sys/dev/nvdimm/nvdimm.c

   1 /*-
   2  * Copyright (c) 2017 The FreeBSD Foundation
   3  * All rights reserved.
   4  * Copyright (c) 2018, 2019 Intel Corporation
   5  *
   6  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
   7  * under sponsorship from the FreeBSD Foundation.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_acpi.h"
  35 #include "opt_ddb.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/bio.h>
  40 #include <sys/bitstring.h>
  41 #include <sys/bus.h>
  42 #include <sys/kernel.h>
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/module.h>
  46 #include <sys/sbuf.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/uuid.h>
  49
  50 #include <contrib/dev/acpica/include/acpi.h>
  51 #include <contrib/dev/acpica/include/accommon.h>
  52 #include <contrib/dev/acpica/include/acuuid.h>
  53 #include <dev/acpica/acpivar.h>
  54
  55 #include <dev/nvdimm/nvdimm_var.h>
  56
  57 #define _COMPONENT      ACPI_OEM
  58 ACPI_MODULE_NAME("NVDIMM")
  59
  60 static struct uuid intel_nvdimm_dsm_uuid =
  61     {0x4309AC30,0x0D11,0x11E4,0x91,0x91,{0x08,0x00,0x20,0x0C,0x9A,0x66}};
  62 #define INTEL_NVDIMM_DSM_REV 1
  63 #define INTEL_NVDIMM_DSM_GET_LABEL_SIZE 4
  64 #define INTEL_NVDIMM_DSM_GET_LABEL_DATA 5
  65
  66 static devclass_t nvdimm_devclass;
  67 MALLOC_DEFINE(M_NVDIMM, "nvdimm", "NVDIMM driver memory");
  68
  69 static int
  70 read_label_area_size(struct nvdimm_dev *nv)
  71 {
  72         ACPI_OBJECT *result_buffer;
  73         ACPI_HANDLE handle;
  74         ACPI_STATUS status;
  75         ACPI_BUFFER result;
  76         uint32_t *out;
  77         int error;
  78
  79         handle = nvdimm_root_get_acpi_handle(nv->nv_dev);
  80         if (handle == NULL)
  81                 return (ENODEV);
  82         result.Length = ACPI_ALLOCATE_BUFFER;
  83         result.Pointer = NULL;
  84         status = acpi_EvaluateDSM(handle, (uint8_t *)&intel_nvdimm_dsm_uuid,
  85             INTEL_NVDIMM_DSM_REV, INTEL_NVDIMM_DSM_GET_LABEL_SIZE, NULL,
  86             &result);
  87         error = ENXIO;
  88         if (ACPI_SUCCESS(status) && result.Pointer != NULL &&
  89             result.Length >= sizeof(ACPI_OBJECT)) {
  90                 result_buffer = result.Pointer;
  91                 if (result_buffer->Type == ACPI_TYPE_BUFFER &&
  92                     result_buffer->Buffer.Length >= 12) {
  93                         out = (uint32_t *)result_buffer->Buffer.Pointer;
  94                         nv->label_area_size = out[1];
  95                         nv->max_label_xfer = out[2];
  96                         error = 0;
  97                 }
  98         }
  99         if (result.Pointer != NULL)
 100                 AcpiOsFree(result.Pointer);
 101         return (error);
 102 }
 103
 104 static int
 105 read_label_area(struct nvdimm_dev *nv, uint8_t *dest, off_t offset,
 106     off_t length)
 107 {
 108         ACPI_BUFFER result;
 109         ACPI_HANDLE handle;
 110         ACPI_OBJECT params_pkg, params_buf, *result_buf;
 111         ACPI_STATUS status;
 112         uint32_t params[2];
 113         off_t to_read;
 114         int error;
 115
 116         error = 0;
 117         handle = nvdimm_root_get_acpi_handle(nv->nv_dev);
 118         if (offset < 0 || length <= 0 ||
 119             offset + length > nv->label_area_size ||
 120             handle == NULL)
 121                 return (ENODEV);
 122         params_pkg.Type = ACPI_TYPE_PACKAGE;
 123         params_pkg.Package.Count = 1;
 124         params_pkg.Package.Elements = &params_buf;
 125         params_buf.Type = ACPI_TYPE_BUFFER;
 126         params_buf.Buffer.Length = sizeof(params);
 127         params_buf.Buffer.Pointer = (UINT8 *)params;
 128         while (length > 0) {
 129                 to_read = MIN(length, nv->max_label_xfer);
 130                 params[0] = offset;
 131                 params[1] = to_read;
 132                 result.Length = ACPI_ALLOCATE_BUFFER;
 133                 result.Pointer = NULL;
 134                 status = acpi_EvaluateDSM(handle,
 135                     (uint8_t *)&intel_nvdimm_dsm_uuid, INTEL_NVDIMM_DSM_REV,
 136                     INTEL_NVDIMM_DSM_GET_LABEL_DATA, &params_pkg, &result);
 137                 if (ACPI_FAILURE(status) ||
 138                     result.Length < sizeof(ACPI_OBJECT) ||
 139                     result.Pointer == NULL) {
 140                         error = ENXIO;
 141                         break;
 142                 }
 143                 result_buf = (ACPI_OBJECT *)result.Pointer;
 144                 if (result_buf->Type != ACPI_TYPE_BUFFER ||
 145                     result_buf->Buffer.Pointer == NULL ||
 146                     result_buf->Buffer.Length != 4 + to_read ||
 147                     ((uint16_t *)result_buf->Buffer.Pointer)[0] != 0) {
 148                         error = ENXIO;
 149                         break;
 150                 }
 151                 bcopy(result_buf->Buffer.Pointer + 4, dest, to_read);
 152                 dest += to_read;
 153                 offset += to_read;
 154                 length -= to_read;
 155                 if (result.Pointer != NULL) {
 156                         AcpiOsFree(result.Pointer);
 157                         result.Pointer = NULL;
 158                 }
 159         }
 160         if (result.Pointer != NULL)
 161                 AcpiOsFree(result.Pointer);
 162         return (error);
 163 }
 164
 165 static uint64_t
 166 fletcher64(const void *data, size_t length)
 167 {
 168         size_t i;
 169         uint32_t a, b;
 170         const uint32_t *d;
 171
 172         a = 0;
 173         b = 0;
 174         d = (const uint32_t *)data;
 175         length = length / sizeof(uint32_t);
 176         for (i = 0; i < length; i++) {
 177                 a += d[i];
 178                 b += a;
 179         }
 180         return ((uint64_t)b << 32 | a);
 181 }
 182
 183 static bool
 184 label_index_is_valid(struct nvdimm_label_index *index, uint32_t max_labels,
 185     size_t size, size_t offset)
 186 {
 187         uint64_t checksum;
 188
 189         index = (struct nvdimm_label_index *)((uint8_t *)index + size * offset);
 190         if (strcmp(index->signature, NVDIMM_INDEX_BLOCK_SIGNATURE) != 0)
 191                 return false;
 192         checksum = index->checksum;
 193         index->checksum = 0;
 194         if (checksum != fletcher64(index, size) ||
 195             index->this_offset != size * offset || index->this_size != size ||
 196             index->other_offset != size * (offset == 0 ? 1 : 0) ||
 197             index->seq == 0 || index->seq > 3 || index->slot_cnt > max_labels ||
 198             index->label_size != 1)
 199                 return false;
 200         return true;
 201 }
 202
 203 static int
 204 read_label(struct nvdimm_dev *nv, int num)
 205 {
 206         struct nvdimm_label_entry *entry, *i, *next;
 207         uint64_t checksum;
 208         off_t offset;
 209         int error;
 210
 211         offset = nv->label_index->label_offset +
 212             num * (128 << nv->label_index->label_size);
 213         entry = malloc(sizeof(*entry), M_NVDIMM, M_WAITOK);
 214         error = read_label_area(nv, (uint8_t *)&entry->label, offset,
 215             sizeof(struct nvdimm_label));
 216         if (error != 0) {
 217                 free(entry, M_NVDIMM);
 218                 return (error);
 219         }
 220         checksum = entry->label.checksum;
 221         entry->label.checksum = 0;
 222         if (checksum != fletcher64(&entry->label, sizeof(entry->label)) ||
 223             entry->label.slot != num) {
 224                 free(entry, M_NVDIMM);
 225                 return (ENXIO);
 226         }
 227
 228         /* Insertion ordered by dimm_phys_addr */
 229         if (SLIST_EMPTY(&nv->labels) ||
 230             entry->label.dimm_phys_addr <=
 231             SLIST_FIRST(&nv->labels)->label.dimm_phys_addr) {
 232                 SLIST_INSERT_HEAD(&nv->labels, entry, link);
 233                 return (0);
 234         }
 235         SLIST_FOREACH_SAFE(i, &nv->labels, link, next) {
 236                 if (next == NULL ||
 237                     entry->label.dimm_phys_addr <= next->label.dimm_phys_addr) {
 238                         SLIST_INSERT_AFTER(i, entry, link);
 239                         return (0);
 240                 }
 241         }
 242         __assert_unreachable();
 243 }
 244
 245 static int
 246 read_labels(struct nvdimm_dev *nv)
 247 {
 248         struct nvdimm_label_index *indices, *index1;
 249         size_t index_size, num_labels;
 250         int error, n;
 251         bool index_0_valid, index_1_valid;
 252
 253         for (index_size = 256; ; index_size += 256) {
 254                 num_labels = 8 * (index_size -
 255                     sizeof(struct nvdimm_label_index));
 256                 if (index_size + num_labels * sizeof(struct nvdimm_label) >=
 257                     nv->label_area_size)
 258                         break;
 259         }
 260         num_labels = (nv->label_area_size - index_size) /
 261             sizeof(struct nvdimm_label);
 262         indices = malloc(2 * index_size, M_NVDIMM, M_WAITOK);
 263         index1 = (void *)((uint8_t *)indices + index_size);
 264         error = read_label_area(nv, (void *)indices, 0, 2 * index_size);
 265         if (error != 0) {
 266                 free(indices, M_NVDIMM);
 267                 return (error);
 268         }
 269         index_0_valid = label_index_is_valid(indices, num_labels, index_size,
 270             0);
 271         index_1_valid = label_index_is_valid(indices, num_labels, index_size,
 272             1);
 273         if (!index_0_valid && !index_1_valid) {
 274                 free(indices, M_NVDIMM);
 275                 return (ENXIO);
 276         }
 277         if (index_0_valid && index_1_valid) {
 278                 if (((int)indices->seq - (int)index1->seq + 3) % 3 == 1) {
 279                         /* index 0 was more recently updated */
 280                         index_1_valid = false;
 281                 } else {
 282                         /*
 283                          * either index 1 was more recently updated,
 284                          * or the sequence numbers are equal, in which
 285                          * case the specification says the block with
 286                          * the higher offset is to be treated as valid
 287                          */
 288                         index_0_valid = false;
 289                 }
 290         }
 291         nv->label_index = malloc(index_size, M_NVDIMM, M_WAITOK);
 292         bcopy(index_0_valid ? indices : index1, nv->label_index, index_size);
 293         free(indices, M_NVDIMM);
 294         bit_ffc_at((bitstr_t *)nv->label_index->free, 0,
 295             nv->label_index->slot_cnt, &n);
 296         while (n >= 0) {
 297                 read_label(nv, n);
 298                 bit_ffc_at((bitstr_t *)nv->label_index->free, n + 1,
 299                     nv->label_index->slot_cnt, &n);
 300         }
 301         return (0);
 302 }
 303
 304 struct nvdimm_dev *
 305 nvdimm_find_by_handle(nfit_handle_t nv_handle)
 306 {
 307         struct nvdimm_dev *res;
 308         device_t *dimms;
 309         int i, error, num_dimms;
 310
 311         res = NULL;
 312         error = devclass_get_devices(nvdimm_devclass, &dimms, &num_dimms);
 313         if (error != 0)
 314                 return (NULL);
 315         for (i = 0; i < num_dimms; i++) {
 316                 if (nvdimm_root_get_device_handle(dimms[i]) == nv_handle) {
 317                         res = device_get_softc(dimms[i]);
 318                         break;
 319                 }
 320         }
 321         free(dimms, M_TEMP);
 322         return (res);
 323 }
 324
 325 static int
 326 nvdimm_probe(device_t dev)
 327 {
 328
 329         return (BUS_PROBE_NOWILDCARD);
 330 }
 331
 332 static int
 333 nvdimm_attach(device_t dev)
 334 {
 335         struct nvdimm_dev *nv;
 336         struct sysctl_ctx_list *ctx;
 337         struct sysctl_oid *oid;
 338         struct sysctl_oid_list *children;
 339         struct sbuf *sb;
 340         ACPI_TABLE_NFIT *nfitbl;
 341         ACPI_HANDLE handle;
 342         ACPI_STATUS status;
 343         ACPI_NFIT_MEMORY_MAP **maps;
 344         int error, i, num_maps;
 345         uint16_t flags;
 346
 347         nv = device_get_softc(dev);
 348         ctx = device_get_sysctl_ctx(dev);
 349         oid = device_get_sysctl_tree(dev);
 350         children = SYSCTL_CHILDREN(oid);
 351         handle = nvdimm_root_get_acpi_handle(dev);
 352         MPASS(handle != NULL);
 353         nv->nv_dev = dev;
 354         nv->nv_handle = nvdimm_root_get_device_handle(dev);
 355
 356         status = AcpiGetTable(ACPI_SIG_NFIT, 1, (ACPI_TABLE_HEADER **)&nfitbl);
 357         if (ACPI_FAILURE(status)) {
 358                 if (bootverbose)
 359                         device_printf(dev, "cannot get NFIT\n");
 360                 return (ENXIO);
 361         }
 362         acpi_nfit_get_flush_addrs(nfitbl, nv->nv_handle, &nv->nv_flush_addr,
 363             &nv->nv_flush_addr_cnt);
 364
 365         /*
 366          * Each NVDIMM should have at least one memory map associated with it.
 367          * If any of the maps have one of the error flags set, reflect that in
 368          * the overall status.
 369          */
 370         acpi_nfit_get_memory_maps_by_dimm(nfitbl, nv->nv_handle, &maps,
 371             &num_maps);
 372         if (num_maps == 0) {
 373                 free(nv->nv_flush_addr, M_NVDIMM);
 374                 free(maps, M_NVDIMM);
 375                 device_printf(dev, "cannot find memory map\n");
 376                 return (ENXIO);
 377         }
 378         flags = 0;
 379         for (i = 0; i < num_maps; i++) {
 380                 flags |= maps[i]->Flags;
 381         }
 382         free(maps, M_NVDIMM);
 383
 384         /* sbuf_new_auto(9) is M_WAITOK; no need to check for NULL. */
 385         sb = sbuf_new_auto();
 386         (void) sbuf_printf(sb, "0x%b", flags,
 387             "\20"
 388             "\001SAVE_FAILED"
 389             "\002RESTORE_FAILED"
 390             "\003FLUSH_FAILED"
 391             "\004NOT_ARMED"
 392             "\005HEALTH_OBSERVED"
 393             "\006HEALTH_ENABLED"
 394             "\007MAP_FAILED");
 395         error = sbuf_finish(sb);
 396         if (error != 0) {
 397                 sbuf_delete(sb);
 398                 free(nv->nv_flush_addr, M_NVDIMM);
 399                 device_printf(dev, "cannot convert flags to string\n");
 400                 return (error);
 401         }
 402         /* strdup(9) is M_WAITOK; no need to check for NULL. */
 403         nv->nv_flags_str = strdup(sbuf_data(sb), M_NVDIMM);
 404         sbuf_delete(sb);
 405         SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "flags",
 406             CTLFLAG_RD | CTLFLAG_MPSAFE, nv->nv_flags_str, 0,
 407             "NVDIMM State Flags");
 408         /*
 409          * Anything other than HEALTH_ENABLED indicates a fault condition of
 410          * some kind, so log if that's seen.
 411          */
 412         if ((flags & ~ACPI_NFIT_MEM_HEALTH_ENABLED) != 0)
 413                 device_printf(dev, "flags: %s\n", nv->nv_flags_str);
 414
 415         AcpiPutTable(&nfitbl->Header);
 416         error = read_label_area_size(nv);
 417         if (error == 0) {
 418                 /*
 419                  * Ignoring errors reading labels. Not all NVDIMMs
 420                  * support labels and namespaces.
 421                  */
 422                 read_labels(nv);
 423         }
 424         return (0);
 425 }
 426
 427 static int
 428 nvdimm_detach(device_t dev)
 429 {
 430         struct nvdimm_dev *nv;
 431         struct nvdimm_label_entry *label, *next;
 432
 433         nv = device_get_softc(dev);
 434         free(nv->nv_flags_str, M_NVDIMM);
 435         free(nv->nv_flush_addr, M_NVDIMM);
 436         free(nv->label_index, M_NVDIMM);
 437         SLIST_FOREACH_SAFE(label, &nv->labels, link, next) {
 438                 SLIST_REMOVE_HEAD(&nv->labels, link);
 439                 free(label, M_NVDIMM);
 440         }
 441         return (0);
 442 }
 443
 444 static int
 445 nvdimm_suspend(device_t dev)
 446 {
 447
 448         return (0);
 449 }
 450
 451 static int
 452 nvdimm_resume(device_t dev)
 453 {
 454
 455         return (0);
 456 }
 457
 458 static device_method_t nvdimm_methods[] = {
 459         DEVMETHOD(device_probe, nvdimm_probe),
 460         DEVMETHOD(device_attach, nvdimm_attach),
 461         DEVMETHOD(device_detach, nvdimm_detach),
 462         DEVMETHOD(device_suspend, nvdimm_suspend),
 463         DEVMETHOD(device_resume, nvdimm_resume),
 464         DEVMETHOD_END
 465 };
 466
 467 static driver_t nvdimm_driver = {
 468         "nvdimm",
 469         nvdimm_methods,
 470         sizeof(struct nvdimm_dev),
 471 };
 472
 473 DRIVER_MODULE(nvdimm, nvdimm_acpi_root, nvdimm_driver, nvdimm_devclass, NULL,
 474     NULL);
 475 MODULE_DEPEND(nvdimm, acpi, 1, 1, 1);