1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (C) 2012-2016 Intel 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 #include "opt_nvme.h" 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/buf.h> 34 #include <sys/bus.h> 35 #include <sys/conf.h> 36 #include <sys/ioccom.h> 37 #include <sys/proc.h> 38 #include <sys/smp.h> 39 #include <sys/uio.h> 40 #include <sys/sbuf.h> 41 #include <sys/endian.h> 42 #include <sys/stdarg.h> 43 #include <vm/vm.h> 44 #include <vm/vm_page.h> 45 #include <vm/vm_extern.h> 46 #include <vm/vm_map.h> 47 48 #include <dev/pci/pcivar.h> 49 50 #include "nvme_private.h" 51 #include "nvme_linux.h" 52 53 #include "nvme_if.h" 54 55 #define B4_CHK_RDY_DELAY_MS 2300 /* work around controller bug */ 56 57 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr, 58 struct nvme_async_event_request *aer); 59 60 static void 61 nvme_ctrlr_barrier(struct nvme_controller *ctrlr, int flags) 62 { 63 bus_barrier(ctrlr->resource, 0, rman_get_size(ctrlr->resource), flags); 64 } 65 66 static void 67 nvme_ctrlr_devctl_va(struct nvme_controller *ctrlr, const char *type, 68 const char *msg, va_list ap) 69 { 70 struct sbuf sb; 71 int error; 72 73 if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL) 74 return; 75 sbuf_printf(&sb, "name=\"%s\" ", device_get_nameunit(ctrlr->dev)); 76 sbuf_vprintf(&sb, msg, ap); 77 error = sbuf_finish(&sb); 78 if (error == 0) 79 devctl_notify("nvme", "controller", type, sbuf_data(&sb)); 80 sbuf_delete(&sb); 81 } 82 83 static void 84 nvme_ctrlr_devctl(struct nvme_controller *ctrlr, const char *type, const char *msg, ...) 85 { 86 va_list ap; 87 88 va_start(ap, msg); 89 nvme_ctrlr_devctl_va(ctrlr, type, msg, ap); 90 va_end(ap); 91 } 92 93 static void 94 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...) 95 { 96 struct sbuf sb; 97 va_list ap; 98 int error; 99 100 if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL) 101 return; 102 sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev)); 103 va_start(ap, msg); 104 sbuf_vprintf(&sb, msg, ap); 105 va_end(ap); 106 error = sbuf_finish(&sb); 107 if (error == 0) 108 printf("%s\n", sbuf_data(&sb)); 109 sbuf_delete(&sb); 110 va_start(ap, msg); 111 nvme_ctrlr_devctl_va(ctrlr, type, msg, ap); 112 va_end(ap); 113 } 114 115 static int 116 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr) 117 { 118 struct nvme_qpair *qpair; 119 uint32_t num_entries; 120 int error; 121 122 qpair = &ctrlr->adminq; 123 qpair->id = 0; 124 qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1; 125 qpair->domain = ctrlr->domain; 126 127 num_entries = NVME_ADMIN_ENTRIES; 128 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries); 129 /* 130 * If admin_entries was overridden to an invalid value, revert it 131 * back to our default value. 132 */ 133 if (num_entries < NVME_MIN_ADMIN_ENTRIES || 134 num_entries > NVME_MAX_ADMIN_ENTRIES) { 135 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d " 136 "specified\n", num_entries); 137 num_entries = NVME_ADMIN_ENTRIES; 138 } 139 140 /* 141 * The admin queue's max xfer size is treated differently than the 142 * max I/O xfer size. 16KB is sufficient here - maybe even less? 143 */ 144 error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS, 145 ctrlr); 146 return (error); 147 } 148 149 #define QP(ctrlr, c) ((c) * (ctrlr)->num_io_queues / mp_ncpus) 150 151 static int 152 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr) 153 { 154 struct nvme_qpair *qpair; 155 uint32_t cap_lo; 156 uint16_t mqes; 157 int c, error, i, n; 158 int num_entries, num_trackers, max_entries; 159 160 /* 161 * NVMe spec sets a hard limit of 64K max entries, but devices may 162 * specify a smaller limit, so we need to check the MQES field in the 163 * capabilities register. We have to cap the number of entries to the 164 * current stride allows for in BAR 0/1, otherwise the remainder entries 165 * are inaccessible. MQES should reflect this, and this is just a 166 * fail-safe. 167 */ 168 max_entries = 169 (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) / 170 (1 << (ctrlr->dstrd + 1)); 171 num_entries = NVME_IO_ENTRIES; 172 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries); 173 cap_lo = nvme_mmio_read_4(ctrlr, cap_lo); 174 mqes = NVME_CAP_LO_MQES(cap_lo); 175 num_entries = min(num_entries, mqes + 1); 176 num_entries = min(num_entries, max_entries); 177 178 /* SHARED_CID_SPACE: IO CIDs must fit within the shared CID table. */ 179 if (ctrlr->quirks & QUIRK_APPLE_SHARED_CID_SPACE) 180 num_entries = min(num_entries, NVME_ADMIN_ENTRIES); 181 182 num_trackers = NVME_IO_TRACKERS; 183 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers); 184 185 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS); 186 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS); 187 /* 188 * No need to have more trackers than entries in the submit queue. Note 189 * also that for a queue size of N, we can only have (N-1) commands 190 * outstanding, hence the "-1" here. 191 */ 192 num_trackers = min(num_trackers, (num_entries-1)); 193 194 if (ctrlr->cdata.maxcmd != 0 && num_trackers > ctrlr->cdata.maxcmd) { 195 nvme_printf(ctrlr, 196 "limiting trackers per I/O queue to MAXCMD (%u -> %u)\n", 197 num_trackers, ctrlr->cdata.maxcmd); 198 num_trackers = ctrlr->cdata.maxcmd; 199 } 200 201 if (ctrlr->quirks & QUIRK_APPLE_SHARED_CID_SPACE) 202 num_trackers = min(num_trackers, 203 NVME_ADMIN_ENTRIES - ctrlr->adminq.num_trackers); 204 if (ctrlr->quirks & QUIRK_APPLE_S3X_SERIALIZE) 205 num_trackers = 1; 206 207 /* 208 * Our best estimate for the maximum number of I/Os that we should 209 * normally have in flight at one time. This should be viewed as a hint, 210 * not a hard limit and will need to be revisited when the upper layers 211 * of the storage system grows multi-queue support. 212 */ 213 ctrlr->max_hw_pend_io = max(1, 214 num_trackers * ctrlr->num_io_queues * 3 / 4); 215 216 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair), 217 M_NVME, M_ZERO | M_WAITOK); 218 219 for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) { 220 qpair = &ctrlr->ioq[i]; 221 222 /* 223 * Admin queue has ID=0. IO queues start at ID=1 - 224 * hence the 'i+1' here. 225 */ 226 qpair->id = i + 1; 227 if (ctrlr->num_io_queues > 1) { 228 /* Find number of CPUs served by this queue. */ 229 for (n = 1; QP(ctrlr, c + n) == i; n++) 230 ; 231 /* Shuffle multiple NVMe devices between CPUs. */ 232 qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n; 233 qpair->domain = pcpu_find(qpair->cpu)->pc_domain; 234 } else { 235 qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1; 236 qpair->domain = ctrlr->domain; 237 } 238 239 /* 240 * For I/O queues, use the controller-wide max_xfer_size 241 * calculated in nvme_attach(). 242 */ 243 error = nvme_qpair_construct(qpair, num_entries, num_trackers, 244 ctrlr); 245 if (error) 246 return (error); 247 248 /* 249 * Do not bother binding interrupts if we only have one I/O 250 * interrupt thread for this controller. 251 */ 252 if (ctrlr->num_io_queues > 1) 253 bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu); 254 } 255 256 return (0); 257 } 258 259 static void 260 nvme_ctrlr_fail(struct nvme_controller *ctrlr, bool admin_also) 261 { 262 int i; 263 264 /* 265 * No need to disable queues before failing them. Failing is a superet 266 * of disabling (though pedantically we'd abort the AERs silently with 267 * a different error, though when we fail, that hardly matters). 268 */ 269 ctrlr->is_failed = true; 270 if (admin_also) { 271 ctrlr->is_failed_admin = true; 272 nvme_qpair_fail(&ctrlr->adminq); 273 } 274 if (ctrlr->ioq != NULL) { 275 for (i = 0; i < ctrlr->num_io_queues; i++) { 276 nvme_qpair_fail(&ctrlr->ioq[i]); 277 } 278 } 279 nvme_notify_fail(ctrlr); 280 } 281 282 /* 283 * Wait for RDY to change. 284 * 285 * Starts sleeping for 1us and geometrically increases it the longer we wait, 286 * capped at 1ms. 287 */ 288 static int 289 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val) 290 { 291 int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms); 292 sbintime_t delta_t = SBT_1US; 293 uint32_t csts; 294 295 while (1) { 296 csts = nvme_mmio_read_4(ctrlr, csts); 297 if (csts == NVME_GONE) /* Hot unplug. */ 298 return (ENXIO); 299 if (NVMEV(NVME_CSTS_REG_RDY, csts) == desired_val) 300 break; 301 if (timeout - ticks < 0) { 302 nvme_printf(ctrlr, "controller ready did not become %d " 303 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms); 304 return (ENXIO); 305 } 306 307 pause_sbt("nvmerdy", delta_t, 0, C_PREL(1)); 308 delta_t = min(SBT_1MS, delta_t * 3 / 2); 309 } 310 311 return (0); 312 } 313 314 static int 315 nvme_ctrlr_disable(struct nvme_controller *ctrlr) 316 { 317 uint32_t cc; 318 uint32_t csts; 319 uint8_t en, rdy; 320 int err; 321 322 cc = nvme_mmio_read_4(ctrlr, cc); 323 csts = nvme_mmio_read_4(ctrlr, csts); 324 325 en = NVMEV(NVME_CC_REG_EN, cc); 326 rdy = NVMEV(NVME_CSTS_REG_RDY, csts); 327 328 /* 329 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1 330 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when 331 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY 332 * isn't the desired value. Short circuit if we're already disabled. 333 */ 334 if (en == 0) { 335 /* Wait for RDY == 0 or timeout & fail */ 336 if (rdy == 0) 337 return (0); 338 return (nvme_ctrlr_wait_for_ready(ctrlr, 0)); 339 } 340 if (rdy == 0) { 341 /* EN == 1, wait for RDY == 1 or timeout & fail */ 342 err = nvme_ctrlr_wait_for_ready(ctrlr, 1); 343 if (err != 0) 344 return (err); 345 } 346 347 cc &= ~NVMEM(NVME_CC_REG_EN); 348 nvme_mmio_write_4(ctrlr, cc, cc); 349 350 /* 351 * A few drives have firmware bugs that freeze the drive if we access 352 * the mmio too soon after we disable. 353 */ 354 if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY) 355 pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS)); 356 return (nvme_ctrlr_wait_for_ready(ctrlr, 0)); 357 } 358 359 static int 360 nvme_ctrlr_enable(struct nvme_controller *ctrlr) 361 { 362 uint32_t cc; 363 uint32_t csts; 364 uint32_t aqa; 365 uint32_t qsize; 366 uint8_t css, en, rdy; 367 int err; 368 369 cc = nvme_mmio_read_4(ctrlr, cc); 370 csts = nvme_mmio_read_4(ctrlr, csts); 371 372 en = NVMEV(NVME_CC_REG_EN, cc); 373 rdy = NVMEV(NVME_CSTS_REG_RDY, csts); 374 375 /* 376 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled. 377 */ 378 if (en == 1) { 379 if (rdy == 1) 380 return (0); 381 return (nvme_ctrlr_wait_for_ready(ctrlr, 1)); 382 } 383 384 /* EN == 0 already wait for RDY == 0 or timeout & fail */ 385 err = nvme_ctrlr_wait_for_ready(ctrlr, 0); 386 if (err != 0) 387 return (err); 388 389 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr); 390 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr); 391 392 /* acqs and asqs are 0-based. */ 393 qsize = ctrlr->adminq.num_entries - 1; 394 395 aqa = 0; 396 aqa |= NVMEF(NVME_AQA_REG_ACQS, qsize); 397 aqa |= NVMEF(NVME_AQA_REG_ASQS, qsize); 398 nvme_mmio_write_4(ctrlr, aqa, aqa); 399 400 /* Initialization values for CC */ 401 cc = 0; 402 cc |= NVMEF(NVME_CC_REG_EN, 1); 403 /* No CSI support; prefer the NVM command set when present. */ 404 css = NVME_CAP_HI_CSS(ctrlr->cap_hi); 405 if ((css & NVME_CAP_CSS_NVM) != 0) 406 cc |= NVMEF(NVME_CC_REG_CSS, NVME_CC_CSS_NVM); 407 else if ((css & NVME_CAP_CSS_NOIOCSS) != 0) 408 cc |= NVMEF(NVME_CC_REG_CSS, NVME_CC_CSS_ADMIN); 409 else if ((css & NVME_CAP_CSS_IOCSS) != 0) 410 cc |= NVMEF(NVME_CC_REG_CSS, NVME_CC_CSS_IOCSS); 411 else 412 cc |= NVMEF(NVME_CC_REG_CSS, NVME_CC_CSS_NVM); 413 cc |= NVMEF(NVME_CC_REG_AMS, 0); 414 cc |= NVMEF(NVME_CC_REG_SHN, 0); 415 cc |= NVMEF(NVME_CC_REG_IOSQES, ctrlr->io_sqes); 416 cc |= NVMEF(NVME_CC_REG_IOCQES, 4); /* CQ entry size == 16 == 2^4 */ 417 418 /* 419 * Use the Memory Page Size selected during device initialization. Note 420 * that value stored in mps is suitable to use here without adjusting by 421 * NVME_MPS_SHIFT. 422 */ 423 cc |= NVMEF(NVME_CC_REG_MPS, ctrlr->mps); 424 425 nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE); 426 nvme_mmio_write_4(ctrlr, cc, cc); 427 428 return (nvme_ctrlr_wait_for_ready(ctrlr, 1)); 429 } 430 431 static void 432 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr) 433 { 434 int i; 435 436 nvme_admin_qpair_disable(&ctrlr->adminq); 437 /* 438 * I/O queues are not allocated before the initial HW 439 * reset, so do not try to disable them. Use is_initialized 440 * to determine if this is the initial HW reset. 441 */ 442 if (ctrlr->is_initialized) { 443 for (i = 0; i < ctrlr->num_io_queues; i++) 444 nvme_io_qpair_disable(&ctrlr->ioq[i]); 445 } 446 } 447 448 static int 449 nvme_ctrlr_pcie_flr(struct nvme_controller *ctrlr, uint32_t csts) 450 { 451 nvme_printf(ctrlr, 452 "fatal status; attempting PCIe function level reset\n"); 453 pci_save_state(ctrlr->dev); 454 if (!pcie_flr(ctrlr->dev, 1000, true)) { 455 pci_restore_state(ctrlr->dev); 456 nvme_printf(ctrlr, "PCIe function level reset failed\n"); 457 nvme_ctrlr_devctl(ctrlr, "FLR_FAILED", "csts=0x%08x", csts); 458 return (ENXIO); 459 } 460 pci_restore_state(ctrlr->dev); 461 nvme_printf(ctrlr, "PCIe function level reset completed\n"); 462 nvme_ctrlr_devctl(ctrlr, "FLR_COMPLETED", "csts=0x%08x", csts); 463 return (0); 464 } 465 466 static int 467 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr) 468 { 469 uint32_t csts; 470 int err; 471 472 TSENTER(); 473 474 ctrlr->is_failed_admin = true; 475 nvme_ctrlr_disable_qpairs(ctrlr); 476 477 csts = nvme_mmio_read_4(ctrlr, csts); 478 if ((ctrlr->quirks & QUIRK_PCIE_FLR_ON_FATAL) != 0 && 479 csts != NVME_GONE && NVMEV(NVME_CSTS_REG_CFS, csts) != 0) { 480 err = nvme_ctrlr_pcie_flr(ctrlr, csts); 481 if (err != 0) 482 goto out; 483 } 484 485 err = nvme_ctrlr_disable(ctrlr); 486 if (err != 0) 487 goto out; 488 489 err = nvme_ctrlr_enable(ctrlr); 490 out: 491 if (err == 0) 492 ctrlr->is_failed_admin = false; 493 494 TSEXIT(); 495 return (err); 496 } 497 498 void 499 nvme_ctrlr_reset(struct nvme_controller *ctrlr) 500 { 501 int cmpset; 502 503 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1); 504 505 if (cmpset == 0) 506 /* 507 * Controller is already resetting. Return immediately since 508 * there is no need to kick off another reset. 509 */ 510 return; 511 512 if (!ctrlr->is_dying) 513 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task); 514 } 515 516 static int 517 nvme_ctrlr_identify(struct nvme_controller *ctrlr) 518 { 519 struct nvme_completion_poll_status status; 520 521 status.done = 0; 522 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata, 523 nvme_completion_poll_cb, &status); 524 nvme_completion_poll(&status); 525 if (nvme_completion_is_error(&status.cpl)) { 526 nvme_printf(ctrlr, "nvme_identify_controller failed!\n"); 527 return (ENXIO); 528 } 529 530 /* Convert data to host endian */ 531 nvme_controller_data_swapbytes(&ctrlr->cdata); 532 533 /* 534 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the 535 * controller supports. 536 */ 537 if (ctrlr->cdata.mdts > 0) 538 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size, 539 1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT + 540 NVME_CAP_HI_MPSMIN(ctrlr->cap_hi))); 541 if (ctrlr->quirks & QUIRK_APPLE_S3X_SERIALIZE) 542 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size, 8192U); 543 544 return (0); 545 } 546 547 static int 548 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr) 549 { 550 struct nvme_completion_poll_status status; 551 int cq_allocated, sq_allocated; 552 553 status.done = 0; 554 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues, 555 nvme_completion_poll_cb, &status); 556 nvme_completion_poll(&status); 557 if (nvme_completion_is_error(&status.cpl)) { 558 nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n"); 559 return (ENXIO); 560 } 561 562 /* 563 * Data in cdw0 is 0-based. 564 * Lower 16-bits indicate number of submission queues allocated. 565 * Upper 16-bits indicate number of completion queues allocated. 566 */ 567 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1; 568 cq_allocated = (status.cpl.cdw0 >> 16) + 1; 569 570 /* 571 * Controller may allocate more queues than we requested, 572 * so use the minimum of the number requested and what was 573 * actually allocated. 574 */ 575 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated); 576 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated); 577 if (ctrlr->num_io_queues > vm_ndomains) 578 ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains; 579 580 return (0); 581 } 582 583 static int 584 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr) 585 { 586 struct nvme_completion_poll_status status; 587 struct nvme_qpair *qpair; 588 int i; 589 590 for (i = 0; i < ctrlr->num_io_queues; i++) { 591 qpair = &ctrlr->ioq[i]; 592 593 status.done = 0; 594 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, 595 nvme_completion_poll_cb, &status); 596 nvme_completion_poll(&status); 597 if (nvme_completion_is_error(&status.cpl)) { 598 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n"); 599 return (ENXIO); 600 } 601 602 status.done = 0; 603 nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair, 604 nvme_completion_poll_cb, &status); 605 nvme_completion_poll(&status); 606 if (nvme_completion_is_error(&status.cpl)) { 607 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n"); 608 return (ENXIO); 609 } 610 } 611 612 return (0); 613 } 614 615 static int 616 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr) 617 { 618 struct nvme_completion_poll_status status; 619 struct nvme_qpair *qpair; 620 621 for (int i = 0; i < ctrlr->num_io_queues; i++) { 622 qpair = &ctrlr->ioq[i]; 623 624 status.done = 0; 625 nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair, 626 nvme_completion_poll_cb, &status); 627 nvme_completion_poll(&status); 628 if (nvme_completion_is_error(&status.cpl)) { 629 nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n"); 630 return (ENXIO); 631 } 632 633 status.done = 0; 634 nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair, 635 nvme_completion_poll_cb, &status); 636 nvme_completion_poll(&status); 637 if (nvme_completion_is_error(&status.cpl)) { 638 nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n"); 639 return (ENXIO); 640 } 641 } 642 643 return (0); 644 } 645 646 static int 647 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr) 648 { 649 struct nvme_namespace *ns; 650 uint32_t i; 651 652 for (i = 0; i < nvme_ctrlr_num_namespaces(ctrlr); i++) { 653 ns = &ctrlr->ns[i]; 654 nvme_ns_construct(ns, i+1, ctrlr); 655 } 656 657 return (0); 658 } 659 660 static bool 661 is_log_page_id_valid(uint8_t page_id) 662 { 663 switch (page_id) { 664 case NVME_LOG_ERROR: 665 case NVME_LOG_HEALTH_INFORMATION: 666 case NVME_LOG_FIRMWARE_SLOT: 667 case NVME_LOG_CHANGED_NAMESPACE: 668 case NVME_LOG_COMMAND_EFFECT: 669 case NVME_LOG_RES_NOTIFICATION: 670 case NVME_LOG_SANITIZE_STATUS: 671 return (true); 672 } 673 674 return (false); 675 } 676 677 static uint32_t 678 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id) 679 { 680 uint32_t log_page_size; 681 682 switch (page_id) { 683 case NVME_LOG_ERROR: 684 log_page_size = min( 685 sizeof(struct nvme_error_information_entry) * 686 (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE); 687 break; 688 case NVME_LOG_HEALTH_INFORMATION: 689 log_page_size = sizeof(struct nvme_health_information_page); 690 break; 691 case NVME_LOG_FIRMWARE_SLOT: 692 log_page_size = sizeof(struct nvme_firmware_page); 693 break; 694 case NVME_LOG_CHANGED_NAMESPACE: 695 log_page_size = sizeof(struct nvme_ns_list); 696 break; 697 case NVME_LOG_COMMAND_EFFECT: 698 log_page_size = sizeof(struct nvme_command_effects_page); 699 break; 700 case NVME_LOG_RES_NOTIFICATION: 701 log_page_size = sizeof(struct nvme_res_notification_page); 702 break; 703 case NVME_LOG_SANITIZE_STATUS: 704 log_page_size = sizeof(struct nvme_sanitize_status_page); 705 break; 706 default: 707 log_page_size = 0; 708 break; 709 } 710 711 return (log_page_size); 712 } 713 714 static void 715 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr, 716 uint8_t state) 717 { 718 if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE) 719 nvme_printf(ctrlr, "SMART WARNING: available spare space below threshold\n"); 720 721 if (state & NVME_CRIT_WARN_ST_TEMPERATURE) 722 nvme_printf(ctrlr, "SMART WARNING: temperature above threshold\n"); 723 724 if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY) 725 nvme_printf(ctrlr, "SMART WARNING: device reliability degraded\n"); 726 727 if (state & NVME_CRIT_WARN_ST_READ_ONLY) 728 nvme_printf(ctrlr, "SMART WARNING: media placed in read only mode\n"); 729 730 if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP) 731 nvme_printf(ctrlr, "SMART WARNING: volatile memory backup device failed\n"); 732 733 if (state & NVME_CRIT_WARN_ST_PERSISTENT_MEMORY_REGION) 734 nvme_printf(ctrlr, "SMART WARNING: persistent memory read only or unreliable\n"); 735 736 if (state & NVME_CRIT_WARN_ST_RESERVED_MASK) 737 nvme_printf(ctrlr, "SMART WARNING: unknown critical warning(s): state = 0x%02x\n", 738 state & NVME_CRIT_WARN_ST_RESERVED_MASK); 739 740 nvme_ctrlr_devctl(ctrlr, "SMART_ERROR", "state=0x%02x", state); 741 } 742 743 static void 744 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl) 745 { 746 struct nvme_async_event_request *aer = arg; 747 748 if (nvme_completion_is_error(cpl)) { 749 /* 750 * Do not retry failed async event requests. This avoids 751 * infinite loops where a new async event request is submitted 752 * to replace the one just failed, only to fail again and 753 * perpetuate the loop. 754 */ 755 return; 756 } 757 758 /* 759 * Save the completion status and associated log page is in bits 23:16 760 * of completion entry dw0. Print a message and queue it for further 761 * processing. 762 */ 763 memcpy(&aer->cpl, cpl, sizeof(*cpl)); 764 aer->log_page_id = NVMEV(NVME_ASYNC_EVENT_LOG_PAGE_ID, cpl->cdw0); 765 nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x," 766 " page 0x%02x)\n", NVMEV(NVME_ASYNC_EVENT_TYPE, cpl->cdw0), 767 NVMEV(NVME_ASYNC_EVENT_INFO, cpl->cdw0), 768 aer->log_page_id); 769 taskqueue_enqueue(aer->ctrlr->taskqueue, &aer->task); 770 } 771 772 static void 773 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr, 774 struct nvme_async_event_request *aer) 775 { 776 struct nvme_request *req; 777 778 /* 779 * We're racing the reset thread, so let that process submit this again. 780 * XXX does this really solve that race? And is that race even possible 781 * since we only reset when we've no theard from the card in a long 782 * time. Why would we get an AER in the middle of that just before we 783 * kick off the reset? 784 */ 785 if (ctrlr->is_resetting) 786 return; 787 788 aer->ctrlr = ctrlr; 789 req = nvme_allocate_request_null(M_WAITOK, nvme_ctrlr_async_event_cb, 790 aer); 791 aer->req = req; 792 aer->log_page_id = 0; /* Not a valid page */ 793 794 /* 795 * Disable timeout here, since asynchronous event requests should by 796 * nature never be timed out. 797 */ 798 req->timeout = false; 799 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST; 800 nvme_ctrlr_submit_admin_request(ctrlr, req); 801 } 802 803 static void 804 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr) 805 { 806 struct nvme_completion_poll_status status; 807 struct nvme_async_event_request *aer; 808 uint32_t i; 809 810 if (ctrlr->quirks & QUIRK_APPLE_NO_ASYNC_EVENT) { 811 ctrlr->num_aers = 0; 812 return; 813 } 814 815 ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE | 816 NVME_CRIT_WARN_ST_DEVICE_RELIABILITY | 817 NVME_CRIT_WARN_ST_READ_ONLY | 818 NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP; 819 if (ctrlr->cdata.ver >= NVME_REV(1, 2)) 820 ctrlr->async_event_config |= 821 ctrlr->cdata.oaes & (NVME_ASYNC_EVENT_NS_ATTRIBUTE | 822 NVME_ASYNC_EVENT_FW_ACTIVATE); 823 824 status.done = 0; 825 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD, 826 0, NULL, 0, nvme_completion_poll_cb, &status); 827 nvme_completion_poll(&status); 828 if (nvme_completion_is_error(&status.cpl) || 829 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF || 830 (status.cpl.cdw0 & 0xFFFF) == 0x0000) { 831 nvme_printf(ctrlr, "temperature threshold not supported\n"); 832 } else 833 ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE; 834 835 nvme_ctrlr_cmd_set_async_event_config(ctrlr, 836 ctrlr->async_event_config, NULL, NULL); 837 838 /* aerl is a zero-based value, so we need to add 1 here. */ 839 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1)); 840 841 for (i = 0; i < ctrlr->num_aers; i++) { 842 aer = &ctrlr->aer[i]; 843 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer); 844 } 845 } 846 847 static void 848 nvme_ctrlr_configure_apst(struct nvme_controller *ctrlr) 849 { 850 struct nvme_completion_poll_status status; 851 uint64_t *data; 852 int data_size, i, read_size; 853 bool enable, error = true; 854 855 if (TUNABLE_BOOL_FETCH("hw.nvme.apst_enable", &enable) == 0 || 856 ctrlr->cdata.apsta == 0) 857 return; 858 859 data_size = 32 * sizeof(*data); 860 data = malloc(data_size, M_NVME, M_WAITOK | M_ZERO); 861 862 if (getenv_array("hw.nvme.apst_data", data, data_size, 863 &read_size, sizeof(*data), GETENV_UNSIGNED) != 0) { 864 for (i = 0; i < read_size / sizeof(*data); ++i) 865 data[i] = htole64(data[i]); 866 } else { 867 status.done = 0; 868 nvme_ctrlr_cmd_get_feature(ctrlr, 869 NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION, 0, 870 data, data_size, nvme_completion_poll_cb, &status); 871 nvme_completion_poll(&status); 872 if (nvme_completion_is_error(&status.cpl)) 873 goto out; 874 } 875 876 status.done = 0; 877 nvme_ctrlr_cmd_set_feature(ctrlr, 878 NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION, enable, 0, 0, 879 0, 0, data, data_size, nvme_completion_poll_cb, &status); 880 nvme_completion_poll(&status); 881 error = nvme_completion_is_error(&status.cpl); 882 out: 883 if (error && bootverbose) 884 nvme_printf(ctrlr, "failed to configure APST\n"); 885 free(data, M_NVME); 886 } 887 888 static void 889 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr) 890 { 891 ctrlr->int_coal_time = 0; 892 TUNABLE_INT_FETCH("hw.nvme.int_coal_time", 893 &ctrlr->int_coal_time); 894 895 ctrlr->int_coal_threshold = 0; 896 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold", 897 &ctrlr->int_coal_threshold); 898 899 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time, 900 ctrlr->int_coal_threshold, NULL, NULL); 901 } 902 903 static void 904 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr) 905 { 906 struct nvme_hmb_chunk *hmbc; 907 int i; 908 909 if (ctrlr->hmb_desc_paddr) { 910 bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map); 911 bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr, 912 ctrlr->hmb_desc_map); 913 ctrlr->hmb_desc_paddr = 0; 914 } 915 if (ctrlr->hmb_desc_tag) { 916 bus_dma_tag_destroy(ctrlr->hmb_desc_tag); 917 ctrlr->hmb_desc_tag = NULL; 918 } 919 for (i = 0; i < ctrlr->hmb_nchunks; i++) { 920 hmbc = &ctrlr->hmb_chunks[i]; 921 bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map); 922 bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr, 923 hmbc->hmbc_map); 924 } 925 ctrlr->hmb_nchunks = 0; 926 if (ctrlr->hmb_tag) { 927 bus_dma_tag_destroy(ctrlr->hmb_tag); 928 ctrlr->hmb_tag = NULL; 929 } 930 if (ctrlr->hmb_chunks) { 931 free(ctrlr->hmb_chunks, M_NVME); 932 ctrlr->hmb_chunks = NULL; 933 } 934 } 935 936 static void 937 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr) 938 { 939 struct nvme_hmb_chunk *hmbc; 940 size_t pref, min, minc, size; 941 int err, i; 942 uint64_t max; 943 944 /* Limit HMB to 5% of RAM size per device by default. */ 945 max = (uint64_t)physmem * PAGE_SIZE / 20; 946 TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max); 947 948 /* 949 * Units of Host Memory Buffer in the Identify info are always in terms 950 * of 4k units. 951 */ 952 min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS; 953 if (max == 0 || max < min) 954 return; 955 pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max); 956 minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size); 957 if (min > 0 && ctrlr->cdata.hmmaxd > 0) 958 minc = MAX(minc, min / ctrlr->cdata.hmmaxd); 959 ctrlr->hmb_chunk = pref; 960 961 again: 962 /* 963 * However, the chunk sizes, number of chunks, and alignment of chunks 964 * are all based on the current MPS (ctrlr->page_size). 965 */ 966 ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size); 967 ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk); 968 if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd) 969 ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd; 970 ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) * 971 ctrlr->hmb_nchunks, M_NVME, M_WAITOK); 972 err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev), 973 ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, 974 ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag); 975 if (err != 0) { 976 nvme_printf(ctrlr, "HMB tag create failed %d\n", err); 977 nvme_ctrlr_hmb_free(ctrlr); 978 return; 979 } 980 981 for (i = 0; i < ctrlr->hmb_nchunks; i++) { 982 hmbc = &ctrlr->hmb_chunks[i]; 983 if (bus_dmamem_alloc(ctrlr->hmb_tag, 984 (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT, 985 &hmbc->hmbc_map)) { 986 nvme_printf(ctrlr, "failed to alloc HMB\n"); 987 break; 988 } 989 if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map, 990 hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map, 991 &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) { 992 bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr, 993 hmbc->hmbc_map); 994 nvme_printf(ctrlr, "failed to load HMB\n"); 995 break; 996 } 997 bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map, 998 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 999 } 1000 1001 if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min && 1002 ctrlr->hmb_chunk / 2 >= minc) { 1003 ctrlr->hmb_nchunks = i; 1004 nvme_ctrlr_hmb_free(ctrlr); 1005 ctrlr->hmb_chunk /= 2; 1006 goto again; 1007 } 1008 ctrlr->hmb_nchunks = i; 1009 if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) { 1010 nvme_ctrlr_hmb_free(ctrlr); 1011 return; 1012 } 1013 1014 size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks; 1015 err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev), 1016 PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, 1017 size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag); 1018 if (err != 0) { 1019 nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err); 1020 nvme_ctrlr_hmb_free(ctrlr); 1021 return; 1022 } 1023 if (bus_dmamem_alloc(ctrlr->hmb_desc_tag, 1024 (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK, 1025 &ctrlr->hmb_desc_map)) { 1026 nvme_printf(ctrlr, "failed to alloc HMB desc\n"); 1027 nvme_ctrlr_hmb_free(ctrlr); 1028 return; 1029 } 1030 if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map, 1031 ctrlr->hmb_desc_vaddr, size, nvme_single_map, 1032 &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) { 1033 bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr, 1034 ctrlr->hmb_desc_map); 1035 nvme_printf(ctrlr, "failed to load HMB desc\n"); 1036 nvme_ctrlr_hmb_free(ctrlr); 1037 return; 1038 } 1039 1040 for (i = 0; i < ctrlr->hmb_nchunks; i++) { 1041 memset(&ctrlr->hmb_desc_vaddr[i], 0, 1042 sizeof(struct nvme_hmb_desc)); 1043 ctrlr->hmb_desc_vaddr[i].addr = 1044 htole64(ctrlr->hmb_chunks[i].hmbc_paddr); 1045 ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size); 1046 } 1047 bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map, 1048 BUS_DMASYNC_PREWRITE); 1049 1050 nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n", 1051 (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk 1052 / 1024 / 1024); 1053 } 1054 1055 static void 1056 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret) 1057 { 1058 struct nvme_completion_poll_status status; 1059 uint32_t cdw11; 1060 1061 cdw11 = 0; 1062 if (enable) 1063 cdw11 |= 1; 1064 if (memret) 1065 cdw11 |= 2; 1066 status.done = 0; 1067 nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11, 1068 ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size, 1069 ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32, 1070 ctrlr->hmb_nchunks, NULL, 0, 1071 nvme_completion_poll_cb, &status); 1072 nvme_completion_poll(&status); 1073 if (nvme_completion_is_error(&status.cpl)) 1074 nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n"); 1075 } 1076 1077 static void 1078 nvme_ctrlr_start(void *ctrlr_arg, bool resetting) 1079 { 1080 struct nvme_controller *ctrlr = ctrlr_arg; 1081 uint32_t old_num_io_queues; 1082 int i; 1083 1084 TSENTER(); 1085 1086 /* 1087 * Only reset adminq here when we are restarting the 1088 * controller after a reset. During initialization, 1089 * we have already submitted admin commands to get 1090 * the number of I/O queues supported, so cannot reset 1091 * the adminq again here. 1092 */ 1093 if (resetting) { 1094 nvme_qpair_reset(&ctrlr->adminq); 1095 nvme_admin_qpair_enable(&ctrlr->adminq); 1096 } 1097 1098 if (ctrlr->ioq != NULL) { 1099 for (i = 0; i < ctrlr->num_io_queues; i++) 1100 nvme_qpair_reset(&ctrlr->ioq[i]); 1101 } 1102 1103 /* 1104 * If it was a reset on initialization command timeout, just 1105 * return here, letting initialization code fail gracefully. 1106 */ 1107 if (resetting && !ctrlr->is_initialized) 1108 return; 1109 1110 if (resetting && nvme_ctrlr_identify(ctrlr) != 0) { 1111 nvme_ctrlr_fail(ctrlr, false); 1112 return; 1113 } 1114 1115 /* 1116 * The number of qpairs are determined during controller initialization, 1117 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the 1118 * HW limit. We call SET_FEATURES again here so that it gets called 1119 * after any reset for controllers that depend on the driver to 1120 * explicit specify how many queues it will use. This value should 1121 * never change between resets, so panic if somehow that does happen. 1122 */ 1123 if (resetting) { 1124 old_num_io_queues = ctrlr->num_io_queues; 1125 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) { 1126 nvme_ctrlr_fail(ctrlr, false); 1127 return; 1128 } 1129 1130 if (old_num_io_queues != ctrlr->num_io_queues) { 1131 panic("num_io_queues changed from %u to %u", 1132 old_num_io_queues, ctrlr->num_io_queues); 1133 } 1134 } 1135 1136 if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) { 1137 nvme_ctrlr_hmb_alloc(ctrlr); 1138 if (ctrlr->hmb_nchunks > 0) 1139 nvme_ctrlr_hmb_enable(ctrlr, true, false); 1140 } else if (ctrlr->hmb_nchunks > 0) 1141 nvme_ctrlr_hmb_enable(ctrlr, true, true); 1142 1143 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) { 1144 nvme_ctrlr_fail(ctrlr, false); 1145 return; 1146 } 1147 1148 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) { 1149 nvme_ctrlr_fail(ctrlr, false); 1150 return; 1151 } 1152 1153 nvme_ctrlr_configure_aer(ctrlr); 1154 nvme_ctrlr_configure_apst(ctrlr); 1155 nvme_ctrlr_configure_int_coalescing(ctrlr); 1156 1157 for (i = 0; i < ctrlr->num_io_queues; i++) 1158 nvme_io_qpair_enable(&ctrlr->ioq[i]); 1159 TSEXIT(); 1160 } 1161 1162 void 1163 nvme_ctrlr_start_config_hook(void *arg) 1164 { 1165 struct nvme_controller *ctrlr = arg; 1166 1167 TSENTER(); 1168 1169 if (nvme_ctrlr_hw_reset(ctrlr) != 0 || ctrlr->fail_on_reset != 0) { 1170 nvme_ctrlr_fail(ctrlr, true); 1171 config_intrhook_disestablish(&ctrlr->config_hook); 1172 return; 1173 } 1174 1175 nvme_qpair_reset(&ctrlr->adminq); 1176 nvme_admin_qpair_enable(&ctrlr->adminq); 1177 1178 if (nvme_ctrlr_identify(ctrlr) == 0 && 1179 nvme_ctrlr_set_num_qpairs(ctrlr) == 0 && 1180 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0) 1181 nvme_ctrlr_start(ctrlr, false); 1182 else 1183 nvme_ctrlr_fail(ctrlr, false); 1184 1185 nvme_sysctl_initialize_ctrlr(ctrlr); 1186 config_intrhook_disestablish(&ctrlr->config_hook); 1187 1188 if (!ctrlr->is_failed) { 1189 device_t child; 1190 1191 if (bootverbose && 1192 (ctrlr->quirks & QUIRK_APPLE_S3X_NS1_ONLY) != 0 && 1193 ctrlr->cdata.nn > nvme_ctrlr_num_namespaces(ctrlr)) 1194 nvme_printf(ctrlr, 1195 "ignoring Apple-internal namespaces above NSID 1\n"); 1196 1197 ctrlr->is_initialized = true; 1198 child = device_add_child(ctrlr->dev, NULL, DEVICE_UNIT_ANY); 1199 device_set_ivars(child, ctrlr); 1200 bus_attach_children(ctrlr->dev); 1201 1202 /* 1203 * Now notify the child of all the known namepsaces 1204 */ 1205 for (int i = 0; i < nvme_ctrlr_num_namespaces(ctrlr); i++) { 1206 struct nvme_namespace *ns = &ctrlr->ns[i]; 1207 1208 if (ns->data.nsze == 0) 1209 continue; 1210 NVME_NS_ADDED(child, ns); 1211 } 1212 } 1213 TSEXIT(); 1214 } 1215 1216 static void 1217 nvme_ctrlr_reset_task(void *arg, int pending) 1218 { 1219 struct nvme_controller *ctrlr = arg; 1220 int status; 1221 1222 nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"start\""); 1223 status = nvme_ctrlr_hw_reset(ctrlr); 1224 if (status == 0) { 1225 nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"success\""); 1226 nvme_ctrlr_start(ctrlr, true); 1227 } else { 1228 nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"timed_out\""); 1229 nvme_ctrlr_fail(ctrlr, true); 1230 } 1231 1232 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0); 1233 } 1234 1235 static void 1236 nvme_ctrlr_aer_done(void *arg, const struct nvme_completion *cpl) 1237 { 1238 struct nvme_async_event_request *aer = arg; 1239 1240 mtx_lock(&aer->mtx); 1241 if (nvme_completion_is_error(cpl)) 1242 aer->log_page_size = (uint32_t)-1; 1243 else 1244 aer->log_page_size = nvme_ctrlr_get_log_page_size( 1245 aer->ctrlr, aer->log_page_id); 1246 wakeup(aer); 1247 mtx_unlock(&aer->mtx); 1248 } 1249 1250 static void 1251 nvme_ctrlr_aer_task(void *arg, int pending) 1252 { 1253 struct nvme_async_event_request *aer = arg; 1254 struct nvme_controller *ctrlr = aer->ctrlr; 1255 uint32_t len; 1256 1257 /* 1258 * We're resetting, so just punt. 1259 */ 1260 if (ctrlr->is_resetting) 1261 return; 1262 1263 if (!is_log_page_id_valid(aer->log_page_id)) { 1264 /* 1265 * Repost another asynchronous event request to replace the one 1266 * that just completed. 1267 */ 1268 nvme_notify_async(ctrlr, &aer->cpl, aer->log_page_id, NULL, 0); 1269 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer); 1270 goto out; 1271 } 1272 1273 nvme_ctrlr_devctl(ctrlr, "aen", "type=0x%x info=0x%x page=0x%x", 1274 NVMEV(NVME_ASYNC_EVENT_TYPE, aer->cpl.cdw0), 1275 NVMEV(NVME_ASYNC_EVENT_INFO, aer->cpl.cdw0), aer->log_page_id); 1276 1277 aer->log_page_size = 0; 1278 len = nvme_ctrlr_get_log_page_size(aer->ctrlr, aer->log_page_id); 1279 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id, 1280 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer, len, 1281 nvme_ctrlr_aer_done, aer); 1282 mtx_lock(&aer->mtx); 1283 while (aer->log_page_size == 0) 1284 mtx_sleep(aer, &aer->mtx, PRIBIO, "nvme_pt", 0); 1285 mtx_unlock(&aer->mtx); 1286 1287 if (aer->log_page_size == (uint32_t)-1) { 1288 /* 1289 * If the log page fetch for some reason completed with an 1290 * error, don't pass log page data to the consumers. In 1291 * practice, this case should never happen. 1292 */ 1293 nvme_notify_async(aer->ctrlr, &aer->cpl, aer->log_page_id, 1294 NULL, 0); 1295 goto out; 1296 } 1297 1298 /* Convert data to host endian */ 1299 switch (aer->log_page_id) { 1300 case NVME_LOG_ERROR: { 1301 struct nvme_error_information_entry *err = 1302 (struct nvme_error_information_entry *)aer->log_page_buffer; 1303 for (uint32_t i = 0; i < aer->log_page_size / sizeof(*err); i++) 1304 nvme_error_information_entry_swapbytes(err++); 1305 break; 1306 } 1307 case NVME_LOG_HEALTH_INFORMATION: 1308 nvme_health_information_page_swapbytes( 1309 (struct nvme_health_information_page *)aer->log_page_buffer); 1310 break; 1311 case NVME_LOG_CHANGED_NAMESPACE: 1312 nvme_ns_list_swapbytes( 1313 (struct nvme_ns_list *)aer->log_page_buffer); 1314 break; 1315 case NVME_LOG_COMMAND_EFFECT: 1316 nvme_command_effects_page_swapbytes( 1317 (struct nvme_command_effects_page *)aer->log_page_buffer); 1318 break; 1319 case NVME_LOG_RES_NOTIFICATION: 1320 nvme_res_notification_page_swapbytes( 1321 (struct nvme_res_notification_page *)aer->log_page_buffer); 1322 break; 1323 case NVME_LOG_SANITIZE_STATUS: 1324 nvme_sanitize_status_page_swapbytes( 1325 (struct nvme_sanitize_status_page *)aer->log_page_buffer); 1326 break; 1327 default: 1328 break; 1329 } 1330 1331 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) { 1332 struct nvme_health_information_page *health_info = 1333 (struct nvme_health_information_page *)aer->log_page_buffer; 1334 1335 /* 1336 * Critical warnings reported through the SMART/health log page 1337 * are persistent, so clear the associated bits in the async 1338 * event config so that we do not receive repeated notifications 1339 * for the same event. 1340 */ 1341 nvme_ctrlr_log_critical_warnings(aer->ctrlr, 1342 health_info->critical_warning); 1343 aer->ctrlr->async_event_config &= 1344 ~health_info->critical_warning; 1345 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr, 1346 aer->ctrlr->async_event_config, NULL, NULL); 1347 } else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE) { 1348 device_t *children; 1349 int n_children; 1350 struct nvme_ns_list *nsl; 1351 1352 if (device_get_children(aer->ctrlr->dev, &children, &n_children) != 0) { 1353 children = NULL; 1354 n_children = 0; 1355 } 1356 nsl = (struct nvme_ns_list *)aer->log_page_buffer; 1357 for (int i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) { 1358 if (!nvme_ctrlr_nsid_visible(ctrlr, nsl->ns[i])) 1359 continue; 1360 /* 1361 * I think we need to query the name space here and see 1362 * if it went away, arrived, or changed in size and call 1363 * the nuanced routine (after constructing or before 1364 * destructing the namespace). XXX needs more work XXX. 1365 */ 1366 for (int j = 0; j < n_children; j++) 1367 NVME_NS_CHANGED(children[j], nsl->ns[i]); 1368 } 1369 if (nsl->ns[0] == 0 && ctrlr->quirks & QUIRK_EMPTY_NAMESPACE_CHANGED_LOG) { 1370 for (int i = 0; i < nvme_ctrlr_num_namespaces(ctrlr); i++) 1371 for (int j = 0; j < n_children; j++) 1372 NVME_NS_CHANGED(children[j], i + 1); 1373 } 1374 free(children, M_TEMP); 1375 } 1376 1377 /* 1378 * Pass the cpl data from the original async event completion, not the 1379 * log page fetch. 1380 */ 1381 nvme_notify_async(aer->ctrlr, &aer->cpl, aer->log_page_id, 1382 aer->log_page_buffer, aer->log_page_size); 1383 1384 /* 1385 * Repost another asynchronous event request to replace the one 1386 * that just completed. 1387 */ 1388 out: 1389 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer); 1390 } 1391 1392 /* 1393 * Poll all the queues enabled on the device for completion. 1394 */ 1395 void 1396 nvme_ctrlr_poll(struct nvme_controller *ctrlr) 1397 { 1398 int i; 1399 1400 nvme_qpair_process_completions(&ctrlr->adminq); 1401 1402 for (i = 0; i < ctrlr->num_io_queues; i++) 1403 if (ctrlr->ioq && ctrlr->ioq[i].cpl) 1404 nvme_qpair_process_completions(&ctrlr->ioq[i]); 1405 } 1406 1407 /* 1408 * Poll the single-vector interrupt case: num_io_queues will be 1 and 1409 * there's only a single vector. While we're polling, we mask further 1410 * interrupts in the controller. 1411 */ 1412 void 1413 nvme_ctrlr_shared_handler(void *arg) 1414 { 1415 struct nvme_controller *ctrlr = arg; 1416 1417 /* INTMS/INTMC are undefined when configured for MSI-X. */ 1418 if (!ctrlr->is_msix) 1419 nvme_mmio_write_4(ctrlr, intms, 1); 1420 nvme_ctrlr_poll(ctrlr); 1421 if (!ctrlr->is_msix) 1422 nvme_mmio_write_4(ctrlr, intmc, 1); 1423 } 1424 1425 #define NVME_MAX_PAGES (int)(1024 / sizeof(vm_page_t)) 1426 1427 static int 1428 nvme_page_count(vm_offset_t start, size_t len) 1429 { 1430 return atop(round_page(start + len) - trunc_page(start)); 1431 } 1432 1433 static int 1434 nvme_user_ioctl_req(vm_offset_t addr, size_t len, bool is_read, 1435 vm_page_t **upages, int max_pages, int *npagesp, struct nvme_request **req, 1436 nvme_cb_fn_t cb_fn, void *cb_arg) 1437 { 1438 vm_prot_t prot = VM_PROT_READ; 1439 int err, npages; 1440 vm_page_t *upages_us; 1441 1442 upages_us = *upages; 1443 npages = nvme_page_count(addr, len); 1444 if (npages > atop(maxphys)) 1445 return (EINVAL); 1446 if (npages > max_pages) 1447 upages_us = malloc(npages * sizeof(vm_page_t), M_NVME, 1448 M_ZERO | M_WAITOK); 1449 1450 if (is_read) 1451 prot |= VM_PROT_WRITE; /* Device will write to host memory */ 1452 err = vm_fault_hold_pages(&curproc->p_vmspace->vm_map, 1453 addr, len, prot, upages_us, npages, npagesp); 1454 if (err != 0) { 1455 if (*upages != upages_us) 1456 free(upages_us, M_NVME); 1457 return (err); 1458 } 1459 *req = nvme_allocate_request_null(M_WAITOK, cb_fn, cb_arg); 1460 (*req)->payload = memdesc_vmpages(upages_us, len, addr & PAGE_MASK); 1461 (*req)->payload_valid = true; 1462 if (*upages != upages_us) 1463 *upages = upages_us; 1464 return (0); 1465 } 1466 1467 static void 1468 nvme_user_ioctl_free(vm_page_t *pages, int npage, bool freeit) 1469 { 1470 vm_page_unhold_pages(pages, npage); 1471 if (freeit) 1472 free(pages, M_NVME); 1473 } 1474 1475 static void 1476 nvme_pt_done(void *arg, const struct nvme_completion *cpl) 1477 { 1478 struct nvme_pt_command *pt = arg; 1479 struct mtx *mtx = pt->driver_lock; 1480 uint16_t status; 1481 1482 bzero(&pt->cpl, sizeof(pt->cpl)); 1483 pt->cpl.cdw0 = cpl->cdw0; 1484 1485 status = cpl->status; 1486 status &= ~NVMEM(NVME_STATUS_P); 1487 pt->cpl.status = status; 1488 1489 mtx_lock(mtx); 1490 pt->driver_lock = NULL; 1491 wakeup(pt); 1492 mtx_unlock(mtx); 1493 } 1494 1495 int 1496 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr, 1497 struct nvme_pt_command *pt, uint32_t nsid, int is_user, 1498 int is_admin_cmd) 1499 { 1500 struct nvme_request *req; 1501 struct mtx *mtx; 1502 int ret = 0; 1503 int npages = 0; 1504 vm_page_t upages_small[NVME_MAX_PAGES]; 1505 vm_page_t *upages = upages_small; 1506 1507 if (pt->len > 0) { 1508 if (pt->len > ctrlr->max_xfer_size) { 1509 nvme_printf(ctrlr, 1510 "len (%d) exceeds max_xfer_size (%d)\n", 1511 pt->len, ctrlr->max_xfer_size); 1512 return (EIO); 1513 } 1514 if (is_user) { 1515 ret = nvme_user_ioctl_req((vm_offset_t)pt->buf, pt->len, 1516 pt->is_read, &upages, nitems(upages_small), &npages, &req, 1517 nvme_pt_done, pt); 1518 if (ret != 0) 1519 return (ret); 1520 } else 1521 req = nvme_allocate_request_vaddr(pt->buf, pt->len, 1522 M_WAITOK, nvme_pt_done, pt); 1523 } else 1524 req = nvme_allocate_request_null(M_WAITOK, nvme_pt_done, pt); 1525 1526 /* Assume user space already converted to little-endian */ 1527 req->cmd.opc = pt->cmd.opc; 1528 req->cmd.fuse = pt->cmd.fuse; 1529 req->cmd.rsvd2 = pt->cmd.rsvd2; 1530 req->cmd.rsvd3 = pt->cmd.rsvd3; 1531 req->cmd.cdw10 = pt->cmd.cdw10; 1532 req->cmd.cdw11 = pt->cmd.cdw11; 1533 req->cmd.cdw12 = pt->cmd.cdw12; 1534 req->cmd.cdw13 = pt->cmd.cdw13; 1535 req->cmd.cdw14 = pt->cmd.cdw14; 1536 req->cmd.cdw15 = pt->cmd.cdw15; 1537 1538 req->cmd.nsid = htole32(nsid); 1539 1540 mtx = mtx_pool_find(mtxpool_sleep, pt); 1541 pt->driver_lock = mtx; 1542 1543 if (is_admin_cmd) 1544 nvme_ctrlr_submit_admin_request(ctrlr, req); 1545 else 1546 nvme_ctrlr_submit_io_request(ctrlr, req); 1547 1548 mtx_lock(mtx); 1549 while (pt->driver_lock != NULL) 1550 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0); 1551 mtx_unlock(mtx); 1552 1553 if (npages > 0) 1554 nvme_user_ioctl_free(upages, npages, upages != upages_small); 1555 1556 return (ret); 1557 } 1558 1559 static void 1560 nvme_npc_done(void *arg, const struct nvme_completion *cpl) 1561 { 1562 struct nvme_passthru_cmd *npc = arg; 1563 struct mtx *mtx = (void *)(uintptr_t)npc->metadata; 1564 1565 npc->result = cpl->cdw0; /* cpl in host order by now */ 1566 mtx_lock(mtx); 1567 npc->metadata = 0; 1568 wakeup(npc); 1569 mtx_unlock(mtx); 1570 } 1571 1572 /* XXX refactor? */ 1573 1574 int 1575 nvme_ctrlr_linux_passthru_cmd(struct nvme_controller *ctrlr, 1576 struct nvme_passthru_cmd *npc, uint32_t nsid, bool is_user, bool is_admin) 1577 { 1578 struct nvme_request *req; 1579 struct mtx *mtx; 1580 int ret = 0; 1581 int npages = 0; 1582 vm_page_t upages_small[NVME_MAX_PAGES]; 1583 vm_page_t *upages = upages_small; 1584 1585 /* 1586 * We don't support metadata. 1587 */ 1588 if (npc->metadata != 0 || npc->metadata_len != 0) 1589 return (EIO); 1590 1591 if (npc->data_len > 0 && npc->addr != 0) { 1592 if (npc->data_len > ctrlr->max_xfer_size) { 1593 nvme_printf(ctrlr, 1594 "data_len (%d) exceeds max_xfer_size (%d)\n", 1595 npc->data_len, ctrlr->max_xfer_size); 1596 return (EIO); 1597 } 1598 if (is_user) { 1599 ret = nvme_user_ioctl_req(npc->addr, npc->data_len, 1600 npc->opcode & 0x1, &upages, nitems(upages_small), 1601 &npages, &req, nvme_npc_done, npc); 1602 if (ret != 0) 1603 return (ret); 1604 } else 1605 req = nvme_allocate_request_vaddr( 1606 (void *)(uintptr_t)npc->addr, npc->data_len, 1607 M_WAITOK, nvme_npc_done, npc); 1608 } else 1609 req = nvme_allocate_request_null(M_WAITOK, nvme_npc_done, npc); 1610 1611 req->cmd.opc = npc->opcode; 1612 req->cmd.fuse = npc->flags; 1613 req->cmd.rsvd2 = htole32(npc->cdw2); 1614 req->cmd.rsvd3 = htole32(npc->cdw3); 1615 req->cmd.cdw10 = htole32(npc->cdw10); 1616 req->cmd.cdw11 = htole32(npc->cdw11); 1617 req->cmd.cdw12 = htole32(npc->cdw12); 1618 req->cmd.cdw13 = htole32(npc->cdw13); 1619 req->cmd.cdw14 = htole32(npc->cdw14); 1620 req->cmd.cdw15 = htole32(npc->cdw15); 1621 1622 req->cmd.nsid = htole32(nsid); 1623 1624 mtx = mtx_pool_find(mtxpool_sleep, npc); 1625 npc->metadata = (uintptr_t) mtx; 1626 1627 /* XXX no timeout passed down */ 1628 if (is_admin) 1629 nvme_ctrlr_submit_admin_request(ctrlr, req); 1630 else 1631 nvme_ctrlr_submit_io_request(ctrlr, req); 1632 1633 mtx_lock(mtx); 1634 while (npc->metadata != 0) 1635 mtx_sleep(npc, mtx, PRIBIO, "nvme_npc", 0); 1636 mtx_unlock(mtx); 1637 1638 if (npages > 0) 1639 nvme_user_ioctl_free(upages, npages, upages != upages_small); 1640 1641 return (ret); 1642 } 1643 1644 static int 1645 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag, 1646 struct thread *td) 1647 { 1648 struct nvme_controller *ctrlr; 1649 struct nvme_pt_command *pt; 1650 1651 ctrlr = cdev->si_drv1; 1652 1653 switch (cmd) { 1654 case NVME_IOCTL_RESET: /* Linux compat */ 1655 case NVME_RESET_CONTROLLER: 1656 nvme_ctrlr_reset(ctrlr); 1657 break; 1658 case NVME_PASSTHROUGH_CMD: 1659 pt = (struct nvme_pt_command *)arg; 1660 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid), 1661 1 /* is_user_buffer */, 1 /* is_admin_cmd */)); 1662 case NVME_GET_NSID: 1663 { 1664 struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg; 1665 strlcpy(gnsid->cdev, device_get_nameunit(ctrlr->dev), 1666 sizeof(gnsid->cdev)); 1667 gnsid->nsid = 0; 1668 break; 1669 } 1670 case NVME_GET_MAX_XFER_SIZE: 1671 *(uint64_t *)arg = ctrlr->max_xfer_size; 1672 break; 1673 case NVME_GET_CONTROLLER_DATA: 1674 memcpy(arg, &ctrlr->cdata, sizeof(ctrlr->cdata)); 1675 break; 1676 case DIOCGIDENT: { 1677 uint8_t *sn = arg; 1678 nvme_cdata_get_disk_ident(&ctrlr->cdata, sn); 1679 break; 1680 } 1681 /* Linux Compatible (see nvme_linux.h) */ 1682 case NVME_IOCTL_ID: 1683 td->td_retval[0] = 0xfffffffful; 1684 return (0); 1685 1686 case NVME_IOCTL_ADMIN_CMD: 1687 case NVME_IOCTL_IO_CMD: { 1688 struct nvme_passthru_cmd *npc = (struct nvme_passthru_cmd *)arg; 1689 1690 return (nvme_ctrlr_linux_passthru_cmd(ctrlr, npc, npc->nsid, true, 1691 cmd == NVME_IOCTL_ADMIN_CMD)); 1692 } 1693 1694 default: 1695 return (ENOTTY); 1696 } 1697 1698 return (0); 1699 } 1700 1701 static struct cdevsw nvme_ctrlr_cdevsw = { 1702 .d_version = D_VERSION, 1703 .d_flags = 0, 1704 .d_ioctl = nvme_ctrlr_ioctl 1705 }; 1706 1707 int 1708 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev) 1709 { 1710 struct make_dev_args md_args; 1711 uint32_t cap_lo; 1712 uint32_t cap_hi; 1713 uint32_t to, vs, pmrcap; 1714 int status, timeout_period; 1715 1716 ctrlr->dev = dev; 1717 ctrlr->io_sqes = 1718 (ctrlr->quirks & QUIRK_APPLE_128_BYTE_SQES) != 0 ? 1719 NVME_IOSQES_128 : NVME_IOSQES_64; 1720 1721 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF); 1722 if (bus_get_domain(dev, &ctrlr->domain) != 0) 1723 ctrlr->domain = 0; 1724 1725 ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo); 1726 if (bootverbose) { 1727 device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n", 1728 cap_lo, NVME_CAP_LO_MQES(cap_lo), 1729 NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "", 1730 NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "", 1731 (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "", 1732 (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "", 1733 NVME_CAP_LO_TO(cap_lo)); 1734 } 1735 ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi); 1736 if (bootverbose) { 1737 device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, " 1738 "CPS %x, MPSMIN %u, MPSMAX %u%s%s%s%s%s\n", cap_hi, 1739 NVME_CAP_HI_DSTRD(cap_hi), 1740 NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "", 1741 NVME_CAP_HI_CSS(cap_hi), 1742 NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "", 1743 NVME_CAP_HI_CPS(cap_hi), 1744 NVME_CAP_HI_MPSMIN(cap_hi), 1745 NVME_CAP_HI_MPSMAX(cap_hi), 1746 NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "", 1747 NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "", 1748 NVME_CAP_HI_NSSS(cap_hi) ? ", NSSS" : "", 1749 NVME_CAP_HI_CRWMS(cap_hi) ? ", CRWMS" : "", 1750 NVME_CAP_HI_CRIMS(cap_hi) ? ", CRIMS" : ""); 1751 } 1752 if (bootverbose) { 1753 vs = nvme_mmio_read_4(ctrlr, vs); 1754 device_printf(dev, "Version: 0x%08x: %d.%d\n", vs, 1755 NVME_MAJOR(vs), NVME_MINOR(vs)); 1756 } 1757 if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) { 1758 pmrcap = nvme_mmio_read_4(ctrlr, pmrcap); 1759 device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, " 1760 "PMRWBM %x, PMRTO %u%s\n", pmrcap, 1761 NVME_PMRCAP_BIR(pmrcap), 1762 NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "", 1763 NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "", 1764 NVME_PMRCAP_PMRTU(pmrcap), 1765 NVME_PMRCAP_PMRWBM(pmrcap), 1766 NVME_PMRCAP_PMRTO(pmrcap), 1767 NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : ""); 1768 } 1769 1770 ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2; 1771 1772 ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi); 1773 ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps); 1774 1775 /* Get ready timeout value from controller, in units of 500ms. */ 1776 to = NVME_CAP_LO_TO(cap_lo) + 1; 1777 ctrlr->ready_timeout_in_ms = to * 500; 1778 1779 timeout_period = NVME_ADMIN_TIMEOUT_PERIOD; 1780 TUNABLE_INT_FETCH("hw.nvme.admin_timeout_period", &timeout_period); 1781 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD); 1782 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD); 1783 ctrlr->admin_timeout_period = timeout_period; 1784 1785 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD; 1786 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period); 1787 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD); 1788 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD); 1789 ctrlr->timeout_period = timeout_period; 1790 1791 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT; 1792 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count); 1793 1794 ctrlr->enable_aborts = 0; 1795 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts); 1796 1797 ctrlr->alignment_splits = counter_u64_alloc(M_WAITOK); 1798 1799 /* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */ 1800 ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size)); 1801 if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0) 1802 return (ENXIO); 1803 1804 /* 1805 * Create 2 threads for the taskqueue. The reset thread will block when 1806 * it detects that the controller has failed until all I/O has been 1807 * failed up the stack. The second thread is used for AER events, which 1808 * can block, but only briefly for memory and log page fetching. 1809 */ 1810 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK, 1811 taskqueue_thread_enqueue, &ctrlr->taskqueue); 1812 taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq"); 1813 1814 ctrlr->is_resetting = 0; 1815 ctrlr->is_initialized = false; 1816 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr); 1817 for (int i = 0; i < NVME_MAX_ASYNC_EVENTS; i++) { 1818 struct nvme_async_event_request *aer = &ctrlr->aer[i]; 1819 1820 TASK_INIT(&aer->task, 0, nvme_ctrlr_aer_task, aer); 1821 mtx_init(&aer->mtx, "AER mutex", NULL, MTX_DEF); 1822 } 1823 ctrlr->is_failed = false; 1824 1825 make_dev_args_init(&md_args); 1826 md_args.mda_devsw = &nvme_ctrlr_cdevsw; 1827 md_args.mda_uid = UID_ROOT; 1828 md_args.mda_gid = GID_WHEEL; 1829 md_args.mda_mode = 0600; 1830 md_args.mda_unit = device_get_unit(dev); 1831 md_args.mda_si_drv1 = (void *)ctrlr; 1832 status = make_dev_s(&md_args, &ctrlr->cdev, "%s", 1833 device_get_nameunit(dev)); 1834 if (status != 0) 1835 return (ENXIO); 1836 1837 return (0); 1838 } 1839 1840 /* 1841 * Called on detach, or on error on attach. The nvme_controller won't be used 1842 * again once we return, so we have to tear everything down (so nothing 1843 * references this, no callbacks, etc), but don't need to reset all the state 1844 * since nvme_controller will be freed soon. 1845 */ 1846 void 1847 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev) 1848 { 1849 int i; 1850 bool gone; 1851 1852 ctrlr->is_dying = true; 1853 1854 if (ctrlr->resource == NULL) 1855 goto nores; 1856 if (!mtx_initialized(&ctrlr->adminq.lock)) 1857 goto noadminq; 1858 1859 /* 1860 * Check whether it is a hot unplug or a clean driver detach. If device 1861 * is not there any more, skip any shutdown commands. Some hotplug 1862 * bridges will return zeros instead of ff's when the device is 1863 * departing, so ask the bridge if the device is gone. Some systems can 1864 * remove the drive w/o the bridge knowing its gone (they don't really 1865 * do hotplug), so failsafe with detecting all ff's (impossible with 1866 * this hardware) as the device being gone. 1867 */ 1868 gone = bus_child_present(dev) == 0 || 1869 (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE); 1870 if (gone) 1871 nvme_ctrlr_fail(ctrlr, true); 1872 else 1873 nvme_notify_fail(ctrlr); 1874 1875 for (i = 0; i < NVME_MAX_NAMESPACES; i++) 1876 nvme_ns_destruct(&ctrlr->ns[i]); 1877 1878 if (ctrlr->cdev) 1879 destroy_dev(ctrlr->cdev); 1880 1881 if (ctrlr->is_initialized) { 1882 if (!gone) { 1883 if (ctrlr->hmb_nchunks > 0) 1884 nvme_ctrlr_hmb_enable(ctrlr, false, false); 1885 nvme_ctrlr_delete_qpairs(ctrlr); 1886 } 1887 nvme_ctrlr_hmb_free(ctrlr); 1888 } 1889 if (ctrlr->ioq != NULL) { 1890 for (i = 0; i < ctrlr->num_io_queues; i++) 1891 nvme_io_qpair_destroy(&ctrlr->ioq[i]); 1892 free(ctrlr->ioq, M_NVME); 1893 } 1894 nvme_admin_qpair_destroy(&ctrlr->adminq); 1895 1896 /* 1897 * Notify the controller of a shutdown, even though this is due to a 1898 * driver unload, not a system shutdown (this path is not invoked uring 1899 * shutdown). This ensures the controller receives a shutdown 1900 * notification in case the system is shutdown before reloading the 1901 * driver. Some NVMe drives need this to flush their cache to stable 1902 * media and consider it a safe shutdown in SMART stats. 1903 */ 1904 if (!gone) { 1905 nvme_ctrlr_shutdown(ctrlr); 1906 nvme_ctrlr_disable(ctrlr); 1907 } 1908 1909 noadminq: 1910 if (ctrlr->taskqueue) { 1911 taskqueue_free(ctrlr->taskqueue); 1912 for (int i = 0; i < NVME_MAX_ASYNC_EVENTS; i++) { 1913 struct nvme_async_event_request *aer = &ctrlr->aer[i]; 1914 1915 mtx_destroy(&aer->mtx); 1916 } 1917 } 1918 1919 if (ctrlr->tag) 1920 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag); 1921 1922 if (ctrlr->res) 1923 bus_release_resource(ctrlr->dev, SYS_RES_IRQ, 1924 rman_get_rid(ctrlr->res), ctrlr->res); 1925 1926 if (ctrlr->msix_table_resource != NULL) { 1927 bus_release_resource(dev, SYS_RES_MEMORY, 1928 ctrlr->msix_table_resource_id, ctrlr->msix_table_resource); 1929 } 1930 1931 if (ctrlr->msix_pba_resource != NULL) { 1932 bus_release_resource(dev, SYS_RES_MEMORY, 1933 ctrlr->msix_pba_resource_id, ctrlr->msix_pba_resource); 1934 } 1935 1936 bus_release_resource(dev, SYS_RES_MEMORY, 1937 ctrlr->resource_id, ctrlr->resource); 1938 1939 nores: 1940 if (ctrlr->alignment_splits) 1941 counter_u64_free(ctrlr->alignment_splits); 1942 1943 mtx_destroy(&ctrlr->lock); 1944 } 1945 1946 void 1947 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr) 1948 { 1949 uint32_t cc; 1950 uint32_t csts; 1951 int timeout; 1952 1953 cc = nvme_mmio_read_4(ctrlr, cc); 1954 cc &= ~NVMEM(NVME_CC_REG_SHN); 1955 cc |= NVMEF(NVME_CC_REG_SHN, NVME_SHN_NORMAL); 1956 nvme_mmio_write_4(ctrlr, cc, cc); 1957 1958 timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz : 1959 ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000); 1960 while (1) { 1961 csts = nvme_mmio_read_4(ctrlr, csts); 1962 if (csts == NVME_GONE) /* Hot unplug. */ 1963 break; 1964 if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE) 1965 break; 1966 if (timeout - ticks < 0) { 1967 nvme_printf(ctrlr, "shutdown timeout\n"); 1968 break; 1969 } 1970 pause("nvmeshut", 1); 1971 } 1972 } 1973 1974 void 1975 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr, 1976 struct nvme_request *req) 1977 { 1978 nvme_qpair_submit_request(&ctrlr->adminq, req); 1979 } 1980 1981 void 1982 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr, 1983 struct nvme_request *req) 1984 { 1985 struct nvme_qpair *qpair; 1986 int32_t ioq; 1987 1988 ioq = req->ioq == NVME_IOQ_DEFAULT ? QP(ctrlr, curcpu) : req->ioq; 1989 qpair = &ctrlr->ioq[ioq]; 1990 nvme_qpair_submit_request(qpair, req); 1991 } 1992 1993 device_t 1994 nvme_ctrlr_get_device(struct nvme_controller *ctrlr) 1995 { 1996 return (ctrlr->dev); 1997 } 1998 1999 const struct nvme_controller_data * 2000 nvme_ctrlr_get_data(struct nvme_controller *ctrlr) 2001 { 2002 return (&ctrlr->cdata); 2003 } 2004 2005 int 2006 nvme_ctrlr_suspend(struct nvme_controller *ctrlr) 2007 { 2008 int to = hz; 2009 2010 /* 2011 * Can't touch failed controllers, so it's already suspended. User will 2012 * need to do an explicit reset to bring it back, if that's even 2013 * possible. 2014 */ 2015 if (ctrlr->is_failed) 2016 return (0); 2017 2018 /* 2019 * We don't want the reset taskqueue running, since it does similar 2020 * things, so prevent it from running after we start. Wait for any reset 2021 * that may have been started to complete. The reset process we follow 2022 * will ensure that any new I/O will queue and be given to the hardware 2023 * after we resume (though there should be none). 2024 */ 2025 while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0) 2026 pause("nvmesusp", 1); 2027 if (to <= 0) { 2028 nvme_printf(ctrlr, 2029 "Competing reset task didn't finish. Try again later.\n"); 2030 return (EWOULDBLOCK); 2031 } 2032 2033 if (ctrlr->hmb_nchunks > 0) 2034 nvme_ctrlr_hmb_enable(ctrlr, false, false); 2035 2036 /* 2037 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to 2038 * delete the hardware I/O queues, and then shutdown. This properly 2039 * flushes any metadata the drive may have stored so it can survive 2040 * having its power removed and prevents the unsafe shutdown count from 2041 * incriminating. Once we delete the qpairs, we have to disable them 2042 * before shutting down. 2043 */ 2044 nvme_ctrlr_delete_qpairs(ctrlr); 2045 nvme_ctrlr_disable_qpairs(ctrlr); 2046 nvme_ctrlr_shutdown(ctrlr); 2047 2048 return (0); 2049 } 2050 2051 int 2052 nvme_ctrlr_resume(struct nvme_controller *ctrlr) 2053 { 2054 /* 2055 * Can't touch failed controllers, so nothing to do to resume. 2056 */ 2057 if (ctrlr->is_failed) 2058 return (0); 2059 2060 if (nvme_ctrlr_hw_reset(ctrlr) != 0) 2061 goto fail; 2062 2063 /* 2064 * Now that we've reset the hardware, we can restart the controller. Any 2065 * I/O that was pending is requeued. Any admin commands are aborted with 2066 * an error. Once we've restarted, stop flagging the controller as being 2067 * in the reset phase. 2068 */ 2069 nvme_ctrlr_start(ctrlr, true); 2070 (void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0); 2071 2072 return (0); 2073 fail: 2074 /* 2075 * Since we can't bring the controller out of reset, announce and fail 2076 * the controller. However, we have to return success for the resume 2077 * itself, due to questionable APIs. 2078 */ 2079 nvme_printf(ctrlr, "Failed to reset on resume, failing.\n"); 2080 nvme_ctrlr_fail(ctrlr, true); 2081 (void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0); 2082 return (0); 2083 } 2084