1 /*- 2 * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions are met: 7 * 8 * 1. Redistributions of source code must retain the above copyright notice, 9 * this list of conditions and the following disclaimer. 10 * 11 * 2. Neither the name of Matthew Macy nor the names of its 12 * contributors may be used to endorse or promote products derived from 13 * this software without specific prior written permission. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 16 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE 19 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 20 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 21 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 22 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 23 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 24 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 25 * POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 #include "opt_inet.h" 30 #include "opt_inet6.h" 31 #include "opt_acpi.h" 32 #include "opt_sched.h" 33 34 #include <sys/param.h> 35 #include <sys/types.h> 36 #include <sys/bus.h> 37 #include <sys/eventhandler.h> 38 #include <sys/kernel.h> 39 #include <sys/lock.h> 40 #include <sys/mutex.h> 41 #include <sys/module.h> 42 #include <sys/kobj.h> 43 #include <sys/rman.h> 44 #include <sys/sbuf.h> 45 #include <sys/smp.h> 46 #include <sys/socket.h> 47 #include <sys/sockio.h> 48 #include <sys/sysctl.h> 49 #include <sys/syslog.h> 50 #include <sys/taskqueue.h> 51 #include <sys/limits.h> 52 53 #include <net/if.h> 54 #include <net/if_var.h> 55 #include <net/if_private.h> 56 #include <net/if_types.h> 57 #include <net/if_media.h> 58 #include <net/bpf.h> 59 #include <net/ethernet.h> 60 #include <net/mp_ring.h> 61 #include <net/debugnet.h> 62 #include <net/pfil.h> 63 #include <net/vnet.h> 64 65 #include <netinet/in.h> 66 #include <netinet/in_pcb.h> 67 #include <netinet/tcp_lro.h> 68 #include <netinet/in_systm.h> 69 #include <netinet/if_ether.h> 70 #include <netinet/ip.h> 71 #include <netinet/ip6.h> 72 #include <netinet/tcp.h> 73 #include <netinet/udp.h> 74 #include <netinet/ip_var.h> 75 #include <netinet6/ip6_var.h> 76 77 #include <machine/bus.h> 78 #include <machine/in_cksum.h> 79 80 #include <vm/vm.h> 81 #include <vm/pmap.h> 82 83 #include <dev/led/led.h> 84 #include <dev/pci/pcireg.h> 85 #include <dev/pci/pcivar.h> 86 #include <dev/pci/pci_private.h> 87 88 #include <net/iflib.h> 89 90 #include "ifdi_if.h" 91 92 #ifdef PCI_IOV 93 #include <dev/pci/pci_iov.h> 94 #endif 95 96 #include <sys/bitstring.h> 97 /* 98 * enable accounting of every mbuf as it comes in to and goes out of 99 * iflib's software descriptor references 100 */ 101 #define MEMORY_LOGGING 0 102 /* 103 * Enable mbuf vectors for compressing long mbuf chains 104 */ 105 106 /* 107 * NB: 108 * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead 109 * we prefetch needs to be determined by the time spent in m_free vis a vis 110 * the cost of a prefetch. This will of course vary based on the workload: 111 * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which 112 * is quite expensive, thus suggesting very little prefetch. 113 * - small packet forwarding which is just returning a single mbuf to 114 * UMA will typically be very fast vis a vis the cost of a memory 115 * access. 116 */ 117 118 /* 119 * File organization: 120 * - private structures 121 * - iflib private utility functions 122 * - ifnet functions 123 * - vlan registry and other exported functions 124 * - iflib public core functions 125 * 126 * 127 */ 128 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library"); 129 130 #define IFLIB_RXEOF_MORE (1U << 0) 131 #define IFLIB_RXEOF_EMPTY (2U << 0) 132 133 struct iflib_txq; 134 typedef struct iflib_txq *iflib_txq_t; 135 struct iflib_rxq; 136 typedef struct iflib_rxq *iflib_rxq_t; 137 struct iflib_fl; 138 typedef struct iflib_fl *iflib_fl_t; 139 140 struct iflib_ctx; 141 142 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid); 143 static void iflib_timer(void *arg); 144 static void iflib_tqg_detach(if_ctx_t ctx); 145 static int iflib_simple_transmit(if_t ifp, struct mbuf *m); 146 147 typedef struct iflib_filter_info { 148 driver_filter_t *ifi_filter; 149 void *ifi_filter_arg; 150 struct grouptask *ifi_task; 151 void *ifi_ctx; 152 } *iflib_filter_info_t; 153 154 struct iflib_ctx { 155 KOBJ_FIELDS; 156 /* 157 * Pointer to hardware driver's softc 158 */ 159 void *ifc_softc; 160 device_t ifc_dev; 161 if_t ifc_ifp; 162 163 cpuset_t ifc_cpus; 164 if_shared_ctx_t ifc_sctx; 165 struct if_softc_ctx ifc_softc_ctx; 166 167 struct sx ifc_ctx_sx; 168 struct mtx ifc_state_mtx; 169 170 iflib_txq_t ifc_txqs; 171 iflib_rxq_t ifc_rxqs; 172 uint32_t ifc_if_flags; 173 uint32_t ifc_flags; 174 uint32_t ifc_max_fl_buf_size; 175 uint32_t ifc_rx_mbuf_sz; 176 177 int ifc_link_state; 178 int ifc_watchdog_events; 179 struct cdev *ifc_led_dev; 180 struct resource *ifc_msix_mem; 181 182 struct if_irq ifc_legacy_irq; 183 struct task ifc_admin_task; 184 struct task ifc_vflr_task; 185 struct taskqueue *ifc_tq; 186 struct iflib_filter_info ifc_filter_info; 187 struct ifmedia ifc_media; 188 struct ifmedia *ifc_mediap; 189 190 struct sysctl_oid *ifc_sysctl_node; 191 uint16_t ifc_sysctl_ntxqs; 192 uint16_t ifc_sysctl_nrxqs; 193 uint16_t ifc_sysctl_qs_eq_override; 194 uint16_t ifc_sysctl_rx_budget; 195 uint16_t ifc_sysctl_tx_abdicate; 196 uint16_t ifc_sysctl_core_offset; 197 #define CORE_OFFSET_UNSPECIFIED 0xffff 198 uint8_t ifc_sysctl_separate_txrx; 199 uint8_t ifc_sysctl_use_logical_cores; 200 uint16_t ifc_sysctl_extra_msix_vectors; 201 bool ifc_cpus_are_physical_cores; 202 bool ifc_sysctl_simple_tx; 203 204 qidx_t ifc_sysctl_ntxds[8]; 205 qidx_t ifc_sysctl_nrxds[8]; 206 struct if_txrx ifc_txrx; 207 #define isc_txd_encap ifc_txrx.ift_txd_encap 208 #define isc_txd_flush ifc_txrx.ift_txd_flush 209 #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update 210 #define isc_rxd_available ifc_txrx.ift_rxd_available 211 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get 212 #define isc_rxd_refill ifc_txrx.ift_rxd_refill 213 #define isc_rxd_flush ifc_txrx.ift_rxd_flush 214 #define isc_legacy_intr ifc_txrx.ift_legacy_intr 215 #define isc_txq_select ifc_txrx.ift_txq_select 216 #define isc_txq_select_v2 ifc_txrx.ift_txq_select_v2 217 218 eventhandler_tag ifc_vlan_attach_event; 219 eventhandler_tag ifc_vlan_detach_event; 220 struct ether_addr ifc_mac; 221 }; 222 223 void * 224 iflib_get_softc(if_ctx_t ctx) 225 { 226 227 return (ctx->ifc_softc); 228 } 229 230 device_t 231 iflib_get_dev(if_ctx_t ctx) 232 { 233 234 return (ctx->ifc_dev); 235 } 236 237 if_t 238 iflib_get_ifp(if_ctx_t ctx) 239 { 240 241 return (ctx->ifc_ifp); 242 } 243 244 struct ifmedia * 245 iflib_get_media(if_ctx_t ctx) 246 { 247 248 return (ctx->ifc_mediap); 249 } 250 251 void 252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN]) 253 { 254 255 bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN); 256 } 257 258 if_softc_ctx_t 259 iflib_get_softc_ctx(if_ctx_t ctx) 260 { 261 262 return (&ctx->ifc_softc_ctx); 263 } 264 265 if_shared_ctx_t 266 iflib_get_sctx(if_ctx_t ctx) 267 { 268 269 return (ctx->ifc_sctx); 270 } 271 272 uint16_t 273 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx) 274 { 275 276 return (ctx->ifc_sysctl_extra_msix_vectors); 277 } 278 279 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2) 280 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE / sizeof(void *)) 281 #define CACHE_PTR_NEXT(ptr) ((void *)(roundup2(ptr, CACHE_LINE_SIZE))) 282 283 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP) 284 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF) 285 286 typedef struct iflib_sw_rx_desc_array { 287 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ 288 struct mbuf **ifsd_m; /* pkthdr mbufs */ 289 caddr_t *ifsd_cl; /* direct cluster pointer for rx */ 290 bus_addr_t *ifsd_ba; /* bus addr of cluster for rx */ 291 } iflib_rxsd_array_t; 292 293 typedef struct iflib_sw_tx_desc_array { 294 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ 295 bus_dmamap_t *ifsd_tso_map; /* bus_dma maps for TSO packet */ 296 struct mbuf **ifsd_m; /* pkthdr mbufs */ 297 } if_txsd_vec_t; 298 299 /* magic number that should be high enough for any hardware */ 300 #define IFLIB_MAX_TX_SEGS 128 301 #define IFLIB_RX_COPY_THRESH 128 302 #define IFLIB_MAX_RX_REFRESH 32 303 /* The minimum descriptors per second before we start coalescing */ 304 #define IFLIB_MIN_DESC_SEC 16384 305 #define IFLIB_DEFAULT_TX_UPDATE_FREQ 16 306 #define IFLIB_QUEUE_IDLE 0 307 #define IFLIB_QUEUE_HUNG 1 308 #define IFLIB_QUEUE_WORKING 2 309 /* maximum number of txqs that can share an rx interrupt */ 310 #define IFLIB_MAX_TX_SHARED_INTR 4 311 312 /* this should really scale with ring size - this is a fairly arbitrary value */ 313 #define TX_BATCH_SIZE 32 314 315 #define IFLIB_RESTART_BUDGET 8 316 317 #define IFC_LEGACY 0x001 318 #define IFC_QFLUSH 0x002 319 #define IFC_MULTISEG 0x004 320 #define IFC_SPARE1 0x008 321 #define IFC_SC_ALLOCATED 0x010 322 #define IFC_INIT_DONE 0x020 323 #define IFC_PREFETCH 0x040 324 #define IFC_DO_RESET 0x080 325 #define IFC_DO_WATCHDOG 0x100 326 #define IFC_SPARE0 0x200 327 #define IFC_SPARE2 0x400 328 #define IFC_IN_DETACH 0x800 329 330 #define IFC_NETMAP_TX_IRQ 0x80000000 331 332 #define CSUM_OFFLOAD (CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \ 333 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \ 334 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP) 335 336 struct iflib_txq { 337 qidx_t ift_in_use; 338 qidx_t ift_cidx; 339 qidx_t ift_cidx_processed; 340 qidx_t ift_pidx; 341 uint8_t ift_gen; 342 uint8_t ift_br_offset; 343 uint16_t ift_npending; 344 uint16_t ift_db_pending; 345 uint16_t ift_rs_pending; 346 /* implicit pad */ 347 uint8_t ift_txd_size[8]; 348 uint64_t ift_processed; 349 uint64_t ift_cleaned; 350 uint64_t ift_cleaned_prev; 351 #if MEMORY_LOGGING 352 uint64_t ift_enqueued; 353 uint64_t ift_dequeued; 354 #endif 355 uint64_t ift_no_tx_dma_setup; 356 uint64_t ift_no_desc_avail; 357 uint64_t ift_mbuf_defrag_failed; 358 uint64_t ift_mbuf_defrag; 359 uint64_t ift_map_failed; 360 uint64_t ift_txd_encap_efbig; 361 uint64_t ift_pullups; 362 uint64_t ift_last_timer_tick; 363 364 struct mtx ift_mtx; 365 struct mtx ift_db_mtx; 366 367 /* constant values */ 368 if_ctx_t ift_ctx; 369 struct ifmp_ring *ift_br; 370 struct grouptask ift_task; 371 qidx_t ift_size; 372 uint16_t ift_id; 373 struct callout ift_timer; 374 #ifdef DEV_NETMAP 375 struct callout ift_netmap_timer; 376 #endif /* DEV_NETMAP */ 377 378 if_txsd_vec_t ift_sds; 379 uint8_t ift_qstatus; 380 uint8_t ift_closed; 381 uint8_t ift_update_freq; 382 struct iflib_filter_info ift_filter_info; 383 bus_dma_tag_t ift_buf_tag; 384 bus_dma_tag_t ift_tso_buf_tag; 385 iflib_dma_info_t ift_ifdi; 386 #define MTX_NAME_LEN 32 387 char ift_mtx_name[MTX_NAME_LEN]; 388 bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE); 389 #ifdef IFLIB_DIAGNOSTICS 390 uint64_t ift_cpu_exec_count[256]; 391 #endif 392 } __aligned(CACHE_LINE_SIZE); 393 394 struct iflib_fl { 395 qidx_t ifl_cidx; 396 qidx_t ifl_pidx; 397 qidx_t ifl_credits; 398 uint8_t ifl_gen; 399 uint8_t ifl_rxd_size; 400 #if MEMORY_LOGGING 401 uint64_t ifl_m_enqueued; 402 uint64_t ifl_m_dequeued; 403 uint64_t ifl_cl_enqueued; 404 uint64_t ifl_cl_dequeued; 405 #endif 406 /* implicit pad */ 407 bitstr_t *ifl_rx_bitmap; 408 qidx_t ifl_fragidx; 409 /* constant */ 410 qidx_t ifl_size; 411 uint16_t ifl_buf_size; 412 uint16_t ifl_cltype; 413 uma_zone_t ifl_zone; 414 iflib_rxsd_array_t ifl_sds; 415 iflib_rxq_t ifl_rxq; 416 uint8_t ifl_id; 417 bus_dma_tag_t ifl_buf_tag; 418 iflib_dma_info_t ifl_ifdi; 419 uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE); 420 qidx_t ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH]; 421 } __aligned(CACHE_LINE_SIZE); 422 423 static inline qidx_t 424 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen) 425 { 426 qidx_t used; 427 428 if (pidx > cidx) 429 used = pidx - cidx; 430 else if (pidx < cidx) 431 used = size - cidx + pidx; 432 else if (gen == 0 && pidx == cidx) 433 used = 0; 434 else if (gen == 1 && pidx == cidx) 435 used = size; 436 else 437 panic("bad state"); 438 439 return (used); 440 } 441 442 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen)) 443 444 #define IDXDIFF(head, tail, wrap) \ 445 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head)) 446 447 struct iflib_rxq { 448 if_ctx_t ifr_ctx; 449 iflib_fl_t ifr_fl; 450 uint64_t ifr_rx_irq; 451 struct pfil_head *pfil; 452 /* 453 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is 454 * the completion queue consumer index. Otherwise it's unused. 455 */ 456 qidx_t ifr_cq_cidx; 457 uint16_t ifr_id; 458 uint8_t ifr_nfl; 459 uint8_t ifr_ntxqirq; 460 uint8_t ifr_txqid[IFLIB_MAX_TX_SHARED_INTR]; 461 uint8_t ifr_fl_offset; 462 struct lro_ctrl ifr_lc; 463 struct grouptask ifr_task; 464 struct callout ifr_watchdog; 465 struct iflib_filter_info ifr_filter_info; 466 iflib_dma_info_t ifr_ifdi; 467 468 /* dynamically allocate if any drivers need a value substantially larger than this */ 469 struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE); 470 #ifdef IFLIB_DIAGNOSTICS 471 uint64_t ifr_cpu_exec_count[256]; 472 #endif 473 } __aligned(CACHE_LINE_SIZE); 474 475 typedef struct if_rxsd { 476 caddr_t *ifsd_cl; 477 iflib_fl_t ifsd_fl; 478 } *if_rxsd_t; 479 480 /* multiple of word size */ 481 #ifdef __LP64__ 482 #define PKT_INFO_SIZE 6 483 #define RXD_INFO_SIZE 5 484 #define PKT_TYPE uint64_t 485 #else 486 #define PKT_INFO_SIZE 11 487 #define RXD_INFO_SIZE 8 488 #define PKT_TYPE uint32_t 489 #endif 490 #define PKT_LOOP_BOUND ((PKT_INFO_SIZE / 3) * 3) 491 #define RXD_LOOP_BOUND ((RXD_INFO_SIZE / 4) * 4) 492 493 typedef struct if_pkt_info_pad { 494 PKT_TYPE pkt_val[PKT_INFO_SIZE]; 495 } *if_pkt_info_pad_t; 496 typedef struct if_rxd_info_pad { 497 PKT_TYPE rxd_val[RXD_INFO_SIZE]; 498 } *if_rxd_info_pad_t; 499 500 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info)); 501 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info)); 502 503 static inline void 504 pkt_info_zero(if_pkt_info_t pi) 505 { 506 if_pkt_info_pad_t pi_pad; 507 508 pi_pad = (if_pkt_info_pad_t)pi; 509 pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0; 510 pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0; 511 #ifndef __LP64__ 512 pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0; 513 pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0; 514 #endif 515 } 516 517 static inline void 518 rxd_info_zero(if_rxd_info_t ri) 519 { 520 if_rxd_info_pad_t ri_pad; 521 int i; 522 523 ri_pad = (if_rxd_info_pad_t)ri; 524 for (i = 0; i < RXD_LOOP_BOUND; i += 4) { 525 ri_pad->rxd_val[i] = 0; 526 ri_pad->rxd_val[i + 1] = 0; 527 ri_pad->rxd_val[i + 2] = 0; 528 ri_pad->rxd_val[i + 3] = 0; 529 } 530 #ifdef __LP64__ 531 ri_pad->rxd_val[RXD_INFO_SIZE - 1] = 0; 532 #endif 533 } 534 535 /* 536 * Only allow a single packet to take up most 1/nth of the tx ring 537 */ 538 #define MAX_SINGLE_PACKET_FRACTION 12 539 #define IF_BAD_DMA ((bus_addr_t)-1) 540 541 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING)) 542 543 #define CTX_LOCK_INIT(_sc) sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock") 544 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx) 545 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx) 546 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx) 547 548 #define STATE_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF) 549 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx) 550 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx) 551 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx) 552 553 #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx) 554 #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx) 555 556 /* Our boot-time initialization hook */ 557 static int iflib_module_event_handler(module_t, int, void *); 558 559 static moduledata_t iflib_moduledata = { 560 "iflib", 561 iflib_module_event_handler, 562 NULL 563 }; 564 565 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY); 566 MODULE_VERSION(iflib, 1); 567 568 MODULE_DEPEND(iflib, pci, 1, 1, 1); 569 MODULE_DEPEND(iflib, ether, 1, 1, 1); 570 571 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1); 572 TASKQGROUP_DEFINE(if_config_tqg, 1, 1); 573 574 #ifndef IFLIB_DEBUG_COUNTERS 575 #ifdef INVARIANTS 576 #define IFLIB_DEBUG_COUNTERS 1 577 #else 578 #define IFLIB_DEBUG_COUNTERS 0 579 #endif /* !INVARIANTS */ 580 #endif 581 582 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 583 "iflib driver parameters"); 584 585 /* 586 * XXX need to ensure that this can't accidentally cause the head to be moved backwards 587 */ 588 static int iflib_min_tx_latency = 0; 589 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW, 590 &iflib_min_tx_latency, 0, 591 "minimize transmit latency at the possible expense of throughput"); 592 static int iflib_no_tx_batch = 0; 593 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW, 594 &iflib_no_tx_batch, 0, 595 "minimize transmit latency at the possible expense of throughput"); 596 static int iflib_timer_default = 1000; 597 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW, 598 &iflib_timer_default, 0, "number of ticks between iflib_timer calls"); 599 600 601 #if IFLIB_DEBUG_COUNTERS 602 603 static int iflib_tx_seen; 604 static int iflib_tx_sent; 605 static int iflib_tx_encap; 606 static int iflib_rx_allocs; 607 static int iflib_fl_refills; 608 static int iflib_fl_refills_large; 609 static int iflib_tx_frees; 610 611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0, 612 "# TX mbufs seen"); 613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0, 614 "# TX mbufs sent"); 615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0, 616 "# TX mbufs encapped"); 617 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0, 618 "# TX frees"); 619 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0, 620 "# RX allocations"); 621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0, 622 "# refills"); 623 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD, 624 &iflib_fl_refills_large, 0, "# large refills"); 625 626 static int iflib_txq_drain_flushing; 627 static int iflib_txq_drain_oactive; 628 static int iflib_txq_drain_notready; 629 630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD, 631 &iflib_txq_drain_flushing, 0, "# drain flushes"); 632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD, 633 &iflib_txq_drain_oactive, 0, "# drain oactives"); 634 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD, 635 &iflib_txq_drain_notready, 0, "# drain notready"); 636 637 static int iflib_encap_load_mbuf_fail; 638 static int iflib_encap_pad_mbuf_fail; 639 static int iflib_encap_txq_avail_fail; 640 static int iflib_encap_txd_encap_fail; 641 642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD, 643 &iflib_encap_load_mbuf_fail, 0, "# busdma load failures"); 644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD, 645 &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures"); 646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD, 647 &iflib_encap_txq_avail_fail, 0, "# txq avail failures"); 648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD, 649 &iflib_encap_txd_encap_fail, 0, "# driver encap failures"); 650 651 static int iflib_task_fn_rxs; 652 static int iflib_rx_intr_enables; 653 static int iflib_fast_intrs; 654 static int iflib_rx_unavail; 655 static int iflib_rx_ctx_inactive; 656 static int iflib_rx_if_input; 657 static int iflib_rxd_flush; 658 659 static int iflib_verbose_debug; 660 661 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0, 662 "# task_fn_rx calls"); 663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD, 664 &iflib_rx_intr_enables, 0, "# RX intr enables"); 665 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0, 666 "# fast_intr calls"); 667 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0, 668 "# times rxeof called with no available data"); 669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD, 670 &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context"); 671 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input, 672 0, "# times rxeof called if_input"); 673 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0, 674 "# times rxd_flush called"); 675 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW, 676 &iflib_verbose_debug, 0, "enable verbose debugging"); 677 678 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1) 679 static void 680 iflib_debug_reset(void) 681 { 682 iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs = 683 iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees = 684 iflib_txq_drain_flushing = iflib_txq_drain_oactive = 685 iflib_txq_drain_notready = 686 iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail = 687 iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail = 688 iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs = 689 iflib_rx_unavail = 690 iflib_rx_ctx_inactive = iflib_rx_if_input = 691 iflib_rxd_flush = 0; 692 } 693 694 #else 695 #define DBG_COUNTER_INC(name) 696 static void iflib_debug_reset(void) {} 697 #endif 698 699 #define IFLIB_DEBUG 0 700 701 static void iflib_tx_structures_free(if_ctx_t ctx); 702 static void iflib_rx_structures_free(if_ctx_t ctx); 703 static int iflib_queues_alloc(if_ctx_t ctx); 704 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); 705 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget); 706 static int iflib_qset_structures_setup(if_ctx_t ctx); 707 static int iflib_msix_init(if_ctx_t ctx); 708 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str); 709 static void iflib_txq_check_drain(iflib_txq_t txq, int budget); 710 static uint32_t iflib_txq_can_drain(struct ifmp_ring *); 711 #ifdef ALTQ 712 static void iflib_altq_if_start(if_t ifp); 713 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m); 714 #endif 715 static void iflib_register(if_ctx_t); 716 static void iflib_deregister(if_ctx_t); 717 static void iflib_unregister_vlan_handlers(if_ctx_t ctx); 718 static uint16_t iflib_get_mbuf_size_for(unsigned int size); 719 static void iflib_init_locked(if_ctx_t ctx); 720 static void iflib_add_device_sysctl_pre(if_ctx_t ctx); 721 static void iflib_add_device_sysctl_post(if_ctx_t ctx); 722 static void iflib_ifmp_purge(iflib_txq_t txq); 723 static void _iflib_pre_assert(if_softc_ctx_t scctx); 724 static void iflib_stop(if_ctx_t ctx); 725 static void iflib_if_init_locked(if_ctx_t ctx); 726 static void iflib_free_intr_mem(if_ctx_t ctx); 727 #ifndef __NO_STRICT_ALIGNMENT 728 static struct mbuf *iflib_fixup_rx(struct mbuf *m); 729 #endif 730 static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh); 731 732 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets = 733 SLIST_HEAD_INITIALIZER(cpu_offsets); 734 struct cpu_offset { 735 SLIST_ENTRY(cpu_offset) entries; 736 cpuset_t set; 737 unsigned int refcount; 738 uint16_t next_cpuid; 739 }; 740 static struct mtx cpu_offset_mtx; 741 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock", 742 MTX_DEF); 743 744 DEBUGNET_DEFINE(iflib); 745 746 static int 747 iflib_num_rx_descs(if_ctx_t ctx) 748 { 749 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 750 if_shared_ctx_t sctx = ctx->ifc_sctx; 751 uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0; 752 753 return (scctx->isc_nrxd[first_rxq]); 754 } 755 756 static int 757 iflib_num_tx_descs(if_ctx_t ctx) 758 { 759 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 760 if_shared_ctx_t sctx = ctx->ifc_sctx; 761 uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0; 762 763 return (scctx->isc_ntxd[first_txq]); 764 } 765 766 #ifdef DEV_NETMAP 767 #include <sys/selinfo.h> 768 #include <net/netmap.h> 769 #include <dev/netmap/netmap_kern.h> 770 771 MODULE_DEPEND(iflib, netmap, 1, 1, 1); 772 773 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init); 774 static void iflib_netmap_timer(void *arg); 775 776 /* 777 * device-specific sysctl variables: 778 * 779 * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it. 780 * During regular operations the CRC is stripped, but on some 781 * hardware reception of frames not multiple of 64 is slower, 782 * so using crcstrip=0 helps in benchmarks. 783 * 784 * iflib_rx_miss, iflib_rx_miss_bufs: 785 * count packets that might be missed due to lost interrupts. 786 */ 787 SYSCTL_DECL(_dev_netmap); 788 /* 789 * The xl driver by default strips CRCs and we do not override it. 790 */ 791 792 int iflib_crcstrip = 1; 793 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip, 794 CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames"); 795 796 int iflib_rx_miss, iflib_rx_miss_bufs; 797 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss, 798 CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr"); 799 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs, 800 CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs"); 801 802 /* 803 * Register/unregister. We are already under netmap lock. 804 * Only called on the first register or the last unregister. 805 */ 806 static int 807 iflib_netmap_register(struct netmap_adapter *na, int onoff) 808 { 809 if_t ifp = na->ifp; 810 if_ctx_t ctx = if_getsoftc(ifp); 811 int status; 812 813 CTX_LOCK(ctx); 814 if (!CTX_IS_VF(ctx)) 815 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); 816 817 iflib_stop(ctx); 818 819 /* 820 * Enable (or disable) netmap flags, and intercept (or restore) 821 * ifp->if_transmit. This is done once the device has been stopped 822 * to prevent race conditions. Also, this must be done after 823 * calling netmap_disable_all_rings() and before calling 824 * netmap_enable_all_rings(), so that these two functions see the 825 * updated state of the NAF_NETMAP_ON bit. 826 */ 827 if (onoff) { 828 nm_set_native_flags(na); 829 } else { 830 nm_clear_native_flags(na); 831 } 832 833 iflib_init_locked(ctx); 834 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ? 835 status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1; 836 if (status) 837 nm_clear_native_flags(na); 838 CTX_UNLOCK(ctx); 839 return (status); 840 } 841 842 static int 843 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info) 844 { 845 if_t ifp = na->ifp; 846 if_ctx_t ctx = if_getsoftc(ifp); 847 iflib_rxq_t rxq = &ctx->ifc_rxqs[0]; 848 iflib_fl_t fl = &rxq->ifr_fl[0]; 849 850 info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; 851 info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; 852 info->num_tx_descs = iflib_num_tx_descs(ctx); 853 info->num_rx_descs = iflib_num_rx_descs(ctx); 854 info->rx_buf_maxsize = fl->ifl_buf_size; 855 nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u", 856 info->num_tx_rings, info->num_rx_rings, info->num_tx_descs, 857 info->num_rx_descs, info->rx_buf_maxsize); 858 859 return (0); 860 } 861 862 static int 863 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init) 864 { 865 struct netmap_adapter *na = kring->na; 866 u_int const lim = kring->nkr_num_slots - 1; 867 struct netmap_ring *ring = kring->ring; 868 bus_dmamap_t *map; 869 struct if_rxd_update iru; 870 if_ctx_t ctx = rxq->ifr_ctx; 871 iflib_fl_t fl = &rxq->ifr_fl[0]; 872 u_int nic_i_first, nic_i; 873 u_int nm_i; 874 int i, n; 875 #if IFLIB_DEBUG_COUNTERS 876 int rf_count = 0; 877 #endif 878 879 /* 880 * This function is used both at initialization and in rxsync. 881 * At initialization we need to prepare (with isc_rxd_refill()) 882 * all the netmap buffers currently owned by the kernel, in 883 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync 884 * (except for kring->nkr_hwofs). These may be less than 885 * kring->nkr_num_slots if netmap_reset() was called while 886 * an application using the kring that still owned some 887 * buffers. 888 * At rxsync time, both indexes point to the next buffer to be 889 * refilled. 890 * In any case we publish (with isc_rxd_flush()) up to 891 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod 892 * pointer to overrun the head/cons pointer, although this is 893 * not necessary for some NICs (e.g. vmx). 894 */ 895 if (__predict_false(init)) { 896 n = kring->nkr_num_slots - nm_kr_rxspace(kring); 897 } else { 898 n = kring->rhead - kring->nr_hwcur; 899 if (n == 0) 900 return (0); /* Nothing to do. */ 901 if (n < 0) 902 n += kring->nkr_num_slots; 903 } 904 905 iru_init(&iru, rxq, 0 /* flid */); 906 map = fl->ifl_sds.ifsd_map; 907 nic_i = fl->ifl_pidx; 908 nm_i = netmap_idx_n2k(kring, nic_i); 909 if (__predict_false(init)) { 910 /* 911 * On init/reset, nic_i must be 0, and we must 912 * start to refill from hwtail (see netmap_reset()). 913 */ 914 MPASS(nic_i == 0); 915 MPASS(nm_i == kring->nr_hwtail); 916 } else 917 MPASS(nm_i == kring->nr_hwcur); 918 DBG_COUNTER_INC(fl_refills); 919 while (n > 0) { 920 #if IFLIB_DEBUG_COUNTERS 921 if (++rf_count == 9) 922 DBG_COUNTER_INC(fl_refills_large); 923 #endif 924 nic_i_first = nic_i; 925 for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) { 926 struct netmap_slot *slot = &ring->slot[nm_i]; 927 uint64_t paddr; 928 void *addr = PNMB(na, slot, &paddr); 929 930 MPASS(i < IFLIB_MAX_RX_REFRESH); 931 932 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */ 933 return (netmap_ring_reinit(kring)); 934 935 fl->ifl_bus_addrs[i] = paddr + 936 nm_get_offset(kring, slot); 937 fl->ifl_rxd_idxs[i] = nic_i; 938 939 if (__predict_false(init)) { 940 netmap_load_map(na, fl->ifl_buf_tag, 941 map[nic_i], addr); 942 } else if (slot->flags & NS_BUF_CHANGED) { 943 /* buffer has changed, reload map */ 944 netmap_reload_map(na, fl->ifl_buf_tag, 945 map[nic_i], addr); 946 } 947 bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i], 948 BUS_DMASYNC_PREREAD); 949 slot->flags &= ~NS_BUF_CHANGED; 950 951 nm_i = nm_next(nm_i, lim); 952 nic_i = nm_next(nic_i, lim); 953 } 954 955 iru.iru_pidx = nic_i_first; 956 iru.iru_count = i; 957 ctx->isc_rxd_refill(ctx->ifc_softc, &iru); 958 } 959 fl->ifl_pidx = nic_i; 960 /* 961 * At the end of the loop we must have refilled everything 962 * we could possibly refill. 963 */ 964 MPASS(nm_i == kring->rhead); 965 kring->nr_hwcur = nm_i; 966 967 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 968 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 969 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, 970 nm_prev(nic_i, lim)); 971 DBG_COUNTER_INC(rxd_flush); 972 973 return (0); 974 } 975 976 #define NETMAP_TX_TIMER_US 90 977 978 /* 979 * Reconcile kernel and user view of the transmit ring. 980 * 981 * All information is in the kring. 982 * Userspace wants to send packets up to the one before kring->rhead, 983 * kernel knows kring->nr_hwcur is the first unsent packet. 984 * 985 * Here we push packets out (as many as possible), and possibly 986 * reclaim buffers from previously completed transmission. 987 * 988 * The caller (netmap) guarantees that there is only one instance 989 * running at any time. Any interference with other driver 990 * methods should be handled by the individual drivers. 991 */ 992 static int 993 iflib_netmap_txsync(struct netmap_kring *kring, int flags) 994 { 995 struct netmap_adapter *na = kring->na; 996 if_t ifp = na->ifp; 997 struct netmap_ring *ring = kring->ring; 998 u_int nm_i; /* index into the netmap kring */ 999 u_int nic_i; /* index into the NIC ring */ 1000 u_int const lim = kring->nkr_num_slots - 1; 1001 u_int const head = kring->rhead; 1002 struct if_pkt_info pi; 1003 int tx_pkts = 0, tx_bytes = 0; 1004 1005 /* 1006 * interrupts on every tx packet are expensive so request 1007 * them every half ring, or where NS_REPORT is set 1008 */ 1009 u_int report_frequency = kring->nkr_num_slots >> 1; 1010 /* device-specific */ 1011 if_ctx_t ctx = if_getsoftc(ifp); 1012 iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id]; 1013 1014 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 1015 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 1016 1017 /* 1018 * First part: process new packets to send. 1019 * nm_i is the current index in the netmap kring, 1020 * nic_i is the corresponding index in the NIC ring. 1021 * 1022 * If we have packets to send (nm_i != head) 1023 * iterate over the netmap ring, fetch length and update 1024 * the corresponding slot in the NIC ring. Some drivers also 1025 * need to update the buffer's physical address in the NIC slot 1026 * even NS_BUF_CHANGED is not set (PNMB computes the addresses). 1027 * 1028 * The netmap_reload_map() calls is especially expensive, 1029 * even when (as in this case) the tag is 0, so do only 1030 * when the buffer has actually changed. 1031 * 1032 * If possible do not set the report/intr bit on all slots, 1033 * but only a few times per ring or when NS_REPORT is set. 1034 * 1035 * Finally, on 10G and faster drivers, it might be useful 1036 * to prefetch the next slot and txr entry. 1037 */ 1038 1039 nm_i = kring->nr_hwcur; 1040 if (nm_i != head) { /* we have new packets to send */ 1041 uint32_t pkt_len = 0, seg_idx = 0; 1042 int nic_i_start = -1, flags = 0; 1043 pkt_info_zero(&pi); 1044 pi.ipi_segs = txq->ift_segs; 1045 pi.ipi_qsidx = kring->ring_id; 1046 nic_i = netmap_idx_k2n(kring, nm_i); 1047 1048 __builtin_prefetch(&ring->slot[nm_i]); 1049 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]); 1050 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]); 1051 1052 while (nm_i != head) { 1053 struct netmap_slot *slot = &ring->slot[nm_i]; 1054 uint64_t offset = nm_get_offset(kring, slot); 1055 u_int len = slot->len; 1056 uint64_t paddr; 1057 void *addr = PNMB(na, slot, &paddr); 1058 1059 flags |= (slot->flags & NS_REPORT || 1060 nic_i == 0 || nic_i == report_frequency) ? 1061 IPI_TX_INTR : 0; 1062 1063 /* 1064 * If this is the first packet fragment, save the 1065 * index of the first NIC slot for later. 1066 */ 1067 if (nic_i_start < 0) 1068 nic_i_start = nic_i; 1069 1070 pi.ipi_segs[seg_idx].ds_addr = paddr + offset; 1071 pi.ipi_segs[seg_idx].ds_len = len; 1072 if (len) { 1073 pkt_len += len; 1074 seg_idx++; 1075 } 1076 1077 if (!(slot->flags & NS_MOREFRAG)) { 1078 pi.ipi_len = pkt_len; 1079 pi.ipi_nsegs = seg_idx; 1080 pi.ipi_pidx = nic_i_start; 1081 pi.ipi_ndescs = 0; 1082 pi.ipi_flags = flags; 1083 1084 /* Prepare the NIC TX ring. */ 1085 ctx->isc_txd_encap(ctx->ifc_softc, &pi); 1086 DBG_COUNTER_INC(tx_encap); 1087 1088 /* Update transmit counters */ 1089 tx_bytes += pi.ipi_len; 1090 tx_pkts++; 1091 1092 /* Reinit per-packet info for the next one. */ 1093 flags = seg_idx = pkt_len = 0; 1094 nic_i_start = -1; 1095 } 1096 1097 /* prefetch for next round */ 1098 __builtin_prefetch(&ring->slot[nm_i + 1]); 1099 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]); 1100 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]); 1101 1102 NM_CHECK_ADDR_LEN_OFF(na, len, offset); 1103 1104 if (slot->flags & NS_BUF_CHANGED) { 1105 /* buffer has changed, reload map */ 1106 netmap_reload_map(na, txq->ift_buf_tag, 1107 txq->ift_sds.ifsd_map[nic_i], addr); 1108 } 1109 /* make sure changes to the buffer are synced */ 1110 bus_dmamap_sync(txq->ift_buf_tag, 1111 txq->ift_sds.ifsd_map[nic_i], 1112 BUS_DMASYNC_PREWRITE); 1113 1114 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG); 1115 nm_i = nm_next(nm_i, lim); 1116 nic_i = nm_next(nic_i, lim); 1117 } 1118 kring->nr_hwcur = nm_i; 1119 1120 /* synchronize the NIC ring */ 1121 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 1122 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 1123 1124 /* (re)start the tx unit up to slot nic_i (excluded) */ 1125 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i); 1126 } 1127 1128 /* 1129 * Second part: reclaim buffers for completed transmissions. 1130 * 1131 * If there are unclaimed buffers, attempt to reclaim them. 1132 * If we don't manage to reclaim them all, and TX IRQs are not in use, 1133 * trigger a per-tx-queue timer to try again later. 1134 */ 1135 if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { 1136 if (iflib_tx_credits_update(ctx, txq)) { 1137 /* some tx completed, increment avail */ 1138 nic_i = txq->ift_cidx_processed; 1139 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim); 1140 } 1141 } 1142 1143 if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) 1144 if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { 1145 callout_reset_sbt_on(&txq->ift_netmap_timer, 1146 NETMAP_TX_TIMER_US * SBT_1US, SBT_1US, 1147 iflib_netmap_timer, txq, 1148 txq->ift_netmap_timer.c_cpu, 0); 1149 } 1150 1151 if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes); 1152 if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts); 1153 1154 return (0); 1155 } 1156 1157 /* 1158 * Reconcile kernel and user view of the receive ring. 1159 * Same as for the txsync, this routine must be efficient. 1160 * The caller guarantees a single invocations, but races against 1161 * the rest of the driver should be handled here. 1162 * 1163 * On call, kring->rhead is the first packet that userspace wants 1164 * to keep, and kring->rcur is the wakeup point. 1165 * The kernel has previously reported packets up to kring->rtail. 1166 * 1167 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective 1168 * of whether or not we received an interrupt. 1169 */ 1170 static int 1171 iflib_netmap_rxsync(struct netmap_kring *kring, int flags) 1172 { 1173 struct netmap_adapter *na = kring->na; 1174 struct netmap_ring *ring = kring->ring; 1175 if_t ifp = na->ifp; 1176 uint32_t nm_i; /* index into the netmap ring */ 1177 uint32_t nic_i; /* index into the NIC ring */ 1178 u_int n; 1179 u_int const lim = kring->nkr_num_slots - 1; 1180 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; 1181 int i = 0, rx_bytes = 0, rx_pkts = 0; 1182 1183 if_ctx_t ctx = if_getsoftc(ifp); 1184 if_shared_ctx_t sctx = ctx->ifc_sctx; 1185 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1186 iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id]; 1187 iflib_fl_t fl = &rxq->ifr_fl[0]; 1188 struct if_rxd_info ri; 1189 qidx_t *cidxp; 1190 1191 /* 1192 * netmap only uses free list 0, to avoid out of order consumption 1193 * of receive buffers 1194 */ 1195 1196 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 1197 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 1198 1199 /* 1200 * First part: import newly received packets. 1201 * 1202 * nm_i is the index of the next free slot in the netmap ring, 1203 * nic_i is the index of the next received packet in the NIC ring 1204 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may 1205 * differ in case if_init() has been called while 1206 * in netmap mode. For the receive ring we have 1207 * 1208 * nic_i = fl->ifl_cidx; 1209 * nm_i = kring->nr_hwtail (previous) 1210 * and 1211 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size 1212 * 1213 * fl->ifl_cidx is set to 0 on a ring reinit 1214 */ 1215 if (netmap_no_pendintr || force_update) { 1216 uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim); 1217 bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ; 1218 int crclen = iflib_crcstrip ? 0 : 4; 1219 int error, avail; 1220 1221 /* 1222 * For the free list consumer index, we use the same 1223 * logic as in iflib_rxeof(). 1224 */ 1225 if (have_rxcq) 1226 cidxp = &rxq->ifr_cq_cidx; 1227 else 1228 cidxp = &fl->ifl_cidx; 1229 avail = ctx->isc_rxd_available(ctx->ifc_softc, 1230 rxq->ifr_id, *cidxp, USHRT_MAX); 1231 1232 nic_i = fl->ifl_cidx; 1233 nm_i = netmap_idx_n2k(kring, nic_i); 1234 MPASS(nm_i == kring->nr_hwtail); 1235 for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) { 1236 rxd_info_zero(&ri); 1237 ri.iri_frags = rxq->ifr_frags; 1238 ri.iri_qsidx = kring->ring_id; 1239 ri.iri_ifp = ctx->ifc_ifp; 1240 ri.iri_cidx = *cidxp; 1241 1242 error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); 1243 for (i = 0; i < ri.iri_nfrags; i++) { 1244 if (error) { 1245 ring->slot[nm_i].len = 0; 1246 ring->slot[nm_i].flags = 0; 1247 } else { 1248 ring->slot[nm_i].len = ri.iri_frags[i].irf_len; 1249 if (i == (ri.iri_nfrags - 1)) { 1250 ring->slot[nm_i].len -= crclen; 1251 ring->slot[nm_i].flags = 0; 1252 1253 /* Update receive counters */ 1254 rx_bytes += ri.iri_len; 1255 rx_pkts++; 1256 } else 1257 ring->slot[nm_i].flags = NS_MOREFRAG; 1258 } 1259 1260 bus_dmamap_sync(fl->ifl_buf_tag, 1261 fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD); 1262 nm_i = nm_next(nm_i, lim); 1263 fl->ifl_cidx = nic_i = nm_next(nic_i, lim); 1264 } 1265 1266 if (have_rxcq) { 1267 *cidxp = ri.iri_cidx; 1268 while (*cidxp >= scctx->isc_nrxd[0]) 1269 *cidxp -= scctx->isc_nrxd[0]; 1270 } 1271 1272 } 1273 if (n) { /* update the state variables */ 1274 if (netmap_no_pendintr && !force_update) { 1275 /* diagnostics */ 1276 iflib_rx_miss++; 1277 iflib_rx_miss_bufs += n; 1278 } 1279 kring->nr_hwtail = nm_i; 1280 } 1281 kring->nr_kflags &= ~NKR_PENDINTR; 1282 } 1283 /* 1284 * Second part: skip past packets that userspace has released. 1285 * (kring->nr_hwcur to head excluded), 1286 * and make the buffers available for reception. 1287 * As usual nm_i is the index in the netmap ring, 1288 * nic_i is the index in the NIC ring, and 1289 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size 1290 */ 1291 netmap_fl_refill(rxq, kring, false); 1292 1293 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); 1294 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); 1295 1296 return (0); 1297 } 1298 1299 static void 1300 iflib_netmap_intr(struct netmap_adapter *na, int onoff) 1301 { 1302 if_ctx_t ctx = if_getsoftc(na->ifp); 1303 1304 CTX_LOCK(ctx); 1305 if (onoff) { 1306 IFDI_INTR_ENABLE(ctx); 1307 } else { 1308 IFDI_INTR_DISABLE(ctx); 1309 } 1310 CTX_UNLOCK(ctx); 1311 } 1312 1313 static int 1314 iflib_netmap_attach(if_ctx_t ctx) 1315 { 1316 struct netmap_adapter na; 1317 1318 bzero(&na, sizeof(na)); 1319 1320 na.ifp = ctx->ifc_ifp; 1321 na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS; 1322 MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); 1323 MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); 1324 1325 na.num_tx_desc = iflib_num_tx_descs(ctx); 1326 na.num_rx_desc = iflib_num_rx_descs(ctx); 1327 na.nm_txsync = iflib_netmap_txsync; 1328 na.nm_rxsync = iflib_netmap_rxsync; 1329 na.nm_register = iflib_netmap_register; 1330 na.nm_intr = iflib_netmap_intr; 1331 na.nm_config = iflib_netmap_config; 1332 na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; 1333 na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; 1334 return (netmap_attach(&na)); 1335 } 1336 1337 static int 1338 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq) 1339 { 1340 struct netmap_adapter *na = NA(ctx->ifc_ifp); 1341 struct netmap_slot *slot; 1342 1343 slot = netmap_reset(na, NR_TX, txq->ift_id, 0); 1344 if (slot == NULL) 1345 return (0); 1346 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) { 1347 /* 1348 * In netmap mode, set the map for the packet buffer. 1349 * NOTE: Some drivers (not this one) also need to set 1350 * the physical buffer address in the NIC ring. 1351 * netmap_idx_n2k() maps a nic index, i, into the corresponding 1352 * netmap slot index, si 1353 */ 1354 int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i); 1355 netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i], 1356 NMB(na, slot + si)); 1357 } 1358 return (1); 1359 } 1360 1361 static int 1362 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq) 1363 { 1364 struct netmap_adapter *na = NA(ctx->ifc_ifp); 1365 struct netmap_kring *kring; 1366 struct netmap_slot *slot; 1367 1368 slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0); 1369 if (slot == NULL) 1370 return (0); 1371 kring = na->rx_rings[rxq->ifr_id]; 1372 netmap_fl_refill(rxq, kring, true); 1373 return (1); 1374 } 1375 1376 static void 1377 iflib_netmap_timer(void *arg) 1378 { 1379 iflib_txq_t txq = arg; 1380 if_ctx_t ctx = txq->ift_ctx; 1381 1382 /* 1383 * Wake up the netmap application, to give it a chance to 1384 * call txsync and reclaim more completed TX buffers. 1385 */ 1386 netmap_tx_irq(ctx->ifc_ifp, txq->ift_id); 1387 } 1388 1389 #define iflib_netmap_detach(ifp) netmap_detach(ifp) 1390 1391 #else 1392 #define iflib_netmap_txq_init(ctx, txq) (0) 1393 #define iflib_netmap_rxq_init(ctx, rxq) (0) 1394 #define iflib_netmap_detach(ifp) 1395 #define netmap_enable_all_rings(ifp) 1396 #define netmap_disable_all_rings(ifp) 1397 1398 #define iflib_netmap_attach(ctx) (0) 1399 #define netmap_rx_irq(ifp, qid, budget) (0) 1400 #endif 1401 1402 #if defined(__i386__) || defined(__amd64__) 1403 static __inline void 1404 prefetch(void *x) 1405 { 1406 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); 1407 } 1408 1409 static __inline void 1410 prefetch2cachelines(void *x) 1411 { 1412 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); 1413 #if (CACHE_LINE_SIZE < 128) 1414 __asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long))))); 1415 #endif 1416 } 1417 #else 1418 static __inline void 1419 prefetch(void *x) 1420 { 1421 } 1422 1423 static __inline void 1424 prefetch2cachelines(void *x) 1425 { 1426 } 1427 #endif 1428 1429 static void 1430 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid) 1431 { 1432 iflib_fl_t fl; 1433 1434 fl = &rxq->ifr_fl[flid]; 1435 iru->iru_paddrs = fl->ifl_bus_addrs; 1436 iru->iru_idxs = fl->ifl_rxd_idxs; 1437 iru->iru_qsidx = rxq->ifr_id; 1438 iru->iru_buf_size = fl->ifl_buf_size; 1439 iru->iru_flidx = fl->ifl_id; 1440 } 1441 1442 static void 1443 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err) 1444 { 1445 if (err) 1446 return; 1447 *(bus_addr_t *) arg = segs[0].ds_addr; 1448 } 1449 1450 #define DMA_WIDTH_TO_BUS_LOWADDR(width) \ 1451 (((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ? \ 1452 BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL) 1453 1454 int 1455 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags) 1456 { 1457 int err; 1458 device_t dev = ctx->ifc_dev; 1459 bus_addr_t lowaddr; 1460 1461 lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width); 1462 1463 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1464 align, 0, /* alignment, bounds */ 1465 lowaddr, /* lowaddr */ 1466 BUS_SPACE_MAXADDR, /* highaddr */ 1467 NULL, NULL, /* filter, filterarg */ 1468 size, /* maxsize */ 1469 1, /* nsegments */ 1470 size, /* maxsegsize */ 1471 BUS_DMA_ALLOCNOW, /* flags */ 1472 NULL, /* lockfunc */ 1473 NULL, /* lockarg */ 1474 &dma->idi_tag); 1475 if (err) { 1476 device_printf(dev, 1477 "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n", 1478 __func__, err, size, align); 1479 goto fail_0; 1480 } 1481 1482 err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr, 1483 BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map); 1484 if (err) { 1485 device_printf(dev, 1486 "%s: bus_dmamem_alloc(%ju) failed: %d\n", 1487 __func__, (uintmax_t)size, err); 1488 goto fail_1; 1489 } 1490 1491 dma->idi_paddr = IF_BAD_DMA; 1492 err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr, 1493 size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT); 1494 if (err || dma->idi_paddr == IF_BAD_DMA) { 1495 device_printf(dev, 1496 "%s: bus_dmamap_load failed: %d\n", 1497 __func__, err); 1498 goto fail_2; 1499 } 1500 1501 dma->idi_size = size; 1502 return (0); 1503 1504 fail_2: 1505 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); 1506 fail_1: 1507 bus_dma_tag_destroy(dma->idi_tag); 1508 fail_0: 1509 dma->idi_tag = NULL; 1510 1511 return (err); 1512 } 1513 1514 int 1515 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags) 1516 { 1517 if_shared_ctx_t sctx = ctx->ifc_sctx; 1518 1519 KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized")); 1520 1521 return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags)); 1522 } 1523 1524 int 1525 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count) 1526 { 1527 int i, err; 1528 iflib_dma_info_t *dmaiter; 1529 1530 dmaiter = dmalist; 1531 for (i = 0; i < count; i++, dmaiter++) { 1532 if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0) 1533 break; 1534 } 1535 if (err) 1536 iflib_dma_free_multi(dmalist, i); 1537 return (err); 1538 } 1539 1540 void 1541 iflib_dma_free(iflib_dma_info_t dma) 1542 { 1543 if (dma->idi_tag == NULL) 1544 return; 1545 if (dma->idi_paddr != IF_BAD_DMA) { 1546 bus_dmamap_sync(dma->idi_tag, dma->idi_map, 1547 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 1548 bus_dmamap_unload(dma->idi_tag, dma->idi_map); 1549 dma->idi_paddr = IF_BAD_DMA; 1550 } 1551 if (dma->idi_vaddr != NULL) { 1552 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); 1553 dma->idi_vaddr = NULL; 1554 } 1555 bus_dma_tag_destroy(dma->idi_tag); 1556 dma->idi_tag = NULL; 1557 } 1558 1559 void 1560 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count) 1561 { 1562 int i; 1563 iflib_dma_info_t *dmaiter = dmalist; 1564 1565 for (i = 0; i < count; i++, dmaiter++) 1566 iflib_dma_free(*dmaiter); 1567 } 1568 1569 static int 1570 iflib_fast_intr(void *arg) 1571 { 1572 iflib_filter_info_t info = arg; 1573 struct grouptask *gtask = info->ifi_task; 1574 int result; 1575 1576 DBG_COUNTER_INC(fast_intrs); 1577 if (info->ifi_filter != NULL) { 1578 result = info->ifi_filter(info->ifi_filter_arg); 1579 if ((result & FILTER_SCHEDULE_THREAD) == 0) 1580 return (result); 1581 } 1582 1583 GROUPTASK_ENQUEUE(gtask); 1584 return (FILTER_HANDLED); 1585 } 1586 1587 static int 1588 iflib_fast_intr_rxtx(void *arg) 1589 { 1590 iflib_filter_info_t info = arg; 1591 struct grouptask *gtask = info->ifi_task; 1592 if_ctx_t ctx; 1593 iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx; 1594 iflib_txq_t txq; 1595 void *sc; 1596 int i, cidx, result; 1597 qidx_t txqid; 1598 bool intr_enable, intr_legacy; 1599 1600 DBG_COUNTER_INC(fast_intrs); 1601 if (info->ifi_filter != NULL) { 1602 result = info->ifi_filter(info->ifi_filter_arg); 1603 if ((result & FILTER_SCHEDULE_THREAD) == 0) 1604 return (result); 1605 } 1606 1607 ctx = rxq->ifr_ctx; 1608 sc = ctx->ifc_softc; 1609 intr_enable = false; 1610 intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY); 1611 MPASS(rxq->ifr_ntxqirq); 1612 for (i = 0; i < rxq->ifr_ntxqirq; i++) { 1613 txqid = rxq->ifr_txqid[i]; 1614 txq = &ctx->ifc_txqs[txqid]; 1615 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 1616 BUS_DMASYNC_POSTREAD); 1617 if (!ctx->isc_txd_credits_update(sc, txqid, false)) { 1618 if (intr_legacy) 1619 intr_enable = true; 1620 else 1621 IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid); 1622 continue; 1623 } 1624 GROUPTASK_ENQUEUE(&txq->ift_task); 1625 } 1626 if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ) 1627 cidx = rxq->ifr_cq_cidx; 1628 else 1629 cidx = rxq->ifr_fl[0].ifl_cidx; 1630 if (iflib_rxd_avail(ctx, rxq, cidx, 1)) 1631 GROUPTASK_ENQUEUE(gtask); 1632 else { 1633 if (intr_legacy) 1634 intr_enable = true; 1635 else 1636 IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); 1637 DBG_COUNTER_INC(rx_intr_enables); 1638 } 1639 if (intr_enable) 1640 IFDI_INTR_ENABLE(ctx); 1641 return (FILTER_HANDLED); 1642 } 1643 1644 static int 1645 iflib_fast_intr_ctx(void *arg) 1646 { 1647 iflib_filter_info_t info = arg; 1648 if_ctx_t ctx = info->ifi_ctx; 1649 int result; 1650 1651 DBG_COUNTER_INC(fast_intrs); 1652 if (info->ifi_filter != NULL) { 1653 result = info->ifi_filter(info->ifi_filter_arg); 1654 if ((result & FILTER_SCHEDULE_THREAD) == 0) 1655 return (result); 1656 } 1657 1658 taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task); 1659 return (FILTER_HANDLED); 1660 } 1661 1662 static int 1663 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, 1664 driver_filter_t filter, driver_intr_t handler, void *arg, 1665 const char *name) 1666 { 1667 struct resource *res; 1668 void *tag = NULL; 1669 device_t dev = ctx->ifc_dev; 1670 int flags, i, rc; 1671 1672 flags = RF_ACTIVE; 1673 if (ctx->ifc_flags & IFC_LEGACY) 1674 flags |= RF_SHAREABLE; 1675 MPASS(rid < 512); 1676 i = rid; 1677 res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags); 1678 if (res == NULL) { 1679 device_printf(dev, 1680 "failed to allocate IRQ for rid %d, name %s.\n", rid, name); 1681 return (ENOMEM); 1682 } 1683 irq->ii_res = res; 1684 KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL")); 1685 rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET, 1686 filter, handler, arg, &tag); 1687 if (rc != 0) { 1688 device_printf(dev, 1689 "failed to setup interrupt for rid %d, name %s: %d\n", 1690 rid, name ? name : "unknown", rc); 1691 return (rc); 1692 } else if (name) 1693 bus_describe_intr(dev, res, tag, "%s", name); 1694 1695 irq->ii_tag = tag; 1696 return (0); 1697 } 1698 1699 /********************************************************************* 1700 * 1701 * Allocate DMA resources for TX buffers as well as memory for the TX 1702 * mbuf map. TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a 1703 * iflib_sw_tx_desc_array structure, storing all the information that 1704 * is needed to transmit a packet on the wire. This is called only 1705 * once at attach, setup is done every reset. 1706 * 1707 **********************************************************************/ 1708 static int 1709 iflib_txsd_alloc(iflib_txq_t txq) 1710 { 1711 if_ctx_t ctx = txq->ift_ctx; 1712 if_shared_ctx_t sctx = ctx->ifc_sctx; 1713 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1714 device_t dev = ctx->ifc_dev; 1715 bus_size_t tsomaxsize; 1716 bus_addr_t lowaddr; 1717 int err, nsegments, ntsosegments; 1718 bool tso; 1719 1720 nsegments = scctx->isc_tx_nsegments; 1721 ntsosegments = scctx->isc_tx_tso_segments_max; 1722 tsomaxsize = scctx->isc_tx_tso_size_max; 1723 if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU) 1724 tsomaxsize += sizeof(struct ether_vlan_header); 1725 MPASS(scctx->isc_ntxd[0] > 0); 1726 MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0); 1727 MPASS(nsegments > 0); 1728 if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) { 1729 MPASS(ntsosegments > 0); 1730 MPASS(sctx->isc_tso_maxsize >= tsomaxsize); 1731 } 1732 1733 lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width); 1734 1735 /* 1736 * Set up DMA tags for TX buffers. 1737 */ 1738 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1739 1, 0, /* alignment, bounds */ 1740 lowaddr, /* lowaddr */ 1741 BUS_SPACE_MAXADDR, /* highaddr */ 1742 NULL, NULL, /* filter, filterarg */ 1743 sctx->isc_tx_maxsize, /* maxsize */ 1744 nsegments, /* nsegments */ 1745 sctx->isc_tx_maxsegsize, /* maxsegsize */ 1746 0, /* flags */ 1747 NULL, /* lockfunc */ 1748 NULL, /* lockfuncarg */ 1749 &txq->ift_buf_tag))) { 1750 device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err); 1751 device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n", 1752 (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize); 1753 goto fail; 1754 } 1755 tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0; 1756 if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev), 1757 1, 0, /* alignment, bounds */ 1758 lowaddr, /* lowaddr */ 1759 BUS_SPACE_MAXADDR, /* highaddr */ 1760 NULL, NULL, /* filter, filterarg */ 1761 tsomaxsize, /* maxsize */ 1762 ntsosegments, /* nsegments */ 1763 sctx->isc_tso_maxsegsize, /* maxsegsize */ 1764 0, /* flags */ 1765 NULL, /* lockfunc */ 1766 NULL, /* lockfuncarg */ 1767 &txq->ift_tso_buf_tag))) { 1768 device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n", 1769 err); 1770 goto fail; 1771 } 1772 1773 /* Allocate memory for the TX mbuf map. */ 1774 if (!(txq->ift_sds.ifsd_m = 1775 (struct mbuf **) malloc(sizeof(struct mbuf *) * 1776 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1777 device_printf(dev, "Unable to allocate TX mbuf map memory\n"); 1778 err = ENOMEM; 1779 goto fail; 1780 } 1781 1782 /* 1783 * Create the DMA maps for TX buffers. 1784 */ 1785 if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc( 1786 sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], 1787 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { 1788 device_printf(dev, 1789 "Unable to allocate TX buffer DMA map memory\n"); 1790 err = ENOMEM; 1791 goto fail; 1792 } 1793 if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc( 1794 sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], 1795 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { 1796 device_printf(dev, 1797 "Unable to allocate TSO TX buffer map memory\n"); 1798 err = ENOMEM; 1799 goto fail; 1800 } 1801 for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) { 1802 err = bus_dmamap_create(txq->ift_buf_tag, 0, 1803 &txq->ift_sds.ifsd_map[i]); 1804 if (err != 0) { 1805 device_printf(dev, "Unable to create TX DMA map\n"); 1806 goto fail; 1807 } 1808 if (!tso) 1809 continue; 1810 err = bus_dmamap_create(txq->ift_tso_buf_tag, 0, 1811 &txq->ift_sds.ifsd_tso_map[i]); 1812 if (err != 0) { 1813 device_printf(dev, "Unable to create TSO TX DMA map\n"); 1814 goto fail; 1815 } 1816 } 1817 return (0); 1818 fail: 1819 /* We free all, it handles case where we are in the middle */ 1820 iflib_tx_structures_free(ctx); 1821 return (err); 1822 } 1823 1824 static void 1825 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i) 1826 { 1827 bus_dmamap_t map; 1828 1829 if (txq->ift_sds.ifsd_map != NULL) { 1830 map = txq->ift_sds.ifsd_map[i]; 1831 bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE); 1832 bus_dmamap_unload(txq->ift_buf_tag, map); 1833 bus_dmamap_destroy(txq->ift_buf_tag, map); 1834 txq->ift_sds.ifsd_map[i] = NULL; 1835 } 1836 1837 if (txq->ift_sds.ifsd_tso_map != NULL) { 1838 map = txq->ift_sds.ifsd_tso_map[i]; 1839 bus_dmamap_sync(txq->ift_tso_buf_tag, map, 1840 BUS_DMASYNC_POSTWRITE); 1841 bus_dmamap_unload(txq->ift_tso_buf_tag, map); 1842 bus_dmamap_destroy(txq->ift_tso_buf_tag, map); 1843 txq->ift_sds.ifsd_tso_map[i] = NULL; 1844 } 1845 } 1846 1847 static void 1848 iflib_txq_destroy(iflib_txq_t txq) 1849 { 1850 if_ctx_t ctx = txq->ift_ctx; 1851 1852 for (int i = 0; i < txq->ift_size; i++) 1853 iflib_txsd_destroy(ctx, txq, i); 1854 1855 if (txq->ift_br != NULL) { 1856 ifmp_ring_free(txq->ift_br); 1857 txq->ift_br = NULL; 1858 } 1859 1860 mtx_destroy(&txq->ift_mtx); 1861 1862 if (txq->ift_sds.ifsd_map != NULL) { 1863 free(txq->ift_sds.ifsd_map, M_IFLIB); 1864 txq->ift_sds.ifsd_map = NULL; 1865 } 1866 if (txq->ift_sds.ifsd_tso_map != NULL) { 1867 free(txq->ift_sds.ifsd_tso_map, M_IFLIB); 1868 txq->ift_sds.ifsd_tso_map = NULL; 1869 } 1870 if (txq->ift_sds.ifsd_m != NULL) { 1871 free(txq->ift_sds.ifsd_m, M_IFLIB); 1872 txq->ift_sds.ifsd_m = NULL; 1873 } 1874 if (txq->ift_buf_tag != NULL) { 1875 bus_dma_tag_destroy(txq->ift_buf_tag); 1876 txq->ift_buf_tag = NULL; 1877 } 1878 if (txq->ift_tso_buf_tag != NULL) { 1879 bus_dma_tag_destroy(txq->ift_tso_buf_tag); 1880 txq->ift_tso_buf_tag = NULL; 1881 } 1882 if (txq->ift_ifdi != NULL) { 1883 free(txq->ift_ifdi, M_IFLIB); 1884 } 1885 } 1886 1887 static void 1888 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i) 1889 { 1890 struct mbuf **mp; 1891 1892 mp = &txq->ift_sds.ifsd_m[i]; 1893 if (*mp == NULL) 1894 return; 1895 1896 if (txq->ift_sds.ifsd_map != NULL) { 1897 bus_dmamap_sync(txq->ift_buf_tag, 1898 txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE); 1899 bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]); 1900 } 1901 if (txq->ift_sds.ifsd_tso_map != NULL) { 1902 bus_dmamap_sync(txq->ift_tso_buf_tag, 1903 txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE); 1904 bus_dmamap_unload(txq->ift_tso_buf_tag, 1905 txq->ift_sds.ifsd_tso_map[i]); 1906 } 1907 m_freem(*mp); 1908 DBG_COUNTER_INC(tx_frees); 1909 *mp = NULL; 1910 } 1911 1912 static int 1913 iflib_txq_setup(iflib_txq_t txq) 1914 { 1915 if_ctx_t ctx = txq->ift_ctx; 1916 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1917 if_shared_ctx_t sctx = ctx->ifc_sctx; 1918 iflib_dma_info_t di; 1919 int i; 1920 1921 /* Set number of descriptors available */ 1922 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 1923 /* XXX make configurable */ 1924 txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ; 1925 1926 /* Reset indices */ 1927 txq->ift_cidx_processed = 0; 1928 txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; 1929 txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset]; 1930 1931 for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) 1932 bzero((void *)di->idi_vaddr, di->idi_size); 1933 1934 IFDI_TXQ_SETUP(ctx, txq->ift_id); 1935 for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) 1936 bus_dmamap_sync(di->idi_tag, di->idi_map, 1937 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 1938 return (0); 1939 } 1940 1941 /********************************************************************* 1942 * 1943 * Allocate DMA resources for RX buffers as well as memory for the RX 1944 * mbuf map, direct RX cluster pointer map and RX cluster bus address 1945 * map. RX DMA map, RX mbuf map, direct RX cluster pointer map and 1946 * RX cluster map are kept in a iflib_sw_rx_desc_array structure. 1947 * Since we use use one entry in iflib_sw_rx_desc_array per received 1948 * packet, the maximum number of entries we'll need is equal to the 1949 * number of hardware receive descriptors that we've allocated. 1950 * 1951 **********************************************************************/ 1952 static int 1953 iflib_rxsd_alloc(iflib_rxq_t rxq) 1954 { 1955 if_ctx_t ctx = rxq->ifr_ctx; 1956 if_shared_ctx_t sctx = ctx->ifc_sctx; 1957 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 1958 device_t dev = ctx->ifc_dev; 1959 iflib_fl_t fl; 1960 bus_addr_t lowaddr; 1961 int err; 1962 1963 MPASS(scctx->isc_nrxd[0] > 0); 1964 MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0); 1965 1966 lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width); 1967 1968 fl = rxq->ifr_fl; 1969 for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { 1970 fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */ 1971 /* Set up DMA tag for RX buffers. */ 1972 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1973 1, 0, /* alignment, bounds */ 1974 lowaddr, /* lowaddr */ 1975 BUS_SPACE_MAXADDR, /* highaddr */ 1976 NULL, NULL, /* filter, filterarg */ 1977 sctx->isc_rx_maxsize, /* maxsize */ 1978 sctx->isc_rx_nsegments, /* nsegments */ 1979 sctx->isc_rx_maxsegsize, /* maxsegsize */ 1980 0, /* flags */ 1981 NULL, /* lockfunc */ 1982 NULL, /* lockarg */ 1983 &fl->ifl_buf_tag); 1984 if (err) { 1985 device_printf(dev, 1986 "Unable to allocate RX DMA tag: %d\n", err); 1987 goto fail; 1988 } 1989 1990 /* Allocate memory for the RX mbuf map. */ 1991 if (!(fl->ifl_sds.ifsd_m = 1992 (struct mbuf **) malloc(sizeof(struct mbuf *) * 1993 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 1994 device_printf(dev, 1995 "Unable to allocate RX mbuf map memory\n"); 1996 err = ENOMEM; 1997 goto fail; 1998 } 1999 2000 /* Allocate memory for the direct RX cluster pointer map. */ 2001 if (!(fl->ifl_sds.ifsd_cl = 2002 (caddr_t *) malloc(sizeof(caddr_t) * 2003 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 2004 device_printf(dev, 2005 "Unable to allocate RX cluster map memory\n"); 2006 err = ENOMEM; 2007 goto fail; 2008 } 2009 2010 /* Allocate memory for the RX cluster bus address map. */ 2011 if (!(fl->ifl_sds.ifsd_ba = 2012 (bus_addr_t *) malloc(sizeof(bus_addr_t) * 2013 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 2014 device_printf(dev, 2015 "Unable to allocate RX bus address map memory\n"); 2016 err = ENOMEM; 2017 goto fail; 2018 } 2019 2020 /* 2021 * Create the DMA maps for RX buffers. 2022 */ 2023 if (!(fl->ifl_sds.ifsd_map = 2024 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { 2025 device_printf(dev, 2026 "Unable to allocate RX buffer DMA map memory\n"); 2027 err = ENOMEM; 2028 goto fail; 2029 } 2030 for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) { 2031 err = bus_dmamap_create(fl->ifl_buf_tag, 0, 2032 &fl->ifl_sds.ifsd_map[i]); 2033 if (err != 0) { 2034 device_printf(dev, "Unable to create RX buffer DMA map\n"); 2035 goto fail; 2036 } 2037 } 2038 } 2039 return (0); 2040 2041 fail: 2042 iflib_rx_structures_free(ctx); 2043 return (err); 2044 } 2045 2046 /* 2047 * Internal service routines 2048 */ 2049 2050 struct rxq_refill_cb_arg { 2051 int error; 2052 bus_dma_segment_t seg; 2053 int nseg; 2054 }; 2055 2056 static void 2057 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) 2058 { 2059 struct rxq_refill_cb_arg *cb_arg = arg; 2060 2061 cb_arg->error = error; 2062 cb_arg->seg = segs[0]; 2063 cb_arg->nseg = nseg; 2064 } 2065 2066 /** 2067 * iflib_fl_refill - refill an rxq free-buffer list 2068 * @ctx: the iflib context 2069 * @fl: the free list to refill 2070 * @count: the number of new buffers to allocate 2071 * 2072 * (Re)populate an rxq free-buffer list with up to @count new packet buffers. 2073 * The caller must assure that @count does not exceed the queue's capacity 2074 * minus one (since we always leave a descriptor unavailable). 2075 */ 2076 static uint8_t 2077 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count) 2078 { 2079 struct if_rxd_update iru; 2080 struct rxq_refill_cb_arg cb_arg; 2081 struct mbuf *m; 2082 caddr_t cl, *sd_cl; 2083 struct mbuf **sd_m; 2084 bus_dmamap_t *sd_map; 2085 bus_addr_t bus_addr, *sd_ba; 2086 int err, frag_idx, i, idx, n, pidx; 2087 qidx_t credits; 2088 2089 MPASS(count <= fl->ifl_size - fl->ifl_credits - 1); 2090 2091 sd_m = fl->ifl_sds.ifsd_m; 2092 sd_map = fl->ifl_sds.ifsd_map; 2093 sd_cl = fl->ifl_sds.ifsd_cl; 2094 sd_ba = fl->ifl_sds.ifsd_ba; 2095 pidx = fl->ifl_pidx; 2096 idx = pidx; 2097 frag_idx = fl->ifl_fragidx; 2098 credits = fl->ifl_credits; 2099 2100 i = 0; 2101 n = count; 2102 MPASS(n > 0); 2103 MPASS(credits + n <= fl->ifl_size); 2104 2105 if (pidx < fl->ifl_cidx) 2106 MPASS(pidx + n <= fl->ifl_cidx); 2107 if (pidx == fl->ifl_cidx && (credits < fl->ifl_size)) 2108 MPASS(fl->ifl_gen == 0); 2109 if (pidx > fl->ifl_cidx) 2110 MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx); 2111 2112 DBG_COUNTER_INC(fl_refills); 2113 if (n > 8) 2114 DBG_COUNTER_INC(fl_refills_large); 2115 iru_init(&iru, fl->ifl_rxq, fl->ifl_id); 2116 while (n-- > 0) { 2117 /* 2118 * We allocate an uninitialized mbuf + cluster, mbuf is 2119 * initialized after rx. 2120 * 2121 * If the cluster is still set then we know a minimum sized 2122 * packet was received 2123 */ 2124 bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size, 2125 &frag_idx); 2126 if (frag_idx < 0) 2127 bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx); 2128 MPASS(frag_idx >= 0); 2129 if ((cl = sd_cl[frag_idx]) == NULL) { 2130 cl = uma_zalloc(fl->ifl_zone, M_NOWAIT); 2131 if (__predict_false(cl == NULL)) 2132 break; 2133 2134 cb_arg.error = 0; 2135 MPASS(sd_map != NULL); 2136 err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx], 2137 cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 2138 BUS_DMA_NOWAIT); 2139 if (__predict_false(err != 0 || cb_arg.error)) { 2140 uma_zfree(fl->ifl_zone, cl); 2141 break; 2142 } 2143 2144 sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr; 2145 sd_cl[frag_idx] = cl; 2146 #if MEMORY_LOGGING 2147 fl->ifl_cl_enqueued++; 2148 #endif 2149 } else { 2150 bus_addr = sd_ba[frag_idx]; 2151 } 2152 bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx], 2153 BUS_DMASYNC_PREREAD); 2154 2155 if (sd_m[frag_idx] == NULL) { 2156 m = m_gethdr_raw(M_NOWAIT, 0); 2157 if (__predict_false(m == NULL)) 2158 break; 2159 sd_m[frag_idx] = m; 2160 } 2161 bit_set(fl->ifl_rx_bitmap, frag_idx); 2162 #if MEMORY_LOGGING 2163 fl->ifl_m_enqueued++; 2164 #endif 2165 2166 DBG_COUNTER_INC(rx_allocs); 2167 fl->ifl_rxd_idxs[i] = frag_idx; 2168 fl->ifl_bus_addrs[i] = bus_addr; 2169 credits++; 2170 i++; 2171 MPASS(credits <= fl->ifl_size); 2172 if (++idx == fl->ifl_size) { 2173 #ifdef INVARIANTS 2174 fl->ifl_gen = 1; 2175 #endif 2176 idx = 0; 2177 } 2178 if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { 2179 iru.iru_pidx = pidx; 2180 iru.iru_count = i; 2181 ctx->isc_rxd_refill(ctx->ifc_softc, &iru); 2182 fl->ifl_pidx = idx; 2183 fl->ifl_credits = credits; 2184 pidx = idx; 2185 i = 0; 2186 } 2187 } 2188 2189 if (n < count - 1) { 2190 if (i != 0) { 2191 iru.iru_pidx = pidx; 2192 iru.iru_count = i; 2193 ctx->isc_rxd_refill(ctx->ifc_softc, &iru); 2194 fl->ifl_pidx = idx; 2195 fl->ifl_credits = credits; 2196 } 2197 DBG_COUNTER_INC(rxd_flush); 2198 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 2199 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 2200 ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, 2201 fl->ifl_id, fl->ifl_pidx); 2202 if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) { 2203 fl->ifl_fragidx = frag_idx + 1; 2204 if (fl->ifl_fragidx == fl->ifl_size) 2205 fl->ifl_fragidx = 0; 2206 } else { 2207 fl->ifl_fragidx = frag_idx; 2208 } 2209 } 2210 2211 return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY); 2212 } 2213 2214 static inline uint8_t 2215 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl) 2216 { 2217 /* 2218 * We leave an unused descriptor to avoid pidx to catch up with cidx. 2219 * This is important as it confuses most NICs. For instance, 2220 * Intel NICs have (per receive ring) RDH and RDT registers, where 2221 * RDH points to the next receive descriptor to be used by the NIC, 2222 * and RDT for the next receive descriptor to be published by the 2223 * driver to the NIC (RDT - 1 is thus the last valid one). 2224 * The condition RDH == RDT means no descriptors are available to 2225 * the NIC, and thus it would be ambiguous if it also meant that 2226 * all the descriptors are available to the NIC. 2227 */ 2228 int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1; 2229 #ifdef INVARIANTS 2230 int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1; 2231 #endif 2232 2233 MPASS(fl->ifl_credits <= fl->ifl_size); 2234 MPASS(reclaimable == delta); 2235 2236 if (reclaimable > 0) 2237 return (iflib_fl_refill(ctx, fl, reclaimable)); 2238 return (0); 2239 } 2240 2241 uint8_t 2242 iflib_in_detach(if_ctx_t ctx) 2243 { 2244 bool in_detach; 2245 2246 STATE_LOCK(ctx); 2247 in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH); 2248 STATE_UNLOCK(ctx); 2249 return (in_detach); 2250 } 2251 2252 static void 2253 iflib_fl_bufs_free(iflib_fl_t fl) 2254 { 2255 iflib_dma_info_t idi = fl->ifl_ifdi; 2256 bus_dmamap_t sd_map; 2257 uint32_t i; 2258 2259 for (i = 0; i < fl->ifl_size; i++) { 2260 struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i]; 2261 caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i]; 2262 2263 if (*sd_cl != NULL) { 2264 sd_map = fl->ifl_sds.ifsd_map[i]; 2265 bus_dmamap_sync(fl->ifl_buf_tag, sd_map, 2266 BUS_DMASYNC_POSTREAD); 2267 bus_dmamap_unload(fl->ifl_buf_tag, sd_map); 2268 uma_zfree(fl->ifl_zone, *sd_cl); 2269 *sd_cl = NULL; 2270 if (*sd_m != NULL) { 2271 m_init(*sd_m, M_NOWAIT, MT_DATA, 0); 2272 m_free_raw(*sd_m); 2273 *sd_m = NULL; 2274 } 2275 } else { 2276 MPASS(*sd_m == NULL); 2277 } 2278 #if MEMORY_LOGGING 2279 fl->ifl_m_dequeued++; 2280 fl->ifl_cl_dequeued++; 2281 #endif 2282 } 2283 #ifdef INVARIANTS 2284 for (i = 0; i < fl->ifl_size; i++) { 2285 MPASS(fl->ifl_sds.ifsd_cl[i] == NULL); 2286 MPASS(fl->ifl_sds.ifsd_m[i] == NULL); 2287 } 2288 #endif 2289 /* 2290 * Reset free list values 2291 */ 2292 fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0; 2293 bzero(idi->idi_vaddr, idi->idi_size); 2294 } 2295 2296 /********************************************************************* 2297 * 2298 * Initialize a free list and its buffers. 2299 * 2300 **********************************************************************/ 2301 static int 2302 iflib_fl_setup(iflib_fl_t fl) 2303 { 2304 iflib_rxq_t rxq = fl->ifl_rxq; 2305 if_ctx_t ctx = rxq->ifr_ctx; 2306 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 2307 int qidx; 2308 2309 bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1); 2310 /* 2311 * Free current RX buffer structs and their mbufs 2312 */ 2313 iflib_fl_bufs_free(fl); 2314 /* Now replenish the mbufs */ 2315 MPASS(fl->ifl_credits == 0); 2316 qidx = rxq->ifr_fl_offset + fl->ifl_id; 2317 if (scctx->isc_rxd_buf_size[qidx] != 0) 2318 fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx]; 2319 else 2320 fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz; 2321 /* 2322 * ifl_buf_size may be a driver-supplied value, so pull it up 2323 * to the selected mbuf size. 2324 */ 2325 fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size); 2326 if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size) 2327 ctx->ifc_max_fl_buf_size = fl->ifl_buf_size; 2328 fl->ifl_cltype = m_gettype(fl->ifl_buf_size); 2329 fl->ifl_zone = m_getzone(fl->ifl_buf_size); 2330 2331 /* 2332 * Avoid pre-allocating zillions of clusters to an idle card 2333 * potentially speeding up attach. In any case make sure 2334 * to leave a descriptor unavailable. See the comment in 2335 * iflib_fl_refill_all(). 2336 */ 2337 MPASS(fl->ifl_size > 0); 2338 (void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1)); 2339 if (min(128, fl->ifl_size - 1) != fl->ifl_credits) 2340 return (ENOBUFS); 2341 /* 2342 * handle failure 2343 */ 2344 MPASS(rxq != NULL); 2345 MPASS(fl->ifl_ifdi != NULL); 2346 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 2347 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 2348 return (0); 2349 } 2350 2351 /********************************************************************* 2352 * 2353 * Free receive ring data structures 2354 * 2355 **********************************************************************/ 2356 static void 2357 iflib_rx_sds_free(iflib_rxq_t rxq) 2358 { 2359 iflib_fl_t fl; 2360 int i, j; 2361 2362 if (rxq->ifr_fl != NULL) { 2363 for (i = 0; i < rxq->ifr_nfl; i++) { 2364 fl = &rxq->ifr_fl[i]; 2365 if (fl->ifl_buf_tag != NULL) { 2366 if (fl->ifl_sds.ifsd_map != NULL) { 2367 for (j = 0; j < fl->ifl_size; j++) { 2368 bus_dmamap_sync( 2369 fl->ifl_buf_tag, 2370 fl->ifl_sds.ifsd_map[j], 2371 BUS_DMASYNC_POSTREAD); 2372 bus_dmamap_unload( 2373 fl->ifl_buf_tag, 2374 fl->ifl_sds.ifsd_map[j]); 2375 bus_dmamap_destroy( 2376 fl->ifl_buf_tag, 2377 fl->ifl_sds.ifsd_map[j]); 2378 } 2379 } 2380 bus_dma_tag_destroy(fl->ifl_buf_tag); 2381 fl->ifl_buf_tag = NULL; 2382 } 2383 free(fl->ifl_sds.ifsd_m, M_IFLIB); 2384 free(fl->ifl_sds.ifsd_cl, M_IFLIB); 2385 free(fl->ifl_sds.ifsd_ba, M_IFLIB); 2386 free(fl->ifl_sds.ifsd_map, M_IFLIB); 2387 free(fl->ifl_rx_bitmap, M_IFLIB); 2388 fl->ifl_sds.ifsd_m = NULL; 2389 fl->ifl_sds.ifsd_cl = NULL; 2390 fl->ifl_sds.ifsd_ba = NULL; 2391 fl->ifl_sds.ifsd_map = NULL; 2392 fl->ifl_rx_bitmap = NULL; 2393 } 2394 free(rxq->ifr_fl, M_IFLIB); 2395 rxq->ifr_fl = NULL; 2396 free(rxq->ifr_ifdi, M_IFLIB); 2397 rxq->ifr_ifdi = NULL; 2398 rxq->ifr_cq_cidx = 0; 2399 } 2400 } 2401 2402 /* 2403 * Timer routine 2404 */ 2405 static void 2406 iflib_timer(void *arg) 2407 { 2408 iflib_txq_t txq = arg; 2409 if_ctx_t ctx = txq->ift_ctx; 2410 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 2411 uint64_t this_tick = ticks; 2412 2413 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) 2414 return; 2415 2416 /* 2417 ** Check on the state of the TX queue(s), this 2418 ** can be done without the lock because its RO 2419 ** and the HUNG state will be static if set. 2420 */ 2421 if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) { 2422 txq->ift_last_timer_tick = this_tick; 2423 IFDI_TIMER(ctx, txq->ift_id); 2424 if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) && 2425 ((txq->ift_cleaned_prev == txq->ift_cleaned) || 2426 (sctx->isc_pause_frames == 0))) 2427 goto hung; 2428 2429 if (txq->ift_qstatus != IFLIB_QUEUE_IDLE && 2430 ifmp_ring_is_stalled(txq->ift_br)) { 2431 KASSERT(ctx->ifc_link_state == LINK_STATE_UP, 2432 ("queue can't be marked as hung if interface is down")); 2433 txq->ift_qstatus = IFLIB_QUEUE_HUNG; 2434 } 2435 txq->ift_cleaned_prev = txq->ift_cleaned; 2436 } 2437 /* handle any laggards */ 2438 if (txq->ift_db_pending) 2439 GROUPTASK_ENQUEUE(&txq->ift_task); 2440 2441 sctx->isc_pause_frames = 0; 2442 if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) 2443 callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, 2444 txq, txq->ift_timer.c_cpu); 2445 return; 2446 2447 hung: 2448 device_printf(ctx->ifc_dev, 2449 "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n", 2450 txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx); 2451 STATE_LOCK(ctx); 2452 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 2453 ctx->ifc_flags |= (IFC_DO_WATCHDOG | IFC_DO_RESET); 2454 iflib_admin_intr_deferred(ctx); 2455 STATE_UNLOCK(ctx); 2456 } 2457 2458 static uint16_t 2459 iflib_get_mbuf_size_for(unsigned int size) 2460 { 2461 2462 if (size <= MCLBYTES) 2463 return (MCLBYTES); 2464 else 2465 return (MJUMPAGESIZE); 2466 } 2467 2468 static void 2469 iflib_calc_rx_mbuf_sz(if_ctx_t ctx) 2470 { 2471 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 2472 2473 /* 2474 * XXX don't set the max_frame_size to larger 2475 * than the hardware can handle 2476 */ 2477 ctx->ifc_rx_mbuf_sz = 2478 iflib_get_mbuf_size_for(sctx->isc_max_frame_size); 2479 } 2480 2481 uint32_t 2482 iflib_get_rx_mbuf_sz(if_ctx_t ctx) 2483 { 2484 2485 return (ctx->ifc_rx_mbuf_sz); 2486 } 2487 2488 static void 2489 iflib_init_locked(if_ctx_t ctx) 2490 { 2491 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 2492 if_t ifp = ctx->ifc_ifp; 2493 iflib_fl_t fl; 2494 iflib_txq_t txq; 2495 iflib_rxq_t rxq; 2496 int i, j, tx_ip_csum_flags, tx_ip6_csum_flags; 2497 2498 if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 2499 IFDI_INTR_DISABLE(ctx); 2500 2501 /* 2502 * See iflib_stop(). Useful in case iflib_init_locked() is 2503 * called without first calling iflib_stop(). 2504 */ 2505 netmap_disable_all_rings(ifp); 2506 2507 tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP); 2508 tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP); 2509 /* Set hardware offload abilities */ 2510 if_clearhwassist(ifp); 2511 if (if_getcapenable(ifp) & IFCAP_TXCSUM) 2512 if_sethwassistbits(ifp, tx_ip_csum_flags, 0); 2513 if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6) 2514 if_sethwassistbits(ifp, tx_ip6_csum_flags, 0); 2515 if (if_getcapenable(ifp) & IFCAP_TSO4) 2516 if_sethwassistbits(ifp, CSUM_IP_TSO, 0); 2517 if (if_getcapenable(ifp) & IFCAP_TSO6) 2518 if_sethwassistbits(ifp, CSUM_IP6_TSO, 0); 2519 2520 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { 2521 CALLOUT_LOCK(txq); 2522 callout_stop(&txq->ift_timer); 2523 #ifdef DEV_NETMAP 2524 callout_stop(&txq->ift_netmap_timer); 2525 #endif /* DEV_NETMAP */ 2526 CALLOUT_UNLOCK(txq); 2527 (void)iflib_netmap_txq_init(ctx, txq); 2528 } 2529 2530 /* 2531 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so 2532 * that drivers can use the value when setting up the hardware receive 2533 * buffers. 2534 */ 2535 iflib_calc_rx_mbuf_sz(ctx); 2536 2537 #ifdef INVARIANTS 2538 i = if_getdrvflags(ifp); 2539 #endif 2540 IFDI_INIT(ctx); 2541 MPASS(if_getdrvflags(ifp) == i); 2542 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { 2543 if (iflib_netmap_rxq_init(ctx, rxq) > 0) { 2544 /* This rxq is in netmap mode. Skip normal init. */ 2545 continue; 2546 } 2547 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { 2548 if (iflib_fl_setup(fl)) { 2549 device_printf(ctx->ifc_dev, 2550 "setting up free list %d failed - " 2551 "check cluster settings\n", j); 2552 goto done; 2553 } 2554 } 2555 } 2556 done: 2557 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 2558 IFDI_INTR_ENABLE(ctx); 2559 txq = ctx->ifc_txqs; 2560 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) 2561 callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq, 2562 txq->ift_timer.c_cpu); 2563 2564 /* Re-enable txsync/rxsync. */ 2565 netmap_enable_all_rings(ifp); 2566 } 2567 2568 static int 2569 iflib_media_change(if_t ifp) 2570 { 2571 if_ctx_t ctx = if_getsoftc(ifp); 2572 int err; 2573 2574 CTX_LOCK(ctx); 2575 if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0) 2576 iflib_if_init_locked(ctx); 2577 CTX_UNLOCK(ctx); 2578 return (err); 2579 } 2580 2581 static void 2582 iflib_media_status(if_t ifp, struct ifmediareq *ifmr) 2583 { 2584 if_ctx_t ctx = if_getsoftc(ifp); 2585 2586 CTX_LOCK(ctx); 2587 IFDI_UPDATE_ADMIN_STATUS(ctx); 2588 IFDI_MEDIA_STATUS(ctx, ifmr); 2589 CTX_UNLOCK(ctx); 2590 } 2591 2592 static void 2593 iflib_stop(if_ctx_t ctx) 2594 { 2595 iflib_txq_t txq = ctx->ifc_txqs; 2596 iflib_rxq_t rxq = ctx->ifc_rxqs; 2597 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 2598 if_shared_ctx_t sctx = ctx->ifc_sctx; 2599 iflib_dma_info_t di; 2600 iflib_fl_t fl; 2601 int i, j; 2602 2603 /* Tell the stack that the interface is no longer active */ 2604 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 2605 2606 IFDI_INTR_DISABLE(ctx); 2607 DELAY(1000); 2608 IFDI_STOP(ctx); 2609 DELAY(1000); 2610 2611 /* 2612 * Stop any pending txsync/rxsync and prevent new ones 2613 * form starting. Processes blocked in poll() will get 2614 * POLLERR. 2615 */ 2616 netmap_disable_all_rings(ctx->ifc_ifp); 2617 2618 iflib_debug_reset(); 2619 /* Wait for current tx queue users to exit to disarm watchdog timer. */ 2620 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) { 2621 /* make sure all transmitters have completed before proceeding XXX */ 2622 2623 CALLOUT_LOCK(txq); 2624 callout_stop(&txq->ift_timer); 2625 #ifdef DEV_NETMAP 2626 callout_stop(&txq->ift_netmap_timer); 2627 #endif /* DEV_NETMAP */ 2628 CALLOUT_UNLOCK(txq); 2629 2630 if (!ctx->ifc_sysctl_simple_tx) { 2631 /* clean any enqueued buffers */ 2632 iflib_ifmp_purge(txq); 2633 } 2634 /* Free any existing tx buffers. */ 2635 for (j = 0; j < txq->ift_size; j++) { 2636 iflib_txsd_free(ctx, txq, j); 2637 } 2638 txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; 2639 txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0; 2640 if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES) 2641 txq->ift_cidx = txq->ift_pidx; 2642 else 2643 txq->ift_cidx = txq->ift_pidx = 0; 2644 2645 txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0; 2646 txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0; 2647 txq->ift_pullups = 0; 2648 ifmp_ring_reset_stats(txq->ift_br); 2649 for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++) 2650 bzero((void *)di->idi_vaddr, di->idi_size); 2651 } 2652 for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) { 2653 if (rxq->ifr_task.gt_taskqueue != NULL) 2654 gtaskqueue_drain(rxq->ifr_task.gt_taskqueue, 2655 &rxq->ifr_task.gt_task); 2656 2657 rxq->ifr_cq_cidx = 0; 2658 for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++) 2659 bzero((void *)di->idi_vaddr, di->idi_size); 2660 /* also resets the free lists pidx/cidx */ 2661 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) 2662 iflib_fl_bufs_free(fl); 2663 } 2664 } 2665 2666 static inline caddr_t 2667 calc_next_rxd(iflib_fl_t fl, int cidx) 2668 { 2669 qidx_t size; 2670 int nrxd; 2671 caddr_t start, end, cur, next; 2672 2673 nrxd = fl->ifl_size; 2674 size = fl->ifl_rxd_size; 2675 start = fl->ifl_ifdi->idi_vaddr; 2676 2677 if (__predict_false(size == 0)) 2678 return (start); 2679 cur = start + size * cidx; 2680 end = start + size * nrxd; 2681 next = CACHE_PTR_NEXT(cur); 2682 return (next < end ? next : start); 2683 } 2684 2685 static inline void 2686 prefetch_pkts(iflib_fl_t fl, int cidx) 2687 { 2688 int nextptr; 2689 int nrxd = fl->ifl_size; 2690 caddr_t next_rxd; 2691 2692 nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1); 2693 prefetch(&fl->ifl_sds.ifsd_m[nextptr]); 2694 prefetch(&fl->ifl_sds.ifsd_cl[nextptr]); 2695 next_rxd = calc_next_rxd(fl, cidx); 2696 prefetch(next_rxd); 2697 prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]); 2698 prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]); 2699 prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]); 2700 prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]); 2701 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]); 2702 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]); 2703 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]); 2704 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]); 2705 } 2706 2707 static struct mbuf * 2708 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd, 2709 int *pf_rv, if_rxd_info_t ri) 2710 { 2711 bus_dmamap_t map; 2712 iflib_fl_t fl; 2713 caddr_t payload; 2714 struct mbuf *m; 2715 int flid, cidx, len, next; 2716 2717 map = NULL; 2718 flid = irf->irf_flid; 2719 cidx = irf->irf_idx; 2720 fl = &rxq->ifr_fl[flid]; 2721 sd->ifsd_fl = fl; 2722 sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx]; 2723 fl->ifl_credits--; 2724 #if MEMORY_LOGGING 2725 fl->ifl_m_dequeued++; 2726 #endif 2727 if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH) 2728 prefetch_pkts(fl, cidx); 2729 next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1); 2730 prefetch(&fl->ifl_sds.ifsd_map[next]); 2731 map = fl->ifl_sds.ifsd_map[cidx]; 2732 2733 bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD); 2734 2735 if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL && 2736 irf->irf_len != 0) { 2737 payload = *sd->ifsd_cl; 2738 payload += ri->iri_pad; 2739 len = ri->iri_len - ri->iri_pad; 2740 *pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m); 2741 switch (*pf_rv) { 2742 case PFIL_DROPPED: 2743 case PFIL_CONSUMED: 2744 /* 2745 * The filter ate it. Everything is recycled. 2746 */ 2747 m = NULL; 2748 unload = 0; 2749 break; 2750 case PFIL_REALLOCED: 2751 /* 2752 * The filter copied it. Everything is recycled. 2753 * 'm' points at new mbuf. 2754 */ 2755 unload = 0; 2756 break; 2757 case PFIL_PASS: 2758 /* 2759 * Filter said it was OK, so receive like 2760 * normal 2761 */ 2762 m = fl->ifl_sds.ifsd_m[cidx]; 2763 fl->ifl_sds.ifsd_m[cidx] = NULL; 2764 break; 2765 default: 2766 MPASS(0); 2767 } 2768 } else { 2769 m = fl->ifl_sds.ifsd_m[cidx]; 2770 fl->ifl_sds.ifsd_m[cidx] = NULL; 2771 if (pf_rv != NULL) 2772 *pf_rv = PFIL_PASS; 2773 } 2774 2775 if (unload && irf->irf_len != 0) 2776 bus_dmamap_unload(fl->ifl_buf_tag, map); 2777 fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1); 2778 if (__predict_false(fl->ifl_cidx == 0)) 2779 fl->ifl_gen = 0; 2780 bit_clear(fl->ifl_rx_bitmap, cidx); 2781 return (m); 2782 } 2783 2784 static struct mbuf * 2785 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv) 2786 { 2787 struct mbuf *m, *mh, *mt; 2788 caddr_t cl; 2789 int *pf_rv_ptr, flags, i, padlen; 2790 bool consumed; 2791 2792 i = 0; 2793 mh = NULL; 2794 consumed = false; 2795 *pf_rv = PFIL_PASS; 2796 pf_rv_ptr = pf_rv; 2797 do { 2798 m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd, 2799 pf_rv_ptr, ri); 2800 2801 MPASS(*sd->ifsd_cl != NULL); 2802 2803 /* 2804 * Exclude zero-length frags & frags from 2805 * packets the filter has consumed or dropped 2806 */ 2807 if (ri->iri_frags[i].irf_len == 0 || consumed || 2808 *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) { 2809 if (mh == NULL) { 2810 /* everything saved here */ 2811 consumed = true; 2812 pf_rv_ptr = NULL; 2813 continue; 2814 } 2815 /* XXX we can save the cluster here, but not the mbuf */ 2816 m_init(m, M_NOWAIT, MT_DATA, 0); 2817 m_free(m); 2818 continue; 2819 } 2820 if (mh == NULL) { 2821 flags = M_PKTHDR | M_EXT; 2822 mh = mt = m; 2823 padlen = ri->iri_pad; 2824 } else { 2825 flags = M_EXT; 2826 mt->m_next = m; 2827 mt = m; 2828 /* assuming padding is only on the first fragment */ 2829 padlen = 0; 2830 } 2831 cl = *sd->ifsd_cl; 2832 *sd->ifsd_cl = NULL; 2833 2834 /* Can these two be made one ? */ 2835 m_init(m, M_NOWAIT, MT_DATA, flags); 2836 m_cljset(m, cl, sd->ifsd_fl->ifl_cltype); 2837 /* 2838 * These must follow m_init and m_cljset 2839 */ 2840 m->m_data += padlen; 2841 ri->iri_len -= padlen; 2842 m->m_len = ri->iri_frags[i].irf_len; 2843 } while (++i < ri->iri_nfrags); 2844 2845 return (mh); 2846 } 2847 2848 /* 2849 * Process one software descriptor 2850 */ 2851 static struct mbuf * 2852 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri) 2853 { 2854 struct if_rxsd sd; 2855 struct mbuf *m; 2856 int pf_rv; 2857 2858 /* should I merge this back in now that the two paths are basically duplicated? */ 2859 if (ri->iri_nfrags == 1 && 2860 ri->iri_frags[0].irf_len != 0 && 2861 ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) { 2862 m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd, 2863 &pf_rv, ri); 2864 if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) 2865 return (m); 2866 if (pf_rv == PFIL_PASS) { 2867 m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); 2868 #ifndef __NO_STRICT_ALIGNMENT 2869 if (!IP_ALIGNED(m) && ri->iri_pad == 0) 2870 m->m_data += 2; 2871 #endif 2872 memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len); 2873 m->m_len = ri->iri_frags[0].irf_len; 2874 m->m_data += ri->iri_pad; 2875 ri->iri_len -= ri->iri_pad; 2876 } 2877 } else { 2878 m = assemble_segments(rxq, ri, &sd, &pf_rv); 2879 if (m == NULL) 2880 return (NULL); 2881 if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) 2882 return (m); 2883 } 2884 m->m_pkthdr.len = ri->iri_len; 2885 m->m_pkthdr.rcvif = ri->iri_ifp; 2886 m->m_flags |= ri->iri_flags; 2887 m->m_pkthdr.ether_vtag = ri->iri_vtag; 2888 m->m_pkthdr.flowid = ri->iri_flowid; 2889 #ifdef NUMA 2890 m->m_pkthdr.numa_domain = if_getnumadomain(ri->iri_ifp); 2891 #endif 2892 M_HASHTYPE_SET(m, ri->iri_rsstype); 2893 m->m_pkthdr.csum_flags = ri->iri_csum_flags; 2894 m->m_pkthdr.csum_data = ri->iri_csum_data; 2895 return (m); 2896 } 2897 2898 static void 2899 _task_fn_rx_watchdog(void *context) 2900 { 2901 iflib_rxq_t rxq = context; 2902 2903 GROUPTASK_ENQUEUE(&rxq->ifr_task); 2904 } 2905 2906 static uint8_t 2907 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget) 2908 { 2909 if_t ifp; 2910 if_ctx_t ctx = rxq->ifr_ctx; 2911 if_shared_ctx_t sctx = ctx->ifc_sctx; 2912 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 2913 int avail, i; 2914 qidx_t *cidxp; 2915 struct if_rxd_info ri; 2916 int err, budget_left, rx_bytes, rx_pkts; 2917 iflib_fl_t fl; 2918 #if defined(INET6) || defined(INET) 2919 int lro_enabled; 2920 #endif 2921 uint8_t retval = 0; 2922 2923 /* 2924 * XXX early demux data packets so that if_input processing only handles 2925 * acks in interrupt context 2926 */ 2927 struct mbuf *m, *mh, *mt; 2928 2929 NET_EPOCH_ASSERT(); 2930 2931 ifp = ctx->ifc_ifp; 2932 mh = mt = NULL; 2933 MPASS(budget > 0); 2934 rx_pkts = rx_bytes = 0; 2935 if (sctx->isc_flags & IFLIB_HAS_RXCQ) 2936 cidxp = &rxq->ifr_cq_cidx; 2937 else 2938 cidxp = &rxq->ifr_fl[0].ifl_cidx; 2939 if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) { 2940 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) 2941 retval |= iflib_fl_refill_all(ctx, fl); 2942 DBG_COUNTER_INC(rx_unavail); 2943 return (retval); 2944 } 2945 2946 #if defined(INET6) || defined(INET) 2947 lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); 2948 #endif 2949 2950 /* pfil needs the vnet to be set */ 2951 CURVNET_SET_QUIET(if_getvnet(ifp)); 2952 for (budget_left = budget; budget_left > 0 && avail > 0;) { 2953 if (__predict_false(!CTX_ACTIVE(ctx))) { 2954 DBG_COUNTER_INC(rx_ctx_inactive); 2955 break; 2956 } 2957 /* 2958 * Reset client set fields to their default values 2959 */ 2960 rxd_info_zero(&ri); 2961 ri.iri_qsidx = rxq->ifr_id; 2962 ri.iri_cidx = *cidxp; 2963 ri.iri_ifp = ifp; 2964 ri.iri_frags = rxq->ifr_frags; 2965 err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); 2966 2967 if (err) 2968 goto err; 2969 rx_pkts += 1; 2970 rx_bytes += ri.iri_len; 2971 if (sctx->isc_flags & IFLIB_HAS_RXCQ) { 2972 *cidxp = ri.iri_cidx; 2973 /* Update our consumer index */ 2974 /* XXX NB: shurd - check if this is still safe */ 2975 while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) 2976 rxq->ifr_cq_cidx -= scctx->isc_nrxd[0]; 2977 /* was this only a completion queue message? */ 2978 if (__predict_false(ri.iri_nfrags == 0)) 2979 continue; 2980 } 2981 MPASS(ri.iri_nfrags != 0); 2982 MPASS(ri.iri_len != 0); 2983 2984 /* will advance the cidx on the corresponding free lists */ 2985 m = iflib_rxd_pkt_get(rxq, &ri); 2986 avail--; 2987 budget_left--; 2988 if (avail == 0 && budget_left) 2989 avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left); 2990 2991 if (__predict_false(m == NULL)) 2992 continue; 2993 2994 #ifndef __NO_STRICT_ALIGNMENT 2995 if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL) 2996 continue; 2997 #endif 2998 #if defined(INET6) || defined(INET) 2999 if (lro_enabled) { 3000 tcp_lro_queue_mbuf(&rxq->ifr_lc, m); 3001 continue; 3002 } 3003 #endif 3004 3005 if (mh == NULL) 3006 mh = mt = m; 3007 else { 3008 mt->m_nextpkt = m; 3009 mt = m; 3010 } 3011 } 3012 CURVNET_RESTORE(); 3013 /* make sure that we can refill faster than drain */ 3014 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) 3015 retval |= iflib_fl_refill_all(ctx, fl); 3016 3017 if (mh != NULL) { 3018 if_input(ifp, mh); 3019 DBG_COUNTER_INC(rx_if_input); 3020 } 3021 3022 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); 3023 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); 3024 3025 /* 3026 * Flush any outstanding LRO work 3027 */ 3028 #if defined(INET6) || defined(INET) 3029 tcp_lro_flush_all(&rxq->ifr_lc); 3030 #endif 3031 if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0) 3032 retval |= IFLIB_RXEOF_MORE; 3033 return (retval); 3034 err: 3035 STATE_LOCK(ctx); 3036 ctx->ifc_flags |= IFC_DO_RESET; 3037 iflib_admin_intr_deferred(ctx); 3038 STATE_UNLOCK(ctx); 3039 return (0); 3040 } 3041 3042 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1) 3043 static inline qidx_t 3044 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use) 3045 { 3046 qidx_t notify_count = TXD_NOTIFY_COUNT(txq); 3047 qidx_t minthresh = txq->ift_size / 8; 3048 if (in_use > 4 * minthresh) 3049 return (notify_count); 3050 if (in_use > 2 * minthresh) 3051 return (notify_count >> 1); 3052 if (in_use > minthresh) 3053 return (notify_count >> 3); 3054 return (0); 3055 } 3056 3057 static inline qidx_t 3058 txq_max_rs_deferred(iflib_txq_t txq) 3059 { 3060 qidx_t notify_count = TXD_NOTIFY_COUNT(txq); 3061 qidx_t minthresh = txq->ift_size / 8; 3062 if (txq->ift_in_use > 4 * minthresh) 3063 return (notify_count); 3064 if (txq->ift_in_use > 2 * minthresh) 3065 return (notify_count >> 1); 3066 if (txq->ift_in_use > minthresh) 3067 return (notify_count >> 2); 3068 return (2); 3069 } 3070 3071 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) 3072 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) 3073 3074 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use)) 3075 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq) 3076 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4) 3077 3078 /* forward compatibility for cxgb */ 3079 #define FIRST_QSET(ctx) 0 3080 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) 3081 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) 3082 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx)) 3083 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) 3084 3085 /* XXX we should be setting this to something other than zero */ 3086 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) 3087 #define MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \ 3088 (ctx)->ifc_softc_ctx.isc_tx_nsegments) 3089 3090 static inline bool 3091 iflib_txd_db_check(iflib_txq_t txq, int ring) 3092 { 3093 if_ctx_t ctx = txq->ift_ctx; 3094 qidx_t dbval, max; 3095 3096 max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use); 3097 3098 /* force || threshold exceeded || at the edge of the ring */ 3099 if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) { 3100 3101 /* 3102 * 'npending' is used if the card's doorbell is in terms of the number of descriptors 3103 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the 3104 * producer index explicitly (INTC). 3105 */ 3106 dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx; 3107 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 3108 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 3109 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval); 3110 3111 /* 3112 * Absent bugs there are zero packets pending so reset pending counts to zero. 3113 */ 3114 txq->ift_db_pending = txq->ift_npending = 0; 3115 return (true); 3116 } 3117 return (false); 3118 } 3119 3120 #ifdef PKT_DEBUG 3121 static void 3122 print_pkt(if_pkt_info_t pi) 3123 { 3124 printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n", 3125 pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx); 3126 printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n", 3127 pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag); 3128 printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n", 3129 pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto); 3130 } 3131 #endif 3132 3133 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO) 3134 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO)) 3135 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO) 3136 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO)) 3137 3138 /** 3139 * Parses out ethernet header information in the given mbuf. 3140 * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length) 3141 * 3142 * This will account for the VLAN header if present. 3143 * 3144 * XXX: This doesn't handle QinQ, which could prevent TX offloads for those 3145 * types of packets. 3146 */ 3147 static int 3148 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups) 3149 { 3150 struct ether_vlan_header *eh; 3151 struct mbuf *m; 3152 3153 m = *mp; 3154 if (__predict_false(m->m_len < sizeof(*eh))) { 3155 (*pullups)++; 3156 if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL)) 3157 return (ENOMEM); 3158 } 3159 eh = mtod(m, struct ether_vlan_header *); 3160 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { 3161 pi->ipi_etype = ntohs(eh->evl_proto); 3162 pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 3163 } else { 3164 pi->ipi_etype = ntohs(eh->evl_encap_proto); 3165 pi->ipi_ehdrlen = ETHER_HDR_LEN; 3166 } 3167 *mp = m; 3168 3169 return (0); 3170 } 3171 3172 /** 3173 * Parse up to the L3 header and extract IPv4/IPv6 header information into pi. 3174 * Currently this information includes: IP ToS value, IP header version/presence 3175 * 3176 * This is missing some checks and doesn't edit the packet content as it goes, 3177 * unlike iflib_parse_header(), in order to keep the amount of code here minimal. 3178 */ 3179 static int 3180 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups) 3181 { 3182 struct mbuf *m; 3183 int err; 3184 3185 *pullups = 0; 3186 m = *mp; 3187 if (!M_WRITABLE(m)) { 3188 if ((m = m_dup(m, M_NOWAIT)) == NULL) { 3189 return (ENOMEM); 3190 } else { 3191 m_freem(*mp); 3192 DBG_COUNTER_INC(tx_frees); 3193 *mp = m; 3194 } 3195 } 3196 3197 /* Fills out pi->ipi_etype */ 3198 err = iflib_parse_ether_header(pi, mp, pullups); 3199 if (err) 3200 return (err); 3201 m = *mp; 3202 3203 switch (pi->ipi_etype) { 3204 #ifdef INET 3205 case ETHERTYPE_IP: 3206 { 3207 struct mbuf *n; 3208 struct ip *ip = NULL; 3209 int miniplen; 3210 3211 miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip)); 3212 if (__predict_false(m->m_len < miniplen)) { 3213 /* 3214 * Check for common case where the first mbuf only contains 3215 * the Ethernet header 3216 */ 3217 if (m->m_len == pi->ipi_ehdrlen) { 3218 n = m->m_next; 3219 MPASS(n); 3220 /* If next mbuf contains at least the minimal IP header, then stop */ 3221 if (n->m_len >= sizeof(*ip)) { 3222 ip = (struct ip *)n->m_data; 3223 } else { 3224 (*pullups)++; 3225 if (__predict_false((m = m_pullup(m, miniplen)) == NULL)) 3226 return (ENOMEM); 3227 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 3228 } 3229 } else { 3230 (*pullups)++; 3231 if (__predict_false((m = m_pullup(m, miniplen)) == NULL)) 3232 return (ENOMEM); 3233 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 3234 } 3235 } else { 3236 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 3237 } 3238 3239 /* Have the IPv4 header w/ no options here */ 3240 pi->ipi_ip_hlen = ip->ip_hl << 2; 3241 pi->ipi_ipproto = ip->ip_p; 3242 pi->ipi_ip_tos = ip->ip_tos; 3243 pi->ipi_flags |= IPI_TX_IPV4; 3244 3245 break; 3246 } 3247 #endif 3248 #ifdef INET6 3249 case ETHERTYPE_IPV6: 3250 { 3251 struct ip6_hdr *ip6; 3252 3253 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { 3254 (*pullups)++; 3255 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) 3256 return (ENOMEM); 3257 } 3258 ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); 3259 3260 /* Have the IPv6 fixed header here */ 3261 pi->ipi_ip_hlen = sizeof(struct ip6_hdr); 3262 pi->ipi_ipproto = ip6->ip6_nxt; 3263 pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6); 3264 pi->ipi_flags |= IPI_TX_IPV6; 3265 3266 break; 3267 } 3268 #endif 3269 default: 3270 pi->ipi_csum_flags &= ~CSUM_OFFLOAD; 3271 pi->ipi_ip_hlen = 0; 3272 break; 3273 } 3274 *mp = m; 3275 3276 return (0); 3277 3278 } 3279 3280 static int 3281 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) 3282 { 3283 if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx; 3284 struct mbuf *m; 3285 int err; 3286 3287 m = *mp; 3288 if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) && 3289 M_WRITABLE(m) == 0) { 3290 if ((m = m_dup(m, M_NOWAIT)) == NULL) { 3291 return (ENOMEM); 3292 } else { 3293 m_freem(*mp); 3294 DBG_COUNTER_INC(tx_frees); 3295 *mp = m; 3296 } 3297 } 3298 3299 /* Fills out pi->ipi_etype */ 3300 err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups); 3301 if (__predict_false(err)) 3302 return (err); 3303 m = *mp; 3304 3305 switch (pi->ipi_etype) { 3306 #ifdef INET 3307 case ETHERTYPE_IP: 3308 { 3309 struct ip *ip; 3310 struct tcphdr *th; 3311 uint8_t hlen; 3312 3313 hlen = pi->ipi_ehdrlen + sizeof(*ip); 3314 if (__predict_false(m->m_len < hlen)) { 3315 txq->ift_pullups++; 3316 if (__predict_false((m = m_pullup(m, hlen)) == NULL)) 3317 return (ENOMEM); 3318 } 3319 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); 3320 hlen = pi->ipi_ehdrlen + (ip->ip_hl << 2); 3321 if (ip->ip_p == IPPROTO_TCP) { 3322 hlen += sizeof(*th); 3323 th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2)); 3324 } else if (ip->ip_p == IPPROTO_UDP) { 3325 hlen += sizeof(struct udphdr); 3326 } 3327 if (__predict_false(m->m_len < hlen)) { 3328 txq->ift_pullups++; 3329 if ((m = m_pullup(m, hlen)) == NULL) 3330 return (ENOMEM); 3331 } 3332 pi->ipi_ip_hlen = ip->ip_hl << 2; 3333 pi->ipi_ipproto = ip->ip_p; 3334 pi->ipi_ip_tos = ip->ip_tos; 3335 pi->ipi_flags |= IPI_TX_IPV4; 3336 3337 /* TCP checksum offload may require TCP header length */ 3338 if (IS_TX_OFFLOAD4(pi)) { 3339 if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) { 3340 pi->ipi_tcp_hflags = tcp_get_flags(th); 3341 pi->ipi_tcp_hlen = th->th_off << 2; 3342 pi->ipi_tcp_seq = th->th_seq; 3343 } 3344 if (IS_TSO4(pi)) { 3345 if (__predict_false(ip->ip_p != IPPROTO_TCP)) 3346 return (ENXIO); 3347 /* 3348 * TSO always requires hardware checksum offload. 3349 */ 3350 pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP); 3351 th->th_sum = in_pseudo(ip->ip_src.s_addr, 3352 ip->ip_dst.s_addr, htons(IPPROTO_TCP)); 3353 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; 3354 if (sctx->isc_flags & IFLIB_TSO_INIT_IP) { 3355 ip->ip_sum = 0; 3356 ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz); 3357 } 3358 } 3359 } 3360 if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP)) 3361 ip->ip_sum = 0; 3362 3363 break; 3364 } 3365 #endif 3366 #ifdef INET6 3367 case ETHERTYPE_IPV6: 3368 { 3369 struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); 3370 struct tcphdr *th; 3371 pi->ipi_ip_hlen = sizeof(struct ip6_hdr); 3372 3373 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { 3374 txq->ift_pullups++; 3375 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) 3376 return (ENOMEM); 3377 } 3378 th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen); 3379 3380 /* XXX-BZ this will go badly in case of ext hdrs. */ 3381 pi->ipi_ipproto = ip6->ip6_nxt; 3382 pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6); 3383 pi->ipi_flags |= IPI_TX_IPV6; 3384 3385 /* TCP checksum offload may require TCP header length */ 3386 if (IS_TX_OFFLOAD6(pi)) { 3387 if (pi->ipi_ipproto == IPPROTO_TCP) { 3388 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) { 3389 txq->ift_pullups++; 3390 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL)) 3391 return (ENOMEM); 3392 } 3393 pi->ipi_tcp_hflags = tcp_get_flags(th); 3394 pi->ipi_tcp_hlen = th->th_off << 2; 3395 pi->ipi_tcp_seq = th->th_seq; 3396 } 3397 if (IS_TSO6(pi)) { 3398 if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP)) 3399 return (ENXIO); 3400 /* 3401 * TSO always requires hardware checksum offload. 3402 */ 3403 pi->ipi_csum_flags |= CSUM_IP6_TCP; 3404 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); 3405 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; 3406 } 3407 } 3408 break; 3409 } 3410 #endif 3411 default: 3412 pi->ipi_csum_flags &= ~CSUM_OFFLOAD; 3413 pi->ipi_ip_hlen = 0; 3414 break; 3415 } 3416 *mp = m; 3417 3418 return (0); 3419 } 3420 3421 /* 3422 * If dodgy hardware rejects the scatter gather chain we've handed it 3423 * we'll need to remove the mbuf chain from ifsg_m[] before we can add the 3424 * m_defrag'd mbufs 3425 */ 3426 static __noinline struct mbuf * 3427 iflib_remove_mbuf(iflib_txq_t txq) 3428 { 3429 int ntxd, pidx; 3430 struct mbuf *m, **ifsd_m; 3431 3432 ifsd_m = txq->ift_sds.ifsd_m; 3433 ntxd = txq->ift_size; 3434 pidx = txq->ift_pidx & (ntxd - 1); 3435 ifsd_m = txq->ift_sds.ifsd_m; 3436 m = ifsd_m[pidx]; 3437 ifsd_m[pidx] = NULL; 3438 bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]); 3439 if (txq->ift_sds.ifsd_tso_map != NULL) 3440 bus_dmamap_unload(txq->ift_tso_buf_tag, 3441 txq->ift_sds.ifsd_tso_map[pidx]); 3442 #if MEMORY_LOGGING 3443 txq->ift_dequeued++; 3444 #endif 3445 return (m); 3446 } 3447 3448 static inline caddr_t 3449 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid) 3450 { 3451 qidx_t size; 3452 int ntxd; 3453 caddr_t start, end, cur, next; 3454 3455 ntxd = txq->ift_size; 3456 size = txq->ift_txd_size[qid]; 3457 start = txq->ift_ifdi[qid].idi_vaddr; 3458 3459 if (__predict_false(size == 0)) 3460 return (start); 3461 cur = start + size * cidx; 3462 end = start + size * ntxd; 3463 next = CACHE_PTR_NEXT(cur); 3464 return (next < end ? next : start); 3465 } 3466 3467 /* 3468 * Pad an mbuf to ensure a minimum ethernet frame size. 3469 * min_frame_size is the frame size (less CRC) to pad the mbuf to 3470 */ 3471 static __noinline int 3472 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size) 3473 { 3474 /* 3475 * 18 is enough bytes to pad an ARP packet to 46 bytes, and 3476 * and ARP message is the smallest common payload I can think of 3477 */ 3478 static char pad[18]; /* just zeros */ 3479 int n; 3480 struct mbuf *new_head; 3481 3482 if (!M_WRITABLE(*m_head)) { 3483 new_head = m_dup(*m_head, M_NOWAIT); 3484 if (new_head == NULL) { 3485 m_freem(*m_head); 3486 device_printf(dev, "cannot pad short frame, m_dup() failed"); 3487 DBG_COUNTER_INC(encap_pad_mbuf_fail); 3488 DBG_COUNTER_INC(tx_frees); 3489 return (ENOMEM); 3490 } 3491 m_freem(*m_head); 3492 *m_head = new_head; 3493 } 3494 3495 for (n = min_frame_size - (*m_head)->m_pkthdr.len; 3496 n > 0; n -= sizeof(pad)) 3497 if (!m_append(*m_head, min(n, sizeof(pad)), pad)) 3498 break; 3499 3500 if (n > 0) { 3501 m_freem(*m_head); 3502 device_printf(dev, "cannot pad short frame\n"); 3503 DBG_COUNTER_INC(encap_pad_mbuf_fail); 3504 DBG_COUNTER_INC(tx_frees); 3505 return (ENOBUFS); 3506 } 3507 3508 return (0); 3509 } 3510 3511 static int 3512 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp) 3513 { 3514 if_ctx_t ctx; 3515 if_shared_ctx_t sctx; 3516 if_softc_ctx_t scctx; 3517 bus_dma_tag_t buf_tag; 3518 bus_dma_segment_t *segs; 3519 struct mbuf *m_head, **ifsd_m; 3520 void *next_txd; 3521 bus_dmamap_t map; 3522 struct if_pkt_info pi; 3523 int remap = 0; 3524 int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd; 3525 3526 ctx = txq->ift_ctx; 3527 sctx = ctx->ifc_sctx; 3528 scctx = &ctx->ifc_softc_ctx; 3529 segs = txq->ift_segs; 3530 ntxd = txq->ift_size; 3531 m_head = *m_headp; 3532 map = NULL; 3533 3534 /* 3535 * If we're doing TSO the next descriptor to clean may be quite far ahead 3536 */ 3537 cidx = txq->ift_cidx; 3538 pidx = txq->ift_pidx; 3539 if (ctx->ifc_flags & IFC_PREFETCH) { 3540 next = (cidx + CACHE_PTR_INCREMENT) & (ntxd - 1); 3541 if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) { 3542 next_txd = calc_next_txd(txq, cidx, 0); 3543 prefetch(next_txd); 3544 } 3545 3546 /* prefetch the next cache line of mbuf pointers and flags */ 3547 prefetch(&txq->ift_sds.ifsd_m[next]); 3548 prefetch(&txq->ift_sds.ifsd_map[next]); 3549 next = (cidx + CACHE_LINE_SIZE) & (ntxd - 1); 3550 } 3551 map = txq->ift_sds.ifsd_map[pidx]; 3552 ifsd_m = txq->ift_sds.ifsd_m; 3553 3554 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) { 3555 buf_tag = txq->ift_tso_buf_tag; 3556 max_segs = scctx->isc_tx_tso_segments_max; 3557 map = txq->ift_sds.ifsd_tso_map[pidx]; 3558 MPASS(buf_tag != NULL); 3559 MPASS(max_segs > 0); 3560 } else { 3561 buf_tag = txq->ift_buf_tag; 3562 max_segs = scctx->isc_tx_nsegments; 3563 map = txq->ift_sds.ifsd_map[pidx]; 3564 } 3565 if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) && 3566 __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) { 3567 err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size); 3568 if (err) { 3569 DBG_COUNTER_INC(encap_txd_encap_fail); 3570 return (err); 3571 } 3572 } 3573 m_head = *m_headp; 3574 3575 pkt_info_zero(&pi); 3576 pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST)); 3577 pi.ipi_pidx = pidx; 3578 pi.ipi_qsidx = txq->ift_id; 3579 pi.ipi_len = m_head->m_pkthdr.len; 3580 pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags; 3581 pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0; 3582 3583 /* deliberate bitwise OR to make one condition */ 3584 if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) { 3585 if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) { 3586 DBG_COUNTER_INC(encap_txd_encap_fail); 3587 return (err); 3588 } 3589 m_head = *m_headp; 3590 } 3591 3592 retry: 3593 err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs, 3594 BUS_DMA_NOWAIT); 3595 defrag: 3596 if (__predict_false(err)) { 3597 switch (err) { 3598 case EFBIG: 3599 /* try collapse once and defrag once */ 3600 if (remap == 0) { 3601 m_head = m_collapse(*m_headp, M_NOWAIT, max_segs); 3602 /* try defrag if collapsing fails */ 3603 if (m_head == NULL) 3604 remap++; 3605 } 3606 if (remap == 1) { 3607 txq->ift_mbuf_defrag++; 3608 m_head = m_defrag(*m_headp, M_NOWAIT); 3609 } 3610 /* 3611 * remap should never be >1 unless bus_dmamap_load_mbuf_sg 3612 * failed to map an mbuf that was run through m_defrag 3613 */ 3614 MPASS(remap <= 1); 3615 if (__predict_false(m_head == NULL || remap > 1)) 3616 goto defrag_failed; 3617 remap++; 3618 *m_headp = m_head; 3619 goto retry; 3620 break; 3621 case ENOMEM: 3622 txq->ift_no_tx_dma_setup++; 3623 break; 3624 default: 3625 txq->ift_no_tx_dma_setup++; 3626 m_freem(*m_headp); 3627 DBG_COUNTER_INC(tx_frees); 3628 *m_headp = NULL; 3629 break; 3630 } 3631 txq->ift_map_failed++; 3632 DBG_COUNTER_INC(encap_load_mbuf_fail); 3633 DBG_COUNTER_INC(encap_txd_encap_fail); 3634 return (err); 3635 } 3636 ifsd_m[pidx] = m_head; 3637 /* 3638 * XXX assumes a 1 to 1 relationship between segments and 3639 * descriptors - this does not hold true on all drivers, e.g. 3640 * cxgb 3641 */ 3642 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { 3643 (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 3644 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { 3645 txq->ift_no_desc_avail++; 3646 bus_dmamap_unload(buf_tag, map); 3647 DBG_COUNTER_INC(encap_txq_avail_fail); 3648 DBG_COUNTER_INC(encap_txd_encap_fail); 3649 if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0) 3650 GROUPTASK_ENQUEUE(&txq->ift_task); 3651 return (ENOBUFS); 3652 } 3653 } 3654 /* 3655 * On Intel cards we can greatly reduce the number of TX interrupts 3656 * we see by only setting report status on every Nth descriptor. 3657 * However, this also means that the driver will need to keep track 3658 * of the descriptors that RS was set on to check them for the DD bit. 3659 */ 3660 txq->ift_rs_pending += nsegs + 1; 3661 if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) || 3662 iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) { 3663 pi.ipi_flags |= IPI_TX_INTR; 3664 txq->ift_rs_pending = 0; 3665 } 3666 3667 pi.ipi_segs = segs; 3668 pi.ipi_nsegs = nsegs; 3669 3670 MPASS(pidx >= 0 && pidx < txq->ift_size); 3671 #ifdef PKT_DEBUG 3672 print_pkt(&pi); 3673 #endif 3674 if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) { 3675 bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE); 3676 DBG_COUNTER_INC(tx_encap); 3677 MPASS(pi.ipi_new_pidx < txq->ift_size); 3678 3679 ndesc = pi.ipi_new_pidx - pi.ipi_pidx; 3680 if (pi.ipi_new_pidx < pi.ipi_pidx) { 3681 ndesc += txq->ift_size; 3682 txq->ift_gen = 1; 3683 } 3684 /* 3685 * drivers can need as many as 3686 * two sentinels 3687 */ 3688 MPASS(ndesc <= pi.ipi_nsegs + 2); 3689 MPASS(pi.ipi_new_pidx != pidx); 3690 MPASS(ndesc > 0); 3691 txq->ift_in_use += ndesc; 3692 txq->ift_db_pending += ndesc; 3693 3694 /* 3695 * We update the last software descriptor again here because there may 3696 * be a sentinel and/or there may be more mbufs than segments 3697 */ 3698 txq->ift_pidx = pi.ipi_new_pidx; 3699 txq->ift_npending += pi.ipi_ndescs; 3700 } else { 3701 *m_headp = m_head = iflib_remove_mbuf(txq); 3702 if (err == EFBIG) { 3703 txq->ift_txd_encap_efbig++; 3704 if (remap < 2) { 3705 remap = 1; 3706 goto defrag; 3707 } 3708 } 3709 goto defrag_failed; 3710 } 3711 /* 3712 * err can't possibly be non-zero here, so we don't neet to test it 3713 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail). 3714 */ 3715 return (err); 3716 3717 defrag_failed: 3718 txq->ift_mbuf_defrag_failed++; 3719 txq->ift_map_failed++; 3720 m_freem(*m_headp); 3721 DBG_COUNTER_INC(tx_frees); 3722 *m_headp = NULL; 3723 DBG_COUNTER_INC(encap_txd_encap_fail); 3724 return (ENOMEM); 3725 } 3726 3727 static void 3728 iflib_tx_desc_free(iflib_txq_t txq, int n) 3729 { 3730 uint32_t qsize, cidx, mask, gen; 3731 struct mbuf *m, **ifsd_m; 3732 bool do_prefetch; 3733 3734 cidx = txq->ift_cidx; 3735 gen = txq->ift_gen; 3736 qsize = txq->ift_size; 3737 mask = qsize - 1; 3738 ifsd_m = txq->ift_sds.ifsd_m; 3739 do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH); 3740 3741 while (n-- > 0) { 3742 if (do_prefetch) { 3743 prefetch(ifsd_m[(cidx + 3) & mask]); 3744 prefetch(ifsd_m[(cidx + 4) & mask]); 3745 } 3746 if ((m = ifsd_m[cidx]) != NULL) { 3747 prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]); 3748 if (m->m_pkthdr.csum_flags & CSUM_TSO) { 3749 bus_dmamap_sync(txq->ift_tso_buf_tag, 3750 txq->ift_sds.ifsd_tso_map[cidx], 3751 BUS_DMASYNC_POSTWRITE); 3752 bus_dmamap_unload(txq->ift_tso_buf_tag, 3753 txq->ift_sds.ifsd_tso_map[cidx]); 3754 } else { 3755 bus_dmamap_sync(txq->ift_buf_tag, 3756 txq->ift_sds.ifsd_map[cidx], 3757 BUS_DMASYNC_POSTWRITE); 3758 bus_dmamap_unload(txq->ift_buf_tag, 3759 txq->ift_sds.ifsd_map[cidx]); 3760 } 3761 /* XXX we don't support any drivers that batch packets yet */ 3762 MPASS(m->m_nextpkt == NULL); 3763 m_freem(m); 3764 ifsd_m[cidx] = NULL; 3765 #if MEMORY_LOGGING 3766 txq->ift_dequeued++; 3767 #endif 3768 DBG_COUNTER_INC(tx_frees); 3769 } 3770 if (__predict_false(++cidx == qsize)) { 3771 cidx = 0; 3772 gen = 0; 3773 } 3774 } 3775 txq->ift_cidx = cidx; 3776 txq->ift_gen = gen; 3777 } 3778 3779 static __inline int 3780 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh) 3781 { 3782 int reclaim; 3783 if_ctx_t ctx = txq->ift_ctx; 3784 3785 KASSERT(thresh >= 0, ("invalid threshold to reclaim")); 3786 MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size); 3787 3788 /* 3789 * Need a rate-limiting check so that this isn't called every time 3790 */ 3791 iflib_tx_credits_update(ctx, txq); 3792 reclaim = DESC_RECLAIMABLE(txq); 3793 3794 if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) { 3795 #ifdef INVARIANTS 3796 if (iflib_verbose_debug) { 3797 printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__, 3798 txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments, 3799 reclaim, thresh); 3800 } 3801 #endif 3802 return (0); 3803 } 3804 iflib_tx_desc_free(txq, reclaim); 3805 txq->ift_cleaned += reclaim; 3806 txq->ift_in_use -= reclaim; 3807 3808 return (reclaim); 3809 } 3810 3811 static struct mbuf ** 3812 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining) 3813 { 3814 int next, size; 3815 struct mbuf **items; 3816 3817 size = r->size; 3818 next = (cidx + CACHE_PTR_INCREMENT) & (size - 1); 3819 items = __DEVOLATILE(struct mbuf **, &r->items[0]); 3820 3821 prefetch(items[(cidx + offset) & (size - 1)]); 3822 if (remaining > 1) { 3823 prefetch2cachelines(&items[next]); 3824 prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]); 3825 prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]); 3826 prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]); 3827 } 3828 return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)])); 3829 } 3830 3831 static void 3832 iflib_txq_check_drain(iflib_txq_t txq, int budget) 3833 { 3834 3835 ifmp_ring_check_drainage(txq->ift_br, budget); 3836 } 3837 3838 static uint32_t 3839 iflib_txq_can_drain(struct ifmp_ring *r) 3840 { 3841 iflib_txq_t txq = r->cookie; 3842 if_ctx_t ctx = txq->ift_ctx; 3843 3844 if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) 3845 return (1); 3846 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 3847 BUS_DMASYNC_POSTREAD); 3848 return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, 3849 false)); 3850 } 3851 3852 static uint32_t 3853 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) 3854 { 3855 iflib_txq_t txq = r->cookie; 3856 if_ctx_t ctx = txq->ift_ctx; 3857 if_t ifp = ctx->ifc_ifp; 3858 struct mbuf *m, **mp; 3859 int avail, bytes_sent, skipped, count, err, i; 3860 int mcast_sent, pkt_sent, reclaimed; 3861 bool do_prefetch, rang, ring; 3862 3863 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) || 3864 !LINK_ACTIVE(ctx))) { 3865 DBG_COUNTER_INC(txq_drain_notready); 3866 return (0); 3867 } 3868 reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 3869 rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending); 3870 avail = IDXDIFF(pidx, cidx, r->size); 3871 3872 if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { 3873 /* 3874 * The driver is unloading so we need to free all pending packets. 3875 */ 3876 DBG_COUNTER_INC(txq_drain_flushing); 3877 for (i = 0; i < avail; i++) { 3878 if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq)) 3879 m_freem(r->items[(cidx + i) & (r->size - 1)]); 3880 r->items[(cidx + i) & (r->size - 1)] = NULL; 3881 } 3882 return (avail); 3883 } 3884 3885 if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) { 3886 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 3887 CALLOUT_LOCK(txq); 3888 callout_stop(&txq->ift_timer); 3889 CALLOUT_UNLOCK(txq); 3890 DBG_COUNTER_INC(txq_drain_oactive); 3891 return (0); 3892 } 3893 3894 /* 3895 * If we've reclaimed any packets this queue cannot be hung. 3896 */ 3897 if (reclaimed) 3898 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 3899 skipped = mcast_sent = bytes_sent = pkt_sent = 0; 3900 count = MIN(avail, TX_BATCH_SIZE); 3901 #ifdef INVARIANTS 3902 if (iflib_verbose_debug) 3903 printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__, 3904 avail, ctx->ifc_flags, TXQ_AVAIL(txq)); 3905 #endif 3906 do_prefetch = (ctx->ifc_flags & IFC_PREFETCH); 3907 err = 0; 3908 for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) { 3909 int rem = do_prefetch ? count - i : 0; 3910 3911 mp = _ring_peek_one(r, cidx, i, rem); 3912 MPASS(mp != NULL && *mp != NULL); 3913 3914 /* 3915 * Completion interrupts will use the address of the txq 3916 * as a sentinel to enqueue _something_ in order to acquire 3917 * the lock on the mp_ring (there's no direct lock call). 3918 * We obviously whave to check for these sentinel cases 3919 * and skip them. 3920 */ 3921 if (__predict_false(*mp == (struct mbuf *)txq)) { 3922 skipped++; 3923 continue; 3924 } 3925 err = iflib_encap(txq, mp); 3926 if (__predict_false(err)) { 3927 /* no room - bail out */ 3928 if (err == ENOBUFS) 3929 break; 3930 skipped++; 3931 /* we can't send this packet - skip it */ 3932 continue; 3933 } 3934 pkt_sent++; 3935 m = *mp; 3936 DBG_COUNTER_INC(tx_sent); 3937 bytes_sent += m->m_pkthdr.len; 3938 mcast_sent += !!(m->m_flags & M_MCAST); 3939 3940 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))) 3941 break; 3942 ETHER_BPF_MTAP(ifp, m); 3943 rang = iflib_txd_db_check(txq, false); 3944 } 3945 3946 /* deliberate use of bitwise or to avoid gratuitous short-circuit */ 3947 ring = rang ? false : (iflib_min_tx_latency | err); 3948 iflib_txd_db_check(txq, ring); 3949 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); 3950 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); 3951 if (mcast_sent) 3952 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); 3953 #ifdef INVARIANTS 3954 if (iflib_verbose_debug) 3955 printf("consumed=%d\n", skipped + pkt_sent); 3956 #endif 3957 return (skipped + pkt_sent); 3958 } 3959 3960 static uint32_t 3961 iflib_txq_drain_always(struct ifmp_ring *r) 3962 { 3963 return (1); 3964 } 3965 3966 static uint32_t 3967 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) 3968 { 3969 int i, avail; 3970 struct mbuf **mp; 3971 iflib_txq_t txq; 3972 3973 txq = r->cookie; 3974 3975 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 3976 CALLOUT_LOCK(txq); 3977 callout_stop(&txq->ift_timer); 3978 CALLOUT_UNLOCK(txq); 3979 3980 avail = IDXDIFF(pidx, cidx, r->size); 3981 for (i = 0; i < avail; i++) { 3982 mp = _ring_peek_one(r, cidx, i, avail - i); 3983 if (__predict_false(*mp == (struct mbuf *)txq)) 3984 continue; 3985 m_freem(*mp); 3986 DBG_COUNTER_INC(tx_frees); 3987 } 3988 MPASS(ifmp_ring_is_stalled(r) == 0); 3989 return (avail); 3990 } 3991 3992 static void 3993 iflib_ifmp_purge(iflib_txq_t txq) 3994 { 3995 struct ifmp_ring *r; 3996 3997 r = txq->ift_br; 3998 r->drain = iflib_txq_drain_free; 3999 r->can_drain = iflib_txq_drain_always; 4000 4001 ifmp_ring_check_drainage(r, r->size); 4002 4003 r->drain = iflib_txq_drain; 4004 r->can_drain = iflib_txq_can_drain; 4005 } 4006 4007 static void 4008 _task_fn_tx(void *context) 4009 { 4010 iflib_txq_t txq = context; 4011 if_ctx_t ctx = txq->ift_ctx; 4012 if_t ifp = ctx->ifc_ifp; 4013 int abdicate = ctx->ifc_sysctl_tx_abdicate; 4014 4015 #ifdef IFLIB_DIAGNOSTICS 4016 txq->ift_cpu_exec_count[curcpu]++; 4017 #endif 4018 if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) 4019 return; 4020 #ifdef DEV_NETMAP 4021 if ((if_getcapenable(ifp) & IFCAP_NETMAP) && 4022 netmap_tx_irq(ifp, txq->ift_id)) 4023 goto skip_ifmp; 4024 #endif 4025 if (ctx->ifc_sysctl_simple_tx) { 4026 mtx_lock(&txq->ift_mtx); 4027 (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 4028 mtx_unlock(&txq->ift_mtx); 4029 goto skip_ifmp; 4030 } 4031 #ifdef ALTQ 4032 if (if_altq_is_enabled(ifp)) 4033 iflib_altq_if_start(ifp); 4034 #endif 4035 if (txq->ift_db_pending) 4036 ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate); 4037 else if (!abdicate) 4038 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); 4039 /* 4040 * When abdicating, we always need to check drainage, not just when we don't enqueue 4041 */ 4042 if (abdicate) 4043 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); 4044 4045 skip_ifmp: 4046 if (ctx->ifc_flags & IFC_LEGACY) 4047 IFDI_INTR_ENABLE(ctx); 4048 else 4049 IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id); 4050 } 4051 4052 static void 4053 _task_fn_rx(void *context) 4054 { 4055 iflib_rxq_t rxq = context; 4056 if_ctx_t ctx = rxq->ifr_ctx; 4057 uint8_t more; 4058 uint16_t budget; 4059 #ifdef DEV_NETMAP 4060 u_int work = 0; 4061 int nmirq; 4062 #endif 4063 4064 #ifdef IFLIB_DIAGNOSTICS 4065 rxq->ifr_cpu_exec_count[curcpu]++; 4066 #endif 4067 DBG_COUNTER_INC(task_fn_rxs); 4068 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) 4069 return; 4070 #ifdef DEV_NETMAP 4071 nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work); 4072 if (nmirq != NM_IRQ_PASS) { 4073 more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0; 4074 goto skip_rxeof; 4075 } 4076 #endif 4077 budget = ctx->ifc_sysctl_rx_budget; 4078 if (budget == 0) 4079 budget = 16; /* XXX */ 4080 more = iflib_rxeof(rxq, budget); 4081 #ifdef DEV_NETMAP 4082 skip_rxeof: 4083 #endif 4084 if ((more & IFLIB_RXEOF_MORE) == 0) { 4085 if (ctx->ifc_flags & IFC_LEGACY) 4086 IFDI_INTR_ENABLE(ctx); 4087 else 4088 IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); 4089 DBG_COUNTER_INC(rx_intr_enables); 4090 } 4091 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) 4092 return; 4093 4094 if (more & IFLIB_RXEOF_MORE) 4095 GROUPTASK_ENQUEUE(&rxq->ifr_task); 4096 else if (more & IFLIB_RXEOF_EMPTY) 4097 callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq); 4098 } 4099 4100 static void 4101 _task_fn_admin(void *context, int pending) 4102 { 4103 if_ctx_t ctx = context; 4104 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; 4105 iflib_txq_t txq; 4106 int i; 4107 bool oactive, running, do_reset, do_watchdog, in_detach; 4108 4109 STATE_LOCK(ctx); 4110 running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING); 4111 oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE); 4112 do_reset = (ctx->ifc_flags & IFC_DO_RESET); 4113 do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG); 4114 in_detach = (ctx->ifc_flags & IFC_IN_DETACH); 4115 ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_WATCHDOG); 4116 STATE_UNLOCK(ctx); 4117 4118 if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) 4119 return; 4120 if (in_detach) 4121 return; 4122 4123 CTX_LOCK(ctx); 4124 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { 4125 CALLOUT_LOCK(txq); 4126 callout_stop(&txq->ift_timer); 4127 CALLOUT_UNLOCK(txq); 4128 } 4129 if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ) 4130 IFDI_ADMIN_COMPLETION_HANDLE(ctx); 4131 if (do_watchdog) { 4132 ctx->ifc_watchdog_events++; 4133 IFDI_WATCHDOG_RESET(ctx); 4134 } 4135 IFDI_UPDATE_ADMIN_STATUS(ctx); 4136 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { 4137 callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq, 4138 txq->ift_timer.c_cpu); 4139 } 4140 IFDI_LINK_INTR_ENABLE(ctx); 4141 if (do_reset) 4142 iflib_if_init_locked(ctx); 4143 CTX_UNLOCK(ctx); 4144 4145 if (LINK_ACTIVE(ctx) == 0) 4146 return; 4147 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) 4148 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); 4149 } 4150 4151 static void 4152 _task_fn_iov(void *context, int pending) 4153 { 4154 if_ctx_t ctx = context; 4155 4156 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) && 4157 !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) 4158 return; 4159 4160 CTX_LOCK(ctx); 4161 IFDI_VFLR_HANDLE(ctx); 4162 CTX_UNLOCK(ctx); 4163 } 4164 4165 static int 4166 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS) 4167 { 4168 int err; 4169 if_int_delay_info_t info; 4170 if_ctx_t ctx; 4171 4172 info = (if_int_delay_info_t)arg1; 4173 ctx = info->iidi_ctx; 4174 info->iidi_req = req; 4175 info->iidi_oidp = oidp; 4176 CTX_LOCK(ctx); 4177 err = IFDI_SYSCTL_INT_DELAY(ctx, info); 4178 CTX_UNLOCK(ctx); 4179 return (err); 4180 } 4181 4182 /********************************************************************* 4183 * 4184 * IFNET FUNCTIONS 4185 * 4186 **********************************************************************/ 4187 4188 static void 4189 iflib_if_init_locked(if_ctx_t ctx) 4190 { 4191 iflib_stop(ctx); 4192 iflib_init_locked(ctx); 4193 } 4194 4195 static void 4196 iflib_if_init(void *arg) 4197 { 4198 if_ctx_t ctx = arg; 4199 4200 CTX_LOCK(ctx); 4201 iflib_if_init_locked(ctx); 4202 CTX_UNLOCK(ctx); 4203 } 4204 4205 static int 4206 iflib_if_transmit(if_t ifp, struct mbuf *m) 4207 { 4208 if_ctx_t ctx = if_getsoftc(ifp); 4209 iflib_txq_t txq; 4210 int err, qidx; 4211 int abdicate; 4212 4213 if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { 4214 DBG_COUNTER_INC(tx_frees); 4215 m_freem(m); 4216 return (ENETDOWN); 4217 } 4218 4219 MPASS(m->m_nextpkt == NULL); 4220 /* ALTQ-enabled interfaces always use queue 0. */ 4221 qidx = 0; 4222 /* Use driver-supplied queue selection method if it exists */ 4223 if (ctx->isc_txq_select_v2) { 4224 struct if_pkt_info pi; 4225 uint64_t early_pullups = 0; 4226 pkt_info_zero(&pi); 4227 4228 err = iflib_parse_header_partial(&pi, &m, &early_pullups); 4229 if (__predict_false(err != 0)) { 4230 /* Assign pullups for bad pkts to default queue */ 4231 ctx->ifc_txqs[0].ift_pullups += early_pullups; 4232 DBG_COUNTER_INC(encap_txd_encap_fail); 4233 return (err); 4234 } 4235 /* Let driver make queueing decision */ 4236 qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi); 4237 ctx->ifc_txqs[qidx].ift_pullups += early_pullups; 4238 } 4239 /* Backwards compatibility w/ simpler queue select */ 4240 else if (ctx->isc_txq_select) 4241 qidx = ctx->isc_txq_select(ctx->ifc_softc, m); 4242 /* If not, use iflib's standard method */ 4243 else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp)) 4244 qidx = QIDX(ctx, m); 4245 4246 /* Set TX queue */ 4247 txq = &ctx->ifc_txqs[qidx]; 4248 4249 #ifdef DRIVER_BACKPRESSURE 4250 if (txq->ift_closed) { 4251 while (m != NULL) { 4252 next = m->m_nextpkt; 4253 m->m_nextpkt = NULL; 4254 m_freem(m); 4255 DBG_COUNTER_INC(tx_frees); 4256 m = next; 4257 } 4258 return (ENOBUFS); 4259 } 4260 #endif 4261 #ifdef notyet 4262 qidx = count = 0; 4263 mp = marr; 4264 next = m; 4265 do { 4266 count++; 4267 next = next->m_nextpkt; 4268 } while (next != NULL); 4269 4270 if (count > nitems(marr)) 4271 if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) { 4272 /* XXX check nextpkt */ 4273 m_freem(m); 4274 /* XXX simplify for now */ 4275 DBG_COUNTER_INC(tx_frees); 4276 return (ENOBUFS); 4277 } 4278 for (next = m, i = 0; next != NULL; i++) { 4279 mp[i] = next; 4280 next = next->m_nextpkt; 4281 mp[i]->m_nextpkt = NULL; 4282 } 4283 #endif 4284 DBG_COUNTER_INC(tx_seen); 4285 abdicate = ctx->ifc_sysctl_tx_abdicate; 4286 4287 err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate); 4288 4289 if (abdicate) 4290 GROUPTASK_ENQUEUE(&txq->ift_task); 4291 if (err) { 4292 if (!abdicate) 4293 GROUPTASK_ENQUEUE(&txq->ift_task); 4294 /* support forthcoming later */ 4295 #ifdef DRIVER_BACKPRESSURE 4296 txq->ift_closed = TRUE; 4297 #endif 4298 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); 4299 m_freem(m); 4300 DBG_COUNTER_INC(tx_frees); 4301 } 4302 4303 return (err); 4304 } 4305 4306 #ifdef ALTQ 4307 /* 4308 * The overall approach to integrating iflib with ALTQ is to continue to use 4309 * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware 4310 * ring. Technically, when using ALTQ, queueing to an intermediate mp_ring 4311 * is redundant/unnecessary, but doing so minimizes the amount of 4312 * ALTQ-specific code required in iflib. It is assumed that the overhead of 4313 * redundantly queueing to an intermediate mp_ring is swamped by the 4314 * performance limitations inherent in using ALTQ. 4315 * 4316 * When ALTQ support is compiled in, all iflib drivers will use a transmit 4317 * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the 4318 * given interface. If ALTQ is enabled for an interface, then all 4319 * transmitted packets for that interface will be submitted to the ALTQ 4320 * subsystem via IFQ_ENQUEUE(). We don't use the legacy if_transmit() 4321 * implementation because it uses IFQ_HANDOFF(), which will duplicatively 4322 * update stats that the iflib machinery handles, and which is sensitve to 4323 * the disused IFF_DRV_OACTIVE flag. Additionally, iflib_altq_if_start() 4324 * will be installed as the start routine for use by ALTQ facilities that 4325 * need to trigger queue drains on a scheduled basis. 4326 * 4327 */ 4328 static void 4329 iflib_altq_if_start(if_t ifp) 4330 { 4331 struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */ 4332 struct mbuf *m; 4333 4334 IFQ_LOCK(ifq); 4335 IFQ_DEQUEUE_NOLOCK(ifq, m); 4336 while (m != NULL) { 4337 iflib_if_transmit(ifp, m); 4338 IFQ_DEQUEUE_NOLOCK(ifq, m); 4339 } 4340 IFQ_UNLOCK(ifq); 4341 } 4342 4343 static int 4344 iflib_altq_if_transmit(if_t ifp, struct mbuf *m) 4345 { 4346 int err; 4347 4348 if (if_altq_is_enabled(ifp)) { 4349 IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */ 4350 if (err == 0) 4351 iflib_altq_if_start(ifp); 4352 } else 4353 err = iflib_if_transmit(ifp, m); 4354 4355 return (err); 4356 } 4357 #endif /* ALTQ */ 4358 4359 static void 4360 iflib_if_qflush(if_t ifp) 4361 { 4362 if_ctx_t ctx = if_getsoftc(ifp); 4363 iflib_txq_t txq = ctx->ifc_txqs; 4364 int i; 4365 4366 STATE_LOCK(ctx); 4367 ctx->ifc_flags |= IFC_QFLUSH; 4368 STATE_UNLOCK(ctx); 4369 for (i = 0; i < NTXQSETS(ctx); i++, txq++) 4370 while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br))) 4371 iflib_txq_check_drain(txq, 0); 4372 STATE_LOCK(ctx); 4373 ctx->ifc_flags &= ~IFC_QFLUSH; 4374 STATE_UNLOCK(ctx); 4375 4376 /* 4377 * When ALTQ is enabled, this will also take care of purging the 4378 * ALTQ queue(s). 4379 */ 4380 if_qflush(ifp); 4381 } 4382 4383 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ 4384 IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \ 4385 IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \ 4386 IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG) 4387 4388 static int 4389 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) 4390 { 4391 if_ctx_t ctx = if_getsoftc(ifp); 4392 struct ifreq *ifr = (struct ifreq *)data; 4393 #if defined(INET) || defined(INET6) 4394 struct ifaddr *ifa = (struct ifaddr *)data; 4395 #endif 4396 bool avoid_reset = false; 4397 int err = 0, reinit = 0, bits; 4398 4399 switch (command) { 4400 case SIOCSIFADDR: 4401 #ifdef INET 4402 if (ifa->ifa_addr->sa_family == AF_INET) 4403 avoid_reset = true; 4404 #endif 4405 #ifdef INET6 4406 if (ifa->ifa_addr->sa_family == AF_INET6) 4407 avoid_reset = true; 4408 #endif 4409 /* 4410 * Calling init results in link renegotiation, 4411 * so we avoid doing it when possible. 4412 */ 4413 if (avoid_reset) { 4414 if_setflagbits(ifp, IFF_UP, 0); 4415 if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) 4416 reinit = 1; 4417 #ifdef INET 4418 if (!(if_getflags(ifp) & IFF_NOARP)) 4419 arp_ifinit(ifp, ifa); 4420 #endif 4421 } else 4422 err = ether_ioctl(ifp, command, data); 4423 break; 4424 case SIOCSIFMTU: 4425 CTX_LOCK(ctx); 4426 if (ifr->ifr_mtu == if_getmtu(ifp)) { 4427 CTX_UNLOCK(ctx); 4428 break; 4429 } 4430 bits = if_getdrvflags(ifp); 4431 /* stop the driver and free any clusters before proceeding */ 4432 iflib_stop(ctx); 4433 4434 if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) { 4435 STATE_LOCK(ctx); 4436 if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size) 4437 ctx->ifc_flags |= IFC_MULTISEG; 4438 else 4439 ctx->ifc_flags &= ~IFC_MULTISEG; 4440 STATE_UNLOCK(ctx); 4441 err = if_setmtu(ifp, ifr->ifr_mtu); 4442 } 4443 iflib_init_locked(ctx); 4444 STATE_LOCK(ctx); 4445 if_setdrvflags(ifp, bits); 4446 STATE_UNLOCK(ctx); 4447 CTX_UNLOCK(ctx); 4448 break; 4449 case SIOCSIFFLAGS: 4450 CTX_LOCK(ctx); 4451 if (if_getflags(ifp) & IFF_UP) { 4452 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 4453 if ((if_getflags(ifp) ^ ctx->ifc_if_flags) & 4454 (IFF_PROMISC | IFF_ALLMULTI)) { 4455 CTX_UNLOCK(ctx); 4456 err = IFDI_PROMISC_SET(ctx, if_getflags(ifp)); 4457 CTX_LOCK(ctx); 4458 } 4459 } else 4460 reinit = 1; 4461 } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 4462 iflib_stop(ctx); 4463 } 4464 ctx->ifc_if_flags = if_getflags(ifp); 4465 CTX_UNLOCK(ctx); 4466 break; 4467 case SIOCADDMULTI: 4468 case SIOCDELMULTI: 4469 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 4470 CTX_LOCK(ctx); 4471 IFDI_INTR_DISABLE(ctx); 4472 IFDI_MULTI_SET(ctx); 4473 IFDI_INTR_ENABLE(ctx); 4474 CTX_UNLOCK(ctx); 4475 } 4476 break; 4477 case SIOCSIFMEDIA: 4478 CTX_LOCK(ctx); 4479 IFDI_MEDIA_SET(ctx); 4480 CTX_UNLOCK(ctx); 4481 /* FALLTHROUGH */ 4482 case SIOCGIFMEDIA: 4483 case SIOCGIFXMEDIA: 4484 err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command); 4485 break; 4486 case SIOCGI2C: 4487 { 4488 struct ifi2creq i2c; 4489 4490 err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c)); 4491 if (err != 0) 4492 break; 4493 if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { 4494 err = EINVAL; 4495 break; 4496 } 4497 if (i2c.len > sizeof(i2c.data)) { 4498 err = EINVAL; 4499 break; 4500 } 4501 4502 if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0) 4503 err = copyout(&i2c, ifr_data_get_ptr(ifr), 4504 sizeof(i2c)); 4505 break; 4506 } 4507 case SIOCSIFCAP: 4508 { 4509 int mask, setmask, oldmask; 4510 4511 oldmask = if_getcapenable(ifp); 4512 mask = ifr->ifr_reqcap ^ oldmask; 4513 mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG; 4514 setmask = 0; 4515 #ifdef TCP_OFFLOAD 4516 setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6); 4517 #endif 4518 setmask |= (mask & IFCAP_FLAGS); 4519 setmask |= (mask & IFCAP_WOL); 4520 4521 /* 4522 * If any RX csum has changed, change all the ones that 4523 * are supported by the driver. 4524 */ 4525 if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) { 4526 setmask |= ctx->ifc_softc_ctx.isc_capabilities & 4527 (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6); 4528 } 4529 4530 /* 4531 * want to ensure that traffic has stopped before we change any of the flags 4532 */ 4533 if (setmask) { 4534 CTX_LOCK(ctx); 4535 bits = if_getdrvflags(ifp); 4536 if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) 4537 iflib_stop(ctx); 4538 STATE_LOCK(ctx); 4539 if_togglecapenable(ifp, setmask); 4540 ctx->ifc_softc_ctx.isc_capenable ^= setmask; 4541 STATE_UNLOCK(ctx); 4542 if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) 4543 iflib_init_locked(ctx); 4544 STATE_LOCK(ctx); 4545 if_setdrvflags(ifp, bits); 4546 STATE_UNLOCK(ctx); 4547 CTX_UNLOCK(ctx); 4548 } 4549 if_vlancap(ifp); 4550 break; 4551 } 4552 case SIOCGPRIVATE_0: 4553 case SIOCSDRVSPEC: 4554 case SIOCGDRVSPEC: 4555 CTX_LOCK(ctx); 4556 err = IFDI_PRIV_IOCTL(ctx, command, data); 4557 CTX_UNLOCK(ctx); 4558 break; 4559 default: 4560 err = ether_ioctl(ifp, command, data); 4561 break; 4562 } 4563 if (reinit) 4564 iflib_if_init(ctx); 4565 return (err); 4566 } 4567 4568 static uint64_t 4569 iflib_if_get_counter(if_t ifp, ift_counter cnt) 4570 { 4571 if_ctx_t ctx = if_getsoftc(ifp); 4572 4573 return (IFDI_GET_COUNTER(ctx, cnt)); 4574 } 4575 4576 /********************************************************************* 4577 * 4578 * OTHER FUNCTIONS EXPORTED TO THE STACK 4579 * 4580 **********************************************************************/ 4581 4582 static void 4583 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag) 4584 { 4585 if_ctx_t ctx = if_getsoftc(ifp); 4586 4587 if ((void *)ctx != arg) 4588 return; 4589 4590 if ((vtag == 0) || (vtag > 4095)) 4591 return; 4592 4593 if (iflib_in_detach(ctx)) 4594 return; 4595 4596 CTX_LOCK(ctx); 4597 /* Driver may need all untagged packets to be flushed */ 4598 if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) 4599 iflib_stop(ctx); 4600 IFDI_VLAN_REGISTER(ctx, vtag); 4601 /* Re-init to load the changes, if required */ 4602 if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) 4603 iflib_init_locked(ctx); 4604 CTX_UNLOCK(ctx); 4605 } 4606 4607 static void 4608 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag) 4609 { 4610 if_ctx_t ctx = if_getsoftc(ifp); 4611 4612 if ((void *)ctx != arg) 4613 return; 4614 4615 if ((vtag == 0) || (vtag > 4095)) 4616 return; 4617 4618 CTX_LOCK(ctx); 4619 /* Driver may need all tagged packets to be flushed */ 4620 if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) 4621 iflib_stop(ctx); 4622 IFDI_VLAN_UNREGISTER(ctx, vtag); 4623 /* Re-init to load the changes, if required */ 4624 if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) 4625 iflib_init_locked(ctx); 4626 CTX_UNLOCK(ctx); 4627 } 4628 4629 static void 4630 iflib_led_func(void *arg, int onoff) 4631 { 4632 if_ctx_t ctx = arg; 4633 4634 CTX_LOCK(ctx); 4635 IFDI_LED_FUNC(ctx, onoff); 4636 CTX_UNLOCK(ctx); 4637 } 4638 4639 /********************************************************************* 4640 * 4641 * BUS FUNCTION DEFINITIONS 4642 * 4643 **********************************************************************/ 4644 4645 int 4646 iflib_device_probe(device_t dev) 4647 { 4648 const pci_vendor_info_t *ent; 4649 if_shared_ctx_t sctx; 4650 uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id; 4651 uint16_t pci_vendor_id; 4652 4653 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) 4654 return (ENOTSUP); 4655 4656 pci_vendor_id = pci_get_vendor(dev); 4657 pci_device_id = pci_get_device(dev); 4658 pci_subvendor_id = pci_get_subvendor(dev); 4659 pci_subdevice_id = pci_get_subdevice(dev); 4660 pci_rev_id = pci_get_revid(dev); 4661 if (sctx->isc_parse_devinfo != NULL) 4662 sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id); 4663 4664 ent = sctx->isc_vendor_info; 4665 while (ent->pvi_vendor_id != 0) { 4666 if (pci_vendor_id != ent->pvi_vendor_id) { 4667 ent++; 4668 continue; 4669 } 4670 if ((pci_device_id == ent->pvi_device_id) && 4671 ((pci_subvendor_id == ent->pvi_subvendor_id) || 4672 (ent->pvi_subvendor_id == 0)) && 4673 ((pci_subdevice_id == ent->pvi_subdevice_id) || 4674 (ent->pvi_subdevice_id == 0)) && 4675 ((pci_rev_id == ent->pvi_rev_id) || 4676 (ent->pvi_rev_id == 0))) { 4677 device_set_desc_copy(dev, ent->pvi_name); 4678 /* this needs to be changed to zero if the bus probing code 4679 * ever stops re-probing on best match because the sctx 4680 * may have its values over written by register calls 4681 * in subsequent probes 4682 */ 4683 return (BUS_PROBE_DEFAULT); 4684 } 4685 ent++; 4686 } 4687 return (ENXIO); 4688 } 4689 4690 int 4691 iflib_device_probe_vendor(device_t dev) 4692 { 4693 int probe; 4694 4695 probe = iflib_device_probe(dev); 4696 if (probe == BUS_PROBE_DEFAULT) 4697 return (BUS_PROBE_VENDOR); 4698 else 4699 return (probe); 4700 } 4701 4702 static void 4703 iflib_reset_qvalues(if_ctx_t ctx) 4704 { 4705 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 4706 if_shared_ctx_t sctx = ctx->ifc_sctx; 4707 device_t dev = ctx->ifc_dev; 4708 int i; 4709 4710 if (ctx->ifc_sysctl_ntxqs != 0) 4711 scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs; 4712 if (ctx->ifc_sysctl_nrxqs != 0) 4713 scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs; 4714 4715 for (i = 0; i < sctx->isc_ntxqs; i++) { 4716 if (ctx->ifc_sysctl_ntxds[i] != 0) 4717 scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i]; 4718 else 4719 scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; 4720 } 4721 4722 for (i = 0; i < sctx->isc_nrxqs; i++) { 4723 if (ctx->ifc_sysctl_nrxds[i] != 0) 4724 scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i]; 4725 else 4726 scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; 4727 } 4728 4729 for (i = 0; i < sctx->isc_nrxqs; i++) { 4730 if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) { 4731 device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n", 4732 i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]); 4733 scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i]; 4734 } 4735 if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) { 4736 device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n", 4737 i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]); 4738 scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i]; 4739 } 4740 if (!powerof2(scctx->isc_nrxd[i])) { 4741 device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n", 4742 i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]); 4743 scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; 4744 } 4745 } 4746 4747 for (i = 0; i < sctx->isc_ntxqs; i++) { 4748 if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) { 4749 device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n", 4750 i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]); 4751 scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i]; 4752 } 4753 if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) { 4754 device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n", 4755 i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]); 4756 scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i]; 4757 } 4758 if (!powerof2(scctx->isc_ntxd[i])) { 4759 device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n", 4760 i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]); 4761 scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; 4762 } 4763 } 4764 } 4765 4766 static void 4767 iflib_add_pfil(if_ctx_t ctx) 4768 { 4769 struct pfil_head *pfil; 4770 struct pfil_head_args pa; 4771 iflib_rxq_t rxq; 4772 int i; 4773 4774 pa.pa_version = PFIL_VERSION; 4775 pa.pa_flags = PFIL_IN; 4776 pa.pa_type = PFIL_TYPE_ETHERNET; 4777 pa.pa_headname = if_name(ctx->ifc_ifp); 4778 pfil = pfil_head_register(&pa); 4779 4780 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { 4781 rxq->pfil = pfil; 4782 } 4783 } 4784 4785 static void 4786 iflib_rem_pfil(if_ctx_t ctx) 4787 { 4788 struct pfil_head *pfil; 4789 iflib_rxq_t rxq; 4790 int i; 4791 4792 rxq = ctx->ifc_rxqs; 4793 pfil = rxq->pfil; 4794 for (i = 0; i < NRXQSETS(ctx); i++, rxq++) { 4795 rxq->pfil = NULL; 4796 } 4797 pfil_head_unregister(pfil); 4798 } 4799 4800 4801 /* 4802 * Advance forward by n members of the cpuset ctx->ifc_cpus starting from 4803 * cpuid and wrapping as necessary. 4804 */ 4805 static unsigned int 4806 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n) 4807 { 4808 unsigned int first_valid; 4809 unsigned int last_valid; 4810 4811 /* cpuid should always be in the valid set */ 4812 MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus)); 4813 4814 /* valid set should never be empty */ 4815 MPASS(!CPU_EMPTY(&ctx->ifc_cpus)); 4816 4817 first_valid = CPU_FFS(&ctx->ifc_cpus) - 1; 4818 last_valid = CPU_FLS(&ctx->ifc_cpus) - 1; 4819 n = n % CPU_COUNT(&ctx->ifc_cpus); 4820 while (n > 0) { 4821 do { 4822 cpuid++; 4823 if (cpuid > last_valid) 4824 cpuid = first_valid; 4825 } while (!CPU_ISSET(cpuid, &ctx->ifc_cpus)); 4826 n--; 4827 } 4828 4829 return (cpuid); 4830 } 4831 4832 #if defined(SMP) && defined(SCHED_ULE) 4833 extern struct cpu_group *cpu_top; /* CPU topology */ 4834 4835 static int 4836 find_child_with_core(int cpu, struct cpu_group *grp) 4837 { 4838 int i; 4839 4840 if (grp->cg_children == 0) 4841 return (-1); 4842 4843 MPASS(grp->cg_child); 4844 for (i = 0; i < grp->cg_children; i++) { 4845 if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask)) 4846 return (i); 4847 } 4848 4849 return (-1); 4850 } 4851 4852 4853 /* 4854 * Find an L2 neighbor of the given CPU or return -1 if none found. This 4855 * does not distinguish among multiple L2 neighbors if the given CPU has 4856 * more than one (it will always return the same result in that case). 4857 */ 4858 static int 4859 find_l2_neighbor(int cpu) 4860 { 4861 struct cpu_group *grp; 4862 int i; 4863 4864 grp = cpu_top; 4865 if (grp == NULL) 4866 return (-1); 4867 4868 /* 4869 * Find the smallest CPU group that contains the given core. 4870 */ 4871 i = 0; 4872 while ((i = find_child_with_core(cpu, grp)) != -1) { 4873 /* 4874 * If the smallest group containing the given CPU has less 4875 * than two members, we conclude the given CPU has no 4876 * L2 neighbor. 4877 */ 4878 if (grp->cg_child[i].cg_count <= 1) 4879 return (-1); 4880 grp = &grp->cg_child[i]; 4881 } 4882 4883 /* Must share L2. */ 4884 if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE) 4885 return (-1); 4886 4887 /* 4888 * Select the first member of the set that isn't the reference 4889 * CPU, which at this point is guaranteed to exist. 4890 */ 4891 for (i = 0; i < CPU_SETSIZE; i++) { 4892 if (CPU_ISSET(i, &grp->cg_mask) && i != cpu) 4893 return (i); 4894 } 4895 4896 /* Should never be reached */ 4897 return (-1); 4898 } 4899 4900 #else 4901 static int 4902 find_l2_neighbor(int cpu) 4903 { 4904 4905 return (-1); 4906 } 4907 #endif 4908 4909 /* 4910 * CPU mapping behaviors 4911 * --------------------- 4912 * 'separate txrx' refers to the separate_txrx sysctl 4913 * 'use logical' refers to the use_logical_cores sysctl 4914 * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded 4915 * 4916 * separate use INTR 4917 * txrx logical CPUS result 4918 * ---------- --------- ------ ------------------------------------------------ 4919 * - - X RX and TX queues mapped to consecutive physical 4920 * cores with RX/TX pairs on same core and excess 4921 * of either following 4922 * - X X RX and TX queues mapped to consecutive cores 4923 * of any type with RX/TX pairs on same core and 4924 * excess of either following 4925 * X - X RX and TX queues mapped to consecutive physical 4926 * cores; all RX then all TX 4927 * X X X RX queues mapped to consecutive physical cores 4928 * first, then TX queues mapped to L2 neighbor of 4929 * the corresponding RX queue if one exists, 4930 * otherwise to consecutive physical cores 4931 * - n/a - RX and TX queues mapped to consecutive cores of 4932 * any type with RX/TX pairs on same core and excess 4933 * of either following 4934 * X n/a - RX and TX queues mapped to consecutive cores of 4935 * any type; all RX then all TX 4936 */ 4937 static unsigned int 4938 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid, 4939 bool is_tx) 4940 { 4941 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 4942 unsigned int core_index; 4943 4944 if (ctx->ifc_sysctl_separate_txrx) { 4945 /* 4946 * When using separate CPUs for TX and RX, the assignment 4947 * will always be of a consecutive CPU out of the set of 4948 * context CPUs, except for the specific case where the 4949 * context CPUs are phsyical cores, the use of logical cores 4950 * has been enabled, the assignment is for TX, the TX qid 4951 * corresponds to an RX qid, and the CPU assigned to the 4952 * corresponding RX queue has an L2 neighbor. 4953 */ 4954 if (ctx->ifc_sysctl_use_logical_cores && 4955 ctx->ifc_cpus_are_physical_cores && 4956 is_tx && qid < scctx->isc_nrxqsets) { 4957 int l2_neighbor; 4958 unsigned int rx_cpuid; 4959 4960 rx_cpuid = cpuid_advance(ctx, base_cpuid, qid); 4961 l2_neighbor = find_l2_neighbor(rx_cpuid); 4962 if (l2_neighbor != -1) { 4963 return (l2_neighbor); 4964 } 4965 /* 4966 * ... else fall through to the normal 4967 * consecutive-after-RX assignment scheme. 4968 * 4969 * Note that we are assuming that all RX queue CPUs 4970 * have an L2 neighbor, or all do not. If a mixed 4971 * scenario is possible, we will have to keep track 4972 * separately of how many queues prior to this one 4973 * were not able to be assigned to an L2 neighbor. 4974 */ 4975 } 4976 if (is_tx) 4977 core_index = scctx->isc_nrxqsets + qid; 4978 else 4979 core_index = qid; 4980 } else { 4981 core_index = qid; 4982 } 4983 4984 return (cpuid_advance(ctx, base_cpuid, core_index)); 4985 } 4986 4987 static uint16_t 4988 get_ctx_core_offset(if_ctx_t ctx) 4989 { 4990 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 4991 struct cpu_offset *op; 4992 cpuset_t assigned_cpus; 4993 unsigned int cores_consumed; 4994 unsigned int base_cpuid = ctx->ifc_sysctl_core_offset; 4995 unsigned int first_valid; 4996 unsigned int last_valid; 4997 unsigned int i; 4998 4999 first_valid = CPU_FFS(&ctx->ifc_cpus) - 1; 5000 last_valid = CPU_FLS(&ctx->ifc_cpus) - 1; 5001 5002 if (base_cpuid != CORE_OFFSET_UNSPECIFIED) { 5003 /* 5004 * Align the user-chosen base CPU ID to the next valid CPU 5005 * for this device. If the chosen base CPU ID is smaller 5006 * than the first valid CPU or larger than the last valid 5007 * CPU, we assume the user does not know what the valid 5008 * range is for this device and is thinking in terms of a 5009 * zero-based reference frame, and so we shift the given 5010 * value into the valid range (and wrap accordingly) so the 5011 * intent is translated to the proper frame of reference. 5012 * If the base CPU ID is within the valid first/last, but 5013 * does not correspond to a valid CPU, it is advanced to the 5014 * next valid CPU (wrapping if necessary). 5015 */ 5016 if (base_cpuid < first_valid || base_cpuid > last_valid) { 5017 /* shift from zero-based to first_valid-based */ 5018 base_cpuid += first_valid; 5019 /* wrap to range [first_valid, last_valid] */ 5020 base_cpuid = (base_cpuid - first_valid) % 5021 (last_valid - first_valid + 1); 5022 } 5023 if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) { 5024 /* 5025 * base_cpuid is in [first_valid, last_valid], but 5026 * not a member of the valid set. In this case, 5027 * there will always be a member of the valid set 5028 * with a CPU ID that is greater than base_cpuid, 5029 * and we simply advance to it. 5030 */ 5031 while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) 5032 base_cpuid++; 5033 } 5034 return (base_cpuid); 5035 } 5036 5037 /* 5038 * Determine how many cores will be consumed by performing the CPU 5039 * assignments and counting how many of the assigned CPUs correspond 5040 * to CPUs in the set of context CPUs. This is done using the CPU 5041 * ID first_valid as the base CPU ID, as the base CPU must be within 5042 * the set of context CPUs. 5043 * 5044 * Note not all assigned CPUs will be in the set of context CPUs 5045 * when separate CPUs are being allocated to TX and RX queues, 5046 * assignment to logical cores has been enabled, the set of context 5047 * CPUs contains only physical CPUs, and TX queues are mapped to L2 5048 * neighbors of CPUs that RX queues have been mapped to - in this 5049 * case we do only want to count how many CPUs in the set of context 5050 * CPUs have been consumed, as that determines the next CPU in that 5051 * set to start allocating at for the next device for which 5052 * core_offset is not set. 5053 */ 5054 CPU_ZERO(&assigned_cpus); 5055 for (i = 0; i < scctx->isc_ntxqsets; i++) 5056 CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true), 5057 &assigned_cpus); 5058 for (i = 0; i < scctx->isc_nrxqsets; i++) 5059 CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false), 5060 &assigned_cpus); 5061 CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus); 5062 cores_consumed = CPU_COUNT(&assigned_cpus); 5063 5064 mtx_lock(&cpu_offset_mtx); 5065 SLIST_FOREACH(op, &cpu_offsets, entries) { 5066 if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { 5067 base_cpuid = op->next_cpuid; 5068 op->next_cpuid = cpuid_advance(ctx, op->next_cpuid, 5069 cores_consumed); 5070 MPASS(op->refcount < UINT_MAX); 5071 op->refcount++; 5072 break; 5073 } 5074 } 5075 if (base_cpuid == CORE_OFFSET_UNSPECIFIED) { 5076 base_cpuid = first_valid; 5077 op = malloc(sizeof(struct cpu_offset), M_IFLIB, 5078 M_NOWAIT | M_ZERO); 5079 if (op == NULL) { 5080 device_printf(ctx->ifc_dev, 5081 "allocation for cpu offset failed.\n"); 5082 } else { 5083 op->next_cpuid = cpuid_advance(ctx, base_cpuid, 5084 cores_consumed); 5085 op->refcount = 1; 5086 CPU_COPY(&ctx->ifc_cpus, &op->set); 5087 SLIST_INSERT_HEAD(&cpu_offsets, op, entries); 5088 } 5089 } 5090 mtx_unlock(&cpu_offset_mtx); 5091 5092 return (base_cpuid); 5093 } 5094 5095 static void 5096 unref_ctx_core_offset(if_ctx_t ctx) 5097 { 5098 struct cpu_offset *op, *top; 5099 5100 mtx_lock(&cpu_offset_mtx); 5101 SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) { 5102 if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { 5103 MPASS(op->refcount > 0); 5104 op->refcount--; 5105 if (op->refcount == 0) { 5106 SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries); 5107 free(op, M_IFLIB); 5108 } 5109 break; 5110 } 5111 } 5112 mtx_unlock(&cpu_offset_mtx); 5113 } 5114 5115 int 5116 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp) 5117 { 5118 if_ctx_t ctx; 5119 if_t ifp; 5120 if_softc_ctx_t scctx; 5121 kobjop_desc_t kobj_desc; 5122 kobj_method_t *kobj_method; 5123 int err, msix, rid; 5124 int num_txd, num_rxd; 5125 char namebuf[TASKQUEUE_NAMELEN]; 5126 5127 ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO); 5128 5129 if (sc == NULL) { 5130 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO); 5131 device_set_softc(dev, ctx); 5132 ctx->ifc_flags |= IFC_SC_ALLOCATED; 5133 } 5134 5135 ctx->ifc_sctx = sctx; 5136 ctx->ifc_dev = dev; 5137 ctx->ifc_softc = sc; 5138 5139 iflib_register(ctx); 5140 iflib_add_device_sysctl_pre(ctx); 5141 5142 scctx = &ctx->ifc_softc_ctx; 5143 ifp = ctx->ifc_ifp; 5144 if (ctx->ifc_sysctl_simple_tx) { 5145 #ifndef ALTQ 5146 if_settransmitfn(ifp, iflib_simple_transmit); 5147 device_printf(dev, "using simple if_transmit\n"); 5148 #else 5149 device_printf(dev, "ALTQ prevents using simple if_transmit\n"); 5150 #endif 5151 } 5152 iflib_reset_qvalues(ctx); 5153 IFNET_WLOCK(); 5154 CTX_LOCK(ctx); 5155 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { 5156 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); 5157 goto fail_unlock; 5158 } 5159 _iflib_pre_assert(scctx); 5160 ctx->ifc_txrx = *scctx->isc_txrx; 5161 5162 MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR)); 5163 5164 if (sctx->isc_flags & IFLIB_DRIVER_MEDIA) 5165 ctx->ifc_mediap = scctx->isc_media; 5166 5167 #ifdef INVARIANTS 5168 if (scctx->isc_capabilities & IFCAP_TXCSUM) 5169 MPASS(scctx->isc_tx_csum_flags); 5170 #endif 5171 5172 if_setcapabilities(ifp, 5173 scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG); 5174 if_setcapenable(ifp, 5175 scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG); 5176 5177 if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) 5178 scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; 5179 if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) 5180 scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; 5181 5182 num_txd = iflib_num_tx_descs(ctx); 5183 num_rxd = iflib_num_rx_descs(ctx); 5184 5185 /* XXX change for per-queue sizes */ 5186 device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n", 5187 num_txd, num_rxd); 5188 5189 if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION) 5190 scctx->isc_tx_nsegments = max(1, num_txd / 5191 MAX_SINGLE_PACKET_FRACTION); 5192 if (scctx->isc_tx_tso_segments_max > num_txd / 5193 MAX_SINGLE_PACKET_FRACTION) 5194 scctx->isc_tx_tso_segments_max = max(1, 5195 num_txd / MAX_SINGLE_PACKET_FRACTION); 5196 5197 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ 5198 if (if_getcapabilities(ifp) & IFCAP_TSO) { 5199 /* 5200 * The stack can't handle a TSO size larger than IP_MAXPACKET, 5201 * but some MACs do. 5202 */ 5203 if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max, 5204 IP_MAXPACKET)); 5205 /* 5206 * Take maximum number of m_pullup(9)'s in iflib_parse_header() 5207 * into account. In the worst case, each of these calls will 5208 * add another mbuf and, thus, the requirement for another DMA 5209 * segment. So for best performance, it doesn't make sense to 5210 * advertize a maximum of TSO segments that typically will 5211 * require defragmentation in iflib_encap(). 5212 */ 5213 if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3); 5214 if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max); 5215 } 5216 if (scctx->isc_rss_table_size == 0) 5217 scctx->isc_rss_table_size = 64; 5218 scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1; 5219 5220 /* Create and start admin taskqueue */ 5221 snprintf(namebuf, TASKQUEUE_NAMELEN, "if_%s_tq", device_get_nameunit(dev)); 5222 ctx->ifc_tq = taskqueue_create_fast(namebuf, M_NOWAIT, 5223 taskqueue_thread_enqueue, &ctx->ifc_tq); 5224 if (ctx->ifc_tq == NULL) { 5225 device_printf(dev, "Unable to create admin taskqueue\n"); 5226 return (ENOMEM); 5227 } 5228 5229 err = taskqueue_start_threads(&ctx->ifc_tq, 1, PI_NET, "%s", namebuf); 5230 if (err) { 5231 device_printf(dev, 5232 "Unable to start admin taskqueue threads error: %d\n", 5233 err); 5234 taskqueue_free(ctx->ifc_tq); 5235 return (err); 5236 } 5237 5238 TASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); 5239 5240 /* Set up cpu set. If it fails, use the set of all CPUs. */ 5241 if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) { 5242 device_printf(dev, "Unable to fetch CPU list\n"); 5243 CPU_COPY(&all_cpus, &ctx->ifc_cpus); 5244 ctx->ifc_cpus_are_physical_cores = false; 5245 } else 5246 ctx->ifc_cpus_are_physical_cores = true; 5247 MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0); 5248 5249 /* 5250 * Now set up MSI or MSI-X, should return us the number of supported 5251 * vectors (will be 1 for a legacy interrupt and MSI). 5252 */ 5253 if (sctx->isc_flags & IFLIB_SKIP_MSIX) { 5254 msix = scctx->isc_vectors; 5255 } else if (scctx->isc_msix_bar != 0) 5256 /* 5257 * The simple fact that isc_msix_bar is not 0 does not mean we 5258 * we have a good value there that is known to work. 5259 */ 5260 msix = iflib_msix_init(ctx); 5261 else { 5262 scctx->isc_vectors = 1; 5263 scctx->isc_ntxqsets = 1; 5264 scctx->isc_nrxqsets = 1; 5265 scctx->isc_intr = IFLIB_INTR_LEGACY; 5266 msix = 0; 5267 } 5268 /* Get memory for the station queues */ 5269 if ((err = iflib_queues_alloc(ctx))) { 5270 device_printf(dev, "Unable to allocate queue memory\n"); 5271 goto fail_intr_free; 5272 } 5273 5274 if ((err = iflib_qset_structures_setup(ctx))) 5275 goto fail_queues; 5276 5277 /* 5278 * Now that we know how many queues there are, get the core offset. 5279 */ 5280 ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx); 5281 5282 if (msix > 1) { 5283 /* 5284 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable 5285 * aren't the default NULL implementation. 5286 */ 5287 kobj_desc = &ifdi_rx_queue_intr_enable_desc; 5288 kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, 5289 kobj_desc); 5290 if (kobj_method == &kobj_desc->deflt) { 5291 device_printf(dev, 5292 "MSI-X requires ifdi_rx_queue_intr_enable method"); 5293 err = EOPNOTSUPP; 5294 goto fail_queues; 5295 } 5296 kobj_desc = &ifdi_tx_queue_intr_enable_desc; 5297 kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, 5298 kobj_desc); 5299 if (kobj_method == &kobj_desc->deflt) { 5300 device_printf(dev, 5301 "MSI-X requires ifdi_tx_queue_intr_enable method"); 5302 err = EOPNOTSUPP; 5303 goto fail_queues; 5304 } 5305 5306 /* 5307 * Assign the MSI-X vectors. 5308 * Note that the default NULL ifdi_msix_intr_assign method will 5309 * fail here, too. 5310 */ 5311 err = IFDI_MSIX_INTR_ASSIGN(ctx, msix); 5312 if (err != 0) { 5313 device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", 5314 err); 5315 goto fail_queues; 5316 } 5317 } else if (scctx->isc_intr != IFLIB_INTR_MSIX) { 5318 rid = 0; 5319 if (scctx->isc_intr == IFLIB_INTR_MSI) { 5320 MPASS(msix == 1); 5321 rid = 1; 5322 } 5323 if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) { 5324 device_printf(dev, "iflib_legacy_setup failed %d\n", err); 5325 goto fail_queues; 5326 } 5327 } else { 5328 device_printf(dev, 5329 "Cannot use iflib with only 1 MSI-X interrupt!\n"); 5330 err = ENODEV; 5331 goto fail_queues; 5332 } 5333 5334 /* 5335 * It prevents a double-locking panic with iflib_media_status when 5336 * the driver loads. 5337 */ 5338 CTX_UNLOCK(ctx); 5339 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); 5340 CTX_LOCK(ctx); 5341 5342 if ((err = IFDI_ATTACH_POST(ctx)) != 0) { 5343 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); 5344 goto fail_detach; 5345 } 5346 5347 /* 5348 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. 5349 * This must appear after the call to ether_ifattach() because 5350 * ether_ifattach() sets if_hdrlen to the default value. 5351 */ 5352 if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) 5353 if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); 5354 5355 if ((err = iflib_netmap_attach(ctx))) { 5356 device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err); 5357 goto fail_detach; 5358 } 5359 *ctxp = ctx; 5360 5361 DEBUGNET_SET(ctx->ifc_ifp, iflib); 5362 5363 iflib_add_device_sysctl_post(ctx); 5364 iflib_add_pfil(ctx); 5365 ctx->ifc_flags |= IFC_INIT_DONE; 5366 CTX_UNLOCK(ctx); 5367 IFNET_WUNLOCK(); 5368 5369 return (0); 5370 5371 fail_detach: 5372 ether_ifdetach(ctx->ifc_ifp); 5373 fail_queues: 5374 taskqueue_free(ctx->ifc_tq); 5375 iflib_tqg_detach(ctx); 5376 iflib_tx_structures_free(ctx); 5377 iflib_rx_structures_free(ctx); 5378 IFDI_DETACH(ctx); 5379 IFDI_QUEUES_FREE(ctx); 5380 fail_intr_free: 5381 iflib_free_intr_mem(ctx); 5382 fail_unlock: 5383 CTX_UNLOCK(ctx); 5384 IFNET_WUNLOCK(); 5385 iflib_deregister(ctx); 5386 device_set_softc(ctx->ifc_dev, NULL); 5387 if (ctx->ifc_flags & IFC_SC_ALLOCATED) 5388 free(ctx->ifc_softc, M_IFLIB); 5389 free(ctx, M_IFLIB); 5390 return (err); 5391 } 5392 5393 int 5394 iflib_device_attach(device_t dev) 5395 { 5396 if_ctx_t ctx; 5397 if_shared_ctx_t sctx; 5398 5399 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) 5400 return (ENOTSUP); 5401 5402 pci_enable_busmaster(dev); 5403 5404 return (iflib_device_register(dev, NULL, sctx, &ctx)); 5405 } 5406 5407 int 5408 iflib_device_deregister(if_ctx_t ctx) 5409 { 5410 if_t ifp = ctx->ifc_ifp; 5411 device_t dev = ctx->ifc_dev; 5412 5413 /* Make sure VLANS are not using driver */ 5414 if (if_vlantrunkinuse(ifp)) { 5415 device_printf(dev, "Vlan in use, detach first\n"); 5416 return (EBUSY); 5417 } 5418 #ifdef PCI_IOV 5419 if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) { 5420 device_printf(dev, "SR-IOV in use; detach first.\n"); 5421 return (EBUSY); 5422 } 5423 #endif 5424 5425 STATE_LOCK(ctx); 5426 ctx->ifc_flags |= IFC_IN_DETACH; 5427 STATE_UNLOCK(ctx); 5428 5429 /* Unregister VLAN handlers before calling iflib_stop() */ 5430 iflib_unregister_vlan_handlers(ctx); 5431 5432 iflib_netmap_detach(ifp); 5433 ether_ifdetach(ifp); 5434 5435 CTX_LOCK(ctx); 5436 iflib_stop(ctx); 5437 CTX_UNLOCK(ctx); 5438 5439 iflib_rem_pfil(ctx); 5440 if (ctx->ifc_led_dev != NULL) 5441 led_destroy(ctx->ifc_led_dev); 5442 5443 iflib_tqg_detach(ctx); 5444 iflib_tx_structures_free(ctx); 5445 iflib_rx_structures_free(ctx); 5446 5447 CTX_LOCK(ctx); 5448 IFDI_DETACH(ctx); 5449 IFDI_QUEUES_FREE(ctx); 5450 CTX_UNLOCK(ctx); 5451 5452 taskqueue_free(ctx->ifc_tq); 5453 ctx->ifc_tq = NULL; 5454 5455 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ 5456 iflib_free_intr_mem(ctx); 5457 5458 bus_generic_detach(dev); 5459 5460 iflib_deregister(ctx); 5461 5462 device_set_softc(ctx->ifc_dev, NULL); 5463 if (ctx->ifc_flags & IFC_SC_ALLOCATED) 5464 free(ctx->ifc_softc, M_IFLIB); 5465 unref_ctx_core_offset(ctx); 5466 free(ctx, M_IFLIB); 5467 return (0); 5468 } 5469 5470 static void 5471 iflib_tqg_detach(if_ctx_t ctx) 5472 { 5473 iflib_txq_t txq; 5474 iflib_rxq_t rxq; 5475 int i; 5476 struct taskqgroup *tqg; 5477 5478 /* XXX drain any dependent tasks */ 5479 tqg = qgroup_if_io_tqg; 5480 for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { 5481 callout_drain(&txq->ift_timer); 5482 #ifdef DEV_NETMAP 5483 callout_drain(&txq->ift_netmap_timer); 5484 #endif /* DEV_NETMAP */ 5485 if (txq->ift_task.gt_uniq != NULL) 5486 taskqgroup_detach(tqg, &txq->ift_task); 5487 } 5488 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { 5489 if (rxq->ifr_task.gt_uniq != NULL) 5490 taskqgroup_detach(tqg, &rxq->ifr_task); 5491 } 5492 } 5493 5494 static void 5495 iflib_free_intr_mem(if_ctx_t ctx) 5496 { 5497 5498 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) { 5499 iflib_irq_free(ctx, &ctx->ifc_legacy_irq); 5500 } 5501 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { 5502 pci_release_msi(ctx->ifc_dev); 5503 } 5504 if (ctx->ifc_msix_mem != NULL) { 5505 bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY, 5506 rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem); 5507 ctx->ifc_msix_mem = NULL; 5508 } 5509 } 5510 5511 int 5512 iflib_device_detach(device_t dev) 5513 { 5514 if_ctx_t ctx = device_get_softc(dev); 5515 5516 return (iflib_device_deregister(ctx)); 5517 } 5518 5519 int 5520 iflib_device_suspend(device_t dev) 5521 { 5522 if_ctx_t ctx = device_get_softc(dev); 5523 5524 CTX_LOCK(ctx); 5525 IFDI_SUSPEND(ctx); 5526 CTX_UNLOCK(ctx); 5527 5528 return (bus_generic_suspend(dev)); 5529 } 5530 int 5531 iflib_device_shutdown(device_t dev) 5532 { 5533 if_ctx_t ctx = device_get_softc(dev); 5534 5535 CTX_LOCK(ctx); 5536 IFDI_SHUTDOWN(ctx); 5537 CTX_UNLOCK(ctx); 5538 5539 return (bus_generic_suspend(dev)); 5540 } 5541 5542 int 5543 iflib_device_resume(device_t dev) 5544 { 5545 if_ctx_t ctx = device_get_softc(dev); 5546 iflib_txq_t txq = ctx->ifc_txqs; 5547 5548 CTX_LOCK(ctx); 5549 IFDI_RESUME(ctx); 5550 iflib_if_init_locked(ctx); 5551 CTX_UNLOCK(ctx); 5552 for (int i = 0; i < NTXQSETS(ctx); i++, txq++) 5553 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); 5554 5555 return (bus_generic_resume(dev)); 5556 } 5557 5558 int 5559 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params) 5560 { 5561 int error; 5562 if_ctx_t ctx = device_get_softc(dev); 5563 5564 CTX_LOCK(ctx); 5565 error = IFDI_IOV_INIT(ctx, num_vfs, params); 5566 CTX_UNLOCK(ctx); 5567 5568 return (error); 5569 } 5570 5571 void 5572 iflib_device_iov_uninit(device_t dev) 5573 { 5574 if_ctx_t ctx = device_get_softc(dev); 5575 5576 CTX_LOCK(ctx); 5577 IFDI_IOV_UNINIT(ctx); 5578 CTX_UNLOCK(ctx); 5579 } 5580 5581 int 5582 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params) 5583 { 5584 int error; 5585 if_ctx_t ctx = device_get_softc(dev); 5586 5587 CTX_LOCK(ctx); 5588 error = IFDI_IOV_VF_ADD(ctx, vfnum, params); 5589 CTX_UNLOCK(ctx); 5590 5591 return (error); 5592 } 5593 5594 /********************************************************************* 5595 * 5596 * MODULE FUNCTION DEFINITIONS 5597 * 5598 **********************************************************************/ 5599 5600 /* 5601 * - Start a fast taskqueue thread for each core 5602 * - Start a taskqueue for control operations 5603 */ 5604 static int 5605 iflib_module_init(void) 5606 { 5607 iflib_timer_default = hz / 2; 5608 return (0); 5609 } 5610 5611 static int 5612 iflib_module_event_handler(module_t mod, int what, void *arg) 5613 { 5614 int err; 5615 5616 switch (what) { 5617 case MOD_LOAD: 5618 if ((err = iflib_module_init()) != 0) 5619 return (err); 5620 break; 5621 case MOD_UNLOAD: 5622 return (EBUSY); 5623 default: 5624 return (EOPNOTSUPP); 5625 } 5626 5627 return (0); 5628 } 5629 5630 /********************************************************************* 5631 * 5632 * PUBLIC FUNCTION DEFINITIONS 5633 * ordered as in iflib.h 5634 * 5635 **********************************************************************/ 5636 5637 static void 5638 _iflib_assert(if_shared_ctx_t sctx) 5639 { 5640 int i; 5641 5642 MPASS(sctx->isc_tx_maxsize); 5643 MPASS(sctx->isc_tx_maxsegsize); 5644 5645 MPASS(sctx->isc_rx_maxsize); 5646 MPASS(sctx->isc_rx_nsegments); 5647 MPASS(sctx->isc_rx_maxsegsize); 5648 5649 MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8); 5650 for (i = 0; i < sctx->isc_nrxqs; i++) { 5651 MPASS(sctx->isc_nrxd_min[i]); 5652 MPASS(powerof2(sctx->isc_nrxd_min[i])); 5653 MPASS(sctx->isc_nrxd_max[i]); 5654 MPASS(powerof2(sctx->isc_nrxd_max[i])); 5655 MPASS(sctx->isc_nrxd_default[i]); 5656 MPASS(powerof2(sctx->isc_nrxd_default[i])); 5657 } 5658 5659 MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8); 5660 for (i = 0; i < sctx->isc_ntxqs; i++) { 5661 MPASS(sctx->isc_ntxd_min[i]); 5662 MPASS(powerof2(sctx->isc_ntxd_min[i])); 5663 MPASS(sctx->isc_ntxd_max[i]); 5664 MPASS(powerof2(sctx->isc_ntxd_max[i])); 5665 MPASS(sctx->isc_ntxd_default[i]); 5666 MPASS(powerof2(sctx->isc_ntxd_default[i])); 5667 } 5668 } 5669 5670 static void 5671 _iflib_pre_assert(if_softc_ctx_t scctx) 5672 { 5673 5674 MPASS(scctx->isc_txrx->ift_txd_encap); 5675 MPASS(scctx->isc_txrx->ift_txd_flush); 5676 MPASS(scctx->isc_txrx->ift_txd_credits_update); 5677 MPASS(scctx->isc_txrx->ift_rxd_available); 5678 MPASS(scctx->isc_txrx->ift_rxd_pkt_get); 5679 MPASS(scctx->isc_txrx->ift_rxd_refill); 5680 MPASS(scctx->isc_txrx->ift_rxd_flush); 5681 } 5682 5683 static void 5684 iflib_register(if_ctx_t ctx) 5685 { 5686 if_shared_ctx_t sctx = ctx->ifc_sctx; 5687 driver_t *driver = sctx->isc_driver; 5688 device_t dev = ctx->ifc_dev; 5689 if_t ifp; 5690 5691 _iflib_assert(sctx); 5692 5693 CTX_LOCK_INIT(ctx); 5694 STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); 5695 ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev); 5696 5697 /* 5698 * Initialize our context's device specific methods 5699 */ 5700 kobj_init((kobj_t) ctx, (kobj_class_t) driver); 5701 kobj_class_compile((kobj_class_t) driver); 5702 5703 if_initname(ifp, device_get_name(dev), device_get_unit(dev)); 5704 if_setsoftc(ifp, ctx); 5705 if_setdev(ifp, dev); 5706 if_setinitfn(ifp, iflib_if_init); 5707 if_setioctlfn(ifp, iflib_if_ioctl); 5708 #ifdef ALTQ 5709 if_setstartfn(ifp, iflib_altq_if_start); 5710 if_settransmitfn(ifp, iflib_altq_if_transmit); 5711 if_setsendqready(ifp); 5712 #else 5713 if_settransmitfn(ifp, iflib_if_transmit); 5714 #endif 5715 if_setqflushfn(ifp, iflib_if_qflush); 5716 if_setgetcounterfn(ifp, iflib_if_get_counter); 5717 if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); 5718 ctx->ifc_vlan_attach_event = 5719 EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx, 5720 EVENTHANDLER_PRI_FIRST); 5721 ctx->ifc_vlan_detach_event = 5722 EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx, 5723 EVENTHANDLER_PRI_FIRST); 5724 5725 if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) { 5726 ctx->ifc_mediap = &ctx->ifc_media; 5727 ifmedia_init(ctx->ifc_mediap, IFM_IMASK, 5728 iflib_media_change, iflib_media_status); 5729 } 5730 } 5731 5732 static void 5733 iflib_unregister_vlan_handlers(if_ctx_t ctx) 5734 { 5735 /* Unregister VLAN events */ 5736 if (ctx->ifc_vlan_attach_event != NULL) { 5737 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event); 5738 ctx->ifc_vlan_attach_event = NULL; 5739 } 5740 if (ctx->ifc_vlan_detach_event != NULL) { 5741 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event); 5742 ctx->ifc_vlan_detach_event = NULL; 5743 } 5744 5745 } 5746 5747 static void 5748 iflib_deregister(if_ctx_t ctx) 5749 { 5750 if_t ifp = ctx->ifc_ifp; 5751 5752 /* Remove all media */ 5753 ifmedia_removeall(&ctx->ifc_media); 5754 5755 /* Ensure that VLAN event handlers are unregistered */ 5756 iflib_unregister_vlan_handlers(ctx); 5757 5758 /* Release kobject reference */ 5759 kobj_delete((kobj_t) ctx, NULL); 5760 5761 /* Free the ifnet structure */ 5762 if_free(ifp); 5763 5764 STATE_LOCK_DESTROY(ctx); 5765 5766 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ 5767 CTX_LOCK_DESTROY(ctx); 5768 } 5769 5770 static int 5771 iflib_queues_alloc(if_ctx_t ctx) 5772 { 5773 if_shared_ctx_t sctx = ctx->ifc_sctx; 5774 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 5775 device_t dev = ctx->ifc_dev; 5776 int nrxqsets = scctx->isc_nrxqsets; 5777 int ntxqsets = scctx->isc_ntxqsets; 5778 iflib_txq_t txq; 5779 iflib_rxq_t rxq; 5780 iflib_fl_t fl = NULL; 5781 int i, j, cpu, err, txconf, rxconf; 5782 iflib_dma_info_t ifdip; 5783 uint32_t *rxqsizes = scctx->isc_rxqsizes; 5784 uint32_t *txqsizes = scctx->isc_txqsizes; 5785 uint8_t nrxqs = sctx->isc_nrxqs; 5786 uint8_t ntxqs = sctx->isc_ntxqs; 5787 int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; 5788 int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0); 5789 caddr_t *vaddrs; 5790 uint64_t *paddrs; 5791 5792 KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1")); 5793 KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1")); 5794 KASSERT(nrxqs >= fl_offset + nfree_lists, 5795 ("there must be at least a rxq for each free list")); 5796 5797 /* Allocate the TX ring struct memory */ 5798 if (!(ctx->ifc_txqs = 5799 (iflib_txq_t) malloc(sizeof(struct iflib_txq) * 5800 ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { 5801 device_printf(dev, "Unable to allocate TX ring memory\n"); 5802 err = ENOMEM; 5803 goto fail; 5804 } 5805 5806 /* Now allocate the RX */ 5807 if (!(ctx->ifc_rxqs = 5808 (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) * 5809 nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { 5810 device_printf(dev, "Unable to allocate RX ring memory\n"); 5811 err = ENOMEM; 5812 goto rx_fail; 5813 } 5814 5815 txq = ctx->ifc_txqs; 5816 rxq = ctx->ifc_rxqs; 5817 5818 /* 5819 * XXX handle allocation failure 5820 */ 5821 for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) { 5822 /* Set up some basics */ 5823 5824 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, 5825 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { 5826 device_printf(dev, 5827 "Unable to allocate TX DMA info memory\n"); 5828 err = ENOMEM; 5829 goto err_tx_desc; 5830 } 5831 txq->ift_ifdi = ifdip; 5832 for (j = 0; j < ntxqs; j++, ifdip++) { 5833 if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) { 5834 device_printf(dev, 5835 "Unable to allocate TX descriptors\n"); 5836 err = ENOMEM; 5837 goto err_tx_desc; 5838 } 5839 txq->ift_txd_size[j] = scctx->isc_txd_size[j]; 5840 bzero((void *)ifdip->idi_vaddr, txqsizes[j]); 5841 } 5842 txq->ift_ctx = ctx; 5843 txq->ift_id = i; 5844 if (sctx->isc_flags & IFLIB_HAS_TXCQ) { 5845 txq->ift_br_offset = 1; 5846 } else { 5847 txq->ift_br_offset = 0; 5848 } 5849 5850 if (iflib_txsd_alloc(txq)) { 5851 device_printf(dev, "Critical Failure setting up TX buffers\n"); 5852 err = ENOMEM; 5853 goto err_tx_desc; 5854 } 5855 5856 /* Initialize the TX lock */ 5857 snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout", 5858 device_get_nameunit(dev), txq->ift_id); 5859 mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF); 5860 callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0); 5861 txq->ift_timer.c_cpu = cpu; 5862 #ifdef DEV_NETMAP 5863 callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0); 5864 txq->ift_netmap_timer.c_cpu = cpu; 5865 #endif /* DEV_NETMAP */ 5866 5867 err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain, 5868 iflib_txq_can_drain, M_IFLIB, M_WAITOK); 5869 if (err) { 5870 /* XXX free any allocated rings */ 5871 device_printf(dev, "Unable to allocate buf_ring\n"); 5872 goto err_tx_desc; 5873 } 5874 } 5875 5876 for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) { 5877 /* Set up some basics */ 5878 callout_init(&rxq->ifr_watchdog, 1); 5879 5880 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, 5881 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { 5882 device_printf(dev, 5883 "Unable to allocate RX DMA info memory\n"); 5884 err = ENOMEM; 5885 goto err_tx_desc; 5886 } 5887 5888 rxq->ifr_ifdi = ifdip; 5889 /* XXX this needs to be changed if #rx queues != #tx queues */ 5890 rxq->ifr_ntxqirq = 1; 5891 rxq->ifr_txqid[0] = i; 5892 for (j = 0; j < nrxqs; j++, ifdip++) { 5893 if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) { 5894 device_printf(dev, 5895 "Unable to allocate RX descriptors\n"); 5896 err = ENOMEM; 5897 goto err_tx_desc; 5898 } 5899 bzero((void *)ifdip->idi_vaddr, rxqsizes[j]); 5900 } 5901 rxq->ifr_ctx = ctx; 5902 rxq->ifr_id = i; 5903 rxq->ifr_fl_offset = fl_offset; 5904 rxq->ifr_nfl = nfree_lists; 5905 if (!(fl = 5906 (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) { 5907 device_printf(dev, "Unable to allocate free list memory\n"); 5908 err = ENOMEM; 5909 goto err_tx_desc; 5910 } 5911 rxq->ifr_fl = fl; 5912 for (j = 0; j < nfree_lists; j++) { 5913 fl[j].ifl_rxq = rxq; 5914 fl[j].ifl_id = j; 5915 fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset]; 5916 fl[j].ifl_rxd_size = scctx->isc_rxd_size[j]; 5917 } 5918 /* Allocate receive buffers for the ring */ 5919 if (iflib_rxsd_alloc(rxq)) { 5920 device_printf(dev, 5921 "Critical Failure setting up receive buffers\n"); 5922 err = ENOMEM; 5923 goto err_rx_desc; 5924 } 5925 5926 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) 5927 fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, 5928 M_WAITOK); 5929 } 5930 5931 /* TXQs */ 5932 vaddrs = malloc(sizeof(caddr_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK); 5933 paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK); 5934 for (i = 0; i < ntxqsets; i++) { 5935 iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi; 5936 5937 for (j = 0; j < ntxqs; j++, di++) { 5938 vaddrs[i * ntxqs + j] = di->idi_vaddr; 5939 paddrs[i * ntxqs + j] = di->idi_paddr; 5940 } 5941 } 5942 if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) { 5943 device_printf(ctx->ifc_dev, 5944 "Unable to allocate device TX queue\n"); 5945 iflib_tx_structures_free(ctx); 5946 free(vaddrs, M_IFLIB); 5947 free(paddrs, M_IFLIB); 5948 goto err_rx_desc; 5949 } 5950 free(vaddrs, M_IFLIB); 5951 free(paddrs, M_IFLIB); 5952 5953 /* RXQs */ 5954 vaddrs = malloc(sizeof(caddr_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK); 5955 paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK); 5956 for (i = 0; i < nrxqsets; i++) { 5957 iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi; 5958 5959 for (j = 0; j < nrxqs; j++, di++) { 5960 vaddrs[i * nrxqs + j] = di->idi_vaddr; 5961 paddrs[i * nrxqs + j] = di->idi_paddr; 5962 } 5963 } 5964 if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) { 5965 device_printf(ctx->ifc_dev, 5966 "Unable to allocate device RX queue\n"); 5967 iflib_tx_structures_free(ctx); 5968 free(vaddrs, M_IFLIB); 5969 free(paddrs, M_IFLIB); 5970 goto err_rx_desc; 5971 } 5972 free(vaddrs, M_IFLIB); 5973 free(paddrs, M_IFLIB); 5974 5975 return (0); 5976 5977 /* XXX handle allocation failure changes */ 5978 err_rx_desc: 5979 err_tx_desc: 5980 rx_fail: 5981 if (ctx->ifc_rxqs != NULL) 5982 free(ctx->ifc_rxqs, M_IFLIB); 5983 ctx->ifc_rxqs = NULL; 5984 if (ctx->ifc_txqs != NULL) 5985 free(ctx->ifc_txqs, M_IFLIB); 5986 ctx->ifc_txqs = NULL; 5987 fail: 5988 return (err); 5989 } 5990 5991 static int 5992 iflib_tx_structures_setup(if_ctx_t ctx) 5993 { 5994 iflib_txq_t txq = ctx->ifc_txqs; 5995 int i; 5996 5997 for (i = 0; i < NTXQSETS(ctx); i++, txq++) 5998 iflib_txq_setup(txq); 5999 6000 return (0); 6001 } 6002 6003 static void 6004 iflib_tx_structures_free(if_ctx_t ctx) 6005 { 6006 iflib_txq_t txq = ctx->ifc_txqs; 6007 if_shared_ctx_t sctx = ctx->ifc_sctx; 6008 int i, j; 6009 6010 for (i = 0; i < NTXQSETS(ctx); i++, txq++) { 6011 for (j = 0; j < sctx->isc_ntxqs; j++) 6012 iflib_dma_free(&txq->ift_ifdi[j]); 6013 iflib_txq_destroy(txq); 6014 } 6015 free(ctx->ifc_txqs, M_IFLIB); 6016 ctx->ifc_txqs = NULL; 6017 } 6018 6019 /********************************************************************* 6020 * 6021 * Initialize all receive rings. 6022 * 6023 **********************************************************************/ 6024 static int 6025 iflib_rx_structures_setup(if_ctx_t ctx) 6026 { 6027 iflib_rxq_t rxq = ctx->ifc_rxqs; 6028 int q; 6029 #if defined(INET6) || defined(INET) 6030 int err, i; 6031 #endif 6032 6033 for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { 6034 #if defined(INET6) || defined(INET) 6035 err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp, 6036 TCP_LRO_ENTRIES, min(1024, 6037 ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset])); 6038 if (err != 0) { 6039 device_printf(ctx->ifc_dev, 6040 "LRO Initialization failed!\n"); 6041 goto fail; 6042 } 6043 #endif 6044 IFDI_RXQ_SETUP(ctx, rxq->ifr_id); 6045 } 6046 return (0); 6047 #if defined(INET6) || defined(INET) 6048 fail: 6049 /* 6050 * Free LRO resources allocated so far, we will only handle 6051 * the rings that completed, the failing case will have 6052 * cleaned up for itself. 'q' failed, so its the terminus. 6053 */ 6054 rxq = ctx->ifc_rxqs; 6055 for (i = 0; i < q; ++i, rxq++) { 6056 tcp_lro_free(&rxq->ifr_lc); 6057 } 6058 return (err); 6059 #endif 6060 } 6061 6062 /********************************************************************* 6063 * 6064 * Free all receive rings. 6065 * 6066 **********************************************************************/ 6067 static void 6068 iflib_rx_structures_free(if_ctx_t ctx) 6069 { 6070 iflib_rxq_t rxq = ctx->ifc_rxqs; 6071 if_shared_ctx_t sctx = ctx->ifc_sctx; 6072 int i, j; 6073 6074 for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) { 6075 for (j = 0; j < sctx->isc_nrxqs; j++) 6076 iflib_dma_free(&rxq->ifr_ifdi[j]); 6077 iflib_rx_sds_free(rxq); 6078 #if defined(INET6) || defined(INET) 6079 tcp_lro_free(&rxq->ifr_lc); 6080 #endif 6081 } 6082 free(ctx->ifc_rxqs, M_IFLIB); 6083 ctx->ifc_rxqs = NULL; 6084 } 6085 6086 static int 6087 iflib_qset_structures_setup(if_ctx_t ctx) 6088 { 6089 int err; 6090 6091 /* 6092 * It is expected that the caller takes care of freeing queues if this 6093 * fails. 6094 */ 6095 if ((err = iflib_tx_structures_setup(ctx)) != 0) { 6096 device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err); 6097 return (err); 6098 } 6099 6100 if ((err = iflib_rx_structures_setup(ctx)) != 0) 6101 device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err); 6102 6103 return (err); 6104 } 6105 6106 int 6107 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, 6108 driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name) 6109 { 6110 6111 return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name)); 6112 } 6113 6114 /* Just to avoid copy/paste */ 6115 static inline int 6116 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, 6117 int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, 6118 const char *name) 6119 { 6120 device_t dev; 6121 unsigned int base_cpuid, cpuid; 6122 int err; 6123 6124 dev = ctx->ifc_dev; 6125 base_cpuid = ctx->ifc_sysctl_core_offset; 6126 cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX); 6127 err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, 6128 irq ? irq->ii_res : NULL, name); 6129 if (err) { 6130 device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err); 6131 return (err); 6132 } 6133 #ifdef notyet 6134 if (cpuid > ctx->ifc_cpuid_highest) 6135 ctx->ifc_cpuid_highest = cpuid; 6136 #endif 6137 return (0); 6138 } 6139 6140 /* 6141 * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware 6142 * resources. 6143 * 6144 * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX 6145 * only. 6146 * 6147 * XXX: Could be removed if subctx's dev has its intr resource allocation 6148 * methods replaced with custom ones? 6149 */ 6150 int 6151 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq, 6152 int rid, iflib_intr_type_t type, 6153 driver_filter_t *filter, void *filter_arg, 6154 int qid, const char *name) 6155 { 6156 device_t dev, subdev; 6157 struct grouptask *gtask; 6158 struct taskqgroup *tqg; 6159 iflib_filter_info_t info; 6160 gtask_fn_t *fn; 6161 int tqrid, err; 6162 driver_filter_t *intr_fast; 6163 void *q; 6164 6165 MPASS(ctx != NULL); 6166 MPASS(subctx != NULL); 6167 6168 tqrid = rid; 6169 dev = ctx->ifc_dev; 6170 subdev = subctx->ifc_dev; 6171 6172 switch (type) { 6173 case IFLIB_INTR_RXTX: 6174 q = &subctx->ifc_rxqs[qid]; 6175 info = &subctx->ifc_rxqs[qid].ifr_filter_info; 6176 gtask = &subctx->ifc_rxqs[qid].ifr_task; 6177 tqg = qgroup_if_io_tqg; 6178 fn = _task_fn_rx; 6179 intr_fast = iflib_fast_intr_rxtx; 6180 NET_GROUPTASK_INIT(gtask, 0, fn, q); 6181 break; 6182 default: 6183 device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n", 6184 __func__, device_get_nameunit(subdev), type); 6185 return (EINVAL); 6186 } 6187 6188 info->ifi_filter = filter; 6189 info->ifi_filter_arg = filter_arg; 6190 info->ifi_task = gtask; 6191 info->ifi_ctx = q; 6192 6193 NET_GROUPTASK_INIT(gtask, 0, fn, q); 6194 6195 /* Allocate interrupts from hardware using parent context */ 6196 err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name); 6197 if (err != 0) { 6198 device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n", 6199 device_get_nameunit(subdev), err); 6200 return (err); 6201 } 6202 6203 if (tqrid != -1) { 6204 err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, 6205 name); 6206 if (err) 6207 return (err); 6208 } else { 6209 taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name); 6210 } 6211 6212 return (0); 6213 } 6214 6215 int 6216 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid, 6217 iflib_intr_type_t type, driver_filter_t *filter, 6218 void *filter_arg, int qid, const char *name) 6219 { 6220 device_t dev; 6221 struct grouptask *gtask; 6222 struct taskqgroup *tqg; 6223 iflib_filter_info_t info; 6224 gtask_fn_t *fn; 6225 int tqrid, err; 6226 driver_filter_t *intr_fast; 6227 void *q; 6228 6229 info = &ctx->ifc_filter_info; 6230 tqrid = rid; 6231 6232 switch (type) { 6233 /* XXX merge tx/rx for netmap? */ 6234 case IFLIB_INTR_TX: 6235 q = &ctx->ifc_txqs[qid]; 6236 info = &ctx->ifc_txqs[qid].ift_filter_info; 6237 gtask = &ctx->ifc_txqs[qid].ift_task; 6238 tqg = qgroup_if_io_tqg; 6239 fn = _task_fn_tx; 6240 intr_fast = iflib_fast_intr; 6241 GROUPTASK_INIT(gtask, 0, fn, q); 6242 ctx->ifc_flags |= IFC_NETMAP_TX_IRQ; 6243 break; 6244 case IFLIB_INTR_RX: 6245 q = &ctx->ifc_rxqs[qid]; 6246 info = &ctx->ifc_rxqs[qid].ifr_filter_info; 6247 gtask = &ctx->ifc_rxqs[qid].ifr_task; 6248 tqg = qgroup_if_io_tqg; 6249 fn = _task_fn_rx; 6250 intr_fast = iflib_fast_intr; 6251 NET_GROUPTASK_INIT(gtask, 0, fn, q); 6252 break; 6253 case IFLIB_INTR_RXTX: 6254 q = &ctx->ifc_rxqs[qid]; 6255 info = &ctx->ifc_rxqs[qid].ifr_filter_info; 6256 gtask = &ctx->ifc_rxqs[qid].ifr_task; 6257 tqg = qgroup_if_io_tqg; 6258 fn = _task_fn_rx; 6259 intr_fast = iflib_fast_intr_rxtx; 6260 NET_GROUPTASK_INIT(gtask, 0, fn, q); 6261 break; 6262 case IFLIB_INTR_ADMIN: 6263 q = ctx; 6264 tqrid = -1; 6265 info = &ctx->ifc_filter_info; 6266 gtask = NULL; 6267 intr_fast = iflib_fast_intr_ctx; 6268 break; 6269 default: 6270 device_printf(ctx->ifc_dev, "%s: unknown net intr type\n", 6271 __func__); 6272 return (EINVAL); 6273 } 6274 6275 info->ifi_filter = filter; 6276 info->ifi_filter_arg = filter_arg; 6277 info->ifi_task = gtask; 6278 info->ifi_ctx = q; 6279 6280 dev = ctx->ifc_dev; 6281 err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name); 6282 if (err != 0) { 6283 device_printf(dev, "_iflib_irq_alloc failed %d\n", err); 6284 return (err); 6285 } 6286 if (type == IFLIB_INTR_ADMIN) 6287 return (0); 6288 6289 if (tqrid != -1) { 6290 err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, 6291 name); 6292 if (err) 6293 return (err); 6294 } else { 6295 taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name); 6296 } 6297 6298 return (0); 6299 } 6300 6301 void 6302 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, 6303 void *arg, int qid, const char *name) 6304 { 6305 device_t dev; 6306 struct grouptask *gtask; 6307 struct taskqgroup *tqg; 6308 gtask_fn_t *fn; 6309 void *q; 6310 int err; 6311 6312 switch (type) { 6313 case IFLIB_INTR_TX: 6314 q = &ctx->ifc_txqs[qid]; 6315 gtask = &ctx->ifc_txqs[qid].ift_task; 6316 tqg = qgroup_if_io_tqg; 6317 fn = _task_fn_tx; 6318 GROUPTASK_INIT(gtask, 0, fn, q); 6319 break; 6320 case IFLIB_INTR_RX: 6321 q = &ctx->ifc_rxqs[qid]; 6322 gtask = &ctx->ifc_rxqs[qid].ifr_task; 6323 tqg = qgroup_if_io_tqg; 6324 fn = _task_fn_rx; 6325 NET_GROUPTASK_INIT(gtask, 0, fn, q); 6326 break; 6327 case IFLIB_INTR_IOV: 6328 TASK_INIT(&ctx->ifc_vflr_task, 0, _task_fn_iov, ctx); 6329 return; 6330 default: 6331 panic("unknown net intr type"); 6332 } 6333 err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name); 6334 if (err) { 6335 dev = ctx->ifc_dev; 6336 taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL, 6337 name); 6338 } 6339 } 6340 6341 void 6342 iflib_irq_free(if_ctx_t ctx, if_irq_t irq) 6343 { 6344 6345 if (irq->ii_tag) 6346 bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag); 6347 6348 if (irq->ii_res) 6349 bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, 6350 rman_get_rid(irq->ii_res), irq->ii_res); 6351 } 6352 6353 static int 6354 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name) 6355 { 6356 iflib_txq_t txq = ctx->ifc_txqs; 6357 iflib_rxq_t rxq = ctx->ifc_rxqs; 6358 if_irq_t irq = &ctx->ifc_legacy_irq; 6359 iflib_filter_info_t info; 6360 device_t dev; 6361 struct grouptask *gtask; 6362 struct resource *res; 6363 int err, tqrid; 6364 bool rx_only; 6365 6366 info = &rxq->ifr_filter_info; 6367 gtask = &rxq->ifr_task; 6368 tqrid = *rid; 6369 rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0; 6370 6371 ctx->ifc_flags |= IFC_LEGACY; 6372 info->ifi_filter = filter; 6373 info->ifi_filter_arg = filter_arg; 6374 info->ifi_task = gtask; 6375 info->ifi_ctx = rxq; 6376 6377 dev = ctx->ifc_dev; 6378 /* We allocate a single interrupt resource */ 6379 err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr : 6380 iflib_fast_intr_rxtx, NULL, info, name); 6381 if (err != 0) 6382 return (err); 6383 NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq); 6384 res = irq->ii_res; 6385 taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name); 6386 6387 GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq); 6388 taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res, 6389 "tx"); 6390 return (0); 6391 } 6392 6393 void 6394 iflib_led_create(if_ctx_t ctx) 6395 { 6396 6397 ctx->ifc_led_dev = led_create(iflib_led_func, ctx, 6398 device_get_nameunit(ctx->ifc_dev)); 6399 } 6400 6401 void 6402 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid) 6403 { 6404 6405 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task); 6406 } 6407 6408 void 6409 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid) 6410 { 6411 6412 GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task); 6413 } 6414 6415 void 6416 iflib_admin_intr_deferred(if_ctx_t ctx) 6417 { 6418 6419 taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task); 6420 } 6421 6422 void 6423 iflib_iov_intr_deferred(if_ctx_t ctx) 6424 { 6425 6426 taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_vflr_task); 6427 } 6428 6429 void 6430 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name) 6431 { 6432 6433 taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL, 6434 name); 6435 } 6436 6437 void 6438 iflib_config_task_init(if_ctx_t ctx, struct task *config_task, task_fn_t *fn) 6439 { 6440 TASK_INIT(config_task, 0, fn, ctx); 6441 } 6442 6443 void 6444 iflib_config_task_enqueue(if_ctx_t ctx, struct task *config_task) 6445 { 6446 taskqueue_enqueue(ctx->ifc_tq, config_task); 6447 } 6448 6449 void 6450 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate) 6451 { 6452 if_t ifp = ctx->ifc_ifp; 6453 iflib_txq_t txq = ctx->ifc_txqs; 6454 6455 if_setbaudrate(ifp, baudrate); 6456 if (baudrate >= IF_Gbps(10)) { 6457 STATE_LOCK(ctx); 6458 ctx->ifc_flags |= IFC_PREFETCH; 6459 STATE_UNLOCK(ctx); 6460 } 6461 /* If link down, disable watchdog */ 6462 if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { 6463 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) 6464 txq->ift_qstatus = IFLIB_QUEUE_IDLE; 6465 } 6466 ctx->ifc_link_state = link_state; 6467 if_link_state_change(ifp, link_state); 6468 } 6469 6470 static int 6471 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq) 6472 { 6473 int credits; 6474 #ifdef INVARIANTS 6475 int credits_pre = txq->ift_cidx_processed; 6476 #endif 6477 6478 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, 6479 BUS_DMASYNC_POSTREAD); 6480 if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0) 6481 return (0); 6482 6483 txq->ift_processed += credits; 6484 txq->ift_cidx_processed += credits; 6485 6486 MPASS(credits_pre + credits == txq->ift_cidx_processed); 6487 if (txq->ift_cidx_processed >= txq->ift_size) 6488 txq->ift_cidx_processed -= txq->ift_size; 6489 return (credits); 6490 } 6491 6492 static int 6493 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget) 6494 { 6495 iflib_fl_t fl; 6496 u_int i; 6497 6498 for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++) 6499 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, 6500 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); 6501 return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx, 6502 budget)); 6503 } 6504 6505 void 6506 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name, 6507 const char *description, if_int_delay_info_t info, 6508 int offset, int value) 6509 { 6510 info->iidi_ctx = ctx; 6511 info->iidi_offset = offset; 6512 info->iidi_value = value; 6513 SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev), 6514 SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)), 6515 OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 6516 info, 0, iflib_sysctl_int_delay, "I", description); 6517 } 6518 6519 struct sx * 6520 iflib_ctx_lock_get(if_ctx_t ctx) 6521 { 6522 6523 return (&ctx->ifc_ctx_sx); 6524 } 6525 6526 static int 6527 iflib_msix_init(if_ctx_t ctx) 6528 { 6529 device_t dev = ctx->ifc_dev; 6530 if_shared_ctx_t sctx = ctx->ifc_sctx; 6531 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 6532 int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues; 6533 int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors; 6534 6535 iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs; 6536 iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs; 6537 6538 if (bootverbose) 6539 device_printf(dev, "msix_init qsets capped at %d\n", 6540 imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets)); 6541 6542 /* Override by tuneable */ 6543 if (scctx->isc_disable_msix) 6544 goto msi; 6545 6546 /* First try MSI-X */ 6547 if ((msgs = pci_msix_count(dev)) == 0) { 6548 if (bootverbose) 6549 device_printf(dev, "MSI-X not supported or disabled\n"); 6550 goto msi; 6551 } 6552 6553 bar = ctx->ifc_softc_ctx.isc_msix_bar; 6554 /* 6555 * bar == -1 => "trust me I know what I'm doing" 6556 * Some drivers are for hardware that is so shoddily 6557 * documented that no one knows which bars are which 6558 * so the developer has to map all bars. This hack 6559 * allows shoddy garbage to use MSI-X in this framework. 6560 */ 6561 if (bar != -1) { 6562 ctx->ifc_msix_mem = bus_alloc_resource_any(dev, 6563 SYS_RES_MEMORY, &bar, RF_ACTIVE); 6564 if (ctx->ifc_msix_mem == NULL) { 6565 device_printf(dev, "Unable to map MSI-X table\n"); 6566 goto msi; 6567 } 6568 } 6569 6570 admincnt = sctx->isc_admin_intrcnt; 6571 #if IFLIB_DEBUG 6572 /* use only 1 qset in debug mode */ 6573 queuemsgs = min(msgs - admincnt, 1); 6574 #else 6575 queuemsgs = msgs - admincnt; 6576 #endif 6577 #ifdef RSS 6578 queues = imin(queuemsgs, rss_getnumbuckets()); 6579 #else 6580 queues = queuemsgs; 6581 #endif 6582 queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues); 6583 if (bootverbose) 6584 device_printf(dev, 6585 "intr CPUs: %d queue msgs: %d admincnt: %d\n", 6586 CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt); 6587 #ifdef RSS 6588 /* If we're doing RSS, clamp at the number of RSS buckets */ 6589 if (queues > rss_getnumbuckets()) 6590 queues = rss_getnumbuckets(); 6591 #endif 6592 if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt) 6593 rx_queues = iflib_num_rx_queues; 6594 else 6595 rx_queues = queues; 6596 6597 if (rx_queues > scctx->isc_nrxqsets) 6598 rx_queues = scctx->isc_nrxqsets; 6599 6600 /* 6601 * We want this to be all logical CPUs by default 6602 */ 6603 if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) 6604 tx_queues = iflib_num_tx_queues; 6605 else 6606 tx_queues = mp_ncpus; 6607 6608 if (tx_queues > scctx->isc_ntxqsets) 6609 tx_queues = scctx->isc_ntxqsets; 6610 6611 if (ctx->ifc_sysctl_qs_eq_override == 0) { 6612 #ifdef INVARIANTS 6613 if (tx_queues != rx_queues) 6614 device_printf(dev, 6615 "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n", 6616 min(rx_queues, tx_queues), min(rx_queues, tx_queues)); 6617 #endif 6618 tx_queues = min(rx_queues, tx_queues); 6619 rx_queues = min(rx_queues, tx_queues); 6620 } 6621 6622 vectors = rx_queues + admincnt; 6623 if (msgs < vectors) { 6624 device_printf(dev, 6625 "insufficient number of MSI-X vectors " 6626 "(supported %d, need %d)\n", msgs, vectors); 6627 goto msi; 6628 } 6629 6630 device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues, 6631 tx_queues); 6632 msgs = vectors; 6633 if ((err = pci_alloc_msix(dev, &vectors)) == 0) { 6634 if (vectors != msgs) { 6635 device_printf(dev, 6636 "Unable to allocate sufficient MSI-X vectors " 6637 "(got %d, need %d)\n", vectors, msgs); 6638 pci_release_msi(dev); 6639 if (bar != -1) { 6640 bus_release_resource(dev, SYS_RES_MEMORY, bar, 6641 ctx->ifc_msix_mem); 6642 ctx->ifc_msix_mem = NULL; 6643 } 6644 goto msi; 6645 } 6646 device_printf(dev, "Using MSI-X interrupts with %d vectors\n", 6647 vectors); 6648 scctx->isc_vectors = vectors; 6649 scctx->isc_nrxqsets = rx_queues; 6650 scctx->isc_ntxqsets = tx_queues; 6651 scctx->isc_intr = IFLIB_INTR_MSIX; 6652 6653 return (vectors); 6654 } else { 6655 device_printf(dev, 6656 "failed to allocate %d MSI-X vectors, err: %d\n", vectors, 6657 err); 6658 if (bar != -1) { 6659 bus_release_resource(dev, SYS_RES_MEMORY, bar, 6660 ctx->ifc_msix_mem); 6661 ctx->ifc_msix_mem = NULL; 6662 } 6663 } 6664 6665 msi: 6666 vectors = pci_msi_count(dev); 6667 scctx->isc_nrxqsets = 1; 6668 scctx->isc_ntxqsets = 1; 6669 scctx->isc_vectors = vectors; 6670 if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) { 6671 device_printf(dev, "Using an MSI interrupt\n"); 6672 scctx->isc_intr = IFLIB_INTR_MSI; 6673 } else { 6674 scctx->isc_vectors = 1; 6675 device_printf(dev, "Using a Legacy interrupt\n"); 6676 scctx->isc_intr = IFLIB_INTR_LEGACY; 6677 } 6678 6679 return (vectors); 6680 } 6681 6682 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" }; 6683 6684 static int 6685 mp_ring_state_handler(SYSCTL_HANDLER_ARGS) 6686 { 6687 int rc; 6688 uint16_t *state = ((uint16_t *)oidp->oid_arg1); 6689 struct sbuf *sb; 6690 const char *ring_state = "UNKNOWN"; 6691 6692 /* XXX needed ? */ 6693 rc = sysctl_wire_old_buffer(req, 0); 6694 MPASS(rc == 0); 6695 if (rc != 0) 6696 return (rc); 6697 sb = sbuf_new_for_sysctl(NULL, NULL, 80, req); 6698 MPASS(sb != NULL); 6699 if (sb == NULL) 6700 return (ENOMEM); 6701 if (state[3] <= 3) 6702 ring_state = ring_states[state[3]]; 6703 6704 sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s", 6705 state[0], state[1], state[2], ring_state); 6706 rc = sbuf_finish(sb); 6707 sbuf_delete(sb); 6708 return (rc); 6709 } 6710 6711 enum iflib_ndesc_handler { 6712 IFLIB_NTXD_HANDLER, 6713 IFLIB_NRXD_HANDLER, 6714 }; 6715 6716 static int 6717 mp_ndesc_handler(SYSCTL_HANDLER_ARGS) 6718 { 6719 if_ctx_t ctx = (void *)arg1; 6720 enum iflib_ndesc_handler type = arg2; 6721 char buf[256] = {0}; 6722 qidx_t *ndesc; 6723 char *p, *next; 6724 int nqs, rc, i; 6725 6726 nqs = 8; 6727 switch (type) { 6728 case IFLIB_NTXD_HANDLER: 6729 ndesc = ctx->ifc_sysctl_ntxds; 6730 if (ctx->ifc_sctx) 6731 nqs = ctx->ifc_sctx->isc_ntxqs; 6732 break; 6733 case IFLIB_NRXD_HANDLER: 6734 ndesc = ctx->ifc_sysctl_nrxds; 6735 if (ctx->ifc_sctx) 6736 nqs = ctx->ifc_sctx->isc_nrxqs; 6737 break; 6738 default: 6739 printf("%s: unhandled type\n", __func__); 6740 return (EINVAL); 6741 } 6742 if (nqs == 0) 6743 nqs = 8; 6744 6745 for (i = 0; i < 8; i++) { 6746 if (i >= nqs) 6747 break; 6748 if (i) 6749 strcat(buf, ","); 6750 sprintf(strchr(buf, 0), "%d", ndesc[i]); 6751 } 6752 6753 rc = sysctl_handle_string(oidp, buf, sizeof(buf), req); 6754 if (rc || req->newptr == NULL) 6755 return (rc); 6756 6757 for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p; 6758 i++, p = strsep(&next, " ,")) { 6759 ndesc[i] = strtoul(p, NULL, 10); 6760 } 6761 6762 return (rc); 6763 } 6764 6765 #define NAME_BUFLEN 32 6766 static void 6767 iflib_add_device_sysctl_pre(if_ctx_t ctx) 6768 { 6769 device_t dev = iflib_get_dev(ctx); 6770 struct sysctl_oid_list *child, *oid_list; 6771 struct sysctl_ctx_list *ctx_list; 6772 struct sysctl_oid *node; 6773 6774 ctx_list = device_get_sysctl_ctx(dev); 6775 child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); 6776 ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, 6777 OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 6778 "IFLIB fields"); 6779 oid_list = SYSCTL_CHILDREN(node); 6780 6781 SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version", 6782 CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version"); 6783 6784 SYSCTL_ADD_BOOL(ctx_list, oid_list, OID_AUTO, "simple_tx", 6785 CTLFLAG_RDTUN, &ctx->ifc_sysctl_simple_tx, 0, 6786 "use simple tx ring"); 6787 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", 6788 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, 6789 "# of txqs to use, 0 => use default #"); 6790 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", 6791 CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0, 6792 "# of rxqs to use, 0 => use default #"); 6793 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable", 6794 CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0, 6795 "permit #txq != #rxq"); 6796 SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix", 6797 CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0, 6798 "disable MSI-X (default 0)"); 6799 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget", 6800 CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget"); 6801 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate", 6802 CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0, 6803 "cause TX to abdicate instead of running to completion"); 6804 ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED; 6805 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset", 6806 CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0, 6807 "offset to start using cores at"); 6808 SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx", 6809 CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0, 6810 "use separate cores for TX and RX"); 6811 SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores", 6812 CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0, 6813 "try to make use of logical cores for TX and RX"); 6814 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "use_extra_msix_vectors", 6815 CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0, 6816 "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later"); 6817 SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "allocated_msix_vectors", 6818 CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0, 6819 "total # of MSI-X vectors allocated by driver"); 6820 6821 /* XXX change for per-queue sizes */ 6822 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds", 6823 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx, 6824 IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A", 6825 "list of # of TX descriptors to use, 0 = use default #"); 6826 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds", 6827 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx, 6828 IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A", 6829 "list of # of RX descriptors to use, 0 = use default #"); 6830 } 6831 6832 static void 6833 iflib_add_device_sysctl_post(if_ctx_t ctx) 6834 { 6835 if_shared_ctx_t sctx = ctx->ifc_sctx; 6836 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; 6837 device_t dev = iflib_get_dev(ctx); 6838 struct sysctl_oid_list *child; 6839 struct sysctl_ctx_list *ctx_list; 6840 iflib_fl_t fl; 6841 iflib_txq_t txq; 6842 iflib_rxq_t rxq; 6843 int i, j; 6844 char namebuf[NAME_BUFLEN]; 6845 char *qfmt; 6846 struct sysctl_oid *queue_node, *fl_node, *node; 6847 struct sysctl_oid_list *queue_list, *fl_list; 6848 ctx_list = device_get_sysctl_ctx(dev); 6849 6850 node = ctx->ifc_sysctl_node; 6851 child = SYSCTL_CHILDREN(node); 6852 6853 if (scctx->isc_ntxqsets > 100) 6854 qfmt = "txq%03d"; 6855 else if (scctx->isc_ntxqsets > 10) 6856 qfmt = "txq%02d"; 6857 else 6858 qfmt = "txq%d"; 6859 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { 6860 snprintf(namebuf, NAME_BUFLEN, qfmt, i); 6861 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, 6862 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name"); 6863 queue_list = SYSCTL_CHILDREN(queue_node); 6864 SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu", 6865 CTLFLAG_RD, &txq->ift_task.gt_cpu, 0, 6866 "cpu this queue is bound to"); 6867 #if MEMORY_LOGGING 6868 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued", 6869 CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed"); 6870 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued", 6871 CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued"); 6872 #endif 6873 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag", 6874 CTLFLAG_RD, &txq->ift_mbuf_defrag, 6875 "# of times m_defrag was called"); 6876 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups", 6877 CTLFLAG_RD, &txq->ift_pullups, 6878 "# of times m_pullup was called"); 6879 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6880 "mbuf_defrag_failed", CTLFLAG_RD, 6881 &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); 6882 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6883 "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail, 6884 "# of times no descriptors were available"); 6885 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6886 "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed, 6887 "# of times DMA map failed"); 6888 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6889 "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig, 6890 "# of times txd_encap returned EFBIG"); 6891 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6892 "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup, 6893 "# of times map failed for other than EFBIG"); 6894 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx", 6895 CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index"); 6896 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx", 6897 CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index"); 6898 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, 6899 "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed, 6900 1, "Consumer Index seen by credit update"); 6901 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use", 6902 CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use"); 6903 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, 6904 "txq_processed", CTLFLAG_RD, &txq->ift_processed, 6905 "descriptors procesed for clean"); 6906 SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned", 6907 CTLFLAG_RD, &txq->ift_cleaned, "total cleaned"); 6908 SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state", 6909 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, 6910 __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0, 6911 mp_ring_state_handler, "A", "soft ring state"); 6912 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6913 "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues, 6914 "# of enqueues to the mp_ring for this queue"); 6915 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6916 "r_drops", CTLFLAG_RD, &txq->ift_br->drops, 6917 "# of drops in the mp_ring for this queue"); 6918 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6919 "r_starts", CTLFLAG_RD, &txq->ift_br->starts, 6920 "# of normal consumer starts in mp_ring for this queue"); 6921 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6922 "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls, 6923 "# of consumer stalls in the mp_ring for this queue"); 6924 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6925 "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts, 6926 "# of consumer restarts in the mp_ring for this queue"); 6927 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, 6928 "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications, 6929 "# of consumer abdications in the mp_ring for this queue"); 6930 } 6931 6932 if (scctx->isc_nrxqsets > 100) 6933 qfmt = "rxq%03d"; 6934 else if (scctx->isc_nrxqsets > 10) 6935 qfmt = "rxq%02d"; 6936 else 6937 qfmt = "rxq%d"; 6938 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { 6939 snprintf(namebuf, NAME_BUFLEN, qfmt, i); 6940 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, 6941 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name"); 6942 queue_list = SYSCTL_CHILDREN(queue_node); 6943 SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu", 6944 CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0, 6945 "cpu this queue is bound to"); 6946 if (sctx->isc_flags & IFLIB_HAS_RXCQ) { 6947 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, 6948 "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1, 6949 "Consumer Index"); 6950 } 6951 6952 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { 6953 snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j); 6954 fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, 6955 OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, 6956 NULL, "freelist Name"); 6957 fl_list = SYSCTL_CHILDREN(fl_node); 6958 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx", 6959 CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index"); 6960 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx", 6961 CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index"); 6962 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits", 6963 CTLFLAG_RD, &fl->ifl_credits, 1, 6964 "credits available"); 6965 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size", 6966 CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size"); 6967 #if MEMORY_LOGGING 6968 SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO, 6969 "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued, 6970 "mbufs allocated"); 6971 SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO, 6972 "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued, 6973 "mbufs freed"); 6974 SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO, 6975 "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued, 6976 "clusters allocated"); 6977 SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO, 6978 "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued, 6979 "clusters freed"); 6980 #endif 6981 } 6982 } 6983 6984 } 6985 6986 void 6987 iflib_request_reset(if_ctx_t ctx) 6988 { 6989 6990 STATE_LOCK(ctx); 6991 ctx->ifc_flags |= IFC_DO_RESET; 6992 STATE_UNLOCK(ctx); 6993 } 6994 6995 #ifndef __NO_STRICT_ALIGNMENT 6996 static struct mbuf * 6997 iflib_fixup_rx(struct mbuf *m) 6998 { 6999 struct mbuf *n; 7000 7001 if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) { 7002 bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len); 7003 m->m_data += ETHER_HDR_LEN; 7004 n = m; 7005 } else { 7006 MGETHDR(n, M_NOWAIT, MT_DATA); 7007 if (n == NULL) { 7008 m_freem(m); 7009 return (NULL); 7010 } 7011 bcopy(m->m_data, n->m_data, ETHER_HDR_LEN); 7012 m->m_data += ETHER_HDR_LEN; 7013 m->m_len -= ETHER_HDR_LEN; 7014 n->m_len = ETHER_HDR_LEN; 7015 M_MOVE_PKTHDR(n, m); 7016 n->m_next = m; 7017 } 7018 return (n); 7019 } 7020 #endif 7021 7022 #ifdef DEBUGNET 7023 static void 7024 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize) 7025 { 7026 if_ctx_t ctx; 7027 7028 ctx = if_getsoftc(ifp); 7029 CTX_LOCK(ctx); 7030 *nrxr = NRXQSETS(ctx); 7031 *ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size; 7032 *clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size; 7033 CTX_UNLOCK(ctx); 7034 } 7035 7036 static void 7037 iflib_debugnet_event(if_t ifp, enum debugnet_ev event) 7038 { 7039 if_ctx_t ctx; 7040 if_softc_ctx_t scctx; 7041 iflib_fl_t fl; 7042 iflib_rxq_t rxq; 7043 int i, j; 7044 7045 ctx = if_getsoftc(ifp); 7046 scctx = &ctx->ifc_softc_ctx; 7047 7048 switch (event) { 7049 case DEBUGNET_START: 7050 for (i = 0; i < scctx->isc_nrxqsets; i++) { 7051 rxq = &ctx->ifc_rxqs[i]; 7052 for (j = 0; j < rxq->ifr_nfl; j++) { 7053 fl = rxq->ifr_fl; 7054 fl->ifl_zone = m_getzone(fl->ifl_buf_size); 7055 } 7056 } 7057 iflib_no_tx_batch = 1; 7058 break; 7059 default: 7060 break; 7061 } 7062 } 7063 7064 static int 7065 iflib_debugnet_transmit(if_t ifp, struct mbuf *m) 7066 { 7067 if_ctx_t ctx; 7068 iflib_txq_t txq; 7069 int error; 7070 7071 ctx = if_getsoftc(ifp); 7072 if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != 7073 IFF_DRV_RUNNING) 7074 return (EBUSY); 7075 7076 txq = &ctx->ifc_txqs[0]; 7077 error = iflib_encap(txq, &m); 7078 if (error == 0) 7079 (void)iflib_txd_db_check(txq, true); 7080 return (error); 7081 } 7082 7083 static int 7084 iflib_debugnet_poll(if_t ifp, int count) 7085 { 7086 struct epoch_tracker et; 7087 if_ctx_t ctx; 7088 if_softc_ctx_t scctx; 7089 iflib_txq_t txq; 7090 int i; 7091 7092 ctx = if_getsoftc(ifp); 7093 scctx = &ctx->ifc_softc_ctx; 7094 7095 if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != 7096 IFF_DRV_RUNNING) 7097 return (EBUSY); 7098 7099 txq = &ctx->ifc_txqs[0]; 7100 (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 7101 7102 NET_EPOCH_ENTER(et); 7103 for (i = 0; i < scctx->isc_nrxqsets; i++) 7104 (void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */); 7105 NET_EPOCH_EXIT(et); 7106 return (0); 7107 } 7108 #endif /* DEBUGNET */ 7109 7110 7111 static inline iflib_txq_t 7112 iflib_simple_select_queue(if_ctx_t ctx, struct mbuf *m) 7113 { 7114 int qidx; 7115 7116 if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m)) 7117 qidx = QIDX(ctx, m); 7118 else 7119 qidx = NTXQSETS(ctx) + FIRST_QSET(ctx) - 1; 7120 return (&ctx->ifc_txqs[qidx]); 7121 } 7122 7123 static int 7124 iflib_simple_transmit(if_t ifp, struct mbuf *m) 7125 { 7126 if_ctx_t ctx; 7127 iflib_txq_t txq; 7128 int error; 7129 int bytes_sent = 0, pkt_sent = 0, mcast_sent = 0; 7130 7131 7132 ctx = if_getsoftc(ifp); 7133 if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != 7134 IFF_DRV_RUNNING) 7135 return (EBUSY); 7136 txq = iflib_simple_select_queue(ctx, m); 7137 mtx_lock(&txq->ift_mtx); 7138 error = iflib_encap(txq, &m); 7139 if (error == 0) { 7140 pkt_sent++; 7141 bytes_sent += m->m_pkthdr.len; 7142 mcast_sent += !!(m->m_flags & M_MCAST); 7143 (void)iflib_txd_db_check(txq, true); 7144 } 7145 (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); 7146 mtx_unlock(&txq->ift_mtx); 7147 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); 7148 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); 7149 if (mcast_sent) 7150 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); 7151 7152 return (error); 7153 } 7154