stand/kboot/hostdisk.c

   1 /*-
   2  * Copyright (C) 2014 Nathan Whitehorn
   3  * All rights reserved.
   4  * Copyright 2022 Netflix, Inc
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  18  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  19  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  20  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
  21  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  22  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  23  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
  24  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/types.h>
  31 #include <sys/disk.h>
  32 #include <stdarg.h>
  33 #include <paths.h>
  34 #include "host_syscall.h"
  35 #include "kboot.h"
  36 #include "bootstrap.h"
  37 #ifdef LOADER_ZFS_SUPPORT
  38 #include "libzfs.h"
  39 #include <sys/zfs_bootenv.h>
  40 #endif
  41
  42 static int hostdisk_init(void);
  43 static int hostdisk_strategy(void *devdata, int flag, daddr_t dblk,
  44     size_t size, char *buf, size_t *rsize);
  45 static int hostdisk_open(struct open_file *f, ...);
  46 static int hostdisk_close(struct open_file *f);
  47 static int hostdisk_ioctl(struct open_file *f, u_long cmd, void *data);
  48 static int hostdisk_print(int verbose);
  49 static char *hostdisk_fmtdev(struct devdesc *vdev);
  50 static bool hostdisk_match(struct devsw *devsw, const char *devspec);
  51 static int hostdisk_parsedev(struct devdesc **idev, const char *devspec, const char **path);
  52
  53 struct devsw hostdisk = {
  54         .dv_name = "/dev",
  55         .dv_type = DEVT_HOSTDISK,
  56         .dv_init = hostdisk_init,
  57         .dv_strategy = hostdisk_strategy,
  58         .dv_open = hostdisk_open,
  59         .dv_close = hostdisk_close,
  60         .dv_ioctl = hostdisk_ioctl,
  61         .dv_print = hostdisk_print,
  62         .dv_cleanup = nullsys,
  63         .dv_fmtdev = hostdisk_fmtdev,
  64         .dv_match = hostdisk_match,
  65         .dv_parsedev = hostdisk_parsedev,
  66 };
  67
  68 /*
  69  * We need to walk through the /sys/block directories looking for
  70  * block devices that we can use.
  71  */
  72 #define SYSBLK "/sys/block"
  73
  74 #define HOSTDISK_MIN_SIZE (16ul << 20)  /* 16MB */
  75
  76 typedef STAILQ_HEAD(, hdinfo) hdinfo_list_t;
  77 typedef struct hdinfo {
  78         STAILQ_ENTRY(hdinfo)    hd_link;        /* link in device list */
  79         hdinfo_list_t   hd_children;
  80         struct hdinfo   *hd_parent;
  81         const char      *hd_dev;
  82         uint64_t        hd_size;                /* In bytes */
  83         uint64_t        hd_sectors;
  84         uint64_t        hd_sectorsize;
  85         int             hd_flags;
  86 #define HDF_HAS_ZPOOL   1                       /* We found a zpool here and uuid valid */
  87         uint64_t        hd_zfs_uuid;
  88 } hdinfo_t;
  89
  90 #define dev2hd(d) ((hdinfo_t *)d->d_opendata)
  91 #define hd_name(hd) ((hd->hd_dev + 5))
  92
  93 static hdinfo_list_t hdinfo = STAILQ_HEAD_INITIALIZER(hdinfo);
  94
  95 typedef bool fef_cb_t(struct host_dirent64 *, void *);
  96 #define FEF_RECURSIVE 1
  97
  98 static bool
  99 foreach_file(const char *dir, fef_cb_t cb, void *argp, u_int flags)
 100 {
 101         char dents[2048];
 102         int fd, dentsize;
 103         struct host_dirent64 *dent;
 104
 105         fd = host_open(dir, O_RDONLY, 0);
 106         if (fd < 0) {
 107                 printf("Can't open %s\n", dir);/* XXX */
 108                 return (false);
 109         }
 110         while (1) {
 111                 dentsize = host_getdents64(fd, dents, sizeof(dents));
 112                 if (dentsize <= 0)
 113                         break;
 114                 for (dent = (struct host_dirent64 *)dents;
 115                      (char *)dent < dents + dentsize;
 116                      dent = (struct host_dirent64 *)((void *)dent + dent->d_reclen)) {
 117                         if (!cb(dent, argp))
 118                                 break;
 119                 }
 120         }
 121         host_close(fd);
 122         return (true);
 123 }
 124
 125 static void
 126 hostdisk_add_part(hdinfo_t *hd, const char *drv, uint64_t secs)
 127 {
 128         hdinfo_t *md;
 129         char *dev;
 130
 131         printf("hd %s adding %s %ju\n", hd->hd_dev, drv, (uintmax_t)secs);
 132         if ((md = calloc(1, sizeof(*md))) == NULL)
 133                 return;
 134         if (asprintf(&dev, "/dev/%s", drv) == -1) {
 135                 printf("hostdisk: no memory\n");
 136                 free(md);
 137                 return;
 138         }
 139         md->hd_dev = dev;
 140         md->hd_sectors = secs;
 141         md->hd_sectorsize = hd->hd_sectorsize;
 142         md->hd_size = md->hd_sectors * md->hd_sectorsize;
 143         md->hd_parent = hd;
 144         STAILQ_INSERT_TAIL(&hd->hd_children, md, hd_link);
 145 }
 146
 147 static bool
 148 hostdisk_one_part(struct host_dirent64 *dent, void *argp)
 149 {
 150         hdinfo_t *hd = argp;
 151         char szfn[1024];
 152         uint64_t sz;
 153
 154         /* Need to skip /dev/ at start of hd_name */
 155         if (strncmp(dent->d_name, hd_name(hd), strlen(hd_name(hd))) != 0)
 156                 return (true);
 157         /* Find out how big this is -- no size not a disk */
 158         snprintf(szfn, sizeof(szfn), "%s/%s/%s/size", SYSBLK,
 159             hd_name(hd), dent->d_name);
 160         if (!file2u64(szfn, &sz))
 161                 return true;
 162         hostdisk_add_part(hd, dent->d_name, sz);
 163         return true;
 164 }
 165
 166 static void
 167 hostdisk_add_parts(hdinfo_t *hd)
 168 {
 169         char fn[1024];
 170
 171         snprintf(fn, sizeof(fn), "%s/%s", SYSBLK, hd_name(hd));
 172         foreach_file(fn, hostdisk_one_part, hd, 0);
 173 }
 174
 175 static void
 176 hostdisk_add_drive(const char *drv, uint64_t secs)
 177 {
 178         hdinfo_t *hd = NULL;
 179         char *dev = NULL;
 180         char fn[1024];
 181
 182         if ((hd = calloc(1, sizeof(*hd))) == NULL)
 183                 return;
 184         if (asprintf(&dev, "/dev/%s", drv) == -1) {
 185                 printf("hostdisk: no memory\n");
 186                 free(hd);
 187                 return;
 188         }
 189         hd->hd_dev = dev;
 190         hd->hd_sectors = secs;
 191         snprintf(fn, sizeof(fn), "%s/%s/queue/hw_sector_size",
 192             SYSBLK, drv);
 193         if (!file2u64(fn, &hd->hd_sectorsize))
 194                 goto err;
 195         hd->hd_size = hd->hd_sectors * hd->hd_sectorsize;
 196         if (hd->hd_size < HOSTDISK_MIN_SIZE)
 197                 goto err;
 198         hd->hd_flags = 0;
 199         STAILQ_INIT(&hd->hd_children);
 200         printf("/dev/%s: %ju %ju %ju\n",
 201             drv, hd->hd_size, hd->hd_sectors, hd->hd_sectorsize);
 202         STAILQ_INSERT_TAIL(&hdinfo, hd, hd_link);
 203         hostdisk_add_parts(hd);
 204         return;
 205 err:
 206         free(dev);
 207         free(hd);
 208         return;
 209 }
 210
 211 /* Find a disk / partition by its filename */
 212
 213 static hdinfo_t *
 214 hostdisk_find(const char *fn)
 215 {
 216         hdinfo_t *hd, *md;
 217
 218         STAILQ_FOREACH(hd, &hdinfo, hd_link) {
 219                 if (strcmp(hd->hd_dev, fn) == 0)
 220                         return (hd);
 221                 STAILQ_FOREACH(md, &hd->hd_children, hd_link) {
 222                         if (strcmp(md->hd_dev, fn) == 0)
 223                                 return (md);
 224                 }
 225         }
 226         return (NULL);
 227 }
 228
 229
 230 static bool
 231 hostdisk_one_disk(struct host_dirent64 *dent, void *argp __unused)
 232 {
 233         char szfn[1024];
 234         uint64_t sz;
 235
 236         /*
 237          * Skip . and ..
 238          */
 239         if (strcmp(dent->d_name, ".") == 0 ||
 240             strcmp(dent->d_name, "..") == 0)
 241                 return (true);
 242
 243         /* Find out how big this is -- no size not a disk */
 244         snprintf(szfn, sizeof(szfn), "%s/%s/size", SYSBLK,
 245             dent->d_name);
 246         if (!file2u64(szfn, &sz))
 247                 return (true);
 248         hostdisk_add_drive(dent->d_name, sz);
 249         return (true);
 250 }
 251
 252 static void
 253 hostdisk_fake_one_disk(char *override)
 254 {
 255         hdinfo_t *hd = NULL;
 256         struct host_kstat sb;
 257
 258         if (host_stat(override, &sb) != 0)
 259                 return;
 260         if (!HOST_S_ISREG(sb.st_mode))
 261                 return;
 262         if (sb.st_size == 0)
 263                 return;
 264         if ((hd = calloc(1, sizeof(*hd))) == NULL)
 265                 return;
 266         if ((hd->hd_dev = strdup(override)) == NULL)
 267                 goto err;
 268         hd->hd_size = sb.st_size;
 269         hd->hd_sectorsize = 512;        /* XXX configurable? */
 270         hd->hd_sectors = hd->hd_size / hd->hd_sectorsize;
 271         if (hd->hd_size < HOSTDISK_MIN_SIZE)
 272                 goto err;
 273         hd->hd_flags = 0;
 274         STAILQ_INIT(&hd->hd_children);
 275         printf("%s: %ju %ju %ju\n",
 276             hd->hd_dev, hd->hd_size, hd->hd_sectors, hd->hd_sectorsize);
 277         STAILQ_INSERT_TAIL(&hdinfo, hd, hd_link);
 278         return;
 279 err:
 280         free(__DECONST(void *, hd->hd_dev));
 281         free(hd);
 282 }
 283
 284 static void
 285 hostdisk_find_block_devices(void)
 286 {
 287         char *override;
 288
 289         override=getenv("hostdisk_override");
 290         if (override != NULL)
 291                 hostdisk_fake_one_disk(override);
 292         else
 293                 foreach_file(SYSBLK, hostdisk_one_disk, NULL, 0);
 294 }
 295
 296 static int
 297 hostdisk_init(void)
 298 {
 299         hostdisk_find_block_devices();
 300
 301         return (0);
 302 }
 303
 304 static int
 305 hostdisk_strategy(void *devdata, int flag, daddr_t dblk, size_t size,
 306     char *buf, size_t *rsize)
 307 {
 308         struct devdesc *desc = devdata;
 309         daddr_t pos;
 310         int n;
 311         int64_t off;
 312         uint64_t res;
 313         uint32_t posl, posh;
 314
 315         pos = dblk * 512;
 316
 317         posl = pos & 0xffffffffu;
 318         posh = (pos >> 32) & 0xffffffffu;
 319         if ((off = host_llseek(desc->d_unit, posh, posl, &res, 0)) < 0) {
 320                 printf("Seek error on fd %d to %ju (dblk %ju) returns %jd\n",
 321                     desc->d_unit, (uintmax_t)pos, (uintmax_t)dblk, (intmax_t)off);
 322                 return (EIO);
 323         }
 324         n = host_read(desc->d_unit, buf, size);
 325
 326         if (n < 0)
 327                 return (EIO);
 328
 329         *rsize = n;
 330         return (0);
 331 }
 332
 333 static int
 334 hostdisk_open(struct open_file *f, ...)
 335 {
 336         struct devdesc *desc;
 337         const char *fn;
 338         va_list vl;
 339
 340         va_start(vl, f);
 341         desc = va_arg(vl, struct devdesc *);
 342         va_end(vl);
 343
 344         fn = dev2hd(desc)->hd_dev;
 345         desc->d_unit = host_open(fn, O_RDONLY, 0);
 346         if (desc->d_unit <= 0) {
 347                 printf("hostdisk_open: couldn't open %s: %d\n", fn, errno);
 348                 return (ENOENT);
 349         }
 350
 351         return (0);
 352 }
 353
 354 static int
 355 hostdisk_close(struct open_file *f)
 356 {
 357         struct devdesc *desc = f->f_devdata;
 358
 359         host_close(desc->d_unit);
 360         return (0);
 361 }
 362
 363 static int
 364 hostdisk_ioctl(struct open_file *f, u_long cmd, void *data)
 365 {
 366         struct devdesc *desc = f->f_devdata;
 367         hdinfo_t *hd = dev2hd(desc);
 368
 369         switch (cmd) {
 370         case DIOCGSECTORSIZE:
 371                 *(u_int *)data = hd->hd_sectorsize;
 372                 break;
 373         case DIOCGMEDIASIZE:
 374                 *(uint64_t *)data = hd->hd_size;
 375                 break;
 376         default:
 377                 return (ENOTTY);
 378         }
 379         return (0);
 380 }
 381
 382 static int
 383 hostdisk_print(int verbose)
 384 {
 385         char line[80];
 386         hdinfo_t *hd, *md;
 387         int ret = 0;
 388
 389         printf("%s devices:", hostdisk.dv_name);
 390         if (pager_output("\n") != 0)
 391                 return (1);
 392
 393         STAILQ_FOREACH(hd, &hdinfo, hd_link) {
 394                 snprintf(line, sizeof(line),
 395                     "   %s: %ju X %ju: %ju bytes\n",
 396                     hd->hd_dev,
 397                     (uintmax_t)hd->hd_sectors,
 398                     (uintmax_t)hd->hd_sectorsize,
 399                     (uintmax_t)hd->hd_size);
 400                 if ((ret = pager_output(line)) != 0)
 401                         break;
 402                 STAILQ_FOREACH(md, &hd->hd_children, hd_link) {
 403                         snprintf(line, sizeof(line),
 404                             "     %s: %ju X %ju: %ju bytes\n",
 405                             md->hd_dev,
 406                             (uintmax_t)md->hd_sectors,
 407                             (uintmax_t)md->hd_sectorsize,
 408                             (uintmax_t)md->hd_size);
 409                         if ((ret = pager_output(line)) != 0)
 410                                 goto done;
 411                 }
 412         }
 413
 414 done:
 415         return (ret);
 416 }
 417
 418 static char *
 419 hostdisk_fmtdev(struct devdesc *vdev)
 420 {
 421         static char name[DEV_DEVLEN];
 422
 423         snprintf(name, sizeof(name), "%s:", dev2hd(vdev)->hd_dev);
 424         return (name);
 425 }
 426
 427 static bool
 428 hostdisk_match(struct devsw *devsw, const char *devspec)
 429 {
 430         hdinfo_t *hd;
 431         const char *colon;
 432         char *cp;
 433
 434         colon = strchr(devspec, ':');
 435         if (colon == NULL)
 436                 return false;
 437         cp = strdup(devspec);
 438         cp[colon - devspec] = '\0';
 439         hd = hostdisk_find(cp);
 440         free(cp);
 441         return (hd != NULL);
 442 }
 443
 444 static int
 445 hostdisk_parsedev(struct devdesc **idev, const char *devspec, const char **path)
 446 {
 447         const char *cp;
 448         struct devdesc *dev;
 449         hdinfo_t *hd;
 450         int len;
 451         char *fn;
 452
 453         /* Must have a : in it */
 454         cp = strchr(devspec, ':');
 455         if (cp == NULL)
 456                 return (EINVAL);
 457         /* XXX Stat the /dev or defer error handling to open(2) call? */
 458         if (path != NULL)
 459                 *path = cp + 1;
 460         len = cp - devspec;
 461         fn = strdup(devspec);
 462         fn[len] = '\0';
 463         hd = hostdisk_find(fn);
 464         if (hd == NULL) {
 465                 printf("Can't find hdinfo for %s\n", fn);
 466                 free(fn);
 467                 return (EINVAL);
 468         }
 469         free(fn);
 470         dev = malloc(sizeof(*dev));
 471         if (dev == NULL)
 472                 return (ENOMEM);
 473         dev->d_unit = 0;
 474         dev->d_dev = &hostdisk;
 475         dev->d_opendata = hd;
 476         *idev = dev;
 477         return (0);
 478 }
 479
 480 /* XXX refactor */
 481 static bool
 482 sanity_check_currdev(void)
 483 {
 484         struct stat st;
 485
 486         return (stat(PATH_DEFAULTS_LOADER_CONF, &st) == 0 ||
 487 #ifdef PATH_BOOTABLE_TOKEN
 488             stat(PATH_BOOTABLE_TOKEN, &st) == 0 || /* non-standard layout */
 489 #endif
 490             stat(PATH_KERNEL, &st) == 0);
 491 }
 492
 493 static const char *
 494 hostdisk_try_one(hdinfo_t *hd)
 495 {
 496         char *fn;
 497
 498         if (asprintf(&fn, "%s:", hd->hd_dev) == -1)
 499                 return (NULL);
 500         set_currdev(fn);
 501         printf("Trying %s\n", fn);
 502         if (sanity_check_currdev())
 503                 return (fn);
 504         printf("Failed %s\n", fn);
 505         free(fn);
 506         return (NULL);
 507 }
 508
 509 const char *
 510 hostdisk_gen_probe(void)
 511 {
 512         hdinfo_t *hd, *md;
 513         const char *rv = NULL;
 514
 515         STAILQ_FOREACH(hd, &hdinfo, hd_link) {
 516                 /* try whole disk */
 517                 if (hd->hd_flags & HDF_HAS_ZPOOL)
 518                         continue;
 519                 rv = hostdisk_try_one(hd);
 520                 if (rv != NULL)
 521                         return (rv);
 522
 523                 /* try all partitions */
 524                 STAILQ_FOREACH(md, &hd->hd_children, hd_link) {
 525                         if (md->hd_flags & HDF_HAS_ZPOOL)
 526                                 continue;
 527                         rv = hostdisk_try_one(md);
 528                         if (rv != NULL)
 529                                 return (rv);
 530                 }
 531         }
 532         return (NULL);
 533 }
 534
 535 #ifdef LOADER_ZFS_SUPPORT
 536 static bool
 537 hostdisk_zfs_check_one(hdinfo_t *hd)
 538 {
 539         char *fn;
 540         bool found = false;
 541         uint64_t pool_uuid;
 542
 543         if (asprintf(&fn, "%s:", hd->hd_dev) == -1)
 544                 return (false);
 545         pool_uuid = 0;
 546         zfs_probe_dev(fn, &pool_uuid, false);
 547         if (pool_uuid != 0) {
 548                 found = true;
 549                 hd->hd_flags |= HDF_HAS_ZPOOL;
 550                 hd->hd_zfs_uuid = pool_uuid;
 551         }
 552         free(fn);
 553
 554         return (found);
 555 }
 556
 557 void
 558 hostdisk_zfs_probe(void)
 559 {
 560         hdinfo_t *hd, *md;
 561
 562         STAILQ_FOREACH(hd, &hdinfo, hd_link) {
 563                 if (hostdisk_zfs_check_one(hd))
 564                         continue;
 565                 STAILQ_FOREACH(md, &hd->hd_children, hd_link) {
 566                         hostdisk_zfs_check_one(md);
 567                 }
 568         }
 569 }
 570
 571 /* This likely shoud move to libsa/zfs/zfs.c and be used by at least EFI booting */
 572 static bool
 573 probe_zfs_currdev(uint64_t pool_guid, uint64_t root_guid, bool setcurrdev)
 574 {
 575         char *devname;
 576         struct zfs_devdesc currdev;
 577         bool bootable;
 578
 579         currdev.dd.d_dev = &zfs_dev;
 580         currdev.dd.d_unit = 0;
 581         currdev.pool_guid = pool_guid;
 582         currdev.root_guid = root_guid;
 583         devname = devformat(&currdev.dd);
 584         if (setcurrdev)
 585                 set_currdev(devname);
 586
 587         bootable = sanity_check_currdev();
 588         if (bootable) {
 589                 char buf[VDEV_PAD_SIZE];
 590
 591                 if (zfs_get_bootonce(&currdev, OS_BOOTONCE, buf, sizeof(buf)) == 0) {
 592                         printf("zfs bootonce: %s\n", buf);
 593                         if (setcurrdev)
 594                                 set_currdev(buf);
 595                         setenv("zfs-bootonce", buf, 1);
 596                 }
 597                 (void)zfs_attach_nvstore(&currdev);
 598                 init_zfs_boot_options(devname);
 599         }
 600         return (bootable);
 601 }
 602
 603 static bool
 604 hostdisk_zfs_try_default(hdinfo_t *hd)
 605 {
 606         return (probe_zfs_currdev(hd->hd_zfs_uuid, 0, true));
 607 }
 608
 609 bool
 610 hostdisk_zfs_find_default(void)
 611 {
 612         hdinfo_t *hd, *md;
 613
 614         STAILQ_FOREACH(hd, &hdinfo, hd_link) {
 615                 if (hd->hd_flags & HDF_HAS_ZPOOL) {
 616                         if (hostdisk_zfs_try_default(hd))
 617                                 return (true);
 618                         continue;
 619                 }
 620                 STAILQ_FOREACH(md, &hd->hd_children, hd_link) {
 621                         if (md->hd_flags & HDF_HAS_ZPOOL) {
 622                                 if (hostdisk_zfs_try_default(md))
 623                                         return (true);
 624                         }
 625                 }
 626         }
 627         return (false);
 628 }
 629 #endif