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: stable/11/sys/dev/ntb/ntb_transport.c 329057 2018-02-09 03:07:12Z mav $"); 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 70static SYSCTL_NODE(_hw, OID_AUTO, ntb_transport, CTLFLAG_RW, 0, "ntb_transport"); 71 72static unsigned g_ntb_transport_debug_level; 73SYSCTL_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 82static unsigned transport_mtu = 0x10000; 83 84static uint64_t max_mw_size; 85SYSCTL_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 89static unsigned enable_xeon_watchdog; 90SYSCTL_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 94STAILQ_HEAD(ntb_queue_list, ntb_queue_entry); 95 96typedef uint32_t ntb_q_idx_t; 97 98struct 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 113struct ntb_rx_info { 114 ntb_q_idx_t entry; 115}; 116 117struct 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 174struct ntb_transport_mw { 175 vm_paddr_t phys_addr; 176 size_t phys_size; 177 size_t xlat_align; 178 size_t xlat_align_size; 179 bus_addr_t addr_limit; 180 /* Tx buff is off vbase / phys_addr */ 181 caddr_t vbase; 182 size_t xlat_size; 183 size_t buff_size; 184 /* Rx buff is off virt_addr / dma_addr */ 185 bus_dma_tag_t dma_tag; 186 bus_dmamap_t dma_map; 187 caddr_t virt_addr; 188 bus_addr_t dma_addr; 189}; 190 191struct ntb_transport_child { 192 device_t dev; 193 int consumer; 194 int qpoff; 195 int qpcnt; 196 struct ntb_transport_child *next; 197}; 198 199struct ntb_transport_ctx { 200 device_t dev; 201 struct ntb_transport_child *child; 202 struct ntb_transport_mw *mw_vec; 203 struct ntb_transport_qp *qp_vec; 204 unsigned mw_count; 205 unsigned qp_count; 206 uint64_t qp_bitmap; 207 volatile bool link_is_up; 208 enum ntb_speed link_speed; 209 enum ntb_width link_width; 210 struct callout link_work; 211 struct callout link_watchdog; 212 struct task link_cleanup; 213}; 214 215enum { 216 NTBT_DESC_DONE_FLAG = 1 << 0, 217 NTBT_LINK_DOWN_FLAG = 1 << 1, 218}; 219 220struct ntb_payload_header { 221 ntb_q_idx_t ver; 222 uint32_t len; 223 uint32_t flags; 224}; 225 226enum { 227 /* 228 * The order of this enum is part of the remote protocol. Do not 229 * reorder without bumping protocol version (and it's probably best 230 * to keep the protocol in lock-step with the Linux NTB driver. 231 */ 232 NTBT_VERSION = 0, 233 NTBT_QP_LINKS, 234 NTBT_NUM_QPS, 235 NTBT_NUM_MWS, 236 /* 237 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2. 238 */ 239 NTBT_MW0_SZ_HIGH, 240 NTBT_MW0_SZ_LOW, 241 NTBT_MW1_SZ_HIGH, 242 NTBT_MW1_SZ_LOW, 243 244 /* 245 * Some NTB-using hardware have a watchdog to work around NTB hangs; if 246 * a register or doorbell isn't written every few seconds, the link is 247 * torn down. Write an otherwise unused register every few seconds to 248 * work around this watchdog. 249 */ 250 NTBT_WATCHDOG_SPAD = 15 251}; 252 253#define QP_TO_MW(nt, qp) ((qp) % nt->mw_count) 254#define NTB_QP_DEF_NUM_ENTRIES 100 255#define NTB_LINK_DOWN_TIMEOUT 100 256 257static int ntb_transport_probe(device_t dev); 258static int ntb_transport_attach(device_t dev); 259static int ntb_transport_detach(device_t dev); 260static void ntb_transport_init_queue(struct ntb_transport_ctx *nt, 261 unsigned int qp_num); 262static int ntb_process_tx(struct ntb_transport_qp *qp, 263 struct ntb_queue_entry *entry); 264static void ntb_transport_rxc_db(void *arg, int pending); 265static int ntb_process_rxc(struct ntb_transport_qp *qp); 266static void ntb_memcpy_rx(struct ntb_transport_qp *qp, 267 struct ntb_queue_entry *entry, void *offset); 268static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp, 269 void *data); 270static void ntb_complete_rxc(struct ntb_transport_qp *qp); 271static void ntb_transport_doorbell_callback(void *data, uint32_t vector); 272static void ntb_transport_event_callback(void *data); 273static void ntb_transport_link_work(void *arg); 274static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size); 275static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw); 276static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, 277 unsigned int qp_num); 278static void ntb_qp_link_work(void *arg); 279static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt); 280static void ntb_transport_link_cleanup_work(void *, int); 281static void ntb_qp_link_down(struct ntb_transport_qp *qp); 282static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp); 283static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp); 284static void ntb_send_link_down(struct ntb_transport_qp *qp); 285static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 286 struct ntb_queue_list *list); 287static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock, 288 struct ntb_queue_list *list); 289static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock, 290 struct ntb_queue_list *from, struct ntb_queue_list *to); 291static void xeon_link_watchdog_hb(void *); 292 293static const struct ntb_ctx_ops ntb_transport_ops = { 294 .link_event = ntb_transport_event_callback, 295 .db_event = ntb_transport_doorbell_callback, 296}; 297 298MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver"); 299 300static inline void 301iowrite32(uint32_t val, void *addr) 302{ 303 304 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr, 305 val); 306} 307 308/* Transport Init and teardown */ 309 310static void 311xeon_link_watchdog_hb(void *arg) 312{ 313 struct ntb_transport_ctx *nt; 314 315 nt = arg; 316 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0); 317 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt); 318} 319 320static int 321ntb_transport_probe(device_t dev) 322{ 323 324 device_set_desc(dev, "NTB Transport"); 325 return (0); 326} 327 328static int 329ntb_transport_attach(device_t dev) 330{ 331 struct ntb_transport_ctx *nt = device_get_softc(dev); 332 struct ntb_transport_child **cpp = &nt->child; 333 struct ntb_transport_child *nc; 334 struct ntb_transport_mw *mw; 335 uint64_t db_bitmap, size; 336 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt; 337 char cfg[128] = ""; 338 char buf[32]; 339 char *n, *np, *c, *name; 340 341 nt->dev = dev; 342 nt->mw_count = ntb_mw_count(dev); 343 spad_count = ntb_spad_count(dev); 344 db_bitmap = ntb_db_valid_mask(dev); 345 db_count = flsll(db_bitmap); 346 KASSERT(db_bitmap == (1 << db_count) - 1, 347 ("Doorbells are not sequential (%jx).\n", db_bitmap)); 348 349 if (nt->mw_count == 0) { 350 device_printf(dev, "At least 1 memory window required.\n"); 351 return (ENXIO); 352 } 353 if (spad_count < 6) { 354 device_printf(dev, "At least 6 scratchpads required.\n"); 355 return (ENXIO); 356 } 357 if (spad_count < 4 + 2 * nt->mw_count) { 358 nt->mw_count = (spad_count - 4) / 2; 359 device_printf(dev, "Scratchpads enough only for %d " 360 "memory windows.\n", nt->mw_count); 361 } 362 if (db_bitmap == 0) { 363 device_printf(dev, "At least one doorbell required.\n"); 364 return (ENXIO); 365 } 366 367 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T, 368 M_WAITOK | M_ZERO); 369 for (i = 0; i < nt->mw_count; i++) { 370 mw = &nt->mw_vec[i]; 371 372 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase, 373 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size, 374 &mw->addr_limit); 375 if (rc != 0) 376 goto err; 377 378 mw->buff_size = 0; 379 mw->xlat_size = 0; 380 mw->virt_addr = NULL; 381 mw->dma_addr = 0; 382 383 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING); 384 if (rc) 385 ntb_printf(0, "Unable to set mw%d caching\n", i); 386 387 /* 388 * Try to preallocate receive memory early, since there may 389 * be not enough contiguous memory later. It is quite likely 390 * that NTB windows are symmetric and this allocation remain, 391 * but even if not, we will just reallocate it later. 392 */ 393 size = mw->phys_size; 394 if (max_mw_size != 0 && size > max_mw_size) 395 size = max_mw_size; 396 ntb_set_mw(nt, i, size); 397 } 398 399 qpu = 0; 400 qpo = imin(db_count, nt->mw_count); 401 qpt = db_count; 402 403 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev), 404 device_get_unit(dev)); 405 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg)); 406 n = cfg; 407 i = 0; 408 while ((c = strsep(&n, ",")) != NULL) { 409 np = c; 410 name = strsep(&np, ":"); 411 if (name != NULL && name[0] == 0) 412 name = NULL; 413 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu; 414 if (qp <= 0) 415 qp = 1; 416 417 if (qp > qpt - qpu) { 418 device_printf(dev, "Not enough resources for config\n"); 419 break; 420 } 421 422 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO); 423 nc->consumer = i; 424 nc->qpoff = qpu; 425 nc->qpcnt = qp; 426 nc->dev = device_add_child(dev, name, -1); 427 if (nc->dev == NULL) { 428 device_printf(dev, "Can not add child.\n"); 429 break; 430 } 431 device_set_ivars(nc->dev, nc); 432 *cpp = nc; 433 cpp = &nc->next; 434 435 if (bootverbose) { 436 device_printf(dev, "%d \"%s\": queues %d", 437 i, name, qpu); 438 if (qp > 1) 439 printf("-%d", qpu + qp - 1); 440 printf("\n"); 441 } 442 443 qpu += qp; 444 i++; 445 } 446 nt->qp_count = qpu; 447 448 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T, 449 M_WAITOK | M_ZERO); 450 451 for (i = 0; i < nt->qp_count; i++) 452 ntb_transport_init_queue(nt, i); 453 454 callout_init(&nt->link_work, 0); 455 callout_init(&nt->link_watchdog, 0); 456 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt); 457 nt->link_is_up = false; 458 459 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops); 460 if (rc != 0) 461 goto err; 462 463 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO); 464 465 if (enable_xeon_watchdog != 0) 466 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt); 467 468 bus_generic_attach(dev); 469 return (0); 470 471err: 472 free(nt->qp_vec, M_NTB_T); 473 free(nt->mw_vec, M_NTB_T); 474 return (rc); 475} 476 477static int 478ntb_transport_detach(device_t dev) 479{ 480 struct ntb_transport_ctx *nt = device_get_softc(dev); 481 struct ntb_transport_child **cpp = &nt->child; 482 struct ntb_transport_child *nc; 483 int error = 0, i; 484 485 while ((nc = *cpp) != NULL) { 486 *cpp = (*cpp)->next; 487 error = device_delete_child(dev, nc->dev); 488 if (error) 489 break; 490 free(nc, M_DEVBUF); 491 } 492 KASSERT(nt->qp_bitmap == 0, 493 ("Some queues not freed on detach (%jx)", nt->qp_bitmap)); 494 495 ntb_transport_link_cleanup(nt); 496 taskqueue_drain(taskqueue_swi, &nt->link_cleanup); 497 callout_drain(&nt->link_work); 498 callout_drain(&nt->link_watchdog); 499 500 ntb_link_disable(dev); 501 ntb_clear_ctx(dev); 502 503 for (i = 0; i < nt->mw_count; i++) 504 ntb_free_mw(nt, i); 505 506 free(nt->qp_vec, M_NTB_T); 507 free(nt->mw_vec, M_NTB_T); 508 return (0); 509} 510 511static int 512ntb_transport_print_child(device_t dev, device_t child) 513{ 514 struct ntb_transport_child *nc = device_get_ivars(child); 515 int retval; 516 517 retval = bus_print_child_header(dev, child); 518 if (nc->qpcnt > 0) { 519 printf(" queue %d", nc->qpoff); 520 if (nc->qpcnt > 1) 521 printf("-%d", nc->qpoff + nc->qpcnt - 1); 522 } 523 retval += printf(" at consumer %d", nc->consumer); 524 retval += bus_print_child_domain(dev, child); 525 retval += bus_print_child_footer(dev, child); 526 527 return (retval); 528} 529 530static int 531ntb_transport_child_location_str(device_t dev, device_t child, char *buf, 532 size_t buflen) 533{ 534 struct ntb_transport_child *nc = device_get_ivars(child); 535 536 snprintf(buf, buflen, "consumer=%d", nc->consumer); 537 return (0); 538} 539 540int 541ntb_transport_queue_count(device_t dev) 542{ 543 struct ntb_transport_child *nc = device_get_ivars(dev); 544 545 return (nc->qpcnt); 546} 547 548static void 549ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num) 550{ 551 struct ntb_transport_mw *mw; 552 struct ntb_transport_qp *qp; 553 vm_paddr_t mw_base; 554 uint64_t mw_size, qp_offset; 555 size_t tx_size; 556 unsigned num_qps_mw, mw_num, mw_count; 557 558 mw_count = nt->mw_count; 559 mw_num = QP_TO_MW(nt, qp_num); 560 mw = &nt->mw_vec[mw_num]; 561 562 qp = &nt->qp_vec[qp_num]; 563 qp->qp_num = qp_num; 564 qp->transport = nt; 565 qp->dev = nt->dev; 566 qp->client_ready = false; 567 qp->event_handler = NULL; 568 ntb_qp_link_down_reset(qp); 569 570 if (mw_num < nt->qp_count % mw_count) 571 num_qps_mw = nt->qp_count / mw_count + 1; 572 else 573 num_qps_mw = nt->qp_count / mw_count; 574 575 mw_base = mw->phys_addr; 576 mw_size = mw->phys_size; 577 578 tx_size = mw_size / num_qps_mw; 579 qp_offset = tx_size * (qp_num / mw_count); 580 581 qp->tx_mw = mw->vbase + qp_offset; 582 KASSERT(qp->tx_mw != NULL, ("uh oh?")); 583 584 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */ 585 qp->tx_mw_phys = mw_base + qp_offset; 586 KASSERT(qp->tx_mw_phys != 0, ("uh oh?")); 587 588 tx_size -= sizeof(struct ntb_rx_info); 589 qp->rx_info = (void *)(qp->tx_mw + tx_size); 590 591 /* Due to house-keeping, there must be at least 2 buffs */ 592 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2); 593 qp->tx_max_entry = tx_size / qp->tx_max_frame; 594 595 callout_init(&qp->link_work, 0); 596 callout_init(&qp->rx_full, 1); 597 598 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN); 599 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN); 600 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF); 601 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp); 602 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK, 603 taskqueue_thread_enqueue, &qp->rxc_tq); 604 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d", 605 device_get_nameunit(nt->dev), qp_num); 606 607 STAILQ_INIT(&qp->rx_post_q); 608 STAILQ_INIT(&qp->rx_pend_q); 609 STAILQ_INIT(&qp->tx_free_q); 610} 611 612void 613ntb_transport_free_queue(struct ntb_transport_qp *qp) 614{ 615 struct ntb_transport_ctx *nt = qp->transport; 616 struct ntb_queue_entry *entry; 617 618 callout_drain(&qp->link_work); 619 620 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 621 taskqueue_drain_all(qp->rxc_tq); 622 taskqueue_free(qp->rxc_tq); 623 624 qp->cb_data = NULL; 625 qp->rx_handler = NULL; 626 qp->tx_handler = NULL; 627 qp->event_handler = NULL; 628 629 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) 630 free(entry, M_NTB_T); 631 632 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) 633 free(entry, M_NTB_T); 634 635 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q))) 636 free(entry, M_NTB_T); 637 638 nt->qp_bitmap &= ~(1 << qp->qp_num); 639} 640 641/** 642 * ntb_transport_create_queue - Create a new NTB transport layer queue 643 * @rx_handler: receive callback function 644 * @tx_handler: transmit callback function 645 * @event_handler: event callback function 646 * 647 * Create a new NTB transport layer queue and provide the queue with a callback 648 * routine for both transmit and receive. The receive callback routine will be 649 * used to pass up data when the transport has received it on the queue. The 650 * transmit callback routine will be called when the transport has completed the 651 * transmission of the data on the queue and the data is ready to be freed. 652 * 653 * RETURNS: pointer to newly created ntb_queue, NULL on error. 654 */ 655struct ntb_transport_qp * 656ntb_transport_create_queue(device_t dev, int q, 657 const struct ntb_queue_handlers *handlers, void *data) 658{ 659 struct ntb_transport_child *nc = device_get_ivars(dev); 660 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev)); 661 struct ntb_queue_entry *entry; 662 struct ntb_transport_qp *qp; 663 int i; 664 665 if (q < 0 || q >= nc->qpcnt) 666 return (NULL); 667 668 qp = &nt->qp_vec[nc->qpoff + q]; 669 nt->qp_bitmap |= (1 << qp->qp_num); 670 qp->cb_data = data; 671 qp->rx_handler = handlers->rx_handler; 672 qp->tx_handler = handlers->tx_handler; 673 qp->event_handler = handlers->event_handler; 674 675 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 676 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 677 entry->cb_data = data; 678 entry->buf = NULL; 679 entry->len = transport_mtu; 680 entry->qp = qp; 681 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q); 682 } 683 684 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 685 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 686 entry->qp = qp; 687 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 688 } 689 690 ntb_db_clear(dev, 1ull << qp->qp_num); 691 return (qp); 692} 693 694/** 695 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue 696 * @qp: NTB transport layer queue to be enabled 697 * 698 * Notify NTB transport layer of client readiness to use queue 699 */ 700void 701ntb_transport_link_up(struct ntb_transport_qp *qp) 702{ 703 struct ntb_transport_ctx *nt = qp->transport; 704 705 qp->client_ready = true; 706 707 ntb_printf(2, "qp %d client ready\n", qp->qp_num); 708 709 if (nt->link_is_up) 710 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 711} 712 713 714 715/* Transport Tx */ 716 717/** 718 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry 719 * @qp: NTB transport layer queue the entry is to be enqueued on 720 * @cb: per buffer pointer for callback function to use 721 * @data: pointer to data buffer that will be sent 722 * @len: length of the data buffer 723 * 724 * Enqueue a new transmit buffer onto the transport queue from which a NTB 725 * payload will be transmitted. This assumes that a lock is being held to 726 * serialize access to the qp. 727 * 728 * RETURNS: An appropriate ERRNO error value on error, or zero for success. 729 */ 730int 731ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data, 732 unsigned int len) 733{ 734 struct ntb_queue_entry *entry; 735 int rc; 736 737 if (!qp->link_is_up || len == 0) { 738 CTR0(KTR_NTB, "TX: link not up"); 739 return (EINVAL); 740 } 741 742 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 743 if (entry == NULL) { 744 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q"); 745 qp->tx_err_no_buf++; 746 return (EBUSY); 747 } 748 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry); 749 750 entry->cb_data = cb; 751 entry->buf = data; 752 entry->len = len; 753 entry->flags = 0; 754 755 mtx_lock(&qp->tx_lock); 756 rc = ntb_process_tx(qp, entry); 757 mtx_unlock(&qp->tx_lock); 758 if (rc != 0) { 759 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 760 CTR1(KTR_NTB, 761 "TX: process_tx failed. Returning entry %p to tx_free_q", 762 entry); 763 } 764 return (rc); 765} 766 767static void 768ntb_tx_copy_callback(void *data) 769{ 770 struct ntb_queue_entry *entry = data; 771 struct ntb_transport_qp *qp = entry->qp; 772 struct ntb_payload_header *hdr = entry->x_hdr; 773 774 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags); 775 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr); 776 777 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num); 778 779 /* 780 * The entry length can only be zero if the packet is intended to be a 781 * "link down" or similar. Since no payload is being sent in these 782 * cases, there is nothing to add to the completion queue. 783 */ 784 if (entry->len > 0) { 785 qp->tx_bytes += entry->len; 786 787 if (qp->tx_handler) 788 qp->tx_handler(qp, qp->cb_data, entry->buf, 789 entry->len); 790 else 791 m_freem(entry->buf); 792 entry->buf = NULL; 793 } 794 795 CTR3(KTR_NTB, 796 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning " 797 "to tx_free_q", entry, hdr->ver, hdr->flags); 798 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 799} 800 801static void 802ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset) 803{ 804 805 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset); 806 if (entry->buf != NULL) { 807 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset); 808 809 /* 810 * Ensure that the data is fully copied before setting the 811 * flags 812 */ 813 wmb(); 814 } 815 816 ntb_tx_copy_callback(entry); 817} 818 819static void 820ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 821{ 822 struct ntb_payload_header *hdr; 823 void *offset; 824 825 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index; 826 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame - 827 sizeof(struct ntb_payload_header)); 828 entry->x_hdr = hdr; 829 830 iowrite32(entry->len, &hdr->len); 831 iowrite32(qp->tx_pkts, &hdr->ver); 832 833 ntb_memcpy_tx(entry, offset); 834} 835 836static int 837ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 838{ 839 840 CTR3(KTR_NTB, 841 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u", 842 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry); 843 if (qp->tx_index == qp->remote_rx_info->entry) { 844 CTR0(KTR_NTB, "TX: ring full"); 845 qp->tx_ring_full++; 846 return (EAGAIN); 847 } 848 849 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) { 850 if (qp->tx_handler != NULL) 851 qp->tx_handler(qp, qp->cb_data, entry->buf, 852 EIO); 853 else 854 m_freem(entry->buf); 855 856 entry->buf = NULL; 857 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 858 CTR1(KTR_NTB, 859 "TX: frame too big. returning entry %p to tx_free_q", 860 entry); 861 return (0); 862 } 863 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index); 864 ntb_async_tx(qp, entry); 865 866 qp->tx_index++; 867 qp->tx_index %= qp->tx_max_entry; 868 869 qp->tx_pkts++; 870 871 return (0); 872} 873 874/* Transport Rx */ 875static void 876ntb_transport_rxc_db(void *arg, int pending __unused) 877{ 878 struct ntb_transport_qp *qp = arg; 879 uint64_t qp_mask = 1ull << qp->qp_num; 880 int rc; 881 882 CTR0(KTR_NTB, "RX: transport_rx"); 883again: 884 while ((rc = ntb_process_rxc(qp)) == 0) 885 ; 886 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc); 887 888 if ((ntb_db_read(qp->dev) & qp_mask) != 0) { 889 /* If db is set, clear it and check queue once more. */ 890 ntb_db_clear(qp->dev, qp_mask); 891 goto again; 892 } 893 if (qp->link_is_up) 894 ntb_db_clear_mask(qp->dev, qp_mask); 895} 896 897static int 898ntb_process_rxc(struct ntb_transport_qp *qp) 899{ 900 struct ntb_payload_header *hdr; 901 struct ntb_queue_entry *entry; 902 caddr_t offset; 903 904 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index; 905 hdr = (void *)(offset + qp->rx_max_frame - 906 sizeof(struct ntb_payload_header)); 907 908 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index); 909 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) { 910 CTR0(KTR_NTB, "RX: hdr not done"); 911 qp->rx_ring_empty++; 912 return (EAGAIN); 913 } 914 915 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) { 916 CTR0(KTR_NTB, "RX: link down"); 917 ntb_qp_link_down(qp); 918 hdr->flags = 0; 919 return (EAGAIN); 920 } 921 922 if (hdr->ver != (uint32_t)qp->rx_pkts) { 923 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). " 924 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts); 925 qp->rx_err_ver++; 926 return (EIO); 927 } 928 929 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q); 930 if (entry == NULL) { 931 qp->rx_err_no_buf++; 932 CTR0(KTR_NTB, "RX: No entries in rx_pend_q"); 933 return (EAGAIN); 934 } 935 callout_stop(&qp->rx_full); 936 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry); 937 938 entry->x_hdr = hdr; 939 entry->index = qp->rx_index; 940 941 if (hdr->len > entry->len) { 942 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju", 943 (uintmax_t)hdr->len, (uintmax_t)entry->len); 944 qp->rx_err_oflow++; 945 946 entry->len = -EIO; 947 entry->flags |= NTBT_DESC_DONE_FLAG; 948 949 ntb_complete_rxc(qp); 950 } else { 951 qp->rx_bytes += hdr->len; 952 qp->rx_pkts++; 953 954 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts); 955 956 entry->len = hdr->len; 957 958 ntb_memcpy_rx(qp, entry, offset); 959 } 960 961 qp->rx_index++; 962 qp->rx_index %= qp->rx_max_entry; 963 return (0); 964} 965 966static void 967ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry, 968 void *offset) 969{ 970 struct ifnet *ifp = entry->cb_data; 971 unsigned int len = entry->len; 972 973 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset); 974 975 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL); 976 if (entry->buf == NULL) 977 entry->len = -ENOMEM; 978 979 /* Ensure that the data is globally visible before clearing the flag */ 980 wmb(); 981 982 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf); 983 ntb_rx_copy_callback(qp, entry); 984} 985 986static inline void 987ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data) 988{ 989 struct ntb_queue_entry *entry; 990 991 entry = data; 992 entry->flags |= NTBT_DESC_DONE_FLAG; 993 ntb_complete_rxc(qp); 994} 995 996static void 997ntb_complete_rxc(struct ntb_transport_qp *qp) 998{ 999 struct ntb_queue_entry *entry; 1000 struct mbuf *m; 1001 unsigned len; 1002 1003 CTR0(KTR_NTB, "RX: rx_completion_task"); 1004 1005 mtx_lock_spin(&qp->ntb_rx_q_lock); 1006 1007 while (!STAILQ_EMPTY(&qp->rx_post_q)) { 1008 entry = STAILQ_FIRST(&qp->rx_post_q); 1009 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0) 1010 break; 1011 1012 entry->x_hdr->flags = 0; 1013 iowrite32(entry->index, &qp->rx_info->entry); 1014 1015 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry); 1016 1017 len = entry->len; 1018 m = entry->buf; 1019 1020 /* 1021 * Re-initialize queue_entry for reuse; rx_handler takes 1022 * ownership of the mbuf. 1023 */ 1024 entry->buf = NULL; 1025 entry->len = transport_mtu; 1026 entry->cb_data = qp->cb_data; 1027 1028 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry); 1029 1030 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1031 1032 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m); 1033 if (qp->rx_handler != NULL && qp->client_ready) 1034 qp->rx_handler(qp, qp->cb_data, m, len); 1035 else 1036 m_freem(m); 1037 1038 mtx_lock_spin(&qp->ntb_rx_q_lock); 1039 } 1040 1041 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1042} 1043 1044static void 1045ntb_transport_doorbell_callback(void *data, uint32_t vector) 1046{ 1047 struct ntb_transport_ctx *nt = data; 1048 struct ntb_transport_qp *qp; 1049 uint64_t vec_mask; 1050 unsigned qp_num; 1051 1052 vec_mask = ntb_db_vector_mask(nt->dev, vector); 1053 vec_mask &= nt->qp_bitmap; 1054 if ((vec_mask & (vec_mask - 1)) != 0) 1055 vec_mask &= ntb_db_read(nt->dev); 1056 if (vec_mask != 0) { 1057 ntb_db_set_mask(nt->dev, vec_mask); 1058 ntb_db_clear(nt->dev, vec_mask); 1059 } 1060 while (vec_mask != 0) { 1061 qp_num = ffsll(vec_mask) - 1; 1062 1063 qp = &nt->qp_vec[qp_num]; 1064 if (qp->link_is_up) 1065 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work); 1066 1067 vec_mask &= ~(1ull << qp_num); 1068 } 1069} 1070 1071/* Link Event handler */ 1072static void 1073ntb_transport_event_callback(void *data) 1074{ 1075 struct ntb_transport_ctx *nt = data; 1076 1077 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) { 1078 ntb_printf(1, "HW link up\n"); 1079 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt); 1080 } else { 1081 ntb_printf(1, "HW link down\n"); 1082 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup); 1083 } 1084} 1085 1086/* Link bring up */ 1087static void 1088ntb_transport_link_work(void *arg) 1089{ 1090 struct ntb_transport_ctx *nt = arg; 1091 device_t dev = nt->dev; 1092 struct ntb_transport_qp *qp; 1093 uint64_t val64, size; 1094 uint32_t val; 1095 unsigned i; 1096 int rc; 1097 1098 /* send the local info, in the opposite order of the way we read it */ 1099 for (i = 0; i < nt->mw_count; i++) { 1100 size = nt->mw_vec[i].phys_size; 1101 1102 if (max_mw_size != 0 && size > max_mw_size) 1103 size = max_mw_size; 1104 1105 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2), 1106 size >> 32); 1107 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size); 1108 } 1109 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count); 1110 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count); 1111 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0); 1112 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION); 1113 1114 /* Query the remote side for its info */ 1115 val = 0; 1116 ntb_spad_read(dev, NTBT_VERSION, &val); 1117 if (val != NTB_TRANSPORT_VERSION) 1118 goto out; 1119 1120 ntb_spad_read(dev, NTBT_NUM_QPS, &val); 1121 if (val != nt->qp_count) 1122 goto out; 1123 1124 ntb_spad_read(dev, NTBT_NUM_MWS, &val); 1125 if (val != nt->mw_count) 1126 goto out; 1127 1128 for (i = 0; i < nt->mw_count; i++) { 1129 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val); 1130 val64 = (uint64_t)val << 32; 1131 1132 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val); 1133 val64 |= val; 1134 1135 rc = ntb_set_mw(nt, i, val64); 1136 if (rc != 0) 1137 goto free_mws; 1138 } 1139 1140 nt->link_is_up = true; 1141 ntb_printf(1, "transport link up\n"); 1142 1143 for (i = 0; i < nt->qp_count; i++) { 1144 qp = &nt->qp_vec[i]; 1145 1146 ntb_transport_setup_qp_mw(nt, i); 1147 1148 if (qp->client_ready) 1149 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 1150 } 1151 1152 return; 1153 1154free_mws: 1155 for (i = 0; i < nt->mw_count; i++) 1156 ntb_free_mw(nt, i); 1157out: 1158 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width)) 1159 callout_reset(&nt->link_work, 1160 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt); 1161} 1162 1163struct ntb_load_cb_args { 1164 bus_addr_t addr; 1165 int error; 1166}; 1167 1168static void 1169ntb_load_cb(void *xsc, bus_dma_segment_t *segs, int nsegs, int error) 1170{ 1171 struct ntb_load_cb_args *cba = (struct ntb_load_cb_args *)xsc; 1172 1173 if (!(cba->error = error)) 1174 cba->addr = segs[0].ds_addr; 1175} 1176 1177static int 1178ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size) 1179{ 1180 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1181 struct ntb_load_cb_args cba; 1182 size_t xlat_size, buff_size; 1183 int rc; 1184 1185 if (size == 0) 1186 return (EINVAL); 1187 1188 xlat_size = roundup(size, mw->xlat_align_size); 1189 buff_size = xlat_size; 1190 1191 /* No need to re-setup */ 1192 if (mw->xlat_size == xlat_size) 1193 return (0); 1194 1195 if (mw->buff_size != 0) 1196 ntb_free_mw(nt, num_mw); 1197 1198 /* Alloc memory for receiving data. Must be aligned */ 1199 mw->xlat_size = xlat_size; 1200 mw->buff_size = buff_size; 1201 1202 if (bus_dma_tag_create(bus_get_dma_tag(nt->dev), mw->xlat_align, 0, 1203 mw->addr_limit, BUS_SPACE_MAXADDR, 1204 NULL, NULL, mw->buff_size, 1, mw->buff_size, 1205 0, NULL, NULL, &mw->dma_tag)) { 1206 ntb_printf(0, "Unable to create MW tag of size %zu/%zu\n", 1207 mw->buff_size, mw->xlat_size); 1208 mw->xlat_size = 0; 1209 mw->buff_size = 0; 1210 return (ENOMEM); 1211 } 1212 if (bus_dmamem_alloc(mw->dma_tag, (void **)&mw->virt_addr, 1213 BUS_DMA_WAITOK | BUS_DMA_ZERO, &mw->dma_map)) { 1214 bus_dma_tag_destroy(mw->dma_tag); 1215 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n", 1216 mw->buff_size, mw->xlat_size); 1217 mw->xlat_size = 0; 1218 mw->buff_size = 0; 1219 return (ENOMEM); 1220 } 1221 if (bus_dmamap_load(mw->dma_tag, mw->dma_map, mw->virt_addr, 1222 mw->buff_size, ntb_load_cb, &cba, BUS_DMA_NOWAIT) || cba.error) { 1223 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1224 bus_dma_tag_destroy(mw->dma_tag); 1225 ntb_printf(0, "Unable to load MW buffer of size %zu/%zu\n", 1226 mw->buff_size, mw->xlat_size); 1227 mw->xlat_size = 0; 1228 mw->buff_size = 0; 1229 return (ENOMEM); 1230 } 1231 mw->dma_addr = cba.addr; 1232 1233 /* Notify HW the memory location of the receive buffer */ 1234 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size); 1235 if (rc) { 1236 ntb_printf(0, "Unable to set mw%d translation\n", num_mw); 1237 ntb_free_mw(nt, num_mw); 1238 return (rc); 1239 } 1240 1241 return (0); 1242} 1243 1244static void 1245ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw) 1246{ 1247 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1248 1249 if (mw->virt_addr == NULL) 1250 return; 1251 1252 ntb_mw_clear_trans(nt->dev, num_mw); 1253 bus_dmamap_unload(mw->dma_tag, mw->dma_map); 1254 bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map); 1255 bus_dma_tag_destroy(mw->dma_tag); 1256 mw->xlat_size = 0; 1257 mw->buff_size = 0; 1258 mw->virt_addr = NULL; 1259} 1260 1261static int 1262ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num) 1263{ 1264 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num]; 1265 struct ntb_transport_mw *mw; 1266 void *offset; 1267 ntb_q_idx_t i; 1268 size_t rx_size; 1269 unsigned num_qps_mw, mw_num, mw_count; 1270 1271 mw_count = nt->mw_count; 1272 mw_num = QP_TO_MW(nt, qp_num); 1273 mw = &nt->mw_vec[mw_num]; 1274 1275 if (mw->virt_addr == NULL) 1276 return (ENOMEM); 1277 1278 if (mw_num < nt->qp_count % mw_count) 1279 num_qps_mw = nt->qp_count / mw_count + 1; 1280 else 1281 num_qps_mw = nt->qp_count / mw_count; 1282 1283 rx_size = mw->xlat_size / num_qps_mw; 1284 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count); 1285 rx_size -= sizeof(struct ntb_rx_info); 1286 1287 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size); 1288 1289 /* Due to house-keeping, there must be at least 2 buffs */ 1290 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2); 1291 qp->rx_max_entry = rx_size / qp->rx_max_frame; 1292 qp->rx_index = 0; 1293 1294 qp->remote_rx_info->entry = qp->rx_max_entry - 1; 1295 1296 /* Set up the hdr offsets with 0s */ 1297 for (i = 0; i < qp->rx_max_entry; i++) { 1298 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) - 1299 sizeof(struct ntb_payload_header)); 1300 memset(offset, 0, sizeof(struct ntb_payload_header)); 1301 } 1302 1303 qp->rx_pkts = 0; 1304 qp->tx_pkts = 0; 1305 qp->tx_index = 0; 1306 1307 return (0); 1308} 1309 1310static void 1311ntb_qp_link_work(void *arg) 1312{ 1313 struct ntb_transport_qp *qp = arg; 1314 device_t dev = qp->dev; 1315 struct ntb_transport_ctx *nt = qp->transport; 1316 int i; 1317 uint32_t val; 1318 1319 /* Report queues that are up on our side */ 1320 for (i = 0, val = 0; i < nt->qp_count; i++) { 1321 if (nt->qp_vec[i].client_ready) 1322 val |= (1 << i); 1323 } 1324 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val); 1325 1326 /* See if the remote side is up */ 1327 ntb_spad_read(dev, NTBT_QP_LINKS, &val); 1328 if ((val & (1ull << qp->qp_num)) != 0) { 1329 ntb_printf(2, "qp %d link up\n", qp->qp_num); 1330 qp->link_is_up = true; 1331 1332 if (qp->event_handler != NULL) 1333 qp->event_handler(qp->cb_data, NTB_LINK_UP); 1334 1335 ntb_db_clear_mask(dev, 1ull << qp->qp_num); 1336 } else if (nt->link_is_up) 1337 callout_reset(&qp->link_work, 1338 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp); 1339} 1340 1341/* Link down event*/ 1342static void 1343ntb_transport_link_cleanup(struct ntb_transport_ctx *nt) 1344{ 1345 struct ntb_transport_qp *qp; 1346 int i; 1347 1348 callout_drain(&nt->link_work); 1349 nt->link_is_up = 0; 1350 1351 /* Pass along the info to any clients */ 1352 for (i = 0; i < nt->qp_count; i++) { 1353 if ((nt->qp_bitmap & (1 << i)) != 0) { 1354 qp = &nt->qp_vec[i]; 1355 ntb_qp_link_cleanup(qp); 1356 callout_drain(&qp->link_work); 1357 } 1358 } 1359 1360 /* 1361 * The scratchpad registers keep the values if the remote side 1362 * goes down, blast them now to give them a sane value the next 1363 * time they are accessed 1364 */ 1365 ntb_spad_clear(nt->dev); 1366} 1367 1368static void 1369ntb_transport_link_cleanup_work(void *arg, int pending __unused) 1370{ 1371 1372 ntb_transport_link_cleanup(arg); 1373} 1374 1375static void 1376ntb_qp_link_down(struct ntb_transport_qp *qp) 1377{ 1378 1379 ntb_qp_link_cleanup(qp); 1380} 1381 1382static void 1383ntb_qp_link_down_reset(struct ntb_transport_qp *qp) 1384{ 1385 1386 qp->link_is_up = false; 1387 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 1388 1389 qp->tx_index = qp->rx_index = 0; 1390 qp->tx_bytes = qp->rx_bytes = 0; 1391 qp->tx_pkts = qp->rx_pkts = 0; 1392 1393 qp->rx_ring_empty = 0; 1394 qp->tx_ring_full = 0; 1395 1396 qp->rx_err_no_buf = qp->tx_err_no_buf = 0; 1397 qp->rx_err_oflow = qp->rx_err_ver = 0; 1398} 1399 1400static void 1401ntb_qp_link_cleanup(struct ntb_transport_qp *qp) 1402{ 1403 1404 callout_drain(&qp->link_work); 1405 ntb_qp_link_down_reset(qp); 1406 1407 if (qp->event_handler != NULL) 1408 qp->event_handler(qp->cb_data, NTB_LINK_DOWN); 1409} 1410 1411/* Link commanded down */ 1412/** 1413 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data 1414 * @qp: NTB transport layer queue to be disabled 1415 * 1416 * Notify NTB transport layer of client's desire to no longer receive data on 1417 * transport queue specified. It is the client's responsibility to ensure all 1418 * entries on queue are purged or otherwise handled appropriately. 1419 */ 1420void 1421ntb_transport_link_down(struct ntb_transport_qp *qp) 1422{ 1423 struct ntb_transport_ctx *nt = qp->transport; 1424 int i; 1425 uint32_t val; 1426 1427 qp->client_ready = false; 1428 for (i = 0, val = 0; i < nt->qp_count; i++) { 1429 if (nt->qp_vec[i].client_ready) 1430 val |= (1 << i); 1431 } 1432 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val); 1433 1434 if (qp->link_is_up) 1435 ntb_send_link_down(qp); 1436 else 1437 callout_drain(&qp->link_work); 1438} 1439 1440/** 1441 * ntb_transport_link_query - Query transport link state 1442 * @qp: NTB transport layer queue to be queried 1443 * 1444 * Query connectivity to the remote system of the NTB transport queue 1445 * 1446 * RETURNS: true for link up or false for link down 1447 */ 1448bool 1449ntb_transport_link_query(struct ntb_transport_qp *qp) 1450{ 1451 1452 return (qp->link_is_up); 1453} 1454 1455/** 1456 * ntb_transport_link_speed - Query transport link speed 1457 * @qp: NTB transport layer queue to be queried 1458 * 1459 * Query connection speed to the remote system of the NTB transport queue 1460 * 1461 * RETURNS: link speed in bits per second 1462 */ 1463uint64_t 1464ntb_transport_link_speed(struct ntb_transport_qp *qp) 1465{ 1466 struct ntb_transport_ctx *nt = qp->transport; 1467 uint64_t rate; 1468 1469 if (!nt->link_is_up) 1470 return (0); 1471 switch (nt->link_speed) { 1472 case NTB_SPEED_GEN1: 1473 rate = 2500000000 * 8 / 10; 1474 break; 1475 case NTB_SPEED_GEN2: 1476 rate = 5000000000 * 8 / 10; 1477 break; 1478 case NTB_SPEED_GEN3: 1479 rate = 8000000000 * 128 / 130; 1480 break; 1481 case NTB_SPEED_GEN4: 1482 rate = 16000000000 * 128 / 130; 1483 break; 1484 default: 1485 return (0); 1486 } 1487 if (nt->link_width <= 0) 1488 return (0); 1489 return (rate * nt->link_width); 1490} 1491 1492static void 1493ntb_send_link_down(struct ntb_transport_qp *qp) 1494{ 1495 struct ntb_queue_entry *entry; 1496 int i, rc; 1497 1498 if (!qp->link_is_up) 1499 return; 1500 1501 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) { 1502 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 1503 if (entry != NULL) 1504 break; 1505 pause("NTB Wait for link down", hz / 10); 1506 } 1507 1508 if (entry == NULL) 1509 return; 1510 1511 entry->cb_data = NULL; 1512 entry->buf = NULL; 1513 entry->len = 0; 1514 entry->flags = NTBT_LINK_DOWN_FLAG; 1515 1516 mtx_lock(&qp->tx_lock); 1517 rc = ntb_process_tx(qp, entry); 1518 mtx_unlock(&qp->tx_lock); 1519 if (rc != 0) 1520 printf("ntb: Failed to send link down\n"); 1521 1522 ntb_qp_link_down_reset(qp); 1523} 1524 1525 1526/* List Management */ 1527 1528static void 1529ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 1530 struct ntb_queue_list *list) 1531{ 1532 1533 mtx_lock_spin(lock); 1534 STAILQ_INSERT_TAIL(list, entry, entry); 1535 mtx_unlock_spin(lock); 1536} 1537 1538static struct ntb_queue_entry * 1539ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list) 1540{ 1541 struct ntb_queue_entry *entry; 1542 1543 mtx_lock_spin(lock); 1544 if (STAILQ_EMPTY(list)) { 1545 entry = NULL; 1546 goto out; 1547 } 1548 entry = STAILQ_FIRST(list); 1549 STAILQ_REMOVE_HEAD(list, entry); 1550out: 1551 mtx_unlock_spin(lock); 1552 1553 return (entry); 1554} 1555 1556static struct ntb_queue_entry * 1557ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from, 1558 struct ntb_queue_list *to) 1559{ 1560 struct ntb_queue_entry *entry; 1561 1562 mtx_lock_spin(lock); 1563 if (STAILQ_EMPTY(from)) { 1564 entry = NULL; 1565 goto out; 1566 } 1567 entry = STAILQ_FIRST(from); 1568 STAILQ_REMOVE_HEAD(from, entry); 1569 STAILQ_INSERT_TAIL(to, entry, entry); 1570 1571out: 1572 mtx_unlock_spin(lock); 1573 return (entry); 1574} 1575 1576/** 1577 * ntb_transport_qp_num - Query the qp number 1578 * @qp: NTB transport layer queue to be queried 1579 * 1580 * Query qp number of the NTB transport queue 1581 * 1582 * RETURNS: a zero based number specifying the qp number 1583 */ 1584unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp) 1585{ 1586 1587 return (qp->qp_num); 1588} 1589 1590/** 1591 * ntb_transport_max_size - Query the max payload size of a qp 1592 * @qp: NTB transport layer queue to be queried 1593 * 1594 * Query the maximum payload size permissible on the given qp 1595 * 1596 * RETURNS: the max payload size of a qp 1597 */ 1598unsigned int 1599ntb_transport_max_size(struct ntb_transport_qp *qp) 1600{ 1601 1602 return (qp->tx_max_frame - sizeof(struct ntb_payload_header)); 1603} 1604 1605unsigned int 1606ntb_transport_tx_free_entry(struct ntb_transport_qp *qp) 1607{ 1608 unsigned int head = qp->tx_index; 1609 unsigned int tail = qp->remote_rx_info->entry; 1610 1611 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head); 1612} 1613 1614static device_method_t ntb_transport_methods[] = { 1615 /* Device interface */ 1616 DEVMETHOD(device_probe, ntb_transport_probe), 1617 DEVMETHOD(device_attach, ntb_transport_attach), 1618 DEVMETHOD(device_detach, ntb_transport_detach), 1619 /* Bus interface */ 1620 DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str), 1621 DEVMETHOD(bus_print_child, ntb_transport_print_child), 1622 DEVMETHOD_END 1623}; 1624 1625devclass_t ntb_transport_devclass; 1626static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver, 1627 ntb_transport_methods, sizeof(struct ntb_transport_ctx)); 1628DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver, 1629 ntb_transport_devclass, NULL, NULL); 1630MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1); 1631MODULE_VERSION(ntb_transport, 1); 1632