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 intent is the ext4_mpage_readpages() function here is intended
  11  * to replace mpage_readpages() 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 #include <linux/cleancache.h>
  47
  48 #include "ext4.h"
  49
  50 #define NUM_PREALLOC_POST_READ_CTXS     128
  51
  52 static struct kmem_cache *bio_post_read_ctx_cache;
  53 static mempool_t *bio_post_read_ctx_pool;
  54
  55 /* postprocessing steps for read bios */
  56 enum bio_post_read_step {
  57         STEP_INITIAL = 0,
  58         STEP_DECRYPT,
  59         STEP_VERITY,
  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 any post_read step 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
 100         fscrypt_decrypt_bio(ctx->bio);
 101
 102         bio_post_read_processing(ctx);
 103 }
 104
 105 static void verity_work(struct work_struct *work)
 106 {
 107         struct bio_post_read_ctx *ctx =
 108                 container_of(work, struct bio_post_read_ctx, work);
 109
 110         fsverity_verify_bio(ctx->bio);
 111
 112         bio_post_read_processing(ctx);
 113 }
 114
 115 static void bio_post_read_processing(struct bio_post_read_ctx *ctx)
 116 {
 117         /*
 118          * We use different work queues for decryption and for verity because
 119          * verity may require reading metadata pages that need decryption, and
 120          * we shouldn't recurse to the same workqueue.
 121          */
 122         switch (++ctx->cur_step) {
 123         case STEP_DECRYPT:
 124                 if (ctx->enabled_steps & (1 << STEP_DECRYPT)) {
 125                         INIT_WORK(&ctx->work, decrypt_work);
 126                         fscrypt_enqueue_decrypt_work(&ctx->work);
 127                         return;
 128                 }
 129                 ctx->cur_step++;
 130                 /* fall-through */
 131         case STEP_VERITY:
 132                 if (ctx->enabled_steps & (1 << STEP_VERITY)) {
 133                         INIT_WORK(&ctx->work, verity_work);
 134                         fsverity_enqueue_verify_work(&ctx->work);
 135                         return;
 136                 }
 137                 ctx->cur_step++;
 138                 /* fall-through */
 139         default:
 140                 __read_end_io(ctx->bio);
 141         }
 142 }
 143
 144 static bool bio_post_read_required(struct bio *bio)
 145 {
 146         return bio->bi_private && !bio->bi_status;
 147 }
 148
 149 /*
 150  * I/O completion handler for multipage BIOs.
 151  *
 152  * The mpage code never puts partial pages into a BIO (except for end-of-file).
 153  * If a page does not map to a contiguous run of blocks then it simply falls
 154  * back to block_read_full_page().
 155  *
 156  * Why is this?  If a page's completion depends on a number of different BIOs
 157  * which can complete in any order (or at the same time) then determining the
 158  * status of that page is hard.  See end_buffer_async_read() for the details.
 159  * There is no point in duplicating all that complexity.
 160  */
 161 static void mpage_end_io(struct bio *bio)
 162 {
 163         if (bio_post_read_required(bio)) {
 164                 struct bio_post_read_ctx *ctx = bio->bi_private;
 165
 166                 ctx->cur_step = STEP_INITIAL;
 167                 bio_post_read_processing(ctx);
 168                 return;
 169         }
 170         __read_end_io(bio);
 171 }
 172
 173 static inline bool ext4_need_verity(const struct inode *inode, pgoff_t idx)
 174 {
 175         return fsverity_active(inode) &&
 176                idx < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
 177 }
 178
 179 static struct bio_post_read_ctx *get_bio_post_read_ctx(struct inode *inode,
 180                                                        struct bio *bio,
 181                                                        pgoff_t first_idx)
 182 {
 183         unsigned int post_read_steps = 0;
 184         struct bio_post_read_ctx *ctx = NULL;
 185
 186         if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode))
 187                 post_read_steps |= 1 << STEP_DECRYPT;
 188
 189         if (ext4_need_verity(inode, first_idx))
 190                 post_read_steps |= 1 << STEP_VERITY;
 191
 192         if (post_read_steps) {
 193                 ctx = mempool_alloc(bio_post_read_ctx_pool, GFP_NOFS);
 194                 if (!ctx)
 195                         return ERR_PTR(-ENOMEM);
 196                 ctx->bio = bio;
 197                 ctx->enabled_steps = post_read_steps;
 198                 bio->bi_private = ctx;
 199         }
 200         return ctx;
 201 }
 202
 203 static inline loff_t ext4_readpage_limit(struct inode *inode)
 204 {
 205         if (IS_ENABLED(CONFIG_FS_VERITY) &&
 206             (IS_VERITY(inode) || ext4_verity_in_progress(inode)))
 207                 return inode->i_sb->s_maxbytes;
 208
 209         return i_size_read(inode);
 210 }
 211
 212 int ext4_mpage_readpages(struct address_space *mapping,
 213                          struct list_head *pages, struct page *page,
 214                          unsigned nr_pages, bool is_readahead)
 215 {
 216         struct bio *bio = NULL;
 217         sector_t last_block_in_bio = 0;
 218
 219         struct inode *inode = mapping->host;
 220         const unsigned blkbits = inode->i_blkbits;
 221         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 222         const unsigned blocksize = 1 << blkbits;
 223         sector_t block_in_file;
 224         sector_t last_block;
 225         sector_t last_block_in_file;
 226         sector_t blocks[MAX_BUF_PER_PAGE];
 227         unsigned page_block;
 228         struct block_device *bdev = inode->i_sb->s_bdev;
 229         int length;
 230         unsigned relative_block = 0;
 231         struct ext4_map_blocks map;
 232
 233         map.m_pblk = 0;
 234         map.m_lblk = 0;
 235         map.m_len = 0;
 236         map.m_flags = 0;
 237
 238         for (; nr_pages; nr_pages--) {
 239                 int fully_mapped = 1;
 240                 unsigned first_hole = blocks_per_page;
 241
 242                 if (pages) {
 243                         page = lru_to_page(pages);
 244
 245                         prefetchw(&page->flags);
 246                         list_del(&page->lru);
 247                         if (add_to_page_cache_lru(page, mapping, page->index,
 248                                   readahead_gfp_mask(mapping)))
 249                                 goto next_page;
 250                 }
 251
 252                 if (page_has_buffers(page))
 253                         goto confused;
 254
 255                 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
 256                 last_block = block_in_file + nr_pages * blocks_per_page;
 257                 last_block_in_file = (ext4_readpage_limit(inode) +
 258                                       blocksize - 1) >> blkbits;
 259                 if (last_block > last_block_in_file)
 260                         last_block = last_block_in_file;
 261                 page_block = 0;
 262
 263                 /*
 264                  * Map blocks using the previous result first.
 265                  */
 266                 if ((map.m_flags & EXT4_MAP_MAPPED) &&
 267                     block_in_file > map.m_lblk &&
 268                     block_in_file < (map.m_lblk + map.m_len)) {
 269                         unsigned map_offset = block_in_file - map.m_lblk;
 270                         unsigned last = map.m_len - map_offset;
 271
 272                         for (relative_block = 0; ; relative_block++) {
 273                                 if (relative_block == last) {
 274                                         /* needed? */
 275                                         map.m_flags &= ~EXT4_MAP_MAPPED;
 276                                         break;
 277                                 }
 278                                 if (page_block == blocks_per_page)
 279                                         break;
 280                                 blocks[page_block] = map.m_pblk + map_offset +
 281                                         relative_block;
 282                                 page_block++;
 283                                 block_in_file++;
 284                         }
 285                 }
 286
 287                 /*
 288                  * Then do more ext4_map_blocks() calls until we are
 289                  * done with this page.
 290                  */
 291                 while (page_block < blocks_per_page) {
 292                         if (block_in_file < last_block) {
 293                                 map.m_lblk = block_in_file;
 294                                 map.m_len = last_block - block_in_file;
 295
 296                                 if (ext4_map_blocks(NULL, inode, &map, 0) < 0) {
 297                                 set_error_page:
 298                                         SetPageError(page);
 299                                         zero_user_segment(page, 0,
 300                                                           PAGE_SIZE);
 301                                         unlock_page(page);
 302                                         goto next_page;
 303                                 }
 304                         }
 305                         if ((map.m_flags & EXT4_MAP_MAPPED) == 0) {
 306                                 fully_mapped = 0;
 307                                 if (first_hole == blocks_per_page)
 308                                         first_hole = page_block;
 309                                 page_block++;
 310                                 block_in_file++;
 311                                 continue;
 312                         }
 313                         if (first_hole != blocks_per_page)
 314                                 goto confused;          /* hole -> non-hole */
 315
 316                         /* Contiguous blocks? */
 317                         if (page_block && blocks[page_block-1] != map.m_pblk-1)
 318                                 goto confused;
 319                         for (relative_block = 0; ; relative_block++) {
 320                                 if (relative_block == map.m_len) {
 321                                         /* needed? */
 322                                         map.m_flags &= ~EXT4_MAP_MAPPED;
 323                                         break;
 324                                 } else if (page_block == blocks_per_page)
 325                                         break;
 326                                 blocks[page_block] = map.m_pblk+relative_block;
 327                                 page_block++;
 328                                 block_in_file++;
 329                         }
 330                 }
 331                 if (first_hole != blocks_per_page) {
 332                         zero_user_segment(page, first_hole << blkbits,
 333                                           PAGE_SIZE);
 334                         if (first_hole == 0) {
 335                                 if (ext4_need_verity(inode, page->index) &&
 336                                     !fsverity_verify_page(page))
 337                                         goto set_error_page;
 338                                 SetPageUptodate(page);
 339                                 unlock_page(page);
 340                                 goto next_page;
 341                         }
 342                 } else if (fully_mapped) {
 343                         SetPageMappedToDisk(page);
 344                 }
 345                 if (fully_mapped && blocks_per_page == 1 &&
 346                     !PageUptodate(page) && cleancache_get_page(page) == 0) {
 347                         SetPageUptodate(page);
 348                         goto confused;
 349                 }
 350
 351                 /*
 352                  * This page will go to BIO.  Do we need to send this
 353                  * BIO off first?
 354                  */
 355                 if (bio && (last_block_in_bio != blocks[0] - 1)) {
 356                 submit_and_realloc:
 357                         submit_bio(bio);
 358                         bio = NULL;
 359                 }
 360                 if (bio == NULL) {
 361                         struct bio_post_read_ctx *ctx;
 362
 363                         /*
 364                          * bio_alloc will _always_ be able to allocate a bio if
 365                          * __GFP_DIRECT_RECLAIM is set, see bio_alloc_bioset().
 366                          */
 367                         bio = bio_alloc(GFP_KERNEL,
 368                                 min_t(int, nr_pages, BIO_MAX_PAGES));
 369                         ctx = get_bio_post_read_ctx(inode, bio, page->index);
 370                         if (IS_ERR(ctx)) {
 371                                 bio_put(bio);
 372                                 bio = NULL;
 373                                 goto set_error_page;
 374                         }
 375                         bio_set_dev(bio, bdev);
 376                         bio->bi_iter.bi_sector = blocks[0] << (blkbits - 9);
 377                         bio->bi_end_io = mpage_end_io;
 378                         bio->bi_private = ctx;
 379                         bio_set_op_attrs(bio, REQ_OP_READ,
 380                                                 is_readahead ? REQ_RAHEAD : 0);
 381                 }
 382
 383                 length = first_hole << blkbits;
 384                 if (bio_add_page(bio, page, length, 0) < length)
 385                         goto submit_and_realloc;
 386
 387                 if (((map.m_flags & EXT4_MAP_BOUNDARY) &&
 388                      (relative_block == map.m_len)) ||
 389                     (first_hole != blocks_per_page)) {
 390                         submit_bio(bio);
 391                         bio = NULL;
 392                 } else
 393                         last_block_in_bio = blocks[blocks_per_page - 1];
 394                 goto next_page;
 395         confused:
 396                 if (bio) {
 397                         submit_bio(bio);
 398                         bio = NULL;
 399                 }
 400                 if (!PageUptodate(page))
 401                         block_read_full_page(page, ext4_get_block);
 402                 else
 403                         unlock_page(page);
 404         next_page:
 405                 if (pages)
 406                         put_page(page);
 407         }
 408         BUG_ON(pages && !list_empty(pages));
 409         if (bio)
 410                 submit_bio(bio);
 411         return 0;
 412 }
 413
 414 int __init ext4_init_post_read_processing(void)
 415 {
 416         bio_post_read_ctx_cache =
 417                 kmem_cache_create("ext4_bio_post_read_ctx",
 418                                   sizeof(struct bio_post_read_ctx), 0, 0, NULL);
 419         if (!bio_post_read_ctx_cache)
 420                 goto fail;
 421         bio_post_read_ctx_pool =
 422                 mempool_create_slab_pool(NUM_PREALLOC_POST_READ_CTXS,
 423                                          bio_post_read_ctx_cache);
 424         if (!bio_post_read_ctx_pool)
 425                 goto fail_free_cache;
 426         return 0;
 427
 428 fail_free_cache:
 429         kmem_cache_destroy(bio_post_read_ctx_cache);
 430 fail:
 431         return -ENOMEM;
 432 }
 433
 434 void ext4_exit_post_read_processing(void)
 435 {
 436         mempool_destroy(bio_post_read_ctx_pool);
 437         kmem_cache_destroy(bio_post_read_ctx_cache);
 438 }