1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2015 Netflix, Inc. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer, 11 * without modification, immediately at the beginning of the file. 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 ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 * 27 * Derived from ata_da.c: 28 * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org> 29 */ 30 31 #include <sys/cdefs.h> 32 __FBSDID("$FreeBSD$"); 33 34 #include <sys/param.h> 35 36 #ifdef _KERNEL 37 #include <sys/systm.h> 38 #include <sys/kernel.h> 39 #include <sys/bio.h> 40 #include <sys/sysctl.h> 41 #include <sys/taskqueue.h> 42 #include <sys/lock.h> 43 #include <sys/mutex.h> 44 #include <sys/conf.h> 45 #include <sys/devicestat.h> 46 #include <sys/eventhandler.h> 47 #include <sys/malloc.h> 48 #include <sys/cons.h> 49 #include <sys/proc.h> 50 #include <sys/reboot.h> 51 #include <geom/geom_disk.h> 52 #endif /* _KERNEL */ 53 54 #ifndef _KERNEL 55 #include <stdio.h> 56 #include <string.h> 57 #endif /* _KERNEL */ 58 59 #include <cam/cam.h> 60 #include <cam/cam_ccb.h> 61 #include <cam/cam_periph.h> 62 #include <cam/cam_xpt_periph.h> 63 #include <cam/cam_sim.h> 64 #include <cam/cam_iosched.h> 65 66 #include <cam/nvme/nvme_all.h> 67 68 typedef enum { 69 NDA_STATE_NORMAL 70 } nda_state; 71 72 typedef enum { 73 NDA_FLAG_OPEN = 0x0001, 74 NDA_FLAG_DIRTY = 0x0002, 75 NDA_FLAG_SCTX_INIT = 0x0004, 76 } nda_flags; 77 78 typedef enum { 79 NDA_Q_4K = 0x01, 80 NDA_Q_NONE = 0x00, 81 } nda_quirks; 82 83 #define NDA_Q_BIT_STRING \ 84 "\020" \ 85 "\001Bit 0" 86 87 typedef enum { 88 NDA_CCB_BUFFER_IO = 0x01, 89 NDA_CCB_DUMP = 0x02, 90 NDA_CCB_TRIM = 0x03, 91 NDA_CCB_TYPE_MASK = 0x0F, 92 } nda_ccb_state; 93 94 /* Offsets into our private area for storing information */ 95 #define ccb_state ccb_h.ppriv_field0 96 #define ccb_bp ccb_h.ppriv_ptr1 /* For NDA_CCB_BUFFER_IO */ 97 #define ccb_trim ccb_h.ppriv_ptr1 /* For NDA_CCB_TRIM */ 98 99 struct nda_softc { 100 struct cam_iosched_softc *cam_iosched; 101 int outstanding_cmds; /* Number of active commands */ 102 int refcount; /* Active xpt_action() calls */ 103 nda_state state; 104 nda_flags flags; 105 nda_quirks quirks; 106 int unmappedio; 107 quad_t deletes; 108 uint32_t nsid; /* Namespace ID for this nda device */ 109 struct disk *disk; 110 struct task sysctl_task; 111 struct sysctl_ctx_list sysctl_ctx; 112 struct sysctl_oid *sysctl_tree; 113 uint64_t trim_count; 114 uint64_t trim_ranges; 115 uint64_t trim_lbas; 116 #ifdef CAM_TEST_FAILURE 117 int force_read_error; 118 int force_write_error; 119 int periodic_read_error; 120 int periodic_read_count; 121 #endif 122 #ifdef CAM_IO_STATS 123 struct sysctl_ctx_list sysctl_stats_ctx; 124 struct sysctl_oid *sysctl_stats_tree; 125 u_int timeouts; 126 u_int errors; 127 u_int invalidations; 128 #endif 129 }; 130 131 struct nda_trim_request { 132 struct nvme_dsm_range dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)]; 133 TAILQ_HEAD(, bio) bps; 134 }; 135 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0, 136 "NVME_MAX_DSM_TRIM must be an integral number of ranges"); 137 138 /* Need quirk table */ 139 140 static disk_strategy_t ndastrategy; 141 static dumper_t ndadump; 142 static periph_init_t ndainit; 143 static void ndaasync(void *callback_arg, u_int32_t code, 144 struct cam_path *path, void *arg); 145 static void ndasysctlinit(void *context, int pending); 146 static periph_ctor_t ndaregister; 147 static periph_dtor_t ndacleanup; 148 static periph_start_t ndastart; 149 static periph_oninv_t ndaoninvalidate; 150 static void ndadone(struct cam_periph *periph, 151 union ccb *done_ccb); 152 static int ndaerror(union ccb *ccb, u_int32_t cam_flags, 153 u_int32_t sense_flags); 154 static void ndashutdown(void *arg, int howto); 155 static void ndasuspend(void *arg); 156 157 #ifndef NDA_DEFAULT_SEND_ORDERED 158 #define NDA_DEFAULT_SEND_ORDERED 1 159 #endif 160 #ifndef NDA_DEFAULT_TIMEOUT 161 #define NDA_DEFAULT_TIMEOUT 30 /* Timeout in seconds */ 162 #endif 163 #ifndef NDA_DEFAULT_RETRY 164 #define NDA_DEFAULT_RETRY 4 165 #endif 166 #ifndef NDA_MAX_TRIM_ENTRIES 167 #define NDA_MAX_TRIM_ENTRIES (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */ 168 #endif 169 170 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD, 0, 171 "CAM Direct Access Disk driver"); 172 173 //static int nda_retry_count = NDA_DEFAULT_RETRY; 174 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED; 175 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT; 176 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES; 177 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN, 178 &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES, 179 "Maximum number of BIO_DELETE to send down as a DSM TRIM."); 180 181 /* 182 * All NVMe media is non-rotational, so all nvme device instances 183 * share this to implement the sysctl. 184 */ 185 static int nda_rotating_media = 0; 186 187 static struct periph_driver ndadriver = 188 { 189 ndainit, "nda", 190 TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0 191 }; 192 193 PERIPHDRIVER_DECLARE(nda, ndadriver); 194 195 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers"); 196 197 /* 198 * nice wrappers. Maybe these belong in nvme_all.c instead of 199 * here, but this is the only place that uses these. Should 200 * we ever grow another NVME periph, we should move them 201 * all there wholesale. 202 */ 203 204 static void 205 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio) 206 { 207 cam_fill_nvmeio(nvmeio, 208 0, /* retries */ 209 ndadone, /* cbfcnp */ 210 CAM_DIR_NONE, /* flags */ 211 NULL, /* data_ptr */ 212 0, /* dxfer_len */ 213 nda_default_timeout * 1000); /* timeout 30s */ 214 nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid); 215 } 216 217 static void 218 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 219 void *payload, uint32_t num_ranges) 220 { 221 cam_fill_nvmeio(nvmeio, 222 0, /* retries */ 223 ndadone, /* cbfcnp */ 224 CAM_DIR_OUT, /* flags */ 225 payload, /* data_ptr */ 226 num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */ 227 nda_default_timeout * 1000); /* timeout 30s */ 228 nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges); 229 } 230 231 static void 232 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 233 void *payload, uint64_t lba, uint32_t len, uint32_t count) 234 { 235 cam_fill_nvmeio(nvmeio, 236 0, /* retries */ 237 ndadone, /* cbfcnp */ 238 CAM_DIR_OUT, /* flags */ 239 payload, /* data_ptr */ 240 len, /* dxfer_len */ 241 nda_default_timeout * 1000); /* timeout 30s */ 242 nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count); 243 } 244 245 static void 246 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 247 struct bio *bp, uint32_t rwcmd) 248 { 249 int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT; 250 void *payload; 251 uint64_t lba; 252 uint32_t count; 253 254 if (bp->bio_flags & BIO_UNMAPPED) { 255 flags |= CAM_DATA_BIO; 256 payload = bp; 257 } else { 258 payload = bp->bio_data; 259 } 260 261 lba = bp->bio_pblkno; 262 count = bp->bio_bcount / softc->disk->d_sectorsize; 263 264 cam_fill_nvmeio(nvmeio, 265 0, /* retries */ 266 ndadone, /* cbfcnp */ 267 flags, /* flags */ 268 payload, /* data_ptr */ 269 bp->bio_bcount, /* dxfer_len */ 270 nda_default_timeout * 1000); /* timeout 30s */ 271 nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count); 272 } 273 274 static int 275 ndaopen(struct disk *dp) 276 { 277 struct cam_periph *periph; 278 struct nda_softc *softc; 279 int error; 280 281 periph = (struct cam_periph *)dp->d_drv1; 282 if (cam_periph_acquire(periph) != 0) { 283 return(ENXIO); 284 } 285 286 cam_periph_lock(periph); 287 if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) { 288 cam_periph_unlock(periph); 289 cam_periph_release(periph); 290 return (error); 291 } 292 293 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 294 ("ndaopen\n")); 295 296 softc = (struct nda_softc *)periph->softc; 297 softc->flags |= NDA_FLAG_OPEN; 298 299 cam_periph_unhold(periph); 300 cam_periph_unlock(periph); 301 return (0); 302 } 303 304 static int 305 ndaclose(struct disk *dp) 306 { 307 struct cam_periph *periph; 308 struct nda_softc *softc; 309 union ccb *ccb; 310 int error; 311 312 periph = (struct cam_periph *)dp->d_drv1; 313 softc = (struct nda_softc *)periph->softc; 314 cam_periph_lock(periph); 315 316 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 317 ("ndaclose\n")); 318 319 if ((softc->flags & NDA_FLAG_DIRTY) != 0 && 320 (periph->flags & CAM_PERIPH_INVALID) == 0 && 321 cam_periph_hold(periph, PRIBIO) == 0) { 322 323 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 324 nda_nvme_flush(softc, &ccb->nvmeio); 325 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 326 /*sense_flags*/0, softc->disk->d_devstat); 327 328 if (error != 0) 329 xpt_print(periph->path, "Synchronize cache failed\n"); 330 else 331 softc->flags &= ~NDA_FLAG_DIRTY; 332 xpt_release_ccb(ccb); 333 cam_periph_unhold(periph); 334 } 335 336 softc->flags &= ~NDA_FLAG_OPEN; 337 338 while (softc->refcount != 0) 339 cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1); 340 KASSERT(softc->outstanding_cmds == 0, 341 ("nda %d outstanding commands", softc->outstanding_cmds)); 342 cam_periph_unlock(periph); 343 cam_periph_release(periph); 344 return (0); 345 } 346 347 static void 348 ndaschedule(struct cam_periph *periph) 349 { 350 struct nda_softc *softc = (struct nda_softc *)periph->softc; 351 352 if (softc->state != NDA_STATE_NORMAL) 353 return; 354 355 cam_iosched_schedule(softc->cam_iosched, periph); 356 } 357 358 /* 359 * Actually translate the requested transfer into one the physical driver 360 * can understand. The transfer is described by a buf and will include 361 * only one physical transfer. 362 */ 363 static void 364 ndastrategy(struct bio *bp) 365 { 366 struct cam_periph *periph; 367 struct nda_softc *softc; 368 369 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 370 softc = (struct nda_softc *)periph->softc; 371 372 cam_periph_lock(periph); 373 374 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp)); 375 376 /* 377 * If the device has been made invalid, error out 378 */ 379 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 380 cam_periph_unlock(periph); 381 biofinish(bp, NULL, ENXIO); 382 return; 383 } 384 385 if (bp->bio_cmd == BIO_DELETE) 386 softc->deletes++; 387 388 /* 389 * Place it in the queue of disk activities for this disk 390 */ 391 cam_iosched_queue_work(softc->cam_iosched, bp); 392 393 /* 394 * Schedule ourselves for performing the work. 395 */ 396 ndaschedule(periph); 397 cam_periph_unlock(periph); 398 399 return; 400 } 401 402 static int 403 ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length) 404 { 405 struct cam_periph *periph; 406 struct nda_softc *softc; 407 u_int secsize; 408 struct ccb_nvmeio nvmeio; 409 struct disk *dp; 410 uint64_t lba; 411 uint32_t count; 412 int error = 0; 413 414 dp = arg; 415 periph = dp->d_drv1; 416 softc = (struct nda_softc *)periph->softc; 417 secsize = softc->disk->d_sectorsize; 418 lba = offset / secsize; 419 count = length / secsize; 420 421 if ((periph->flags & CAM_PERIPH_INVALID) != 0) 422 return (ENXIO); 423 424 /* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */ 425 memset(&nvmeio, 0, sizeof(nvmeio)); 426 if (length > 0) { 427 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 428 nvmeio.ccb_state = NDA_CCB_DUMP; 429 nda_nvme_write(softc, &nvmeio, virtual, lba, length, count); 430 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 431 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 432 if (error != 0) 433 printf("Aborting dump due to I/O error %d.\n", error); 434 435 return (error); 436 } 437 438 /* Flush */ 439 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 440 441 nvmeio.ccb_state = NDA_CCB_DUMP; 442 nda_nvme_flush(softc, &nvmeio); 443 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 444 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 445 if (error != 0) 446 xpt_print(periph->path, "flush cmd failed\n"); 447 return (error); 448 } 449 450 static void 451 ndainit(void) 452 { 453 cam_status status; 454 455 /* 456 * Install a global async callback. This callback will 457 * receive async callbacks like "new device found". 458 */ 459 status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL); 460 461 if (status != CAM_REQ_CMP) { 462 printf("nda: Failed to attach master async callback " 463 "due to status 0x%x!\n", status); 464 } else if (nda_send_ordered) { 465 466 /* Register our event handlers */ 467 if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend, 468 NULL, EVENTHANDLER_PRI_LAST)) == NULL) 469 printf("ndainit: power event registration failed!\n"); 470 if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown, 471 NULL, SHUTDOWN_PRI_DEFAULT)) == NULL) 472 printf("ndainit: shutdown event registration failed!\n"); 473 } 474 } 475 476 /* 477 * Callback from GEOM, called when it has finished cleaning up its 478 * resources. 479 */ 480 static void 481 ndadiskgonecb(struct disk *dp) 482 { 483 struct cam_periph *periph; 484 485 periph = (struct cam_periph *)dp->d_drv1; 486 487 cam_periph_release(periph); 488 } 489 490 static void 491 ndaoninvalidate(struct cam_periph *periph) 492 { 493 struct nda_softc *softc; 494 495 softc = (struct nda_softc *)periph->softc; 496 497 /* 498 * De-register any async callbacks. 499 */ 500 xpt_register_async(0, ndaasync, periph, periph->path); 501 #ifdef CAM_IO_STATS 502 softc->invalidations++; 503 #endif 504 505 /* 506 * Return all queued I/O with ENXIO. 507 * XXX Handle any transactions queued to the card 508 * with XPT_ABORT_CCB. 509 */ 510 cam_iosched_flush(softc->cam_iosched, NULL, ENXIO); 511 512 disk_gone(softc->disk); 513 } 514 515 static void 516 ndacleanup(struct cam_periph *periph) 517 { 518 struct nda_softc *softc; 519 520 softc = (struct nda_softc *)periph->softc; 521 522 cam_periph_unlock(periph); 523 524 cam_iosched_fini(softc->cam_iosched); 525 526 /* 527 * If we can't free the sysctl tree, oh well... 528 */ 529 if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) { 530 #ifdef CAM_IO_STATS 531 if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0) 532 xpt_print(periph->path, 533 "can't remove sysctl stats context\n"); 534 #endif 535 if (sysctl_ctx_free(&softc->sysctl_ctx) != 0) 536 xpt_print(periph->path, 537 "can't remove sysctl context\n"); 538 } 539 540 disk_destroy(softc->disk); 541 free(softc, M_DEVBUF); 542 cam_periph_lock(periph); 543 } 544 545 static void 546 ndaasync(void *callback_arg, u_int32_t code, 547 struct cam_path *path, void *arg) 548 { 549 struct cam_periph *periph; 550 551 periph = (struct cam_periph *)callback_arg; 552 switch (code) { 553 case AC_FOUND_DEVICE: 554 { 555 struct ccb_getdev *cgd; 556 cam_status status; 557 558 cgd = (struct ccb_getdev *)arg; 559 if (cgd == NULL) 560 break; 561 562 if (cgd->protocol != PROTO_NVME) 563 break; 564 565 /* 566 * Allocate a peripheral instance for 567 * this device and start the probe 568 * process. 569 */ 570 status = cam_periph_alloc(ndaregister, ndaoninvalidate, 571 ndacleanup, ndastart, 572 "nda", CAM_PERIPH_BIO, 573 path, ndaasync, 574 AC_FOUND_DEVICE, cgd); 575 576 if (status != CAM_REQ_CMP 577 && status != CAM_REQ_INPROG) 578 printf("ndaasync: Unable to attach to new device " 579 "due to status 0x%x\n", status); 580 break; 581 } 582 case AC_ADVINFO_CHANGED: 583 { 584 uintptr_t buftype; 585 586 buftype = (uintptr_t)arg; 587 if (buftype == CDAI_TYPE_PHYS_PATH) { 588 struct nda_softc *softc; 589 590 softc = periph->softc; 591 disk_attr_changed(softc->disk, "GEOM::physpath", 592 M_NOWAIT); 593 } 594 break; 595 } 596 case AC_LOST_DEVICE: 597 default: 598 cam_periph_async(periph, code, path, arg); 599 break; 600 } 601 } 602 603 static void 604 ndasysctlinit(void *context, int pending) 605 { 606 struct cam_periph *periph; 607 struct nda_softc *softc; 608 char tmpstr[32], tmpstr2[16]; 609 610 periph = (struct cam_periph *)context; 611 612 /* periph was held for us when this task was enqueued */ 613 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 614 cam_periph_release(periph); 615 return; 616 } 617 618 softc = (struct nda_softc *)periph->softc; 619 snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number); 620 snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number); 621 622 sysctl_ctx_init(&softc->sysctl_ctx); 623 softc->flags |= NDA_FLAG_SCTX_INIT; 624 softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx, 625 SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2, 626 CTLFLAG_RD, 0, tmpstr, "device_index"); 627 if (softc->sysctl_tree == NULL) { 628 printf("ndasysctlinit: unable to allocate sysctl tree\n"); 629 cam_periph_release(periph); 630 return; 631 } 632 633 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 634 OID_AUTO, "unmapped_io", CTLFLAG_RD, 635 &softc->unmappedio, 0, "Unmapped I/O leaf"); 636 637 SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 638 OID_AUTO, "deletes", CTLFLAG_RD, 639 &softc->deletes, "Number of BIO_DELETE requests"); 640 641 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 642 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 643 "trim_count", CTLFLAG_RD, &softc->trim_count, 644 "Total number of unmap/dsm commands sent"); 645 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 646 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 647 "trim_ranges", CTLFLAG_RD, &softc->trim_ranges, 648 "Total number of ranges in unmap/dsm commands"); 649 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 650 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 651 "trim_lbas", CTLFLAG_RD, &softc->trim_lbas, 652 "Total lbas in the unmap/dsm commands sent"); 653 654 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 655 OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1, 656 "Rotating media"); 657 658 #ifdef CAM_IO_STATS 659 softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx, 660 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats", 661 CTLFLAG_RD, 0, "Statistics"); 662 if (softc->sysctl_stats_tree == NULL) { 663 printf("ndasysctlinit: unable to allocate sysctl tree for stats\n"); 664 cam_periph_release(periph); 665 return; 666 } 667 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 668 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 669 OID_AUTO, "timeouts", CTLFLAG_RD, 670 &softc->timeouts, 0, 671 "Device timeouts reported by the SIM"); 672 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 673 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 674 OID_AUTO, "errors", CTLFLAG_RD, 675 &softc->errors, 0, 676 "Transport errors reported by the SIM."); 677 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 678 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 679 OID_AUTO, "pack_invalidations", CTLFLAG_RD, 680 &softc->invalidations, 0, 681 "Device pack invalidations."); 682 #endif 683 684 #ifdef CAM_TEST_FAILURE 685 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 686 OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 687 periph, 0, cam_periph_invalidate_sysctl, "I", 688 "Write 1 to invalidate the drive immediately"); 689 #endif 690 691 cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx, 692 softc->sysctl_tree); 693 694 cam_periph_release(periph); 695 } 696 697 static int 698 ndagetattr(struct bio *bp) 699 { 700 int ret; 701 struct cam_periph *periph; 702 703 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 704 cam_periph_lock(periph); 705 ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute, 706 periph->path); 707 cam_periph_unlock(periph); 708 if (ret == 0) 709 bp->bio_completed = bp->bio_length; 710 return ret; 711 } 712 713 static cam_status 714 ndaregister(struct cam_periph *periph, void *arg) 715 { 716 struct nda_softc *softc; 717 struct disk *disk; 718 struct ccb_pathinq cpi; 719 const struct nvme_namespace_data *nsd; 720 const struct nvme_controller_data *cd; 721 char announce_buf[80]; 722 uint8_t flbas_fmt, lbads, vwc_present; 723 u_int maxio; 724 int quirks; 725 726 nsd = nvme_get_identify_ns(periph); 727 cd = nvme_get_identify_cntrl(periph); 728 729 softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF, 730 M_NOWAIT | M_ZERO); 731 732 if (softc == NULL) { 733 printf("ndaregister: Unable to probe new device. " 734 "Unable to allocate softc\n"); 735 return(CAM_REQ_CMP_ERR); 736 } 737 738 if (cam_iosched_init(&softc->cam_iosched, periph) != 0) { 739 printf("ndaregister: Unable to probe new device. " 740 "Unable to allocate iosched memory\n"); 741 free(softc, M_DEVBUF); 742 return(CAM_REQ_CMP_ERR); 743 } 744 745 /* ident_data parsing */ 746 747 periph->softc = softc; 748 749 softc->quirks = NDA_Q_NONE; 750 751 xpt_path_inq(&cpi, periph->path); 752 753 TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph); 754 755 /* 756 * The name space ID is the lun, save it for later I/O 757 */ 758 softc->nsid = (uint32_t)xpt_path_lun_id(periph->path); 759 760 /* 761 * Register this media as a disk 762 */ 763 (void)cam_periph_hold(periph, PRIBIO); 764 cam_periph_unlock(periph); 765 snprintf(announce_buf, sizeof(announce_buf), 766 "kern.cam.nda.%d.quirks", periph->unit_number); 767 quirks = softc->quirks; 768 TUNABLE_INT_FETCH(announce_buf, &quirks); 769 softc->quirks = quirks; 770 cam_iosched_set_sort_queue(softc->cam_iosched, 0); 771 softc->disk = disk = disk_alloc(); 772 disk->d_rotation_rate = DISK_RR_NON_ROTATING; 773 disk->d_open = ndaopen; 774 disk->d_close = ndaclose; 775 disk->d_strategy = ndastrategy; 776 disk->d_getattr = ndagetattr; 777 disk->d_dump = ndadump; 778 disk->d_gone = ndadiskgonecb; 779 disk->d_name = "nda"; 780 disk->d_drv1 = periph; 781 disk->d_unit = periph->unit_number; 782 maxio = cpi.maxio; /* Honor max I/O size of SIM */ 783 if (maxio == 0) 784 maxio = DFLTPHYS; /* traditional default */ 785 else if (maxio > MAXPHYS) 786 maxio = MAXPHYS; /* for safety */ 787 disk->d_maxsize = maxio; 788 flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) & 789 NVME_NS_DATA_FLBAS_FORMAT_MASK; 790 lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) & 791 NVME_NS_DATA_LBAF_LBADS_MASK; 792 disk->d_sectorsize = 1 << lbads; 793 disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze); 794 disk->d_delmaxsize = disk->d_mediasize; 795 disk->d_flags = DISKFLAG_DIRECT_COMPLETION; 796 if (nvme_ctrlr_has_dataset_mgmt(cd)) 797 disk->d_flags |= DISKFLAG_CANDELETE; 798 vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) & 799 NVME_CTRLR_DATA_VWC_PRESENT_MASK; 800 if (vwc_present) 801 disk->d_flags |= DISKFLAG_CANFLUSHCACHE; 802 if ((cpi.hba_misc & PIM_UNMAPPED) != 0) { 803 disk->d_flags |= DISKFLAG_UNMAPPED_BIO; 804 softc->unmappedio = 1; 805 } 806 /* 807 * d_ident and d_descr are both far bigger than the length of either 808 * the serial or model number strings. 809 */ 810 cam_strvis(disk->d_descr, cd->mn, 811 sizeof(disk->d_descr), NVME_MODEL_NUMBER_LENGTH); 812 cam_strvis(disk->d_ident, cd->sn, 813 sizeof(disk->d_ident), NVME_SERIAL_NUMBER_LENGTH); 814 disk->d_hba_vendor = cpi.hba_vendor; 815 disk->d_hba_device = cpi.hba_device; 816 disk->d_hba_subvendor = cpi.hba_subvendor; 817 disk->d_hba_subdevice = cpi.hba_subdevice; 818 disk->d_stripesize = disk->d_sectorsize; 819 disk->d_stripeoffset = 0; 820 disk->d_devstat = devstat_new_entry(periph->periph_name, 821 periph->unit_number, disk->d_sectorsize, 822 DEVSTAT_ALL_SUPPORTED, 823 DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport), 824 DEVSTAT_PRIORITY_DISK); 825 /* 826 * Add alias for older nvd drives to ease transition. 827 */ 828 /* disk_add_alias(disk, "nvd"); Have reports of this causing problems */ 829 830 /* 831 * Acquire a reference to the periph before we register with GEOM. 832 * We'll release this reference once GEOM calls us back (via 833 * ndadiskgonecb()) telling us that our provider has been freed. 834 */ 835 if (cam_periph_acquire(periph) != 0) { 836 xpt_print(periph->path, "%s: lost periph during " 837 "registration!\n", __func__); 838 cam_periph_lock(periph); 839 return (CAM_REQ_CMP_ERR); 840 } 841 disk_create(softc->disk, DISK_VERSION); 842 cam_periph_lock(periph); 843 cam_periph_unhold(periph); 844 845 snprintf(announce_buf, sizeof(announce_buf), 846 "%juMB (%ju %u byte sectors)", 847 (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)), 848 (uintmax_t)disk->d_mediasize / disk->d_sectorsize, 849 disk->d_sectorsize); 850 xpt_announce_periph(periph, announce_buf); 851 xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING); 852 853 /* 854 * Create our sysctl variables, now that we know 855 * we have successfully attached. 856 */ 857 if (cam_periph_acquire(periph) == 0) 858 taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task); 859 860 /* 861 * Register for device going away and info about the drive 862 * changing (though with NVMe, it can't) 863 */ 864 xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED, 865 ndaasync, periph, periph->path); 866 867 softc->state = NDA_STATE_NORMAL; 868 return(CAM_REQ_CMP); 869 } 870 871 static void 872 ndastart(struct cam_periph *periph, union ccb *start_ccb) 873 { 874 struct nda_softc *softc = (struct nda_softc *)periph->softc; 875 struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio; 876 877 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n")); 878 879 switch (softc->state) { 880 case NDA_STATE_NORMAL: 881 { 882 struct bio *bp; 883 884 bp = cam_iosched_next_bio(softc->cam_iosched); 885 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp)); 886 if (bp == NULL) { 887 xpt_release_ccb(start_ccb); 888 break; 889 } 890 891 switch (bp->bio_cmd) { 892 case BIO_WRITE: 893 softc->flags |= NDA_FLAG_DIRTY; 894 /* FALLTHROUGH */ 895 case BIO_READ: 896 { 897 #ifdef CAM_TEST_FAILURE 898 int fail = 0; 899 900 /* 901 * Support the failure ioctls. If the command is a 902 * read, and there are pending forced read errors, or 903 * if a write and pending write errors, then fail this 904 * operation with EIO. This is useful for testing 905 * purposes. Also, support having every Nth read fail. 906 * 907 * This is a rather blunt tool. 908 */ 909 if (bp->bio_cmd == BIO_READ) { 910 if (softc->force_read_error) { 911 softc->force_read_error--; 912 fail = 1; 913 } 914 if (softc->periodic_read_error > 0) { 915 if (++softc->periodic_read_count >= 916 softc->periodic_read_error) { 917 softc->periodic_read_count = 0; 918 fail = 1; 919 } 920 } 921 } else { 922 if (softc->force_write_error) { 923 softc->force_write_error--; 924 fail = 1; 925 } 926 } 927 if (fail) { 928 biofinish(bp, NULL, EIO); 929 xpt_release_ccb(start_ccb); 930 ndaschedule(periph); 931 return; 932 } 933 #endif 934 KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 || 935 round_page(bp->bio_bcount + bp->bio_ma_offset) / 936 PAGE_SIZE == bp->bio_ma_n, 937 ("Short bio %p", bp)); 938 nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ? 939 NVME_OPC_READ : NVME_OPC_WRITE); 940 break; 941 } 942 case BIO_DELETE: 943 { 944 struct nvme_dsm_range *dsm_range, *dsm_end; 945 struct nda_trim_request *trim; 946 struct bio *bp1; 947 int ents; 948 uint32_t totalcount = 0, ranges = 0; 949 950 trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT); 951 if (trim == NULL) { 952 biofinish(bp, NULL, ENOMEM); 953 xpt_release_ccb(start_ccb); 954 ndaschedule(periph); 955 return; 956 } 957 TAILQ_INIT(&trim->bps); 958 bp1 = bp; 959 ents = min(nitems(trim->dsm), nda_max_trim_entries); 960 dsm_range = trim->dsm; 961 dsm_end = dsm_range + ents; 962 do { 963 TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue); 964 dsm_range->length = 965 htole32(bp1->bio_bcount / softc->disk->d_sectorsize); 966 dsm_range->starting_lba = 967 htole64(bp1->bio_offset / softc->disk->d_sectorsize); 968 ranges++; 969 totalcount += dsm_range->length; 970 dsm_range++; 971 if (dsm_range >= dsm_end) 972 break; 973 bp1 = cam_iosched_next_trim(softc->cam_iosched); 974 /* XXX -- Could collapse adjacent ranges, but we don't for now */ 975 /* XXX -- Could limit based on total payload size */ 976 } while (bp1 != NULL); 977 start_ccb->ccb_trim = trim; 978 nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm, 979 dsm_range - trim->dsm); 980 start_ccb->ccb_state = NDA_CCB_TRIM; 981 softc->trim_count++; 982 softc->trim_ranges += ranges; 983 softc->trim_lbas += totalcount; 984 /* 985 * Note: We can have multiple TRIMs in flight, so we don't call 986 * cam_iosched_submit_trim(softc->cam_iosched); 987 * since that forces the I/O scheduler to only schedule one at a time. 988 * On NVMe drives, this is a performance disaster. 989 */ 990 goto out; 991 } 992 case BIO_FLUSH: 993 nda_nvme_flush(softc, nvmeio); 994 break; 995 } 996 start_ccb->ccb_state = NDA_CCB_BUFFER_IO; 997 start_ccb->ccb_bp = bp; 998 out: 999 start_ccb->ccb_h.flags |= CAM_UNLOCKED; 1000 softc->outstanding_cmds++; 1001 softc->refcount++; /* For submission only */ 1002 cam_periph_unlock(periph); 1003 xpt_action(start_ccb); 1004 cam_periph_lock(periph); 1005 softc->refcount--; /* Submission done */ 1006 1007 /* May have more work to do, so ensure we stay scheduled */ 1008 ndaschedule(periph); 1009 break; 1010 } 1011 } 1012 } 1013 1014 static void 1015 ndadone(struct cam_periph *periph, union ccb *done_ccb) 1016 { 1017 struct nda_softc *softc; 1018 struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio; 1019 struct cam_path *path; 1020 int state; 1021 1022 softc = (struct nda_softc *)periph->softc; 1023 path = done_ccb->ccb_h.path; 1024 1025 CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n")); 1026 1027 state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK; 1028 switch (state) { 1029 case NDA_CCB_BUFFER_IO: 1030 case NDA_CCB_TRIM: 1031 { 1032 int error; 1033 1034 cam_periph_lock(periph); 1035 if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) { 1036 error = ndaerror(done_ccb, 0, 0); 1037 if (error == ERESTART) { 1038 /* A retry was scheduled, so just return. */ 1039 cam_periph_unlock(periph); 1040 return; 1041 } 1042 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1043 cam_release_devq(path, 1044 /*relsim_flags*/0, 1045 /*reduction*/0, 1046 /*timeout*/0, 1047 /*getcount_only*/0); 1048 } else { 1049 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1050 panic("REQ_CMP with QFRZN"); 1051 error = 0; 1052 } 1053 if (state == NDA_CCB_BUFFER_IO) { 1054 struct bio *bp; 1055 1056 bp = (struct bio *)done_ccb->ccb_bp; 1057 bp->bio_error = error; 1058 if (error != 0) { 1059 bp->bio_resid = bp->bio_bcount; 1060 bp->bio_flags |= BIO_ERROR; 1061 } else { 1062 bp->bio_resid = 0; 1063 } 1064 softc->outstanding_cmds--; 1065 1066 /* 1067 * We need to call cam_iosched before we call biodone so that we 1068 * don't measure any activity that happens in the completion 1069 * routine, which in the case of sendfile can be quite 1070 * extensive. 1071 */ 1072 cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb); 1073 xpt_release_ccb(done_ccb); 1074 ndaschedule(periph); 1075 cam_periph_unlock(periph); 1076 biodone(bp); 1077 } else { /* state == NDA_CCB_TRIM */ 1078 struct nda_trim_request *trim; 1079 struct bio *bp1, *bp2; 1080 TAILQ_HEAD(, bio) queue; 1081 1082 trim = nvmeio->ccb_trim; 1083 TAILQ_INIT(&queue); 1084 TAILQ_CONCAT(&queue, &trim->bps, bio_queue); 1085 free(trim, M_NVMEDA); 1086 1087 /* 1088 * Since we can have multiple trims in flight, we don't 1089 * need to call this here. 1090 * cam_iosched_trim_done(softc->cam_iosched); 1091 */ 1092 /* 1093 * The the I/O scheduler that we're finishing the I/O 1094 * so we can keep book. The first one we pass in the CCB 1095 * which has the timing information. The rest we pass in NULL 1096 * so we can keep proper counts. 1097 */ 1098 bp1 = TAILQ_FIRST(&queue); 1099 cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb); 1100 xpt_release_ccb(done_ccb); 1101 softc->outstanding_cmds--; 1102 ndaschedule(periph); 1103 cam_periph_unlock(periph); 1104 while ((bp2 = TAILQ_FIRST(&queue)) != NULL) { 1105 TAILQ_REMOVE(&queue, bp2, bio_queue); 1106 bp2->bio_error = error; 1107 if (error != 0) { 1108 bp2->bio_flags |= BIO_ERROR; 1109 bp2->bio_resid = bp1->bio_bcount; 1110 } else 1111 bp2->bio_resid = 0; 1112 if (bp1 != bp2) 1113 cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL); 1114 biodone(bp2); 1115 } 1116 } 1117 return; 1118 } 1119 case NDA_CCB_DUMP: 1120 /* No-op. We're polling */ 1121 return; 1122 default: 1123 break; 1124 } 1125 xpt_release_ccb(done_ccb); 1126 } 1127 1128 static int 1129 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags) 1130 { 1131 struct nda_softc *softc; 1132 struct cam_periph *periph; 1133 1134 periph = xpt_path_periph(ccb->ccb_h.path); 1135 softc = (struct nda_softc *)periph->softc; 1136 1137 switch (ccb->ccb_h.status & CAM_STATUS_MASK) { 1138 case CAM_CMD_TIMEOUT: 1139 #ifdef CAM_IO_STATS 1140 softc->timeouts++; 1141 #endif 1142 break; 1143 case CAM_REQ_ABORTED: 1144 case CAM_REQ_CMP_ERR: 1145 case CAM_REQ_TERMIO: 1146 case CAM_UNREC_HBA_ERROR: 1147 case CAM_DATA_RUN_ERR: 1148 case CAM_ATA_STATUS_ERROR: 1149 #ifdef CAM_IO_STATS 1150 softc->errors++; 1151 #endif 1152 break; 1153 default: 1154 break; 1155 } 1156 1157 return(cam_periph_error(ccb, cam_flags, sense_flags)); 1158 } 1159 1160 /* 1161 * Step through all NDA peripheral drivers, and if the device is still open, 1162 * sync the disk cache to physical media. 1163 */ 1164 static void 1165 ndaflush(void) 1166 { 1167 struct cam_periph *periph; 1168 struct nda_softc *softc; 1169 union ccb *ccb; 1170 int error; 1171 1172 CAM_PERIPH_FOREACH(periph, &ndadriver) { 1173 softc = (struct nda_softc *)periph->softc; 1174 1175 if (SCHEDULER_STOPPED()) { 1176 /* 1177 * If we paniced with the lock held or the periph is not 1178 * open, do not recurse. Otherwise, call ndadump since 1179 * that avoids the sleeping cam_periph_getccb does if no 1180 * CCBs are available. 1181 */ 1182 if (!cam_periph_owned(periph) && 1183 (softc->flags & NDA_FLAG_OPEN)) { 1184 ndadump(softc->disk, NULL, 0, 0, 0); 1185 } 1186 continue; 1187 } 1188 1189 /* 1190 * We only sync the cache if the drive is still open 1191 */ 1192 cam_periph_lock(periph); 1193 if ((softc->flags & NDA_FLAG_OPEN) == 0) { 1194 cam_periph_unlock(periph); 1195 continue; 1196 } 1197 1198 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 1199 nda_nvme_flush(softc, &ccb->nvmeio); 1200 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 1201 /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY, 1202 softc->disk->d_devstat); 1203 if (error != 0) 1204 xpt_print(periph->path, "Synchronize cache failed\n"); 1205 xpt_release_ccb(ccb); 1206 cam_periph_unlock(periph); 1207 } 1208 } 1209 1210 static void 1211 ndashutdown(void *arg, int howto) 1212 { 1213 1214 ndaflush(); 1215 } 1216 1217 static void 1218 ndasuspend(void *arg) 1219 { 1220 1221 ndaflush(); 1222 } 1223