xref: /freebsd/sys/net/iflib.c (revision 06aaca43fe05d22cd289af9b6fa751b9bd08b560)
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/fail.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/module.h>
42 #include <sys/kobj.h>
43 #include <sys/proc.h>
44 #include <sys/rman.h>
45 #include <sys/sbuf.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_private.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/debugnet.h>
64 #include <net/pfil.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/udp.h>
76 #include <netinet/ip_var.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 
92 #include "ifdi_if.h"
93 
94 #ifdef PCI_IOV
95 #include <dev/pci/pci_iov.h>
96 #endif
97 
98 #include <sys/bitstring.h>
99 /*
100  * enable accounting of every mbuf as it comes in to and goes out of
101  * iflib's software descriptor references
102  */
103 #define MEMORY_LOGGING 0
104 /*
105  * Enable mbuf vectors for compressing long mbuf chains
106  */
107 
108 /*
109  * NB:
110  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
111  *   we prefetch needs to be determined by the time spent in m_free vis a vis
112  *   the cost of a prefetch. This will of course vary based on the workload:
113  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
114  *        is quite expensive, thus suggesting very little prefetch.
115  *      - small packet forwarding which is just returning a single mbuf to
116  *        UMA will typically be very fast vis a vis the cost of a memory
117  *        access.
118  */
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE	(1U << 0)
133 #define	IFLIB_RXEOF_EMPTY	(2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 static void iflib_tqg_detach(if_ctx_t ctx);
147 #ifndef ALTQ
148 static int  iflib_simple_transmit(if_t ifp, struct mbuf *m);
149 #endif
150 
151 typedef struct iflib_filter_info {
152 	driver_filter_t *ifi_filter;
153 	void *ifi_filter_arg;
154 	struct grouptask *ifi_task;
155 	void *ifi_ctx;
156 } *iflib_filter_info_t;
157 
158 struct iflib_ctx {
159 	KOBJ_FIELDS;
160 	/*
161 	 * Pointer to hardware driver's softc
162 	 */
163 	void *ifc_softc;
164 	device_t ifc_dev;
165 	if_t ifc_ifp;
166 
167 	cpuset_t ifc_cpus;
168 	if_shared_ctx_t ifc_sctx;
169 	struct if_softc_ctx ifc_softc_ctx;
170 
171 	struct sx ifc_ctx_sx;
172 	struct mtx ifc_state_mtx;
173 
174 	iflib_txq_t ifc_txqs;
175 	iflib_rxq_t ifc_rxqs;
176 	uint32_t ifc_if_flags;
177 	uint32_t ifc_flags;
178 	uint32_t ifc_max_fl_buf_size;
179 	uint32_t ifc_rx_mbuf_sz;
180 
181 	int ifc_link_state;
182 	uint32_t ifc_tx_watchdog_events;
183 	struct cdev *ifc_led_dev;
184 	int ifc_led_state;
185 	struct resource *ifc_msix_mem;
186 
187 	struct if_irq ifc_legacy_irq;
188 	struct task ifc_admin_task;
189 	struct task ifc_led_task;
190 	struct task ifc_vflr_task;
191 	struct taskqueue *ifc_tq;
192 	struct iflib_filter_info ifc_filter_info;
193 	struct ifmedia	ifc_media;
194 	struct ifmedia	*ifc_mediap;
195 
196 	struct sysctl_ctx_list ifc_sysctl_ctx;
197 	struct sysctl_oid *ifc_sysctl_node;
198 	uint16_t ifc_sysctl_ntxqs;
199 	uint16_t ifc_sysctl_nrxqs;
200 	uint16_t ifc_sysctl_qs_eq_override;
201 	uint16_t ifc_sysctl_rx_budget;
202 	uint16_t ifc_sysctl_tx_abdicate;
203 	uint16_t ifc_sysctl_core_offset;
204 #define	CORE_OFFSET_UNSPECIFIED	0xffff
205 	uint8_t  ifc_sysctl_separate_txrx;
206 	uint8_t  ifc_sysctl_use_logical_cores;
207 	uint16_t ifc_sysctl_extra_msix_vectors;
208 	bool     ifc_cpus_are_physical_cores;
209 	bool     ifc_core_offset_ref;
210 	bool     ifc_sysctl_simple_tx;
211 	bool     ifc_sysctl_tx_defer_mfree;
212 	uint16_t ifc_sysctl_tx_reclaim_thresh;
213 	uint16_t ifc_sysctl_tx_reclaim_ticks;
214 
215 	qidx_t ifc_sysctl_ntxds[8];
216 	qidx_t ifc_sysctl_nrxds[8];
217 	struct if_txrx ifc_txrx;
218 #define isc_txd_encap		ifc_txrx.ift_txd_encap
219 #define isc_txd_flush		ifc_txrx.ift_txd_flush
220 #define isc_txd_credits_update	ifc_txrx.ift_txd_credits_update
221 #define isc_rxd_available	ifc_txrx.ift_rxd_available
222 #define isc_rxd_pkt_get		ifc_txrx.ift_rxd_pkt_get
223 #define isc_rxd_refill		ifc_txrx.ift_rxd_refill
224 #define isc_rxd_flush		ifc_txrx.ift_rxd_flush
225 #define isc_legacy_intr		ifc_txrx.ift_legacy_intr
226 #define isc_txq_select		ifc_txrx.ift_txq_select
227 #define isc_txq_select_v2	ifc_txrx.ift_txq_select_v2
228 
229 	eventhandler_tag ifc_vlan_attach_event;
230 	eventhandler_tag ifc_vlan_detach_event;
231 	struct ether_addr ifc_mac;
232 };
233 
234 void *
iflib_get_softc(if_ctx_t ctx)235 iflib_get_softc(if_ctx_t ctx)
236 {
237 
238 	return (ctx->ifc_softc);
239 }
240 
241 device_t
iflib_get_dev(if_ctx_t ctx)242 iflib_get_dev(if_ctx_t ctx)
243 {
244 
245 	return (ctx->ifc_dev);
246 }
247 
248 if_t
iflib_get_ifp(if_ctx_t ctx)249 iflib_get_ifp(if_ctx_t ctx)
250 {
251 
252 	return (ctx->ifc_ifp);
253 }
254 
255 struct ifmedia *
iflib_get_media(if_ctx_t ctx)256 iflib_get_media(if_ctx_t ctx)
257 {
258 
259 	return (ctx->ifc_mediap);
260 }
261 
262 void
iflib_set_mac(if_ctx_t ctx,uint8_t mac[ETHER_ADDR_LEN])263 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
264 {
265 
266 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
267 }
268 
269 if_softc_ctx_t
iflib_get_softc_ctx(if_ctx_t ctx)270 iflib_get_softc_ctx(if_ctx_t ctx)
271 {
272 
273 	return (&ctx->ifc_softc_ctx);
274 }
275 
276 if_shared_ctx_t
iflib_get_sctx(if_ctx_t ctx)277 iflib_get_sctx(if_ctx_t ctx)
278 {
279 
280 	return (ctx->ifc_sctx);
281 }
282 
283 uint16_t
iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)284 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
285 {
286 
287 	return (ctx->ifc_sysctl_extra_msix_vectors);
288 }
289 
290 #define IP_ALIGNED(m)		((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
291 #define CACHE_PTR_INCREMENT	(CACHE_LINE_SIZE / sizeof(void *))
292 #define CACHE_PTR_NEXT(ptr)	((void *)(roundup2(ptr, CACHE_LINE_SIZE)))
293 
294 #define LINK_ACTIVE(ctx)	((ctx)->ifc_link_state == LINK_STATE_UP)
295 #define CTX_IS_VF(ctx)		((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
296 
297 typedef struct iflib_sw_rx_desc_array {
298 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
299 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
300 	caddr_t		*ifsd_cl;	/* direct cluster pointer for rx */
301 	bus_addr_t	*ifsd_ba;	/* bus addr of cluster for rx */
302 } iflib_rxsd_array_t;
303 
304 typedef struct iflib_sw_tx_desc_array {
305 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
306 	bus_dmamap_t	*ifsd_tso_map;	/* bus_dma maps for TSO packet */
307 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
308 	struct mbuf	**ifsd_m_defer;	/* deferred mbuf ptr */
309 	struct mbuf	**ifsd_m_deferb;/* deferred mbuf backing ptr */
310 } if_txsd_vec_t;
311 
312 /* magic number that should be high enough for any hardware */
313 #define IFLIB_MAX_TX_SEGS		128
314 #define IFLIB_RX_COPY_THRESH		128
315 #define IFLIB_MAX_RX_REFRESH		32
316 /* The minimum descriptors per second before we start coalescing */
317 #define IFLIB_MIN_DESC_SEC		16384
318 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
319 /* maximum number of txqs that can share an rx interrupt */
320 #define IFLIB_MAX_TX_SHARED_INTR	4
321 
322 /* this should really scale with ring size - this is a fairly arbitrary value */
323 #define TX_BATCH_SIZE			32
324 
325 #define IFLIB_RESTART_BUDGET		8
326 
327 
328 /*
329  * Encode TSO or !TSO in the low bits of the tx ifsd_m pointer so as
330  * to avoid defref'ing the mbuf to determine the correct busdma resources
331  * to release
332  */
333 #define IFLIB_TSO		(1ULL << 0)
334 #define IFLIB_NO_TSO		(2ULL << 0)
335 #define IFLIB_FLAGS_MASK	(0x3ULL)
336 #define IFLIB_SAVE_MBUF(mbuf, flags)	((void *)(((uintptr_t)mbuf) | flags))
337 #define IFLIB_GET_FLAGS(a)	((uintptr_t)a & IFLIB_FLAGS_MASK)
338 #define IFLIB_GET_MBUF(a)	((struct mbuf *)((uintptr_t)a & ~IFLIB_FLAGS_MASK))
339 
340 
341 #define	IFC_LEGACY		0x001
342 #define	IFC_QFLUSH		0x002
343 #define	IFC_MULTISEG		0x004
344 #define	IFC_INIT_FAILED		0x008
345 #define	IFC_SC_ALLOCATED	0x010
346 #define	IFC_INIT_DONE		0x020
347 #define	IFC_PREFETCH		0x040
348 #define	IFC_DO_RESET		0x080
349 #define	IFC_DO_WATCHDOG		0x100
350 #define	IFC_DO_RESET_IF_UP	0x200
351 #define	IFC_SPARE2		0x400
352 #define	IFC_IN_DETACH		0x800
353 
354 #define	IFC_NETMAP_TX_IRQ	0x80000000
355 
356 #define CSUM_OFFLOAD		(CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \
357 				 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \
358 				 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP)
359 
360 struct iflib_txq {
361 	qidx_t		ift_in_use;
362 	qidx_t		ift_cidx;
363 	qidx_t		ift_cidx_processed;
364 	qidx_t		ift_pidx;
365 	uint8_t		ift_gen;
366 	uint8_t		ift_br_offset:1,
367 			ift_defer_mfree:1,
368 			ift_spare_bits0:6;
369 	uint16_t	ift_npending;
370 	uint16_t	ift_db_pending;
371 	uint16_t	ift_rs_pending;
372 	uint32_t	ift_last_reclaim;
373 	uint16_t	ift_reclaim_thresh;
374 	uint16_t	ift_reclaim_ticks;
375 	uint8_t		ift_txd_size[8];
376 	uint64_t	ift_processed;
377 	uint64_t	ift_cleaned;
378 	uint64_t	ift_processed_prev;
379 #if MEMORY_LOGGING
380 	uint64_t	ift_enqueued;
381 	uint64_t	ift_dequeued;
382 #endif
383 	uint64_t	ift_no_tx_dma_setup;
384 	uint64_t	ift_no_desc_avail;
385 	uint64_t	ift_mbuf_defrag_failed;
386 	uint64_t	ift_mbuf_defrag;
387 	uint64_t	ift_map_failed;
388 	uint64_t	ift_txd_encap_efbig;
389 	uint64_t	ift_pullups;
390 	uint64_t	ift_last_timer_tick;
391 
392 	struct mtx	ift_mtx;
393 	struct mtx	ift_db_mtx;
394 
395 	/* constant values */
396 	if_ctx_t	ift_ctx;
397 	struct ifmp_ring	*ift_br;
398 	struct grouptask	ift_task;
399 	qidx_t		ift_size;
400 	qidx_t		ift_pad;
401 	uint16_t	ift_id;
402 	struct callout	ift_timer;
403 #ifdef DEV_NETMAP
404 	struct callout	ift_netmap_timer;
405 #endif /* DEV_NETMAP */
406 
407 	if_txsd_vec_t	ift_sds;
408 	/*
409 	 * TX watchdog state, updated once per iflib_timer period.  The
410 	 * period count saturates instead of wrapping, and is 16 bits
411 	 * wide so that it still reaches any value
412 	 * net.iflib.tx_watchdog_periods is plausibly set to; an 8-bit
413 	 * counter would silently disable the check for a threshold
414 	 * above 255.
415 	 */
416 	qidx_t		ift_outstanding_prev;
417 	uint16_t	ift_wdog_armed;
418 	uint8_t		ift_closed;
419 	uint8_t		ift_update_freq;
420 	uint8_t		ift_spare0[2];	/* pad to the next pointer boundary */
421 	struct iflib_filter_info ift_filter_info;
422 	bus_dma_tag_t	ift_buf_tag;
423 	bus_dma_tag_t	ift_tso_buf_tag;
424 	iflib_dma_info_t	ift_ifdi;
425 #define	MTX_NAME_LEN	32
426 	char		ift_mtx_name[MTX_NAME_LEN];
427 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
428 #ifdef IFLIB_DIAGNOSTICS
429 	uint64_t ift_cpu_exec_count[256];
430 #endif
431 } __aligned(CACHE_LINE_SIZE);
432 
433 struct iflib_fl {
434 	qidx_t		ifl_cidx;
435 	qidx_t		ifl_pidx;
436 	qidx_t		ifl_credits;
437 	uint8_t		ifl_gen;
438 	uint8_t		ifl_rxd_size;
439 #if MEMORY_LOGGING
440 	uint64_t	ifl_m_enqueued;
441 	uint64_t	ifl_m_dequeued;
442 	uint64_t	ifl_cl_enqueued;
443 	uint64_t	ifl_cl_dequeued;
444 #endif
445 	/* implicit pad */
446 	bitstr_t	*ifl_rx_bitmap;
447 	qidx_t		ifl_fragidx;
448 	/* constant */
449 	qidx_t		ifl_size;
450 	uint16_t	ifl_buf_size;
451 	uint16_t	ifl_cltype;
452 	uma_zone_t	ifl_zone;
453 	iflib_rxsd_array_t	ifl_sds;
454 	iflib_rxq_t	ifl_rxq;
455 	uint8_t		ifl_id;
456 	bus_dma_tag_t	ifl_buf_tag;
457 	iflib_dma_info_t	ifl_ifdi;
458 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
459 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
460 }  __aligned(CACHE_LINE_SIZE);
461 
462 static inline qidx_t
get_inuse(int size,qidx_t cidx,qidx_t pidx,uint8_t gen)463 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
464 {
465 	qidx_t used;
466 
467 	if (pidx > cidx)
468 		used = pidx - cidx;
469 	else if (pidx < cidx)
470 		used = size - cidx + pidx;
471 	else if (gen == 0 && pidx == cidx)
472 		used = 0;
473 	else if (gen == 1 && pidx == cidx)
474 		used = size;
475 	else
476 		panic("bad state");
477 
478 	return (used);
479 }
480 
481 #define TXQ_AVAIL(txq) ((txq->ift_size - txq->ift_pad) -\
482 	    get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
483 
484 #define IDXDIFF(head, tail, wrap) \
485 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
486 
487 struct iflib_rxq {
488 	if_ctx_t	ifr_ctx;
489 	iflib_fl_t	ifr_fl;
490 	struct pfil_head	*pfil;
491 	/*
492 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
493 	 * the completion queue consumer index.  Otherwise it's unused.
494 	 */
495 	qidx_t		ifr_cq_cidx;
496 	uint16_t	ifr_id;
497 	uint8_t		ifr_nfl;
498 	uint8_t		ifr_ntxqirq;
499 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
500 	uint8_t		ifr_fl_offset;
501 	struct lro_ctrl		ifr_lc;
502 	struct grouptask	ifr_task;
503 	struct callout		ifr_watchdog;
504 	struct iflib_filter_info ifr_filter_info;
505 	iflib_dma_info_t	ifr_ifdi;
506 
507 	/* dynamically allocate if any drivers need a value substantially larger than this */
508 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
509 #ifdef IFLIB_DIAGNOSTICS
510 	uint64_t ifr_cpu_exec_count[256];
511 #endif
512 }  __aligned(CACHE_LINE_SIZE);
513 
514 typedef struct if_rxsd {
515 	caddr_t *ifsd_cl;
516 	iflib_fl_t ifsd_fl;
517 } *if_rxsd_t;
518 
519 /*
520  * Only allow a single packet to take up most 1/nth of the tx ring
521  */
522 #define MAX_SINGLE_PACKET_FRACTION 12
523 #define IF_BAD_DMA	((bus_addr_t)-1)
524 
525 #define CTX_ACTIVE(ctx)	((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
526 
527 #define CTX_LOCK_INIT(_sc)	sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
528 #define CTX_LOCK(ctx)		sx_xlock(&(ctx)->ifc_ctx_sx)
529 #define CTX_UNLOCK(ctx)		sx_xunlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_LOCK_DESTROY(ctx)	sx_destroy(&(ctx)->ifc_ctx_sx)
531 
532 #define STATE_LOCK_INIT(_sc, _name)	mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
533 #define STATE_LOCK(ctx)		mtx_lock(&(ctx)->ifc_state_mtx)
534 #define STATE_UNLOCK(ctx)	mtx_unlock(&(ctx)->ifc_state_mtx)
535 #define STATE_LOCK_DESTROY(ctx)	mtx_destroy(&(ctx)->ifc_state_mtx)
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq)	mtx_unlock(&txq->ift_mtx)
539 
540 /* Our boot-time initialization hook */
541 static int	iflib_module_event_handler(module_t, int, void *);
542 
543 static moduledata_t iflib_moduledata = {
544 	"iflib",
545 	iflib_module_event_handler,
546 	NULL
547 };
548 
549 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
550 MODULE_VERSION(iflib, 1);
551 
552 MODULE_DEPEND(iflib, pci, 1, 1, 1);
553 MODULE_DEPEND(iflib, ether, 1, 1, 1);
554 
555 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
556 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
557 
558 #ifndef IFLIB_DEBUG_COUNTERS
559 #ifdef INVARIANTS
560 #define IFLIB_DEBUG_COUNTERS 1
561 #else
562 #define IFLIB_DEBUG_COUNTERS 0
563 #endif /* !INVARIANTS */
564 #endif
565 
566 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
567     "iflib driver parameters");
568 
569 static SYSCTL_NODE(_debug_fail_point, OID_AUTO, iflib,
570     CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "iflib fail points");
571 
572 static char iflib_register_fail_device[32];
573 SYSCTL_STRING(_debug_fail_point_iflib, OID_AUTO, register_device,
574     CTLFLAG_RW | CTLFLAG_MPSAFE,
575     iflib_register_fail_device, sizeof(iflib_register_fail_device),
576     "device name eligible for registration fail points");
577 
578 static char iflib_admin_task_fail_device[32];
579 SYSCTL_STRING(_debug_fail_point_iflib, OID_AUTO, admin_task_device,
580     CTLFLAG_RW | CTLFLAG_MPSAFE,
581     iflib_admin_task_fail_device, sizeof(iflib_admin_task_fail_device),
582     "device name eligible for admin task fail points");
583 
584 /*
585  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
586  */
587 static int iflib_min_tx_latency = 0;
588 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
589     &iflib_min_tx_latency, 0,
590     "minimize transmit latency at the possible expense of throughput");
591 static int iflib_no_tx_batch = 0;
592 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
593     &iflib_no_tx_batch, 0,
594     "minimize transmit latency at the possible expense of throughput");
595 static int iflib_timer_default = 1000;
596 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
597     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
598 /*
599  * Consecutive timer periods a TX queue must stay frozen - see
600  * iflib_timer(), which defines that state - before the hardware is
601  * asked whether it has completions pending.  Four periods is roughly
602  * two seconds with the default timer interval: a healthy queue on
603  * hardware that coalesces completion reports (e.g. 8254x,
604  * TXDCTL.WTHRESH) stays frozen for at most two (measured on 82541PI),
605  * a wedged one until it is reset.
606  */
607 static int iflib_tx_watchdog_periods = 4;
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_watchdog_periods, CTLFLAG_RWTUN,
609     &iflib_tx_watchdog_periods, 0,
610     "consecutive frozen timer periods before a TX queue is checked for "
611     "a hang (0 disables the check)");
612 
613 
614 #if IFLIB_DEBUG_COUNTERS
615 
616 static int iflib_tx_seen;
617 static int iflib_tx_sent;
618 static int iflib_tx_encap;
619 static int iflib_rx_allocs;
620 static int iflib_fl_refills;
621 static int iflib_fl_refills_large;
622 static int iflib_tx_frees;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
625     "# TX mbufs seen");
626 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
627     "# TX mbufs sent");
628 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
629     "# TX mbufs encapped");
630 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
631     "# TX frees");
632 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
633     "# RX allocations");
634 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
635     "# refills");
636 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
637     &iflib_fl_refills_large, 0, "# large refills");
638 
639 static int iflib_txq_drain_flushing;
640 static int iflib_txq_drain_oactive;
641 static int iflib_txq_drain_notready;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
644     &iflib_txq_drain_flushing, 0, "# drain flushes");
645 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
646     &iflib_txq_drain_oactive, 0, "# drain oactives");
647 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
648     &iflib_txq_drain_notready, 0, "# drain notready");
649 
650 static int iflib_encap_load_mbuf_fail;
651 static int iflib_encap_pad_mbuf_fail;
652 static int iflib_encap_txq_avail_fail;
653 static int iflib_encap_txd_encap_fail;
654 
655 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
656     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
657 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
658     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
659 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
660     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
661 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
662     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
663 
664 static int iflib_task_fn_rxs;
665 static int iflib_rx_intr_enables;
666 static int iflib_fast_intrs;
667 static int iflib_rx_unavail;
668 static int iflib_rx_ctx_inactive;
669 static int iflib_rx_if_input;
670 static int iflib_rxd_flush;
671 
672 static int iflib_verbose_debug;
673 
674 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
675     "# task_fn_rx calls");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
677     &iflib_rx_intr_enables, 0, "# RX intr enables");
678 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
679     "# fast_intr calls");
680 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
681     "# times rxeof called with no available data");
682 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
683     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
684 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
685     0, "# times rxeof called if_input");
686 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
687     "# times rxd_flush called");
688 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
689     &iflib_verbose_debug, 0, "enable verbose debugging");
690 
691 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
692 static void
iflib_debug_reset(void)693 iflib_debug_reset(void)
694 {
695 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
696 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
697 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
698 		iflib_txq_drain_notready =
699 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
700 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
701 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
702 		iflib_rx_unavail =
703 		iflib_rx_ctx_inactive = iflib_rx_if_input =
704 		iflib_rxd_flush = 0;
705 }
706 
707 #else
708 #define DBG_COUNTER_INC(name)
iflib_debug_reset(void)709 static void iflib_debug_reset(void) {}
710 #endif
711 
712 #define IFLIB_DEBUG 0
713 
714 static void iflib_tx_structures_free(if_ctx_t ctx);
715 static void iflib_rx_structures_free(if_ctx_t ctx);
716 static int iflib_queues_alloc(if_ctx_t ctx);
717 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
718 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
719 static int iflib_qset_structures_setup(if_ctx_t ctx);
720 static int iflib_msix_init(if_ctx_t ctx);
721 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
722 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
723 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
724 #ifdef ALTQ
725 static void iflib_altq_if_start(if_t ifp);
726 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
727 #endif
728 static void iflib_register(if_ctx_t);
729 static void iflib_deregister(if_ctx_t);
730 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
731 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
732 static void iflib_init_locked(if_ctx_t ctx);
733 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
734 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
735 static void iflib_ifmp_purge(iflib_txq_t txq);
736 static void _iflib_pre_assert(if_softc_ctx_t scctx);
737 static void iflib_stop(if_ctx_t ctx);
738 static void iflib_if_init_locked(if_ctx_t ctx);
739 static void iflib_free_intr_mem(if_ctx_t ctx);
740 #ifndef __NO_STRICT_ALIGNMENT
741 static struct mbuf *iflib_fixup_rx(struct mbuf *m);
742 #endif
743 static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq,
744     struct mbuf **m_defer);
745 static __inline void iflib_completed_tx_reclaim_force(iflib_txq_t txq);
746 
747 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
748     SLIST_HEAD_INITIALIZER(cpu_offsets);
749 struct cpu_offset {
750 	SLIST_ENTRY(cpu_offset) entries;
751 	cpuset_t	set;
752 	unsigned int	refcount;
753 	uint16_t	next_cpuid;
754 };
755 static struct mtx cpu_offset_mtx;
756 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
757     MTX_DEF);
758 
759 DEBUGNET_DEFINE(iflib);
760 
761 static int
iflib_num_rx_descs(if_ctx_t ctx)762 iflib_num_rx_descs(if_ctx_t ctx)
763 {
764 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
765 	if_shared_ctx_t sctx = ctx->ifc_sctx;
766 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
767 
768 	return (scctx->isc_nrxd[first_rxq]);
769 }
770 
771 static int
iflib_num_tx_descs(if_ctx_t ctx)772 iflib_num_tx_descs(if_ctx_t ctx)
773 {
774 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
775 	if_shared_ctx_t sctx = ctx->ifc_sctx;
776 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
777 
778 	return (scctx->isc_ntxd[first_txq]);
779 }
780 
781 #ifdef DEV_NETMAP
782 #include <sys/selinfo.h>
783 #include <net/netmap.h>
784 #include <dev/netmap/netmap_kern.h>
785 
786 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
787 
788 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
789 static void iflib_netmap_timer(void *arg);
790 
791 /*
792  * device-specific sysctl variables:
793  *
794  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
795  *	During regular operations the CRC is stripped, but on some
796  *	hardware reception of frames not multiple of 64 is slower,
797  *	so using crcstrip=0 helps in benchmarks.
798  *
799  * iflib_rx_miss, iflib_rx_miss_bufs:
800  *	count packets that might be missed due to lost interrupts.
801  */
802 SYSCTL_DECL(_dev_netmap);
803 /*
804  * The xl driver by default strips CRCs and we do not override it.
805  */
806 
807 int iflib_crcstrip = 1;
808 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
809     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
810 
811 int iflib_rx_miss, iflib_rx_miss_bufs;
812 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
813     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
814 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
815     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
816 
817 /*
818  * Register/unregister. We are already under netmap lock.
819  * Only called on the first register or the last unregister.
820  */
821 static int
iflib_netmap_register(struct netmap_adapter * na,int onoff)822 iflib_netmap_register(struct netmap_adapter *na, int onoff)
823 {
824 	if_t ifp = na->ifp;
825 	if_ctx_t ctx = if_getsoftc(ifp);
826 	int status;
827 
828 	CTX_LOCK(ctx);
829 	if (!CTX_IS_VF(ctx))
830 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
831 
832 	iflib_stop(ctx);
833 
834 	/*
835 	 * Enable (or disable) netmap flags, and intercept (or restore)
836 	 * ifp->if_transmit. This is done once the device has been stopped
837 	 * to prevent race conditions. Also, this must be done after
838 	 * calling netmap_disable_all_rings() and before calling
839 	 * netmap_enable_all_rings(), so that these two functions see the
840 	 * updated state of the NAF_NETMAP_ON bit.
841 	 */
842 	if (onoff) {
843 		nm_set_native_flags(na);
844 	} else {
845 		nm_clear_native_flags(na);
846 	}
847 
848 	iflib_init_locked(ctx);
849 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
850 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
851 	if (status)
852 		nm_clear_native_flags(na);
853 	CTX_UNLOCK(ctx);
854 	return (status);
855 }
856 
857 static int
iflib_netmap_config(struct netmap_adapter * na,struct nm_config_info * info)858 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
859 {
860 	if_t ifp = na->ifp;
861 	if_ctx_t ctx = if_getsoftc(ifp);
862 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
863 	iflib_fl_t fl = &rxq->ifr_fl[0];
864 
865 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
866 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
867 	info->num_tx_descs = iflib_num_tx_descs(ctx);
868 	info->num_rx_descs = iflib_num_rx_descs(ctx);
869 	info->rx_buf_maxsize = fl->ifl_buf_size;
870 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
871 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
872 		info->num_rx_descs, info->rx_buf_maxsize);
873 
874 	return (0);
875 }
876 
877 static int
netmap_fl_refill(iflib_rxq_t rxq,struct netmap_kring * kring,bool init)878 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
879 {
880 	struct netmap_adapter *na = kring->na;
881 	u_int const lim = kring->nkr_num_slots - 1;
882 	struct netmap_ring *ring = kring->ring;
883 	bus_dmamap_t *map;
884 	struct if_rxd_update iru;
885 	if_ctx_t ctx = rxq->ifr_ctx;
886 	iflib_fl_t fl = &rxq->ifr_fl[0];
887 	u_int nic_i_first, nic_i;
888 	u_int nm_i;
889 	int i, n;
890 #if IFLIB_DEBUG_COUNTERS
891 	int rf_count = 0;
892 #endif
893 
894 	/*
895 	 * This function is used both at initialization and in rxsync.
896 	 * At initialization we need to prepare (with isc_rxd_refill())
897 	 * all the netmap buffers currently owned by the kernel, in
898 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
899 	 * (except for kring->nkr_hwofs). These may be less than
900 	 * kring->nkr_num_slots if netmap_reset() was called while
901 	 * an application using the kring that still owned some
902 	 * buffers.
903 	 * At rxsync time, both indexes point to the next buffer to be
904 	 * refilled.
905 	 * In any case we publish (with isc_rxd_flush()) up to
906 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
907 	 * pointer to overrun the head/cons pointer, although this is
908 	 * not necessary for some NICs (e.g. vmx).
909 	 */
910 	if (__predict_false(init)) {
911 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
912 	} else {
913 		n = kring->rhead - kring->nr_hwcur;
914 		if (n == 0)
915 			return (0); /* Nothing to do. */
916 		if (n < 0)
917 			n += kring->nkr_num_slots;
918 	}
919 
920 	iru_init(&iru, rxq, 0 /* flid */);
921 	map = fl->ifl_sds.ifsd_map;
922 	nic_i = fl->ifl_pidx;
923 	nm_i = netmap_idx_n2k(kring, nic_i);
924 	if (__predict_false(init)) {
925 		/*
926 		 * On init/reset, nic_i must be 0, and we must
927 		 * start to refill from hwtail (see netmap_reset()).
928 		 */
929 		MPASS(nic_i == 0);
930 		MPASS(nm_i == kring->nr_hwtail);
931 	} else
932 		MPASS(nm_i == kring->nr_hwcur);
933 	DBG_COUNTER_INC(fl_refills);
934 	while (n > 0) {
935 #if IFLIB_DEBUG_COUNTERS
936 		if (++rf_count == 9)
937 			DBG_COUNTER_INC(fl_refills_large);
938 #endif
939 		nic_i_first = nic_i;
940 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
941 			struct netmap_slot *slot = &ring->slot[nm_i];
942 			uint64_t paddr;
943 			void *addr = PNMB(na, slot, &paddr);
944 
945 			MPASS(i < IFLIB_MAX_RX_REFRESH);
946 
947 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
948 				return (netmap_ring_reinit(kring));
949 
950 			fl->ifl_bus_addrs[i] = paddr +
951 			    nm_get_offset(kring, slot);
952 			fl->ifl_rxd_idxs[i] = nic_i;
953 
954 			if (__predict_false(init)) {
955 				netmap_load_map(na, fl->ifl_buf_tag,
956 				    map[nic_i], addr);
957 			} else if (slot->flags & NS_BUF_CHANGED) {
958 				/* buffer has changed, reload map */
959 				netmap_reload_map(na, fl->ifl_buf_tag,
960 				    map[nic_i], addr);
961 			}
962 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
963 			    BUS_DMASYNC_PREREAD);
964 			slot->flags &= ~NS_BUF_CHANGED;
965 
966 			nm_i = nm_next(nm_i, lim);
967 			nic_i = nm_next(nic_i, lim);
968 		}
969 
970 		iru.iru_pidx = nic_i_first;
971 		iru.iru_count = i;
972 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
973 	}
974 	fl->ifl_pidx = nic_i;
975 	/*
976 	 * At the end of the loop we must have refilled everything
977 	 * we could possibly refill.
978 	 */
979 	MPASS(nm_i == kring->rhead);
980 	kring->nr_hwcur = nm_i;
981 
982 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
983 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
984 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
985 	    nm_prev(nic_i, lim));
986 	DBG_COUNTER_INC(rxd_flush);
987 
988 	return (0);
989 }
990 
991 #define NETMAP_TX_TIMER_US	90
992 
993 /*
994  * Reconcile kernel and user view of the transmit ring.
995  *
996  * All information is in the kring.
997  * Userspace wants to send packets up to the one before kring->rhead,
998  * kernel knows kring->nr_hwcur is the first unsent packet.
999  *
1000  * Here we push packets out (as many as possible), and possibly
1001  * reclaim buffers from previously completed transmission.
1002  *
1003  * The caller (netmap) guarantees that there is only one instance
1004  * running at any time. Any interference with other driver
1005  * methods should be handled by the individual drivers.
1006  */
1007 static int
iflib_netmap_txsync(struct netmap_kring * kring,int flags)1008 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
1009 {
1010 	struct netmap_adapter *na = kring->na;
1011 	if_t ifp = na->ifp;
1012 	struct netmap_ring *ring = kring->ring;
1013 	u_int nm_i;	/* index into the netmap kring */
1014 	u_int nic_i;	/* index into the NIC ring */
1015 	u_int const lim = kring->nkr_num_slots - 1;
1016 	u_int const head = kring->rhead;
1017 	struct if_pkt_info pi;
1018 	int tx_pkts = 0, tx_bytes = 0;
1019 
1020 	/*
1021 	 * interrupts on every tx packet are expensive so request
1022 	 * them every half ring, or where NS_REPORT is set
1023 	 */
1024 	u_int report_frequency = kring->nkr_num_slots >> 1;
1025 	/* device-specific */
1026 	if_ctx_t ctx = if_getsoftc(ifp);
1027 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1028 
1029 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1030 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1031 
1032 	/*
1033 	 * First part: process new packets to send.
1034 	 * nm_i is the current index in the netmap kring,
1035 	 * nic_i is the corresponding index in the NIC ring.
1036 	 *
1037 	 * If we have packets to send (nm_i != head)
1038 	 * iterate over the netmap ring, fetch length and update
1039 	 * the corresponding slot in the NIC ring. Some drivers also
1040 	 * need to update the buffer's physical address in the NIC slot
1041 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1042 	 *
1043 	 * The netmap_reload_map() calls is especially expensive,
1044 	 * even when (as in this case) the tag is 0, so do only
1045 	 * when the buffer has actually changed.
1046 	 *
1047 	 * If possible do not set the report/intr bit on all slots,
1048 	 * but only a few times per ring or when NS_REPORT is set.
1049 	 *
1050 	 * Finally, on 10G and faster drivers, it might be useful
1051 	 * to prefetch the next slot and txr entry.
1052 	 */
1053 
1054 	nm_i = kring->nr_hwcur;
1055 	if (nm_i != head) {	/* we have new packets to send */
1056 		uint32_t pkt_len = 0, seg_idx = 0;
1057 		int nic_i_start = -1, flags = 0;
1058 		memset(&pi, 0, sizeof(pi));
1059 		pi.ipi_segs = txq->ift_segs;
1060 		pi.ipi_qsidx = kring->ring_id;
1061 		nic_i = netmap_idx_k2n(kring, nm_i);
1062 
1063 		__builtin_prefetch(&ring->slot[nm_i]);
1064 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1065 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1066 
1067 		while (nm_i != head) {
1068 			struct netmap_slot *slot = &ring->slot[nm_i];
1069 			uint64_t offset = nm_get_offset(kring, slot);
1070 			u_int len = slot->len;
1071 			uint64_t paddr;
1072 			void *addr = PNMB(na, slot, &paddr);
1073 
1074 			flags |= (slot->flags & NS_REPORT ||
1075 				nic_i == 0 || nic_i == report_frequency) ?
1076 				IPI_TX_INTR : 0;
1077 
1078 			/*
1079 			 * If this is the first packet fragment, save the
1080 			 * index of the first NIC slot for later.
1081 			 */
1082 			if (nic_i_start < 0)
1083 				nic_i_start = nic_i;
1084 
1085 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1086 			pi.ipi_segs[seg_idx].ds_len = len;
1087 			if (len) {
1088 				pkt_len += len;
1089 				seg_idx++;
1090 			}
1091 
1092 			if (!(slot->flags & NS_MOREFRAG)) {
1093 				pi.ipi_len = pkt_len;
1094 				pi.ipi_nsegs = seg_idx;
1095 				pi.ipi_pidx = nic_i_start;
1096 				pi.ipi_ndescs = 0;
1097 				pi.ipi_flags = flags;
1098 
1099 				/* Prepare the NIC TX ring. */
1100 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1101 				DBG_COUNTER_INC(tx_encap);
1102 
1103 				/* Update transmit counters */
1104 				tx_bytes += pi.ipi_len;
1105 				tx_pkts++;
1106 
1107 				/* Reinit per-packet info for the next one. */
1108 				flags = seg_idx = pkt_len = 0;
1109 				nic_i_start = -1;
1110 			}
1111 
1112 			/* prefetch for next round */
1113 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1114 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1115 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1116 
1117 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1118 
1119 			if (slot->flags & NS_BUF_CHANGED) {
1120 				/* buffer has changed, reload map */
1121 				netmap_reload_map(na, txq->ift_buf_tag,
1122 				    txq->ift_sds.ifsd_map[nic_i], addr);
1123 			}
1124 			/* make sure changes to the buffer are synced */
1125 			bus_dmamap_sync(txq->ift_buf_tag,
1126 			    txq->ift_sds.ifsd_map[nic_i],
1127 			    BUS_DMASYNC_PREWRITE);
1128 
1129 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1130 			nm_i = nm_next(nm_i, lim);
1131 			nic_i = nm_next(nic_i, lim);
1132 		}
1133 		kring->nr_hwcur = nm_i;
1134 
1135 		/* synchronize the NIC ring */
1136 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1137 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1138 
1139 		/* (re)start the tx unit up to slot nic_i (excluded) */
1140 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1141 	}
1142 
1143 	/*
1144 	 * Second part: reclaim buffers for completed transmissions.
1145 	 *
1146 	 * If there are unclaimed buffers, attempt to reclaim them.
1147 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1148 	 * trigger a per-tx-queue timer to try again later.
1149 	 */
1150 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1151 		if (iflib_tx_credits_update(ctx, txq)) {
1152 			/* some tx completed, increment avail */
1153 			nic_i = txq->ift_cidx_processed;
1154 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1155 		}
1156 	}
1157 
1158 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1159 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1160 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1161 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1162 			    iflib_netmap_timer, txq,
1163 			    txq->ift_netmap_timer.c_cpu, 0);
1164 		}
1165 
1166 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1167 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1168 
1169 	return (0);
1170 }
1171 
1172 /*
1173  * Reconcile kernel and user view of the receive ring.
1174  * Same as for the txsync, this routine must be efficient.
1175  * The caller guarantees a single invocations, but races against
1176  * the rest of the driver should be handled here.
1177  *
1178  * On call, kring->rhead is the first packet that userspace wants
1179  * to keep, and kring->rcur is the wakeup point.
1180  * The kernel has previously reported packets up to kring->rtail.
1181  *
1182  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1183  * of whether or not we received an interrupt.
1184  */
1185 static int
iflib_netmap_rxsync(struct netmap_kring * kring,int flags)1186 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1187 {
1188 	struct netmap_adapter *na = kring->na;
1189 	struct netmap_ring *ring = kring->ring;
1190 	if_t ifp = na->ifp;
1191 	uint32_t nm_i;	/* index into the netmap ring */
1192 	uint32_t nic_i;	/* index into the NIC ring */
1193 	u_int n;
1194 	u_int const lim = kring->nkr_num_slots - 1;
1195 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1196 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1197 
1198 	if_ctx_t ctx = if_getsoftc(ifp);
1199 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1200 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1201 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1202 	iflib_fl_t fl = &rxq->ifr_fl[0];
1203 	struct if_rxd_info ri;
1204 	qidx_t *cidxp;
1205 
1206 	/*
1207 	 * netmap only uses free list 0, to avoid out of order consumption
1208 	 * of receive buffers
1209 	 */
1210 
1211 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1212 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1213 
1214 	/*
1215 	 * First part: import newly received packets.
1216 	 *
1217 	 * nm_i is the index of the next free slot in the netmap ring,
1218 	 * nic_i is the index of the next received packet in the NIC ring
1219 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1220 	 * differ in case if_init() has been called while
1221 	 * in netmap mode. For the receive ring we have
1222 	 *
1223 	 *	nic_i = fl->ifl_cidx;
1224 	 *	nm_i = kring->nr_hwtail (previous)
1225 	 * and
1226 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1227 	 *
1228 	 * fl->ifl_cidx is set to 0 on a ring reinit
1229 	 */
1230 	if (netmap_no_pendintr || force_update) {
1231 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1232 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1233 		int crclen = iflib_crcstrip ? 0 : 4;
1234 		int error, avail;
1235 
1236 		/*
1237 		 * For the free list consumer index, we use the same
1238 		 * logic as in iflib_rxeof().
1239 		 */
1240 		if (have_rxcq)
1241 			cidxp = &rxq->ifr_cq_cidx;
1242 		else
1243 			cidxp = &fl->ifl_cidx;
1244 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1245 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1246 
1247 		nic_i = fl->ifl_cidx;
1248 		nm_i = netmap_idx_n2k(kring, nic_i);
1249 		MPASS(nm_i == kring->nr_hwtail);
1250 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1251 			memset(&ri, 0, sizeof(ri));
1252 			ri.iri_frags = rxq->ifr_frags;
1253 			ri.iri_qsidx = kring->ring_id;
1254 			ri.iri_ifp = ctx->ifc_ifp;
1255 			ri.iri_cidx = *cidxp;
1256 
1257 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1258 			for (i = 0; i < ri.iri_nfrags; i++) {
1259 				if (error) {
1260 					ring->slot[nm_i].len = 0;
1261 					ring->slot[nm_i].flags = 0;
1262 				} else {
1263 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1264 					if (i == (ri.iri_nfrags - 1)) {
1265 						ring->slot[nm_i].len -= crclen;
1266 						ring->slot[nm_i].flags = 0;
1267 
1268 						/* Update receive counters */
1269 						rx_bytes += ri.iri_len;
1270 						rx_pkts++;
1271 					} else
1272 						ring->slot[nm_i].flags = NS_MOREFRAG;
1273 				}
1274 
1275 				bus_dmamap_sync(fl->ifl_buf_tag,
1276 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1277 				nm_i = nm_next(nm_i, lim);
1278 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1279 			}
1280 
1281 			if (have_rxcq) {
1282 				*cidxp = ri.iri_cidx;
1283 				while (*cidxp >= scctx->isc_nrxd[0])
1284 					*cidxp -= scctx->isc_nrxd[0];
1285 			}
1286 
1287 		}
1288 		if (n) { /* update the state variables */
1289 			if (netmap_no_pendintr && !force_update) {
1290 				/* diagnostics */
1291 				iflib_rx_miss++;
1292 				iflib_rx_miss_bufs += n;
1293 			}
1294 			kring->nr_hwtail = nm_i;
1295 		}
1296 		kring->nr_kflags &= ~NKR_PENDINTR;
1297 	}
1298 	/*
1299 	 * Second part: skip past packets that userspace has released.
1300 	 * (kring->nr_hwcur to head excluded),
1301 	 * and make the buffers available for reception.
1302 	 * As usual nm_i is the index in the netmap ring,
1303 	 * nic_i is the index in the NIC ring, and
1304 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1305 	 */
1306 	netmap_fl_refill(rxq, kring, false);
1307 
1308 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1309 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1310 
1311 	return (0);
1312 }
1313 
1314 static void
iflib_netmap_intr(struct netmap_adapter * na,int onoff)1315 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1316 {
1317 	if_ctx_t ctx = if_getsoftc(na->ifp);
1318 
1319 	CTX_LOCK(ctx);
1320 	if (onoff) {
1321 		IFDI_INTR_ENABLE(ctx);
1322 	} else {
1323 		IFDI_INTR_DISABLE(ctx);
1324 	}
1325 	CTX_UNLOCK(ctx);
1326 }
1327 
1328 static int
iflib_netmap_attach(if_ctx_t ctx)1329 iflib_netmap_attach(if_ctx_t ctx)
1330 {
1331 	struct netmap_adapter na;
1332 
1333 	bzero(&na, sizeof(na));
1334 
1335 	na.ifp = ctx->ifc_ifp;
1336 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1337 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1338 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1339 
1340 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1341 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1342 	na.nm_txsync = iflib_netmap_txsync;
1343 	na.nm_rxsync = iflib_netmap_rxsync;
1344 	na.nm_register = iflib_netmap_register;
1345 	na.nm_intr = iflib_netmap_intr;
1346 	na.nm_config = iflib_netmap_config;
1347 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1348 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1349 	return (netmap_attach(&na));
1350 }
1351 
1352 static int
iflib_netmap_txq_init(if_ctx_t ctx,iflib_txq_t txq)1353 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1354 {
1355 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1356 	struct netmap_slot *slot;
1357 
1358 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1359 	if (slot == NULL)
1360 		return (0);
1361 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1362 		/*
1363 		 * In netmap mode, set the map for the packet buffer.
1364 		 * NOTE: Some drivers (not this one) also need to set
1365 		 * the physical buffer address in the NIC ring.
1366 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1367 		 * netmap slot index, si
1368 		 */
1369 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1370 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1371 		    NMB(na, slot + si));
1372 	}
1373 	return (1);
1374 }
1375 
1376 static int
iflib_netmap_rxq_init(if_ctx_t ctx,iflib_rxq_t rxq)1377 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1378 {
1379 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1380 	struct netmap_kring *kring;
1381 	struct netmap_slot *slot;
1382 
1383 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1384 	if (slot == NULL)
1385 		return (0);
1386 	kring = na->rx_rings[rxq->ifr_id];
1387 	netmap_fl_refill(rxq, kring, true);
1388 	return (1);
1389 }
1390 
1391 static void
iflib_netmap_timer(void * arg)1392 iflib_netmap_timer(void *arg)
1393 {
1394 	iflib_txq_t txq = arg;
1395 	if_ctx_t ctx = txq->ift_ctx;
1396 
1397 	/*
1398 	 * Wake up the netmap application, to give it a chance to
1399 	 * call txsync and reclaim more completed TX buffers.
1400 	 */
1401 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1402 }
1403 
1404 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1405 
1406 #else
1407 #define iflib_netmap_txq_init(ctx, txq) (0)
1408 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1409 #define iflib_netmap_detach(ifp)
1410 #define netmap_enable_all_rings(ifp)
1411 #define netmap_disable_all_rings(ifp)
1412 
1413 #define iflib_netmap_attach(ctx) (0)
1414 #define netmap_rx_irq(ifp, qid, budget) (0)
1415 #endif
1416 
1417 #if defined(__i386__) || defined(__amd64__)
1418 static __inline void
prefetch(void * x)1419 prefetch(void *x)
1420 {
1421 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1422 }
1423 
1424 static __inline void
prefetch2cachelines(void * x)1425 prefetch2cachelines(void *x)
1426 {
1427 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1428 #if (CACHE_LINE_SIZE < 128)
1429 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long)))));
1430 #endif
1431 }
1432 #else
1433 static __inline void
prefetch(void * x)1434 prefetch(void *x)
1435 {
1436 }
1437 
1438 static __inline void
prefetch2cachelines(void * x)1439 prefetch2cachelines(void *x)
1440 {
1441 }
1442 #endif
1443 
1444 static void
iru_init(if_rxd_update_t iru,iflib_rxq_t rxq,uint8_t flid)1445 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1446 {
1447 	iflib_fl_t fl;
1448 
1449 	fl = &rxq->ifr_fl[flid];
1450 	iru->iru_paddrs = fl->ifl_bus_addrs;
1451 	iru->iru_idxs = fl->ifl_rxd_idxs;
1452 	iru->iru_qsidx = rxq->ifr_id;
1453 	iru->iru_buf_size = fl->ifl_buf_size;
1454 	iru->iru_flidx = fl->ifl_id;
1455 }
1456 
1457 static void
_iflib_dmamap_cb(void * arg,bus_dma_segment_t * segs,int nseg,int err)1458 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1459 {
1460 	if (err)
1461 		return;
1462 	*(bus_addr_t *) arg = segs[0].ds_addr;
1463 }
1464 
1465 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1466 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1467 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1468 
1469 int
iflib_dma_alloc_align(if_ctx_t ctx,int size,int align,iflib_dma_info_t dma,int mapflags)1470 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1471 {
1472 	int err;
1473 	device_t dev = ctx->ifc_dev;
1474 	bus_addr_t lowaddr;
1475 
1476 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1477 
1478 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1479 		    align, 0,		/* alignment, bounds */
1480 		    lowaddr,		/* lowaddr */
1481 		    BUS_SPACE_MAXADDR,	/* highaddr */
1482 		    NULL, NULL,		/* filter, filterarg */
1483 		    size,		/* maxsize */
1484 		    1,			/* nsegments */
1485 		    size,		/* maxsegsize */
1486 		    BUS_DMA_ALLOCNOW,	/* flags */
1487 		    NULL,		/* lockfunc */
1488 		    NULL,		/* lockarg */
1489 		    &dma->idi_tag);
1490 	if (err) {
1491 		device_printf(dev,
1492 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1493 		    __func__, err, size, align);
1494 		goto fail_0;
1495 	}
1496 
1497 	err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr,
1498 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1499 	if (err) {
1500 		device_printf(dev,
1501 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1502 		    __func__, (uintmax_t)size, err);
1503 		goto fail_1;
1504 	}
1505 
1506 	dma->idi_paddr = IF_BAD_DMA;
1507 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1508 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1509 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1510 		device_printf(dev,
1511 		    "%s: bus_dmamap_load failed: %d\n",
1512 		    __func__, err);
1513 		goto fail_2;
1514 	}
1515 
1516 	dma->idi_size = size;
1517 	return (0);
1518 
1519 fail_2:
1520 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1521 fail_1:
1522 	bus_dma_tag_destroy(dma->idi_tag);
1523 fail_0:
1524 	dma->idi_tag = NULL;
1525 
1526 	return (err);
1527 }
1528 
1529 int
iflib_dma_alloc(if_ctx_t ctx,int size,iflib_dma_info_t dma,int mapflags)1530 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1531 {
1532 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1533 
1534 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1535 
1536 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1537 }
1538 
1539 int
iflib_dma_alloc_multi(if_ctx_t ctx,int * sizes,iflib_dma_info_t * dmalist,int mapflags,int count)1540 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1541 {
1542 	int i, err;
1543 	iflib_dma_info_t *dmaiter;
1544 
1545 	dmaiter = dmalist;
1546 	for (i = 0; i < count; i++, dmaiter++) {
1547 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1548 			break;
1549 	}
1550 	if (err)
1551 		iflib_dma_free_multi(dmalist, i);
1552 	return (err);
1553 }
1554 
1555 void
iflib_dma_free(iflib_dma_info_t dma)1556 iflib_dma_free(iflib_dma_info_t dma)
1557 {
1558 	if (dma->idi_tag == NULL)
1559 		return;
1560 	if (dma->idi_paddr != IF_BAD_DMA) {
1561 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1562 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1563 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1564 		dma->idi_paddr = IF_BAD_DMA;
1565 	}
1566 	if (dma->idi_vaddr != NULL) {
1567 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1568 		dma->idi_vaddr = NULL;
1569 	}
1570 	bus_dma_tag_destroy(dma->idi_tag);
1571 	dma->idi_tag = NULL;
1572 }
1573 
1574 void
iflib_dma_free_multi(iflib_dma_info_t * dmalist,int count)1575 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1576 {
1577 	int i;
1578 	iflib_dma_info_t *dmaiter = dmalist;
1579 
1580 	for (i = 0; i < count; i++, dmaiter++)
1581 		iflib_dma_free(*dmaiter);
1582 }
1583 
1584 static int
iflib_fast_intr(void * arg)1585 iflib_fast_intr(void *arg)
1586 {
1587 	iflib_filter_info_t info = arg;
1588 	struct grouptask *gtask = info->ifi_task;
1589 	int result;
1590 
1591 	DBG_COUNTER_INC(fast_intrs);
1592 	if (info->ifi_filter != NULL) {
1593 		result = info->ifi_filter(info->ifi_filter_arg);
1594 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1595 			return (result);
1596 	}
1597 
1598 	GROUPTASK_ENQUEUE(gtask);
1599 	return (FILTER_HANDLED);
1600 }
1601 
1602 static int
iflib_fast_intr_rxtx(void * arg)1603 iflib_fast_intr_rxtx(void *arg)
1604 {
1605 	iflib_filter_info_t info = arg;
1606 	struct grouptask *gtask = info->ifi_task;
1607 	if_ctx_t ctx;
1608 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1609 	iflib_txq_t txq;
1610 	void *sc;
1611 	int i, cidx, result;
1612 	qidx_t txqid;
1613 	bool intr_enable, intr_legacy;
1614 
1615 	DBG_COUNTER_INC(fast_intrs);
1616 	if (info->ifi_filter != NULL) {
1617 		result = info->ifi_filter(info->ifi_filter_arg);
1618 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1619 			return (result);
1620 	}
1621 
1622 	ctx = rxq->ifr_ctx;
1623 	sc = ctx->ifc_softc;
1624 	intr_enable = false;
1625 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1626 	MPASS(rxq->ifr_ntxqirq);
1627 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1628 		txqid = rxq->ifr_txqid[i];
1629 		txq = &ctx->ifc_txqs[txqid];
1630 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1631 		    BUS_DMASYNC_POSTREAD);
1632 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1633 			if (intr_legacy)
1634 				intr_enable = true;
1635 			else
1636 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1637 			continue;
1638 		}
1639 		GROUPTASK_ENQUEUE(&txq->ift_task);
1640 	}
1641 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1642 		cidx = rxq->ifr_cq_cidx;
1643 	else
1644 		cidx = rxq->ifr_fl[0].ifl_cidx;
1645 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1646 		GROUPTASK_ENQUEUE(gtask);
1647 	else {
1648 		if (intr_legacy)
1649 			intr_enable = true;
1650 		else
1651 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1652 		DBG_COUNTER_INC(rx_intr_enables);
1653 	}
1654 	if (intr_enable)
1655 		IFDI_INTR_ENABLE(ctx);
1656 	return (FILTER_HANDLED);
1657 }
1658 
1659 static int
iflib_fast_intr_ctx(void * arg)1660 iflib_fast_intr_ctx(void *arg)
1661 {
1662 	iflib_filter_info_t info = arg;
1663 	if_ctx_t ctx = info->ifi_ctx;
1664 	int result;
1665 
1666 	DBG_COUNTER_INC(fast_intrs);
1667 	if (info->ifi_filter != NULL) {
1668 		result = info->ifi_filter(info->ifi_filter_arg);
1669 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1670 			return (result);
1671 	}
1672 
1673 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
1674 	return (FILTER_HANDLED);
1675 }
1676 
1677 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)1678 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1679 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1680 		 const char *name)
1681 {
1682 	struct resource *res;
1683 	void *tag = NULL;
1684 	device_t dev = ctx->ifc_dev;
1685 	int flags, i, rc;
1686 
1687 	flags = RF_ACTIVE;
1688 	if (ctx->ifc_flags & IFC_LEGACY)
1689 		flags |= RF_SHAREABLE;
1690 	MPASS(rid < 512);
1691 	i = rid;
1692 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1693 	if (res == NULL) {
1694 		device_printf(dev,
1695 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1696 		return (ENOMEM);
1697 	}
1698 	irq->ii_res = res;
1699 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1700 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1701 		    filter, handler, arg, &tag);
1702 	if (rc != 0) {
1703 		device_printf(dev,
1704 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1705 		    rid, name ? name : "unknown", rc);
1706 		return (rc);
1707 	} else if (name)
1708 		bus_describe_intr(dev, res, tag, "%s", name);
1709 
1710 	irq->ii_tag = tag;
1711 	return (0);
1712 }
1713 
1714 /*********************************************************************
1715  *
1716  *  Allocate DMA resources for TX buffers as well as memory for the TX
1717  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1718  *  iflib_sw_tx_desc_array structure, storing all the information that
1719  *  is needed to transmit a packet on the wire.  This is called only
1720  *  once at attach, setup is done every reset.
1721  *
1722  **********************************************************************/
1723 static int
iflib_txsd_alloc(iflib_txq_t txq)1724 iflib_txsd_alloc(iflib_txq_t txq)
1725 {
1726 	if_ctx_t ctx = txq->ift_ctx;
1727 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1728 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1729 	device_t dev = ctx->ifc_dev;
1730 	bus_size_t tsomaxsize;
1731 	bus_addr_t lowaddr;
1732 	int err, nsegments, ntsosegments;
1733 	bool tso;
1734 
1735 	nsegments = scctx->isc_tx_nsegments;
1736 	ntsosegments = scctx->isc_tx_tso_segments_max;
1737 	tsomaxsize = scctx->isc_tx_tso_size_max;
1738 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1739 		tsomaxsize += sizeof(struct ether_vlan_header);
1740 	MPASS(scctx->isc_ntxd[0] > 0);
1741 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1742 	MPASS(nsegments > 0);
1743 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1744 		MPASS(ntsosegments > 0);
1745 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1746 	}
1747 
1748 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1749 
1750 	/*
1751 	 * Set up DMA tags for TX buffers.
1752 	 */
1753 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1754 		    1, 0,			/* alignment, bounds */
1755 		    lowaddr,			/* lowaddr */
1756 		    BUS_SPACE_MAXADDR,		/* highaddr */
1757 		    NULL, NULL,			/* filter, filterarg */
1758 		    sctx->isc_tx_maxsize,	/* maxsize */
1759 		    nsegments,			/* nsegments */
1760 		    sctx->isc_tx_maxsegsize,	/* maxsegsize */
1761 		    0,				/* flags */
1762 		    NULL,			/* lockfunc */
1763 		    NULL,			/* lockfuncarg */
1764 		    &txq->ift_buf_tag))) {
1765 		device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err);
1766 		device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1767 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1768 		goto fail;
1769 	}
1770 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1771 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1772 		    1, 0,			/* alignment, bounds */
1773 		    lowaddr,			/* lowaddr */
1774 		    BUS_SPACE_MAXADDR,		/* highaddr */
1775 		    NULL, NULL,			/* filter, filterarg */
1776 		    tsomaxsize,			/* maxsize */
1777 		    ntsosegments,		/* nsegments */
1778 		    sctx->isc_tso_maxsegsize,	/* maxsegsize */
1779 		    0,				/* flags */
1780 		    NULL,			/* lockfunc */
1781 		    NULL,			/* lockfuncarg */
1782 		    &txq->ift_tso_buf_tag))) {
1783 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1784 		    err);
1785 		goto fail;
1786 	}
1787 
1788 	/* Allocate memory for the TX mbuf map. */
1789 	if (!(txq->ift_sds.ifsd_m =
1790 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1791 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1792 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1793 		err = ENOMEM;
1794 		goto fail;
1795 	}
1796 	if (ctx->ifc_sysctl_simple_tx) {
1797 		if (!(txq->ift_sds.ifsd_m_defer =
1798 			(struct mbuf **) malloc(sizeof(struct mbuf *) *
1799 			    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1800 			device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1801 			err = ENOMEM;
1802 			goto fail;
1803 		}
1804 	}
1805 	txq->ift_sds.ifsd_m_deferb = txq->ift_sds.ifsd_m_defer;
1806 	/*
1807 	 * Create the DMA maps for TX buffers.
1808 	 */
1809 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1810 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1811 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1812 		device_printf(dev,
1813 		    "Unable to allocate TX buffer DMA map memory\n");
1814 		err = ENOMEM;
1815 		goto fail;
1816 	}
1817 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1818 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1819 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1820 		device_printf(dev,
1821 		    "Unable to allocate TSO TX buffer map memory\n");
1822 		err = ENOMEM;
1823 		goto fail;
1824 	}
1825 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1826 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1827 		    &txq->ift_sds.ifsd_map[i]);
1828 		if (err != 0) {
1829 			device_printf(dev, "Unable to create TX DMA map\n");
1830 			goto fail;
1831 		}
1832 		if (!tso)
1833 			continue;
1834 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1835 		    &txq->ift_sds.ifsd_tso_map[i]);
1836 		if (err != 0) {
1837 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1838 			goto fail;
1839 		}
1840 	}
1841 	return (0);
1842 fail:
1843 	/* We free all, it handles case where we are in the middle */
1844 	iflib_tx_structures_free(ctx);
1845 	return (err);
1846 }
1847 
1848 static void
iflib_txsd_destroy(if_ctx_t ctx,iflib_txq_t txq,int i)1849 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1850 {
1851 	bus_dmamap_t map;
1852 
1853 	if (txq->ift_sds.ifsd_map != NULL) {
1854 		map = txq->ift_sds.ifsd_map[i];
1855 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1856 		bus_dmamap_unload(txq->ift_buf_tag, map);
1857 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1858 		txq->ift_sds.ifsd_map[i] = NULL;
1859 	}
1860 
1861 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1862 		map = txq->ift_sds.ifsd_tso_map[i];
1863 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1864 		    BUS_DMASYNC_POSTWRITE);
1865 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1866 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1867 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1868 	}
1869 }
1870 
1871 static void
iflib_txq_destroy(iflib_txq_t txq)1872 iflib_txq_destroy(iflib_txq_t txq)
1873 {
1874 	if_ctx_t ctx = txq->ift_ctx;
1875 
1876 	for (int i = 0; i < txq->ift_size; i++)
1877 		iflib_txsd_destroy(ctx, txq, i);
1878 
1879 	if (txq->ift_br != NULL) {
1880 		ifmp_ring_free(txq->ift_br);
1881 		txq->ift_br = NULL;
1882 	}
1883 
1884 	mtx_destroy(&txq->ift_mtx);
1885 
1886 	if (txq->ift_sds.ifsd_map != NULL) {
1887 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1888 		txq->ift_sds.ifsd_map = NULL;
1889 	}
1890 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1891 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1892 		txq->ift_sds.ifsd_tso_map = NULL;
1893 	}
1894 	if (txq->ift_sds.ifsd_m != NULL) {
1895 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1896 		txq->ift_sds.ifsd_m = NULL;
1897 	}
1898 	if (txq->ift_sds.ifsd_m_defer != NULL) {
1899 		free(txq->ift_sds.ifsd_m_defer, M_IFLIB);
1900 		txq->ift_sds.ifsd_m_defer = NULL;
1901 	}
1902 	if (txq->ift_buf_tag != NULL) {
1903 		bus_dma_tag_destroy(txq->ift_buf_tag);
1904 		txq->ift_buf_tag = NULL;
1905 	}
1906 	if (txq->ift_tso_buf_tag != NULL) {
1907 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1908 		txq->ift_tso_buf_tag = NULL;
1909 	}
1910 	if (txq->ift_ifdi != NULL) {
1911 		free(txq->ift_ifdi, M_IFLIB);
1912 	}
1913 }
1914 
1915 static void
iflib_txsd_free(if_ctx_t ctx,iflib_txq_t txq,int i)1916 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1917 {
1918 	struct mbuf *m;
1919 
1920 	m = IFLIB_GET_MBUF(txq->ift_sds.ifsd_m[i]);
1921 	if (m == NULL)
1922 		return;
1923 
1924 	if (txq->ift_sds.ifsd_map != NULL) {
1925 		bus_dmamap_sync(txq->ift_buf_tag,
1926 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1927 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1928 	}
1929 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1930 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1931 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1932 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1933 		    txq->ift_sds.ifsd_tso_map[i]);
1934 	}
1935 	txq->ift_sds.ifsd_m[i] = NULL;
1936 	m_freem(m);
1937 	DBG_COUNTER_INC(tx_frees);
1938 }
1939 
1940 static int
iflib_txq_setup(iflib_txq_t txq)1941 iflib_txq_setup(iflib_txq_t txq)
1942 {
1943 	if_ctx_t ctx = txq->ift_ctx;
1944 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1945 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1946 	iflib_dma_info_t di;
1947 	int i;
1948 
1949 	/* XXX make configurable */
1950 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1951 
1952 	/* Reset indices */
1953 	txq->ift_cidx_processed = 0;
1954 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1955 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1956 	txq->ift_pad = scctx->isc_tx_pad;
1957 
1958 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1959 		bzero((void *)di->idi_vaddr, di->idi_size);
1960 
1961 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1962 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1963 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1964 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1965 	return (0);
1966 }
1967 
1968 /*********************************************************************
1969  *
1970  *  Allocate DMA resources for RX buffers as well as memory for the RX
1971  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1972  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1973  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1974  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1975  *  packet, the maximum number of entries we'll need is equal to the
1976  *  number of hardware receive descriptors that we've allocated.
1977  *
1978  **********************************************************************/
1979 static int
iflib_rxsd_alloc(iflib_rxq_t rxq)1980 iflib_rxsd_alloc(iflib_rxq_t rxq)
1981 {
1982 	if_ctx_t ctx = rxq->ifr_ctx;
1983 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1984 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1985 	device_t dev = ctx->ifc_dev;
1986 	iflib_fl_t fl;
1987 	bus_addr_t lowaddr;
1988 	int err;
1989 
1990 	MPASS(scctx->isc_nrxd[0] > 0);
1991 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1992 
1993 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1994 
1995 	fl = rxq->ifr_fl;
1996 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1997 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1998 		/* Set up DMA tag for RX buffers. */
1999 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
2000 			    1, 0,			/* alignment, bounds */
2001 			    lowaddr,			/* lowaddr */
2002 			    BUS_SPACE_MAXADDR,		/* highaddr */
2003 			    NULL, NULL,			/* filter, filterarg */
2004 			    sctx->isc_rx_maxsize,	/* maxsize */
2005 			    sctx->isc_rx_nsegments,	/* nsegments */
2006 			    sctx->isc_rx_maxsegsize,	/* maxsegsize */
2007 			    0,				/* flags */
2008 			    NULL,			/* lockfunc */
2009 			    NULL,			/* lockarg */
2010 			    &fl->ifl_buf_tag);
2011 		if (err) {
2012 			device_printf(dev,
2013 			    "Unable to allocate RX DMA tag: %d\n", err);
2014 			goto fail;
2015 		}
2016 
2017 		/* Allocate memory for the RX mbuf map. */
2018 		if (!(fl->ifl_sds.ifsd_m =
2019 		    (struct mbuf **) malloc(sizeof(struct mbuf *) *
2020 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2021 			device_printf(dev,
2022 			    "Unable to allocate RX mbuf map memory\n");
2023 			err = ENOMEM;
2024 			goto fail;
2025 		}
2026 
2027 		/* Allocate memory for the direct RX cluster pointer map. */
2028 		if (!(fl->ifl_sds.ifsd_cl =
2029 		    (caddr_t *) malloc(sizeof(caddr_t) *
2030 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2031 			device_printf(dev,
2032 			    "Unable to allocate RX cluster map memory\n");
2033 			err = ENOMEM;
2034 			goto fail;
2035 		}
2036 
2037 		/* Allocate memory for the RX cluster bus address map. */
2038 		if (!(fl->ifl_sds.ifsd_ba =
2039 		    (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2040 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2041 			device_printf(dev,
2042 			    "Unable to allocate RX bus address map memory\n");
2043 			err = ENOMEM;
2044 			goto fail;
2045 		}
2046 
2047 		/*
2048 		 * Create the DMA maps for RX buffers.
2049 		 */
2050 		if (!(fl->ifl_sds.ifsd_map =
2051 		    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2052 			device_printf(dev,
2053 			    "Unable to allocate RX buffer DMA map memory\n");
2054 			err = ENOMEM;
2055 			goto fail;
2056 		}
2057 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2058 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2059 			    &fl->ifl_sds.ifsd_map[i]);
2060 			if (err != 0) {
2061 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2062 				goto fail;
2063 			}
2064 		}
2065 	}
2066 	return (0);
2067 
2068 fail:
2069 	iflib_rx_structures_free(ctx);
2070 	return (err);
2071 }
2072 
2073 /*
2074  * Internal service routines
2075  */
2076 
2077 struct rxq_refill_cb_arg {
2078 	int               error;
2079 	bus_dma_segment_t seg;
2080 	int               nseg;
2081 };
2082 
2083 static void
_rxq_refill_cb(void * arg,bus_dma_segment_t * segs,int nseg,int error)2084 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2085 {
2086 	struct rxq_refill_cb_arg *cb_arg = arg;
2087 
2088 	cb_arg->error = error;
2089 	cb_arg->seg = segs[0];
2090 	cb_arg->nseg = nseg;
2091 }
2092 
2093 /**
2094  * iflib_fl_refill - refill an rxq free-buffer list
2095  * @ctx: the iflib context
2096  * @fl: the free list to refill
2097  * @count: the number of new buffers to allocate
2098  *
2099  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2100  * The caller must assure that @count does not exceed the queue's capacity
2101  * minus one (since we always leave a descriptor unavailable).
2102  */
2103 static uint8_t
iflib_fl_refill(if_ctx_t ctx,iflib_fl_t fl,int count)2104 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2105 {
2106 	struct if_rxd_update iru;
2107 	struct rxq_refill_cb_arg cb_arg;
2108 	struct mbuf *m;
2109 	caddr_t cl, *sd_cl;
2110 	struct mbuf **sd_m;
2111 	bus_dmamap_t *sd_map;
2112 	bus_addr_t bus_addr, *sd_ba;
2113 	int err, frag_idx, i, idx, n, pidx;
2114 	qidx_t credits;
2115 
2116 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2117 
2118 	sd_m = fl->ifl_sds.ifsd_m;
2119 	sd_map = fl->ifl_sds.ifsd_map;
2120 	sd_cl = fl->ifl_sds.ifsd_cl;
2121 	sd_ba = fl->ifl_sds.ifsd_ba;
2122 	pidx = fl->ifl_pidx;
2123 	idx = pidx;
2124 	frag_idx = fl->ifl_fragidx;
2125 	credits = fl->ifl_credits;
2126 
2127 	i = 0;
2128 	n = count;
2129 	MPASS(n > 0);
2130 	MPASS(credits + n <= fl->ifl_size);
2131 
2132 	if (pidx < fl->ifl_cidx)
2133 		MPASS(pidx + n <= fl->ifl_cidx);
2134 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2135 		MPASS(fl->ifl_gen == 0);
2136 	if (pidx > fl->ifl_cidx)
2137 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2138 
2139 	DBG_COUNTER_INC(fl_refills);
2140 	if (n > 8)
2141 		DBG_COUNTER_INC(fl_refills_large);
2142 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2143 	while (n-- > 0) {
2144 		/*
2145 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2146 		 * initialized after rx.
2147 		 *
2148 		 * If the cluster is still set then we know a minimum sized
2149 		 * packet was received
2150 		 */
2151 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2152 		    &frag_idx);
2153 		if (frag_idx < 0)
2154 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2155 		MPASS(frag_idx >= 0);
2156 		if ((cl = sd_cl[frag_idx]) == NULL) {
2157 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2158 			if (__predict_false(cl == NULL))
2159 				break;
2160 
2161 			cb_arg.error = 0;
2162 			MPASS(sd_map != NULL);
2163 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2164 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2165 			    BUS_DMA_NOWAIT);
2166 			if (__predict_false(err != 0 || cb_arg.error)) {
2167 				uma_zfree(fl->ifl_zone, cl);
2168 				break;
2169 			}
2170 
2171 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2172 			sd_cl[frag_idx] = cl;
2173 #if MEMORY_LOGGING
2174 			fl->ifl_cl_enqueued++;
2175 #endif
2176 		} else {
2177 			bus_addr = sd_ba[frag_idx];
2178 		}
2179 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2180 		    BUS_DMASYNC_PREREAD);
2181 
2182 		if (sd_m[frag_idx] == NULL) {
2183 			m = m_gethdr_raw(M_NOWAIT, 0);
2184 			if (__predict_false(m == NULL))
2185 				break;
2186 			sd_m[frag_idx] = m;
2187 		}
2188 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2189 #if MEMORY_LOGGING
2190 		fl->ifl_m_enqueued++;
2191 #endif
2192 
2193 		DBG_COUNTER_INC(rx_allocs);
2194 		fl->ifl_rxd_idxs[i] = frag_idx;
2195 		fl->ifl_bus_addrs[i] = bus_addr;
2196 		credits++;
2197 		i++;
2198 		MPASS(credits <= fl->ifl_size);
2199 		if (++idx == fl->ifl_size) {
2200 #ifdef INVARIANTS
2201 			fl->ifl_gen = 1;
2202 #endif
2203 			idx = 0;
2204 		}
2205 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2206 			iru.iru_pidx = pidx;
2207 			iru.iru_count = i;
2208 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2209 			fl->ifl_pidx = idx;
2210 			fl->ifl_credits = credits;
2211 			pidx = idx;
2212 			i = 0;
2213 		}
2214 	}
2215 
2216 	if (n < count - 1) {
2217 		if (i != 0) {
2218 			iru.iru_pidx = pidx;
2219 			iru.iru_count = i;
2220 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2221 			fl->ifl_pidx = idx;
2222 			fl->ifl_credits = credits;
2223 		}
2224 		DBG_COUNTER_INC(rxd_flush);
2225 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2226 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2227 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2228 		    fl->ifl_id, fl->ifl_pidx);
2229 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2230 			fl->ifl_fragidx = frag_idx + 1;
2231 			if (fl->ifl_fragidx == fl->ifl_size)
2232 				fl->ifl_fragidx = 0;
2233 		} else {
2234 			fl->ifl_fragidx = frag_idx;
2235 		}
2236 	}
2237 
2238 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2239 }
2240 
2241 static inline uint8_t
iflib_fl_refill_all(if_ctx_t ctx,iflib_fl_t fl)2242 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2243 {
2244 	/*
2245 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2246 	 * This is important as it confuses most NICs. For instance,
2247 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2248 	 * RDH points to the next receive descriptor to be used by the NIC,
2249 	 * and RDT for the next receive descriptor to be published by the
2250 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2251 	 * The condition RDH == RDT means no descriptors are available to
2252 	 * the NIC, and thus it would be ambiguous if it also meant that
2253 	 * all the descriptors are available to the NIC.
2254 	 */
2255 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2256 #ifdef INVARIANTS
2257 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2258 #endif
2259 
2260 	MPASS(fl->ifl_credits <= fl->ifl_size);
2261 	MPASS(reclaimable == delta);
2262 
2263 	if (reclaimable > 0)
2264 		return (iflib_fl_refill(ctx, fl, reclaimable));
2265 	return (0);
2266 }
2267 
2268 uint8_t
iflib_in_detach(if_ctx_t ctx)2269 iflib_in_detach(if_ctx_t ctx)
2270 {
2271 	bool in_detach;
2272 
2273 	STATE_LOCK(ctx);
2274 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2275 	STATE_UNLOCK(ctx);
2276 	return (in_detach);
2277 }
2278 
2279 static void
iflib_fl_bufs_free(iflib_fl_t fl)2280 iflib_fl_bufs_free(iflib_fl_t fl)
2281 {
2282 	iflib_dma_info_t idi = fl->ifl_ifdi;
2283 	bus_dmamap_t sd_map;
2284 	uint32_t i;
2285 
2286 	for (i = 0; i < fl->ifl_size; i++) {
2287 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2288 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2289 
2290 		if (*sd_cl != NULL) {
2291 			sd_map = fl->ifl_sds.ifsd_map[i];
2292 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2293 			    BUS_DMASYNC_POSTREAD);
2294 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2295 			uma_zfree(fl->ifl_zone, *sd_cl);
2296 			*sd_cl = NULL;
2297 			if (*sd_m != NULL) {
2298 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2299 				m_free_raw(*sd_m);
2300 				*sd_m = NULL;
2301 			}
2302 		} else {
2303 			MPASS(*sd_m == NULL);
2304 		}
2305 #if MEMORY_LOGGING
2306 		fl->ifl_m_dequeued++;
2307 		fl->ifl_cl_dequeued++;
2308 #endif
2309 	}
2310 #ifdef INVARIANTS
2311 	for (i = 0; i < fl->ifl_size; i++) {
2312 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2313 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2314 	}
2315 #endif
2316 	/*
2317 	 * Reset free list values
2318 	 */
2319 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2320 	bzero(idi->idi_vaddr, idi->idi_size);
2321 }
2322 
2323 /*********************************************************************
2324  *
2325  *  Initialize a free list and its buffers.
2326  *
2327  **********************************************************************/
2328 static int
iflib_fl_setup(iflib_fl_t fl)2329 iflib_fl_setup(iflib_fl_t fl)
2330 {
2331 	iflib_rxq_t rxq = fl->ifl_rxq;
2332 	if_ctx_t ctx = rxq->ifr_ctx;
2333 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2334 	int qidx;
2335 
2336 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2337 	/*
2338 	 * Free current RX buffer structs and their mbufs
2339 	 */
2340 	iflib_fl_bufs_free(fl);
2341 	/* Now replenish the mbufs */
2342 	MPASS(fl->ifl_credits == 0);
2343 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2344 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2345 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2346 	else
2347 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2348 	/*
2349 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2350 	 * to the selected mbuf size.
2351 	 */
2352 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2353 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2354 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2355 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2356 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2357 
2358 	/*
2359 	 * Avoid pre-allocating zillions of clusters to an idle card
2360 	 * potentially speeding up attach. In any case make sure
2361 	 * to leave a descriptor unavailable. See the comment in
2362 	 * iflib_fl_refill_all().
2363 	 */
2364 	MPASS(fl->ifl_size > 0);
2365 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2366 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2367 		return (ENOBUFS);
2368 	/*
2369 	 * handle failure
2370 	 */
2371 	MPASS(rxq != NULL);
2372 	MPASS(fl->ifl_ifdi != NULL);
2373 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2374 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2375 	return (0);
2376 }
2377 
2378 /*********************************************************************
2379  *
2380  *  Free receive ring data structures
2381  *
2382  **********************************************************************/
2383 static void
iflib_rx_sds_free(iflib_rxq_t rxq)2384 iflib_rx_sds_free(iflib_rxq_t rxq)
2385 {
2386 	iflib_fl_t fl;
2387 	int i, j;
2388 
2389 	if (rxq->ifr_fl != NULL) {
2390 		for (i = 0; i < rxq->ifr_nfl; i++) {
2391 			fl = &rxq->ifr_fl[i];
2392 			if (fl->ifl_buf_tag != NULL) {
2393 				if (fl->ifl_sds.ifsd_map != NULL) {
2394 					for (j = 0; j < fl->ifl_size; j++) {
2395 						bus_dmamap_sync(
2396 						    fl->ifl_buf_tag,
2397 						    fl->ifl_sds.ifsd_map[j],
2398 						    BUS_DMASYNC_POSTREAD);
2399 						bus_dmamap_unload(
2400 						    fl->ifl_buf_tag,
2401 						    fl->ifl_sds.ifsd_map[j]);
2402 						bus_dmamap_destroy(
2403 						    fl->ifl_buf_tag,
2404 						    fl->ifl_sds.ifsd_map[j]);
2405 					}
2406 				}
2407 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2408 				fl->ifl_buf_tag = NULL;
2409 			}
2410 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2411 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2412 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2413 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2414 			free(fl->ifl_rx_bitmap, M_IFLIB);
2415 			fl->ifl_sds.ifsd_m = NULL;
2416 			fl->ifl_sds.ifsd_cl = NULL;
2417 			fl->ifl_sds.ifsd_ba = NULL;
2418 			fl->ifl_sds.ifsd_map = NULL;
2419 			fl->ifl_rx_bitmap = NULL;
2420 		}
2421 		free(rxq->ifr_fl, M_IFLIB);
2422 		rxq->ifr_fl = NULL;
2423 		free(rxq->ifr_ifdi, M_IFLIB);
2424 		rxq->ifr_ifdi = NULL;
2425 		rxq->ifr_cq_cidx = 0;
2426 	}
2427 }
2428 
2429 /*
2430  * Timer routine
2431  */
2432 static void
iflib_timer(void * arg)2433 iflib_timer(void *arg)
2434 {
2435 	iflib_txq_t txq = arg;
2436 	if_ctx_t ctx = txq->ift_ctx;
2437 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2438 	uint64_t this_tick = ticks;
2439 
2440 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2441 		return;
2442 
2443 	/*
2444 	 * Check on the state of the TX queue(s); this can be done
2445 	 * without the lock: the counters the check reads are only
2446 	 * advanced by the queue's tx task and a stale read just
2447 	 * delays the verdict by one timer period.
2448 	 */
2449 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2450 		qidx_t outstanding;
2451 		bool frozen;
2452 
2453 		txq->ift_last_timer_tick = this_tick;
2454 		IFDI_TIMER(ctx, txq->ift_id);
2455 
2456 		/*
2457 		 * Descriptors the hardware has not reported as
2458 		 * completed: neither harvested as credits
2459 		 * (ift_processed) nor reclaimed (ift_cleaned accounts
2460 		 * the difference to ift_in_use).  The tail whose
2461 		 * report-status request is still deferred is never
2462 		 * reported and must not count (ift_rs_pending
2463 		 * over-counts it by one per packet).
2464 		 */
2465 		outstanding = txq->ift_in_use -
2466 		    (qidx_t)(txq->ift_processed - txq->ift_cleaned);
2467 
2468 		/*
2469 		 * The queue is frozen while it has descriptors the
2470 		 * hardware has not reported as completed and none
2471 		 * were reclaimed over the period; the link must be
2472 		 * up, with no pause frames and no pending doorbell
2473 		 * (the laggard check below rings it).
2474 		 *
2475 		 * Being frozen is not a fault - the hardware may
2476 		 * defer marking descriptors as completed
2477 		 * indefinitely, and 8254x hardware does so for a
2478 		 * quiet queue - therefore the check arms only when a
2479 		 * frozen queue also takes on new work, and acts only
2480 		 * once it has stayed frozen for
2481 		 * net.iflib.tx_watchdog_periods consecutive periods.
2482 		 */
2483 		frozen = outstanding > txq->ift_rs_pending &&
2484 		    txq->ift_processed == txq->ift_processed_prev &&
2485 		    txq->ift_db_pending == 0 &&
2486 		    sctx->isc_pause_frames == 0 &&
2487 		    ctx->ifc_link_state == LINK_STATE_UP;
2488 		if (!frozen)
2489 			txq->ift_wdog_armed = 0;
2490 		else if (txq->ift_wdog_armed > 0 ||
2491 		    outstanding > txq->ift_outstanding_prev) {
2492 			if (txq->ift_wdog_armed < UINT16_MAX)
2493 				txq->ift_wdog_armed++;
2494 		}
2495 
2496 		/*
2497 		 * Frozen long enough: ask the hardware.  Completions
2498 		 * ready but unharvested for this long mean the
2499 		 * completion interrupt went missing - kick the
2500 		 * queue's task.  Nothing ready, although the queue
2501 		 * kept taking on work, means it is hung.
2502 		 */
2503 		if (iflib_tx_watchdog_periods > 0 &&
2504 		    txq->ift_wdog_armed >= iflib_tx_watchdog_periods) {
2505 			bus_dmamap_sync(txq->ift_ifdi->idi_tag,
2506 			    txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD);
2507 			if (ctx->isc_txd_credits_update(ctx->ifc_softc,
2508 			    txq->ift_id, false) == 0) {
2509 				device_printf(ctx->ifc_dev,
2510 				    "Watchdog timeout (TX: %d desc "
2511 				    "avail: %d pidx: %d) -- resetting\n",
2512 				    txq->ift_id, TXQ_AVAIL(txq),
2513 				    txq->ift_pidx);
2514 				STATE_LOCK(ctx);
2515 				if_setdrvflagbits(ctx->ifc_ifp,
2516 				    IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2517 				ctx->ifc_flags |=
2518 				    (IFC_DO_WATCHDOG | IFC_DO_RESET);
2519 				iflib_admin_intr_deferred(ctx);
2520 				STATE_UNLOCK(ctx);
2521 				return;
2522 			}
2523 			GROUPTASK_ENQUEUE(&txq->ift_task);
2524 		}
2525 		txq->ift_outstanding_prev = outstanding;
2526 		txq->ift_processed_prev = txq->ift_processed;
2527 	}
2528 	/* handle any laggards */
2529 	if (txq->ift_db_pending)
2530 		GROUPTASK_ENQUEUE(&txq->ift_task);
2531 
2532 	sctx->isc_pause_frames = 0;
2533 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2534 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2535 		    txq, txq->ift_timer.c_cpu);
2536 }
2537 
2538 static uint16_t
iflib_get_mbuf_size_for(unsigned int size)2539 iflib_get_mbuf_size_for(unsigned int size)
2540 {
2541 
2542 	if (size <= MCLBYTES)
2543 		return (MCLBYTES);
2544 	else
2545 		return (MJUMPAGESIZE);
2546 }
2547 
2548 static void
iflib_calc_rx_mbuf_sz(if_ctx_t ctx)2549 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2550 {
2551 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2552 
2553 	/*
2554 	 * XXX don't set the max_frame_size to larger
2555 	 * than the hardware can handle
2556 	 */
2557 	ctx->ifc_rx_mbuf_sz =
2558 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2559 }
2560 
2561 uint32_t
iflib_get_rx_mbuf_sz(if_ctx_t ctx)2562 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2563 {
2564 
2565 	return (ctx->ifc_rx_mbuf_sz);
2566 }
2567 
2568 static void
iflib_init_locked(if_ctx_t ctx)2569 iflib_init_locked(if_ctx_t ctx)
2570 {
2571 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2572 	if_t ifp = ctx->ifc_ifp;
2573 	iflib_fl_t fl;
2574 	iflib_txq_t txq;
2575 	iflib_rxq_t rxq;
2576 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2577 	bool init_failed;
2578 
2579 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2580 	IFDI_INTR_DISABLE(ctx);
2581 
2582 	/*
2583 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2584 	 * called without first calling iflib_stop().
2585 	 */
2586 	netmap_disable_all_rings(ifp);
2587 
2588 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2589 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2590 	/* Set hardware offload abilities */
2591 	if_clearhwassist(ifp);
2592 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2593 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2594 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2595 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2596 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2597 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2598 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2599 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2600 
2601 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2602 		CALLOUT_LOCK(txq);
2603 		callout_stop(&txq->ift_timer);
2604 #ifdef DEV_NETMAP
2605 		callout_stop(&txq->ift_netmap_timer);
2606 #endif /* DEV_NETMAP */
2607 		CALLOUT_UNLOCK(txq);
2608 		(void)iflib_netmap_txq_init(ctx, txq);
2609 	}
2610 
2611 	/*
2612 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2613 	 * that drivers can use the value when setting up the hardware receive
2614 	 * buffers.
2615 	 */
2616 	iflib_calc_rx_mbuf_sz(ctx);
2617 
2618 #ifdef INVARIANTS
2619 	i = if_getdrvflags(ifp);
2620 #endif
2621 	STATE_LOCK(ctx);
2622 	ctx->ifc_flags &= ~IFC_INIT_FAILED;
2623 	STATE_UNLOCK(ctx);
2624 	IFDI_INIT(ctx);
2625 	MPASS(if_getdrvflags(ifp) == i);
2626 	STATE_LOCK(ctx);
2627 	init_failed = (ctx->ifc_flags & IFC_INIT_FAILED) != 0;
2628 	STATE_UNLOCK(ctx);
2629 	if (init_failed)
2630 		return;
2631 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2632 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2633 			/* This rxq is in netmap mode. Skip normal init. */
2634 			continue;
2635 		}
2636 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2637 			if (iflib_fl_setup(fl)) {
2638 				device_printf(ctx->ifc_dev,
2639 				    "setting up free list %d failed - "
2640 				    "check cluster settings\n", j);
2641 				goto done;
2642 			}
2643 		}
2644 	}
2645 done:
2646 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2647 	IFDI_INTR_ENABLE(ctx);
2648 	txq = ctx->ifc_txqs;
2649 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2650 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2651 			txq->ift_timer.c_cpu);
2652 
2653 	/* Re-enable txsync/rxsync. */
2654 	netmap_enable_all_rings(ifp);
2655 }
2656 
2657 static int
iflib_media_change(if_t ifp)2658 iflib_media_change(if_t ifp)
2659 {
2660 	if_ctx_t ctx = if_getsoftc(ifp);
2661 	int err;
2662 
2663 	CTX_LOCK(ctx);
2664 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2665 		iflib_if_init_locked(ctx);
2666 	CTX_UNLOCK(ctx);
2667 	return (err);
2668 }
2669 
2670 static void
iflib_media_status(if_t ifp,struct ifmediareq * ifmr)2671 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2672 {
2673 	if_ctx_t ctx = if_getsoftc(ifp);
2674 
2675 	CTX_LOCK(ctx);
2676 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2677 	IFDI_MEDIA_STATUS(ctx, ifmr);
2678 	CTX_UNLOCK(ctx);
2679 }
2680 
2681 static void
iflib_stop(if_ctx_t ctx)2682 iflib_stop(if_ctx_t ctx)
2683 {
2684 	iflib_txq_t txq = ctx->ifc_txqs;
2685 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2686 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2687 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2688 	iflib_dma_info_t di;
2689 	iflib_fl_t fl;
2690 	int i, j;
2691 
2692 	/* Tell the stack that the interface is no longer active */
2693 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2694 
2695 	IFDI_INTR_DISABLE(ctx);
2696 	DELAY(1000);
2697 	IFDI_STOP(ctx);
2698 	DELAY(1000);
2699 
2700 	/*
2701 	 * Stop any pending txsync/rxsync and prevent new ones
2702 	 * form starting. Processes blocked in poll() will get
2703 	 * POLLERR.
2704 	 */
2705 	netmap_disable_all_rings(ctx->ifc_ifp);
2706 
2707 	iflib_debug_reset();
2708 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2709 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2710 		/* make sure all transmitters have completed before proceeding XXX */
2711 
2712 		CALLOUT_LOCK(txq);
2713 		callout_stop(&txq->ift_timer);
2714 #ifdef DEV_NETMAP
2715 		callout_stop(&txq->ift_netmap_timer);
2716 #endif /* DEV_NETMAP */
2717 		CALLOUT_UNLOCK(txq);
2718 
2719 		if (!ctx->ifc_sysctl_simple_tx) {
2720 			/* clean any enqueued buffers */
2721 			iflib_ifmp_purge(txq);
2722 		}
2723 		/* Free any existing tx buffers. */
2724 		for (j = 0; j < txq->ift_size; j++) {
2725 			iflib_txsd_free(ctx, txq, j);
2726 		}
2727 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2728 		txq->ift_processed_prev = 0;
2729 		txq->ift_outstanding_prev = 0;
2730 		txq->ift_wdog_armed = 0;
2731 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2732 		txq->ift_npending = txq->ift_db_pending = 0;
2733 		txq->ift_rs_pending = 0;
2734 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2735 			txq->ift_cidx = txq->ift_pidx;
2736 		else
2737 			txq->ift_cidx = txq->ift_pidx = 0;
2738 
2739 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2740 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2741 		txq->ift_pullups = 0;
2742 		ifmp_ring_reset_stats(txq->ift_br);
2743 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2744 			bzero((void *)di->idi_vaddr, di->idi_size);
2745 	}
2746 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2747 		if (rxq->ifr_task.gt_taskqueue != NULL)
2748 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2749 				 &rxq->ifr_task.gt_task);
2750 
2751 		rxq->ifr_cq_cidx = 0;
2752 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2753 			bzero((void *)di->idi_vaddr, di->idi_size);
2754 		/* also resets the free lists pidx/cidx */
2755 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2756 			iflib_fl_bufs_free(fl);
2757 	}
2758 }
2759 
2760 static inline caddr_t
calc_next_rxd(iflib_fl_t fl,int cidx)2761 calc_next_rxd(iflib_fl_t fl, int cidx)
2762 {
2763 	qidx_t size;
2764 	int nrxd;
2765 	caddr_t start, end, cur, next;
2766 
2767 	nrxd = fl->ifl_size;
2768 	size = fl->ifl_rxd_size;
2769 	start = fl->ifl_ifdi->idi_vaddr;
2770 
2771 	if (__predict_false(size == 0))
2772 		return (start);
2773 	cur = start + size * cidx;
2774 	end = start + size * nrxd;
2775 	next = CACHE_PTR_NEXT(cur);
2776 	return (next < end ? next : start);
2777 }
2778 
2779 static inline void
prefetch_pkts(iflib_fl_t fl,int cidx)2780 prefetch_pkts(iflib_fl_t fl, int cidx)
2781 {
2782 	int nextptr;
2783 	int nrxd = fl->ifl_size;
2784 	caddr_t next_rxd;
2785 
2786 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1);
2787 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2788 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2789 	next_rxd = calc_next_rxd(fl, cidx);
2790 	prefetch(next_rxd);
2791 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]);
2792 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]);
2793 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]);
2794 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]);
2795 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]);
2796 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]);
2797 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]);
2798 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]);
2799 }
2800 
2801 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)2802 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2803     int *pf_rv, if_rxd_info_t ri)
2804 {
2805 	bus_dmamap_t map;
2806 	iflib_fl_t fl;
2807 	caddr_t payload;
2808 	struct mbuf *m;
2809 	int flid, cidx, len, next;
2810 
2811 	map = NULL;
2812 	flid = irf->irf_flid;
2813 	cidx = irf->irf_idx;
2814 	fl = &rxq->ifr_fl[flid];
2815 	sd->ifsd_fl = fl;
2816 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2817 	fl->ifl_credits--;
2818 #if MEMORY_LOGGING
2819 	fl->ifl_m_dequeued++;
2820 #endif
2821 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2822 		prefetch_pkts(fl, cidx);
2823 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1);
2824 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2825 	map = fl->ifl_sds.ifsd_map[cidx];
2826 
2827 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2828 
2829 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2830 	    irf->irf_len != 0) {
2831 		payload  = *sd->ifsd_cl;
2832 		payload +=  ri->iri_pad;
2833 		len = ri->iri_len - ri->iri_pad;
2834 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2835 		switch (*pf_rv) {
2836 		case PFIL_DROPPED:
2837 		case PFIL_CONSUMED:
2838 			/*
2839 			 * The filter ate it.  Everything is recycled.
2840 			 */
2841 			m = NULL;
2842 			unload = 0;
2843 			break;
2844 		case PFIL_REALLOCED:
2845 			/*
2846 			 * The filter copied it.  Everything is recycled.
2847 			 * 'm' points at new mbuf.
2848 			 */
2849 			unload = 0;
2850 			break;
2851 		case PFIL_PASS:
2852 			/*
2853 			 * Filter said it was OK, so receive like
2854 			 * normal
2855 			 */
2856 			m = fl->ifl_sds.ifsd_m[cidx];
2857 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2858 			break;
2859 		default:
2860 			MPASS(0);
2861 		}
2862 	} else {
2863 		m = fl->ifl_sds.ifsd_m[cidx];
2864 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2865 		if (pf_rv != NULL)
2866 			*pf_rv = PFIL_PASS;
2867 	}
2868 
2869 	if (unload && irf->irf_len != 0)
2870 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2871 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1);
2872 	if (__predict_false(fl->ifl_cidx == 0))
2873 		fl->ifl_gen = 0;
2874 	bit_clear(fl->ifl_rx_bitmap, cidx);
2875 	return (m);
2876 }
2877 
2878 static struct mbuf *
assemble_segments(iflib_rxq_t rxq,if_rxd_info_t ri,if_rxsd_t sd,int * pf_rv)2879 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2880 {
2881 	struct mbuf *m, *mh, *mt;
2882 	caddr_t cl;
2883 	int  *pf_rv_ptr, flags, i, padlen;
2884 	bool consumed;
2885 
2886 	i = 0;
2887 	mh = NULL;
2888 	consumed = false;
2889 	*pf_rv = PFIL_PASS;
2890 	pf_rv_ptr = pf_rv;
2891 	do {
2892 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2893 		    pf_rv_ptr, ri);
2894 
2895 		MPASS(*sd->ifsd_cl != NULL);
2896 
2897 		/*
2898 		 * Exclude zero-length frags & frags from
2899 		 * packets the filter has consumed or dropped
2900 		 */
2901 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2902 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2903 			if (mh == NULL) {
2904 				consumed = true;
2905 				pf_rv_ptr = NULL;
2906 			}
2907 			/* XXX we can save the cluster here, but not the mbuf */
2908 			if (m != NULL) {
2909 				m_init(m, M_NOWAIT, MT_DATA, 0);
2910 				m_free(m);
2911 			}
2912 			continue;
2913 		}
2914 		if (mh == NULL) {
2915 			flags = M_PKTHDR | M_EXT;
2916 			mh = mt = m;
2917 			padlen = ri->iri_pad;
2918 		} else {
2919 			flags = M_EXT;
2920 			mt->m_next = m;
2921 			mt = m;
2922 			/* assuming padding is only on the first fragment */
2923 			padlen = 0;
2924 		}
2925 		cl = *sd->ifsd_cl;
2926 		*sd->ifsd_cl = NULL;
2927 
2928 		/* Can these two be made one ? */
2929 		m_init(m, M_NOWAIT, MT_DATA, flags);
2930 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2931 		/*
2932 		 * These must follow m_init and m_cljset
2933 		 */
2934 		m->m_data += padlen;
2935 		ri->iri_len -= padlen;
2936 		m->m_len = ri->iri_frags[i].irf_len;
2937 	} while (++i < ri->iri_nfrags);
2938 
2939 	return (mh);
2940 }
2941 
2942 /*
2943  * Process one software descriptor
2944  */
2945 static struct mbuf *
iflib_rxd_pkt_get(iflib_rxq_t rxq,if_rxd_info_t ri)2946 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2947 {
2948 	struct if_rxsd sd;
2949 	struct mbuf *m;
2950 	int pf_rv;
2951 
2952 	/* should I merge this back in now that the two paths are basically duplicated? */
2953 	if (ri->iri_nfrags == 1 &&
2954 	    ri->iri_frags[0].irf_len != 0 &&
2955 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2956 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2957 		    &pf_rv, ri);
2958 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2959 			return (m);
2960 		if (pf_rv == PFIL_PASS) {
2961 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2962 #ifndef __NO_STRICT_ALIGNMENT
2963 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2964 				m->m_data += 2;
2965 #endif
2966 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2967 			m->m_len = ri->iri_frags[0].irf_len;
2968 			m->m_data += ri->iri_pad;
2969 			ri->iri_len -= ri->iri_pad;
2970 		}
2971 	} else {
2972 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2973 		if (m == NULL)
2974 			return (NULL);
2975 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2976 			return (m);
2977 	}
2978 	m->m_pkthdr.len = ri->iri_len;
2979 	m->m_pkthdr.rcvif = ri->iri_ifp;
2980 	m->m_flags |= ri->iri_flags;
2981 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2982 	m->m_pkthdr.flowid = ri->iri_flowid;
2983 #ifdef NUMA
2984 	m->m_pkthdr.numa_domain = if_getnumadomain(ri->iri_ifp);
2985 #endif
2986 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2987 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2988 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2989 	return (m);
2990 }
2991 
2992 static void
_task_fn_rx_watchdog(void * context)2993 _task_fn_rx_watchdog(void *context)
2994 {
2995 	iflib_rxq_t rxq = context;
2996 
2997 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2998 }
2999 
3000 static uint8_t
iflib_rxeof(iflib_rxq_t rxq,qidx_t budget)3001 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
3002 {
3003 	if_t ifp;
3004 	if_ctx_t ctx = rxq->ifr_ctx;
3005 	if_shared_ctx_t sctx = ctx->ifc_sctx;
3006 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
3007 	int avail, i;
3008 	qidx_t *cidxp;
3009 	struct if_rxd_info ri;
3010 	int err, budget_left, rx_bytes, rx_pkts;
3011 	iflib_fl_t fl;
3012 #if defined(INET6) || defined(INET)
3013 	int lro_enabled;
3014 #endif
3015 	uint8_t retval = 0;
3016 
3017 	/*
3018 	 * XXX early demux data packets so that if_input processing only handles
3019 	 * acks in interrupt context
3020 	 */
3021 	struct mbuf *m, *mh, *mt;
3022 
3023 	NET_EPOCH_ASSERT();
3024 
3025 	ifp = ctx->ifc_ifp;
3026 	mh = mt = NULL;
3027 	MPASS(budget > 0);
3028 	rx_pkts	= rx_bytes = 0;
3029 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
3030 		cidxp = &rxq->ifr_cq_cidx;
3031 	else
3032 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
3033 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
3034 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3035 			retval |= iflib_fl_refill_all(ctx, fl);
3036 		DBG_COUNTER_INC(rx_unavail);
3037 		return (retval);
3038 	}
3039 
3040 #if defined(INET6) || defined(INET)
3041 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3042 #endif
3043 
3044 	/* pfil needs the vnet to be set */
3045 	CURVNET_SET_QUIET(if_getvnet(ifp));
3046 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
3047 		if (__predict_false(!CTX_ACTIVE(ctx))) {
3048 			DBG_COUNTER_INC(rx_ctx_inactive);
3049 			break;
3050 		}
3051 		/*
3052 		 * Reset client set fields to their default values
3053 		 */
3054 		memset(&ri, 0, sizeof(ri));
3055 		ri.iri_qsidx = rxq->ifr_id;
3056 		ri.iri_cidx = *cidxp;
3057 		ri.iri_ifp = ifp;
3058 		ri.iri_frags = rxq->ifr_frags;
3059 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
3060 
3061 		if (err) {
3062 			CURVNET_RESTORE();
3063 			goto err;
3064 		}
3065 		rx_pkts += 1;
3066 		rx_bytes += ri.iri_len;
3067 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3068 			*cidxp = ri.iri_cidx;
3069 			/* Update our consumer index */
3070 			/* XXX NB: shurd - check if this is still safe */
3071 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3072 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3073 			/* was this only a completion queue message? */
3074 			if (__predict_false(ri.iri_nfrags == 0))
3075 				continue;
3076 		}
3077 		MPASS(ri.iri_nfrags != 0);
3078 		MPASS(ri.iri_len != 0);
3079 
3080 		/* will advance the cidx on the corresponding free lists */
3081 		m = iflib_rxd_pkt_get(rxq, &ri);
3082 		avail--;
3083 		budget_left--;
3084 		if (avail == 0 && budget_left)
3085 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3086 
3087 		if (__predict_false(m == NULL))
3088 			continue;
3089 
3090 #ifndef __NO_STRICT_ALIGNMENT
3091 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3092 			continue;
3093 #endif
3094 #if defined(INET6) || defined(INET)
3095 		if (lro_enabled) {
3096 			tcp_lro_queue_mbuf(&rxq->ifr_lc, m);
3097 			continue;
3098 		}
3099 #endif
3100 
3101 		if (mh == NULL)
3102 			mh = mt = m;
3103 		else {
3104 			mt->m_nextpkt = m;
3105 			mt = m;
3106 		}
3107 	}
3108 	CURVNET_RESTORE();
3109 	/* make sure that we can refill faster than drain */
3110 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3111 		retval |= iflib_fl_refill_all(ctx, fl);
3112 
3113 	if (mh != NULL) {
3114 		if_input(ifp, mh);
3115 		DBG_COUNTER_INC(rx_if_input);
3116 	}
3117 
3118 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3119 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3120 
3121 	/*
3122 	 * Flush any outstanding LRO work
3123 	 */
3124 #if defined(INET6) || defined(INET)
3125 	tcp_lro_flush_all(&rxq->ifr_lc);
3126 #endif
3127 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3128 		retval |= IFLIB_RXEOF_MORE;
3129 	return (retval);
3130 err:
3131 	STATE_LOCK(ctx);
3132 	ctx->ifc_flags |= IFC_DO_RESET;
3133 	iflib_admin_intr_deferred(ctx);
3134 	STATE_UNLOCK(ctx);
3135 	return (0);
3136 }
3137 
3138 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1)
3139 static inline qidx_t
txq_max_db_deferred(iflib_txq_t txq,qidx_t in_use)3140 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3141 {
3142 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3143 	qidx_t minthresh = txq->ift_size / 8;
3144 	if (in_use > 4 * minthresh)
3145 		return (notify_count);
3146 	if (in_use > 2 * minthresh)
3147 		return (notify_count >> 1);
3148 	if (in_use > minthresh)
3149 		return (notify_count >> 3);
3150 	return (0);
3151 }
3152 
3153 static inline qidx_t
txq_max_rs_deferred(iflib_txq_t txq)3154 txq_max_rs_deferred(iflib_txq_t txq)
3155 {
3156 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3157 	qidx_t minthresh = txq->ift_size / 8;
3158 	if (txq->ift_in_use > 4 * minthresh)
3159 		return (notify_count);
3160 	if (txq->ift_in_use > 2 * minthresh)
3161 		return (notify_count >> 1);
3162 	if (txq->ift_in_use > minthresh)
3163 		return (notify_count >> 2);
3164 	return (2);
3165 }
3166 
3167 #define M_CSUM_FLAGS(m)		((m)->m_pkthdr.csum_flags)
3168 #define M_HAS_VLANTAG(m)	(m->m_flags & M_VLANTAG)
3169 
3170 #define TXQ_MAX_DB_DEFERRED(txq, in_use)	txq_max_db_deferred((txq), (in_use))
3171 #define TXQ_MAX_RS_DEFERRED(txq)	txq_max_rs_deferred(txq)
3172 #define TXQ_MAX_DB_CONSUMED(size)	(size >> 4)
3173 
3174 /* forward compatibility for cxgb */
3175 #define FIRST_QSET(ctx) 0
3176 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3177 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3178 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3179 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3180 
3181 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3182     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3183 
3184 static inline bool
iflib_txd_db_check(iflib_txq_t txq,int ring)3185 iflib_txd_db_check(iflib_txq_t txq, int ring)
3186 {
3187 	if_ctx_t ctx = txq->ift_ctx;
3188 	qidx_t dbval, max;
3189 
3190 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3191 
3192 	/* force || threshold exceeded || at the edge of the ring */
3193 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx))) {
3194 
3195 		/*
3196 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3197 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3198 		 * producer index explicitly (INTC).
3199 		 */
3200 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3201 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3202 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3203 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3204 
3205 		/*
3206 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3207 		 */
3208 		txq->ift_db_pending = txq->ift_npending = 0;
3209 		return (true);
3210 	}
3211 	return (false);
3212 }
3213 
3214 #ifdef PKT_DEBUG
3215 static void
print_pkt(if_pkt_info_t pi)3216 print_pkt(if_pkt_info_t pi)
3217 {
3218 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3219 	    pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3220 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3221 	    pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3222 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3223 	    pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3224 }
3225 #endif
3226 
3227 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3228 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3229 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3230 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3231 
3232 /**
3233  * Parses out ethernet header information in the given mbuf.
3234  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3235  *
3236  * This will account for the VLAN header if present.
3237  *
3238  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3239  * types of packets.
3240  */
3241 static int
iflib_parse_ether_header(if_pkt_info_t pi,struct mbuf ** mp,uint64_t * pullups)3242 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3243 {
3244 	struct ether_vlan_header *eh;
3245 	struct mbuf *m;
3246 
3247 	m = *mp;
3248 	if (__predict_false(m->m_len < sizeof(*eh))) {
3249 		(*pullups)++;
3250 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3251 			return (ENOMEM);
3252 	}
3253 	eh = mtod(m, struct ether_vlan_header *);
3254 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3255 		pi->ipi_etype = ntohs(eh->evl_proto);
3256 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3257 	} else {
3258 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3259 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3260 	}
3261 	*mp = m;
3262 
3263 	return (0);
3264 }
3265 
3266 /**
3267  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3268  * Currently this information includes: IP ToS value, IP header version/presence
3269  *
3270  * This is missing some checks and doesn't edit the packet content as it goes,
3271  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3272  */
3273 static int
iflib_parse_header_partial(if_pkt_info_t pi,struct mbuf ** mp,uint64_t * pullups)3274 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3275 {
3276 	struct mbuf *m;
3277 	int err;
3278 
3279 	*pullups = 0;
3280 	m = *mp;
3281 	if (!M_WRITABLE(m)) {
3282 		m = m_dup(m, M_NOWAIT);
3283 		m_freem(*mp);
3284 		DBG_COUNTER_INC(tx_frees);
3285 		*mp = m;
3286 		if (m == NULL)
3287 			return (ENOMEM);
3288 	}
3289 
3290 	/* Fills out pi->ipi_etype */
3291 	err = iflib_parse_ether_header(pi, mp, pullups);
3292 	if (err)
3293 		return (err);
3294 	m = *mp;
3295 
3296 	switch (pi->ipi_etype) {
3297 #ifdef INET
3298 	case ETHERTYPE_IP:
3299 	{
3300 		struct mbuf *n;
3301 		struct ip *ip = NULL;
3302 		int miniplen;
3303 
3304 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3305 		if (__predict_false(m->m_len < miniplen)) {
3306 			/*
3307 			 * Check for common case where the first mbuf only contains
3308 			 * the Ethernet header
3309 			 */
3310 			if (m->m_len == pi->ipi_ehdrlen) {
3311 				n = m->m_next;
3312 				MPASS(n);
3313 				/* If next mbuf contains at least the minimal IP header, then stop */
3314 				if (n->m_len >= sizeof(*ip)) {
3315 					ip = (struct ip *)n->m_data;
3316 				} else {
3317 					(*pullups)++;
3318 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3319 						return (ENOMEM);
3320 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3321 				}
3322 			} else {
3323 				(*pullups)++;
3324 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3325 					return (ENOMEM);
3326 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3327 			}
3328 		} else {
3329 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3330 		}
3331 
3332 		/* Have the IPv4 header w/ no options here */
3333 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3334 		pi->ipi_ipproto = ip->ip_p;
3335 		pi->ipi_ip_tos = ip->ip_tos;
3336 		pi->ipi_flags |= IPI_TX_IPV4;
3337 
3338 		break;
3339 	}
3340 #endif
3341 #ifdef INET6
3342 	case ETHERTYPE_IPV6:
3343 	{
3344 		struct ip6_hdr *ip6;
3345 
3346 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3347 			(*pullups)++;
3348 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3349 				return (ENOMEM);
3350 		}
3351 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3352 
3353 		/* Have the IPv6 fixed header here */
3354 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3355 		pi->ipi_ipproto = ip6->ip6_nxt;
3356 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3357 		pi->ipi_flags |= IPI_TX_IPV6;
3358 
3359 		break;
3360 	}
3361 #endif
3362 	default:
3363 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3364 		pi->ipi_ip_hlen = 0;
3365 		break;
3366 	}
3367 	*mp = m;
3368 
3369 	return (0);
3370 
3371 }
3372 
3373 static int
iflib_parse_header(iflib_txq_t txq,if_pkt_info_t pi,struct mbuf ** mp)3374 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3375 {
3376 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3377 	struct mbuf *m;
3378 	int err;
3379 
3380 	m = *mp;
3381 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3382 	    M_WRITABLE(m) == 0) {
3383 		m = m_dup(m, M_NOWAIT);
3384 		m_freem(*mp);
3385 		DBG_COUNTER_INC(tx_frees);
3386 		*mp = m;
3387 		if (m == NULL)
3388 			return (ENOMEM);
3389 	}
3390 
3391 	/* Fills out pi->ipi_etype */
3392 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3393 	if (__predict_false(err))
3394 		return (err);
3395 	m = *mp;
3396 
3397 	switch (pi->ipi_etype) {
3398 #ifdef INET
3399 	case ETHERTYPE_IP:
3400 	{
3401 		struct ip *ip;
3402 		struct tcphdr *th;
3403 		uint8_t hlen;
3404 
3405 		hlen = pi->ipi_ehdrlen + sizeof(*ip);
3406 		if (__predict_false(m->m_len < hlen)) {
3407 			txq->ift_pullups++;
3408 			if (__predict_false((m = m_pullup(m, hlen)) == NULL))
3409 				return (ENOMEM);
3410 		}
3411 		ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3412 		hlen = pi->ipi_ehdrlen + (ip->ip_hl << 2);
3413 		if (ip->ip_p == IPPROTO_TCP) {
3414 			hlen += sizeof(*th);
3415 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3416 		} else if (ip->ip_p == IPPROTO_UDP) {
3417 			hlen += sizeof(struct udphdr);
3418 		}
3419 		if (__predict_false(m->m_len < hlen)) {
3420 			txq->ift_pullups++;
3421 			if ((m = m_pullup(m, hlen)) == NULL)
3422 				return (ENOMEM);
3423 			/* reset pointers after pullup */
3424 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3425 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3426 		}
3427 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3428 		pi->ipi_ipproto = ip->ip_p;
3429 		pi->ipi_ip_tos = ip->ip_tos;
3430 		pi->ipi_flags |= IPI_TX_IPV4;
3431 
3432 		/* TCP checksum offload may require TCP header length */
3433 		if (IS_TX_OFFLOAD4(pi)) {
3434 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3435 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3436 				pi->ipi_tcp_hlen = th->th_off << 2;
3437 				pi->ipi_tcp_seq = th->th_seq;
3438 			}
3439 			if (IS_TSO4(pi)) {
3440 				MPASS(ip->ip_p == IPPROTO_TCP);
3441 				/*
3442 				 * TSO always requires hardware checksum offload.
3443 				 */
3444 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3445 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3446 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3447 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3448 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3449 					ip->ip_sum = 0;
3450 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3451 				}
3452 			}
3453 		}
3454 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3455 			ip->ip_sum = 0;
3456 
3457 		break;
3458 	}
3459 #endif
3460 #ifdef INET6
3461 	case ETHERTYPE_IPV6:
3462 	{
3463 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3464 		struct tcphdr *th;
3465 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3466 
3467 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3468 			txq->ift_pullups++;
3469 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3470 				return (ENOMEM);
3471 			/* reset pointers after pullup */
3472 			ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3473 		}
3474 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3475 
3476 		/* XXX-BZ this will go badly in case of ext hdrs. */
3477 		pi->ipi_ipproto = ip6->ip6_nxt;
3478 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3479 		pi->ipi_flags |= IPI_TX_IPV6;
3480 
3481 		/* TCP checksum offload may require TCP header length */
3482 		if (IS_TX_OFFLOAD6(pi)) {
3483 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3484 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3485 					txq->ift_pullups++;
3486 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3487 						return (ENOMEM);
3488 					/* reset pointers after pullup */
3489 					ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3490 					th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3491 				}
3492 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3493 				pi->ipi_tcp_hlen = th->th_off << 2;
3494 				pi->ipi_tcp_seq = th->th_seq;
3495 			}
3496 			if (IS_TSO6(pi)) {
3497 				MPASS(ip6->ip6_nxt == IPPROTO_TCP);
3498 				/*
3499 				 * TSO always requires hardware checksum offload.
3500 				 */
3501 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3502 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3503 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3504 			}
3505 		}
3506 		break;
3507 	}
3508 #endif
3509 	default:
3510 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3511 		pi->ipi_ip_hlen = 0;
3512 		break;
3513 	}
3514 	*mp = m;
3515 
3516 	return (0);
3517 }
3518 
3519 /*
3520  * If dodgy hardware rejects the scatter gather chain we've handed it
3521  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3522  * m_defrag'd mbufs
3523  */
3524 static __noinline struct mbuf *
iflib_remove_mbuf(iflib_txq_t txq)3525 iflib_remove_mbuf(iflib_txq_t txq)
3526 {
3527 	int ntxd, pidx;
3528 	struct mbuf *m, **ifsd_m;
3529 
3530 	ifsd_m = txq->ift_sds.ifsd_m;
3531 	ntxd = txq->ift_size;
3532 	pidx = txq->ift_pidx & (ntxd - 1);
3533 	ifsd_m = txq->ift_sds.ifsd_m;
3534 	m = IFLIB_GET_MBUF(ifsd_m[pidx]);
3535 	ifsd_m[pidx] = NULL;
3536 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3537 	if (txq->ift_sds.ifsd_tso_map != NULL)
3538 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3539 		    txq->ift_sds.ifsd_tso_map[pidx]);
3540 #if MEMORY_LOGGING
3541 	txq->ift_dequeued++;
3542 #endif
3543 	return (m);
3544 }
3545 
3546 /*
3547  * Pad an mbuf to ensure a minimum ethernet frame size.
3548  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3549  */
3550 static __noinline int
iflib_ether_pad(device_t dev,struct mbuf ** m_head,uint16_t min_frame_size)3551 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3552 {
3553 	/*
3554 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3555 	 * and ARP message is the smallest common payload I can think of
3556 	 */
3557 	static char pad[18];	/* just zeros */
3558 	int n;
3559 	struct mbuf *new_head;
3560 
3561 	if (!M_WRITABLE(*m_head)) {
3562 		new_head = m_dup(*m_head, M_NOWAIT);
3563 		m_freem(*m_head);
3564 		*m_head = new_head;
3565 		if (new_head == NULL) {
3566 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3567 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3568 			DBG_COUNTER_INC(tx_frees);
3569 			return (ENOMEM);
3570 		}
3571 	}
3572 
3573 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3574 	     n > 0; n -= sizeof(pad))
3575 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3576 			break;
3577 
3578 	if (n > 0) {
3579 		m_freem(*m_head);
3580 		*m_head = NULL;
3581 		device_printf(dev, "cannot pad short frame\n");
3582 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3583 		DBG_COUNTER_INC(tx_frees);
3584 		return (ENOMEM);
3585 	}
3586 
3587 	return (0);
3588 }
3589 
3590 static int
iflib_encap(iflib_txq_t txq,struct mbuf ** m_headp,int * obytes,int * opkts)3591 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp, int *obytes, int *opkts)
3592 {
3593 	if_ctx_t		ctx;
3594 	if_shared_ctx_t		sctx;
3595 	if_softc_ctx_t		scctx;
3596 	bus_dma_tag_t		buf_tag;
3597 	bus_dma_segment_t	*segs;
3598 	struct mbuf		*m_head, **ifsd_m;
3599 	bus_dmamap_t		map;
3600 	struct if_pkt_info	pi;
3601 	uintptr_t		flags;
3602 	int remap = 0;
3603 	int err, nsegs, ndesc, max_segs, pidx;
3604 
3605 	ctx = txq->ift_ctx;
3606 	sctx = ctx->ifc_sctx;
3607 	scctx = &ctx->ifc_softc_ctx;
3608 	segs = txq->ift_segs;
3609 	m_head = *m_headp;
3610 	map = NULL;
3611 
3612 	/*
3613 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3614 	 */
3615 	pidx = txq->ift_pidx;
3616 	map = txq->ift_sds.ifsd_map[pidx];
3617 	ifsd_m = txq->ift_sds.ifsd_m;
3618 
3619 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3620 		buf_tag = txq->ift_tso_buf_tag;
3621 		max_segs = scctx->isc_tx_tso_segments_max;
3622 		map = txq->ift_sds.ifsd_tso_map[pidx];
3623 		MPASS(buf_tag != NULL);
3624 		MPASS(max_segs > 0);
3625 		flags = IFLIB_TSO;
3626 	} else {
3627 		buf_tag = txq->ift_buf_tag;
3628 		max_segs = scctx->isc_tx_nsegments;
3629 		map = txq->ift_sds.ifsd_map[pidx];
3630 		flags = IFLIB_NO_TSO;
3631 	}
3632 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3633 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3634 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3635 		if (err) {
3636 			DBG_COUNTER_INC(encap_txd_encap_fail);
3637 			return (err);
3638 		}
3639 	}
3640 	m_head = *m_headp;
3641 
3642 	memset(&pi, 0, sizeof(pi));
3643 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST));
3644 	pi.ipi_pidx = pidx;
3645 	pi.ipi_qsidx = txq->ift_id;
3646 	pi.ipi_len = m_head->m_pkthdr.len;
3647 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3648 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3649 
3650 	/* deliberate bitwise OR to make one condition */
3651 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3652 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3653 			DBG_COUNTER_INC(encap_txd_encap_fail);
3654 			return (err);
3655 		}
3656 		m_head = *m_headp;
3657 	}
3658 
3659 retry:
3660 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3661 	    BUS_DMA_NOWAIT);
3662 defrag:
3663 	if (__predict_false(err)) {
3664 		switch (err) {
3665 		case EFBIG:
3666 			/* try collapse once and defrag once */
3667 			if (remap == 0) {
3668 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3669 				/* try defrag if collapsing fails */
3670 				if (m_head == NULL)
3671 					remap++;
3672 			}
3673 			if (remap == 1) {
3674 				txq->ift_mbuf_defrag++;
3675 				m_head = m_defrag(*m_headp, M_NOWAIT);
3676 			}
3677 			/*
3678 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3679 			 * failed to map an mbuf that was run through m_defrag
3680 			 */
3681 			MPASS(remap <= 1);
3682 			if (__predict_false(m_head == NULL || remap > 1))
3683 				goto defrag_failed;
3684 			remap++;
3685 			*m_headp = m_head;
3686 			goto retry;
3687 			break;
3688 		case ENOMEM:
3689 			/* FALLTHROUGH */
3690 		default:
3691 			txq->ift_no_tx_dma_setup++;
3692 			m_freem(*m_headp);
3693 			DBG_COUNTER_INC(tx_frees);
3694 			*m_headp = NULL;
3695 			break;
3696 		}
3697 		txq->ift_map_failed++;
3698 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3699 		DBG_COUNTER_INC(encap_txd_encap_fail);
3700 		return (err);
3701 	}
3702 	ifsd_m[pidx] = IFLIB_SAVE_MBUF(m_head, flags);
3703 	if (m_head->m_pkthdr.csum_flags & CSUM_SND_TAG)
3704 		pi.ipi_mbuf = m_head;
3705 	else
3706 		pi.ipi_mbuf = NULL;
3707 	/*
3708 	 * XXX assumes a 1 to 1 relationship between segments and
3709 	 *        descriptors - this does not hold true on all drivers, e.g.
3710 	 *        cxgb
3711 	 */
3712 	if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3713 		iflib_completed_tx_reclaim_force(txq);
3714 		if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3715 			txq->ift_no_desc_avail++;
3716 			bus_dmamap_unload(buf_tag, map);
3717 			DBG_COUNTER_INC(encap_txq_avail_fail);
3718 			DBG_COUNTER_INC(encap_txd_encap_fail);
3719 			if (ctx->ifc_sysctl_simple_tx) {
3720 				*m_headp = m_head = iflib_remove_mbuf(txq);
3721 				m_freem(*m_headp);
3722 				DBG_COUNTER_INC(tx_frees);
3723 				*m_headp = NULL;
3724 			}
3725 			if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3726 				GROUPTASK_ENQUEUE(&txq->ift_task);
3727 			return (ENOBUFS);
3728 		}
3729 	}
3730 	/*
3731 	 * On Intel cards we can greatly reduce the number of TX interrupts
3732 	 * we see by only setting report status on every Nth descriptor.
3733 	 * However, this also means that the driver will need to keep track
3734 	 * of the descriptors that RS was set on to check them for the DD bit.
3735 	 */
3736 	txq->ift_rs_pending += nsegs + 1;
3737 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3738 	    iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx)) {
3739 		pi.ipi_flags |= IPI_TX_INTR;
3740 		txq->ift_rs_pending = 0;
3741 	}
3742 
3743 	pi.ipi_segs = segs;
3744 	pi.ipi_nsegs = nsegs;
3745 
3746 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3747 #ifdef PKT_DEBUG
3748 	print_pkt(&pi);
3749 #endif
3750 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3751 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3752 		DBG_COUNTER_INC(tx_encap);
3753 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3754 
3755 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3756 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3757 			ndesc += txq->ift_size;
3758 			txq->ift_gen = 1;
3759 		}
3760 		/*
3761 		 * drivers can need up to ift_pad sentinels
3762 		 */
3763 		MPASS(ndesc <= pi.ipi_nsegs + txq->ift_pad);
3764 		MPASS(pi.ipi_new_pidx != pidx);
3765 		MPASS(ndesc > 0);
3766 		txq->ift_in_use += ndesc;
3767 		txq->ift_db_pending += ndesc;
3768 
3769 		/*
3770 		 * We update the last software descriptor again here because there may
3771 		 * be a sentinel and/or there may be more mbufs than segments
3772 		 */
3773 		txq->ift_pidx = pi.ipi_new_pidx;
3774 		txq->ift_npending += pi.ipi_ndescs;
3775 
3776 		/*
3777 		 * Update packets / bytes sent
3778 		 */
3779 		if (flags & IFLIB_TSO) {
3780 			int hlen = pi.ipi_ehdrlen + pi.ipi_ip_hlen + pi.ipi_tcp_hlen;
3781 			int tsolen = pi.ipi_len - hlen;
3782 			int nsegs = (tsolen + pi.ipi_tso_segsz - 1) / pi.ipi_tso_segsz;
3783 			*obytes += tsolen + nsegs * hlen;
3784 			*opkts += nsegs;
3785 		} else {
3786 			*obytes += pi.ipi_len;
3787 			*opkts += 1;
3788 		}
3789 	} else {
3790 		*m_headp = m_head = iflib_remove_mbuf(txq);
3791 		if (err == EFBIG) {
3792 			txq->ift_txd_encap_efbig++;
3793 			if (remap < 2) {
3794 				remap = 1;
3795 				goto defrag;
3796 			}
3797 			goto defrag_failed;
3798 		}
3799 		/* mp_ring assumes ENOBUFS means we didn't consume the mbuf */
3800 		if (err == ENOBUFS && !ctx->ifc_sysctl_simple_tx)
3801 			err = ENOMEM;
3802 		goto out_with_error;
3803 	}
3804 	/*
3805 	 * err can't possibly be non-zero here, so we don't neet to test it
3806 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3807 	 */
3808 	return (err);
3809 
3810 defrag_failed:
3811 	err = ENOMEM;
3812 	txq->ift_mbuf_defrag_failed++;
3813 out_with_error:
3814 	txq->ift_map_failed++;
3815 	m_freem(*m_headp);
3816 	DBG_COUNTER_INC(tx_frees);
3817 	*m_headp = NULL;
3818 	DBG_COUNTER_INC(encap_txd_encap_fail);
3819 	return (err);
3820 }
3821 
3822 static void
iflib_tx_desc_free(iflib_txq_t txq,int n,struct mbuf ** m_defer)3823 iflib_tx_desc_free(iflib_txq_t txq, int n, struct mbuf **m_defer)
3824 {
3825 	uint32_t qsize, cidx, gen;
3826 	struct mbuf *m, **ifsd_m;
3827 	uintptr_t flags;
3828 
3829 	cidx = txq->ift_cidx;
3830 	gen = txq->ift_gen;
3831 	qsize = txq->ift_size;
3832 	ifsd_m =txq->ift_sds.ifsd_m;
3833 
3834 	while (n-- > 0) {
3835 		if ((m = IFLIB_GET_MBUF(ifsd_m[cidx])) != NULL) {
3836 			flags = IFLIB_GET_FLAGS(ifsd_m[cidx]);
3837 			MPASS(flags != 0);
3838 			if (flags & IFLIB_TSO) {
3839 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3840 				    txq->ift_sds.ifsd_tso_map[cidx],
3841 				    BUS_DMASYNC_POSTWRITE);
3842 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3843 				    txq->ift_sds.ifsd_tso_map[cidx]);
3844 			} else {
3845 				bus_dmamap_sync(txq->ift_buf_tag,
3846 				    txq->ift_sds.ifsd_map[cidx],
3847 				    BUS_DMASYNC_POSTWRITE);
3848 				bus_dmamap_unload(txq->ift_buf_tag,
3849 				    txq->ift_sds.ifsd_map[cidx]);
3850 			}
3851 			/* XXX we don't support any drivers that batch packets yet */
3852 			MPASS(m->m_nextpkt == NULL);
3853 			if (m_defer == NULL) {
3854 				m_freem(m);
3855 			} else if (m != NULL) {
3856 				*m_defer = m;
3857 				m_defer++;
3858 			}
3859 			ifsd_m[cidx] = NULL;
3860 #if MEMORY_LOGGING
3861 			txq->ift_dequeued++;
3862 #endif
3863 			DBG_COUNTER_INC(tx_frees);
3864 		}
3865 		if (__predict_false(++cidx == qsize)) {
3866 			cidx = 0;
3867 			gen = 0;
3868 		}
3869 	}
3870 	txq->ift_cidx = cidx;
3871 	txq->ift_gen = gen;
3872 }
3873 
3874 static __inline int
iflib_txq_can_reclaim(iflib_txq_t txq)3875 iflib_txq_can_reclaim(iflib_txq_t txq)
3876 {
3877 	int reclaim, thresh;
3878 
3879 	thresh = txq->ift_reclaim_thresh;
3880 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3881 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3882 
3883 	if (ticks <= (txq->ift_last_reclaim + txq->ift_reclaim_ticks) &&
3884 	    txq->ift_in_use < thresh)
3885 		return (false);
3886 	iflib_tx_credits_update(txq->ift_ctx, txq);
3887 	reclaim = DESC_RECLAIMABLE(txq);
3888 	if (reclaim <= thresh) {
3889 #ifdef INVARIANTS
3890 		if (iflib_verbose_debug) {
3891 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__,
3892 			    txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3893 			    reclaim, thresh);
3894 		}
3895 #endif
3896 		return (0);
3897 	}
3898 	return (reclaim);
3899 }
3900 
3901 static __inline void
_iflib_completed_tx_reclaim(iflib_txq_t txq,struct mbuf ** m_defer,int reclaim)3902 _iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer, int reclaim)
3903 {
3904 	txq->ift_last_reclaim = ticks;
3905 	iflib_tx_desc_free(txq, reclaim, m_defer);
3906 	txq->ift_cleaned += reclaim;
3907 	txq->ift_in_use -= reclaim;
3908 }
3909 
3910 static __inline int
iflib_completed_tx_reclaim(iflib_txq_t txq,struct mbuf ** m_defer)3911 iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer)
3912 {
3913 	int reclaim;
3914 
3915 	reclaim = iflib_txq_can_reclaim(txq);
3916 	if (reclaim == 0)
3917 		return (0);
3918 	_iflib_completed_tx_reclaim(txq, m_defer, reclaim);
3919 	return (reclaim);
3920 }
3921 
3922 /*
3923  * Reclaim any transmit descriptors possible, ignoring coalescing
3924  */
3925 static __inline void
iflib_completed_tx_reclaim_force(iflib_txq_t txq)3926 iflib_completed_tx_reclaim_force(iflib_txq_t txq)
3927 {
3928 	int reclaim;
3929 
3930 	iflib_tx_credits_update(txq->ift_ctx, txq);
3931 	reclaim = DESC_RECLAIMABLE(txq);
3932 	if (reclaim != 0)
3933 		_iflib_completed_tx_reclaim(txq, NULL, reclaim);
3934 }
3935 
3936 static struct mbuf **
_ring_peek_one(struct ifmp_ring * r,int cidx,int offset,int remaining)3937 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3938 {
3939 	int next, size;
3940 	struct mbuf **items;
3941 
3942 	size = r->size;
3943 	next = (cidx + CACHE_PTR_INCREMENT) & (size - 1);
3944 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3945 
3946 	prefetch(items[(cidx + offset) & (size - 1)]);
3947 	if (remaining > 1) {
3948 		prefetch2cachelines(&items[next]);
3949 		prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]);
3950 		prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]);
3951 		prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]);
3952 	}
3953 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)]));
3954 }
3955 
3956 static void
iflib_txq_check_drain(iflib_txq_t txq,int budget)3957 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3958 {
3959 
3960 	ifmp_ring_check_drainage(txq->ift_br, budget);
3961 }
3962 
3963 static uint32_t
iflib_txq_can_drain(struct ifmp_ring * r)3964 iflib_txq_can_drain(struct ifmp_ring *r)
3965 {
3966 	iflib_txq_t txq = r->cookie;
3967 	if_ctx_t ctx = txq->ift_ctx;
3968 
3969 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx))
3970 		return (1);
3971 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3972 	    BUS_DMASYNC_POSTREAD);
3973 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3974 	    false));
3975 }
3976 
3977 static uint32_t
iflib_txq_drain(struct ifmp_ring * r,uint32_t cidx,uint32_t pidx)3978 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3979 {
3980 	iflib_txq_t txq = r->cookie;
3981 	if_ctx_t ctx = txq->ift_ctx;
3982 	if_t ifp = ctx->ifc_ifp;
3983 	struct mbuf *m, **mp;
3984 	int avail, bytes_sent, consumed, count, err, i;
3985 	int mcast_sent, pkt_sent, reclaimed;
3986 	bool do_prefetch, rang, ring;
3987 
3988 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3989 			    !LINK_ACTIVE(ctx))) {
3990 		DBG_COUNTER_INC(txq_drain_notready);
3991 		return (0);
3992 	}
3993 	reclaimed = iflib_completed_tx_reclaim(txq, NULL);
3994 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3995 	avail = IDXDIFF(pidx, cidx, r->size);
3996 
3997 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3998 		/*
3999 		 * The driver is unloading so we need to free all pending packets.
4000 		 */
4001 		DBG_COUNTER_INC(txq_drain_flushing);
4002 		for (i = 0; i < avail; i++) {
4003 			if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq))
4004 				m_freem(r->items[(cidx + i) & (r->size - 1)]);
4005 			r->items[(cidx + i) & (r->size - 1)] = NULL;
4006 		}
4007 		return (avail);
4008 	}
4009 
4010 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
4011 		CALLOUT_LOCK(txq);
4012 		callout_stop(&txq->ift_timer);
4013 		CALLOUT_UNLOCK(txq);
4014 		DBG_COUNTER_INC(txq_drain_oactive);
4015 		return (0);
4016 	}
4017 
4018 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
4019 	count = MIN(avail, TX_BATCH_SIZE);
4020 #ifdef INVARIANTS
4021 	if (iflib_verbose_debug)
4022 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__,
4023 		    avail, ctx->ifc_flags, TXQ_AVAIL(txq));
4024 #endif
4025 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
4026 	err = 0;
4027 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx); i++) {
4028 		int rem = do_prefetch ? count - i : 0;
4029 
4030 		mp = _ring_peek_one(r, cidx, i, rem);
4031 		MPASS(mp != NULL && *mp != NULL);
4032 
4033 		/*
4034 		 * Completion interrupts will use the address of the txq
4035 		 * as a sentinel to enqueue _something_ in order to acquire
4036 		 * the lock on the mp_ring (there's no direct lock call).
4037 		 * We obviously whave to check for these sentinel cases
4038 		 * and skip them.
4039 		 */
4040 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4041 			consumed++;
4042 			continue;
4043 		}
4044 		err = iflib_encap(txq, mp, &bytes_sent, &pkt_sent);
4045 		if (__predict_false(err)) {
4046 			/* no room - bail out */
4047 			if (err == ENOBUFS)
4048 				break;
4049 			consumed++;
4050 			/* we can't send this packet - skip it */
4051 			continue;
4052 		}
4053 		consumed++;
4054 		m = *mp;
4055 		DBG_COUNTER_INC(tx_sent);
4056 		mcast_sent += !!(m->m_flags & M_MCAST);
4057 
4058 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4059 			break;
4060 		ETHER_BPF_MTAP(ifp, m);
4061 		rang = iflib_txd_db_check(txq, false);
4062 	}
4063 
4064 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4065 	ring = rang ? false  : (iflib_min_tx_latency | err | (!!txq->ift_reclaim_thresh));
4066 	iflib_txd_db_check(txq, ring);
4067 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4068 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4069 	if (mcast_sent)
4070 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4071 #ifdef INVARIANTS
4072 	if (iflib_verbose_debug)
4073 		printf("consumed=%d\n", consumed);
4074 #endif
4075 	return (consumed);
4076 }
4077 
4078 static uint32_t
iflib_txq_drain_always(struct ifmp_ring * r)4079 iflib_txq_drain_always(struct ifmp_ring *r)
4080 {
4081 	return (1);
4082 }
4083 
4084 static uint32_t
iflib_txq_drain_free(struct ifmp_ring * r,uint32_t cidx,uint32_t pidx)4085 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4086 {
4087 	int i, avail;
4088 	struct mbuf **mp;
4089 	iflib_txq_t txq;
4090 
4091 	txq = r->cookie;
4092 
4093 	CALLOUT_LOCK(txq);
4094 	callout_stop(&txq->ift_timer);
4095 	CALLOUT_UNLOCK(txq);
4096 
4097 	avail = IDXDIFF(pidx, cidx, r->size);
4098 	for (i = 0; i < avail; i++) {
4099 		mp = _ring_peek_one(r, cidx, i, avail - i);
4100 		if (__predict_false(*mp == (struct mbuf *)txq))
4101 			continue;
4102 		m_freem(*mp);
4103 		DBG_COUNTER_INC(tx_frees);
4104 	}
4105 	MPASS(ifmp_ring_is_stalled(r) == 0);
4106 	return (avail);
4107 }
4108 
4109 static void
iflib_ifmp_purge(iflib_txq_t txq)4110 iflib_ifmp_purge(iflib_txq_t txq)
4111 {
4112 	struct ifmp_ring *r;
4113 
4114 	r = txq->ift_br;
4115 	r->drain = iflib_txq_drain_free;
4116 	r->can_drain = iflib_txq_drain_always;
4117 
4118 	ifmp_ring_check_drainage(r, r->size);
4119 
4120 	r->drain = iflib_txq_drain;
4121 	r->can_drain = iflib_txq_can_drain;
4122 }
4123 
4124 static void
_task_fn_tx(void * context)4125 _task_fn_tx(void *context)
4126 {
4127 	iflib_txq_t txq = context;
4128 	if_ctx_t ctx = txq->ift_ctx;
4129 	if_t ifp = ctx->ifc_ifp;
4130 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4131 
4132 #ifdef IFLIB_DIAGNOSTICS
4133 	txq->ift_cpu_exec_count[curcpu]++;
4134 #endif
4135 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4136 		return;
4137 #ifdef DEV_NETMAP
4138 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4139 	    netmap_tx_irq(ifp, txq->ift_id))
4140 		goto skip_ifmp;
4141 #endif
4142         if (ctx->ifc_sysctl_simple_tx) {
4143                 mtx_lock(&txq->ift_mtx);
4144                 (void)iflib_completed_tx_reclaim(txq, NULL);
4145                 mtx_unlock(&txq->ift_mtx);
4146                 goto skip_ifmp;
4147         }
4148 #ifdef ALTQ
4149 	if (if_altq_is_enabled(ifp))
4150 		iflib_altq_if_start(ifp);
4151 #endif
4152 	if (txq->ift_db_pending)
4153 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4154 	else if (!abdicate)
4155 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4156 	/*
4157 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4158 	 */
4159 	if (abdicate)
4160 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4161 
4162 skip_ifmp:
4163 	if (ctx->ifc_flags & IFC_LEGACY)
4164 		IFDI_INTR_ENABLE(ctx);
4165 	else
4166 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4167 }
4168 
4169 static void
_task_fn_rx(void * context)4170 _task_fn_rx(void *context)
4171 {
4172 	iflib_rxq_t rxq = context;
4173 	if_ctx_t ctx = rxq->ifr_ctx;
4174 	uint8_t more;
4175 	uint16_t budget;
4176 #ifdef DEV_NETMAP
4177 	u_int work = 0;
4178 	int nmirq;
4179 #endif
4180 
4181 #ifdef IFLIB_DIAGNOSTICS
4182 	rxq->ifr_cpu_exec_count[curcpu]++;
4183 #endif
4184 	DBG_COUNTER_INC(task_fn_rxs);
4185 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4186 		return;
4187 #ifdef DEV_NETMAP
4188 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4189 	if (nmirq != NM_IRQ_PASS) {
4190 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4191 		goto skip_rxeof;
4192 	}
4193 #endif
4194 	budget = ctx->ifc_sysctl_rx_budget;
4195 	if (budget == 0)
4196 		budget = 16;	/* XXX */
4197 	more = iflib_rxeof(rxq, budget);
4198 #ifdef DEV_NETMAP
4199 skip_rxeof:
4200 #endif
4201 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4202 		if (ctx->ifc_flags & IFC_LEGACY)
4203 			IFDI_INTR_ENABLE(ctx);
4204 		else
4205 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4206 		DBG_COUNTER_INC(rx_intr_enables);
4207 	}
4208 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4209 		return;
4210 
4211 	if (more & IFLIB_RXEOF_MORE)
4212 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4213 	else if (more & IFLIB_RXEOF_EMPTY)
4214 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4215 }
4216 
4217 static void
_task_fn_admin(void * context,int pending)4218 _task_fn_admin(void *context, int pending)
4219 {
4220 	if_ctx_t ctx = context;
4221 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4222 	iflib_txq_t txq;
4223 	int i;
4224 	bool oactive, running, do_reset, do_reset_if_up, do_watchdog;
4225 	bool in_detach;
4226 
4227 	STATE_LOCK(ctx);
4228 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4229 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4230 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4231 	do_reset_if_up = (ctx->ifc_flags & IFC_DO_RESET_IF_UP);
4232 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4233 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4234 	ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_RESET_IF_UP |
4235 	    IFC_DO_WATCHDOG);
4236 	STATE_UNLOCK(ctx);
4237 
4238 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4239 		return;
4240 	if (in_detach)
4241 		return;
4242 	KFAIL_POINT_CODE_COND(_debug_fail_point_iflib,
4243 	    admin_task_after_detach_check,
4244 	    iflib_admin_task_fail_device[0] != '\0' &&
4245 	    strcmp(device_get_nameunit(ctx->ifc_dev),
4246 	    iflib_admin_task_fail_device) == 0, FAIL_POINT_NONSLEEPABLE, {});
4247 
4248 	CTX_LOCK(ctx);
4249 	if (!do_reset && do_reset_if_up &&
4250 	    (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0)
4251 		do_reset = true;
4252 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4253 		CALLOUT_LOCK(txq);
4254 		callout_stop(&txq->ift_timer);
4255 		CALLOUT_UNLOCK(txq);
4256 	}
4257 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4258 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4259 	if (do_watchdog) {
4260 		ctx->ifc_tx_watchdog_events++;
4261 		IFDI_WATCHDOG_RESET(ctx);
4262 	}
4263 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4264 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4265 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4266 		    txq->ift_timer.c_cpu);
4267 	}
4268 	IFDI_LINK_INTR_ENABLE(ctx);
4269 	if (do_reset)
4270 		iflib_if_init_locked(ctx);
4271 	CTX_UNLOCK(ctx);
4272 
4273 	if (LINK_ACTIVE(ctx) == 0)
4274 		return;
4275 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4276 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4277 }
4278 
4279 static void
_task_fn_iov(void * context,int pending)4280 _task_fn_iov(void *context, int pending)
4281 {
4282 	if_ctx_t ctx = context;
4283 
4284 	if (iflib_in_detach(ctx))
4285 		return;
4286 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4287 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4288 		return;
4289 
4290 	CTX_LOCK(ctx);
4291 	IFDI_VFLR_HANDLE(ctx);
4292 	CTX_UNLOCK(ctx);
4293 }
4294 
4295 static int
iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)4296 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4297 {
4298 	int err;
4299 	if_int_delay_info_t info;
4300 	if_ctx_t ctx;
4301 
4302 	info = (if_int_delay_info_t)arg1;
4303 	ctx = info->iidi_ctx;
4304 	info->iidi_req = req;
4305 	info->iidi_oidp = oidp;
4306 	CTX_LOCK(ctx);
4307 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4308 	CTX_UNLOCK(ctx);
4309 	return (err);
4310 }
4311 
4312 /*********************************************************************
4313  *
4314  *  IFNET FUNCTIONS
4315  *
4316  **********************************************************************/
4317 
4318 static void
iflib_if_init_locked(if_ctx_t ctx)4319 iflib_if_init_locked(if_ctx_t ctx)
4320 {
4321 	iflib_stop(ctx);
4322 	iflib_init_locked(ctx);
4323 }
4324 
4325 static void
iflib_if_init(void * arg)4326 iflib_if_init(void *arg)
4327 {
4328 	if_ctx_t ctx = arg;
4329 
4330 	CTX_LOCK(ctx);
4331 	iflib_if_init_locked(ctx);
4332 	CTX_UNLOCK(ctx);
4333 }
4334 
4335 static int
iflib_if_transmit(if_t ifp,struct mbuf * m)4336 iflib_if_transmit(if_t ifp, struct mbuf *m)
4337 {
4338 	if_ctx_t ctx = if_getsoftc(ifp);
4339 	iflib_txq_t txq;
4340 	int err, qidx;
4341 	int abdicate;
4342 
4343 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4344 		DBG_COUNTER_INC(tx_frees);
4345 		m_freem(m);
4346 		return (ENETDOWN);
4347 	}
4348 
4349 	MPASS(m->m_nextpkt == NULL);
4350 	/* ALTQ-enabled interfaces always use queue 0. */
4351 	qidx = 0;
4352 	/* Use driver-supplied queue selection method if it exists */
4353 	if (ctx->isc_txq_select_v2) {
4354 		struct if_pkt_info pi;
4355 		uint64_t early_pullups = 0;
4356 		memset(&pi, 0, sizeof(pi));
4357 
4358 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4359 		if (__predict_false(err != 0)) {
4360 			/* Assign pullups for bad pkts to default queue */
4361 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4362 			DBG_COUNTER_INC(encap_txd_encap_fail);
4363 			return (err);
4364 		}
4365 		/* Let driver make queueing decision */
4366 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4367 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4368 	}
4369 	/* Backwards compatibility w/ simpler queue select */
4370 	else if (ctx->isc_txq_select)
4371 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4372 	/* If not, use iflib's standard method */
4373 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4374 		qidx = QIDX(ctx, m);
4375 
4376 	/* Set TX queue */
4377 	txq = &ctx->ifc_txqs[qidx];
4378 
4379 #ifdef DRIVER_BACKPRESSURE
4380 	if (txq->ift_closed) {
4381 		while (m != NULL) {
4382 			next = m->m_nextpkt;
4383 			m->m_nextpkt = NULL;
4384 			m_freem(m);
4385 			DBG_COUNTER_INC(tx_frees);
4386 			m = next;
4387 		}
4388 		return (ENOBUFS);
4389 	}
4390 #endif
4391 #ifdef notyet
4392 	qidx = count = 0;
4393 	mp = marr;
4394 	next = m;
4395 	do {
4396 		count++;
4397 		next = next->m_nextpkt;
4398 	} while (next != NULL);
4399 
4400 	if (count > nitems(marr))
4401 		if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4402 			/* XXX check nextpkt */
4403 			m_freem(m);
4404 			/* XXX simplify for now */
4405 			DBG_COUNTER_INC(tx_frees);
4406 			return (ENOBUFS);
4407 		}
4408 	for (next = m, i = 0; next != NULL; i++) {
4409 		mp[i] = next;
4410 		next = next->m_nextpkt;
4411 		mp[i]->m_nextpkt = NULL;
4412 	}
4413 #endif
4414 	DBG_COUNTER_INC(tx_seen);
4415 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4416 
4417 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4418 
4419 	if (abdicate)
4420 		GROUPTASK_ENQUEUE(&txq->ift_task);
4421 	if (err) {
4422 		if (!abdicate)
4423 			GROUPTASK_ENQUEUE(&txq->ift_task);
4424 		/* support forthcoming later */
4425 #ifdef DRIVER_BACKPRESSURE
4426 		txq->ift_closed = TRUE;
4427 #endif
4428 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4429 		m_freem(m);
4430 		DBG_COUNTER_INC(tx_frees);
4431 		if (err == ENOBUFS)
4432 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
4433 		else
4434 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
4435 	}
4436 
4437 	return (err);
4438 }
4439 
4440 #ifdef ALTQ
4441 /*
4442  * The overall approach to integrating iflib with ALTQ is to continue to use
4443  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4444  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4445  * is redundant/unnecessary, but doing so minimizes the amount of
4446  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4447  * redundantly queueing to an intermediate mp_ring is swamped by the
4448  * performance limitations inherent in using ALTQ.
4449  *
4450  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4451  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4452  * given interface.  If ALTQ is enabled for an interface, then all
4453  * transmitted packets for that interface will be submitted to the ALTQ
4454  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4455  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4456  * update stats that the iflib machinery handles, and which is sensitve to
4457  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4458  * will be installed as the start routine for use by ALTQ facilities that
4459  * need to trigger queue drains on a scheduled basis.
4460  *
4461  */
4462 static void
iflib_altq_if_start(if_t ifp)4463 iflib_altq_if_start(if_t ifp)
4464 {
4465 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4466 	struct mbuf *m;
4467 
4468 	IFQ_LOCK(ifq);
4469 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4470 	while (m != NULL) {
4471 		iflib_if_transmit(ifp, m);
4472 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4473 	}
4474 	IFQ_UNLOCK(ifq);
4475 }
4476 
4477 static int
iflib_altq_if_transmit(if_t ifp,struct mbuf * m)4478 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4479 {
4480 	int err;
4481 
4482 	if (if_altq_is_enabled(ifp)) {
4483 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4484 		if (err == 0)
4485 			iflib_altq_if_start(ifp);
4486 	} else
4487 		err = iflib_if_transmit(ifp, m);
4488 
4489 	return (err);
4490 }
4491 #endif /* ALTQ */
4492 
4493 static void
iflib_if_qflush(if_t ifp)4494 iflib_if_qflush(if_t ifp)
4495 {
4496 	if_ctx_t ctx = if_getsoftc(ifp);
4497 	iflib_txq_t txq = ctx->ifc_txqs;
4498 	int i;
4499 
4500 	STATE_LOCK(ctx);
4501 	ctx->ifc_flags |= IFC_QFLUSH;
4502 	STATE_UNLOCK(ctx);
4503 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4504 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4505 			iflib_txq_check_drain(txq, 0);
4506 	STATE_LOCK(ctx);
4507 	ctx->ifc_flags &= ~IFC_QFLUSH;
4508 	STATE_UNLOCK(ctx);
4509 
4510 	/*
4511 	 * When ALTQ is enabled, this will also take care of purging the
4512 	 * ALTQ queue(s).
4513 	 */
4514 	if_qflush(ifp);
4515 }
4516 
4517 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4518 		    IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4519 		    IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4520 		    IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4521 
4522 static int
iflib_if_ioctl(if_t ifp,u_long command,caddr_t data)4523 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4524 {
4525 	if_ctx_t ctx = if_getsoftc(ifp);
4526 	struct ifreq	*ifr = (struct ifreq *)data;
4527 #if defined(INET) || defined(INET6)
4528 	struct ifaddr	*ifa = (struct ifaddr *)data;
4529 #endif
4530 	bool		avoid_reset = false;
4531 	int		err = 0, reinit = 0, bits;
4532 
4533 	switch (command) {
4534 	case SIOCSIFADDR:
4535 #ifdef INET
4536 		if (ifa->ifa_addr->sa_family == AF_INET)
4537 			avoid_reset = true;
4538 #endif
4539 #ifdef INET6
4540 		if (ifa->ifa_addr->sa_family == AF_INET6)
4541 			avoid_reset = true;
4542 #endif
4543 		/*
4544 		 * Calling init results in link renegotiation,
4545 		 * so we avoid doing it when possible.
4546 		 */
4547 		if (avoid_reset) {
4548 			if_setflagbits(ifp, IFF_UP, 0);
4549 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4550 				reinit = 1;
4551 #ifdef INET
4552 			if (!(if_getflags(ifp) & IFF_NOARP))
4553 				arp_ifinit(ifp, ifa);
4554 #endif
4555 		} else
4556 			err = ether_ioctl(ifp, command, data);
4557 		break;
4558 	case SIOCSIFMTU:
4559 		CTX_LOCK(ctx);
4560 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4561 			CTX_UNLOCK(ctx);
4562 			break;
4563 		}
4564 		bits = if_getdrvflags(ifp);
4565 		/* stop the driver and free any clusters before proceeding */
4566 		iflib_stop(ctx);
4567 
4568 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4569 			STATE_LOCK(ctx);
4570 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4571 				ctx->ifc_flags |= IFC_MULTISEG;
4572 			else
4573 				ctx->ifc_flags &= ~IFC_MULTISEG;
4574 			STATE_UNLOCK(ctx);
4575 			err = if_setmtu(ifp, ifr->ifr_mtu);
4576 		}
4577 		iflib_init_locked(ctx);
4578 		STATE_LOCK(ctx);
4579 		/* Preserve the stopped state reported by iflib_init_failed(). */
4580 		if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4581 			if_setdrvflags(ifp, bits);
4582 		STATE_UNLOCK(ctx);
4583 		CTX_UNLOCK(ctx);
4584 		break;
4585 	case SIOCSIFFLAGS:
4586 		CTX_LOCK(ctx);
4587 		if (if_getflags(ifp) & IFF_UP) {
4588 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4589 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4590 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4591 					CTX_UNLOCK(ctx);
4592 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4593 					CTX_LOCK(ctx);
4594 				}
4595 			} else
4596 				reinit = 1;
4597 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4598 			iflib_stop(ctx);
4599 		}
4600 		ctx->ifc_if_flags = if_getflags(ifp);
4601 		CTX_UNLOCK(ctx);
4602 		break;
4603 	case SIOCADDMULTI:
4604 	case SIOCDELMULTI:
4605 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4606 			CTX_LOCK(ctx);
4607 			IFDI_INTR_DISABLE(ctx);
4608 			IFDI_MULTI_SET(ctx);
4609 			IFDI_INTR_ENABLE(ctx);
4610 			CTX_UNLOCK(ctx);
4611 		}
4612 		break;
4613 	case SIOCSIFMEDIA:
4614 		CTX_LOCK(ctx);
4615 		IFDI_MEDIA_SET(ctx);
4616 		CTX_UNLOCK(ctx);
4617 		/* FALLTHROUGH */
4618 	case SIOCGIFMEDIA:
4619 	case SIOCGIFXMEDIA:
4620 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4621 		break;
4622 	case SIOCGI2C:
4623 		/* FALLTHROUGH */
4624 	case SIOCGI2CPB:
4625 	{
4626 		struct ifi2creq i2c;
4627 		if_shared_ctx_t sctx = ctx->ifc_sctx;
4628 
4629 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4630 		if (err != 0)
4631 			break;
4632 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4633 			err = EINVAL;
4634 			break;
4635 		}
4636 		if (i2c.len > sizeof(i2c.data)) {
4637 			err = EINVAL;
4638 			break;
4639 		}
4640 		if (command == SIOCGI2C) {
4641 			i2c.page = i2c.bank = 0;
4642 		} else if ((sctx->isc_flags & IFLIB_I2C_PAGE_BANK) == 0) {
4643 			err = EINVAL;
4644 			break;
4645 		}
4646 
4647 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4648 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4649 			    sizeof(i2c));
4650 		break;
4651 	}
4652 	case SIOCSIFCAP:
4653 	{
4654 		int mask, setmask, oldmask;
4655 
4656 		oldmask = if_getcapenable(ifp);
4657 		mask = ifr->ifr_reqcap ^ oldmask;
4658 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4659 		setmask = 0;
4660 #ifdef TCP_OFFLOAD
4661 		setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6);
4662 #endif
4663 		setmask |= (mask & IFCAP_FLAGS);
4664 		setmask |= (mask & IFCAP_WOL);
4665 
4666 		/*
4667 		 * If any RX csum has changed, change all the ones that
4668 		 * are supported by the driver.
4669 		 */
4670 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4671 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4672 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4673 		}
4674 
4675 		/*
4676 		 * want to ensure that traffic has stopped before we change any of the flags
4677 		 */
4678 		if (setmask) {
4679 			CTX_LOCK(ctx);
4680 			bits = if_getdrvflags(ifp);
4681 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4682 				iflib_stop(ctx);
4683 			STATE_LOCK(ctx);
4684 			if_togglecapenable(ifp, setmask);
4685 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4686 			STATE_UNLOCK(ctx);
4687 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4688 				iflib_init_locked(ctx);
4689 			STATE_LOCK(ctx);
4690 			if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4691 				if_setdrvflags(ifp, bits);
4692 			STATE_UNLOCK(ctx);
4693 			CTX_UNLOCK(ctx);
4694 		}
4695 		if_vlancap(ifp);
4696 		break;
4697 	}
4698 	case SIOCGPRIVATE_0:
4699 	case SIOCSDRVSPEC:
4700 	case SIOCGDRVSPEC:
4701 		CTX_LOCK(ctx);
4702 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4703 		CTX_UNLOCK(ctx);
4704 		break;
4705 	case SIOCGIFDOWNREASON:
4706 		CTX_LOCK(ctx);
4707 		err = IFDI_GET_DOWNREASON(ctx, (struct ifdownreason *)data);
4708 		CTX_UNLOCK(ctx);
4709 		break;
4710 	default:
4711 		err = ether_ioctl(ifp, command, data);
4712 		break;
4713 	}
4714 	if (reinit)
4715 		iflib_if_init(ctx);
4716 	return (err);
4717 }
4718 
4719 static uint64_t
iflib_if_get_counter(if_t ifp,ift_counter cnt)4720 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4721 {
4722 	if_ctx_t ctx = if_getsoftc(ifp);
4723 
4724 	return (IFDI_GET_COUNTER(ctx, cnt));
4725 }
4726 
4727 /*********************************************************************
4728  *
4729  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4730  *
4731  **********************************************************************/
4732 
4733 static void
iflib_vlan_register(void * arg,if_t ifp,uint16_t vtag)4734 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4735 {
4736 	if_ctx_t ctx = if_getsoftc(ifp);
4737 
4738 	if ((void *)ctx != arg)
4739 		return;
4740 
4741 	if ((vtag == 0) || (vtag > 4095))
4742 		return;
4743 
4744 	if (iflib_in_detach(ctx))
4745 		return;
4746 
4747 	CTX_LOCK(ctx);
4748 	/* Driver may need all untagged packets to be flushed */
4749 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4750 		iflib_stop(ctx);
4751 	IFDI_VLAN_REGISTER(ctx, vtag);
4752 	/* Re-init to load the changes, if required */
4753 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4754 		iflib_init_locked(ctx);
4755 	CTX_UNLOCK(ctx);
4756 }
4757 
4758 static void
iflib_vlan_unregister(void * arg,if_t ifp,uint16_t vtag)4759 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4760 {
4761 	if_ctx_t ctx = if_getsoftc(ifp);
4762 
4763 	if ((void *)ctx != arg)
4764 		return;
4765 
4766 	if ((vtag == 0) || (vtag > 4095))
4767 		return;
4768 
4769 	CTX_LOCK(ctx);
4770 	/* Driver may need all tagged packets to be flushed */
4771 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4772 		iflib_stop(ctx);
4773 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4774 	/* Re-init to load the changes, if required */
4775 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4776 		iflib_init_locked(ctx);
4777 	CTX_UNLOCK(ctx);
4778 }
4779 
4780 static void
_task_fn_led(void * context,int pending __unused)4781 _task_fn_led(void *context, int pending __unused)
4782 {
4783 	if_ctx_t ctx = context;
4784 	bool in_detach;
4785 	int onoff;
4786 
4787 	STATE_LOCK(ctx);
4788 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4789 	onoff = ctx->ifc_led_state;
4790 	STATE_UNLOCK(ctx);
4791 	if (in_detach)
4792 		return;
4793 
4794 	CTX_LOCK(ctx);
4795 	IFDI_LED_FUNC(ctx, onoff);
4796 	CTX_UNLOCK(ctx);
4797 }
4798 
4799 static void
iflib_led_func(void * arg,int onoff)4800 iflib_led_func(void *arg, int onoff)
4801 {
4802 	if_ctx_t ctx = arg;
4803 	bool in_detach;
4804 
4805 	/* led(4) may invoke this callback from a non-sleepable callout. */
4806 	STATE_LOCK(ctx);
4807 	ctx->ifc_led_state = onoff;
4808 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4809 	STATE_UNLOCK(ctx);
4810 	if (!in_detach)
4811 		taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_led_task);
4812 }
4813 
4814 /*********************************************************************
4815  *
4816  *  BUS FUNCTION DEFINITIONS
4817  *
4818  **********************************************************************/
4819 
4820 int
iflib_device_probe(device_t dev)4821 iflib_device_probe(device_t dev)
4822 {
4823 	const pci_vendor_info_t *ent;
4824 	if_shared_ctx_t sctx;
4825 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4826 	uint16_t pci_vendor_id;
4827 
4828 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4829 		return (ENOTSUP);
4830 
4831 	pci_vendor_id = pci_get_vendor(dev);
4832 	pci_device_id = pci_get_device(dev);
4833 	pci_subvendor_id = pci_get_subvendor(dev);
4834 	pci_subdevice_id = pci_get_subdevice(dev);
4835 	pci_rev_id = pci_get_revid(dev);
4836 	if (sctx->isc_parse_devinfo != NULL)
4837 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4838 
4839 	ent = sctx->isc_vendor_info;
4840 	while (ent->pvi_vendor_id != 0) {
4841 		if (pci_vendor_id != ent->pvi_vendor_id) {
4842 			ent++;
4843 			continue;
4844 		}
4845 		if ((pci_device_id == ent->pvi_device_id) &&
4846 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4847 		     (ent->pvi_subvendor_id == 0)) &&
4848 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4849 		     (ent->pvi_subdevice_id == 0)) &&
4850 		    ((pci_rev_id == ent->pvi_rev_id) ||
4851 		     (ent->pvi_rev_id == 0))) {
4852 			device_set_desc_copy(dev, ent->pvi_name);
4853 			/* this needs to be changed to zero if the bus probing code
4854 			 * ever stops re-probing on best match because the sctx
4855 			 * may have its values over written by register calls
4856 			 * in subsequent probes
4857 			 */
4858 			return (BUS_PROBE_DEFAULT);
4859 		}
4860 		ent++;
4861 	}
4862 	return (ENXIO);
4863 }
4864 
4865 int
iflib_device_probe_vendor(device_t dev)4866 iflib_device_probe_vendor(device_t dev)
4867 {
4868 	int probe;
4869 
4870 	probe = iflib_device_probe(dev);
4871 	if (probe == BUS_PROBE_DEFAULT)
4872 		return (BUS_PROBE_VENDOR);
4873 	else
4874 		return (probe);
4875 }
4876 
4877 static void
iflib_reset_qvalues(if_ctx_t ctx)4878 iflib_reset_qvalues(if_ctx_t ctx)
4879 {
4880 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4881 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4882 	device_t dev = ctx->ifc_dev;
4883 	int i;
4884 
4885 	if (ctx->ifc_sysctl_ntxqs != 0)
4886 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4887 	if (ctx->ifc_sysctl_nrxqs != 0)
4888 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4889 
4890 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4891 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4892 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4893 		else
4894 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4895 	}
4896 
4897 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4898 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4899 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4900 		else
4901 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4902 	}
4903 
4904 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4905 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4906 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4907 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4908 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4909 		}
4910 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4911 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4912 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4913 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4914 		}
4915 		if (!powerof2(scctx->isc_nrxd[i])) {
4916 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4917 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4918 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4919 		}
4920 	}
4921 
4922 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4923 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4924 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4925 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4926 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4927 		}
4928 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4929 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4930 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4931 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4932 		}
4933 		if (!powerof2(scctx->isc_ntxd[i])) {
4934 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4935 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4936 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4937 		}
4938 	}
4939 	scctx->isc_tx_pad = 2;
4940 }
4941 
4942 static void
iflib_add_pfil(if_ctx_t ctx)4943 iflib_add_pfil(if_ctx_t ctx)
4944 {
4945 	struct pfil_head *pfil;
4946 	struct pfil_head_args pa;
4947 	iflib_rxq_t rxq;
4948 	int i;
4949 
4950 	pa.pa_version = PFIL_VERSION;
4951 	pa.pa_flags = PFIL_IN;
4952 	pa.pa_type = PFIL_TYPE_ETHERNET;
4953 	pa.pa_headname = if_name(ctx->ifc_ifp);
4954 	pfil = pfil_head_register(&pa);
4955 
4956 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4957 		rxq->pfil = pfil;
4958 	}
4959 }
4960 
4961 static void
iflib_rem_pfil(if_ctx_t ctx)4962 iflib_rem_pfil(if_ctx_t ctx)
4963 {
4964 	struct pfil_head *pfil;
4965 	iflib_rxq_t rxq;
4966 	int i;
4967 
4968 	rxq = ctx->ifc_rxqs;
4969 	pfil = rxq->pfil;
4970 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4971 		rxq->pfil = NULL;
4972 	}
4973 	pfil_head_unregister(pfil);
4974 }
4975 
4976 
4977 /*
4978  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4979  * cpuid and wrapping as necessary.
4980  */
4981 static unsigned int
cpuid_advance(if_ctx_t ctx,unsigned int cpuid,unsigned int n)4982 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4983 {
4984 	unsigned int first_valid;
4985 	unsigned int last_valid;
4986 
4987 	/* cpuid should always be in the valid set */
4988 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4989 
4990 	/* valid set should never be empty */
4991 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4992 
4993 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4994 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4995 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4996 	while (n > 0) {
4997 		do {
4998 			cpuid++;
4999 			if (cpuid > last_valid)
5000 				cpuid = first_valid;
5001 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
5002 		n--;
5003 	}
5004 
5005 	return (cpuid);
5006 }
5007 
5008 /*
5009  * CPU mapping behaviors
5010  * ---------------------
5011  * 'separate txrx' refers to the separate_txrx sysctl
5012  * 'use logical' refers to the use_logical_cores sysctl
5013  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
5014  *
5015  *  separate     use     INTR
5016  *    txrx     logical   CPUS   result
5017  * ---------- --------- ------ ------------------------------------------------
5018  *     -          -       X     RX and TX queues mapped to consecutive physical
5019  *                              cores with RX/TX pairs on same core and excess
5020  *                              of either following
5021  *     -          X       X     RX and TX queues mapped to consecutive cores
5022  *                              of any type with RX/TX pairs on same core and
5023  *                              excess of either following
5024  *     X          -       X     RX and TX queues mapped to consecutive physical
5025  *                              cores; all RX then all TX
5026  *     X          X       X     RX queues mapped to consecutive physical cores
5027  *                              first, then TX queues mapped to L2 neighbor of
5028  *                              the corresponding RX queue if one exists,
5029  *                              otherwise to consecutive physical cores
5030  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5031  *                              any type with RX/TX pairs on same core and excess
5032  *                              of either following
5033  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5034  *                              any type; all RX then all TX
5035  */
5036 static unsigned int
get_cpuid_for_queue(if_ctx_t ctx,unsigned int base_cpuid,unsigned int qid,bool is_tx)5037 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5038     bool is_tx)
5039 {
5040 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5041 	unsigned int core_index;
5042 
5043 	if (ctx->ifc_sysctl_separate_txrx) {
5044 		/*
5045 		 * When using separate CPUs for TX and RX, the assignment
5046 		 * will always be of a consecutive CPU out of the set of
5047 		 * context CPUs, except for the specific case where the
5048 		 * context CPUs are phsyical cores, the use of logical cores
5049 		 * has been enabled, the assignment is for TX, the TX qid
5050 		 * corresponds to an RX qid, and the CPU assigned to the
5051 		 * corresponding RX queue has an L2 neighbor.
5052 		 */
5053 		if (ctx->ifc_sysctl_use_logical_cores &&
5054 		    ctx->ifc_cpus_are_physical_cores &&
5055 		    is_tx && qid < scctx->isc_nrxqsets) {
5056 			int l2_neighbor;
5057 			unsigned int rx_cpuid;
5058 
5059 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5060 			l2_neighbor = sched_find_l2_neighbor(rx_cpuid);
5061 			if (l2_neighbor != -1) {
5062 				return (l2_neighbor);
5063 			}
5064 			/*
5065 			 * ... else fall through to the normal
5066 			 * consecutive-after-RX assignment scheme.
5067 			 *
5068 			 * Note that we are assuming that all RX queue CPUs
5069 			 * have an L2 neighbor, or all do not.  If a mixed
5070 			 * scenario is possible, we will have to keep track
5071 			 * separately of how many queues prior to this one
5072 			 * were not able to be assigned to an L2 neighbor.
5073 			 */
5074 		}
5075 		if (is_tx)
5076 			core_index = scctx->isc_nrxqsets + qid;
5077 		else
5078 			core_index = qid;
5079 	} else {
5080 		core_index = qid;
5081 	}
5082 
5083 	return (cpuid_advance(ctx, base_cpuid, core_index));
5084 }
5085 
5086 static uint16_t
get_ctx_core_offset(if_ctx_t ctx)5087 get_ctx_core_offset(if_ctx_t ctx)
5088 {
5089 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5090 	struct cpu_offset *op;
5091 	cpuset_t assigned_cpus;
5092 	unsigned int cores_consumed;
5093 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5094 	unsigned int first_valid;
5095 	unsigned int last_valid;
5096 	unsigned int i;
5097 
5098 	MPASS(!ctx->ifc_core_offset_ref);
5099 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5100 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5101 
5102 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5103 		/*
5104 		 * Align the user-chosen base CPU ID to the next valid CPU
5105 		 * for this device.  If the chosen base CPU ID is smaller
5106 		 * than the first valid CPU or larger than the last valid
5107 		 * CPU, we assume the user does not know what the valid
5108 		 * range is for this device and is thinking in terms of a
5109 		 * zero-based reference frame, and so we shift the given
5110 		 * value into the valid range (and wrap accordingly) so the
5111 		 * intent is translated to the proper frame of reference.
5112 		 * If the base CPU ID is within the valid first/last, but
5113 		 * does not correspond to a valid CPU, it is advanced to the
5114 		 * next valid CPU (wrapping if necessary).
5115 		 */
5116 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5117 			/* shift from zero-based to first_valid-based */
5118 			base_cpuid += first_valid;
5119 			/* wrap to range [first_valid, last_valid] */
5120 			base_cpuid = (base_cpuid - first_valid) %
5121 			    (last_valid - first_valid + 1);
5122 		}
5123 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5124 			/*
5125 			 * base_cpuid is in [first_valid, last_valid], but
5126 			 * not a member of the valid set.  In this case,
5127 			 * there will always be a member of the valid set
5128 			 * with a CPU ID that is greater than base_cpuid,
5129 			 * and we simply advance to it.
5130 			 */
5131 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5132 				base_cpuid++;
5133 		}
5134 		return (base_cpuid);
5135 	}
5136 
5137 	/*
5138 	 * Determine how many cores will be consumed by performing the CPU
5139 	 * assignments and counting how many of the assigned CPUs correspond
5140 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5141 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5142 	 * the set of context CPUs.
5143 	 *
5144 	 * Note not all assigned CPUs will be in the set of context CPUs
5145 	 * when separate CPUs are being allocated to TX and RX queues,
5146 	 * assignment to logical cores has been enabled, the set of context
5147 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5148 	 * neighbors of CPUs that RX queues have been mapped to - in this
5149 	 * case we do only want to count how many CPUs in the set of context
5150 	 * CPUs have been consumed, as that determines the next CPU in that
5151 	 * set to start allocating at for the next device for which
5152 	 * core_offset is not set.
5153 	 */
5154 	CPU_ZERO(&assigned_cpus);
5155 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5156 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5157 		    &assigned_cpus);
5158 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5159 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5160 		    &assigned_cpus);
5161 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5162 	cores_consumed = CPU_COUNT(&assigned_cpus);
5163 
5164 	mtx_lock(&cpu_offset_mtx);
5165 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5166 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5167 			base_cpuid = op->next_cpuid;
5168 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5169 			    cores_consumed);
5170 			MPASS(op->refcount < UINT_MAX);
5171 			op->refcount++;
5172 			ctx->ifc_core_offset_ref = true;
5173 			break;
5174 		}
5175 	}
5176 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5177 		base_cpuid = first_valid;
5178 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5179 		    M_NOWAIT | M_ZERO);
5180 		if (op == NULL) {
5181 			device_printf(ctx->ifc_dev,
5182 			    "allocation for cpu offset failed.\n");
5183 		} else {
5184 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5185 			    cores_consumed);
5186 			op->refcount = 1;
5187 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5188 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5189 			ctx->ifc_core_offset_ref = true;
5190 		}
5191 	}
5192 	mtx_unlock(&cpu_offset_mtx);
5193 
5194 	return (base_cpuid);
5195 }
5196 
5197 static void
unref_ctx_core_offset(if_ctx_t ctx)5198 unref_ctx_core_offset(if_ctx_t ctx)
5199 {
5200 	struct cpu_offset *op, *top;
5201 
5202 	if (!ctx->ifc_core_offset_ref)
5203 		return;
5204 
5205 	mtx_lock(&cpu_offset_mtx);
5206 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5207 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5208 			MPASS(op->refcount > 0);
5209 			op->refcount--;
5210 			if (op->refcount == 0) {
5211 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5212 				free(op, M_IFLIB);
5213 			}
5214 			ctx->ifc_core_offset_ref = false;
5215 			break;
5216 		}
5217 	}
5218 	mtx_unlock(&cpu_offset_mtx);
5219 	MPASS(!ctx->ifc_core_offset_ref);
5220 }
5221 
5222 static bool
iflib_register_fail_device_matches(device_t dev)5223 iflib_register_fail_device_matches(device_t dev)
5224 {
5225 	const char *nameunit;
5226 
5227 	nameunit = device_get_nameunit(dev);
5228 	return (iflib_register_fail_device[0] != '\0' && nameunit != NULL &&
5229 	    strcmp(nameunit, iflib_register_fail_device) == 0);
5230 }
5231 
5232 #define	IFLIB_REGISTER_FAIL_POINT(_dev, _name, _error, _label) do { \
5233 	KFAIL_POINT_CODE_COND(_debug_fail_point_iflib, _name, \
5234 	    iflib_register_fail_device_matches((_dev)), \
5235 	    FAIL_POINT_NONSLEEPABLE, { \
5236 		(_error) = RETURN_VALUE; \
5237 		if ((_error) <= 0) \
5238 			(_error) = EIO; \
5239 		device_printf((_dev), \
5240 		    "injecting iflib registration failure at %s: %d\n", \
5241 		    #_name, (_error)); \
5242 		goto _label; \
5243 	}); \
5244 } while (0)
5245 
5246 int
iflib_device_register(device_t dev,void * sc,if_shared_ctx_t sctx,if_ctx_t * ctxp)5247 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5248 {
5249 	if_ctx_t ctx;
5250 	if_t ifp;
5251 	if_softc_ctx_t scctx;
5252 	kobjop_desc_t kobj_desc;
5253 	kobj_method_t *kobj_method;
5254 	bool attach_pre_succeeded, intr_allocated, queues_allocated;
5255 	int err, msix, rid;
5256 #ifdef PCI_IOV
5257 	int iov_error;
5258 #endif
5259 	int num_txd, num_rxd;
5260 	char namebuf[TASKQUEUE_NAMELEN];
5261 
5262 	attach_pre_succeeded = false;
5263 	intr_allocated = false;
5264 	queues_allocated = false;
5265 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO);
5266 
5267 	if (sc == NULL) {
5268 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO);
5269 		device_set_softc(dev, ctx);
5270 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5271 	}
5272 
5273 	ctx->ifc_sctx = sctx;
5274 	ctx->ifc_dev = dev;
5275 	ctx->ifc_softc = sc;
5276 
5277 	iflib_register(ctx);
5278 	iflib_add_device_sysctl_pre(ctx);
5279 
5280 	scctx = &ctx->ifc_softc_ctx;
5281 	ifp = ctx->ifc_ifp;
5282 	if (ctx->ifc_sysctl_simple_tx) {
5283 #ifndef ALTQ
5284 		if_settransmitfn(ifp, iflib_simple_transmit);
5285 		device_printf(dev, "using simple if_transmit\n");
5286 #else
5287 		device_printf(dev, "ALTQ prevents using simple if_transmit\n");
5288 		ctx->ifc_sysctl_simple_tx = 0;
5289 #endif
5290 	}
5291 	iflib_reset_qvalues(ctx);
5292 	IFNET_WLOCK();
5293 	CTX_LOCK(ctx);
5294 	IFLIB_REGISTER_FAIL_POINT(dev, register_before_attach_pre, err,
5295 	    fail_cleanup);
5296 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5297 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5298 		goto fail_cleanup;
5299 	}
5300 	attach_pre_succeeded = true;
5301 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_attach_pre, err,
5302 	    fail_cleanup);
5303 	_iflib_pre_assert(scctx);
5304 	ctx->ifc_txrx = *scctx->isc_txrx;
5305 
5306 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5307 
5308 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5309 		ctx->ifc_mediap = scctx->isc_media;
5310 
5311 #ifdef INVARIANTS
5312 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5313 		MPASS(scctx->isc_tx_csum_flags);
5314 #endif
5315 
5316 	if_setcapabilities(ifp,
5317 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5318 	if_setcapenable(ifp,
5319 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5320 
5321 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5322 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5323 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5324 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5325 
5326 	num_txd = iflib_num_tx_descs(ctx);
5327 	num_rxd = iflib_num_rx_descs(ctx);
5328 
5329 	/* XXX change for per-queue sizes */
5330 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5331 	    num_txd, num_rxd);
5332 
5333 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5334 		scctx->isc_tx_nsegments = max(1, num_txd /
5335 		    MAX_SINGLE_PACKET_FRACTION);
5336 	if (scctx->isc_tx_tso_segments_max > num_txd /
5337 	    MAX_SINGLE_PACKET_FRACTION)
5338 		scctx->isc_tx_tso_segments_max = max(1,
5339 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5340 
5341 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5342 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5343 		/*
5344 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5345 		 * but some MACs do.
5346 		 */
5347 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5348 		    IP_MAXPACKET));
5349 		/*
5350 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5351 		 * into account.  In the worst case, each of these calls will
5352 		 * add another mbuf and, thus, the requirement for another DMA
5353 		 * segment.  So for best performance, it doesn't make sense to
5354 		 * advertize a maximum of TSO segments that typically will
5355 		 * require defragmentation in iflib_encap().
5356 		 */
5357 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5358 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5359 	}
5360 	if (scctx->isc_rss_table_size == 0)
5361 		scctx->isc_rss_table_size = 64;
5362 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1;
5363 
5364 	/* Create and start admin taskqueue */
5365 	snprintf(namebuf, TASKQUEUE_NAMELEN, "if_%s_tq", device_get_nameunit(dev));
5366 	ctx->ifc_tq = taskqueue_create_fast(namebuf, M_NOWAIT,
5367 	    taskqueue_thread_enqueue, &ctx->ifc_tq);
5368 	if (ctx->ifc_tq == NULL) {
5369 		device_printf(dev, "Unable to create admin taskqueue\n");
5370 		err = ENOMEM;
5371 		goto fail_cleanup;
5372 	}
5373 
5374 	err = taskqueue_start_threads(&ctx->ifc_tq, 1, PI_NET, "%s", namebuf);
5375 	if (err) {
5376 		device_printf(dev,
5377 		    "Unable to start admin taskqueue threads error: %d\n",
5378 		    err);
5379 		taskqueue_free(ctx->ifc_tq);
5380 		ctx->ifc_tq = NULL;
5381 		goto fail_cleanup;
5382 	}
5383 
5384 	TASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5385 	TASK_INIT(&ctx->ifc_led_task, 0, _task_fn_led, ctx);
5386 	TASK_INIT(&ctx->ifc_vflr_task, 0, _task_fn_iov, ctx);
5387 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_taskqueue, err,
5388 	    fail_cleanup);
5389 
5390 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5391 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5392 		device_printf(dev, "Unable to fetch CPU list\n");
5393 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5394 		ctx->ifc_cpus_are_physical_cores = false;
5395 	} else
5396 		ctx->ifc_cpus_are_physical_cores = true;
5397 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5398 
5399 	/*
5400 	 * Now set up MSI or MSI-X, should return us the number of supported
5401 	 * vectors (will be 1 for a legacy interrupt and MSI).
5402 	 */
5403 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5404 		msix = scctx->isc_vectors;
5405 	} else if (scctx->isc_msix_bar != 0)
5406 		/*
5407 		 * The simple fact that isc_msix_bar is not 0 does not mean we
5408 		 * we have a good value there that is known to work.
5409 		 */
5410 		msix = iflib_msix_init(ctx);
5411 	else {
5412 		scctx->isc_vectors = 1;
5413 		scctx->isc_ntxqsets = 1;
5414 		scctx->isc_nrxqsets = 1;
5415 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5416 		msix = 0;
5417 	}
5418 	intr_allocated = true;
5419 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_interrupts, err,
5420 	    fail_cleanup);
5421 	/* Get memory for the station queues */
5422 	if ((err = iflib_queues_alloc(ctx))) {
5423 		device_printf(dev, "Unable to allocate queue memory\n");
5424 		goto fail_cleanup;
5425 	}
5426 	queues_allocated = true;
5427 
5428 	if ((err = iflib_qset_structures_setup(ctx)))
5429 		goto fail_cleanup;
5430 
5431 	/*
5432 	 * Now that we know how many queues there are, get the core offset.
5433 	 */
5434 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5435 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_queues, err,
5436 	    fail_cleanup);
5437 
5438 	if (msix > 1) {
5439 		/*
5440 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5441 		 * aren't the default NULL implementation.
5442 		 */
5443 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5444 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5445 		    kobj_desc);
5446 		if (kobj_method == &kobj_desc->deflt) {
5447 			device_printf(dev,
5448 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5449 			err = EOPNOTSUPP;
5450 			goto fail_cleanup;
5451 		}
5452 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5453 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5454 		    kobj_desc);
5455 		if (kobj_method == &kobj_desc->deflt) {
5456 			device_printf(dev,
5457 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5458 			err = EOPNOTSUPP;
5459 			goto fail_cleanup;
5460 		}
5461 
5462 		/*
5463 		 * Assign the MSI-X vectors.
5464 		 * Note that the default NULL ifdi_msix_intr_assign method will
5465 		 * fail here, too.
5466 		 */
5467 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5468 		if (err != 0) {
5469 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5470 			    err);
5471 			goto fail_cleanup;
5472 		}
5473 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5474 		rid = 0;
5475 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5476 			MPASS(msix == 1);
5477 			rid = 1;
5478 		}
5479 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5480 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5481 			goto fail_cleanup;
5482 		}
5483 	} else {
5484 		device_printf(dev,
5485 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5486 		err = ENODEV;
5487 		goto fail_cleanup;
5488 	}
5489 
5490 	/*
5491 	 * It prevents a double-locking panic with iflib_media_status when
5492 	 * the driver loads.
5493 	 */
5494 	CTX_UNLOCK(ctx);
5495 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5496 	CTX_LOCK(ctx);
5497 
5498 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5499 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5500 		goto fail_detach;
5501 	}
5502 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_attach_post, err,
5503 	    fail_detach);
5504 
5505 	/*
5506 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5507 	 * This must appear after the call to ether_ifattach() because
5508 	 * ether_ifattach() sets if_hdrlen to the default value.
5509 	 */
5510 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5511 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5512 
5513 	if ((err = iflib_netmap_attach(ctx))) {
5514 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5515 		goto fail_detach;
5516 	}
5517 	*ctxp = ctx;
5518 
5519 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5520 
5521 	iflib_add_device_sysctl_post(ctx);
5522 	iflib_add_pfil(ctx);
5523 	ctx->ifc_flags |= IFC_INIT_DONE;
5524 	CTX_UNLOCK(ctx);
5525 	IFNET_WUNLOCK();
5526 
5527 	/* Create led(4) devices if the driver defined the method */
5528 	kobj_desc = &ifdi_led_func_desc;
5529 	kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5530 	    kobj_desc);
5531 	if (kobj_method != &kobj_desc->deflt && IFDI_LED_SUPPORTED(ctx))
5532 		iflib_led_create(ctx);
5533 
5534 	return (0);
5535 
5536 fail_detach:
5537 	STATE_LOCK(ctx);
5538 	ctx->ifc_flags |= IFC_IN_DETACH;
5539 	STATE_UNLOCK(ctx);
5540 	/* Tasks may need either lock; ether_ifdetach() takes ifnet_detach_sx. */
5541 	CTX_UNLOCK(ctx);
5542 	IFNET_WUNLOCK();
5543 	taskqueue_drain_all(ctx->ifc_tq);
5544 #ifdef PCI_IOV
5545 	/*
5546 	 * IFDI_ATTACH_POST may have registered an SR-IOV schema.  Match the
5547 	 * normal deregistration order so a failed attach cannot leave a stale
5548 	 * /dev/iov node behind.  device_attach() holds Giant throughout this
5549 	 * path, so an IOV configuration cannot race the detach.
5550 	 */
5551 	if (!CTX_IS_VF(ctx)) {
5552 		iov_error = pci_iov_detach(dev);
5553 		if (iov_error != 0)
5554 			device_printf(dev, "Could not detach SR-IOV after "
5555 			    "attach failure: %d\n", iov_error);
5556 	}
5557 #endif
5558 	ether_ifdetach(ctx->ifc_ifp);
5559 	IFNET_WLOCK();
5560 	CTX_LOCK(ctx);
5561 	goto fail_cleanup_detaching;
5562 
5563 fail_cleanup:
5564 	STATE_LOCK(ctx);
5565 	ctx->ifc_flags |= IFC_IN_DETACH;
5566 	STATE_UNLOCK(ctx);
5567 
5568 fail_cleanup_detaching:
5569 	/*
5570 	 * The pre-attach sysctls contain pointers into ctx.  Remove them on
5571 	 * every registration failure before iflib_deregister() frees ctx.
5572 	 */
5573 	if (ctx->ifc_sysctl_node != NULL) {
5574 		sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5575 		ctx->ifc_sysctl_node = NULL;
5576 	}
5577 
5578 	if (ctx->ifc_tq != NULL) {
5579 		/*
5580 		 * Drain without holding the ifnet or context locks so configuration
5581 		 * tasks can run to completion.  On fail_detach a second drain also
5582 		 * catches tasks queued during the first drain.
5583 		 */
5584 		CTX_UNLOCK(ctx);
5585 		IFNET_WUNLOCK();
5586 		taskqueue_drain_all(ctx->ifc_tq);
5587 		IFNET_WLOCK();
5588 		CTX_LOCK(ctx);
5589 	}
5590 
5591 	if (queues_allocated) {
5592 		iflib_tqg_detach(ctx);
5593 		iflib_tx_structures_free(ctx);
5594 		iflib_rx_structures_free(ctx);
5595 	}
5596 
5597 	/*
5598 	 * A successful IFDI_ATTACH_PRE must be matched by IFDI_DETACH, even
5599 	 * when registration fails before queue allocation.  Match
5600 	 * iflib_device_deregister by detaching before taskqueue_free, and avoid
5601 	 * holding IFNET_WLOCK across driver detach (LinuxKPI workqueue drain).
5602 	 */
5603 	if (attach_pre_succeeded) {
5604 		IFNET_WUNLOCK();
5605 		IFDI_DETACH(ctx);
5606 		if (queues_allocated)
5607 			IFDI_QUEUES_FREE(ctx);
5608 		/* Reacquire the global lock before the context lock. */
5609 		CTX_UNLOCK(ctx);
5610 		IFNET_WLOCK();
5611 		CTX_LOCK(ctx);
5612 	}
5613 	if (ctx->ifc_tq != NULL) {
5614 		taskqueue_free(ctx->ifc_tq);
5615 		ctx->ifc_tq = NULL;
5616 	}
5617 	if (intr_allocated)
5618 		iflib_free_intr_mem(ctx);
5619 
5620 	CTX_UNLOCK(ctx);
5621 	IFNET_WUNLOCK();
5622 	iflib_deregister(ctx);
5623 	device_set_softc(ctx->ifc_dev, NULL);
5624 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5625 		free(ctx->ifc_softc, M_IFLIB);
5626 	unref_ctx_core_offset(ctx);
5627 	free(ctx, M_IFLIB);
5628 	return (err);
5629 }
5630 
5631 int
iflib_device_attach(device_t dev)5632 iflib_device_attach(device_t dev)
5633 {
5634 	if_ctx_t ctx;
5635 	if_shared_ctx_t sctx;
5636 
5637 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5638 		return (ENOTSUP);
5639 
5640 	pci_enable_busmaster(dev);
5641 
5642 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5643 }
5644 
5645 int
iflib_device_deregister(if_ctx_t ctx)5646 iflib_device_deregister(if_ctx_t ctx)
5647 {
5648 	if_t ifp = ctx->ifc_ifp;
5649 	device_t dev = ctx->ifc_dev;
5650 
5651 	sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5652 	ctx->ifc_sysctl_node = NULL;
5653 
5654 	/* Make sure VLANS are not using driver */
5655 	if (if_vlantrunkinuse(ifp)) {
5656 		device_printf(dev, "Vlan in use, detach first\n");
5657 		return (EBUSY);
5658 	}
5659 #ifdef PCI_IOV
5660 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5661 		device_printf(dev, "SR-IOV in use; detach first.\n");
5662 		return (EBUSY);
5663 	}
5664 #endif
5665 
5666 	STATE_LOCK(ctx);
5667 	ctx->ifc_flags |= IFC_IN_DETACH;
5668 	STATE_UNLOCK(ctx);
5669 
5670 	/* Unregister VLAN handlers before calling iflib_stop() */
5671 	iflib_unregister_vlan_handlers(ctx);
5672 
5673 	iflib_netmap_detach(ifp);
5674 	/*
5675 	 * A task that passed its IFC_IN_DETACH check before the flag was set
5676 	 * can still report a link change.  Drain every private task before
5677 	 * ether_ifdetach() performs the final if_linktask drain.  Drivers may
5678 	 * register their own link-related tasks on this taskqueue.
5679 	 */
5680 	taskqueue_drain_all(ctx->ifc_tq);
5681 	ether_ifdetach(ifp);
5682 
5683 	CTX_LOCK(ctx);
5684 	iflib_stop(ctx);
5685 	CTX_UNLOCK(ctx);
5686 
5687 	iflib_rem_pfil(ctx);
5688 	if (ctx->ifc_led_dev != NULL) {
5689 		led_destroy(ctx->ifc_led_dev);
5690 		taskqueue_drain(ctx->ifc_tq, &ctx->ifc_led_task);
5691 	}
5692 
5693 	iflib_tqg_detach(ctx);
5694 	iflib_tx_structures_free(ctx);
5695 	iflib_rx_structures_free(ctx);
5696 
5697 	CTX_LOCK(ctx);
5698 	IFDI_DETACH(ctx);
5699 	IFDI_QUEUES_FREE(ctx);
5700 	CTX_UNLOCK(ctx);
5701 
5702 	taskqueue_free(ctx->ifc_tq);
5703 	ctx->ifc_tq = NULL;
5704 
5705 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5706 	iflib_free_intr_mem(ctx);
5707 
5708 	bus_generic_detach(dev);
5709 
5710 	iflib_deregister(ctx);
5711 
5712 	device_set_softc(ctx->ifc_dev, NULL);
5713 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5714 		free(ctx->ifc_softc, M_IFLIB);
5715 	unref_ctx_core_offset(ctx);
5716 	free(ctx, M_IFLIB);
5717 	return (0);
5718 }
5719 
5720 static void
iflib_tqg_detach(if_ctx_t ctx)5721 iflib_tqg_detach(if_ctx_t ctx)
5722 {
5723 	iflib_txq_t txq;
5724 	iflib_rxq_t rxq;
5725 	int i;
5726 	struct taskqgroup *tqg;
5727 
5728 	/* XXX drain any dependent tasks */
5729 	tqg = qgroup_if_io_tqg;
5730 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5731 		callout_drain(&txq->ift_timer);
5732 #ifdef DEV_NETMAP
5733 		callout_drain(&txq->ift_netmap_timer);
5734 #endif /* DEV_NETMAP */
5735 		if (txq->ift_task.gt_uniq != NULL)
5736 			taskqgroup_detach(tqg, &txq->ift_task);
5737 	}
5738 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5739 		if (rxq->ifr_task.gt_uniq != NULL)
5740 			taskqgroup_detach(tqg, &rxq->ifr_task);
5741 	}
5742 }
5743 
5744 static void
iflib_free_intr_mem(if_ctx_t ctx)5745 iflib_free_intr_mem(if_ctx_t ctx)
5746 {
5747 
5748 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5749 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5750 	}
5751 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5752 		pci_release_msi(ctx->ifc_dev);
5753 	}
5754 	if (ctx->ifc_msix_mem != NULL) {
5755 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5756 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5757 		ctx->ifc_msix_mem = NULL;
5758 	}
5759 }
5760 
5761 int
iflib_device_detach(device_t dev)5762 iflib_device_detach(device_t dev)
5763 {
5764 	if_ctx_t ctx = device_get_softc(dev);
5765 
5766 	return (iflib_device_deregister(ctx));
5767 }
5768 
5769 int
iflib_device_suspend(device_t dev)5770 iflib_device_suspend(device_t dev)
5771 {
5772 	if_ctx_t ctx = device_get_softc(dev);
5773 
5774 	CTX_LOCK(ctx);
5775 	IFDI_SUSPEND(ctx);
5776 	CTX_UNLOCK(ctx);
5777 
5778 	return (bus_generic_suspend(dev));
5779 }
5780 int
iflib_device_shutdown(device_t dev)5781 iflib_device_shutdown(device_t dev)
5782 {
5783 	if_ctx_t ctx = device_get_softc(dev);
5784 
5785 	CTX_LOCK(ctx);
5786 	IFDI_SHUTDOWN(ctx);
5787 	CTX_UNLOCK(ctx);
5788 
5789 	return (bus_generic_suspend(dev));
5790 }
5791 
5792 int
iflib_device_resume(device_t dev)5793 iflib_device_resume(device_t dev)
5794 {
5795 	if_ctx_t ctx = device_get_softc(dev);
5796 	iflib_txq_t txq = ctx->ifc_txqs;
5797 
5798 	CTX_LOCK(ctx);
5799 	IFDI_RESUME(ctx);
5800 	iflib_if_init_locked(ctx);
5801 	CTX_UNLOCK(ctx);
5802 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5803 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5804 
5805 	return (bus_generic_resume(dev));
5806 }
5807 
5808 int
iflib_device_iov_init(device_t dev,uint16_t num_vfs,const nvlist_t * params)5809 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5810 {
5811 	int error;
5812 	if_ctx_t ctx = device_get_softc(dev);
5813 
5814 	CTX_LOCK(ctx);
5815 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5816 	CTX_UNLOCK(ctx);
5817 
5818 	return (error);
5819 }
5820 
5821 int
iflib_device_iov_init_restart(device_t dev,uint16_t num_vfs,const nvlist_t * params)5822 iflib_device_iov_init_restart(device_t dev, uint16_t num_vfs,
5823     const nvlist_t *params)
5824 {
5825 	if_ctx_t ctx;
5826 	if_t ifp;
5827 	bool restart, running;
5828 	int error;
5829 
5830 	ctx = device_get_softc(dev);
5831 	ifp = ctx->ifc_ifp;
5832 
5833 	CTX_LOCK(ctx);
5834 	/*
5835 	 * Drivers which change the PF queue layout need the complete iflib
5836 	 * stop/init sequence around their IOV callback when the interface is
5837 	 * active.  An administratively-down interface has no live queues to
5838 	 * quiesce, and must remain down after the new layout is installed.
5839 	 * Keep the transition within one context-lock critical section.
5840 	 */
5841 	restart = (if_getflags(ifp) & IFF_UP) != 0;
5842 	running = (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0;
5843 	if (restart || running)
5844 		iflib_stop(ctx);
5845 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5846 	if (restart)
5847 		iflib_init_locked(ctx);
5848 	CTX_UNLOCK(ctx);
5849 	return (error);
5850 }
5851 
5852 void
iflib_device_iov_uninit(device_t dev)5853 iflib_device_iov_uninit(device_t dev)
5854 {
5855 	if_ctx_t ctx = device_get_softc(dev);
5856 
5857 	CTX_LOCK(ctx);
5858 	IFDI_IOV_UNINIT(ctx);
5859 	CTX_UNLOCK(ctx);
5860 }
5861 
5862 void
iflib_device_iov_uninit_restart(device_t dev)5863 iflib_device_iov_uninit_restart(device_t dev)
5864 {
5865 	if_ctx_t ctx;
5866 	bool restart;
5867 
5868 	ctx = device_get_softc(dev);
5869 
5870 	CTX_LOCK(ctx);
5871 	/*
5872 	 * RUNNING can be clear while a watchdog reset is pending but the
5873 	 * hardware is still live.  Always stop before the driver changes its
5874 	 * queue layout, and use IFF_UP only to preserve administrative state.
5875 	 */
5876 	restart = (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0;
5877 	iflib_stop(ctx);
5878 	IFDI_IOV_UNINIT(ctx);
5879 	if (restart)
5880 		iflib_init_locked(ctx);
5881 	CTX_UNLOCK(ctx);
5882 }
5883 
5884 int
iflib_device_iov_add_vf(device_t dev,uint16_t vfnum,const nvlist_t * params)5885 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5886 {
5887 	int error;
5888 	if_ctx_t ctx = device_get_softc(dev);
5889 
5890 	CTX_LOCK(ctx);
5891 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5892 	CTX_UNLOCK(ctx);
5893 
5894 	return (error);
5895 }
5896 
5897 /*********************************************************************
5898  *
5899  *  MODULE FUNCTION DEFINITIONS
5900  *
5901  **********************************************************************/
5902 
5903 /*
5904  * - Start a fast taskqueue thread for each core
5905  * - Start a taskqueue for control operations
5906  */
5907 static int
iflib_module_init(void)5908 iflib_module_init(void)
5909 {
5910 	iflib_timer_default = hz / 2;
5911 	return (0);
5912 }
5913 
5914 static int
iflib_module_event_handler(module_t mod,int what,void * arg)5915 iflib_module_event_handler(module_t mod, int what, void *arg)
5916 {
5917 	int err;
5918 
5919 	switch (what) {
5920 	case MOD_LOAD:
5921 		if ((err = iflib_module_init()) != 0)
5922 			return (err);
5923 		break;
5924 	case MOD_UNLOAD:
5925 		return (EBUSY);
5926 	default:
5927 		return (EOPNOTSUPP);
5928 	}
5929 
5930 	return (0);
5931 }
5932 
5933 /*********************************************************************
5934  *
5935  *  PUBLIC FUNCTION DEFINITIONS
5936  *     ordered as in iflib.h
5937  *
5938  **********************************************************************/
5939 
5940 static void
_iflib_assert(if_shared_ctx_t sctx)5941 _iflib_assert(if_shared_ctx_t sctx)
5942 {
5943 	int i;
5944 
5945 	MPASS(sctx->isc_tx_maxsize);
5946 	MPASS(sctx->isc_tx_maxsegsize);
5947 
5948 	MPASS(sctx->isc_rx_maxsize);
5949 	MPASS(sctx->isc_rx_nsegments);
5950 	MPASS(sctx->isc_rx_maxsegsize);
5951 
5952 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5953 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5954 		MPASS(sctx->isc_nrxd_min[i]);
5955 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5956 		MPASS(sctx->isc_nrxd_max[i]);
5957 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5958 		MPASS(sctx->isc_nrxd_default[i]);
5959 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5960 	}
5961 
5962 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5963 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5964 		MPASS(sctx->isc_ntxd_min[i]);
5965 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5966 		MPASS(sctx->isc_ntxd_max[i]);
5967 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5968 		MPASS(sctx->isc_ntxd_default[i]);
5969 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5970 	}
5971 }
5972 
5973 static void
_iflib_pre_assert(if_softc_ctx_t scctx)5974 _iflib_pre_assert(if_softc_ctx_t scctx)
5975 {
5976 
5977 	MPASS(scctx->isc_txrx->ift_txd_encap);
5978 	MPASS(scctx->isc_txrx->ift_txd_flush);
5979 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5980 	MPASS(scctx->isc_txrx->ift_rxd_available);
5981 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5982 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5983 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5984 }
5985 
5986 static void
iflib_register(if_ctx_t ctx)5987 iflib_register(if_ctx_t ctx)
5988 {
5989 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5990 	driver_t *driver = sctx->isc_driver;
5991 	device_t dev = ctx->ifc_dev;
5992 	if_t ifp;
5993 
5994 	_iflib_assert(sctx);
5995 
5996 	CTX_LOCK_INIT(ctx);
5997 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5998 	ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev);
5999 
6000 	/*
6001 	 * Initialize our context's device specific methods
6002 	 */
6003 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
6004 	kobj_class_compile((kobj_class_t) driver);
6005 
6006 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
6007 	if_setsoftc(ifp, ctx);
6008 	if_setdev(ifp, dev);
6009 	if_setinitfn(ifp, iflib_if_init);
6010 	if_setioctlfn(ifp, iflib_if_ioctl);
6011 #ifdef ALTQ
6012 	if_setstartfn(ifp, iflib_altq_if_start);
6013 	if_settransmitfn(ifp, iflib_altq_if_transmit);
6014 	if_setsendqready(ifp);
6015 #else
6016 	if_settransmitfn(ifp, iflib_if_transmit);
6017 #endif
6018 	if_setqflushfn(ifp, iflib_if_qflush);
6019 	if_setgetcounterfn(ifp, iflib_if_get_counter);
6020 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
6021 	ctx->ifc_vlan_attach_event =
6022 	    EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
6023 		    EVENTHANDLER_PRI_FIRST);
6024 	ctx->ifc_vlan_detach_event =
6025 	    EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
6026 		    EVENTHANDLER_PRI_FIRST);
6027 
6028 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
6029 		ctx->ifc_mediap = &ctx->ifc_media;
6030 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
6031 		    iflib_media_change, iflib_media_status);
6032 	}
6033 }
6034 
6035 static void
iflib_unregister_vlan_handlers(if_ctx_t ctx)6036 iflib_unregister_vlan_handlers(if_ctx_t ctx)
6037 {
6038 	/* Unregister VLAN events */
6039 	if (ctx->ifc_vlan_attach_event != NULL) {
6040 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
6041 		ctx->ifc_vlan_attach_event = NULL;
6042 	}
6043 	if (ctx->ifc_vlan_detach_event != NULL) {
6044 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
6045 		ctx->ifc_vlan_detach_event = NULL;
6046 	}
6047 
6048 }
6049 
6050 static void
iflib_deregister(if_ctx_t ctx)6051 iflib_deregister(if_ctx_t ctx)
6052 {
6053 	if_t ifp = ctx->ifc_ifp;
6054 
6055 	/* Remove all media */
6056 	ifmedia_removeall(&ctx->ifc_media);
6057 
6058 	/* Ensure that VLAN event handlers are unregistered */
6059 	iflib_unregister_vlan_handlers(ctx);
6060 
6061 	/* Release kobject reference */
6062 	kobj_delete((kobj_t) ctx, NULL);
6063 
6064 	/* Free the ifnet structure */
6065 	if_free(ifp);
6066 
6067 	STATE_LOCK_DESTROY(ctx);
6068 
6069 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
6070 	CTX_LOCK_DESTROY(ctx);
6071 }
6072 
6073 static int
iflib_queues_alloc(if_ctx_t ctx)6074 iflib_queues_alloc(if_ctx_t ctx)
6075 {
6076 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6077 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6078 	device_t dev = ctx->ifc_dev;
6079 	int nrxqsets = scctx->isc_nrxqsets;
6080 	int ntxqsets = scctx->isc_ntxqsets;
6081 	iflib_txq_t txq;
6082 	iflib_rxq_t rxq;
6083 	iflib_fl_t fl = NULL;
6084 	int i, j, cpu, err, txconf, rxconf;
6085 	iflib_dma_info_t ifdip;
6086 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
6087 	uint32_t *txqsizes = scctx->isc_txqsizes;
6088 	uint8_t nrxqs = sctx->isc_nrxqs;
6089 	uint8_t ntxqs = sctx->isc_ntxqs;
6090 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
6091 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
6092 	caddr_t *vaddrs;
6093 	uint64_t *paddrs;
6094 
6095 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
6096 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
6097 	KASSERT(nrxqs >= fl_offset + nfree_lists,
6098 	    ("there must be at least a rxq for each free list"));
6099 
6100 	/* Allocate the TX ring struct memory */
6101 	if (!(ctx->ifc_txqs =
6102 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
6103 		    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6104 		device_printf(dev, "Unable to allocate TX ring memory\n");
6105 		err = ENOMEM;
6106 		goto fail;
6107 	}
6108 
6109 	/* Now allocate the RX */
6110 	if (!(ctx->ifc_rxqs =
6111 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
6112 		    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6113 		device_printf(dev, "Unable to allocate RX ring memory\n");
6114 		err = ENOMEM;
6115 		goto rx_fail;
6116 	}
6117 
6118 	txq = ctx->ifc_txqs;
6119 	rxq = ctx->ifc_rxqs;
6120 
6121 	/*
6122 	 * XXX handle allocation failure
6123 	 */
6124 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
6125 		/* Set up some basics */
6126 
6127 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
6128 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6129 			device_printf(dev,
6130 			    "Unable to allocate TX DMA info memory\n");
6131 			err = ENOMEM;
6132 			goto err_tx_desc;
6133 		}
6134 		txq->ift_ifdi = ifdip;
6135 		for (j = 0; j < ntxqs; j++, ifdip++) {
6136 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
6137 				device_printf(dev,
6138 				    "Unable to allocate TX descriptors\n");
6139 				err = ENOMEM;
6140 				goto err_tx_desc;
6141 			}
6142 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
6143 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
6144 		}
6145 		txq->ift_ctx = ctx;
6146 		txq->ift_id = i;
6147 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
6148 			txq->ift_br_offset = 1;
6149 		} else {
6150 			txq->ift_br_offset = 0;
6151 		}
6152 
6153 		if (iflib_txsd_alloc(txq)) {
6154 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6155 			err = ENOMEM;
6156 			goto err_tx_desc;
6157 		}
6158 
6159 		/* Initialize the TX lock */
6160 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6161 		    device_get_nameunit(dev), txq->ift_id);
6162 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6163 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6164 		txq->ift_timer.c_cpu = cpu;
6165 #ifdef DEV_NETMAP
6166 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6167 		txq->ift_netmap_timer.c_cpu = cpu;
6168 #endif /* DEV_NETMAP */
6169 
6170 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6171 		    iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6172 		if (err) {
6173 			/* XXX free any allocated rings */
6174 			device_printf(dev, "Unable to allocate buf_ring\n");
6175 			goto err_tx_desc;
6176 		}
6177 		txq->ift_reclaim_thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
6178 	}
6179 
6180 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6181 		/* Set up some basics */
6182 		callout_init(&rxq->ifr_watchdog, 1);
6183 
6184 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6185 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6186 			device_printf(dev,
6187 			    "Unable to allocate RX DMA info memory\n");
6188 			err = ENOMEM;
6189 			goto err_tx_desc;
6190 		}
6191 
6192 		rxq->ifr_ifdi = ifdip;
6193 		/* XXX this needs to be changed if #rx queues != #tx queues */
6194 		rxq->ifr_ntxqirq = 1;
6195 		rxq->ifr_txqid[0] = i;
6196 		for (j = 0; j < nrxqs; j++, ifdip++) {
6197 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6198 				device_printf(dev,
6199 				    "Unable to allocate RX descriptors\n");
6200 				err = ENOMEM;
6201 				goto err_tx_desc;
6202 			}
6203 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6204 		}
6205 		rxq->ifr_ctx = ctx;
6206 		rxq->ifr_id = i;
6207 		rxq->ifr_fl_offset = fl_offset;
6208 		rxq->ifr_nfl = nfree_lists;
6209 		if (!(fl =
6210 		    (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6211 			device_printf(dev, "Unable to allocate free list memory\n");
6212 			err = ENOMEM;
6213 			goto err_tx_desc;
6214 		}
6215 		rxq->ifr_fl = fl;
6216 		for (j = 0; j < nfree_lists; j++) {
6217 			fl[j].ifl_rxq = rxq;
6218 			fl[j].ifl_id = j;
6219 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6220 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6221 		}
6222 		/* Allocate receive buffers for the ring */
6223 		if (iflib_rxsd_alloc(rxq)) {
6224 			device_printf(dev,
6225 			    "Critical Failure setting up receive buffers\n");
6226 			err = ENOMEM;
6227 			goto err_rx_desc;
6228 		}
6229 
6230 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6231 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6232 			    M_WAITOK);
6233 	}
6234 
6235 	/* TXQs */
6236 	vaddrs = malloc(sizeof(caddr_t)  * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6237 	paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6238 	for (i = 0; i < ntxqsets; i++) {
6239 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6240 
6241 		for (j = 0; j < ntxqs; j++, di++) {
6242 			vaddrs[i * ntxqs + j] = di->idi_vaddr;
6243 			paddrs[i * ntxqs + j] = di->idi_paddr;
6244 		}
6245 	}
6246 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6247 		device_printf(ctx->ifc_dev,
6248 		    "Unable to allocate device TX queue\n");
6249 		iflib_tx_structures_free(ctx);
6250 		free(vaddrs, M_IFLIB);
6251 		free(paddrs, M_IFLIB);
6252 		goto err_rx_desc;
6253 	}
6254 	free(vaddrs, M_IFLIB);
6255 	free(paddrs, M_IFLIB);
6256 
6257 	/* RXQs */
6258 	vaddrs = malloc(sizeof(caddr_t)  * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6259 	paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6260 	for (i = 0; i < nrxqsets; i++) {
6261 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6262 
6263 		for (j = 0; j < nrxqs; j++, di++) {
6264 			vaddrs[i * nrxqs + j] = di->idi_vaddr;
6265 			paddrs[i * nrxqs + j] = di->idi_paddr;
6266 		}
6267 	}
6268 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6269 		device_printf(ctx->ifc_dev,
6270 		    "Unable to allocate device RX queue\n");
6271 		iflib_tx_structures_free(ctx);
6272 		free(vaddrs, M_IFLIB);
6273 		free(paddrs, M_IFLIB);
6274 		goto err_rx_desc;
6275 	}
6276 	free(vaddrs, M_IFLIB);
6277 	free(paddrs, M_IFLIB);
6278 
6279 	return (0);
6280 
6281 /* XXX handle allocation failure changes */
6282 err_rx_desc:
6283 err_tx_desc:
6284 rx_fail:
6285 	if (ctx->ifc_rxqs != NULL)
6286 		free(ctx->ifc_rxqs, M_IFLIB);
6287 	ctx->ifc_rxqs = NULL;
6288 	if (ctx->ifc_txqs != NULL)
6289 		free(ctx->ifc_txqs, M_IFLIB);
6290 	ctx->ifc_txqs = NULL;
6291 fail:
6292 	return (err);
6293 }
6294 
6295 static int
iflib_tx_structures_setup(if_ctx_t ctx)6296 iflib_tx_structures_setup(if_ctx_t ctx)
6297 {
6298 	iflib_txq_t txq = ctx->ifc_txqs;
6299 	int i;
6300 
6301 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6302 		iflib_txq_setup(txq);
6303 
6304 	return (0);
6305 }
6306 
6307 static void
iflib_tx_structures_free(if_ctx_t ctx)6308 iflib_tx_structures_free(if_ctx_t ctx)
6309 {
6310 	iflib_txq_t txq = ctx->ifc_txqs;
6311 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6312 	int i, j;
6313 
6314 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6315 		for (j = 0; j < sctx->isc_ntxqs; j++)
6316 			iflib_dma_free(&txq->ift_ifdi[j]);
6317 		iflib_txq_destroy(txq);
6318 	}
6319 	free(ctx->ifc_txqs, M_IFLIB);
6320 	ctx->ifc_txqs = NULL;
6321 }
6322 
6323 /*********************************************************************
6324  *
6325  *  Initialize all receive rings.
6326  *
6327  **********************************************************************/
6328 static int
iflib_rx_structures_setup(if_ctx_t ctx)6329 iflib_rx_structures_setup(if_ctx_t ctx)
6330 {
6331 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6332 	int q;
6333 #if defined(INET6) || defined(INET)
6334 	int err, i;
6335 #endif
6336 
6337 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6338 #if defined(INET6) || defined(INET)
6339 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6340 		    TCP_LRO_ENTRIES, min(1024,
6341 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6342 		if (err != 0) {
6343 			device_printf(ctx->ifc_dev,
6344 			    "LRO Initialization failed!\n");
6345 			goto fail;
6346 		}
6347 #endif
6348 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6349 	}
6350 	return (0);
6351 #if defined(INET6) || defined(INET)
6352 fail:
6353 	/*
6354 	 * Free LRO resources allocated so far, we will only handle
6355 	 * the rings that completed, the failing case will have
6356 	 * cleaned up for itself.  'q' failed, so its the terminus.
6357 	 */
6358 	rxq = ctx->ifc_rxqs;
6359 	for (i = 0; i < q; ++i, rxq++) {
6360 		tcp_lro_free(&rxq->ifr_lc);
6361 	}
6362 	return (err);
6363 #endif
6364 }
6365 
6366 /*********************************************************************
6367  *
6368  *  Free all receive rings.
6369  *
6370  **********************************************************************/
6371 static void
iflib_rx_structures_free(if_ctx_t ctx)6372 iflib_rx_structures_free(if_ctx_t ctx)
6373 {
6374 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6375 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6376 	int i, j;
6377 
6378 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6379 		for (j = 0; j < sctx->isc_nrxqs; j++)
6380 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6381 		iflib_rx_sds_free(rxq);
6382 #if defined(INET6) || defined(INET)
6383 		tcp_lro_free(&rxq->ifr_lc);
6384 #endif
6385 	}
6386 	free(ctx->ifc_rxqs, M_IFLIB);
6387 	ctx->ifc_rxqs = NULL;
6388 }
6389 
6390 static int
iflib_qset_structures_setup(if_ctx_t ctx)6391 iflib_qset_structures_setup(if_ctx_t ctx)
6392 {
6393 	int err;
6394 
6395 	/*
6396 	 * It is expected that the caller takes care of freeing queues if this
6397 	 * fails.
6398 	 */
6399 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6400 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6401 		return (err);
6402 	}
6403 
6404 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6405 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6406 
6407 	return (err);
6408 }
6409 
6410 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)6411 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6412 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6413 {
6414 
6415 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6416 }
6417 
6418 /* Just to avoid copy/paste */
6419 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)6420 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6421     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6422     const char *name)
6423 {
6424 	device_t dev;
6425 	unsigned int base_cpuid, cpuid;
6426 	int err;
6427 
6428 	dev = ctx->ifc_dev;
6429 	base_cpuid = ctx->ifc_sysctl_core_offset;
6430 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6431 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6432 	    irq ? irq->ii_res : NULL, name);
6433 	if (err) {
6434 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6435 		return (err);
6436 	}
6437 #ifdef notyet
6438 	if (cpuid > ctx->ifc_cpuid_highest)
6439 		ctx->ifc_cpuid_highest = cpuid;
6440 #endif
6441 	return (0);
6442 }
6443 
6444 /*
6445  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6446  * resources.
6447  *
6448  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6449  * only.
6450  *
6451  * XXX: Could be removed if subctx's dev has its intr resource allocation
6452  * methods replaced with custom ones?
6453  */
6454 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)6455 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6456 			       int rid, iflib_intr_type_t type,
6457 			       driver_filter_t *filter, void *filter_arg,
6458 			       int qid, const char *name)
6459 {
6460 	device_t dev, subdev;
6461 	struct grouptask *gtask;
6462 	struct taskqgroup *tqg;
6463 	iflib_filter_info_t info;
6464 	gtask_fn_t *fn;
6465 	int tqrid, err;
6466 	driver_filter_t *intr_fast;
6467 	void *q;
6468 
6469 	MPASS(ctx != NULL);
6470 	MPASS(subctx != NULL);
6471 
6472 	tqrid = rid;
6473 	dev = ctx->ifc_dev;
6474 	subdev = subctx->ifc_dev;
6475 
6476 	switch (type) {
6477 	case IFLIB_INTR_RXTX:
6478 		q = &subctx->ifc_rxqs[qid];
6479 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6480 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6481 		tqg = qgroup_if_io_tqg;
6482 		fn = _task_fn_rx;
6483 		intr_fast = iflib_fast_intr_rxtx;
6484 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6485 		break;
6486 	default:
6487 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6488 		    __func__, device_get_nameunit(subdev), type);
6489 		return (EINVAL);
6490 	}
6491 
6492 	info->ifi_filter = filter;
6493 	info->ifi_filter_arg = filter_arg;
6494 	info->ifi_task = gtask;
6495 	info->ifi_ctx = q;
6496 
6497 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6498 
6499 	/* Allocate interrupts from hardware using parent context */
6500 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6501 	if (err != 0) {
6502 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6503 		    device_get_nameunit(subdev), err);
6504 		return (err);
6505 	}
6506 
6507 	if (tqrid != -1) {
6508 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6509 		    name);
6510 		if (err)
6511 			return (err);
6512 	} else {
6513 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6514 	}
6515 
6516 	return (0);
6517 }
6518 
6519 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)6520 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6521 			iflib_intr_type_t type, driver_filter_t *filter,
6522 			void *filter_arg, int qid, const char *name)
6523 {
6524 	device_t dev;
6525 	struct grouptask *gtask;
6526 	struct taskqgroup *tqg;
6527 	iflib_filter_info_t info;
6528 	gtask_fn_t *fn;
6529 	int tqrid, err;
6530 	driver_filter_t *intr_fast;
6531 	void *q;
6532 
6533 	info = &ctx->ifc_filter_info;
6534 	tqrid = rid;
6535 
6536 	switch (type) {
6537 	/* XXX merge tx/rx for netmap? */
6538 	case IFLIB_INTR_TX:
6539 		q = &ctx->ifc_txqs[qid];
6540 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6541 		gtask = &ctx->ifc_txqs[qid].ift_task;
6542 		tqg = qgroup_if_io_tqg;
6543 		fn = _task_fn_tx;
6544 		intr_fast = iflib_fast_intr;
6545 		GROUPTASK_INIT(gtask, 0, fn, q);
6546 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6547 		break;
6548 	case IFLIB_INTR_RX:
6549 		q = &ctx->ifc_rxqs[qid];
6550 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6551 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6552 		tqg = qgroup_if_io_tqg;
6553 		fn = _task_fn_rx;
6554 		intr_fast = iflib_fast_intr;
6555 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6556 		break;
6557 	case IFLIB_INTR_RXTX:
6558 		q = &ctx->ifc_rxqs[qid];
6559 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6560 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6561 		tqg = qgroup_if_io_tqg;
6562 		fn = _task_fn_rx;
6563 		intr_fast = iflib_fast_intr_rxtx;
6564 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6565 		break;
6566 	case IFLIB_INTR_ADMIN:
6567 		q = ctx;
6568 		tqrid = -1;
6569 		info = &ctx->ifc_filter_info;
6570 		gtask = NULL;
6571 		intr_fast = iflib_fast_intr_ctx;
6572 		break;
6573 	default:
6574 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6575 		    __func__);
6576 		return (EINVAL);
6577 	}
6578 
6579 	info->ifi_filter = filter;
6580 	info->ifi_filter_arg = filter_arg;
6581 	info->ifi_task = gtask;
6582 	info->ifi_ctx = q;
6583 
6584 	dev = ctx->ifc_dev;
6585 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6586 	if (err != 0) {
6587 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6588 		return (err);
6589 	}
6590 	if (type == IFLIB_INTR_ADMIN)
6591 		return (0);
6592 
6593 	if (tqrid != -1) {
6594 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6595 		    name);
6596 		if (err)
6597 			return (err);
6598 	} else {
6599 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6600 	}
6601 
6602 	return (0);
6603 }
6604 
6605 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)6606 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6607 			    void *arg, int qid, const char *name)
6608 {
6609 	device_t dev;
6610 	struct grouptask *gtask;
6611 	struct taskqgroup *tqg;
6612 	gtask_fn_t *fn;
6613 	void *q;
6614 	int err;
6615 
6616 	switch (type) {
6617 	case IFLIB_INTR_TX:
6618 		q = &ctx->ifc_txqs[qid];
6619 		gtask = &ctx->ifc_txqs[qid].ift_task;
6620 		tqg = qgroup_if_io_tqg;
6621 		fn = _task_fn_tx;
6622 		GROUPTASK_INIT(gtask, 0, fn, q);
6623 		break;
6624 	case IFLIB_INTR_RX:
6625 		q = &ctx->ifc_rxqs[qid];
6626 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6627 		tqg = qgroup_if_io_tqg;
6628 		fn = _task_fn_rx;
6629 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6630 		break;
6631 	case IFLIB_INTR_IOV:
6632 		return;
6633 	default:
6634 		panic("unknown net intr type");
6635 	}
6636 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6637 	if (err) {
6638 		dev = ctx->ifc_dev;
6639 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6640 		    name);
6641 	}
6642 }
6643 
6644 void
iflib_irq_free(if_ctx_t ctx,if_irq_t irq)6645 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6646 {
6647 
6648 	if (irq->ii_tag)
6649 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6650 
6651 	if (irq->ii_res)
6652 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6653 		    rman_get_rid(irq->ii_res), irq->ii_res);
6654 }
6655 
6656 static int
iflib_legacy_setup(if_ctx_t ctx,driver_filter_t filter,void * filter_arg,int * rid,const char * name)6657 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6658 {
6659 	iflib_txq_t txq = ctx->ifc_txqs;
6660 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6661 	if_irq_t irq = &ctx->ifc_legacy_irq;
6662 	iflib_filter_info_t info;
6663 	device_t dev;
6664 	struct grouptask *gtask;
6665 	struct resource *res;
6666 	int err, tqrid;
6667 	bool rx_only;
6668 
6669 	info = &rxq->ifr_filter_info;
6670 	gtask = &rxq->ifr_task;
6671 	tqrid = *rid;
6672 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6673 
6674 	ctx->ifc_flags |= IFC_LEGACY;
6675 	info->ifi_filter = filter;
6676 	info->ifi_filter_arg = filter_arg;
6677 	info->ifi_task = gtask;
6678 	info->ifi_ctx = rxq;
6679 
6680 	dev = ctx->ifc_dev;
6681 	/* We allocate a single interrupt resource */
6682 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr :
6683 	    iflib_fast_intr_rxtx, NULL, info, name);
6684 	if (err != 0)
6685 		return (err);
6686 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq);
6687 	res = irq->ii_res;
6688 	taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name);
6689 
6690 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6691 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6692 	    "tx");
6693 	return (0);
6694 }
6695 
6696 void
iflib_led_create(if_ctx_t ctx)6697 iflib_led_create(if_ctx_t ctx)
6698 {
6699 
6700 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6701 	    device_get_nameunit(ctx->ifc_dev));
6702 }
6703 
6704 void
iflib_tx_intr_deferred(if_ctx_t ctx,int txqid)6705 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6706 {
6707 
6708 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6709 }
6710 
6711 void
iflib_rx_intr_deferred(if_ctx_t ctx,int rxqid)6712 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6713 {
6714 
6715 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6716 }
6717 
6718 void
iflib_admin_intr_deferred(if_ctx_t ctx)6719 iflib_admin_intr_deferred(if_ctx_t ctx)
6720 {
6721 
6722 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
6723 }
6724 
6725 void
iflib_iov_intr_deferred(if_ctx_t ctx)6726 iflib_iov_intr_deferred(if_ctx_t ctx)
6727 {
6728 
6729 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_vflr_task);
6730 }
6731 
6732 void
iflib_io_tqg_attach(struct grouptask * gt,void * uniq,int cpu,const char * name)6733 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6734 {
6735 
6736 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6737 	    name);
6738 }
6739 
6740 void
iflib_config_task_init(if_ctx_t ctx,struct task * config_task,task_fn_t * fn)6741 iflib_config_task_init(if_ctx_t ctx, struct task *config_task, task_fn_t *fn)
6742 {
6743 	TASK_INIT(config_task, 0, fn, ctx);
6744 }
6745 
6746 void
iflib_config_task_enqueue(if_ctx_t ctx,struct task * config_task)6747 iflib_config_task_enqueue(if_ctx_t ctx, struct task *config_task)
6748 {
6749 	taskqueue_enqueue(ctx->ifc_tq, config_task);
6750 }
6751 
6752 void
iflib_link_state_change(if_ctx_t ctx,int link_state,uint64_t baudrate)6753 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6754 {
6755 	if_t ifp = ctx->ifc_ifp;
6756 
6757 	if_setbaudrate(ifp, baudrate);
6758 	if (baudrate >= IF_Gbps(10)) {
6759 		STATE_LOCK(ctx);
6760 		ctx->ifc_flags |= IFC_PREFETCH;
6761 		STATE_UNLOCK(ctx);
6762 	}
6763 	ctx->ifc_link_state = link_state;
6764 	if_link_state_change(ifp, link_state);
6765 }
6766 
6767 static int
iflib_tx_credits_update(if_ctx_t ctx,iflib_txq_t txq)6768 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6769 {
6770 	int credits;
6771 #ifdef INVARIANTS
6772 	int credits_pre = txq->ift_cidx_processed;
6773 #endif
6774 
6775 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6776 	    BUS_DMASYNC_POSTREAD);
6777 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6778 		return (0);
6779 
6780 	txq->ift_processed += credits;
6781 	txq->ift_cidx_processed += credits;
6782 
6783 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6784 	if (txq->ift_cidx_processed >= txq->ift_size)
6785 		txq->ift_cidx_processed -= txq->ift_size;
6786 	return (credits);
6787 }
6788 
6789 static int
iflib_rxd_avail(if_ctx_t ctx,iflib_rxq_t rxq,qidx_t cidx,qidx_t budget)6790 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6791 {
6792 	iflib_fl_t fl;
6793 	u_int i;
6794 
6795 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6796 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6797 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6798 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6799 	    budget));
6800 }
6801 
6802 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)6803 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6804 	const char *description, if_int_delay_info_t info,
6805 	int offset, int value)
6806 {
6807 	info->iidi_ctx = ctx;
6808 	info->iidi_offset = offset;
6809 	info->iidi_value = value;
6810 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6811 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6812 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6813 	    info, 0, iflib_sysctl_int_delay, "I", description);
6814 }
6815 
6816 struct sx *
iflib_ctx_lock_get(if_ctx_t ctx)6817 iflib_ctx_lock_get(if_ctx_t ctx)
6818 {
6819 
6820 	return (&ctx->ifc_ctx_sx);
6821 }
6822 
6823 static int
iflib_msix_init(if_ctx_t ctx)6824 iflib_msix_init(if_ctx_t ctx)
6825 {
6826 	device_t dev = ctx->ifc_dev;
6827 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6828 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6829 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6830 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6831 
6832 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6833 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6834 
6835 	if (bootverbose)
6836 		device_printf(dev, "msix_init qsets capped at %d\n",
6837 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6838 
6839 	/* Override by tuneable */
6840 	if (scctx->isc_disable_msix)
6841 		goto msi;
6842 
6843 	/* First try MSI-X */
6844 	if ((msgs = pci_msix_count(dev)) == 0) {
6845 		if (bootverbose)
6846 			device_printf(dev, "MSI-X not supported or disabled\n");
6847 		goto msi;
6848 	}
6849 
6850 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6851 	/*
6852 	 * bar == -1 => "trust me I know what I'm doing"
6853 	 * Some drivers are for hardware that is so shoddily
6854 	 * documented that no one knows which bars are which
6855 	 * so the developer has to map all bars. This hack
6856 	 * allows shoddy garbage to use MSI-X in this framework.
6857 	 */
6858 	if (bar != -1) {
6859 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6860 		    SYS_RES_MEMORY, &bar, RF_ACTIVE);
6861 		if (ctx->ifc_msix_mem == NULL) {
6862 			device_printf(dev, "Unable to map MSI-X table\n");
6863 			goto msi;
6864 		}
6865 	}
6866 
6867 	admincnt = sctx->isc_admin_intrcnt;
6868 #if IFLIB_DEBUG
6869 	/* use only 1 qset in debug mode */
6870 	queuemsgs = min(msgs - admincnt, 1);
6871 #else
6872 	queuemsgs = msgs - admincnt;
6873 #endif
6874 #ifdef RSS
6875 	queues = imin(queuemsgs, rss_getnumbuckets());
6876 #else
6877 	queues = queuemsgs;
6878 #endif
6879 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6880 	if (bootverbose)
6881 		device_printf(dev,
6882 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6883 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6884 #ifdef  RSS
6885 	/* If we're doing RSS, clamp at the number of RSS buckets */
6886 	if (queues > rss_getnumbuckets())
6887 		queues = rss_getnumbuckets();
6888 #endif
6889 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6890 		rx_queues = iflib_num_rx_queues;
6891 	else
6892 		rx_queues = queues;
6893 
6894 	if (rx_queues > scctx->isc_nrxqsets)
6895 		rx_queues = scctx->isc_nrxqsets;
6896 
6897 	/*
6898 	 * We want this to be all logical CPUs by default
6899 	 */
6900 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6901 		tx_queues = iflib_num_tx_queues;
6902 	else
6903 		tx_queues = mp_ncpus;
6904 
6905 	if (tx_queues > scctx->isc_ntxqsets)
6906 		tx_queues = scctx->isc_ntxqsets;
6907 
6908 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6909 #ifdef INVARIANTS
6910 		if (tx_queues != rx_queues)
6911 			device_printf(dev,
6912 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6913 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6914 #endif
6915 		tx_queues = min(rx_queues, tx_queues);
6916 		rx_queues = min(rx_queues, tx_queues);
6917 	}
6918 
6919 	vectors = rx_queues + admincnt;
6920 	if (msgs < vectors) {
6921 		device_printf(dev,
6922 		    "insufficient number of MSI-X vectors "
6923 		    "(supported %d, need %d)\n", msgs, vectors);
6924 		goto msi;
6925 	}
6926 
6927 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6928 	    tx_queues);
6929 	msgs = vectors;
6930 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6931 		if (vectors != msgs) {
6932 			device_printf(dev,
6933 			    "Unable to allocate sufficient MSI-X vectors "
6934 			    "(got %d, need %d)\n", vectors, msgs);
6935 			pci_release_msi(dev);
6936 			if (bar != -1) {
6937 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6938 				    ctx->ifc_msix_mem);
6939 				ctx->ifc_msix_mem = NULL;
6940 			}
6941 			goto msi;
6942 		}
6943 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6944 		    vectors);
6945 		scctx->isc_vectors = vectors;
6946 		scctx->isc_nrxqsets = rx_queues;
6947 		scctx->isc_ntxqsets = tx_queues;
6948 		scctx->isc_intr = IFLIB_INTR_MSIX;
6949 
6950 		return (vectors);
6951 	} else {
6952 		device_printf(dev,
6953 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6954 		    err);
6955 		if (bar != -1) {
6956 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6957 			    ctx->ifc_msix_mem);
6958 			ctx->ifc_msix_mem = NULL;
6959 		}
6960 	}
6961 
6962 msi:
6963 	vectors = pci_msi_count(dev);
6964 	scctx->isc_nrxqsets = 1;
6965 	scctx->isc_ntxqsets = 1;
6966 	scctx->isc_vectors = vectors;
6967 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6968 		device_printf(dev, "Using an MSI interrupt\n");
6969 		scctx->isc_intr = IFLIB_INTR_MSI;
6970 	} else {
6971 		scctx->isc_vectors = 1;
6972 		device_printf(dev, "Using a Legacy interrupt\n");
6973 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6974 	}
6975 
6976 	return (vectors);
6977 }
6978 
6979 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6980 
6981 static int
mp_ring_state_handler(SYSCTL_HANDLER_ARGS)6982 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6983 {
6984 	int rc;
6985 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6986 	struct sbuf *sb;
6987 	const char *ring_state = "UNKNOWN";
6988 
6989 	/* XXX needed ? */
6990 	rc = sysctl_wire_old_buffer(req, 0);
6991 	MPASS(rc == 0);
6992 	if (rc != 0)
6993 		return (rc);
6994 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6995 	MPASS(sb != NULL);
6996 	if (sb == NULL)
6997 		return (ENOMEM);
6998 	if (state[3] <= 3)
6999 		ring_state = ring_states[state[3]];
7000 
7001 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
7002 		    state[0], state[1], state[2], ring_state);
7003 	rc = sbuf_finish(sb);
7004 	sbuf_delete(sb);
7005 	return (rc);
7006 }
7007 
7008 enum iflib_ndesc_handler {
7009 	IFLIB_NTXD_HANDLER,
7010 	IFLIB_NRXD_HANDLER,
7011 };
7012 
7013 static int
mp_ndesc_handler(SYSCTL_HANDLER_ARGS)7014 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
7015 {
7016 	if_ctx_t ctx = (void *)arg1;
7017 	enum iflib_ndesc_handler type = arg2;
7018 	char buf[256] = {0};
7019 	qidx_t *ndesc;
7020 	char *p, *next;
7021 	int nqs, rc, i;
7022 
7023 	nqs = 8;
7024 	switch (type) {
7025 	case IFLIB_NTXD_HANDLER:
7026 		ndesc = ctx->ifc_sysctl_ntxds;
7027 		if (ctx->ifc_sctx)
7028 			nqs = ctx->ifc_sctx->isc_ntxqs;
7029 		break;
7030 	case IFLIB_NRXD_HANDLER:
7031 		ndesc = ctx->ifc_sysctl_nrxds;
7032 		if (ctx->ifc_sctx)
7033 			nqs = ctx->ifc_sctx->isc_nrxqs;
7034 		break;
7035 	default:
7036 		printf("%s: unhandled type\n", __func__);
7037 		return (EINVAL);
7038 	}
7039 	if (nqs == 0)
7040 		nqs = 8;
7041 
7042 	for (i = 0; i < 8; i++) {
7043 		if (i >= nqs)
7044 			break;
7045 		if (i)
7046 			strcat(buf, ",");
7047 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
7048 	}
7049 
7050 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
7051 	if (rc || req->newptr == NULL)
7052 		return (rc);
7053 
7054 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
7055 	    i++, p = strsep(&next, " ,")) {
7056 		ndesc[i] = strtoul(p, NULL, 10);
7057 	}
7058 
7059 	return (rc);
7060 }
7061 
7062 static int
iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)7063 iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)
7064 {
7065 	if_ctx_t ctx = (void *)arg1;
7066 	iflib_txq_t txq;
7067 	int i, err;
7068 	int thresh;
7069 
7070 	thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
7071 	err = sysctl_handle_int(oidp, &thresh, arg2, req);
7072 	if (err != 0) {
7073 		return err;
7074 	}
7075 
7076 	if (thresh == ctx->ifc_sysctl_tx_reclaim_thresh)
7077 		return 0;
7078 
7079 	if (thresh > ctx->ifc_softc_ctx.isc_ntxd[0] / 2) {
7080 		device_printf(ctx->ifc_dev, "TX Reclaim thresh must be <= %d\n",
7081 		    ctx->ifc_softc_ctx.isc_ntxd[0] / 2);
7082 		return (EINVAL);
7083 	}
7084 
7085 	ctx->ifc_sysctl_tx_reclaim_thresh = thresh;
7086 	if (ctx->ifc_txqs == NULL)
7087 		return (err);
7088 
7089 	txq = &ctx->ifc_txqs[0];
7090 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7091 		txq->ift_reclaim_thresh = thresh;
7092 	}
7093 	return (err);
7094 }
7095 
7096 static int
iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)7097 iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)
7098 {
7099 	if_ctx_t ctx = (void *)arg1;
7100 	iflib_txq_t txq;
7101 	int i, err;
7102 	int ticks;
7103 
7104 	ticks = ctx->ifc_sysctl_tx_reclaim_ticks;
7105 	err = sysctl_handle_int(oidp, &ticks, arg2, req);
7106 	if (err != 0) {
7107 		return err;
7108 	}
7109 
7110 	if (ticks == ctx->ifc_sysctl_tx_reclaim_ticks)
7111 		return 0;
7112 
7113 	if (ticks > hz) {
7114 		device_printf(ctx->ifc_dev,
7115 		    "TX Reclaim ticks must be <= hz (%d)\n", hz);
7116 		return (EINVAL);
7117 	}
7118 
7119 	ctx->ifc_sysctl_tx_reclaim_ticks = ticks;
7120 	if (ctx->ifc_txqs == NULL)
7121 		return (err);
7122 
7123 	txq = &ctx->ifc_txqs[0];
7124 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7125 		txq->ift_reclaim_ticks = ticks;
7126 	}
7127 	return (err);
7128 }
7129 
7130 static int
iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)7131 iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)
7132 {
7133 	if_ctx_t ctx = (void *)arg1;
7134 	iflib_txq_t txq;
7135 	int i, err;
7136 	int defer;
7137 
7138 	defer = ctx->ifc_sysctl_tx_defer_mfree;
7139 	err = sysctl_handle_int(oidp, &defer, arg2, req);
7140 	if (err != 0) {
7141 		return err;
7142 	}
7143 
7144 	if (defer == ctx->ifc_sysctl_tx_defer_mfree)
7145 		return 0;
7146 
7147 	ctx->ifc_sysctl_tx_defer_mfree = defer;
7148 	if (ctx->ifc_txqs == NULL)
7149 		return (err);
7150 
7151 	txq = &ctx->ifc_txqs[0];
7152 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7153 		txq->ift_defer_mfree = defer;
7154 	}
7155 	return (err);
7156 }
7157 
7158 #define NAME_BUFLEN 32
7159 static void
iflib_add_device_sysctl_pre(if_ctx_t ctx)7160 iflib_add_device_sysctl_pre(if_ctx_t ctx)
7161 {
7162 	device_t dev = iflib_get_dev(ctx);
7163 	struct sysctl_oid_list *child, *oid_list;
7164 	struct sysctl_oid *node;
7165 
7166 	sysctl_ctx_init(&ctx->ifc_sysctl_ctx);
7167 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
7168 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(&ctx->ifc_sysctl_ctx, child,
7169 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
7170 	    "IFLIB fields");
7171 	oid_list = SYSCTL_CHILDREN(node);
7172 
7173 	SYSCTL_ADD_CONST_STRING(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "driver_version",
7174 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
7175 	SYSCTL_ADD_U32(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO,
7176 	    "tx_watchdog_events", CTLFLAG_RD, &ctx->ifc_tx_watchdog_events, 0,
7177 	    "TX watchdog resets initiated by iflib");
7178 
7179 	SYSCTL_ADD_BOOL(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "simple_tx",
7180 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_simple_tx, 0,
7181 	    "use simple tx ring");
7182 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxqs",
7183 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
7184 	    "# of txqs to use, 0 => use default #");
7185 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxqs",
7186 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
7187 	    "# of rxqs to use, 0 => use default #");
7188 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_qs_enable",
7189 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
7190 	    "permit #txq != #rxq");
7191 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "disable_msix",
7192 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
7193 	    "disable MSI-X (default 0)");
7194 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "rx_budget",
7195 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
7196 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "tx_abdicate",
7197 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7198 	    "cause TX to abdicate instead of running to completion");
7199 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7200 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "core_offset",
7201 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7202 	    "offset to start using cores at");
7203 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "separate_txrx",
7204 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7205 	    "use separate cores for TX and RX");
7206 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_logical_cores",
7207 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7208 	    "try to make use of logical cores for TX and RX");
7209 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_extra_msix_vectors",
7210 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
7211 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
7212 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "allocated_msix_vectors",
7213 	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
7214 	    "total # of MSI-X vectors allocated by driver");
7215 
7216 	/* XXX change for per-queue sizes */
7217 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxds",
7218 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7219 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7220 	    "list of # of TX descriptors to use, 0 = use default #");
7221 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxds",
7222 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7223 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7224 	    "list of # of RX descriptors to use, 0 = use default #");
7225 }
7226 
7227 static void
iflib_add_device_sysctl_post(if_ctx_t ctx)7228 iflib_add_device_sysctl_post(if_ctx_t ctx)
7229 {
7230 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7231 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7232 	struct sysctl_oid_list *child;
7233 	struct sysctl_ctx_list *ctx_list = &ctx->ifc_sysctl_ctx;
7234 	iflib_fl_t fl;
7235 	iflib_txq_t txq;
7236 	iflib_rxq_t rxq;
7237 	int i, j;
7238 	char namebuf[NAME_BUFLEN];
7239 	char *qfmt;
7240 	struct sysctl_oid *queue_node, *fl_node, *node;
7241 	struct sysctl_oid_list *queue_list, *fl_list;
7242 
7243 	node = ctx->ifc_sysctl_node;
7244 	child = SYSCTL_CHILDREN(node);
7245 
7246        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_thresh",
7247            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7248            0, iflib_handle_tx_reclaim_thresh, "I",
7249            "Number of TX descs outstanding before reclaim is called");
7250 
7251        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_ticks",
7252            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7253            0, iflib_handle_tx_reclaim_ticks, "I",
7254            "Number of ticks before a TX reclaim is forced");
7255 
7256        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_defer_mfree",
7257            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7258            0, iflib_handle_tx_defer_mfree, "I",
7259            "Free completed transmits outside of TX ring lock");
7260 
7261 	if (scctx->isc_ntxqsets > 100)
7262 		qfmt = "txq%03d";
7263 	else if (scctx->isc_ntxqsets > 10)
7264 		qfmt = "txq%02d";
7265 	else
7266 		qfmt = "txq%d";
7267 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7268 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7269 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7270 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7271 		queue_list = SYSCTL_CHILDREN(queue_node);
7272 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7273 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
7274 		    "cpu this queue is bound to");
7275 #if MEMORY_LOGGING
7276 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7277 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
7278 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7279 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
7280 #endif
7281 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7282 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
7283 		    "# of times m_defrag was called");
7284 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7285 		    CTLFLAG_RD, &txq->ift_pullups,
7286 		    "# of times m_pullup was called");
7287 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7288 		    "mbuf_defrag_failed", CTLFLAG_RD,
7289 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7290 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7291 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
7292 		    "# of times no descriptors were available");
7293 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7294 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
7295 		    "# of times DMA map failed");
7296 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7297 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
7298 		    "# of times txd_encap returned EFBIG");
7299 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7300 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
7301 		    "# of times map failed for other than EFBIG");
7302 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7303 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
7304 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7305 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
7306 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7307 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
7308 		    1, "Consumer Index seen by credit update");
7309 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7310 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
7311 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7312 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
7313 		    "descriptors procesed for clean");
7314 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7315 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
7316 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7317 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7318 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7319 		    mp_ring_state_handler, "A", "soft ring state");
7320 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7321 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
7322 		    "# of enqueues to the mp_ring for this queue");
7323 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7324 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
7325 		    "# of drops in the mp_ring for this queue");
7326 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7327 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
7328 		    "# of normal consumer starts in mp_ring for this queue");
7329 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7330 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
7331 		    "# of consumer stalls in the mp_ring for this queue");
7332 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7333 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
7334 		    "# of consumer restarts in the mp_ring for this queue");
7335 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7336 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
7337 		    "# of consumer abdications in the mp_ring for this queue");
7338 	}
7339 
7340 	if (scctx->isc_nrxqsets > 100)
7341 		qfmt = "rxq%03d";
7342 	else if (scctx->isc_nrxqsets > 10)
7343 		qfmt = "rxq%02d";
7344 	else
7345 		qfmt = "rxq%d";
7346 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7347 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7348 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7349 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7350 		queue_list = SYSCTL_CHILDREN(queue_node);
7351 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7352 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
7353 		    "cpu this queue is bound to");
7354 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7355 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7356 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
7357 			    "Consumer Index");
7358 		}
7359 
7360 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7361 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7362 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
7363 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
7364 			    NULL, "freelist Name");
7365 			fl_list = SYSCTL_CHILDREN(fl_node);
7366 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7367 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
7368 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7369 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
7370 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7371 			    CTLFLAG_RD, &fl->ifl_credits, 1,
7372 			    "credits available");
7373 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7374 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
7375 #if MEMORY_LOGGING
7376 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7377 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
7378 			    "mbufs allocated");
7379 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7380 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
7381 			    "mbufs freed");
7382 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7383 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
7384 			    "clusters allocated");
7385 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7386 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
7387 			    "clusters freed");
7388 #endif
7389 		}
7390 	}
7391 
7392 }
7393 
7394 void
iflib_request_reset(if_ctx_t ctx)7395 iflib_request_reset(if_ctx_t ctx)
7396 {
7397 
7398 	STATE_LOCK(ctx);
7399 	ctx->ifc_flags |= IFC_DO_RESET;
7400 	STATE_UNLOCK(ctx);
7401 }
7402 
7403 void
iflib_request_reset_if_up(if_ctx_t ctx)7404 iflib_request_reset_if_up(if_ctx_t ctx)
7405 {
7406 
7407 	STATE_LOCK(ctx);
7408 	ctx->ifc_flags |= IFC_DO_RESET_IF_UP;
7409 	STATE_UNLOCK(ctx);
7410 }
7411 
7412 void
iflib_init_failed(if_ctx_t ctx)7413 iflib_init_failed(if_ctx_t ctx)
7414 {
7415 
7416 	sx_assert(&ctx->ifc_ctx_sx, SA_XLOCKED);
7417 	STATE_LOCK(ctx);
7418 	ctx->ifc_flags |= IFC_INIT_FAILED;
7419 	STATE_UNLOCK(ctx);
7420 }
7421 
7422 #ifndef __NO_STRICT_ALIGNMENT
7423 static struct mbuf *
iflib_fixup_rx(struct mbuf * m)7424 iflib_fixup_rx(struct mbuf *m)
7425 {
7426 	struct mbuf *n;
7427 
7428 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7429 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7430 		m->m_data += ETHER_HDR_LEN;
7431 		n = m;
7432 	} else {
7433 		MGETHDR(n, M_NOWAIT, MT_DATA);
7434 		if (n == NULL) {
7435 			m_freem(m);
7436 			return (NULL);
7437 		}
7438 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7439 		m->m_data += ETHER_HDR_LEN;
7440 		m->m_len -= ETHER_HDR_LEN;
7441 		n->m_len = ETHER_HDR_LEN;
7442 		M_MOVE_PKTHDR(n, m);
7443 		n->m_next = m;
7444 	}
7445 	return (n);
7446 }
7447 #endif
7448 
7449 #ifdef DEBUGNET
7450 static void
iflib_debugnet_init(if_t ifp,int * nrxr,int * ncl,int * clsize)7451 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7452 {
7453 	if_ctx_t ctx;
7454 
7455 	ctx = if_getsoftc(ifp);
7456 	CTX_LOCK(ctx);
7457 	*nrxr = NRXQSETS(ctx);
7458 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7459 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7460 	CTX_UNLOCK(ctx);
7461 }
7462 
7463 static void
iflib_debugnet_event(if_t ifp,enum debugnet_ev event)7464 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7465 {
7466 	if_ctx_t ctx;
7467 	if_softc_ctx_t scctx;
7468 	iflib_fl_t fl;
7469 	iflib_rxq_t rxq;
7470 	int i, j;
7471 
7472 	ctx = if_getsoftc(ifp);
7473 	scctx = &ctx->ifc_softc_ctx;
7474 
7475 	switch (event) {
7476 	case DEBUGNET_START:
7477 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7478 			rxq = &ctx->ifc_rxqs[i];
7479 			for (j = 0; j < rxq->ifr_nfl; j++) {
7480 				fl = rxq->ifr_fl;
7481 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7482 			}
7483 		}
7484 		iflib_no_tx_batch = 1;
7485 		break;
7486 	default:
7487 		break;
7488 	}
7489 }
7490 
7491 static int
iflib_debugnet_transmit(if_t ifp,struct mbuf * m)7492 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7493 {
7494 	if_ctx_t ctx;
7495 	iflib_txq_t txq;
7496 	int error;
7497 	int bytes_sent = 0;
7498 	int pkt_sent = 0;
7499 
7500 	ctx = if_getsoftc(ifp);
7501 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7502 	    IFF_DRV_RUNNING)
7503 		return (EBUSY);
7504 
7505 	txq = &ctx->ifc_txqs[0];
7506 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7507 	if (error == 0)
7508 		(void)iflib_txd_db_check(txq, true);
7509 	return (error);
7510 }
7511 
7512 static int
iflib_debugnet_poll(if_t ifp,int count)7513 iflib_debugnet_poll(if_t ifp, int count)
7514 {
7515 	struct epoch_tracker et;
7516 	if_ctx_t ctx;
7517 	if_softc_ctx_t scctx;
7518 	iflib_txq_t txq;
7519 	int i;
7520 
7521 	ctx = if_getsoftc(ifp);
7522 	scctx = &ctx->ifc_softc_ctx;
7523 
7524 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7525 	    IFF_DRV_RUNNING)
7526 		return (EBUSY);
7527 
7528 	txq = &ctx->ifc_txqs[0];
7529 	(void)iflib_completed_tx_reclaim(txq, NULL);
7530 
7531 	NET_EPOCH_ENTER(et);
7532 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7533 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7534 	NET_EPOCH_EXIT(et);
7535 	return (0);
7536 }
7537 #endif /* DEBUGNET */
7538 
7539 #ifndef ALTQ
7540 static inline iflib_txq_t
iflib_simple_select_queue(if_ctx_t ctx,struct mbuf * m)7541 iflib_simple_select_queue(if_ctx_t ctx, struct mbuf *m)
7542 {
7543 	int qidx;
7544 
7545 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
7546 		qidx = QIDX(ctx, m);
7547 	else
7548 		qidx = NTXQSETS(ctx) + FIRST_QSET(ctx) - 1;
7549 	return (&ctx->ifc_txqs[qidx]);
7550 }
7551 
7552 static int
iflib_simple_transmit(if_t ifp,struct mbuf * m)7553 iflib_simple_transmit(if_t ifp, struct mbuf *m)
7554 {
7555 	if_ctx_t ctx;
7556 	iflib_txq_t txq;
7557 	struct mbuf **m_defer;
7558 	int error, i, reclaimable;
7559 	int bytes_sent = 0, pkt_sent = 0, mcast_sent = 0;
7560 
7561 
7562 	ctx = if_getsoftc(ifp);
7563 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0
7564 		|| !LINK_ACTIVE(ctx))) {
7565 		DBG_COUNTER_INC(tx_frees);
7566 		m_freem(m);
7567 		return (ENETDOWN);
7568 	}
7569 
7570 	txq = iflib_simple_select_queue(ctx, m);
7571 	mtx_lock(&txq->ift_mtx);
7572 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7573 	if (error == 0) {
7574 		mcast_sent += !!(m->m_flags & M_MCAST);
7575 		(void)iflib_txd_db_check(txq, true);
7576 	} else {
7577 		if (error == ENOBUFS)
7578 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
7579 		else
7580 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
7581 	}
7582 	m_defer = NULL;
7583 	reclaimable = iflib_txq_can_reclaim(txq);
7584 	if (reclaimable != 0) {
7585 		/*
7586 		 * Try to set m_defer to the deferred mbuf reclaim array.  If
7587 		 * we can, the frees will happen outside the tx lock.  If we
7588 		 * can't, it means another thread is still proccessing frees.
7589 		 */
7590 		if (txq->ift_defer_mfree &&
7591 		    atomic_cmpset_acq_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7592 			(uintptr_t )txq->ift_sds.ifsd_m_deferb, 0)) {
7593 			m_defer = txq->ift_sds.ifsd_m_deferb;
7594 		}
7595 		_iflib_completed_tx_reclaim(txq, m_defer, reclaimable);
7596 	}
7597 	mtx_unlock(&txq->ift_mtx);
7598 
7599 	/*
7600 	 * Process mbuf frees outside the tx lock
7601 	 */
7602 	if (m_defer != NULL) {
7603 		for (i = 0; m_defer[i] != NULL; i++) {
7604 			m_freem(m_defer[i]);
7605 			m_defer[i] = NULL;
7606 		}
7607 		atomic_store_rel_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7608 		    (uintptr_t)m_defer);
7609 	}
7610 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
7611 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
7612 	if (mcast_sent)
7613 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
7614 
7615 	return (error);
7616 }
7617 #endif
7618