7201 		    (uintptr_t)m_defer);
7202 	}
7203 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
7204 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
7205 	if (mcast_sent)
7206 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
7207 
7208 	return (error);
7209 }
7210 #endif
7211