1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2014-2016 Christoph Hellwig. 4 */ 5 #include <linux/sunrpc/svc.h> 6 #include <linux/blkdev.h> 7 #include <linux/fs_struct.h> 8 #include <linux/nfs4.h> 9 #include <linux/nfs_fs.h> 10 #include <linux/nfs_xdr.h> 11 #include <linux/pr.h> 12 13 #include "blocklayout.h" 14 #include "../nfs4trace.h" 15 16 #define NFSDBG_FACILITY NFSDBG_PNFS_LD 17 18 static void bl_unregister_scsi(struct pnfs_block_dev *dev) 19 { 20 struct block_device *bdev = file_bdev(dev->bdev_file); 21 const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops; 22 int status; 23 24 status = ops->pr_register(bdev, dev->pr_key, 0, false); 25 if (status) 26 trace_bl_pr_key_unreg_err(bdev, dev->pr_key, status); 27 else 28 trace_bl_pr_key_unreg(bdev, dev->pr_key); 29 } 30 31 static bool bl_register_scsi(struct pnfs_block_dev *dev) 32 { 33 struct block_device *bdev = file_bdev(dev->bdev_file); 34 const struct pr_ops *ops = bdev->bd_disk->fops->pr_ops; 35 int status; 36 37 if (test_and_set_bit(PNFS_BDEV_REGISTERED, &dev->flags)) 38 return true; 39 40 status = ops->pr_register(bdev, 0, dev->pr_key, true); 41 if (status) { 42 trace_bl_pr_key_reg_err(bdev, dev->pr_key, status); 43 return false; 44 } 45 trace_bl_pr_key_reg(bdev, dev->pr_key); 46 return true; 47 } 48 49 static void bl_unregister_dev(struct pnfs_block_dev *dev) 50 { 51 u32 i; 52 53 if (dev->nr_children) { 54 for (i = 0; i < dev->nr_children; i++) 55 bl_unregister_dev(&dev->children[i]); 56 return; 57 } 58 59 if (dev->type == PNFS_BLOCK_VOLUME_SCSI && 60 test_and_clear_bit(PNFS_BDEV_REGISTERED, &dev->flags)) 61 bl_unregister_scsi(dev); 62 } 63 64 bool bl_register_dev(struct pnfs_block_dev *dev) 65 { 66 u32 i; 67 68 if (dev->nr_children) { 69 for (i = 0; i < dev->nr_children; i++) { 70 if (!bl_register_dev(&dev->children[i])) { 71 while (i > 0) 72 bl_unregister_dev(&dev->children[--i]); 73 return false; 74 } 75 } 76 return true; 77 } 78 79 if (dev->type == PNFS_BLOCK_VOLUME_SCSI) 80 return bl_register_scsi(dev); 81 return true; 82 } 83 84 static void 85 bl_free_device(struct pnfs_block_dev *dev) 86 { 87 bl_unregister_dev(dev); 88 89 if (dev->nr_children) { 90 int i; 91 92 for (i = 0; i < dev->nr_children; i++) 93 bl_free_device(&dev->children[i]); 94 kfree(dev->children); 95 } else { 96 if (dev->bdev_file) 97 fput(dev->bdev_file); 98 } 99 } 100 101 void 102 bl_free_deviceid_node(struct nfs4_deviceid_node *d) 103 { 104 struct pnfs_block_dev *dev = 105 container_of(d, struct pnfs_block_dev, node); 106 107 bl_free_device(dev); 108 kfree_rcu(dev, node.rcu); 109 } 110 111 static int 112 nfs4_block_decode_volume(struct xdr_stream *xdr, struct pnfs_block_volume *b) 113 { 114 __be32 *p; 115 int i; 116 117 p = xdr_inline_decode(xdr, 4); 118 if (!p) 119 return -EIO; 120 b->type = be32_to_cpup(p++); 121 122 switch (b->type) { 123 case PNFS_BLOCK_VOLUME_SIMPLE: 124 p = xdr_inline_decode(xdr, 4); 125 if (!p) 126 return -EIO; 127 b->simple.nr_sigs = be32_to_cpup(p++); 128 if (!b->simple.nr_sigs || b->simple.nr_sigs > PNFS_BLOCK_MAX_UUIDS) { 129 dprintk("Bad signature count: %d\n", b->simple.nr_sigs); 130 return -EIO; 131 } 132 133 b->simple.len = 4 + 4; 134 for (i = 0; i < b->simple.nr_sigs; i++) { 135 p = xdr_inline_decode(xdr, 8 + 4); 136 if (!p) 137 return -EIO; 138 p = xdr_decode_hyper(p, &b->simple.sigs[i].offset); 139 b->simple.sigs[i].sig_len = be32_to_cpup(p++); 140 if (b->simple.sigs[i].sig_len > PNFS_BLOCK_UUID_LEN) { 141 pr_info("signature too long: %d\n", 142 b->simple.sigs[i].sig_len); 143 return -EIO; 144 } 145 146 p = xdr_inline_decode(xdr, b->simple.sigs[i].sig_len); 147 if (!p) 148 return -EIO; 149 memcpy(&b->simple.sigs[i].sig, p, 150 b->simple.sigs[i].sig_len); 151 152 b->simple.len += 8 + 4 + \ 153 (XDR_QUADLEN(b->simple.sigs[i].sig_len) << 2); 154 } 155 break; 156 case PNFS_BLOCK_VOLUME_SLICE: 157 p = xdr_inline_decode(xdr, 8 + 8 + 4); 158 if (!p) 159 return -EIO; 160 p = xdr_decode_hyper(p, &b->slice.start); 161 p = xdr_decode_hyper(p, &b->slice.len); 162 b->slice.volume = be32_to_cpup(p++); 163 break; 164 case PNFS_BLOCK_VOLUME_CONCAT: 165 p = xdr_inline_decode(xdr, 4); 166 if (!p) 167 return -EIO; 168 169 b->concat.volumes_count = be32_to_cpup(p++); 170 if (b->concat.volumes_count > PNFS_BLOCK_MAX_DEVICES) { 171 dprintk("Too many volumes: %d\n", b->concat.volumes_count); 172 return -EIO; 173 } 174 175 p = xdr_inline_decode(xdr, b->concat.volumes_count * 4); 176 if (!p) 177 return -EIO; 178 for (i = 0; i < b->concat.volumes_count; i++) 179 b->concat.volumes[i] = be32_to_cpup(p++); 180 break; 181 case PNFS_BLOCK_VOLUME_STRIPE: 182 p = xdr_inline_decode(xdr, 8 + 4); 183 if (!p) 184 return -EIO; 185 186 p = xdr_decode_hyper(p, &b->stripe.chunk_size); 187 b->stripe.volumes_count = be32_to_cpup(p++); 188 if (b->stripe.volumes_count > PNFS_BLOCK_MAX_DEVICES) { 189 dprintk("Too many volumes: %d\n", b->stripe.volumes_count); 190 return -EIO; 191 } 192 193 p = xdr_inline_decode(xdr, b->stripe.volumes_count * 4); 194 if (!p) 195 return -EIO; 196 for (i = 0; i < b->stripe.volumes_count; i++) 197 b->stripe.volumes[i] = be32_to_cpup(p++); 198 break; 199 case PNFS_BLOCK_VOLUME_SCSI: 200 p = xdr_inline_decode(xdr, 4 + 4 + 4); 201 if (!p) 202 return -EIO; 203 b->scsi.code_set = be32_to_cpup(p++); 204 b->scsi.designator_type = be32_to_cpup(p++); 205 b->scsi.designator_len = be32_to_cpup(p++); 206 p = xdr_inline_decode(xdr, b->scsi.designator_len); 207 if (!p) 208 return -EIO; 209 if (b->scsi.designator_len > 256) 210 return -EIO; 211 memcpy(&b->scsi.designator, p, b->scsi.designator_len); 212 p = xdr_inline_decode(xdr, 8); 213 if (!p) 214 return -EIO; 215 p = xdr_decode_hyper(p, &b->scsi.pr_key); 216 break; 217 default: 218 dprintk("unknown volume type!\n"); 219 return -EIO; 220 } 221 222 return 0; 223 } 224 225 static bool bl_map_simple(struct pnfs_block_dev *dev, u64 offset, 226 struct pnfs_block_dev_map *map) 227 { 228 map->start = dev->start; 229 map->len = dev->len; 230 map->disk_offset = dev->disk_offset; 231 map->bdev = file_bdev(dev->bdev_file); 232 return true; 233 } 234 235 static bool bl_map_concat(struct pnfs_block_dev *dev, u64 offset, 236 struct pnfs_block_dev_map *map) 237 { 238 int i; 239 240 for (i = 0; i < dev->nr_children; i++) { 241 struct pnfs_block_dev *child = &dev->children[i]; 242 243 if (child->start > offset || 244 child->start + child->len <= offset) 245 continue; 246 247 child->map(child, offset - child->start, map); 248 return true; 249 } 250 251 dprintk("%s: ran off loop!\n", __func__); 252 return false; 253 } 254 255 static bool bl_map_stripe(struct pnfs_block_dev *dev, u64 offset, 256 struct pnfs_block_dev_map *map) 257 { 258 struct pnfs_block_dev *child; 259 u64 chunk; 260 u32 chunk_idx; 261 u64 disk_chunk; 262 u64 disk_offset; 263 264 chunk = div_u64(offset, dev->chunk_size); 265 disk_chunk = div_u64_rem(chunk, dev->nr_children, &chunk_idx); 266 267 if (chunk_idx >= dev->nr_children) { 268 dprintk("%s: invalid chunk idx %d (%lld/%lld)\n", 269 __func__, chunk_idx, offset, dev->chunk_size); 270 /* error, should not happen */ 271 return false; 272 } 273 274 /* truncate offset to the beginning of the stripe */ 275 offset = chunk * dev->chunk_size; 276 277 /* disk offset of the stripe */ 278 disk_offset = disk_chunk * dev->chunk_size; 279 280 child = &dev->children[chunk_idx]; 281 child->map(child, disk_offset, map); 282 283 map->start += offset; 284 map->disk_offset += disk_offset; 285 map->len = dev->chunk_size; 286 return true; 287 } 288 289 static int 290 bl_parse_deviceid(struct nfs_server *server, struct pnfs_block_dev *d, 291 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask); 292 293 294 static int 295 bl_parse_simple(struct nfs_server *server, struct pnfs_block_dev *d, 296 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 297 { 298 struct pnfs_block_volume *v = &volumes[idx]; 299 struct file *bdev_file; 300 dev_t dev; 301 302 dev = bl_resolve_deviceid(server, v, gfp_mask); 303 if (!dev) 304 return -EIO; 305 306 bdev_file = bdev_file_open_by_dev(dev, BLK_OPEN_READ | BLK_OPEN_WRITE, 307 NULL, NULL); 308 if (IS_ERR(bdev_file)) { 309 printk(KERN_WARNING "pNFS: failed to open device %d:%d (%ld)\n", 310 MAJOR(dev), MINOR(dev), PTR_ERR(bdev_file)); 311 return PTR_ERR(bdev_file); 312 } 313 d->bdev_file = bdev_file; 314 d->len = bdev_nr_bytes(file_bdev(bdev_file)); 315 d->map = bl_map_simple; 316 317 printk(KERN_INFO "pNFS: using block device %s\n", 318 file_bdev(bdev_file)->bd_disk->disk_name); 319 return 0; 320 } 321 322 static bool 323 bl_validate_designator(struct pnfs_block_volume *v) 324 { 325 switch (v->scsi.designator_type) { 326 case PS_DESIGNATOR_EUI64: 327 if (v->scsi.code_set != PS_CODE_SET_BINARY) 328 return false; 329 330 if (v->scsi.designator_len != 8 && 331 v->scsi.designator_len != 10 && 332 v->scsi.designator_len != 16) 333 return false; 334 335 return true; 336 case PS_DESIGNATOR_NAA: 337 if (v->scsi.code_set != PS_CODE_SET_BINARY) 338 return false; 339 340 if (v->scsi.designator_len != 8 && 341 v->scsi.designator_len != 16) 342 return false; 343 344 return true; 345 case PS_DESIGNATOR_T10: 346 case PS_DESIGNATOR_NAME: 347 pr_err("pNFS: unsupported designator " 348 "(code set %d, type %d, len %d.\n", 349 v->scsi.code_set, 350 v->scsi.designator_type, 351 v->scsi.designator_len); 352 return false; 353 default: 354 pr_err("pNFS: invalid designator " 355 "(code set %d, type %d, len %d.\n", 356 v->scsi.code_set, 357 v->scsi.designator_type, 358 v->scsi.designator_len); 359 return false; 360 } 361 } 362 363 static struct file * 364 bl_open_path(struct pnfs_block_volume *v, const char *prefix) 365 { 366 struct file *bdev_file; 367 const char *devname __free(kfree) = NULL; 368 369 devname = kasprintf(GFP_KERNEL, "/dev/disk/by-id/%s%*phN", 370 prefix, v->scsi.designator_len, v->scsi.designator); 371 if (!devname) 372 return ERR_PTR(-ENOMEM); 373 374 if (tsk_is_kthread(current)) { 375 scoped_with_init_fs() 376 bdev_file = bdev_file_open_by_path(devname, 377 BLK_OPEN_READ | BLK_OPEN_WRITE, 378 NULL, NULL); 379 } else { 380 bdev_file = bdev_file_open_by_path(devname, 381 BLK_OPEN_READ | BLK_OPEN_WRITE, NULL, NULL); 382 } 383 if (IS_ERR(bdev_file)) { 384 dprintk("failed to open device %s (%ld)\n", 385 devname, PTR_ERR(bdev_file)); 386 } else { 387 pr_info("pNFS: using block device %s\n", 388 file_bdev(bdev_file)->bd_disk->disk_name); 389 } 390 391 return bdev_file; 392 } 393 394 static int 395 bl_parse_scsi(struct nfs_server *server, struct pnfs_block_dev *d, 396 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 397 { 398 struct pnfs_block_volume *v = &volumes[idx]; 399 struct block_device *bdev; 400 const struct pr_ops *ops; 401 struct file *bdev_file; 402 int error; 403 404 if (!bl_validate_designator(v)) 405 return -EINVAL; 406 407 /* 408 * Try to open the RH/Fedora specific dm-mpath udev path first, as the 409 * wwn- links will only point to the first discovered SCSI device there. 410 * On other distributions like Debian, the default SCSI by-id path will 411 * point to the dm-multipath device if one exists. 412 */ 413 bdev_file = bl_open_path(v, "dm-uuid-mpath-0x"); 414 if (IS_ERR(bdev_file)) 415 bdev_file = bl_open_path(v, "wwn-0x"); 416 if (IS_ERR(bdev_file)) 417 bdev_file = bl_open_path(v, "nvme-eui."); 418 if (IS_ERR(bdev_file)) { 419 pr_warn("pNFS: no device found for volume %*phN\n", 420 v->scsi.designator_len, v->scsi.designator); 421 return PTR_ERR(bdev_file); 422 } 423 d->bdev_file = bdev_file; 424 bdev = file_bdev(bdev_file); 425 426 d->len = bdev_nr_bytes(bdev); 427 d->map = bl_map_simple; 428 d->pr_key = v->scsi.pr_key; 429 430 if (d->len == 0) { 431 error = -ENODEV; 432 goto out_blkdev_put; 433 } 434 435 ops = bdev->bd_disk->fops->pr_ops; 436 if (!ops) { 437 pr_err("pNFS: block device %s does not support reservations.", 438 bdev->bd_disk->disk_name); 439 error = -EINVAL; 440 goto out_blkdev_put; 441 } 442 443 return 0; 444 445 out_blkdev_put: 446 fput(d->bdev_file); 447 return error; 448 } 449 450 static int 451 bl_parse_slice(struct nfs_server *server, struct pnfs_block_dev *d, 452 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 453 { 454 struct pnfs_block_volume *v = &volumes[idx]; 455 int ret; 456 457 ret = bl_parse_deviceid(server, d, volumes, v->slice.volume, gfp_mask); 458 if (ret) 459 return ret; 460 461 d->disk_offset = v->slice.start; 462 d->len = v->slice.len; 463 return 0; 464 } 465 466 static int 467 bl_parse_concat(struct nfs_server *server, struct pnfs_block_dev *d, 468 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 469 { 470 struct pnfs_block_volume *v = &volumes[idx]; 471 u64 len = 0; 472 int ret, i; 473 474 d->children = kzalloc_objs(struct pnfs_block_dev, 475 v->concat.volumes_count, gfp_mask); 476 if (!d->children) 477 return -ENOMEM; 478 479 for (i = 0; i < v->concat.volumes_count; i++) { 480 ret = bl_parse_deviceid(server, &d->children[i], 481 volumes, v->concat.volumes[i], gfp_mask); 482 if (ret) 483 return ret; 484 485 d->nr_children++; 486 d->children[i].start += len; 487 len += d->children[i].len; 488 } 489 490 d->len = len; 491 d->map = bl_map_concat; 492 return 0; 493 } 494 495 static int 496 bl_parse_stripe(struct nfs_server *server, struct pnfs_block_dev *d, 497 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 498 { 499 struct pnfs_block_volume *v = &volumes[idx]; 500 u64 len = 0; 501 int ret, i; 502 503 d->children = kzalloc_objs(struct pnfs_block_dev, 504 v->stripe.volumes_count, gfp_mask); 505 if (!d->children) 506 return -ENOMEM; 507 508 for (i = 0; i < v->stripe.volumes_count; i++) { 509 ret = bl_parse_deviceid(server, &d->children[i], 510 volumes, v->stripe.volumes[i], gfp_mask); 511 if (ret) 512 return ret; 513 514 d->nr_children++; 515 len += d->children[i].len; 516 } 517 518 d->len = len; 519 d->chunk_size = v->stripe.chunk_size; 520 d->map = bl_map_stripe; 521 return 0; 522 } 523 524 static int 525 bl_parse_deviceid(struct nfs_server *server, struct pnfs_block_dev *d, 526 struct pnfs_block_volume *volumes, int idx, gfp_t gfp_mask) 527 { 528 d->type = volumes[idx].type; 529 530 switch (d->type) { 531 case PNFS_BLOCK_VOLUME_SIMPLE: 532 return bl_parse_simple(server, d, volumes, idx, gfp_mask); 533 case PNFS_BLOCK_VOLUME_SLICE: 534 return bl_parse_slice(server, d, volumes, idx, gfp_mask); 535 case PNFS_BLOCK_VOLUME_CONCAT: 536 return bl_parse_concat(server, d, volumes, idx, gfp_mask); 537 case PNFS_BLOCK_VOLUME_STRIPE: 538 return bl_parse_stripe(server, d, volumes, idx, gfp_mask); 539 case PNFS_BLOCK_VOLUME_SCSI: 540 return bl_parse_scsi(server, d, volumes, idx, gfp_mask); 541 default: 542 dprintk("unsupported volume type: %d\n", d->type); 543 return -EIO; 544 } 545 } 546 547 struct nfs4_deviceid_node * 548 bl_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev, 549 gfp_t gfp_mask) 550 { 551 struct nfs4_deviceid_node *node = NULL; 552 struct pnfs_block_volume *volumes; 553 struct pnfs_block_dev *top; 554 struct xdr_stream xdr; 555 struct xdr_buf buf; 556 struct folio *scratch; 557 int nr_volumes, ret, i; 558 __be32 *p; 559 560 scratch = folio_alloc(gfp_mask, 0); 561 if (!scratch) 562 goto out; 563 564 xdr_init_decode_pages(&xdr, &buf, pdev->pages, pdev->pglen); 565 xdr_set_scratch_folio(&xdr, scratch); 566 567 p = xdr_inline_decode(&xdr, sizeof(__be32)); 568 if (!p) 569 goto out_free_scratch; 570 nr_volumes = be32_to_cpup(p++); 571 572 volumes = kzalloc_objs(struct pnfs_block_volume, nr_volumes, gfp_mask); 573 if (!volumes) 574 goto out_free_scratch; 575 576 for (i = 0; i < nr_volumes; i++) { 577 ret = nfs4_block_decode_volume(&xdr, &volumes[i]); 578 if (ret < 0) 579 goto out_free_volumes; 580 } 581 582 top = kzalloc_obj(*top, gfp_mask); 583 if (!top) 584 goto out_free_volumes; 585 586 ret = bl_parse_deviceid(server, top, volumes, nr_volumes - 1, gfp_mask); 587 588 node = &top->node; 589 nfs4_init_deviceid_node(node, server, &pdev->dev_id); 590 if (ret) 591 nfs4_mark_deviceid_unavailable(node); 592 593 out_free_volumes: 594 kfree(volumes); 595 out_free_scratch: 596 folio_put(scratch); 597 out: 598 return node; 599 } 600