1 /*- 2 * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions are met: 7 * 8 * 1. Redistributions of source code must retain the above copyright notice, 9 * this list of conditions and the following disclaimer. 10 * 11 * 2. Neither the name of Matthew Macy nor the names of its 12 * contributors may be used to endorse or promote products derived from 13 * this software without specific prior written permission. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 16 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE 19 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 20 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 21 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 22 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 23 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 24 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 25 * POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_acpi.h" 34 35 #include <sys/param.h> 36 #include <sys/types.h> 37 #include <sys/bus.h> 38 #include <sys/eventhandler.h> 39 #include <sys/sockio.h> 40 #include <sys/kernel.h> 41 #include <sys/lock.h> 42 #include <sys/mutex.h> 43 #include <sys/module.h> 44 #include <sys/kobj.h> 45 #include <sys/rman.h> 46 #include <sys/sbuf.h> 47 #include <sys/smp.h> 48 #include <sys/socket.h> 49 #include <sys/sysctl.h> 50 #include <sys/syslog.h> 51 #include <sys/taskqueue.h> 52 #include <sys/limits.h> 53 54 55 #include <net/if.h> 56 #include <net/if_var.h> 57 #include <net/if_types.h> 58 #include <net/if_media.h> 59 #include <net/bpf.h> 60 #include <net/ethernet.h> 61 #include <net/mp_ring.h> 62 63 #include <netinet/in.h> 64 #include <netinet/in_pcb.h> 65 #include <netinet/tcp_lro.h> 66 #include <netinet/in_systm.h> 67 #include <netinet/if_ether.h> 68 #include <netinet/ip.h> 69 #include <netinet/ip6.h> 70 #include <netinet/tcp.h> 71 72 #include <machine/bus.h> 73 #include <machine/in_cksum.h> 74 75 #include <vm/vm.h> 76 #include <vm/pmap.h> 77 78 #include <dev/led/led.h> 79 #include <dev/pci/pcireg.h> 80 #include <dev/pci/pcivar.h> 81 #include <dev/pci/pci_private.h> 82 83 #include <net/iflib.h> 84 85 #include "ifdi_if.h" 86 87 #if defined(__i386__) || defined(__amd64__) 88 #include <sys/memdesc.h> 89 #include <machine/bus.h> 90 #include <machine/md_var.h> 91 #include <machine/specialreg.h> 92 #include <x86/include/busdma_impl.h> 93 #include <x86/iommu/busdma_dmar.h> 94 #endif 95 96 /* 97 * enable accounting of every mbuf as it comes in to and goes out of iflib's software descriptor references 98 */ 99 #define MEMORY_LOGGING 0 100 /* 101 * Enable mbuf vectors for compressing long mbuf chains 102 */ 103 104 /* 105 * NB: 106 * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead 107 * we prefetch needs to be determined by the time spent in m_free vis a vis 108 * the cost of a prefetch. This will of course vary based on the workload: 109 * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which 110 * is quite expensive, thus suggesting very little prefetch. 111 * - small packet forwarding which is just returning a single mbuf to 112 * UMA will typically be very fast vis a vis the cost of a memory 113 * access. 114 */ 115 116 117 /* 118 * File organization: 119 * - private structures 120 * - iflib private utility functions 121 * - ifnet functions 122 * - vlan registry and other exported functions 123 * - iflib public core functions 124 * 125 * 126 */ 127 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library"); 128 129 struct iflib_txq; 130 typedef struct iflib_txq *iflib_txq_t; 131 struct iflib_rxq; 132 typedef struct iflib_rxq *iflib_rxq_t; 133 struct iflib_fl; 134 typedef struct iflib_fl *iflib_fl_t; 135 136 struct iflib_ctx; 137 138 typedef struct iflib_filter_info { 139 driver_filter_t *ifi_filter; 140 void *ifi_filter_arg; 141 struct grouptask *ifi_task; 142 struct iflib_ctx *ifi_ctx; 143 } *iflib_filter_info_t; 144 145 struct iflib_ctx { 146 KOBJ_FIELDS; 147 /* 148 * Pointer to hardware driver's softc 149 */ 150 void *ifc_softc; 151 device_t ifc_dev; 152 if_t ifc_ifp; 153 154 cpuset_t ifc_cpus; 155 if_shared_ctx_t ifc_sctx; 156 struct if_softc_ctx ifc_softc_ctx; 157 158 struct mtx ifc_mtx; 159 160 uint16_t ifc_nhwtxqs; 161 uint16_t ifc_nhwrxqs; 162 163 iflib_txq_t ifc_txqs; 164 iflib_rxq_t ifc_rxqs; 165 uint32_t ifc_if_flags; 166 uint32_t ifc_flags; 167 uint32_t ifc_max_fl_buf_size; 168 int ifc_in_detach; 169 170 int ifc_link_state; 171 int ifc_link_irq; 172 int ifc_pause_frames; 173 int ifc_watchdog_events; 174 struct cdev *ifc_led_dev; 175 struct resource *ifc_msix_mem; 176 177 struct if_irq ifc_legacy_irq; 178 struct grouptask ifc_admin_task; 179 struct grouptask ifc_vflr_task; 180 struct iflib_filter_info ifc_filter_info; 181 struct ifmedia ifc_media; 182 183 struct sysctl_oid *ifc_sysctl_node; 184 uint16_t ifc_sysctl_ntxqs; 185 uint16_t ifc_sysctl_nrxqs; 186 uint16_t ifc_sysctl_qs_eq_override; 187 188 uint16_t ifc_sysctl_ntxds[8]; 189 uint16_t ifc_sysctl_nrxds[8]; 190 struct if_txrx ifc_txrx; 191 #define isc_txd_encap ifc_txrx.ift_txd_encap 192 #define isc_txd_flush ifc_txrx.ift_txd_flush 193 #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update 194 #define isc_rxd_available ifc_txrx.ift_rxd_available 195 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get 196 #define isc_rxd_refill ifc_txrx.ift_rxd_refill 197 #define isc_rxd_flush ifc_txrx.ift_rxd_flush 198 #define isc_rxd_refill ifc_txrx.ift_rxd_refill 199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill 200 #define isc_legacy_intr ifc_txrx.ift_legacy_intr 201 eventhandler_tag ifc_vlan_attach_event; 202 eventhandler_tag ifc_vlan_detach_event; 203 uint8_t ifc_mac[ETHER_ADDR_LEN]; 204 char ifc_mtx_name[16]; 205 }; 206 207 208 void * 209 iflib_get_softc(if_ctx_t ctx) 210 { 211 212 return (ctx->ifc_softc); 213 } 214 215 device_t 216 iflib_get_dev(if_ctx_t ctx) 217 { 218 219 return (ctx->ifc_dev); 220 } 221 222 if_t 223 iflib_get_ifp(if_ctx_t ctx) 224 { 225 226 return (ctx->ifc_ifp); 227 } 228 229 struct ifmedia * 230 iflib_get_media(if_ctx_t ctx) 231 { 232 233 return (&ctx->ifc_media); 234 } 235 236 void 237 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN]) 238 { 239 240 bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN); 241 } 242 243 if_softc_ctx_t 244 iflib_get_softc_ctx(if_ctx_t ctx) 245 { 246 247 return (&ctx->ifc_softc_ctx); 248 } 249 250 if_shared_ctx_t 251 iflib_get_sctx(if_ctx_t ctx) 252 { 253 254 return (ctx->ifc_sctx); 255 } 256 257 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*)) 258 259 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP) 260 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF) 261 262 #define RX_SW_DESC_MAP_CREATED (1 << 0) 263 #define TX_SW_DESC_MAP_CREATED (1 << 1) 264 #define RX_SW_DESC_INUSE (1 << 3) 265 #define TX_SW_DESC_MAPPED (1 << 4) 266 267 typedef struct iflib_sw_rx_desc_array { 268 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ 269 struct mbuf **ifsd_m; /* pkthdr mbufs */ 270 caddr_t *ifsd_cl; /* direct cluster pointer for rx */ 271 uint8_t *ifsd_flags; 272 } iflib_rxsd_array_t; 273 274 typedef struct iflib_sw_tx_desc_array { 275 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ 276 struct mbuf **ifsd_m; /* pkthdr mbufs */ 277 uint8_t *ifsd_flags; 278 } iflib_txsd_array_t; 279 280 281 /* magic number that should be high enough for any hardware */ 282 #define IFLIB_MAX_TX_SEGS 128 283 #define IFLIB_MAX_RX_SEGS 32 284 #define IFLIB_RX_COPY_THRESH 63 285 #define IFLIB_MAX_RX_REFRESH 32 286 #define IFLIB_QUEUE_IDLE 0 287 #define IFLIB_QUEUE_HUNG 1 288 #define IFLIB_QUEUE_WORKING 2 289 290 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */ 291 #define TX_BATCH_SIZE 16 292 293 #define IFLIB_RESTART_BUDGET 8 294 295 #define IFC_LEGACY 0x01 296 #define IFC_QFLUSH 0x02 297 #define IFC_MULTISEG 0x04 298 #define IFC_DMAR 0x08 299 #define IFC_SC_ALLOCATED 0x10 300 #define IFC_INIT_DONE 0x20 301 302 303 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \ 304 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \ 305 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP) 306 struct iflib_txq { 307 uint16_t ift_in_use; 308 uint16_t ift_cidx; 309 uint16_t ift_cidx_processed; 310 uint16_t ift_pidx; 311 uint8_t ift_gen; 312 uint8_t ift_db_pending; 313 uint8_t ift_db_pending_queued; 314 uint8_t ift_npending; 315 uint8_t ift_br_offset; 316 /* implicit pad */ 317 uint64_t ift_processed; 318 uint64_t ift_cleaned; 319 #if MEMORY_LOGGING 320 uint64_t ift_enqueued; 321 uint64_t ift_dequeued; 322 #endif 323 uint64_t ift_no_tx_dma_setup; 324 uint64_t ift_no_desc_avail; 325 uint64_t ift_mbuf_defrag_failed; 326 uint64_t ift_mbuf_defrag; 327 uint64_t ift_map_failed; 328 uint64_t ift_txd_encap_efbig; 329 uint64_t ift_pullups; 330 331 struct mtx ift_mtx; 332 struct mtx ift_db_mtx; 333 334 /* constant values */ 335 if_ctx_t ift_ctx; 336 struct ifmp_ring **ift_br; 337 struct grouptask ift_task; 338 uint16_t ift_size; 339 uint16_t ift_id; 340 struct callout ift_timer; 341 struct callout ift_db_check; 342 343 iflib_txsd_array_t ift_sds; 344 uint8_t ift_nbr; 345 uint8_t ift_qstatus; 346 uint8_t ift_active; 347 uint8_t ift_closed; 348 int ift_watchdog_time; 349 struct iflib_filter_info ift_filter_info; 350 bus_dma_tag_t ift_desc_tag; 351 bus_dma_tag_t ift_tso_desc_tag; 352 iflib_dma_info_t ift_ifdi; 353 #define MTX_NAME_LEN 16 354 char ift_mtx_name[MTX_NAME_LEN]; 355 char ift_db_mtx_name[MTX_NAME_LEN]; 356 bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE); 357 #ifdef IFLIB_DIAGNOSTICS 358 uint64_t ift_cpu_exec_count[256]; 359 #endif 360 } __aligned(CACHE_LINE_SIZE); 361 362 struct iflib_fl { 363 uint16_t ifl_cidx; 364 uint16_t ifl_pidx; 365 uint16_t ifl_credits; 366 uint8_t ifl_gen; 367 #if MEMORY_LOGGING 368 uint64_t ifl_m_enqueued; 369 uint64_t ifl_m_dequeued; 370 uint64_t ifl_cl_enqueued; 371 uint64_t ifl_cl_dequeued; 372 #endif 373 /* implicit pad */ 374 375 /* constant */ 376 uint16_t ifl_size; 377 uint16_t ifl_buf_size; 378 uint16_t ifl_cltype; 379 uma_zone_t ifl_zone; 380 iflib_rxsd_array_t ifl_sds; 381 iflib_rxq_t ifl_rxq; 382 uint8_t ifl_id; 383 bus_dma_tag_t ifl_desc_tag; 384 iflib_dma_info_t ifl_ifdi; 385 uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE); 386 caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH]; 387 } __aligned(CACHE_LINE_SIZE); 388 389 static inline int 390 get_inuse(int size, int cidx, int pidx, int gen) 391 { 392 int used; 393 394 if (pidx > cidx) 395 used = pidx - cidx; 396 else if (pidx < cidx) 397 used = size - cidx + pidx; 398 else if (gen == 0 && pidx == cidx) 399 used = 0; 400 else if (gen == 1 && pidx == cidx) 401 used = size; 402 else 403 panic("bad state"); 404 405 return (used); 406 } 407 408 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen)) 409 410 #define IDXDIFF(head, tail, wrap) \ 411 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head)) 412 413 struct iflib_rxq { 414 /* If there is a separate completion queue - 415 * these are the cq cidx and pidx. Otherwise 416 * these are unused. 417 */ 418 uint16_t ifr_size; 419 uint16_t ifr_cq_cidx; 420 uint16_t ifr_cq_pidx; 421 uint8_t ifr_cq_gen; 422 uint8_t ifr_fl_offset; 423 424 if_ctx_t ifr_ctx; 425 iflib_fl_t ifr_fl; 426 uint64_t ifr_rx_irq; 427 uint16_t ifr_id; 428 uint8_t ifr_lro_enabled; 429 uint8_t ifr_nfl; 430 struct lro_ctrl ifr_lc; 431 struct grouptask ifr_task; 432 struct iflib_filter_info ifr_filter_info; 433 iflib_dma_info_t ifr_ifdi; 434 /* dynamically allocate if any drivers need a value substantially larger than this */ 435 struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE); 436 #ifdef IFLIB_DIAGNOSTICS 437 uint64_t ifr_cpu_exec_count[256]; 438 #endif 439 } __aligned(CACHE_LINE_SIZE); 440 441 /* 442 * Only allow a single packet to take up most 1/nth of the tx ring 443 */ 444 #define MAX_SINGLE_PACKET_FRACTION 12 445 #define IF_BAD_DMA (bus_addr_t)-1 446 447 static int enable_msix = 1; 448 449 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING)) 450 451 #define CTX_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF) 452 453 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx) 454 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx) 455 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx) 456 457 458 #define TXDB_LOCK_INIT(txq) mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF) 459 #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx) 460 #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx) 461 #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx) 462 #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx) 463 464 #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx) 465 #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx) 466 467 468 /* Our boot-time initialization hook */ 469 static int iflib_module_event_handler(module_t, int, void *); 470 471 static moduledata_t iflib_moduledata = { 472 "iflib", 473 iflib_module_event_handler, 474 NULL 475 }; 476 477 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY); 478 MODULE_VERSION(iflib, 1); 479 480 MODULE_DEPEND(iflib, pci, 1, 1, 1); 481 MODULE_DEPEND(iflib, ether, 1, 1, 1); 482 483 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1); 484 TASKQGROUP_DEFINE(if_config_tqg, 1, 1); 485 486 #ifndef IFLIB_DEBUG_COUNTERS 487 #ifdef INVARIANTS 488 #define IFLIB_DEBUG_COUNTERS 1 489 #else 490 #define IFLIB_DEBUG_COUNTERS 0 491 #endif /* !INVARIANTS */ 492 #endif 493 494 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0, 495 "iflib driver parameters"); 496 497 /* 498 * XXX need to ensure that this can't accidentally cause the head to be moved backwards 499 */ 500 static int iflib_min_tx_latency = 0; 501 502 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW, 503 &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput"); 504 505 506 #if IFLIB_DEBUG_COUNTERS 507 508 static int iflib_tx_seen; 509 static int iflib_tx_sent; 510 static int iflib_tx_encap; 511 static int iflib_rx_allocs; 512 static int iflib_fl_refills; 513 static int iflib_fl_refills_large; 514 static int iflib_tx_frees; 515 516 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, 517 &iflib_tx_seen, 0, "# tx mbufs seen"); 518 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, 519 &iflib_tx_sent, 0, "# tx mbufs sent"); 520 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, 521 &iflib_tx_encap, 0, "# tx mbufs encapped"); 522 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, 523 &iflib_tx_frees, 0, "# tx frees"); 524 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, 525 &iflib_rx_allocs, 0, "# rx allocations"); 526 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, 527 &iflib_fl_refills, 0, "# refills"); 528 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD, 529 &iflib_fl_refills_large, 0, "# large refills"); 530 531 532 static int iflib_txq_drain_flushing; 533 static int iflib_txq_drain_oactive; 534 static int iflib_txq_drain_notready; 535 static int iflib_txq_drain_encapfail; 536 537 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD, 538 &iflib_txq_drain_flushing, 0, "# drain flushes"); 539 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD, 540 &iflib_txq_drain_oactive, 0, "# drain oactives"); 541 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD, 542 &iflib_txq_drain_notready, 0, "# drain notready"); 543 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD, 544 &iflib_txq_drain_encapfail, 0, "# drain encap fails"); 545 546 547 static int iflib_encap_load_mbuf_fail; 548 static int iflib_encap_txq_avail_fail; 549 static int iflib_encap_txd_encap_fail; 550 551 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD, 552 &iflib_encap_load_mbuf_fail, 0, "# busdma load failures"); 553 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD, 554 &iflib_encap_txq_avail_fail, 0, "# txq avail failures"); 555 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD, 556 &iflib_encap_txd_encap_fail, 0, "# driver encap failures"); 557 558 static int iflib_task_fn_rxs; 559 static int iflib_rx_intr_enables; 560 static int iflib_fast_intrs; 561 static int iflib_intr_link; 562 static int iflib_intr_msix; 563 static int iflib_rx_unavail; 564 static int iflib_rx_ctx_inactive; 565 static int iflib_rx_zero_len; 566 static int iflib_rx_if_input; 567 static int iflib_rx_mbuf_null; 568 static int iflib_rxd_flush; 569 570 static int iflib_verbose_debug; 571 572 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD, 573 &iflib_intr_link, 0, "# intr link calls"); 574 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD, 575 &iflib_intr_msix, 0, "# intr msix calls"); 576 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, 577 &iflib_task_fn_rxs, 0, "# task_fn_rx calls"); 578 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD, 579 &iflib_rx_intr_enables, 0, "# rx intr enables"); 580 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, 581 &iflib_fast_intrs, 0, "# fast_intr calls"); 582 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, 583 &iflib_rx_unavail, 0, "# times rxeof called with no available data"); 584 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD, 585 &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context"); 586 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD, 587 &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf"); 588 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, 589 &iflib_rx_if_input, 0, "# times rxeof called if_input"); 590 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD, 591 &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf"); 592 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, 593 &iflib_rxd_flush, 0, "# times rxd_flush called"); 594 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW, 595 &iflib_verbose_debug, 0, "enable verbose debugging"); 596 597 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1) 598 static void 599 iflib_debug_reset(void) 600 { 601 iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs = 602 iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees = 603 iflib_txq_drain_flushing = iflib_txq_drain_oactive = 604 iflib_txq_drain_notready = iflib_txq_drain_encapfail = 605 iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail = 606 iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables = 607 iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail = 608 iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input = 609 iflib_rx_mbuf_null = iflib_rxd_flush = 0; 610 } 611 612 #else 613 #define DBG_COUNTER_INC(name) 614 static void iflib_debug_reset(void) {} 615 #endif 616 617 618 619 #define IFLIB_DEBUG 0 620 621 static void iflib_tx_structures_free(if_ctx_t ctx); 622 static void iflib_rx_structures_free(if_ctx_t ctx); 623 static int iflib_queues_alloc(if_ctx_t ctx); 624 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); 625 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget); 626 static int iflib_qset_structures_setup(if_ctx_t ctx); 627 static int iflib_msix_init(if_ctx_t ctx); 628 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str); 629 static void iflib_txq_check_drain(iflib_txq_t txq, int budget); 630 static uint32_t iflib_txq_can_drain(struct ifmp_ring *); 631 static int iflib_register(if_ctx_t); 632 static void iflib_init_locked(if_ctx_t ctx); 633 static void iflib_add_device_sysctl_pre(if_ctx_t ctx); 634 static void iflib_add_device_sysctl_post(if_ctx_t ctx); 635 static void iflib_ifmp_purge(iflib_txq_t txq); 636 static void _iflib_pre_assert(if_softc_ctx_t scctx); 637 638 #ifdef DEV_NETMAP 639 #include <sys/selinfo.h> 640 #include <net/netmap.h> 641 #include <dev/netmap/netmap_kern.h> 642 643 MODULE_DEPEND(iflib, netmap, 1, 1, 1); 644 645 /* 646 * device-specific sysctl variables: 647 * 648 * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it. 649 * During regular operations the CRC is stripped, but on some 650 * hardware reception of frames not multiple of 64 is slower, 651 * so using crcstrip=0 helps in benchmarks. 652 * 653 * iflib_rx_miss, iflib_rx_miss_bufs: 654 * count packets that might be missed due to lost interrupts. 655 */ 656 SYSCTL_DECL(_dev_netmap); 657 /* 658 * The xl driver by default strips CRCs and we do not override it. 659 */ 660 661 int iflib_crcstrip = 1; 662 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip, 663 CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames"); 664 665 int iflib_rx_miss, iflib_rx_miss_bufs; 666 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss, 667 CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr"); 668 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs, 669 CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs"); 670 671 /* 672 * Register/unregister. We are already under netmap lock. 673 * Only called on the first register or the last unregister. 674 */ 675 static int 676 iflib_netmap_register(struct netmap_adapter *na, int onoff) 677 { 678 struct ifnet *ifp = na->ifp; 679 if_ctx_t ctx = ifp->if_softc; 680 681 CTX_LOCK(ctx); 682 IFDI_INTR_DISABLE(ctx); 683 684 /* Tell the stack that the interface is no longer active */ 685 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE); 686 687 if (!CTX_IS_VF(ctx)) 688 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); 689 690 /* enable or disable flags and callbacks in na and ifp */ 691 if (onoff) { 692 nm_set_native_flags(na); 693 } else { 694 nm_clear_native_flags(na); 695 } 696 IFDI_INIT(ctx); 697 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ? 698 CTX_UNLOCK(ctx); 699 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1); 700 } 701 702 /* 703 * Reconcile kernel and user view of the transmit ring. 704 * 705 * All information is in the kring. 706 * Userspace wants to send packets up to the one before kring->rhead, 707 * kernel knows kring->nr_hwcur is the first unsent packet. 708 * 709 * Here we push packets out (as many as possible), and possibly 710 * reclaim buffers from previously completed transmission. 711 * 712 * The caller (netmap) guarantees that there is only one instance 713 * running at any time. Any interference with other driver 714 * methods should be handled by the individual drivers. 715 */ 716 static int 717 iflib_netmap_txsync(struct netmap_kring *kring, int flags) 718 { 719 struct netmap_adapter *na = kring->na; 720 struct ifnet *ifp = na->ifp; 721 struct netmap_ring *ring = kring->ring; 722 u_int nm_i; /* index into the netmap ring */ 723 u_int nic_i; /* index into the NIC ring */ 724 u_int n; 725 u_int const lim = kring->nkr_num_slots - 1; 726 u_int const head = kring->rhead; 727 struct if_pkt_info pi; 728 729 /* 730 * interrupts on every tx packet are expensive so request 731 * them every half ring, or where NS_REPORT is set 732 */ 733 u_int report_frequency = kring->nkr_num_slots >> 1; 734 /* device-specific */ 735 if_ctx_t ctx = ifp->if_softc; 736 iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id]; 737 738 pi.ipi_segs = txq->ift_segs; 739 pi.ipi_qsidx = kring->ring_id; 740 pi.ipi_ndescs = 0; 741 742 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map, 743 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 744 745 746 /* 747 * First part: process new packets to send. 748 * nm_i is the current index in the netmap ring, 749 * nic_i is the corresponding index in the NIC ring. 750 * 751 * If we have packets to send (nm_i != head) 752 * iterate over the netmap ring, fetch length and update 753 * the corresponding slot in the NIC ring. Some drivers also 754 * need to update the buffer's physical address in the NIC slot 755 * even NS_BUF_CHANGED is not set (PNMB computes the addresses). 756 * 757 * The netmap_reload_map() calls is especially expensive, 758 * even when (as in this case) the tag is 0, so do only 759 * when the buffer has actually changed. 760 * 761 * If possible do not set the report/intr bit on all slots, 762 * but only a few times per ring or when NS_REPORT is set. 763 * 764 * Finally, on 10G and faster drivers, it might be useful 765 * to prefetch the next slot and txr entry. 766 */ 767 768 nm_i = kring->nr_hwcur; 769 if (nm_i != head) { /* we have new packets to send */ 770 nic_i = netmap_idx_k2n(kring, nm_i); 771 772 __builtin_prefetch(&ring->slot[nm_i]); 773 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]); 774 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]); 775 776 for (n = 0; nm_i != head; n++) { 777 struct netmap_slot *slot = &ring->slot[nm_i]; 778 u_int len = slot->len; 779 uint64_t paddr; 780 void *addr = PNMB(na, slot, &paddr); 781 int flags = (slot->flags & NS_REPORT || 782 nic_i == 0 || nic_i == report_frequency) ? 783 IPI_TX_INTR : 0; 784 785 /* device-specific */ 786 pi.ipi_pidx = nic_i; 787 pi.ipi_flags = flags; 788 789 /* Fill the slot in the NIC ring. */ 790 ctx->isc_txd_encap(ctx->ifc_softc, &pi); 791 792 /* prefetch for next round */ 793 __builtin_prefetch(&ring->slot[nm_i + 1]); 794 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]); 795 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]); 796 797 NM_CHECK_ADDR_LEN(na, addr, len); 798 799 if (slot->flags & NS_BUF_CHANGED) { 800 /* buffer has changed, reload map */ 801 netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr); 802 } 803 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED); 804 805 /* make sure changes to the buffer are synced */ 806 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i], 807 BUS_DMASYNC_PREWRITE); 808 809 nm_i = nm_next(nm_i, lim); 810 nic_i = nm_next(nic_i, lim); 811 } 812 kring->nr_hwcur = head; 813 814 /* synchronize the NIC ring */ 815 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map, 816 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 817 818 /* (re)start the tx unit up to slot nic_i (excluded) */ 819 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i); 820 } 821 822 /* 823 * Second part: reclaim buffers for completed transmissions. 824 */ 825 if (iflib_tx_credits_update(ctx, txq)) { 826 /* some tx completed, increment avail */ 827 nic_i = txq->ift_cidx_processed; 828 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim); 829 } 830 return (0); 831 } 832 833 /* 834 * Reconcile kernel and user view of the receive ring. 835 * Same as for the txsync, this routine must be efficient. 836 * The caller guarantees a single invocations, but races against 837 * the rest of the driver should be handled here. 838 * 839 * On call, kring->rhead is the first packet that userspace wants 840 * to keep, and kring->rcur is the wakeup point. 841 * The kernel has previously reported packets up to kring->rtail. 842 * 843 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective 844 * of whether or not we received an interrupt. 845 */ 846 static int 847 iflib_netmap_rxsync(struct netmap_kring *kring, int flags) 848 { 849 struct netmap_adapter *na = kring->na; 850 struct ifnet *ifp = na->ifp; 851 struct netmap_ring *ring = kring->ring; 852 u_int nm_i; /* index into the netmap ring */ 853 u_int nic_i; /* index into the NIC ring */ 854 u_int i, n; 855 u_int const lim = kring->nkr_num_slots - 1; 856 u_int const head = kring->rhead; 857 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; 858 struct if_rxd_info ri; 859 /* device-specific */ 860 if_ctx_t ctx = ifp->if_softc; 861 iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id]; 862 iflib_fl_t fl = rxq->ifr_fl; 863 if (head > lim) 864 return netmap_ring_reinit(kring); 865 866 bzero(&ri, sizeof(ri)); 867 ri.iri_qsidx = kring->ring_id; 868 ri.iri_ifp = ctx->ifc_ifp; 869 /* XXX check sync modes */ 870 for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) 871 bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map, 872 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 873 874 /* 875 * First part: import newly received packets. 876 * 877 * nm_i is the index of the next free slot in the netmap ring, 878 * nic_i is the index of the next received packet in the NIC ring, 879 * and they may differ in case if_init() has been called while 880 * in netmap mode. For the receive ring we have 881 * 882 * nic_i = rxr->next_check; 883 * nm_i = kring->nr_hwtail (previous) 884 * and 885 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size 886 * 887 * rxr->next_check is set to 0 on a ring reinit 888 */ 889 if (netmap_no_pendintr || force_update) { 890 int crclen = iflib_crcstrip ? 0 : 4; 891 int error, avail; 892 uint16_t slot_flags = kring->nkr_slot_flags; 893 894 for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) { 895 nic_i = fl->ifl_cidx; 896 nm_i = netmap_idx_n2k(kring, nic_i); 897 avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i, INT_MAX); 898 for (n = 0; avail > 0; n++, avail--) { 899 error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); 900 if (error) 901 ring->slot[nm_i].len = 0; 902 else 903 ring->slot[nm_i].len = ri.iri_len - crclen; 904 ring->slot[nm_i].flags = slot_flags; 905 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, 906 fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD); 907 nm_i = nm_next(nm_i, lim); 908 nic_i = nm_next(nic_i, lim); 909 } 910 if (n) { /* update the state variables */ 911 if (netmap_no_pendintr && !force_update) { 912 /* diagnostics */ 913 iflib_rx_miss ++; 914 iflib_rx_miss_bufs += n; 915 } 916 fl->ifl_cidx = nic_i; 917 kring->nr_hwtail = nm_i; 918 } 919 kring->nr_kflags &= ~NKR_PENDINTR; 920 } 921 } 922 /* 923 * Second part: skip past packets that userspace has released. 924 * (kring->nr_hwcur to head excluded), 925 * and make the buffers available for reception. 926 * As usual nm_i is the index in the netmap ring, 927 * nic_i is the index in the NIC ring, and 928 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size 929 */ 930 /* XXX not sure how this will work with multiple free lists */ 931 nm_i = kring->nr_hwcur; 932 if (nm_i != head) { 933 nic_i = netmap_idx_k2n(kring, nm_i); 934 for (n = 0; nm_i != head; n++) { 935 struct netmap_slot *slot = &ring->slot[nm_i]; 936 uint64_t paddr; 937 caddr_t vaddr; 938 void *addr = PNMB(na, slot, &paddr); 939 940 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */ 941 goto ring_reset; 942 943 vaddr = addr; 944 if (slot->flags & NS_BUF_CHANGED) { 945 /* buffer has changed, reload map */ 946 netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr); 947 slot->flags &= ~NS_BUF_CHANGED; 948 } 949 /* 950 * XXX we should be batching this operation - TODO 951 */ 952 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1, fl->ifl_buf_size); 953 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], 954 BUS_DMASYNC_PREREAD); 955 nm_i = nm_next(nm_i, lim); 956 nic_i = nm_next(nic_i, lim); 957 } 958 kring->nr_hwcur = head; 959 960 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 961 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 962 /* 963 * IMPORTANT: we must leave one free slot in the ring, 964 * so move nic_i back by one unit 965 */ 966 nic_i = nm_prev(nic_i, lim); 967 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i); 968 } 969 970 return 0; 971 972 ring_reset: 973 return netmap_ring_reinit(kring); 974 } 975 976 static int 977 iflib_netmap_attach(if_ctx_t ctx) 978 { 979 struct netmap_adapter na; 980 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 981 982 bzero(&na, sizeof(na)); 983 984 na.ifp = ctx->ifc_ifp; 985 na.na_flags = NAF_BDG_MAYSLEEP; 986 MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); 987 MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); 988 989 na.num_tx_desc = scctx->isc_ntxd[0]; 990 na.num_rx_desc = scctx->isc_nrxd[0]; 991 na.nm_txsync = iflib_netmap_txsync; 992 na.nm_rxsync = iflib_netmap_rxsync; 993 na.nm_register = iflib_netmap_register; 994 na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; 995 na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; 996 return (netmap_attach(&na)); 997 } 998 999 static void 1000 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq) 1001 { 1002 struct netmap_adapter *na = NA(ctx->ifc_ifp); 1003 struct netmap_slot *slot; 1004 1005 slot = netmap_reset(na, NR_TX, txq->ift_id, 0); 1006 if (slot == 0) 1007 return; 1008 1009 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) { 1010 1011 /* 1012 * In netmap mode, set the map for the packet buffer. 1013 * NOTE: Some drivers (not this one) also need to set 1014 * the physical buffer address in the NIC ring. 1015 * netmap_idx_n2k() maps a nic index, i, into the corresponding 1016 * netmap slot index, si 1017 */ 1018 int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i); 1019 netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si)); 1020 } 1021 } 1022 static void 1023 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq) 1024 { 1025 struct netmap_adapter *na = NA(ctx->ifc_ifp); 1026 struct netmap_slot *slot; 1027 bus_dmamap_t *map; 1028 int nrxd; 1029 1030 slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0); 1031 if (slot == 0) 1032 return; 1033 map = rxq->ifr_fl[0].ifl_sds.ifsd_map; 1034 nrxd = ctx->ifc_softc_ctx.isc_nrxd[0]; 1035 for (int i = 0; i < nrxd; i++, map++) { 1036 int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i); 1037 uint64_t paddr; 1038 void *addr; 1039 caddr_t vaddr; 1040 1041 vaddr = addr = PNMB(na, slot + sj, &paddr); 1042 netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr); 1043 /* Update descriptor and the cached value */ 1044 ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1, rxq->ifr_fl[0].ifl_buf_size); 1045 } 1046 /* preserve queue */ 1047 if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) { 1048 struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id]; 1049 int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring); 1050 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t); 1051 } else 1052 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1); 1053 } 1054 1055 #define iflib_netmap_detach(ifp) netmap_detach(ifp) 1056 1057 #else 1058 #define iflib_netmap_txq_init(ctx, txq) 1059 #define iflib_netmap_rxq_init(ctx, rxq) 1060 #define iflib_netmap_detach(ifp) 1061 1062 #define iflib_netmap_attach(ctx) (0) 1063 #define netmap_rx_irq(ifp, qid, budget) (0) 1064 1065 #endif 1066 1067 #if defined(__i386__) || defined(__amd64__) 1068 static __inline void 1069 prefetch(void *x) 1070 { 1071 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); 1072 } 1073 #else 1074 #define prefetch(x) 1075 #endif 1076 1077 static void 1078 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err) 1079 { 1080 if (err) 1081 return; 1082 *(bus_addr_t *) arg = segs[0].ds_addr; 1083 } 1084 1085 int 1086 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags) 1087 { 1088 int err; 1089 if_shared_ctx_t sctx = ctx->ifc_sctx; 1090 device_t dev = ctx->ifc_dev; 1091 1092 KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized")); 1093 1094 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1095 sctx->isc_q_align, 0, /* alignment, bounds */ 1096 BUS_SPACE_MAXADDR, /* lowaddr */ 1097 BUS_SPACE_MAXADDR, /* highaddr */ 1098 NULL, NULL, /* filter, filterarg */ 1099 size, /* maxsize */ 1100 1, /* nsegments */ 1101 size, /* maxsegsize */ 1102 BUS_DMA_ALLOCNOW, /* flags */ 1103 NULL, /* lockfunc */ 1104 NULL, /* lockarg */ 1105 &dma->idi_tag); 1106 if (err) { 1107 device_printf(dev, 1108 "%s: bus_dma_tag_create failed: %d\n", 1109 __func__, err); 1110 goto fail_0; 1111 } 1112 1113 err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr, 1114 BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map); 1115 if (err) { 1116 device_printf(dev, 1117 "%s: bus_dmamem_alloc(%ju) failed: %d\n", 1118 __func__, (uintmax_t)size, err); 1119 goto fail_1; 1120 } 1121 1122 dma->idi_paddr = IF_BAD_DMA; 1123 err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr, 1124 size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT); 1125 if (err || dma->idi_paddr == IF_BAD_DMA) { 1126 device_printf(dev, 1127 "%s: bus_dmamap_load failed: %d\n", 1128 __func__, err); 1129 goto fail_2; 1130 } 1131 1132 dma->idi_size = size; 1133 return (0); 1134 1135 fail_2: 1136 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); 1137 fail_1: 1138 bus_dma_tag_destroy(dma->idi_tag); 1139 fail_0: 1140 dma->idi_tag = NULL; 1141 1142 return (err); 1143 } 1144 1145 int 1146 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count) 1147 { 1148 int i, err; 1149 iflib_dma_info_t *dmaiter; 1150 1151 dmaiter = dmalist; 1152 for (i = 0; i < count; i++, dmaiter++) { 1153 if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0) 1154 break; 1155 } 1156 if (err) 1157 iflib_dma_free_multi(dmalist, i); 1158 return (err); 1159 } 1160 1161 void 1162 iflib_dma_free(iflib_dma_info_t dma) 1163 { 1164 if (dma->idi_tag == NULL) 1165 return; 1166 if (dma->idi_paddr != IF_BAD_DMA) { 1167 bus_dmamap_sync(dma->idi_tag, dma->idi_map, 1168 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 1169 bus_dmamap_unload(dma->idi_tag, dma->idi_map); 1170 dma->idi_paddr = IF_BAD_DMA; 1171 } 1172 if (dma->idi_vaddr != NULL) { 1173 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); 1174 dma->idi_vaddr = NULL; 1175 } 1176 bus_dma_tag_destroy(dma->idi_tag); 1177 dma->idi_tag = NULL; 1178 } 1179 1180 void 1181 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count) 1182 { 1183 int i; 1184 iflib_dma_info_t *dmaiter = dmalist; 1185 1186 for (i = 0; i < count; i++, dmaiter++) 1187 iflib_dma_free(*dmaiter); 1188 } 1189 1190 #ifdef EARLY_AP_STARTUP 1191 static const int iflib_started = 1; 1192 #else 1193 /* 1194 * We used to abuse the smp_started flag to decide if the queues have been 1195 * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()). 1196 * That gave bad races, since the SYSINIT() runs strictly after smp_started 1197 * is set. Run a SYSINIT() strictly after that to just set a usable 1198 * completion flag. 1199 */ 1200 1201 static int iflib_started; 1202 1203 static void 1204 iflib_record_started(void *arg) 1205 { 1206 iflib_started = 1; 1207 } 1208 1209 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST, 1210 iflib_record_started, NULL); 1211 #endif 1212 1213 static int 1214 iflib_fast_intr(void *arg) 1215 { 1216 iflib_filter_info_t info = arg; 1217 struct grouptask *gtask = info->ifi_task; 1218 1219 if (!iflib_started) 1220 return (FILTER_HANDLED); 1221 1222 DBG_COUNTER_INC(fast_intrs); 1223 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED) 1224 return (FILTER_HANDLED); 1225 1226 GROUPTASK_ENQUEUE(gtask); 1227 return (FILTER_HANDLED); 1228 } 1229 1230 static int 1231 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, 1232 driver_filter_t filter, driver_intr_t handler, void *arg, 1233 char *name) 1234 { 1235 int rc; 1236 struct resource *res; 1237 void *tag; 1238 device_t dev = ctx->ifc_dev; 1239 1240 MPASS(rid < 512); 1241 irq->ii_rid = rid; 1242 res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, 1243 RF_SHAREABLE | RF_ACTIVE); 1244 if (res == NULL) { 1245 device_printf(dev, 1246 "failed to allocate IRQ for rid %d, name %s.\n", rid, name); 1247 return (ENOMEM); 1248 } 1249 irq->ii_res = res; 1250 KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL")); 1251 rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET, 1252 filter, handler, arg, &tag); 1253 if (rc != 0) { 1254 device_printf(dev, 1255 "failed to setup interrupt for rid %d, name %s: %d\n", 1256 rid, name ? name : "unknown", rc); 1257 return (rc); 1258 } else if (name) 1259 bus_describe_intr(dev, res, tag, "%s", name); 1260 1261 irq->ii_tag = tag; 1262 return (0); 1263 } 1264 1265 1266 /********************************************************************* 1267 * 1268 * Allocate memory for tx_buffer structures. The tx_buffer stores all 1269 * the information needed to transmit a packet on the wire. This is 1270 * called only once at attach, setup is done every reset. 1271 * 1272 **********************************************************************/ 1273 1274 static int 1275 iflib_txsd_alloc(iflib_txq_t txq) 1276 { 1277 if_ctx_t ctx = txq->ift_ctx; 1278 if_shared_ctx_t sctx = ctx->ifc_sctx; 1279 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1280 device_t dev = ctx->ifc_dev; 1281 int err, nsegments, ntsosegments; 1282 1283 nsegments = scctx->isc_tx_nsegments; 1284 ntsosegments = scctx->isc_tx_tso_segments_max; 1285 MPASS(scctx->isc_ntxd[0] > 0); 1286 MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0); 1287 MPASS(nsegments > 0); 1288 MPASS(ntsosegments > 0); 1289 /* 1290 * Setup DMA descriptor areas. 1291 */ 1292 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1293 1, 0, /* alignment, bounds */ 1294 BUS_SPACE_MAXADDR, /* lowaddr */ 1295 BUS_SPACE_MAXADDR, /* highaddr */ 1296 NULL, NULL, /* filter, filterarg */ 1297 sctx->isc_tx_maxsize, /* maxsize */ 1298 nsegments, /* nsegments */ 1299 sctx->isc_tx_maxsegsize, /* maxsegsize */ 1300 0, /* flags */ 1301 NULL, /* lockfunc */ 1302 NULL, /* lockfuncarg */ 1303 &txq->ift_desc_tag))) { 1304 device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err); 1305 device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n", 1306 sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize); 1307 goto fail; 1308 } 1309 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1310 1, 0, /* alignment, bounds */ 1311 BUS_SPACE_MAXADDR, /* lowaddr */ 1312 BUS_SPACE_MAXADDR, /* highaddr */ 1313 NULL, NULL, /* filter, filterarg */ 1314 scctx->isc_tx_tso_size_max, /* maxsize */ 1315 ntsosegments, /* nsegments */ 1316 scctx->isc_tx_tso_segsize_max, /* maxsegsize */ 1317 0, /* flags */ 1318 NULL, /* lockfunc */ 1319 NULL, /* lockfuncarg */ 1320 &txq->ift_tso_desc_tag))) { 1321 device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err); 1322 1323 goto fail; 1324 } 1325 if (!(txq->ift_sds.ifsd_flags = 1326 (uint8_t *) malloc(sizeof(uint8_t) * 1327 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1328 device_printf(dev, "Unable to allocate tx_buffer memory\n"); 1329 err = ENOMEM; 1330 goto fail; 1331 } 1332 if (!(txq->ift_sds.ifsd_m = 1333 (struct mbuf **) malloc(sizeof(struct mbuf *) * 1334 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1335 device_printf(dev, "Unable to allocate tx_buffer memory\n"); 1336 err = ENOMEM; 1337 goto fail; 1338 } 1339 1340 /* Create the descriptor buffer dma maps */ 1341 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__))) 1342 if ((ctx->ifc_flags & IFC_DMAR) == 0) 1343 return (0); 1344 1345 if (!(txq->ift_sds.ifsd_map = 1346 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1347 device_printf(dev, "Unable to allocate tx_buffer map memory\n"); 1348 err = ENOMEM; 1349 goto fail; 1350 } 1351 1352 for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) { 1353 err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]); 1354 if (err != 0) { 1355 device_printf(dev, "Unable to create TX DMA map\n"); 1356 goto fail; 1357 } 1358 } 1359 #endif 1360 return (0); 1361 fail: 1362 /* We free all, it handles case where we are in the middle */ 1363 iflib_tx_structures_free(ctx); 1364 return (err); 1365 } 1366 1367 static void 1368 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i) 1369 { 1370 bus_dmamap_t map; 1371 1372 map = NULL; 1373 if (txq->ift_sds.ifsd_map != NULL) 1374 map = txq->ift_sds.ifsd_map[i]; 1375 if (map != NULL) { 1376 bus_dmamap_unload(txq->ift_desc_tag, map); 1377 bus_dmamap_destroy(txq->ift_desc_tag, map); 1378 txq->ift_sds.ifsd_map[i] = NULL; 1379 } 1380 } 1381 1382 static void 1383 iflib_txq_destroy(iflib_txq_t txq) 1384 { 1385 if_ctx_t ctx = txq->ift_ctx; 1386 1387 for (int i = 0; i < txq->ift_size; i++) 1388 iflib_txsd_destroy(ctx, txq, i); 1389 if (txq->ift_sds.ifsd_map != NULL) { 1390 free(txq->ift_sds.ifsd_map, M_IFLIB); 1391 txq->ift_sds.ifsd_map = NULL; 1392 } 1393 if (txq->ift_sds.ifsd_m != NULL) { 1394 free(txq->ift_sds.ifsd_m, M_IFLIB); 1395 txq->ift_sds.ifsd_m = NULL; 1396 } 1397 if (txq->ift_sds.ifsd_flags != NULL) { 1398 free(txq->ift_sds.ifsd_flags, M_IFLIB); 1399 txq->ift_sds.ifsd_flags = NULL; 1400 } 1401 if (txq->ift_desc_tag != NULL) { 1402 bus_dma_tag_destroy(txq->ift_desc_tag); 1403 txq->ift_desc_tag = NULL; 1404 } 1405 if (txq->ift_tso_desc_tag != NULL) { 1406 bus_dma_tag_destroy(txq->ift_tso_desc_tag); 1407 txq->ift_tso_desc_tag = NULL; 1408 } 1409 } 1410 1411 static void 1412 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i) 1413 { 1414 struct mbuf **mp; 1415 1416 mp = &txq->ift_sds.ifsd_m[i]; 1417 if (*mp == NULL) 1418 return; 1419 1420 if (txq->ift_sds.ifsd_map != NULL) { 1421 bus_dmamap_sync(txq->ift_desc_tag, 1422 txq->ift_sds.ifsd_map[i], 1423 BUS_DMASYNC_POSTWRITE); 1424 bus_dmamap_unload(txq->ift_desc_tag, 1425 txq->ift_sds.ifsd_map[i]); 1426 } 1427 m_free(*mp); 1428 DBG_COUNTER_INC(tx_frees); 1429 *mp = NULL; 1430 } 1431 1432 static int 1433 iflib_txq_setup(iflib_txq_t txq) 1434 { 1435 if_ctx_t ctx = txq->ift_ctx; 1436 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1437 iflib_dma_info_t di; 1438 int i; 1439 1440 /* Set number of descriptors available */ 1441 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 1442 1443 /* Reset indices */ 1444 txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; 1445 txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset]; 1446 1447 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++) 1448 bzero((void *)di->idi_vaddr, di->idi_size); 1449 1450 IFDI_TXQ_SETUP(ctx, txq->ift_id); 1451 for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++) 1452 bus_dmamap_sync(di->idi_tag, di->idi_map, 1453 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 1454 return (0); 1455 } 1456 1457 /********************************************************************* 1458 * 1459 * Allocate memory for rx_buffer structures. Since we use one 1460 * rx_buffer per received packet, the maximum number of rx_buffer's 1461 * that we'll need is equal to the number of receive descriptors 1462 * that we've allocated. 1463 * 1464 **********************************************************************/ 1465 static int 1466 iflib_rxsd_alloc(iflib_rxq_t rxq) 1467 { 1468 if_ctx_t ctx = rxq->ifr_ctx; 1469 if_shared_ctx_t sctx = ctx->ifc_sctx; 1470 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1471 device_t dev = ctx->ifc_dev; 1472 iflib_fl_t fl; 1473 int err; 1474 1475 MPASS(scctx->isc_nrxd[0] > 0); 1476 MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0); 1477 1478 fl = rxq->ifr_fl; 1479 for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { 1480 fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */ 1481 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1482 1, 0, /* alignment, bounds */ 1483 BUS_SPACE_MAXADDR, /* lowaddr */ 1484 BUS_SPACE_MAXADDR, /* highaddr */ 1485 NULL, NULL, /* filter, filterarg */ 1486 sctx->isc_rx_maxsize, /* maxsize */ 1487 sctx->isc_rx_nsegments, /* nsegments */ 1488 sctx->isc_rx_maxsegsize, /* maxsegsize */ 1489 0, /* flags */ 1490 NULL, /* lockfunc */ 1491 NULL, /* lockarg */ 1492 &fl->ifl_desc_tag); 1493 if (err) { 1494 device_printf(dev, "%s: bus_dma_tag_create failed %d\n", 1495 __func__, err); 1496 goto fail; 1497 } 1498 if (!(fl->ifl_sds.ifsd_flags = 1499 (uint8_t *) malloc(sizeof(uint8_t) * 1500 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1501 device_printf(dev, "Unable to allocate tx_buffer memory\n"); 1502 err = ENOMEM; 1503 goto fail; 1504 } 1505 if (!(fl->ifl_sds.ifsd_m = 1506 (struct mbuf **) malloc(sizeof(struct mbuf *) * 1507 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1508 device_printf(dev, "Unable to allocate tx_buffer memory\n"); 1509 err = ENOMEM; 1510 goto fail; 1511 } 1512 if (!(fl->ifl_sds.ifsd_cl = 1513 (caddr_t *) malloc(sizeof(caddr_t) * 1514 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1515 device_printf(dev, "Unable to allocate tx_buffer memory\n"); 1516 err = ENOMEM; 1517 goto fail; 1518 } 1519 1520 /* Create the descriptor buffer dma maps */ 1521 #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__))) 1522 if ((ctx->ifc_flags & IFC_DMAR) == 0) 1523 continue; 1524 1525 if (!(fl->ifl_sds.ifsd_map = 1526 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1527 device_printf(dev, "Unable to allocate tx_buffer map memory\n"); 1528 err = ENOMEM; 1529 goto fail; 1530 } 1531 1532 for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) { 1533 err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]); 1534 if (err != 0) { 1535 device_printf(dev, "Unable to create TX DMA map\n"); 1536 goto fail; 1537 } 1538 } 1539 #endif 1540 } 1541 return (0); 1542 1543 fail: 1544 iflib_rx_structures_free(ctx); 1545 return (err); 1546 } 1547 1548 1549 /* 1550 * Internal service routines 1551 */ 1552 1553 struct rxq_refill_cb_arg { 1554 int error; 1555 bus_dma_segment_t seg; 1556 int nseg; 1557 }; 1558 1559 static void 1560 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) 1561 { 1562 struct rxq_refill_cb_arg *cb_arg = arg; 1563 1564 cb_arg->error = error; 1565 cb_arg->seg = segs[0]; 1566 cb_arg->nseg = nseg; 1567 } 1568 1569 1570 #ifdef ACPI_DMAR 1571 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR) 1572 #else 1573 #define IS_DMAR(ctx) (0) 1574 #endif 1575 1576 /** 1577 * rxq_refill - refill an rxq free-buffer list 1578 * @ctx: the iflib context 1579 * @rxq: the free-list to refill 1580 * @n: the number of new buffers to allocate 1581 * 1582 * (Re)populate an rxq free-buffer list with up to @n new packet buffers. 1583 * The caller must assure that @n does not exceed the queue's capacity. 1584 */ 1585 static void 1586 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count) 1587 { 1588 struct mbuf *m; 1589 int idx, pidx = fl->ifl_pidx; 1590 caddr_t cl, *sd_cl; 1591 struct mbuf **sd_m; 1592 uint8_t *sd_flags; 1593 bus_dmamap_t *sd_map; 1594 int n, i = 0; 1595 uint64_t bus_addr; 1596 int err; 1597 1598 sd_m = fl->ifl_sds.ifsd_m; 1599 sd_map = fl->ifl_sds.ifsd_map; 1600 sd_cl = fl->ifl_sds.ifsd_cl; 1601 sd_flags = fl->ifl_sds.ifsd_flags; 1602 idx = pidx; 1603 1604 n = count; 1605 MPASS(n > 0); 1606 MPASS(fl->ifl_credits + n <= fl->ifl_size); 1607 1608 if (pidx < fl->ifl_cidx) 1609 MPASS(pidx + n <= fl->ifl_cidx); 1610 if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size)) 1611 MPASS(fl->ifl_gen == 0); 1612 if (pidx > fl->ifl_cidx) 1613 MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx); 1614 1615 DBG_COUNTER_INC(fl_refills); 1616 if (n > 8) 1617 DBG_COUNTER_INC(fl_refills_large); 1618 1619 while (n--) { 1620 /* 1621 * We allocate an uninitialized mbuf + cluster, mbuf is 1622 * initialized after rx. 1623 * 1624 * If the cluster is still set then we know a minimum sized packet was received 1625 */ 1626 if ((cl = sd_cl[idx]) == NULL) { 1627 if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL) 1628 break; 1629 #if MEMORY_LOGGING 1630 fl->ifl_cl_enqueued++; 1631 #endif 1632 } 1633 if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) { 1634 break; 1635 } 1636 #if MEMORY_LOGGING 1637 fl->ifl_m_enqueued++; 1638 #endif 1639 1640 DBG_COUNTER_INC(rx_allocs); 1641 #ifdef notyet 1642 if ((sd_flags[pidx] & RX_SW_DESC_MAP_CREATED) == 0) { 1643 int err; 1644 1645 if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &sd_map[idx]))) { 1646 log(LOG_WARNING, "bus_dmamap_create failed %d\n", err); 1647 uma_zfree(fl->ifl_zone, cl); 1648 n = 0; 1649 goto done; 1650 } 1651 sd_flags[idx] |= RX_SW_DESC_MAP_CREATED; 1652 } 1653 #endif 1654 #if defined(__i386__) || defined(__amd64__) 1655 if (!IS_DMAR(ctx)) { 1656 bus_addr = pmap_kextract((vm_offset_t)cl); 1657 } else 1658 #endif 1659 { 1660 struct rxq_refill_cb_arg cb_arg; 1661 iflib_rxq_t q; 1662 1663 cb_arg.error = 0; 1664 q = fl->ifl_rxq; 1665 err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx], 1666 cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0); 1667 1668 if (err != 0 || cb_arg.error) { 1669 /* 1670 * !zone_pack ? 1671 */ 1672 if (fl->ifl_zone == zone_pack) 1673 uma_zfree(fl->ifl_zone, cl); 1674 m_free(m); 1675 n = 0; 1676 goto done; 1677 } 1678 bus_addr = cb_arg.seg.ds_addr; 1679 } 1680 sd_flags[idx] |= RX_SW_DESC_INUSE; 1681 1682 MPASS(sd_m[idx] == NULL); 1683 sd_cl[idx] = cl; 1684 sd_m[idx] = m; 1685 fl->ifl_bus_addrs[i] = bus_addr; 1686 fl->ifl_vm_addrs[i] = cl; 1687 fl->ifl_credits++; 1688 i++; 1689 MPASS(fl->ifl_credits <= fl->ifl_size); 1690 if (++idx == fl->ifl_size) { 1691 fl->ifl_gen = 1; 1692 idx = 0; 1693 } 1694 if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { 1695 ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx, 1696 fl->ifl_bus_addrs, fl->ifl_vm_addrs, i, fl->ifl_buf_size); 1697 i = 0; 1698 pidx = idx; 1699 } 1700 fl->ifl_pidx = idx; 1701 1702 } 1703 done: 1704 DBG_COUNTER_INC(rxd_flush); 1705 if (fl->ifl_pidx == 0) 1706 pidx = fl->ifl_size - 1; 1707 else 1708 pidx = fl->ifl_pidx - 1; 1709 ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx); 1710 } 1711 1712 static __inline void 1713 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max) 1714 { 1715 /* we avoid allowing pidx to catch up with cidx as it confuses ixl */ 1716 int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1; 1717 #ifdef INVARIANTS 1718 int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1; 1719 #endif 1720 1721 MPASS(fl->ifl_credits <= fl->ifl_size); 1722 MPASS(reclaimable == delta); 1723 1724 if (reclaimable > 0) 1725 _iflib_fl_refill(ctx, fl, min(max, reclaimable)); 1726 } 1727 1728 static void 1729 iflib_fl_bufs_free(iflib_fl_t fl) 1730 { 1731 iflib_dma_info_t idi = fl->ifl_ifdi; 1732 uint32_t i; 1733 1734 for (i = 0; i < fl->ifl_size; i++) { 1735 struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i]; 1736 uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i]; 1737 caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i]; 1738 1739 if (*sd_flags & RX_SW_DESC_INUSE) { 1740 if (fl->ifl_sds.ifsd_map != NULL) { 1741 bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i]; 1742 bus_dmamap_unload(fl->ifl_desc_tag, sd_map); 1743 bus_dmamap_destroy(fl->ifl_desc_tag, sd_map); 1744 } 1745 if (*sd_m != NULL) { 1746 m_init(*sd_m, M_NOWAIT, MT_DATA, 0); 1747 uma_zfree(zone_mbuf, *sd_m); 1748 } 1749 if (*sd_cl != NULL) 1750 uma_zfree(fl->ifl_zone, *sd_cl); 1751 *sd_flags = 0; 1752 } else { 1753 MPASS(*sd_cl == NULL); 1754 MPASS(*sd_m == NULL); 1755 } 1756 #if MEMORY_LOGGING 1757 fl->ifl_m_dequeued++; 1758 fl->ifl_cl_dequeued++; 1759 #endif 1760 *sd_cl = NULL; 1761 *sd_m = NULL; 1762 } 1763 /* 1764 * Reset free list values 1765 */ 1766 fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;; 1767 bzero(idi->idi_vaddr, idi->idi_size); 1768 } 1769 1770 /********************************************************************* 1771 * 1772 * Initialize a receive ring and its buffers. 1773 * 1774 **********************************************************************/ 1775 static int 1776 iflib_fl_setup(iflib_fl_t fl) 1777 { 1778 iflib_rxq_t rxq = fl->ifl_rxq; 1779 if_ctx_t ctx = rxq->ifr_ctx; 1780 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 1781 1782 /* 1783 ** Free current RX buffer structs and their mbufs 1784 */ 1785 iflib_fl_bufs_free(fl); 1786 /* Now replenish the mbufs */ 1787 MPASS(fl->ifl_credits == 0); 1788 /* 1789 * XXX don't set the max_frame_size to larger 1790 * than the hardware can handle 1791 */ 1792 if (sctx->isc_max_frame_size <= 2048) 1793 fl->ifl_buf_size = MCLBYTES; 1794 else if (sctx->isc_max_frame_size <= 4096) 1795 fl->ifl_buf_size = MJUMPAGESIZE; 1796 else if (sctx->isc_max_frame_size <= 9216) 1797 fl->ifl_buf_size = MJUM9BYTES; 1798 else 1799 fl->ifl_buf_size = MJUM16BYTES; 1800 if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size) 1801 ctx->ifc_max_fl_buf_size = fl->ifl_buf_size; 1802 fl->ifl_cltype = m_gettype(fl->ifl_buf_size); 1803 fl->ifl_zone = m_getzone(fl->ifl_buf_size); 1804 1805 1806 /* avoid pre-allocating zillions of clusters to an idle card 1807 * potentially speeding up attach 1808 */ 1809 _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size)); 1810 MPASS(min(128, fl->ifl_size) == fl->ifl_credits); 1811 if (min(128, fl->ifl_size) != fl->ifl_credits) 1812 return (ENOBUFS); 1813 /* 1814 * handle failure 1815 */ 1816 MPASS(rxq != NULL); 1817 MPASS(fl->ifl_ifdi != NULL); 1818 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 1819 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 1820 return (0); 1821 } 1822 1823 /********************************************************************* 1824 * 1825 * Free receive ring data structures 1826 * 1827 **********************************************************************/ 1828 static void 1829 iflib_rx_sds_free(iflib_rxq_t rxq) 1830 { 1831 iflib_fl_t fl; 1832 int i; 1833 1834 if (rxq->ifr_fl != NULL) { 1835 for (i = 0; i < rxq->ifr_nfl; i++) { 1836 fl = &rxq->ifr_fl[i]; 1837 if (fl->ifl_desc_tag != NULL) { 1838 bus_dma_tag_destroy(fl->ifl_desc_tag); 1839 fl->ifl_desc_tag = NULL; 1840 } 1841 free(fl->ifl_sds.ifsd_m, M_IFLIB); 1842 free(fl->ifl_sds.ifsd_cl, M_IFLIB); 1843 /* XXX destroy maps first */ 1844 free(fl->ifl_sds.ifsd_map, M_IFLIB); 1845 fl->ifl_sds.ifsd_m = NULL; 1846 fl->ifl_sds.ifsd_cl = NULL; 1847 fl->ifl_sds.ifsd_map = NULL; 1848 } 1849 free(rxq->ifr_fl, M_IFLIB); 1850 rxq->ifr_fl = NULL; 1851 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0; 1852 } 1853 } 1854 1855 /* 1856 * MI independent logic 1857 * 1858 */ 1859 static void 1860 iflib_timer(void *arg) 1861 { 1862 iflib_txq_t txq = arg; 1863 if_ctx_t ctx = txq->ift_ctx; 1864 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1865 1866 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) 1867 return; 1868 /* 1869 ** Check on the state of the TX queue(s), this 1870 ** can be done without the lock because its RO 1871 ** and the HUNG state will be static if set. 1872 */ 1873 IFDI_TIMER(ctx, txq->ift_id); 1874 if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) && 1875 (ctx->ifc_pause_frames == 0)) 1876 goto hung; 1877 1878 if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments || 1879 ifmp_ring_is_stalled(txq->ift_br[0])) 1880 GROUPTASK_ENQUEUE(&txq->ift_task); 1881 1882 ctx->ifc_pause_frames = 0; 1883 if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) 1884 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); 1885 return; 1886 hung: 1887 CTX_LOCK(ctx); 1888 if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING); 1889 device_printf(ctx->ifc_dev, "TX(%d) desc avail = %d, pidx = %d\n", 1890 txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx); 1891 1892 IFDI_WATCHDOG_RESET(ctx); 1893 ctx->ifc_watchdog_events++; 1894 ctx->ifc_pause_frames = 0; 1895 1896 iflib_init_locked(ctx); 1897 CTX_UNLOCK(ctx); 1898 } 1899 1900 static void 1901 iflib_init_locked(if_ctx_t ctx) 1902 { 1903 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 1904 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1905 if_t ifp = ctx->ifc_ifp; 1906 iflib_fl_t fl; 1907 iflib_txq_t txq; 1908 iflib_rxq_t rxq; 1909 int i, j, tx_ip_csum_flags, tx_ip6_csum_flags; 1910 1911 1912 if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 1913 IFDI_INTR_DISABLE(ctx); 1914 1915 tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP); 1916 tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP); 1917 /* Set hardware offload abilities */ 1918 if_clearhwassist(ifp); 1919 if (if_getcapenable(ifp) & IFCAP_TXCSUM) 1920 if_sethwassistbits(ifp, tx_ip_csum_flags, 0); 1921 if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6) 1922 if_sethwassistbits(ifp, tx_ip6_csum_flags, 0); 1923 if (if_getcapenable(ifp) & IFCAP_TSO4) 1924 if_sethwassistbits(ifp, CSUM_IP_TSO, 0); 1925 if (if_getcapenable(ifp) & IFCAP_TSO6) 1926 if_sethwassistbits(ifp, CSUM_IP6_TSO, 0); 1927 1928 for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) { 1929 CALLOUT_LOCK(txq); 1930 callout_stop(&txq->ift_timer); 1931 callout_stop(&txq->ift_db_check); 1932 CALLOUT_UNLOCK(txq); 1933 iflib_netmap_txq_init(ctx, txq); 1934 } 1935 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { 1936 iflib_netmap_rxq_init(ctx, rxq); 1937 } 1938 #ifdef INVARIANTS 1939 i = if_getdrvflags(ifp); 1940 #endif 1941 IFDI_INIT(ctx); 1942 MPASS(if_getdrvflags(ifp) == i); 1943 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { 1944 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { 1945 if (iflib_fl_setup(fl)) { 1946 device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n"); 1947 goto done; 1948 } 1949 } 1950 } 1951 done: 1952 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 1953 IFDI_INTR_ENABLE(ctx); 1954 txq = ctx->ifc_txqs; 1955 for (i = 0; i < sctx->isc_ntxqsets; i++, txq++) 1956 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, 1957 txq->ift_timer.c_cpu); 1958 } 1959 1960 static int 1961 iflib_media_change(if_t ifp) 1962 { 1963 if_ctx_t ctx = if_getsoftc(ifp); 1964 int err; 1965 1966 CTX_LOCK(ctx); 1967 if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0) 1968 iflib_init_locked(ctx); 1969 CTX_UNLOCK(ctx); 1970 return (err); 1971 } 1972 1973 static void 1974 iflib_media_status(if_t ifp, struct ifmediareq *ifmr) 1975 { 1976 if_ctx_t ctx = if_getsoftc(ifp); 1977 1978 CTX_LOCK(ctx); 1979 IFDI_UPDATE_ADMIN_STATUS(ctx); 1980 IFDI_MEDIA_STATUS(ctx, ifmr); 1981 CTX_UNLOCK(ctx); 1982 } 1983 1984 static void 1985 iflib_stop(if_ctx_t ctx) 1986 { 1987 iflib_txq_t txq = ctx->ifc_txqs; 1988 iflib_rxq_t rxq = ctx->ifc_rxqs; 1989 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1990 iflib_dma_info_t di; 1991 iflib_fl_t fl; 1992 int i, j; 1993 1994 /* Tell the stack that the interface is no longer active */ 1995 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 1996 1997 IFDI_INTR_DISABLE(ctx); 1998 DELAY(100000); 1999 IFDI_STOP(ctx); 2000 DELAY(100000); 2001 2002 iflib_debug_reset(); 2003 /* Wait for current tx queue users to exit to disarm watchdog timer. */ 2004 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) { 2005 /* make sure all transmitters have completed before proceeding XXX */ 2006 2007 /* clean any enqueued buffers */ 2008 iflib_ifmp_purge(txq); 2009 /* Free any existing tx buffers. */ 2010 for (j = 0; j < txq->ift_size; j++) { 2011 iflib_txsd_free(ctx, txq, j); 2012 } 2013 txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; 2014 txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0; 2015 txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0; 2016 txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0; 2017 txq->ift_pullups = 0; 2018 ifmp_ring_reset_stats(txq->ift_br[0]); 2019 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++) 2020 bzero((void *)di->idi_vaddr, di->idi_size); 2021 } 2022 for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) { 2023 /* make sure all transmitters have completed before proceeding XXX */ 2024 2025 for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++) 2026 bzero((void *)di->idi_vaddr, di->idi_size); 2027 /* also resets the free lists pidx/cidx */ 2028 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) 2029 iflib_fl_bufs_free(fl); 2030 } 2031 } 2032 2033 static inline void 2034 prefetch_pkts(iflib_fl_t fl, int cidx) 2035 { 2036 int nextptr; 2037 int nrxd = fl->ifl_size; 2038 2039 nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1); 2040 prefetch(&fl->ifl_sds.ifsd_m[nextptr]); 2041 prefetch(&fl->ifl_sds.ifsd_cl[nextptr]); 2042 prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]); 2043 prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]); 2044 prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]); 2045 prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]); 2046 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]); 2047 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]); 2048 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]); 2049 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]); 2050 } 2051 2052 static void 2053 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload, iflib_fl_t *pfl, int *pcidx) 2054 { 2055 int flid, cidx; 2056 bus_dmamap_t map; 2057 iflib_fl_t fl; 2058 iflib_dma_info_t di; 2059 int next; 2060 2061 flid = irf->irf_flid; 2062 cidx = irf->irf_idx; 2063 fl = &rxq->ifr_fl[flid]; 2064 fl->ifl_credits--; 2065 #if MEMORY_LOGGING 2066 fl->ifl_m_dequeued++; 2067 if (cltype) 2068 fl->ifl_cl_dequeued++; 2069 #endif 2070 prefetch_pkts(fl, cidx); 2071 if (fl->ifl_sds.ifsd_map != NULL) { 2072 next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1); 2073 prefetch(&fl->ifl_sds.ifsd_map[next]); 2074 map = fl->ifl_sds.ifsd_map[cidx]; 2075 di = fl->ifl_ifdi; 2076 next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1); 2077 prefetch(&fl->ifl_sds.ifsd_flags[next]); 2078 bus_dmamap_sync(di->idi_tag, di->idi_map, 2079 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 2080 2081 /* not valid assert if bxe really does SGE from non-contiguous elements */ 2082 MPASS(fl->ifl_cidx == cidx); 2083 if (unload) 2084 bus_dmamap_unload(fl->ifl_desc_tag, map); 2085 } 2086 if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) { 2087 fl->ifl_cidx = 0; 2088 fl->ifl_gen = 0; 2089 } 2090 /* YES ick */ 2091 if (cltype) 2092 *cltype = fl->ifl_cltype; 2093 *pfl = fl; 2094 *pcidx = cidx; 2095 } 2096 2097 static struct mbuf * 2098 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri) 2099 { 2100 int i, padlen , flags, cltype; 2101 struct mbuf *m, *mh, *mt, *sd_m; 2102 iflib_fl_t fl; 2103 int cidx; 2104 caddr_t cl, sd_cl; 2105 2106 i = 0; 2107 mh = NULL; 2108 do { 2109 rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE, &fl, &cidx); 2110 sd_m = fl->ifl_sds.ifsd_m[cidx]; 2111 sd_cl = fl->ifl_sds.ifsd_cl[cidx]; 2112 2113 MPASS(sd_cl != NULL); 2114 MPASS(sd_m != NULL); 2115 2116 /* Don't include zero-length frags */ 2117 if (ri->iri_frags[i].irf_len == 0) { 2118 /* XXX we can save the cluster here, but not the mbuf */ 2119 m_init(sd_m, M_NOWAIT, MT_DATA, 0); 2120 m_free(sd_m); 2121 fl->ifl_sds.ifsd_m[cidx] = NULL; 2122 continue; 2123 } 2124 m = sd_m; 2125 if (mh == NULL) { 2126 flags = M_PKTHDR|M_EXT; 2127 mh = mt = m; 2128 padlen = ri->iri_pad; 2129 } else { 2130 flags = M_EXT; 2131 mt->m_next = m; 2132 mt = m; 2133 /* assuming padding is only on the first fragment */ 2134 padlen = 0; 2135 } 2136 fl->ifl_sds.ifsd_m[cidx] = NULL; 2137 cl = fl->ifl_sds.ifsd_cl[cidx]; 2138 fl->ifl_sds.ifsd_cl[cidx] = NULL; 2139 2140 /* Can these two be made one ? */ 2141 m_init(m, M_NOWAIT, MT_DATA, flags); 2142 m_cljset(m, cl, cltype); 2143 /* 2144 * These must follow m_init and m_cljset 2145 */ 2146 m->m_data += padlen; 2147 ri->iri_len -= padlen; 2148 m->m_len = ri->iri_frags[i].irf_len; 2149 } while (++i < ri->iri_nfrags); 2150 2151 return (mh); 2152 } 2153 2154 /* 2155 * Process one software descriptor 2156 */ 2157 static struct mbuf * 2158 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri) 2159 { 2160 struct mbuf *m; 2161 iflib_fl_t fl; 2162 caddr_t sd_cl; 2163 int cidx; 2164 2165 /* should I merge this back in now that the two paths are basically duplicated? */ 2166 if (ri->iri_nfrags == 1 && 2167 ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) { 2168 rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE, &fl, &cidx); 2169 m = fl->ifl_sds.ifsd_m[cidx]; 2170 fl->ifl_sds.ifsd_m[cidx] = NULL; 2171 sd_cl = fl->ifl_sds.ifsd_cl[cidx]; 2172 m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); 2173 memcpy(m->m_data, sd_cl, ri->iri_len); 2174 m->m_len = ri->iri_frags[0].irf_len; 2175 } else { 2176 m = assemble_segments(rxq, ri); 2177 } 2178 m->m_pkthdr.len = ri->iri_len; 2179 m->m_pkthdr.rcvif = ri->iri_ifp; 2180 m->m_flags |= ri->iri_flags; 2181 m->m_pkthdr.ether_vtag = ri->iri_vtag; 2182 m->m_pkthdr.flowid = ri->iri_flowid; 2183 M_HASHTYPE_SET(m, ri->iri_rsstype); 2184 m->m_pkthdr.csum_flags = ri->iri_csum_flags; 2185 m->m_pkthdr.csum_data = ri->iri_csum_data; 2186 return (m); 2187 } 2188 2189 static bool 2190 iflib_rxeof(iflib_rxq_t rxq, int budget) 2191 { 2192 if_ctx_t ctx = rxq->ifr_ctx; 2193 if_shared_ctx_t sctx = ctx->ifc_sctx; 2194 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 2195 int avail, i; 2196 uint16_t *cidxp; 2197 struct if_rxd_info ri; 2198 int err, budget_left, rx_bytes, rx_pkts; 2199 iflib_fl_t fl; 2200 struct ifnet *ifp; 2201 int lro_enabled; 2202 /* 2203 * XXX early demux data packets so that if_input processing only handles 2204 * acks in interrupt context 2205 */ 2206 struct mbuf *m, *mh, *mt; 2207 2208 if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) { 2209 return (FALSE); 2210 } 2211 2212 mh = mt = NULL; 2213 MPASS(budget > 0); 2214 rx_pkts = rx_bytes = 0; 2215 if (sctx->isc_flags & IFLIB_HAS_RXCQ) 2216 cidxp = &rxq->ifr_cq_cidx; 2217 else 2218 cidxp = &rxq->ifr_fl[0].ifl_cidx; 2219 if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) { 2220 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) 2221 __iflib_fl_refill_lt(ctx, fl, budget + 8); 2222 DBG_COUNTER_INC(rx_unavail); 2223 return (false); 2224 } 2225 2226 for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) { 2227 if (__predict_false(!CTX_ACTIVE(ctx))) { 2228 DBG_COUNTER_INC(rx_ctx_inactive); 2229 break; 2230 } 2231 /* 2232 * Reset client set fields to their default values 2233 */ 2234 bzero(&ri, sizeof(ri)); 2235 ri.iri_qsidx = rxq->ifr_id; 2236 ri.iri_cidx = *cidxp; 2237 ri.iri_ifp = ctx->ifc_ifp; 2238 ri.iri_frags = rxq->ifr_frags; 2239 err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); 2240 2241 /* in lieu of handling correctly - make sure it isn't being unhandled */ 2242 MPASS(err == 0); 2243 if (sctx->isc_flags & IFLIB_HAS_RXCQ) { 2244 *cidxp = ri.iri_cidx; 2245 /* Update our consumer index */ 2246 while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) { 2247 rxq->ifr_cq_cidx -= scctx->isc_nrxd[0]; 2248 rxq->ifr_cq_gen = 0; 2249 } 2250 /* was this only a completion queue message? */ 2251 if (__predict_false(ri.iri_nfrags == 0)) 2252 continue; 2253 } 2254 MPASS(ri.iri_nfrags != 0); 2255 MPASS(ri.iri_len != 0); 2256 2257 /* will advance the cidx on the corresponding free lists */ 2258 m = iflib_rxd_pkt_get(rxq, &ri); 2259 if (avail == 0 && budget_left) 2260 avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left); 2261 2262 if (__predict_false(m == NULL)) { 2263 DBG_COUNTER_INC(rx_mbuf_null); 2264 continue; 2265 } 2266 /* imm_pkt: -- cxgb */ 2267 if (mh == NULL) 2268 mh = mt = m; 2269 else { 2270 mt->m_nextpkt = m; 2271 mt = m; 2272 } 2273 } 2274 /* make sure that we can refill faster than drain */ 2275 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) 2276 __iflib_fl_refill_lt(ctx, fl, budget + 8); 2277 2278 ifp = ctx->ifc_ifp; 2279 lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); 2280 while (mh != NULL) { 2281 m = mh; 2282 mh = mh->m_nextpkt; 2283 m->m_nextpkt = NULL; 2284 rx_bytes += m->m_pkthdr.len; 2285 rx_pkts++; 2286 #if defined(INET6) || defined(INET) 2287 if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0) 2288 continue; 2289 #endif 2290 DBG_COUNTER_INC(rx_if_input); 2291 ifp->if_input(ifp, m); 2292 } 2293 2294 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); 2295 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); 2296 2297 /* 2298 * Flush any outstanding LRO work 2299 */ 2300 #if defined(INET6) || defined(INET) 2301 tcp_lro_flush_all(&rxq->ifr_lc); 2302 #endif 2303 if (avail) 2304 return true; 2305 return (iflib_rxd_avail(ctx, rxq, *cidxp, 1)); 2306 } 2307 2308 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) 2309 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) 2310 #define TXQ_MAX_DB_DEFERRED(size) (size >> 5) 2311 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4) 2312 2313 static __inline void 2314 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring) 2315 { 2316 uint32_t dbval; 2317 2318 if (ring || txq->ift_db_pending >= 2319 TXQ_MAX_DB_DEFERRED(txq->ift_size)) { 2320 2321 /* the lock will only ever be contended in the !min_latency case */ 2322 if (!TXDB_TRYLOCK(txq)) 2323 return; 2324 dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx; 2325 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval); 2326 txq->ift_db_pending = txq->ift_npending = 0; 2327 TXDB_UNLOCK(txq); 2328 } 2329 } 2330 2331 static void 2332 iflib_txd_deferred_db_check(void * arg) 2333 { 2334 iflib_txq_t txq = arg; 2335 2336 /* simple non-zero boolean so use bitwise OR */ 2337 if ((txq->ift_db_pending | txq->ift_npending) && 2338 txq->ift_db_pending >= txq->ift_db_pending_queued) 2339 iflib_txd_db_check(txq->ift_ctx, txq, TRUE); 2340 txq->ift_db_pending_queued = 0; 2341 if (ifmp_ring_is_stalled(txq->ift_br[0])) 2342 iflib_txq_check_drain(txq, 4); 2343 } 2344 2345 #ifdef PKT_DEBUG 2346 static void 2347 print_pkt(if_pkt_info_t pi) 2348 { 2349 printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n", 2350 pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx); 2351 printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n", 2352 pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag); 2353 printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n", 2354 pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto); 2355 } 2356 #endif 2357 2358 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO) 2359 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO) 2360 2361 static int 2362 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) 2363 { 2364 if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx; 2365 struct ether_vlan_header *eh; 2366 struct mbuf *m, *n; 2367 2368 n = m = *mp; 2369 if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) && 2370 M_WRITABLE(m) == 0) { 2371 if ((m = m_dup(m, M_NOWAIT)) == NULL) { 2372 return (ENOMEM); 2373 } else { 2374 m_freem(*mp); 2375 n = *mp = m; 2376 } 2377 } 2378 2379 /* 2380 * Determine where frame payload starts. 2381 * Jump over vlan headers if already present, 2382 * helpful for QinQ too. 2383 */ 2384 if (__predict_false(m->m_len < sizeof(*eh))) { 2385 txq->ift_pullups++; 2386 if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL)) 2387 return (ENOMEM); 2388 } 2389 eh = mtod(m, struct ether_vlan_header *); 2390 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { 2391 pi->ipi_etype = ntohs(eh->evl_proto); 2392 pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 2393 } else { 2394 pi->ipi_etype = ntohs(eh->evl_encap_proto); 2395 pi->ipi_ehdrlen = ETHER_HDR_LEN; 2396 } 2397 2398 switch (pi->ipi_etype) { 2399 #ifdef INET 2400 case ETHERTYPE_IP: 2401 { 2402 struct ip *ip = NULL; 2403 struct tcphdr *th = NULL; 2404 int minthlen; 2405 2406 minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th)); 2407 if (__predict_false(m->m_len < minthlen)) { 2408 /* 2409 * if this code bloat is causing too much of a hit 2410 * move it to a separate function and mark it noinline 2411 */ 2412 if (m->m_len == pi->ipi_ehdrlen) { 2413 n = m->m_next; 2414 MPASS(n); 2415 if (n->m_len >= sizeof(*ip)) { 2416 ip = (struct ip *)n->m_data; 2417 if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th)) 2418 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); 2419 } else { 2420 txq->ift_pullups++; 2421 if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) 2422 return (ENOMEM); 2423 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 2424 } 2425 } else { 2426 txq->ift_pullups++; 2427 if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) 2428 return (ENOMEM); 2429 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 2430 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) 2431 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); 2432 } 2433 } else { 2434 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 2435 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) 2436 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); 2437 } 2438 pi->ipi_ip_hlen = ip->ip_hl << 2; 2439 pi->ipi_ipproto = ip->ip_p; 2440 pi->ipi_flags |= IPI_TX_IPV4; 2441 2442 if (pi->ipi_csum_flags & CSUM_IP) 2443 ip->ip_sum = 0; 2444 2445 if (pi->ipi_ipproto == IPPROTO_TCP) { 2446 if (__predict_false(th == NULL)) { 2447 txq->ift_pullups++; 2448 if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL)) 2449 return (ENOMEM); 2450 th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen); 2451 } 2452 pi->ipi_tcp_hflags = th->th_flags; 2453 pi->ipi_tcp_hlen = th->th_off << 2; 2454 pi->ipi_tcp_seq = th->th_seq; 2455 } 2456 if (IS_TSO4(pi)) { 2457 if (__predict_false(ip->ip_p != IPPROTO_TCP)) 2458 return (ENXIO); 2459 th->th_sum = in_pseudo(ip->ip_src.s_addr, 2460 ip->ip_dst.s_addr, htons(IPPROTO_TCP)); 2461 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; 2462 if (sctx->isc_flags & IFLIB_TSO_INIT_IP) { 2463 ip->ip_sum = 0; 2464 ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz); 2465 } 2466 } 2467 break; 2468 } 2469 #endif 2470 #ifdef INET6 2471 case ETHERTYPE_IPV6: 2472 { 2473 struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); 2474 struct tcphdr *th; 2475 pi->ipi_ip_hlen = sizeof(struct ip6_hdr); 2476 2477 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { 2478 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) 2479 return (ENOMEM); 2480 } 2481 th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen); 2482 2483 /* XXX-BZ this will go badly in case of ext hdrs. */ 2484 pi->ipi_ipproto = ip6->ip6_nxt; 2485 pi->ipi_flags |= IPI_TX_IPV6; 2486 2487 if (pi->ipi_ipproto == IPPROTO_TCP) { 2488 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) { 2489 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL)) 2490 return (ENOMEM); 2491 } 2492 pi->ipi_tcp_hflags = th->th_flags; 2493 pi->ipi_tcp_hlen = th->th_off << 2; 2494 } 2495 if (IS_TSO6(pi)) { 2496 2497 if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP)) 2498 return (ENXIO); 2499 /* 2500 * The corresponding flag is set by the stack in the IPv4 2501 * TSO case, but not in IPv6 (at least in FreeBSD 10.2). 2502 * So, set it here because the rest of the flow requires it. 2503 */ 2504 pi->ipi_csum_flags |= CSUM_TCP_IPV6; 2505 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); 2506 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; 2507 } 2508 break; 2509 } 2510 #endif 2511 default: 2512 pi->ipi_csum_flags &= ~CSUM_OFFLOAD; 2513 pi->ipi_ip_hlen = 0; 2514 break; 2515 } 2516 *mp = m; 2517 2518 return (0); 2519 } 2520 2521 static __noinline struct mbuf * 2522 collapse_pkthdr(struct mbuf *m0) 2523 { 2524 struct mbuf *m, *m_next, *tmp; 2525 2526 m = m0; 2527 m_next = m->m_next; 2528 while (m_next != NULL && m_next->m_len == 0) { 2529 m = m_next; 2530 m->m_next = NULL; 2531 m_free(m); 2532 m_next = m_next->m_next; 2533 } 2534 m = m0; 2535 m->m_next = m_next; 2536 if ((m_next->m_flags & M_EXT) == 0) { 2537 m = m_defrag(m, M_NOWAIT); 2538 } else { 2539 tmp = m_next->m_next; 2540 memcpy(m_next, m, MPKTHSIZE); 2541 m = m_next; 2542 m->m_next = tmp; 2543 } 2544 return (m); 2545 } 2546 2547 /* 2548 * If dodgy hardware rejects the scatter gather chain we've handed it 2549 * we'll need to remove the mbuf chain from ifsg_m[] before we can add the 2550 * m_defrag'd mbufs 2551 */ 2552 static __noinline struct mbuf * 2553 iflib_remove_mbuf(iflib_txq_t txq) 2554 { 2555 int ntxd, i, pidx; 2556 struct mbuf *m, *mh, **ifsd_m; 2557 2558 pidx = txq->ift_pidx; 2559 ifsd_m = txq->ift_sds.ifsd_m; 2560 ntxd = txq->ift_size; 2561 mh = m = ifsd_m[pidx]; 2562 ifsd_m[pidx] = NULL; 2563 #if MEMORY_LOGGING 2564 txq->ift_dequeued++; 2565 #endif 2566 i = 1; 2567 2568 while (m) { 2569 ifsd_m[(pidx + i) & (ntxd -1)] = NULL; 2570 #if MEMORY_LOGGING 2571 txq->ift_dequeued++; 2572 #endif 2573 m = m->m_next; 2574 i++; 2575 } 2576 return (mh); 2577 } 2578 2579 static int 2580 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map, 2581 struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs, 2582 int max_segs, int flags) 2583 { 2584 if_ctx_t ctx; 2585 if_shared_ctx_t sctx; 2586 if_softc_ctx_t scctx; 2587 int i, next, pidx, mask, err, maxsegsz, ntxd, count; 2588 struct mbuf *m, *tmp, **ifsd_m, **mp; 2589 2590 m = *m0; 2591 2592 /* 2593 * Please don't ever do this 2594 */ 2595 if (__predict_false(m->m_len == 0)) 2596 *m0 = m = collapse_pkthdr(m); 2597 2598 ctx = txq->ift_ctx; 2599 sctx = ctx->ifc_sctx; 2600 scctx = &ctx->ifc_softc_ctx; 2601 ifsd_m = txq->ift_sds.ifsd_m; 2602 ntxd = txq->ift_size; 2603 pidx = txq->ift_pidx; 2604 if (map != NULL) { 2605 uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags; 2606 2607 err = bus_dmamap_load_mbuf_sg(tag, map, 2608 *m0, segs, nsegs, BUS_DMA_NOWAIT); 2609 if (err) 2610 return (err); 2611 ifsd_flags[pidx] |= TX_SW_DESC_MAPPED; 2612 i = 0; 2613 next = pidx; 2614 mask = (txq->ift_size-1); 2615 m = *m0; 2616 do { 2617 mp = &ifsd_m[next]; 2618 *mp = m; 2619 m = m->m_next; 2620 if (__predict_false((*mp)->m_len == 0)) { 2621 m_free(*mp); 2622 *mp = NULL; 2623 } else 2624 next = (pidx + i) & (ntxd-1); 2625 } while (m != NULL); 2626 } else { 2627 int buflen, sgsize, max_sgsize; 2628 vm_offset_t vaddr; 2629 vm_paddr_t curaddr; 2630 2631 count = i = 0; 2632 maxsegsz = sctx->isc_tx_maxsize; 2633 m = *m0; 2634 do { 2635 if (__predict_false(m->m_len <= 0)) { 2636 tmp = m; 2637 m = m->m_next; 2638 tmp->m_next = NULL; 2639 m_free(tmp); 2640 continue; 2641 } 2642 buflen = m->m_len; 2643 vaddr = (vm_offset_t)m->m_data; 2644 /* 2645 * see if we can't be smarter about physically 2646 * contiguous mappings 2647 */ 2648 next = (pidx + count) & (ntxd-1); 2649 MPASS(ifsd_m[next] == NULL); 2650 #if MEMORY_LOGGING 2651 txq->ift_enqueued++; 2652 #endif 2653 ifsd_m[next] = m; 2654 while (buflen > 0) { 2655 max_sgsize = MIN(buflen, maxsegsz); 2656 curaddr = pmap_kextract(vaddr); 2657 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK); 2658 sgsize = MIN(sgsize, max_sgsize); 2659 segs[i].ds_addr = curaddr; 2660 segs[i].ds_len = sgsize; 2661 vaddr += sgsize; 2662 buflen -= sgsize; 2663 i++; 2664 if (i >= max_segs) 2665 goto err; 2666 } 2667 count++; 2668 tmp = m; 2669 m = m->m_next; 2670 } while (m != NULL); 2671 *nsegs = i; 2672 } 2673 return (0); 2674 err: 2675 *m0 = iflib_remove_mbuf(txq); 2676 return (EFBIG); 2677 } 2678 2679 static int 2680 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp) 2681 { 2682 if_ctx_t ctx; 2683 if_shared_ctx_t sctx; 2684 if_softc_ctx_t scctx; 2685 bus_dma_segment_t *segs; 2686 struct mbuf *m_head; 2687 bus_dmamap_t map; 2688 struct if_pkt_info pi; 2689 int remap = 0; 2690 int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd; 2691 bus_dma_tag_t desc_tag; 2692 2693 segs = txq->ift_segs; 2694 ctx = txq->ift_ctx; 2695 sctx = ctx->ifc_sctx; 2696 scctx = &ctx->ifc_softc_ctx; 2697 segs = txq->ift_segs; 2698 ntxd = txq->ift_size; 2699 m_head = *m_headp; 2700 map = NULL; 2701 2702 /* 2703 * If we're doing TSO the next descriptor to clean may be quite far ahead 2704 */ 2705 cidx = txq->ift_cidx; 2706 pidx = txq->ift_pidx; 2707 next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1); 2708 2709 /* prefetch the next cache line of mbuf pointers and flags */ 2710 prefetch(&txq->ift_sds.ifsd_m[next]); 2711 if (txq->ift_sds.ifsd_map != NULL) { 2712 prefetch(&txq->ift_sds.ifsd_map[next]); 2713 map = txq->ift_sds.ifsd_map[pidx]; 2714 next = (cidx + CACHE_LINE_SIZE) & (ntxd-1); 2715 prefetch(&txq->ift_sds.ifsd_flags[next]); 2716 } 2717 2718 2719 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) { 2720 desc_tag = txq->ift_tso_desc_tag; 2721 max_segs = scctx->isc_tx_tso_segments_max; 2722 } else { 2723 desc_tag = txq->ift_desc_tag; 2724 max_segs = scctx->isc_tx_nsegments; 2725 } 2726 m_head = *m_headp; 2727 bzero(&pi, sizeof(pi)); 2728 pi.ipi_len = m_head->m_pkthdr.len; 2729 pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST)); 2730 pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags; 2731 pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0; 2732 pi.ipi_pidx = pidx; 2733 pi.ipi_qsidx = txq->ift_id; 2734 2735 /* deliberate bitwise OR to make one condition */ 2736 if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) { 2737 if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) 2738 return (err); 2739 m_head = *m_headp; 2740 } 2741 2742 retry: 2743 err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT); 2744 defrag: 2745 if (__predict_false(err)) { 2746 switch (err) { 2747 case EFBIG: 2748 /* try collapse once and defrag once */ 2749 if (remap == 0) 2750 m_head = m_collapse(*m_headp, M_NOWAIT, max_segs); 2751 if (remap == 1) 2752 m_head = m_defrag(*m_headp, M_NOWAIT); 2753 remap++; 2754 if (__predict_false(m_head == NULL)) 2755 goto defrag_failed; 2756 txq->ift_mbuf_defrag++; 2757 *m_headp = m_head; 2758 goto retry; 2759 break; 2760 case ENOMEM: 2761 txq->ift_no_tx_dma_setup++; 2762 break; 2763 default: 2764 txq->ift_no_tx_dma_setup++; 2765 m_freem(*m_headp); 2766 DBG_COUNTER_INC(tx_frees); 2767 *m_headp = NULL; 2768 break; 2769 } 2770 txq->ift_map_failed++; 2771 DBG_COUNTER_INC(encap_load_mbuf_fail); 2772 return (err); 2773 } 2774 2775 /* 2776 * XXX assumes a 1 to 1 relationship between segments and 2777 * descriptors - this does not hold true on all drivers, e.g. 2778 * cxgb 2779 */ 2780 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { 2781 txq->ift_no_desc_avail++; 2782 if (map != NULL) 2783 bus_dmamap_unload(desc_tag, map); 2784 DBG_COUNTER_INC(encap_txq_avail_fail); 2785 if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0) 2786 GROUPTASK_ENQUEUE(&txq->ift_task); 2787 return (ENOBUFS); 2788 } 2789 pi.ipi_segs = segs; 2790 pi.ipi_nsegs = nsegs; 2791 2792 MPASS(pidx >= 0 && pidx < txq->ift_size); 2793 #ifdef PKT_DEBUG 2794 print_pkt(&pi); 2795 #endif 2796 if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) { 2797 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 2798 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 2799 2800 DBG_COUNTER_INC(tx_encap); 2801 MPASS(pi.ipi_new_pidx >= 0 && 2802 pi.ipi_new_pidx < txq->ift_size); 2803 2804 ndesc = pi.ipi_new_pidx - pi.ipi_pidx; 2805 if (pi.ipi_new_pidx < pi.ipi_pidx) { 2806 ndesc += txq->ift_size; 2807 txq->ift_gen = 1; 2808 } 2809 /* 2810 * drivers can need as many as 2811 * two sentinels 2812 */ 2813 MPASS(ndesc <= pi.ipi_nsegs + 2); 2814 MPASS(pi.ipi_new_pidx != pidx); 2815 MPASS(ndesc > 0); 2816 txq->ift_in_use += ndesc; 2817 /* 2818 * We update the last software descriptor again here because there may 2819 * be a sentinel and/or there may be more mbufs than segments 2820 */ 2821 txq->ift_pidx = pi.ipi_new_pidx; 2822 txq->ift_npending += pi.ipi_ndescs; 2823 } else if (__predict_false(err == EFBIG && remap < 2)) { 2824 *m_headp = m_head = iflib_remove_mbuf(txq); 2825 remap = 1; 2826 txq->ift_txd_encap_efbig++; 2827 goto defrag; 2828 } else 2829 DBG_COUNTER_INC(encap_txd_encap_fail); 2830 return (err); 2831 2832 defrag_failed: 2833 txq->ift_mbuf_defrag_failed++; 2834 txq->ift_map_failed++; 2835 m_freem(*m_headp); 2836 DBG_COUNTER_INC(tx_frees); 2837 *m_headp = NULL; 2838 return (ENOMEM); 2839 } 2840 2841 /* forward compatibility for cxgb */ 2842 #define FIRST_QSET(ctx) 0 2843 2844 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) 2845 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) 2846 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx)) 2847 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) 2848 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) 2849 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max) 2850 2851 2852 2853 /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring 2854 * doorbell writes 2855 * 2856 * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic 2857 */ 2858 #define TXQ_MIN_OCCUPANCY(size) ((size >> 6)| 0x2) 2859 2860 static inline int 2861 iflib_txq_min_occupancy(iflib_txq_t txq) 2862 { 2863 if_ctx_t ctx; 2864 2865 ctx = txq->ift_ctx; 2866 return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, 2867 txq->ift_gen) < TXQ_MIN_OCCUPANCY(txq->ift_size) + 2868 MAX_TX_DESC(ctx)); 2869 } 2870 2871 static void 2872 iflib_tx_desc_free(iflib_txq_t txq, int n) 2873 { 2874 int hasmap; 2875 uint32_t qsize, cidx, mask, gen; 2876 struct mbuf *m, **ifsd_m; 2877 uint8_t *ifsd_flags; 2878 bus_dmamap_t *ifsd_map; 2879 2880 cidx = txq->ift_cidx; 2881 gen = txq->ift_gen; 2882 qsize = txq->ift_size; 2883 mask = qsize-1; 2884 hasmap = txq->ift_sds.ifsd_map != NULL; 2885 ifsd_flags = txq->ift_sds.ifsd_flags; 2886 ifsd_m = txq->ift_sds.ifsd_m; 2887 ifsd_map = txq->ift_sds.ifsd_map; 2888 2889 while (n--) { 2890 prefetch(ifsd_m[(cidx + 3) & mask]); 2891 prefetch(ifsd_m[(cidx + 4) & mask]); 2892 2893 if (ifsd_m[cidx] != NULL) { 2894 prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]); 2895 prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]); 2896 if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) { 2897 /* 2898 * does it matter if it's not the TSO tag? If so we'll 2899 * have to add the type to flags 2900 */ 2901 bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]); 2902 ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED; 2903 } 2904 if ((m = ifsd_m[cidx]) != NULL) { 2905 /* XXX we don't support any drivers that batch packets yet */ 2906 MPASS(m->m_nextpkt == NULL); 2907 2908 m_free(m); 2909 ifsd_m[cidx] = NULL; 2910 #if MEMORY_LOGGING 2911 txq->ift_dequeued++; 2912 #endif 2913 DBG_COUNTER_INC(tx_frees); 2914 } 2915 } 2916 if (__predict_false(++cidx == qsize)) { 2917 cidx = 0; 2918 gen = 0; 2919 } 2920 } 2921 txq->ift_cidx = cidx; 2922 txq->ift_gen = gen; 2923 } 2924 2925 static __inline int 2926 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh) 2927 { 2928 int reclaim; 2929 if_ctx_t ctx = txq->ift_ctx; 2930 2931 KASSERT(thresh >= 0, ("invalid threshold to reclaim")); 2932 MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size); 2933 2934 /* 2935 * Need a rate-limiting check so that this isn't called every time 2936 */ 2937 iflib_tx_credits_update(ctx, txq); 2938 reclaim = DESC_RECLAIMABLE(txq); 2939 2940 if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) { 2941 #ifdef INVARIANTS 2942 if (iflib_verbose_debug) { 2943 printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__, 2944 txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments, 2945 reclaim, thresh); 2946 2947 } 2948 #endif 2949 return (0); 2950 } 2951 iflib_tx_desc_free(txq, reclaim); 2952 txq->ift_cleaned += reclaim; 2953 txq->ift_in_use -= reclaim; 2954 2955 if (txq->ift_active == FALSE) 2956 txq->ift_active = TRUE; 2957 2958 return (reclaim); 2959 } 2960 2961 static struct mbuf ** 2962 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset) 2963 { 2964 2965 return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)])); 2966 } 2967 2968 static void 2969 iflib_txq_check_drain(iflib_txq_t txq, int budget) 2970 { 2971 2972 ifmp_ring_check_drainage(txq->ift_br[0], budget); 2973 } 2974 2975 static uint32_t 2976 iflib_txq_can_drain(struct ifmp_ring *r) 2977 { 2978 iflib_txq_t txq = r->cookie; 2979 if_ctx_t ctx = txq->ift_ctx; 2980 2981 return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) || 2982 ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false)); 2983 } 2984 2985 static uint32_t 2986 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) 2987 { 2988 iflib_txq_t txq = r->cookie; 2989 if_ctx_t ctx = txq->ift_ctx; 2990 if_t ifp = ctx->ifc_ifp; 2991 struct mbuf **mp, *m; 2992 int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used; 2993 2994 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) || 2995 !LINK_ACTIVE(ctx))) { 2996 DBG_COUNTER_INC(txq_drain_notready); 2997 return (0); 2998 } 2999 3000 avail = IDXDIFF(pidx, cidx, r->size); 3001 if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { 3002 DBG_COUNTER_INC(txq_drain_flushing); 3003 for (i = 0; i < avail; i++) { 3004 m_free(r->items[(cidx + i) & (r->size-1)]); 3005 r->items[(cidx + i) & (r->size-1)] = NULL; 3006 } 3007 return (avail); 3008 } 3009 iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 3010 if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) { 3011 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 3012 CALLOUT_LOCK(txq); 3013 callout_stop(&txq->ift_timer); 3014 callout_stop(&txq->ift_db_check); 3015 CALLOUT_UNLOCK(txq); 3016 DBG_COUNTER_INC(txq_drain_oactive); 3017 return (0); 3018 } 3019 consumed = mcast_sent = bytes_sent = pkt_sent = 0; 3020 count = MIN(avail, TX_BATCH_SIZE); 3021 #ifdef INVARIANTS 3022 if (iflib_verbose_debug) 3023 printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__, 3024 avail, ctx->ifc_flags, TXQ_AVAIL(txq)); 3025 #endif 3026 3027 for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) { 3028 mp = _ring_peek_one(r, cidx, i); 3029 MPASS(mp != NULL && *mp != NULL); 3030 in_use_prev = txq->ift_in_use; 3031 if ((err = iflib_encap(txq, mp)) == ENOBUFS) { 3032 DBG_COUNTER_INC(txq_drain_encapfail); 3033 /* no room - bail out */ 3034 break; 3035 } 3036 consumed++; 3037 if (err) { 3038 DBG_COUNTER_INC(txq_drain_encapfail); 3039 /* we can't send this packet - skip it */ 3040 continue; 3041 } 3042 pkt_sent++; 3043 m = *mp; 3044 DBG_COUNTER_INC(tx_sent); 3045 bytes_sent += m->m_pkthdr.len; 3046 if (m->m_flags & M_MCAST) 3047 mcast_sent++; 3048 3049 txq->ift_db_pending += (txq->ift_in_use - in_use_prev); 3050 desc_used += (txq->ift_in_use - in_use_prev); 3051 iflib_txd_db_check(ctx, txq, FALSE); 3052 ETHER_BPF_MTAP(ifp, m); 3053 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) 3054 break; 3055 3056 if (desc_used >= TXQ_MAX_DB_CONSUMED(txq->ift_size)) 3057 break; 3058 } 3059 3060 if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending) 3061 iflib_txd_db_check(ctx, txq, TRUE); 3062 else if ((txq->ift_db_pending || TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2) && 3063 (callout_pending(&txq->ift_db_check) == 0)) { 3064 txq->ift_db_pending_queued = txq->ift_db_pending; 3065 callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check, 3066 txq, txq->ift_db_check.c_cpu); 3067 } 3068 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); 3069 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); 3070 if (mcast_sent) 3071 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); 3072 #ifdef INVARIANTS 3073 if (iflib_verbose_debug) 3074 printf("consumed=%d\n", consumed); 3075 #endif 3076 return (consumed); 3077 } 3078 3079 static uint32_t 3080 iflib_txq_drain_always(struct ifmp_ring *r) 3081 { 3082 return (1); 3083 } 3084 3085 static uint32_t 3086 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) 3087 { 3088 int i, avail; 3089 struct mbuf **mp; 3090 iflib_txq_t txq; 3091 3092 txq = r->cookie; 3093 3094 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 3095 CALLOUT_LOCK(txq); 3096 callout_stop(&txq->ift_timer); 3097 callout_stop(&txq->ift_db_check); 3098 CALLOUT_UNLOCK(txq); 3099 3100 avail = IDXDIFF(pidx, cidx, r->size); 3101 for (i = 0; i < avail; i++) { 3102 mp = _ring_peek_one(r, cidx, i); 3103 m_freem(*mp); 3104 } 3105 MPASS(ifmp_ring_is_stalled(r) == 0); 3106 return (avail); 3107 } 3108 3109 static void 3110 iflib_ifmp_purge(iflib_txq_t txq) 3111 { 3112 struct ifmp_ring *r; 3113 3114 r = txq->ift_br[0]; 3115 r->drain = iflib_txq_drain_free; 3116 r->can_drain = iflib_txq_drain_always; 3117 3118 ifmp_ring_check_drainage(r, r->size); 3119 3120 r->drain = iflib_txq_drain; 3121 r->can_drain = iflib_txq_can_drain; 3122 } 3123 3124 static void 3125 _task_fn_tx(void *context) 3126 { 3127 iflib_txq_t txq = context; 3128 if_ctx_t ctx = txq->ift_ctx; 3129 3130 #ifdef IFLIB_DIAGNOSTICS 3131 txq->ift_cpu_exec_count[curcpu]++; 3132 #endif 3133 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) 3134 return; 3135 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE); 3136 } 3137 3138 static void 3139 _task_fn_rx(void *context) 3140 { 3141 iflib_rxq_t rxq = context; 3142 if_ctx_t ctx = rxq->ifr_ctx; 3143 bool more; 3144 int rc; 3145 3146 #ifdef IFLIB_DIAGNOSTICS 3147 rxq->ifr_cpu_exec_count[curcpu]++; 3148 #endif 3149 DBG_COUNTER_INC(task_fn_rxs); 3150 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) 3151 return; 3152 3153 if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) { 3154 if (ctx->ifc_flags & IFC_LEGACY) 3155 IFDI_INTR_ENABLE(ctx); 3156 else { 3157 DBG_COUNTER_INC(rx_intr_enables); 3158 rc = IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); 3159 KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver")); 3160 } 3161 } 3162 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) 3163 return; 3164 if (more) 3165 GROUPTASK_ENQUEUE(&rxq->ifr_task); 3166 } 3167 3168 static void 3169 _task_fn_admin(void *context) 3170 { 3171 if_ctx_t ctx = context; 3172 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 3173 iflib_txq_t txq; 3174 int i; 3175 3176 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) 3177 return; 3178 3179 CTX_LOCK(ctx); 3180 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { 3181 CALLOUT_LOCK(txq); 3182 callout_stop(&txq->ift_timer); 3183 CALLOUT_UNLOCK(txq); 3184 } 3185 IFDI_UPDATE_ADMIN_STATUS(ctx); 3186 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) 3187 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); 3188 IFDI_LINK_INTR_ENABLE(ctx); 3189 CTX_UNLOCK(ctx); 3190 3191 if (LINK_ACTIVE(ctx) == 0) 3192 return; 3193 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) 3194 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); 3195 } 3196 3197 3198 static void 3199 _task_fn_iov(void *context) 3200 { 3201 if_ctx_t ctx = context; 3202 3203 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) 3204 return; 3205 3206 CTX_LOCK(ctx); 3207 IFDI_VFLR_HANDLE(ctx); 3208 CTX_UNLOCK(ctx); 3209 } 3210 3211 static int 3212 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS) 3213 { 3214 int err; 3215 if_int_delay_info_t info; 3216 if_ctx_t ctx; 3217 3218 info = (if_int_delay_info_t)arg1; 3219 ctx = info->iidi_ctx; 3220 info->iidi_req = req; 3221 info->iidi_oidp = oidp; 3222 CTX_LOCK(ctx); 3223 err = IFDI_SYSCTL_INT_DELAY(ctx, info); 3224 CTX_UNLOCK(ctx); 3225 return (err); 3226 } 3227 3228 /********************************************************************* 3229 * 3230 * IFNET FUNCTIONS 3231 * 3232 **********************************************************************/ 3233 3234 static void 3235 iflib_if_init_locked(if_ctx_t ctx) 3236 { 3237 iflib_stop(ctx); 3238 iflib_init_locked(ctx); 3239 } 3240 3241 3242 static void 3243 iflib_if_init(void *arg) 3244 { 3245 if_ctx_t ctx = arg; 3246 3247 CTX_LOCK(ctx); 3248 iflib_if_init_locked(ctx); 3249 CTX_UNLOCK(ctx); 3250 } 3251 3252 static int 3253 iflib_if_transmit(if_t ifp, struct mbuf *m) 3254 { 3255 if_ctx_t ctx = if_getsoftc(ifp); 3256 3257 iflib_txq_t txq; 3258 int err, qidx; 3259 3260 if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { 3261 DBG_COUNTER_INC(tx_frees); 3262 m_freem(m); 3263 return (ENOBUFS); 3264 } 3265 3266 MPASS(m->m_nextpkt == NULL); 3267 qidx = 0; 3268 if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m)) 3269 qidx = QIDX(ctx, m); 3270 /* 3271 * XXX calculate buf_ring based on flowid (divvy up bits?) 3272 */ 3273 txq = &ctx->ifc_txqs[qidx]; 3274 3275 #ifdef DRIVER_BACKPRESSURE 3276 if (txq->ift_closed) { 3277 while (m != NULL) { 3278 next = m->m_nextpkt; 3279 m->m_nextpkt = NULL; 3280 m_freem(m); 3281 m = next; 3282 } 3283 return (ENOBUFS); 3284 } 3285 #endif 3286 #ifdef notyet 3287 qidx = count = 0; 3288 mp = marr; 3289 next = m; 3290 do { 3291 count++; 3292 next = next->m_nextpkt; 3293 } while (next != NULL); 3294 3295 if (count > nitems(marr)) 3296 if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) { 3297 /* XXX check nextpkt */ 3298 m_freem(m); 3299 /* XXX simplify for now */ 3300 DBG_COUNTER_INC(tx_frees); 3301 return (ENOBUFS); 3302 } 3303 for (next = m, i = 0; next != NULL; i++) { 3304 mp[i] = next; 3305 next = next->m_nextpkt; 3306 mp[i]->m_nextpkt = NULL; 3307 } 3308 #endif 3309 DBG_COUNTER_INC(tx_seen); 3310 err = ifmp_ring_enqueue(txq->ift_br[0], (void **)&m, 1, TX_BATCH_SIZE); 3311 3312 if (err) { 3313 GROUPTASK_ENQUEUE(&txq->ift_task); 3314 /* support forthcoming later */ 3315 #ifdef DRIVER_BACKPRESSURE 3316 txq->ift_closed = TRUE; 3317 #endif 3318 ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE); 3319 m_freem(m); 3320 } else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) { 3321 GROUPTASK_ENQUEUE(&txq->ift_task); 3322 } 3323 3324 return (err); 3325 } 3326 3327 static void 3328 iflib_if_qflush(if_t ifp) 3329 { 3330 if_ctx_t ctx = if_getsoftc(ifp); 3331 iflib_txq_t txq = ctx->ifc_txqs; 3332 int i; 3333 3334 CTX_LOCK(ctx); 3335 ctx->ifc_flags |= IFC_QFLUSH; 3336 CTX_UNLOCK(ctx); 3337 for (i = 0; i < NTXQSETS(ctx); i++, txq++) 3338 while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0]))) 3339 iflib_txq_check_drain(txq, 0); 3340 CTX_LOCK(ctx); 3341 ctx->ifc_flags &= ~IFC_QFLUSH; 3342 CTX_UNLOCK(ctx); 3343 3344 if_qflush(ifp); 3345 } 3346 3347 3348 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ 3349 IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | \ 3350 IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO) 3351 3352 static int 3353 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) 3354 { 3355 if_ctx_t ctx = if_getsoftc(ifp); 3356 struct ifreq *ifr = (struct ifreq *)data; 3357 #if defined(INET) || defined(INET6) 3358 struct ifaddr *ifa = (struct ifaddr *)data; 3359 #endif 3360 bool avoid_reset = FALSE; 3361 int err = 0, reinit = 0, bits; 3362 3363 switch (command) { 3364 case SIOCSIFADDR: 3365 #ifdef INET 3366 if (ifa->ifa_addr->sa_family == AF_INET) 3367 avoid_reset = TRUE; 3368 #endif 3369 #ifdef INET6 3370 if (ifa->ifa_addr->sa_family == AF_INET6) 3371 avoid_reset = TRUE; 3372 #endif 3373 /* 3374 ** Calling init results in link renegotiation, 3375 ** so we avoid doing it when possible. 3376 */ 3377 if (avoid_reset) { 3378 if_setflagbits(ifp, IFF_UP,0); 3379 if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING)) 3380 reinit = 1; 3381 #ifdef INET 3382 if (!(if_getflags(ifp) & IFF_NOARP)) 3383 arp_ifinit(ifp, ifa); 3384 #endif 3385 } else 3386 err = ether_ioctl(ifp, command, data); 3387 break; 3388 case SIOCSIFMTU: 3389 CTX_LOCK(ctx); 3390 if (ifr->ifr_mtu == if_getmtu(ifp)) { 3391 CTX_UNLOCK(ctx); 3392 break; 3393 } 3394 bits = if_getdrvflags(ifp); 3395 /* stop the driver and free any clusters before proceeding */ 3396 iflib_stop(ctx); 3397 3398 if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) { 3399 if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size) 3400 ctx->ifc_flags |= IFC_MULTISEG; 3401 else 3402 ctx->ifc_flags &= ~IFC_MULTISEG; 3403 err = if_setmtu(ifp, ifr->ifr_mtu); 3404 } 3405 iflib_init_locked(ctx); 3406 if_setdrvflags(ifp, bits); 3407 CTX_UNLOCK(ctx); 3408 break; 3409 case SIOCSIFFLAGS: 3410 CTX_LOCK(ctx); 3411 if (if_getflags(ifp) & IFF_UP) { 3412 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 3413 if ((if_getflags(ifp) ^ ctx->ifc_if_flags) & 3414 (IFF_PROMISC | IFF_ALLMULTI)) { 3415 err = IFDI_PROMISC_SET(ctx, if_getflags(ifp)); 3416 } 3417 } else 3418 reinit = 1; 3419 } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 3420 iflib_stop(ctx); 3421 } 3422 ctx->ifc_if_flags = if_getflags(ifp); 3423 CTX_UNLOCK(ctx); 3424 break; 3425 3426 break; 3427 case SIOCADDMULTI: 3428 case SIOCDELMULTI: 3429 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 3430 CTX_LOCK(ctx); 3431 IFDI_INTR_DISABLE(ctx); 3432 IFDI_MULTI_SET(ctx); 3433 IFDI_INTR_ENABLE(ctx); 3434 CTX_UNLOCK(ctx); 3435 } 3436 break; 3437 case SIOCSIFMEDIA: 3438 CTX_LOCK(ctx); 3439 IFDI_MEDIA_SET(ctx); 3440 CTX_UNLOCK(ctx); 3441 /* falls thru */ 3442 case SIOCGIFMEDIA: 3443 err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command); 3444 break; 3445 case SIOCGI2C: 3446 { 3447 struct ifi2creq i2c; 3448 3449 err = copyin(ifr->ifr_data, &i2c, sizeof(i2c)); 3450 if (err != 0) 3451 break; 3452 if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { 3453 err = EINVAL; 3454 break; 3455 } 3456 if (i2c.len > sizeof(i2c.data)) { 3457 err = EINVAL; 3458 break; 3459 } 3460 3461 if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0) 3462 err = copyout(&i2c, ifr->ifr_data, sizeof(i2c)); 3463 break; 3464 } 3465 case SIOCSIFCAP: 3466 { 3467 int mask, setmask; 3468 3469 mask = ifr->ifr_reqcap ^ if_getcapenable(ifp); 3470 setmask = 0; 3471 #ifdef TCP_OFFLOAD 3472 setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6); 3473 #endif 3474 setmask |= (mask & IFCAP_FLAGS); 3475 3476 if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) 3477 setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6); 3478 if ((mask & IFCAP_WOL) && 3479 (if_getcapabilities(ifp) & IFCAP_WOL) != 0) 3480 setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC)); 3481 if_vlancap(ifp); 3482 /* 3483 * want to ensure that traffic has stopped before we change any of the flags 3484 */ 3485 if (setmask) { 3486 CTX_LOCK(ctx); 3487 bits = if_getdrvflags(ifp); 3488 if (bits & IFF_DRV_RUNNING) 3489 iflib_stop(ctx); 3490 if_togglecapenable(ifp, setmask); 3491 if (bits & IFF_DRV_RUNNING) 3492 iflib_init_locked(ctx); 3493 if_setdrvflags(ifp, bits); 3494 CTX_UNLOCK(ctx); 3495 } 3496 break; 3497 } 3498 case SIOCGPRIVATE_0: 3499 case SIOCSDRVSPEC: 3500 case SIOCGDRVSPEC: 3501 CTX_LOCK(ctx); 3502 err = IFDI_PRIV_IOCTL(ctx, command, data); 3503 CTX_UNLOCK(ctx); 3504 break; 3505 default: 3506 err = ether_ioctl(ifp, command, data); 3507 break; 3508 } 3509 if (reinit) 3510 iflib_if_init(ctx); 3511 return (err); 3512 } 3513 3514 static uint64_t 3515 iflib_if_get_counter(if_t ifp, ift_counter cnt) 3516 { 3517 if_ctx_t ctx = if_getsoftc(ifp); 3518 3519 return (IFDI_GET_COUNTER(ctx, cnt)); 3520 } 3521 3522 /********************************************************************* 3523 * 3524 * OTHER FUNCTIONS EXPORTED TO THE STACK 3525 * 3526 **********************************************************************/ 3527 3528 static void 3529 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag) 3530 { 3531 if_ctx_t ctx = if_getsoftc(ifp); 3532 3533 if ((void *)ctx != arg) 3534 return; 3535 3536 if ((vtag == 0) || (vtag > 4095)) 3537 return; 3538 3539 CTX_LOCK(ctx); 3540 IFDI_VLAN_REGISTER(ctx, vtag); 3541 /* Re-init to load the changes */ 3542 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) 3543 iflib_init_locked(ctx); 3544 CTX_UNLOCK(ctx); 3545 } 3546 3547 static void 3548 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag) 3549 { 3550 if_ctx_t ctx = if_getsoftc(ifp); 3551 3552 if ((void *)ctx != arg) 3553 return; 3554 3555 if ((vtag == 0) || (vtag > 4095)) 3556 return; 3557 3558 CTX_LOCK(ctx); 3559 IFDI_VLAN_UNREGISTER(ctx, vtag); 3560 /* Re-init to load the changes */ 3561 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) 3562 iflib_init_locked(ctx); 3563 CTX_UNLOCK(ctx); 3564 } 3565 3566 static void 3567 iflib_led_func(void *arg, int onoff) 3568 { 3569 if_ctx_t ctx = arg; 3570 3571 CTX_LOCK(ctx); 3572 IFDI_LED_FUNC(ctx, onoff); 3573 CTX_UNLOCK(ctx); 3574 } 3575 3576 /********************************************************************* 3577 * 3578 * BUS FUNCTION DEFINITIONS 3579 * 3580 **********************************************************************/ 3581 3582 int 3583 iflib_device_probe(device_t dev) 3584 { 3585 pci_vendor_info_t *ent; 3586 3587 uint16_t pci_vendor_id, pci_device_id; 3588 uint16_t pci_subvendor_id, pci_subdevice_id; 3589 uint16_t pci_rev_id; 3590 if_shared_ctx_t sctx; 3591 3592 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) 3593 return (ENOTSUP); 3594 3595 pci_vendor_id = pci_get_vendor(dev); 3596 pci_device_id = pci_get_device(dev); 3597 pci_subvendor_id = pci_get_subvendor(dev); 3598 pci_subdevice_id = pci_get_subdevice(dev); 3599 pci_rev_id = pci_get_revid(dev); 3600 if (sctx->isc_parse_devinfo != NULL) 3601 sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id); 3602 3603 ent = sctx->isc_vendor_info; 3604 while (ent->pvi_vendor_id != 0) { 3605 if (pci_vendor_id != ent->pvi_vendor_id) { 3606 ent++; 3607 continue; 3608 } 3609 if ((pci_device_id == ent->pvi_device_id) && 3610 ((pci_subvendor_id == ent->pvi_subvendor_id) || 3611 (ent->pvi_subvendor_id == 0)) && 3612 ((pci_subdevice_id == ent->pvi_subdevice_id) || 3613 (ent->pvi_subdevice_id == 0)) && 3614 ((pci_rev_id == ent->pvi_rev_id) || 3615 (ent->pvi_rev_id == 0))) { 3616 3617 device_set_desc_copy(dev, ent->pvi_name); 3618 /* this needs to be changed to zero if the bus probing code 3619 * ever stops re-probing on best match because the sctx 3620 * may have its values over written by register calls 3621 * in subsequent probes 3622 */ 3623 return (BUS_PROBE_DEFAULT); 3624 } 3625 ent++; 3626 } 3627 return (ENXIO); 3628 } 3629 3630 int 3631 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp) 3632 { 3633 int err, rid, msix, msix_bar; 3634 if_ctx_t ctx; 3635 if_t ifp; 3636 if_softc_ctx_t scctx; 3637 int i; 3638 uint16_t main_txq; 3639 uint16_t main_rxq; 3640 3641 3642 ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO); 3643 3644 if (sc == NULL) { 3645 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); 3646 device_set_softc(dev, ctx); 3647 ctx->ifc_flags |= IFC_SC_ALLOCATED; 3648 } 3649 3650 ctx->ifc_sctx = sctx; 3651 ctx->ifc_dev = dev; 3652 ctx->ifc_softc = sc; 3653 3654 if ((err = iflib_register(ctx)) != 0) { 3655 device_printf(dev, "iflib_register failed %d\n", err); 3656 return (err); 3657 } 3658 iflib_add_device_sysctl_pre(ctx); 3659 3660 scctx = &ctx->ifc_softc_ctx; 3661 ifp = ctx->ifc_ifp; 3662 3663 /* 3664 * XXX sanity check that ntxd & nrxd are a power of 2 3665 */ 3666 if (ctx->ifc_sysctl_ntxqs != 0) 3667 scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs; 3668 if (ctx->ifc_sysctl_nrxqs != 0) 3669 scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs; 3670 3671 for (i = 0; i < sctx->isc_ntxqs; i++) { 3672 if (ctx->ifc_sysctl_ntxds[i] != 0) 3673 scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i]; 3674 else 3675 scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; 3676 } 3677 3678 for (i = 0; i < sctx->isc_nrxqs; i++) { 3679 if (ctx->ifc_sysctl_nrxds[i] != 0) 3680 scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i]; 3681 else 3682 scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; 3683 } 3684 3685 for (i = 0; i < sctx->isc_nrxqs; i++) { 3686 if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) { 3687 device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n", 3688 i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]); 3689 scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i]; 3690 } 3691 if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) { 3692 device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n", 3693 i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]); 3694 scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i]; 3695 } 3696 } 3697 3698 for (i = 0; i < sctx->isc_ntxqs; i++) { 3699 if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) { 3700 device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n", 3701 i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]); 3702 scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i]; 3703 } 3704 if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) { 3705 device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n", 3706 i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]); 3707 scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i]; 3708 } 3709 } 3710 3711 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { 3712 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); 3713 return (err); 3714 } 3715 _iflib_pre_assert(scctx); 3716 ctx->ifc_txrx = *scctx->isc_txrx; 3717 3718 #ifdef INVARIANTS 3719 MPASS(scctx->isc_capenable); 3720 if (scctx->isc_capenable & IFCAP_TXCSUM) 3721 MPASS(scctx->isc_tx_csum_flags); 3722 #endif 3723 3724 if_setcapabilities(ifp, scctx->isc_capenable); 3725 if_setcapenable(ifp, scctx->isc_capenable); 3726 3727 if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) 3728 scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; 3729 if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) 3730 scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; 3731 3732 #ifdef ACPI_DMAR 3733 if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL) 3734 ctx->ifc_flags |= IFC_DMAR; 3735 #endif 3736 3737 msix_bar = scctx->isc_msix_bar; 3738 3739 if(sctx->isc_flags & IFLIB_HAS_TXCQ) 3740 main_txq = 1; 3741 else 3742 main_txq = 0; 3743 3744 if(sctx->isc_flags & IFLIB_HAS_RXCQ) 3745 main_rxq = 1; 3746 else 3747 main_rxq = 0; 3748 3749 /* XXX change for per-queue sizes */ 3750 device_printf(dev, "using %d tx descriptors and %d rx descriptors\n", 3751 scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]); 3752 for (i = 0; i < sctx->isc_nrxqs; i++) { 3753 if (!powerof2(scctx->isc_nrxd[i])) { 3754 /* round down instead? */ 3755 device_printf(dev, "# rx descriptors must be a power of 2\n"); 3756 err = EINVAL; 3757 goto fail; 3758 } 3759 } 3760 for (i = 0; i < sctx->isc_ntxqs; i++) { 3761 if (!powerof2(scctx->isc_ntxd[i])) { 3762 device_printf(dev, 3763 "# tx descriptors must be a power of 2"); 3764 err = EINVAL; 3765 goto fail; 3766 } 3767 } 3768 3769 if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] / 3770 MAX_SINGLE_PACKET_FRACTION) 3771 scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] / 3772 MAX_SINGLE_PACKET_FRACTION); 3773 if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] / 3774 MAX_SINGLE_PACKET_FRACTION) 3775 scctx->isc_tx_tso_segments_max = max(1, 3776 scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); 3777 3778 /* 3779 * Protect the stack against modern hardware 3780 */ 3781 if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX) 3782 scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX; 3783 3784 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ 3785 ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max; 3786 ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max; 3787 ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max; 3788 if (scctx->isc_rss_table_size == 0) 3789 scctx->isc_rss_table_size = 64; 3790 scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; 3791 3792 GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); 3793 /* XXX format name */ 3794 taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin"); 3795 /* 3796 ** Now setup MSI or MSI/X, should 3797 ** return us the number of supported 3798 ** vectors. (Will be 1 for MSI) 3799 */ 3800 if (sctx->isc_flags & IFLIB_SKIP_MSIX) { 3801 msix = scctx->isc_vectors; 3802 } else if (scctx->isc_msix_bar != 0) 3803 /* 3804 * The simple fact that isc_msix_bar is not 0 does not mean we 3805 * we have a good value there that is known to work. 3806 */ 3807 msix = iflib_msix_init(ctx); 3808 else { 3809 scctx->isc_vectors = 1; 3810 scctx->isc_ntxqsets = 1; 3811 scctx->isc_nrxqsets = 1; 3812 scctx->isc_intr = IFLIB_INTR_LEGACY; 3813 msix = 0; 3814 } 3815 /* Get memory for the station queues */ 3816 if ((err = iflib_queues_alloc(ctx))) { 3817 device_printf(dev, "Unable to allocate queue memory\n"); 3818 goto fail; 3819 } 3820 3821 if ((err = iflib_qset_structures_setup(ctx))) { 3822 device_printf(dev, "qset structure setup failed %d\n", err); 3823 goto fail_queues; 3824 } 3825 3826 /* 3827 * Group taskqueues aren't properly set up until SMP is started, 3828 * so we disable interrupts until we can handle them post 3829 * SI_SUB_SMP. 3830 * 3831 * XXX: disabling interrupts doesn't actually work, at least for 3832 * the non-MSI case. When they occur before SI_SUB_SMP completes, 3833 * we do null handling and depend on this not causing too large an 3834 * interrupt storm. 3835 */ 3836 IFDI_INTR_DISABLE(ctx); 3837 if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) { 3838 device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err); 3839 goto fail_intr_free; 3840 } 3841 if (msix <= 1) { 3842 rid = 0; 3843 if (scctx->isc_intr == IFLIB_INTR_MSI) { 3844 MPASS(msix == 1); 3845 rid = 1; 3846 } 3847 if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) { 3848 device_printf(dev, "iflib_legacy_setup failed %d\n", err); 3849 goto fail_intr_free; 3850 } 3851 } 3852 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac); 3853 if ((err = IFDI_ATTACH_POST(ctx)) != 0) { 3854 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); 3855 goto fail_detach; 3856 } 3857 if ((err = iflib_netmap_attach(ctx))) { 3858 device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err); 3859 goto fail_detach; 3860 } 3861 *ctxp = ctx; 3862 3863 if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); 3864 iflib_add_device_sysctl_post(ctx); 3865 ctx->ifc_flags |= IFC_INIT_DONE; 3866 return (0); 3867 fail_detach: 3868 ether_ifdetach(ctx->ifc_ifp); 3869 fail_intr_free: 3870 if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI) 3871 pci_release_msi(ctx->ifc_dev); 3872 fail_queues: 3873 /* XXX free queues */ 3874 fail: 3875 IFDI_DETACH(ctx); 3876 return (err); 3877 } 3878 3879 int 3880 iflib_device_attach(device_t dev) 3881 { 3882 if_ctx_t ctx; 3883 if_shared_ctx_t sctx; 3884 3885 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) 3886 return (ENOTSUP); 3887 3888 pci_enable_busmaster(dev); 3889 3890 return (iflib_device_register(dev, NULL, sctx, &ctx)); 3891 } 3892 3893 int 3894 iflib_device_deregister(if_ctx_t ctx) 3895 { 3896 if_t ifp = ctx->ifc_ifp; 3897 iflib_txq_t txq; 3898 iflib_rxq_t rxq; 3899 device_t dev = ctx->ifc_dev; 3900 int i; 3901 struct taskqgroup *tqg; 3902 3903 /* Make sure VLANS are not using driver */ 3904 if (if_vlantrunkinuse(ifp)) { 3905 device_printf(dev,"Vlan in use, detach first\n"); 3906 return (EBUSY); 3907 } 3908 3909 CTX_LOCK(ctx); 3910 ctx->ifc_in_detach = 1; 3911 iflib_stop(ctx); 3912 CTX_UNLOCK(ctx); 3913 3914 /* Unregister VLAN events */ 3915 if (ctx->ifc_vlan_attach_event != NULL) 3916 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event); 3917 if (ctx->ifc_vlan_detach_event != NULL) 3918 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event); 3919 3920 iflib_netmap_detach(ifp); 3921 ether_ifdetach(ifp); 3922 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ 3923 CTX_LOCK_DESTROY(ctx); 3924 if (ctx->ifc_led_dev != NULL) 3925 led_destroy(ctx->ifc_led_dev); 3926 /* XXX drain any dependent tasks */ 3927 tqg = qgroup_if_io_tqg; 3928 for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { 3929 callout_drain(&txq->ift_timer); 3930 callout_drain(&txq->ift_db_check); 3931 if (txq->ift_task.gt_uniq != NULL) 3932 taskqgroup_detach(tqg, &txq->ift_task); 3933 } 3934 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { 3935 if (rxq->ifr_task.gt_uniq != NULL) 3936 taskqgroup_detach(tqg, &rxq->ifr_task); 3937 } 3938 tqg = qgroup_if_config_tqg; 3939 if (ctx->ifc_admin_task.gt_uniq != NULL) 3940 taskqgroup_detach(tqg, &ctx->ifc_admin_task); 3941 if (ctx->ifc_vflr_task.gt_uniq != NULL) 3942 taskqgroup_detach(tqg, &ctx->ifc_vflr_task); 3943 3944 IFDI_DETACH(ctx); 3945 device_set_softc(ctx->ifc_dev, NULL); 3946 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { 3947 pci_release_msi(dev); 3948 } 3949 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) { 3950 iflib_irq_free(ctx, &ctx->ifc_legacy_irq); 3951 } 3952 if (ctx->ifc_msix_mem != NULL) { 3953 bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY, 3954 ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem); 3955 ctx->ifc_msix_mem = NULL; 3956 } 3957 3958 bus_generic_detach(dev); 3959 if_free(ifp); 3960 3961 iflib_tx_structures_free(ctx); 3962 iflib_rx_structures_free(ctx); 3963 if (ctx->ifc_flags & IFC_SC_ALLOCATED) 3964 free(ctx->ifc_softc, M_IFLIB); 3965 free(ctx, M_IFLIB); 3966 return (0); 3967 } 3968 3969 3970 int 3971 iflib_device_detach(device_t dev) 3972 { 3973 if_ctx_t ctx = device_get_softc(dev); 3974 3975 return (iflib_device_deregister(ctx)); 3976 } 3977 3978 int 3979 iflib_device_suspend(device_t dev) 3980 { 3981 if_ctx_t ctx = device_get_softc(dev); 3982 3983 CTX_LOCK(ctx); 3984 IFDI_SUSPEND(ctx); 3985 CTX_UNLOCK(ctx); 3986 3987 return bus_generic_suspend(dev); 3988 } 3989 int 3990 iflib_device_shutdown(device_t dev) 3991 { 3992 if_ctx_t ctx = device_get_softc(dev); 3993 3994 CTX_LOCK(ctx); 3995 IFDI_SHUTDOWN(ctx); 3996 CTX_UNLOCK(ctx); 3997 3998 return bus_generic_suspend(dev); 3999 } 4000 4001 4002 int 4003 iflib_device_resume(device_t dev) 4004 { 4005 if_ctx_t ctx = device_get_softc(dev); 4006 iflib_txq_t txq = ctx->ifc_txqs; 4007 4008 CTX_LOCK(ctx); 4009 IFDI_RESUME(ctx); 4010 iflib_init_locked(ctx); 4011 CTX_UNLOCK(ctx); 4012 for (int i = 0; i < NTXQSETS(ctx); i++, txq++) 4013 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); 4014 4015 return (bus_generic_resume(dev)); 4016 } 4017 4018 int 4019 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params) 4020 { 4021 int error; 4022 if_ctx_t ctx = device_get_softc(dev); 4023 4024 CTX_LOCK(ctx); 4025 error = IFDI_IOV_INIT(ctx, num_vfs, params); 4026 CTX_UNLOCK(ctx); 4027 4028 return (error); 4029 } 4030 4031 void 4032 iflib_device_iov_uninit(device_t dev) 4033 { 4034 if_ctx_t ctx = device_get_softc(dev); 4035 4036 CTX_LOCK(ctx); 4037 IFDI_IOV_UNINIT(ctx); 4038 CTX_UNLOCK(ctx); 4039 } 4040 4041 int 4042 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params) 4043 { 4044 int error; 4045 if_ctx_t ctx = device_get_softc(dev); 4046 4047 CTX_LOCK(ctx); 4048 error = IFDI_IOV_VF_ADD(ctx, vfnum, params); 4049 CTX_UNLOCK(ctx); 4050 4051 return (error); 4052 } 4053 4054 /********************************************************************* 4055 * 4056 * MODULE FUNCTION DEFINITIONS 4057 * 4058 **********************************************************************/ 4059 4060 /* 4061 * - Start a fast taskqueue thread for each core 4062 * - Start a taskqueue for control operations 4063 */ 4064 static int 4065 iflib_module_init(void) 4066 { 4067 return (0); 4068 } 4069 4070 static int 4071 iflib_module_event_handler(module_t mod, int what, void *arg) 4072 { 4073 int err; 4074 4075 switch (what) { 4076 case MOD_LOAD: 4077 if ((err = iflib_module_init()) != 0) 4078 return (err); 4079 break; 4080 case MOD_UNLOAD: 4081 return (EBUSY); 4082 default: 4083 return (EOPNOTSUPP); 4084 } 4085 4086 return (0); 4087 } 4088 4089 /********************************************************************* 4090 * 4091 * PUBLIC FUNCTION DEFINITIONS 4092 * ordered as in iflib.h 4093 * 4094 **********************************************************************/ 4095 4096 4097 static void 4098 _iflib_assert(if_shared_ctx_t sctx) 4099 { 4100 MPASS(sctx->isc_tx_maxsize); 4101 MPASS(sctx->isc_tx_maxsegsize); 4102 4103 MPASS(sctx->isc_rx_maxsize); 4104 MPASS(sctx->isc_rx_nsegments); 4105 MPASS(sctx->isc_rx_maxsegsize); 4106 4107 MPASS(sctx->isc_nrxd_min[0]); 4108 MPASS(sctx->isc_nrxd_max[0]); 4109 MPASS(sctx->isc_nrxd_default[0]); 4110 MPASS(sctx->isc_ntxd_min[0]); 4111 MPASS(sctx->isc_ntxd_max[0]); 4112 MPASS(sctx->isc_ntxd_default[0]); 4113 } 4114 4115 static void 4116 _iflib_pre_assert(if_softc_ctx_t scctx) 4117 { 4118 4119 MPASS(scctx->isc_txrx->ift_txd_encap); 4120 MPASS(scctx->isc_txrx->ift_txd_flush); 4121 MPASS(scctx->isc_txrx->ift_txd_credits_update); 4122 MPASS(scctx->isc_txrx->ift_rxd_available); 4123 MPASS(scctx->isc_txrx->ift_rxd_pkt_get); 4124 MPASS(scctx->isc_txrx->ift_rxd_refill); 4125 MPASS(scctx->isc_txrx->ift_rxd_flush); 4126 } 4127 4128 static int 4129 iflib_register(if_ctx_t ctx) 4130 { 4131 if_shared_ctx_t sctx = ctx->ifc_sctx; 4132 driver_t *driver = sctx->isc_driver; 4133 device_t dev = ctx->ifc_dev; 4134 if_t ifp; 4135 4136 _iflib_assert(sctx); 4137 4138 CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); 4139 4140 ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER); 4141 if (ifp == NULL) { 4142 device_printf(dev, "can not allocate ifnet structure\n"); 4143 return (ENOMEM); 4144 } 4145 4146 /* 4147 * Initialize our context's device specific methods 4148 */ 4149 kobj_init((kobj_t) ctx, (kobj_class_t) driver); 4150 kobj_class_compile((kobj_class_t) driver); 4151 driver->refs++; 4152 4153 if_initname(ifp, device_get_name(dev), device_get_unit(dev)); 4154 if_setsoftc(ifp, ctx); 4155 if_setdev(ifp, dev); 4156 if_setinitfn(ifp, iflib_if_init); 4157 if_setioctlfn(ifp, iflib_if_ioctl); 4158 if_settransmitfn(ifp, iflib_if_transmit); 4159 if_setqflushfn(ifp, iflib_if_qflush); 4160 if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); 4161 4162 ctx->ifc_vlan_attach_event = 4163 EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx, 4164 EVENTHANDLER_PRI_FIRST); 4165 ctx->ifc_vlan_detach_event = 4166 EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx, 4167 EVENTHANDLER_PRI_FIRST); 4168 4169 ifmedia_init(&ctx->ifc_media, IFM_IMASK, 4170 iflib_media_change, iflib_media_status); 4171 4172 return (0); 4173 } 4174 4175 4176 static int 4177 iflib_queues_alloc(if_ctx_t ctx) 4178 { 4179 if_shared_ctx_t sctx = ctx->ifc_sctx; 4180 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 4181 device_t dev = ctx->ifc_dev; 4182 int nrxqsets = scctx->isc_nrxqsets; 4183 int ntxqsets = scctx->isc_ntxqsets; 4184 iflib_txq_t txq; 4185 iflib_rxq_t rxq; 4186 iflib_fl_t fl = NULL; 4187 int i, j, cpu, err, txconf, rxconf; 4188 iflib_dma_info_t ifdip; 4189 uint32_t *rxqsizes = scctx->isc_rxqsizes; 4190 uint32_t *txqsizes = scctx->isc_txqsizes; 4191 uint8_t nrxqs = sctx->isc_nrxqs; 4192 uint8_t ntxqs = sctx->isc_ntxqs; 4193 int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; 4194 caddr_t *vaddrs; 4195 uint64_t *paddrs; 4196 struct ifmp_ring **brscp; 4197 int nbuf_rings = 1; /* XXX determine dynamically */ 4198 4199 KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1")); 4200 KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1")); 4201 4202 brscp = NULL; 4203 txq = NULL; 4204 rxq = NULL; 4205 4206 /* Allocate the TX ring struct memory */ 4207 if (!(txq = 4208 (iflib_txq_t) malloc(sizeof(struct iflib_txq) * 4209 ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { 4210 device_printf(dev, "Unable to allocate TX ring memory\n"); 4211 err = ENOMEM; 4212 goto fail; 4213 } 4214 4215 /* Now allocate the RX */ 4216 if (!(rxq = 4217 (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) * 4218 nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { 4219 device_printf(dev, "Unable to allocate RX ring memory\n"); 4220 err = ENOMEM; 4221 goto rx_fail; 4222 } 4223 if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { 4224 device_printf(dev, "Unable to buf_ring_sc * memory\n"); 4225 err = ENOMEM; 4226 goto rx_fail; 4227 } 4228 4229 ctx->ifc_txqs = txq; 4230 ctx->ifc_rxqs = rxq; 4231 4232 /* 4233 * XXX handle allocation failure 4234 */ 4235 for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) { 4236 /* Set up some basics */ 4237 4238 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) { 4239 device_printf(dev, "failed to allocate iflib_dma_info\n"); 4240 err = ENOMEM; 4241 goto err_tx_desc; 4242 } 4243 txq->ift_ifdi = ifdip; 4244 for (j = 0; j < ntxqs; j++, ifdip++) { 4245 if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) { 4246 device_printf(dev, "Unable to allocate Descriptor memory\n"); 4247 err = ENOMEM; 4248 goto err_tx_desc; 4249 } 4250 bzero((void *)ifdip->idi_vaddr, txqsizes[j]); 4251 } 4252 txq->ift_ctx = ctx; 4253 txq->ift_id = i; 4254 if (sctx->isc_flags & IFLIB_HAS_TXCQ) { 4255 txq->ift_br_offset = 1; 4256 } else { 4257 txq->ift_br_offset = 0; 4258 } 4259 /* XXX fix this */ 4260 txq->ift_timer.c_cpu = cpu; 4261 txq->ift_db_check.c_cpu = cpu; 4262 txq->ift_nbr = nbuf_rings; 4263 4264 if (iflib_txsd_alloc(txq)) { 4265 device_printf(dev, "Critical Failure setting up TX buffers\n"); 4266 err = ENOMEM; 4267 goto err_tx_desc; 4268 } 4269 4270 /* Initialize the TX lock */ 4271 snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout", 4272 device_get_nameunit(dev), txq->ift_id); 4273 mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF); 4274 callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0); 4275 callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0); 4276 4277 snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db", 4278 device_get_nameunit(dev), txq->ift_id); 4279 TXDB_LOCK_INIT(txq); 4280 4281 txq->ift_br = brscp + i*nbuf_rings; 4282 for (j = 0; j < nbuf_rings; j++) { 4283 err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain, 4284 iflib_txq_can_drain, M_IFLIB, M_WAITOK); 4285 if (err) { 4286 /* XXX free any allocated rings */ 4287 device_printf(dev, "Unable to allocate buf_ring\n"); 4288 goto err_tx_desc; 4289 } 4290 } 4291 } 4292 4293 for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) { 4294 /* Set up some basics */ 4295 4296 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) { 4297 device_printf(dev, "failed to allocate iflib_dma_info\n"); 4298 err = ENOMEM; 4299 goto err_tx_desc; 4300 } 4301 4302 rxq->ifr_ifdi = ifdip; 4303 for (j = 0; j < nrxqs; j++, ifdip++) { 4304 if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) { 4305 device_printf(dev, "Unable to allocate Descriptor memory\n"); 4306 err = ENOMEM; 4307 goto err_tx_desc; 4308 } 4309 bzero((void *)ifdip->idi_vaddr, rxqsizes[j]); 4310 } 4311 rxq->ifr_ctx = ctx; 4312 rxq->ifr_id = i; 4313 if (sctx->isc_flags & IFLIB_HAS_RXCQ) { 4314 rxq->ifr_fl_offset = 1; 4315 } else { 4316 rxq->ifr_fl_offset = 0; 4317 } 4318 rxq->ifr_nfl = nfree_lists; 4319 if (!(fl = 4320 (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) { 4321 device_printf(dev, "Unable to allocate free list memory\n"); 4322 err = ENOMEM; 4323 goto err_tx_desc; 4324 } 4325 rxq->ifr_fl = fl; 4326 for (j = 0; j < nfree_lists; j++) { 4327 rxq->ifr_fl[j].ifl_rxq = rxq; 4328 rxq->ifr_fl[j].ifl_id = j; 4329 rxq->ifr_fl[j].ifl_ifdi = 4330 &rxq->ifr_ifdi[j + rxq->ifr_fl_offset]; 4331 } 4332 /* Allocate receive buffers for the ring*/ 4333 if (iflib_rxsd_alloc(rxq)) { 4334 device_printf(dev, 4335 "Critical Failure setting up receive buffers\n"); 4336 err = ENOMEM; 4337 goto err_rx_desc; 4338 } 4339 } 4340 4341 /* TXQs */ 4342 vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); 4343 paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); 4344 for (i = 0; i < ntxqsets; i++) { 4345 iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi; 4346 4347 for (j = 0; j < ntxqs; j++, di++) { 4348 vaddrs[i*ntxqs + j] = di->idi_vaddr; 4349 paddrs[i*ntxqs + j] = di->idi_paddr; 4350 } 4351 } 4352 if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) { 4353 device_printf(ctx->ifc_dev, "device queue allocation failed\n"); 4354 iflib_tx_structures_free(ctx); 4355 free(vaddrs, M_IFLIB); 4356 free(paddrs, M_IFLIB); 4357 goto err_rx_desc; 4358 } 4359 free(vaddrs, M_IFLIB); 4360 free(paddrs, M_IFLIB); 4361 4362 /* RXQs */ 4363 vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); 4364 paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); 4365 for (i = 0; i < nrxqsets; i++) { 4366 iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi; 4367 4368 for (j = 0; j < nrxqs; j++, di++) { 4369 vaddrs[i*nrxqs + j] = di->idi_vaddr; 4370 paddrs[i*nrxqs + j] = di->idi_paddr; 4371 } 4372 } 4373 if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) { 4374 device_printf(ctx->ifc_dev, "device queue allocation failed\n"); 4375 iflib_tx_structures_free(ctx); 4376 free(vaddrs, M_IFLIB); 4377 free(paddrs, M_IFLIB); 4378 goto err_rx_desc; 4379 } 4380 free(vaddrs, M_IFLIB); 4381 free(paddrs, M_IFLIB); 4382 4383 return (0); 4384 4385 /* XXX handle allocation failure changes */ 4386 err_rx_desc: 4387 err_tx_desc: 4388 if (ctx->ifc_rxqs != NULL) 4389 free(ctx->ifc_rxqs, M_IFLIB); 4390 ctx->ifc_rxqs = NULL; 4391 if (ctx->ifc_txqs != NULL) 4392 free(ctx->ifc_txqs, M_IFLIB); 4393 ctx->ifc_txqs = NULL; 4394 rx_fail: 4395 if (brscp != NULL) 4396 free(brscp, M_IFLIB); 4397 if (rxq != NULL) 4398 free(rxq, M_IFLIB); 4399 if (txq != NULL) 4400 free(txq, M_IFLIB); 4401 fail: 4402 return (err); 4403 } 4404 4405 static int 4406 iflib_tx_structures_setup(if_ctx_t ctx) 4407 { 4408 iflib_txq_t txq = ctx->ifc_txqs; 4409 int i; 4410 4411 for (i = 0; i < NTXQSETS(ctx); i++, txq++) 4412 iflib_txq_setup(txq); 4413 4414 return (0); 4415 } 4416 4417 static void 4418 iflib_tx_structures_free(if_ctx_t ctx) 4419 { 4420 iflib_txq_t txq = ctx->ifc_txqs; 4421 int i, j; 4422 4423 for (i = 0; i < NTXQSETS(ctx); i++, txq++) { 4424 iflib_txq_destroy(txq); 4425 for (j = 0; j < ctx->ifc_nhwtxqs; j++) 4426 iflib_dma_free(&txq->ift_ifdi[j]); 4427 } 4428 free(ctx->ifc_txqs, M_IFLIB); 4429 ctx->ifc_txqs = NULL; 4430 IFDI_QUEUES_FREE(ctx); 4431 } 4432 4433 /********************************************************************* 4434 * 4435 * Initialize all receive rings. 4436 * 4437 **********************************************************************/ 4438 static int 4439 iflib_rx_structures_setup(if_ctx_t ctx) 4440 { 4441 iflib_rxq_t rxq = ctx->ifc_rxqs; 4442 int q; 4443 #if defined(INET6) || defined(INET) 4444 int i, err; 4445 #endif 4446 4447 for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { 4448 #if defined(INET6) || defined(INET) 4449 tcp_lro_free(&rxq->ifr_lc); 4450 if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp, 4451 TCP_LRO_ENTRIES, min(1024, 4452 ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) { 4453 device_printf(ctx->ifc_dev, "LRO Initialization failed!\n"); 4454 goto fail; 4455 } 4456 rxq->ifr_lro_enabled = TRUE; 4457 #endif 4458 IFDI_RXQ_SETUP(ctx, rxq->ifr_id); 4459 } 4460 return (0); 4461 #if defined(INET6) || defined(INET) 4462 fail: 4463 /* 4464 * Free RX software descriptors allocated so far, we will only handle 4465 * the rings that completed, the failing case will have 4466 * cleaned up for itself. 'q' failed, so its the terminus. 4467 */ 4468 rxq = ctx->ifc_rxqs; 4469 for (i = 0; i < q; ++i, rxq++) { 4470 iflib_rx_sds_free(rxq); 4471 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0; 4472 } 4473 return (err); 4474 #endif 4475 } 4476 4477 /********************************************************************* 4478 * 4479 * Free all receive rings. 4480 * 4481 **********************************************************************/ 4482 static void 4483 iflib_rx_structures_free(if_ctx_t ctx) 4484 { 4485 iflib_rxq_t rxq = ctx->ifc_rxqs; 4486 4487 for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) { 4488 iflib_rx_sds_free(rxq); 4489 } 4490 } 4491 4492 static int 4493 iflib_qset_structures_setup(if_ctx_t ctx) 4494 { 4495 int err; 4496 4497 if ((err = iflib_tx_structures_setup(ctx)) != 0) 4498 return (err); 4499 4500 if ((err = iflib_rx_structures_setup(ctx)) != 0) { 4501 device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err); 4502 iflib_tx_structures_free(ctx); 4503 iflib_rx_structures_free(ctx); 4504 } 4505 return (err); 4506 } 4507 4508 int 4509 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, 4510 driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name) 4511 { 4512 4513 return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name)); 4514 } 4515 4516 static int 4517 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid) 4518 { 4519 int i, cpuid, eqid, count; 4520 4521 CPU_COPY(&ctx->ifc_cpus, cpus); 4522 count = CPU_COUNT(&ctx->ifc_cpus); 4523 eqid = qid % count; 4524 /* clear up to the qid'th bit */ 4525 for (i = 0; i < eqid; i++) { 4526 cpuid = CPU_FFS(cpus); 4527 MPASS(cpuid != 0); 4528 CPU_CLR(cpuid-1, cpus); 4529 } 4530 cpuid = CPU_FFS(cpus); 4531 MPASS(cpuid != 0); 4532 return (cpuid-1); 4533 } 4534 4535 int 4536 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid, 4537 iflib_intr_type_t type, driver_filter_t *filter, 4538 void *filter_arg, int qid, char *name) 4539 { 4540 struct grouptask *gtask; 4541 struct taskqgroup *tqg; 4542 iflib_filter_info_t info; 4543 cpuset_t cpus; 4544 gtask_fn_t *fn; 4545 int tqrid, err, cpuid; 4546 void *q; 4547 4548 info = &ctx->ifc_filter_info; 4549 tqrid = rid; 4550 4551 switch (type) { 4552 /* XXX merge tx/rx for netmap? */ 4553 case IFLIB_INTR_TX: 4554 q = &ctx->ifc_txqs[qid]; 4555 info = &ctx->ifc_txqs[qid].ift_filter_info; 4556 gtask = &ctx->ifc_txqs[qid].ift_task; 4557 tqg = qgroup_if_io_tqg; 4558 fn = _task_fn_tx; 4559 GROUPTASK_INIT(gtask, 0, fn, q); 4560 break; 4561 case IFLIB_INTR_RX: 4562 q = &ctx->ifc_rxqs[qid]; 4563 info = &ctx->ifc_rxqs[qid].ifr_filter_info; 4564 gtask = &ctx->ifc_rxqs[qid].ifr_task; 4565 tqg = qgroup_if_io_tqg; 4566 fn = _task_fn_rx; 4567 GROUPTASK_INIT(gtask, 0, fn, q); 4568 break; 4569 case IFLIB_INTR_ADMIN: 4570 q = ctx; 4571 tqrid = -1; 4572 info = &ctx->ifc_filter_info; 4573 gtask = &ctx->ifc_admin_task; 4574 tqg = qgroup_if_config_tqg; 4575 fn = _task_fn_admin; 4576 break; 4577 default: 4578 panic("unknown net intr type"); 4579 } 4580 4581 info->ifi_filter = filter; 4582 info->ifi_filter_arg = filter_arg; 4583 info->ifi_task = gtask; 4584 info->ifi_ctx = ctx; 4585 4586 err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info, name); 4587 if (err != 0) { 4588 device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err); 4589 return (err); 4590 } 4591 if (type == IFLIB_INTR_ADMIN) 4592 return (0); 4593 4594 if (tqrid != -1) { 4595 cpuid = find_nth(ctx, &cpus, qid); 4596 taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name); 4597 } else { 4598 taskqgroup_attach(tqg, gtask, q, tqrid, name); 4599 } 4600 4601 return (0); 4602 } 4603 4604 void 4605 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type, void *arg, int qid, char *name) 4606 { 4607 struct grouptask *gtask; 4608 struct taskqgroup *tqg; 4609 gtask_fn_t *fn; 4610 void *q; 4611 4612 switch (type) { 4613 case IFLIB_INTR_TX: 4614 q = &ctx->ifc_txqs[qid]; 4615 gtask = &ctx->ifc_txqs[qid].ift_task; 4616 tqg = qgroup_if_io_tqg; 4617 fn = _task_fn_tx; 4618 break; 4619 case IFLIB_INTR_RX: 4620 q = &ctx->ifc_rxqs[qid]; 4621 gtask = &ctx->ifc_rxqs[qid].ifr_task; 4622 tqg = qgroup_if_io_tqg; 4623 fn = _task_fn_rx; 4624 break; 4625 case IFLIB_INTR_IOV: 4626 q = ctx; 4627 gtask = &ctx->ifc_vflr_task; 4628 tqg = qgroup_if_config_tqg; 4629 rid = -1; 4630 fn = _task_fn_iov; 4631 break; 4632 default: 4633 panic("unknown net intr type"); 4634 } 4635 GROUPTASK_INIT(gtask, 0, fn, q); 4636 taskqgroup_attach(tqg, gtask, q, rid, name); 4637 } 4638 4639 void 4640 iflib_irq_free(if_ctx_t ctx, if_irq_t irq) 4641 { 4642 if (irq->ii_tag) 4643 bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag); 4644 4645 if (irq->ii_res) 4646 bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res); 4647 } 4648 4649 static int 4650 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name) 4651 { 4652 iflib_txq_t txq = ctx->ifc_txqs; 4653 iflib_rxq_t rxq = ctx->ifc_rxqs; 4654 if_irq_t irq = &ctx->ifc_legacy_irq; 4655 iflib_filter_info_t info; 4656 struct grouptask *gtask; 4657 struct taskqgroup *tqg; 4658 gtask_fn_t *fn; 4659 int tqrid; 4660 void *q; 4661 int err; 4662 4663 q = &ctx->ifc_rxqs[0]; 4664 info = &rxq[0].ifr_filter_info; 4665 gtask = &rxq[0].ifr_task; 4666 tqg = qgroup_if_io_tqg; 4667 tqrid = irq->ii_rid = *rid; 4668 fn = _task_fn_rx; 4669 4670 ctx->ifc_flags |= IFC_LEGACY; 4671 info->ifi_filter = filter; 4672 info->ifi_filter_arg = filter_arg; 4673 info->ifi_task = gtask; 4674 info->ifi_ctx = ctx; 4675 4676 /* We allocate a single interrupt resource */ 4677 if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0) 4678 return (err); 4679 GROUPTASK_INIT(gtask, 0, fn, q); 4680 taskqgroup_attach(tqg, gtask, q, tqrid, name); 4681 4682 GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq); 4683 taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx"); 4684 return (0); 4685 } 4686 4687 void 4688 iflib_led_create(if_ctx_t ctx) 4689 { 4690 4691 ctx->ifc_led_dev = led_create(iflib_led_func, ctx, 4692 device_get_nameunit(ctx->ifc_dev)); 4693 } 4694 4695 void 4696 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid) 4697 { 4698 4699 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task); 4700 } 4701 4702 void 4703 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid) 4704 { 4705 4706 GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task); 4707 } 4708 4709 void 4710 iflib_admin_intr_deferred(if_ctx_t ctx) 4711 { 4712 #ifdef INVARIANTS 4713 struct grouptask *gtask; 4714 4715 gtask = &ctx->ifc_admin_task; 4716 MPASS(gtask->gt_taskqueue != NULL); 4717 #endif 4718 4719 GROUPTASK_ENQUEUE(&ctx->ifc_admin_task); 4720 } 4721 4722 void 4723 iflib_iov_intr_deferred(if_ctx_t ctx) 4724 { 4725 4726 GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task); 4727 } 4728 4729 void 4730 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name) 4731 { 4732 4733 taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name); 4734 } 4735 4736 void 4737 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn, 4738 char *name) 4739 { 4740 4741 GROUPTASK_INIT(gtask, 0, fn, ctx); 4742 taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name); 4743 } 4744 4745 void 4746 iflib_config_gtask_deinit(struct grouptask *gtask) 4747 { 4748 4749 taskqgroup_detach(qgroup_if_config_tqg, gtask); 4750 } 4751 4752 void 4753 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate) 4754 { 4755 if_t ifp = ctx->ifc_ifp; 4756 iflib_txq_t txq = ctx->ifc_txqs; 4757 4758 if_setbaudrate(ifp, baudrate); 4759 4760 /* If link down, disable watchdog */ 4761 if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { 4762 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) 4763 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 4764 } 4765 ctx->ifc_link_state = link_state; 4766 if_link_state_change(ifp, link_state); 4767 } 4768 4769 static int 4770 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq) 4771 { 4772 int credits; 4773 #ifdef INVARIANTS 4774 int credits_pre = txq->ift_cidx_processed; 4775 #endif 4776 4777 if (ctx->isc_txd_credits_update == NULL) 4778 return (0); 4779 4780 if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0) 4781 return (0); 4782 4783 txq->ift_processed += credits; 4784 txq->ift_cidx_processed += credits; 4785 4786 MPASS(credits_pre + credits == txq->ift_cidx_processed); 4787 if (txq->ift_cidx_processed >= txq->ift_size) 4788 txq->ift_cidx_processed -= txq->ift_size; 4789 return (credits); 4790 } 4791 4792 static int 4793 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx, int budget) 4794 { 4795 4796 return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx, 4797 budget)); 4798 } 4799 4800 void 4801 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name, 4802 const char *description, if_int_delay_info_t info, 4803 int offset, int value) 4804 { 4805 info->iidi_ctx = ctx; 4806 info->iidi_offset = offset; 4807 info->iidi_value = value; 4808 SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev), 4809 SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)), 4810 OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW, 4811 info, 0, iflib_sysctl_int_delay, "I", description); 4812 } 4813 4814 struct mtx * 4815 iflib_ctx_lock_get(if_ctx_t ctx) 4816 { 4817 4818 return (&ctx->ifc_mtx); 4819 } 4820 4821 static int 4822 iflib_msix_init(if_ctx_t ctx) 4823 { 4824 device_t dev = ctx->ifc_dev; 4825 if_shared_ctx_t sctx = ctx->ifc_sctx; 4826 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 4827 int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs; 4828 int iflib_num_tx_queues, iflib_num_rx_queues; 4829 int err, admincnt, bar; 4830 4831 iflib_num_tx_queues = scctx->isc_ntxqsets; 4832 iflib_num_rx_queues = scctx->isc_nrxqsets; 4833 4834 device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues); 4835 4836 bar = ctx->ifc_softc_ctx.isc_msix_bar; 4837 admincnt = sctx->isc_admin_intrcnt; 4838 /* Override by tuneable */ 4839 if (enable_msix == 0) 4840 goto msi; 4841 4842 /* 4843 ** When used in a virtualized environment 4844 ** PCI BUSMASTER capability may not be set 4845 ** so explicity set it here and rewrite 4846 ** the ENABLE in the MSIX control register 4847 ** at this point to cause the host to 4848 ** successfully initialize us. 4849 */ 4850 { 4851 int msix_ctrl, rid; 4852 4853 pci_enable_busmaster(dev); 4854 rid = 0; 4855 if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) { 4856 rid += PCIR_MSIX_CTRL; 4857 msix_ctrl = pci_read_config(dev, rid, 2); 4858 msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE; 4859 pci_write_config(dev, rid, msix_ctrl, 2); 4860 } else { 4861 device_printf(dev, "PCIY_MSIX capability not found; " 4862 "or rid %d == 0.\n", rid); 4863 goto msi; 4864 } 4865 } 4866 4867 /* 4868 * bar == -1 => "trust me I know what I'm doing" 4869 * https://www.youtube.com/watch?v=nnwWKkNau4I 4870 * Some drivers are for hardware that is so shoddily 4871 * documented that no one knows which bars are which 4872 * so the developer has to map all bars. This hack 4873 * allows shoddy garbage to use msix in this framework. 4874 */ 4875 if (bar != -1) { 4876 ctx->ifc_msix_mem = bus_alloc_resource_any(dev, 4877 SYS_RES_MEMORY, &bar, RF_ACTIVE); 4878 if (ctx->ifc_msix_mem == NULL) { 4879 /* May not be enabled */ 4880 device_printf(dev, "Unable to map MSIX table \n"); 4881 goto msi; 4882 } 4883 } 4884 /* First try MSI/X */ 4885 if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */ 4886 device_printf(dev, "System has MSIX disabled \n"); 4887 bus_release_resource(dev, SYS_RES_MEMORY, 4888 bar, ctx->ifc_msix_mem); 4889 ctx->ifc_msix_mem = NULL; 4890 goto msi; 4891 } 4892 #if IFLIB_DEBUG 4893 /* use only 1 qset in debug mode */ 4894 queuemsgs = min(msgs - admincnt, 1); 4895 #else 4896 queuemsgs = msgs - admincnt; 4897 #endif 4898 if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) { 4899 #ifdef RSS 4900 queues = imin(queuemsgs, rss_getnumbuckets()); 4901 #else 4902 queues = queuemsgs; 4903 #endif 4904 queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues); 4905 device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n", 4906 CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt); 4907 } else { 4908 device_printf(dev, "Unable to fetch CPU list\n"); 4909 /* Figure out a reasonable auto config value */ 4910 queues = min(queuemsgs, mp_ncpus); 4911 } 4912 #ifdef RSS 4913 /* If we're doing RSS, clamp at the number of RSS buckets */ 4914 if (queues > rss_getnumbuckets()) 4915 queues = rss_getnumbuckets(); 4916 #endif 4917 if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt) 4918 rx_queues = iflib_num_rx_queues; 4919 else 4920 rx_queues = queues; 4921 /* 4922 * We want this to be all logical CPUs by default 4923 */ 4924 if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) 4925 tx_queues = iflib_num_tx_queues; 4926 else 4927 tx_queues = mp_ncpus; 4928 4929 if (ctx->ifc_sysctl_qs_eq_override == 0) { 4930 #ifdef INVARIANTS 4931 if (tx_queues != rx_queues) 4932 device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n", 4933 min(rx_queues, tx_queues), min(rx_queues, tx_queues)); 4934 #endif 4935 tx_queues = min(rx_queues, tx_queues); 4936 rx_queues = min(rx_queues, tx_queues); 4937 } 4938 4939 device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues); 4940 4941 vectors = rx_queues + admincnt; 4942 if ((err = pci_alloc_msix(dev, &vectors)) == 0) { 4943 device_printf(dev, 4944 "Using MSIX interrupts with %d vectors\n", vectors); 4945 scctx->isc_vectors = vectors; 4946 scctx->isc_nrxqsets = rx_queues; 4947 scctx->isc_ntxqsets = tx_queues; 4948 scctx->isc_intr = IFLIB_INTR_MSIX; 4949 4950 return (vectors); 4951 } else { 4952 device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err); 4953 } 4954 msi: 4955 vectors = pci_msi_count(dev); 4956 scctx->isc_nrxqsets = 1; 4957 scctx->isc_ntxqsets = 1; 4958 scctx->isc_vectors = vectors; 4959 if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) { 4960 device_printf(dev,"Using an MSI interrupt\n"); 4961 scctx->isc_intr = IFLIB_INTR_MSI; 4962 } else { 4963 device_printf(dev,"Using a Legacy interrupt\n"); 4964 scctx->isc_intr = IFLIB_INTR_LEGACY; 4965 } 4966 4967 return (vectors); 4968 } 4969 4970 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" }; 4971 4972 static int 4973 mp_ring_state_handler(SYSCTL_HANDLER_ARGS) 4974 { 4975 int rc; 4976 uint16_t *state = ((uint16_t *)oidp->oid_arg1); 4977 struct sbuf *sb; 4978 char *ring_state = "UNKNOWN"; 4979 4980 /* XXX needed ? */ 4981 rc = sysctl_wire_old_buffer(req, 0); 4982 MPASS(rc == 0); 4983 if (rc != 0) 4984 return (rc); 4985 sb = sbuf_new_for_sysctl(NULL, NULL, 80, req); 4986 MPASS(sb != NULL); 4987 if (sb == NULL) 4988 return (ENOMEM); 4989 if (state[3] <= 3) 4990 ring_state = ring_states[state[3]]; 4991 4992 sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s", 4993 state[0], state[1], state[2], ring_state); 4994 rc = sbuf_finish(sb); 4995 sbuf_delete(sb); 4996 return(rc); 4997 } 4998 4999 enum iflib_ndesc_handler { 5000 IFLIB_NTXD_HANDLER, 5001 IFLIB_NRXD_HANDLER, 5002 }; 5003 5004 static int 5005 mp_ndesc_handler(SYSCTL_HANDLER_ARGS) 5006 { 5007 if_ctx_t ctx = (void *)arg1; 5008 enum iflib_ndesc_handler type = arg2; 5009 char buf[256] = {0}; 5010 uint16_t *ndesc; 5011 char *p, *next; 5012 int nqs, rc, i; 5013 5014 MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER); 5015 5016 nqs = 8; 5017 switch(type) { 5018 case IFLIB_NTXD_HANDLER: 5019 ndesc = ctx->ifc_sysctl_ntxds; 5020 if (ctx->ifc_sctx) 5021 nqs = ctx->ifc_sctx->isc_ntxqs; 5022 break; 5023 case IFLIB_NRXD_HANDLER: 5024 ndesc = ctx->ifc_sysctl_nrxds; 5025 if (ctx->ifc_sctx) 5026 nqs = ctx->ifc_sctx->isc_nrxqs; 5027 break; 5028 } 5029 if (nqs == 0) 5030 nqs = 8; 5031 5032 for (i=0; i<8; i++) { 5033 if (i >= nqs) 5034 break; 5035 if (i) 5036 strcat(buf, ","); 5037 sprintf(strchr(buf, 0), "%d", ndesc[i]); 5038 } 5039 5040 rc = sysctl_handle_string(oidp, buf, sizeof(buf), req); 5041 if (rc || req->newptr == NULL) 5042 return rc; 5043 5044 for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p; 5045 i++, p = strsep(&next, " ,")) { 5046 ndesc[i] = strtoul(p, NULL, 10); 5047 } 5048 5049 return(rc); 5050 } 5051 5052 #define NAME_BUFLEN 32 5053 static void 5054 iflib_add_device_sysctl_pre(if_ctx_t ctx) 5055 { 5056 device_t dev = iflib_get_dev(ctx); 5057 struct sysctl_oid_list *child, *oid_list; 5058 struct sysctl_ctx_list *ctx_list; 5059 struct sysctl_oid *node; 5060 5061 ctx_list = device_get_sysctl_ctx(dev); 5062 child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); 5063 ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib", 5064 CTLFLAG_RD, NULL, "IFLIB fields"); 5065 oid_list = SYSCTL_CHILDREN(node); 5066 5067 SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version", 5068 CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0, 5069 "driver version"); 5070 5071 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", 5072 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, 5073 "# of txqs to use, 0 => use default #"); 5074 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", 5075 CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0, 5076 "# of rxqs to use, 0 => use default #"); 5077 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable", 5078 CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0, 5079 "permit #txq != #rxq"); 5080 5081 /* XXX change for per-queue sizes */ 5082 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds", 5083 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER, 5084 mp_ndesc_handler, "A", 5085 "list of # of tx descriptors to use, 0 = use default #"); 5086 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds", 5087 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER, 5088 mp_ndesc_handler, "A", 5089 "list of # of rx descriptors to use, 0 = use default #"); 5090 } 5091 5092 static void 5093 iflib_add_device_sysctl_post(if_ctx_t ctx) 5094 { 5095 if_shared_ctx_t sctx = ctx->ifc_sctx; 5096 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 5097 device_t dev = iflib_get_dev(ctx); 5098 struct sysctl_oid_list *child; 5099 struct sysctl_ctx_list *ctx_list; 5100 iflib_fl_t fl; 5101 iflib_txq_t txq; 5102 iflib_rxq_t rxq; 5103 int i, j; 5104 char namebuf[NAME_BUFLEN]; 5105 char *qfmt; 5106 struct sysctl_oid *queue_node, *fl_node, *node; 5107 struct sysctl_oid_list *queue_list, *fl_list; 5108 ctx_list = device_get_sysctl_ctx(dev); 5109 5110 node = ctx->ifc_sysctl_node; 5111 child = SYSCTL_CHILDREN(node); 5112 5113 if (scctx->isc_ntxqsets > 100) 5114 qfmt = "txq%03d"; 5115 else if (scctx->isc_ntxqsets > 10) 5116 qfmt = "txq%02d"; 5117 else 5118 qfmt = "txq%d"; 5119 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { 5120 snprintf(namebuf, NAME_BUFLEN, qfmt, i); 5121 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, 5122 CTLFLAG_RD, NULL, "Queue Name"); 5123 queue_list = SYSCTL_CHILDREN(queue_node); 5124 #if MEMORY_LOGGING 5125 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued", 5126 CTLFLAG_RD, 5127 &txq->ift_dequeued, "total mbufs freed"); 5128 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued", 5129 CTLFLAG_RD, 5130 &txq->ift_enqueued, "total mbufs enqueued"); 5131 #endif 5132 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag", 5133 CTLFLAG_RD, 5134 &txq->ift_mbuf_defrag, "# of times m_defrag was called"); 5135 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups", 5136 CTLFLAG_RD, 5137 &txq->ift_pullups, "# of times m_pullup was called"); 5138 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed", 5139 CTLFLAG_RD, 5140 &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); 5141 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail", 5142 CTLFLAG_RD, 5143 &txq->ift_no_desc_avail, "# of times no descriptors were available"); 5144 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed", 5145 CTLFLAG_RD, 5146 &txq->ift_map_failed, "# of times dma map failed"); 5147 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig", 5148 CTLFLAG_RD, 5149 &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG"); 5150 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup", 5151 CTLFLAG_RD, 5152 &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG"); 5153 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx", 5154 CTLFLAG_RD, 5155 &txq->ift_pidx, 1, "Producer Index"); 5156 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx", 5157 CTLFLAG_RD, 5158 &txq->ift_cidx, 1, "Consumer Index"); 5159 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed", 5160 CTLFLAG_RD, 5161 &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update"); 5162 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use", 5163 CTLFLAG_RD, 5164 &txq->ift_in_use, 1, "descriptors in use"); 5165 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed", 5166 CTLFLAG_RD, 5167 &txq->ift_processed, "descriptors procesed for clean"); 5168 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned", 5169 CTLFLAG_RD, 5170 &txq->ift_cleaned, "total cleaned"); 5171 SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state", 5172 CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state), 5173 0, mp_ring_state_handler, "A", "soft ring state"); 5174 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues", 5175 CTLFLAG_RD, &txq->ift_br[0]->enqueues, 5176 "# of enqueues to the mp_ring for this queue"); 5177 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops", 5178 CTLFLAG_RD, &txq->ift_br[0]->drops, 5179 "# of drops in the mp_ring for this queue"); 5180 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts", 5181 CTLFLAG_RD, &txq->ift_br[0]->starts, 5182 "# of normal consumer starts in the mp_ring for this queue"); 5183 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls", 5184 CTLFLAG_RD, &txq->ift_br[0]->stalls, 5185 "# of consumer stalls in the mp_ring for this queue"); 5186 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts", 5187 CTLFLAG_RD, &txq->ift_br[0]->restarts, 5188 "# of consumer restarts in the mp_ring for this queue"); 5189 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications", 5190 CTLFLAG_RD, &txq->ift_br[0]->abdications, 5191 "# of consumer abdications in the mp_ring for this queue"); 5192 } 5193 5194 if (scctx->isc_nrxqsets > 100) 5195 qfmt = "rxq%03d"; 5196 else if (scctx->isc_nrxqsets > 10) 5197 qfmt = "rxq%02d"; 5198 else 5199 qfmt = "rxq%d"; 5200 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { 5201 snprintf(namebuf, NAME_BUFLEN, qfmt, i); 5202 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, 5203 CTLFLAG_RD, NULL, "Queue Name"); 5204 queue_list = SYSCTL_CHILDREN(queue_node); 5205 if (sctx->isc_flags & IFLIB_HAS_RXCQ) { 5206 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx", 5207 CTLFLAG_RD, 5208 &rxq->ifr_cq_pidx, 1, "Producer Index"); 5209 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx", 5210 CTLFLAG_RD, 5211 &rxq->ifr_cq_cidx, 1, "Consumer Index"); 5212 } 5213 5214 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { 5215 snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j); 5216 fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf, 5217 CTLFLAG_RD, NULL, "freelist Name"); 5218 fl_list = SYSCTL_CHILDREN(fl_node); 5219 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx", 5220 CTLFLAG_RD, 5221 &fl->ifl_pidx, 1, "Producer Index"); 5222 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx", 5223 CTLFLAG_RD, 5224 &fl->ifl_cidx, 1, "Consumer Index"); 5225 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits", 5226 CTLFLAG_RD, 5227 &fl->ifl_credits, 1, "credits available"); 5228 #if MEMORY_LOGGING 5229 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued", 5230 CTLFLAG_RD, 5231 &fl->ifl_m_enqueued, "mbufs allocated"); 5232 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued", 5233 CTLFLAG_RD, 5234 &fl->ifl_m_dequeued, "mbufs freed"); 5235 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued", 5236 CTLFLAG_RD, 5237 &fl->ifl_cl_enqueued, "clusters allocated"); 5238 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued", 5239 CTLFLAG_RD, 5240 &fl->ifl_cl_dequeued, "clusters freed"); 5241 #endif 5242 5243 } 5244 } 5245 5246 } 5247