fs/ext4/readpage.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/fs/ext4/readpage.c
   4  *
   5  * Copyright (C) 2002, Linus Torvalds.
   6  * Copyright (C) 2015, Google, Inc.
   7  *
   8  * This was originally taken from fs/mpage.c
   9  *
  10  * The ext4_mpage_readpages() function here is intended to
  11  * replace mpage_readahead() in the general case, not just for
  12  * encrypted files.  It has some limitations (see below), where it
  13  * will fall back to read_block_full_page(), but these limitations
  14  * should only be hit when page_size != block_size.
  15  *
  16  * This will allow us to attach a callback function to support ext4
  17  * encryption.
  18  *
  19  * If anything unusual happens, such as:
  20  *
  21  * - encountering a page which has buffers
  22  * - encountering a page which has a non-hole after a hole
  23  * - encountering a page with non-contiguous blocks
  24  *
  25  * then this code just gives up and calls the buffer_head-based read function.
  26  * It does handle a page which has holes at the end - that is a common case:
  27  * the end-of-file on blocksize < PAGE_SIZE setups.
  28  *
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/export.h>
  33 #include <linux/mm.h>
  34 #include <linux/kdev_t.h>
  35 #include <linux/gfp.h>
  36 #include <linux/bio.h>
  37 #include <linux/fs.h>
  38 #include <linux/buffer_head.h>
  39 #include <linux/blkdev.h>
  40 #include <linux/highmem.h>
  41 #include <linux/prefetch.h>
  42 #include <linux/mpage.h>
  43 #include <linux/writeback.h>
  44 #include <linux/backing-dev.h>
  45 #include <linux/pagevec.h>
  46
  47 #include "ext4.h"
  48
  49 #define NUM_PREALLOC_POST_READ_CTXS     128
  50
  51 static struct kmem_cache *bio_post_read_ctx_cache;
  52 static mempool_t *bio_post_read_ctx_pool;
  53
  54 /* postprocessing steps for read bios */
  55 enum bio_post_read_step {
  56         STEP_INITIAL = 0,
  57         STEP_DECRYPT,
  58         STEP_VERITY,
  59         STEP_MAX,
  60 };
  61
  62 struct bio_post_read_ctx {
  63         struct bio *bio;
  64         struct work_struct work;
  65         unsigned int cur_step;
  66         unsigned int enabled_steps;
  67 };
  68
  69 static void __read_end_io(struct bio *bio)
  70 {
  71         struct page *page;
  72         struct bio_vec *bv;
  73         struct bvec_iter_all iter_all;
  74
  75         bio_for_each_segment_all(bv, bio, iter_all) {
  76                 page = bv->bv_page;
  77
  78                 /* PG_error was set if verity failed. */
  79                 if (bio->bi_status || PageError(page)) {
  80                         ClearPageUptodate(page);
  81                         /* will re-read again later */
  82                         ClearPageError(page);
  83                 } else {
  84                         SetPageUptodate(page);
  85                 }
  86                 unlock_page(page);
  87         }
  88         if (bio->bi_private)
  89                 mempool_free(bio->bi_private, bio_post_read_ctx_pool);
  90         bio_put(bio);
  91 }
  92
  93 static void bio_post_read_processing(struct bio_post_read_ctx *ctx);
  94
  95 static void decrypt_work(struct work_struct *work)
  96 {
  97         struct bio_post_read_ctx *ctx =
  98                 container_of(work, struct bio_post_read_ctx, work);
  99         struct bio *bio = ctx->bio;
 100
 101         if (fscrypt_decrypt_bio(bio))
 102                 bio_post_read_processing(ctx);
 103         else
 104                 __read_end_io(bio);
 105 }
 106
 107 static void verity_work(struct work_struct *work)
 108 {
 109         struct bio_post_read_ctx *ctx =
 110                 container_of(work, struct bio_post_read_ctx, work);
 111         struct bio *bio = ctx->bio;
 112
 113         /*
 114          * fsverity_verify_bio() may call readahead() again, and although verity
 115          * will be disabled for that, decryption may still be needed, causing
 116          * another bio_post_read_ctx to be allocated.  So to guarantee that
 117          * mempool_alloc() never deadlocks we must free the current ctx first.
 118          * This is safe because verity is the last post-read step.
 119          */
 120         BUILD_BUG_ON(STEP_VERITY + 1 != STEP_MAX);
 121         mempool_free(ctx, bio_post_read_ctx_pool);
 122         bio->bi_private = NULL;
 123
 124         fsverity_verify_bio(bio);
 125
 126         __read_end_io(bio);
 127 }
 128
 129 static void bio_post_read_processing(struct bio_post_read_ctx *ctx)
 130 {
 131         /*
 132          * We use different work queues for decryption and for verity because
 133          * verity may require reading metadata pages that need decryption, and
 134          * we shouldn't recurse to the same workqueue.
 135          */
 136         switch (++ctx->cur_step) {
 137         case STEP_DECRYPT:
 138                 if (ctx->enabled_steps & (1 << STEP_DECRYPT)) {
 139                         INIT_WORK(&ctx->work, decrypt_work);
 140                         fscrypt_enqueue_decrypt_work(&ctx->work);
 141                         return;
 142                 }
 143                 ctx->cur_step++;
 144                 fallthrough;
 145         case STEP_VERITY:
 146                 if (ctx->enabled_steps & (1 << STEP_VERITY)) {
 147                         INIT_WORK(&ctx->work, verity_work);
 148                         fsverity_enqueue_verify_work(&ctx->work);
 149                         return;
 150                 }
 151                 ctx->cur_step++;
 152                 fallthrough;
 153         default:
 154                 __read_end_io(ctx->bio);
 155         }
 156 }
 157
 158 static bool bio_post_read_required(struct bio *bio)
 159 {
 160         return bio->bi_private && !bio->bi_status;
 161 }
 162
 163 /*
 164  * I/O completion handler for multipage BIOs.
 165  *
 166  * The mpage code never puts partial pages into a BIO (except for end-of-file).
 167  * If a page does not map to a contiguous run of blocks then it simply falls
 168  * back to block_read_full_folio().
 169  *
 170  * Why is this?  If a page's completion depends on a number of different BIOs
 171  * which can complete in any order (or at the same time) then determining the
 172  * status of that page is hard.  See end_buffer_async_read() for the details.
 173  * There is no point in duplicating all that complexity.
 174  */
 175 static void mpage_end_io(struct bio *bio)
 176 {
 177         if (bio_post_read_required(bio)) {
 178                 struct bio_post_read_ctx *ctx = bio->bi_private;
 179
 180                 ctx->cur_step = STEP_INITIAL;
 181                 bio_post_read_processing(ctx);
 182                 return;
 183         }
 184         __read_end_io(bio);
 185 }
 186
 187 static inline bool ext4_need_verity(const struct inode *inode, pgoff_t idx)
 188 {
 189         return fsverity_active(inode) &&
 190                idx < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
 191 }
 192
 193 static void ext4_set_bio_post_read_ctx(struct bio *bio,
 194                                        const struct inode *inode,
 195                                        pgoff_t first_idx)
 196 {
 197         unsigned int post_read_steps = 0;
 198
 199         if (fscrypt_inode_uses_fs_layer_crypto(inode))
 200                 post_read_steps |= 1 << STEP_DECRYPT;
 201
 202         if (ext4_need_verity(inode, first_idx))
 203                 post_read_steps |= 1 << STEP_VERITY;
 204
 205         if (post_read_steps) {
 206                 /* Due to the mempool, this never fails. */
 207                 struct bio_post_read_ctx *ctx =
 208                         mempool_alloc(bio_post_read_ctx_pool, GFP_NOFS);
 209
 210                 ctx->bio = bio;
 211                 ctx->enabled_steps = post_read_steps;
 212                 bio->bi_private = ctx;
 213         }
 214 }
 215
 216 static inline loff_t ext4_readpage_limit(struct inode *inode)
 217 {
 218         if (IS_ENABLED(CONFIG_FS_VERITY) &&
 219             (IS_VERITY(inode) || ext4_verity_in_progress(inode)))
 220                 return inode->i_sb->s_maxbytes;
 221
 222         return i_size_read(inode);
 223 }
 224
 225 int ext4_mpage_readpages(struct inode *inode,
 226                 struct readahead_control *rac, struct page *page)
 227 {
 228         struct bio *bio = NULL;
 229         sector_t last_block_in_bio = 0;
 230
 231         const unsigned blkbits = inode->i_blkbits;
 232         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 233         const unsigned blocksize = 1 << blkbits;
 234         sector_t next_block;
 235         sector_t block_in_file;
 236         sector_t last_block;
 237         sector_t last_block_in_file;
 238         sector_t blocks[MAX_BUF_PER_PAGE];
 239         unsigned page_block;
 240         struct block_device *bdev = inode->i_sb->s_bdev;
 241         int length;
 242         unsigned relative_block = 0;
 243         struct ext4_map_blocks map;
 244         unsigned int nr_pages = rac ? readahead_count(rac) : 1;
 245
 246         map.m_pblk = 0;
 247         map.m_lblk = 0;
 248         map.m_len = 0;
 249         map.m_flags = 0;
 250
 251         for (; nr_pages; nr_pages--) {
 252                 int fully_mapped = 1;
 253                 unsigned first_hole = blocks_per_page;
 254
 255                 if (rac) {
 256                         page = readahead_page(rac);
 257                         prefetchw(&page->flags);
 258                 }
 259
 260                 if (page_has_buffers(page))
 261                         goto confused;
 262
 263                 block_in_file = next_block =
 264                         (sector_t)page->index << (PAGE_SHIFT - blkbits);
 265                 last_block = block_in_file + nr_pages * blocks_per_page;
 266                 last_block_in_file = (ext4_readpage_limit(inode) +
 267                                       blocksize - 1) >> blkbits;
 268                 if (last_block > last_block_in_file)
 269                         last_block = last_block_in_file;
 270                 page_block = 0;
 271
 272                 /*
 273                  * Map blocks using the previous result first.
 274                  */
 275                 if ((map.m_flags & EXT4_MAP_MAPPED) &&
 276                     block_in_file > map.m_lblk &&
 277                     block_in_file < (map.m_lblk + map.m_len)) {
 278                         unsigned map_offset = block_in_file - map.m_lblk;
 279                         unsigned last = map.m_len - map_offset;
 280
 281                         for (relative_block = 0; ; relative_block++) {
 282                                 if (relative_block == last) {
 283                                         /* needed? */
 284                                         map.m_flags &= ~EXT4_MAP_MAPPED;
 285                                         break;
 286                                 }
 287                                 if (page_block == blocks_per_page)
 288                                         break;
 289                                 blocks[page_block] = map.m_pblk + map_offset +
 290                                         relative_block;
 291                                 page_block++;
 292                                 block_in_file++;
 293                         }
 294                 }
 295
 296                 /*
 297                  * Then do more ext4_map_blocks() calls until we are
 298                  * done with this page.
 299                  */
 300                 while (page_block < blocks_per_page) {
 301                         if (block_in_file < last_block) {
 302                                 map.m_lblk = block_in_file;
 303                                 map.m_len = last_block - block_in_file;
 304
 305                                 if (ext4_map_blocks(NULL, inode, &map, 0) < 0) {
 306                                 set_error_page:
 307                                         SetPageError(page);
 308                                         zero_user_segment(page, 0,
 309                                                           PAGE_SIZE);
 310                                         unlock_page(page);
 311                                         goto next_page;
 312                                 }
 313                         }
 314                         if ((map.m_flags & EXT4_MAP_MAPPED) == 0) {
 315                                 fully_mapped = 0;
 316                                 if (first_hole == blocks_per_page)
 317                                         first_hole = page_block;
 318                                 page_block++;
 319                                 block_in_file++;
 320                                 continue;
 321                         }
 322                         if (first_hole != blocks_per_page)
 323                                 goto confused;          /* hole -> non-hole */
 324
 325                         /* Contiguous blocks? */
 326                         if (page_block && blocks[page_block-1] != map.m_pblk-1)
 327                                 goto confused;
 328                         for (relative_block = 0; ; relative_block++) {
 329                                 if (relative_block == map.m_len) {
 330                                         /* needed? */
 331                                         map.m_flags &= ~EXT4_MAP_MAPPED;
 332                                         break;
 333                                 } else if (page_block == blocks_per_page)
 334                                         break;
 335                                 blocks[page_block] = map.m_pblk+relative_block;
 336                                 page_block++;
 337                                 block_in_file++;
 338                         }
 339                 }
 340                 if (first_hole != blocks_per_page) {
 341                         zero_user_segment(page, first_hole << blkbits,
 342                                           PAGE_SIZE);
 343                         if (first_hole == 0) {
 344                                 if (ext4_need_verity(inode, page->index) &&
 345                                     !fsverity_verify_page(page))
 346                                         goto set_error_page;
 347                                 SetPageUptodate(page);
 348                                 unlock_page(page);
 349                                 goto next_page;
 350                         }
 351                 } else if (fully_mapped) {
 352                         SetPageMappedToDisk(page);
 353                 }
 354
 355                 /*
 356                  * This page will go to BIO.  Do we need to send this
 357                  * BIO off first?
 358                  */
 359                 if (bio && (last_block_in_bio != blocks[0] - 1 ||
 360                             !fscrypt_mergeable_bio(bio, inode, next_block))) {
 361                 submit_and_realloc:
 362                         submit_bio(bio);
 363                         bio = NULL;
 364                 }
 365                 if (bio == NULL) {
 366                         /*
 367                          * bio_alloc will _always_ be able to allocate a bio if
 368                          * __GFP_DIRECT_RECLAIM is set, see bio_alloc_bioset().
 369                          */
 370                         bio = bio_alloc(bdev, bio_max_segs(nr_pages),
 371                                         REQ_OP_READ, GFP_KERNEL);
 372                         fscrypt_set_bio_crypt_ctx(bio, inode, next_block,
 373                                                   GFP_KERNEL);
 374                         ext4_set_bio_post_read_ctx(bio, inode, page->index);
 375                         bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
 376                         bio->bi_end_io = mpage_end_io;
 377                         if (rac)
 378                                 bio->bi_opf |= REQ_RAHEAD;
 379                 }
 380
 381                 length = first_hole << blkbits;
 382                 if (bio_add_page(bio, page, length, 0) < length)
 383                         goto submit_and_realloc;
 384
 385                 if (((map.m_flags & EXT4_MAP_BOUNDARY) &&
 386                      (relative_block == map.m_len)) ||
 387                     (first_hole != blocks_per_page)) {
 388                         submit_bio(bio);
 389                         bio = NULL;
 390                 } else
 391                         last_block_in_bio = blocks[blocks_per_page - 1];
 392                 goto next_page;
 393         confused:
 394                 if (bio) {
 395                         submit_bio(bio);
 396                         bio = NULL;
 397                 }
 398                 if (!PageUptodate(page))
 399                         block_read_full_folio(page_folio(page), ext4_get_block);
 400                 else
 401                         unlock_page(page);
 402         next_page:
 403                 if (rac)
 404                         put_page(page);
 405         }
 406         if (bio)
 407                 submit_bio(bio);
 408         return 0;
 409 }
 410
 411 int __init ext4_init_post_read_processing(void)
 412 {
 413         bio_post_read_ctx_cache =
 414                 kmem_cache_create("ext4_bio_post_read_ctx",
 415                                   sizeof(struct bio_post_read_ctx), 0, 0, NULL);
 416         if (!bio_post_read_ctx_cache)
 417                 goto fail;
 418         bio_post_read_ctx_pool =
 419                 mempool_create_slab_pool(NUM_PREALLOC_POST_READ_CTXS,
 420                                          bio_post_read_ctx_cache);
 421         if (!bio_post_read_ctx_pool)
 422                 goto fail_free_cache;
 423         return 0;
 424
 425 fail_free_cache:
 426         kmem_cache_destroy(bio_post_read_ctx_cache);
 427 fail:
 428         return -ENOMEM;
 429 }
 430
 431 void ext4_exit_post_read_processing(void)
 432 {
 433         mempool_destroy(bio_post_read_ctx_pool);
 434         kmem_cache_destroy(bio_post_read_ctx_cache);
 435 }