2 * Copyright (c) 2014-2016, 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>
55 #include <net/if_var.h>
56 #include <net/if_types.h>
57 #include <net/if_media.h>
59 #include <net/ethernet.h>
60 #include <net/mp_ring.h>
62 #include <netinet/in.h>
63 #include <netinet/in_pcb.h>
64 #include <netinet/tcp_lro.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/if_ether.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip6.h>
69 #include <netinet/tcp.h>
71 #include <machine/bus.h>
72 #include <machine/in_cksum.h>
77 #include <dev/led/led.h>
78 #include <dev/pci/pcireg.h>
79 #include <dev/pci/pcivar.h>
80 #include <dev/pci/pci_private.h>
82 #include <net/iflib.h>
86 #if defined(__i386__) || defined(__amd64__)
87 #include <sys/memdesc.h>
88 #include <machine/bus.h>
89 #include <machine/md_var.h>
90 #include <machine/specialreg.h>
91 #include <x86/include/busdma_impl.h>
92 #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
107 * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108 * we prefetch needs to be determined by the time spent in m_free vis a vis
109 * the cost of a prefetch. This will of course vary based on the workload:
110 * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111 * is quite expensive, thus suggesting very little prefetch.
112 * - small packet forwarding which is just returning a single mbuf to
113 * UMA will typically be very fast vis a vis the cost of a memory
120 * - private structures
121 * - iflib private utility functions
123 * - vlan registry and other exported functions
124 * - iflib public core functions
128 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 typedef struct iflib_txq *iflib_txq_t;
133 typedef struct iflib_rxq *iflib_rxq_t;
135 typedef struct iflib_fl *iflib_fl_t;
137 typedef struct iflib_filter_info {
138 driver_filter_t *ifi_filter;
139 void *ifi_filter_arg;
140 struct grouptask *ifi_task;
141 } *iflib_filter_info_t;
146 * Pointer to hardware driver's softc
153 if_shared_ctx_t ifc_sctx;
154 struct if_softc_ctx ifc_softc_ctx;
158 uint16_t ifc_nhwtxqs;
159 uint16_t ifc_nhwrxqs;
161 iflib_txq_t ifc_txqs;
162 iflib_rxq_t ifc_rxqs;
163 uint32_t ifc_if_flags;
165 uint32_t ifc_max_fl_buf_size;
170 int ifc_pause_frames;
171 int ifc_watchdog_events;
172 struct cdev *ifc_led_dev;
173 struct resource *ifc_msix_mem;
175 struct if_irq ifc_legacy_irq;
176 struct grouptask ifc_admin_task;
177 struct grouptask ifc_vflr_task;
178 struct iflib_filter_info ifc_filter_info;
179 struct ifmedia ifc_media;
181 struct sysctl_oid *ifc_sysctl_node;
182 uint16_t ifc_sysctl_ntxqs;
183 uint16_t ifc_sysctl_nrxqs;
184 uint16_t ifc_sysctl_ntxds;
185 uint16_t ifc_sysctl_nrxds;
186 struct if_txrx ifc_txrx;
187 #define isc_txd_encap ifc_txrx.ift_txd_encap
188 #define isc_txd_flush ifc_txrx.ift_txd_flush
189 #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update
190 #define isc_rxd_available ifc_txrx.ift_rxd_available
191 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
192 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
193 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
194 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
195 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
196 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
197 eventhandler_tag ifc_vlan_attach_event;
198 eventhandler_tag ifc_vlan_detach_event;
199 uint8_t ifc_mac[ETHER_ADDR_LEN];
200 char ifc_mtx_name[16];
205 iflib_get_softc(if_ctx_t ctx)
208 return (ctx->ifc_softc);
212 iflib_get_dev(if_ctx_t ctx)
215 return (ctx->ifc_dev);
219 iflib_get_ifp(if_ctx_t ctx)
222 return (ctx->ifc_ifp);
226 iflib_get_media(if_ctx_t ctx)
229 return (&ctx->ifc_media);
233 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
236 bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
240 iflib_get_softc_ctx(if_ctx_t ctx)
243 return (&ctx->ifc_softc_ctx);
247 iflib_get_sctx(if_ctx_t ctx)
250 return (ctx->ifc_sctx);
253 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
255 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
256 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
258 #define RX_SW_DESC_MAP_CREATED (1 << 0)
259 #define TX_SW_DESC_MAP_CREATED (1 << 1)
260 #define RX_SW_DESC_INUSE (1 << 3)
261 #define TX_SW_DESC_MAPPED (1 << 4)
263 typedef struct iflib_sw_rx_desc {
264 bus_dmamap_t ifsd_map; /* bus_dma map for packet */
265 struct mbuf *ifsd_m; /* rx: uninitialized mbuf */
266 caddr_t ifsd_cl; /* direct cluster pointer for rx */
270 typedef struct iflib_sw_tx_desc_val {
271 bus_dmamap_t ifsd_map; /* bus_dma map for packet */
272 struct mbuf *ifsd_m; /* pkthdr mbuf */
276 typedef struct iflib_sw_tx_desc_array {
277 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */
278 struct mbuf **ifsd_m; /* pkthdr mbufs */
280 } iflib_txsd_array_t;
283 /* magic number that should be high enough for any hardware */
284 #define IFLIB_MAX_TX_SEGS 128
285 #define IFLIB_MAX_RX_SEGS 32
286 #define IFLIB_RX_COPY_THRESH 128
287 #define IFLIB_MAX_RX_REFRESH 32
288 #define IFLIB_QUEUE_IDLE 0
289 #define IFLIB_QUEUE_HUNG 1
290 #define IFLIB_QUEUE_WORKING 2
292 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */
293 #define TX_BATCH_SIZE 16
295 #define IFLIB_RESTART_BUDGET 8
297 #define IFC_LEGACY 0x1
298 #define IFC_QFLUSH 0x2
299 #define IFC_MULTISEG 0x4
302 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
303 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
304 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
308 uint16_t ift_cidx_processed;
311 uint8_t ift_db_pending;
312 uint8_t ift_db_pending_queued;
313 uint8_t ift_npending;
315 uint64_t ift_processed;
316 uint64_t ift_cleaned;
318 uint64_t ift_enqueued;
319 uint64_t ift_dequeued;
321 uint64_t ift_no_tx_dma_setup;
322 uint64_t ift_no_desc_avail;
323 uint64_t ift_mbuf_defrag_failed;
324 uint64_t ift_mbuf_defrag;
325 uint64_t ift_map_failed;
326 uint64_t ift_txd_encap_efbig;
327 uint64_t ift_pullups;
330 struct mtx ift_db_mtx;
332 /* constant values */
334 struct ifmp_ring **ift_br;
335 struct grouptask ift_task;
338 struct callout ift_timer;
339 struct callout ift_db_check;
341 iflib_txsd_array_t ift_sds;
346 int ift_watchdog_time;
347 struct iflib_filter_info ift_filter_info;
348 bus_dma_tag_t ift_desc_tag;
349 bus_dma_tag_t ift_tso_desc_tag;
350 iflib_dma_info_t ift_ifdi;
351 #define MTX_NAME_LEN 16
352 char ift_mtx_name[MTX_NAME_LEN];
353 char ift_db_mtx_name[MTX_NAME_LEN];
354 bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE);
355 } __aligned(CACHE_LINE_SIZE);
360 uint16_t ifl_credits;
363 uint64_t ifl_m_enqueued;
364 uint64_t ifl_m_dequeued;
365 uint64_t ifl_cl_enqueued;
366 uint64_t ifl_cl_dequeued;
372 uint16_t ifl_buf_size;
375 iflib_rxsd_t ifl_sds;
378 bus_dma_tag_t ifl_desc_tag;
379 iflib_dma_info_t ifl_ifdi;
380 uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
381 caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
382 } __aligned(CACHE_LINE_SIZE);
385 get_inuse(int size, int cidx, int pidx, int gen)
391 else if (pidx < cidx)
392 used = size - cidx + pidx;
393 else if (gen == 0 && pidx == cidx)
395 else if (gen == 1 && pidx == cidx)
403 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
405 #define IDXDIFF(head, tail, wrap) \
406 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
409 /* If there is a separate completion queue -
410 * these are the cq cidx and pidx. Otherwise
414 uint16_t ifr_cq_cidx;
415 uint16_t ifr_cq_pidx;
422 uint8_t ifr_lro_enabled;
424 struct lro_ctrl ifr_lc;
425 struct grouptask ifr_task;
426 struct iflib_filter_info ifr_filter_info;
427 iflib_dma_info_t ifr_ifdi;
428 /* dynamically allocate if any drivers need a value substantially larger than this */
429 struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
430 } __aligned(CACHE_LINE_SIZE);
433 * Only allow a single packet to take up most 1/nth of the tx ring
435 #define MAX_SINGLE_PACKET_FRACTION 12
436 #define IF_BAD_DMA (bus_addr_t)-1
438 static int enable_msix = 1;
440 #define mtx_held(m) (((m)->mtx_lock & ~MTX_FLAGMASK) != (uintptr_t)0)
444 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
446 #define CTX_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
448 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
449 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
450 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
453 #define TXDB_LOCK_INIT(txq) mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF)
454 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx)
455 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx)
456 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx)
457 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx)
459 #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx)
460 #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx)
463 /* Our boot-time initialization hook */
464 static int iflib_module_event_handler(module_t, int, void *);
466 static moduledata_t iflib_moduledata = {
468 iflib_module_event_handler,
472 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
473 MODULE_VERSION(iflib, 1);
475 MODULE_DEPEND(iflib, pci, 1, 1, 1);
476 MODULE_DEPEND(iflib, ether, 1, 1, 1);
478 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
479 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
481 #ifndef IFLIB_DEBUG_COUNTERS
483 #define IFLIB_DEBUG_COUNTERS 1
485 #define IFLIB_DEBUG_COUNTERS 0
486 #endif /* !INVARIANTS */
489 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
490 "iflib driver parameters");
493 * XXX need to ensure that this can't accidentally cause the head to be moved backwards
495 static int iflib_min_tx_latency = 0;
497 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
498 &iflib_min_tx_latency, 0, "minimize transmit latency at the possibel expense of throughput");
501 #if IFLIB_DEBUG_COUNTERS
503 static int iflib_tx_seen;
504 static int iflib_tx_sent;
505 static int iflib_tx_encap;
506 static int iflib_rx_allocs;
507 static int iflib_fl_refills;
508 static int iflib_fl_refills_large;
509 static int iflib_tx_frees;
511 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
512 &iflib_tx_seen, 0, "# tx mbufs seen");
513 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
514 &iflib_tx_sent, 0, "# tx mbufs sent");
515 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
516 &iflib_tx_encap, 0, "# tx mbufs encapped");
517 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
518 &iflib_tx_frees, 0, "# tx frees");
519 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
520 &iflib_rx_allocs, 0, "# rx allocations");
521 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
522 &iflib_fl_refills, 0, "# refills");
523 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
524 &iflib_fl_refills_large, 0, "# large refills");
527 static int iflib_txq_drain_flushing;
528 static int iflib_txq_drain_oactive;
529 static int iflib_txq_drain_notready;
530 static int iflib_txq_drain_encapfail;
532 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
533 &iflib_txq_drain_flushing, 0, "# drain flushes");
534 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
535 &iflib_txq_drain_oactive, 0, "# drain oactives");
536 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
537 &iflib_txq_drain_notready, 0, "# drain notready");
538 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
539 &iflib_txq_drain_encapfail, 0, "# drain encap fails");
542 static int iflib_encap_load_mbuf_fail;
543 static int iflib_encap_txq_avail_fail;
544 static int iflib_encap_txd_encap_fail;
546 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
547 &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
548 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
549 &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
550 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
551 &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
553 static int iflib_task_fn_rxs;
554 static int iflib_rx_intr_enables;
555 static int iflib_fast_intrs;
556 static int iflib_intr_link;
557 static int iflib_intr_msix;
558 static int iflib_rx_unavail;
559 static int iflib_rx_ctx_inactive;
560 static int iflib_rx_zero_len;
561 static int iflib_rx_if_input;
562 static int iflib_rx_mbuf_null;
563 static int iflib_rxd_flush;
565 static int iflib_verbose_debug;
567 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
568 &iflib_intr_link, 0, "# intr link calls");
569 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
570 &iflib_intr_msix, 0, "# intr msix calls");
571 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
572 &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
573 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
574 &iflib_rx_intr_enables, 0, "# rx intr enables");
575 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
576 &iflib_fast_intrs, 0, "# fast_intr calls");
577 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
578 &iflib_rx_unavail, 0, "# times rxeof called with no available data");
579 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
580 &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
581 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
582 &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
583 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
584 &iflib_rx_if_input, 0, "# times rxeof called if_input");
585 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
586 &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
587 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
588 &iflib_rxd_flush, 0, "# times rxd_flush called");
589 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
590 &iflib_verbose_debug, 0, "enable verbose debugging");
592 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
595 #define DBG_COUNTER_INC(name)
601 #define IFLIB_DEBUG 0
603 static void iflib_tx_structures_free(if_ctx_t ctx);
604 static void iflib_rx_structures_free(if_ctx_t ctx);
605 static int iflib_queues_alloc(if_ctx_t ctx);
606 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
607 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx);
608 static int iflib_qset_structures_setup(if_ctx_t ctx);
609 static int iflib_msix_init(if_ctx_t ctx);
610 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
611 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
612 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
613 static int iflib_register(if_ctx_t);
614 static void iflib_init_locked(if_ctx_t ctx);
615 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
616 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
620 #include <sys/selinfo.h>
621 #include <net/netmap.h>
622 #include <dev/netmap/netmap_kern.h>
624 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
627 * device-specific sysctl variables:
629 * ixl_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
630 * During regular operations the CRC is stripped, but on some
631 * hardware reception of frames not multiple of 64 is slower,
632 * so using crcstrip=0 helps in benchmarks.
634 * ixl_rx_miss, ixl_rx_miss_bufs:
635 * count packets that might be missed due to lost interrupts.
637 SYSCTL_DECL(_dev_netmap);
639 * The xl driver by default strips CRCs and we do not override it.
642 int iflib_crcstrip = 1;
643 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
644 CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
646 int iflib_rx_miss, iflib_rx_miss_bufs;
647 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
648 CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
649 SYSCTL_INT(_dev_netmap, OID_AUTO, ixl_rx_miss_bufs,
650 CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
653 * Register/unregister. We are already under netmap lock.
654 * Only called on the first register or the last unregister.
657 iflib_netmap_register(struct netmap_adapter *na, int onoff)
659 struct ifnet *ifp = na->ifp;
660 if_ctx_t ctx = ifp->if_softc;
663 IFDI_INTR_DISABLE(ctx);
665 /* Tell the stack that the interface is no longer active */
666 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
669 IFDI_CRCSTRIP_SET(ctx, onoff);
671 /* enable or disable flags and callbacks in na and ifp */
673 nm_set_native_flags(na);
675 nm_clear_native_flags(na);
678 IFDI_CRCSTRIP_SET(ctx, onoff); // XXX why twice ?
680 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
684 * Reconcile kernel and user view of the transmit ring.
686 * All information is in the kring.
687 * Userspace wants to send packets up to the one before kring->rhead,
688 * kernel knows kring->nr_hwcur is the first unsent packet.
690 * Here we push packets out (as many as possible), and possibly
691 * reclaim buffers from previously completed transmission.
693 * The caller (netmap) guarantees that there is only one instance
694 * running at any time. Any interference with other driver
695 * methods should be handled by the individual drivers.
698 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
700 struct netmap_adapter *na = kring->na;
701 struct ifnet *ifp = na->ifp;
702 struct netmap_ring *ring = kring->ring;
703 u_int nm_i; /* index into the netmap ring */
704 u_int nic_i; /* index into the NIC ring */
706 u_int const lim = kring->nkr_num_slots - 1;
707 u_int const head = kring->rhead;
708 struct if_pkt_info pi;
711 * interrupts on every tx packet are expensive so request
712 * them every half ring, or where NS_REPORT is set
714 u_int report_frequency = kring->nkr_num_slots >> 1;
715 /* device-specific */
716 if_ctx_t ctx = ifp->if_softc;
717 iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
719 pi.ipi_segs = txq->ift_segs;
720 pi.ipi_qsidx = kring->ring_id;
723 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
724 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
728 * First part: process new packets to send.
729 * nm_i is the current index in the netmap ring,
730 * nic_i is the corresponding index in the NIC ring.
732 * If we have packets to send (nm_i != head)
733 * iterate over the netmap ring, fetch length and update
734 * the corresponding slot in the NIC ring. Some drivers also
735 * need to update the buffer's physical address in the NIC slot
736 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
738 * The netmap_reload_map() calls is especially expensive,
739 * even when (as in this case) the tag is 0, so do only
740 * when the buffer has actually changed.
742 * If possible do not set the report/intr bit on all slots,
743 * but only a few times per ring or when NS_REPORT is set.
745 * Finally, on 10G and faster drivers, it might be useful
746 * to prefetch the next slot and txr entry.
749 nm_i = kring->nr_hwcur;
750 if (nm_i != head) { /* we have new packets to send */
751 nic_i = netmap_idx_k2n(kring, nm_i);
753 __builtin_prefetch(&ring->slot[nm_i]);
754 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
755 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
757 for (n = 0; nm_i != head; n++) {
758 struct netmap_slot *slot = &ring->slot[nm_i];
759 u_int len = slot->len;
761 void *addr = PNMB(na, slot, &paddr);
762 int flags = (slot->flags & NS_REPORT ||
763 nic_i == 0 || nic_i == report_frequency) ?
766 /* device-specific */
768 pi.ipi_flags = flags;
770 /* Fill the slot in the NIC ring. */
771 ctx->isc_txd_encap(ctx->ifc_softc, &pi);
773 /* prefetch for next round */
774 __builtin_prefetch(&ring->slot[nm_i + 1]);
775 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
776 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
778 NM_CHECK_ADDR_LEN(na, addr, len);
780 if (slot->flags & NS_BUF_CHANGED) {
781 /* buffer has changed, reload map */
782 netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
784 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
786 /* make sure changes to the buffer are synced */
787 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
788 BUS_DMASYNC_PREWRITE);
790 nm_i = nm_next(nm_i, lim);
791 nic_i = nm_next(nic_i, lim);
793 kring->nr_hwcur = head;
795 /* synchronize the NIC ring */
796 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
797 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
799 /* (re)start the tx unit up to slot nic_i (excluded) */
800 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
804 * Second part: reclaim buffers for completed transmissions.
806 if (iflib_tx_credits_update(ctx, txq)) {
807 /* some tx completed, increment avail */
808 nic_i = txq->ift_cidx_processed;
809 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
815 * Reconcile kernel and user view of the receive ring.
816 * Same as for the txsync, this routine must be efficient.
817 * The caller guarantees a single invocations, but races against
818 * the rest of the driver should be handled here.
820 * On call, kring->rhead is the first packet that userspace wants
821 * to keep, and kring->rcur is the wakeup point.
822 * The kernel has previously reported packets up to kring->rtail.
824 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
825 * of whether or not we received an interrupt.
828 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
830 struct netmap_adapter *na = kring->na;
831 struct ifnet *ifp = na->ifp;
832 struct netmap_ring *ring = kring->ring;
833 u_int nm_i; /* index into the netmap ring */
834 u_int nic_i; /* index into the NIC ring */
836 u_int const lim = kring->nkr_num_slots - 1;
837 u_int const head = kring->rhead;
838 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
839 struct if_rxd_info ri;
840 /* device-specific */
841 if_ctx_t ctx = ifp->if_softc;
842 iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
843 iflib_fl_t fl = rxq->ifr_fl;
845 return netmap_ring_reinit(kring);
847 bzero(&ri, sizeof(ri));
848 ri.iri_qsidx = kring->ring_id;
849 ri.iri_ifp = ctx->ifc_ifp;
850 /* XXX check sync modes */
851 for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++)
852 bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
853 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
856 * First part: import newly received packets.
858 * nm_i is the index of the next free slot in the netmap ring,
859 * nic_i is the index of the next received packet in the NIC ring,
860 * and they may differ in case if_init() has been called while
861 * in netmap mode. For the receive ring we have
863 * nic_i = rxr->next_check;
864 * nm_i = kring->nr_hwtail (previous)
866 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
868 * rxr->next_check is set to 0 on a ring reinit
870 if (netmap_no_pendintr || force_update) {
871 int crclen = iflib_crcstrip ? 0 : 4;
873 uint16_t slot_flags = kring->nkr_slot_flags;
875 for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
876 nic_i = fl->ifl_cidx;
877 nm_i = netmap_idx_n2k(kring, nic_i);
878 avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i);
879 for (n = 0; avail > 0; n++, avail--) {
880 error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
882 ring->slot[nm_i].len = 0;
884 ring->slot[nm_i].len = ri.iri_len - crclen;
885 ring->slot[nm_i].flags = slot_flags;
886 bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
887 fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_POSTREAD);
888 nm_i = nm_next(nm_i, lim);
889 nic_i = nm_next(nic_i, lim);
891 if (n) { /* update the state variables */
892 if (netmap_no_pendintr && !force_update) {
895 iflib_rx_miss_bufs += n;
897 fl->ifl_cidx = nic_i;
898 kring->nr_hwtail = nm_i;
900 kring->nr_kflags &= ~NKR_PENDINTR;
904 * Second part: skip past packets that userspace has released.
905 * (kring->nr_hwcur to head excluded),
906 * and make the buffers available for reception.
907 * As usual nm_i is the index in the netmap ring,
908 * nic_i is the index in the NIC ring, and
909 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
911 /* XXX not sure how this will work with multiple free lists */
912 nm_i = kring->nr_hwcur;
914 nic_i = netmap_idx_k2n(kring, nm_i);
915 for (n = 0; nm_i != head; n++) {
916 struct netmap_slot *slot = &ring->slot[nm_i];
919 void *addr = PNMB(na, slot, &paddr);
921 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
925 if (slot->flags & NS_BUF_CHANGED) {
926 /* buffer has changed, reload map */
927 netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, addr);
928 slot->flags &= ~NS_BUF_CHANGED;
931 * XXX we should be batching this operation - TODO
933 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1);
934 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map,
935 BUS_DMASYNC_PREREAD);
936 nm_i = nm_next(nm_i, lim);
937 nic_i = nm_next(nic_i, lim);
939 kring->nr_hwcur = head;
941 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
942 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
944 * IMPORTANT: we must leave one free slot in the ring,
945 * so move nic_i back by one unit
947 nic_i = nm_prev(nic_i, lim);
948 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
954 return netmap_ring_reinit(kring);
958 iflib_netmap_attach(if_ctx_t ctx)
960 struct netmap_adapter na;
962 bzero(&na, sizeof(na));
964 na.ifp = ctx->ifc_ifp;
965 na.na_flags = NAF_BDG_MAYSLEEP;
966 MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
967 MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
969 na.num_tx_desc = ctx->ifc_sctx->isc_ntxd;
970 na.num_rx_desc = ctx->ifc_sctx->isc_ntxd;
971 na.nm_txsync = iflib_netmap_txsync;
972 na.nm_rxsync = iflib_netmap_rxsync;
973 na.nm_register = iflib_netmap_register;
974 na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
975 na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
976 return (netmap_attach(&na));
980 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
982 struct netmap_adapter *na = NA(ctx->ifc_ifp);
983 struct netmap_slot *slot;
985 slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
989 for (int i = 0; i < ctx->ifc_sctx->isc_ntxd; i++) {
992 * In netmap mode, set the map for the packet buffer.
993 * NOTE: Some drivers (not this one) also need to set
994 * the physical buffer address in the NIC ring.
995 * netmap_idx_n2k() maps a nic index, i, into the corresponding
996 * netmap slot index, si
998 int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
999 netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1003 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1005 struct netmap_adapter *na = NA(ctx->ifc_ifp);
1006 struct netmap_slot *slot;
1010 slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1013 sd = rxq->ifr_fl[0].ifl_sds;
1014 nrxd = ctx->ifc_sctx->isc_nrxd;
1015 for (int i = 0; i < nrxd; i++, sd++) {
1016 int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1021 vaddr = addr = PNMB(na, slot + sj, &paddr);
1022 netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, sd->ifsd_map, addr);
1023 /* Update descriptor and the cached value */
1024 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1);
1026 /* preserve queue */
1027 if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1028 struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1029 int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1030 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1032 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1035 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1038 #define iflib_netmap_txq_init(ctx, txq)
1039 #define iflib_netmap_rxq_init(ctx, rxq)
1040 #define iflib_netmap_detach(ifp)
1042 #define iflib_netmap_attach(ctx) (0)
1043 #define netmap_rx_irq(ifp, qid, budget) (0)
1047 #if defined(__i386__) || defined(__amd64__)
1048 static __inline void
1051 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1058 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1062 *(bus_addr_t *) arg = segs[0].ds_addr;
1066 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1069 if_shared_ctx_t sctx = ctx->ifc_sctx;
1070 device_t dev = ctx->ifc_dev;
1072 KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1074 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1075 sctx->isc_q_align, 0, /* alignment, bounds */
1076 BUS_SPACE_MAXADDR, /* lowaddr */
1077 BUS_SPACE_MAXADDR, /* highaddr */
1078 NULL, NULL, /* filter, filterarg */
1081 size, /* maxsegsize */
1082 BUS_DMA_ALLOCNOW, /* flags */
1083 NULL, /* lockfunc */
1088 "%s: bus_dma_tag_create failed: %d\n",
1093 err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1094 BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1097 "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1098 __func__, (uintmax_t)size, err);
1102 dma->idi_paddr = IF_BAD_DMA;
1103 err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1104 size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1105 if (err || dma->idi_paddr == IF_BAD_DMA) {
1107 "%s: bus_dmamap_load failed: %d\n",
1112 dma->idi_size = size;
1116 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1118 bus_dma_tag_destroy(dma->idi_tag);
1120 dma->idi_tag = NULL;
1126 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1129 iflib_dma_info_t *dmaiter;
1132 for (i = 0; i < count; i++, dmaiter++) {
1133 if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1137 iflib_dma_free_multi(dmalist, i);
1142 iflib_dma_free(iflib_dma_info_t dma)
1144 if (dma->idi_tag == NULL)
1146 if (dma->idi_paddr != IF_BAD_DMA) {
1147 bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1148 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1149 bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1150 dma->idi_paddr = IF_BAD_DMA;
1152 if (dma->idi_vaddr != NULL) {
1153 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1154 dma->idi_vaddr = NULL;
1156 bus_dma_tag_destroy(dma->idi_tag);
1157 dma->idi_tag = NULL;
1161 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1164 iflib_dma_info_t *dmaiter = dmalist;
1166 for (i = 0; i < count; i++, dmaiter++)
1167 iflib_dma_free(*dmaiter);
1171 iflib_fast_intr(void *arg)
1173 iflib_filter_info_t info = arg;
1174 struct grouptask *gtask = info->ifi_task;
1176 DBG_COUNTER_INC(fast_intrs);
1177 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1178 return (FILTER_HANDLED);
1180 GROUPTASK_ENQUEUE(gtask);
1181 return (FILTER_HANDLED);
1185 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1186 driver_filter_t filter, driver_intr_t handler, void *arg,
1190 struct resource *res;
1192 device_t dev = ctx->ifc_dev;
1196 res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1197 RF_SHAREABLE | RF_ACTIVE);
1200 "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1204 KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1205 rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1206 filter, handler, arg, &tag);
1209 "failed to setup interrupt for rid %d, name %s: %d\n",
1210 rid, name ? name : "unknown", rc);
1213 bus_describe_intr(dev, res, tag, name);
1220 /*********************************************************************
1222 * Allocate memory for tx_buffer structures. The tx_buffer stores all
1223 * the information needed to transmit a packet on the wire. This is
1224 * called only once at attach, setup is done every reset.
1226 **********************************************************************/
1229 iflib_txsd_alloc(iflib_txq_t txq)
1231 if_ctx_t ctx = txq->ift_ctx;
1232 if_shared_ctx_t sctx = ctx->ifc_sctx;
1233 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1234 device_t dev = ctx->ifc_dev;
1235 int err, nsegments, ntsosegments;
1237 nsegments = scctx->isc_tx_nsegments;
1238 ntsosegments = scctx->isc_tx_tso_segments_max;
1239 MPASS(sctx->isc_ntxd > 0);
1240 MPASS(nsegments > 0);
1241 MPASS(ntsosegments > 0);
1243 * Setup DMA descriptor areas.
1245 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1246 1, 0, /* alignment, bounds */
1247 BUS_SPACE_MAXADDR, /* lowaddr */
1248 BUS_SPACE_MAXADDR, /* highaddr */
1249 NULL, NULL, /* filter, filterarg */
1250 sctx->isc_tx_maxsize, /* maxsize */
1251 nsegments, /* nsegments */
1252 sctx->isc_tx_maxsegsize, /* maxsegsize */
1254 NULL, /* lockfunc */
1255 NULL, /* lockfuncarg */
1256 &txq->ift_desc_tag))) {
1257 device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1258 device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1259 sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1263 device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1264 sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1266 device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n",
1267 scctx->isc_tx_tso_size_max, ntsosegments,
1268 scctx->isc_tx_tso_segsize_max);
1269 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1270 1, 0, /* alignment, bounds */
1271 BUS_SPACE_MAXADDR, /* lowaddr */
1272 BUS_SPACE_MAXADDR, /* highaddr */
1273 NULL, NULL, /* filter, filterarg */
1274 scctx->isc_tx_tso_size_max, /* maxsize */
1275 ntsosegments, /* nsegments */
1276 scctx->isc_tx_tso_segsize_max, /* maxsegsize */
1278 NULL, /* lockfunc */
1279 NULL, /* lockfuncarg */
1280 &txq->ift_tso_desc_tag))) {
1281 device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1286 device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n",
1287 scctx->isc_tx_tso_size_max, ntsosegments,
1288 scctx->isc_tx_tso_segsize_max);
1290 if (!(txq->ift_sds.ifsd_flags =
1291 (uint8_t *) malloc(sizeof(uint8_t) *
1292 sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) {
1293 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1297 if (!(txq->ift_sds.ifsd_m =
1298 (struct mbuf **) malloc(sizeof(struct mbuf *) *
1299 sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) {
1300 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1305 /* Create the descriptor buffer dma maps */
1306 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__)))
1307 if ((ctx->ifc_flags & IFC_DMAR) == 0)
1310 if (!(txq->ift_sds.ifsd_map =
1311 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) {
1312 device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1317 for (int i = 0; i < sctx->isc_ntxd; i++) {
1318 err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1320 device_printf(dev, "Unable to create TX DMA map\n");
1327 /* We free all, it handles case where we are in the middle */
1328 iflib_tx_structures_free(ctx);
1333 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1338 if (txq->ift_sds.ifsd_map != NULL)
1339 map = txq->ift_sds.ifsd_map[i];
1341 bus_dmamap_unload(txq->ift_desc_tag, map);
1342 bus_dmamap_destroy(txq->ift_desc_tag, map);
1343 txq->ift_sds.ifsd_map[i] = NULL;
1348 iflib_txq_destroy(iflib_txq_t txq)
1350 if_ctx_t ctx = txq->ift_ctx;
1351 if_shared_ctx_t sctx = ctx->ifc_sctx;
1353 for (int i = 0; i < sctx->isc_ntxd; i++)
1354 iflib_txsd_destroy(ctx, txq, i);
1355 if (txq->ift_sds.ifsd_map != NULL) {
1356 free(txq->ift_sds.ifsd_map, M_IFLIB);
1357 txq->ift_sds.ifsd_map = NULL;
1359 if (txq->ift_sds.ifsd_m != NULL) {
1360 free(txq->ift_sds.ifsd_m, M_IFLIB);
1361 txq->ift_sds.ifsd_m = NULL;
1363 if (txq->ift_sds.ifsd_flags != NULL) {
1364 free(txq->ift_sds.ifsd_flags, M_IFLIB);
1365 txq->ift_sds.ifsd_flags = NULL;
1367 if (txq->ift_desc_tag != NULL) {
1368 bus_dma_tag_destroy(txq->ift_desc_tag);
1369 txq->ift_desc_tag = NULL;
1371 if (txq->ift_tso_desc_tag != NULL) {
1372 bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1373 txq->ift_tso_desc_tag = NULL;
1378 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1382 mp = &txq->ift_sds.ifsd_m[i];
1386 if (txq->ift_sds.ifsd_map != NULL) {
1387 bus_dmamap_sync(txq->ift_desc_tag,
1388 txq->ift_sds.ifsd_map[i],
1389 BUS_DMASYNC_POSTWRITE);
1390 bus_dmamap_unload(txq->ift_desc_tag,
1391 txq->ift_sds.ifsd_map[i]);
1394 DBG_COUNTER_INC(tx_frees);
1399 iflib_txq_setup(iflib_txq_t txq)
1401 if_ctx_t ctx = txq->ift_ctx;
1402 if_shared_ctx_t sctx = ctx->ifc_sctx;
1403 iflib_dma_info_t di;
1406 /* Set number of descriptors available */
1407 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1410 txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1411 txq->ift_size = sctx->isc_ntxd;
1413 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1414 bzero((void *)di->idi_vaddr, di->idi_size);
1416 IFDI_TXQ_SETUP(ctx, txq->ift_id);
1417 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1418 bus_dmamap_sync(di->idi_tag, di->idi_map,
1419 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1423 /*********************************************************************
1425 * Allocate memory for rx_buffer structures. Since we use one
1426 * rx_buffer per received packet, the maximum number of rx_buffer's
1427 * that we'll need is equal to the number of receive descriptors
1428 * that we've allocated.
1430 **********************************************************************/
1432 iflib_rxsd_alloc(iflib_rxq_t rxq)
1434 if_ctx_t ctx = rxq->ifr_ctx;
1435 if_shared_ctx_t sctx = ctx->ifc_sctx;
1436 device_t dev = ctx->ifc_dev;
1441 MPASS(sctx->isc_nrxd > 0);
1444 for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1445 fl->ifl_sds = malloc(sizeof(struct iflib_sw_rx_desc) *
1446 sctx->isc_nrxd, M_IFLIB, M_WAITOK | M_ZERO);
1447 if (fl->ifl_sds == NULL) {
1448 device_printf(dev, "Unable to allocate rx sw desc memory\n");
1451 fl->ifl_size = sctx->isc_nrxd; /* this isn't necessarily the same */
1452 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1453 1, 0, /* alignment, bounds */
1454 BUS_SPACE_MAXADDR, /* lowaddr */
1455 BUS_SPACE_MAXADDR, /* highaddr */
1456 NULL, NULL, /* filter, filterarg */
1457 sctx->isc_rx_maxsize, /* maxsize */
1458 sctx->isc_rx_nsegments, /* nsegments */
1459 sctx->isc_rx_maxsegsize, /* maxsegsize */
1461 NULL, /* lockfunc */
1465 device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1471 for (int i = 0; i < sctx->isc_nrxd; i++, rxsd++) {
1472 err = bus_dmamap_create(fl->ifl_desc_tag, 0, &rxsd->ifsd_map);
1474 device_printf(dev, "%s: bus_dmamap_create failed: %d\n",
1483 iflib_rx_structures_free(ctx);
1489 * Internal service routines
1492 struct rxq_refill_cb_arg {
1494 bus_dma_segment_t seg;
1499 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1501 struct rxq_refill_cb_arg *cb_arg = arg;
1503 cb_arg->error = error;
1504 cb_arg->seg = segs[0];
1505 cb_arg->nseg = nseg;
1510 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1512 #define IS_DMAR(ctx) (0)
1516 * rxq_refill - refill an rxq free-buffer list
1517 * @ctx: the iflib context
1518 * @rxq: the free-list to refill
1519 * @n: the number of new buffers to allocate
1521 * (Re)populate an rxq free-buffer list with up to @n new packet buffers.
1522 * The caller must assure that @n does not exceed the queue's capacity.
1525 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1528 int pidx = fl->ifl_pidx;
1529 iflib_rxsd_t rxsd = &fl->ifl_sds[pidx];
1537 MPASS(fl->ifl_credits + n <= fl->ifl_size);
1539 if (pidx < fl->ifl_cidx)
1540 MPASS(pidx + n <= fl->ifl_cidx);
1541 if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1542 MPASS(fl->ifl_gen == 0);
1543 if (pidx > fl->ifl_cidx)
1544 MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1546 DBG_COUNTER_INC(fl_refills);
1548 DBG_COUNTER_INC(fl_refills_large);
1552 * We allocate an uninitialized mbuf + cluster, mbuf is
1553 * initialized after rx.
1555 * If the cluster is still set then we know a minimum sized packet was received
1557 if ((cl = rxsd->ifsd_cl) == NULL) {
1558 if ((cl = rxsd->ifsd_cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1561 fl->ifl_cl_enqueued++;
1564 if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1568 fl->ifl_m_enqueued++;
1571 DBG_COUNTER_INC(rx_allocs);
1573 if ((rxsd->ifsd_flags & RX_SW_DESC_MAP_CREATED) == 0) {
1576 if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &rxsd->ifsd_map))) {
1577 log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
1578 uma_zfree(fl->ifl_zone, cl);
1582 rxsd->ifsd_flags |= RX_SW_DESC_MAP_CREATED;
1585 #if defined(__i386__) || defined(__amd64__)
1586 if (!IS_DMAR(ctx)) {
1587 bus_addr = pmap_kextract((vm_offset_t)cl);
1591 struct rxq_refill_cb_arg cb_arg;
1596 err = bus_dmamap_load(fl->ifl_desc_tag, rxsd->ifsd_map,
1597 cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1599 if (err != 0 || cb_arg.error) {
1603 if (fl->ifl_zone == zone_pack)
1604 uma_zfree(fl->ifl_zone, cl);
1609 bus_addr = cb_arg.seg.ds_addr;
1611 rxsd->ifsd_flags |= RX_SW_DESC_INUSE;
1613 MPASS(rxsd->ifsd_m == NULL);
1616 fl->ifl_bus_addrs[i] = bus_addr;
1617 fl->ifl_vm_addrs[i] = cl;
1621 MPASS(fl->ifl_credits <= fl->ifl_size);
1622 if (++fl->ifl_pidx == fl->ifl_size) {
1627 if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1628 ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx,
1629 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i);
1631 pidx = fl->ifl_pidx;
1635 DBG_COUNTER_INC(rxd_flush);
1636 if (fl->ifl_pidx == 0)
1637 pidx = fl->ifl_size - 1;
1639 pidx = fl->ifl_pidx - 1;
1640 ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1643 static __inline void
1644 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1646 /* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1647 int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1649 int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1652 MPASS(fl->ifl_credits <= fl->ifl_size);
1653 MPASS(reclaimable == delta);
1655 if (reclaimable > 0)
1656 _iflib_fl_refill(ctx, fl, min(max, reclaimable));
1660 iflib_fl_bufs_free(iflib_fl_t fl)
1662 iflib_dma_info_t idi = fl->ifl_ifdi;
1665 for (i = 0; i < fl->ifl_size; i++) {
1666 iflib_rxsd_t d = &fl->ifl_sds[i];
1668 if (d->ifsd_flags & RX_SW_DESC_INUSE) {
1669 bus_dmamap_unload(fl->ifl_desc_tag, d->ifsd_map);
1670 bus_dmamap_destroy(fl->ifl_desc_tag, d->ifsd_map);
1671 if (d->ifsd_m != NULL) {
1672 m_init(d->ifsd_m, M_NOWAIT, MT_DATA, 0);
1673 uma_zfree(zone_mbuf, d->ifsd_m);
1675 if (d->ifsd_cl != NULL)
1676 uma_zfree(fl->ifl_zone, d->ifsd_cl);
1679 MPASS(d->ifsd_cl == NULL);
1680 MPASS(d->ifsd_m == NULL);
1683 fl->ifl_m_dequeued++;
1684 fl->ifl_cl_dequeued++;
1690 * Reset free list values
1692 fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1693 bzero(idi->idi_vaddr, idi->idi_size);
1696 /*********************************************************************
1698 * Initialize a receive ring and its buffers.
1700 **********************************************************************/
1702 iflib_fl_setup(iflib_fl_t fl)
1704 iflib_rxq_t rxq = fl->ifl_rxq;
1705 if_ctx_t ctx = rxq->ifr_ctx;
1706 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1709 ** Free current RX buffer structs and their mbufs
1711 iflib_fl_bufs_free(fl);
1712 /* Now replenish the mbufs */
1713 MPASS(fl->ifl_credits == 0);
1715 * XXX don't set the max_frame_size to larger
1716 * than the hardware can handle
1718 if (sctx->isc_max_frame_size <= 2048)
1719 fl->ifl_buf_size = MCLBYTES;
1720 else if (sctx->isc_max_frame_size <= 4096)
1721 fl->ifl_buf_size = MJUMPAGESIZE;
1722 else if (sctx->isc_max_frame_size <= 9216)
1723 fl->ifl_buf_size = MJUM9BYTES;
1725 fl->ifl_buf_size = MJUM16BYTES;
1726 if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
1727 ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
1728 fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
1729 fl->ifl_zone = m_getzone(fl->ifl_buf_size);
1732 /* avoid pre-allocating zillions of clusters to an idle card
1733 * potentially speeding up attach
1735 _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
1736 MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
1737 if (min(128, fl->ifl_size) != fl->ifl_credits)
1743 MPASS(fl->ifl_ifdi != NULL);
1744 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1745 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1749 /*********************************************************************
1751 * Free receive ring data structures
1753 **********************************************************************/
1755 iflib_rx_sds_free(iflib_rxq_t rxq)
1760 if (rxq->ifr_fl != NULL) {
1761 for (i = 0; i < rxq->ifr_nfl; i++) {
1762 fl = &rxq->ifr_fl[i];
1763 if (fl->ifl_desc_tag != NULL) {
1764 bus_dma_tag_destroy(fl->ifl_desc_tag);
1765 fl->ifl_desc_tag = NULL;
1768 if (rxq->ifr_fl->ifl_sds != NULL)
1769 free(rxq->ifr_fl->ifl_sds, M_IFLIB);
1771 free(rxq->ifr_fl, M_IFLIB);
1773 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
1778 * MI independent logic
1782 iflib_timer(void *arg)
1784 iflib_txq_t txq = arg;
1785 if_ctx_t ctx = txq->ift_ctx;
1786 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1788 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
1791 ** Check on the state of the TX queue(s), this
1792 ** can be done without the lock because its RO
1793 ** and the HUNG state will be static if set.
1795 IFDI_TIMER(ctx, txq->ift_id);
1796 if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
1797 (ctx->ifc_pause_frames == 0))
1800 if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments ||
1801 ifmp_ring_is_stalled(txq->ift_br[0]))
1802 GROUPTASK_ENQUEUE(&txq->ift_task);
1804 ctx->ifc_pause_frames = 0;
1805 if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
1806 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
1810 if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
1811 device_printf(ctx->ifc_dev, "TX(%d) desc avail = %d, pidx = %d\n",
1812 txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
1814 IFDI_WATCHDOG_RESET(ctx);
1815 ctx->ifc_watchdog_events++;
1816 ctx->ifc_pause_frames = 0;
1818 iflib_init_locked(ctx);
1823 iflib_init_locked(if_ctx_t ctx)
1825 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
1826 if_t ifp = ctx->ifc_ifp;
1833 if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1834 IFDI_INTR_DISABLE(ctx);
1836 /* Set hardware offload abilities */
1837 if_clearhwassist(ifp);
1838 if (if_getcapenable(ifp) & IFCAP_TXCSUM)
1839 if_sethwassistbits(ifp, CSUM_IP | CSUM_TCP | CSUM_UDP, 0);
1840 if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
1841 if_sethwassistbits(ifp, (CSUM_TCP_IPV6 | CSUM_UDP_IPV6), 0);
1842 if (if_getcapenable(ifp) & IFCAP_TSO4)
1843 if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
1844 if (if_getcapenable(ifp) & IFCAP_TSO6)
1845 if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
1847 for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
1849 callout_stop(&txq->ift_timer);
1850 callout_stop(&txq->ift_db_check);
1851 CALLOUT_UNLOCK(txq);
1852 iflib_netmap_txq_init(ctx, txq);
1854 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1855 iflib_netmap_rxq_init(ctx, rxq);
1858 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
1859 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
1860 if (iflib_fl_setup(fl)) {
1861 device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
1867 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
1868 IFDI_INTR_ENABLE(ctx);
1869 txq = ctx->ifc_txqs;
1870 for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
1871 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
1872 txq->ift_timer.c_cpu);
1876 iflib_media_change(if_t ifp)
1878 if_ctx_t ctx = if_getsoftc(ifp);
1882 if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
1883 iflib_init_locked(ctx);
1889 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
1891 if_ctx_t ctx = if_getsoftc(ifp);
1894 IFDI_UPDATE_ADMIN_STATUS(ctx);
1895 IFDI_MEDIA_STATUS(ctx, ifmr);
1900 iflib_stop(if_ctx_t ctx)
1902 iflib_txq_t txq = ctx->ifc_txqs;
1903 iflib_rxq_t rxq = ctx->ifc_rxqs;
1904 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1905 if_shared_ctx_t sctx = ctx->ifc_sctx;
1906 iflib_dma_info_t di;
1910 /* Tell the stack that the interface is no longer active */
1911 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
1913 IFDI_INTR_DISABLE(ctx);
1914 msleep(ctx, &ctx->ifc_mtx, PUSER, "iflib_init", hz);
1916 /* Wait for current tx queue users to exit to disarm watchdog timer. */
1917 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
1918 /* make sure all transmitters have completed before proceeding XXX */
1920 /* clean any enqueued buffers */
1921 iflib_txq_check_drain(txq, 0);
1922 /* Free any existing tx buffers. */
1923 for (j = 0; j < sctx->isc_ntxd; j++) {
1924 iflib_txsd_free(ctx, txq, j);
1926 txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
1927 txq->ift_in_use = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
1928 txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
1929 txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
1930 txq->ift_pullups = 0;
1931 ifmp_ring_reset_stats(txq->ift_br[0]);
1932 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
1933 bzero((void *)di->idi_vaddr, di->idi_size);
1935 for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
1936 /* make sure all transmitters have completed before proceeding XXX */
1938 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
1939 bzero((void *)di->idi_vaddr, di->idi_size);
1940 /* also resets the free lists pidx/cidx */
1941 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
1942 iflib_fl_bufs_free(fl);
1948 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload)
1953 iflib_dma_info_t di;
1955 flid = irf->irf_flid;
1956 cidx = irf->irf_idx;
1957 fl = &rxq->ifr_fl[flid];
1960 fl->ifl_m_dequeued++;
1962 fl->ifl_cl_dequeued++;
1964 sd = &fl->ifl_sds[cidx];
1966 bus_dmamap_sync(di->idi_tag, di->idi_map,
1967 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1969 /* not valid assert if bxe really does SGE from non-contiguous elements */
1970 MPASS(fl->ifl_cidx == cidx);
1972 bus_dmamap_unload(fl->ifl_desc_tag, sd->ifsd_map);
1974 if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) {
1980 *cltype = fl->ifl_cltype;
1984 static struct mbuf *
1985 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri)
1987 int i, padlen , flags, cltype;
1988 struct mbuf *m, *mh, *mt;
1994 sd = rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE);
1996 MPASS(sd->ifsd_cl != NULL);
1997 MPASS(sd->ifsd_m != NULL);
2000 flags = M_PKTHDR|M_EXT;
2002 padlen = ri->iri_pad;
2007 /* assuming padding is only on the first fragment */
2014 /* Can these two be made one ? */
2015 m_init(m, M_NOWAIT, MT_DATA, flags);
2016 m_cljset(m, cl, cltype);
2018 * These must follow m_init and m_cljset
2020 m->m_data += padlen;
2021 ri->iri_len -= padlen;
2022 m->m_len = ri->iri_len;
2023 } while (++i < ri->iri_nfrags);
2031 * Process one software descriptor
2033 static struct mbuf *
2034 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2039 /* should I merge this back in now that the two paths are basically duplicated? */
2040 if (ri->iri_len <= IFLIB_RX_COPY_THRESH) {
2041 sd = rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE);
2044 m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2045 memcpy(m->m_data, sd->ifsd_cl, ri->iri_len);
2046 m->m_len = ri->iri_len;
2048 m = assemble_segments(rxq, ri);
2050 m->m_pkthdr.len = ri->iri_len;
2051 m->m_pkthdr.rcvif = ri->iri_ifp;
2052 m->m_flags |= ri->iri_flags;
2053 m->m_pkthdr.ether_vtag = ri->iri_vtag;
2054 m->m_pkthdr.flowid = ri->iri_flowid;
2055 M_HASHTYPE_SET(m, ri->iri_rsstype);
2056 m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2057 m->m_pkthdr.csum_data = ri->iri_csum_data;
2062 iflib_rxeof(iflib_rxq_t rxq, int budget)
2064 if_ctx_t ctx = rxq->ifr_ctx;
2065 if_shared_ctx_t sctx = ctx->ifc_sctx;
2068 struct if_rxd_info ri;
2069 int err, budget_left, rx_bytes, rx_pkts;
2072 struct lro_entry *queued;
2075 * XXX early demux data packets so that if_input processing only handles
2076 * acks in interrupt context
2078 struct mbuf *m, *mh, *mt;
2080 if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) {
2086 rx_pkts = rx_bytes = 0;
2087 if (sctx->isc_flags & IFLIB_HAS_CQ)
2088 cidxp = &rxq->ifr_cq_cidx;
2090 cidxp = &rxq->ifr_fl[0].ifl_cidx;
2091 if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp)) == 0) {
2092 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2093 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2094 DBG_COUNTER_INC(rx_unavail);
2098 for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2099 if (__predict_false(!CTX_ACTIVE(ctx))) {
2100 DBG_COUNTER_INC(rx_ctx_inactive);
2104 * Reset client set fields to their default values
2106 bzero(&ri, sizeof(ri));
2107 ri.iri_qsidx = rxq->ifr_id;
2108 ri.iri_cidx = *cidxp;
2109 ri.iri_ifp = ctx->ifc_ifp;
2110 ri.iri_frags = rxq->ifr_frags;
2111 err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2113 /* in lieu of handling correctly - make sure it isn't being unhandled */
2115 if (sctx->isc_flags & IFLIB_HAS_CQ) {
2116 /* we know we consumed _one_ CQ entry */
2117 if (++rxq->ifr_cq_cidx == sctx->isc_nrxd) {
2118 rxq->ifr_cq_cidx = 0;
2119 rxq->ifr_cq_gen = 0;
2121 /* was this only a completion queue message? */
2122 if (__predict_false(ri.iri_nfrags == 0))
2125 MPASS(ri.iri_nfrags != 0);
2126 MPASS(ri.iri_len != 0);
2128 /* will advance the cidx on the corresponding free lists */
2129 m = iflib_rxd_pkt_get(rxq, &ri);
2130 if (avail == 0 && budget_left)
2131 avail = iflib_rxd_avail(ctx, rxq, *cidxp);
2133 if (__predict_false(m == NULL)) {
2134 DBG_COUNTER_INC(rx_mbuf_null);
2137 /* imm_pkt: -- cxgb */
2145 /* make sure that we can refill faster than drain */
2146 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2147 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2150 lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2152 while (mh != NULL) {
2155 m->m_nextpkt = NULL;
2156 rx_bytes += m->m_pkthdr.len;
2158 #if defined(INET6) || defined(INET)
2159 if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2162 DBG_COUNTER_INC(rx_if_input);
2163 ifp->if_input(ifp, m);
2165 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2166 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2169 * Flush any outstanding LRO work
2171 while ((queued = LIST_FIRST(&rxq->ifr_lc.lro_active)) != NULL) {
2172 LIST_REMOVE(queued, next);
2173 #if defined(INET6) || defined(INET)
2174 tcp_lro_flush(&rxq->ifr_lc, queued);
2177 return (iflib_rxd_avail(ctx, rxq, *cidxp));
2180 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2181 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2182 #define TXQ_MAX_DB_DEFERRED(ctx) (ctx->ifc_sctx->isc_ntxd >> 5)
2183 #define TXQ_MAX_DB_CONSUMED(ctx) (ctx->ifc_sctx->isc_ntxd >> 4)
2185 static __inline void
2186 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring)
2190 if (ring || txq->ift_db_pending >= TXQ_MAX_DB_DEFERRED(ctx)) {
2192 /* the lock will only ever be contended in the !min_latency case */
2193 if (!TXDB_TRYLOCK(txq))
2195 dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2196 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2197 txq->ift_db_pending = txq->ift_npending = 0;
2203 iflib_txd_deferred_db_check(void * arg)
2205 iflib_txq_t txq = arg;
2207 /* simple non-zero boolean so use bitwise OR */
2208 if ((txq->ift_db_pending | txq->ift_npending) &&
2209 txq->ift_db_pending >= txq->ift_db_pending_queued)
2210 iflib_txd_db_check(txq->ift_ctx, txq, TRUE);
2211 txq->ift_db_pending_queued = 0;
2212 if (ifmp_ring_is_stalled(txq->ift_br[0]))
2213 iflib_txq_check_drain(txq, 4);
2218 print_pkt(if_pkt_info_t pi)
2220 printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2221 pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2222 printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2223 pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2224 printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2225 pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2229 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2230 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2233 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2235 struct ether_vlan_header *eh;
2240 * Determine where frame payload starts.
2241 * Jump over vlan headers if already present,
2242 * helpful for QinQ too.
2244 if (__predict_false(m->m_len < sizeof(*eh))) {
2246 if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2249 eh = mtod(m, struct ether_vlan_header *);
2250 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2251 pi->ipi_etype = ntohs(eh->evl_proto);
2252 pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2254 pi->ipi_etype = ntohs(eh->evl_encap_proto);
2255 pi->ipi_ehdrlen = ETHER_HDR_LEN;
2258 switch (pi->ipi_etype) {
2262 struct ip *ip = NULL;
2263 struct tcphdr *th = NULL;
2267 minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2268 if (__predict_false(m->m_len < minthlen)) {
2270 * if this code bloat is causing too much of a hit
2271 * move it to a separate function and mark it noinline
2273 if (m->m_len == pi->ipi_ehdrlen) {
2276 if (n->m_len >= sizeof(*ip)) {
2277 ip = (struct ip *)n->m_data;
2278 if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2279 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2282 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2284 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2288 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2290 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2291 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2292 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2295 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2296 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2297 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2299 pi->ipi_ip_hlen = ip->ip_hl << 2;
2300 pi->ipi_ipproto = ip->ip_p;
2301 pi->ipi_flags |= IPI_TX_IPV4;
2303 if (pi->ipi_csum_flags & CSUM_IP)
2306 if (pi->ipi_ipproto == IPPROTO_TCP) {
2307 if (__predict_false(th == NULL)) {
2309 if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2311 th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2313 pi->ipi_tcp_hflags = th->th_flags;
2314 pi->ipi_tcp_hlen = th->th_off << 2;
2315 pi->ipi_tcp_seq = th->th_seq;
2318 if (__predict_false(ip->ip_p != IPPROTO_TCP))
2320 th->th_sum = in_pseudo(ip->ip_src.s_addr,
2321 ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2322 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2328 case ETHERTYPE_IPV6:
2330 struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2332 pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2334 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2335 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2338 th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2340 /* XXX-BZ this will go badly in case of ext hdrs. */
2341 pi->ipi_ipproto = ip6->ip6_nxt;
2342 pi->ipi_flags |= IPI_TX_IPV6;
2344 if (pi->ipi_ipproto == IPPROTO_TCP) {
2345 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2346 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2349 pi->ipi_tcp_hflags = th->th_flags;
2350 pi->ipi_tcp_hlen = th->th_off << 2;
2354 if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2357 * The corresponding flag is set by the stack in the IPv4
2358 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2359 * So, set it here because the rest of the flow requires it.
2361 pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2362 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2363 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2369 pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2370 pi->ipi_ip_hlen = 0;
2378 static __noinline struct mbuf *
2379 collapse_pkthdr(struct mbuf *m0)
2381 struct mbuf *m, *m_next, *tmp;
2385 while (m_next != NULL && m_next->m_len == 0) {
2389 m_next = m_next->m_next;
2393 if ((m_next->m_flags & M_EXT) == 0) {
2394 m = m_defrag(m, M_NOWAIT);
2396 tmp = m_next->m_next;
2397 memcpy(m_next, m, MPKTHSIZE);
2405 * If dodgy hardware rejects the scatter gather chain we've handed it
2406 * we'll need to rebuild the mbuf chain before we can call m_defrag
2408 static __noinline struct mbuf *
2409 iflib_rebuild_mbuf(iflib_txq_t txq)
2412 int ntxd, mhlen, len, i, pidx;
2413 struct mbuf *m, *mh, **ifsd_m;
2414 if_shared_ctx_t sctx;
2416 pidx = txq->ift_pidx;
2417 ifsd_m = txq->ift_sds.ifsd_m;
2418 sctx = txq->ift_ctx->ifc_sctx;
2419 ntxd = sctx->isc_ntxd;
2420 mh = m = ifsd_m[pidx];
2421 ifsd_m[pidx] = NULL;
2423 txq->ift_dequeued++;
2426 mhlen = m->m_pkthdr.len;
2429 while (len < mhlen && (m->m_next == NULL)) {
2430 m->m_next = ifsd_m[(pidx + i) & (ntxd-1)];
2431 ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2433 txq->ift_dequeued++;
2443 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2444 struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2445 int max_segs, int flags)
2448 if_shared_ctx_t sctx;
2449 int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2450 struct mbuf *m, *tmp, **ifsd_m, **mp;
2455 * Please don't ever do this
2457 if (__predict_false(m->m_len == 0))
2458 *m0 = m = collapse_pkthdr(m);
2461 sctx = ctx->ifc_sctx;
2462 ifsd_m = txq->ift_sds.ifsd_m;
2463 ntxd = sctx->isc_ntxd;
2464 pidx = txq->ift_pidx;
2466 uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2468 err = bus_dmamap_load_mbuf_sg(tag, map,
2469 *m0, segs, nsegs, BUS_DMA_NOWAIT);
2472 ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2475 mask = (sctx->isc_ntxd-1);
2481 (*mp)->m_next = NULL;
2482 if (__predict_false((*mp)->m_len == 0)) {
2486 next = (pidx + i) & (ntxd-1);
2487 } while (m != NULL);
2489 int buflen, sgsize, max_sgsize;
2494 maxsegsz = sctx->isc_tx_maxsize;
2497 if (__predict_false(m->m_len <= 0)) {
2505 vaddr = (vm_offset_t)m->m_data;
2507 * see if we can't be smarter about physically
2508 * contiguous mappings
2510 next = (pidx + count) & (ntxd-1);
2511 MPASS(ifsd_m[next] == NULL);
2513 txq->ift_enqueued++;
2516 while (buflen > 0) {
2517 max_sgsize = MIN(buflen, maxsegsz);
2518 curaddr = pmap_kextract(vaddr);
2519 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2520 sgsize = MIN(sgsize, max_sgsize);
2521 segs[i].ds_addr = curaddr;
2522 segs[i].ds_len = sgsize;
2533 } while (m != NULL);
2538 *m0 = iflib_rebuild_mbuf(txq);
2543 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2546 if_shared_ctx_t sctx;
2547 if_softc_ctx_t scctx;
2548 bus_dma_segment_t *segs;
2549 struct mbuf *m_head;
2551 struct if_pkt_info pi;
2553 int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
2554 bus_dma_tag_t desc_tag;
2556 segs = txq->ift_segs;
2558 sctx = ctx->ifc_sctx;
2559 scctx = &ctx->ifc_softc_ctx;
2560 segs = txq->ift_segs;
2561 ntxd = sctx->isc_ntxd;
2566 * If we're doing TSO the next descriptor to clean may be quite far ahead
2568 cidx = txq->ift_cidx;
2569 pidx = txq->ift_pidx;
2570 next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
2572 /* prefetch the next cache line of mbuf pointers and flags */
2573 prefetch(&txq->ift_sds.ifsd_m[next]);
2574 if (txq->ift_sds.ifsd_map != NULL) {
2575 prefetch(&txq->ift_sds.ifsd_map[next]);
2576 map = txq->ift_sds.ifsd_map[pidx];
2577 next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
2578 prefetch(&txq->ift_sds.ifsd_flags[next]);
2582 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
2583 desc_tag = txq->ift_tso_desc_tag;
2584 max_segs = scctx->isc_tx_tso_segments_max;
2586 desc_tag = txq->ift_desc_tag;
2587 max_segs = scctx->isc_tx_nsegments;
2590 bzero(&pi, sizeof(pi));
2591 pi.ipi_len = m_head->m_pkthdr.len;
2592 pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
2593 pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
2594 pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
2596 pi.ipi_qsidx = txq->ift_id;
2598 /* deliberate bitwise OR to make one condition */
2599 if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
2600 if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
2606 err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
2608 if (__predict_false(err)) {
2611 /* try collapse once and defrag once */
2613 m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
2615 m_head = m_defrag(*m_headp, M_NOWAIT);
2617 if (__predict_false(m_head == NULL))
2619 txq->ift_mbuf_defrag++;
2624 txq->ift_no_tx_dma_setup++;
2627 txq->ift_no_tx_dma_setup++;
2629 DBG_COUNTER_INC(tx_frees);
2633 txq->ift_map_failed++;
2634 DBG_COUNTER_INC(encap_load_mbuf_fail);
2639 * XXX assumes a 1 to 1 relationship between segments and
2640 * descriptors - this does not hold true on all drivers, e.g.
2643 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
2644 txq->ift_no_desc_avail++;
2646 bus_dmamap_unload(desc_tag, map);
2647 DBG_COUNTER_INC(encap_txq_avail_fail);
2648 if (txq->ift_task.gt_task.ta_pending == 0)
2649 GROUPTASK_ENQUEUE(&txq->ift_task);
2653 pi.ipi_nsegs = nsegs;
2655 MPASS(pidx >= 0 && pidx < sctx->isc_ntxd);
2659 if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
2660 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2661 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2663 DBG_COUNTER_INC(tx_encap);
2664 MPASS(pi.ipi_new_pidx >= 0 && pi.ipi_new_pidx < sctx->isc_ntxd);
2666 ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
2667 if (pi.ipi_new_pidx < pi.ipi_pidx) {
2668 ndesc += sctx->isc_ntxd;
2671 MPASS(pi.ipi_new_pidx != pidx);
2673 txq->ift_in_use += ndesc;
2675 * We update the last software descriptor again here because there may
2676 * be a sentinel and/or there may be more mbufs than segments
2678 txq->ift_pidx = pi.ipi_new_pidx;
2679 txq->ift_npending += pi.ipi_ndescs;
2680 } else if (__predict_false(err == EFBIG && remap < 2)) {
2681 *m_headp = m_head = iflib_rebuild_mbuf(txq);
2683 txq->ift_txd_encap_efbig++;
2686 DBG_COUNTER_INC(encap_txd_encap_fail);
2690 txq->ift_mbuf_defrag_failed++;
2691 txq->ift_map_failed++;
2693 DBG_COUNTER_INC(tx_frees);
2698 /* forward compatibility for cxgb */
2699 #define FIRST_QSET(ctx) 0
2701 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2702 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2703 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NRXQSETS(ctx)) + FIRST_QSET(ctx))
2704 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2705 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2706 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2710 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring
2713 * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic
2715 #define TXQ_MIN_OCCUPANCY(ctx) ((ctx->ifc_sctx->isc_ntxd >> 6)| 0x2)
2718 iflib_txq_min_occupancy(iflib_txq_t txq)
2723 return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen) < TXQ_MIN_OCCUPANCY(ctx) + MAX_TX_DESC(ctx));
2727 iflib_tx_desc_free(iflib_txq_t txq, int n)
2730 uint32_t qsize, cidx, mask, gen;
2731 struct mbuf *m, **ifsd_m;
2732 uint8_t *ifsd_flags;
2733 bus_dmamap_t *ifsd_map;
2735 cidx = txq->ift_cidx;
2737 qsize = txq->ift_ctx->ifc_sctx->isc_ntxd;
2739 hasmap = txq->ift_sds.ifsd_map != NULL;
2740 ifsd_flags = txq->ift_sds.ifsd_flags;
2741 ifsd_m = txq->ift_sds.ifsd_m;
2742 ifsd_map = txq->ift_sds.ifsd_map;
2745 prefetch(ifsd_m[(cidx + 3) & mask]);
2746 prefetch(ifsd_m[(cidx + 4) & mask]);
2748 if (ifsd_m[cidx] != NULL) {
2749 prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
2750 prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
2751 if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
2753 * does it matter if it's not the TSO tag? If so we'll
2754 * have to add the type to flags
2756 bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
2757 ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
2759 if ((m = ifsd_m[cidx]) != NULL) {
2760 /* XXX we don't support any drivers that batch packets yet */
2761 MPASS(m->m_nextpkt == NULL);
2764 ifsd_m[cidx] = NULL;
2766 txq->ift_dequeued++;
2768 DBG_COUNTER_INC(tx_frees);
2771 if (__predict_false(++cidx == qsize)) {
2776 txq->ift_cidx = cidx;
2781 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
2784 if_ctx_t ctx = txq->ift_ctx;
2786 KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
2787 MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
2790 * Need a rate-limiting check so that this isn't called every time
2792 iflib_tx_credits_update(ctx, txq);
2793 reclaim = DESC_RECLAIMABLE(txq);
2795 if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
2797 if (iflib_verbose_debug) {
2798 printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
2799 txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
2806 iflib_tx_desc_free(txq, reclaim);
2807 txq->ift_cleaned += reclaim;
2808 txq->ift_in_use -= reclaim;
2810 if (txq->ift_active == FALSE)
2811 txq->ift_active = TRUE;
2816 static struct mbuf **
2817 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset)
2820 return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)]));
2824 iflib_txq_check_drain(iflib_txq_t txq, int budget)
2827 ifmp_ring_check_drainage(txq->ift_br[0], budget);
2831 iflib_txq_can_drain(struct ifmp_ring *r)
2833 iflib_txq_t txq = r->cookie;
2834 if_ctx_t ctx = txq->ift_ctx;
2836 return ((TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx)) ||
2837 ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false));
2841 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
2843 iflib_txq_t txq = r->cookie;
2844 if_ctx_t ctx = txq->ift_ctx;
2845 if_t ifp = ctx->ifc_ifp;
2846 struct mbuf **mp, *m;
2847 int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used;
2849 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
2850 !LINK_ACTIVE(ctx))) {
2851 DBG_COUNTER_INC(txq_drain_notready);
2855 avail = IDXDIFF(pidx, cidx, r->size);
2856 if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
2857 DBG_COUNTER_INC(txq_drain_flushing);
2858 for (i = 0; i < avail; i++) {
2859 m_freem(r->items[(cidx + i) & (r->size-1)]);
2860 r->items[(cidx + i) & (r->size-1)] = NULL;
2864 iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
2865 if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
2866 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
2868 callout_stop(&txq->ift_timer);
2869 callout_stop(&txq->ift_db_check);
2870 CALLOUT_UNLOCK(txq);
2871 DBG_COUNTER_INC(txq_drain_oactive);
2874 consumed = mcast_sent = bytes_sent = pkt_sent = 0;
2875 count = MIN(avail, TX_BATCH_SIZE);
2877 for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) {
2878 mp = _ring_peek_one(r, cidx, i);
2879 in_use_prev = txq->ift_in_use;
2880 err = iflib_encap(txq, mp);
2882 * What other errors should we bail out for?
2884 if (err == ENOBUFS) {
2885 DBG_COUNTER_INC(txq_drain_encapfail);
2894 DBG_COUNTER_INC(tx_sent);
2895 bytes_sent += m->m_pkthdr.len;
2896 if (m->m_flags & M_MCAST)
2899 txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
2900 desc_used += (txq->ift_in_use - in_use_prev);
2901 iflib_txd_db_check(ctx, txq, FALSE);
2902 ETHER_BPF_MTAP(ifp, m);
2903 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
2906 if (desc_used > TXQ_MAX_DB_CONSUMED(ctx))
2910 if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending)
2911 iflib_txd_db_check(ctx, txq, TRUE);
2912 else if ((txq->ift_db_pending || TXQ_AVAIL(txq) < MAX_TX_DESC(ctx)) &&
2913 (callout_pending(&txq->ift_db_check) == 0)) {
2914 txq->ift_db_pending_queued = txq->ift_db_pending;
2915 callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check,
2916 txq, txq->ift_db_check.c_cpu);
2918 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
2919 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
2921 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
2927 _task_fn_tx(void *context, int pending)
2929 iflib_txq_t txq = context;
2930 if_ctx_t ctx = txq->ift_ctx;
2932 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2934 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
2938 _task_fn_rx(void *context, int pending)
2940 iflib_rxq_t rxq = context;
2941 if_ctx_t ctx = rxq->ifr_ctx;
2944 DBG_COUNTER_INC(task_fn_rxs);
2945 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
2948 if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
2949 if (ctx->ifc_flags & IFC_LEGACY)
2950 IFDI_INTR_ENABLE(ctx);
2952 DBG_COUNTER_INC(rx_intr_enables);
2953 IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
2956 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
2959 GROUPTASK_ENQUEUE(&rxq->ifr_task);
2963 _task_fn_admin(void *context, int pending)
2965 if_ctx_t ctx = context;
2966 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2970 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2974 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
2976 callout_stop(&txq->ift_timer);
2977 CALLOUT_UNLOCK(txq);
2979 IFDI_UPDATE_ADMIN_STATUS(ctx);
2980 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2981 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2982 IFDI_LINK_INTR_ENABLE(ctx);
2985 if (LINK_ACTIVE(ctx) == 0)
2987 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2988 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
2993 _task_fn_iov(void *context, int pending)
2995 if_ctx_t ctx = context;
2997 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3001 IFDI_VFLR_HANDLE(ctx);
3006 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3009 if_int_delay_info_t info;
3012 info = (if_int_delay_info_t)arg1;
3013 ctx = info->iidi_ctx;
3014 info->iidi_req = req;
3015 info->iidi_oidp = oidp;
3017 err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3022 /*********************************************************************
3026 **********************************************************************/
3029 iflib_if_init_locked(if_ctx_t ctx)
3032 iflib_init_locked(ctx);
3037 iflib_if_init(void *arg)
3042 iflib_if_init_locked(ctx);
3047 iflib_if_transmit(if_t ifp, struct mbuf *m)
3049 if_ctx_t ctx = if_getsoftc(ifp);
3052 struct mbuf *marr[8], **mp, *next;
3053 int err, i, count, qidx;
3055 if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3056 DBG_COUNTER_INC(tx_frees);
3062 if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3063 qidx = QIDX(ctx, m);
3065 * XXX calculate buf_ring based on flowid (divvy up bits?)
3067 txq = &ctx->ifc_txqs[qidx];
3069 #ifdef DRIVER_BACKPRESSURE
3070 if (txq->ift_closed) {
3072 next = m->m_nextpkt;
3073 m->m_nextpkt = NULL;
3085 next = next->m_nextpkt;
3086 } while (next != NULL);
3089 if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3090 /* XXX check nextpkt */
3092 /* XXX simplify for now */
3093 DBG_COUNTER_INC(tx_frees);
3096 for (next = m, i = 0; next != NULL; i++) {
3098 next = next->m_nextpkt;
3099 mp[i]->m_nextpkt = NULL;
3101 DBG_COUNTER_INC(tx_seen);
3102 err = ifmp_ring_enqueue(txq->ift_br[0], (void **)mp, count, TX_BATCH_SIZE);
3104 if (iflib_txq_can_drain(txq->ift_br[0]))
3105 GROUPTASK_ENQUEUE(&txq->ift_task);
3107 /* support forthcoming later */
3108 #ifdef DRIVER_BACKPRESSURE
3109 txq->ift_closed = TRUE;
3111 for (i = 0; i < count; i++)
3113 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE);
3122 iflib_if_qflush(if_t ifp)
3124 if_ctx_t ctx = if_getsoftc(ifp);
3125 iflib_txq_t txq = ctx->ifc_txqs;
3129 ctx->ifc_flags |= IFC_QFLUSH;
3131 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3132 while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0])))
3133 iflib_txq_check_drain(txq, 0);
3135 ctx->ifc_flags &= ~IFC_QFLUSH;
3141 #define IFCAP_REINIT (IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_TSO6|IFCAP_VLAN_HWTAGGING|IFCAP_VLAN_MTU | \
3142 IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3144 #define IFCAP_FLAGS (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3145 IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | \
3146 IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3149 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3151 if_ctx_t ctx = if_getsoftc(ifp);
3152 struct ifreq *ifr = (struct ifreq *)data;
3153 #if defined(INET) || defined(INET6)
3154 struct ifaddr *ifa = (struct ifaddr *)data;
3156 bool avoid_reset = FALSE;
3157 int err = 0, reinit = 0, bits;
3162 if (ifa->ifa_addr->sa_family == AF_INET)
3166 if (ifa->ifa_addr->sa_family == AF_INET6)
3170 ** Calling init results in link renegotiation,
3171 ** so we avoid doing it when possible.
3174 if_setflagbits(ifp, IFF_UP,0);
3175 if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3178 if (!(if_getflags(ifp) & IFF_NOARP))
3179 arp_ifinit(ifp, ifa);
3182 err = ether_ioctl(ifp, command, data);
3186 if (ifr->ifr_mtu == if_getmtu(ifp)) {
3190 bits = if_getdrvflags(ifp);
3191 /* stop the driver and free any clusters before proceeding */
3194 if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3195 if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3196 ctx->ifc_flags |= IFC_MULTISEG;
3198 ctx->ifc_flags &= ~IFC_MULTISEG;
3199 err = if_setmtu(ifp, ifr->ifr_mtu);
3201 iflib_init_locked(ctx);
3202 if_setdrvflags(ifp, bits);
3207 if (if_getflags(ifp) & IFF_UP) {
3208 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3209 if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3210 (IFF_PROMISC | IFF_ALLMULTI)) {
3211 err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3215 } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3218 ctx->ifc_if_flags = if_getflags(ifp);
3225 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3227 IFDI_INTR_DISABLE(ctx);
3228 IFDI_MULTI_SET(ctx);
3229 IFDI_INTR_ENABLE(ctx);
3235 IFDI_MEDIA_SET(ctx);
3239 err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3243 struct ifi2creq i2c;
3245 err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3248 if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3252 if (i2c.len > sizeof(i2c.data)) {
3257 if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3258 err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3265 mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3268 setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3270 setmask |= (mask & IFCAP_FLAGS);
3272 if ((mask & IFCAP_WOL) &&
3273 (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3274 setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3277 * want to ensure that traffic has stopped before we change any of the flags
3281 bits = if_getdrvflags(ifp);
3282 if (setmask & IFCAP_REINIT)
3284 if_togglecapenable(ifp, setmask);
3285 if (setmask & IFCAP_REINIT)
3286 iflib_init_locked(ctx);
3287 if_setdrvflags(ifp, bits);
3292 case SIOCGPRIVATE_0:
3296 err = IFDI_PRIV_IOCTL(ctx, command, data);
3300 err = ether_ioctl(ifp, command, data);
3309 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3311 if_ctx_t ctx = if_getsoftc(ifp);
3313 return (IFDI_GET_COUNTER(ctx, cnt));
3316 /*********************************************************************
3318 * OTHER FUNCTIONS EXPORTED TO THE STACK
3320 **********************************************************************/
3323 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3325 if_ctx_t ctx = if_getsoftc(ifp);
3327 if ((void *)ctx != arg)
3330 if ((vtag == 0) || (vtag > 4095))
3334 IFDI_VLAN_REGISTER(ctx, vtag);
3335 /* Re-init to load the changes */
3336 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3337 iflib_init_locked(ctx);
3342 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3344 if_ctx_t ctx = if_getsoftc(ifp);
3346 if ((void *)ctx != arg)
3349 if ((vtag == 0) || (vtag > 4095))
3353 IFDI_VLAN_UNREGISTER(ctx, vtag);
3354 /* Re-init to load the changes */
3355 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3356 iflib_init_locked(ctx);
3361 iflib_led_func(void *arg, int onoff)
3366 IFDI_LED_FUNC(ctx, onoff);
3370 /*********************************************************************
3372 * BUS FUNCTION DEFINITIONS
3374 **********************************************************************/
3377 iflib_device_probe(device_t dev)
3379 pci_vendor_info_t *ent;
3381 uint16_t pci_vendor_id, pci_device_id;
3382 uint16_t pci_subvendor_id, pci_subdevice_id;
3383 uint16_t pci_rev_id;
3384 if_shared_ctx_t sctx;
3386 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3389 pci_vendor_id = pci_get_vendor(dev);
3390 pci_device_id = pci_get_device(dev);
3391 pci_subvendor_id = pci_get_subvendor(dev);
3392 pci_subdevice_id = pci_get_subdevice(dev);
3393 pci_rev_id = pci_get_revid(dev);
3394 if (sctx->isc_parse_devinfo != NULL)
3395 sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3397 ent = sctx->isc_vendor_info;
3398 while (ent->pvi_vendor_id != 0) {
3399 if (pci_vendor_id != ent->pvi_vendor_id) {
3403 if ((pci_device_id == ent->pvi_device_id) &&
3404 ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3405 (ent->pvi_subvendor_id == 0)) &&
3406 ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3407 (ent->pvi_subdevice_id == 0)) &&
3408 ((pci_rev_id == ent->pvi_rev_id) ||
3409 (ent->pvi_rev_id == 0))) {
3411 device_set_desc_copy(dev, ent->pvi_name);
3412 /* this needs to be changed to zero if the bus probing code
3413 * ever stops re-probing on best match because the sctx
3414 * may have its values over written by register calls
3415 * in subsequent probes
3417 return (BUS_PROBE_DEFAULT);
3425 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3427 int err, rid, msix, msix_bar;
3430 if_softc_ctx_t scctx;
3433 ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3436 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3437 device_set_softc(dev, ctx);
3440 ctx->ifc_sctx = sctx;
3442 ctx->ifc_txrx = *sctx->isc_txrx;
3443 ctx->ifc_softc = sc;
3445 if ((err = iflib_register(ctx)) != 0) {
3446 device_printf(dev, "iflib_register failed %d\n", err);
3449 iflib_add_device_sysctl_pre(ctx);
3450 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
3451 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
3455 if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
3456 ctx->ifc_flags |= IFC_DMAR;
3459 scctx = &ctx->ifc_softc_ctx;
3460 msix_bar = scctx->isc_msix_bar;
3462 if (scctx->isc_tx_nsegments > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION)
3463 scctx->isc_tx_nsegments = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION);
3464 if (scctx->isc_tx_tso_segments_max > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION)
3465 scctx->isc_tx_tso_segments_max = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION);
3470 * XXX sanity check that ntxd & nrxd are a power of 2
3474 * Protect the stack against modern hardware
3476 if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
3477 scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
3479 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
3480 ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
3481 ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
3482 ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
3483 if (scctx->isc_rss_table_size == 0)
3484 scctx->isc_rss_table_size = 64;
3485 scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;;
3487 ** Now setup MSI or MSI/X, should
3488 ** return us the number of supported
3489 ** vectors. (Will be 1 for MSI)
3491 if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
3492 msix = scctx->isc_vectors;
3493 } else if (scctx->isc_msix_bar != 0)
3494 msix = iflib_msix_init(ctx);
3496 scctx->isc_vectors = 1;
3497 scctx->isc_ntxqsets = 1;
3498 scctx->isc_nrxqsets = 1;
3499 scctx->isc_intr = IFLIB_INTR_LEGACY;
3502 /* Get memory for the station queues */
3503 if ((err = iflib_queues_alloc(ctx))) {
3504 device_printf(dev, "Unable to allocate queue memory\n");
3508 if ((err = iflib_qset_structures_setup(ctx))) {
3509 device_printf(dev, "qset structure setup failed %d\n", err);
3513 if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
3514 device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
3515 goto fail_intr_free;
3519 if (scctx->isc_intr == IFLIB_INTR_MSI) {
3523 if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx, &rid, "irq0")) != 0) {
3524 device_printf(dev, "iflib_legacy_setup failed %d\n", err);
3525 goto fail_intr_free;
3528 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
3529 if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
3530 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
3533 if ((err = iflib_netmap_attach(ctx))) {
3534 device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
3539 iflib_add_device_sysctl_post(ctx);
3542 ether_ifdetach(ctx->ifc_ifp);
3544 if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
3545 pci_release_msi(ctx->ifc_dev);
3547 /* XXX free queues */
3554 iflib_device_attach(device_t dev)
3557 if_shared_ctx_t sctx;
3559 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3562 pci_enable_busmaster(dev);
3564 return (iflib_device_register(dev, NULL, sctx, &ctx));
3568 iflib_device_deregister(if_ctx_t ctx)
3570 if_t ifp = ctx->ifc_ifp;
3573 device_t dev = ctx->ifc_dev;
3575 struct taskqgroup *tqg;
3577 /* Make sure VLANS are not using driver */
3578 if (if_vlantrunkinuse(ifp)) {
3579 device_printf(dev,"Vlan in use, detach first\n");
3584 ctx->ifc_in_detach = 1;
3588 /* Unregister VLAN events */
3589 if (ctx->ifc_vlan_attach_event != NULL)
3590 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
3591 if (ctx->ifc_vlan_detach_event != NULL)
3592 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
3594 iflib_netmap_detach(ifp);
3595 ether_ifdetach(ifp);
3596 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
3597 CTX_LOCK_DESTROY(ctx);
3598 if (ctx->ifc_led_dev != NULL)
3599 led_destroy(ctx->ifc_led_dev);
3600 /* XXX drain any dependent tasks */
3601 tqg = qgroup_if_io_tqg;
3602 for (txq = ctx->ifc_txqs, i = 0, rxq = ctx->ifc_rxqs; i < NTXQSETS(ctx); i++, txq++) {
3603 callout_drain(&txq->ift_timer);
3604 callout_drain(&txq->ift_db_check);
3605 if (txq->ift_task.gt_uniq != NULL)
3606 taskqgroup_detach(tqg, &txq->ift_task);
3608 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
3609 if (rxq->ifr_task.gt_uniq != NULL)
3610 taskqgroup_detach(tqg, &rxq->ifr_task);
3612 tqg = qgroup_if_config_tqg;
3613 if (ctx->ifc_admin_task.gt_uniq != NULL)
3614 taskqgroup_detach(tqg, &ctx->ifc_admin_task);
3615 if (ctx->ifc_vflr_task.gt_uniq != NULL)
3616 taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
3619 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
3620 pci_release_msi(dev);
3622 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
3623 iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
3625 if (ctx->ifc_msix_mem != NULL) {
3626 bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
3627 ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
3628 ctx->ifc_msix_mem = NULL;
3631 bus_generic_detach(dev);
3634 iflib_tx_structures_free(ctx);
3635 iflib_rx_structures_free(ctx);
3641 iflib_device_detach(device_t dev)
3643 if_ctx_t ctx = device_get_softc(dev);
3645 return (iflib_device_deregister(ctx));
3649 iflib_device_suspend(device_t dev)
3651 if_ctx_t ctx = device_get_softc(dev);
3657 return bus_generic_suspend(dev);
3660 iflib_device_shutdown(device_t dev)
3662 if_ctx_t ctx = device_get_softc(dev);
3668 return bus_generic_suspend(dev);
3673 iflib_device_resume(device_t dev)
3675 if_ctx_t ctx = device_get_softc(dev);
3676 iflib_txq_t txq = ctx->ifc_txqs;
3680 iflib_init_locked(ctx);
3682 for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
3683 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3685 return (bus_generic_resume(dev));
3689 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
3692 if_ctx_t ctx = device_get_softc(dev);
3695 error = IFDI_IOV_INIT(ctx, num_vfs, params);
3702 iflib_device_iov_uninit(device_t dev)
3704 if_ctx_t ctx = device_get_softc(dev);
3707 IFDI_IOV_UNINIT(ctx);
3712 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
3715 if_ctx_t ctx = device_get_softc(dev);
3718 error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
3724 /*********************************************************************
3726 * MODULE FUNCTION DEFINITIONS
3728 **********************************************************************/
3731 * - Start a fast taskqueue thread for each core
3732 * - Start a taskqueue for control operations
3735 iflib_module_init(void)
3741 iflib_module_event_handler(module_t mod, int what, void *arg)
3747 if ((err = iflib_module_init()) != 0)
3753 return (EOPNOTSUPP);
3759 /*********************************************************************
3761 * PUBLIC FUNCTION DEFINITIONS
3762 * ordered as in iflib.h
3764 **********************************************************************/
3768 _iflib_assert(if_shared_ctx_t sctx)
3770 MPASS(sctx->isc_tx_maxsize);
3771 MPASS(sctx->isc_tx_maxsegsize);
3773 MPASS(sctx->isc_rx_maxsize);
3774 MPASS(sctx->isc_rx_nsegments);
3775 MPASS(sctx->isc_rx_maxsegsize);
3778 MPASS(sctx->isc_txrx->ift_txd_encap);
3779 MPASS(sctx->isc_txrx->ift_txd_flush);
3780 MPASS(sctx->isc_txrx->ift_txd_credits_update);
3781 MPASS(sctx->isc_txrx->ift_rxd_available);
3782 MPASS(sctx->isc_txrx->ift_rxd_pkt_get);
3783 MPASS(sctx->isc_txrx->ift_rxd_refill);
3784 MPASS(sctx->isc_txrx->ift_rxd_flush);
3785 MPASS(sctx->isc_nrxd);
3789 iflib_register(if_ctx_t ctx)
3791 if_shared_ctx_t sctx = ctx->ifc_sctx;
3792 driver_t *driver = sctx->isc_driver;
3793 device_t dev = ctx->ifc_dev;
3796 _iflib_assert(sctx);
3798 CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
3799 MPASS(ctx->ifc_flags == 0);
3801 ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
3803 device_printf(dev, "can not allocate ifnet structure\n");
3808 * Initialize our context's device specific methods
3810 kobj_init((kobj_t) ctx, (kobj_class_t) driver);
3811 kobj_class_compile((kobj_class_t) driver);
3814 if_initname(ifp, device_get_name(dev), device_get_unit(dev));
3815 if_setsoftc(ifp, ctx);
3816 if_setdev(ifp, dev);
3817 if_setinitfn(ifp, iflib_if_init);
3818 if_setioctlfn(ifp, iflib_if_ioctl);
3819 if_settransmitfn(ifp, iflib_if_transmit);
3820 if_setqflushfn(ifp, iflib_if_qflush);
3821 if_setgetcounterfn(ifp, iflib_if_get_counter);
3822 if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
3824 if_setcapabilities(ifp, 0);
3825 if_setcapenable(ifp, 0);
3827 ctx->ifc_vlan_attach_event =
3828 EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
3829 EVENTHANDLER_PRI_FIRST);
3830 ctx->ifc_vlan_detach_event =
3831 EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
3832 EVENTHANDLER_PRI_FIRST);
3834 ifmedia_init(&ctx->ifc_media, IFM_IMASK,
3835 iflib_media_change, iflib_media_status);
3842 iflib_queues_alloc(if_ctx_t ctx)
3844 if_shared_ctx_t sctx = ctx->ifc_sctx;
3845 device_t dev = ctx->ifc_dev;
3846 int nrxqsets = ctx->ifc_softc_ctx.isc_nrxqsets;
3847 int ntxqsets = ctx->ifc_softc_ctx.isc_ntxqsets;
3850 iflib_fl_t fl = NULL;
3851 int i, j, err, txconf, rxconf, fl_ifdi_offset;
3852 iflib_dma_info_t ifdip;
3853 uint32_t *rxqsizes = sctx->isc_rxqsizes;
3854 uint32_t *txqsizes = sctx->isc_txqsizes;
3855 uint8_t nrxqs = sctx->isc_nrxqs;
3856 uint8_t ntxqs = sctx->isc_ntxqs;
3857 int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
3860 struct ifmp_ring **brscp;
3861 int nbuf_rings = 1; /* XXX determine dynamically */
3863 KASSERT(ntxqs > 0, ("number of queues must be at least 1"));
3864 KASSERT(nrxqs > 0, ("number of queues must be at least 1"));
3866 /* Allocate the TX ring struct memory */
3868 (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
3869 ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
3870 device_printf(dev, "Unable to allocate TX ring memory\n");
3875 /* Now allocate the RX */
3877 (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
3878 nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
3879 device_printf(dev, "Unable to allocate RX ring memory\n");
3883 if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
3884 device_printf(dev, "Unable to buf_ring_sc * memory\n");
3889 ctx->ifc_txqs = txq;
3890 ctx->ifc_rxqs = rxq;
3893 * XXX handle allocation failure
3895 for (txconf = i = 0; i < ntxqsets; i++, txconf++, txq++) {
3896 /* Set up some basics */
3898 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
3899 device_printf(dev, "failed to allocate iflib_dma_info\n");
3903 txq->ift_ifdi = ifdip;
3904 for (j = 0; j < ntxqs; j++, ifdip++) {
3905 if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
3906 device_printf(dev, "Unable to allocate Descriptor memory\n");
3910 bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
3915 txq->ift_timer.c_cpu = i % mp_ncpus;
3916 txq->ift_db_check.c_cpu = i % mp_ncpus;
3917 txq->ift_nbr = nbuf_rings;
3919 if (iflib_txsd_alloc(txq)) {
3920 device_printf(dev, "Critical Failure setting up TX buffers\n");
3925 /* Initialize the TX lock */
3926 snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
3927 device_get_nameunit(dev), txq->ift_id);
3928 mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
3929 callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
3930 callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0);
3932 snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
3933 device_get_nameunit(dev), txq->ift_id);
3934 TXDB_LOCK_INIT(txq);
3936 txq->ift_br = brscp + i*nbuf_rings;
3937 for (j = 0; j < nbuf_rings; j++) {
3938 err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain,
3939 iflib_txq_can_drain, M_IFLIB, M_WAITOK);
3941 /* XXX free any allocated rings */
3942 device_printf(dev, "Unable to allocate buf_ring\n");
3948 for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
3949 /* Set up some basics */
3951 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
3952 device_printf(dev, "failed to allocate iflib_dma_info\n");
3957 rxq->ifr_ifdi = ifdip;
3958 for (j = 0; j < nrxqs; j++, ifdip++) {
3959 if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
3960 device_printf(dev, "Unable to allocate Descriptor memory\n");
3964 bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
3968 if (sctx->isc_flags & IFLIB_HAS_CQ) {
3973 rxq->ifr_nfl = nfree_lists;
3975 (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
3976 device_printf(dev, "Unable to allocate free list memory\n");
3981 for (j = 0; j < nfree_lists; j++) {
3982 rxq->ifr_fl[j].ifl_rxq = rxq;
3983 rxq->ifr_fl[j].ifl_id = j;
3984 rxq->ifr_fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + fl_ifdi_offset];
3986 /* Allocate receive buffers for the ring*/
3987 if (iflib_rxsd_alloc(rxq)) {
3989 "Critical Failure setting up receive buffers\n");
3996 vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
3997 paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
3998 for (i = 0; i < ntxqsets; i++) {
3999 iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4001 for (j = 0; j < ntxqs; j++, di++) {
4002 vaddrs[i*ntxqs + j] = di->idi_vaddr;
4003 paddrs[i*ntxqs + j] = di->idi_paddr;
4006 if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4007 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4008 iflib_tx_structures_free(ctx);
4009 free(vaddrs, M_IFLIB);
4010 free(paddrs, M_IFLIB);
4013 free(vaddrs, M_IFLIB);
4014 free(paddrs, M_IFLIB);
4017 vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4018 paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4019 for (i = 0; i < nrxqsets; i++) {
4020 iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4022 for (j = 0; j < nrxqs; j++, di++) {
4023 vaddrs[i*nrxqs + j] = di->idi_vaddr;
4024 paddrs[i*nrxqs + j] = di->idi_paddr;
4027 if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4028 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4029 iflib_tx_structures_free(ctx);
4030 free(vaddrs, M_IFLIB);
4031 free(paddrs, M_IFLIB);
4034 free(vaddrs, M_IFLIB);
4035 free(paddrs, M_IFLIB);
4039 /* XXX handle allocation failure changes */
4042 if (ctx->ifc_rxqs != NULL)
4043 free(ctx->ifc_rxqs, M_IFLIB);
4044 ctx->ifc_rxqs = NULL;
4046 if (ctx->ifc_txqs != NULL)
4047 free(ctx->ifc_txqs, M_IFLIB);
4048 ctx->ifc_txqs = NULL;
4054 iflib_tx_structures_setup(if_ctx_t ctx)
4056 iflib_txq_t txq = ctx->ifc_txqs;
4059 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4060 iflib_txq_setup(txq);
4066 iflib_tx_structures_free(if_ctx_t ctx)
4068 iflib_txq_t txq = ctx->ifc_txqs;
4071 for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4072 iflib_txq_destroy(txq);
4073 for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4074 iflib_dma_free(&txq->ift_ifdi[j]);
4076 free(ctx->ifc_txqs, M_IFLIB);
4077 ctx->ifc_txqs = NULL;
4078 IFDI_QUEUES_FREE(ctx);
4081 /*********************************************************************
4083 * Initialize all receive rings.
4085 **********************************************************************/
4087 iflib_rx_structures_setup(if_ctx_t ctx)
4089 iflib_rxq_t rxq = ctx->ifc_rxqs;
4091 #if defined(INET6) || defined(INET)
4095 for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4096 #if defined(INET6) || defined(INET)
4097 tcp_lro_free(&rxq->ifr_lc);
4098 if ((err = tcp_lro_init(&rxq->ifr_lc)) != 0) {
4099 device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4102 rxq->ifr_lro_enabled = TRUE;
4103 rxq->ifr_lc.ifp = ctx->ifc_ifp;
4105 IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4108 #if defined(INET6) || defined(INET)
4111 * Free RX software descriptors allocated so far, we will only handle
4112 * the rings that completed, the failing case will have
4113 * cleaned up for itself. 'q' failed, so its the terminus.
4115 rxq = ctx->ifc_rxqs;
4116 for (i = 0; i < q; ++i, rxq++) {
4117 iflib_rx_sds_free(rxq);
4118 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4124 /*********************************************************************
4126 * Free all receive rings.
4128 **********************************************************************/
4130 iflib_rx_structures_free(if_ctx_t ctx)
4132 iflib_rxq_t rxq = ctx->ifc_rxqs;
4134 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, rxq++) {
4135 iflib_rx_sds_free(rxq);
4140 iflib_qset_structures_setup(if_ctx_t ctx)
4144 if ((err = iflib_tx_structures_setup(ctx)) != 0)
4147 if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4148 device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4149 iflib_tx_structures_free(ctx);
4150 iflib_rx_structures_free(ctx);
4156 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4157 driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4160 return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4164 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4168 CPU_COPY(&ctx->ifc_cpus, cpus);
4169 /* clear up to the qid'th bit */
4170 for (i = 0; i < qid; i++) {
4171 cpuid = CPU_FFS(cpus);
4172 CPU_CLR(cpuid, cpus);
4177 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4178 iflib_intr_type_t type, driver_filter_t *filter,
4179 void *filter_arg, int qid, char *name)
4181 struct grouptask *gtask;
4182 struct taskqgroup *tqg;
4183 iflib_filter_info_t info;
4189 info = &ctx->ifc_filter_info;
4192 /* XXX merge tx/rx for netmap? */
4194 q = &ctx->ifc_txqs[qid];
4195 info = &ctx->ifc_txqs[qid].ift_filter_info;
4196 gtask = &ctx->ifc_txqs[qid].ift_task;
4197 tqg = qgroup_if_io_tqg;
4198 tqrid = irq->ii_rid;
4202 q = &ctx->ifc_rxqs[qid];
4203 info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4204 gtask = &ctx->ifc_rxqs[qid].ifr_task;
4205 tqg = qgroup_if_io_tqg;
4206 tqrid = irq->ii_rid;
4209 case IFLIB_INTR_ADMIN:
4211 info = &ctx->ifc_filter_info;
4212 gtask = &ctx->ifc_admin_task;
4213 tqg = qgroup_if_config_tqg;
4215 fn = _task_fn_admin;
4218 panic("unknown net intr type");
4220 GROUPTASK_INIT(gtask, 0, fn, q);
4222 info->ifi_filter = filter;
4223 info->ifi_filter_arg = filter_arg;
4224 info->ifi_task = gtask;
4226 /* XXX query cpu that rid belongs to */
4228 err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info, name);
4232 find_nth(ctx, &cpus, qid);
4233 taskqgroup_attach_cpu(tqg, gtask, q, CPU_FFS(&cpus), irq->ii_rid, name);
4235 taskqgroup_attach(tqg, gtask, q, tqrid, name);
4242 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type, void *arg, int qid, char *name)
4244 struct grouptask *gtask;
4245 struct taskqgroup *tqg;
4251 q = &ctx->ifc_txqs[qid];
4252 gtask = &ctx->ifc_txqs[qid].ift_task;
4253 tqg = qgroup_if_io_tqg;
4257 q = &ctx->ifc_rxqs[qid];
4258 gtask = &ctx->ifc_rxqs[qid].ifr_task;
4259 tqg = qgroup_if_io_tqg;
4262 case IFLIB_INTR_ADMIN:
4264 gtask = &ctx->ifc_admin_task;
4265 tqg = qgroup_if_config_tqg;
4267 fn = _task_fn_admin;
4269 case IFLIB_INTR_IOV:
4271 gtask = &ctx->ifc_vflr_task;
4272 tqg = qgroup_if_config_tqg;
4277 panic("unknown net intr type");
4279 GROUPTASK_INIT(gtask, 0, fn, q);
4280 taskqgroup_attach(tqg, gtask, q, rid, name);
4284 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4287 bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4290 bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4294 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4296 iflib_txq_t txq = ctx->ifc_txqs;
4297 iflib_rxq_t rxq = ctx->ifc_rxqs;
4298 if_irq_t irq = &ctx->ifc_legacy_irq;
4299 iflib_filter_info_t info;
4300 struct grouptask *gtask;
4301 struct taskqgroup *tqg;
4307 q = &ctx->ifc_rxqs[0];
4308 info = &rxq[0].ifr_filter_info;
4309 gtask = &rxq[0].ifr_task;
4310 tqg = qgroup_if_io_tqg;
4311 tqrid = irq->ii_rid = *rid;
4314 ctx->ifc_flags |= IFC_LEGACY;
4315 info->ifi_filter = filter;
4316 info->ifi_filter_arg = filter_arg;
4317 info->ifi_task = gtask;
4319 /* We allocate a single interrupt resource */
4320 if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0)
4322 GROUPTASK_INIT(gtask, 0, fn, q);
4323 taskqgroup_attach(tqg, gtask, q, tqrid, name);
4325 GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
4326 taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
4327 GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4328 taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin/link");
4334 iflib_led_create(if_ctx_t ctx)
4337 ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
4338 device_get_nameunit(ctx->ifc_dev));
4342 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
4345 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
4349 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
4352 GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
4356 iflib_admin_intr_deferred(if_ctx_t ctx)
4359 GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
4363 iflib_iov_intr_deferred(if_ctx_t ctx)
4366 GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
4370 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
4373 taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
4377 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, task_fn_t *fn,
4381 GROUPTASK_INIT(gtask, 0, fn, ctx);
4382 taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
4386 iflib_link_state_change(if_ctx_t ctx, int link_state)
4388 if_t ifp = ctx->ifc_ifp;
4389 iflib_txq_t txq = ctx->ifc_txqs;
4392 if_setbaudrate(ifp, baudrate);
4394 /* If link down, disable watchdog */
4395 if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
4396 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
4397 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4399 ctx->ifc_link_state = link_state;
4400 if_link_state_change(ifp, link_state);
4404 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
4408 if (ctx->isc_txd_credits_update == NULL)
4411 if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0)
4414 txq->ift_processed += credits;
4415 txq->ift_cidx_processed += credits;
4417 if (txq->ift_cidx_processed >= txq->ift_size)
4418 txq->ift_cidx_processed -= txq->ift_size;
4423 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx)
4426 return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx));
4430 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
4431 const char *description, if_int_delay_info_t info,
4432 int offset, int value)
4434 info->iidi_ctx = ctx;
4435 info->iidi_offset = offset;
4436 info->iidi_value = value;
4437 SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
4438 SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
4439 OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
4440 info, 0, iflib_sysctl_int_delay, "I", description);
4444 iflib_ctx_lock_get(if_ctx_t ctx)
4447 return (&ctx->ifc_mtx);
4451 iflib_msix_init(if_ctx_t ctx)
4453 device_t dev = ctx->ifc_dev;
4454 if_shared_ctx_t sctx = ctx->ifc_sctx;
4455 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4456 int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
4457 int iflib_num_tx_queues, iflib_num_rx_queues;
4458 int err, admincnt, bar;
4460 iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
4461 iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
4462 bar = ctx->ifc_softc_ctx.isc_msix_bar;
4463 admincnt = sctx->isc_admin_intrcnt;
4464 /* Override by tuneable */
4465 if (enable_msix == 0)
4469 ** When used in a virtualized environment
4470 ** PCI BUSMASTER capability may not be set
4471 ** so explicity set it here and rewrite
4472 ** the ENABLE in the MSIX control register
4473 ** at this point to cause the host to
4474 ** successfully initialize us.
4477 uint16_t pci_cmd_word;
4481 pci_cmd_word = pci_read_config(dev, PCIR_COMMAND, 2);
4482 pci_cmd_word |= PCIM_CMD_BUSMASTEREN;
4483 pci_write_config(dev, PCIR_COMMAND, pci_cmd_word, 2);
4484 pci_find_cap(dev, PCIY_MSIX, &rid);
4485 rid += PCIR_MSIX_CTRL;
4486 msix_ctrl = pci_read_config(dev, rid, 2);
4487 msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
4488 pci_write_config(dev, rid, msix_ctrl, 2);
4492 * bar == -1 => "trust me I know what I'm doing"
4493 * https://www.youtube.com/watch?v=nnwWKkNau4I
4494 * Some drivers are for hardware that is so shoddily
4495 * documented that no one knows which bars are which
4496 * so the developer has to map all bars. This hack
4497 * allows shoddy garbage to use msix in this framework.
4500 ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
4501 SYS_RES_MEMORY, &bar, RF_ACTIVE);
4502 if (ctx->ifc_msix_mem == NULL) {
4503 /* May not be enabled */
4504 device_printf(dev, "Unable to map MSIX table \n");
4508 /* First try MSI/X */
4509 if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
4510 device_printf(dev, "System has MSIX disabled \n");
4511 bus_release_resource(dev, SYS_RES_MEMORY,
4512 bar, ctx->ifc_msix_mem);
4513 ctx->ifc_msix_mem = NULL;
4517 /* use only 1 qset in debug mode */
4518 queuemsgs = min(msgs - admincnt, 1);
4520 queuemsgs = msgs - admincnt;
4522 if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
4524 queues = imin(queuemsgs, rss_getnumbuckets());
4528 queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
4529 device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
4530 CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
4532 device_printf(dev, "Unable to fetch CPU list\n");
4533 /* Figure out a reasonable auto config value */
4534 queues = min(queuemsgs, mp_ncpus);
4537 /* If we're doing RSS, clamp at the number of RSS buckets */
4538 if (queues > rss_getnumbuckets())
4539 queues = rss_getnumbuckets();
4541 if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queues)
4542 queues = rx_queues = iflib_num_rx_queues;
4545 if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
4546 tx_queues = iflib_num_tx_queues;
4550 device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
4552 vectors = queues + admincnt;
4553 if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
4555 "Using MSIX interrupts with %d vectors\n", vectors);
4556 scctx->isc_vectors = vectors;
4557 scctx->isc_nrxqsets = rx_queues;
4558 scctx->isc_ntxqsets = tx_queues;
4559 scctx->isc_intr = IFLIB_INTR_MSIX;
4562 device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
4565 vectors = pci_msi_count(dev);
4566 scctx->isc_nrxqsets = 1;
4567 scctx->isc_ntxqsets = 1;
4568 scctx->isc_vectors = vectors;
4569 if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
4570 device_printf(dev,"Using an MSI interrupt\n");
4571 scctx->isc_intr = IFLIB_INTR_MSI;
4573 device_printf(dev,"Using a Legacy interrupt\n");
4574 scctx->isc_intr = IFLIB_INTR_LEGACY;
4580 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
4583 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
4586 uint16_t *state = ((uint16_t *)oidp->oid_arg1);
4588 char *ring_state = "UNKNOWN";
4591 rc = sysctl_wire_old_buffer(req, 0);
4595 sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
4600 ring_state = ring_states[state[3]];
4602 sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
4603 state[0], state[1], state[2], ring_state);
4604 rc = sbuf_finish(sb);
4611 #define NAME_BUFLEN 32
4613 iflib_add_device_sysctl_pre(if_ctx_t ctx)
4615 device_t dev = iflib_get_dev(ctx);
4616 struct sysctl_oid_list *child, *oid_list;
4617 struct sysctl_ctx_list *ctx_list;
4618 struct sysctl_oid *node;
4620 ctx_list = device_get_sysctl_ctx(dev);
4621 child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
4622 ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
4623 CTLFLAG_RD, NULL, "IFLIB fields");
4624 oid_list = SYSCTL_CHILDREN(node);
4626 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
4627 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
4628 "# of txqs to use, 0 => use default #");
4629 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
4630 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
4631 "# of txqs to use, 0 => use default #");
4632 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxds",
4633 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxds, 0,
4634 "# of tx descriptors to use, 0 => use default #");
4635 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxds",
4636 CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxds, 0,
4637 "# of rx descriptors to use, 0 => use default #");
4642 iflib_add_device_sysctl_post(if_ctx_t ctx)
4644 if_shared_ctx_t sctx = ctx->ifc_sctx;
4645 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4646 device_t dev = iflib_get_dev(ctx);
4647 struct sysctl_oid_list *child;
4648 struct sysctl_ctx_list *ctx_list;
4653 char namebuf[NAME_BUFLEN];
4655 struct sysctl_oid *queue_node, *fl_node, *node;
4656 struct sysctl_oid_list *queue_list, *fl_list;
4657 ctx_list = device_get_sysctl_ctx(dev);
4659 node = ctx->ifc_sysctl_node;
4660 child = SYSCTL_CHILDREN(node);
4662 if (scctx->isc_ntxqsets > 100)
4664 else if (scctx->isc_ntxqsets > 10)
4668 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
4669 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
4670 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
4671 CTLFLAG_RD, NULL, "Queue Name");
4672 queue_list = SYSCTL_CHILDREN(queue_node);
4674 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
4676 &txq->ift_dequeued, "total mbufs freed");
4677 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
4679 &txq->ift_enqueued, "total mbufs enqueued");
4681 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
4683 &txq->ift_mbuf_defrag, "# of times m_defrag was called");
4684 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
4686 &txq->ift_pullups, "# of times m_pullup was called");
4687 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
4689 &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
4690 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
4692 &txq->ift_mbuf_defrag_failed, "# of times no descriptors were available");
4693 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
4695 &txq->ift_map_failed, "# of times dma map failed");
4696 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
4698 &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
4699 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
4701 &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
4702 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
4704 &txq->ift_pidx, 1, "Producer Index");
4705 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
4707 &txq->ift_cidx, 1, "Consumer Index");
4708 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
4710 &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
4711 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
4713 &txq->ift_in_use, 1, "descriptors in use");
4714 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
4716 &txq->ift_processed, "descriptors procesed for clean");
4717 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
4719 &txq->ift_cleaned, "total cleaned");
4720 SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
4721 CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state),
4722 0, mp_ring_state_handler, "A", "soft ring state");
4723 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
4724 CTLFLAG_RD, &txq->ift_br[0]->enqueues,
4725 "# of enqueues to the mp_ring for this queue");
4726 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
4727 CTLFLAG_RD, &txq->ift_br[0]->drops,
4728 "# of drops in the mp_ring for this queue");
4729 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
4730 CTLFLAG_RD, &txq->ift_br[0]->starts,
4731 "# of normal consumer starts in the mp_ring for this queue");
4732 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
4733 CTLFLAG_RD, &txq->ift_br[0]->stalls,
4734 "# of consumer stalls in the mp_ring for this queue");
4735 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
4736 CTLFLAG_RD, &txq->ift_br[0]->restarts,
4737 "# of consumer restarts in the mp_ring for this queue");
4738 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
4739 CTLFLAG_RD, &txq->ift_br[0]->abdications,
4740 "# of consumer abdications in the mp_ring for this queue");
4744 if (scctx->isc_nrxqsets > 100)
4746 else if (scctx->isc_nrxqsets > 10)
4750 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
4751 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
4752 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
4753 CTLFLAG_RD, NULL, "Queue Name");
4754 queue_list = SYSCTL_CHILDREN(queue_node);
4755 if (sctx->isc_flags & IFLIB_HAS_CQ) {
4756 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
4758 &rxq->ifr_cq_pidx, 1, "Producer Index");
4759 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
4761 &rxq->ifr_cq_cidx, 1, "Consumer Index");
4763 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
4764 snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
4765 fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
4766 CTLFLAG_RD, NULL, "freelist Name");
4767 fl_list = SYSCTL_CHILDREN(fl_node);
4768 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
4770 &fl->ifl_pidx, 1, "Producer Index");
4771 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
4773 &fl->ifl_cidx, 1, "Consumer Index");
4774 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
4776 &fl->ifl_credits, 1, "credits available");
4778 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
4780 &fl->ifl_m_enqueued, "mbufs allocated");
4781 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
4783 &fl->ifl_m_dequeued, "mbufs freed");
4784 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
4786 &fl->ifl_cl_enqueued, "clusters allocated");
4787 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
4789 &fl->ifl_cl_dequeued, "clusters freed");