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, 71 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 72 "ntb_transport"); 73 74 static unsigned g_ntb_transport_debug_level; 75 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, debug_level, CTLFLAG_RWTUN, 76 &g_ntb_transport_debug_level, 0, 77 "ntb_transport log level -- higher is more verbose"); 78 #define ntb_printf(lvl, ...) do { \ 79 if ((lvl) <= g_ntb_transport_debug_level) { \ 80 printf(__VA_ARGS__); \ 81 } \ 82 } while (0) 83 84 static unsigned transport_mtu = 0x10000; 85 86 static uint64_t max_mw_size = 256*1024*1024; 87 SYSCTL_UQUAD(_hw_ntb_transport, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0, 88 "If enabled (non-zero), limit the size of large memory windows. " 89 "Both sides of the NTB MUST set the same value here."); 90 91 static unsigned enable_xeon_watchdog; 92 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN, 93 &enable_xeon_watchdog, 0, "If non-zero, write a register every second to " 94 "keep a watchdog from tearing down the NTB link"); 95 96 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry); 97 98 typedef uint32_t ntb_q_idx_t; 99 100 struct ntb_queue_entry { 101 /* ntb_queue list reference */ 102 STAILQ_ENTRY(ntb_queue_entry) entry; 103 104 /* info on data to be transferred */ 105 void *cb_data; 106 void *buf; 107 uint32_t len; 108 uint32_t flags; 109 110 struct ntb_transport_qp *qp; 111 struct ntb_payload_header *x_hdr; 112 ntb_q_idx_t index; 113 }; 114 115 struct ntb_rx_info { 116 ntb_q_idx_t entry; 117 }; 118 119 struct ntb_transport_qp { 120 struct ntb_transport_ctx *transport; 121 device_t dev; 122 123 void *cb_data; 124 125 bool client_ready; 126 volatile bool link_is_up; 127 uint8_t qp_num; /* Only 64 QPs are allowed. 0-63 */ 128 129 struct ntb_rx_info *rx_info; 130 struct ntb_rx_info *remote_rx_info; 131 132 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data, 133 void *data, int len); 134 struct ntb_queue_list tx_free_q; 135 struct mtx ntb_tx_free_q_lock; 136 caddr_t tx_mw; 137 bus_addr_t tx_mw_phys; 138 ntb_q_idx_t tx_index; 139 ntb_q_idx_t tx_max_entry; 140 uint64_t tx_max_frame; 141 142 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data, 143 void *data, int len); 144 struct ntb_queue_list rx_post_q; 145 struct ntb_queue_list rx_pend_q; 146 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */ 147 struct mtx ntb_rx_q_lock; 148 struct task rxc_db_work; 149 struct taskqueue *rxc_tq; 150 caddr_t rx_buff; 151 ntb_q_idx_t rx_index; 152 ntb_q_idx_t rx_max_entry; 153 uint64_t rx_max_frame; 154 155 void (*event_handler)(void *data, enum ntb_link_event status); 156 struct callout link_work; 157 struct callout rx_full; 158 159 uint64_t last_rx_no_buf; 160 161 /* Stats */ 162 uint64_t rx_bytes; 163 uint64_t rx_pkts; 164 uint64_t rx_ring_empty; 165 uint64_t rx_err_no_buf; 166 uint64_t rx_err_oflow; 167 uint64_t rx_err_ver; 168 uint64_t tx_bytes; 169 uint64_t tx_pkts; 170 uint64_t tx_ring_full; 171 uint64_t tx_err_no_buf; 172 173 struct mtx tx_lock; 174 }; 175 176 struct ntb_transport_mw { 177 vm_paddr_t phys_addr; 178 size_t phys_size; 179 size_t xlat_align; 180 size_t xlat_align_size; 181 bus_addr_t addr_limit; 182 /* Tx buff is vbase / phys_addr / tx_size */ 183 caddr_t vbase; 184 size_t tx_size; 185 /* Rx buff is virt_addr / dma_addr / rx_size */ 186 bus_dma_tag_t dma_tag; 187 bus_dmamap_t dma_map; 188 caddr_t virt_addr; 189 bus_addr_t dma_addr; 190 size_t rx_size; 191 /* rx_size increased to size alignment requirements of the hardware. */ 192 size_t buff_size; 193 }; 194 195 struct ntb_transport_child { 196 device_t dev; 197 int consumer; 198 int qpoff; 199 int qpcnt; 200 struct ntb_transport_child *next; 201 }; 202 203 struct ntb_transport_ctx { 204 device_t dev; 205 struct ntb_transport_child *child; 206 struct ntb_transport_mw *mw_vec; 207 struct ntb_transport_qp *qp_vec; 208 int compact; 209 unsigned mw_count; 210 unsigned qp_count; 211 uint64_t qp_bitmap; 212 volatile bool link_is_up; 213 enum ntb_speed link_speed; 214 enum ntb_width link_width; 215 struct callout link_work; 216 struct callout link_watchdog; 217 struct task link_cleanup; 218 }; 219 220 enum { 221 NTBT_DESC_DONE_FLAG = 1 << 0, 222 NTBT_LINK_DOWN_FLAG = 1 << 1, 223 }; 224 225 struct ntb_payload_header { 226 ntb_q_idx_t ver; 227 uint32_t len; 228 uint32_t flags; 229 }; 230 231 enum { 232 /* 233 * The order of this enum is part of the remote protocol. Do not 234 * reorder without bumping protocol version (and it's probably best 235 * to keep the protocol in lock-step with the Linux NTB driver. 236 */ 237 NTBT_VERSION = 0, 238 NTBT_QP_LINKS, 239 NTBT_NUM_QPS, 240 NTBT_NUM_MWS, 241 /* 242 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2. 243 */ 244 NTBT_MW0_SZ_HIGH, 245 NTBT_MW0_SZ_LOW, 246 NTBT_MW1_SZ_HIGH, 247 NTBT_MW1_SZ_LOW, 248 249 /* 250 * Some NTB-using hardware have a watchdog to work around NTB hangs; if 251 * a register or doorbell isn't written every few seconds, the link is 252 * torn down. Write an otherwise unused register every few seconds to 253 * work around this watchdog. 254 */ 255 NTBT_WATCHDOG_SPAD = 15 256 }; 257 258 /* 259 * Compart version of sratchpad protocol, using twice less registers. 260 */ 261 enum { 262 NTBTC_PARAMS = 0, /* NUM_QPS << 24 + NUM_MWS << 16 + VERSION */ 263 NTBTC_QP_LINKS, /* QP links status */ 264 NTBTC_MW0_SZ, /* MW size limited to 32 bits. */ 265 }; 266 267 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count) 268 #define NTB_QP_DEF_NUM_ENTRIES 100 269 #define NTB_LINK_DOWN_TIMEOUT 100 270 271 static int ntb_transport_probe(device_t dev); 272 static int ntb_transport_attach(device_t dev); 273 static int ntb_transport_detach(device_t dev); 274 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt, 275 unsigned int qp_num); 276 static int ntb_process_tx(struct ntb_transport_qp *qp, 277 struct ntb_queue_entry *entry); 278 static void ntb_transport_rxc_db(void *arg, int pending); 279 static int ntb_process_rxc(struct ntb_transport_qp *qp); 280 static void ntb_memcpy_rx(struct ntb_transport_qp *qp, 281 struct ntb_queue_entry *entry, void *offset); 282 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp, 283 void *data); 284 static void ntb_complete_rxc(struct ntb_transport_qp *qp); 285 static void ntb_transport_doorbell_callback(void *data, uint32_t vector); 286 static void ntb_transport_event_callback(void *data); 287 static void ntb_transport_link_work(void *arg); 288 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size); 289 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw); 290 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, 291 unsigned int qp_num); 292 static void ntb_qp_link_work(void *arg); 293 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt); 294 static void ntb_transport_link_cleanup_work(void *, int); 295 static void ntb_qp_link_down(struct ntb_transport_qp *qp); 296 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp); 297 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp); 298 static void ntb_send_link_down(struct ntb_transport_qp *qp); 299 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 300 struct ntb_queue_list *list); 301 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock, 302 struct ntb_queue_list *list); 303 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock, 304 struct ntb_queue_list *from, struct ntb_queue_list *to); 305 static void xeon_link_watchdog_hb(void *); 306 307 static const struct ntb_ctx_ops ntb_transport_ops = { 308 .link_event = ntb_transport_event_callback, 309 .db_event = ntb_transport_doorbell_callback, 310 }; 311 312 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver"); 313 314 static inline void 315 iowrite32(uint32_t val, void *addr) 316 { 317 318 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr, 319 val); 320 } 321 322 /* Transport Init and teardown */ 323 324 static void 325 xeon_link_watchdog_hb(void *arg) 326 { 327 struct ntb_transport_ctx *nt; 328 329 nt = arg; 330 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0); 331 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt); 332 } 333 334 static int 335 ntb_transport_probe(device_t dev) 336 { 337 338 device_set_desc(dev, "NTB Transport"); 339 return (0); 340 } 341 342 static int 343 ntb_transport_attach(device_t dev) 344 { 345 struct ntb_transport_ctx *nt = device_get_softc(dev); 346 struct ntb_transport_child **cpp = &nt->child; 347 struct ntb_transport_child *nc; 348 struct ntb_transport_mw *mw; 349 uint64_t db_bitmap; 350 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt; 351 char cfg[128] = ""; 352 char buf[32]; 353 char *n, *np, *c, *name; 354 355 nt->dev = dev; 356 nt->mw_count = ntb_mw_count(dev); 357 spad_count = ntb_spad_count(dev); 358 db_bitmap = ntb_db_valid_mask(dev); 359 db_count = flsll(db_bitmap); 360 KASSERT(db_bitmap == (1 << db_count) - 1, 361 ("Doorbells are not sequential (%jx).\n", db_bitmap)); 362 363 if (nt->mw_count == 0) { 364 device_printf(dev, "At least 1 memory window required.\n"); 365 return (ENXIO); 366 } 367 nt->compact = (spad_count < 4 + 2 * nt->mw_count); 368 snprintf(buf, sizeof(buf), "hint.%s.%d.compact", device_get_name(dev), 369 device_get_unit(dev)); 370 TUNABLE_INT_FETCH(buf, &nt->compact); 371 if (nt->compact) { 372 if (spad_count < 3) { 373 device_printf(dev, "At least 3 scratchpads required.\n"); 374 return (ENXIO); 375 } 376 if (spad_count < 2 + nt->mw_count) { 377 nt->mw_count = spad_count - 2; 378 device_printf(dev, "Scratchpads enough only for %d " 379 "memory windows.\n", nt->mw_count); 380 } 381 } else { 382 if (spad_count < 6) { 383 device_printf(dev, "At least 6 scratchpads required.\n"); 384 return (ENXIO); 385 } 386 if (spad_count < 4 + 2 * nt->mw_count) { 387 nt->mw_count = (spad_count - 4) / 2; 388 device_printf(dev, "Scratchpads enough only for %d " 389 "memory windows.\n", nt->mw_count); 390 } 391 } 392 if (db_bitmap == 0) { 393 device_printf(dev, "At least one doorbell required.\n"); 394 return (ENXIO); 395 } 396 397 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T, 398 M_WAITOK | M_ZERO); 399 for (i = 0; i < nt->mw_count; i++) { 400 mw = &nt->mw_vec[i]; 401 402 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase, 403 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size, 404 &mw->addr_limit); 405 if (rc != 0) 406 goto err; 407 408 mw->tx_size = mw->phys_size; 409 if (max_mw_size != 0 && mw->tx_size > max_mw_size) { 410 device_printf(dev, "Memory window %d limited from " 411 "%ju to %ju\n", i, (uintmax_t)mw->tx_size, 412 max_mw_size); 413 mw->tx_size = max_mw_size; 414 } 415 if (nt->compact && mw->tx_size > UINT32_MAX) { 416 device_printf(dev, "Memory window %d is too big " 417 "(%ju)\n", i, (uintmax_t)mw->tx_size); 418 rc = ENXIO; 419 goto err; 420 } 421 422 mw->rx_size = 0; 423 mw->buff_size = 0; 424 mw->virt_addr = NULL; 425 mw->dma_addr = 0; 426 427 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING); 428 if (rc) 429 ntb_printf(0, "Unable to set mw%d caching\n", i); 430 431 /* 432 * Try to preallocate receive memory early, since there may 433 * be not enough contiguous memory later. It is quite likely 434 * that NTB windows are symmetric and this allocation remain, 435 * but even if not, we will just reallocate it later. 436 */ 437 ntb_set_mw(nt, i, mw->tx_size); 438 } 439 440 qpu = 0; 441 qpo = imin(db_count, nt->mw_count); 442 qpt = db_count; 443 444 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev), 445 device_get_unit(dev)); 446 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg)); 447 n = cfg; 448 i = 0; 449 while ((c = strsep(&n, ",")) != NULL) { 450 np = c; 451 name = strsep(&np, ":"); 452 if (name != NULL && name[0] == 0) 453 name = NULL; 454 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu; 455 if (qp <= 0) 456 qp = 1; 457 458 if (qp > qpt - qpu) { 459 device_printf(dev, "Not enough resources for config\n"); 460 break; 461 } 462 463 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO); 464 nc->consumer = i; 465 nc->qpoff = qpu; 466 nc->qpcnt = qp; 467 nc->dev = device_add_child(dev, name, -1); 468 if (nc->dev == NULL) { 469 device_printf(dev, "Can not add child.\n"); 470 break; 471 } 472 device_set_ivars(nc->dev, nc); 473 *cpp = nc; 474 cpp = &nc->next; 475 476 if (bootverbose) { 477 device_printf(dev, "%d \"%s\": queues %d", 478 i, name, qpu); 479 if (qp > 1) 480 printf("-%d", qpu + qp - 1); 481 printf("\n"); 482 } 483 484 qpu += qp; 485 i++; 486 } 487 nt->qp_count = qpu; 488 489 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T, 490 M_WAITOK | M_ZERO); 491 492 for (i = 0; i < nt->qp_count; i++) 493 ntb_transport_init_queue(nt, i); 494 495 callout_init(&nt->link_work, 0); 496 callout_init(&nt->link_watchdog, 0); 497 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt); 498 nt->link_is_up = false; 499 500 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops); 501 if (rc != 0) 502 goto err; 503 504 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO); 505 506 for (i = 0; i < nt->mw_count; i++) { 507 mw = &nt->mw_vec[i]; 508 rc = ntb_mw_set_trans(nt->dev, i, mw->dma_addr, mw->buff_size); 509 if (rc != 0) 510 ntb_printf(0, "load time mw%d xlat fails, rc %d\n", i, rc); 511 } 512 513 if (enable_xeon_watchdog != 0) 514 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt); 515 516 bus_generic_attach(dev); 517 return (0); 518 519 err: 520 free(nt->qp_vec, M_NTB_T); 521 free(nt->mw_vec, M_NTB_T); 522 return (rc); 523 } 524 525 static int 526 ntb_transport_detach(device_t dev) 527 { 528 struct ntb_transport_ctx *nt = device_get_softc(dev); 529 struct ntb_transport_child **cpp = &nt->child; 530 struct ntb_transport_child *nc; 531 int error = 0, i; 532 533 while ((nc = *cpp) != NULL) { 534 *cpp = (*cpp)->next; 535 error = device_delete_child(dev, nc->dev); 536 if (error) 537 break; 538 free(nc, M_DEVBUF); 539 } 540 KASSERT(nt->qp_bitmap == 0, 541 ("Some queues not freed on detach (%jx)", nt->qp_bitmap)); 542 543 ntb_transport_link_cleanup(nt); 544 taskqueue_drain(taskqueue_swi, &nt->link_cleanup); 545 callout_drain(&nt->link_work); 546 callout_drain(&nt->link_watchdog); 547 548 ntb_link_disable(dev); 549 ntb_clear_ctx(dev); 550 551 for (i = 0; i < nt->mw_count; i++) 552 ntb_free_mw(nt, i); 553 554 free(nt->qp_vec, M_NTB_T); 555 free(nt->mw_vec, M_NTB_T); 556 return (0); 557 } 558 559 static int 560 ntb_transport_print_child(device_t dev, device_t child) 561 { 562 struct ntb_transport_child *nc = device_get_ivars(child); 563 int retval; 564 565 retval = bus_print_child_header(dev, child); 566 if (nc->qpcnt > 0) { 567 printf(" queue %d", nc->qpoff); 568 if (nc->qpcnt > 1) 569 printf("-%d", nc->qpoff + nc->qpcnt - 1); 570 } 571 retval += printf(" at consumer %d", nc->consumer); 572 retval += bus_print_child_domain(dev, child); 573 retval += bus_print_child_footer(dev, child); 574 575 return (retval); 576 } 577 578 static int 579 ntb_transport_child_location_str(device_t dev, device_t child, char *buf, 580 size_t buflen) 581 { 582 struct ntb_transport_child *nc = device_get_ivars(child); 583 584 snprintf(buf, buflen, "consumer=%d", nc->consumer); 585 return (0); 586 } 587 588 int 589 ntb_transport_queue_count(device_t dev) 590 { 591 struct ntb_transport_child *nc = device_get_ivars(dev); 592 593 return (nc->qpcnt); 594 } 595 596 static void 597 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num) 598 { 599 struct ntb_transport_mw *mw; 600 struct ntb_transport_qp *qp; 601 vm_paddr_t mw_base; 602 uint64_t qp_offset; 603 size_t tx_size; 604 unsigned num_qps_mw, mw_num, mw_count; 605 606 mw_count = nt->mw_count; 607 mw_num = QP_TO_MW(nt, qp_num); 608 mw = &nt->mw_vec[mw_num]; 609 610 qp = &nt->qp_vec[qp_num]; 611 qp->qp_num = qp_num; 612 qp->transport = nt; 613 qp->dev = nt->dev; 614 qp->client_ready = false; 615 qp->event_handler = NULL; 616 ntb_qp_link_down_reset(qp); 617 618 if (mw_num < nt->qp_count % mw_count) 619 num_qps_mw = nt->qp_count / mw_count + 1; 620 else 621 num_qps_mw = nt->qp_count / mw_count; 622 623 mw_base = mw->phys_addr; 624 625 tx_size = mw->tx_size / num_qps_mw; 626 qp_offset = tx_size * (qp_num / mw_count); 627 628 qp->tx_mw = mw->vbase + qp_offset; 629 KASSERT(qp->tx_mw != NULL, ("uh oh?")); 630 631 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */ 632 qp->tx_mw_phys = mw_base + qp_offset; 633 KASSERT(qp->tx_mw_phys != 0, ("uh oh?")); 634 635 tx_size -= sizeof(struct ntb_rx_info); 636 qp->rx_info = (void *)(qp->tx_mw + tx_size); 637 638 /* Due to house-keeping, there must be at least 2 buffs */ 639 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2); 640 qp->tx_max_entry = tx_size / qp->tx_max_frame; 641 642 callout_init(&qp->link_work, 0); 643 callout_init(&qp->rx_full, 1); 644 645 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN); 646 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN); 647 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF); 648 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp); 649 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK, 650 taskqueue_thread_enqueue, &qp->rxc_tq); 651 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d", 652 device_get_nameunit(nt->dev), qp_num); 653 654 STAILQ_INIT(&qp->rx_post_q); 655 STAILQ_INIT(&qp->rx_pend_q); 656 STAILQ_INIT(&qp->tx_free_q); 657 } 658 659 void 660 ntb_transport_free_queue(struct ntb_transport_qp *qp) 661 { 662 struct ntb_transport_ctx *nt = qp->transport; 663 struct ntb_queue_entry *entry; 664 665 callout_drain(&qp->link_work); 666 667 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 668 taskqueue_drain_all(qp->rxc_tq); 669 taskqueue_free(qp->rxc_tq); 670 671 qp->cb_data = NULL; 672 qp->rx_handler = NULL; 673 qp->tx_handler = NULL; 674 qp->event_handler = NULL; 675 676 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) 677 free(entry, M_NTB_T); 678 679 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) 680 free(entry, M_NTB_T); 681 682 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q))) 683 free(entry, M_NTB_T); 684 685 nt->qp_bitmap &= ~(1 << qp->qp_num); 686 } 687 688 /** 689 * ntb_transport_create_queue - Create a new NTB transport layer queue 690 * @rx_handler: receive callback function 691 * @tx_handler: transmit callback function 692 * @event_handler: event callback function 693 * 694 * Create a new NTB transport layer queue and provide the queue with a callback 695 * routine for both transmit and receive. The receive callback routine will be 696 * used to pass up data when the transport has received it on the queue. The 697 * transmit callback routine will be called when the transport has completed the 698 * transmission of the data on the queue and the data is ready to be freed. 699 * 700 * RETURNS: pointer to newly created ntb_queue, NULL on error. 701 */ 702 struct ntb_transport_qp * 703 ntb_transport_create_queue(device_t dev, int q, 704 const struct ntb_queue_handlers *handlers, void *data) 705 { 706 struct ntb_transport_child *nc = device_get_ivars(dev); 707 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev)); 708 struct ntb_queue_entry *entry; 709 struct ntb_transport_qp *qp; 710 int i; 711 712 if (q < 0 || q >= nc->qpcnt) 713 return (NULL); 714 715 qp = &nt->qp_vec[nc->qpoff + q]; 716 nt->qp_bitmap |= (1 << qp->qp_num); 717 qp->cb_data = data; 718 qp->rx_handler = handlers->rx_handler; 719 qp->tx_handler = handlers->tx_handler; 720 qp->event_handler = handlers->event_handler; 721 722 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 723 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 724 entry->cb_data = data; 725 entry->buf = NULL; 726 entry->len = transport_mtu; 727 entry->qp = qp; 728 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q); 729 } 730 731 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 732 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 733 entry->qp = qp; 734 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 735 } 736 737 ntb_db_clear(dev, 1ull << qp->qp_num); 738 return (qp); 739 } 740 741 /** 742 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue 743 * @qp: NTB transport layer queue to be enabled 744 * 745 * Notify NTB transport layer of client readiness to use queue 746 */ 747 void 748 ntb_transport_link_up(struct ntb_transport_qp *qp) 749 { 750 struct ntb_transport_ctx *nt = qp->transport; 751 752 qp->client_ready = true; 753 754 ntb_printf(2, "qp %d client ready\n", qp->qp_num); 755 756 if (nt->link_is_up) 757 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 758 } 759 760 761 762 /* Transport Tx */ 763 764 /** 765 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry 766 * @qp: NTB transport layer queue the entry is to be enqueued on 767 * @cb: per buffer pointer for callback function to use 768 * @data: pointer to data buffer that will be sent 769 * @len: length of the data buffer 770 * 771 * Enqueue a new transmit buffer onto the transport queue from which a NTB 772 * payload will be transmitted. This assumes that a lock is being held to 773 * serialize access to the qp. 774 * 775 * RETURNS: An appropriate ERRNO error value on error, or zero for success. 776 */ 777 int 778 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data, 779 unsigned int len) 780 { 781 struct ntb_queue_entry *entry; 782 int rc; 783 784 if (!qp->link_is_up || len == 0) { 785 CTR0(KTR_NTB, "TX: link not up"); 786 return (EINVAL); 787 } 788 789 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 790 if (entry == NULL) { 791 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q"); 792 qp->tx_err_no_buf++; 793 return (EBUSY); 794 } 795 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry); 796 797 entry->cb_data = cb; 798 entry->buf = data; 799 entry->len = len; 800 entry->flags = 0; 801 802 mtx_lock(&qp->tx_lock); 803 rc = ntb_process_tx(qp, entry); 804 mtx_unlock(&qp->tx_lock); 805 if (rc != 0) { 806 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 807 CTR1(KTR_NTB, 808 "TX: process_tx failed. Returning entry %p to tx_free_q", 809 entry); 810 } 811 return (rc); 812 } 813 814 static void 815 ntb_tx_copy_callback(void *data) 816 { 817 struct ntb_queue_entry *entry = data; 818 struct ntb_transport_qp *qp = entry->qp; 819 struct ntb_payload_header *hdr = entry->x_hdr; 820 821 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags); 822 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr); 823 824 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num); 825 826 /* 827 * The entry length can only be zero if the packet is intended to be a 828 * "link down" or similar. Since no payload is being sent in these 829 * cases, there is nothing to add to the completion queue. 830 */ 831 if (entry->len > 0) { 832 qp->tx_bytes += entry->len; 833 834 if (qp->tx_handler) 835 qp->tx_handler(qp, qp->cb_data, entry->buf, 836 entry->len); 837 else 838 m_freem(entry->buf); 839 entry->buf = NULL; 840 } 841 842 CTR3(KTR_NTB, 843 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning " 844 "to tx_free_q", entry, hdr->ver, hdr->flags); 845 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 846 } 847 848 static void 849 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset) 850 { 851 852 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset); 853 if (entry->buf != NULL) { 854 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset); 855 856 /* 857 * Ensure that the data is fully copied before setting the 858 * flags 859 */ 860 wmb(); 861 } 862 863 ntb_tx_copy_callback(entry); 864 } 865 866 static void 867 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 868 { 869 struct ntb_payload_header *hdr; 870 void *offset; 871 872 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index; 873 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame - 874 sizeof(struct ntb_payload_header)); 875 entry->x_hdr = hdr; 876 877 iowrite32(entry->len, &hdr->len); 878 iowrite32(qp->tx_pkts, &hdr->ver); 879 880 ntb_memcpy_tx(entry, offset); 881 } 882 883 static int 884 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 885 { 886 887 CTR3(KTR_NTB, 888 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u", 889 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry); 890 if (qp->tx_index == qp->remote_rx_info->entry) { 891 CTR0(KTR_NTB, "TX: ring full"); 892 qp->tx_ring_full++; 893 return (EAGAIN); 894 } 895 896 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) { 897 if (qp->tx_handler != NULL) 898 qp->tx_handler(qp, qp->cb_data, entry->buf, 899 EIO); 900 else 901 m_freem(entry->buf); 902 903 entry->buf = NULL; 904 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 905 CTR1(KTR_NTB, 906 "TX: frame too big. returning entry %p to tx_free_q", 907 entry); 908 return (0); 909 } 910 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index); 911 ntb_async_tx(qp, entry); 912 913 qp->tx_index++; 914 qp->tx_index %= qp->tx_max_entry; 915 916 qp->tx_pkts++; 917 918 return (0); 919 } 920 921 /* Transport Rx */ 922 static void 923 ntb_transport_rxc_db(void *arg, int pending __unused) 924 { 925 struct ntb_transport_qp *qp = arg; 926 uint64_t qp_mask = 1ull << qp->qp_num; 927 int rc; 928 929 CTR0(KTR_NTB, "RX: transport_rx"); 930 again: 931 while ((rc = ntb_process_rxc(qp)) == 0) 932 ; 933 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc); 934 935 if ((ntb_db_read(qp->dev) & qp_mask) != 0) { 936 /* If db is set, clear it and check queue once more. */ 937 ntb_db_clear(qp->dev, qp_mask); 938 goto again; 939 } 940 if (qp->link_is_up) 941 ntb_db_clear_mask(qp->dev, qp_mask); 942 } 943 944 static int 945 ntb_process_rxc(struct ntb_transport_qp *qp) 946 { 947 struct ntb_payload_header *hdr; 948 struct ntb_queue_entry *entry; 949 caddr_t offset; 950 951 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index; 952 hdr = (void *)(offset + qp->rx_max_frame - 953 sizeof(struct ntb_payload_header)); 954 955 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index); 956 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) { 957 CTR0(KTR_NTB, "RX: hdr not done"); 958 qp->rx_ring_empty++; 959 return (EAGAIN); 960 } 961 962 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) { 963 CTR0(KTR_NTB, "RX: link down"); 964 ntb_qp_link_down(qp); 965 hdr->flags = 0; 966 return (EAGAIN); 967 } 968 969 if (hdr->ver != (uint32_t)qp->rx_pkts) { 970 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). " 971 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts); 972 qp->rx_err_ver++; 973 return (EIO); 974 } 975 976 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q); 977 if (entry == NULL) { 978 qp->rx_err_no_buf++; 979 CTR0(KTR_NTB, "RX: No entries in rx_pend_q"); 980 return (EAGAIN); 981 } 982 callout_stop(&qp->rx_full); 983 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry); 984 985 entry->x_hdr = hdr; 986 entry->index = qp->rx_index; 987 988 if (hdr->len > entry->len) { 989 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju", 990 (uintmax_t)hdr->len, (uintmax_t)entry->len); 991 qp->rx_err_oflow++; 992 993 entry->len = -EIO; 994 entry->flags |= NTBT_DESC_DONE_FLAG; 995 996 ntb_complete_rxc(qp); 997 } else { 998 qp->rx_bytes += hdr->len; 999 qp->rx_pkts++; 1000 1001 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts); 1002 1003 entry->len = hdr->len; 1004 1005 ntb_memcpy_rx(qp, entry, offset); 1006 } 1007 1008 qp->rx_index++; 1009 qp->rx_index %= qp->rx_max_entry; 1010 return (0); 1011 } 1012 1013 static void 1014 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry, 1015 void *offset) 1016 { 1017 struct ifnet *ifp = entry->cb_data; 1018 unsigned int len = entry->len; 1019 1020 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset); 1021 1022 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL); 1023 if (entry->buf == NULL) 1024 entry->len = -ENOMEM; 1025 1026 /* Ensure that the data is globally visible before clearing the flag */ 1027 wmb(); 1028 1029 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf); 1030 ntb_rx_copy_callback(qp, entry); 1031 } 1032 1033 static inline void 1034 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data) 1035 { 1036 struct ntb_queue_entry *entry; 1037 1038 entry = data; 1039 entry->flags |= NTBT_DESC_DONE_FLAG; 1040 ntb_complete_rxc(qp); 1041 } 1042 1043 static void 1044 ntb_complete_rxc(struct ntb_transport_qp *qp) 1045 { 1046 struct ntb_queue_entry *entry; 1047 struct mbuf *m; 1048 unsigned len; 1049 1050 CTR0(KTR_NTB, "RX: rx_completion_task"); 1051 1052 mtx_lock_spin(&qp->ntb_rx_q_lock); 1053 1054 while (!STAILQ_EMPTY(&qp->rx_post_q)) { 1055 entry = STAILQ_FIRST(&qp->rx_post_q); 1056 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0) 1057 break; 1058 1059 entry->x_hdr->flags = 0; 1060 iowrite32(entry->index, &qp->rx_info->entry); 1061 1062 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry); 1063 1064 len = entry->len; 1065 m = entry->buf; 1066 1067 /* 1068 * Re-initialize queue_entry for reuse; rx_handler takes 1069 * ownership of the mbuf. 1070 */ 1071 entry->buf = NULL; 1072 entry->len = transport_mtu; 1073 entry->cb_data = qp->cb_data; 1074 1075 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry); 1076 1077 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1078 1079 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m); 1080 if (qp->rx_handler != NULL && qp->client_ready) 1081 qp->rx_handler(qp, qp->cb_data, m, len); 1082 else 1083 m_freem(m); 1084 1085 mtx_lock_spin(&qp->ntb_rx_q_lock); 1086 } 1087 1088 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1089 } 1090 1091 static void 1092 ntb_transport_doorbell_callback(void *data, uint32_t vector) 1093 { 1094 struct ntb_transport_ctx *nt = data; 1095 struct ntb_transport_qp *qp; 1096 uint64_t vec_mask; 1097 unsigned qp_num; 1098 1099 vec_mask = ntb_db_vector_mask(nt->dev, vector); 1100 vec_mask &= nt->qp_bitmap; 1101 if ((vec_mask & (vec_mask - 1)) != 0) 1102 vec_mask &= ntb_db_read(nt->dev); 1103 if (vec_mask != 0) { 1104 ntb_db_set_mask(nt->dev, vec_mask); 1105 ntb_db_clear(nt->dev, vec_mask); 1106 } 1107 while (vec_mask != 0) { 1108 qp_num = ffsll(vec_mask) - 1; 1109 1110 qp = &nt->qp_vec[qp_num]; 1111 if (qp->link_is_up) 1112 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work); 1113 1114 vec_mask &= ~(1ull << qp_num); 1115 } 1116 } 1117 1118 /* Link Event handler */ 1119 static void 1120 ntb_transport_event_callback(void *data) 1121 { 1122 struct ntb_transport_ctx *nt = data; 1123 1124 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) { 1125 ntb_printf(1, "HW link up\n"); 1126 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt); 1127 } else { 1128 ntb_printf(1, "HW link down\n"); 1129 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup); 1130 } 1131 } 1132 1133 /* Link bring up */ 1134 static void 1135 ntb_transport_link_work(void *arg) 1136 { 1137 struct ntb_transport_ctx *nt = arg; 1138 struct ntb_transport_mw *mw; 1139 device_t dev = nt->dev; 1140 struct ntb_transport_qp *qp; 1141 uint64_t val64, size; 1142 uint32_t val; 1143 unsigned i; 1144 int rc; 1145 1146 /* send the local info, in the opposite order of the way we read it */ 1147 if (nt->compact) { 1148 for (i = 0; i < nt->mw_count; i++) { 1149 size = nt->mw_vec[i].tx_size; 1150 KASSERT(size <= UINT32_MAX, ("size too big (%jx)", size)); 1151 ntb_peer_spad_write(dev, NTBTC_MW0_SZ + i, size); 1152 } 1153 ntb_peer_spad_write(dev, NTBTC_QP_LINKS, 0); 1154 ntb_peer_spad_write(dev, NTBTC_PARAMS, 1155 (nt->qp_count << 24) | (nt->mw_count << 16) | 1156 NTB_TRANSPORT_VERSION); 1157 } else { 1158 for (i = 0; i < nt->mw_count; i++) { 1159 size = nt->mw_vec[i].tx_size; 1160 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2), 1161 size >> 32); 1162 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size); 1163 } 1164 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count); 1165 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count); 1166 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0); 1167 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION); 1168 } 1169 1170 /* Query the remote side for its info */ 1171 val = 0; 1172 if (nt->compact) { 1173 ntb_spad_read(dev, NTBTC_PARAMS, &val); 1174 if (val != ((nt->qp_count << 24) | (nt->mw_count << 16) | 1175 NTB_TRANSPORT_VERSION)) 1176 goto out; 1177 } else { 1178 ntb_spad_read(dev, NTBT_VERSION, &val); 1179 if (val != NTB_TRANSPORT_VERSION) 1180 goto out; 1181 1182 ntb_spad_read(dev, NTBT_NUM_QPS, &val); 1183 if (val != nt->qp_count) 1184 goto out; 1185 1186 ntb_spad_read(dev, NTBT_NUM_MWS, &val); 1187 if (val != nt->mw_count) 1188 goto out; 1189 } 1190 1191 for (i = 0; i < nt->mw_count; i++) { 1192 if (nt->compact) { 1193 ntb_spad_read(dev, NTBTC_MW0_SZ + i, &val); 1194 val64 = val; 1195 } else { 1196 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val); 1197 val64 = (uint64_t)val << 32; 1198 1199 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val); 1200 val64 |= val; 1201 } 1202 1203 mw = &nt->mw_vec[i]; 1204 mw->rx_size = val64; 1205 val64 = roundup(val64, mw->xlat_align_size); 1206 if (mw->buff_size != val64) { 1207 1208 rc = ntb_set_mw(nt, i, val64); 1209 if (rc != 0) { 1210 ntb_printf(0, "link up set mw%d fails, rc %d\n", 1211 i, rc); 1212 goto free_mws; 1213 } 1214 1215 /* Notify HW the memory location of the receive buffer */ 1216 rc = ntb_mw_set_trans(nt->dev, i, mw->dma_addr, 1217 mw->buff_size); 1218 if (rc != 0) { 1219 ntb_printf(0, "link up mw%d xlat fails, rc %d\n", 1220 i, rc); 1221 goto free_mws; 1222 } 1223 } 1224 } 1225 1226 nt->link_is_up = true; 1227 ntb_printf(1, "transport link up\n"); 1228 1229 for (i = 0; i < nt->qp_count; i++) { 1230 qp = &nt->qp_vec[i]; 1231 1232 ntb_transport_setup_qp_mw(nt, i); 1233 1234 if (qp->client_ready) 1235 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 1236 } 1237 1238 return; 1239 1240 free_mws: 1241 for (i = 0; i < nt->mw_count; i++) 1242 ntb_free_mw(nt, i); 1243 out: 1244 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width)) 1245 callout_reset(&nt->link_work, 1246 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt); 1247 } 1248 1249 struct ntb_load_cb_args { 1250 bus_addr_t addr; 1251 int error; 1252 }; 1253 1254 static void 1255 ntb_load_cb(void *xsc, bus_dma_segment_t *segs, int nsegs, int error) 1256 { 1257 struct ntb_load_cb_args *cba = (struct ntb_load_cb_args *)xsc; 1258 1259 if (!(cba->error = error)) 1260 cba->addr = segs[0].ds_addr; 1261 } 1262 1263 static int 1264 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size) 1265 { 1266 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1267 struct ntb_load_cb_args cba; 1268 size_t buff_size; 1269 1270 if (size == 0) 1271 return (EINVAL); 1272 1273 buff_size = roundup(size, mw->xlat_align_size); 1274 1275 /* No need to re-setup */ 1276 if (mw->buff_size == buff_size) 1277 return (0); 1278 1279 if (mw->buff_size != 0) 1280 ntb_free_mw(nt, num_mw); 1281 1282 /* Alloc memory for receiving data. Must be aligned */ 1283 mw->buff_size = buff_size; 1284 1285 if (bus_dma_tag_create(bus_get_dma_tag(nt->dev), mw->xlat_align, 0, 1286 mw->addr_limit, BUS_SPACE_MAXADDR, 1287 NULL, NULL, mw->buff_size, 1, mw->buff_size, 1288 0, NULL, NULL, &mw->dma_tag)) { 1289 ntb_printf(0, "Unable to create MW tag of size %zu\n", 1290 mw->buff_size); 1291 mw->buff_size = 0; 1292 return (ENOMEM); 1293 } 1294 if (bus_dmamem_alloc(mw->dma_tag, (void **)&mw->virt_addr, 1295 BUS_DMA_WAITOK | BUS_DMA_ZERO, &mw->dma_map)) { 1296 bus_dma_tag_destroy(mw->dma_tag); 1297 ntb_printf(0, "Unable to allocate MW buffer of size %zu\n", 1298 mw->buff_size); 1299 mw->buff_size = 0; 1300 return (ENOMEM); 1301 } 1302 if (bus_dmamap_load(mw->dma_tag, mw->dma_map, mw->virt_addr, 1303 mw->buff_size, ntb_load_cb, &cba, BUS_DMA_NOWAIT) || cba.error) { 1304 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1305 bus_dma_tag_destroy(mw->dma_tag); 1306 ntb_printf(0, "Unable to load MW buffer of size %zu\n", 1307 mw->buff_size); 1308 mw->buff_size = 0; 1309 return (ENOMEM); 1310 } 1311 mw->dma_addr = cba.addr; 1312 1313 return (0); 1314 } 1315 1316 static void 1317 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw) 1318 { 1319 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1320 1321 if (mw->virt_addr == NULL) 1322 return; 1323 1324 ntb_mw_clear_trans(nt->dev, num_mw); 1325 bus_dmamap_unload(mw->dma_tag, mw->dma_map); 1326 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1327 bus_dma_tag_destroy(mw->dma_tag); 1328 mw->buff_size = 0; 1329 mw->virt_addr = NULL; 1330 } 1331 1332 static int 1333 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num) 1334 { 1335 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num]; 1336 struct ntb_transport_mw *mw; 1337 void *offset; 1338 ntb_q_idx_t i; 1339 size_t rx_size; 1340 unsigned num_qps_mw, mw_num, mw_count; 1341 1342 mw_count = nt->mw_count; 1343 mw_num = QP_TO_MW(nt, qp_num); 1344 mw = &nt->mw_vec[mw_num]; 1345 1346 if (mw->virt_addr == NULL) 1347 return (ENOMEM); 1348 1349 if (mw_num < nt->qp_count % mw_count) 1350 num_qps_mw = nt->qp_count / mw_count + 1; 1351 else 1352 num_qps_mw = nt->qp_count / mw_count; 1353 1354 rx_size = mw->rx_size / num_qps_mw; 1355 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count); 1356 rx_size -= sizeof(struct ntb_rx_info); 1357 1358 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size); 1359 1360 /* Due to house-keeping, there must be at least 2 buffs */ 1361 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2); 1362 qp->rx_max_entry = rx_size / qp->rx_max_frame; 1363 qp->rx_index = 0; 1364 1365 qp->remote_rx_info->entry = qp->rx_max_entry - 1; 1366 1367 /* Set up the hdr offsets with 0s */ 1368 for (i = 0; i < qp->rx_max_entry; i++) { 1369 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) - 1370 sizeof(struct ntb_payload_header)); 1371 memset(offset, 0, sizeof(struct ntb_payload_header)); 1372 } 1373 1374 qp->rx_pkts = 0; 1375 qp->tx_pkts = 0; 1376 qp->tx_index = 0; 1377 1378 return (0); 1379 } 1380 1381 static void 1382 ntb_qp_link_work(void *arg) 1383 { 1384 struct ntb_transport_qp *qp = arg; 1385 device_t dev = qp->dev; 1386 struct ntb_transport_ctx *nt = qp->transport; 1387 int i; 1388 uint32_t val; 1389 1390 /* Report queues that are up on our side */ 1391 for (i = 0, val = 0; i < nt->qp_count; i++) { 1392 if (nt->qp_vec[i].client_ready) 1393 val |= (1 << i); 1394 } 1395 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val); 1396 1397 /* See if the remote side is up */ 1398 ntb_spad_read(dev, NTBT_QP_LINKS, &val); 1399 if ((val & (1ull << qp->qp_num)) != 0) { 1400 ntb_printf(2, "qp %d link up\n", qp->qp_num); 1401 qp->link_is_up = true; 1402 1403 if (qp->event_handler != NULL) 1404 qp->event_handler(qp->cb_data, NTB_LINK_UP); 1405 1406 ntb_db_clear_mask(dev, 1ull << qp->qp_num); 1407 } else if (nt->link_is_up) 1408 callout_reset(&qp->link_work, 1409 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp); 1410 } 1411 1412 /* Link down event*/ 1413 static void 1414 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt) 1415 { 1416 struct ntb_transport_qp *qp; 1417 int i; 1418 1419 callout_drain(&nt->link_work); 1420 nt->link_is_up = 0; 1421 1422 /* Pass along the info to any clients */ 1423 for (i = 0; i < nt->qp_count; i++) { 1424 if ((nt->qp_bitmap & (1 << i)) != 0) { 1425 qp = &nt->qp_vec[i]; 1426 ntb_qp_link_cleanup(qp); 1427 callout_drain(&qp->link_work); 1428 } 1429 } 1430 1431 /* 1432 * The scratchpad registers keep the values if the remote side 1433 * goes down, blast them now to give them a sane value the next 1434 * time they are accessed 1435 */ 1436 ntb_spad_clear(nt->dev); 1437 } 1438 1439 static void 1440 ntb_transport_link_cleanup_work(void *arg, int pending __unused) 1441 { 1442 1443 ntb_transport_link_cleanup(arg); 1444 } 1445 1446 static void 1447 ntb_qp_link_down(struct ntb_transport_qp *qp) 1448 { 1449 1450 ntb_qp_link_cleanup(qp); 1451 } 1452 1453 static void 1454 ntb_qp_link_down_reset(struct ntb_transport_qp *qp) 1455 { 1456 1457 qp->link_is_up = false; 1458 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 1459 1460 qp->tx_index = qp->rx_index = 0; 1461 qp->tx_bytes = qp->rx_bytes = 0; 1462 qp->tx_pkts = qp->rx_pkts = 0; 1463 1464 qp->rx_ring_empty = 0; 1465 qp->tx_ring_full = 0; 1466 1467 qp->rx_err_no_buf = qp->tx_err_no_buf = 0; 1468 qp->rx_err_oflow = qp->rx_err_ver = 0; 1469 } 1470 1471 static void 1472 ntb_qp_link_cleanup(struct ntb_transport_qp *qp) 1473 { 1474 1475 callout_drain(&qp->link_work); 1476 ntb_qp_link_down_reset(qp); 1477 1478 if (qp->event_handler != NULL) 1479 qp->event_handler(qp->cb_data, NTB_LINK_DOWN); 1480 } 1481 1482 /* Link commanded down */ 1483 /** 1484 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data 1485 * @qp: NTB transport layer queue to be disabled 1486 * 1487 * Notify NTB transport layer of client's desire to no longer receive data on 1488 * transport queue specified. It is the client's responsibility to ensure all 1489 * entries on queue are purged or otherwise handled appropriately. 1490 */ 1491 void 1492 ntb_transport_link_down(struct ntb_transport_qp *qp) 1493 { 1494 struct ntb_transport_ctx *nt = qp->transport; 1495 int i; 1496 uint32_t val; 1497 1498 qp->client_ready = false; 1499 for (i = 0, val = 0; i < nt->qp_count; i++) { 1500 if (nt->qp_vec[i].client_ready) 1501 val |= (1 << i); 1502 } 1503 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val); 1504 1505 if (qp->link_is_up) 1506 ntb_send_link_down(qp); 1507 else 1508 callout_drain(&qp->link_work); 1509 } 1510 1511 /** 1512 * ntb_transport_link_query - Query transport link state 1513 * @qp: NTB transport layer queue to be queried 1514 * 1515 * Query connectivity to the remote system of the NTB transport queue 1516 * 1517 * RETURNS: true for link up or false for link down 1518 */ 1519 bool 1520 ntb_transport_link_query(struct ntb_transport_qp *qp) 1521 { 1522 1523 return (qp->link_is_up); 1524 } 1525 1526 /** 1527 * ntb_transport_link_speed - Query transport link speed 1528 * @qp: NTB transport layer queue to be queried 1529 * 1530 * Query connection speed to the remote system of the NTB transport queue 1531 * 1532 * RETURNS: link speed in bits per second 1533 */ 1534 uint64_t 1535 ntb_transport_link_speed(struct ntb_transport_qp *qp) 1536 { 1537 struct ntb_transport_ctx *nt = qp->transport; 1538 uint64_t rate; 1539 1540 if (!nt->link_is_up) 1541 return (0); 1542 switch (nt->link_speed) { 1543 case NTB_SPEED_GEN1: 1544 rate = 2500000000 * 8 / 10; 1545 break; 1546 case NTB_SPEED_GEN2: 1547 rate = 5000000000 * 8 / 10; 1548 break; 1549 case NTB_SPEED_GEN3: 1550 rate = 8000000000 * 128 / 130; 1551 break; 1552 case NTB_SPEED_GEN4: 1553 rate = 16000000000 * 128 / 130; 1554 break; 1555 default: 1556 return (0); 1557 } 1558 if (nt->link_width <= 0) 1559 return (0); 1560 return (rate * nt->link_width); 1561 } 1562 1563 static void 1564 ntb_send_link_down(struct ntb_transport_qp *qp) 1565 { 1566 struct ntb_queue_entry *entry; 1567 int i, rc; 1568 1569 if (!qp->link_is_up) 1570 return; 1571 1572 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) { 1573 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 1574 if (entry != NULL) 1575 break; 1576 pause("NTB Wait for link down", hz / 10); 1577 } 1578 1579 if (entry == NULL) 1580 return; 1581 1582 entry->cb_data = NULL; 1583 entry->buf = NULL; 1584 entry->len = 0; 1585 entry->flags = NTBT_LINK_DOWN_FLAG; 1586 1587 mtx_lock(&qp->tx_lock); 1588 rc = ntb_process_tx(qp, entry); 1589 mtx_unlock(&qp->tx_lock); 1590 if (rc != 0) 1591 printf("ntb: Failed to send link down\n"); 1592 1593 ntb_qp_link_down_reset(qp); 1594 } 1595 1596 1597 /* List Management */ 1598 1599 static void 1600 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 1601 struct ntb_queue_list *list) 1602 { 1603 1604 mtx_lock_spin(lock); 1605 STAILQ_INSERT_TAIL(list, entry, entry); 1606 mtx_unlock_spin(lock); 1607 } 1608 1609 static struct ntb_queue_entry * 1610 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list) 1611 { 1612 struct ntb_queue_entry *entry; 1613 1614 mtx_lock_spin(lock); 1615 if (STAILQ_EMPTY(list)) { 1616 entry = NULL; 1617 goto out; 1618 } 1619 entry = STAILQ_FIRST(list); 1620 STAILQ_REMOVE_HEAD(list, entry); 1621 out: 1622 mtx_unlock_spin(lock); 1623 1624 return (entry); 1625 } 1626 1627 static struct ntb_queue_entry * 1628 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from, 1629 struct ntb_queue_list *to) 1630 { 1631 struct ntb_queue_entry *entry; 1632 1633 mtx_lock_spin(lock); 1634 if (STAILQ_EMPTY(from)) { 1635 entry = NULL; 1636 goto out; 1637 } 1638 entry = STAILQ_FIRST(from); 1639 STAILQ_REMOVE_HEAD(from, entry); 1640 STAILQ_INSERT_TAIL(to, entry, entry); 1641 1642 out: 1643 mtx_unlock_spin(lock); 1644 return (entry); 1645 } 1646 1647 /** 1648 * ntb_transport_qp_num - Query the qp number 1649 * @qp: NTB transport layer queue to be queried 1650 * 1651 * Query qp number of the NTB transport queue 1652 * 1653 * RETURNS: a zero based number specifying the qp number 1654 */ 1655 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp) 1656 { 1657 1658 return (qp->qp_num); 1659 } 1660 1661 /** 1662 * ntb_transport_max_size - Query the max payload size of a qp 1663 * @qp: NTB transport layer queue to be queried 1664 * 1665 * Query the maximum payload size permissible on the given qp 1666 * 1667 * RETURNS: the max payload size of a qp 1668 */ 1669 unsigned int 1670 ntb_transport_max_size(struct ntb_transport_qp *qp) 1671 { 1672 1673 return (qp->tx_max_frame - sizeof(struct ntb_payload_header)); 1674 } 1675 1676 unsigned int 1677 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp) 1678 { 1679 unsigned int head = qp->tx_index; 1680 unsigned int tail = qp->remote_rx_info->entry; 1681 1682 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head); 1683 } 1684 1685 static device_method_t ntb_transport_methods[] = { 1686 /* Device interface */ 1687 DEVMETHOD(device_probe, ntb_transport_probe), 1688 DEVMETHOD(device_attach, ntb_transport_attach), 1689 DEVMETHOD(device_detach, ntb_transport_detach), 1690 /* Bus interface */ 1691 DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str), 1692 DEVMETHOD(bus_print_child, ntb_transport_print_child), 1693 DEVMETHOD_END 1694 }; 1695 1696 devclass_t ntb_transport_devclass; 1697 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver, 1698 ntb_transport_methods, sizeof(struct ntb_transport_ctx)); 1699 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver, 1700 ntb_transport_devclass, NULL, NULL); 1701 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1); 1702 MODULE_VERSION(ntb_transport, 1); 1703