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 union { 133 struct nvme_dsm_range dsm; 134 uint8_t data[NVME_MAX_DSM_TRIM]; 135 }; 136 TAILQ_HEAD(, bio) bps; 137 }; 138 139 /* Need quirk table */ 140 141 static disk_strategy_t ndastrategy; 142 static dumper_t ndadump; 143 static periph_init_t ndainit; 144 static void ndaasync(void *callback_arg, u_int32_t code, 145 struct cam_path *path, void *arg); 146 static void ndasysctlinit(void *context, int pending); 147 static periph_ctor_t ndaregister; 148 static periph_dtor_t ndacleanup; 149 static periph_start_t ndastart; 150 static periph_oninv_t ndaoninvalidate; 151 static void ndadone(struct cam_periph *periph, 152 union ccb *done_ccb); 153 static int ndaerror(union ccb *ccb, u_int32_t cam_flags, 154 u_int32_t sense_flags); 155 static void ndashutdown(void *arg, int howto); 156 static void ndasuspend(void *arg); 157 158 #ifndef NDA_DEFAULT_SEND_ORDERED 159 #define NDA_DEFAULT_SEND_ORDERED 1 160 #endif 161 #ifndef NDA_DEFAULT_TIMEOUT 162 #define NDA_DEFAULT_TIMEOUT 30 /* Timeout in seconds */ 163 #endif 164 #ifndef NDA_DEFAULT_RETRY 165 #define NDA_DEFAULT_RETRY 4 166 #endif 167 #ifndef NDA_MAX_TRIM_ENTRIES 168 #define NDA_MAX_TRIM_ENTRIES (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */ 169 #endif 170 171 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD, 0, 172 "CAM Direct Access Disk driver"); 173 174 //static int nda_retry_count = NDA_DEFAULT_RETRY; 175 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED; 176 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT; 177 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES; 178 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN, 179 &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES, 180 "Maximum number of BIO_DELETE to send down as a DSM TRIM."); 181 182 /* 183 * All NVMe media is non-rotational, so all nvme device instances 184 * share this to implement the sysctl. 185 */ 186 static int nda_rotating_media = 0; 187 188 static struct periph_driver ndadriver = 189 { 190 ndainit, "nda", 191 TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0 192 }; 193 194 PERIPHDRIVER_DECLARE(nda, ndadriver); 195 196 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers"); 197 198 /* 199 * nice wrappers. Maybe these belong in nvme_all.c instead of 200 * here, but this is the only place that uses these. Should 201 * we ever grow another NVME periph, we should move them 202 * all there wholesale. 203 */ 204 205 static void 206 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio) 207 { 208 cam_fill_nvmeio(nvmeio, 209 0, /* retries */ 210 ndadone, /* cbfcnp */ 211 CAM_DIR_NONE, /* flags */ 212 NULL, /* data_ptr */ 213 0, /* dxfer_len */ 214 nda_default_timeout * 1000); /* timeout 30s */ 215 nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid); 216 } 217 218 static void 219 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 220 void *payload, uint32_t num_ranges) 221 { 222 cam_fill_nvmeio(nvmeio, 223 0, /* retries */ 224 ndadone, /* cbfcnp */ 225 CAM_DIR_OUT, /* flags */ 226 payload, /* data_ptr */ 227 num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */ 228 nda_default_timeout * 1000); /* timeout 30s */ 229 nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges); 230 } 231 232 static void 233 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 234 void *payload, uint64_t lba, uint32_t len, uint32_t count) 235 { 236 cam_fill_nvmeio(nvmeio, 237 0, /* retries */ 238 ndadone, /* cbfcnp */ 239 CAM_DIR_OUT, /* flags */ 240 payload, /* data_ptr */ 241 len, /* dxfer_len */ 242 nda_default_timeout * 1000); /* timeout 30s */ 243 nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count); 244 } 245 246 static void 247 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 248 struct bio *bp, uint32_t rwcmd) 249 { 250 int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT; 251 void *payload; 252 uint64_t lba; 253 uint32_t count; 254 255 if (bp->bio_flags & BIO_UNMAPPED) { 256 flags |= CAM_DATA_BIO; 257 payload = bp; 258 } else { 259 payload = bp->bio_data; 260 } 261 262 lba = bp->bio_pblkno; 263 count = bp->bio_bcount / softc->disk->d_sectorsize; 264 265 cam_fill_nvmeio(nvmeio, 266 0, /* retries */ 267 ndadone, /* cbfcnp */ 268 flags, /* flags */ 269 payload, /* data_ptr */ 270 bp->bio_bcount, /* dxfer_len */ 271 nda_default_timeout * 1000); /* timeout 30s */ 272 nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count); 273 } 274 275 static int 276 ndaopen(struct disk *dp) 277 { 278 struct cam_periph *periph; 279 struct nda_softc *softc; 280 int error; 281 282 periph = (struct cam_periph *)dp->d_drv1; 283 if (cam_periph_acquire(periph) != 0) { 284 return(ENXIO); 285 } 286 287 cam_periph_lock(periph); 288 if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) { 289 cam_periph_unlock(periph); 290 cam_periph_release(periph); 291 return (error); 292 } 293 294 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 295 ("ndaopen\n")); 296 297 softc = (struct nda_softc *)periph->softc; 298 softc->flags |= NDA_FLAG_OPEN; 299 300 cam_periph_unhold(periph); 301 cam_periph_unlock(periph); 302 return (0); 303 } 304 305 static int 306 ndaclose(struct disk *dp) 307 { 308 struct cam_periph *periph; 309 struct nda_softc *softc; 310 union ccb *ccb; 311 int error; 312 313 periph = (struct cam_periph *)dp->d_drv1; 314 softc = (struct nda_softc *)periph->softc; 315 cam_periph_lock(periph); 316 317 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 318 ("ndaclose\n")); 319 320 if ((softc->flags & NDA_FLAG_DIRTY) != 0 && 321 (periph->flags & CAM_PERIPH_INVALID) == 0 && 322 cam_periph_hold(periph, PRIBIO) == 0) { 323 324 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 325 nda_nvme_flush(softc, &ccb->nvmeio); 326 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 327 /*sense_flags*/0, softc->disk->d_devstat); 328 329 if (error != 0) 330 xpt_print(periph->path, "Synchronize cache failed\n"); 331 else 332 softc->flags &= ~NDA_FLAG_DIRTY; 333 xpt_release_ccb(ccb); 334 cam_periph_unhold(periph); 335 } 336 337 softc->flags &= ~NDA_FLAG_OPEN; 338 339 while (softc->refcount != 0) 340 cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1); 341 KASSERT(softc->outstanding_cmds == 0, 342 ("nda %d outstanding commands", softc->outstanding_cmds)); 343 cam_periph_unlock(periph); 344 cam_periph_release(periph); 345 return (0); 346 } 347 348 static void 349 ndaschedule(struct cam_periph *periph) 350 { 351 struct nda_softc *softc = (struct nda_softc *)periph->softc; 352 353 if (softc->state != NDA_STATE_NORMAL) 354 return; 355 356 cam_iosched_schedule(softc->cam_iosched, periph); 357 } 358 359 /* 360 * Actually translate the requested transfer into one the physical driver 361 * can understand. The transfer is described by a buf and will include 362 * only one physical transfer. 363 */ 364 static void 365 ndastrategy(struct bio *bp) 366 { 367 struct cam_periph *periph; 368 struct nda_softc *softc; 369 370 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 371 softc = (struct nda_softc *)periph->softc; 372 373 cam_periph_lock(periph); 374 375 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp)); 376 377 /* 378 * If the device has been made invalid, error out 379 */ 380 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 381 cam_periph_unlock(periph); 382 biofinish(bp, NULL, ENXIO); 383 return; 384 } 385 386 if (bp->bio_cmd == BIO_DELETE) 387 softc->deletes++; 388 389 /* 390 * Place it in the queue of disk activities for this disk 391 */ 392 cam_iosched_queue_work(softc->cam_iosched, bp); 393 394 /* 395 * Schedule ourselves for performing the work. 396 */ 397 ndaschedule(periph); 398 cam_periph_unlock(periph); 399 400 return; 401 } 402 403 static int 404 ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length) 405 { 406 struct cam_periph *periph; 407 struct nda_softc *softc; 408 u_int secsize; 409 struct ccb_nvmeio nvmeio; 410 struct disk *dp; 411 uint64_t lba; 412 uint32_t count; 413 int error = 0; 414 415 dp = arg; 416 periph = dp->d_drv1; 417 softc = (struct nda_softc *)periph->softc; 418 secsize = softc->disk->d_sectorsize; 419 lba = offset / secsize; 420 count = length / secsize; 421 422 if ((periph->flags & CAM_PERIPH_INVALID) != 0) 423 return (ENXIO); 424 425 /* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */ 426 memset(&nvmeio, 0, sizeof(nvmeio)); 427 if (length > 0) { 428 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 429 nvmeio.ccb_state = NDA_CCB_DUMP; 430 nda_nvme_write(softc, &nvmeio, virtual, lba, length, count); 431 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 432 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 433 if (error != 0) 434 printf("Aborting dump due to I/O error %d.\n", error); 435 436 return (error); 437 } 438 439 /* Flush */ 440 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 441 442 nvmeio.ccb_state = NDA_CCB_DUMP; 443 nda_nvme_flush(softc, &nvmeio); 444 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 445 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 446 if (error != 0) 447 xpt_print(periph->path, "flush cmd failed\n"); 448 return (error); 449 } 450 451 static void 452 ndainit(void) 453 { 454 cam_status status; 455 456 /* 457 * Install a global async callback. This callback will 458 * receive async callbacks like "new device found". 459 */ 460 status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL); 461 462 if (status != CAM_REQ_CMP) { 463 printf("nda: Failed to attach master async callback " 464 "due to status 0x%x!\n", status); 465 } else if (nda_send_ordered) { 466 467 /* Register our event handlers */ 468 if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend, 469 NULL, EVENTHANDLER_PRI_LAST)) == NULL) 470 printf("ndainit: power event registration failed!\n"); 471 if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown, 472 NULL, SHUTDOWN_PRI_DEFAULT)) == NULL) 473 printf("ndainit: shutdown event registration failed!\n"); 474 } 475 } 476 477 /* 478 * Callback from GEOM, called when it has finished cleaning up its 479 * resources. 480 */ 481 static void 482 ndadiskgonecb(struct disk *dp) 483 { 484 struct cam_periph *periph; 485 486 periph = (struct cam_periph *)dp->d_drv1; 487 488 cam_periph_release(periph); 489 } 490 491 static void 492 ndaoninvalidate(struct cam_periph *periph) 493 { 494 struct nda_softc *softc; 495 496 softc = (struct nda_softc *)periph->softc; 497 498 /* 499 * De-register any async callbacks. 500 */ 501 xpt_register_async(0, ndaasync, periph, periph->path); 502 #ifdef CAM_IO_STATS 503 softc->invalidations++; 504 #endif 505 506 /* 507 * Return all queued I/O with ENXIO. 508 * XXX Handle any transactions queued to the card 509 * with XPT_ABORT_CCB. 510 */ 511 cam_iosched_flush(softc->cam_iosched, NULL, ENXIO); 512 513 disk_gone(softc->disk); 514 } 515 516 static void 517 ndacleanup(struct cam_periph *periph) 518 { 519 struct nda_softc *softc; 520 521 softc = (struct nda_softc *)periph->softc; 522 523 cam_periph_unlock(periph); 524 525 cam_iosched_fini(softc->cam_iosched); 526 527 /* 528 * If we can't free the sysctl tree, oh well... 529 */ 530 if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) { 531 #ifdef CAM_IO_STATS 532 if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0) 533 xpt_print(periph->path, 534 "can't remove sysctl stats context\n"); 535 #endif 536 if (sysctl_ctx_free(&softc->sysctl_ctx) != 0) 537 xpt_print(periph->path, 538 "can't remove sysctl context\n"); 539 } 540 541 disk_destroy(softc->disk); 542 free(softc, M_DEVBUF); 543 cam_periph_lock(periph); 544 } 545 546 static void 547 ndaasync(void *callback_arg, u_int32_t code, 548 struct cam_path *path, void *arg) 549 { 550 struct cam_periph *periph; 551 552 periph = (struct cam_periph *)callback_arg; 553 switch (code) { 554 case AC_FOUND_DEVICE: 555 { 556 struct ccb_getdev *cgd; 557 cam_status status; 558 559 cgd = (struct ccb_getdev *)arg; 560 if (cgd == NULL) 561 break; 562 563 if (cgd->protocol != PROTO_NVME) 564 break; 565 566 /* 567 * Allocate a peripheral instance for 568 * this device and start the probe 569 * process. 570 */ 571 status = cam_periph_alloc(ndaregister, ndaoninvalidate, 572 ndacleanup, ndastart, 573 "nda", CAM_PERIPH_BIO, 574 path, ndaasync, 575 AC_FOUND_DEVICE, cgd); 576 577 if (status != CAM_REQ_CMP 578 && status != CAM_REQ_INPROG) 579 printf("ndaasync: Unable to attach to new device " 580 "due to status 0x%x\n", status); 581 break; 582 } 583 case AC_ADVINFO_CHANGED: 584 { 585 uintptr_t buftype; 586 587 buftype = (uintptr_t)arg; 588 if (buftype == CDAI_TYPE_PHYS_PATH) { 589 struct nda_softc *softc; 590 591 softc = periph->softc; 592 disk_attr_changed(softc->disk, "GEOM::physpath", 593 M_NOWAIT); 594 } 595 break; 596 } 597 case AC_LOST_DEVICE: 598 default: 599 cam_periph_async(periph, code, path, arg); 600 break; 601 } 602 } 603 604 static void 605 ndasysctlinit(void *context, int pending) 606 { 607 struct cam_periph *periph; 608 struct nda_softc *softc; 609 char tmpstr[32], tmpstr2[16]; 610 611 periph = (struct cam_periph *)context; 612 613 /* periph was held for us when this task was enqueued */ 614 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 615 cam_periph_release(periph); 616 return; 617 } 618 619 softc = (struct nda_softc *)periph->softc; 620 snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number); 621 snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number); 622 623 sysctl_ctx_init(&softc->sysctl_ctx); 624 softc->flags |= NDA_FLAG_SCTX_INIT; 625 softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx, 626 SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2, 627 CTLFLAG_RD, 0, tmpstr, "device_index"); 628 if (softc->sysctl_tree == NULL) { 629 printf("ndasysctlinit: unable to allocate sysctl tree\n"); 630 cam_periph_release(periph); 631 return; 632 } 633 634 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 635 OID_AUTO, "unmapped_io", CTLFLAG_RD, 636 &softc->unmappedio, 0, "Unmapped I/O leaf"); 637 638 SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 639 OID_AUTO, "deletes", CTLFLAG_RD, 640 &softc->deletes, "Number of BIO_DELETE requests"); 641 642 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 643 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 644 "trim_count", CTLFLAG_RD, &softc->trim_count, 645 "Total number of unmap/dsm commands sent"); 646 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 647 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 648 "trim_ranges", CTLFLAG_RD, &softc->trim_ranges, 649 "Total number of ranges in unmap/dsm commands"); 650 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 651 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 652 "trim_lbas", CTLFLAG_RD, &softc->trim_lbas, 653 "Total lbas in the unmap/dsm commands sent"); 654 655 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 656 OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1, 657 "Rotating media"); 658 659 #ifdef CAM_IO_STATS 660 softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx, 661 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats", 662 CTLFLAG_RD, 0, "Statistics"); 663 if (softc->sysctl_stats_tree == NULL) { 664 printf("ndasysctlinit: unable to allocate sysctl tree for stats\n"); 665 cam_periph_release(periph); 666 return; 667 } 668 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 669 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 670 OID_AUTO, "timeouts", CTLFLAG_RD, 671 &softc->timeouts, 0, 672 "Device timeouts reported by the SIM"); 673 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 674 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 675 OID_AUTO, "errors", CTLFLAG_RD, 676 &softc->errors, 0, 677 "Transport errors reported by the SIM."); 678 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 679 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 680 OID_AUTO, "pack_invalidations", CTLFLAG_RD, 681 &softc->invalidations, 0, 682 "Device pack invalidations."); 683 #endif 684 685 #ifdef CAM_TEST_FAILURE 686 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 687 OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 688 periph, 0, cam_periph_invalidate_sysctl, "I", 689 "Write 1 to invalidate the drive immediately"); 690 #endif 691 692 cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx, 693 softc->sysctl_tree); 694 695 cam_periph_release(periph); 696 } 697 698 static int 699 ndagetattr(struct bio *bp) 700 { 701 int ret; 702 struct cam_periph *periph; 703 704 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 705 cam_periph_lock(periph); 706 ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute, 707 periph->path); 708 cam_periph_unlock(periph); 709 if (ret == 0) 710 bp->bio_completed = bp->bio_length; 711 return ret; 712 } 713 714 static cam_status 715 ndaregister(struct cam_periph *periph, void *arg) 716 { 717 struct nda_softc *softc; 718 struct disk *disk; 719 struct ccb_pathinq cpi; 720 const struct nvme_namespace_data *nsd; 721 const struct nvme_controller_data *cd; 722 char announce_buf[80]; 723 uint8_t flbas_fmt, lbads, vwc_present; 724 u_int maxio; 725 int quirks; 726 727 nsd = nvme_get_identify_ns(periph); 728 cd = nvme_get_identify_cntrl(periph); 729 730 softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF, 731 M_NOWAIT | M_ZERO); 732 733 if (softc == NULL) { 734 printf("ndaregister: Unable to probe new device. " 735 "Unable to allocate softc\n"); 736 return(CAM_REQ_CMP_ERR); 737 } 738 739 if (cam_iosched_init(&softc->cam_iosched, periph) != 0) { 740 printf("ndaregister: Unable to probe new device. " 741 "Unable to allocate iosched memory\n"); 742 free(softc, M_DEVBUF); 743 return(CAM_REQ_CMP_ERR); 744 } 745 746 /* ident_data parsing */ 747 748 periph->softc = softc; 749 750 softc->quirks = NDA_Q_NONE; 751 752 xpt_path_inq(&cpi, periph->path); 753 754 TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph); 755 756 /* 757 * The name space ID is the lun, save it for later I/O 758 */ 759 softc->nsid = (uint32_t)xpt_path_lun_id(periph->path); 760 761 /* 762 * Register this media as a disk 763 */ 764 (void)cam_periph_hold(periph, PRIBIO); 765 cam_periph_unlock(periph); 766 snprintf(announce_buf, sizeof(announce_buf), 767 "kern.cam.nda.%d.quirks", periph->unit_number); 768 quirks = softc->quirks; 769 TUNABLE_INT_FETCH(announce_buf, &quirks); 770 softc->quirks = quirks; 771 cam_iosched_set_sort_queue(softc->cam_iosched, 0); 772 softc->disk = disk = disk_alloc(); 773 strlcpy(softc->disk->d_descr, cd->mn, 774 MIN(sizeof(softc->disk->d_descr), sizeof(cd->mn))); 775 strlcpy(softc->disk->d_ident, cd->sn, 776 MIN(sizeof(softc->disk->d_ident), sizeof(cd->sn))); 777 disk->d_rotation_rate = DISK_RR_NON_ROTATING; 778 disk->d_open = ndaopen; 779 disk->d_close = ndaclose; 780 disk->d_strategy = ndastrategy; 781 disk->d_getattr = ndagetattr; 782 disk->d_dump = ndadump; 783 disk->d_gone = ndadiskgonecb; 784 disk->d_name = "nda"; 785 disk->d_drv1 = periph; 786 disk->d_unit = periph->unit_number; 787 maxio = cpi.maxio; /* Honor max I/O size of SIM */ 788 if (maxio == 0) 789 maxio = DFLTPHYS; /* traditional default */ 790 else if (maxio > MAXPHYS) 791 maxio = MAXPHYS; /* for safety */ 792 disk->d_maxsize = maxio; 793 flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) & 794 NVME_NS_DATA_FLBAS_FORMAT_MASK; 795 lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) & 796 NVME_NS_DATA_LBAF_LBADS_MASK; 797 disk->d_sectorsize = 1 << lbads; 798 disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze); 799 disk->d_delmaxsize = disk->d_mediasize; 800 disk->d_flags = DISKFLAG_DIRECT_COMPLETION; 801 if (nvme_ctrlr_has_dataset_mgmt(cd)) 802 disk->d_flags |= DISKFLAG_CANDELETE; 803 vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) & 804 NVME_CTRLR_DATA_VWC_PRESENT_MASK; 805 if (vwc_present) 806 disk->d_flags |= DISKFLAG_CANFLUSHCACHE; 807 if ((cpi.hba_misc & PIM_UNMAPPED) != 0) { 808 disk->d_flags |= DISKFLAG_UNMAPPED_BIO; 809 softc->unmappedio = 1; 810 } 811 /* 812 * d_ident and d_descr are both far bigger than the length of either 813 * the serial or model number strings. 814 */ 815 nvme_strvis(disk->d_descr, cd->mn, 816 sizeof(disk->d_descr), NVME_MODEL_NUMBER_LENGTH); 817 nvme_strvis(disk->d_ident, cd->sn, 818 sizeof(disk->d_ident), NVME_SERIAL_NUMBER_LENGTH); 819 disk->d_hba_vendor = cpi.hba_vendor; 820 disk->d_hba_device = cpi.hba_device; 821 disk->d_hba_subvendor = cpi.hba_subvendor; 822 disk->d_hba_subdevice = cpi.hba_subdevice; 823 disk->d_stripesize = disk->d_sectorsize; 824 disk->d_stripeoffset = 0; 825 disk->d_devstat = devstat_new_entry(periph->periph_name, 826 periph->unit_number, disk->d_sectorsize, 827 DEVSTAT_ALL_SUPPORTED, 828 DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport), 829 DEVSTAT_PRIORITY_DISK); 830 /* 831 * Add alias for older nvd drives to ease transition. 832 */ 833 /* disk_add_alias(disk, "nvd"); Have reports of this causing problems */ 834 835 /* 836 * Acquire a reference to the periph before we register with GEOM. 837 * We'll release this reference once GEOM calls us back (via 838 * ndadiskgonecb()) telling us that our provider has been freed. 839 */ 840 if (cam_periph_acquire(periph) != 0) { 841 xpt_print(periph->path, "%s: lost periph during " 842 "registration!\n", __func__); 843 cam_periph_lock(periph); 844 return (CAM_REQ_CMP_ERR); 845 } 846 disk_create(softc->disk, DISK_VERSION); 847 cam_periph_lock(periph); 848 cam_periph_unhold(periph); 849 850 snprintf(announce_buf, sizeof(announce_buf), 851 "%juMB (%ju %u byte sectors)", 852 (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)), 853 (uintmax_t)disk->d_mediasize / disk->d_sectorsize, 854 disk->d_sectorsize); 855 xpt_announce_periph(periph, announce_buf); 856 xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING); 857 858 /* 859 * Create our sysctl variables, now that we know 860 * we have successfully attached. 861 */ 862 if (cam_periph_acquire(periph) == 0) 863 taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task); 864 865 /* 866 * Register for device going away and info about the drive 867 * changing (though with NVMe, it can't) 868 */ 869 xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED, 870 ndaasync, periph, periph->path); 871 872 softc->state = NDA_STATE_NORMAL; 873 return(CAM_REQ_CMP); 874 } 875 876 static void 877 ndastart(struct cam_periph *periph, union ccb *start_ccb) 878 { 879 struct nda_softc *softc = (struct nda_softc *)periph->softc; 880 struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio; 881 882 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n")); 883 884 switch (softc->state) { 885 case NDA_STATE_NORMAL: 886 { 887 struct bio *bp; 888 889 bp = cam_iosched_next_bio(softc->cam_iosched); 890 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp)); 891 if (bp == NULL) { 892 xpt_release_ccb(start_ccb); 893 break; 894 } 895 896 switch (bp->bio_cmd) { 897 case BIO_WRITE: 898 softc->flags |= NDA_FLAG_DIRTY; 899 /* FALLTHROUGH */ 900 case BIO_READ: 901 { 902 #ifdef CAM_TEST_FAILURE 903 int fail = 0; 904 905 /* 906 * Support the failure ioctls. If the command is a 907 * read, and there are pending forced read errors, or 908 * if a write and pending write errors, then fail this 909 * operation with EIO. This is useful for testing 910 * purposes. Also, support having every Nth read fail. 911 * 912 * This is a rather blunt tool. 913 */ 914 if (bp->bio_cmd == BIO_READ) { 915 if (softc->force_read_error) { 916 softc->force_read_error--; 917 fail = 1; 918 } 919 if (softc->periodic_read_error > 0) { 920 if (++softc->periodic_read_count >= 921 softc->periodic_read_error) { 922 softc->periodic_read_count = 0; 923 fail = 1; 924 } 925 } 926 } else { 927 if (softc->force_write_error) { 928 softc->force_write_error--; 929 fail = 1; 930 } 931 } 932 if (fail) { 933 biofinish(bp, NULL, EIO); 934 xpt_release_ccb(start_ccb); 935 ndaschedule(periph); 936 return; 937 } 938 #endif 939 KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 || 940 round_page(bp->bio_bcount + bp->bio_ma_offset) / 941 PAGE_SIZE == bp->bio_ma_n, 942 ("Short bio %p", bp)); 943 nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ? 944 NVME_OPC_READ : NVME_OPC_WRITE); 945 break; 946 } 947 case BIO_DELETE: 948 { 949 struct nvme_dsm_range *dsm_range, *dsm_end; 950 struct nda_trim_request *trim; 951 struct bio *bp1; 952 int ents; 953 uint32_t totalcount = 0, ranges = 0; 954 955 trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT); 956 if (trim == NULL) { 957 biofinish(bp, NULL, ENOMEM); 958 xpt_release_ccb(start_ccb); 959 ndaschedule(periph); 960 return; 961 } 962 TAILQ_INIT(&trim->bps); 963 bp1 = bp; 964 ents = sizeof(trim->data) / sizeof(struct nvme_dsm_range); 965 ents = min(ents, nda_max_trim_entries); 966 dsm_range = &trim->dsm; 967 dsm_end = dsm_range + ents; 968 do { 969 TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue); 970 dsm_range->length = 971 htole32(bp1->bio_bcount / softc->disk->d_sectorsize); 972 dsm_range->starting_lba = 973 htole64(bp1->bio_offset / softc->disk->d_sectorsize); 974 ranges++; 975 totalcount += dsm_range->length; 976 dsm_range++; 977 if (dsm_range >= dsm_end) 978 break; 979 bp1 = cam_iosched_next_trim(softc->cam_iosched); 980 /* XXX -- Could collapse adjacent ranges, but we don't for now */ 981 /* XXX -- Could limit based on total payload size */ 982 } while (bp1 != NULL); 983 start_ccb->ccb_trim = trim; 984 nda_nvme_trim(softc, &start_ccb->nvmeio, &trim->dsm, 985 dsm_range - &trim->dsm); 986 start_ccb->ccb_state = NDA_CCB_TRIM; 987 softc->trim_count++; 988 softc->trim_ranges += ranges; 989 softc->trim_lbas += totalcount; 990 /* 991 * Note: We can have multiple TRIMs in flight, so we don't call 992 * cam_iosched_submit_trim(softc->cam_iosched); 993 * since that forces the I/O scheduler to only schedule one at a time. 994 * On NVMe drives, this is a performance disaster. 995 */ 996 goto out; 997 } 998 case BIO_FLUSH: 999 nda_nvme_flush(softc, nvmeio); 1000 break; 1001 } 1002 start_ccb->ccb_state = NDA_CCB_BUFFER_IO; 1003 start_ccb->ccb_bp = bp; 1004 out: 1005 start_ccb->ccb_h.flags |= CAM_UNLOCKED; 1006 softc->outstanding_cmds++; 1007 softc->refcount++; /* For submission only */ 1008 cam_periph_unlock(periph); 1009 xpt_action(start_ccb); 1010 cam_periph_lock(periph); 1011 softc->refcount--; /* Submission done */ 1012 1013 /* May have more work to do, so ensure we stay scheduled */ 1014 ndaschedule(periph); 1015 break; 1016 } 1017 } 1018 } 1019 1020 static void 1021 ndadone(struct cam_periph *periph, union ccb *done_ccb) 1022 { 1023 struct nda_softc *softc; 1024 struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio; 1025 struct cam_path *path; 1026 int state; 1027 1028 softc = (struct nda_softc *)periph->softc; 1029 path = done_ccb->ccb_h.path; 1030 1031 CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n")); 1032 1033 state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK; 1034 switch (state) { 1035 case NDA_CCB_BUFFER_IO: 1036 case NDA_CCB_TRIM: 1037 { 1038 int error; 1039 1040 cam_periph_lock(periph); 1041 if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) { 1042 error = ndaerror(done_ccb, 0, 0); 1043 if (error == ERESTART) { 1044 /* A retry was scheduled, so just return. */ 1045 cam_periph_unlock(periph); 1046 return; 1047 } 1048 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1049 cam_release_devq(path, 1050 /*relsim_flags*/0, 1051 /*reduction*/0, 1052 /*timeout*/0, 1053 /*getcount_only*/0); 1054 } else { 1055 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1056 panic("REQ_CMP with QFRZN"); 1057 error = 0; 1058 } 1059 if (state == NDA_CCB_BUFFER_IO) { 1060 struct bio *bp; 1061 1062 bp = (struct bio *)done_ccb->ccb_bp; 1063 bp->bio_error = error; 1064 if (error != 0) { 1065 bp->bio_resid = bp->bio_bcount; 1066 bp->bio_flags |= BIO_ERROR; 1067 } else { 1068 bp->bio_resid = 0; 1069 } 1070 softc->outstanding_cmds--; 1071 1072 /* 1073 * We need to call cam_iosched before we call biodone so that we 1074 * don't measure any activity that happens in the completion 1075 * routine, which in the case of sendfile can be quite 1076 * extensive. 1077 */ 1078 cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb); 1079 xpt_release_ccb(done_ccb); 1080 ndaschedule(periph); 1081 cam_periph_unlock(periph); 1082 biodone(bp); 1083 } else { /* state == NDA_CCB_TRIM */ 1084 struct nda_trim_request *trim; 1085 struct bio *bp1, *bp2; 1086 TAILQ_HEAD(, bio) queue; 1087 1088 trim = nvmeio->ccb_trim; 1089 TAILQ_INIT(&queue); 1090 TAILQ_CONCAT(&queue, &trim->bps, bio_queue); 1091 free(trim, M_NVMEDA); 1092 1093 /* 1094 * Since we can have multiple trims in flight, we don't 1095 * need to call this here. 1096 * cam_iosched_trim_done(softc->cam_iosched); 1097 */ 1098 /* 1099 * The the I/O scheduler that we're finishing the I/O 1100 * so we can keep book. The first one we pass in the CCB 1101 * which has the timing information. The rest we pass in NULL 1102 * so we can keep proper counts. 1103 */ 1104 bp1 = TAILQ_FIRST(&queue); 1105 cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb); 1106 xpt_release_ccb(done_ccb); 1107 softc->outstanding_cmds--; 1108 ndaschedule(periph); 1109 cam_periph_unlock(periph); 1110 while ((bp2 = TAILQ_FIRST(&queue)) != NULL) { 1111 TAILQ_REMOVE(&queue, bp2, bio_queue); 1112 bp2->bio_error = error; 1113 if (error != 0) { 1114 bp2->bio_flags |= BIO_ERROR; 1115 bp2->bio_resid = bp1->bio_bcount; 1116 } else 1117 bp2->bio_resid = 0; 1118 if (bp1 != bp2) 1119 cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL); 1120 biodone(bp2); 1121 } 1122 } 1123 return; 1124 } 1125 case NDA_CCB_DUMP: 1126 /* No-op. We're polling */ 1127 return; 1128 default: 1129 break; 1130 } 1131 xpt_release_ccb(done_ccb); 1132 } 1133 1134 static int 1135 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags) 1136 { 1137 struct nda_softc *softc; 1138 struct cam_periph *periph; 1139 1140 periph = xpt_path_periph(ccb->ccb_h.path); 1141 softc = (struct nda_softc *)periph->softc; 1142 1143 switch (ccb->ccb_h.status & CAM_STATUS_MASK) { 1144 case CAM_CMD_TIMEOUT: 1145 #ifdef CAM_IO_STATS 1146 softc->timeouts++; 1147 #endif 1148 break; 1149 case CAM_REQ_ABORTED: 1150 case CAM_REQ_CMP_ERR: 1151 case CAM_REQ_TERMIO: 1152 case CAM_UNREC_HBA_ERROR: 1153 case CAM_DATA_RUN_ERR: 1154 case CAM_ATA_STATUS_ERROR: 1155 #ifdef CAM_IO_STATS 1156 softc->errors++; 1157 #endif 1158 break; 1159 default: 1160 break; 1161 } 1162 1163 return(cam_periph_error(ccb, cam_flags, sense_flags)); 1164 } 1165 1166 /* 1167 * Step through all NDA peripheral drivers, and if the device is still open, 1168 * sync the disk cache to physical media. 1169 */ 1170 static void 1171 ndaflush(void) 1172 { 1173 struct cam_periph *periph; 1174 struct nda_softc *softc; 1175 union ccb *ccb; 1176 int error; 1177 1178 CAM_PERIPH_FOREACH(periph, &ndadriver) { 1179 softc = (struct nda_softc *)periph->softc; 1180 1181 if (SCHEDULER_STOPPED()) { 1182 /* 1183 * If we paniced with the lock held or the periph is not 1184 * open, do not recurse. Otherwise, call ndadump since 1185 * that avoids the sleeping cam_periph_getccb does if no 1186 * CCBs are available. 1187 */ 1188 if (!cam_periph_owned(periph) && 1189 (softc->flags & NDA_FLAG_OPEN)) { 1190 ndadump(softc->disk, NULL, 0, 0, 0); 1191 } 1192 continue; 1193 } 1194 1195 /* 1196 * We only sync the cache if the drive is still open 1197 */ 1198 cam_periph_lock(periph); 1199 if ((softc->flags & NDA_FLAG_OPEN) == 0) { 1200 cam_periph_unlock(periph); 1201 continue; 1202 } 1203 1204 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 1205 nda_nvme_flush(softc, &ccb->nvmeio); 1206 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 1207 /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY, 1208 softc->disk->d_devstat); 1209 if (error != 0) 1210 xpt_print(periph->path, "Synchronize cache failed\n"); 1211 xpt_release_ccb(ccb); 1212 cam_periph_unlock(periph); 1213 } 1214 } 1215 1216 static void 1217 ndashutdown(void *arg, int howto) 1218 { 1219 1220 ndaflush(); 1221 } 1222 1223 static void 1224 ndasuspend(void *arg) 1225 { 1226 1227 ndaflush(); 1228 } 1229