1 /*- 2 * Copyright (c) 2016-2017 Alexander Motin <mav@FreeBSD.org> 3 * Copyright (C) 2013 Intel Corporation 4 * Copyright (C) 2015 EMC Corporation 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 /* 30 * The Non-Transparent Bridge (NTB) is a device that allows you to connect 31 * two or more systems using a PCI-e links, providing remote memory access. 32 * 33 * This module contains a transport for sending and receiving messages by 34 * writing to remote memory window(s) provided by underlying NTB device. 35 * 36 * NOTE: Much of the code in this module is shared with Linux. Any patches may 37 * be picked up and redistributed in Linux with a dual GPL/BSD license. 38 */ 39 40 #include <sys/cdefs.h> 41 __FBSDID("$FreeBSD$"); 42 43 #include <sys/param.h> 44 #include <sys/kernel.h> 45 #include <sys/systm.h> 46 #include <sys/bus.h> 47 #include <sys/ktr.h> 48 #include <sys/limits.h> 49 #include <sys/lock.h> 50 #include <sys/malloc.h> 51 #include <sys/mbuf.h> 52 #include <sys/module.h> 53 #include <sys/mutex.h> 54 #include <sys/queue.h> 55 #include <sys/sysctl.h> 56 #include <sys/taskqueue.h> 57 58 #include <vm/vm.h> 59 #include <vm/pmap.h> 60 61 #include <machine/bus.h> 62 63 #include "ntb.h" 64 #include "ntb_transport.h" 65 66 #define KTR_NTB KTR_SPARE3 67 68 #define NTB_TRANSPORT_VERSION 4 69 70 static SYSCTL_NODE(_hw, OID_AUTO, ntb_transport, CTLFLAG_RW, 0, "ntb_transport"); 71 72 static unsigned g_ntb_transport_debug_level; 73 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, debug_level, CTLFLAG_RWTUN, 74 &g_ntb_transport_debug_level, 0, 75 "ntb_transport log level -- higher is more verbose"); 76 #define ntb_printf(lvl, ...) do { \ 77 if ((lvl) <= g_ntb_transport_debug_level) { \ 78 printf(__VA_ARGS__); \ 79 } \ 80 } while (0) 81 82 static unsigned transport_mtu = 0x10000; 83 84 static uint64_t max_mw_size; 85 SYSCTL_UQUAD(_hw_ntb_transport, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0, 86 "If enabled (non-zero), limit the size of large memory windows. " 87 "Both sides of the NTB MUST set the same value here."); 88 89 static unsigned enable_xeon_watchdog; 90 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN, 91 &enable_xeon_watchdog, 0, "If non-zero, write a register every second to " 92 "keep a watchdog from tearing down the NTB link"); 93 94 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry); 95 96 typedef uint32_t ntb_q_idx_t; 97 98 struct ntb_queue_entry { 99 /* ntb_queue list reference */ 100 STAILQ_ENTRY(ntb_queue_entry) entry; 101 102 /* info on data to be transferred */ 103 void *cb_data; 104 void *buf; 105 uint32_t len; 106 uint32_t flags; 107 108 struct ntb_transport_qp *qp; 109 struct ntb_payload_header *x_hdr; 110 ntb_q_idx_t index; 111 }; 112 113 struct ntb_rx_info { 114 ntb_q_idx_t entry; 115 }; 116 117 struct ntb_transport_qp { 118 struct ntb_transport_ctx *transport; 119 device_t dev; 120 121 void *cb_data; 122 123 bool client_ready; 124 volatile bool link_is_up; 125 uint8_t qp_num; /* Only 64 QPs are allowed. 0-63 */ 126 127 struct ntb_rx_info *rx_info; 128 struct ntb_rx_info *remote_rx_info; 129 130 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data, 131 void *data, int len); 132 struct ntb_queue_list tx_free_q; 133 struct mtx ntb_tx_free_q_lock; 134 caddr_t tx_mw; 135 bus_addr_t tx_mw_phys; 136 ntb_q_idx_t tx_index; 137 ntb_q_idx_t tx_max_entry; 138 uint64_t tx_max_frame; 139 140 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data, 141 void *data, int len); 142 struct ntb_queue_list rx_post_q; 143 struct ntb_queue_list rx_pend_q; 144 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */ 145 struct mtx ntb_rx_q_lock; 146 struct task rxc_db_work; 147 struct taskqueue *rxc_tq; 148 caddr_t rx_buff; 149 ntb_q_idx_t rx_index; 150 ntb_q_idx_t rx_max_entry; 151 uint64_t rx_max_frame; 152 153 void (*event_handler)(void *data, enum ntb_link_event status); 154 struct callout link_work; 155 struct callout rx_full; 156 157 uint64_t last_rx_no_buf; 158 159 /* Stats */ 160 uint64_t rx_bytes; 161 uint64_t rx_pkts; 162 uint64_t rx_ring_empty; 163 uint64_t rx_err_no_buf; 164 uint64_t rx_err_oflow; 165 uint64_t rx_err_ver; 166 uint64_t tx_bytes; 167 uint64_t tx_pkts; 168 uint64_t tx_ring_full; 169 uint64_t tx_err_no_buf; 170 171 struct mtx tx_lock; 172 }; 173 174 struct ntb_transport_mw { 175 vm_paddr_t phys_addr; 176 size_t phys_size; 177 size_t xlat_align; 178 size_t xlat_align_size; 179 bus_addr_t addr_limit; 180 /* Tx buff is off vbase / phys_addr */ 181 caddr_t vbase; 182 size_t xlat_size; 183 size_t buff_size; 184 /* Rx buff is off virt_addr / dma_addr */ 185 bus_dma_tag_t dma_tag; 186 bus_dmamap_t dma_map; 187 caddr_t virt_addr; 188 bus_addr_t dma_addr; 189 }; 190 191 struct ntb_transport_child { 192 device_t dev; 193 int consumer; 194 int qpoff; 195 int qpcnt; 196 struct ntb_transport_child *next; 197 }; 198 199 struct ntb_transport_ctx { 200 device_t dev; 201 struct ntb_transport_child *child; 202 struct ntb_transport_mw *mw_vec; 203 struct ntb_transport_qp *qp_vec; 204 unsigned mw_count; 205 unsigned qp_count; 206 uint64_t qp_bitmap; 207 volatile bool link_is_up; 208 enum ntb_speed link_speed; 209 enum ntb_width link_width; 210 struct callout link_work; 211 struct callout link_watchdog; 212 struct task link_cleanup; 213 }; 214 215 enum { 216 NTBT_DESC_DONE_FLAG = 1 << 0, 217 NTBT_LINK_DOWN_FLAG = 1 << 1, 218 }; 219 220 struct ntb_payload_header { 221 ntb_q_idx_t ver; 222 uint32_t len; 223 uint32_t flags; 224 }; 225 226 enum { 227 /* 228 * The order of this enum is part of the remote protocol. Do not 229 * reorder without bumping protocol version (and it's probably best 230 * to keep the protocol in lock-step with the Linux NTB driver. 231 */ 232 NTBT_VERSION = 0, 233 NTBT_QP_LINKS, 234 NTBT_NUM_QPS, 235 NTBT_NUM_MWS, 236 /* 237 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2. 238 */ 239 NTBT_MW0_SZ_HIGH, 240 NTBT_MW0_SZ_LOW, 241 NTBT_MW1_SZ_HIGH, 242 NTBT_MW1_SZ_LOW, 243 244 /* 245 * Some NTB-using hardware have a watchdog to work around NTB hangs; if 246 * a register or doorbell isn't written every few seconds, the link is 247 * torn down. Write an otherwise unused register every few seconds to 248 * work around this watchdog. 249 */ 250 NTBT_WATCHDOG_SPAD = 15 251 }; 252 253 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count) 254 #define NTB_QP_DEF_NUM_ENTRIES 100 255 #define NTB_LINK_DOWN_TIMEOUT 100 256 257 static int ntb_transport_probe(device_t dev); 258 static int ntb_transport_attach(device_t dev); 259 static int ntb_transport_detach(device_t dev); 260 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt, 261 unsigned int qp_num); 262 static int ntb_process_tx(struct ntb_transport_qp *qp, 263 struct ntb_queue_entry *entry); 264 static void ntb_transport_rxc_db(void *arg, int pending); 265 static int ntb_process_rxc(struct ntb_transport_qp *qp); 266 static void ntb_memcpy_rx(struct ntb_transport_qp *qp, 267 struct ntb_queue_entry *entry, void *offset); 268 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp, 269 void *data); 270 static void ntb_complete_rxc(struct ntb_transport_qp *qp); 271 static void ntb_transport_doorbell_callback(void *data, uint32_t vector); 272 static void ntb_transport_event_callback(void *data); 273 static void ntb_transport_link_work(void *arg); 274 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size); 275 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw); 276 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, 277 unsigned int qp_num); 278 static void ntb_qp_link_work(void *arg); 279 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt); 280 static void ntb_transport_link_cleanup_work(void *, int); 281 static void ntb_qp_link_down(struct ntb_transport_qp *qp); 282 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp); 283 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp); 284 static void ntb_send_link_down(struct ntb_transport_qp *qp); 285 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 286 struct ntb_queue_list *list); 287 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock, 288 struct ntb_queue_list *list); 289 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock, 290 struct ntb_queue_list *from, struct ntb_queue_list *to); 291 static void xeon_link_watchdog_hb(void *); 292 293 static const struct ntb_ctx_ops ntb_transport_ops = { 294 .link_event = ntb_transport_event_callback, 295 .db_event = ntb_transport_doorbell_callback, 296 }; 297 298 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver"); 299 300 static inline void 301 iowrite32(uint32_t val, void *addr) 302 { 303 304 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr, 305 val); 306 } 307 308 /* Transport Init and teardown */ 309 310 static void 311 xeon_link_watchdog_hb(void *arg) 312 { 313 struct ntb_transport_ctx *nt; 314 315 nt = arg; 316 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0); 317 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt); 318 } 319 320 static int 321 ntb_transport_probe(device_t dev) 322 { 323 324 device_set_desc(dev, "NTB Transport"); 325 return (0); 326 } 327 328 static int 329 ntb_transport_attach(device_t dev) 330 { 331 struct ntb_transport_ctx *nt = device_get_softc(dev); 332 struct ntb_transport_child **cpp = &nt->child; 333 struct ntb_transport_child *nc; 334 struct ntb_transport_mw *mw; 335 uint64_t db_bitmap; 336 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt; 337 char cfg[128] = ""; 338 char buf[32]; 339 char *n, *np, *c, *name; 340 341 nt->dev = dev; 342 nt->mw_count = ntb_mw_count(dev); 343 spad_count = ntb_spad_count(dev); 344 db_bitmap = ntb_db_valid_mask(dev); 345 db_count = flsll(db_bitmap); 346 KASSERT(db_bitmap == (1 << db_count) - 1, 347 ("Doorbells are not sequential (%jx).\n", db_bitmap)); 348 349 if (nt->mw_count == 0) { 350 device_printf(dev, "At least 1 memory window required.\n"); 351 return (ENXIO); 352 } 353 if (spad_count < 6) { 354 device_printf(dev, "At least 6 scratchpads required.\n"); 355 return (ENXIO); 356 } 357 if (spad_count < 4 + 2 * nt->mw_count) { 358 nt->mw_count = (spad_count - 4) / 2; 359 device_printf(dev, "Scratchpads enough only for %d " 360 "memory windows.\n", nt->mw_count); 361 } 362 if (db_bitmap == 0) { 363 device_printf(dev, "At least one doorbell required.\n"); 364 return (ENXIO); 365 } 366 367 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T, 368 M_WAITOK | M_ZERO); 369 for (i = 0; i < nt->mw_count; i++) { 370 mw = &nt->mw_vec[i]; 371 372 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase, 373 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size, 374 &mw->addr_limit); 375 if (rc != 0) 376 goto err; 377 378 mw->buff_size = 0; 379 mw->xlat_size = 0; 380 mw->virt_addr = NULL; 381 mw->dma_addr = 0; 382 383 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING); 384 if (rc) 385 ntb_printf(0, "Unable to set mw%d caching\n", i); 386 } 387 388 qpu = 0; 389 qpo = imin(db_count, nt->mw_count); 390 qpt = db_count; 391 392 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev), 393 device_get_unit(dev)); 394 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg)); 395 n = cfg; 396 i = 0; 397 while ((c = strsep(&n, ",")) != NULL) { 398 np = c; 399 name = strsep(&np, ":"); 400 if (name != NULL && name[0] == 0) 401 name = NULL; 402 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu; 403 if (qp <= 0) 404 qp = 1; 405 406 if (qp > qpt - qpu) { 407 device_printf(dev, "Not enough resources for config\n"); 408 break; 409 } 410 411 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO); 412 nc->consumer = i; 413 nc->qpoff = qpu; 414 nc->qpcnt = qp; 415 nc->dev = device_add_child(dev, name, -1); 416 if (nc->dev == NULL) { 417 device_printf(dev, "Can not add child.\n"); 418 break; 419 } 420 device_set_ivars(nc->dev, nc); 421 *cpp = nc; 422 cpp = &nc->next; 423 424 if (bootverbose) { 425 device_printf(dev, "%d \"%s\": queues %d", 426 i, name, qpu); 427 if (qp > 1) 428 printf("-%d", qpu + qp - 1); 429 printf("\n"); 430 } 431 432 qpu += qp; 433 i++; 434 } 435 nt->qp_count = qpu; 436 437 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T, 438 M_WAITOK | M_ZERO); 439 440 for (i = 0; i < nt->qp_count; i++) 441 ntb_transport_init_queue(nt, i); 442 443 callout_init(&nt->link_work, 0); 444 callout_init(&nt->link_watchdog, 0); 445 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt); 446 nt->link_is_up = false; 447 448 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops); 449 if (rc != 0) 450 goto err; 451 452 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO); 453 454 if (enable_xeon_watchdog != 0) 455 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt); 456 457 bus_generic_attach(dev); 458 return (0); 459 460 err: 461 free(nt->qp_vec, M_NTB_T); 462 free(nt->mw_vec, M_NTB_T); 463 return (rc); 464 } 465 466 static int 467 ntb_transport_detach(device_t dev) 468 { 469 struct ntb_transport_ctx *nt = device_get_softc(dev); 470 struct ntb_transport_child **cpp = &nt->child; 471 struct ntb_transport_child *nc; 472 int error = 0, i; 473 474 while ((nc = *cpp) != NULL) { 475 *cpp = (*cpp)->next; 476 error = device_delete_child(dev, nc->dev); 477 if (error) 478 break; 479 free(nc, M_DEVBUF); 480 } 481 KASSERT(nt->qp_bitmap == 0, 482 ("Some queues not freed on detach (%jx)", nt->qp_bitmap)); 483 484 ntb_transport_link_cleanup(nt); 485 taskqueue_drain(taskqueue_swi, &nt->link_cleanup); 486 callout_drain(&nt->link_work); 487 callout_drain(&nt->link_watchdog); 488 489 ntb_link_disable(dev); 490 ntb_clear_ctx(dev); 491 492 for (i = 0; i < nt->mw_count; i++) 493 ntb_free_mw(nt, i); 494 495 free(nt->qp_vec, M_NTB_T); 496 free(nt->mw_vec, M_NTB_T); 497 return (0); 498 } 499 500 static int 501 ntb_transport_print_child(device_t dev, device_t child) 502 { 503 struct ntb_transport_child *nc = device_get_ivars(child); 504 int retval; 505 506 retval = bus_print_child_header(dev, child); 507 if (nc->qpcnt > 0) { 508 printf(" queue %d", nc->qpoff); 509 if (nc->qpcnt > 1) 510 printf("-%d", nc->qpoff + nc->qpcnt - 1); 511 } 512 retval += printf(" at consumer %d", nc->consumer); 513 retval += bus_print_child_domain(dev, child); 514 retval += bus_print_child_footer(dev, child); 515 516 return (retval); 517 } 518 519 static int 520 ntb_transport_child_location_str(device_t dev, device_t child, char *buf, 521 size_t buflen) 522 { 523 struct ntb_transport_child *nc = device_get_ivars(child); 524 525 snprintf(buf, buflen, "consumer=%d", nc->consumer); 526 return (0); 527 } 528 529 int 530 ntb_transport_queue_count(device_t dev) 531 { 532 struct ntb_transport_child *nc = device_get_ivars(dev); 533 534 return (nc->qpcnt); 535 } 536 537 static void 538 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num) 539 { 540 struct ntb_transport_mw *mw; 541 struct ntb_transport_qp *qp; 542 vm_paddr_t mw_base; 543 uint64_t mw_size, qp_offset; 544 size_t tx_size; 545 unsigned num_qps_mw, mw_num, mw_count; 546 547 mw_count = nt->mw_count; 548 mw_num = QP_TO_MW(nt, qp_num); 549 mw = &nt->mw_vec[mw_num]; 550 551 qp = &nt->qp_vec[qp_num]; 552 qp->qp_num = qp_num; 553 qp->transport = nt; 554 qp->dev = nt->dev; 555 qp->client_ready = false; 556 qp->event_handler = NULL; 557 ntb_qp_link_down_reset(qp); 558 559 if (mw_num < nt->qp_count % mw_count) 560 num_qps_mw = nt->qp_count / mw_count + 1; 561 else 562 num_qps_mw = nt->qp_count / mw_count; 563 564 mw_base = mw->phys_addr; 565 mw_size = mw->phys_size; 566 567 tx_size = mw_size / num_qps_mw; 568 qp_offset = tx_size * (qp_num / mw_count); 569 570 qp->tx_mw = mw->vbase + qp_offset; 571 KASSERT(qp->tx_mw != NULL, ("uh oh?")); 572 573 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */ 574 qp->tx_mw_phys = mw_base + qp_offset; 575 KASSERT(qp->tx_mw_phys != 0, ("uh oh?")); 576 577 tx_size -= sizeof(struct ntb_rx_info); 578 qp->rx_info = (void *)(qp->tx_mw + tx_size); 579 580 /* Due to house-keeping, there must be at least 2 buffs */ 581 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2); 582 qp->tx_max_entry = tx_size / qp->tx_max_frame; 583 584 callout_init(&qp->link_work, 0); 585 callout_init(&qp->rx_full, 1); 586 587 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN); 588 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN); 589 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF); 590 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp); 591 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK, 592 taskqueue_thread_enqueue, &qp->rxc_tq); 593 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d", 594 device_get_nameunit(nt->dev), qp_num); 595 596 STAILQ_INIT(&qp->rx_post_q); 597 STAILQ_INIT(&qp->rx_pend_q); 598 STAILQ_INIT(&qp->tx_free_q); 599 } 600 601 void 602 ntb_transport_free_queue(struct ntb_transport_qp *qp) 603 { 604 struct ntb_transport_ctx *nt = qp->transport; 605 struct ntb_queue_entry *entry; 606 607 callout_drain(&qp->link_work); 608 609 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 610 taskqueue_drain_all(qp->rxc_tq); 611 taskqueue_free(qp->rxc_tq); 612 613 qp->cb_data = NULL; 614 qp->rx_handler = NULL; 615 qp->tx_handler = NULL; 616 qp->event_handler = NULL; 617 618 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) 619 free(entry, M_NTB_T); 620 621 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) 622 free(entry, M_NTB_T); 623 624 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q))) 625 free(entry, M_NTB_T); 626 627 nt->qp_bitmap &= ~(1 << qp->qp_num); 628 } 629 630 /** 631 * ntb_transport_create_queue - Create a new NTB transport layer queue 632 * @rx_handler: receive callback function 633 * @tx_handler: transmit callback function 634 * @event_handler: event callback function 635 * 636 * Create a new NTB transport layer queue and provide the queue with a callback 637 * routine for both transmit and receive. The receive callback routine will be 638 * used to pass up data when the transport has received it on the queue. The 639 * transmit callback routine will be called when the transport has completed the 640 * transmission of the data on the queue and the data is ready to be freed. 641 * 642 * RETURNS: pointer to newly created ntb_queue, NULL on error. 643 */ 644 struct ntb_transport_qp * 645 ntb_transport_create_queue(device_t dev, int q, 646 const struct ntb_queue_handlers *handlers, void *data) 647 { 648 struct ntb_transport_child *nc = device_get_ivars(dev); 649 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev)); 650 struct ntb_queue_entry *entry; 651 struct ntb_transport_qp *qp; 652 int i; 653 654 if (q < 0 || q >= nc->qpcnt) 655 return (NULL); 656 657 qp = &nt->qp_vec[nc->qpoff + q]; 658 nt->qp_bitmap |= (1 << qp->qp_num); 659 qp->cb_data = data; 660 qp->rx_handler = handlers->rx_handler; 661 qp->tx_handler = handlers->tx_handler; 662 qp->event_handler = handlers->event_handler; 663 664 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 665 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 666 entry->cb_data = data; 667 entry->buf = NULL; 668 entry->len = transport_mtu; 669 entry->qp = qp; 670 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q); 671 } 672 673 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 674 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 675 entry->qp = qp; 676 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 677 } 678 679 ntb_db_clear(dev, 1ull << qp->qp_num); 680 return (qp); 681 } 682 683 /** 684 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue 685 * @qp: NTB transport layer queue to be enabled 686 * 687 * Notify NTB transport layer of client readiness to use queue 688 */ 689 void 690 ntb_transport_link_up(struct ntb_transport_qp *qp) 691 { 692 struct ntb_transport_ctx *nt = qp->transport; 693 694 qp->client_ready = true; 695 696 ntb_printf(2, "qp %d client ready\n", qp->qp_num); 697 698 if (nt->link_is_up) 699 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 700 } 701 702 703 704 /* Transport Tx */ 705 706 /** 707 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry 708 * @qp: NTB transport layer queue the entry is to be enqueued on 709 * @cb: per buffer pointer for callback function to use 710 * @data: pointer to data buffer that will be sent 711 * @len: length of the data buffer 712 * 713 * Enqueue a new transmit buffer onto the transport queue from which a NTB 714 * payload will be transmitted. This assumes that a lock is being held to 715 * serialize access to the qp. 716 * 717 * RETURNS: An appropriate ERRNO error value on error, or zero for success. 718 */ 719 int 720 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data, 721 unsigned int len) 722 { 723 struct ntb_queue_entry *entry; 724 int rc; 725 726 if (!qp->link_is_up || len == 0) { 727 CTR0(KTR_NTB, "TX: link not up"); 728 return (EINVAL); 729 } 730 731 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 732 if (entry == NULL) { 733 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q"); 734 qp->tx_err_no_buf++; 735 return (EBUSY); 736 } 737 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry); 738 739 entry->cb_data = cb; 740 entry->buf = data; 741 entry->len = len; 742 entry->flags = 0; 743 744 mtx_lock(&qp->tx_lock); 745 rc = ntb_process_tx(qp, entry); 746 mtx_unlock(&qp->tx_lock); 747 if (rc != 0) { 748 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 749 CTR1(KTR_NTB, 750 "TX: process_tx failed. Returning entry %p to tx_free_q", 751 entry); 752 } 753 return (rc); 754 } 755 756 static void 757 ntb_tx_copy_callback(void *data) 758 { 759 struct ntb_queue_entry *entry = data; 760 struct ntb_transport_qp *qp = entry->qp; 761 struct ntb_payload_header *hdr = entry->x_hdr; 762 763 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags); 764 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr); 765 766 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num); 767 768 /* 769 * The entry length can only be zero if the packet is intended to be a 770 * "link down" or similar. Since no payload is being sent in these 771 * cases, there is nothing to add to the completion queue. 772 */ 773 if (entry->len > 0) { 774 qp->tx_bytes += entry->len; 775 776 if (qp->tx_handler) 777 qp->tx_handler(qp, qp->cb_data, entry->buf, 778 entry->len); 779 else 780 m_freem(entry->buf); 781 entry->buf = NULL; 782 } 783 784 CTR3(KTR_NTB, 785 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning " 786 "to tx_free_q", entry, hdr->ver, hdr->flags); 787 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 788 } 789 790 static void 791 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset) 792 { 793 794 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset); 795 if (entry->buf != NULL) { 796 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset); 797 798 /* 799 * Ensure that the data is fully copied before setting the 800 * flags 801 */ 802 wmb(); 803 } 804 805 ntb_tx_copy_callback(entry); 806 } 807 808 static void 809 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 810 { 811 struct ntb_payload_header *hdr; 812 void *offset; 813 814 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index; 815 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame - 816 sizeof(struct ntb_payload_header)); 817 entry->x_hdr = hdr; 818 819 iowrite32(entry->len, &hdr->len); 820 iowrite32(qp->tx_pkts, &hdr->ver); 821 822 ntb_memcpy_tx(entry, offset); 823 } 824 825 static int 826 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 827 { 828 829 CTR3(KTR_NTB, 830 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u", 831 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry); 832 if (qp->tx_index == qp->remote_rx_info->entry) { 833 CTR0(KTR_NTB, "TX: ring full"); 834 qp->tx_ring_full++; 835 return (EAGAIN); 836 } 837 838 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) { 839 if (qp->tx_handler != NULL) 840 qp->tx_handler(qp, qp->cb_data, entry->buf, 841 EIO); 842 else 843 m_freem(entry->buf); 844 845 entry->buf = NULL; 846 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 847 CTR1(KTR_NTB, 848 "TX: frame too big. returning entry %p to tx_free_q", 849 entry); 850 return (0); 851 } 852 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index); 853 ntb_async_tx(qp, entry); 854 855 qp->tx_index++; 856 qp->tx_index %= qp->tx_max_entry; 857 858 qp->tx_pkts++; 859 860 return (0); 861 } 862 863 /* Transport Rx */ 864 static void 865 ntb_transport_rxc_db(void *arg, int pending __unused) 866 { 867 struct ntb_transport_qp *qp = arg; 868 uint64_t qp_mask = 1ull << qp->qp_num; 869 int rc; 870 871 CTR0(KTR_NTB, "RX: transport_rx"); 872 again: 873 while ((rc = ntb_process_rxc(qp)) == 0) 874 ; 875 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc); 876 877 if ((ntb_db_read(qp->dev) & qp_mask) != 0) { 878 /* If db is set, clear it and check queue once more. */ 879 ntb_db_clear(qp->dev, qp_mask); 880 goto again; 881 } 882 if (qp->link_is_up) 883 ntb_db_clear_mask(qp->dev, qp_mask); 884 } 885 886 static int 887 ntb_process_rxc(struct ntb_transport_qp *qp) 888 { 889 struct ntb_payload_header *hdr; 890 struct ntb_queue_entry *entry; 891 caddr_t offset; 892 893 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index; 894 hdr = (void *)(offset + qp->rx_max_frame - 895 sizeof(struct ntb_payload_header)); 896 897 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index); 898 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) { 899 CTR0(KTR_NTB, "RX: hdr not done"); 900 qp->rx_ring_empty++; 901 return (EAGAIN); 902 } 903 904 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) { 905 CTR0(KTR_NTB, "RX: link down"); 906 ntb_qp_link_down(qp); 907 hdr->flags = 0; 908 return (EAGAIN); 909 } 910 911 if (hdr->ver != (uint32_t)qp->rx_pkts) { 912 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). " 913 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts); 914 qp->rx_err_ver++; 915 return (EIO); 916 } 917 918 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q); 919 if (entry == NULL) { 920 qp->rx_err_no_buf++; 921 CTR0(KTR_NTB, "RX: No entries in rx_pend_q"); 922 return (EAGAIN); 923 } 924 callout_stop(&qp->rx_full); 925 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry); 926 927 entry->x_hdr = hdr; 928 entry->index = qp->rx_index; 929 930 if (hdr->len > entry->len) { 931 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju", 932 (uintmax_t)hdr->len, (uintmax_t)entry->len); 933 qp->rx_err_oflow++; 934 935 entry->len = -EIO; 936 entry->flags |= NTBT_DESC_DONE_FLAG; 937 938 ntb_complete_rxc(qp); 939 } else { 940 qp->rx_bytes += hdr->len; 941 qp->rx_pkts++; 942 943 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts); 944 945 entry->len = hdr->len; 946 947 ntb_memcpy_rx(qp, entry, offset); 948 } 949 950 qp->rx_index++; 951 qp->rx_index %= qp->rx_max_entry; 952 return (0); 953 } 954 955 static void 956 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry, 957 void *offset) 958 { 959 struct ifnet *ifp = entry->cb_data; 960 unsigned int len = entry->len; 961 962 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset); 963 964 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL); 965 if (entry->buf == NULL) 966 entry->len = -ENOMEM; 967 968 /* Ensure that the data is globally visible before clearing the flag */ 969 wmb(); 970 971 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf); 972 ntb_rx_copy_callback(qp, entry); 973 } 974 975 static inline void 976 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data) 977 { 978 struct ntb_queue_entry *entry; 979 980 entry = data; 981 entry->flags |= NTBT_DESC_DONE_FLAG; 982 ntb_complete_rxc(qp); 983 } 984 985 static void 986 ntb_complete_rxc(struct ntb_transport_qp *qp) 987 { 988 struct ntb_queue_entry *entry; 989 struct mbuf *m; 990 unsigned len; 991 992 CTR0(KTR_NTB, "RX: rx_completion_task"); 993 994 mtx_lock_spin(&qp->ntb_rx_q_lock); 995 996 while (!STAILQ_EMPTY(&qp->rx_post_q)) { 997 entry = STAILQ_FIRST(&qp->rx_post_q); 998 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0) 999 break; 1000 1001 entry->x_hdr->flags = 0; 1002 iowrite32(entry->index, &qp->rx_info->entry); 1003 1004 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry); 1005 1006 len = entry->len; 1007 m = entry->buf; 1008 1009 /* 1010 * Re-initialize queue_entry for reuse; rx_handler takes 1011 * ownership of the mbuf. 1012 */ 1013 entry->buf = NULL; 1014 entry->len = transport_mtu; 1015 entry->cb_data = qp->cb_data; 1016 1017 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry); 1018 1019 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1020 1021 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m); 1022 if (qp->rx_handler != NULL && qp->client_ready) 1023 qp->rx_handler(qp, qp->cb_data, m, len); 1024 else 1025 m_freem(m); 1026 1027 mtx_lock_spin(&qp->ntb_rx_q_lock); 1028 } 1029 1030 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1031 } 1032 1033 static void 1034 ntb_transport_doorbell_callback(void *data, uint32_t vector) 1035 { 1036 struct ntb_transport_ctx *nt = data; 1037 struct ntb_transport_qp *qp; 1038 uint64_t vec_mask; 1039 unsigned qp_num; 1040 1041 vec_mask = ntb_db_vector_mask(nt->dev, vector); 1042 vec_mask &= nt->qp_bitmap; 1043 if ((vec_mask & (vec_mask - 1)) != 0) 1044 vec_mask &= ntb_db_read(nt->dev); 1045 if (vec_mask != 0) { 1046 ntb_db_set_mask(nt->dev, vec_mask); 1047 ntb_db_clear(nt->dev, vec_mask); 1048 } 1049 while (vec_mask != 0) { 1050 qp_num = ffsll(vec_mask) - 1; 1051 1052 qp = &nt->qp_vec[qp_num]; 1053 if (qp->link_is_up) 1054 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work); 1055 1056 vec_mask &= ~(1ull << qp_num); 1057 } 1058 } 1059 1060 /* Link Event handler */ 1061 static void 1062 ntb_transport_event_callback(void *data) 1063 { 1064 struct ntb_transport_ctx *nt = data; 1065 1066 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) { 1067 ntb_printf(1, "HW link up\n"); 1068 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt); 1069 } else { 1070 ntb_printf(1, "HW link down\n"); 1071 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup); 1072 } 1073 } 1074 1075 /* Link bring up */ 1076 static void 1077 ntb_transport_link_work(void *arg) 1078 { 1079 struct ntb_transport_ctx *nt = arg; 1080 device_t dev = nt->dev; 1081 struct ntb_transport_qp *qp; 1082 uint64_t val64, size; 1083 uint32_t val; 1084 unsigned i; 1085 int rc; 1086 1087 /* send the local info, in the opposite order of the way we read it */ 1088 for (i = 0; i < nt->mw_count; i++) { 1089 size = nt->mw_vec[i].phys_size; 1090 1091 if (max_mw_size != 0 && size > max_mw_size) 1092 size = max_mw_size; 1093 1094 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2), 1095 size >> 32); 1096 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size); 1097 } 1098 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count); 1099 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count); 1100 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0); 1101 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION); 1102 1103 /* Query the remote side for its info */ 1104 val = 0; 1105 ntb_spad_read(dev, NTBT_VERSION, &val); 1106 if (val != NTB_TRANSPORT_VERSION) 1107 goto out; 1108 1109 ntb_spad_read(dev, NTBT_NUM_QPS, &val); 1110 if (val != nt->qp_count) 1111 goto out; 1112 1113 ntb_spad_read(dev, NTBT_NUM_MWS, &val); 1114 if (val != nt->mw_count) 1115 goto out; 1116 1117 for (i = 0; i < nt->mw_count; i++) { 1118 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val); 1119 val64 = (uint64_t)val << 32; 1120 1121 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val); 1122 val64 |= val; 1123 1124 rc = ntb_set_mw(nt, i, val64); 1125 if (rc != 0) 1126 goto free_mws; 1127 } 1128 1129 nt->link_is_up = true; 1130 ntb_printf(1, "transport link up\n"); 1131 1132 for (i = 0; i < nt->qp_count; i++) { 1133 qp = &nt->qp_vec[i]; 1134 1135 ntb_transport_setup_qp_mw(nt, i); 1136 1137 if (qp->client_ready) 1138 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 1139 } 1140 1141 return; 1142 1143 free_mws: 1144 for (i = 0; i < nt->mw_count; i++) 1145 ntb_free_mw(nt, i); 1146 out: 1147 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width)) 1148 callout_reset(&nt->link_work, 1149 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt); 1150 } 1151 1152 struct ntb_load_cb_args { 1153 bus_addr_t addr; 1154 int error; 1155 }; 1156 1157 static void 1158 ntb_load_cb(void *xsc, bus_dma_segment_t *segs, int nsegs, int error) 1159 { 1160 struct ntb_load_cb_args *cba = (struct ntb_load_cb_args *)xsc; 1161 1162 if (!(cba->error = error)) 1163 cba->addr = segs[0].ds_addr; 1164 } 1165 1166 static int 1167 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size) 1168 { 1169 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1170 struct ntb_load_cb_args cba; 1171 size_t xlat_size, buff_size; 1172 int rc; 1173 1174 if (size == 0) 1175 return (EINVAL); 1176 1177 xlat_size = roundup(size, mw->xlat_align_size); 1178 buff_size = xlat_size; 1179 1180 /* No need to re-setup */ 1181 if (mw->xlat_size == xlat_size) 1182 return (0); 1183 1184 if (mw->buff_size != 0) 1185 ntb_free_mw(nt, num_mw); 1186 1187 /* Alloc memory for receiving data. Must be aligned */ 1188 mw->xlat_size = xlat_size; 1189 mw->buff_size = buff_size; 1190 1191 if (bus_dma_tag_create(bus_get_dma_tag(nt->dev), mw->xlat_align, 0, 1192 mw->addr_limit, BUS_SPACE_MAXADDR, 1193 NULL, NULL, mw->buff_size, 1, mw->buff_size, 1194 0, NULL, NULL, &mw->dma_tag)) { 1195 ntb_printf(0, "Unable to create MW tag of size %zu/%zu\n", 1196 mw->buff_size, mw->xlat_size); 1197 mw->xlat_size = 0; 1198 mw->buff_size = 0; 1199 return (ENOMEM); 1200 } 1201 if (bus_dmamem_alloc(mw->dma_tag, (void **)&mw->virt_addr, 1202 BUS_DMA_WAITOK | BUS_DMA_ZERO, &mw->dma_map)) { 1203 bus_dma_tag_destroy(mw->dma_tag); 1204 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n", 1205 mw->buff_size, mw->xlat_size); 1206 mw->xlat_size = 0; 1207 mw->buff_size = 0; 1208 return (ENOMEM); 1209 } 1210 if (bus_dmamap_load(mw->dma_tag, mw->dma_map, mw->virt_addr, 1211 mw->buff_size, ntb_load_cb, &cba, BUS_DMA_NOWAIT) || cba.error) { 1212 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1213 bus_dma_tag_destroy(mw->dma_tag); 1214 ntb_printf(0, "Unable to load MW buffer of size %zu/%zu\n", 1215 mw->buff_size, mw->xlat_size); 1216 mw->xlat_size = 0; 1217 mw->buff_size = 0; 1218 return (ENOMEM); 1219 } 1220 mw->dma_addr = cba.addr; 1221 1222 /* Notify HW the memory location of the receive buffer */ 1223 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size); 1224 if (rc) { 1225 ntb_printf(0, "Unable to set mw%d translation\n", num_mw); 1226 ntb_free_mw(nt, num_mw); 1227 return (rc); 1228 } 1229 1230 return (0); 1231 } 1232 1233 static void 1234 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw) 1235 { 1236 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1237 1238 if (mw->virt_addr == NULL) 1239 return; 1240 1241 ntb_mw_clear_trans(nt->dev, num_mw); 1242 bus_dmamap_unload(mw->dma_tag, mw->dma_map); 1243 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1244 bus_dma_tag_destroy(mw->dma_tag); 1245 mw->xlat_size = 0; 1246 mw->buff_size = 0; 1247 mw->virt_addr = NULL; 1248 } 1249 1250 static int 1251 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num) 1252 { 1253 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num]; 1254 struct ntb_transport_mw *mw; 1255 void *offset; 1256 ntb_q_idx_t i; 1257 size_t rx_size; 1258 unsigned num_qps_mw, mw_num, mw_count; 1259 1260 mw_count = nt->mw_count; 1261 mw_num = QP_TO_MW(nt, qp_num); 1262 mw = &nt->mw_vec[mw_num]; 1263 1264 if (mw->virt_addr == NULL) 1265 return (ENOMEM); 1266 1267 if (mw_num < nt->qp_count % mw_count) 1268 num_qps_mw = nt->qp_count / mw_count + 1; 1269 else 1270 num_qps_mw = nt->qp_count / mw_count; 1271 1272 rx_size = mw->xlat_size / num_qps_mw; 1273 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count); 1274 rx_size -= sizeof(struct ntb_rx_info); 1275 1276 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size); 1277 1278 /* Due to house-keeping, there must be at least 2 buffs */ 1279 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2); 1280 qp->rx_max_entry = rx_size / qp->rx_max_frame; 1281 qp->rx_index = 0; 1282 1283 qp->remote_rx_info->entry = qp->rx_max_entry - 1; 1284 1285 /* Set up the hdr offsets with 0s */ 1286 for (i = 0; i < qp->rx_max_entry; i++) { 1287 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) - 1288 sizeof(struct ntb_payload_header)); 1289 memset(offset, 0, sizeof(struct ntb_payload_header)); 1290 } 1291 1292 qp->rx_pkts = 0; 1293 qp->tx_pkts = 0; 1294 qp->tx_index = 0; 1295 1296 return (0); 1297 } 1298 1299 static void 1300 ntb_qp_link_work(void *arg) 1301 { 1302 struct ntb_transport_qp *qp = arg; 1303 device_t dev = qp->dev; 1304 struct ntb_transport_ctx *nt = qp->transport; 1305 int i; 1306 uint32_t val; 1307 1308 /* Report queues that are up on our side */ 1309 for (i = 0, val = 0; i < nt->qp_count; i++) { 1310 if (nt->qp_vec[i].client_ready) 1311 val |= (1 << i); 1312 } 1313 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val); 1314 1315 /* See if the remote side is up */ 1316 ntb_spad_read(dev, NTBT_QP_LINKS, &val); 1317 if ((val & (1ull << qp->qp_num)) != 0) { 1318 ntb_printf(2, "qp %d link up\n", qp->qp_num); 1319 qp->link_is_up = true; 1320 1321 if (qp->event_handler != NULL) 1322 qp->event_handler(qp->cb_data, NTB_LINK_UP); 1323 1324 ntb_db_clear_mask(dev, 1ull << qp->qp_num); 1325 } else if (nt->link_is_up) 1326 callout_reset(&qp->link_work, 1327 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp); 1328 } 1329 1330 /* Link down event*/ 1331 static void 1332 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt) 1333 { 1334 struct ntb_transport_qp *qp; 1335 int i; 1336 1337 callout_drain(&nt->link_work); 1338 nt->link_is_up = 0; 1339 1340 /* Pass along the info to any clients */ 1341 for (i = 0; i < nt->qp_count; i++) { 1342 if ((nt->qp_bitmap & (1 << i)) != 0) { 1343 qp = &nt->qp_vec[i]; 1344 ntb_qp_link_cleanup(qp); 1345 callout_drain(&qp->link_work); 1346 } 1347 } 1348 1349 /* 1350 * The scratchpad registers keep the values if the remote side 1351 * goes down, blast them now to give them a sane value the next 1352 * time they are accessed 1353 */ 1354 ntb_spad_clear(nt->dev); 1355 } 1356 1357 static void 1358 ntb_transport_link_cleanup_work(void *arg, int pending __unused) 1359 { 1360 1361 ntb_transport_link_cleanup(arg); 1362 } 1363 1364 static void 1365 ntb_qp_link_down(struct ntb_transport_qp *qp) 1366 { 1367 1368 ntb_qp_link_cleanup(qp); 1369 } 1370 1371 static void 1372 ntb_qp_link_down_reset(struct ntb_transport_qp *qp) 1373 { 1374 1375 qp->link_is_up = false; 1376 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 1377 1378 qp->tx_index = qp->rx_index = 0; 1379 qp->tx_bytes = qp->rx_bytes = 0; 1380 qp->tx_pkts = qp->rx_pkts = 0; 1381 1382 qp->rx_ring_empty = 0; 1383 qp->tx_ring_full = 0; 1384 1385 qp->rx_err_no_buf = qp->tx_err_no_buf = 0; 1386 qp->rx_err_oflow = qp->rx_err_ver = 0; 1387 } 1388 1389 static void 1390 ntb_qp_link_cleanup(struct ntb_transport_qp *qp) 1391 { 1392 1393 callout_drain(&qp->link_work); 1394 ntb_qp_link_down_reset(qp); 1395 1396 if (qp->event_handler != NULL) 1397 qp->event_handler(qp->cb_data, NTB_LINK_DOWN); 1398 } 1399 1400 /* Link commanded down */ 1401 /** 1402 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data 1403 * @qp: NTB transport layer queue to be disabled 1404 * 1405 * Notify NTB transport layer of client's desire to no longer receive data on 1406 * transport queue specified. It is the client's responsibility to ensure all 1407 * entries on queue are purged or otherwise handled appropriately. 1408 */ 1409 void 1410 ntb_transport_link_down(struct ntb_transport_qp *qp) 1411 { 1412 struct ntb_transport_ctx *nt = qp->transport; 1413 int i; 1414 uint32_t val; 1415 1416 qp->client_ready = false; 1417 for (i = 0, val = 0; i < nt->qp_count; i++) { 1418 if (nt->qp_vec[i].client_ready) 1419 val |= (1 << i); 1420 } 1421 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val); 1422 1423 if (qp->link_is_up) 1424 ntb_send_link_down(qp); 1425 else 1426 callout_drain(&qp->link_work); 1427 } 1428 1429 /** 1430 * ntb_transport_link_query - Query transport link state 1431 * @qp: NTB transport layer queue to be queried 1432 * 1433 * Query connectivity to the remote system of the NTB transport queue 1434 * 1435 * RETURNS: true for link up or false for link down 1436 */ 1437 bool 1438 ntb_transport_link_query(struct ntb_transport_qp *qp) 1439 { 1440 1441 return (qp->link_is_up); 1442 } 1443 1444 /** 1445 * ntb_transport_link_speed - Query transport link speed 1446 * @qp: NTB transport layer queue to be queried 1447 * 1448 * Query connection speed to the remote system of the NTB transport queue 1449 * 1450 * RETURNS: link speed in bits per second 1451 */ 1452 uint64_t 1453 ntb_transport_link_speed(struct ntb_transport_qp *qp) 1454 { 1455 struct ntb_transport_ctx *nt = qp->transport; 1456 uint64_t rate; 1457 1458 if (!nt->link_is_up) 1459 return (0); 1460 switch (nt->link_speed) { 1461 case NTB_SPEED_GEN1: 1462 rate = 2500000000 * 8 / 10; 1463 break; 1464 case NTB_SPEED_GEN2: 1465 rate = 5000000000 * 8 / 10; 1466 break; 1467 case NTB_SPEED_GEN3: 1468 rate = 8000000000 * 128 / 130; 1469 break; 1470 case NTB_SPEED_GEN4: 1471 rate = 16000000000 * 128 / 130; 1472 break; 1473 default: 1474 return (0); 1475 } 1476 if (nt->link_width <= 0) 1477 return (0); 1478 return (rate * nt->link_width); 1479 } 1480 1481 static void 1482 ntb_send_link_down(struct ntb_transport_qp *qp) 1483 { 1484 struct ntb_queue_entry *entry; 1485 int i, rc; 1486 1487 if (!qp->link_is_up) 1488 return; 1489 1490 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) { 1491 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 1492 if (entry != NULL) 1493 break; 1494 pause("NTB Wait for link down", hz / 10); 1495 } 1496 1497 if (entry == NULL) 1498 return; 1499 1500 entry->cb_data = NULL; 1501 entry->buf = NULL; 1502 entry->len = 0; 1503 entry->flags = NTBT_LINK_DOWN_FLAG; 1504 1505 mtx_lock(&qp->tx_lock); 1506 rc = ntb_process_tx(qp, entry); 1507 mtx_unlock(&qp->tx_lock); 1508 if (rc != 0) 1509 printf("ntb: Failed to send link down\n"); 1510 1511 ntb_qp_link_down_reset(qp); 1512 } 1513 1514 1515 /* List Management */ 1516 1517 static void 1518 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 1519 struct ntb_queue_list *list) 1520 { 1521 1522 mtx_lock_spin(lock); 1523 STAILQ_INSERT_TAIL(list, entry, entry); 1524 mtx_unlock_spin(lock); 1525 } 1526 1527 static struct ntb_queue_entry * 1528 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list) 1529 { 1530 struct ntb_queue_entry *entry; 1531 1532 mtx_lock_spin(lock); 1533 if (STAILQ_EMPTY(list)) { 1534 entry = NULL; 1535 goto out; 1536 } 1537 entry = STAILQ_FIRST(list); 1538 STAILQ_REMOVE_HEAD(list, entry); 1539 out: 1540 mtx_unlock_spin(lock); 1541 1542 return (entry); 1543 } 1544 1545 static struct ntb_queue_entry * 1546 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from, 1547 struct ntb_queue_list *to) 1548 { 1549 struct ntb_queue_entry *entry; 1550 1551 mtx_lock_spin(lock); 1552 if (STAILQ_EMPTY(from)) { 1553 entry = NULL; 1554 goto out; 1555 } 1556 entry = STAILQ_FIRST(from); 1557 STAILQ_REMOVE_HEAD(from, entry); 1558 STAILQ_INSERT_TAIL(to, entry, entry); 1559 1560 out: 1561 mtx_unlock_spin(lock); 1562 return (entry); 1563 } 1564 1565 /** 1566 * ntb_transport_qp_num - Query the qp number 1567 * @qp: NTB transport layer queue to be queried 1568 * 1569 * Query qp number of the NTB transport queue 1570 * 1571 * RETURNS: a zero based number specifying the qp number 1572 */ 1573 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp) 1574 { 1575 1576 return (qp->qp_num); 1577 } 1578 1579 /** 1580 * ntb_transport_max_size - Query the max payload size of a qp 1581 * @qp: NTB transport layer queue to be queried 1582 * 1583 * Query the maximum payload size permissible on the given qp 1584 * 1585 * RETURNS: the max payload size of a qp 1586 */ 1587 unsigned int 1588 ntb_transport_max_size(struct ntb_transport_qp *qp) 1589 { 1590 1591 return (qp->tx_max_frame - sizeof(struct ntb_payload_header)); 1592 } 1593 1594 unsigned int 1595 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp) 1596 { 1597 unsigned int head = qp->tx_index; 1598 unsigned int tail = qp->remote_rx_info->entry; 1599 1600 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head); 1601 } 1602 1603 static device_method_t ntb_transport_methods[] = { 1604 /* Device interface */ 1605 DEVMETHOD(device_probe, ntb_transport_probe), 1606 DEVMETHOD(device_attach, ntb_transport_attach), 1607 DEVMETHOD(device_detach, ntb_transport_detach), 1608 /* Bus interface */ 1609 DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str), 1610 DEVMETHOD(bus_print_child, ntb_transport_print_child), 1611 DEVMETHOD_END 1612 }; 1613 1614 devclass_t ntb_transport_devclass; 1615 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver, 1616 ntb_transport_methods, sizeof(struct ntb_transport_ctx)); 1617 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver, 1618 ntb_transport_devclass, NULL, NULL); 1619 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1); 1620 MODULE_VERSION(ntb_transport, 1); 1621