2 * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are met:
8 * 1. Redistributions of source code must retain the above copyright notice,
9 * this list of conditions and the following disclaimer.
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.
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.
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
35 #include <sys/param.h>
36 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.h>
40 #include <sys/kernel.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
83 #include <net/iflib.h>
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
97 * enable accounting of every mbuf as it comes in to and goes out of iflib's software descriptor references
99 #define MEMORY_LOGGING 0
101 * Enable mbuf vectors for compressing long mbuf chains
106 * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
107 * we prefetch needs to be determined by the time spent in m_free vis a vis
108 * the cost of a prefetch. This will of course vary based on the workload:
109 * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
110 * is quite expensive, thus suggesting very little prefetch.
111 * - small packet forwarding which is just returning a single mbuf to
112 * UMA will typically be very fast vis a vis the cost of a memory
119 * - private structures
120 * - iflib private utility functions
122 * - vlan registry and other exported functions
123 * - iflib public core functions
127 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 typedef struct iflib_txq *iflib_txq_t;
132 typedef struct iflib_rxq *iflib_rxq_t;
134 typedef struct iflib_fl *iflib_fl_t;
138 typedef struct iflib_filter_info {
139 driver_filter_t *ifi_filter;
140 void *ifi_filter_arg;
141 struct grouptask *ifi_task;
142 struct iflib_ctx *ifi_ctx;
143 } *iflib_filter_info_t;
148 * Pointer to hardware driver's softc
155 if_shared_ctx_t ifc_sctx;
156 struct if_softc_ctx ifc_softc_ctx;
160 uint16_t ifc_nhwtxqs;
161 uint16_t ifc_nhwrxqs;
163 iflib_txq_t ifc_txqs;
164 iflib_rxq_t ifc_rxqs;
165 uint32_t ifc_if_flags;
167 uint32_t ifc_max_fl_buf_size;
172 int ifc_pause_frames;
173 int ifc_watchdog_events;
174 struct cdev *ifc_led_dev;
175 struct resource *ifc_msix_mem;
177 struct if_irq ifc_legacy_irq;
178 struct grouptask ifc_admin_task;
179 struct grouptask ifc_vflr_task;
180 struct iflib_filter_info ifc_filter_info;
181 struct ifmedia ifc_media;
183 struct sysctl_oid *ifc_sysctl_node;
184 uint16_t ifc_sysctl_ntxqs;
185 uint16_t ifc_sysctl_nrxqs;
186 uint16_t ifc_sysctl_qs_eq_override;
188 uint16_t ifc_sysctl_ntxds[8];
189 uint16_t ifc_sysctl_nrxds[8];
190 struct if_txrx ifc_txrx;
191 #define isc_txd_encap ifc_txrx.ift_txd_encap
192 #define isc_txd_flush ifc_txrx.ift_txd_flush
193 #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update
194 #define isc_rxd_available ifc_txrx.ift_rxd_available
195 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
197 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
198 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
201 eventhandler_tag ifc_vlan_attach_event;
202 eventhandler_tag ifc_vlan_detach_event;
203 uint8_t ifc_mac[ETHER_ADDR_LEN];
204 char ifc_mtx_name[16];
209 iflib_get_softc(if_ctx_t ctx)
212 return (ctx->ifc_softc);
216 iflib_get_dev(if_ctx_t ctx)
219 return (ctx->ifc_dev);
223 iflib_get_ifp(if_ctx_t ctx)
226 return (ctx->ifc_ifp);
230 iflib_get_media(if_ctx_t ctx)
233 return (&ctx->ifc_media);
237 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
240 bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
244 iflib_get_softc_ctx(if_ctx_t ctx)
247 return (&ctx->ifc_softc_ctx);
251 iflib_get_sctx(if_ctx_t ctx)
254 return (ctx->ifc_sctx);
257 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
259 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
260 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
262 #define RX_SW_DESC_MAP_CREATED (1 << 0)
263 #define TX_SW_DESC_MAP_CREATED (1 << 1)
264 #define RX_SW_DESC_INUSE (1 << 3)
265 #define TX_SW_DESC_MAPPED (1 << 4)
267 typedef struct iflib_sw_rx_desc_array {
268 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */
269 struct mbuf **ifsd_m; /* pkthdr mbufs */
270 caddr_t *ifsd_cl; /* direct cluster pointer for rx */
272 } iflib_rxsd_array_t;
274 typedef struct iflib_sw_tx_desc_array {
275 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */
276 struct mbuf **ifsd_m; /* pkthdr mbufs */
278 } iflib_txsd_array_t;
281 /* magic number that should be high enough for any hardware */
282 #define IFLIB_MAX_TX_SEGS 128
283 #define IFLIB_MAX_RX_SEGS 32
284 #define IFLIB_RX_COPY_THRESH 63
285 #define IFLIB_MAX_RX_REFRESH 32
286 #define IFLIB_QUEUE_IDLE 0
287 #define IFLIB_QUEUE_HUNG 1
288 #define IFLIB_QUEUE_WORKING 2
290 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */
291 #define TX_BATCH_SIZE 16
293 #define IFLIB_RESTART_BUDGET 8
295 #define IFC_LEGACY 0x01
296 #define IFC_QFLUSH 0x02
297 #define IFC_MULTISEG 0x04
298 #define IFC_DMAR 0x08
299 #define IFC_SC_ALLOCATED 0x10
300 #define IFC_INIT_DONE 0x20
303 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
304 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
305 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
309 uint16_t ift_cidx_processed;
312 uint8_t ift_db_pending;
313 uint8_t ift_db_pending_queued;
314 uint8_t ift_npending;
315 uint8_t ift_br_offset;
317 uint64_t ift_processed;
318 uint64_t ift_cleaned;
320 uint64_t ift_enqueued;
321 uint64_t ift_dequeued;
323 uint64_t ift_no_tx_dma_setup;
324 uint64_t ift_no_desc_avail;
325 uint64_t ift_mbuf_defrag_failed;
326 uint64_t ift_mbuf_defrag;
327 uint64_t ift_map_failed;
328 uint64_t ift_txd_encap_efbig;
329 uint64_t ift_pullups;
332 struct mtx ift_db_mtx;
334 /* constant values */
336 struct ifmp_ring **ift_br;
337 struct grouptask ift_task;
340 struct callout ift_timer;
341 struct callout ift_db_check;
343 iflib_txsd_array_t ift_sds;
348 int ift_watchdog_time;
349 struct iflib_filter_info ift_filter_info;
350 bus_dma_tag_t ift_desc_tag;
351 bus_dma_tag_t ift_tso_desc_tag;
352 iflib_dma_info_t ift_ifdi;
353 #define MTX_NAME_LEN 16
354 char ift_mtx_name[MTX_NAME_LEN];
355 char ift_db_mtx_name[MTX_NAME_LEN];
356 bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE);
357 #ifdef IFLIB_DIAGNOSTICS
358 uint64_t ift_cpu_exec_count[256];
360 } __aligned(CACHE_LINE_SIZE);
365 uint16_t ifl_credits;
368 uint64_t ifl_m_enqueued;
369 uint64_t ifl_m_dequeued;
370 uint64_t ifl_cl_enqueued;
371 uint64_t ifl_cl_dequeued;
377 uint16_t ifl_buf_size;
380 iflib_rxsd_array_t ifl_sds;
383 bus_dma_tag_t ifl_desc_tag;
384 iflib_dma_info_t ifl_ifdi;
385 uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
386 caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
387 } __aligned(CACHE_LINE_SIZE);
390 get_inuse(int size, int cidx, int pidx, int gen)
396 else if (pidx < cidx)
397 used = size - cidx + pidx;
398 else if (gen == 0 && pidx == cidx)
400 else if (gen == 1 && pidx == cidx)
408 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
410 #define IDXDIFF(head, tail, wrap) \
411 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
414 /* If there is a separate completion queue -
415 * these are the cq cidx and pidx. Otherwise
419 uint16_t ifr_cq_cidx;
420 uint16_t ifr_cq_pidx;
422 uint8_t ifr_fl_offset;
428 uint8_t ifr_lro_enabled;
430 struct lro_ctrl ifr_lc;
431 struct grouptask ifr_task;
432 struct iflib_filter_info ifr_filter_info;
433 iflib_dma_info_t ifr_ifdi;
434 /* dynamically allocate if any drivers need a value substantially larger than this */
435 struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
436 #ifdef IFLIB_DIAGNOSTICS
437 uint64_t ifr_cpu_exec_count[256];
439 } __aligned(CACHE_LINE_SIZE);
442 * Only allow a single packet to take up most 1/nth of the tx ring
444 #define MAX_SINGLE_PACKET_FRACTION 12
445 #define IF_BAD_DMA (bus_addr_t)-1
447 static int enable_msix = 1;
449 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
451 #define CTX_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
453 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
454 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
455 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
458 #define TXDB_LOCK_INIT(txq) mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF)
459 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx)
460 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx)
461 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx)
462 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx)
464 #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx)
465 #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx)
468 /* Our boot-time initialization hook */
469 static int iflib_module_event_handler(module_t, int, void *);
471 static moduledata_t iflib_moduledata = {
473 iflib_module_event_handler,
477 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
478 MODULE_VERSION(iflib, 1);
480 MODULE_DEPEND(iflib, pci, 1, 1, 1);
481 MODULE_DEPEND(iflib, ether, 1, 1, 1);
483 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
484 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
486 #ifndef IFLIB_DEBUG_COUNTERS
488 #define IFLIB_DEBUG_COUNTERS 1
490 #define IFLIB_DEBUG_COUNTERS 0
491 #endif /* !INVARIANTS */
494 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
495 "iflib driver parameters");
498 * XXX need to ensure that this can't accidentally cause the head to be moved backwards
500 static int iflib_min_tx_latency = 0;
502 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
503 &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
506 #if IFLIB_DEBUG_COUNTERS
508 static int iflib_tx_seen;
509 static int iflib_tx_sent;
510 static int iflib_tx_encap;
511 static int iflib_rx_allocs;
512 static int iflib_fl_refills;
513 static int iflib_fl_refills_large;
514 static int iflib_tx_frees;
516 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
517 &iflib_tx_seen, 0, "# tx mbufs seen");
518 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
519 &iflib_tx_sent, 0, "# tx mbufs sent");
520 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
521 &iflib_tx_encap, 0, "# tx mbufs encapped");
522 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
523 &iflib_tx_frees, 0, "# tx frees");
524 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
525 &iflib_rx_allocs, 0, "# rx allocations");
526 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
527 &iflib_fl_refills, 0, "# refills");
528 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
529 &iflib_fl_refills_large, 0, "# large refills");
532 static int iflib_txq_drain_flushing;
533 static int iflib_txq_drain_oactive;
534 static int iflib_txq_drain_notready;
535 static int iflib_txq_drain_encapfail;
537 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
538 &iflib_txq_drain_flushing, 0, "# drain flushes");
539 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
540 &iflib_txq_drain_oactive, 0, "# drain oactives");
541 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
542 &iflib_txq_drain_notready, 0, "# drain notready");
543 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
544 &iflib_txq_drain_encapfail, 0, "# drain encap fails");
547 static int iflib_encap_load_mbuf_fail;
548 static int iflib_encap_txq_avail_fail;
549 static int iflib_encap_txd_encap_fail;
551 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
552 &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
553 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
554 &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
555 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
556 &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
558 static int iflib_task_fn_rxs;
559 static int iflib_rx_intr_enables;
560 static int iflib_fast_intrs;
561 static int iflib_intr_link;
562 static int iflib_intr_msix;
563 static int iflib_rx_unavail;
564 static int iflib_rx_ctx_inactive;
565 static int iflib_rx_zero_len;
566 static int iflib_rx_if_input;
567 static int iflib_rx_mbuf_null;
568 static int iflib_rxd_flush;
570 static int iflib_verbose_debug;
572 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
573 &iflib_intr_link, 0, "# intr link calls");
574 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
575 &iflib_intr_msix, 0, "# intr msix calls");
576 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
577 &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
578 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
579 &iflib_rx_intr_enables, 0, "# rx intr enables");
580 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
581 &iflib_fast_intrs, 0, "# fast_intr calls");
582 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
583 &iflib_rx_unavail, 0, "# times rxeof called with no available data");
584 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
585 &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
586 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
587 &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
588 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
589 &iflib_rx_if_input, 0, "# times rxeof called if_input");
590 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
591 &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
592 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
593 &iflib_rxd_flush, 0, "# times rxd_flush called");
594 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
595 &iflib_verbose_debug, 0, "enable verbose debugging");
597 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
599 iflib_debug_reset(void)
601 iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
602 iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
603 iflib_txq_drain_flushing = iflib_txq_drain_oactive =
604 iflib_txq_drain_notready = iflib_txq_drain_encapfail =
605 iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
606 iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
607 iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
608 iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
609 iflib_rx_mbuf_null = iflib_rxd_flush = 0;
613 #define DBG_COUNTER_INC(name)
614 static void iflib_debug_reset(void) {}
619 #define IFLIB_DEBUG 0
621 static void iflib_tx_structures_free(if_ctx_t ctx);
622 static void iflib_rx_structures_free(if_ctx_t ctx);
623 static int iflib_queues_alloc(if_ctx_t ctx);
624 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
625 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget);
626 static int iflib_qset_structures_setup(if_ctx_t ctx);
627 static int iflib_msix_init(if_ctx_t ctx);
628 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
629 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
630 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
631 static int iflib_register(if_ctx_t);
632 static void iflib_init_locked(if_ctx_t ctx);
633 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
634 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
635 static void iflib_ifmp_purge(iflib_txq_t txq);
636 static void _iflib_pre_assert(if_softc_ctx_t scctx);
639 #include <sys/selinfo.h>
640 #include <net/netmap.h>
641 #include <dev/netmap/netmap_kern.h>
643 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
646 * device-specific sysctl variables:
648 * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
649 * During regular operations the CRC is stripped, but on some
650 * hardware reception of frames not multiple of 64 is slower,
651 * so using crcstrip=0 helps in benchmarks.
653 * iflib_rx_miss, iflib_rx_miss_bufs:
654 * count packets that might be missed due to lost interrupts.
656 SYSCTL_DECL(_dev_netmap);
658 * The xl driver by default strips CRCs and we do not override it.
661 int iflib_crcstrip = 1;
662 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
663 CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
665 int iflib_rx_miss, iflib_rx_miss_bufs;
666 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
667 CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
668 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
669 CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
672 * Register/unregister. We are already under netmap lock.
673 * Only called on the first register or the last unregister.
676 iflib_netmap_register(struct netmap_adapter *na, int onoff)
678 struct ifnet *ifp = na->ifp;
679 if_ctx_t ctx = ifp->if_softc;
682 IFDI_INTR_DISABLE(ctx);
684 /* Tell the stack that the interface is no longer active */
685 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
688 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
690 /* enable or disable flags and callbacks in na and ifp */
692 nm_set_native_flags(na);
694 nm_clear_native_flags(na);
697 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
699 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
703 * Reconcile kernel and user view of the transmit ring.
705 * All information is in the kring.
706 * Userspace wants to send packets up to the one before kring->rhead,
707 * kernel knows kring->nr_hwcur is the first unsent packet.
709 * Here we push packets out (as many as possible), and possibly
710 * reclaim buffers from previously completed transmission.
712 * The caller (netmap) guarantees that there is only one instance
713 * running at any time. Any interference with other driver
714 * methods should be handled by the individual drivers.
717 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
719 struct netmap_adapter *na = kring->na;
720 struct ifnet *ifp = na->ifp;
721 struct netmap_ring *ring = kring->ring;
722 u_int nm_i; /* index into the netmap ring */
723 u_int nic_i; /* index into the NIC ring */
725 u_int const lim = kring->nkr_num_slots - 1;
726 u_int const head = kring->rhead;
727 struct if_pkt_info pi;
730 * interrupts on every tx packet are expensive so request
731 * them every half ring, or where NS_REPORT is set
733 u_int report_frequency = kring->nkr_num_slots >> 1;
734 /* device-specific */
735 if_ctx_t ctx = ifp->if_softc;
736 iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
738 pi.ipi_segs = txq->ift_segs;
739 pi.ipi_qsidx = kring->ring_id;
742 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
743 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
747 * First part: process new packets to send.
748 * nm_i is the current index in the netmap ring,
749 * nic_i is the corresponding index in the NIC ring.
751 * If we have packets to send (nm_i != head)
752 * iterate over the netmap ring, fetch length and update
753 * the corresponding slot in the NIC ring. Some drivers also
754 * need to update the buffer's physical address in the NIC slot
755 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
757 * The netmap_reload_map() calls is especially expensive,
758 * even when (as in this case) the tag is 0, so do only
759 * when the buffer has actually changed.
761 * If possible do not set the report/intr bit on all slots,
762 * but only a few times per ring or when NS_REPORT is set.
764 * Finally, on 10G and faster drivers, it might be useful
765 * to prefetch the next slot and txr entry.
768 nm_i = kring->nr_hwcur;
769 if (nm_i != head) { /* we have new packets to send */
770 nic_i = netmap_idx_k2n(kring, nm_i);
772 __builtin_prefetch(&ring->slot[nm_i]);
773 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
774 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
776 for (n = 0; nm_i != head; n++) {
777 struct netmap_slot *slot = &ring->slot[nm_i];
778 u_int len = slot->len;
780 void *addr = PNMB(na, slot, &paddr);
781 int flags = (slot->flags & NS_REPORT ||
782 nic_i == 0 || nic_i == report_frequency) ?
785 /* device-specific */
787 pi.ipi_flags = flags;
789 /* Fill the slot in the NIC ring. */
790 ctx->isc_txd_encap(ctx->ifc_softc, &pi);
792 /* prefetch for next round */
793 __builtin_prefetch(&ring->slot[nm_i + 1]);
794 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
795 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
797 NM_CHECK_ADDR_LEN(na, addr, len);
799 if (slot->flags & NS_BUF_CHANGED) {
800 /* buffer has changed, reload map */
801 netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
803 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
805 /* make sure changes to the buffer are synced */
806 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
807 BUS_DMASYNC_PREWRITE);
809 nm_i = nm_next(nm_i, lim);
810 nic_i = nm_next(nic_i, lim);
812 kring->nr_hwcur = head;
814 /* synchronize the NIC ring */
815 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
816 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
818 /* (re)start the tx unit up to slot nic_i (excluded) */
819 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
823 * Second part: reclaim buffers for completed transmissions.
825 if (iflib_tx_credits_update(ctx, txq)) {
826 /* some tx completed, increment avail */
827 nic_i = txq->ift_cidx_processed;
828 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
834 * Reconcile kernel and user view of the receive ring.
835 * Same as for the txsync, this routine must be efficient.
836 * The caller guarantees a single invocations, but races against
837 * the rest of the driver should be handled here.
839 * On call, kring->rhead is the first packet that userspace wants
840 * to keep, and kring->rcur is the wakeup point.
841 * The kernel has previously reported packets up to kring->rtail.
843 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
844 * of whether or not we received an interrupt.
847 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
849 struct netmap_adapter *na = kring->na;
850 struct ifnet *ifp = na->ifp;
851 struct netmap_ring *ring = kring->ring;
852 u_int nm_i; /* index into the netmap ring */
853 u_int nic_i; /* index into the NIC ring */
855 u_int const lim = kring->nkr_num_slots - 1;
856 u_int const head = kring->rhead;
857 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
858 struct if_rxd_info ri;
859 /* device-specific */
860 if_ctx_t ctx = ifp->if_softc;
861 iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
862 iflib_fl_t fl = rxq->ifr_fl;
864 return netmap_ring_reinit(kring);
866 bzero(&ri, sizeof(ri));
867 ri.iri_qsidx = kring->ring_id;
868 ri.iri_ifp = ctx->ifc_ifp;
869 /* XXX check sync modes */
870 for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++)
871 bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
872 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
875 * First part: import newly received packets.
877 * nm_i is the index of the next free slot in the netmap ring,
878 * nic_i is the index of the next received packet in the NIC ring,
879 * and they may differ in case if_init() has been called while
880 * in netmap mode. For the receive ring we have
882 * nic_i = rxr->next_check;
883 * nm_i = kring->nr_hwtail (previous)
885 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
887 * rxr->next_check is set to 0 on a ring reinit
889 if (netmap_no_pendintr || force_update) {
890 int crclen = iflib_crcstrip ? 0 : 4;
892 uint16_t slot_flags = kring->nkr_slot_flags;
894 for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
895 nic_i = fl->ifl_cidx;
896 nm_i = netmap_idx_n2k(kring, nic_i);
897 avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX);
898 for (n = 0; avail > 0; n++, avail--) {
899 error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
901 ring->slot[nm_i].len = 0;
903 ring->slot[nm_i].len = ri.iri_len - crclen;
904 ring->slot[nm_i].flags = slot_flags;
905 bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
906 fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
907 nm_i = nm_next(nm_i, lim);
908 nic_i = nm_next(nic_i, lim);
910 if (n) { /* update the state variables */
911 if (netmap_no_pendintr && !force_update) {
914 iflib_rx_miss_bufs += n;
916 fl->ifl_cidx = nic_i;
917 kring->nr_hwtail = nm_i;
919 kring->nr_kflags &= ~NKR_PENDINTR;
923 * Second part: skip past packets that userspace has released.
924 * (kring->nr_hwcur to head excluded),
925 * and make the buffers available for reception.
926 * As usual nm_i is the index in the netmap ring,
927 * nic_i is the index in the NIC ring, and
928 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
930 /* XXX not sure how this will work with multiple free lists */
931 nm_i = kring->nr_hwcur;
933 nic_i = netmap_idx_k2n(kring, nm_i);
934 for (n = 0; nm_i != head; n++) {
935 struct netmap_slot *slot = &ring->slot[nm_i];
938 void *addr = PNMB(na, slot, &paddr);
940 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
944 if (slot->flags & NS_BUF_CHANGED) {
945 /* buffer has changed, reload map */
946 netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
947 slot->flags &= ~NS_BUF_CHANGED;
950 * XXX we should be batching this operation - TODO
952 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size);
953 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
954 BUS_DMASYNC_PREREAD);
955 nm_i = nm_next(nm_i, lim);
956 nic_i = nm_next(nic_i, lim);
958 kring->nr_hwcur = head;
960 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
961 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
963 * IMPORTANT: we must leave one free slot in the ring,
964 * so move nic_i back by one unit
966 nic_i = nm_prev(nic_i, lim);
967 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
973 return netmap_ring_reinit(kring);
977 iflib_netmap_attach(if_ctx_t ctx)
979 struct netmap_adapter na;
980 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
982 bzero(&na, sizeof(na));
984 na.ifp = ctx->ifc_ifp;
985 na.na_flags = NAF_BDG_MAYSLEEP;
986 MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
987 MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
989 na.num_tx_desc = scctx->isc_ntxd[0];
990 na.num_rx_desc = scctx->isc_nrxd[0];
991 na.nm_txsync = iflib_netmap_txsync;
992 na.nm_rxsync = iflib_netmap_rxsync;
993 na.nm_register = iflib_netmap_register;
994 na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
995 na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
996 return (netmap_attach(&na));
1000 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1002 struct netmap_adapter *na = NA(ctx->ifc_ifp);
1003 struct netmap_slot *slot;
1005 slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1009 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1012 * In netmap mode, set the map for the packet buffer.
1013 * NOTE: Some drivers (not this one) also need to set
1014 * the physical buffer address in the NIC ring.
1015 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1016 * netmap slot index, si
1018 int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1019 netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1023 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1025 struct netmap_adapter *na = NA(ctx->ifc_ifp);
1026 struct netmap_slot *slot;
1030 slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1033 map = rxq->ifr_fl[0].ifl_sds.ifsd_map;
1034 nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1035 for (int i = 0; i < nrxd; i++, map++) {
1036 int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1041 vaddr = addr = PNMB(na, slot + sj, &paddr);
1042 netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1043 /* Update descriptor and the cached value */
1044 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size);
1046 /* preserve queue */
1047 if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1048 struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1049 int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1050 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1052 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1055 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1058 #define iflib_netmap_txq_init(ctx, txq)
1059 #define iflib_netmap_rxq_init(ctx, rxq)
1060 #define iflib_netmap_detach(ifp)
1062 #define iflib_netmap_attach(ctx) (0)
1063 #define netmap_rx_irq(ifp, qid, budget) (0)
1067 #if defined(__i386__) || defined(__amd64__)
1068 static __inline void
1071 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1078 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1082 *(bus_addr_t *) arg = segs[0].ds_addr;
1086 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1089 if_shared_ctx_t sctx = ctx->ifc_sctx;
1090 device_t dev = ctx->ifc_dev;
1092 KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1094 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1095 sctx->isc_q_align, 0, /* alignment, bounds */
1096 BUS_SPACE_MAXADDR, /* lowaddr */
1097 BUS_SPACE_MAXADDR, /* highaddr */
1098 NULL, NULL, /* filter, filterarg */
1101 size, /* maxsegsize */
1102 BUS_DMA_ALLOCNOW, /* flags */
1103 NULL, /* lockfunc */
1108 "%s: bus_dma_tag_create failed: %d\n",
1113 err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1114 BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1117 "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1118 __func__, (uintmax_t)size, err);
1122 dma->idi_paddr = IF_BAD_DMA;
1123 err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1124 size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1125 if (err || dma->idi_paddr == IF_BAD_DMA) {
1127 "%s: bus_dmamap_load failed: %d\n",
1132 dma->idi_size = size;
1136 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1138 bus_dma_tag_destroy(dma->idi_tag);
1140 dma->idi_tag = NULL;
1146 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1149 iflib_dma_info_t *dmaiter;
1152 for (i = 0; i < count; i++, dmaiter++) {
1153 if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1157 iflib_dma_free_multi(dmalist, i);
1162 iflib_dma_free(iflib_dma_info_t dma)
1164 if (dma->idi_tag == NULL)
1166 if (dma->idi_paddr != IF_BAD_DMA) {
1167 bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1168 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1169 bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1170 dma->idi_paddr = IF_BAD_DMA;
1172 if (dma->idi_vaddr != NULL) {
1173 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1174 dma->idi_vaddr = NULL;
1176 bus_dma_tag_destroy(dma->idi_tag);
1177 dma->idi_tag = NULL;
1181 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1184 iflib_dma_info_t *dmaiter = dmalist;
1186 for (i = 0; i < count; i++, dmaiter++)
1187 iflib_dma_free(*dmaiter);
1190 #ifdef EARLY_AP_STARTUP
1191 static const int iflib_started = 1;
1194 * We used to abuse the smp_started flag to decide if the queues have been
1195 * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1196 * That gave bad races, since the SYSINIT() runs strictly after smp_started
1197 * is set. Run a SYSINIT() strictly after that to just set a usable
1201 static int iflib_started;
1204 iflib_record_started(void *arg)
1209 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1210 iflib_record_started, NULL);
1214 iflib_fast_intr(void *arg)
1216 iflib_filter_info_t info = arg;
1217 struct grouptask *gtask = info->ifi_task;
1220 return (FILTER_HANDLED);
1222 DBG_COUNTER_INC(fast_intrs);
1223 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1224 return (FILTER_HANDLED);
1226 GROUPTASK_ENQUEUE(gtask);
1227 return (FILTER_HANDLED);
1231 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1232 driver_filter_t filter, driver_intr_t handler, void *arg,
1236 struct resource *res;
1238 device_t dev = ctx->ifc_dev;
1242 res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1243 RF_SHAREABLE | RF_ACTIVE);
1246 "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1250 KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1251 rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1252 filter, handler, arg, &tag);
1255 "failed to setup interrupt for rid %d, name %s: %d\n",
1256 rid, name ? name : "unknown", rc);
1259 bus_describe_intr(dev, res, tag, "%s", name);
1266 /*********************************************************************
1268 * Allocate memory for tx_buffer structures. The tx_buffer stores all
1269 * the information needed to transmit a packet on the wire. This is
1270 * called only once at attach, setup is done every reset.
1272 **********************************************************************/
1275 iflib_txsd_alloc(iflib_txq_t txq)
1277 if_ctx_t ctx = txq->ift_ctx;
1278 if_shared_ctx_t sctx = ctx->ifc_sctx;
1279 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1280 device_t dev = ctx->ifc_dev;
1281 int err, nsegments, ntsosegments;
1283 nsegments = scctx->isc_tx_nsegments;
1284 ntsosegments = scctx->isc_tx_tso_segments_max;
1285 MPASS(scctx->isc_ntxd[0] > 0);
1286 MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1287 MPASS(nsegments > 0);
1288 MPASS(ntsosegments > 0);
1290 * Setup DMA descriptor areas.
1292 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1293 1, 0, /* alignment, bounds */
1294 BUS_SPACE_MAXADDR, /* lowaddr */
1295 BUS_SPACE_MAXADDR, /* highaddr */
1296 NULL, NULL, /* filter, filterarg */
1297 sctx->isc_tx_maxsize, /* maxsize */
1298 nsegments, /* nsegments */
1299 sctx->isc_tx_maxsegsize, /* maxsegsize */
1301 NULL, /* lockfunc */
1302 NULL, /* lockfuncarg */
1303 &txq->ift_desc_tag))) {
1304 device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1305 device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1306 sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1309 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1310 1, 0, /* alignment, bounds */
1311 BUS_SPACE_MAXADDR, /* lowaddr */
1312 BUS_SPACE_MAXADDR, /* highaddr */
1313 NULL, NULL, /* filter, filterarg */
1314 scctx->isc_tx_tso_size_max, /* maxsize */
1315 ntsosegments, /* nsegments */
1316 scctx->isc_tx_tso_segsize_max, /* maxsegsize */
1318 NULL, /* lockfunc */
1319 NULL, /* lockfuncarg */
1320 &txq->ift_tso_desc_tag))) {
1321 device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1325 if (!(txq->ift_sds.ifsd_flags =
1326 (uint8_t *) malloc(sizeof(uint8_t) *
1327 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1328 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1332 if (!(txq->ift_sds.ifsd_m =
1333 (struct mbuf **) malloc(sizeof(struct mbuf *) *
1334 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1335 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1340 /* Create the descriptor buffer dma maps */
1341 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1342 if ((ctx->ifc_flags & IFC_DMAR) == 0)
1345 if (!(txq->ift_sds.ifsd_map =
1346 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1347 device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1352 for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1353 err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1355 device_printf(dev, "Unable to create TX DMA map\n");
1362 /* We free all, it handles case where we are in the middle */
1363 iflib_tx_structures_free(ctx);
1368 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1373 if (txq->ift_sds.ifsd_map != NULL)
1374 map = txq->ift_sds.ifsd_map[i];
1376 bus_dmamap_unload(txq->ift_desc_tag, map);
1377 bus_dmamap_destroy(txq->ift_desc_tag, map);
1378 txq->ift_sds.ifsd_map[i] = NULL;
1383 iflib_txq_destroy(iflib_txq_t txq)
1385 if_ctx_t ctx = txq->ift_ctx;
1387 for (int i = 0; i < txq->ift_size; i++)
1388 iflib_txsd_destroy(ctx, txq, i);
1389 if (txq->ift_sds.ifsd_map != NULL) {
1390 free(txq->ift_sds.ifsd_map, M_IFLIB);
1391 txq->ift_sds.ifsd_map = NULL;
1393 if (txq->ift_sds.ifsd_m != NULL) {
1394 free(txq->ift_sds.ifsd_m, M_IFLIB);
1395 txq->ift_sds.ifsd_m = NULL;
1397 if (txq->ift_sds.ifsd_flags != NULL) {
1398 free(txq->ift_sds.ifsd_flags, M_IFLIB);
1399 txq->ift_sds.ifsd_flags = NULL;
1401 if (txq->ift_desc_tag != NULL) {
1402 bus_dma_tag_destroy(txq->ift_desc_tag);
1403 txq->ift_desc_tag = NULL;
1405 if (txq->ift_tso_desc_tag != NULL) {
1406 bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1407 txq->ift_tso_desc_tag = NULL;
1412 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1416 mp = &txq->ift_sds.ifsd_m[i];
1420 if (txq->ift_sds.ifsd_map != NULL) {
1421 bus_dmamap_sync(txq->ift_desc_tag,
1422 txq->ift_sds.ifsd_map[i],
1423 BUS_DMASYNC_POSTWRITE);
1424 bus_dmamap_unload(txq->ift_desc_tag,
1425 txq->ift_sds.ifsd_map[i]);
1428 DBG_COUNTER_INC(tx_frees);
1433 iflib_txq_setup(iflib_txq_t txq)
1435 if_ctx_t ctx = txq->ift_ctx;
1436 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1437 iflib_dma_info_t di;
1440 /* Set number of descriptors available */
1441 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1444 txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1445 txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1447 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1448 bzero((void *)di->idi_vaddr, di->idi_size);
1450 IFDI_TXQ_SETUP(ctx, txq->ift_id);
1451 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1452 bus_dmamap_sync(di->idi_tag, di->idi_map,
1453 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1457 /*********************************************************************
1459 * Allocate memory for rx_buffer structures. Since we use one
1460 * rx_buffer per received packet, the maximum number of rx_buffer's
1461 * that we'll need is equal to the number of receive descriptors
1462 * that we've allocated.
1464 **********************************************************************/
1466 iflib_rxsd_alloc(iflib_rxq_t rxq)
1468 if_ctx_t ctx = rxq->ifr_ctx;
1469 if_shared_ctx_t sctx = ctx->ifc_sctx;
1470 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1471 device_t dev = ctx->ifc_dev;
1475 MPASS(scctx->isc_nrxd[0] > 0);
1476 MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1479 for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1480 fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1481 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1482 1, 0, /* alignment, bounds */
1483 BUS_SPACE_MAXADDR, /* lowaddr */
1484 BUS_SPACE_MAXADDR, /* highaddr */
1485 NULL, NULL, /* filter, filterarg */
1486 sctx->isc_rx_maxsize, /* maxsize */
1487 sctx->isc_rx_nsegments, /* nsegments */
1488 sctx->isc_rx_maxsegsize, /* maxsegsize */
1490 NULL, /* lockfunc */
1494 device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1498 if (!(fl->ifl_sds.ifsd_flags =
1499 (uint8_t *) malloc(sizeof(uint8_t) *
1500 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1501 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1505 if (!(fl->ifl_sds.ifsd_m =
1506 (struct mbuf **) malloc(sizeof(struct mbuf *) *
1507 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1508 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1512 if (!(fl->ifl_sds.ifsd_cl =
1513 (caddr_t *) malloc(sizeof(caddr_t) *
1514 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1515 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1520 /* Create the descriptor buffer dma maps */
1521 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1522 if ((ctx->ifc_flags & IFC_DMAR) == 0)
1525 if (!(fl->ifl_sds.ifsd_map =
1526 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1527 device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1532 for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1533 err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1535 device_printf(dev, "Unable to create TX DMA map\n");
1544 iflib_rx_structures_free(ctx);
1550 * Internal service routines
1553 struct rxq_refill_cb_arg {
1555 bus_dma_segment_t seg;
1560 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1562 struct rxq_refill_cb_arg *cb_arg = arg;
1564 cb_arg->error = error;
1565 cb_arg->seg = segs[0];
1566 cb_arg->nseg = nseg;
1571 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1573 #define IS_DMAR(ctx) (0)
1577 * rxq_refill - refill an rxq free-buffer list
1578 * @ctx: the iflib context
1579 * @rxq: the free-list to refill
1580 * @n: the number of new buffers to allocate
1582 * (Re)populate an rxq free-buffer list with up to @n new packet buffers.
1583 * The caller must assure that @n does not exceed the queue's capacity.
1586 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1589 int idx, pidx = fl->ifl_pidx;
1593 bus_dmamap_t *sd_map;
1598 sd_m = fl->ifl_sds.ifsd_m;
1599 sd_map = fl->ifl_sds.ifsd_map;
1600 sd_cl = fl->ifl_sds.ifsd_cl;
1601 sd_flags = fl->ifl_sds.ifsd_flags;
1606 MPASS(fl->ifl_credits + n <= fl->ifl_size);
1608 if (pidx < fl->ifl_cidx)
1609 MPASS(pidx + n <= fl->ifl_cidx);
1610 if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1611 MPASS(fl->ifl_gen == 0);
1612 if (pidx > fl->ifl_cidx)
1613 MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1615 DBG_COUNTER_INC(fl_refills);
1617 DBG_COUNTER_INC(fl_refills_large);
1621 * We allocate an uninitialized mbuf + cluster, mbuf is
1622 * initialized after rx.
1624 * If the cluster is still set then we know a minimum sized packet was received
1626 if ((cl = sd_cl[idx]) == NULL) {
1627 if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1630 fl->ifl_cl_enqueued++;
1633 if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1637 fl->ifl_m_enqueued++;
1640 DBG_COUNTER_INC(rx_allocs);
1642 if ((sd_flags[pidx] & RX_SW_DESC_MAP_CREATED) == 0) {
1645 if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &sd_map[idx]))) {
1646 log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
1647 uma_zfree(fl->ifl_zone, cl);
1651 sd_flags[idx] |= RX_SW_DESC_MAP_CREATED;
1654 #if defined(__i386__) || defined(__amd64__)
1655 if (!IS_DMAR(ctx)) {
1656 bus_addr = pmap_kextract((vm_offset_t)cl);
1660 struct rxq_refill_cb_arg cb_arg;
1665 err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1666 cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1668 if (err != 0 || cb_arg.error) {
1672 if (fl->ifl_zone == zone_pack)
1673 uma_zfree(fl->ifl_zone, cl);
1678 bus_addr = cb_arg.seg.ds_addr;
1680 sd_flags[idx] |= RX_SW_DESC_INUSE;
1682 MPASS(sd_m[idx] == NULL);
1685 fl->ifl_bus_addrs[i] = bus_addr;
1686 fl->ifl_vm_addrs[i] = cl;
1689 MPASS(fl->ifl_credits <= fl->ifl_size);
1690 if (++idx == fl->ifl_size) {
1694 if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1695 ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx,
1696 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size);
1704 DBG_COUNTER_INC(rxd_flush);
1705 if (fl->ifl_pidx == 0)
1706 pidx = fl->ifl_size - 1;
1708 pidx = fl->ifl_pidx - 1;
1709 ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1712 static __inline void
1713 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1715 /* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1716 int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1718 int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1721 MPASS(fl->ifl_credits <= fl->ifl_size);
1722 MPASS(reclaimable == delta);
1724 if (reclaimable > 0)
1725 _iflib_fl_refill(ctx, fl, min(max, reclaimable));
1729 iflib_fl_bufs_free(iflib_fl_t fl)
1731 iflib_dma_info_t idi = fl->ifl_ifdi;
1734 for (i = 0; i < fl->ifl_size; i++) {
1735 struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1736 uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1737 caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1739 if (*sd_flags & RX_SW_DESC_INUSE) {
1740 if (fl->ifl_sds.ifsd_map != NULL) {
1741 bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1742 bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1743 bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1745 if (*sd_m != NULL) {
1746 m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1747 uma_zfree(zone_mbuf, *sd_m);
1750 uma_zfree(fl->ifl_zone, *sd_cl);
1753 MPASS(*sd_cl == NULL);
1754 MPASS(*sd_m == NULL);
1757 fl->ifl_m_dequeued++;
1758 fl->ifl_cl_dequeued++;
1764 * Reset free list values
1766 fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1767 bzero(idi->idi_vaddr, idi->idi_size);
1770 /*********************************************************************
1772 * Initialize a receive ring and its buffers.
1774 **********************************************************************/
1776 iflib_fl_setup(iflib_fl_t fl)
1778 iflib_rxq_t rxq = fl->ifl_rxq;
1779 if_ctx_t ctx = rxq->ifr_ctx;
1780 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1783 ** Free current RX buffer structs and their mbufs
1785 iflib_fl_bufs_free(fl);
1786 /* Now replenish the mbufs */
1787 MPASS(fl->ifl_credits == 0);
1789 * XXX don't set the max_frame_size to larger
1790 * than the hardware can handle
1792 if (sctx->isc_max_frame_size <= 2048)
1793 fl->ifl_buf_size = MCLBYTES;
1794 else if (sctx->isc_max_frame_size <= 4096)
1795 fl->ifl_buf_size = MJUMPAGESIZE;
1796 else if (sctx->isc_max_frame_size <= 9216)
1797 fl->ifl_buf_size = MJUM9BYTES;
1799 fl->ifl_buf_size = MJUM16BYTES;
1800 if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
1801 ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
1802 fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
1803 fl->ifl_zone = m_getzone(fl->ifl_buf_size);
1806 /* avoid pre-allocating zillions of clusters to an idle card
1807 * potentially speeding up attach
1809 _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
1810 MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
1811 if (min(128, fl->ifl_size) != fl->ifl_credits)
1817 MPASS(fl->ifl_ifdi != NULL);
1818 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1819 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1823 /*********************************************************************
1825 * Free receive ring data structures
1827 **********************************************************************/
1829 iflib_rx_sds_free(iflib_rxq_t rxq)
1834 if (rxq->ifr_fl != NULL) {
1835 for (i = 0; i < rxq->ifr_nfl; i++) {
1836 fl = &rxq->ifr_fl[i];
1837 if (fl->ifl_desc_tag != NULL) {
1838 bus_dma_tag_destroy(fl->ifl_desc_tag);
1839 fl->ifl_desc_tag = NULL;
1841 free(fl->ifl_sds.ifsd_m, M_IFLIB);
1842 free(fl->ifl_sds.ifsd_cl, M_IFLIB);
1843 /* XXX destroy maps first */
1844 free(fl->ifl_sds.ifsd_map, M_IFLIB);
1845 fl->ifl_sds.ifsd_m = NULL;
1846 fl->ifl_sds.ifsd_cl = NULL;
1847 fl->ifl_sds.ifsd_map = NULL;
1849 free(rxq->ifr_fl, M_IFLIB);
1851 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
1856 * MI independent logic
1860 iflib_timer(void *arg)
1862 iflib_txq_t txq = arg;
1863 if_ctx_t ctx = txq->ift_ctx;
1864 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1866 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
1869 ** Check on the state of the TX queue(s), this
1870 ** can be done without the lock because its RO
1871 ** and the HUNG state will be static if set.
1873 IFDI_TIMER(ctx, txq->ift_id);
1874 if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
1875 (ctx->ifc_pause_frames == 0))
1878 if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments ||
1879 ifmp_ring_is_stalled(txq->ift_br[0]))
1880 GROUPTASK_ENQUEUE(&txq->ift_task);
1882 ctx->ifc_pause_frames = 0;
1883 if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
1884 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
1888 if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
1889 device_printf(ctx->ifc_dev, "TX(%d) desc avail = %d, pidx = %d\n",
1890 txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
1892 IFDI_WATCHDOG_RESET(ctx);
1893 ctx->ifc_watchdog_events++;
1894 ctx->ifc_pause_frames = 0;
1896 iflib_init_locked(ctx);
1901 iflib_init_locked(if_ctx_t ctx)
1903 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1904 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1905 if_t ifp = ctx->ifc_ifp;
1909 int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
1912 if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1913 IFDI_INTR_DISABLE(ctx);
1915 tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
1916 tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
1917 /* Set hardware offload abilities */
1918 if_clearhwassist(ifp);
1919 if (if_getcapenable(ifp) & IFCAP_TXCSUM)
1920 if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
1921 if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
1922 if_sethwassistbits(ifp, tx_ip6_csum_flags, 0);
1923 if (if_getcapenable(ifp) & IFCAP_TSO4)
1924 if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
1925 if (if_getcapenable(ifp) & IFCAP_TSO6)
1926 if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
1928 for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
1930 callout_stop(&txq->ift_timer);
1931 callout_stop(&txq->ift_db_check);
1932 CALLOUT_UNLOCK(txq);
1933 iflib_netmap_txq_init(ctx, txq);
1935 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1936 iflib_netmap_rxq_init(ctx, rxq);
1939 i = if_getdrvflags(ifp);
1942 MPASS(if_getdrvflags(ifp) == i);
1943 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1944 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
1945 if (iflib_fl_setup(fl)) {
1946 device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
1952 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
1953 IFDI_INTR_ENABLE(ctx);
1954 txq = ctx->ifc_txqs;
1955 for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
1956 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
1957 txq->ift_timer.c_cpu);
1961 iflib_media_change(if_t ifp)
1963 if_ctx_t ctx = if_getsoftc(ifp);
1967 if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
1968 iflib_init_locked(ctx);
1974 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
1976 if_ctx_t ctx = if_getsoftc(ifp);
1979 IFDI_UPDATE_ADMIN_STATUS(ctx);
1980 IFDI_MEDIA_STATUS(ctx, ifmr);
1985 iflib_stop(if_ctx_t ctx)
1987 iflib_txq_t txq = ctx->ifc_txqs;
1988 iflib_rxq_t rxq = ctx->ifc_rxqs;
1989 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1990 iflib_dma_info_t di;
1994 /* Tell the stack that the interface is no longer active */
1995 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1997 IFDI_INTR_DISABLE(ctx);
2002 iflib_debug_reset();
2003 /* Wait for current tx queue users to exit to disarm watchdog timer. */
2004 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2005 /* make sure all transmitters have completed before proceeding XXX */
2007 /* clean any enqueued buffers */
2008 iflib_ifmp_purge(txq);
2009 /* Free any existing tx buffers. */
2010 for (j = 0; j < txq->ift_size; j++) {
2011 iflib_txsd_free(ctx, txq, j);
2013 txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2014 txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2015 txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2016 txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2017 txq->ift_pullups = 0;
2018 ifmp_ring_reset_stats(txq->ift_br[0]);
2019 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2020 bzero((void *)di->idi_vaddr, di->idi_size);
2022 for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2023 /* make sure all transmitters have completed before proceeding XXX */
2025 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2026 bzero((void *)di->idi_vaddr, di->idi_size);
2027 /* also resets the free lists pidx/cidx */
2028 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2029 iflib_fl_bufs_free(fl);
2034 prefetch_pkts(iflib_fl_t fl, int cidx)
2037 int nrxd = fl->ifl_size;
2039 nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2040 prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2041 prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2042 prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2043 prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2044 prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2045 prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2046 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2047 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2048 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2049 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2053 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload, iflib_fl_t *pfl, int *pcidx)
2058 iflib_dma_info_t di;
2061 flid = irf->irf_flid;
2062 cidx = irf->irf_idx;
2063 fl = &rxq->ifr_fl[flid];
2066 fl->ifl_m_dequeued++;
2068 fl->ifl_cl_dequeued++;
2070 prefetch_pkts(fl, cidx);
2071 if (fl->ifl_sds.ifsd_map != NULL) {
2072 next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2073 prefetch(&fl->ifl_sds.ifsd_map[next]);
2074 map = fl->ifl_sds.ifsd_map[cidx];
2076 next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2077 prefetch(&fl->ifl_sds.ifsd_flags[next]);
2078 bus_dmamap_sync(di->idi_tag, di->idi_map,
2079 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2081 /* not valid assert if bxe really does SGE from non-contiguous elements */
2082 MPASS(fl->ifl_cidx == cidx);
2084 bus_dmamap_unload(fl->ifl_desc_tag, map);
2086 if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) {
2092 *cltype = fl->ifl_cltype;
2097 static struct mbuf *
2098 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri)
2100 int i, padlen , flags, cltype;
2101 struct mbuf *m, *mh, *mt, *sd_m;
2109 rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE, &fl, &cidx);
2110 sd_m = fl->ifl_sds.ifsd_m[cidx];
2111 sd_cl = fl->ifl_sds.ifsd_cl[cidx];
2113 MPASS(sd_cl != NULL);
2114 MPASS(sd_m != NULL);
2116 /* Don't include zero-length frags */
2117 if (ri->iri_frags[i].irf_len == 0) {
2118 /* XXX we can save the cluster here, but not the mbuf */
2119 m_init(sd_m, M_NOWAIT, MT_DATA, 0);
2121 fl->ifl_sds.ifsd_m[cidx] = NULL;
2126 flags = M_PKTHDR|M_EXT;
2128 padlen = ri->iri_pad;
2133 /* assuming padding is only on the first fragment */
2136 fl->ifl_sds.ifsd_m[cidx] = NULL;
2137 cl = fl->ifl_sds.ifsd_cl[cidx];
2138 fl->ifl_sds.ifsd_cl[cidx] = NULL;
2140 /* Can these two be made one ? */
2141 m_init(m, M_NOWAIT, MT_DATA, flags);
2142 m_cljset(m, cl, cltype);
2144 * These must follow m_init and m_cljset
2146 m->m_data += padlen;
2147 ri->iri_len -= padlen;
2148 m->m_len = ri->iri_frags[i].irf_len;
2149 } while (++i < ri->iri_nfrags);
2155 * Process one software descriptor
2157 static struct mbuf *
2158 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2165 /* should I merge this back in now that the two paths are basically duplicated? */
2166 if (ri->iri_nfrags == 1 &&
2167 ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2168 rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE, &fl, &cidx);
2169 m = fl->ifl_sds.ifsd_m[cidx];
2170 fl->ifl_sds.ifsd_m[cidx] = NULL;
2171 sd_cl = fl->ifl_sds.ifsd_cl[cidx];
2172 m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2173 memcpy(m->m_data, sd_cl, ri->iri_len);
2174 m->m_len = ri->iri_frags[0].irf_len;
2176 m = assemble_segments(rxq, ri);
2178 m->m_pkthdr.len = ri->iri_len;
2179 m->m_pkthdr.rcvif = ri->iri_ifp;
2180 m->m_flags |= ri->iri_flags;
2181 m->m_pkthdr.ether_vtag = ri->iri_vtag;
2182 m->m_pkthdr.flowid = ri->iri_flowid;
2183 M_HASHTYPE_SET(m, ri->iri_rsstype);
2184 m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2185 m->m_pkthdr.csum_data = ri->iri_csum_data;
2190 iflib_rxeof(iflib_rxq_t rxq, int budget)
2192 if_ctx_t ctx = rxq->ifr_ctx;
2193 if_shared_ctx_t sctx = ctx->ifc_sctx;
2194 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2197 struct if_rxd_info ri;
2198 int err, budget_left, rx_bytes, rx_pkts;
2203 * XXX early demux data packets so that if_input processing only handles
2204 * acks in interrupt context
2206 struct mbuf *m, *mh, *mt;
2208 if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) {
2214 rx_pkts = rx_bytes = 0;
2215 if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2216 cidxp = &rxq->ifr_cq_cidx;
2218 cidxp = &rxq->ifr_fl[0].ifl_cidx;
2219 if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2220 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2221 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2222 DBG_COUNTER_INC(rx_unavail);
2226 for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2227 if (__predict_false(!CTX_ACTIVE(ctx))) {
2228 DBG_COUNTER_INC(rx_ctx_inactive);
2232 * Reset client set fields to their default values
2234 bzero(&ri, sizeof(ri));
2235 ri.iri_qsidx = rxq->ifr_id;
2236 ri.iri_cidx = *cidxp;
2237 ri.iri_ifp = ctx->ifc_ifp;
2238 ri.iri_frags = rxq->ifr_frags;
2239 err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2241 /* in lieu of handling correctly - make sure it isn't being unhandled */
2243 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2244 *cidxp = ri.iri_cidx;
2245 /* Update our consumer index */
2246 while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2247 rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2248 rxq->ifr_cq_gen = 0;
2250 /* was this only a completion queue message? */
2251 if (__predict_false(ri.iri_nfrags == 0))
2254 MPASS(ri.iri_nfrags != 0);
2255 MPASS(ri.iri_len != 0);
2257 /* will advance the cidx on the corresponding free lists */
2258 m = iflib_rxd_pkt_get(rxq, &ri);
2259 if (avail == 0 && budget_left)
2260 avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2262 if (__predict_false(m == NULL)) {
2263 DBG_COUNTER_INC(rx_mbuf_null);
2266 /* imm_pkt: -- cxgb */
2274 /* make sure that we can refill faster than drain */
2275 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2276 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2279 lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2280 while (mh != NULL) {
2283 m->m_nextpkt = NULL;
2284 rx_bytes += m->m_pkthdr.len;
2286 #if defined(INET6) || defined(INET)
2287 if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2290 DBG_COUNTER_INC(rx_if_input);
2291 ifp->if_input(ifp, m);
2294 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2295 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2298 * Flush any outstanding LRO work
2300 #if defined(INET6) || defined(INET)
2301 tcp_lro_flush_all(&rxq->ifr_lc);
2305 return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2308 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2309 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2310 #define TXQ_MAX_DB_DEFERRED(size) (size >> 5)
2311 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2313 static __inline void
2314 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring)
2318 if (ring || txq->ift_db_pending >=
2319 TXQ_MAX_DB_DEFERRED(txq->ift_size)) {
2321 /* the lock will only ever be contended in the !min_latency case */
2322 if (!TXDB_TRYLOCK(txq))
2324 dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2325 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2326 txq->ift_db_pending = txq->ift_npending = 0;
2332 iflib_txd_deferred_db_check(void * arg)
2334 iflib_txq_t txq = arg;
2336 /* simple non-zero boolean so use bitwise OR */
2337 if ((txq->ift_db_pending | txq->ift_npending) &&
2338 txq->ift_db_pending >= txq->ift_db_pending_queued)
2339 iflib_txd_db_check(txq->ift_ctx, txq, TRUE);
2340 txq->ift_db_pending_queued = 0;
2341 if (ifmp_ring_is_stalled(txq->ift_br[0]))
2342 iflib_txq_check_drain(txq, 4);
2347 print_pkt(if_pkt_info_t pi)
2349 printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2350 pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2351 printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2352 pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2353 printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2354 pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2358 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2359 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2362 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2364 if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2365 struct ether_vlan_header *eh;
2369 if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2370 M_WRITABLE(m) == 0) {
2371 if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2380 * Determine where frame payload starts.
2381 * Jump over vlan headers if already present,
2382 * helpful for QinQ too.
2384 if (__predict_false(m->m_len < sizeof(*eh))) {
2386 if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2389 eh = mtod(m, struct ether_vlan_header *);
2390 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2391 pi->ipi_etype = ntohs(eh->evl_proto);
2392 pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2394 pi->ipi_etype = ntohs(eh->evl_encap_proto);
2395 pi->ipi_ehdrlen = ETHER_HDR_LEN;
2398 switch (pi->ipi_etype) {
2402 struct ip *ip = NULL;
2403 struct tcphdr *th = NULL;
2406 minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2407 if (__predict_false(m->m_len < minthlen)) {
2409 * if this code bloat is causing too much of a hit
2410 * move it to a separate function and mark it noinline
2412 if (m->m_len == pi->ipi_ehdrlen) {
2415 if (n->m_len >= sizeof(*ip)) {
2416 ip = (struct ip *)n->m_data;
2417 if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2418 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2421 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2423 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2427 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2429 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2430 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2431 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2434 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2435 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2436 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2438 pi->ipi_ip_hlen = ip->ip_hl << 2;
2439 pi->ipi_ipproto = ip->ip_p;
2440 pi->ipi_flags |= IPI_TX_IPV4;
2442 if (pi->ipi_csum_flags & CSUM_IP)
2445 if (pi->ipi_ipproto == IPPROTO_TCP) {
2446 if (__predict_false(th == NULL)) {
2448 if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2450 th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2452 pi->ipi_tcp_hflags = th->th_flags;
2453 pi->ipi_tcp_hlen = th->th_off << 2;
2454 pi->ipi_tcp_seq = th->th_seq;
2457 if (__predict_false(ip->ip_p != IPPROTO_TCP))
2459 th->th_sum = in_pseudo(ip->ip_src.s_addr,
2460 ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2461 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2462 if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2464 ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2471 case ETHERTYPE_IPV6:
2473 struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2475 pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2477 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2478 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2481 th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2483 /* XXX-BZ this will go badly in case of ext hdrs. */
2484 pi->ipi_ipproto = ip6->ip6_nxt;
2485 pi->ipi_flags |= IPI_TX_IPV6;
2487 if (pi->ipi_ipproto == IPPROTO_TCP) {
2488 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2489 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2492 pi->ipi_tcp_hflags = th->th_flags;
2493 pi->ipi_tcp_hlen = th->th_off << 2;
2497 if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2500 * The corresponding flag is set by the stack in the IPv4
2501 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2502 * So, set it here because the rest of the flow requires it.
2504 pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2505 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2506 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2512 pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2513 pi->ipi_ip_hlen = 0;
2521 static __noinline struct mbuf *
2522 collapse_pkthdr(struct mbuf *m0)
2524 struct mbuf *m, *m_next, *tmp;
2528 while (m_next != NULL && m_next->m_len == 0) {
2532 m_next = m_next->m_next;
2536 if ((m_next->m_flags & M_EXT) == 0) {
2537 m = m_defrag(m, M_NOWAIT);
2539 tmp = m_next->m_next;
2540 memcpy(m_next, m, MPKTHSIZE);
2548 * If dodgy hardware rejects the scatter gather chain we've handed it
2549 * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2552 static __noinline struct mbuf *
2553 iflib_remove_mbuf(iflib_txq_t txq)
2556 struct mbuf *m, *mh, **ifsd_m;
2558 pidx = txq->ift_pidx;
2559 ifsd_m = txq->ift_sds.ifsd_m;
2560 ntxd = txq->ift_size;
2561 mh = m = ifsd_m[pidx];
2562 ifsd_m[pidx] = NULL;
2564 txq->ift_dequeued++;
2569 ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2571 txq->ift_dequeued++;
2580 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2581 struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2582 int max_segs, int flags)
2585 if_shared_ctx_t sctx;
2586 if_softc_ctx_t scctx;
2587 int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2588 struct mbuf *m, *tmp, **ifsd_m, **mp;
2593 * Please don't ever do this
2595 if (__predict_false(m->m_len == 0))
2596 *m0 = m = collapse_pkthdr(m);
2599 sctx = ctx->ifc_sctx;
2600 scctx = &ctx->ifc_softc_ctx;
2601 ifsd_m = txq->ift_sds.ifsd_m;
2602 ntxd = txq->ift_size;
2603 pidx = txq->ift_pidx;
2605 uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2607 err = bus_dmamap_load_mbuf_sg(tag, map,
2608 *m0, segs, nsegs, BUS_DMA_NOWAIT);
2611 ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2614 mask = (txq->ift_size-1);
2620 if (__predict_false((*mp)->m_len == 0)) {
2624 next = (pidx + i) & (ntxd-1);
2625 } while (m != NULL);
2627 int buflen, sgsize, max_sgsize;
2632 maxsegsz = sctx->isc_tx_maxsize;
2635 if (__predict_false(m->m_len <= 0)) {
2643 vaddr = (vm_offset_t)m->m_data;
2645 * see if we can't be smarter about physically
2646 * contiguous mappings
2648 next = (pidx + count) & (ntxd-1);
2649 MPASS(ifsd_m[next] == NULL);
2651 txq->ift_enqueued++;
2654 while (buflen > 0) {
2655 max_sgsize = MIN(buflen, maxsegsz);
2656 curaddr = pmap_kextract(vaddr);
2657 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2658 sgsize = MIN(sgsize, max_sgsize);
2659 segs[i].ds_addr = curaddr;
2660 segs[i].ds_len = sgsize;
2670 } while (m != NULL);
2675 *m0 = iflib_remove_mbuf(txq);
2680 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2683 if_shared_ctx_t sctx;
2684 if_softc_ctx_t scctx;
2685 bus_dma_segment_t *segs;
2686 struct mbuf *m_head;
2688 struct if_pkt_info pi;
2690 int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
2691 bus_dma_tag_t desc_tag;
2693 segs = txq->ift_segs;
2695 sctx = ctx->ifc_sctx;
2696 scctx = &ctx->ifc_softc_ctx;
2697 segs = txq->ift_segs;
2698 ntxd = txq->ift_size;
2703 * If we're doing TSO the next descriptor to clean may be quite far ahead
2705 cidx = txq->ift_cidx;
2706 pidx = txq->ift_pidx;
2707 next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
2709 /* prefetch the next cache line of mbuf pointers and flags */
2710 prefetch(&txq->ift_sds.ifsd_m[next]);
2711 if (txq->ift_sds.ifsd_map != NULL) {
2712 prefetch(&txq->ift_sds.ifsd_map[next]);
2713 map = txq->ift_sds.ifsd_map[pidx];
2714 next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
2715 prefetch(&txq->ift_sds.ifsd_flags[next]);
2719 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
2720 desc_tag = txq->ift_tso_desc_tag;
2721 max_segs = scctx->isc_tx_tso_segments_max;
2723 desc_tag = txq->ift_desc_tag;
2724 max_segs = scctx->isc_tx_nsegments;
2727 bzero(&pi, sizeof(pi));
2728 pi.ipi_len = m_head->m_pkthdr.len;
2729 pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
2730 pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
2731 pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
2733 pi.ipi_qsidx = txq->ift_id;
2735 /* deliberate bitwise OR to make one condition */
2736 if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
2737 if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
2743 err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
2745 if (__predict_false(err)) {
2748 /* try collapse once and defrag once */
2750 m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
2752 m_head = m_defrag(*m_headp, M_NOWAIT);
2754 if (__predict_false(m_head == NULL))
2756 txq->ift_mbuf_defrag++;
2761 txq->ift_no_tx_dma_setup++;
2764 txq->ift_no_tx_dma_setup++;
2766 DBG_COUNTER_INC(tx_frees);
2770 txq->ift_map_failed++;
2771 DBG_COUNTER_INC(encap_load_mbuf_fail);
2776 * XXX assumes a 1 to 1 relationship between segments and
2777 * descriptors - this does not hold true on all drivers, e.g.
2780 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
2781 txq->ift_no_desc_avail++;
2783 bus_dmamap_unload(desc_tag, map);
2784 DBG_COUNTER_INC(encap_txq_avail_fail);
2785 if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
2786 GROUPTASK_ENQUEUE(&txq->ift_task);
2790 pi.ipi_nsegs = nsegs;
2792 MPASS(pidx >= 0 && pidx < txq->ift_size);
2796 if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
2797 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2798 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2800 DBG_COUNTER_INC(tx_encap);
2801 MPASS(pi.ipi_new_pidx >= 0 &&
2802 pi.ipi_new_pidx < txq->ift_size);
2804 ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
2805 if (pi.ipi_new_pidx < pi.ipi_pidx) {
2806 ndesc += txq->ift_size;
2810 * drivers can need as many as
2813 MPASS(ndesc <= pi.ipi_nsegs + 2);
2814 MPASS(pi.ipi_new_pidx != pidx);
2816 txq->ift_in_use += ndesc;
2818 * We update the last software descriptor again here because there may
2819 * be a sentinel and/or there may be more mbufs than segments
2821 txq->ift_pidx = pi.ipi_new_pidx;
2822 txq->ift_npending += pi.ipi_ndescs;
2823 } else if (__predict_false(err == EFBIG && remap < 2)) {
2824 *m_headp = m_head = iflib_remove_mbuf(txq);
2826 txq->ift_txd_encap_efbig++;
2829 DBG_COUNTER_INC(encap_txd_encap_fail);
2833 txq->ift_mbuf_defrag_failed++;
2834 txq->ift_map_failed++;
2836 DBG_COUNTER_INC(tx_frees);
2841 /* forward compatibility for cxgb */
2842 #define FIRST_QSET(ctx) 0
2844 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2845 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2846 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2847 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2848 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2849 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2853 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring
2856 * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic
2858 #define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2)
2861 iflib_txq_min_occupancy(iflib_txq_t txq)
2866 return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx,
2867 txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) +
2872 iflib_tx_desc_free(iflib_txq_t txq, int n)
2875 uint32_t qsize, cidx, mask, gen;
2876 struct mbuf *m, **ifsd_m;
2877 uint8_t *ifsd_flags;
2878 bus_dmamap_t *ifsd_map;
2880 cidx = txq->ift_cidx;
2882 qsize = txq->ift_size;
2884 hasmap = txq->ift_sds.ifsd_map != NULL;
2885 ifsd_flags = txq->ift_sds.ifsd_flags;
2886 ifsd_m = txq->ift_sds.ifsd_m;
2887 ifsd_map = txq->ift_sds.ifsd_map;
2890 prefetch(ifsd_m[(cidx + 3) & mask]);
2891 prefetch(ifsd_m[(cidx + 4) & mask]);
2893 if (ifsd_m[cidx] != NULL) {
2894 prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
2895 prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
2896 if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
2898 * does it matter if it's not the TSO tag? If so we'll
2899 * have to add the type to flags
2901 bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
2902 ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
2904 if ((m = ifsd_m[cidx]) != NULL) {
2905 /* XXX we don't support any drivers that batch packets yet */
2906 MPASS(m->m_nextpkt == NULL);
2909 ifsd_m[cidx] = NULL;
2911 txq->ift_dequeued++;
2913 DBG_COUNTER_INC(tx_frees);
2916 if (__predict_false(++cidx == qsize)) {
2921 txq->ift_cidx = cidx;
2926 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
2929 if_ctx_t ctx = txq->ift_ctx;
2931 KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
2932 MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
2935 * Need a rate-limiting check so that this isn't called every time
2937 iflib_tx_credits_update(ctx, txq);
2938 reclaim = DESC_RECLAIMABLE(txq);
2940 if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
2942 if (iflib_verbose_debug) {
2943 printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
2944 txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
2951 iflib_tx_desc_free(txq, reclaim);
2952 txq->ift_cleaned += reclaim;
2953 txq->ift_in_use -= reclaim;
2955 if (txq->ift_active == FALSE)
2956 txq->ift_active = TRUE;
2961 static struct mbuf **
2962 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset)
2965 return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)]));
2969 iflib_txq_check_drain(iflib_txq_t txq, int budget)
2972 ifmp_ring_check_drainage(txq->ift_br[0], budget);
2976 iflib_txq_can_drain(struct ifmp_ring *r)
2978 iflib_txq_t txq = r->cookie;
2979 if_ctx_t ctx = txq->ift_ctx;
2981 return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
2982 ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false));
2986 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2988 iflib_txq_t txq = r->cookie;
2989 if_ctx_t ctx = txq->ift_ctx;
2990 if_t ifp = ctx->ifc_ifp;
2991 struct mbuf **mp, *m;
2992 int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used;
2994 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
2995 !LINK_ACTIVE(ctx))) {
2996 DBG_COUNTER_INC(txq_drain_notready);
3000 avail = IDXDIFF(pidx, cidx, r->size);
3001 if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3002 DBG_COUNTER_INC(txq_drain_flushing);
3003 for (i = 0; i < avail; i++) {
3004 m_free(r->items[(cidx + i) & (r->size-1)]);
3005 r->items[(cidx + i) & (r->size-1)] = NULL;
3009 iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3010 if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3011 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3013 callout_stop(&txq->ift_timer);
3014 callout_stop(&txq->ift_db_check);
3015 CALLOUT_UNLOCK(txq);
3016 DBG_COUNTER_INC(txq_drain_oactive);
3019 consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3020 count = MIN(avail, TX_BATCH_SIZE);
3022 if (iflib_verbose_debug)
3023 printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3024 avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3027 for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) {
3028 mp = _ring_peek_one(r, cidx, i);
3029 MPASS(mp != NULL && *mp != NULL);
3030 in_use_prev = txq->ift_in_use;
3031 if ((err = iflib_encap(txq, mp)) == ENOBUFS) {
3032 DBG_COUNTER_INC(txq_drain_encapfail);
3033 /* no room - bail out */
3038 DBG_COUNTER_INC(txq_drain_encapfail);
3039 /* we can't send this packet - skip it */
3044 DBG_COUNTER_INC(tx_sent);
3045 bytes_sent += m->m_pkthdr.len;
3046 if (m->m_flags & M_MCAST)
3049 txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3050 desc_used += (txq->ift_in_use - in_use_prev);
3051 iflib_txd_db_check(ctx, txq, FALSE);
3052 ETHER_BPF_MTAP(ifp, m);
3053 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3056 if (desc_used >= TXQ_MAX_DB_CONSUMED(txq->ift_size))
3060 if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending)
3061 iflib_txd_db_check(ctx, txq, TRUE);
3062 else if ((txq->ift_db_pending || TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2) &&
3063 (callout_pending(&txq->ift_db_check) == 0)) {
3064 txq->ift_db_pending_queued = txq->ift_db_pending;
3065 callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check,
3066 txq, txq->ift_db_check.c_cpu);
3068 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3069 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3071 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3073 if (iflib_verbose_debug)
3074 printf("consumed=%d\n", consumed);
3080 iflib_txq_drain_always(struct ifmp_ring *r)
3086 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3094 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3096 callout_stop(&txq->ift_timer);
3097 callout_stop(&txq->ift_db_check);
3098 CALLOUT_UNLOCK(txq);
3100 avail = IDXDIFF(pidx, cidx, r->size);
3101 for (i = 0; i < avail; i++) {
3102 mp = _ring_peek_one(r, cidx, i);
3105 MPASS(ifmp_ring_is_stalled(r) == 0);
3110 iflib_ifmp_purge(iflib_txq_t txq)
3112 struct ifmp_ring *r;
3115 r->drain = iflib_txq_drain_free;
3116 r->can_drain = iflib_txq_drain_always;
3118 ifmp_ring_check_drainage(r, r->size);
3120 r->drain = iflib_txq_drain;
3121 r->can_drain = iflib_txq_can_drain;
3125 _task_fn_tx(void *context)
3127 iflib_txq_t txq = context;
3128 if_ctx_t ctx = txq->ift_ctx;
3130 #ifdef IFLIB_DIAGNOSTICS
3131 txq->ift_cpu_exec_count[curcpu]++;
3133 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3135 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3139 _task_fn_rx(void *context)
3141 iflib_rxq_t rxq = context;
3142 if_ctx_t ctx = rxq->ifr_ctx;
3146 #ifdef IFLIB_DIAGNOSTICS
3147 rxq->ifr_cpu_exec_count[curcpu]++;
3149 DBG_COUNTER_INC(task_fn_rxs);
3150 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3153 if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3154 if (ctx->ifc_flags & IFC_LEGACY)
3155 IFDI_INTR_ENABLE(ctx);
3157 DBG_COUNTER_INC(rx_intr_enables);
3158 rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3159 KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3162 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3165 GROUPTASK_ENQUEUE(&rxq->ifr_task);
3169 _task_fn_admin(void *context)
3171 if_ctx_t ctx = context;
3172 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3176 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3180 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3182 callout_stop(&txq->ift_timer);
3183 CALLOUT_UNLOCK(txq);
3185 IFDI_UPDATE_ADMIN_STATUS(ctx);
3186 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3187 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3188 IFDI_LINK_INTR_ENABLE(ctx);
3191 if (LINK_ACTIVE(ctx) == 0)
3193 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3194 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3199 _task_fn_iov(void *context)
3201 if_ctx_t ctx = context;
3203 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3207 IFDI_VFLR_HANDLE(ctx);
3212 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3215 if_int_delay_info_t info;
3218 info = (if_int_delay_info_t)arg1;
3219 ctx = info->iidi_ctx;
3220 info->iidi_req = req;
3221 info->iidi_oidp = oidp;
3223 err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3228 /*********************************************************************
3232 **********************************************************************/
3235 iflib_if_init_locked(if_ctx_t ctx)
3238 iflib_init_locked(ctx);
3243 iflib_if_init(void *arg)
3248 iflib_if_init_locked(ctx);
3253 iflib_if_transmit(if_t ifp, struct mbuf *m)
3255 if_ctx_t ctx = if_getsoftc(ifp);
3260 if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3261 DBG_COUNTER_INC(tx_frees);
3266 MPASS(m->m_nextpkt == NULL);
3268 if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3269 qidx = QIDX(ctx, m);
3271 * XXX calculate buf_ring based on flowid (divvy up bits?)
3273 txq = &ctx->ifc_txqs[qidx];
3275 #ifdef DRIVER_BACKPRESSURE
3276 if (txq->ift_closed) {
3278 next = m->m_nextpkt;
3279 m->m_nextpkt = NULL;
3292 next = next->m_nextpkt;
3293 } while (next != NULL);
3295 if (count > nitems(marr))
3296 if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3297 /* XXX check nextpkt */
3299 /* XXX simplify for now */
3300 DBG_COUNTER_INC(tx_frees);
3303 for (next = m, i = 0; next != NULL; i++) {
3305 next = next->m_nextpkt;
3306 mp[i]->m_nextpkt = NULL;
3309 DBG_COUNTER_INC(tx_seen);
3310 err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE);
3313 GROUPTASK_ENQUEUE(&txq->ift_task);
3314 /* support forthcoming later */
3315 #ifdef DRIVER_BACKPRESSURE
3316 txq->ift_closed = TRUE;
3318 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3320 } else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3321 GROUPTASK_ENQUEUE(&txq->ift_task);
3328 iflib_if_qflush(if_t ifp)
3330 if_ctx_t ctx = if_getsoftc(ifp);
3331 iflib_txq_t txq = ctx->ifc_txqs;
3335 ctx->ifc_flags |= IFC_QFLUSH;
3337 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3338 while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0])))
3339 iflib_txq_check_drain(txq, 0);
3341 ctx->ifc_flags &= ~IFC_QFLUSH;
3348 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3349 IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | \
3350 IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3353 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3355 if_ctx_t ctx = if_getsoftc(ifp);
3356 struct ifreq *ifr = (struct ifreq *)data;
3357 #if defined(INET) || defined(INET6)
3358 struct ifaddr *ifa = (struct ifaddr *)data;
3360 bool avoid_reset = FALSE;
3361 int err = 0, reinit = 0, bits;
3366 if (ifa->ifa_addr->sa_family == AF_INET)
3370 if (ifa->ifa_addr->sa_family == AF_INET6)
3374 ** Calling init results in link renegotiation,
3375 ** so we avoid doing it when possible.
3378 if_setflagbits(ifp, IFF_UP,0);
3379 if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3382 if (!(if_getflags(ifp) & IFF_NOARP))
3383 arp_ifinit(ifp, ifa);
3386 err = ether_ioctl(ifp, command, data);
3390 if (ifr->ifr_mtu == if_getmtu(ifp)) {
3394 bits = if_getdrvflags(ifp);
3395 /* stop the driver and free any clusters before proceeding */
3398 if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3399 if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3400 ctx->ifc_flags |= IFC_MULTISEG;
3402 ctx->ifc_flags &= ~IFC_MULTISEG;
3403 err = if_setmtu(ifp, ifr->ifr_mtu);
3405 iflib_init_locked(ctx);
3406 if_setdrvflags(ifp, bits);
3411 if (if_getflags(ifp) & IFF_UP) {
3412 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3413 if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3414 (IFF_PROMISC | IFF_ALLMULTI)) {
3415 err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3419 } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3422 ctx->ifc_if_flags = if_getflags(ifp);
3429 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3431 IFDI_INTR_DISABLE(ctx);
3432 IFDI_MULTI_SET(ctx);
3433 IFDI_INTR_ENABLE(ctx);
3439 IFDI_MEDIA_SET(ctx);
3443 err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3447 struct ifi2creq i2c;
3449 err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3452 if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3456 if (i2c.len > sizeof(i2c.data)) {
3461 if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3462 err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3469 mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3472 setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3474 setmask |= (mask & IFCAP_FLAGS);
3476 if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3477 setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3478 if ((mask & IFCAP_WOL) &&
3479 (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3480 setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3483 * want to ensure that traffic has stopped before we change any of the flags
3487 bits = if_getdrvflags(ifp);
3488 if (bits & IFF_DRV_RUNNING)
3490 if_togglecapenable(ifp, setmask);
3491 if (bits & IFF_DRV_RUNNING)
3492 iflib_init_locked(ctx);
3493 if_setdrvflags(ifp, bits);
3498 case SIOCGPRIVATE_0:
3502 err = IFDI_PRIV_IOCTL(ctx, command, data);
3506 err = ether_ioctl(ifp, command, data);
3515 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3517 if_ctx_t ctx = if_getsoftc(ifp);
3519 return (IFDI_GET_COUNTER(ctx, cnt));
3522 /*********************************************************************
3524 * OTHER FUNCTIONS EXPORTED TO THE STACK
3526 **********************************************************************/
3529 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3531 if_ctx_t ctx = if_getsoftc(ifp);
3533 if ((void *)ctx != arg)
3536 if ((vtag == 0) || (vtag > 4095))
3540 IFDI_VLAN_REGISTER(ctx, vtag);
3541 /* Re-init to load the changes */
3542 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3543 iflib_init_locked(ctx);
3548 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3550 if_ctx_t ctx = if_getsoftc(ifp);
3552 if ((void *)ctx != arg)
3555 if ((vtag == 0) || (vtag > 4095))
3559 IFDI_VLAN_UNREGISTER(ctx, vtag);
3560 /* Re-init to load the changes */
3561 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3562 iflib_init_locked(ctx);
3567 iflib_led_func(void *arg, int onoff)
3572 IFDI_LED_FUNC(ctx, onoff);
3576 /*********************************************************************
3578 * BUS FUNCTION DEFINITIONS
3580 **********************************************************************/
3583 iflib_device_probe(device_t dev)
3585 pci_vendor_info_t *ent;
3587 uint16_t pci_vendor_id, pci_device_id;
3588 uint16_t pci_subvendor_id, pci_subdevice_id;
3589 uint16_t pci_rev_id;
3590 if_shared_ctx_t sctx;
3592 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3595 pci_vendor_id = pci_get_vendor(dev);
3596 pci_device_id = pci_get_device(dev);
3597 pci_subvendor_id = pci_get_subvendor(dev);
3598 pci_subdevice_id = pci_get_subdevice(dev);
3599 pci_rev_id = pci_get_revid(dev);
3600 if (sctx->isc_parse_devinfo != NULL)
3601 sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3603 ent = sctx->isc_vendor_info;
3604 while (ent->pvi_vendor_id != 0) {
3605 if (pci_vendor_id != ent->pvi_vendor_id) {
3609 if ((pci_device_id == ent->pvi_device_id) &&
3610 ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3611 (ent->pvi_subvendor_id == 0)) &&
3612 ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3613 (ent->pvi_subdevice_id == 0)) &&
3614 ((pci_rev_id == ent->pvi_rev_id) ||
3615 (ent->pvi_rev_id == 0))) {
3617 device_set_desc_copy(dev, ent->pvi_name);
3618 /* this needs to be changed to zero if the bus probing code
3619 * ever stops re-probing on best match because the sctx
3620 * may have its values over written by register calls
3621 * in subsequent probes
3623 return (BUS_PROBE_DEFAULT);
3631 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3633 int err, rid, msix, msix_bar;
3636 if_softc_ctx_t scctx;
3642 ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3645 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3646 device_set_softc(dev, ctx);
3647 ctx->ifc_flags |= IFC_SC_ALLOCATED;
3650 ctx->ifc_sctx = sctx;
3652 ctx->ifc_softc = sc;
3654 if ((err = iflib_register(ctx)) != 0) {
3655 device_printf(dev, "iflib_register failed %d\n", err);
3658 iflib_add_device_sysctl_pre(ctx);
3660 scctx = &ctx->ifc_softc_ctx;
3664 * XXX sanity check that ntxd & nrxd are a power of 2
3666 if (ctx->ifc_sysctl_ntxqs != 0)
3667 scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
3668 if (ctx->ifc_sysctl_nrxqs != 0)
3669 scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
3671 for (i = 0; i < sctx->isc_ntxqs; i++) {
3672 if (ctx->ifc_sysctl_ntxds[i] != 0)
3673 scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
3675 scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
3678 for (i = 0; i < sctx->isc_nrxqs; i++) {
3679 if (ctx->ifc_sysctl_nrxds[i] != 0)
3680 scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
3682 scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
3685 for (i = 0; i < sctx->isc_nrxqs; i++) {
3686 if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
3687 device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
3688 i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
3689 scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
3691 if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
3692 device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
3693 i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
3694 scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
3698 for (i = 0; i < sctx->isc_ntxqs; i++) {
3699 if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
3700 device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
3701 i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
3702 scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
3704 if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
3705 device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
3706 i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
3707 scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
3711 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
3712 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
3715 _iflib_pre_assert(scctx);
3716 ctx->ifc_txrx = *scctx->isc_txrx;
3719 MPASS(scctx->isc_capenable);
3720 if (scctx->isc_capenable & IFCAP_TXCSUM)
3721 MPASS(scctx->isc_tx_csum_flags);
3724 if_setcapabilities(ifp, scctx->isc_capenable);
3725 if_setcapenable(ifp, scctx->isc_capenable);
3727 if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
3728 scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
3729 if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
3730 scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
3733 if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
3734 ctx->ifc_flags |= IFC_DMAR;
3737 msix_bar = scctx->isc_msix_bar;
3739 if(sctx->isc_flags & IFLIB_HAS_TXCQ)
3744 if(sctx->isc_flags & IFLIB_HAS_RXCQ)
3749 /* XXX change for per-queue sizes */
3750 device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
3751 scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
3752 for (i = 0; i < sctx->isc_nrxqs; i++) {
3753 if (!powerof2(scctx->isc_nrxd[i])) {
3754 /* round down instead? */
3755 device_printf(dev, "# rx descriptors must be a power of 2\n");
3760 for (i = 0; i < sctx->isc_ntxqs; i++) {
3761 if (!powerof2(scctx->isc_ntxd[i])) {
3763 "# tx descriptors must be a power of 2");
3769 if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
3770 MAX_SINGLE_PACKET_FRACTION)
3771 scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
3772 MAX_SINGLE_PACKET_FRACTION);
3773 if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
3774 MAX_SINGLE_PACKET_FRACTION)
3775 scctx->isc_tx_tso_segments_max = max(1,
3776 scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
3779 * Protect the stack against modern hardware
3781 if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
3782 scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
3784 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
3785 ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
3786 ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
3787 ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
3788 if (scctx->isc_rss_table_size == 0)
3789 scctx->isc_rss_table_size = 64;
3790 scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
3792 GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
3793 /* XXX format name */
3794 taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
3796 ** Now setup MSI or MSI/X, should
3797 ** return us the number of supported
3798 ** vectors. (Will be 1 for MSI)
3800 if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
3801 msix = scctx->isc_vectors;
3802 } else if (scctx->isc_msix_bar != 0)
3804 * The simple fact that isc_msix_bar is not 0 does not mean we
3805 * we have a good value there that is known to work.
3807 msix = iflib_msix_init(ctx);
3809 scctx->isc_vectors = 1;
3810 scctx->isc_ntxqsets = 1;
3811 scctx->isc_nrxqsets = 1;
3812 scctx->isc_intr = IFLIB_INTR_LEGACY;
3815 /* Get memory for the station queues */
3816 if ((err = iflib_queues_alloc(ctx))) {
3817 device_printf(dev, "Unable to allocate queue memory\n");
3821 if ((err = iflib_qset_structures_setup(ctx))) {
3822 device_printf(dev, "qset structure setup failed %d\n", err);
3827 * Group taskqueues aren't properly set up until SMP is started,
3828 * so we disable interrupts until we can handle them post
3831 * XXX: disabling interrupts doesn't actually work, at least for
3832 * the non-MSI case. When they occur before SI_SUB_SMP completes,
3833 * we do null handling and depend on this not causing too large an
3836 IFDI_INTR_DISABLE(ctx);
3837 if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
3838 device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
3839 goto fail_intr_free;
3843 if (scctx->isc_intr == IFLIB_INTR_MSI) {
3847 if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
3848 device_printf(dev, "iflib_legacy_setup failed %d\n", err);
3849 goto fail_intr_free;
3852 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
3853 if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
3854 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
3857 if ((err = iflib_netmap_attach(ctx))) {
3858 device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
3863 if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
3864 iflib_add_device_sysctl_post(ctx);
3865 ctx->ifc_flags |= IFC_INIT_DONE;
3868 ether_ifdetach(ctx->ifc_ifp);
3870 if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
3871 pci_release_msi(ctx->ifc_dev);
3873 /* XXX free queues */
3880 iflib_device_attach(device_t dev)
3883 if_shared_ctx_t sctx;
3885 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3888 pci_enable_busmaster(dev);
3890 return (iflib_device_register(dev, NULL, sctx, &ctx));
3894 iflib_device_deregister(if_ctx_t ctx)
3896 if_t ifp = ctx->ifc_ifp;
3899 device_t dev = ctx->ifc_dev;
3901 struct taskqgroup *tqg;
3903 /* Make sure VLANS are not using driver */
3904 if (if_vlantrunkinuse(ifp)) {
3905 device_printf(dev,"Vlan in use, detach first\n");
3910 ctx->ifc_in_detach = 1;
3914 /* Unregister VLAN events */
3915 if (ctx->ifc_vlan_attach_event != NULL)
3916 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
3917 if (ctx->ifc_vlan_detach_event != NULL)
3918 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
3920 iflib_netmap_detach(ifp);
3921 ether_ifdetach(ifp);
3922 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
3923 CTX_LOCK_DESTROY(ctx);
3924 if (ctx->ifc_led_dev != NULL)
3925 led_destroy(ctx->ifc_led_dev);
3926 /* XXX drain any dependent tasks */
3927 tqg = qgroup_if_io_tqg;
3928 for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
3929 callout_drain(&txq->ift_timer);
3930 callout_drain(&txq->ift_db_check);
3931 if (txq->ift_task.gt_uniq != NULL)
3932 taskqgroup_detach(tqg, &txq->ift_task);
3934 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
3935 if (rxq->ifr_task.gt_uniq != NULL)
3936 taskqgroup_detach(tqg, &rxq->ifr_task);
3938 tqg = qgroup_if_config_tqg;
3939 if (ctx->ifc_admin_task.gt_uniq != NULL)
3940 taskqgroup_detach(tqg, &ctx->ifc_admin_task);
3941 if (ctx->ifc_vflr_task.gt_uniq != NULL)
3942 taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
3945 device_set_softc(ctx->ifc_dev, NULL);
3946 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
3947 pci_release_msi(dev);
3949 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
3950 iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
3952 if (ctx->ifc_msix_mem != NULL) {
3953 bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
3954 ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
3955 ctx->ifc_msix_mem = NULL;
3958 bus_generic_detach(dev);
3961 iflib_tx_structures_free(ctx);
3962 iflib_rx_structures_free(ctx);
3963 if (ctx->ifc_flags & IFC_SC_ALLOCATED)
3964 free(ctx->ifc_softc, M_IFLIB);
3971 iflib_device_detach(device_t dev)
3973 if_ctx_t ctx = device_get_softc(dev);
3975 return (iflib_device_deregister(ctx));
3979 iflib_device_suspend(device_t dev)
3981 if_ctx_t ctx = device_get_softc(dev);
3987 return bus_generic_suspend(dev);
3990 iflib_device_shutdown(device_t dev)
3992 if_ctx_t ctx = device_get_softc(dev);
3998 return bus_generic_suspend(dev);
4003 iflib_device_resume(device_t dev)
4005 if_ctx_t ctx = device_get_softc(dev);
4006 iflib_txq_t txq = ctx->ifc_txqs;
4010 iflib_init_locked(ctx);
4012 for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4013 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4015 return (bus_generic_resume(dev));
4019 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4022 if_ctx_t ctx = device_get_softc(dev);
4025 error = IFDI_IOV_INIT(ctx, num_vfs, params);
4032 iflib_device_iov_uninit(device_t dev)
4034 if_ctx_t ctx = device_get_softc(dev);
4037 IFDI_IOV_UNINIT(ctx);
4042 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4045 if_ctx_t ctx = device_get_softc(dev);
4048 error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4054 /*********************************************************************
4056 * MODULE FUNCTION DEFINITIONS
4058 **********************************************************************/
4061 * - Start a fast taskqueue thread for each core
4062 * - Start a taskqueue for control operations
4065 iflib_module_init(void)
4071 iflib_module_event_handler(module_t mod, int what, void *arg)
4077 if ((err = iflib_module_init()) != 0)
4083 return (EOPNOTSUPP);
4089 /*********************************************************************
4091 * PUBLIC FUNCTION DEFINITIONS
4092 * ordered as in iflib.h
4094 **********************************************************************/
4098 _iflib_assert(if_shared_ctx_t sctx)
4100 MPASS(sctx->isc_tx_maxsize);
4101 MPASS(sctx->isc_tx_maxsegsize);
4103 MPASS(sctx->isc_rx_maxsize);
4104 MPASS(sctx->isc_rx_nsegments);
4105 MPASS(sctx->isc_rx_maxsegsize);
4107 MPASS(sctx->isc_nrxd_min[0]);
4108 MPASS(sctx->isc_nrxd_max[0]);
4109 MPASS(sctx->isc_nrxd_default[0]);
4110 MPASS(sctx->isc_ntxd_min[0]);
4111 MPASS(sctx->isc_ntxd_max[0]);
4112 MPASS(sctx->isc_ntxd_default[0]);
4116 _iflib_pre_assert(if_softc_ctx_t scctx)
4119 MPASS(scctx->isc_txrx->ift_txd_encap);
4120 MPASS(scctx->isc_txrx->ift_txd_flush);
4121 MPASS(scctx->isc_txrx->ift_txd_credits_update);
4122 MPASS(scctx->isc_txrx->ift_rxd_available);
4123 MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4124 MPASS(scctx->isc_txrx->ift_rxd_refill);
4125 MPASS(scctx->isc_txrx->ift_rxd_flush);
4129 iflib_register(if_ctx_t ctx)
4131 if_shared_ctx_t sctx = ctx->ifc_sctx;
4132 driver_t *driver = sctx->isc_driver;
4133 device_t dev = ctx->ifc_dev;
4136 _iflib_assert(sctx);
4138 CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4140 ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4142 device_printf(dev, "can not allocate ifnet structure\n");
4147 * Initialize our context's device specific methods
4149 kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4150 kobj_class_compile((kobj_class_t) driver);
4153 if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4154 if_setsoftc(ifp, ctx);
4155 if_setdev(ifp, dev);
4156 if_setinitfn(ifp, iflib_if_init);
4157 if_setioctlfn(ifp, iflib_if_ioctl);
4158 if_settransmitfn(ifp, iflib_if_transmit);
4159 if_setqflushfn(ifp, iflib_if_qflush);
4160 if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4162 ctx->ifc_vlan_attach_event =
4163 EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4164 EVENTHANDLER_PRI_FIRST);
4165 ctx->ifc_vlan_detach_event =
4166 EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4167 EVENTHANDLER_PRI_FIRST);
4169 ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4170 iflib_media_change, iflib_media_status);
4177 iflib_queues_alloc(if_ctx_t ctx)
4179 if_shared_ctx_t sctx = ctx->ifc_sctx;
4180 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4181 device_t dev = ctx->ifc_dev;
4182 int nrxqsets = scctx->isc_nrxqsets;
4183 int ntxqsets = scctx->isc_ntxqsets;
4186 iflib_fl_t fl = NULL;
4187 int i, j, cpu, err, txconf, rxconf;
4188 iflib_dma_info_t ifdip;
4189 uint32_t *rxqsizes = scctx->isc_rxqsizes;
4190 uint32_t *txqsizes = scctx->isc_txqsizes;
4191 uint8_t nrxqs = sctx->isc_nrxqs;
4192 uint8_t ntxqs = sctx->isc_ntxqs;
4193 int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4196 struct ifmp_ring **brscp;
4197 int nbuf_rings = 1; /* XXX determine dynamically */
4199 KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4200 KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4206 /* Allocate the TX ring struct memory */
4208 (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4209 ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4210 device_printf(dev, "Unable to allocate TX ring memory\n");
4215 /* Now allocate the RX */
4217 (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4218 nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4219 device_printf(dev, "Unable to allocate RX ring memory\n");
4223 if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4224 device_printf(dev, "Unable to buf_ring_sc * memory\n");
4229 ctx->ifc_txqs = txq;
4230 ctx->ifc_rxqs = rxq;
4233 * XXX handle allocation failure
4235 for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4236 /* Set up some basics */
4238 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4239 device_printf(dev, "failed to allocate iflib_dma_info\n");
4243 txq->ift_ifdi = ifdip;
4244 for (j = 0; j < ntxqs; j++, ifdip++) {
4245 if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4246 device_printf(dev, "Unable to allocate Descriptor memory\n");
4250 bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4254 if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4255 txq->ift_br_offset = 1;
4257 txq->ift_br_offset = 0;
4260 txq->ift_timer.c_cpu = cpu;
4261 txq->ift_db_check.c_cpu = cpu;
4262 txq->ift_nbr = nbuf_rings;
4264 if (iflib_txsd_alloc(txq)) {
4265 device_printf(dev, "Critical Failure setting up TX buffers\n");
4270 /* Initialize the TX lock */
4271 snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4272 device_get_nameunit(dev), txq->ift_id);
4273 mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4274 callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4275 callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0);
4277 snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4278 device_get_nameunit(dev), txq->ift_id);
4279 TXDB_LOCK_INIT(txq);
4281 txq->ift_br = brscp + i*nbuf_rings;
4282 for (j = 0; j < nbuf_rings; j++) {
4283 err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain,
4284 iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4286 /* XXX free any allocated rings */
4287 device_printf(dev, "Unable to allocate buf_ring\n");
4293 for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4294 /* Set up some basics */
4296 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4297 device_printf(dev, "failed to allocate iflib_dma_info\n");
4302 rxq->ifr_ifdi = ifdip;
4303 for (j = 0; j < nrxqs; j++, ifdip++) {
4304 if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4305 device_printf(dev, "Unable to allocate Descriptor memory\n");
4309 bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4313 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4314 rxq->ifr_fl_offset = 1;
4316 rxq->ifr_fl_offset = 0;
4318 rxq->ifr_nfl = nfree_lists;
4320 (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4321 device_printf(dev, "Unable to allocate free list memory\n");
4326 for (j = 0; j < nfree_lists; j++) {
4327 rxq->ifr_fl[j].ifl_rxq = rxq;
4328 rxq->ifr_fl[j].ifl_id = j;
4329 rxq->ifr_fl[j].ifl_ifdi =
4330 &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4332 /* Allocate receive buffers for the ring*/
4333 if (iflib_rxsd_alloc(rxq)) {
4335 "Critical Failure setting up receive buffers\n");
4342 vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4343 paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4344 for (i = 0; i < ntxqsets; i++) {
4345 iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4347 for (j = 0; j < ntxqs; j++, di++) {
4348 vaddrs[i*ntxqs + j] = di->idi_vaddr;
4349 paddrs[i*ntxqs + j] = di->idi_paddr;
4352 if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4353 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4354 iflib_tx_structures_free(ctx);
4355 free(vaddrs, M_IFLIB);
4356 free(paddrs, M_IFLIB);
4359 free(vaddrs, M_IFLIB);
4360 free(paddrs, M_IFLIB);
4363 vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4364 paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4365 for (i = 0; i < nrxqsets; i++) {
4366 iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4368 for (j = 0; j < nrxqs; j++, di++) {
4369 vaddrs[i*nrxqs + j] = di->idi_vaddr;
4370 paddrs[i*nrxqs + j] = di->idi_paddr;
4373 if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4374 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4375 iflib_tx_structures_free(ctx);
4376 free(vaddrs, M_IFLIB);
4377 free(paddrs, M_IFLIB);
4380 free(vaddrs, M_IFLIB);
4381 free(paddrs, M_IFLIB);
4385 /* XXX handle allocation failure changes */
4388 if (ctx->ifc_rxqs != NULL)
4389 free(ctx->ifc_rxqs, M_IFLIB);
4390 ctx->ifc_rxqs = NULL;
4391 if (ctx->ifc_txqs != NULL)
4392 free(ctx->ifc_txqs, M_IFLIB);
4393 ctx->ifc_txqs = NULL;
4396 free(brscp, M_IFLIB);
4406 iflib_tx_structures_setup(if_ctx_t ctx)
4408 iflib_txq_t txq = ctx->ifc_txqs;
4411 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4412 iflib_txq_setup(txq);
4418 iflib_tx_structures_free(if_ctx_t ctx)
4420 iflib_txq_t txq = ctx->ifc_txqs;
4423 for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4424 iflib_txq_destroy(txq);
4425 for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4426 iflib_dma_free(&txq->ift_ifdi[j]);
4428 free(ctx->ifc_txqs, M_IFLIB);
4429 ctx->ifc_txqs = NULL;
4430 IFDI_QUEUES_FREE(ctx);
4433 /*********************************************************************
4435 * Initialize all receive rings.
4437 **********************************************************************/
4439 iflib_rx_structures_setup(if_ctx_t ctx)
4441 iflib_rxq_t rxq = ctx->ifc_rxqs;
4443 #if defined(INET6) || defined(INET)
4447 for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4448 #if defined(INET6) || defined(INET)
4449 tcp_lro_free(&rxq->ifr_lc);
4450 if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4451 TCP_LRO_ENTRIES, min(1024,
4452 ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4453 device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4456 rxq->ifr_lro_enabled = TRUE;
4458 IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4461 #if defined(INET6) || defined(INET)
4464 * Free RX software descriptors allocated so far, we will only handle
4465 * the rings that completed, the failing case will have
4466 * cleaned up for itself. 'q' failed, so its the terminus.
4468 rxq = ctx->ifc_rxqs;
4469 for (i = 0; i < q; ++i, rxq++) {
4470 iflib_rx_sds_free(rxq);
4471 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4477 /*********************************************************************
4479 * Free all receive rings.
4481 **********************************************************************/
4483 iflib_rx_structures_free(if_ctx_t ctx)
4485 iflib_rxq_t rxq = ctx->ifc_rxqs;
4487 for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4488 iflib_rx_sds_free(rxq);
4493 iflib_qset_structures_setup(if_ctx_t ctx)
4497 if ((err = iflib_tx_structures_setup(ctx)) != 0)
4500 if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4501 device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4502 iflib_tx_structures_free(ctx);
4503 iflib_rx_structures_free(ctx);
4509 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4510 driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4513 return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4517 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4519 int i, cpuid, eqid, count;
4521 CPU_COPY(&ctx->ifc_cpus, cpus);
4522 count = CPU_COUNT(&ctx->ifc_cpus);
4524 /* clear up to the qid'th bit */
4525 for (i = 0; i < eqid; i++) {
4526 cpuid = CPU_FFS(cpus);
4528 CPU_CLR(cpuid-1, cpus);
4530 cpuid = CPU_FFS(cpus);
4536 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4537 iflib_intr_type_t type, driver_filter_t *filter,
4538 void *filter_arg, int qid, char *name)
4540 struct grouptask *gtask;
4541 struct taskqgroup *tqg;
4542 iflib_filter_info_t info;
4545 int tqrid, err, cpuid;
4548 info = &ctx->ifc_filter_info;
4552 /* XXX merge tx/rx for netmap? */
4554 q = &ctx->ifc_txqs[qid];
4555 info = &ctx->ifc_txqs[qid].ift_filter_info;
4556 gtask = &ctx->ifc_txqs[qid].ift_task;
4557 tqg = qgroup_if_io_tqg;
4559 GROUPTASK_INIT(gtask, 0, fn, q);
4562 q = &ctx->ifc_rxqs[qid];
4563 info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4564 gtask = &ctx->ifc_rxqs[qid].ifr_task;
4565 tqg = qgroup_if_io_tqg;
4567 GROUPTASK_INIT(gtask, 0, fn, q);
4569 case IFLIB_INTR_ADMIN:
4572 info = &ctx->ifc_filter_info;
4573 gtask = &ctx->ifc_admin_task;
4574 tqg = qgroup_if_config_tqg;
4575 fn = _task_fn_admin;
4578 panic("unknown net intr type");
4581 info->ifi_filter = filter;
4582 info->ifi_filter_arg = filter_arg;
4583 info->ifi_task = gtask;
4584 info->ifi_ctx = ctx;
4586 err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info, name);
4588 device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4591 if (type == IFLIB_INTR_ADMIN)
4595 cpuid = find_nth(ctx, &cpus, qid);
4596 taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4598 taskqgroup_attach(tqg, gtask, q, tqrid, name);
4605 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type, void *arg, int qid, char *name)
4607 struct grouptask *gtask;
4608 struct taskqgroup *tqg;
4614 q = &ctx->ifc_txqs[qid];
4615 gtask = &ctx->ifc_txqs[qid].ift_task;
4616 tqg = qgroup_if_io_tqg;
4620 q = &ctx->ifc_rxqs[qid];
4621 gtask = &ctx->ifc_rxqs[qid].ifr_task;
4622 tqg = qgroup_if_io_tqg;
4625 case IFLIB_INTR_IOV:
4627 gtask = &ctx->ifc_vflr_task;
4628 tqg = qgroup_if_config_tqg;
4633 panic("unknown net intr type");
4635 GROUPTASK_INIT(gtask, 0, fn, q);
4636 taskqgroup_attach(tqg, gtask, q, rid, name);
4640 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4643 bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4646 bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4650 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4652 iflib_txq_t txq = ctx->ifc_txqs;
4653 iflib_rxq_t rxq = ctx->ifc_rxqs;
4654 if_irq_t irq = &ctx->ifc_legacy_irq;
4655 iflib_filter_info_t info;
4656 struct grouptask *gtask;
4657 struct taskqgroup *tqg;
4663 q = &ctx->ifc_rxqs[0];
4664 info = &rxq[0].ifr_filter_info;
4665 gtask = &rxq[0].ifr_task;
4666 tqg = qgroup_if_io_tqg;
4667 tqrid = irq->ii_rid = *rid;
4670 ctx->ifc_flags |= IFC_LEGACY;
4671 info->ifi_filter = filter;
4672 info->ifi_filter_arg = filter_arg;
4673 info->ifi_task = gtask;
4674 info->ifi_ctx = ctx;
4676 /* We allocate a single interrupt resource */
4677 if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0)
4679 GROUPTASK_INIT(gtask, 0, fn, q);
4680 taskqgroup_attach(tqg, gtask, q, tqrid, name);
4682 GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
4683 taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
4688 iflib_led_create(if_ctx_t ctx)
4691 ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
4692 device_get_nameunit(ctx->ifc_dev));
4696 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
4699 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
4703 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
4706 GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
4710 iflib_admin_intr_deferred(if_ctx_t ctx)
4713 struct grouptask *gtask;
4715 gtask = &ctx->ifc_admin_task;
4716 MPASS(gtask->gt_taskqueue != NULL);
4719 GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
4723 iflib_iov_intr_deferred(if_ctx_t ctx)
4726 GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
4730 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
4733 taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
4737 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
4741 GROUPTASK_INIT(gtask, 0, fn, ctx);
4742 taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
4746 iflib_config_gtask_deinit(struct grouptask *gtask)
4749 taskqgroup_detach(qgroup_if_config_tqg, gtask);
4753 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
4755 if_t ifp = ctx->ifc_ifp;
4756 iflib_txq_t txq = ctx->ifc_txqs;
4758 if_setbaudrate(ifp, baudrate);
4760 /* If link down, disable watchdog */
4761 if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
4762 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
4763 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4765 ctx->ifc_link_state = link_state;
4766 if_link_state_change(ifp, link_state);
4770 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
4774 int credits_pre = txq->ift_cidx_processed;
4777 if (ctx->isc_txd_credits_update == NULL)
4780 if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0)
4783 txq->ift_processed += credits;
4784 txq->ift_cidx_processed += credits;
4786 MPASS(credits_pre + credits == txq->ift_cidx_processed);
4787 if (txq->ift_cidx_processed >= txq->ift_size)
4788 txq->ift_cidx_processed -= txq->ift_size;
4793 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget)
4796 return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
4801 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
4802 const char *description, if_int_delay_info_t info,
4803 int offset, int value)
4805 info->iidi_ctx = ctx;
4806 info->iidi_offset = offset;
4807 info->iidi_value = value;
4808 SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
4809 SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
4810 OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
4811 info, 0, iflib_sysctl_int_delay, "I", description);
4815 iflib_ctx_lock_get(if_ctx_t ctx)
4818 return (&ctx->ifc_mtx);
4822 iflib_msix_init(if_ctx_t ctx)
4824 device_t dev = ctx->ifc_dev;
4825 if_shared_ctx_t sctx = ctx->ifc_sctx;
4826 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4827 int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
4828 int iflib_num_tx_queues, iflib_num_rx_queues;
4829 int err, admincnt, bar;
4831 iflib_num_tx_queues = scctx->isc_ntxqsets;
4832 iflib_num_rx_queues = scctx->isc_nrxqsets;
4834 device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
4836 bar = ctx->ifc_softc_ctx.isc_msix_bar;
4837 admincnt = sctx->isc_admin_intrcnt;
4838 /* Override by tuneable */
4839 if (enable_msix == 0)
4843 ** When used in a virtualized environment
4844 ** PCI BUSMASTER capability may not be set
4845 ** so explicity set it here and rewrite
4846 ** the ENABLE in the MSIX control register
4847 ** at this point to cause the host to
4848 ** successfully initialize us.
4853 pci_enable_busmaster(dev);
4855 if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
4856 rid += PCIR_MSIX_CTRL;
4857 msix_ctrl = pci_read_config(dev, rid, 2);
4858 msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
4859 pci_write_config(dev, rid, msix_ctrl, 2);
4861 device_printf(dev, "PCIY_MSIX capability not found; "
4862 "or rid %d == 0.\n", rid);
4868 * bar == -1 => "trust me I know what I'm doing"
4869 * https://www.youtube.com/watch?v=nnwWKkNau4I
4870 * Some drivers are for hardware that is so shoddily
4871 * documented that no one knows which bars are which
4872 * so the developer has to map all bars. This hack
4873 * allows shoddy garbage to use msix in this framework.
4876 ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
4877 SYS_RES_MEMORY, &bar, RF_ACTIVE);
4878 if (ctx->ifc_msix_mem == NULL) {
4879 /* May not be enabled */
4880 device_printf(dev, "Unable to map MSIX table \n");
4884 /* First try MSI/X */
4885 if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
4886 device_printf(dev, "System has MSIX disabled \n");
4887 bus_release_resource(dev, SYS_RES_MEMORY,
4888 bar, ctx->ifc_msix_mem);
4889 ctx->ifc_msix_mem = NULL;
4893 /* use only 1 qset in debug mode */
4894 queuemsgs = min(msgs - admincnt, 1);
4896 queuemsgs = msgs - admincnt;
4898 if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
4900 queues = imin(queuemsgs, rss_getnumbuckets());
4904 queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
4905 device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
4906 CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
4908 device_printf(dev, "Unable to fetch CPU list\n");
4909 /* Figure out a reasonable auto config value */
4910 queues = min(queuemsgs, mp_ncpus);
4913 /* If we're doing RSS, clamp at the number of RSS buckets */
4914 if (queues > rss_getnumbuckets())
4915 queues = rss_getnumbuckets();
4917 if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
4918 rx_queues = iflib_num_rx_queues;
4922 * We want this to be all logical CPUs by default
4924 if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
4925 tx_queues = iflib_num_tx_queues;
4927 tx_queues = mp_ncpus;
4929 if (ctx->ifc_sysctl_qs_eq_override == 0) {
4931 if (tx_queues != rx_queues)
4932 device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
4933 min(rx_queues, tx_queues), min(rx_queues, tx_queues));
4935 tx_queues = min(rx_queues, tx_queues);
4936 rx_queues = min(rx_queues, tx_queues);
4939 device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
4941 vectors = rx_queues + admincnt;
4942 if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
4944 "Using MSIX interrupts with %d vectors\n", vectors);
4945 scctx->isc_vectors = vectors;
4946 scctx->isc_nrxqsets = rx_queues;
4947 scctx->isc_ntxqsets = tx_queues;
4948 scctx->isc_intr = IFLIB_INTR_MSIX;
4952 device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
4955 vectors = pci_msi_count(dev);
4956 scctx->isc_nrxqsets = 1;
4957 scctx->isc_ntxqsets = 1;
4958 scctx->isc_vectors = vectors;
4959 if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
4960 device_printf(dev,"Using an MSI interrupt\n");
4961 scctx->isc_intr = IFLIB_INTR_MSI;
4963 device_printf(dev,"Using a Legacy interrupt\n");
4964 scctx->isc_intr = IFLIB_INTR_LEGACY;
4970 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
4973 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
4976 uint16_t *state = ((uint16_t *)oidp->oid_arg1);
4978 char *ring_state = "UNKNOWN";
4981 rc = sysctl_wire_old_buffer(req, 0);
4985 sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
4990 ring_state = ring_states[state[3]];
4992 sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
4993 state[0], state[1], state[2], ring_state);
4994 rc = sbuf_finish(sb);
4999 enum iflib_ndesc_handler {
5005 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5007 if_ctx_t ctx = (void *)arg1;
5008 enum iflib_ndesc_handler type = arg2;
5009 char buf[256] = {0};
5014 MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5018 case IFLIB_NTXD_HANDLER:
5019 ndesc = ctx->ifc_sysctl_ntxds;
5021 nqs = ctx->ifc_sctx->isc_ntxqs;
5023 case IFLIB_NRXD_HANDLER:
5024 ndesc = ctx->ifc_sysctl_nrxds;
5026 nqs = ctx->ifc_sctx->isc_nrxqs;
5032 for (i=0; i<8; i++) {
5037 sprintf(strchr(buf, 0), "%d", ndesc[i]);
5040 rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5041 if (rc || req->newptr == NULL)
5044 for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5045 i++, p = strsep(&next, " ,")) {
5046 ndesc[i] = strtoul(p, NULL, 10);
5052 #define NAME_BUFLEN 32
5054 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5056 device_t dev = iflib_get_dev(ctx);
5057 struct sysctl_oid_list *child, *oid_list;
5058 struct sysctl_ctx_list *ctx_list;
5059 struct sysctl_oid *node;
5061 ctx_list = device_get_sysctl_ctx(dev);
5062 child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5063 ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5064 CTLFLAG_RD, NULL, "IFLIB fields");
5065 oid_list = SYSCTL_CHILDREN(node);
5067 SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5068 CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5071 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5072 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5073 "# of txqs to use, 0 => use default #");
5074 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5075 CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5076 "# of rxqs to use, 0 => use default #");
5077 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5078 CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5079 "permit #txq != #rxq");
5081 /* XXX change for per-queue sizes */
5082 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5083 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5084 mp_ndesc_handler, "A",
5085 "list of # of tx descriptors to use, 0 = use default #");
5086 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5087 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5088 mp_ndesc_handler, "A",
5089 "list of # of rx descriptors to use, 0 = use default #");
5093 iflib_add_device_sysctl_post(if_ctx_t ctx)
5095 if_shared_ctx_t sctx = ctx->ifc_sctx;
5096 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5097 device_t dev = iflib_get_dev(ctx);
5098 struct sysctl_oid_list *child;
5099 struct sysctl_ctx_list *ctx_list;
5104 char namebuf[NAME_BUFLEN];
5106 struct sysctl_oid *queue_node, *fl_node, *node;
5107 struct sysctl_oid_list *queue_list, *fl_list;
5108 ctx_list = device_get_sysctl_ctx(dev);
5110 node = ctx->ifc_sysctl_node;
5111 child = SYSCTL_CHILDREN(node);
5113 if (scctx->isc_ntxqsets > 100)
5115 else if (scctx->isc_ntxqsets > 10)
5119 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5120 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5121 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5122 CTLFLAG_RD, NULL, "Queue Name");
5123 queue_list = SYSCTL_CHILDREN(queue_node);
5125 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5127 &txq->ift_dequeued, "total mbufs freed");
5128 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5130 &txq->ift_enqueued, "total mbufs enqueued");
5132 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5134 &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5135 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5137 &txq->ift_pullups, "# of times m_pullup was called");
5138 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5140 &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5141 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5143 &txq->ift_no_desc_avail, "# of times no descriptors were available");
5144 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5146 &txq->ift_map_failed, "# of times dma map failed");
5147 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5149 &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5150 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5152 &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5153 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5155 &txq->ift_pidx, 1, "Producer Index");
5156 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5158 &txq->ift_cidx, 1, "Consumer Index");
5159 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5161 &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5162 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5164 &txq->ift_in_use, 1, "descriptors in use");
5165 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5167 &txq->ift_processed, "descriptors procesed for clean");
5168 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5170 &txq->ift_cleaned, "total cleaned");
5171 SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5172 CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state),
5173 0, mp_ring_state_handler, "A", "soft ring state");
5174 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5175 CTLFLAG_RD, &txq->ift_br[0]->enqueues,
5176 "# of enqueues to the mp_ring for this queue");
5177 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5178 CTLFLAG_RD, &txq->ift_br[0]->drops,
5179 "# of drops in the mp_ring for this queue");
5180 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5181 CTLFLAG_RD, &txq->ift_br[0]->starts,
5182 "# of normal consumer starts in the mp_ring for this queue");
5183 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5184 CTLFLAG_RD, &txq->ift_br[0]->stalls,
5185 "# of consumer stalls in the mp_ring for this queue");
5186 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5187 CTLFLAG_RD, &txq->ift_br[0]->restarts,
5188 "# of consumer restarts in the mp_ring for this queue");
5189 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5190 CTLFLAG_RD, &txq->ift_br[0]->abdications,
5191 "# of consumer abdications in the mp_ring for this queue");
5194 if (scctx->isc_nrxqsets > 100)
5196 else if (scctx->isc_nrxqsets > 10)
5200 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5201 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5202 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5203 CTLFLAG_RD, NULL, "Queue Name");
5204 queue_list = SYSCTL_CHILDREN(queue_node);
5205 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5206 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5208 &rxq->ifr_cq_pidx, 1, "Producer Index");
5209 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5211 &rxq->ifr_cq_cidx, 1, "Consumer Index");
5214 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5215 snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5216 fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5217 CTLFLAG_RD, NULL, "freelist Name");
5218 fl_list = SYSCTL_CHILDREN(fl_node);
5219 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5221 &fl->ifl_pidx, 1, "Producer Index");
5222 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5224 &fl->ifl_cidx, 1, "Consumer Index");
5225 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5227 &fl->ifl_credits, 1, "credits available");
5229 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5231 &fl->ifl_m_enqueued, "mbufs allocated");
5232 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5234 &fl->ifl_m_dequeued, "mbufs freed");
5235 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5237 &fl->ifl_cl_enqueued, "clusters allocated");
5238 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5240 &fl->ifl_cl_dequeued, "clusters freed");