1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * 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 <sys/sbuf.h> 52 #include <geom/geom.h> 53 #include <geom/geom_disk.h> 54 #endif /* _KERNEL */ 55 56 #ifndef _KERNEL 57 #include <stdio.h> 58 #include <string.h> 59 #endif /* _KERNEL */ 60 61 #include <cam/cam.h> 62 #include <cam/cam_ccb.h> 63 #include <cam/cam_periph.h> 64 #include <cam/cam_xpt_periph.h> 65 #include <cam/cam_sim.h> 66 #include <cam/cam_iosched.h> 67 68 #include <cam/nvme/nvme_all.h> 69 70 typedef enum { 71 NDA_STATE_NORMAL 72 } nda_state; 73 74 typedef enum { 75 NDA_FLAG_OPEN = 0x0001, 76 NDA_FLAG_DIRTY = 0x0002, 77 NDA_FLAG_SCTX_INIT = 0x0004, 78 } nda_flags; 79 #define NDA_FLAG_STRING \ 80 "\020" \ 81 "\001OPEN" \ 82 "\002DIRTY" \ 83 "\003SCTX_INIT" 84 85 typedef enum { 86 NDA_Q_4K = 0x01, 87 NDA_Q_NONE = 0x00, 88 } nda_quirks; 89 90 #define NDA_Q_BIT_STRING \ 91 "\020" \ 92 "\001Bit 0" 93 94 typedef enum { 95 NDA_CCB_BUFFER_IO = 0x01, 96 NDA_CCB_DUMP = 0x02, 97 NDA_CCB_TRIM = 0x03, 98 NDA_CCB_PASS = 0x04, 99 NDA_CCB_TYPE_MASK = 0x0F, 100 } nda_ccb_state; 101 102 /* Offsets into our private area for storing information */ 103 #define ccb_state ccb_h.ppriv_field0 104 #define ccb_bp ccb_h.ppriv_ptr1 /* For NDA_CCB_BUFFER_IO */ 105 #define ccb_trim ccb_h.ppriv_ptr1 /* For NDA_CCB_TRIM */ 106 107 struct nda_softc { 108 struct cam_iosched_softc *cam_iosched; 109 int outstanding_cmds; /* Number of active commands */ 110 int refcount; /* Active xpt_action() calls */ 111 nda_state state; 112 nda_flags flags; 113 nda_quirks quirks; 114 int unmappedio; 115 quad_t deletes; 116 uint32_t nsid; /* Namespace ID for this nda device */ 117 struct disk *disk; 118 struct task sysctl_task; 119 struct sysctl_ctx_list sysctl_ctx; 120 struct sysctl_oid *sysctl_tree; 121 uint64_t trim_count; 122 uint64_t trim_ranges; 123 uint64_t trim_lbas; 124 #ifdef CAM_TEST_FAILURE 125 int force_read_error; 126 int force_write_error; 127 int periodic_read_error; 128 int periodic_read_count; 129 #endif 130 #ifdef CAM_IO_STATS 131 struct sysctl_ctx_list sysctl_stats_ctx; 132 struct sysctl_oid *sysctl_stats_tree; 133 u_int timeouts; 134 u_int errors; 135 u_int invalidations; 136 #endif 137 }; 138 139 struct nda_trim_request { 140 struct nvme_dsm_range dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)]; 141 TAILQ_HEAD(, bio) bps; 142 }; 143 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0, 144 "NVME_MAX_DSM_TRIM must be an integral number of ranges"); 145 146 /* Need quirk table */ 147 148 static disk_ioctl_t ndaioctl; 149 static disk_strategy_t ndastrategy; 150 static dumper_t ndadump; 151 static periph_init_t ndainit; 152 static void ndaasync(void *callback_arg, u_int32_t code, 153 struct cam_path *path, void *arg); 154 static void ndasysctlinit(void *context, int pending); 155 static int ndaflagssysctl(SYSCTL_HANDLER_ARGS); 156 static periph_ctor_t ndaregister; 157 static periph_dtor_t ndacleanup; 158 static periph_start_t ndastart; 159 static periph_oninv_t ndaoninvalidate; 160 static void ndadone(struct cam_periph *periph, 161 union ccb *done_ccb); 162 static int ndaerror(union ccb *ccb, u_int32_t cam_flags, 163 u_int32_t sense_flags); 164 static void ndashutdown(void *arg, int howto); 165 static void ndasuspend(void *arg); 166 167 #ifndef NDA_DEFAULT_SEND_ORDERED 168 #define NDA_DEFAULT_SEND_ORDERED 1 169 #endif 170 #ifndef NDA_DEFAULT_TIMEOUT 171 #define NDA_DEFAULT_TIMEOUT 30 /* Timeout in seconds */ 172 #endif 173 #ifndef NDA_DEFAULT_RETRY 174 #define NDA_DEFAULT_RETRY 4 175 #endif 176 #ifndef NDA_MAX_TRIM_ENTRIES 177 #define NDA_MAX_TRIM_ENTRIES (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */ 178 #endif 179 180 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 181 "CAM Direct Access Disk driver"); 182 183 //static int nda_retry_count = NDA_DEFAULT_RETRY; 184 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED; 185 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT; 186 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES; 187 static int nda_enable_biospeedup = 1; 188 static int nda_nvd_compat = 1; 189 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN, 190 &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES, 191 "Maximum number of BIO_DELETE to send down as a DSM TRIM."); 192 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN, 193 &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing."); 194 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN, 195 &nda_nvd_compat, 1, "Enable creation of nvd aliases."); 196 197 /* 198 * All NVMe media is non-rotational, so all nvme device instances 199 * share this to implement the sysctl. 200 */ 201 static int nda_rotating_media = 0; 202 203 static struct periph_driver ndadriver = 204 { 205 ndainit, "nda", 206 TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0 207 }; 208 209 PERIPHDRIVER_DECLARE(nda, ndadriver); 210 211 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers"); 212 213 /* 214 * nice wrappers. Maybe these belong in nvme_all.c instead of 215 * here, but this is the only place that uses these. Should 216 * we ever grow another NVME periph, we should move them 217 * all there wholesale. 218 */ 219 220 static void 221 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio) 222 { 223 cam_fill_nvmeio(nvmeio, 224 0, /* retries */ 225 ndadone, /* cbfcnp */ 226 CAM_DIR_NONE, /* flags */ 227 NULL, /* data_ptr */ 228 0, /* dxfer_len */ 229 nda_default_timeout * 1000); /* timeout 30s */ 230 nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid); 231 } 232 233 static void 234 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 235 void *payload, uint32_t num_ranges) 236 { 237 cam_fill_nvmeio(nvmeio, 238 0, /* retries */ 239 ndadone, /* cbfcnp */ 240 CAM_DIR_OUT, /* flags */ 241 payload, /* data_ptr */ 242 num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */ 243 nda_default_timeout * 1000); /* timeout 30s */ 244 nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges); 245 } 246 247 static void 248 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 249 void *payload, uint64_t lba, uint32_t len, uint32_t count) 250 { 251 cam_fill_nvmeio(nvmeio, 252 0, /* retries */ 253 ndadone, /* cbfcnp */ 254 CAM_DIR_OUT, /* flags */ 255 payload, /* data_ptr */ 256 len, /* dxfer_len */ 257 nda_default_timeout * 1000); /* timeout 30s */ 258 nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count); 259 } 260 261 static void 262 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio, 263 struct bio *bp, uint32_t rwcmd) 264 { 265 int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT; 266 void *payload; 267 uint64_t lba; 268 uint32_t count; 269 270 if (bp->bio_flags & BIO_UNMAPPED) { 271 flags |= CAM_DATA_BIO; 272 payload = bp; 273 } else { 274 payload = bp->bio_data; 275 } 276 277 lba = bp->bio_pblkno; 278 count = bp->bio_bcount / softc->disk->d_sectorsize; 279 280 cam_fill_nvmeio(nvmeio, 281 0, /* retries */ 282 ndadone, /* cbfcnp */ 283 flags, /* flags */ 284 payload, /* data_ptr */ 285 bp->bio_bcount, /* dxfer_len */ 286 nda_default_timeout * 1000); /* timeout 30s */ 287 nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count); 288 } 289 290 static int 291 ndaopen(struct disk *dp) 292 { 293 struct cam_periph *periph; 294 struct nda_softc *softc; 295 int error; 296 297 periph = (struct cam_periph *)dp->d_drv1; 298 if (cam_periph_acquire(periph) != 0) { 299 return(ENXIO); 300 } 301 302 cam_periph_lock(periph); 303 if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) { 304 cam_periph_unlock(periph); 305 cam_periph_release(periph); 306 return (error); 307 } 308 309 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 310 ("ndaopen\n")); 311 312 softc = (struct nda_softc *)periph->softc; 313 softc->flags |= NDA_FLAG_OPEN; 314 315 cam_periph_unhold(periph); 316 cam_periph_unlock(periph); 317 return (0); 318 } 319 320 static int 321 ndaclose(struct disk *dp) 322 { 323 struct cam_periph *periph; 324 struct nda_softc *softc; 325 union ccb *ccb; 326 int error; 327 328 periph = (struct cam_periph *)dp->d_drv1; 329 softc = (struct nda_softc *)periph->softc; 330 cam_periph_lock(periph); 331 332 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH, 333 ("ndaclose\n")); 334 335 if ((softc->flags & NDA_FLAG_DIRTY) != 0 && 336 (periph->flags & CAM_PERIPH_INVALID) == 0 && 337 cam_periph_hold(periph, PRIBIO) == 0) { 338 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 339 nda_nvme_flush(softc, &ccb->nvmeio); 340 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 341 /*sense_flags*/0, softc->disk->d_devstat); 342 343 if (error != 0) 344 xpt_print(periph->path, "Synchronize cache failed\n"); 345 else 346 softc->flags &= ~NDA_FLAG_DIRTY; 347 xpt_release_ccb(ccb); 348 cam_periph_unhold(periph); 349 } 350 351 softc->flags &= ~NDA_FLAG_OPEN; 352 353 while (softc->refcount != 0) 354 cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1); 355 KASSERT(softc->outstanding_cmds == 0, 356 ("nda %d outstanding commands", softc->outstanding_cmds)); 357 cam_periph_unlock(periph); 358 cam_periph_release(periph); 359 return (0); 360 } 361 362 static void 363 ndaschedule(struct cam_periph *periph) 364 { 365 struct nda_softc *softc = (struct nda_softc *)periph->softc; 366 367 if (softc->state != NDA_STATE_NORMAL) 368 return; 369 370 cam_iosched_schedule(softc->cam_iosched, periph); 371 } 372 373 static int 374 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag, 375 struct thread *td) 376 { 377 struct cam_periph *periph; 378 379 periph = (struct cam_periph *)dp->d_drv1; 380 381 switch (cmd) { 382 case NVME_IO_TEST: 383 case NVME_BIO_TEST: 384 /* 385 * These don't map well to the underlying CCBs, so 386 * they are usupported via CAM. 387 */ 388 return (ENOTTY); 389 case NVME_GET_NSID: 390 { 391 struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data; 392 struct ccb_pathinq cpi; 393 394 xpt_path_inq(&cpi, periph->path); 395 strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name, 396 sizeof(gnsid->cdev)); 397 gnsid->nsid = cpi.xport_specific.nvme.nsid; 398 return (0); 399 } 400 case NVME_PASSTHROUGH_CMD: 401 { 402 struct nvme_pt_command *pt; 403 union ccb *ccb; 404 struct cam_periph_map_info mapinfo; 405 u_int maxmap = dp->d_maxsize; 406 int error; 407 408 /* 409 * Create a NVME_IO CCB to do the passthrough command. 410 */ 411 pt = (struct nvme_pt_command *)data; 412 ccb = xpt_alloc_ccb(); 413 xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL); 414 ccb->ccb_state = NDA_CCB_PASS; 415 cam_fill_nvmeio(&ccb->nvmeio, 416 0, /* Retries */ 417 ndadone, 418 (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR, 419 pt->buf, 420 pt->len, 421 nda_default_timeout * 1000); 422 memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd)); 423 424 /* 425 * Wire the user memory in this request for the I/O 426 */ 427 memset(&mapinfo, 0, sizeof(mapinfo)); 428 error = cam_periph_mapmem(ccb, &mapinfo, maxmap); 429 if (error) 430 goto out; 431 432 /* 433 * Lock the periph and run the command. 434 */ 435 cam_periph_lock(periph); 436 cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO, 437 SF_RETRY_UA | SF_NO_PRINT, NULL); 438 439 /* 440 * Tear down mapping and return status. 441 */ 442 cam_periph_unlock(periph); 443 cam_periph_unmapmem(ccb, &mapinfo); 444 error = (ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP ? 445 0 : EIO; 446 out: 447 cam_periph_lock(periph); 448 xpt_release_ccb(ccb); 449 cam_periph_unlock(periph); 450 return (error); 451 } 452 default: 453 break; 454 } 455 return (ENOTTY); 456 } 457 458 /* 459 * Actually translate the requested transfer into one the physical driver 460 * can understand. The transfer is described by a buf and will include 461 * only one physical transfer. 462 */ 463 static void 464 ndastrategy(struct bio *bp) 465 { 466 struct cam_periph *periph; 467 struct nda_softc *softc; 468 469 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 470 softc = (struct nda_softc *)periph->softc; 471 472 cam_periph_lock(periph); 473 474 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp)); 475 476 /* 477 * If the device has been made invalid, error out 478 */ 479 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 480 cam_periph_unlock(periph); 481 biofinish(bp, NULL, ENXIO); 482 return; 483 } 484 485 if (bp->bio_cmd == BIO_DELETE) 486 softc->deletes++; 487 488 /* 489 * Place it in the queue of disk activities for this disk 490 */ 491 cam_iosched_queue_work(softc->cam_iosched, bp); 492 493 /* 494 * Schedule ourselves for performing the work. 495 */ 496 ndaschedule(periph); 497 cam_periph_unlock(periph); 498 499 return; 500 } 501 502 static int 503 ndadump(void *arg, void *virtual, off_t offset, size_t length) 504 { 505 struct cam_periph *periph; 506 struct nda_softc *softc; 507 u_int secsize; 508 struct ccb_nvmeio nvmeio; 509 struct disk *dp; 510 uint64_t lba; 511 uint32_t count; 512 int error = 0; 513 514 dp = arg; 515 periph = dp->d_drv1; 516 softc = (struct nda_softc *)periph->softc; 517 secsize = softc->disk->d_sectorsize; 518 lba = offset / secsize; 519 count = length / secsize; 520 521 if ((periph->flags & CAM_PERIPH_INVALID) != 0) 522 return (ENXIO); 523 524 /* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */ 525 memset(&nvmeio, 0, sizeof(nvmeio)); 526 if (length > 0) { 527 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 528 nvmeio.ccb_state = NDA_CCB_DUMP; 529 nda_nvme_write(softc, &nvmeio, virtual, lba, length, count); 530 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 531 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 532 if (error != 0) 533 printf("Aborting dump due to I/O error %d.\n", error); 534 535 return (error); 536 } 537 538 /* Flush */ 539 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL); 540 541 nvmeio.ccb_state = NDA_CCB_DUMP; 542 nda_nvme_flush(softc, &nvmeio); 543 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error, 544 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL); 545 if (error != 0) 546 xpt_print(periph->path, "flush cmd failed\n"); 547 return (error); 548 } 549 550 static void 551 ndainit(void) 552 { 553 cam_status status; 554 555 /* 556 * Install a global async callback. This callback will 557 * receive async callbacks like "new device found". 558 */ 559 status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL); 560 561 if (status != CAM_REQ_CMP) { 562 printf("nda: Failed to attach master async callback " 563 "due to status 0x%x!\n", status); 564 } else if (nda_send_ordered) { 565 /* Register our event handlers */ 566 if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend, 567 NULL, EVENTHANDLER_PRI_LAST)) == NULL) 568 printf("ndainit: power event registration failed!\n"); 569 if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown, 570 NULL, SHUTDOWN_PRI_DEFAULT)) == NULL) 571 printf("ndainit: shutdown event registration failed!\n"); 572 } 573 } 574 575 /* 576 * Callback from GEOM, called when it has finished cleaning up its 577 * resources. 578 */ 579 static void 580 ndadiskgonecb(struct disk *dp) 581 { 582 struct cam_periph *periph; 583 584 periph = (struct cam_periph *)dp->d_drv1; 585 586 cam_periph_release(periph); 587 } 588 589 static void 590 ndaoninvalidate(struct cam_periph *periph) 591 { 592 struct nda_softc *softc; 593 594 softc = (struct nda_softc *)periph->softc; 595 596 /* 597 * De-register any async callbacks. 598 */ 599 xpt_register_async(0, ndaasync, periph, periph->path); 600 #ifdef CAM_IO_STATS 601 softc->invalidations++; 602 #endif 603 604 /* 605 * Return all queued I/O with ENXIO. Transactions may be queued up here 606 * for retry (since we are called while there's other transactions 607 * pending). Any requests in the hardware will drain before ndacleanup 608 * is called. 609 */ 610 cam_iosched_flush(softc->cam_iosched, NULL, ENXIO); 611 612 /* 613 * Tell GEOM that we've gone away, we'll get a callback when it is 614 * done cleaning up its resources. 615 */ 616 disk_gone(softc->disk); 617 } 618 619 static void 620 ndacleanup(struct cam_periph *periph) 621 { 622 struct nda_softc *softc; 623 624 softc = (struct nda_softc *)periph->softc; 625 626 cam_periph_unlock(periph); 627 628 cam_iosched_fini(softc->cam_iosched); 629 630 /* 631 * If we can't free the sysctl tree, oh well... 632 */ 633 if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) { 634 #ifdef CAM_IO_STATS 635 if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0) 636 xpt_print(periph->path, 637 "can't remove sysctl stats context\n"); 638 #endif 639 if (sysctl_ctx_free(&softc->sysctl_ctx) != 0) 640 xpt_print(periph->path, 641 "can't remove sysctl context\n"); 642 } 643 644 disk_destroy(softc->disk); 645 free(softc, M_DEVBUF); 646 cam_periph_lock(periph); 647 } 648 649 static void 650 ndaasync(void *callback_arg, u_int32_t code, 651 struct cam_path *path, void *arg) 652 { 653 struct cam_periph *periph; 654 655 periph = (struct cam_periph *)callback_arg; 656 switch (code) { 657 case AC_FOUND_DEVICE: 658 { 659 struct ccb_getdev *cgd; 660 cam_status status; 661 662 cgd = (struct ccb_getdev *)arg; 663 if (cgd == NULL) 664 break; 665 666 if (cgd->protocol != PROTO_NVME) 667 break; 668 669 /* 670 * Allocate a peripheral instance for 671 * this device and start the probe 672 * process. 673 */ 674 status = cam_periph_alloc(ndaregister, ndaoninvalidate, 675 ndacleanup, ndastart, 676 "nda", CAM_PERIPH_BIO, 677 path, ndaasync, 678 AC_FOUND_DEVICE, cgd); 679 680 if (status != CAM_REQ_CMP 681 && status != CAM_REQ_INPROG) 682 printf("ndaasync: Unable to attach to new device " 683 "due to status 0x%x\n", status); 684 break; 685 } 686 case AC_ADVINFO_CHANGED: 687 { 688 uintptr_t buftype; 689 690 buftype = (uintptr_t)arg; 691 if (buftype == CDAI_TYPE_PHYS_PATH) { 692 struct nda_softc *softc; 693 694 softc = periph->softc; 695 disk_attr_changed(softc->disk, "GEOM::physpath", 696 M_NOWAIT); 697 } 698 break; 699 } 700 case AC_LOST_DEVICE: 701 default: 702 break; 703 } 704 cam_periph_async(periph, code, path, arg); 705 } 706 707 static void 708 ndasysctlinit(void *context, int pending) 709 { 710 struct cam_periph *periph; 711 struct nda_softc *softc; 712 char tmpstr[32], tmpstr2[16]; 713 714 periph = (struct cam_periph *)context; 715 716 /* periph was held for us when this task was enqueued */ 717 if ((periph->flags & CAM_PERIPH_INVALID) != 0) { 718 cam_periph_release(periph); 719 return; 720 } 721 722 softc = (struct nda_softc *)periph->softc; 723 snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number); 724 snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number); 725 726 sysctl_ctx_init(&softc->sysctl_ctx); 727 softc->flags |= NDA_FLAG_SCTX_INIT; 728 softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx, 729 SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2, 730 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index"); 731 if (softc->sysctl_tree == NULL) { 732 printf("ndasysctlinit: unable to allocate sysctl tree\n"); 733 cam_periph_release(periph); 734 return; 735 } 736 737 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 738 OID_AUTO, "unmapped_io", CTLFLAG_RD, 739 &softc->unmappedio, 0, "Unmapped I/O leaf"); 740 741 SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 742 OID_AUTO, "deletes", CTLFLAG_RD, 743 &softc->deletes, "Number of BIO_DELETE requests"); 744 745 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 746 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 747 "trim_count", CTLFLAG_RD, &softc->trim_count, 748 "Total number of unmap/dsm commands sent"); 749 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 750 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 751 "trim_ranges", CTLFLAG_RD, &softc->trim_ranges, 752 "Total number of ranges in unmap/dsm commands"); 753 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx, 754 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, 755 "trim_lbas", CTLFLAG_RD, &softc->trim_lbas, 756 "Total lbas in the unmap/dsm commands sent"); 757 758 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 759 OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1, 760 "Rotating media"); 761 762 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 763 OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 764 softc, 0, ndaflagssysctl, "A", 765 "Flags for drive"); 766 767 #ifdef CAM_IO_STATS 768 softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx, 769 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats", 770 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics"); 771 if (softc->sysctl_stats_tree == NULL) { 772 printf("ndasysctlinit: unable to allocate sysctl tree for stats\n"); 773 cam_periph_release(periph); 774 return; 775 } 776 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 777 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 778 OID_AUTO, "timeouts", CTLFLAG_RD, 779 &softc->timeouts, 0, 780 "Device timeouts reported by the SIM"); 781 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 782 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 783 OID_AUTO, "errors", CTLFLAG_RD, 784 &softc->errors, 0, 785 "Transport errors reported by the SIM."); 786 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx, 787 SYSCTL_CHILDREN(softc->sysctl_stats_tree), 788 OID_AUTO, "pack_invalidations", CTLFLAG_RD, 789 &softc->invalidations, 0, 790 "Device pack invalidations."); 791 #endif 792 793 #ifdef CAM_TEST_FAILURE 794 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree), 795 OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, 796 periph, 0, cam_periph_invalidate_sysctl, "I", 797 "Write 1 to invalidate the drive immediately"); 798 #endif 799 800 cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx, 801 softc->sysctl_tree); 802 803 cam_periph_release(periph); 804 } 805 806 static int 807 ndaflagssysctl(SYSCTL_HANDLER_ARGS) 808 { 809 struct sbuf sbuf; 810 struct nda_softc *softc = arg1; 811 int error; 812 813 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 814 if (softc->flags != 0) 815 sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING); 816 else 817 sbuf_printf(&sbuf, "0"); 818 error = sbuf_finish(&sbuf); 819 sbuf_delete(&sbuf); 820 821 return (error); 822 } 823 824 static int 825 ndagetattr(struct bio *bp) 826 { 827 int ret; 828 struct cam_periph *periph; 829 830 if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup)) 831 return (EJUSTRETURN); 832 833 periph = (struct cam_periph *)bp->bio_disk->d_drv1; 834 cam_periph_lock(periph); 835 ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute, 836 periph->path); 837 cam_periph_unlock(periph); 838 if (ret == 0) 839 bp->bio_completed = bp->bio_length; 840 return ret; 841 } 842 843 static cam_status 844 ndaregister(struct cam_periph *periph, void *arg) 845 { 846 struct nda_softc *softc; 847 struct disk *disk; 848 struct ccb_pathinq cpi; 849 const struct nvme_namespace_data *nsd; 850 const struct nvme_controller_data *cd; 851 char announce_buf[80]; 852 uint8_t flbas_fmt, lbads, vwc_present; 853 u_int maxio; 854 int quirks; 855 856 nsd = nvme_get_identify_ns(periph); 857 cd = nvme_get_identify_cntrl(periph); 858 859 softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF, 860 M_NOWAIT | M_ZERO); 861 862 if (softc == NULL) { 863 printf("ndaregister: Unable to probe new device. " 864 "Unable to allocate softc\n"); 865 return(CAM_REQ_CMP_ERR); 866 } 867 868 if (cam_iosched_init(&softc->cam_iosched, periph) != 0) { 869 printf("ndaregister: Unable to probe new device. " 870 "Unable to allocate iosched memory\n"); 871 free(softc, M_DEVBUF); 872 return(CAM_REQ_CMP_ERR); 873 } 874 875 /* ident_data parsing */ 876 877 periph->softc = softc; 878 softc->quirks = NDA_Q_NONE; 879 xpt_path_inq(&cpi, periph->path); 880 TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph); 881 882 /* 883 * The name space ID is the lun, save it for later I/O 884 */ 885 softc->nsid = (uint32_t)xpt_path_lun_id(periph->path); 886 887 /* 888 * Register this media as a disk 889 */ 890 (void)cam_periph_acquire(periph); 891 cam_periph_unlock(periph); 892 snprintf(announce_buf, sizeof(announce_buf), 893 "kern.cam.nda.%d.quirks", periph->unit_number); 894 quirks = softc->quirks; 895 TUNABLE_INT_FETCH(announce_buf, &quirks); 896 softc->quirks = quirks; 897 cam_iosched_set_sort_queue(softc->cam_iosched, 0); 898 softc->disk = disk = disk_alloc(); 899 disk->d_rotation_rate = DISK_RR_NON_ROTATING; 900 disk->d_open = ndaopen; 901 disk->d_close = ndaclose; 902 disk->d_strategy = ndastrategy; 903 disk->d_ioctl = ndaioctl; 904 disk->d_getattr = ndagetattr; 905 if (cam_sim_pollable(periph->sim)) 906 disk->d_dump = ndadump; 907 disk->d_gone = ndadiskgonecb; 908 disk->d_name = "nda"; 909 disk->d_drv1 = periph; 910 disk->d_unit = periph->unit_number; 911 maxio = cpi.maxio; /* Honor max I/O size of SIM */ 912 if (maxio == 0) 913 maxio = DFLTPHYS; /* traditional default */ 914 else if (maxio > maxphys) 915 maxio = maxphys; /* for safety */ 916 disk->d_maxsize = maxio; 917 flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) & 918 NVME_NS_DATA_FLBAS_FORMAT_MASK; 919 lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) & 920 NVME_NS_DATA_LBAF_LBADS_MASK; 921 disk->d_sectorsize = 1 << lbads; 922 disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze); 923 disk->d_delmaxsize = disk->d_mediasize; 924 disk->d_flags = DISKFLAG_DIRECT_COMPLETION; 925 if (nvme_ctrlr_has_dataset_mgmt(cd)) 926 disk->d_flags |= DISKFLAG_CANDELETE; 927 vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) & 928 NVME_CTRLR_DATA_VWC_PRESENT_MASK; 929 if (vwc_present) 930 disk->d_flags |= DISKFLAG_CANFLUSHCACHE; 931 if ((cpi.hba_misc & PIM_UNMAPPED) != 0) { 932 disk->d_flags |= DISKFLAG_UNMAPPED_BIO; 933 softc->unmappedio = 1; 934 } 935 /* 936 * d_ident and d_descr are both far bigger than the length of either 937 * the serial or model number strings. 938 */ 939 cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH, 940 sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC); 941 942 cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH, 943 sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC); 944 945 disk->d_hba_vendor = cpi.hba_vendor; 946 disk->d_hba_device = cpi.hba_device; 947 disk->d_hba_subvendor = cpi.hba_subvendor; 948 disk->d_hba_subdevice = cpi.hba_subdevice; 949 snprintf(disk->d_attachment, sizeof(disk->d_attachment), 950 "%s%d", cpi.dev_name, cpi.unit_number); 951 if (((nsd->nsfeat >> NVME_NS_DATA_NSFEAT_NPVALID_SHIFT) & 952 NVME_NS_DATA_NSFEAT_NPVALID_MASK) != 0 && nsd->npwg != 0) 953 disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize); 954 else 955 disk->d_stripesize = nsd->noiob * disk->d_sectorsize; 956 disk->d_stripeoffset = 0; 957 disk->d_devstat = devstat_new_entry(periph->periph_name, 958 periph->unit_number, disk->d_sectorsize, 959 DEVSTAT_ALL_SUPPORTED, 960 DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport), 961 DEVSTAT_PRIORITY_DISK); 962 /* 963 * Add alias for older nvd drives to ease transition. 964 */ 965 if (nda_nvd_compat) 966 disk_add_alias(disk, "nvd"); 967 968 cam_periph_lock(periph); 969 970 snprintf(announce_buf, sizeof(announce_buf), 971 "%juMB (%ju %u byte sectors)", 972 (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)), 973 (uintmax_t)disk->d_mediasize / disk->d_sectorsize, 974 disk->d_sectorsize); 975 xpt_announce_periph(periph, announce_buf); 976 xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING); 977 978 /* 979 * Create our sysctl variables, now that we know 980 * we have successfully attached. 981 */ 982 if (cam_periph_acquire(periph) == 0) 983 taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task); 984 985 /* 986 * Register for device going away and info about the drive 987 * changing (though with NVMe, it can't) 988 */ 989 xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED, 990 ndaasync, periph, periph->path); 991 992 softc->state = NDA_STATE_NORMAL; 993 994 /* 995 * We'll release this reference once GEOM calls us back via 996 * ndadiskgonecb(), telling us that our provider has been freed. 997 */ 998 if (cam_periph_acquire(periph) == 0) 999 disk_create(softc->disk, DISK_VERSION); 1000 1001 cam_periph_release_locked(periph); 1002 return(CAM_REQ_CMP); 1003 } 1004 1005 static void 1006 ndastart(struct cam_periph *periph, union ccb *start_ccb) 1007 { 1008 struct nda_softc *softc = (struct nda_softc *)periph->softc; 1009 struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio; 1010 1011 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n")); 1012 1013 switch (softc->state) { 1014 case NDA_STATE_NORMAL: 1015 { 1016 struct bio *bp; 1017 1018 bp = cam_iosched_next_bio(softc->cam_iosched); 1019 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp)); 1020 if (bp == NULL) { 1021 xpt_release_ccb(start_ccb); 1022 break; 1023 } 1024 1025 switch (bp->bio_cmd) { 1026 case BIO_WRITE: 1027 softc->flags |= NDA_FLAG_DIRTY; 1028 /* FALLTHROUGH */ 1029 case BIO_READ: 1030 { 1031 #ifdef CAM_TEST_FAILURE 1032 int fail = 0; 1033 1034 /* 1035 * Support the failure ioctls. If the command is a 1036 * read, and there are pending forced read errors, or 1037 * if a write and pending write errors, then fail this 1038 * operation with EIO. This is useful for testing 1039 * purposes. Also, support having every Nth read fail. 1040 * 1041 * This is a rather blunt tool. 1042 */ 1043 if (bp->bio_cmd == BIO_READ) { 1044 if (softc->force_read_error) { 1045 softc->force_read_error--; 1046 fail = 1; 1047 } 1048 if (softc->periodic_read_error > 0) { 1049 if (++softc->periodic_read_count >= 1050 softc->periodic_read_error) { 1051 softc->periodic_read_count = 0; 1052 fail = 1; 1053 } 1054 } 1055 } else { 1056 if (softc->force_write_error) { 1057 softc->force_write_error--; 1058 fail = 1; 1059 } 1060 } 1061 if (fail) { 1062 biofinish(bp, NULL, EIO); 1063 xpt_release_ccb(start_ccb); 1064 ndaschedule(periph); 1065 return; 1066 } 1067 #endif 1068 KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 || 1069 round_page(bp->bio_bcount + bp->bio_ma_offset) / 1070 PAGE_SIZE == bp->bio_ma_n, 1071 ("Short bio %p", bp)); 1072 nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ? 1073 NVME_OPC_READ : NVME_OPC_WRITE); 1074 break; 1075 } 1076 case BIO_DELETE: 1077 { 1078 struct nvme_dsm_range *dsm_range, *dsm_end; 1079 struct nda_trim_request *trim; 1080 struct bio *bp1; 1081 int ents; 1082 uint32_t totalcount = 0, ranges = 0; 1083 1084 trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT); 1085 if (trim == NULL) { 1086 biofinish(bp, NULL, ENOMEM); 1087 xpt_release_ccb(start_ccb); 1088 ndaschedule(periph); 1089 return; 1090 } 1091 TAILQ_INIT(&trim->bps); 1092 bp1 = bp; 1093 ents = min(nitems(trim->dsm), nda_max_trim_entries); 1094 ents = max(ents, 1); 1095 dsm_range = trim->dsm; 1096 dsm_end = dsm_range + ents; 1097 do { 1098 TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue); 1099 dsm_range->length = 1100 htole32(bp1->bio_bcount / softc->disk->d_sectorsize); 1101 dsm_range->starting_lba = 1102 htole64(bp1->bio_offset / softc->disk->d_sectorsize); 1103 ranges++; 1104 totalcount += dsm_range->length; 1105 dsm_range++; 1106 if (dsm_range >= dsm_end) 1107 break; 1108 bp1 = cam_iosched_next_trim(softc->cam_iosched); 1109 /* XXX -- Could collapse adjacent ranges, but we don't for now */ 1110 /* XXX -- Could limit based on total payload size */ 1111 } while (bp1 != NULL); 1112 start_ccb->ccb_trim = trim; 1113 nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm, 1114 dsm_range - trim->dsm); 1115 start_ccb->ccb_state = NDA_CCB_TRIM; 1116 softc->trim_count++; 1117 softc->trim_ranges += ranges; 1118 softc->trim_lbas += totalcount; 1119 /* 1120 * Note: We can have multiple TRIMs in flight, so we don't call 1121 * cam_iosched_submit_trim(softc->cam_iosched); 1122 * since that forces the I/O scheduler to only schedule one at a time. 1123 * On NVMe drives, this is a performance disaster. 1124 */ 1125 goto out; 1126 } 1127 case BIO_FLUSH: 1128 nda_nvme_flush(softc, nvmeio); 1129 break; 1130 default: 1131 biofinish(bp, NULL, EOPNOTSUPP); 1132 xpt_release_ccb(start_ccb); 1133 ndaschedule(periph); 1134 return; 1135 } 1136 start_ccb->ccb_state = NDA_CCB_BUFFER_IO; 1137 start_ccb->ccb_bp = bp; 1138 out: 1139 start_ccb->ccb_h.flags |= CAM_UNLOCKED; 1140 softc->outstanding_cmds++; 1141 softc->refcount++; /* For submission only */ 1142 cam_periph_unlock(periph); 1143 xpt_action(start_ccb); 1144 cam_periph_lock(periph); 1145 softc->refcount--; /* Submission done */ 1146 1147 /* May have more work to do, so ensure we stay scheduled */ 1148 ndaschedule(periph); 1149 break; 1150 } 1151 } 1152 } 1153 1154 static void 1155 ndadone(struct cam_periph *periph, union ccb *done_ccb) 1156 { 1157 struct nda_softc *softc; 1158 struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio; 1159 struct cam_path *path; 1160 int state; 1161 1162 softc = (struct nda_softc *)periph->softc; 1163 path = done_ccb->ccb_h.path; 1164 1165 CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n")); 1166 1167 state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK; 1168 switch (state) { 1169 case NDA_CCB_BUFFER_IO: 1170 case NDA_CCB_TRIM: 1171 { 1172 int error; 1173 1174 cam_periph_lock(periph); 1175 if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) { 1176 error = ndaerror(done_ccb, 0, 0); 1177 if (error == ERESTART) { 1178 /* A retry was scheduled, so just return. */ 1179 cam_periph_unlock(periph); 1180 return; 1181 } 1182 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1183 cam_release_devq(path, 1184 /*relsim_flags*/0, 1185 /*reduction*/0, 1186 /*timeout*/0, 1187 /*getcount_only*/0); 1188 } else { 1189 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) 1190 panic("REQ_CMP with QFRZN"); 1191 error = 0; 1192 } 1193 if (state == NDA_CCB_BUFFER_IO) { 1194 struct bio *bp; 1195 1196 bp = (struct bio *)done_ccb->ccb_bp; 1197 bp->bio_error = error; 1198 if (error != 0) { 1199 bp->bio_resid = bp->bio_bcount; 1200 bp->bio_flags |= BIO_ERROR; 1201 } else { 1202 bp->bio_resid = 0; 1203 } 1204 softc->outstanding_cmds--; 1205 1206 /* 1207 * We need to call cam_iosched before we call biodone so that we 1208 * don't measure any activity that happens in the completion 1209 * routine, which in the case of sendfile can be quite 1210 * extensive. 1211 */ 1212 cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb); 1213 xpt_release_ccb(done_ccb); 1214 ndaschedule(periph); 1215 cam_periph_unlock(periph); 1216 biodone(bp); 1217 } else { /* state == NDA_CCB_TRIM */ 1218 struct nda_trim_request *trim; 1219 struct bio *bp1, *bp2; 1220 TAILQ_HEAD(, bio) queue; 1221 1222 trim = nvmeio->ccb_trim; 1223 TAILQ_INIT(&queue); 1224 TAILQ_CONCAT(&queue, &trim->bps, bio_queue); 1225 free(trim, M_NVMEDA); 1226 1227 /* 1228 * Since we can have multiple trims in flight, we don't 1229 * need to call this here. 1230 * cam_iosched_trim_done(softc->cam_iosched); 1231 */ 1232 /* 1233 * The the I/O scheduler that we're finishing the I/O 1234 * so we can keep book. The first one we pass in the CCB 1235 * which has the timing information. The rest we pass in NULL 1236 * so we can keep proper counts. 1237 */ 1238 bp1 = TAILQ_FIRST(&queue); 1239 cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb); 1240 xpt_release_ccb(done_ccb); 1241 softc->outstanding_cmds--; 1242 ndaschedule(periph); 1243 cam_periph_unlock(periph); 1244 while ((bp2 = TAILQ_FIRST(&queue)) != NULL) { 1245 TAILQ_REMOVE(&queue, bp2, bio_queue); 1246 bp2->bio_error = error; 1247 if (error != 0) { 1248 bp2->bio_flags |= BIO_ERROR; 1249 bp2->bio_resid = bp1->bio_bcount; 1250 } else 1251 bp2->bio_resid = 0; 1252 if (bp1 != bp2) 1253 cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL); 1254 biodone(bp2); 1255 } 1256 } 1257 return; 1258 } 1259 case NDA_CCB_DUMP: 1260 /* No-op. We're polling */ 1261 return; 1262 case NDA_CCB_PASS: 1263 /* NVME_PASSTHROUGH_CMD runs this CCB and releases it */ 1264 return; 1265 default: 1266 break; 1267 } 1268 xpt_release_ccb(done_ccb); 1269 } 1270 1271 static int 1272 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags) 1273 { 1274 #ifdef CAM_IO_STATS 1275 struct nda_softc *softc; 1276 struct cam_periph *periph; 1277 1278 periph = xpt_path_periph(ccb->ccb_h.path); 1279 softc = (struct nda_softc *)periph->softc; 1280 #endif 1281 1282 switch (ccb->ccb_h.status & CAM_STATUS_MASK) { 1283 case CAM_CMD_TIMEOUT: 1284 #ifdef CAM_IO_STATS 1285 softc->timeouts++; 1286 #endif 1287 break; 1288 case CAM_REQ_ABORTED: 1289 case CAM_REQ_CMP_ERR: 1290 case CAM_REQ_TERMIO: 1291 case CAM_UNREC_HBA_ERROR: 1292 case CAM_DATA_RUN_ERR: 1293 case CAM_ATA_STATUS_ERROR: 1294 #ifdef CAM_IO_STATS 1295 softc->errors++; 1296 #endif 1297 break; 1298 default: 1299 break; 1300 } 1301 1302 return(cam_periph_error(ccb, cam_flags, sense_flags)); 1303 } 1304 1305 /* 1306 * Step through all NDA peripheral drivers, and if the device is still open, 1307 * sync the disk cache to physical media. 1308 */ 1309 static void 1310 ndaflush(void) 1311 { 1312 struct cam_periph *periph; 1313 struct nda_softc *softc; 1314 union ccb *ccb; 1315 int error; 1316 1317 CAM_PERIPH_FOREACH(periph, &ndadriver) { 1318 softc = (struct nda_softc *)periph->softc; 1319 1320 if (SCHEDULER_STOPPED()) { 1321 /* 1322 * If we panicked with the lock held or the periph is not 1323 * open, do not recurse. Otherwise, call ndadump since 1324 * that avoids the sleeping cam_periph_getccb does if no 1325 * CCBs are available. 1326 */ 1327 if (!cam_periph_owned(periph) && 1328 (softc->flags & NDA_FLAG_OPEN)) { 1329 ndadump(softc->disk, NULL, 0, 0); 1330 } 1331 continue; 1332 } 1333 1334 /* 1335 * We only sync the cache if the drive is still open 1336 */ 1337 cam_periph_lock(periph); 1338 if ((softc->flags & NDA_FLAG_OPEN) == 0) { 1339 cam_periph_unlock(periph); 1340 continue; 1341 } 1342 1343 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL); 1344 nda_nvme_flush(softc, &ccb->nvmeio); 1345 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0, 1346 /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY, 1347 softc->disk->d_devstat); 1348 if (error != 0) 1349 xpt_print(periph->path, "Synchronize cache failed\n"); 1350 xpt_release_ccb(ccb); 1351 cam_periph_unlock(periph); 1352 } 1353 } 1354 1355 static void 1356 ndashutdown(void *arg, int howto) 1357 { 1358 1359 if ((howto & RB_NOSYNC) != 0) 1360 return; 1361 1362 ndaflush(); 1363 } 1364 1365 static void 1366 ndasuspend(void *arg) 1367 { 1368 1369 ndaflush(); 1370 } 1371