fs/ext4/page-io.c

   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/module.h>
  10 #include <linux/fs.h>
  11 #include <linux/time.h>
  12 #include <linux/jbd2.h>
  13 #include <linux/highuid.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/quotaops.h>
  16 #include <linux/string.h>
  17 #include <linux/buffer_head.h>
  18 #include <linux/writeback.h>
  19 #include <linux/pagevec.h>
  20 #include <linux/mpage.h>
  21 #include <linux/namei.h>
  22 #include <linux/uio.h>
  23 #include <linux/bio.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/kernel.h>
  26 #include <linux/slab.h>
  27
  28 #include "ext4_jbd2.h"
  29 #include "xattr.h"
  30 #include "acl.h"
  31 #include "ext4_extents.h"
  32
  33 static struct kmem_cache *io_page_cachep, *io_end_cachep;
  34
  35 int __init ext4_init_pageio(void)
  36 {
  37         io_page_cachep = KMEM_CACHE(ext4_io_page, SLAB_RECLAIM_ACCOUNT);
  38         if (io_page_cachep == NULL)
  39                 return -ENOMEM;
  40         io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
  41         if (io_end_cachep == NULL) {
  42                 kmem_cache_destroy(io_page_cachep);
  43                 return -ENOMEM;
  44         }
  45         return 0;
  46 }
  47
  48 void ext4_exit_pageio(void)
  49 {
  50         kmem_cache_destroy(io_end_cachep);
  51         kmem_cache_destroy(io_page_cachep);
  52 }
  53
  54 void ext4_ioend_wait(struct inode *inode)
  55 {
  56         wait_queue_head_t *wq = ext4_ioend_wq(inode);
  57
  58         wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_ioend_count) == 0));
  59 }
  60
  61 static void put_io_page(struct ext4_io_page *io_page)
  62 {
  63         if (atomic_dec_and_test(&io_page->p_count)) {
  64                 end_page_writeback(io_page->p_page);
  65                 put_page(io_page->p_page);
  66                 kmem_cache_free(io_page_cachep, io_page);
  67         }
  68 }
  69
  70 void ext4_free_io_end(ext4_io_end_t *io)
  71 {
  72         int i;
  73         wait_queue_head_t *wq;
  74
  75         BUG_ON(!io);
  76         if (io->page)
  77                 put_page(io->page);
  78         for (i = 0; i < io->num_io_pages; i++)
  79                 put_io_page(io->pages[i]);
  80         io->num_io_pages = 0;
  81         wq = ext4_ioend_wq(io->inode);
  82         if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count) &&
  83             waitqueue_active(wq))
  84                 wake_up_all(wq);
  85         kmem_cache_free(io_end_cachep, io);
  86 }
  87
  88 /*
  89  * check a range of space and convert unwritten extents to written.
  90  */
  91 int ext4_end_io_nolock(ext4_io_end_t *io)
  92 {
  93         struct inode *inode = io->inode;
  94         loff_t offset = io->offset;
  95         ssize_t size = io->size;
  96         wait_queue_head_t *wq;
  97         int ret = 0;
  98
  99         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
 100                    "list->prev 0x%p\n",
 101                    io, inode->i_ino, io->list.next, io->list.prev);
 102
 103         if (list_empty(&io->list))
 104                 return ret;
 105
 106         if (!(io->flag & EXT4_IO_END_UNWRITTEN))
 107                 return ret;
 108
 109         ret = ext4_convert_unwritten_extents(inode, offset, size);
 110         if (ret < 0) {
 111                 printk(KERN_EMERG "%s: failed to convert unwritten "
 112                         "extents to written extents, error is %d "
 113                         "io is still on inode %lu aio dio list\n",
 114                        __func__, ret, inode->i_ino);
 115                 return ret;
 116         }
 117
 118         if (io->iocb)
 119                 aio_complete(io->iocb, io->result, 0);
 120         /* clear the DIO AIO unwritten flag */
 121         if (io->flag & EXT4_IO_END_UNWRITTEN) {
 122                 io->flag &= ~EXT4_IO_END_UNWRITTEN;
 123                 /* Wake up anyone waiting on unwritten extent conversion */
 124                 wq = ext4_ioend_wq(io->inode);
 125                 if (atomic_dec_and_test(&EXT4_I(inode)->i_aiodio_unwritten) &&
 126                     waitqueue_active(wq)) {
 127                         wake_up_all(wq);
 128                 }
 129         }
 130
 131         return ret;
 132 }
 133
 134 /*
 135  * work on completed aio dio IO, to convert unwritten extents to extents
 136  */
 137 static void ext4_end_io_work(struct work_struct *work)
 138 {
 139         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
 140         struct inode            *inode = io->inode;
 141         struct ext4_inode_info  *ei = EXT4_I(inode);
 142         unsigned long           flags;
 143         int                     ret;
 144
 145         if (!mutex_trylock(&inode->i_mutex)) {
 146                 /*
 147                  * Requeue the work instead of waiting so that the work
 148                  * items queued after this can be processed.
 149                  */
 150                 queue_work(EXT4_SB(inode->i_sb)->dio_unwritten_wq, &io->work);
 151                 /*
 152                  * To prevent the ext4-dio-unwritten thread from keeping
 153                  * requeueing end_io requests and occupying cpu for too long,
 154                  * yield the cpu if it sees an end_io request that has already
 155                  * been requeued.
 156                  */
 157                 if (io->flag & EXT4_IO_END_QUEUED)
 158                         yield();
 159                 io->flag |= EXT4_IO_END_QUEUED;
 160                 return;
 161         }
 162         ret = ext4_end_io_nolock(io);
 163         if (ret < 0) {
 164                 mutex_unlock(&inode->i_mutex);
 165                 return;
 166         }
 167
 168         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
 169         if (!list_empty(&io->list))
 170                 list_del_init(&io->list);
 171         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
 172         mutex_unlock(&inode->i_mutex);
 173         ext4_free_io_end(io);
 174 }
 175
 176 ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
 177 {
 178         ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
 179         if (io) {
 180                 atomic_inc(&EXT4_I(inode)->i_ioend_count);
 181                 io->inode = inode;
 182                 INIT_WORK(&io->work, ext4_end_io_work);
 183                 INIT_LIST_HEAD(&io->list);
 184         }
 185         return io;
 186 }
 187
 188 /*
 189  * Print an buffer I/O error compatible with the fs/buffer.c.  This
 190  * provides compatibility with dmesg scrapers that look for a specific
 191  * buffer I/O error message.  We really need a unified error reporting
 192  * structure to userspace ala Digital Unix's uerf system, but it's
 193  * probably not going to happen in my lifetime, due to LKML politics...
 194  */
 195 static void buffer_io_error(struct buffer_head *bh)
 196 {
 197         char b[BDEVNAME_SIZE];
 198         printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
 199                         bdevname(bh->b_bdev, b),
 200                         (unsigned long long)bh->b_blocknr);
 201 }
 202
 203 static void ext4_end_bio(struct bio *bio, int error)
 204 {
 205         ext4_io_end_t *io_end = bio->bi_private;
 206         struct workqueue_struct *wq;
 207         struct inode *inode;
 208         unsigned long flags;
 209         int i;
 210         sector_t bi_sector = bio->bi_sector;
 211
 212         BUG_ON(!io_end);
 213         bio->bi_private = NULL;
 214         bio->bi_end_io = NULL;
 215         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 216                 error = 0;
 217         bio_put(bio);
 218
 219         for (i = 0; i < io_end->num_io_pages; i++) {
 220                 struct page *page = io_end->pages[i]->p_page;
 221                 struct buffer_head *bh, *head;
 222                 loff_t offset;
 223                 loff_t io_end_offset;
 224
 225                 if (error) {
 226                         SetPageError(page);
 227                         set_bit(AS_EIO, &page->mapping->flags);
 228                         head = page_buffers(page);
 229                         BUG_ON(!head);
 230
 231                         io_end_offset = io_end->offset + io_end->size;
 232
 233                         offset = (sector_t) page->index << PAGE_CACHE_SHIFT;
 234                         bh = head;
 235                         do {
 236                                 if ((offset >= io_end->offset) &&
 237                                     (offset+bh->b_size <= io_end_offset))
 238                                         buffer_io_error(bh);
 239
 240                                 offset += bh->b_size;
 241                                 bh = bh->b_this_page;
 242                         } while (bh != head);
 243                 }
 244
 245                 put_io_page(io_end->pages[i]);
 246         }
 247         io_end->num_io_pages = 0;
 248         inode = io_end->inode;
 249
 250         if (error) {
 251                 io_end->flag |= EXT4_IO_END_ERROR;
 252                 ext4_warning(inode->i_sb, "I/O error writing to inode %lu "
 253                              "(offset %llu size %ld starting block %llu)",
 254                              inode->i_ino,
 255                              (unsigned long long) io_end->offset,
 256                              (long) io_end->size,
 257                              (unsigned long long)
 258                              bi_sector >> (inode->i_blkbits - 9));
 259         }
 260
 261         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
 262                 ext4_free_io_end(io_end);
 263                 return;
 264         }
 265
 266         /* Add the io_end to per-inode completed io list*/
 267         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
 268         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
 269         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
 270
 271         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
 272         /* queue the work to convert unwritten extents to written */
 273         queue_work(wq, &io_end->work);
 274 }
 275
 276 void ext4_io_submit(struct ext4_io_submit *io)
 277 {
 278         struct bio *bio = io->io_bio;
 279
 280         if (bio) {
 281                 bio_get(io->io_bio);
 282                 submit_bio(io->io_op, io->io_bio);
 283                 BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
 284                 bio_put(io->io_bio);
 285         }
 286         io->io_bio = NULL;
 287         io->io_op = 0;
 288         io->io_end = NULL;
 289 }
 290
 291 static int io_submit_init(struct ext4_io_submit *io,
 292                           struct inode *inode,
 293                           struct writeback_control *wbc,
 294                           struct buffer_head *bh)
 295 {
 296         ext4_io_end_t *io_end;
 297         struct page *page = bh->b_page;
 298         int nvecs = bio_get_nr_vecs(bh->b_bdev);
 299         struct bio *bio;
 300
 301         io_end = ext4_init_io_end(inode, GFP_NOFS);
 302         if (!io_end)
 303                 return -ENOMEM;
 304         do {
 305                 bio = bio_alloc(GFP_NOIO, nvecs);
 306                 nvecs >>= 1;
 307         } while (bio == NULL);
 308
 309         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
 310         bio->bi_bdev = bh->b_bdev;
 311         bio->bi_private = io->io_end = io_end;
 312         bio->bi_end_io = ext4_end_bio;
 313
 314         io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
 315
 316         io->io_bio = bio;
 317         io->io_op = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
 318         io->io_next_block = bh->b_blocknr;
 319         return 0;
 320 }
 321
 322 static int io_submit_add_bh(struct ext4_io_submit *io,
 323                             struct ext4_io_page *io_page,
 324                             struct inode *inode,
 325                             struct writeback_control *wbc,
 326                             struct buffer_head *bh)
 327 {
 328         ext4_io_end_t *io_end;
 329         int ret;
 330
 331         if (buffer_new(bh)) {
 332                 clear_buffer_new(bh);
 333                 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 334         }
 335
 336         if (!buffer_mapped(bh) || buffer_delay(bh)) {
 337                 if (!buffer_mapped(bh))
 338                         clear_buffer_dirty(bh);
 339                 if (io->io_bio)
 340                         ext4_io_submit(io);
 341                 return 0;
 342         }
 343
 344         if (io->io_bio && bh->b_blocknr != io->io_next_block) {
 345 submit_and_retry:
 346                 ext4_io_submit(io);
 347         }
 348         if (io->io_bio == NULL) {
 349                 ret = io_submit_init(io, inode, wbc, bh);
 350                 if (ret)
 351                         return ret;
 352         }
 353         io_end = io->io_end;
 354         if ((io_end->num_io_pages >= MAX_IO_PAGES) &&
 355             (io_end->pages[io_end->num_io_pages-1] != io_page))
 356                 goto submit_and_retry;
 357         if (buffer_uninit(bh) && !(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
 358                 io_end->flag |= EXT4_IO_END_UNWRITTEN;
 359                 atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
 360         }
 361         io->io_end->size += bh->b_size;
 362         io->io_next_block++;
 363         ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
 364         if (ret != bh->b_size)
 365                 goto submit_and_retry;
 366         if ((io_end->num_io_pages == 0) ||
 367             (io_end->pages[io_end->num_io_pages-1] != io_page)) {
 368                 io_end->pages[io_end->num_io_pages++] = io_page;
 369                 atomic_inc(&io_page->p_count);
 370         }
 371         return 0;
 372 }
 373
 374 int ext4_bio_write_page(struct ext4_io_submit *io,
 375                         struct page *page,
 376                         int len,
 377                         struct writeback_control *wbc)
 378 {
 379         struct inode *inode = page->mapping->host;
 380         unsigned block_start, block_end, blocksize;
 381         struct ext4_io_page *io_page;
 382         struct buffer_head *bh, *head;
 383         int ret = 0;
 384
 385         blocksize = 1 << inode->i_blkbits;
 386
 387         BUG_ON(!PageLocked(page));
 388         BUG_ON(PageWriteback(page));
 389
 390         io_page = kmem_cache_alloc(io_page_cachep, GFP_NOFS);
 391         if (!io_page) {
 392                 set_page_dirty(page);
 393                 unlock_page(page);
 394                 return -ENOMEM;
 395         }
 396         io_page->p_page = page;
 397         atomic_set(&io_page->p_count, 1);
 398         get_page(page);
 399         set_page_writeback(page);
 400         ClearPageError(page);
 401
 402         for (bh = head = page_buffers(page), block_start = 0;
 403              bh != head || !block_start;
 404              block_start = block_end, bh = bh->b_this_page) {
 405
 406                 block_end = block_start + blocksize;
 407                 if (block_start >= len) {
 408                         /*
 409                          * Comments copied from block_write_full_page_endio:
 410                          *
 411                          * The page straddles i_size.  It must be zeroed out on
 412                          * each and every writepage invocation because it may
 413                          * be mmapped.  "A file is mapped in multiples of the
 414                          * page size.  For a file that is not a multiple of
 415                          * the  page size, the remaining memory is zeroed when
 416                          * mapped, and writes to that region are not written
 417                          * out to the file."
 418                          */
 419                         zero_user_segment(page, block_start, block_end);
 420                         clear_buffer_dirty(bh);
 421                         set_buffer_uptodate(bh);
 422                         continue;
 423                 }
 424                 clear_buffer_dirty(bh);
 425                 ret = io_submit_add_bh(io, io_page, inode, wbc, bh);
 426                 if (ret) {
 427                         /*
 428                          * We only get here on ENOMEM.  Not much else
 429                          * we can do but mark the page as dirty, and
 430                          * better luck next time.
 431                          */
 432                         set_page_dirty(page);
 433                         break;
 434                 }
 435         }
 436         unlock_page(page);
 437         /*
 438          * If the page was truncated before we could do the writeback,
 439          * or we had a memory allocation error while trying to write
 440          * the first buffer head, we won't have submitted any pages for
 441          * I/O.  In that case we need to make sure we've cleared the
 442          * PageWriteback bit from the page to prevent the system from
 443          * wedging later on.
 444          */
 445         put_io_page(io_page);
 446         return ret;
 447 }