ntb_transport.c revision 323455
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 323455 2017-09-11 18:50:36Z 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 caddr_t virt_addr; 186 bus_addr_t dma_addr; 187}; 188 189struct ntb_transport_child { 190 device_t dev; 191 int consumer; 192 int qpoff; 193 int qpcnt; 194 struct ntb_transport_child *next; 195}; 196 197struct ntb_transport_ctx { 198 device_t dev; 199 struct ntb_transport_child *child; 200 struct ntb_transport_mw *mw_vec; 201 struct ntb_transport_qp *qp_vec; 202 unsigned mw_count; 203 unsigned qp_count; 204 uint64_t qp_bitmap; 205 volatile bool link_is_up; 206 enum ntb_speed link_speed; 207 enum ntb_width link_width; 208 struct callout link_work; 209 struct callout link_watchdog; 210 struct task link_cleanup; 211}; 212 213enum { 214 NTBT_DESC_DONE_FLAG = 1 << 0, 215 NTBT_LINK_DOWN_FLAG = 1 << 1, 216}; 217 218struct ntb_payload_header { 219 ntb_q_idx_t ver; 220 uint32_t len; 221 uint32_t flags; 222}; 223 224enum { 225 /* 226 * The order of this enum is part of the remote protocol. Do not 227 * reorder without bumping protocol version (and it's probably best 228 * to keep the protocol in lock-step with the Linux NTB driver. 229 */ 230 NTBT_VERSION = 0, 231 NTBT_QP_LINKS, 232 NTBT_NUM_QPS, 233 NTBT_NUM_MWS, 234 /* 235 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2. 236 */ 237 NTBT_MW0_SZ_HIGH, 238 NTBT_MW0_SZ_LOW, 239 NTBT_MW1_SZ_HIGH, 240 NTBT_MW1_SZ_LOW, 241 242 /* 243 * Some NTB-using hardware have a watchdog to work around NTB hangs; if 244 * a register or doorbell isn't written every few seconds, the link is 245 * torn down. Write an otherwise unused register every few seconds to 246 * work around this watchdog. 247 */ 248 NTBT_WATCHDOG_SPAD = 15 249}; 250 251#define QP_TO_MW(nt, qp) ((qp) % nt->mw_count) 252#define NTB_QP_DEF_NUM_ENTRIES 100 253#define NTB_LINK_DOWN_TIMEOUT 10 254 255static int ntb_transport_probe(device_t dev); 256static int ntb_transport_attach(device_t dev); 257static int ntb_transport_detach(device_t dev); 258static void ntb_transport_init_queue(struct ntb_transport_ctx *nt, 259 unsigned int qp_num); 260static int ntb_process_tx(struct ntb_transport_qp *qp, 261 struct ntb_queue_entry *entry); 262static void ntb_transport_rxc_db(void *arg, int pending); 263static int ntb_process_rxc(struct ntb_transport_qp *qp); 264static void ntb_memcpy_rx(struct ntb_transport_qp *qp, 265 struct ntb_queue_entry *entry, void *offset); 266static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp, 267 void *data); 268static void ntb_complete_rxc(struct ntb_transport_qp *qp); 269static void ntb_transport_doorbell_callback(void *data, uint32_t vector); 270static void ntb_transport_event_callback(void *data); 271static void ntb_transport_link_work(void *arg); 272static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size); 273static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw); 274static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, 275 unsigned int qp_num); 276static void ntb_qp_link_work(void *arg); 277static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt); 278static void ntb_transport_link_cleanup_work(void *, int); 279static void ntb_qp_link_down(struct ntb_transport_qp *qp); 280static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp); 281static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp); 282static void ntb_send_link_down(struct ntb_transport_qp *qp); 283static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 284 struct ntb_queue_list *list); 285static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock, 286 struct ntb_queue_list *list); 287static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock, 288 struct ntb_queue_list *from, struct ntb_queue_list *to); 289static void xeon_link_watchdog_hb(void *); 290 291static const struct ntb_ctx_ops ntb_transport_ops = { 292 .link_event = ntb_transport_event_callback, 293 .db_event = ntb_transport_doorbell_callback, 294}; 295 296MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver"); 297 298static inline void 299iowrite32(uint32_t val, void *addr) 300{ 301 302 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr, 303 val); 304} 305 306/* Transport Init and teardown */ 307 308static void 309xeon_link_watchdog_hb(void *arg) 310{ 311 struct ntb_transport_ctx *nt; 312 313 nt = arg; 314 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0); 315 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt); 316} 317 318static int 319ntb_transport_probe(device_t dev) 320{ 321 322 device_set_desc(dev, "NTB Transport"); 323 return (0); 324} 325 326static int 327ntb_transport_attach(device_t dev) 328{ 329 struct ntb_transport_ctx *nt = device_get_softc(dev); 330 struct ntb_transport_child **cpp = &nt->child; 331 struct ntb_transport_child *nc; 332 struct ntb_transport_mw *mw; 333 uint64_t db_bitmap; 334 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt; 335 char cfg[128] = ""; 336 char buf[32]; 337 char *n, *np, *c, *name; 338 339 nt->dev = dev; 340 nt->mw_count = ntb_mw_count(dev); 341 spad_count = ntb_spad_count(dev); 342 db_bitmap = ntb_db_valid_mask(dev); 343 db_count = flsll(db_bitmap); 344 KASSERT(db_bitmap == (1 << db_count) - 1, 345 ("Doorbells are not sequential (%jx).\n", db_bitmap)); 346 347 if (nt->mw_count == 0) { 348 device_printf(dev, "At least 1 memory window required.\n"); 349 return (ENXIO); 350 } 351 if (spad_count < 6) { 352 device_printf(dev, "At least 6 scratchpads required.\n"); 353 return (ENXIO); 354 } 355 if (spad_count < 4 + 2 * nt->mw_count) { 356 nt->mw_count = (spad_count - 4) / 2; 357 device_printf(dev, "Scratchpads enough only for %d " 358 "memory windows.\n", nt->mw_count); 359 } 360 if (db_bitmap == 0) { 361 device_printf(dev, "At least one doorbell required.\n"); 362 return (ENXIO); 363 } 364 365 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T, 366 M_WAITOK | M_ZERO); 367 for (i = 0; i < nt->mw_count; i++) { 368 mw = &nt->mw_vec[i]; 369 370 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase, 371 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size, 372 &mw->addr_limit); 373 if (rc != 0) 374 goto err; 375 376 mw->buff_size = 0; 377 mw->xlat_size = 0; 378 mw->virt_addr = NULL; 379 mw->dma_addr = 0; 380 381 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING); 382 if (rc) 383 ntb_printf(0, "Unable to set mw%d caching\n", i); 384 } 385 386 qpu = 0; 387 qpo = imin(db_count, nt->mw_count); 388 qpt = db_count; 389 390 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev), 391 device_get_unit(dev)); 392 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg)); 393 n = cfg; 394 i = 0; 395 while ((c = strsep(&n, ",")) != NULL) { 396 np = c; 397 name = strsep(&np, ":"); 398 if (name != NULL && name[0] == 0) 399 name = NULL; 400 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu; 401 if (qp <= 0) 402 qp = 1; 403 404 if (qp > qpt - qpu) { 405 device_printf(dev, "Not enough resources for config\n"); 406 break; 407 } 408 409 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO); 410 nc->consumer = i; 411 nc->qpoff = qpu; 412 nc->qpcnt = qp; 413 nc->dev = device_add_child(dev, name, -1); 414 if (nc->dev == NULL) { 415 device_printf(dev, "Can not add child.\n"); 416 break; 417 } 418 device_set_ivars(nc->dev, nc); 419 *cpp = nc; 420 cpp = &nc->next; 421 422 if (bootverbose) { 423 device_printf(dev, "%d \"%s\": queues %d", 424 i, name, qpu); 425 if (qp > 1) 426 printf("-%d", qpu + qp - 1); 427 printf("\n"); 428 } 429 430 qpu += qp; 431 i++; 432 } 433 nt->qp_count = qpu; 434 435 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T, 436 M_WAITOK | M_ZERO); 437 438 for (i = 0; i < nt->qp_count; i++) 439 ntb_transport_init_queue(nt, i); 440 441 callout_init(&nt->link_work, 0); 442 callout_init(&nt->link_watchdog, 0); 443 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt); 444 nt->link_is_up = false; 445 446 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops); 447 if (rc != 0) 448 goto err; 449 450 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO); 451 452 if (enable_xeon_watchdog != 0) 453 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt); 454 455 bus_generic_attach(dev); 456 return (0); 457 458err: 459 free(nt->qp_vec, M_NTB_T); 460 free(nt->mw_vec, M_NTB_T); 461 return (rc); 462} 463 464static int 465ntb_transport_detach(device_t dev) 466{ 467 struct ntb_transport_ctx *nt = device_get_softc(dev); 468 struct ntb_transport_child **cpp = &nt->child; 469 struct ntb_transport_child *nc; 470 int error = 0, i; 471 472 while ((nc = *cpp) != NULL) { 473 *cpp = (*cpp)->next; 474 error = device_delete_child(dev, nc->dev); 475 if (error) 476 break; 477 free(nc, M_DEVBUF); 478 } 479 KASSERT(nt->qp_bitmap == 0, 480 ("Some queues not freed on detach (%jx)", nt->qp_bitmap)); 481 482 ntb_transport_link_cleanup(nt); 483 taskqueue_drain(taskqueue_swi, &nt->link_cleanup); 484 callout_drain(&nt->link_work); 485 callout_drain(&nt->link_watchdog); 486 487 ntb_link_disable(dev); 488 ntb_clear_ctx(dev); 489 490 for (i = 0; i < nt->mw_count; i++) 491 ntb_free_mw(nt, i); 492 493 free(nt->qp_vec, M_NTB_T); 494 free(nt->mw_vec, M_NTB_T); 495 return (0); 496} 497 498static int 499ntb_transport_print_child(device_t dev, device_t child) 500{ 501 struct ntb_transport_child *nc = device_get_ivars(child); 502 int retval; 503 504 retval = bus_print_child_header(dev, child); 505 if (nc->qpcnt > 0) { 506 printf(" queue %d", nc->qpoff); 507 if (nc->qpcnt > 1) 508 printf("-%d", nc->qpoff + nc->qpcnt - 1); 509 } 510 retval += printf(" at consumer %d", nc->consumer); 511 retval += bus_print_child_domain(dev, child); 512 retval += bus_print_child_footer(dev, child); 513 514 return (retval); 515} 516 517static int 518ntb_transport_child_location_str(device_t dev, device_t child, char *buf, 519 size_t buflen) 520{ 521 struct ntb_transport_child *nc = device_get_ivars(child); 522 523 snprintf(buf, buflen, "consumer=%d", nc->consumer); 524 return (0); 525} 526 527int 528ntb_transport_queue_count(device_t dev) 529{ 530 struct ntb_transport_child *nc = device_get_ivars(dev); 531 532 return (nc->qpcnt); 533} 534 535static void 536ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num) 537{ 538 struct ntb_transport_mw *mw; 539 struct ntb_transport_qp *qp; 540 vm_paddr_t mw_base; 541 uint64_t mw_size, qp_offset; 542 size_t tx_size; 543 unsigned num_qps_mw, mw_num, mw_count; 544 545 mw_count = nt->mw_count; 546 mw_num = QP_TO_MW(nt, qp_num); 547 mw = &nt->mw_vec[mw_num]; 548 549 qp = &nt->qp_vec[qp_num]; 550 qp->qp_num = qp_num; 551 qp->transport = nt; 552 qp->dev = nt->dev; 553 qp->client_ready = false; 554 qp->event_handler = NULL; 555 ntb_qp_link_down_reset(qp); 556 557 if (mw_num < nt->qp_count % mw_count) 558 num_qps_mw = nt->qp_count / mw_count + 1; 559 else 560 num_qps_mw = nt->qp_count / mw_count; 561 562 mw_base = mw->phys_addr; 563 mw_size = mw->phys_size; 564 565 tx_size = mw_size / num_qps_mw; 566 qp_offset = tx_size * (qp_num / mw_count); 567 568 qp->tx_mw = mw->vbase + qp_offset; 569 KASSERT(qp->tx_mw != NULL, ("uh oh?")); 570 571 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */ 572 qp->tx_mw_phys = mw_base + qp_offset; 573 KASSERT(qp->tx_mw_phys != 0, ("uh oh?")); 574 575 tx_size -= sizeof(struct ntb_rx_info); 576 qp->rx_info = (void *)(qp->tx_mw + tx_size); 577 578 /* Due to house-keeping, there must be at least 2 buffs */ 579 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2); 580 qp->tx_max_entry = tx_size / qp->tx_max_frame; 581 582 callout_init(&qp->link_work, 0); 583 callout_init(&qp->rx_full, 1); 584 585 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN); 586 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN); 587 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF); 588 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp); 589 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK, 590 taskqueue_thread_enqueue, &qp->rxc_tq); 591 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d", 592 device_get_nameunit(nt->dev), qp_num); 593 594 STAILQ_INIT(&qp->rx_post_q); 595 STAILQ_INIT(&qp->rx_pend_q); 596 STAILQ_INIT(&qp->tx_free_q); 597} 598 599void 600ntb_transport_free_queue(struct ntb_transport_qp *qp) 601{ 602 struct ntb_transport_ctx *nt = qp->transport; 603 struct ntb_queue_entry *entry; 604 605 callout_drain(&qp->link_work); 606 607 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 608 taskqueue_drain_all(qp->rxc_tq); 609 taskqueue_free(qp->rxc_tq); 610 611 qp->cb_data = NULL; 612 qp->rx_handler = NULL; 613 qp->tx_handler = NULL; 614 qp->event_handler = NULL; 615 616 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) 617 free(entry, M_NTB_T); 618 619 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) 620 free(entry, M_NTB_T); 621 622 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q))) 623 free(entry, M_NTB_T); 624 625 nt->qp_bitmap &= ~(1 << qp->qp_num); 626} 627 628/** 629 * ntb_transport_create_queue - Create a new NTB transport layer queue 630 * @rx_handler: receive callback function 631 * @tx_handler: transmit callback function 632 * @event_handler: event callback function 633 * 634 * Create a new NTB transport layer queue and provide the queue with a callback 635 * routine for both transmit and receive. The receive callback routine will be 636 * used to pass up data when the transport has received it on the queue. The 637 * transmit callback routine will be called when the transport has completed the 638 * transmission of the data on the queue and the data is ready to be freed. 639 * 640 * RETURNS: pointer to newly created ntb_queue, NULL on error. 641 */ 642struct ntb_transport_qp * 643ntb_transport_create_queue(device_t dev, int q, 644 const struct ntb_queue_handlers *handlers, void *data) 645{ 646 struct ntb_transport_child *nc = device_get_ivars(dev); 647 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev)); 648 struct ntb_queue_entry *entry; 649 struct ntb_transport_qp *qp; 650 int i; 651 652 if (q < 0 || q >= nc->qpcnt) 653 return (NULL); 654 655 qp = &nt->qp_vec[nc->qpoff + q]; 656 nt->qp_bitmap |= (1 << qp->qp_num); 657 qp->cb_data = data; 658 qp->rx_handler = handlers->rx_handler; 659 qp->tx_handler = handlers->tx_handler; 660 qp->event_handler = handlers->event_handler; 661 662 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 663 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 664 entry->cb_data = data; 665 entry->buf = NULL; 666 entry->len = transport_mtu; 667 entry->qp = qp; 668 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q); 669 } 670 671 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) { 672 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO); 673 entry->qp = qp; 674 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 675 } 676 677 ntb_db_clear(dev, 1ull << qp->qp_num); 678 return (qp); 679} 680 681/** 682 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue 683 * @qp: NTB transport layer queue to be enabled 684 * 685 * Notify NTB transport layer of client readiness to use queue 686 */ 687void 688ntb_transport_link_up(struct ntb_transport_qp *qp) 689{ 690 struct ntb_transport_ctx *nt = qp->transport; 691 692 qp->client_ready = true; 693 694 ntb_printf(2, "qp %d client ready\n", qp->qp_num); 695 696 if (nt->link_is_up) 697 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 698} 699 700 701 702/* Transport Tx */ 703 704/** 705 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry 706 * @qp: NTB transport layer queue the entry is to be enqueued on 707 * @cb: per buffer pointer for callback function to use 708 * @data: pointer to data buffer that will be sent 709 * @len: length of the data buffer 710 * 711 * Enqueue a new transmit buffer onto the transport queue from which a NTB 712 * payload will be transmitted. This assumes that a lock is being held to 713 * serialize access to the qp. 714 * 715 * RETURNS: An appropriate ERRNO error value on error, or zero for success. 716 */ 717int 718ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data, 719 unsigned int len) 720{ 721 struct ntb_queue_entry *entry; 722 int rc; 723 724 if (!qp->link_is_up || len == 0) { 725 CTR0(KTR_NTB, "TX: link not up"); 726 return (EINVAL); 727 } 728 729 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 730 if (entry == NULL) { 731 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q"); 732 qp->tx_err_no_buf++; 733 return (EBUSY); 734 } 735 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry); 736 737 entry->cb_data = cb; 738 entry->buf = data; 739 entry->len = len; 740 entry->flags = 0; 741 742 mtx_lock(&qp->tx_lock); 743 rc = ntb_process_tx(qp, entry); 744 mtx_unlock(&qp->tx_lock); 745 if (rc != 0) { 746 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 747 CTR1(KTR_NTB, 748 "TX: process_tx failed. Returning entry %p to tx_free_q", 749 entry); 750 } 751 return (rc); 752} 753 754static void 755ntb_tx_copy_callback(void *data) 756{ 757 struct ntb_queue_entry *entry = data; 758 struct ntb_transport_qp *qp = entry->qp; 759 struct ntb_payload_header *hdr = entry->x_hdr; 760 761 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags); 762 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr); 763 764 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num); 765 766 /* 767 * The entry length can only be zero if the packet is intended to be a 768 * "link down" or similar. Since no payload is being sent in these 769 * cases, there is nothing to add to the completion queue. 770 */ 771 if (entry->len > 0) { 772 qp->tx_bytes += entry->len; 773 774 if (qp->tx_handler) 775 qp->tx_handler(qp, qp->cb_data, entry->buf, 776 entry->len); 777 else 778 m_freem(entry->buf); 779 entry->buf = NULL; 780 } 781 782 CTR3(KTR_NTB, 783 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning " 784 "to tx_free_q", entry, hdr->ver, hdr->flags); 785 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 786} 787 788static void 789ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset) 790{ 791 792 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset); 793 if (entry->buf != NULL) { 794 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset); 795 796 /* 797 * Ensure that the data is fully copied before setting the 798 * flags 799 */ 800 wmb(); 801 } 802 803 ntb_tx_copy_callback(entry); 804} 805 806static void 807ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 808{ 809 struct ntb_payload_header *hdr; 810 void *offset; 811 812 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index; 813 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame - 814 sizeof(struct ntb_payload_header)); 815 entry->x_hdr = hdr; 816 817 iowrite32(entry->len, &hdr->len); 818 iowrite32(qp->tx_pkts, &hdr->ver); 819 820 ntb_memcpy_tx(entry, offset); 821} 822 823static int 824ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry) 825{ 826 827 CTR3(KTR_NTB, 828 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u", 829 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry); 830 if (qp->tx_index == qp->remote_rx_info->entry) { 831 CTR0(KTR_NTB, "TX: ring full"); 832 qp->tx_ring_full++; 833 return (EAGAIN); 834 } 835 836 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) { 837 if (qp->tx_handler != NULL) 838 qp->tx_handler(qp, qp->cb_data, entry->buf, 839 EIO); 840 else 841 m_freem(entry->buf); 842 843 entry->buf = NULL; 844 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q); 845 CTR1(KTR_NTB, 846 "TX: frame too big. returning entry %p to tx_free_q", 847 entry); 848 return (0); 849 } 850 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index); 851 ntb_async_tx(qp, entry); 852 853 qp->tx_index++; 854 qp->tx_index %= qp->tx_max_entry; 855 856 qp->tx_pkts++; 857 858 return (0); 859} 860 861/* Transport Rx */ 862static void 863ntb_transport_rxc_db(void *arg, int pending __unused) 864{ 865 struct ntb_transport_qp *qp = arg; 866 uint64_t qp_mask = 1ull << qp->qp_num; 867 int rc; 868 869 CTR0(KTR_NTB, "RX: transport_rx"); 870again: 871 while ((rc = ntb_process_rxc(qp)) == 0) 872 ; 873 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc); 874 875 if ((ntb_db_read(qp->dev) & qp_mask) != 0) { 876 /* If db is set, clear it and check queue once more. */ 877 ntb_db_clear(qp->dev, qp_mask); 878 goto again; 879 } 880 if (qp->link_is_up) 881 ntb_db_clear_mask(qp->dev, qp_mask); 882} 883 884static int 885ntb_process_rxc(struct ntb_transport_qp *qp) 886{ 887 struct ntb_payload_header *hdr; 888 struct ntb_queue_entry *entry; 889 caddr_t offset; 890 891 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index; 892 hdr = (void *)(offset + qp->rx_max_frame - 893 sizeof(struct ntb_payload_header)); 894 895 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index); 896 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) { 897 CTR0(KTR_NTB, "RX: hdr not done"); 898 qp->rx_ring_empty++; 899 return (EAGAIN); 900 } 901 902 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) { 903 CTR0(KTR_NTB, "RX: link down"); 904 ntb_qp_link_down(qp); 905 hdr->flags = 0; 906 return (EAGAIN); 907 } 908 909 if (hdr->ver != (uint32_t)qp->rx_pkts) { 910 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). " 911 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts); 912 qp->rx_err_ver++; 913 return (EIO); 914 } 915 916 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q); 917 if (entry == NULL) { 918 qp->rx_err_no_buf++; 919 CTR0(KTR_NTB, "RX: No entries in rx_pend_q"); 920 return (EAGAIN); 921 } 922 callout_stop(&qp->rx_full); 923 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry); 924 925 entry->x_hdr = hdr; 926 entry->index = qp->rx_index; 927 928 if (hdr->len > entry->len) { 929 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju", 930 (uintmax_t)hdr->len, (uintmax_t)entry->len); 931 qp->rx_err_oflow++; 932 933 entry->len = -EIO; 934 entry->flags |= NTBT_DESC_DONE_FLAG; 935 936 ntb_complete_rxc(qp); 937 } else { 938 qp->rx_bytes += hdr->len; 939 qp->rx_pkts++; 940 941 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts); 942 943 entry->len = hdr->len; 944 945 ntb_memcpy_rx(qp, entry, offset); 946 } 947 948 qp->rx_index++; 949 qp->rx_index %= qp->rx_max_entry; 950 return (0); 951} 952 953static void 954ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry, 955 void *offset) 956{ 957 struct ifnet *ifp = entry->cb_data; 958 unsigned int len = entry->len; 959 960 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset); 961 962 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL); 963 if (entry->buf == NULL) 964 entry->len = -ENOMEM; 965 966 /* Ensure that the data is globally visible before clearing the flag */ 967 wmb(); 968 969 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf); 970 ntb_rx_copy_callback(qp, entry); 971} 972 973static inline void 974ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data) 975{ 976 struct ntb_queue_entry *entry; 977 978 entry = data; 979 entry->flags |= NTBT_DESC_DONE_FLAG; 980 ntb_complete_rxc(qp); 981} 982 983static void 984ntb_complete_rxc(struct ntb_transport_qp *qp) 985{ 986 struct ntb_queue_entry *entry; 987 struct mbuf *m; 988 unsigned len; 989 990 CTR0(KTR_NTB, "RX: rx_completion_task"); 991 992 mtx_lock_spin(&qp->ntb_rx_q_lock); 993 994 while (!STAILQ_EMPTY(&qp->rx_post_q)) { 995 entry = STAILQ_FIRST(&qp->rx_post_q); 996 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0) 997 break; 998 999 entry->x_hdr->flags = 0; 1000 iowrite32(entry->index, &qp->rx_info->entry); 1001 1002 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry); 1003 1004 len = entry->len; 1005 m = entry->buf; 1006 1007 /* 1008 * Re-initialize queue_entry for reuse; rx_handler takes 1009 * ownership of the mbuf. 1010 */ 1011 entry->buf = NULL; 1012 entry->len = transport_mtu; 1013 entry->cb_data = qp->cb_data; 1014 1015 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry); 1016 1017 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1018 1019 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m); 1020 if (qp->rx_handler != NULL && qp->client_ready) 1021 qp->rx_handler(qp, qp->cb_data, m, len); 1022 else 1023 m_freem(m); 1024 1025 mtx_lock_spin(&qp->ntb_rx_q_lock); 1026 } 1027 1028 mtx_unlock_spin(&qp->ntb_rx_q_lock); 1029} 1030 1031static void 1032ntb_transport_doorbell_callback(void *data, uint32_t vector) 1033{ 1034 struct ntb_transport_ctx *nt = data; 1035 struct ntb_transport_qp *qp; 1036 uint64_t vec_mask; 1037 unsigned qp_num; 1038 1039 vec_mask = ntb_db_vector_mask(nt->dev, vector); 1040 vec_mask &= nt->qp_bitmap; 1041 if ((vec_mask & (vec_mask - 1)) != 0) 1042 vec_mask &= ntb_db_read(nt->dev); 1043 if (vec_mask != 0) { 1044 ntb_db_set_mask(nt->dev, vec_mask); 1045 ntb_db_clear(nt->dev, vec_mask); 1046 } 1047 while (vec_mask != 0) { 1048 qp_num = ffsll(vec_mask) - 1; 1049 1050 qp = &nt->qp_vec[qp_num]; 1051 if (qp->link_is_up) 1052 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work); 1053 1054 vec_mask &= ~(1ull << qp_num); 1055 } 1056} 1057 1058/* Link Event handler */ 1059static void 1060ntb_transport_event_callback(void *data) 1061{ 1062 struct ntb_transport_ctx *nt = data; 1063 1064 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) { 1065 ntb_printf(1, "HW link up\n"); 1066 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt); 1067 } else { 1068 ntb_printf(1, "HW link down\n"); 1069 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup); 1070 } 1071} 1072 1073/* Link bring up */ 1074static void 1075ntb_transport_link_work(void *arg) 1076{ 1077 struct ntb_transport_ctx *nt = arg; 1078 device_t dev = nt->dev; 1079 struct ntb_transport_qp *qp; 1080 uint64_t val64, size; 1081 uint32_t val; 1082 unsigned i; 1083 int rc; 1084 1085 /* send the local info, in the opposite order of the way we read it */ 1086 for (i = 0; i < nt->mw_count; i++) { 1087 size = nt->mw_vec[i].phys_size; 1088 1089 if (max_mw_size != 0 && size > max_mw_size) 1090 size = max_mw_size; 1091 1092 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2), 1093 size >> 32); 1094 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size); 1095 } 1096 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count); 1097 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count); 1098 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0); 1099 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION); 1100 1101 /* Query the remote side for its info */ 1102 val = 0; 1103 ntb_spad_read(dev, NTBT_VERSION, &val); 1104 if (val != NTB_TRANSPORT_VERSION) 1105 goto out; 1106 1107 ntb_spad_read(dev, NTBT_NUM_QPS, &val); 1108 if (val != nt->qp_count) 1109 goto out; 1110 1111 ntb_spad_read(dev, NTBT_NUM_MWS, &val); 1112 if (val != nt->mw_count) 1113 goto out; 1114 1115 for (i = 0; i < nt->mw_count; i++) { 1116 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val); 1117 val64 = (uint64_t)val << 32; 1118 1119 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val); 1120 val64 |= val; 1121 1122 rc = ntb_set_mw(nt, i, val64); 1123 if (rc != 0) 1124 goto free_mws; 1125 } 1126 1127 nt->link_is_up = true; 1128 ntb_printf(1, "transport link up\n"); 1129 1130 for (i = 0; i < nt->qp_count; i++) { 1131 qp = &nt->qp_vec[i]; 1132 1133 ntb_transport_setup_qp_mw(nt, i); 1134 1135 if (qp->client_ready) 1136 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp); 1137 } 1138 1139 return; 1140 1141free_mws: 1142 for (i = 0; i < nt->mw_count; i++) 1143 ntb_free_mw(nt, i); 1144out: 1145 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width)) 1146 callout_reset(&nt->link_work, 1147 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt); 1148} 1149 1150static int 1151ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size) 1152{ 1153 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1154 size_t xlat_size, buff_size; 1155 int rc; 1156 1157 if (size == 0) 1158 return (EINVAL); 1159 1160 xlat_size = roundup(size, mw->xlat_align_size); 1161 buff_size = xlat_size; 1162 1163 /* No need to re-setup */ 1164 if (mw->xlat_size == xlat_size) 1165 return (0); 1166 1167 if (mw->buff_size != 0) 1168 ntb_free_mw(nt, num_mw); 1169 1170 /* Alloc memory for receiving data. Must be aligned */ 1171 mw->xlat_size = xlat_size; 1172 mw->buff_size = buff_size; 1173 1174 mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_T, M_ZERO, 0, 1175 mw->addr_limit, mw->xlat_align, 0); 1176 if (mw->virt_addr == NULL) { 1177 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n", 1178 mw->buff_size, mw->xlat_size); 1179 mw->xlat_size = 0; 1180 mw->buff_size = 0; 1181 return (ENOMEM); 1182 } 1183 /* TODO: replace with bus_space_* functions */ 1184 mw->dma_addr = vtophys(mw->virt_addr); 1185 1186 /* 1187 * Ensure that the allocation from contigmalloc is aligned as 1188 * requested. XXX: This may not be needed -- brought in for parity 1189 * with the Linux driver. 1190 */ 1191 if (mw->dma_addr % mw->xlat_align != 0) { 1192 ntb_printf(0, 1193 "DMA memory 0x%jx not aligned to BAR size 0x%zx\n", 1194 (uintmax_t)mw->dma_addr, size); 1195 ntb_free_mw(nt, num_mw); 1196 return (ENOMEM); 1197 } 1198 1199 /* Notify HW the memory location of the receive buffer */ 1200 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size); 1201 if (rc) { 1202 ntb_printf(0, "Unable to set mw%d translation\n", num_mw); 1203 ntb_free_mw(nt, num_mw); 1204 return (rc); 1205 } 1206 1207 return (0); 1208} 1209 1210static void 1211ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw) 1212{ 1213 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw]; 1214 1215 if (mw->virt_addr == NULL) 1216 return; 1217 1218 ntb_mw_clear_trans(nt->dev, num_mw); 1219 contigfree(mw->virt_addr, mw->xlat_size, M_NTB_T); 1220 mw->xlat_size = 0; 1221 mw->buff_size = 0; 1222 mw->virt_addr = NULL; 1223} 1224 1225static int 1226ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num) 1227{ 1228 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num]; 1229 struct ntb_transport_mw *mw; 1230 void *offset; 1231 ntb_q_idx_t i; 1232 size_t rx_size; 1233 unsigned num_qps_mw, mw_num, mw_count; 1234 1235 mw_count = nt->mw_count; 1236 mw_num = QP_TO_MW(nt, qp_num); 1237 mw = &nt->mw_vec[mw_num]; 1238 1239 if (mw->virt_addr == NULL) 1240 return (ENOMEM); 1241 1242 if (mw_num < nt->qp_count % mw_count) 1243 num_qps_mw = nt->qp_count / mw_count + 1; 1244 else 1245 num_qps_mw = nt->qp_count / mw_count; 1246 1247 rx_size = mw->xlat_size / num_qps_mw; 1248 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count); 1249 rx_size -= sizeof(struct ntb_rx_info); 1250 1251 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size); 1252 1253 /* Due to house-keeping, there must be at least 2 buffs */ 1254 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2); 1255 qp->rx_max_entry = rx_size / qp->rx_max_frame; 1256 qp->rx_index = 0; 1257 1258 qp->remote_rx_info->entry = qp->rx_max_entry - 1; 1259 1260 /* Set up the hdr offsets with 0s */ 1261 for (i = 0; i < qp->rx_max_entry; i++) { 1262 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) - 1263 sizeof(struct ntb_payload_header)); 1264 memset(offset, 0, sizeof(struct ntb_payload_header)); 1265 } 1266 1267 qp->rx_pkts = 0; 1268 qp->tx_pkts = 0; 1269 qp->tx_index = 0; 1270 1271 return (0); 1272} 1273 1274static void 1275ntb_qp_link_work(void *arg) 1276{ 1277 struct ntb_transport_qp *qp = arg; 1278 device_t dev = qp->dev; 1279 struct ntb_transport_ctx *nt = qp->transport; 1280 int i; 1281 uint32_t val; 1282 1283 /* Report queues that are up on our side */ 1284 for (i = 0, val = 0; i < nt->qp_count; i++) { 1285 if (nt->qp_vec[i].client_ready) 1286 val |= (1 << i); 1287 } 1288 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val); 1289 1290 /* See if the remote side is up */ 1291 ntb_spad_read(dev, NTBT_QP_LINKS, &val); 1292 if ((val & (1ull << qp->qp_num)) != 0) { 1293 ntb_printf(2, "qp %d link up\n", qp->qp_num); 1294 qp->link_is_up = true; 1295 1296 if (qp->event_handler != NULL) 1297 qp->event_handler(qp->cb_data, NTB_LINK_UP); 1298 1299 ntb_db_clear_mask(dev, 1ull << qp->qp_num); 1300 } else if (nt->link_is_up) 1301 callout_reset(&qp->link_work, 1302 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp); 1303} 1304 1305/* Link down event*/ 1306static void 1307ntb_transport_link_cleanup(struct ntb_transport_ctx *nt) 1308{ 1309 struct ntb_transport_qp *qp; 1310 int i; 1311 1312 callout_drain(&nt->link_work); 1313 nt->link_is_up = 0; 1314 1315 /* Pass along the info to any clients */ 1316 for (i = 0; i < nt->qp_count; i++) { 1317 if ((nt->qp_bitmap & (1 << i)) != 0) { 1318 qp = &nt->qp_vec[i]; 1319 ntb_qp_link_cleanup(qp); 1320 callout_drain(&qp->link_work); 1321 } 1322 } 1323 1324 /* 1325 * The scratchpad registers keep the values if the remote side 1326 * goes down, blast them now to give them a sane value the next 1327 * time they are accessed 1328 */ 1329 ntb_spad_clear(nt->dev); 1330} 1331 1332static void 1333ntb_transport_link_cleanup_work(void *arg, int pending __unused) 1334{ 1335 1336 ntb_transport_link_cleanup(arg); 1337} 1338 1339static void 1340ntb_qp_link_down(struct ntb_transport_qp *qp) 1341{ 1342 1343 ntb_qp_link_cleanup(qp); 1344} 1345 1346static void 1347ntb_qp_link_down_reset(struct ntb_transport_qp *qp) 1348{ 1349 1350 qp->link_is_up = false; 1351 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num); 1352 1353 qp->tx_index = qp->rx_index = 0; 1354 qp->tx_bytes = qp->rx_bytes = 0; 1355 qp->tx_pkts = qp->rx_pkts = 0; 1356 1357 qp->rx_ring_empty = 0; 1358 qp->tx_ring_full = 0; 1359 1360 qp->rx_err_no_buf = qp->tx_err_no_buf = 0; 1361 qp->rx_err_oflow = qp->rx_err_ver = 0; 1362} 1363 1364static void 1365ntb_qp_link_cleanup(struct ntb_transport_qp *qp) 1366{ 1367 1368 callout_drain(&qp->link_work); 1369 ntb_qp_link_down_reset(qp); 1370 1371 if (qp->event_handler != NULL) 1372 qp->event_handler(qp->cb_data, NTB_LINK_DOWN); 1373} 1374 1375/* Link commanded down */ 1376/** 1377 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data 1378 * @qp: NTB transport layer queue to be disabled 1379 * 1380 * Notify NTB transport layer of client's desire to no longer receive data on 1381 * transport queue specified. It is the client's responsibility to ensure all 1382 * entries on queue are purged or otherwise handled appropriately. 1383 */ 1384void 1385ntb_transport_link_down(struct ntb_transport_qp *qp) 1386{ 1387 struct ntb_transport_ctx *nt = qp->transport; 1388 int i; 1389 uint32_t val; 1390 1391 qp->client_ready = false; 1392 for (i = 0, val = 0; i < nt->qp_count; i++) { 1393 if (nt->qp_vec[i].client_ready) 1394 val |= (1 << i); 1395 } 1396 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val); 1397 1398 if (qp->link_is_up) 1399 ntb_send_link_down(qp); 1400 else 1401 callout_drain(&qp->link_work); 1402} 1403 1404/** 1405 * ntb_transport_link_query - Query transport link state 1406 * @qp: NTB transport layer queue to be queried 1407 * 1408 * Query connectivity to the remote system of the NTB transport queue 1409 * 1410 * RETURNS: true for link up or false for link down 1411 */ 1412bool 1413ntb_transport_link_query(struct ntb_transport_qp *qp) 1414{ 1415 1416 return (qp->link_is_up); 1417} 1418 1419/** 1420 * ntb_transport_link_speed - Query transport link speed 1421 * @qp: NTB transport layer queue to be queried 1422 * 1423 * Query connection speed to the remote system of the NTB transport queue 1424 * 1425 * RETURNS: link speed in bits per second 1426 */ 1427uint64_t 1428ntb_transport_link_speed(struct ntb_transport_qp *qp) 1429{ 1430 struct ntb_transport_ctx *nt = qp->transport; 1431 uint64_t rate; 1432 1433 if (!nt->link_is_up) 1434 return (0); 1435 switch (nt->link_speed) { 1436 case NTB_SPEED_GEN1: 1437 rate = 2500000000 * 8 / 10; 1438 break; 1439 case NTB_SPEED_GEN2: 1440 rate = 5000000000 * 8 / 10; 1441 break; 1442 case NTB_SPEED_GEN3: 1443 rate = 8000000000 * 128 / 130; 1444 break; 1445 case NTB_SPEED_GEN4: 1446 rate = 16000000000 * 128 / 130; 1447 break; 1448 default: 1449 return (0); 1450 } 1451 if (nt->link_width <= 0) 1452 return (0); 1453 return (rate * nt->link_width); 1454} 1455 1456static void 1457ntb_send_link_down(struct ntb_transport_qp *qp) 1458{ 1459 struct ntb_queue_entry *entry; 1460 int i, rc; 1461 1462 if (!qp->link_is_up) 1463 return; 1464 1465 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) { 1466 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q); 1467 if (entry != NULL) 1468 break; 1469 pause("NTB Wait for link down", hz / 10); 1470 } 1471 1472 if (entry == NULL) 1473 return; 1474 1475 entry->cb_data = NULL; 1476 entry->buf = NULL; 1477 entry->len = 0; 1478 entry->flags = NTBT_LINK_DOWN_FLAG; 1479 1480 mtx_lock(&qp->tx_lock); 1481 rc = ntb_process_tx(qp, entry); 1482 mtx_unlock(&qp->tx_lock); 1483 if (rc != 0) 1484 printf("ntb: Failed to send link down\n"); 1485 1486 ntb_qp_link_down_reset(qp); 1487} 1488 1489 1490/* List Management */ 1491 1492static void 1493ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry, 1494 struct ntb_queue_list *list) 1495{ 1496 1497 mtx_lock_spin(lock); 1498 STAILQ_INSERT_TAIL(list, entry, entry); 1499 mtx_unlock_spin(lock); 1500} 1501 1502static struct ntb_queue_entry * 1503ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list) 1504{ 1505 struct ntb_queue_entry *entry; 1506 1507 mtx_lock_spin(lock); 1508 if (STAILQ_EMPTY(list)) { 1509 entry = NULL; 1510 goto out; 1511 } 1512 entry = STAILQ_FIRST(list); 1513 STAILQ_REMOVE_HEAD(list, entry); 1514out: 1515 mtx_unlock_spin(lock); 1516 1517 return (entry); 1518} 1519 1520static struct ntb_queue_entry * 1521ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from, 1522 struct ntb_queue_list *to) 1523{ 1524 struct ntb_queue_entry *entry; 1525 1526 mtx_lock_spin(lock); 1527 if (STAILQ_EMPTY(from)) { 1528 entry = NULL; 1529 goto out; 1530 } 1531 entry = STAILQ_FIRST(from); 1532 STAILQ_REMOVE_HEAD(from, entry); 1533 STAILQ_INSERT_TAIL(to, entry, entry); 1534 1535out: 1536 mtx_unlock_spin(lock); 1537 return (entry); 1538} 1539 1540/** 1541 * ntb_transport_qp_num - Query the qp number 1542 * @qp: NTB transport layer queue to be queried 1543 * 1544 * Query qp number of the NTB transport queue 1545 * 1546 * RETURNS: a zero based number specifying the qp number 1547 */ 1548unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp) 1549{ 1550 1551 return (qp->qp_num); 1552} 1553 1554/** 1555 * ntb_transport_max_size - Query the max payload size of a qp 1556 * @qp: NTB transport layer queue to be queried 1557 * 1558 * Query the maximum payload size permissible on the given qp 1559 * 1560 * RETURNS: the max payload size of a qp 1561 */ 1562unsigned int 1563ntb_transport_max_size(struct ntb_transport_qp *qp) 1564{ 1565 1566 return (qp->tx_max_frame - sizeof(struct ntb_payload_header)); 1567} 1568 1569unsigned int 1570ntb_transport_tx_free_entry(struct ntb_transport_qp *qp) 1571{ 1572 unsigned int head = qp->tx_index; 1573 unsigned int tail = qp->remote_rx_info->entry; 1574 1575 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head); 1576} 1577 1578static device_method_t ntb_transport_methods[] = { 1579 /* Device interface */ 1580 DEVMETHOD(device_probe, ntb_transport_probe), 1581 DEVMETHOD(device_attach, ntb_transport_attach), 1582 DEVMETHOD(device_detach, ntb_transport_detach), 1583 /* Bus interface */ 1584 DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str), 1585 DEVMETHOD(bus_print_child, ntb_transport_print_child), 1586 DEVMETHOD_END 1587}; 1588 1589devclass_t ntb_transport_devclass; 1590static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver, 1591 ntb_transport_methods, sizeof(struct ntb_transport_ctx)); 1592DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver, 1593 ntb_transport_devclass, NULL, NULL); 1594MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1); 1595MODULE_VERSION(ntb_transport, 1); 1596