1 /* $NetBSD: nvme.c,v 1.49 2020/07/28 15:59:19 jdolecek Exp $ */ 2 /* $OpenBSD: nvme.c,v 1.49 2016/04/18 05:59:50 dlg Exp $ */ 3 4 /* 5 * Copyright (c) 2014 David Gwynne <dlg@openbsd.org> 6 * 7 * Permission to use, copy, modify, and distribute this software for any 8 * purpose with or without fee is hereby granted, provided that the above 9 * copyright notice and this permission notice appear in all copies. 10 * 11 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 12 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 13 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 14 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 15 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 16 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 17 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 18 */ 19 20 #include <sys/cdefs.h> 21 __KERNEL_RCSID(0, "$NetBSD: nvme.c,v 1.49 2020/07/28 15:59:19 jdolecek Exp $"); 22 23 #include <sys/param.h> 24 #include <sys/systm.h> 25 #include <sys/kernel.h> 26 #include <sys/atomic.h> 27 #include <sys/bus.h> 28 #include <sys/buf.h> 29 #include <sys/conf.h> 30 #include <sys/device.h> 31 #include <sys/kmem.h> 32 #include <sys/once.h> 33 #include <sys/proc.h> 34 #include <sys/queue.h> 35 #include <sys/mutex.h> 36 37 #include <uvm/uvm_extern.h> 38 39 #include <dev/ic/nvmereg.h> 40 #include <dev/ic/nvmevar.h> 41 #include <dev/ic/nvmeio.h> 42 43 #include "ioconf.h" 44 45 #define B4_CHK_RDY_DELAY_MS 2300 /* workaround controller bug */ 46 47 int nvme_adminq_size = 32; 48 int nvme_ioq_size = 1024; 49 50 static int nvme_print(void *, const char *); 51 52 static int nvme_ready(struct nvme_softc *, uint32_t); 53 static int nvme_enable(struct nvme_softc *, u_int); 54 static int nvme_disable(struct nvme_softc *); 55 static int nvme_shutdown(struct nvme_softc *); 56 57 #ifdef NVME_DEBUG 58 static void nvme_dumpregs(struct nvme_softc *); 59 #endif 60 static int nvme_identify(struct nvme_softc *, u_int); 61 static void nvme_fill_identify(struct nvme_queue *, struct nvme_ccb *, 62 void *); 63 64 static int nvme_ccbs_alloc(struct nvme_queue *, uint16_t); 65 static void nvme_ccbs_free(struct nvme_queue *); 66 67 static struct nvme_ccb * 68 nvme_ccb_get(struct nvme_queue *, bool); 69 static void nvme_ccb_put(struct nvme_queue *, struct nvme_ccb *); 70 71 static int nvme_poll(struct nvme_softc *, struct nvme_queue *, 72 struct nvme_ccb *, void (*)(struct nvme_queue *, 73 struct nvme_ccb *, void *), int); 74 static void nvme_poll_fill(struct nvme_queue *, struct nvme_ccb *, void *); 75 static void nvme_poll_done(struct nvme_queue *, struct nvme_ccb *, 76 struct nvme_cqe *); 77 static void nvme_sqe_fill(struct nvme_queue *, struct nvme_ccb *, void *); 78 static void nvme_empty_done(struct nvme_queue *, struct nvme_ccb *, 79 struct nvme_cqe *); 80 81 static struct nvme_queue * 82 nvme_q_alloc(struct nvme_softc *, uint16_t, u_int, u_int); 83 static int nvme_q_create(struct nvme_softc *, struct nvme_queue *); 84 static int nvme_q_delete(struct nvme_softc *, struct nvme_queue *); 85 static void nvme_q_submit(struct nvme_softc *, struct nvme_queue *, 86 struct nvme_ccb *, void (*)(struct nvme_queue *, 87 struct nvme_ccb *, void *)); 88 static int nvme_q_complete(struct nvme_softc *, struct nvme_queue *q); 89 static void nvme_q_free(struct nvme_softc *, struct nvme_queue *); 90 static void nvme_q_wait_complete(struct nvme_softc *, struct nvme_queue *, 91 bool (*)(void *), void *); 92 93 static struct nvme_dmamem * 94 nvme_dmamem_alloc(struct nvme_softc *, size_t); 95 static void nvme_dmamem_free(struct nvme_softc *, struct nvme_dmamem *); 96 static void nvme_dmamem_sync(struct nvme_softc *, struct nvme_dmamem *, 97 int); 98 99 static void nvme_ns_io_fill(struct nvme_queue *, struct nvme_ccb *, 100 void *); 101 static void nvme_ns_io_done(struct nvme_queue *, struct nvme_ccb *, 102 struct nvme_cqe *); 103 static void nvme_ns_sync_fill(struct nvme_queue *, struct nvme_ccb *, 104 void *); 105 static void nvme_ns_sync_done(struct nvme_queue *, struct nvme_ccb *, 106 struct nvme_cqe *); 107 static void nvme_getcache_fill(struct nvme_queue *, struct nvme_ccb *, 108 void *); 109 static void nvme_getcache_done(struct nvme_queue *, struct nvme_ccb *, 110 struct nvme_cqe *); 111 112 static void nvme_pt_fill(struct nvme_queue *, struct nvme_ccb *, 113 void *); 114 static void nvme_pt_done(struct nvme_queue *, struct nvme_ccb *, 115 struct nvme_cqe *); 116 static int nvme_command_passthrough(struct nvme_softc *, 117 struct nvme_pt_command *, uint16_t, struct lwp *, bool); 118 119 static int nvme_set_number_of_queues(struct nvme_softc *, u_int, u_int *, 120 u_int *); 121 122 #define NVME_TIMO_QOP 5 /* queue create and delete timeout */ 123 #define NVME_TIMO_IDENT 10 /* probe identify timeout */ 124 #define NVME_TIMO_PT -1 /* passthrough cmd timeout */ 125 #define NVME_TIMO_SY 60 /* sync cache timeout */ 126 127 #define nvme_read4(_s, _r) \ 128 bus_space_read_4((_s)->sc_iot, (_s)->sc_ioh, (_r)) 129 #define nvme_write4(_s, _r, _v) \ 130 bus_space_write_4((_s)->sc_iot, (_s)->sc_ioh, (_r), (_v)) 131 /* 132 * Some controllers, at least Apple NVMe, always require split 133 * transfers, so don't use bus_space_{read,write}_8() on LP64. 134 */ 135 static inline uint64_t 136 nvme_read8(struct nvme_softc *sc, bus_size_t r) 137 { 138 uint64_t v; 139 uint32_t *a = (uint32_t *)&v; 140 141 #if _BYTE_ORDER == _LITTLE_ENDIAN 142 a[0] = nvme_read4(sc, r); 143 a[1] = nvme_read4(sc, r + 4); 144 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */ 145 a[1] = nvme_read4(sc, r); 146 a[0] = nvme_read4(sc, r + 4); 147 #endif 148 149 return v; 150 } 151 152 static inline void 153 nvme_write8(struct nvme_softc *sc, bus_size_t r, uint64_t v) 154 { 155 uint32_t *a = (uint32_t *)&v; 156 157 #if _BYTE_ORDER == _LITTLE_ENDIAN 158 nvme_write4(sc, r, a[0]); 159 nvme_write4(sc, r + 4, a[1]); 160 #else /* _BYTE_ORDER == _LITTLE_ENDIAN */ 161 nvme_write4(sc, r, a[1]); 162 nvme_write4(sc, r + 4, a[0]); 163 #endif 164 } 165 #define nvme_barrier(_s, _r, _l, _f) \ 166 bus_space_barrier((_s)->sc_iot, (_s)->sc_ioh, (_r), (_l), (_f)) 167 168 #ifdef NVME_DEBUG 169 static __used void 170 nvme_dumpregs(struct nvme_softc *sc) 171 { 172 uint64_t r8; 173 uint32_t r4; 174 175 #define DEVNAME(_sc) device_xname((_sc)->sc_dev) 176 r8 = nvme_read8(sc, NVME_CAP); 177 printf("%s: cap 0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_CAP)); 178 printf("%s: mpsmax %u (%u)\n", DEVNAME(sc), 179 (u_int)NVME_CAP_MPSMAX(r8), (1 << NVME_CAP_MPSMAX(r8))); 180 printf("%s: mpsmin %u (%u)\n", DEVNAME(sc), 181 (u_int)NVME_CAP_MPSMIN(r8), (1 << NVME_CAP_MPSMIN(r8))); 182 printf("%s: css %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CSS(r8)); 183 printf("%s: nssrs %"PRIu64"\n", DEVNAME(sc), NVME_CAP_NSSRS(r8)); 184 printf("%s: dstrd %"PRIu64"\n", DEVNAME(sc), NVME_CAP_DSTRD(r8)); 185 printf("%s: to %"PRIu64" msec\n", DEVNAME(sc), NVME_CAP_TO(r8)); 186 printf("%s: ams %"PRIu64"\n", DEVNAME(sc), NVME_CAP_AMS(r8)); 187 printf("%s: cqr %"PRIu64"\n", DEVNAME(sc), NVME_CAP_CQR(r8)); 188 printf("%s: mqes %"PRIu64"\n", DEVNAME(sc), NVME_CAP_MQES(r8)); 189 190 printf("%s: vs 0x%04x\n", DEVNAME(sc), nvme_read4(sc, NVME_VS)); 191 192 r4 = nvme_read4(sc, NVME_CC); 193 printf("%s: cc 0x%04x\n", DEVNAME(sc), r4); 194 printf("%s: iocqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOCQES_R(r4), 195 (1 << NVME_CC_IOCQES_R(r4))); 196 printf("%s: iosqes %u (%u)\n", DEVNAME(sc), NVME_CC_IOSQES_R(r4), 197 (1 << NVME_CC_IOSQES_R(r4))); 198 printf("%s: shn %u\n", DEVNAME(sc), NVME_CC_SHN_R(r4)); 199 printf("%s: ams %u\n", DEVNAME(sc), NVME_CC_AMS_R(r4)); 200 printf("%s: mps %u (%u)\n", DEVNAME(sc), NVME_CC_MPS_R(r4), 201 (1 << NVME_CC_MPS_R(r4))); 202 printf("%s: css %u\n", DEVNAME(sc), NVME_CC_CSS_R(r4)); 203 printf("%s: en %u\n", DEVNAME(sc), ISSET(r4, NVME_CC_EN) ? 1 : 0); 204 205 r4 = nvme_read4(sc, NVME_CSTS); 206 printf("%s: csts 0x%08x\n", DEVNAME(sc), r4); 207 printf("%s: rdy %u\n", DEVNAME(sc), r4 & NVME_CSTS_RDY); 208 printf("%s: cfs %u\n", DEVNAME(sc), r4 & NVME_CSTS_CFS); 209 printf("%s: shst %x\n", DEVNAME(sc), r4 & NVME_CSTS_SHST_MASK); 210 211 r4 = nvme_read4(sc, NVME_AQA); 212 printf("%s: aqa 0x%08x\n", DEVNAME(sc), r4); 213 printf("%s: acqs %u\n", DEVNAME(sc), NVME_AQA_ACQS_R(r4)); 214 printf("%s: asqs %u\n", DEVNAME(sc), NVME_AQA_ASQS_R(r4)); 215 216 printf("%s: asq 0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ASQ)); 217 printf("%s: acq 0x%016"PRIx64"\n", DEVNAME(sc), nvme_read8(sc, NVME_ACQ)); 218 #undef DEVNAME 219 } 220 #endif /* NVME_DEBUG */ 221 222 static int 223 nvme_ready(struct nvme_softc *sc, uint32_t rdy) 224 { 225 u_int i = 0; 226 227 while ((nvme_read4(sc, NVME_CSTS) & NVME_CSTS_RDY) != rdy) { 228 if (i++ > sc->sc_rdy_to) 229 return ENXIO; 230 231 delay(1000); 232 nvme_barrier(sc, NVME_CSTS, 4, BUS_SPACE_BARRIER_READ); 233 } 234 235 return 0; 236 } 237 238 static int 239 nvme_enable(struct nvme_softc *sc, u_int mps) 240 { 241 uint32_t cc, csts; 242 int error; 243 244 cc = nvme_read4(sc, NVME_CC); 245 csts = nvme_read4(sc, NVME_CSTS); 246 247 /* 248 * See note in nvme_disable. Short circuit if we're already enabled. 249 */ 250 if (ISSET(cc, NVME_CC_EN)) { 251 if (ISSET(csts, NVME_CSTS_RDY)) 252 return 0; 253 254 goto waitready; 255 } else { 256 /* EN == 0 already wait for RDY == 0 or fail */ 257 error = nvme_ready(sc, 0); 258 if (error) 259 return error; 260 } 261 262 nvme_write8(sc, NVME_ASQ, NVME_DMA_DVA(sc->sc_admin_q->q_sq_dmamem)); 263 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 264 delay(5000); 265 nvme_write8(sc, NVME_ACQ, NVME_DMA_DVA(sc->sc_admin_q->q_cq_dmamem)); 266 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 267 delay(5000); 268 269 nvme_write4(sc, NVME_AQA, NVME_AQA_ACQS(sc->sc_admin_q->q_entries) | 270 NVME_AQA_ASQS(sc->sc_admin_q->q_entries)); 271 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_WRITE); 272 delay(5000); 273 274 CLR(cc, NVME_CC_IOCQES_MASK | NVME_CC_IOSQES_MASK | NVME_CC_SHN_MASK | 275 NVME_CC_AMS_MASK | NVME_CC_MPS_MASK | NVME_CC_CSS_MASK); 276 SET(cc, NVME_CC_IOSQES(ffs(64) - 1) | NVME_CC_IOCQES(ffs(16) - 1)); 277 SET(cc, NVME_CC_SHN(NVME_CC_SHN_NONE)); 278 SET(cc, NVME_CC_CSS(NVME_CC_CSS_NVM)); 279 SET(cc, NVME_CC_AMS(NVME_CC_AMS_RR)); 280 SET(cc, NVME_CC_MPS(mps)); 281 SET(cc, NVME_CC_EN); 282 283 nvme_write4(sc, NVME_CC, cc); 284 nvme_barrier(sc, 0, sc->sc_ios, 285 BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE); 286 287 waitready: 288 return nvme_ready(sc, NVME_CSTS_RDY); 289 } 290 291 static int 292 nvme_disable(struct nvme_softc *sc) 293 { 294 uint32_t cc, csts; 295 int error; 296 297 cc = nvme_read4(sc, NVME_CC); 298 csts = nvme_read4(sc, NVME_CSTS); 299 300 /* 301 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1 302 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when 303 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY 304 * isn't the desired value. Short circuit if we're already disabled. 305 */ 306 if (ISSET(cc, NVME_CC_EN)) { 307 if (!ISSET(csts, NVME_CSTS_RDY)) { 308 /* EN == 1, wait for RDY == 1 or fail */ 309 error = nvme_ready(sc, NVME_CSTS_RDY); 310 if (error) 311 return error; 312 } 313 } else { 314 /* EN == 0 already wait for RDY == 0 */ 315 if (!ISSET(csts, NVME_CSTS_RDY)) 316 return 0; 317 318 goto waitready; 319 } 320 321 CLR(cc, NVME_CC_EN); 322 nvme_write4(sc, NVME_CC, cc); 323 nvme_barrier(sc, 0, sc->sc_ios, BUS_SPACE_BARRIER_READ); 324 325 /* 326 * Some drives have issues with accessing the mmio after we disable, 327 * so delay for a bit after we write the bit to cope with these issues. 328 */ 329 if (ISSET(sc->sc_quirks, NVME_QUIRK_DELAY_B4_CHK_RDY)) 330 delay(B4_CHK_RDY_DELAY_MS); 331 332 waitready: 333 return nvme_ready(sc, 0); 334 } 335 336 int 337 nvme_attach(struct nvme_softc *sc) 338 { 339 uint64_t cap; 340 uint32_t reg; 341 u_int dstrd; 342 u_int mps = PAGE_SHIFT; 343 u_int ncq, nsq; 344 uint16_t adminq_entries = nvme_adminq_size; 345 uint16_t ioq_entries = nvme_ioq_size; 346 int i; 347 348 reg = nvme_read4(sc, NVME_VS); 349 if (reg == 0xffffffff) { 350 aprint_error_dev(sc->sc_dev, "invalid mapping\n"); 351 return 1; 352 } 353 354 if (NVME_VS_TER(reg) == 0) 355 aprint_normal_dev(sc->sc_dev, "NVMe %d.%d\n", NVME_VS_MJR(reg), 356 NVME_VS_MNR(reg)); 357 else 358 aprint_normal_dev(sc->sc_dev, "NVMe %d.%d.%d\n", NVME_VS_MJR(reg), 359 NVME_VS_MNR(reg), NVME_VS_TER(reg)); 360 361 cap = nvme_read8(sc, NVME_CAP); 362 dstrd = NVME_CAP_DSTRD(cap); 363 if (NVME_CAP_MPSMIN(cap) > PAGE_SHIFT) { 364 aprint_error_dev(sc->sc_dev, "NVMe minimum page size %u " 365 "is greater than CPU page size %u\n", 366 1 << NVME_CAP_MPSMIN(cap), 1 << PAGE_SHIFT); 367 return 1; 368 } 369 if (NVME_CAP_MPSMAX(cap) < mps) 370 mps = NVME_CAP_MPSMAX(cap); 371 if (ioq_entries > NVME_CAP_MQES(cap)) 372 ioq_entries = NVME_CAP_MQES(cap); 373 374 /* set initial values to be used for admin queue during probe */ 375 sc->sc_rdy_to = NVME_CAP_TO(cap); 376 sc->sc_mps = 1 << mps; 377 sc->sc_mdts = MAXPHYS; 378 sc->sc_max_sgl = btoc(round_page(sc->sc_mdts)); 379 380 if (nvme_disable(sc) != 0) { 381 aprint_error_dev(sc->sc_dev, "unable to disable controller\n"); 382 return 1; 383 } 384 385 sc->sc_admin_q = nvme_q_alloc(sc, NVME_ADMIN_Q, adminq_entries, dstrd); 386 if (sc->sc_admin_q == NULL) { 387 aprint_error_dev(sc->sc_dev, 388 "unable to allocate admin queue\n"); 389 return 1; 390 } 391 if (sc->sc_intr_establish(sc, NVME_ADMIN_Q, sc->sc_admin_q)) 392 goto free_admin_q; 393 394 if (nvme_enable(sc, mps) != 0) { 395 aprint_error_dev(sc->sc_dev, "unable to enable controller\n"); 396 goto disestablish_admin_q; 397 } 398 399 if (nvme_identify(sc, NVME_CAP_MPSMIN(cap)) != 0) { 400 aprint_error_dev(sc->sc_dev, "unable to identify controller\n"); 401 goto disable; 402 } 403 if (sc->sc_nn == 0) { 404 aprint_error_dev(sc->sc_dev, "namespace not found\n"); 405 goto disable; 406 } 407 408 /* we know how big things are now */ 409 sc->sc_max_sgl = sc->sc_mdts / sc->sc_mps; 410 411 /* reallocate ccbs of admin queue with new max sgl. */ 412 nvme_ccbs_free(sc->sc_admin_q); 413 nvme_ccbs_alloc(sc->sc_admin_q, sc->sc_admin_q->q_entries); 414 415 if (sc->sc_use_mq) { 416 /* Limit the number of queues to the number allocated in HW */ 417 if (nvme_set_number_of_queues(sc, sc->sc_nq, &ncq, &nsq) != 0) { 418 aprint_error_dev(sc->sc_dev, 419 "unable to get number of queues\n"); 420 goto disable; 421 } 422 if (sc->sc_nq > ncq) 423 sc->sc_nq = ncq; 424 if (sc->sc_nq > nsq) 425 sc->sc_nq = nsq; 426 } 427 428 sc->sc_q = kmem_zalloc(sizeof(*sc->sc_q) * sc->sc_nq, KM_SLEEP); 429 for (i = 0; i < sc->sc_nq; i++) { 430 sc->sc_q[i] = nvme_q_alloc(sc, i + 1, ioq_entries, dstrd); 431 if (sc->sc_q[i] == NULL) { 432 aprint_error_dev(sc->sc_dev, 433 "unable to allocate io queue\n"); 434 goto free_q; 435 } 436 if (nvme_q_create(sc, sc->sc_q[i]) != 0) { 437 aprint_error_dev(sc->sc_dev, 438 "unable to create io queue\n"); 439 nvme_q_free(sc, sc->sc_q[i]); 440 goto free_q; 441 } 442 } 443 444 if (!sc->sc_use_mq) 445 nvme_write4(sc, NVME_INTMC, 1); 446 447 /* probe subdevices */ 448 sc->sc_namespaces = kmem_zalloc(sizeof(*sc->sc_namespaces) * sc->sc_nn, 449 KM_SLEEP); 450 nvme_rescan(sc->sc_dev, "nvme", &i); 451 452 return 0; 453 454 free_q: 455 while (--i >= 0) { 456 nvme_q_delete(sc, sc->sc_q[i]); 457 nvme_q_free(sc, sc->sc_q[i]); 458 } 459 disable: 460 nvme_disable(sc); 461 disestablish_admin_q: 462 sc->sc_intr_disestablish(sc, NVME_ADMIN_Q); 463 free_admin_q: 464 nvme_q_free(sc, sc->sc_admin_q); 465 466 return 1; 467 } 468 469 int 470 nvme_rescan(device_t self, const char *attr, const int *flags) 471 { 472 struct nvme_softc *sc = device_private(self); 473 struct nvme_attach_args naa; 474 uint64_t cap; 475 int ioq_entries = nvme_ioq_size; 476 int i; 477 478 cap = nvme_read8(sc, NVME_CAP); 479 if (ioq_entries > NVME_CAP_MQES(cap)) 480 ioq_entries = NVME_CAP_MQES(cap); 481 482 for (i = 0; i < sc->sc_nn; i++) { 483 if (sc->sc_namespaces[i].dev) 484 continue; 485 memset(&naa, 0, sizeof(naa)); 486 naa.naa_nsid = i + 1; 487 naa.naa_qentries = (ioq_entries - 1) * sc->sc_nq; 488 naa.naa_maxphys = sc->sc_mdts; 489 naa.naa_typename = sc->sc_modelname; 490 sc->sc_namespaces[i].dev = config_found(sc->sc_dev, &naa, 491 nvme_print); 492 } 493 return 0; 494 } 495 496 static int 497 nvme_print(void *aux, const char *pnp) 498 { 499 struct nvme_attach_args *naa = aux; 500 501 if (pnp) 502 aprint_normal("ld at %s", pnp); 503 504 if (naa->naa_nsid > 0) 505 aprint_normal(" nsid %d", naa->naa_nsid); 506 507 return UNCONF; 508 } 509 510 int 511 nvme_detach(struct nvme_softc *sc, int flags) 512 { 513 int i, error; 514 515 error = config_detach_children(sc->sc_dev, flags); 516 if (error) 517 return error; 518 519 error = nvme_shutdown(sc); 520 if (error) 521 return error; 522 523 /* from now on we are committed to detach, following will never fail */ 524 for (i = 0; i < sc->sc_nq; i++) 525 nvme_q_free(sc, sc->sc_q[i]); 526 kmem_free(sc->sc_q, sizeof(*sc->sc_q) * sc->sc_nq); 527 nvme_q_free(sc, sc->sc_admin_q); 528 529 return 0; 530 } 531 532 static int 533 nvme_shutdown(struct nvme_softc *sc) 534 { 535 uint32_t cc, csts; 536 bool disabled = false; 537 int i; 538 539 if (!sc->sc_use_mq) 540 nvme_write4(sc, NVME_INTMS, 1); 541 542 for (i = 0; i < sc->sc_nq; i++) { 543 if (nvme_q_delete(sc, sc->sc_q[i]) != 0) { 544 aprint_error_dev(sc->sc_dev, 545 "unable to delete io queue %d, disabling\n", i + 1); 546 disabled = true; 547 } 548 } 549 sc->sc_intr_disestablish(sc, NVME_ADMIN_Q); 550 if (disabled) 551 goto disable; 552 553 cc = nvme_read4(sc, NVME_CC); 554 CLR(cc, NVME_CC_SHN_MASK); 555 SET(cc, NVME_CC_SHN(NVME_CC_SHN_NORMAL)); 556 nvme_write4(sc, NVME_CC, cc); 557 558 for (i = 0; i < 4000; i++) { 559 nvme_barrier(sc, 0, sc->sc_ios, 560 BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE); 561 csts = nvme_read4(sc, NVME_CSTS); 562 if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_DONE) 563 return 0; 564 565 delay(1000); 566 } 567 568 aprint_error_dev(sc->sc_dev, "unable to shudown, disabling\n"); 569 570 disable: 571 nvme_disable(sc); 572 return 0; 573 } 574 575 void 576 nvme_childdet(device_t self, device_t child) 577 { 578 struct nvme_softc *sc = device_private(self); 579 int i; 580 581 for (i = 0; i < sc->sc_nn; i++) { 582 if (sc->sc_namespaces[i].dev == child) { 583 /* Already freed ns->ident. */ 584 sc->sc_namespaces[i].dev = NULL; 585 break; 586 } 587 } 588 } 589 590 int 591 nvme_ns_identify(struct nvme_softc *sc, uint16_t nsid) 592 { 593 struct nvme_sqe sqe; 594 struct nvm_identify_namespace *identify; 595 struct nvme_dmamem *mem; 596 struct nvme_ccb *ccb; 597 struct nvme_namespace *ns; 598 int rv; 599 600 KASSERT(nsid > 0); 601 602 ccb = nvme_ccb_get(sc->sc_admin_q, false); 603 KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */ 604 605 mem = nvme_dmamem_alloc(sc, sizeof(*identify)); 606 if (mem == NULL) { 607 nvme_ccb_put(sc->sc_admin_q, ccb); 608 return ENOMEM; 609 } 610 611 memset(&sqe, 0, sizeof(sqe)); 612 sqe.opcode = NVM_ADMIN_IDENTIFY; 613 htolem32(&sqe.nsid, nsid); 614 htolem64(&sqe.entry.prp[0], NVME_DMA_DVA(mem)); 615 htolem32(&sqe.cdw10, 0); 616 617 ccb->ccb_done = nvme_empty_done; 618 ccb->ccb_cookie = &sqe; 619 620 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD); 621 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_IDENT); 622 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD); 623 624 nvme_ccb_put(sc->sc_admin_q, ccb); 625 626 if (rv != 0) { 627 rv = EIO; 628 goto done; 629 } 630 631 /* commit */ 632 633 identify = kmem_zalloc(sizeof(*identify), KM_SLEEP); 634 *identify = *((volatile struct nvm_identify_namespace *)NVME_DMA_KVA(mem)); 635 636 /* Convert data to host endian */ 637 nvme_identify_namespace_swapbytes(identify); 638 639 ns = nvme_ns_get(sc, nsid); 640 KASSERT(ns); 641 KASSERT(ns->ident == NULL); 642 ns->ident = identify; 643 644 done: 645 nvme_dmamem_free(sc, mem); 646 647 return rv; 648 } 649 650 int 651 nvme_ns_dobio(struct nvme_softc *sc, uint16_t nsid, void *cookie, 652 struct buf *bp, void *data, size_t datasize, 653 int secsize, daddr_t blkno, int flags, nvme_nnc_done nnc_done) 654 { 655 struct nvme_queue *q = nvme_get_q(sc, bp, false); 656 struct nvme_ccb *ccb; 657 bus_dmamap_t dmap; 658 int i, error; 659 660 ccb = nvme_ccb_get(q, false); 661 if (ccb == NULL) 662 return EAGAIN; 663 664 ccb->ccb_done = nvme_ns_io_done; 665 ccb->ccb_cookie = cookie; 666 667 /* namespace context */ 668 ccb->nnc_nsid = nsid; 669 ccb->nnc_flags = flags; 670 ccb->nnc_buf = bp; 671 ccb->nnc_datasize = datasize; 672 ccb->nnc_secsize = secsize; 673 ccb->nnc_blkno = blkno; 674 ccb->nnc_done = nnc_done; 675 676 dmap = ccb->ccb_dmamap; 677 error = bus_dmamap_load(sc->sc_dmat, dmap, data, 678 datasize, NULL, 679 (ISSET(flags, NVME_NS_CTX_F_POLL) ? 680 BUS_DMA_NOWAIT : BUS_DMA_WAITOK) | 681 (ISSET(flags, NVME_NS_CTX_F_READ) ? 682 BUS_DMA_READ : BUS_DMA_WRITE)); 683 if (error) { 684 nvme_ccb_put(q, ccb); 685 return error; 686 } 687 688 bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize, 689 ISSET(flags, NVME_NS_CTX_F_READ) ? 690 BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE); 691 692 if (dmap->dm_nsegs > 2) { 693 for (i = 1; i < dmap->dm_nsegs; i++) { 694 htolem64(&ccb->ccb_prpl[i - 1], 695 dmap->dm_segs[i].ds_addr); 696 } 697 bus_dmamap_sync(sc->sc_dmat, 698 NVME_DMA_MAP(q->q_ccb_prpls), 699 ccb->ccb_prpl_off, 700 sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1), 701 BUS_DMASYNC_PREWRITE); 702 } 703 704 if (ISSET(flags, NVME_NS_CTX_F_POLL)) { 705 if (nvme_poll(sc, q, ccb, nvme_ns_io_fill, NVME_TIMO_PT) != 0) 706 return EIO; 707 return 0; 708 } 709 710 nvme_q_submit(sc, q, ccb, nvme_ns_io_fill); 711 return 0; 712 } 713 714 static void 715 nvme_ns_io_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 716 { 717 struct nvme_sqe_io *sqe = slot; 718 bus_dmamap_t dmap = ccb->ccb_dmamap; 719 720 sqe->opcode = ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ? 721 NVM_CMD_READ : NVM_CMD_WRITE; 722 htolem32(&sqe->nsid, ccb->nnc_nsid); 723 724 htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr); 725 switch (dmap->dm_nsegs) { 726 case 1: 727 break; 728 case 2: 729 htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr); 730 break; 731 default: 732 /* the prp list is already set up and synced */ 733 htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva); 734 break; 735 } 736 737 htolem64(&sqe->slba, ccb->nnc_blkno); 738 739 if (ISSET(ccb->nnc_flags, NVME_NS_CTX_F_FUA)) 740 htolem16(&sqe->ioflags, NVM_SQE_IO_FUA); 741 742 /* guaranteed by upper layers, but check just in case */ 743 KASSERT((ccb->nnc_datasize % ccb->nnc_secsize) == 0); 744 htolem16(&sqe->nlb, (ccb->nnc_datasize / ccb->nnc_secsize) - 1); 745 } 746 747 static void 748 nvme_ns_io_done(struct nvme_queue *q, struct nvme_ccb *ccb, 749 struct nvme_cqe *cqe) 750 { 751 struct nvme_softc *sc = q->q_sc; 752 bus_dmamap_t dmap = ccb->ccb_dmamap; 753 void *nnc_cookie = ccb->ccb_cookie; 754 nvme_nnc_done nnc_done = ccb->nnc_done; 755 struct buf *bp = ccb->nnc_buf; 756 757 if (dmap->dm_nsegs > 2) { 758 bus_dmamap_sync(sc->sc_dmat, 759 NVME_DMA_MAP(q->q_ccb_prpls), 760 ccb->ccb_prpl_off, 761 sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1), 762 BUS_DMASYNC_POSTWRITE); 763 } 764 765 bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize, 766 ISSET(ccb->nnc_flags, NVME_NS_CTX_F_READ) ? 767 BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE); 768 769 bus_dmamap_unload(sc->sc_dmat, dmap); 770 nvme_ccb_put(q, ccb); 771 772 nnc_done(nnc_cookie, bp, lemtoh16(&cqe->flags), lemtoh32(&cqe->cdw0)); 773 } 774 775 /* 776 * If there is no volatile write cache, it makes no sense to issue 777 * flush commands or query for the status. 778 */ 779 static bool 780 nvme_has_volatile_write_cache(struct nvme_softc *sc) 781 { 782 /* sc_identify is filled during attachment */ 783 return ((sc->sc_identify.vwc & NVME_ID_CTRLR_VWC_PRESENT) != 0); 784 } 785 786 static bool 787 nvme_ns_sync_finished(void *cookie) 788 { 789 int *result = cookie; 790 791 return (*result != 0); 792 } 793 794 int 795 nvme_ns_sync(struct nvme_softc *sc, uint16_t nsid, int flags) 796 { 797 struct nvme_queue *q = nvme_get_q(sc, NULL, true); 798 struct nvme_ccb *ccb; 799 int result = 0; 800 801 if (!nvme_has_volatile_write_cache(sc)) { 802 /* cache not present, no value in trying to flush it */ 803 return 0; 804 } 805 806 ccb = nvme_ccb_get(q, true); 807 KASSERT(ccb != NULL); 808 809 ccb->ccb_done = nvme_ns_sync_done; 810 ccb->ccb_cookie = &result; 811 812 /* namespace context */ 813 ccb->nnc_nsid = nsid; 814 ccb->nnc_flags = flags; 815 ccb->nnc_done = NULL; 816 817 if (ISSET(flags, NVME_NS_CTX_F_POLL)) { 818 if (nvme_poll(sc, q, ccb, nvme_ns_sync_fill, NVME_TIMO_SY) != 0) 819 return EIO; 820 return 0; 821 } 822 823 nvme_q_submit(sc, q, ccb, nvme_ns_sync_fill); 824 825 /* wait for completion */ 826 nvme_q_wait_complete(sc, q, nvme_ns_sync_finished, &result); 827 KASSERT(result != 0); 828 829 return (result > 0) ? 0 : EIO; 830 } 831 832 static void 833 nvme_ns_sync_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 834 { 835 struct nvme_sqe *sqe = slot; 836 837 sqe->opcode = NVM_CMD_FLUSH; 838 htolem32(&sqe->nsid, ccb->nnc_nsid); 839 } 840 841 static void 842 nvme_ns_sync_done(struct nvme_queue *q, struct nvme_ccb *ccb, 843 struct nvme_cqe *cqe) 844 { 845 int *result = ccb->ccb_cookie; 846 uint16_t status = NVME_CQE_SC(lemtoh16(&cqe->flags)); 847 848 if (status == NVME_CQE_SC_SUCCESS) 849 *result = 1; 850 else 851 *result = -1; 852 853 nvme_ccb_put(q, ccb); 854 } 855 856 static bool 857 nvme_getcache_finished(void *xc) 858 { 859 int *addr = xc; 860 861 return (*addr != 0); 862 } 863 864 /* 865 * Get status of volatile write cache. Always asynchronous. 866 */ 867 int 868 nvme_admin_getcache(struct nvme_softc *sc, int *addr) 869 { 870 struct nvme_ccb *ccb; 871 struct nvme_queue *q = sc->sc_admin_q; 872 int result = 0, error; 873 874 if (!nvme_has_volatile_write_cache(sc)) { 875 /* cache simply not present */ 876 *addr = 0; 877 return 0; 878 } 879 880 ccb = nvme_ccb_get(q, true); 881 KASSERT(ccb != NULL); 882 883 ccb->ccb_done = nvme_getcache_done; 884 ccb->ccb_cookie = &result; 885 886 /* namespace context */ 887 ccb->nnc_flags = 0; 888 ccb->nnc_done = NULL; 889 890 nvme_q_submit(sc, q, ccb, nvme_getcache_fill); 891 892 /* wait for completion */ 893 nvme_q_wait_complete(sc, q, nvme_getcache_finished, &result); 894 KASSERT(result != 0); 895 896 if (result > 0) { 897 *addr = result; 898 error = 0; 899 } else 900 error = EINVAL; 901 902 return error; 903 } 904 905 static void 906 nvme_getcache_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 907 { 908 struct nvme_sqe *sqe = slot; 909 910 sqe->opcode = NVM_ADMIN_GET_FEATURES; 911 htolem32(&sqe->cdw10, NVM_FEATURE_VOLATILE_WRITE_CACHE); 912 htolem32(&sqe->cdw11, NVM_VOLATILE_WRITE_CACHE_WCE); 913 } 914 915 static void 916 nvme_getcache_done(struct nvme_queue *q, struct nvme_ccb *ccb, 917 struct nvme_cqe *cqe) 918 { 919 int *addr = ccb->ccb_cookie; 920 uint16_t status = NVME_CQE_SC(lemtoh16(&cqe->flags)); 921 uint32_t cdw0 = lemtoh32(&cqe->cdw0); 922 int result; 923 924 if (status == NVME_CQE_SC_SUCCESS) { 925 result = 0; 926 927 /* 928 * DPO not supported, Dataset Management (DSM) field doesn't 929 * specify the same semantics. FUA is always supported. 930 */ 931 result = DKCACHE_FUA; 932 933 if (cdw0 & NVM_VOLATILE_WRITE_CACHE_WCE) 934 result |= DKCACHE_WRITE; 935 936 /* 937 * If volatile write cache is present, the flag shall also be 938 * settable. 939 */ 940 result |= DKCACHE_WCHANGE; 941 942 /* 943 * ONCS field indicates whether the optional SAVE is also 944 * supported for Set Features. According to spec v1.3, 945 * Volatile Write Cache however doesn't support persistency 946 * across power cycle/reset. 947 */ 948 949 } else { 950 result = -1; 951 } 952 953 *addr = result; 954 955 nvme_ccb_put(q, ccb); 956 } 957 958 struct nvme_setcache_state { 959 int dkcache; 960 int result; 961 }; 962 963 static bool 964 nvme_setcache_finished(void *xc) 965 { 966 struct nvme_setcache_state *st = xc; 967 968 return (st->result != 0); 969 } 970 971 static void 972 nvme_setcache_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 973 { 974 struct nvme_sqe *sqe = slot; 975 struct nvme_setcache_state *st = ccb->ccb_cookie; 976 977 sqe->opcode = NVM_ADMIN_SET_FEATURES; 978 htolem32(&sqe->cdw10, NVM_FEATURE_VOLATILE_WRITE_CACHE); 979 if (st->dkcache & DKCACHE_WRITE) 980 htolem32(&sqe->cdw11, NVM_VOLATILE_WRITE_CACHE_WCE); 981 } 982 983 static void 984 nvme_setcache_done(struct nvme_queue *q, struct nvme_ccb *ccb, 985 struct nvme_cqe *cqe) 986 { 987 struct nvme_setcache_state *st = ccb->ccb_cookie; 988 uint16_t status = NVME_CQE_SC(lemtoh16(&cqe->flags)); 989 990 if (status == NVME_CQE_SC_SUCCESS) { 991 st->result = 1; 992 } else { 993 st->result = -1; 994 } 995 996 nvme_ccb_put(q, ccb); 997 } 998 999 /* 1000 * Set status of volatile write cache. Always asynchronous. 1001 */ 1002 int 1003 nvme_admin_setcache(struct nvme_softc *sc, int dkcache) 1004 { 1005 struct nvme_ccb *ccb; 1006 struct nvme_queue *q = sc->sc_admin_q; 1007 int error; 1008 struct nvme_setcache_state st; 1009 1010 if (!nvme_has_volatile_write_cache(sc)) { 1011 /* cache simply not present */ 1012 return EOPNOTSUPP; 1013 } 1014 1015 if (dkcache & ~(DKCACHE_WRITE)) { 1016 /* unsupported parameters */ 1017 return EOPNOTSUPP; 1018 } 1019 1020 ccb = nvme_ccb_get(q, true); 1021 KASSERT(ccb != NULL); 1022 1023 memset(&st, 0, sizeof(st)); 1024 st.dkcache = dkcache; 1025 1026 ccb->ccb_done = nvme_setcache_done; 1027 ccb->ccb_cookie = &st; 1028 1029 /* namespace context */ 1030 ccb->nnc_flags = 0; 1031 ccb->nnc_done = NULL; 1032 1033 nvme_q_submit(sc, q, ccb, nvme_setcache_fill); 1034 1035 /* wait for completion */ 1036 nvme_q_wait_complete(sc, q, nvme_setcache_finished, &st); 1037 KASSERT(st.result != 0); 1038 1039 if (st.result > 0) 1040 error = 0; 1041 else 1042 error = EINVAL; 1043 1044 return error; 1045 } 1046 1047 void 1048 nvme_ns_free(struct nvme_softc *sc, uint16_t nsid) 1049 { 1050 struct nvme_namespace *ns; 1051 struct nvm_identify_namespace *identify; 1052 1053 ns = nvme_ns_get(sc, nsid); 1054 KASSERT(ns); 1055 1056 identify = ns->ident; 1057 ns->ident = NULL; 1058 if (identify != NULL) 1059 kmem_free(identify, sizeof(*identify)); 1060 } 1061 1062 struct nvme_pt_state { 1063 struct nvme_pt_command *pt; 1064 bool finished; 1065 }; 1066 1067 static void 1068 nvme_pt_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 1069 { 1070 struct nvme_softc *sc = q->q_sc; 1071 struct nvme_sqe *sqe = slot; 1072 struct nvme_pt_state *state = ccb->ccb_cookie; 1073 struct nvme_pt_command *pt = state->pt; 1074 bus_dmamap_t dmap = ccb->ccb_dmamap; 1075 int i; 1076 1077 sqe->opcode = pt->cmd.opcode; 1078 htolem32(&sqe->nsid, pt->cmd.nsid); 1079 1080 if (pt->buf != NULL && pt->len > 0) { 1081 htolem64(&sqe->entry.prp[0], dmap->dm_segs[0].ds_addr); 1082 switch (dmap->dm_nsegs) { 1083 case 1: 1084 break; 1085 case 2: 1086 htolem64(&sqe->entry.prp[1], dmap->dm_segs[1].ds_addr); 1087 break; 1088 default: 1089 for (i = 1; i < dmap->dm_nsegs; i++) { 1090 htolem64(&ccb->ccb_prpl[i - 1], 1091 dmap->dm_segs[i].ds_addr); 1092 } 1093 bus_dmamap_sync(sc->sc_dmat, 1094 NVME_DMA_MAP(q->q_ccb_prpls), 1095 ccb->ccb_prpl_off, 1096 sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1), 1097 BUS_DMASYNC_PREWRITE); 1098 htolem64(&sqe->entry.prp[1], ccb->ccb_prpl_dva); 1099 break; 1100 } 1101 } 1102 1103 htolem32(&sqe->cdw10, pt->cmd.cdw10); 1104 htolem32(&sqe->cdw11, pt->cmd.cdw11); 1105 htolem32(&sqe->cdw12, pt->cmd.cdw12); 1106 htolem32(&sqe->cdw13, pt->cmd.cdw13); 1107 htolem32(&sqe->cdw14, pt->cmd.cdw14); 1108 htolem32(&sqe->cdw15, pt->cmd.cdw15); 1109 } 1110 1111 static void 1112 nvme_pt_done(struct nvme_queue *q, struct nvme_ccb *ccb, struct nvme_cqe *cqe) 1113 { 1114 struct nvme_softc *sc = q->q_sc; 1115 struct nvme_pt_state *state = ccb->ccb_cookie; 1116 struct nvme_pt_command *pt = state->pt; 1117 bus_dmamap_t dmap = ccb->ccb_dmamap; 1118 1119 if (pt->buf != NULL && pt->len > 0) { 1120 if (dmap->dm_nsegs > 2) { 1121 bus_dmamap_sync(sc->sc_dmat, 1122 NVME_DMA_MAP(q->q_ccb_prpls), 1123 ccb->ccb_prpl_off, 1124 sizeof(*ccb->ccb_prpl) * (dmap->dm_nsegs - 1), 1125 BUS_DMASYNC_POSTWRITE); 1126 } 1127 1128 bus_dmamap_sync(sc->sc_dmat, dmap, 0, dmap->dm_mapsize, 1129 pt->is_read ? BUS_DMASYNC_POSTREAD : BUS_DMASYNC_POSTWRITE); 1130 bus_dmamap_unload(sc->sc_dmat, dmap); 1131 } 1132 1133 pt->cpl.cdw0 = lemtoh32(&cqe->cdw0); 1134 pt->cpl.flags = lemtoh16(&cqe->flags) & ~NVME_CQE_PHASE; 1135 1136 state->finished = true; 1137 1138 nvme_ccb_put(q, ccb); 1139 } 1140 1141 static bool 1142 nvme_pt_finished(void *cookie) 1143 { 1144 struct nvme_pt_state *state = cookie; 1145 1146 return state->finished; 1147 } 1148 1149 static int 1150 nvme_command_passthrough(struct nvme_softc *sc, struct nvme_pt_command *pt, 1151 uint16_t nsid, struct lwp *l, bool is_adminq) 1152 { 1153 struct nvme_queue *q; 1154 struct nvme_ccb *ccb; 1155 void *buf = NULL; 1156 struct nvme_pt_state state; 1157 int error; 1158 1159 /* limit command size to maximum data transfer size */ 1160 if ((pt->buf == NULL && pt->len > 0) || 1161 (pt->buf != NULL && (pt->len == 0 || pt->len > sc->sc_mdts))) 1162 return EINVAL; 1163 1164 q = is_adminq ? sc->sc_admin_q : nvme_get_q(sc, NULL, true); 1165 ccb = nvme_ccb_get(q, true); 1166 KASSERT(ccb != NULL); 1167 1168 if (pt->buf != NULL) { 1169 KASSERT(pt->len > 0); 1170 buf = kmem_alloc(pt->len, KM_SLEEP); 1171 if (!pt->is_read) { 1172 error = copyin(pt->buf, buf, pt->len); 1173 if (error) 1174 goto kmem_free; 1175 } 1176 error = bus_dmamap_load(sc->sc_dmat, ccb->ccb_dmamap, buf, 1177 pt->len, NULL, 1178 BUS_DMA_WAITOK | 1179 (pt->is_read ? BUS_DMA_READ : BUS_DMA_WRITE)); 1180 if (error) 1181 goto kmem_free; 1182 bus_dmamap_sync(sc->sc_dmat, ccb->ccb_dmamap, 1183 0, ccb->ccb_dmamap->dm_mapsize, 1184 pt->is_read ? BUS_DMASYNC_PREREAD : BUS_DMASYNC_PREWRITE); 1185 } 1186 1187 memset(&state, 0, sizeof(state)); 1188 state.pt = pt; 1189 state.finished = false; 1190 1191 ccb->ccb_done = nvme_pt_done; 1192 ccb->ccb_cookie = &state; 1193 1194 pt->cmd.nsid = nsid; 1195 1196 nvme_q_submit(sc, q, ccb, nvme_pt_fill); 1197 1198 /* wait for completion */ 1199 nvme_q_wait_complete(sc, q, nvme_pt_finished, &state); 1200 KASSERT(state.finished); 1201 1202 error = 0; 1203 1204 if (buf != NULL) { 1205 if (error == 0 && pt->is_read) 1206 error = copyout(buf, pt->buf, pt->len); 1207 kmem_free: 1208 kmem_free(buf, pt->len); 1209 } 1210 1211 return error; 1212 } 1213 1214 static void 1215 nvme_q_submit(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb, 1216 void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *)) 1217 { 1218 struct nvme_sqe *sqe = NVME_DMA_KVA(q->q_sq_dmamem); 1219 uint32_t tail; 1220 1221 mutex_enter(&q->q_sq_mtx); 1222 tail = q->q_sq_tail; 1223 if (++q->q_sq_tail >= q->q_entries) 1224 q->q_sq_tail = 0; 1225 1226 sqe += tail; 1227 1228 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 1229 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_POSTWRITE); 1230 memset(sqe, 0, sizeof(*sqe)); 1231 (*fill)(q, ccb, sqe); 1232 htolem16(&sqe->cid, ccb->ccb_id); 1233 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(q->q_sq_dmamem), 1234 sizeof(*sqe) * tail, sizeof(*sqe), BUS_DMASYNC_PREWRITE); 1235 1236 nvme_write4(sc, q->q_sqtdbl, q->q_sq_tail); 1237 mutex_exit(&q->q_sq_mtx); 1238 } 1239 1240 struct nvme_poll_state { 1241 struct nvme_sqe s; 1242 struct nvme_cqe c; 1243 void *cookie; 1244 void (*done)(struct nvme_queue *, struct nvme_ccb *, struct nvme_cqe *); 1245 }; 1246 1247 static int 1248 nvme_poll(struct nvme_softc *sc, struct nvme_queue *q, struct nvme_ccb *ccb, 1249 void (*fill)(struct nvme_queue *, struct nvme_ccb *, void *), int timo_sec) 1250 { 1251 struct nvme_poll_state state; 1252 uint16_t flags; 1253 int step = 10; 1254 int maxloop = timo_sec * 1000000 / step; 1255 int error = 0; 1256 1257 memset(&state, 0, sizeof(state)); 1258 (*fill)(q, ccb, &state.s); 1259 1260 state.done = ccb->ccb_done; 1261 state.cookie = ccb->ccb_cookie; 1262 1263 ccb->ccb_done = nvme_poll_done; 1264 ccb->ccb_cookie = &state; 1265 1266 nvme_q_submit(sc, q, ccb, nvme_poll_fill); 1267 while (!ISSET(state.c.flags, htole16(NVME_CQE_PHASE))) { 1268 if (nvme_q_complete(sc, q) == 0) 1269 delay(step); 1270 1271 if (timo_sec >= 0 && --maxloop <= 0) { 1272 error = ETIMEDOUT; 1273 break; 1274 } 1275 } 1276 1277 if (error == 0) { 1278 flags = lemtoh16(&state.c.flags); 1279 return flags & ~NVME_CQE_PHASE; 1280 } else { 1281 /* 1282 * If it succeds later, it would hit ccb which will have been 1283 * already reused for something else. Not good. Cross 1284 * fingers and hope for best. XXX do controller reset? 1285 */ 1286 aprint_error_dev(sc->sc_dev, "polled command timed out\n"); 1287 1288 /* Invoke the callback to clean state anyway */ 1289 struct nvme_cqe cqe; 1290 memset(&cqe, 0, sizeof(cqe)); 1291 ccb->ccb_done(q, ccb, &cqe); 1292 1293 return 1; 1294 } 1295 } 1296 1297 static void 1298 nvme_poll_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 1299 { 1300 struct nvme_sqe *sqe = slot; 1301 struct nvme_poll_state *state = ccb->ccb_cookie; 1302 1303 *sqe = state->s; 1304 } 1305 1306 static void 1307 nvme_poll_done(struct nvme_queue *q, struct nvme_ccb *ccb, 1308 struct nvme_cqe *cqe) 1309 { 1310 struct nvme_poll_state *state = ccb->ccb_cookie; 1311 1312 state->c = *cqe; 1313 SET(state->c.flags, htole16(NVME_CQE_PHASE)); 1314 1315 ccb->ccb_cookie = state->cookie; 1316 state->done(q, ccb, &state->c); 1317 } 1318 1319 static void 1320 nvme_sqe_fill(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 1321 { 1322 struct nvme_sqe *src = ccb->ccb_cookie; 1323 struct nvme_sqe *dst = slot; 1324 1325 *dst = *src; 1326 } 1327 1328 static void 1329 nvme_empty_done(struct nvme_queue *q, struct nvme_ccb *ccb, 1330 struct nvme_cqe *cqe) 1331 { 1332 } 1333 1334 static int 1335 nvme_q_complete(struct nvme_softc *sc, struct nvme_queue *q) 1336 { 1337 struct nvme_ccb *ccb; 1338 struct nvme_cqe *ring = NVME_DMA_KVA(q->q_cq_dmamem), *cqe; 1339 uint16_t flags; 1340 int rv = 0; 1341 1342 mutex_enter(&q->q_cq_mtx); 1343 1344 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD); 1345 for (;;) { 1346 cqe = &ring[q->q_cq_head]; 1347 flags = lemtoh16(&cqe->flags); 1348 if ((flags & NVME_CQE_PHASE) != q->q_cq_phase) 1349 break; 1350 1351 ccb = &q->q_ccbs[cqe->cid]; 1352 1353 if (++q->q_cq_head >= q->q_entries) { 1354 q->q_cq_head = 0; 1355 q->q_cq_phase ^= NVME_CQE_PHASE; 1356 } 1357 1358 #ifdef DEBUG 1359 /* 1360 * If we get spurious completion notification, something 1361 * is seriously hosed up. Very likely DMA to some random 1362 * memory place happened, so just bail out. 1363 */ 1364 if ((intptr_t)ccb->ccb_cookie == NVME_CCB_FREE) { 1365 panic("%s: invalid ccb detected", 1366 device_xname(sc->sc_dev)); 1367 /* NOTREACHED */ 1368 } 1369 #endif 1370 1371 rv++; 1372 1373 /* 1374 * Unlock the mutex before calling the ccb_done callback 1375 * and re-lock afterwards. The callback triggers lddone() 1376 * which schedules another i/o, and also calls nvme_ccb_put(). 1377 * Unlock/relock avoids possibility of deadlock. 1378 */ 1379 mutex_exit(&q->q_cq_mtx); 1380 ccb->ccb_done(q, ccb, cqe); 1381 mutex_enter(&q->q_cq_mtx); 1382 } 1383 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD); 1384 1385 if (rv) 1386 nvme_write4(sc, q->q_cqhdbl, q->q_cq_head); 1387 1388 mutex_exit(&q->q_cq_mtx); 1389 1390 return rv; 1391 } 1392 1393 static void 1394 nvme_q_wait_complete(struct nvme_softc *sc, 1395 struct nvme_queue *q, bool (*finished)(void *), void *cookie) 1396 { 1397 mutex_enter(&q->q_ccb_mtx); 1398 if (finished(cookie)) 1399 goto out; 1400 1401 for(;;) { 1402 q->q_ccb_waiting = true; 1403 cv_wait(&q->q_ccb_wait, &q->q_ccb_mtx); 1404 1405 if (finished(cookie)) 1406 break; 1407 } 1408 1409 out: 1410 mutex_exit(&q->q_ccb_mtx); 1411 } 1412 1413 static int 1414 nvme_identify(struct nvme_softc *sc, u_int mps) 1415 { 1416 char sn[41], mn[81], fr[17]; 1417 struct nvm_identify_controller *identify; 1418 struct nvme_dmamem *mem; 1419 struct nvme_ccb *ccb; 1420 u_int mdts; 1421 int rv = 1; 1422 1423 ccb = nvme_ccb_get(sc->sc_admin_q, false); 1424 KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */ 1425 1426 mem = nvme_dmamem_alloc(sc, sizeof(*identify)); 1427 if (mem == NULL) 1428 return 1; 1429 1430 ccb->ccb_done = nvme_empty_done; 1431 ccb->ccb_cookie = mem; 1432 1433 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_PREREAD); 1434 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_fill_identify, 1435 NVME_TIMO_IDENT); 1436 nvme_dmamem_sync(sc, mem, BUS_DMASYNC_POSTREAD); 1437 1438 nvme_ccb_put(sc->sc_admin_q, ccb); 1439 1440 if (rv != 0) 1441 goto done; 1442 1443 identify = NVME_DMA_KVA(mem); 1444 sc->sc_identify = *identify; 1445 identify = NULL; 1446 1447 /* Convert data to host endian */ 1448 nvme_identify_controller_swapbytes(&sc->sc_identify); 1449 1450 strnvisx(sn, sizeof(sn), (const char *)sc->sc_identify.sn, 1451 sizeof(sc->sc_identify.sn), VIS_TRIM|VIS_SAFE|VIS_OCTAL); 1452 strnvisx(mn, sizeof(mn), (const char *)sc->sc_identify.mn, 1453 sizeof(sc->sc_identify.mn), VIS_TRIM|VIS_SAFE|VIS_OCTAL); 1454 strnvisx(fr, sizeof(fr), (const char *)sc->sc_identify.fr, 1455 sizeof(sc->sc_identify.fr), VIS_TRIM|VIS_SAFE|VIS_OCTAL); 1456 aprint_normal_dev(sc->sc_dev, "%s, firmware %s, serial %s\n", mn, fr, 1457 sn); 1458 1459 strlcpy(sc->sc_modelname, mn, sizeof(sc->sc_modelname)); 1460 1461 if (sc->sc_identify.mdts > 0) { 1462 mdts = (1 << sc->sc_identify.mdts) * (1 << mps); 1463 if (mdts < sc->sc_mdts) 1464 sc->sc_mdts = mdts; 1465 } 1466 1467 sc->sc_nn = sc->sc_identify.nn; 1468 1469 done: 1470 nvme_dmamem_free(sc, mem); 1471 1472 return rv; 1473 } 1474 1475 static int 1476 nvme_q_create(struct nvme_softc *sc, struct nvme_queue *q) 1477 { 1478 struct nvme_sqe_q sqe; 1479 struct nvme_ccb *ccb; 1480 int rv; 1481 1482 if (sc->sc_use_mq && sc->sc_intr_establish(sc, q->q_id, q) != 0) 1483 return 1; 1484 1485 ccb = nvme_ccb_get(sc->sc_admin_q, false); 1486 KASSERT(ccb != NULL); 1487 1488 ccb->ccb_done = nvme_empty_done; 1489 ccb->ccb_cookie = &sqe; 1490 1491 memset(&sqe, 0, sizeof(sqe)); 1492 sqe.opcode = NVM_ADMIN_ADD_IOCQ; 1493 htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_cq_dmamem)); 1494 htolem16(&sqe.qsize, q->q_entries - 1); 1495 htolem16(&sqe.qid, q->q_id); 1496 sqe.qflags = NVM_SQE_CQ_IEN | NVM_SQE_Q_PC; 1497 if (sc->sc_use_mq) 1498 htolem16(&sqe.cqid, q->q_id); /* qid == vector */ 1499 1500 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP); 1501 if (rv != 0) 1502 goto fail; 1503 1504 ccb->ccb_done = nvme_empty_done; 1505 ccb->ccb_cookie = &sqe; 1506 1507 memset(&sqe, 0, sizeof(sqe)); 1508 sqe.opcode = NVM_ADMIN_ADD_IOSQ; 1509 htolem64(&sqe.prp1, NVME_DMA_DVA(q->q_sq_dmamem)); 1510 htolem16(&sqe.qsize, q->q_entries - 1); 1511 htolem16(&sqe.qid, q->q_id); 1512 htolem16(&sqe.cqid, q->q_id); 1513 sqe.qflags = NVM_SQE_Q_PC; 1514 1515 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP); 1516 if (rv != 0) 1517 goto fail; 1518 1519 nvme_ccb_put(sc->sc_admin_q, ccb); 1520 return 0; 1521 1522 fail: 1523 if (sc->sc_use_mq) 1524 sc->sc_intr_disestablish(sc, q->q_id); 1525 1526 nvme_ccb_put(sc->sc_admin_q, ccb); 1527 return rv; 1528 } 1529 1530 static int 1531 nvme_q_delete(struct nvme_softc *sc, struct nvme_queue *q) 1532 { 1533 struct nvme_sqe_q sqe; 1534 struct nvme_ccb *ccb; 1535 int rv; 1536 1537 ccb = nvme_ccb_get(sc->sc_admin_q, false); 1538 KASSERT(ccb != NULL); 1539 1540 ccb->ccb_done = nvme_empty_done; 1541 ccb->ccb_cookie = &sqe; 1542 1543 memset(&sqe, 0, sizeof(sqe)); 1544 sqe.opcode = NVM_ADMIN_DEL_IOSQ; 1545 htolem16(&sqe.qid, q->q_id); 1546 1547 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP); 1548 if (rv != 0) 1549 goto fail; 1550 1551 ccb->ccb_done = nvme_empty_done; 1552 ccb->ccb_cookie = &sqe; 1553 1554 memset(&sqe, 0, sizeof(sqe)); 1555 sqe.opcode = NVM_ADMIN_DEL_IOCQ; 1556 htolem16(&sqe.qid, q->q_id); 1557 1558 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_sqe_fill, NVME_TIMO_QOP); 1559 if (rv != 0) 1560 goto fail; 1561 1562 fail: 1563 nvme_ccb_put(sc->sc_admin_q, ccb); 1564 1565 if (rv == 0 && sc->sc_use_mq) { 1566 if (sc->sc_intr_disestablish(sc, q->q_id)) 1567 rv = 1; 1568 } 1569 1570 return rv; 1571 } 1572 1573 static void 1574 nvme_fill_identify(struct nvme_queue *q, struct nvme_ccb *ccb, void *slot) 1575 { 1576 struct nvme_sqe *sqe = slot; 1577 struct nvme_dmamem *mem = ccb->ccb_cookie; 1578 1579 sqe->opcode = NVM_ADMIN_IDENTIFY; 1580 htolem64(&sqe->entry.prp[0], NVME_DMA_DVA(mem)); 1581 htolem32(&sqe->cdw10, 1); 1582 } 1583 1584 static int 1585 nvme_set_number_of_queues(struct nvme_softc *sc, u_int nq, u_int *ncqa, 1586 u_int *nsqa) 1587 { 1588 struct nvme_pt_state state; 1589 struct nvme_pt_command pt; 1590 struct nvme_ccb *ccb; 1591 int rv; 1592 1593 ccb = nvme_ccb_get(sc->sc_admin_q, false); 1594 KASSERT(ccb != NULL); /* it's a bug if we don't have spare ccb here */ 1595 1596 memset(&pt, 0, sizeof(pt)); 1597 pt.cmd.opcode = NVM_ADMIN_SET_FEATURES; 1598 htolem32(&pt.cmd.cdw10, NVM_FEATURE_NUMBER_OF_QUEUES); 1599 htolem32(&pt.cmd.cdw11, ((nq - 1) << 16) | (nq - 1)); 1600 1601 memset(&state, 0, sizeof(state)); 1602 state.pt = &pt; 1603 state.finished = false; 1604 1605 ccb->ccb_done = nvme_pt_done; 1606 ccb->ccb_cookie = &state; 1607 1608 rv = nvme_poll(sc, sc->sc_admin_q, ccb, nvme_pt_fill, NVME_TIMO_QOP); 1609 1610 if (rv != 0) { 1611 *ncqa = *nsqa = 0; 1612 return EIO; 1613 } 1614 1615 *ncqa = (pt.cpl.cdw0 >> 16) + 1; 1616 *nsqa = (pt.cpl.cdw0 & 0xffff) + 1; 1617 1618 return 0; 1619 } 1620 1621 static int 1622 nvme_ccbs_alloc(struct nvme_queue *q, uint16_t nccbs) 1623 { 1624 struct nvme_softc *sc = q->q_sc; 1625 struct nvme_ccb *ccb; 1626 bus_addr_t off; 1627 uint64_t *prpl; 1628 u_int i; 1629 1630 mutex_init(&q->q_ccb_mtx, MUTEX_DEFAULT, IPL_BIO); 1631 cv_init(&q->q_ccb_wait, "nvmeqw"); 1632 q->q_ccb_waiting = false; 1633 SIMPLEQ_INIT(&q->q_ccb_list); 1634 1635 q->q_ccbs = kmem_alloc(sizeof(*ccb) * nccbs, KM_SLEEP); 1636 1637 q->q_nccbs = nccbs; 1638 q->q_ccb_prpls = nvme_dmamem_alloc(sc, 1639 sizeof(*prpl) * sc->sc_max_sgl * nccbs); 1640 1641 prpl = NVME_DMA_KVA(q->q_ccb_prpls); 1642 off = 0; 1643 1644 for (i = 0; i < nccbs; i++) { 1645 ccb = &q->q_ccbs[i]; 1646 1647 if (bus_dmamap_create(sc->sc_dmat, sc->sc_mdts, 1648 sc->sc_max_sgl + 1 /* we get a free prp in the sqe */, 1649 sc->sc_mps, sc->sc_mps, BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, 1650 &ccb->ccb_dmamap) != 0) 1651 goto free_maps; 1652 1653 ccb->ccb_id = i; 1654 ccb->ccb_prpl = prpl; 1655 ccb->ccb_prpl_off = off; 1656 ccb->ccb_prpl_dva = NVME_DMA_DVA(q->q_ccb_prpls) + off; 1657 1658 SIMPLEQ_INSERT_TAIL(&q->q_ccb_list, ccb, ccb_entry); 1659 1660 prpl += sc->sc_max_sgl; 1661 off += sizeof(*prpl) * sc->sc_max_sgl; 1662 } 1663 1664 return 0; 1665 1666 free_maps: 1667 nvme_ccbs_free(q); 1668 return 1; 1669 } 1670 1671 static struct nvme_ccb * 1672 nvme_ccb_get(struct nvme_queue *q, bool wait) 1673 { 1674 struct nvme_ccb *ccb = NULL; 1675 1676 mutex_enter(&q->q_ccb_mtx); 1677 again: 1678 ccb = SIMPLEQ_FIRST(&q->q_ccb_list); 1679 if (ccb != NULL) { 1680 SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry); 1681 #ifdef DEBUG 1682 ccb->ccb_cookie = NULL; 1683 #endif 1684 } else { 1685 if (__predict_false(wait)) { 1686 q->q_ccb_waiting = true; 1687 cv_wait(&q->q_ccb_wait, &q->q_ccb_mtx); 1688 goto again; 1689 } 1690 } 1691 mutex_exit(&q->q_ccb_mtx); 1692 1693 return ccb; 1694 } 1695 1696 static void 1697 nvme_ccb_put(struct nvme_queue *q, struct nvme_ccb *ccb) 1698 { 1699 1700 mutex_enter(&q->q_ccb_mtx); 1701 #ifdef DEBUG 1702 ccb->ccb_cookie = (void *)NVME_CCB_FREE; 1703 #endif 1704 SIMPLEQ_INSERT_HEAD(&q->q_ccb_list, ccb, ccb_entry); 1705 1706 /* It's unlikely there are any waiters, it's not used for regular I/O */ 1707 if (__predict_false(q->q_ccb_waiting)) { 1708 q->q_ccb_waiting = false; 1709 cv_broadcast(&q->q_ccb_wait); 1710 } 1711 1712 mutex_exit(&q->q_ccb_mtx); 1713 } 1714 1715 static void 1716 nvme_ccbs_free(struct nvme_queue *q) 1717 { 1718 struct nvme_softc *sc = q->q_sc; 1719 struct nvme_ccb *ccb; 1720 1721 mutex_enter(&q->q_ccb_mtx); 1722 while ((ccb = SIMPLEQ_FIRST(&q->q_ccb_list)) != NULL) { 1723 SIMPLEQ_REMOVE_HEAD(&q->q_ccb_list, ccb_entry); 1724 /* 1725 * bus_dmamap_destroy() may call vm_map_lock() and rw_enter() 1726 * internally. don't hold spin mutex 1727 */ 1728 mutex_exit(&q->q_ccb_mtx); 1729 bus_dmamap_destroy(sc->sc_dmat, ccb->ccb_dmamap); 1730 mutex_enter(&q->q_ccb_mtx); 1731 } 1732 mutex_exit(&q->q_ccb_mtx); 1733 1734 nvme_dmamem_free(sc, q->q_ccb_prpls); 1735 kmem_free(q->q_ccbs, sizeof(*ccb) * q->q_nccbs); 1736 q->q_ccbs = NULL; 1737 cv_destroy(&q->q_ccb_wait); 1738 mutex_destroy(&q->q_ccb_mtx); 1739 } 1740 1741 static struct nvme_queue * 1742 nvme_q_alloc(struct nvme_softc *sc, uint16_t id, u_int entries, u_int dstrd) 1743 { 1744 struct nvme_queue *q; 1745 1746 q = kmem_alloc(sizeof(*q), KM_SLEEP); 1747 q->q_sc = sc; 1748 q->q_sq_dmamem = nvme_dmamem_alloc(sc, 1749 sizeof(struct nvme_sqe) * entries); 1750 if (q->q_sq_dmamem == NULL) 1751 goto free; 1752 1753 q->q_cq_dmamem = nvme_dmamem_alloc(sc, 1754 sizeof(struct nvme_cqe) * entries); 1755 if (q->q_cq_dmamem == NULL) 1756 goto free_sq; 1757 1758 memset(NVME_DMA_KVA(q->q_sq_dmamem), 0, NVME_DMA_LEN(q->q_sq_dmamem)); 1759 memset(NVME_DMA_KVA(q->q_cq_dmamem), 0, NVME_DMA_LEN(q->q_cq_dmamem)); 1760 1761 mutex_init(&q->q_sq_mtx, MUTEX_DEFAULT, IPL_BIO); 1762 mutex_init(&q->q_cq_mtx, MUTEX_DEFAULT, IPL_BIO); 1763 q->q_sqtdbl = NVME_SQTDBL(id, dstrd); 1764 q->q_cqhdbl = NVME_CQHDBL(id, dstrd); 1765 q->q_id = id; 1766 q->q_entries = entries; 1767 q->q_sq_tail = 0; 1768 q->q_cq_head = 0; 1769 q->q_cq_phase = NVME_CQE_PHASE; 1770 1771 nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_PREWRITE); 1772 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_PREREAD); 1773 1774 /* 1775 * Due to definition of full and empty queue (queue is empty 1776 * when head == tail, full when tail is one less then head), 1777 * we can actually only have (entries - 1) in-flight commands. 1778 */ 1779 if (nvme_ccbs_alloc(q, entries - 1) != 0) { 1780 aprint_error_dev(sc->sc_dev, "unable to allocate ccbs\n"); 1781 goto free_cq; 1782 } 1783 1784 return q; 1785 1786 free_cq: 1787 nvme_dmamem_free(sc, q->q_cq_dmamem); 1788 free_sq: 1789 nvme_dmamem_free(sc, q->q_sq_dmamem); 1790 free: 1791 kmem_free(q, sizeof(*q)); 1792 1793 return NULL; 1794 } 1795 1796 static void 1797 nvme_q_free(struct nvme_softc *sc, struct nvme_queue *q) 1798 { 1799 nvme_ccbs_free(q); 1800 mutex_destroy(&q->q_sq_mtx); 1801 mutex_destroy(&q->q_cq_mtx); 1802 nvme_dmamem_sync(sc, q->q_cq_dmamem, BUS_DMASYNC_POSTREAD); 1803 nvme_dmamem_sync(sc, q->q_sq_dmamem, BUS_DMASYNC_POSTWRITE); 1804 nvme_dmamem_free(sc, q->q_cq_dmamem); 1805 nvme_dmamem_free(sc, q->q_sq_dmamem); 1806 kmem_free(q, sizeof(*q)); 1807 } 1808 1809 int 1810 nvme_intr(void *xsc) 1811 { 1812 struct nvme_softc *sc = xsc; 1813 1814 /* 1815 * INTx is level triggered, controller deasserts the interrupt only 1816 * when we advance command queue head via write to the doorbell. 1817 * Tell the controller to block the interrupts while we process 1818 * the queue(s). 1819 */ 1820 nvme_write4(sc, NVME_INTMS, 1); 1821 1822 softint_schedule(sc->sc_softih[0]); 1823 1824 /* don't know, might not have been for us */ 1825 return 1; 1826 } 1827 1828 void 1829 nvme_softintr_intx(void *xq) 1830 { 1831 struct nvme_queue *q = xq; 1832 struct nvme_softc *sc = q->q_sc; 1833 1834 nvme_q_complete(sc, sc->sc_admin_q); 1835 if (sc->sc_q != NULL) 1836 nvme_q_complete(sc, sc->sc_q[0]); 1837 1838 /* 1839 * Processing done, tell controller to issue interrupts again. There 1840 * is no race, as NVMe spec requires the controller to maintain state, 1841 * and assert the interrupt whenever there are unacknowledged 1842 * completion queue entries. 1843 */ 1844 nvme_write4(sc, NVME_INTMC, 1); 1845 } 1846 1847 int 1848 nvme_intr_msi(void *xq) 1849 { 1850 struct nvme_queue *q = xq; 1851 1852 KASSERT(q && q->q_sc && q->q_sc->sc_softih 1853 && q->q_sc->sc_softih[q->q_id]); 1854 1855 /* 1856 * MSI/MSI-X are edge triggered, so can handover processing to softint 1857 * without masking the interrupt. 1858 */ 1859 softint_schedule(q->q_sc->sc_softih[q->q_id]); 1860 1861 return 1; 1862 } 1863 1864 void 1865 nvme_softintr_msi(void *xq) 1866 { 1867 struct nvme_queue *q = xq; 1868 struct nvme_softc *sc = q->q_sc; 1869 1870 nvme_q_complete(sc, q); 1871 } 1872 1873 static struct nvme_dmamem * 1874 nvme_dmamem_alloc(struct nvme_softc *sc, size_t size) 1875 { 1876 struct nvme_dmamem *ndm; 1877 int nsegs; 1878 1879 ndm = kmem_zalloc(sizeof(*ndm), KM_SLEEP); 1880 if (ndm == NULL) 1881 return NULL; 1882 1883 ndm->ndm_size = size; 1884 1885 if (bus_dmamap_create(sc->sc_dmat, size, btoc(round_page(size)), size, 0, 1886 BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &ndm->ndm_map) != 0) 1887 goto ndmfree; 1888 1889 if (bus_dmamem_alloc(sc->sc_dmat, size, sc->sc_mps, 0, &ndm->ndm_seg, 1890 1, &nsegs, BUS_DMA_WAITOK) != 0) 1891 goto destroy; 1892 1893 if (bus_dmamem_map(sc->sc_dmat, &ndm->ndm_seg, nsegs, size, 1894 &ndm->ndm_kva, BUS_DMA_WAITOK) != 0) 1895 goto free; 1896 memset(ndm->ndm_kva, 0, size); 1897 1898 if (bus_dmamap_load(sc->sc_dmat, ndm->ndm_map, ndm->ndm_kva, size, 1899 NULL, BUS_DMA_WAITOK) != 0) 1900 goto unmap; 1901 1902 return ndm; 1903 1904 unmap: 1905 bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, size); 1906 free: 1907 bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1); 1908 destroy: 1909 bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map); 1910 ndmfree: 1911 kmem_free(ndm, sizeof(*ndm)); 1912 return NULL; 1913 } 1914 1915 static void 1916 nvme_dmamem_sync(struct nvme_softc *sc, struct nvme_dmamem *mem, int ops) 1917 { 1918 bus_dmamap_sync(sc->sc_dmat, NVME_DMA_MAP(mem), 1919 0, NVME_DMA_LEN(mem), ops); 1920 } 1921 1922 void 1923 nvme_dmamem_free(struct nvme_softc *sc, struct nvme_dmamem *ndm) 1924 { 1925 bus_dmamap_unload(sc->sc_dmat, ndm->ndm_map); 1926 bus_dmamem_unmap(sc->sc_dmat, ndm->ndm_kva, ndm->ndm_size); 1927 bus_dmamem_free(sc->sc_dmat, &ndm->ndm_seg, 1); 1928 bus_dmamap_destroy(sc->sc_dmat, ndm->ndm_map); 1929 kmem_free(ndm, sizeof(*ndm)); 1930 } 1931 1932 /* 1933 * ioctl 1934 */ 1935 1936 dev_type_open(nvmeopen); 1937 dev_type_close(nvmeclose); 1938 dev_type_ioctl(nvmeioctl); 1939 1940 const struct cdevsw nvme_cdevsw = { 1941 .d_open = nvmeopen, 1942 .d_close = nvmeclose, 1943 .d_read = noread, 1944 .d_write = nowrite, 1945 .d_ioctl = nvmeioctl, 1946 .d_stop = nostop, 1947 .d_tty = notty, 1948 .d_poll = nopoll, 1949 .d_mmap = nommap, 1950 .d_kqfilter = nokqfilter, 1951 .d_discard = nodiscard, 1952 .d_flag = D_OTHER, 1953 }; 1954 1955 /* 1956 * Accept an open operation on the control device. 1957 */ 1958 int 1959 nvmeopen(dev_t dev, int flag, int mode, struct lwp *l) 1960 { 1961 struct nvme_softc *sc; 1962 int unit = minor(dev) / 0x10000; 1963 int nsid = minor(dev) & 0xffff; 1964 int nsidx; 1965 1966 if ((sc = device_lookup_private(&nvme_cd, unit)) == NULL) 1967 return ENXIO; 1968 if ((sc->sc_flags & NVME_F_ATTACHED) == 0) 1969 return ENXIO; 1970 1971 if (nsid == 0) { 1972 /* controller */ 1973 if (ISSET(sc->sc_flags, NVME_F_OPEN)) 1974 return EBUSY; 1975 SET(sc->sc_flags, NVME_F_OPEN); 1976 } else { 1977 /* namespace */ 1978 nsidx = nsid - 1; 1979 if (nsidx >= sc->sc_nn || sc->sc_namespaces[nsidx].dev == NULL) 1980 return ENXIO; 1981 if (ISSET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN)) 1982 return EBUSY; 1983 SET(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN); 1984 } 1985 return 0; 1986 } 1987 1988 /* 1989 * Accept the last close on the control device. 1990 */ 1991 int 1992 nvmeclose(dev_t dev, int flag, int mode, struct lwp *l) 1993 { 1994 struct nvme_softc *sc; 1995 int unit = minor(dev) / 0x10000; 1996 int nsid = minor(dev) & 0xffff; 1997 int nsidx; 1998 1999 sc = device_lookup_private(&nvme_cd, unit); 2000 if (sc == NULL) 2001 return ENXIO; 2002 2003 if (nsid == 0) { 2004 /* controller */ 2005 CLR(sc->sc_flags, NVME_F_OPEN); 2006 } else { 2007 /* namespace */ 2008 nsidx = nsid - 1; 2009 if (nsidx >= sc->sc_nn) 2010 return ENXIO; 2011 CLR(sc->sc_namespaces[nsidx].flags, NVME_NS_F_OPEN); 2012 } 2013 2014 return 0; 2015 } 2016 2017 /* 2018 * Handle control operations. 2019 */ 2020 int 2021 nvmeioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l) 2022 { 2023 struct nvme_softc *sc; 2024 int unit = minor(dev) / 0x10000; 2025 int nsid = minor(dev) & 0xffff; 2026 struct nvme_pt_command *pt; 2027 2028 sc = device_lookup_private(&nvme_cd, unit); 2029 if (sc == NULL) 2030 return ENXIO; 2031 2032 switch (cmd) { 2033 case NVME_PASSTHROUGH_CMD: 2034 pt = data; 2035 return nvme_command_passthrough(sc, data, 2036 nsid == 0 ? pt->cmd.nsid : nsid, l, nsid == 0); 2037 } 2038 2039 return ENOTTY; 2040 } 2041