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