1 /* 2 * linux/fs/ext4/page-io.c 3 * 4 * This contains the new page_io functions for ext4 5 * 6 * Written by Theodore Ts'o, 2010. 7 */ 8 9 #include <linux/fs.h> 10 #include <linux/time.h> 11 #include <linux/jbd2.h> 12 #include <linux/highuid.h> 13 #include <linux/pagemap.h> 14 #include <linux/quotaops.h> 15 #include <linux/string.h> 16 #include <linux/buffer_head.h> 17 #include <linux/writeback.h> 18 #include <linux/pagevec.h> 19 #include <linux/mpage.h> 20 #include <linux/namei.h> 21 #include <linux/uio.h> 22 #include <linux/bio.h> 23 #include <linux/workqueue.h> 24 #include <linux/kernel.h> 25 #include <linux/slab.h> 26 27 #include "ext4_jbd2.h" 28 #include "xattr.h" 29 #include "acl.h" 30 #include "ext4_extents.h" 31 32 static struct kmem_cache *io_page_cachep, *io_end_cachep; 33 34 int __init ext4_init_pageio(void) 35 { 36 io_page_cachep = KMEM_CACHE(ext4_io_page, SLAB_RECLAIM_ACCOUNT); 37 if (io_page_cachep == NULL) 38 return -ENOMEM; 39 io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT); 40 if (io_end_cachep == NULL) { 41 kmem_cache_destroy(io_page_cachep); 42 return -ENOMEM; 43 } 44 return 0; 45 } 46 47 void ext4_exit_pageio(void) 48 { 49 kmem_cache_destroy(io_end_cachep); 50 kmem_cache_destroy(io_page_cachep); 51 } 52 53 void ext4_ioend_wait(struct inode *inode) 54 { 55 wait_queue_head_t *wq = ext4_ioend_wq(inode); 56 57 wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_ioend_count) == 0)); 58 } 59 60 static void put_io_page(struct ext4_io_page *io_page) 61 { 62 if (atomic_dec_and_test(&io_page->p_count)) { 63 end_page_writeback(io_page->p_page); 64 put_page(io_page->p_page); 65 kmem_cache_free(io_page_cachep, io_page); 66 } 67 } 68 69 void ext4_free_io_end(ext4_io_end_t *io) 70 { 71 int i; 72 73 BUG_ON(!io); 74 if (io->page) 75 put_page(io->page); 76 for (i = 0; i < io->num_io_pages; i++) 77 put_io_page(io->pages[i]); 78 io->num_io_pages = 0; 79 if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count)) 80 wake_up_all(ext4_ioend_wq(io->inode)); 81 kmem_cache_free(io_end_cachep, io); 82 } 83 84 /* 85 * check a range of space and convert unwritten extents to written. 86 * 87 * Called with inode->i_mutex; we depend on this when we manipulate 88 * io->flag, since we could otherwise race with ext4_flush_completed_IO() 89 */ 90 int ext4_end_io_nolock(ext4_io_end_t *io) 91 { 92 struct inode *inode = io->inode; 93 loff_t offset = io->offset; 94 ssize_t size = io->size; 95 int ret = 0; 96 97 ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p," 98 "list->prev 0x%p\n", 99 io, inode->i_ino, io->list.next, io->list.prev); 100 101 ret = ext4_convert_unwritten_extents(inode, offset, size); 102 if (ret < 0) { 103 ext4_msg(inode->i_sb, KERN_EMERG, 104 "failed to convert unwritten extents to written " 105 "extents -- potential data loss! " 106 "(inode %lu, offset %llu, size %zd, error %d)", 107 inode->i_ino, offset, size, ret); 108 } 109 110 if (io->iocb) 111 aio_complete(io->iocb, io->result, 0); 112 113 /* Wake up anyone waiting on unwritten extent conversion */ 114 if (atomic_dec_and_test(&EXT4_I(inode)->i_aiodio_unwritten)) 115 wake_up_all(ext4_ioend_wq(io->inode)); 116 return ret; 117 } 118 119 /* 120 * work on completed aio dio IO, to convert unwritten extents to extents 121 */ 122 static void ext4_end_io_work(struct work_struct *work) 123 { 124 ext4_io_end_t *io = container_of(work, ext4_io_end_t, work); 125 struct inode *inode = io->inode; 126 struct ext4_inode_info *ei = EXT4_I(inode); 127 unsigned long flags; 128 129 spin_lock_irqsave(&ei->i_completed_io_lock, flags); 130 if (list_empty(&io->list)) { 131 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags); 132 goto free; 133 } 134 135 if (!mutex_trylock(&inode->i_mutex)) { 136 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags); 137 /* 138 * Requeue the work instead of waiting so that the work 139 * items queued after this can be processed. 140 */ 141 queue_work(EXT4_SB(inode->i_sb)->dio_unwritten_wq, &io->work); 142 /* 143 * To prevent the ext4-dio-unwritten thread from keeping 144 * requeueing end_io requests and occupying cpu for too long, 145 * yield the cpu if it sees an end_io request that has already 146 * been requeued. 147 */ 148 if (io->flag & EXT4_IO_END_QUEUED) 149 yield(); 150 io->flag |= EXT4_IO_END_QUEUED; 151 return; 152 } 153 list_del_init(&io->list); 154 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags); 155 (void) ext4_end_io_nolock(io); 156 mutex_unlock(&inode->i_mutex); 157 free: 158 ext4_free_io_end(io); 159 } 160 161 ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags) 162 { 163 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags); 164 if (io) { 165 atomic_inc(&EXT4_I(inode)->i_ioend_count); 166 io->inode = inode; 167 INIT_WORK(&io->work, ext4_end_io_work); 168 INIT_LIST_HEAD(&io->list); 169 } 170 return io; 171 } 172 173 /* 174 * Print an buffer I/O error compatible with the fs/buffer.c. This 175 * provides compatibility with dmesg scrapers that look for a specific 176 * buffer I/O error message. We really need a unified error reporting 177 * structure to userspace ala Digital Unix's uerf system, but it's 178 * probably not going to happen in my lifetime, due to LKML politics... 179 */ 180 static void buffer_io_error(struct buffer_head *bh) 181 { 182 char b[BDEVNAME_SIZE]; 183 printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n", 184 bdevname(bh->b_bdev, b), 185 (unsigned long long)bh->b_blocknr); 186 } 187 188 static void ext4_end_bio(struct bio *bio, int error) 189 { 190 ext4_io_end_t *io_end = bio->bi_private; 191 struct workqueue_struct *wq; 192 struct inode *inode; 193 unsigned long flags; 194 int i; 195 sector_t bi_sector = bio->bi_sector; 196 197 BUG_ON(!io_end); 198 bio->bi_private = NULL; 199 bio->bi_end_io = NULL; 200 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 201 error = 0; 202 bio_put(bio); 203 204 for (i = 0; i < io_end->num_io_pages; i++) { 205 struct page *page = io_end->pages[i]->p_page; 206 struct buffer_head *bh, *head; 207 loff_t offset; 208 loff_t io_end_offset; 209 210 if (error) { 211 SetPageError(page); 212 set_bit(AS_EIO, &page->mapping->flags); 213 head = page_buffers(page); 214 BUG_ON(!head); 215 216 io_end_offset = io_end->offset + io_end->size; 217 218 offset = (sector_t) page->index << PAGE_CACHE_SHIFT; 219 bh = head; 220 do { 221 if ((offset >= io_end->offset) && 222 (offset+bh->b_size <= io_end_offset)) 223 buffer_io_error(bh); 224 225 offset += bh->b_size; 226 bh = bh->b_this_page; 227 } while (bh != head); 228 } 229 230 put_io_page(io_end->pages[i]); 231 } 232 io_end->num_io_pages = 0; 233 inode = io_end->inode; 234 235 if (error) { 236 io_end->flag |= EXT4_IO_END_ERROR; 237 ext4_warning(inode->i_sb, "I/O error writing to inode %lu " 238 "(offset %llu size %ld starting block %llu)", 239 inode->i_ino, 240 (unsigned long long) io_end->offset, 241 (long) io_end->size, 242 (unsigned long long) 243 bi_sector >> (inode->i_blkbits - 9)); 244 } 245 246 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) { 247 ext4_free_io_end(io_end); 248 return; 249 } 250 251 /* Add the io_end to per-inode completed io list*/ 252 spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags); 253 list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list); 254 spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags); 255 256 wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq; 257 /* queue the work to convert unwritten extents to written */ 258 queue_work(wq, &io_end->work); 259 } 260 261 void ext4_io_submit(struct ext4_io_submit *io) 262 { 263 struct bio *bio = io->io_bio; 264 265 if (bio) { 266 bio_get(io->io_bio); 267 submit_bio(io->io_op, io->io_bio); 268 BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP)); 269 bio_put(io->io_bio); 270 } 271 io->io_bio = NULL; 272 io->io_op = 0; 273 io->io_end = NULL; 274 } 275 276 static int io_submit_init(struct ext4_io_submit *io, 277 struct inode *inode, 278 struct writeback_control *wbc, 279 struct buffer_head *bh) 280 { 281 ext4_io_end_t *io_end; 282 struct page *page = bh->b_page; 283 int nvecs = bio_get_nr_vecs(bh->b_bdev); 284 struct bio *bio; 285 286 io_end = ext4_init_io_end(inode, GFP_NOFS); 287 if (!io_end) 288 return -ENOMEM; 289 bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES)); 290 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9); 291 bio->bi_bdev = bh->b_bdev; 292 bio->bi_private = io->io_end = io_end; 293 bio->bi_end_io = ext4_end_bio; 294 295 io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh); 296 297 io->io_bio = bio; 298 io->io_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE); 299 io->io_next_block = bh->b_blocknr; 300 return 0; 301 } 302 303 static int io_submit_add_bh(struct ext4_io_submit *io, 304 struct ext4_io_page *io_page, 305 struct inode *inode, 306 struct writeback_control *wbc, 307 struct buffer_head *bh) 308 { 309 ext4_io_end_t *io_end; 310 int ret; 311 312 if (buffer_new(bh)) { 313 clear_buffer_new(bh); 314 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 315 } 316 317 if (!buffer_mapped(bh) || buffer_delay(bh)) { 318 if (!buffer_mapped(bh)) 319 clear_buffer_dirty(bh); 320 if (io->io_bio) 321 ext4_io_submit(io); 322 return 0; 323 } 324 325 if (io->io_bio && bh->b_blocknr != io->io_next_block) { 326 submit_and_retry: 327 ext4_io_submit(io); 328 } 329 if (io->io_bio == NULL) { 330 ret = io_submit_init(io, inode, wbc, bh); 331 if (ret) 332 return ret; 333 } 334 io_end = io->io_end; 335 if ((io_end->num_io_pages >= MAX_IO_PAGES) && 336 (io_end->pages[io_end->num_io_pages-1] != io_page)) 337 goto submit_and_retry; 338 if (buffer_uninit(bh)) 339 ext4_set_io_unwritten_flag(inode, io_end); 340 io->io_end->size += bh->b_size; 341 io->io_next_block++; 342 ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh)); 343 if (ret != bh->b_size) 344 goto submit_and_retry; 345 if ((io_end->num_io_pages == 0) || 346 (io_end->pages[io_end->num_io_pages-1] != io_page)) { 347 io_end->pages[io_end->num_io_pages++] = io_page; 348 atomic_inc(&io_page->p_count); 349 } 350 return 0; 351 } 352 353 int ext4_bio_write_page(struct ext4_io_submit *io, 354 struct page *page, 355 int len, 356 struct writeback_control *wbc) 357 { 358 struct inode *inode = page->mapping->host; 359 unsigned block_start, block_end, blocksize; 360 struct ext4_io_page *io_page; 361 struct buffer_head *bh, *head; 362 int ret = 0; 363 364 blocksize = 1 << inode->i_blkbits; 365 366 BUG_ON(!PageLocked(page)); 367 BUG_ON(PageWriteback(page)); 368 369 io_page = kmem_cache_alloc(io_page_cachep, GFP_NOFS); 370 if (!io_page) { 371 set_page_dirty(page); 372 unlock_page(page); 373 return -ENOMEM; 374 } 375 io_page->p_page = page; 376 atomic_set(&io_page->p_count, 1); 377 get_page(page); 378 set_page_writeback(page); 379 ClearPageError(page); 380 381 for (bh = head = page_buffers(page), block_start = 0; 382 bh != head || !block_start; 383 block_start = block_end, bh = bh->b_this_page) { 384 385 block_end = block_start + blocksize; 386 if (block_start >= len) { 387 /* 388 * Comments copied from block_write_full_page_endio: 389 * 390 * The page straddles i_size. It must be zeroed out on 391 * each and every writepage invocation because it may 392 * be mmapped. "A file is mapped in multiples of the 393 * page size. For a file that is not a multiple of 394 * the page size, the remaining memory is zeroed when 395 * mapped, and writes to that region are not written 396 * out to the file." 397 */ 398 zero_user_segment(page, block_start, block_end); 399 clear_buffer_dirty(bh); 400 set_buffer_uptodate(bh); 401 continue; 402 } 403 clear_buffer_dirty(bh); 404 ret = io_submit_add_bh(io, io_page, inode, wbc, bh); 405 if (ret) { 406 /* 407 * We only get here on ENOMEM. Not much else 408 * we can do but mark the page as dirty, and 409 * better luck next time. 410 */ 411 set_page_dirty(page); 412 break; 413 } 414 } 415 unlock_page(page); 416 /* 417 * If the page was truncated before we could do the writeback, 418 * or we had a memory allocation error while trying to write 419 * the first buffer head, we won't have submitted any pages for 420 * I/O. In that case we need to make sure we've cleared the 421 * PageWriteback bit from the page to prevent the system from 422 * wedging later on. 423 */ 424 put_io_page(io_page); 425 return ret; 426 } 427