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