1 /* $OpenBSD: nvme.c,v 1.83 2020/07/20 14:41:13 krw Exp $ */ 2 3 /* 4 * Copyright (c) 2014 David Gwynne <dlg@openbsd.org> 5 * 6 * Permission to use, copy, modify, and distribute this software for any 7 * purpose with or without fee is hereby granted, provided that the above 8 * copyright notice and this permission notice appear in all copies. 9 * 10 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 11 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 12 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 13 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 14 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 15 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 16 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 17 */ 18 19 #include <sys/param.h> 20 #include <sys/systm.h> 21 #include <sys/buf.h> 22 #include <sys/kernel.h> 23 #include <sys/malloc.h> 24 #include <sys/device.h> 25 #include <sys/queue.h> 26 #include <sys/mutex.h> 27 #include <sys/pool.h> 28 29 #include <sys/atomic.h> 30 31 #include <machine/bus.h> 32 33 #include <scsi/scsi_all.h> 34 #include <scsi/scsi_disk.h> 35 #include <scsi/scsiconf.h> 36 37 #include <dev/ic/nvmereg.h> 38 #include <dev/ic/nvmevar.h> 39 40 struct cfdriver nvme_cd = { 41 NULL, 42 "nvme", 43 DV_DULL 44 }; 45 46 int nvme_ready(struct nvme_softc *, u_int32_t); 47 int nvme_enable(struct nvme_softc *); 48 int nvme_disable(struct nvme_softc *); 49 int nvme_shutdown(struct nvme_softc *); 50 int nvme_resume(struct nvme_softc *); 51 52 void nvme_dumpregs(struct nvme_softc *); 53 int nvme_identify(struct nvme_softc *, u_int); 54 void nvme_fill_identify(struct nvme_softc *, struct nvme_ccb *, void *); 55 56 int nvme_ccbs_alloc(struct nvme_softc *, u_int); 57 void nvme_ccbs_free(struct nvme_softc *, u_int); 58 59 void * nvme_ccb_get(void *); 60 void nvme_ccb_put(void *, void *); 61 62 int nvme_poll(struct nvme_softc *, struct nvme_queue *, struct nvme_ccb *, 63 void (*)(struct nvme_softc *, struct nvme_ccb *, void *)); 64 void nvme_poll_fill(struct nvme_softc *, struct nvme_ccb *, void *); 65 void nvme_poll_done(struct nvme_softc *, struct nvme_ccb *, 66 struct nvme_cqe *); 67 void nvme_sqe_fill(struct nvme_softc *, struct nvme_ccb *, void *); 68 void nvme_empty_done(struct nvme_softc *, struct nvme_ccb *, 69 struct nvme_cqe *); 70 71 struct nvme_queue * 72 nvme_q_alloc(struct nvme_softc *, u_int16_t, u_int, u_int); 73 int nvme_q_create(struct nvme_softc *, struct nvme_queue *); 74 int nvme_q_reset(struct nvme_softc *, struct nvme_queue *); 75 int nvme_q_delete(struct nvme_softc *, struct nvme_queue *); 76 void nvme_q_submit(struct nvme_softc *, 77 struct nvme_queue *, struct nvme_ccb *, 78 void (*)(struct nvme_softc *, struct nvme_ccb *, void *)); 79 int nvme_q_complete(struct nvme_softc *, struct nvme_queue *); 80 void nvme_q_free(struct nvme_softc *, struct nvme_queue *); 81 82 struct nvme_dmamem * 83 nvme_dmamem_alloc(struct nvme_softc *, size_t); 84 void nvme_dmamem_free(struct nvme_softc *, struct nvme_dmamem *); 85 void nvme_dmamem_sync(struct nvme_softc *, struct nvme_dmamem *, int); 86 87 void nvme_scsi_cmd(struct scsi_xfer *); 88 void nvme_minphys(struct buf *, struct scsi_link *); 89 int nvme_scsi_probe(struct scsi_link *); 90 void nvme_scsi_free(struct scsi_link *); 91 92 #ifdef HIBERNATE 93 #include <uvm/uvm_extern.h> 94 #include <sys/hibernate.h> 95 #include <sys/disk.h> 96 #include <sys/disklabel.h> 97 98 int nvme_hibernate_io(dev_t, daddr_t, vaddr_t, size_t, int, void *); 99 #endif 100 101 struct scsi_adapter nvme_switch = { 102 nvme_scsi_cmd, nvme_minphys, nvme_scsi_probe, nvme_scsi_free, NULL 103 }; 104 105 void nvme_scsi_io(struct scsi_xfer *, int); 106 void nvme_scsi_io_fill(struct nvme_softc *, struct nvme_ccb *, void *); 107 void nvme_scsi_io_done(struct nvme_softc *, struct nvme_ccb *, 108 struct nvme_cqe *); 109 110 void nvme_scsi_sync(struct scsi_xfer *); 111 void nvme_scsi_sync_fill(struct nvme_softc *, struct nvme_ccb *, void *); 112 void nvme_scsi_sync_done(struct nvme_softc *, struct nvme_ccb *, 113 struct nvme_cqe *); 114 115 void nvme_scsi_inq(struct scsi_xfer *); 116 void nvme_scsi_inquiry(struct scsi_xfer *); 117 void nvme_scsi_capacity16(struct scsi_xfer *); 118 void nvme_scsi_capacity(struct scsi_xfer *); 119 120 #define nvme_read4(_s, _r) \ 121 bus_space_read_4((_s)->sc_iot, (_s)->sc_ioh, (_r)) 122 #define nvme_write4(_s, _r, _v) \ 123 bus_space_write_4((_s)->sc_iot, (_s)->sc_ioh, (_r), (_v)) 124 /* 125 * Some controllers, at least Apple NVMe, always require split 126 * transfers, so don't use bus_space_{read,write}_8() on LP64. 127 */ 128 static inline u_int64_t 129 nvme_read8(struct nvme_softc *sc, bus_size_t r) 130 { 131 u_int64_t v; 132 u_int32_t *a = (u_int32_t *)&v; 133 134 #if _BYTE_ORDER == _LITTLE_ENDIAN 135 a[0] = nvme_read4(sc, r); 136 a[1] = nvme_read4(sc, r + 4); 137 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */ 138 a[1] = nvme_read4(sc, r); 139 a[0] = nvme_read4(sc, r + 4); 140 #endif 141 142 return (v); 143 } 144 145 static inline void 146 nvme_write8(struct nvme_softc *sc, bus_size_t r, u_int64_t v) 147 { 148 u_int32_t *a = (u_int32_t *)&v; 149 150 #if _BYTE_ORDER == _LITTLE_ENDIAN 151 nvme_write4(sc, r, a[0]); 152 nvme_write4(sc, r + 4, a[1]); 153 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */ 154 nvme_write4(sc, r, a[1]); 155 nvme_write4(sc, r + 4, a[0]); 156 #endif 157 } 158 #define nvme_barrier(_s, _r, _l, _f) \ 159 bus_space_barrier((_s)->sc_iot, (_s)->sc_ioh, (_r), (_l), (_f)) 160 161 void 162 nvme_dumpregs(struct nvme_softc *sc) 163 { 164 u_int64_t r8; 165 u_int32_t r4; 166 167 r8 = nvme_read8(sc, NVME_CAP); 168 printf("%s: cap 0x%016llx\n", DEVNAME(sc), nvme_read8(sc, NVME_CAP)); 169 printf("%s: mpsmax %u (%u)\n", DEVNAME(sc), 170 (u_int)NVME_CAP_MPSMAX(r8), (1 << NVME_CAP_MPSMAX(r8))); 171 printf("%s: mpsmin %u (%u)\n", DEVNAME(sc), 172 (u_int)NVME_CAP_MPSMIN(r8), (1 << NVME_CAP_MPSMIN(r8))); 173 printf("%s: css %llu\n", DEVNAME(sc), NVME_CAP_CSS(r8)); 174 printf("%s: nssrs %llu\n", DEVNAME(sc), NVME_CAP_NSSRS(r8)); 175 printf("%s: dstrd %u\n", DEVNAME(sc), NVME_CAP_DSTRD(r8)); 176 printf("%s: to %llu msec\n", DEVNAME(sc), NVME_CAP_TO(r8)); 177 printf("%s: ams %llu\n", DEVNAME(sc), NVME_CAP_AMS(r8)); 178 printf("%s: cqr %llu\n", DEVNAME(sc), NVME_CAP_CQR(r8)); 179 printf("%s: mqes %llu\n", DEVNAME(sc), NVME_CAP_MQES(r8)); 180 181 printf("%s: vs 0x%04x\n", DEVNAME(sc), nvme_read4(sc, NVME_VS)); 182 183 r4 = nvme_read4(sc, NVME_CC); 184 printf("%s: cc 0x%04x\n", DEVNAME(sc), r4); 185 printf("%s: iocqes %u\n", DEVNAME(sc), NVME_CC_IOCQES_R(r4)); 186 printf("%s: iosqes %u\n", DEVNAME(sc), NVME_CC_IOSQES_R(r4)); 187 printf("%s: shn %u\n", DEVNAME(sc), NVME_CC_SHN_R(r4)); 188 printf("%s: ams %u\n", DEVNAME(sc), NVME_CC_AMS_R(r4)); 189 printf("%s: mps %u\n", DEVNAME(sc), NVME_CC_MPS_R(r4)); 190 printf("%s: css %u\n", DEVNAME(sc), NVME_CC_CSS_R(r4)); 191 printf("%s: en %u\n", DEVNAME(sc), ISSET(r4, NVME_CC_EN)); 192 193 printf("%s: csts 0x%08x\n", DEVNAME(sc), nvme_read4(sc, NVME_CSTS)); 194 printf("%s: aqa 0x%08x\n", DEVNAME(sc), nvme_read4(sc, NVME_AQA)); 195 printf("%s: asq 0x%016llx\n", DEVNAME(sc), nvme_read8(sc, NVME_ASQ)); 196 printf("%s: acq 0x%016llx\n", DEVNAME(sc), nvme_read8(sc, NVME_ACQ)); 197 } 198 199 int 200 nvme_ready(struct nvme_softc *sc, u_int32_t rdy) 201 { 202 u_int i = 0; 203 204 while ((nvme_read4(sc, NVME_CSTS) & NVME_CSTS_RDY) != rdy) { 205 if (i++ > sc->sc_rdy_to) 206 return (1); 207 208 delay(1000); 209 nvme_barrier(sc, NVME_CSTS, 4, BUS_SPACE_BARRIER_READ); 210 } 211 212 return (0); 213 } 214 215 int 216 nvme_enable(struct nvme_softc *sc) 217 { 218 u_int32_t cc; 219 220 cc = nvme_read4(sc, NVME_CC); 221 if (ISSET(cc, NVME_CC_EN)) 222 return (nvme_ready(sc, NVME_CSTS_RDY)); 223 224 nvme_write4(sc, NVME_AQA, NVME_AQA_ACQS(sc->sc_admin_q->q_entries) | 225 NVME_AQA_ASQS(sc->sc_admin_q->q_entries)); 226 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 227 228 nvme_write8(sc, NVME_ASQ, NVME_DMA_DVA(sc->sc_admin_q->q_sq_dmamem)); 229 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 230 nvme_write8(sc, NVME_ACQ, NVME_DMA_DVA(sc->sc_admin_q->q_cq_dmamem)); 231 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 232 233 CLR(cc, NVME_CC_IOCQES_MASK | NVME_CC_IOSQES_MASK | NVME_CC_SHN_MASK | 234 NVME_CC_AMS_MASK | NVME_CC_MPS_MASK | NVME_CC_CSS_MASK); 235 SET(cc, NVME_CC_IOSQES(6)); /* Submission queue size == 2**6 (64) */ 236 SET(cc, NVME_CC_IOCQES(4)); /* Completion queue size == 2**4 (16) */ 237 SET(cc, NVME_CC_SHN(NVME_CC_SHN_NONE)); 238 SET(cc, NVME_CC_CSS(NVME_CC_CSS_NVM)); 239 SET(cc, NVME_CC_AMS(NVME_CC_AMS_RR)); 240 SET(cc, NVME_CC_MPS(ffs(sc->sc_mps) - 1)); 241 SET(cc, NVME_CC_EN); 242 243 nvme_write4(sc, NVME_CC, cc); 244 nvme_barrier(sc, 0, sc->sc_ios, 245 BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE); 246 247 return (nvme_ready(sc, NVME_CSTS_RDY)); 248 } 249 250 int 251 nvme_disable(struct nvme_softc *sc) 252 { 253 u_int32_t cc, csts; 254 255 cc = nvme_read4(sc, NVME_CC); 256 if (ISSET(cc, NVME_CC_EN)) { 257 csts = nvme_read4(sc, NVME_CSTS); 258 if (!ISSET(csts, NVME_CSTS_CFS) && 259 nvme_ready(sc, NVME_CSTS_RDY) != 0) 260 return (1); 261 } 262 263 CLR(cc, NVME_CC_EN); 264 265 nvme_write4(sc, NVME_CC, cc); 266 nvme_barrier(sc, 0, sc->sc_ios, 267 BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE); 268 269 return (nvme_ready(sc, 0)); 270 } 271 272 int 273 nvme_attach(struct nvme_softc *sc) 274 { 275 struct scsibus_attach_args saa; 276 u_int64_t cap; 277 u_int32_t reg; 278 u_int nccbs = 0; 279 280 mtx_init(&sc->sc_ccb_mtx, IPL_BIO); 281 SIMPLEQ_INIT(&sc->sc_ccb_list); 282 scsi_iopool_init(&sc->sc_iopool, sc, nvme_ccb_get, nvme_ccb_put); 283 284 reg = nvme_read4(sc, NVME_VS); 285 if (reg == 0xffffffff) { 286 printf(", invalid mapping\n"); 287 return (1); 288 } 289 290 printf(", NVMe %d.%d\n", NVME_VS_MJR(reg), NVME_VS_MNR(reg)); 291 292 cap = nvme_read8(sc, NVME_CAP); 293 sc->sc_dstrd = NVME_CAP_DSTRD(cap); 294 if (NVME_CAP_MPSMIN(cap) > PAGE_SHIFT) { 295 printf("%s: NVMe minimum page size %u " 296 "is greater than CPU page size %u\n", DEVNAME(sc), 297 1 << NVME_CAP_MPSMIN(cap), 1 << PAGE_SHIFT); 298 return (1); 299 } 300 if (NVME_CAP_MPSMAX(cap) < PAGE_SHIFT) 301 sc->sc_mps = 1 << NVME_CAP_MPSMAX(cap); 302 else 303 sc->sc_mps = 1 << PAGE_SHIFT; 304 305 sc->sc_rdy_to = NVME_CAP_TO(cap); 306 sc->sc_mdts = MAXPHYS; 307 sc->sc_max_prpl = sc->sc_mdts / sc->sc_mps; 308 309 if (nvme_disable(sc) != 0) { 310 printf("%s: unable to disable controller\n", DEVNAME(sc)); 311 return (1); 312 } 313 314 sc->sc_admin_q = nvme_q_alloc(sc, NVME_ADMIN_Q, 128, sc->sc_dstrd); 315 if (sc->sc_admin_q == NULL) { 316 printf("%s: unable to allocate admin queue\n", DEVNAME(sc)); 317 return (1); 318 } 319 320 if (nvme_ccbs_alloc(sc, 16) != 0) { 321 printf("%s: unable to allocate initial ccbs\n", DEVNAME(sc)); 322 goto free_admin_q; 323 } 324 nccbs = 16; 325 326 if (nvme_enable(sc) != 0) { 327 printf("%s: unable to enable controller\n", DEVNAME(sc)); 328 goto free_ccbs; 329 } 330 331 if (nvme_identify(sc, NVME_CAP_MPSMIN(cap)) != 0) { 332 printf("%s: unable to identify controller\n", DEVNAME(sc)); 333 goto disable; 334 } 335 336 /* We now know the real values of sc_mdts and sc_max_prpl. */ 337 nvme_ccbs_free(sc, nccbs); 338 if (nvme_ccbs_alloc(sc, 64) != 0) { 339 printf("%s: unable to allocate ccbs\n", DEVNAME(sc)); 340 goto free_admin_q; 341 } 342 nccbs = 64; 343 344 sc->sc_q = nvme_q_alloc(sc, NVME_IO_Q, 128, sc->sc_dstrd); 345 if (sc->sc_q == NULL) { 346 printf("%s: unable to allocate io q\n", DEVNAME(sc)); 347 goto disable; 348 } 349 350 if (nvme_q_create(sc, sc->sc_q) != 0) { 351 printf("%s: unable to create io q\n", DEVNAME(sc)); 352 goto free_q; 353 } 354 355 sc->sc_hib_q = nvme_q_alloc(sc, NVME_HIB_Q, 4, sc->sc_dstrd); 356 if (sc->sc_hib_q == NULL) { 357 printf("%s: unable to allocate hibernate io queue\n", DEVNAME(sc)); 358 goto free_q; 359 } 360 361 nvme_write4(sc, NVME_INTMC, 1); 362 363 sc->sc_namespaces = mallocarray(sc->sc_nn + 1, 364 sizeof(*sc->sc_namespaces), M_DEVBUF, M_WAITOK|M_ZERO); 365 366 saa.saa_adapter = &nvme_switch; 367 saa.saa_adapter_softc = sc; 368 saa.saa_adapter_buswidth = sc->sc_nn + 1; 369 saa.saa_luns = 1; 370 saa.saa_adapter_target = 0; 371 saa.saa_openings = 64; 372 saa.saa_pool = &sc->sc_iopool; 373 saa.saa_quirks = saa.saa_flags = 0; 374 saa.saa_wwpn = saa.saa_wwnn = 0; 375 376 config_found(&sc->sc_dev, &saa, scsiprint); 377 378 return (0); 379 380 free_q: 381 nvme_q_free(sc, sc->sc_q); 382 disable: 383 nvme_disable(sc); 384 free_ccbs: 385 nvme_ccbs_free(sc, nccbs); 386 free_admin_q: 387 nvme_q_free(sc, sc->sc_admin_q); 388 389 return (1); 390 } 391 392 int 393 nvme_resume(struct nvme_softc *sc) 394 { 395 if (nvme_disable(sc) != 0) { 396 printf("%s: unable to disable controller\n", DEVNAME(sc)); 397 return (1); 398 } 399 400 if (nvme_q_reset(sc, sc->sc_admin_q) != 0) { 401 printf("%s: unable to reset admin queue\n", DEVNAME(sc)); 402 return (1); 403 } 404 405 if (nvme_enable(sc) != 0) { 406 printf("%s: unable to enable controller\n", DEVNAME(sc)); 407 return (1); 408 } 409 410 sc->sc_q = nvme_q_alloc(sc, NVME_IO_Q, 128, sc->sc_dstrd); 411 if (sc->sc_q == NULL) { 412 printf("%s: unable to allocate io q\n", DEVNAME(sc)); 413 goto disable; 414 } 415 416 if (nvme_q_create(sc, sc->sc_q) != 0) { 417 printf("%s: unable to create io q\n", DEVNAME(sc)); 418 goto free_q; 419 } 420 421 nvme_write4(sc, NVME_INTMC, 1); 422 423 return (0); 424 425 free_q: 426 nvme_q_free(sc, sc->sc_q); 427 disable: 428 nvme_disable(sc); 429 430 return (1); 431 } 432 433 int 434 nvme_scsi_probe(struct scsi_link *link) 435 { 436 struct nvme_softc *sc = link->bus->sb_adapter_softc; 437 struct nvme_sqe sqe; 438 struct nvm_identify_namespace *identify; 439 struct nvme_dmamem *mem; 440 struct nvme_ccb *ccb; 441 int rv; 442 443 ccb = scsi_io_get(&sc->sc_iopool, 0); 444 KASSERT(ccb != NULL); 445 446 mem = nvme_dmamem_alloc(sc, sizeof(*identify)); 447 if (mem == NULL) 448 return (ENOMEM); 449 450 memset(&sqe, 0, sizeof(sqe)); 451 sqe.opcode = NVM_ADMIN_IDENTIFY; 452 htolem32(&sqe.nsid, link->target); 453 htolem64(&sqe.entry.prp[0], NVME_DMA_DVA(mem)); 454 htolem32(&sqe.cdw10, 0); 455 456 ccb->ccb_done = nvme_empty_done; 457 ccb->ccb_cookie = &sqe; 458 459 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD); 460 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill); 461 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD); 462 463 scsi_io_put(&sc->sc_iopool, ccb); 464 465 identify = NVME_DMA_KVA(mem); 466 if (rv == 0 && lemtoh64(&identify->nsze) > 0) { 467 /* Commit namespace if it has a size greater than zero. */ 468 identify = malloc(sizeof(*identify), M_DEVBUF, M_WAITOK); 469 memcpy(identify, NVME_DMA_KVA(mem), sizeof(*identify)); 470 sc->sc_namespaces[link->target].ident = identify; 471 } 472 473 nvme_dmamem_free(sc, mem); 474 475 return (rv); 476 } 477 478 int 479 nvme_shutdown(struct nvme_softc *sc) 480 { 481 u_int32_t cc, csts; 482 int i; 483 484 nvme_write4(sc, NVME_INTMC, 0); 485 486 if (nvme_q_delete(sc, sc->sc_q) != 0) { 487 printf("%s: unable to delete q, disabling\n", DEVNAME(sc)); 488 goto disable; 489 } 490 491 cc = nvme_read4(sc, NVME_CC); 492 CLR(cc, NVME_CC_SHN_MASK); 493 SET(cc, NVME_CC_SHN(NVME_CC_SHN_NORMAL)); 494 nvme_write4(sc, NVME_CC, cc); 495 496 for (i = 0; i < 4000; i++) { 497 nvme_barrier(sc, 0, sc->sc_ios, 498 BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE); 499 csts = nvme_read4(sc, NVME_CSTS); 500 if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_DONE) 501 return (0); 502 503 delay(1000); 504 } 505 506 printf("%s: unable to shutdown, disabling\n", DEVNAME(sc)); 507 508 disable: 509 nvme_disable(sc); 510 return (0); 511 } 512 513 int 514 nvme_activate(struct nvme_softc *sc, int act) 515 { 516 int rv; 517 518 switch (act) { 519 case DVACT_POWERDOWN: 520 rv = config_activate_children(&sc->sc_dev, act); 521 nvme_shutdown(sc); 522 break; 523 case DVACT_RESUME: 524 rv = nvme_resume(sc); 525 if (rv == 0) 526 rv = config_activate_children(&sc->sc_dev, act); 527 break; 528 default: 529 rv = config_activate_children(&sc->sc_dev, act); 530 break; 531 } 532 533 return (rv); 534 } 535 536 void 537 nvme_scsi_cmd(struct scsi_xfer *xs) 538 { 539 switch (xs->cmd->opcode) { 540 case READ_COMMAND: 541 case READ_BIG: 542 case READ_12: 543 case READ_16: 544 nvme_scsi_io(xs, SCSI_DATA_IN); 545 return; 546 case WRITE_COMMAND: 547 case WRITE_BIG: 548 case WRITE_12: 549 case WRITE_16: 550 nvme_scsi_io(xs, SCSI_DATA_OUT); 551 return; 552 553 case SYNCHRONIZE_CACHE: 554 nvme_scsi_sync(xs); 555 return; 556 557 case INQUIRY: 558 nvme_scsi_inq(xs); 559 return; 560 case READ_CAPACITY_16: 561 nvme_scsi_capacity16(xs); 562 return; 563 case READ_CAPACITY: 564 nvme_scsi_capacity(xs); 565 return; 566 567 case TEST_UNIT_READY: 568 case PREVENT_ALLOW: 569 case START_STOP: 570 xs->error = XS_NOERROR; 571 scsi_done(xs); 572 return; 573 574 default: 575 break; 576 } 577 578 xs->error = XS_DRIVER_STUFFUP; 579 scsi_done(xs); 580 } 581 582 void 583 nvme_minphys(struct buf *bp, struct scsi_link *link) 584 { 585 struct nvme_softc *sc = link->bus->sb_adapter_softc; 586 587 if (bp->b_bcount > sc->sc_mdts) 588 bp->b_bcount = sc->sc_mdts; 589 } 590 591 void 592 nvme_scsi_io(struct scsi_xfer *xs, int dir) 593 { 594 struct scsi_link *link = xs->sc_link; 595 struct nvme_softc *sc = link->bus->sb_adapter_softc; 596 struct nvme_ccb *ccb = xs->io; 597 bus_dmamap_t dmap = ccb->ccb_dmamap; 598 int i; 599 600 if ((xs->flags & (SCSI_DATA_IN|SCSI_DATA_OUT)) != dir) 601 goto stuffup; 602 603 ccb->ccb_done = nvme_scsi_io_done; 604 ccb->ccb_cookie = xs; 605 606 if (bus_dmamap_load(sc->sc_dmat, dmap, 607 xs->data, xs->datalen, NULL, ISSET(xs->flags, SCSI_NOSLEEP) ? 608 BUS_DMA_NOWAIT : BUS_DMA_WAITOK) != 0) 609 goto stuffup; 610 611 bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize, 612 ISSET(xs->flags, SCSI_DATA_IN) ? 613 BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE); 614 615 if (dmap->dm_nsegs > 2) { 616 for (i = 1; i < dmap->dm_nsegs; i++) { 617 htolem64(&ccb->ccb_prpl[i - 1], 618 dmap->dm_segs[i].ds_addr); 619 } 620 bus_dmamap_sync(sc->sc_dmat, 621 NVME_DMA_MAP(sc->sc_ccb_prpls), 622 ccb->ccb_prpl_off, 623 sizeof(*ccb->ccb_prpl) * dmap->dm_nsegs - 1, 624 BUS_DMASYNC_PREWRITE); 625 } 626 627 if (ISSET(xs->flags, SCSI_POLL)) { 628 nvme_poll(sc, sc->sc_q, ccb, nvme_scsi_io_fill); 629 return; 630 } 631 632 nvme_q_submit(sc, sc->sc_q, ccb, nvme_scsi_io_fill); 633 return; 634 635 stuffup: 636 xs->error = XS_DRIVER_STUFFUP; 637 scsi_done(xs); 638 } 639 640 void 641 nvme_scsi_io_fill(struct nvme_softc *sc, struct nvme_ccb *ccb, void *slot) 642 { 643 struct nvme_sqe_io *sqe = slot; 644 struct scsi_xfer *xs = ccb->ccb_cookie; 645 struct scsi_link *link = xs->sc_link; 646 bus_dmamap_t dmap = ccb->ccb_dmamap; 647 u_int64_t lba; 648 u_int32_t blocks; 649 650 scsi_cmd_rw_decode(xs->cmd, &lba, &blocks); 651 652 sqe->opcode = ISSET(xs->flags, SCSI_DATA_IN) ? 653 NVM_CMD_READ : NVM_CMD_WRITE; 654 htolem32(&sqe->nsid, link->target); 655 656 htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr); 657 switch (dmap->dm_nsegs) { 658 case 1: 659 break; 660 case 2: 661 htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr); 662 break; 663 default: 664 /* the prp list is already set up and synced */ 665 htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva); 666 break; 667 } 668 669 htolem64(&sqe->slba, lba); 670 htolem16(&sqe->nlb, blocks - 1); 671 } 672 673 void 674 nvme_scsi_io_done(struct nvme_softc *sc, struct nvme_ccb *ccb, 675 struct nvme_cqe *cqe) 676 { 677 struct scsi_xfer *xs = ccb->ccb_cookie; 678 bus_dmamap_t dmap = ccb->ccb_dmamap; 679 u_int16_t flags; 680 681 if (dmap->dm_nsegs > 2) { 682 bus_dmamap_sync(sc->sc_dmat, 683 NVME_DMA_MAP(sc->sc_ccb_prpls), 684 ccb->ccb_prpl_off, 685 sizeof(*ccb->ccb_prpl) * dmap->dm_nsegs - 1, 686 BUS_DMASYNC_POSTWRITE); 687 } 688 689 bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize, 690 ISSET(xs->flags, SCSI_DATA_IN) ? 691 BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE); 692 693 bus_dmamap_unload(sc->sc_dmat, dmap); 694 695 flags = lemtoh16(&cqe->flags); 696 697 xs->error = (NVME_CQE_SC(flags) == NVME_CQE_SC_SUCCESS) ? 698 XS_NOERROR : XS_DRIVER_STUFFUP; 699 xs->status = SCSI_OK; 700 xs->resid = 0; 701 scsi_done(xs); 702 } 703 704 void 705 nvme_scsi_sync(struct scsi_xfer *xs) 706 { 707 struct scsi_link *link = xs->sc_link; 708 struct nvme_softc *sc = link->bus->sb_adapter_softc; 709 struct nvme_ccb *ccb = xs->io; 710 711 ccb->ccb_done = nvme_scsi_sync_done; 712 ccb->ccb_cookie = xs; 713 714 if (ISSET(xs->flags, SCSI_POLL)) { 715 nvme_poll(sc, sc->sc_q, ccb, nvme_scsi_sync_fill); 716 return; 717 } 718 719 nvme_q_submit(sc, sc->sc_q, ccb, nvme_scsi_sync_fill); 720 } 721 722 void 723 nvme_scsi_sync_fill(struct nvme_softc *sc, struct nvme_ccb *ccb, void *slot) 724 { 725 struct nvme_sqe *sqe = slot; 726 struct scsi_xfer *xs = ccb->ccb_cookie; 727 struct scsi_link *link = xs->sc_link; 728 729 sqe->opcode = NVM_CMD_FLUSH; 730 htolem32(&sqe->nsid, link->target); 731 } 732 733 void 734 nvme_scsi_sync_done(struct nvme_softc *sc, struct nvme_ccb *ccb, 735 struct nvme_cqe *cqe) 736 { 737 struct scsi_xfer *xs = ccb->ccb_cookie; 738 u_int16_t flags; 739 740 flags = lemtoh16(&cqe->flags); 741 742 xs->error = (NVME_CQE_SC(flags) == NVME_CQE_SC_SUCCESS) ? 743 XS_NOERROR : XS_DRIVER_STUFFUP; 744 xs->status = SCSI_OK; 745 xs->resid = 0; 746 scsi_done(xs); 747 } 748 749 void 750 nvme_scsi_inq(struct scsi_xfer *xs) 751 { 752 struct scsi_inquiry *inq = (struct scsi_inquiry *)xs->cmd; 753 754 if (!ISSET(inq->flags, SI_EVPD)) { 755 nvme_scsi_inquiry(xs); 756 return; 757 } 758 759 switch (inq->pagecode) { 760 default: 761 /* printf("%s: %d\n", __func__, inq->pagecode); */ 762 break; 763 } 764 765 xs->error = XS_DRIVER_STUFFUP; 766 scsi_done(xs); 767 } 768 769 void 770 nvme_scsi_inquiry(struct scsi_xfer *xs) 771 { 772 struct scsi_inquiry_data inq; 773 struct scsi_link *link = xs->sc_link; 774 struct nvme_softc *sc = link->bus->sb_adapter_softc; 775 struct nvm_identify_namespace *ns; 776 777 ns = sc->sc_namespaces[link->target].ident; 778 779 memset(&inq, 0, sizeof(inq)); 780 781 inq.device = T_DIRECT; 782 inq.version = 0x06; /* SPC-4 */ 783 inq.response_format = 2; 784 inq.additional_length = 32; 785 inq.flags |= SID_CmdQue; 786 memcpy(inq.vendor, "NVMe ", sizeof(inq.vendor)); 787 memcpy(inq.product, sc->sc_identify.mn, sizeof(inq.product)); 788 memcpy(inq.revision, sc->sc_identify.fr, sizeof(inq.revision)); 789 790 memcpy(xs->data, &inq, MIN(sizeof(inq), xs->datalen)); 791 792 xs->error = XS_NOERROR; 793 scsi_done(xs); 794 } 795 796 void 797 nvme_scsi_capacity16(struct scsi_xfer *xs) 798 { 799 struct scsi_read_cap_data_16 rcd; 800 struct scsi_link *link = xs->sc_link; 801 struct nvme_softc *sc = link->bus->sb_adapter_softc; 802 struct nvm_identify_namespace *ns; 803 struct nvm_namespace_format *f; 804 u_int64_t nsze; 805 u_int16_t tpe = READ_CAP_16_TPE; 806 807 ns = sc->sc_namespaces[link->target].ident; 808 809 if (xs->cmdlen != sizeof(struct scsi_read_capacity_16)) { 810 xs->error = XS_DRIVER_STUFFUP; 811 scsi_done(xs); 812 return; 813 } 814 815 /* sd_read_cap_16() will add one */ 816 nsze = lemtoh64(&ns->nsze) - 1; 817 f = &ns->lbaf[NVME_ID_NS_FLBAS(ns->flbas)]; 818 819 memset(&rcd, 0, sizeof(rcd)); 820 _lto8b(nsze, rcd.addr); 821 _lto4b(1 << f->lbads, rcd.length); 822 _lto2b(tpe, rcd.lowest_aligned); 823 824 memcpy(xs->data, &rcd, MIN(sizeof(rcd), xs->datalen)); 825 826 xs->error = XS_NOERROR; 827 scsi_done(xs); 828 } 829 830 void 831 nvme_scsi_capacity(struct scsi_xfer *xs) 832 { 833 struct scsi_read_cap_data rcd; 834 struct scsi_link *link = xs->sc_link; 835 struct nvme_softc *sc = link->bus->sb_adapter_softc; 836 struct nvm_identify_namespace *ns; 837 struct nvm_namespace_format *f; 838 u_int64_t nsze; 839 840 ns = sc->sc_namespaces[link->target].ident; 841 842 if (xs->cmdlen != sizeof(struct scsi_read_capacity)) { 843 xs->error = XS_DRIVER_STUFFUP; 844 scsi_done(xs); 845 return; 846 } 847 848 /* sd_read_cap_10() will add one */ 849 nsze = lemtoh64(&ns->nsze) - 1; 850 if (nsze > 0xffffffff) 851 nsze = 0xffffffff; 852 853 f = &ns->lbaf[NVME_ID_NS_FLBAS(ns->flbas)]; 854 855 memset(&rcd, 0, sizeof(rcd)); 856 _lto4b(nsze, rcd.addr); 857 _lto4b(1 << f->lbads, rcd.length); 858 859 memcpy(xs->data, &rcd, MIN(sizeof(rcd), xs->datalen)); 860 861 xs->error = XS_NOERROR; 862 scsi_done(xs); 863 } 864 865 void 866 nvme_scsi_free(struct scsi_link *link) 867 { 868 struct nvme_softc *sc = link->bus->sb_adapter_softc; 869 struct nvm_identify_namespace *identify; 870 871 identify = sc->sc_namespaces[link->target].ident; 872 sc->sc_namespaces[link->target].ident = NULL; 873 874 free(identify, M_DEVBUF, sizeof(*identify)); 875 } 876 877 void 878 nvme_q_submit(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb, 879 void (*fill)(struct nvme_softc *, struct nvme_ccb *, void *)) 880 { 881 struct nvme_sqe *sqe = NVME_DMA_KVA(q->q_sq_dmamem); 882 u_int32_t tail; 883 884 mtx_enter(&q->q_sq_mtx); 885 tail = q->q_sq_tail; 886 if (++q->q_sq_tail >= q->q_entries) 887 q->q_sq_tail = 0; 888 889 sqe += tail; 890 891 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 892 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE); 893 memset(sqe, 0, sizeof(*sqe)); 894 (*fill)(sc, ccb, sqe); 895 sqe->cid = ccb->ccb_id; 896 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 897 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE); 898 899 nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail); 900 mtx_leave(&q->q_sq_mtx); 901 } 902 903 struct nvme_poll_state { 904 struct nvme_sqe s; 905 struct nvme_cqe c; 906 }; 907 908 int 909 nvme_poll(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb, 910 void (*fill)(struct nvme_softc *, struct nvme_ccb *, void *)) 911 { 912 struct nvme_poll_state state; 913 void (*done)(struct nvme_softc *, struct nvme_ccb *, struct nvme_cqe *); 914 void *cookie; 915 u_int16_t flags; 916 917 memset(&state, 0, sizeof(state)); 918 (*fill)(sc, ccb, &state.s); 919 920 done = ccb->ccb_done; 921 cookie = ccb->ccb_cookie; 922 923 ccb->ccb_done = nvme_poll_done; 924 ccb->ccb_cookie = &state; 925 926 nvme_q_submit(sc, q, ccb, nvme_poll_fill); 927 while (!ISSET(state.c.flags, htole16(NVME_CQE_PHASE))) { 928 if (nvme_q_complete(sc, q) == 0) 929 delay(10); 930 931 /* XXX no timeout? */ 932 } 933 934 ccb->ccb_cookie = cookie; 935 done(sc, ccb, &state.c); 936 937 flags = lemtoh16(&state.c.flags); 938 939 return (flags & ~NVME_CQE_PHASE); 940 } 941 942 void 943 nvme_poll_fill(struct nvme_softc *sc, struct nvme_ccb *ccb, void *slot) 944 { 945 struct nvme_sqe *sqe = slot; 946 struct nvme_poll_state *state = ccb->ccb_cookie; 947 948 *sqe = state->s; 949 } 950 951 void 952 nvme_poll_done(struct nvme_softc *sc, struct nvme_ccb *ccb, 953 struct nvme_cqe *cqe) 954 { 955 struct nvme_poll_state *state = ccb->ccb_cookie; 956 957 state->c = *cqe; 958 SET(state->c.flags, htole16(NVME_CQE_PHASE)); 959 } 960 961 void 962 nvme_sqe_fill(struct nvme_softc *sc, struct nvme_ccb *ccb, void *slot) 963 { 964 struct nvme_sqe *src = ccb->ccb_cookie; 965 struct nvme_sqe *dst = slot; 966 967 *dst = *src; 968 } 969 970 void 971 nvme_empty_done(struct nvme_softc *sc, struct nvme_ccb *ccb, 972 struct nvme_cqe *cqe) 973 { 974 } 975 976 int 977 nvme_q_complete(struct nvme_softc *sc, struct nvme_queue *q) 978 { 979 struct nvme_ccb *ccb; 980 struct nvme_cqe *ring = NVME_DMA_KVA(q->q_cq_dmamem), *cqe; 981 u_int32_t head; 982 u_int16_t flags; 983 int rv = 0; 984 985 if (!mtx_enter_try(&q->q_cq_mtx)) 986 return (-1); 987 988 head = q->q_cq_head; 989 990 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD); 991 for (;;) { 992 cqe = &ring[head]; 993 flags = lemtoh16(&cqe->flags); 994 if ((flags & NVME_CQE_PHASE) != q->q_cq_phase) 995 break; 996 997 ccb = &sc->sc_ccbs[cqe->cid]; 998 ccb->ccb_done(sc, ccb, cqe); 999 1000 if (++head >= q->q_entries) { 1001 head = 0; 1002 q->q_cq_phase ^= NVME_CQE_PHASE; 1003 } 1004 1005 rv = 1; 1006 } 1007 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD); 1008 1009 if (rv) 1010 nvme_write4(sc, q->q_cqhdbl, q->q_cq_head = head); 1011 mtx_leave(&q->q_cq_mtx); 1012 1013 return (rv); 1014 } 1015 1016 int 1017 nvme_identify(struct nvme_softc *sc, u_int mpsmin) 1018 { 1019 char sn[41], mn[81], fr[17]; 1020 struct nvm_identify_controller *identify; 1021 struct nvme_dmamem *mem; 1022 struct nvme_ccb *ccb; 1023 int rv = 1; 1024 1025 ccb = nvme_ccb_get(sc); 1026 if (ccb == NULL) 1027 panic("nvme_identify: nvme_ccb_get returned NULL"); 1028 1029 mem = nvme_dmamem_alloc(sc, sizeof(*identify)); 1030 if (mem == NULL) 1031 return (1); 1032 1033 ccb->ccb_done = nvme_empty_done; 1034 ccb->ccb_cookie = mem; 1035 1036 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD); 1037 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_fill_identify); 1038 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD); 1039 1040 nvme_ccb_put(sc, ccb); 1041 1042 if (rv != 0) 1043 goto done; 1044 1045 identify = NVME_DMA_KVA(mem); 1046 1047 scsi_strvis(sn, identify->sn, sizeof(identify->sn)); 1048 scsi_strvis(mn, identify->mn, sizeof(identify->mn)); 1049 scsi_strvis(fr, identify->fr, sizeof(identify->fr)); 1050 1051 printf("%s: %s, firmware %s, serial %s\n", DEVNAME(sc), mn, fr, sn); 1052 1053 if (identify->mdts > 0) { 1054 sc->sc_mdts = (1 << identify->mdts) * (1 << mpsmin); 1055 if (sc->sc_mdts > NVME_MAXPHYS) 1056 sc->sc_mdts = NVME_MAXPHYS; 1057 sc->sc_max_prpl = sc->sc_mdts / sc->sc_mps; 1058 } 1059 1060 sc->sc_nn = lemtoh32(&identify->nn); 1061 1062 /* 1063 * At least one Apple NVMe device presents a second, bogus disk that is 1064 * inaccessible, so cap targets at 1. 1065 * 1066 * sd1 at scsibus1 targ 2 lun 0: <NVMe, APPLE SSD AP0512, 16.1> [..] 1067 * sd1: 0MB, 4096 bytes/sector, 2 sectors 1068 */ 1069 if (sc->sc_nn > 1 && 1070 mn[0] == 'A' && mn[1] == 'P' && mn[2] == 'P' && mn[3] == 'L' && 1071 mn[4] == 'E') 1072 sc->sc_nn = 1; 1073 1074 memcpy(&sc->sc_identify, identify, sizeof(sc->sc_identify)); 1075 1076 done: 1077 nvme_dmamem_free(sc, mem); 1078 1079 return (rv); 1080 } 1081 1082 int 1083 nvme_q_create(struct nvme_softc *sc, struct nvme_queue *q) 1084 { 1085 struct nvme_sqe_q sqe; 1086 struct nvme_ccb *ccb; 1087 int rv; 1088 1089 ccb = scsi_io_get(&sc->sc_iopool, 0); 1090 KASSERT(ccb != NULL); 1091 1092 ccb->ccb_done = nvme_empty_done; 1093 ccb->ccb_cookie = &sqe; 1094 1095 memset(&sqe, 0, sizeof(sqe)); 1096 sqe.opcode = NVM_ADMIN_ADD_IOCQ; 1097 htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_cq_dmamem)); 1098 htolem16(&sqe.qsize, q->q_entries - 1); 1099 htolem16(&sqe.qid, q->q_id); 1100 sqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC; 1101 1102 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill); 1103 if (rv != 0) 1104 goto fail; 1105 1106 ccb->ccb_done = nvme_empty_done; 1107 ccb->ccb_cookie = &sqe; 1108 1109 memset(&sqe, 0, sizeof(sqe)); 1110 sqe.opcode = NVM_ADMIN_ADD_IOSQ; 1111 htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem)); 1112 htolem16(&sqe.qsize, q->q_entries - 1); 1113 htolem16(&sqe.qid, q->q_id); 1114 htolem16(&sqe.cqid, q->q_id); 1115 sqe.qflags = NVM_SQE_Q_PC; 1116 1117 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill); 1118 if (rv != 0) 1119 goto fail; 1120 1121 fail: 1122 scsi_io_put(&sc->sc_iopool, ccb); 1123 return (rv); 1124 } 1125 1126 int 1127 nvme_q_delete(struct nvme_softc *sc, struct nvme_queue *q) 1128 { 1129 struct nvme_sqe_q sqe; 1130 struct nvme_ccb *ccb; 1131 int rv; 1132 1133 ccb = scsi_io_get(&sc->sc_iopool, 0); 1134 KASSERT(ccb != NULL); 1135 1136 ccb->ccb_done = nvme_empty_done; 1137 ccb->ccb_cookie = &sqe; 1138 1139 memset(&sqe, 0, sizeof(sqe)); 1140 sqe.opcode = NVM_ADMIN_DEL_IOSQ; 1141 htolem16(&sqe.qid, q->q_id); 1142 1143 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill); 1144 if (rv != 0) 1145 goto fail; 1146 1147 ccb->ccb_done = nvme_empty_done; 1148 ccb->ccb_cookie = &sqe; 1149 1150 memset(&sqe, 0, sizeof(sqe)); 1151 sqe.opcode = NVM_ADMIN_DEL_IOCQ; 1152 htolem16(&sqe.qid, q->q_id); 1153 1154 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill); 1155 if (rv != 0) 1156 goto fail; 1157 1158 nvme_q_free(sc, q); 1159 1160 fail: 1161 scsi_io_put(&sc->sc_iopool, ccb); 1162 return (rv); 1163 1164 } 1165 1166 void 1167 nvme_fill_identify(struct nvme_softc *sc, struct nvme_ccb *ccb, void *slot) 1168 { 1169 struct nvme_sqe *sqe = slot; 1170 struct nvme_dmamem *mem = ccb->ccb_cookie; 1171 1172 sqe->opcode = NVM_ADMIN_IDENTIFY; 1173 htolem64(&sqe->entry.prp[0], NVME_DMA_DVA(mem)); 1174 htolem32(&sqe->cdw10, 1); 1175 } 1176 1177 int 1178 nvme_ccbs_alloc(struct nvme_softc *sc, u_int nccbs) 1179 { 1180 struct nvme_ccb *ccb; 1181 bus_addr_t off; 1182 u_int64_t *prpl; 1183 u_int i; 1184 1185 sc->sc_ccbs = mallocarray(nccbs, sizeof(*ccb), M_DEVBUF, 1186 M_WAITOK | M_CANFAIL); 1187 if (sc->sc_ccbs == NULL) 1188 return (1); 1189 1190 sc->sc_ccb_prpls = nvme_dmamem_alloc(sc, 1191 sizeof(*prpl) * sc->sc_max_prpl * nccbs); 1192 1193 prpl = NVME_DMA_KVA(sc->sc_ccb_prpls); 1194 off = 0; 1195 1196 for (i = 0; i < nccbs; i++) { 1197 ccb = &sc->sc_ccbs[i]; 1198 1199 if (bus_dmamap_create(sc->sc_dmat, sc->sc_mdts, 1200 sc->sc_max_prpl + 1, /* we get a free prp in the sqe */ 1201 sc->sc_mps, sc->sc_mps, BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, 1202 &ccb->ccb_dmamap) != 0) 1203 goto free_maps; 1204 1205 ccb->ccb_id = i; 1206 ccb->ccb_prpl = prpl; 1207 ccb->ccb_prpl_off = off; 1208 ccb->ccb_prpl_dva = NVME_DMA_DVA(sc->sc_ccb_prpls) + off; 1209 1210 SIMPLEQ_INSERT_TAIL(&sc->sc_ccb_list, ccb, ccb_entry); 1211 1212 prpl += sc->sc_max_prpl; 1213 off += sizeof(*prpl) * sc->sc_max_prpl; 1214 } 1215 1216 return (0); 1217 1218 free_maps: 1219 nvme_ccbs_free(sc, nccbs); 1220 return (1); 1221 } 1222 1223 void * 1224 nvme_ccb_get(void *cookie) 1225 { 1226 struct nvme_softc *sc = cookie; 1227 struct nvme_ccb *ccb; 1228 1229 mtx_enter(&sc->sc_ccb_mtx); 1230 ccb = SIMPLEQ_FIRST(&sc->sc_ccb_list); 1231 if (ccb != NULL) 1232 SIMPLEQ_REMOVE_HEAD(&sc->sc_ccb_list, ccb_entry); 1233 mtx_leave(&sc->sc_ccb_mtx); 1234 1235 return (ccb); 1236 } 1237 1238 void 1239 nvme_ccb_put(void *cookie, void *io) 1240 { 1241 struct nvme_softc *sc = cookie; 1242 struct nvme_ccb *ccb = io; 1243 1244 mtx_enter(&sc->sc_ccb_mtx); 1245 SIMPLEQ_INSERT_HEAD(&sc->sc_ccb_list, ccb, ccb_entry); 1246 mtx_leave(&sc->sc_ccb_mtx); 1247 } 1248 1249 void 1250 nvme_ccbs_free(struct nvme_softc *sc, unsigned int nccbs) 1251 { 1252 struct nvme_ccb *ccb; 1253 1254 while ((ccb = SIMPLEQ_FIRST(&sc->sc_ccb_list)) != NULL) { 1255 SIMPLEQ_REMOVE_HEAD(&sc->sc_ccb_list, ccb_entry); 1256 bus_dmamap_destroy(sc->sc_dmat, ccb->ccb_dmamap); 1257 } 1258 1259 nvme_dmamem_free(sc, sc->sc_ccb_prpls); 1260 free(sc->sc_ccbs, M_DEVBUF, nccbs * sizeof(*ccb)); 1261 } 1262 1263 struct nvme_queue * 1264 nvme_q_alloc(struct nvme_softc *sc, u_int16_t id, u_int entries, u_int dstrd) 1265 { 1266 struct nvme_queue *q; 1267 1268 q = malloc(sizeof(*q), M_DEVBUF, M_WAITOK | M_CANFAIL); 1269 if (q == NULL) 1270 return (NULL); 1271 1272 q->q_sq_dmamem = nvme_dmamem_alloc(sc, 1273 sizeof(struct nvme_sqe) * entries); 1274 if (q->q_sq_dmamem == NULL) 1275 goto free; 1276 1277 q->q_cq_dmamem = nvme_dmamem_alloc(sc, 1278 sizeof(struct nvme_cqe) * entries); 1279 if (q->q_cq_dmamem == NULL) 1280 goto free_sq; 1281 1282 memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem)); 1283 memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem)); 1284 1285 mtx_init(&q->q_sq_mtx, IPL_BIO); 1286 mtx_init(&q->q_cq_mtx, IPL_BIO); 1287 q->q_sqtdbl = NVME_SQTDBL(id, dstrd); 1288 q->q_cqhdbl = NVME_CQHDBL(id, dstrd); 1289 1290 q->q_id = id; 1291 q->q_entries = entries; 1292 q->q_sq_tail = 0; 1293 q->q_cq_head = 0; 1294 q->q_cq_phase = NVME_CQE_PHASE; 1295 1296 nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE); 1297 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD); 1298 1299 return (q); 1300 1301 free_sq: 1302 nvme_dmamem_free(sc, q->q_sq_dmamem); 1303 free: 1304 free(q, M_DEVBUF, sizeof *q); 1305 1306 return (NULL); 1307 } 1308 1309 int 1310 nvme_q_reset(struct nvme_softc *sc, struct nvme_queue *q) 1311 { 1312 memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem)); 1313 memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem)); 1314 1315 q->q_sqtdbl = NVME_SQTDBL(q->q_id, sc->sc_dstrd); 1316 q->q_cqhdbl = NVME_CQHDBL(q->q_id, sc->sc_dstrd); 1317 1318 q->q_sq_tail = 0; 1319 q->q_cq_head = 0; 1320 q->q_cq_phase = NVME_CQE_PHASE; 1321 1322 nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE); 1323 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD); 1324 1325 return (0); 1326 } 1327 1328 void 1329 nvme_q_free(struct nvme_softc *sc, struct nvme_queue *q) 1330 { 1331 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD); 1332 nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_POSTWRITE); 1333 nvme_dmamem_free(sc, q->q_cq_dmamem); 1334 nvme_dmamem_free(sc, q->q_sq_dmamem); 1335 free(q, M_DEVBUF, sizeof *q); 1336 } 1337 1338 int 1339 nvme_intr(void *xsc) 1340 { 1341 struct nvme_softc *sc = xsc; 1342 int rv = 0; 1343 1344 if (nvme_q_complete(sc, sc->sc_q)) 1345 rv = 1; 1346 if (nvme_q_complete(sc, sc->sc_admin_q)) 1347 rv = 1; 1348 1349 return (rv); 1350 } 1351 1352 int 1353 nvme_intr_intx(void *xsc) 1354 { 1355 struct nvme_softc *sc = xsc; 1356 int rv; 1357 1358 nvme_write4(sc, NVME_INTMS, 1); 1359 rv = nvme_intr(sc); 1360 nvme_write4(sc, NVME_INTMC, 1); 1361 1362 return (rv); 1363 } 1364 1365 struct nvme_dmamem * 1366 nvme_dmamem_alloc(struct nvme_softc *sc, size_t size) 1367 { 1368 struct nvme_dmamem *ndm; 1369 int nsegs; 1370 1371 ndm = malloc(sizeof(*ndm), M_DEVBUF, M_WAITOK | M_ZERO); 1372 if (ndm == NULL) 1373 return (NULL); 1374 1375 ndm->ndm_size = size; 1376 1377 if (bus_dmamap_create(sc->sc_dmat, size, 1, size, 0, 1378 BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &ndm->ndm_map) != 0) 1379 goto ndmfree; 1380 1381 if (bus_dmamem_alloc(sc->sc_dmat, size, sc->sc_mps, 0, &ndm->ndm_seg, 1382 1, &nsegs, BUS_DMA_WAITOK | BUS_DMA_ZERO) != 0) 1383 goto destroy; 1384 1385 if (bus_dmamem_map(sc->sc_dmat, &ndm->ndm_seg, nsegs, size, 1386 &ndm->ndm_kva, BUS_DMA_WAITOK) != 0) 1387 goto free; 1388 1389 if (bus_dmamap_load(sc->sc_dmat, ndm->ndm_map, ndm->ndm_kva, size, 1390 NULL, BUS_DMA_WAITOK) != 0) 1391 goto unmap; 1392 1393 return (ndm); 1394 1395 unmap: 1396 bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, size); 1397 free: 1398 bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1); 1399 destroy: 1400 bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map); 1401 ndmfree: 1402 free(ndm, M_DEVBUF, sizeof *ndm); 1403 1404 return (NULL); 1405 } 1406 1407 void 1408 nvme_dmamem_sync(struct nvme_softc *sc, struct nvme_dmamem *mem, int ops) 1409 { 1410 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(mem), 1411 0, NVME_DMA_LEN(mem), ops); 1412 } 1413 1414 void 1415 nvme_dmamem_free(struct nvme_softc *sc, struct nvme_dmamem *ndm) 1416 { 1417 bus_dmamap_unload(sc->sc_dmat, ndm->ndm_map); 1418 bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, ndm->ndm_size); 1419 bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1); 1420 bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map); 1421 free(ndm, M_DEVBUF, sizeof *ndm); 1422 } 1423 1424 #ifdef HIBERNATE 1425 1426 int 1427 nvme_hibernate_admin_cmd(struct nvme_softc *sc, struct nvme_sqe *sqe, 1428 struct nvme_cqe *cqe, int cid) 1429 { 1430 struct nvme_sqe *asqe = NVME_DMA_KVA(sc->sc_admin_q->q_sq_dmamem); 1431 struct nvme_cqe *acqe = NVME_DMA_KVA(sc->sc_admin_q->q_cq_dmamem); 1432 struct nvme_queue *q = sc->sc_admin_q; 1433 int tail; 1434 u_int16_t flags; 1435 1436 /* submit command */ 1437 tail = q->q_sq_tail; 1438 if (++q->q_sq_tail >= q->q_entries) 1439 q->q_sq_tail = 0; 1440 1441 asqe += tail; 1442 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 1443 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE); 1444 *asqe = *sqe; 1445 asqe->cid = cid; 1446 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 1447 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE); 1448 1449 nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail); 1450 1451 /* wait for completion */ 1452 acqe += q->q_cq_head; 1453 for (;;) { 1454 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD); 1455 flags = lemtoh16(&acqe->flags); 1456 if ((flags & NVME_CQE_PHASE) == q->q_cq_phase) 1457 break; 1458 1459 delay(10); 1460 } 1461 1462 if (++q->q_cq_head >= q->q_entries) { 1463 q->q_cq_head = 0; 1464 q->q_cq_phase ^= NVME_CQE_PHASE; 1465 } 1466 nvme_write4(sc, q->q_cqhdbl, q->q_cq_head); 1467 if ((NVME_CQE_SC(flags) != NVME_CQE_SC_SUCCESS) || (acqe->cid != cid)) 1468 return (EIO); 1469 1470 return (0); 1471 } 1472 1473 int 1474 nvme_hibernate_io(dev_t dev, daddr_t blkno, vaddr_t addr, size_t size, 1475 int op, void *page) 1476 { 1477 struct nvme_hibernate_page { 1478 u_int64_t prpl[MAXPHYS / PAGE_SIZE]; 1479 1480 struct nvme_softc *sc; 1481 int nsid; 1482 int sq_tail; 1483 int cq_head; 1484 int cqe_phase; 1485 1486 daddr_t poffset; 1487 size_t psize; 1488 } *my = page; 1489 struct nvme_sqe_io *isqe; 1490 struct nvme_cqe *icqe; 1491 paddr_t data_phys, page_phys; 1492 u_int64_t data_bus_phys, page_bus_phys; 1493 u_int16_t flags; 1494 int i; 1495 1496 if (op == HIB_INIT) { 1497 struct device *disk; 1498 struct device *scsibus; 1499 extern struct cfdriver sd_cd; 1500 struct scsi_link *link; 1501 struct scsibus_softc *bus_sc; 1502 struct nvme_sqe_q qsqe; 1503 struct nvme_cqe qcqe; 1504 1505 /* find nvme softc */ 1506 disk = disk_lookup(&sd_cd, DISKUNIT(dev)); 1507 scsibus = disk->dv_parent; 1508 my->sc = (struct nvme_softc *)disk->dv_parent->dv_parent; 1509 1510 /* find scsi_link, which tells us the target */ 1511 my->nsid = 0; 1512 bus_sc = (struct scsibus_softc *)scsibus; 1513 SLIST_FOREACH(link, &bus_sc->sc_link_list, bus_list) { 1514 if (link->device_softc == disk) { 1515 my->nsid = link->target; 1516 break; 1517 } 1518 } 1519 if (my->nsid == 0) 1520 return (EIO); 1521 1522 my->poffset = blkno; 1523 my->psize = size; 1524 1525 memset(NVME_DMA_KVA(my->sc->sc_hib_q->q_cq_dmamem), 0, 1526 my->sc->sc_hib_q->q_entries * sizeof(struct nvme_cqe)); 1527 memset(NVME_DMA_KVA(my->sc->sc_hib_q->q_sq_dmamem), 0, 1528 my->sc->sc_hib_q->q_entries * sizeof(struct nvme_sqe)); 1529 1530 my->sq_tail = 0; 1531 my->cq_head = 0; 1532 my->cqe_phase = NVME_CQE_PHASE; 1533 1534 pmap_extract(pmap_kernel(), (vaddr_t)page, &page_phys); 1535 1536 memset(&qsqe, 0, sizeof(qsqe)); 1537 qsqe.opcode = NVM_ADMIN_ADD_IOCQ; 1538 htolem64(&qsqe.prp1, 1539 NVME_DMA_DVA(my->sc->sc_hib_q->q_cq_dmamem)); 1540 htolem16(&qsqe.qsize, my->sc->sc_hib_q->q_entries - 1); 1541 htolem16(&qsqe.qid, my->sc->sc_hib_q->q_id); 1542 qsqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC; 1543 if (nvme_hibernate_admin_cmd(my->sc, (struct nvme_sqe *)&qsqe, 1544 &qcqe, 1) != 0) 1545 return (EIO); 1546 1547 memset(&qsqe, 0, sizeof(qsqe)); 1548 qsqe.opcode = NVM_ADMIN_ADD_IOSQ; 1549 htolem64(&qsqe.prp1, 1550 NVME_DMA_DVA(my->sc->sc_hib_q->q_sq_dmamem)); 1551 htolem16(&qsqe.qsize, my->sc->sc_hib_q->q_entries - 1); 1552 htolem16(&qsqe.qid, my->sc->sc_hib_q->q_id); 1553 htolem16(&qsqe.cqid, my->sc->sc_hib_q->q_id); 1554 qsqe.qflags = NVM_SQE_Q_PC; 1555 if (nvme_hibernate_admin_cmd(my->sc, (struct nvme_sqe *)&qsqe, 1556 &qcqe, 2) != 0) 1557 return (EIO); 1558 1559 return (0); 1560 } 1561 1562 if (op != HIB_W) 1563 return (0); 1564 1565 isqe = NVME_DMA_KVA(my->sc->sc_hib_q->q_sq_dmamem); 1566 isqe += my->sq_tail; 1567 if (++my->sq_tail == my->sc->sc_hib_q->q_entries) 1568 my->sq_tail = 0; 1569 1570 memset(isqe, 0, sizeof(*isqe)); 1571 isqe->opcode = NVM_CMD_WRITE; 1572 htolem32(&isqe->nsid, my->nsid); 1573 1574 pmap_extract(pmap_kernel(), addr, &data_phys); 1575 data_bus_phys = data_phys; 1576 htolem64(&isqe->entry.prp[0], data_bus_phys); 1577 if ((size > my->sc->sc_mps) && (size <= my->sc->sc_mps * 2)) { 1578 htolem64(&isqe->entry.prp[1], data_bus_phys + my->sc->sc_mps); 1579 } else if (size > my->sc->sc_mps * 2) { 1580 pmap_extract(pmap_kernel(), (vaddr_t)page, &page_phys); 1581 page_bus_phys = page_phys; 1582 htolem64(&isqe->entry.prp[1], page_bus_phys + 1583 offsetof(struct nvme_hibernate_page, prpl)); 1584 for (i = 1; i < (size / my->sc->sc_mps); i++) { 1585 htolem64(&my->prpl[i - 1], data_bus_phys + 1586 (i * my->sc->sc_mps)); 1587 } 1588 } 1589 1590 isqe->slba = blkno + my->poffset; 1591 isqe->nlb = (size / DEV_BSIZE) - 1; 1592 isqe->cid = blkno % 0xffff; 1593 1594 nvme_write4(my->sc, NVME_SQTDBL(NVME_HIB_Q, my->sc->sc_dstrd), 1595 my->sq_tail); 1596 1597 icqe = NVME_DMA_KVA(my->sc->sc_hib_q->q_cq_dmamem); 1598 icqe += my->cq_head; 1599 for (;;) { 1600 flags = lemtoh16(&icqe->flags); 1601 if ((flags & NVME_CQE_PHASE) == my->cqe_phase) 1602 break; 1603 1604 delay(10); 1605 } 1606 1607 if (++my->cq_head == my->sc->sc_hib_q->q_entries) { 1608 my->cq_head = 0; 1609 my->cqe_phase ^= NVME_CQE_PHASE; 1610 } 1611 nvme_write4(my->sc, NVME_CQHDBL(NVME_HIB_Q, my->sc->sc_dstrd), 1612 my->cq_head); 1613 if ((NVME_CQE_SC(flags) != NVME_CQE_SC_SUCCESS) || 1614 (icqe->cid != blkno % 0xffff)) 1615 return (EIO); 1616 1617 return (0); 1618 } 1619 1620 #endif 1621