1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2017-2018 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 * Copyright (C) 2021, Alibaba Cloud 6 */ 7 #include "internal.h" 8 #include <linux/filelock.h> 9 #include <linux/sched/mm.h> 10 #include <trace/events/erofs.h> 11 12 void erofs_unmap_metabuf(struct erofs_buf *buf) 13 { 14 if (!buf->base) 15 return; 16 kunmap_local(buf->base); 17 buf->base = NULL; 18 } 19 20 void erofs_put_metabuf(struct erofs_buf *buf) 21 { 22 if (!buf->page) 23 return; 24 erofs_unmap_metabuf(buf); 25 folio_put(page_folio(buf->page)); 26 buf->page = NULL; 27 } 28 29 void *erofs_bread(struct erofs_buf *buf, erofs_off_t offset, bool need_kmap) 30 { 31 pgoff_t index = (buf->off + offset) >> PAGE_SHIFT; 32 struct folio *folio = NULL; 33 loff_t fpos; 34 int err; 35 36 /* 37 * Metadata access for file-backed mounts reuses page cache of backing 38 * fs inodes (only folio data will be needed) to prevent double caching. 39 * However, the data access range must be verified here in advance. 40 */ 41 if (buf->file) { 42 fpos = (loff_t)index << PAGE_SHIFT; 43 err = rw_verify_area(READ, buf->file, &fpos, PAGE_SIZE); 44 if (err < 0) 45 return ERR_PTR(err); 46 } 47 48 if (buf->page) { 49 folio = page_folio(buf->page); 50 if (folio_file_page(folio, index) != buf->page) 51 erofs_unmap_metabuf(buf); 52 } 53 if (!folio || !folio_contains(folio, index)) { 54 erofs_put_metabuf(buf); 55 folio = read_mapping_folio(buf->mapping, index, buf->file); 56 if (IS_ERR(folio)) 57 return folio; 58 } 59 buf->page = folio_file_page(folio, index); 60 if (!need_kmap) 61 return NULL; 62 if (!buf->base) 63 buf->base = kmap_local_page(buf->page); 64 return buf->base + (offset & ~PAGE_MASK); 65 } 66 67 int erofs_init_metabuf(struct erofs_buf *buf, struct super_block *sb, 68 bool in_metabox) 69 { 70 struct erofs_sb_info *sbi = EROFS_SB(sb); 71 72 buf->file = NULL; 73 if (in_metabox) { 74 if (unlikely(!sbi->metabox_inode)) 75 return -EFSCORRUPTED; 76 buf->mapping = sbi->metabox_inode->i_mapping; 77 return 0; 78 } 79 buf->off = sbi->dif0.fsoff; 80 if (erofs_is_fileio_mode(sbi)) { 81 buf->file = sbi->dif0.file; /* some fs like FUSE needs it */ 82 buf->mapping = buf->file->f_mapping; 83 } else 84 buf->mapping = sb->s_bdev->bd_mapping; 85 return 0; 86 } 87 88 void *erofs_read_metabuf(struct erofs_buf *buf, struct super_block *sb, 89 erofs_off_t offset, bool in_metabox) 90 { 91 int err; 92 93 err = erofs_init_metabuf(buf, sb, in_metabox); 94 if (err) 95 return ERR_PTR(err); 96 return erofs_bread(buf, offset, true); 97 } 98 99 static int erofs_map_chunks(struct inode *inode, struct erofs_map_blocks *map) 100 { 101 struct erofs_buf buf = __EROFS_BUF_INITIALIZER; 102 struct super_block *sb = inode->i_sb; 103 struct erofs_inode *vi = EROFS_I(inode); 104 struct erofs_inode_chunk_index *idx; 105 unsigned int unit = vi->chunkformat & EROFS_CHUNK_FORMAT_INDEXES ? 106 sizeof(*idx) : EROFS_BLOCK_MAP_ENTRY_SIZE; 107 erofs_blk_t addrmask = (vi->chunkformat & EROFS_CHUNK_FORMAT_48BIT) ? 108 BIT_ULL(48) - 1 : BIT_ULL(32) - 1; 109 u64 nr = map->m_la >> vi->chunkbits, chunksize = 1ULL << vi->chunkbits; 110 erofs_off_t pos = ALIGN(erofs_iloc(inode) + vi->inode_isize + 111 vi->xattr_isize, unit) + unit * nr; 112 /* m_llen will be clamped to EOF in the end */ 113 erofs_off_t endpos = round_up(pos + 1, sb->s_blocksize); 114 u64 last, addr; 115 116 idx = erofs_read_metabuf(&buf, sb, pos, erofs_inode_in_metabox(inode)); 117 if (IS_ERR(idx)) 118 return PTR_ERR(idx); 119 120 map->m_la = nr << vi->chunkbits; 121 map->m_llen = 0; 122 nr = 0; 123 do { 124 if (unit == EROFS_BLOCK_MAP_ENTRY_SIZE) { 125 addr = le32_to_cpu(((__le32 *)idx)[nr]); 126 if (addr == (u32)EROFS_NULL_ADDR) 127 addr = EROFS_NULL_ADDR; 128 } else { 129 addr = (((u64)le16_to_cpu(idx[nr].startblk_hi) << 32) | 130 le32_to_cpu(idx[nr].startblk_lo)) & addrmask; 131 if (addr ^ (EROFS_NULL_ADDR & addrmask)) 132 addr |= (u64)(le16_to_cpu(idx[nr].device_id) & 133 EROFS_SB(sb)->device_id_mask) << 48; 134 else 135 addr = EROFS_NULL_ADDR; 136 } 137 if (!nr) { 138 last = addr; 139 continue; 140 } 141 /* expand and account the prior chunk here */ 142 map->m_llen += chunksize; 143 if (last != EROFS_NULL_ADDR) 144 last += erofs_blknr(sb, chunksize); 145 } while (addr == last && pos + (++nr) * unit < endpos); 146 147 if (last != EROFS_NULL_ADDR) { 148 map->m_pa = erofs_pos(sb, last & addrmask) - map->m_llen; 149 map->m_deviceid = last >> 48; 150 map->m_flags = EROFS_MAP_MAPPED; 151 } 152 if (addr == last) 153 map->m_llen += chunksize; 154 map->m_llen = min_t(erofs_off_t, map->m_llen, 155 round_up(inode->i_size - map->m_la, sb->s_blocksize)); 156 erofs_put_metabuf(&buf); 157 return 0; 158 } 159 160 int erofs_map_blocks(struct inode *inode, struct erofs_map_blocks *map) 161 { 162 struct super_block *sb = inode->i_sb; 163 struct erofs_inode *vi = EROFS_I(inode); 164 bool tailinline = (vi->datalayout == EROFS_INODE_FLAT_INLINE); 165 erofs_off_t pos; 166 int err = 0; 167 168 trace_erofs_map_blocks_enter(inode, map, 0); 169 map->m_deviceid = 0; 170 map->m_flags = 0; 171 if (map->m_la >= inode->i_size) 172 goto out; 173 if (vi->datalayout == EROFS_INODE_CHUNK_BASED) { 174 err = erofs_map_chunks(inode, map); 175 } else if (tailinline || vi->startblk != EROFS_NULL_ADDR) { 176 pos = erofs_pos(sb, erofs_iblks(inode) - tailinline); 177 map->m_flags = EROFS_MAP_MAPPED; 178 if (map->m_la < pos) { 179 map->m_pa = erofs_pos(sb, vi->startblk) + map->m_la; 180 map->m_llen = pos - map->m_la; 181 } else { 182 map->m_pa = erofs_iloc(inode) + vi->inode_isize + 183 vi->xattr_isize + erofs_blkoff(sb, map->m_la); 184 map->m_llen = inode->i_size - map->m_la; 185 map->m_flags |= EROFS_MAP_META; 186 if (erofs_blkoff(sb, map->m_pa) + map->m_llen > 187 sb->s_blocksize) { 188 erofs_err(sb, "inline data across blocks @ nid %llu", vi->nid); 189 return -EFSCORRUPTED; 190 } 191 } 192 } 193 out: 194 map->m_plen = err ? 0 : map->m_llen; 195 trace_erofs_map_blocks_exit(inode, map, 0, err); 196 return err; 197 } 198 199 static void erofs_fill_from_devinfo(struct erofs_map_dev *map, 200 struct super_block *sb, struct erofs_device_info *dif) 201 { 202 map->m_sb = sb; 203 map->m_dif = dif; 204 map->m_bdev = NULL; 205 if (dif->file && S_ISBLK(file_inode(dif->file)->i_mode)) 206 map->m_bdev = file_bdev(dif->file); 207 } 208 209 int erofs_map_dev(struct super_block *sb, struct erofs_map_dev *map) 210 { 211 struct erofs_dev_context *devs = EROFS_SB(sb)->devs; 212 struct erofs_device_info *dif; 213 erofs_off_t startoff; 214 int id; 215 216 erofs_fill_from_devinfo(map, sb, &EROFS_SB(sb)->dif0); 217 map->m_bdev = sb->s_bdev; /* use s_bdev for the primary device */ 218 if (map->m_deviceid) { 219 down_read(&devs->rwsem); 220 dif = idr_find(&devs->tree, map->m_deviceid - 1); 221 if (!dif) { 222 up_read(&devs->rwsem); 223 return -ENODEV; 224 } 225 if (devs->flatdev) { 226 map->m_pa += erofs_pos(sb, dif->uniaddr); 227 up_read(&devs->rwsem); 228 return 0; 229 } 230 erofs_fill_from_devinfo(map, sb, dif); 231 up_read(&devs->rwsem); 232 } else if (devs->extra_devices && !devs->flatdev) { 233 down_read(&devs->rwsem); 234 idr_for_each_entry(&devs->tree, dif, id) { 235 if (!dif->uniaddr) 236 continue; 237 238 startoff = erofs_pos(sb, dif->uniaddr); 239 if (map->m_pa >= startoff && 240 map->m_pa < startoff + erofs_pos(sb, dif->blocks)) { 241 map->m_pa -= startoff; 242 erofs_fill_from_devinfo(map, sb, dif); 243 break; 244 } 245 } 246 up_read(&devs->rwsem); 247 } 248 return 0; 249 } 250 251 /* 252 * bit 30: I/O error occurred on this folio 253 * bit 29: CPU has dirty data in D-cache (needs aliasing handling); 254 * bit 0 - 29: remaining parts to complete this folio 255 */ 256 #define EROFS_ONLINEFOLIO_EIO 30 257 #define EROFS_ONLINEFOLIO_DIRTY 29 258 259 void erofs_onlinefolio_init(struct folio *folio) 260 { 261 union { 262 atomic_t o; 263 void *v; 264 } u = { .o = ATOMIC_INIT(1) }; 265 266 folio->private = u.v; /* valid only if file-backed folio is locked */ 267 } 268 269 void erofs_onlinefolio_split(struct folio *folio) 270 { 271 atomic_inc((atomic_t *)&folio->private); 272 } 273 274 void erofs_onlinefolio_end(struct folio *folio, int err, bool dirty) 275 { 276 int orig, v; 277 278 do { 279 orig = atomic_read((atomic_t *)&folio->private); 280 DBG_BUGON(orig <= 0); 281 v = dirty << EROFS_ONLINEFOLIO_DIRTY; 282 v |= (orig - 1) | (!!err << EROFS_ONLINEFOLIO_EIO); 283 } while (atomic_cmpxchg((atomic_t *)&folio->private, orig, v) != orig); 284 285 if (v & (BIT(EROFS_ONLINEFOLIO_DIRTY) - 1)) 286 return; 287 folio->private = 0; 288 if (v & BIT(EROFS_ONLINEFOLIO_DIRTY)) 289 flush_dcache_folio(folio); 290 folio_end_read(folio, !(v & BIT(EROFS_ONLINEFOLIO_EIO))); 291 } 292 293 struct erofs_iomap_iter_ctx { 294 struct page *page; 295 void *base; 296 struct inode *realinode; 297 }; 298 299 static int erofs_iomap_begin(struct inode *inode, loff_t offset, loff_t length, 300 unsigned int flags, struct iomap *iomap, struct iomap *srcmap) 301 { 302 struct iomap_iter *iter = container_of(iomap, struct iomap_iter, iomap); 303 struct erofs_iomap_iter_ctx *ctx = iter->private; 304 struct inode *realinode = ctx ? ctx->realinode : inode; 305 struct super_block *sb = realinode->i_sb; 306 struct erofs_map_blocks map; 307 struct erofs_map_dev mdev; 308 int ret; 309 310 map.m_la = offset; 311 map.m_llen = length; 312 ret = erofs_map_blocks(realinode, &map); 313 if (ret < 0) 314 return ret; 315 316 iomap->offset = map.m_la; 317 iomap->length = map.m_llen; 318 iomap->flags = 0; 319 iomap->addr = IOMAP_NULL_ADDR; 320 if (!(map.m_flags & EROFS_MAP_MAPPED)) { 321 iomap->type = IOMAP_HOLE; 322 return 0; 323 } 324 325 if (!(map.m_flags & EROFS_MAP_META) || !erofs_inode_in_metabox(realinode)) { 326 mdev = (struct erofs_map_dev) { 327 .m_deviceid = map.m_deviceid, 328 .m_pa = map.m_pa, 329 }; 330 ret = erofs_map_dev(sb, &mdev); 331 if (ret) 332 return ret; 333 334 if (flags & IOMAP_DAX) 335 iomap->dax_dev = mdev.m_dif->dax_dev; 336 else 337 iomap->bdev = mdev.m_bdev; 338 iomap->addr = mdev.m_dif->fsoff + mdev.m_pa; 339 if (flags & IOMAP_DAX) 340 iomap->addr += mdev.m_dif->dax_part_off; 341 } 342 343 if (map.m_flags & EROFS_MAP_META) { 344 iomap->type = IOMAP_INLINE; 345 /* read context should read the inlined data */ 346 if (ctx) { 347 struct erofs_buf buf = __EROFS_BUF_INITIALIZER; 348 void *ptr; 349 350 ptr = erofs_read_metabuf(&buf, sb, map.m_pa, 351 erofs_inode_in_metabox(realinode)); 352 if (IS_ERR(ptr)) 353 return PTR_ERR(ptr); 354 iomap->inline_data = ptr; 355 ctx->page = buf.page; 356 ctx->base = buf.base; 357 } 358 } else { 359 iomap->type = IOMAP_MAPPED; 360 } 361 return 0; 362 } 363 364 static int erofs_iomap_end(struct inode *inode, loff_t pos, loff_t length, 365 ssize_t written, unsigned int flags, struct iomap *iomap) 366 { 367 struct iomap_iter *iter = container_of(iomap, struct iomap_iter, iomap); 368 struct erofs_iomap_iter_ctx *ctx = iter->private; 369 370 if (ctx && ctx->base) { 371 struct erofs_buf buf = { 372 .page = ctx->page, 373 .base = ctx->base, 374 }; 375 376 DBG_BUGON(iomap->type != IOMAP_INLINE); 377 erofs_put_metabuf(&buf); 378 ctx->base = NULL; 379 } 380 return written; 381 } 382 383 static const struct iomap_ops erofs_iomap_ops = { 384 .iomap_begin = erofs_iomap_begin, 385 .iomap_end = erofs_iomap_end, 386 }; 387 388 int erofs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, 389 u64 start, u64 len) 390 { 391 if (erofs_inode_is_data_compressed(EROFS_I(inode)->datalayout)) { 392 if (!IS_ENABLED(CONFIG_EROFS_FS_ZIP)) 393 return -EOPNOTSUPP; 394 return iomap_fiemap(inode, fieinfo, start, len, 395 &z_erofs_iomap_report_ops); 396 } 397 return iomap_fiemap(inode, fieinfo, start, len, &erofs_iomap_ops); 398 } 399 400 /* 401 * since we dont have write or truncate flows, so no inode 402 * locking needs to be held at the moment. 403 */ 404 static int erofs_read_folio(struct file *file, struct folio *folio) 405 { 406 struct iomap_read_folio_ctx read_ctx = { 407 .ops = &iomap_bio_read_ops, 408 .cur_folio = folio, 409 }; 410 bool need_iput; 411 struct erofs_iomap_iter_ctx iter_ctx = { 412 .realinode = erofs_real_inode(folio_inode(folio), &need_iput), 413 }; 414 415 trace_erofs_read_folio(iter_ctx.realinode, folio, true); 416 iomap_read_folio(&erofs_iomap_ops, &read_ctx, &iter_ctx); 417 if (need_iput) 418 iput(iter_ctx.realinode); 419 return 0; 420 } 421 422 static void erofs_readahead(struct readahead_control *rac) 423 { 424 struct iomap_read_folio_ctx read_ctx = { 425 .ops = &iomap_bio_read_ops, 426 .rac = rac, 427 }; 428 bool need_iput; 429 struct erofs_iomap_iter_ctx iter_ctx = { 430 .realinode = erofs_real_inode(rac->mapping->host, &need_iput), 431 }; 432 433 trace_erofs_readahead(iter_ctx.realinode, readahead_index(rac), 434 readahead_count(rac), true); 435 iomap_readahead(&erofs_iomap_ops, &read_ctx, &iter_ctx); 436 if (need_iput) 437 iput(iter_ctx.realinode); 438 } 439 440 static sector_t erofs_bmap(struct address_space *mapping, sector_t block) 441 { 442 return iomap_bmap(mapping, block, &erofs_iomap_ops); 443 } 444 445 static ssize_t erofs_file_read_iter(struct kiocb *iocb, struct iov_iter *to) 446 { 447 struct inode *inode = file_inode(iocb->ki_filp); 448 449 /* no need taking (shared) inode lock since it's a ro filesystem */ 450 if (!iov_iter_count(to)) 451 return 0; 452 453 if (IS_ENABLED(CONFIG_FS_DAX) && IS_DAX(inode)) 454 return dax_iomap_rw(iocb, to, &erofs_iomap_ops); 455 456 if ((iocb->ki_flags & IOCB_DIRECT) && inode->i_sb->s_bdev) { 457 struct erofs_iomap_iter_ctx iter_ctx = { 458 .realinode = inode, 459 }; 460 461 return iomap_dio_rw(iocb, to, &erofs_iomap_ops, 462 NULL, 0, &iter_ctx, 0); 463 } 464 return filemap_read(iocb, to, 0); 465 } 466 467 /* for uncompressed (aligned) files and raw access for other files */ 468 const struct address_space_operations erofs_aops = { 469 .read_folio = erofs_read_folio, 470 .readahead = erofs_readahead, 471 .bmap = erofs_bmap, 472 .direct_IO = noop_direct_IO, 473 .release_folio = iomap_release_folio, 474 .invalidate_folio = iomap_invalidate_folio, 475 }; 476 477 #ifdef CONFIG_FS_DAX 478 static vm_fault_t erofs_dax_huge_fault(struct vm_fault *vmf, 479 unsigned int order) 480 { 481 return dax_iomap_fault(vmf, order, NULL, NULL, &erofs_iomap_ops); 482 } 483 484 static vm_fault_t erofs_dax_fault(struct vm_fault *vmf) 485 { 486 return erofs_dax_huge_fault(vmf, 0); 487 } 488 489 static const struct vm_operations_struct erofs_dax_vm_ops = { 490 .fault = erofs_dax_fault, 491 .huge_fault = erofs_dax_huge_fault, 492 }; 493 494 static int erofs_file_mmap_prepare(struct vm_area_desc *desc) 495 { 496 if (!IS_DAX(file_inode(desc->file))) 497 return generic_file_readonly_mmap_prepare(desc); 498 499 if (vma_desc_test_all(desc, VMA_SHARED_BIT, VMA_MAYWRITE_BIT)) 500 return -EINVAL; 501 502 desc->vm_ops = &erofs_dax_vm_ops; 503 vma_desc_set_flags(desc, VMA_HUGEPAGE_BIT); 504 return 0; 505 } 506 #else 507 #define erofs_file_mmap_prepare generic_file_readonly_mmap_prepare 508 #endif 509 510 static loff_t erofs_file_llseek(struct file *file, loff_t offset, int whence) 511 { 512 struct inode *inode = file->f_mapping->host; 513 const struct iomap_ops *ops = &erofs_iomap_ops; 514 515 if (erofs_inode_is_data_compressed(EROFS_I(inode)->datalayout)) { 516 if (!IS_ENABLED(CONFIG_EROFS_FS_ZIP)) 517 return generic_file_llseek(file, offset, whence); 518 ops = &z_erofs_iomap_report_ops; 519 } 520 521 if (whence == SEEK_HOLE) 522 offset = iomap_seek_hole(inode, offset, ops); 523 else if (whence == SEEK_DATA) 524 offset = iomap_seek_data(inode, offset, ops); 525 else 526 return generic_file_llseek(file, offset, whence); 527 528 if (offset < 0) 529 return offset; 530 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes); 531 } 532 533 const struct file_operations erofs_file_fops = { 534 .llseek = erofs_file_llseek, 535 .read_iter = erofs_file_read_iter, 536 .unlocked_ioctl = erofs_ioctl, 537 #ifdef CONFIG_COMPAT 538 .compat_ioctl = erofs_compat_ioctl, 539 #endif 540 .mmap_prepare = erofs_file_mmap_prepare, 541 .get_unmapped_area = thp_get_unmapped_area, 542 .splice_read = filemap_splice_read, 543 .setlease = generic_setlease, 544 }; 545