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/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         BUG_ON(io->flag & EXT4_IO_END_UNWRITTEN);
  75
  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         if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count))
  82                 wake_up_all(ext4_ioend_wq(io->inode));
  83         kmem_cache_free(io_end_cachep, io);
  84 }
  85
  86 /*
  87  * check a range of space and convert unwritten extents to written.
  88  *
  89  * Called with inode->i_mutex; we depend on this when we manipulate
  90  * io->flag, since we could otherwise race with ext4_flush_completed_IO()
  91  */
  92 int ext4_end_io_nolock(ext4_io_end_t *io)
  93 {
  94         struct inode *inode = io->inode;
  95         loff_t offset = io->offset;
  96         ssize_t size = io->size;
  97         int ret = 0;
  98
  99         BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
 100
 101         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
 102                    "list->prev 0x%p\n",
 103                    io, inode->i_ino, io->list.next, io->list.prev);
 104
 105         ret = ext4_convert_unwritten_extents(inode, offset, size);
 106         if (ret < 0) {
 107                 ext4_msg(inode->i_sb, KERN_EMERG,
 108                          "failed to convert unwritten extents to written "
 109                          "extents -- potential data loss!  "
 110                          "(inode %lu, offset %llu, size %zd, error %d)",
 111                          inode->i_ino, offset, size, ret);
 112         }
 113         io->flag &= ~EXT4_IO_END_UNWRITTEN;
 114         if (io->iocb)
 115                 aio_complete(io->iocb, io->result, 0);
 116
 117         if (io->flag & EXT4_IO_END_DIRECT)
 118                 inode_dio_done(inode);
 119         /* Wake up anyone waiting on unwritten extent conversion */
 120         if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
 121                 wake_up_all(ext4_ioend_wq(io->inode));
 122         return ret;
 123 }
 124
 125 /*
 126  * work on completed aio dio IO, to convert unwritten extents to extents
 127  */
 128 static void ext4_end_io_work(struct work_struct *work)
 129 {
 130         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
 131         struct inode            *inode = io->inode;
 132         struct ext4_inode_info  *ei = EXT4_I(inode);
 133         unsigned long           flags;
 134
 135         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
 136         if (io->flag & EXT4_IO_END_IN_FSYNC)
 137                 goto requeue;
 138         if (list_empty(&io->list)) {
 139                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
 140                 goto free;
 141         }
 142
 143         if (!mutex_trylock(&inode->i_mutex)) {
 144                 bool was_queued;
 145 requeue:
 146                 was_queued = !!(io->flag & EXT4_IO_END_QUEUED);
 147                 io->flag |= EXT4_IO_END_QUEUED;
 148                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
 149                 /*
 150                  * Requeue the work instead of waiting so that the work
 151                  * items queued after this can be processed.
 152                  */
 153                 queue_work(EXT4_SB(inode->i_sb)->dio_unwritten_wq, &io->work);
 154                 /*
 155                  * To prevent the ext4-dio-unwritten thread from keeping
 156                  * requeueing end_io requests and occupying cpu for too long,
 157                  * yield the cpu if it sees an end_io request that has already
 158                  * been requeued.
 159                  */
 160                 if (was_queued)
 161                         yield();
 162                 return;
 163         }
 164         list_del_init(&io->list);
 165         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
 166         (void) ext4_end_io_nolock(io);
 167         mutex_unlock(&inode->i_mutex);
 168 free:
 169         ext4_free_io_end(io);
 170 }
 171
 172 ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
 173 {
 174         ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
 175         if (io) {
 176                 atomic_inc(&EXT4_I(inode)->i_ioend_count);
 177                 io->inode = inode;
 178                 INIT_WORK(&io->work, ext4_end_io_work);
 179                 INIT_LIST_HEAD(&io->list);
 180         }
 181         return io;
 182 }
 183
 184 /*
 185  * Print an buffer I/O error compatible with the fs/buffer.c.  This
 186  * provides compatibility with dmesg scrapers that look for a specific
 187  * buffer I/O error message.  We really need a unified error reporting
 188  * structure to userspace ala Digital Unix's uerf system, but it's
 189  * probably not going to happen in my lifetime, due to LKML politics...
 190  */
 191 static void buffer_io_error(struct buffer_head *bh)
 192 {
 193         char b[BDEVNAME_SIZE];
 194         printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
 195                         bdevname(bh->b_bdev, b),
 196                         (unsigned long long)bh->b_blocknr);
 197 }
 198
 199 static void ext4_end_bio(struct bio *bio, int error)
 200 {
 201         ext4_io_end_t *io_end = bio->bi_private;
 202         struct workqueue_struct *wq;
 203         struct inode *inode;
 204         unsigned long flags;
 205         int i;
 206         sector_t bi_sector = bio->bi_sector;
 207
 208         BUG_ON(!io_end);
 209         bio->bi_private = NULL;
 210         bio->bi_end_io = NULL;
 211         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 212                 error = 0;
 213         bio_put(bio);
 214
 215         for (i = 0; i < io_end->num_io_pages; i++) {
 216                 struct page *page = io_end->pages[i]->p_page;
 217                 struct buffer_head *bh, *head;
 218                 loff_t offset;
 219                 loff_t io_end_offset;
 220
 221                 if (error) {
 222                         SetPageError(page);
 223                         set_bit(AS_EIO, &page->mapping->flags);
 224                         head = page_buffers(page);
 225                         BUG_ON(!head);
 226
 227                         io_end_offset = io_end->offset + io_end->size;
 228
 229                         offset = (sector_t) page->index << PAGE_CACHE_SHIFT;
 230                         bh = head;
 231                         do {
 232                                 if ((offset >= io_end->offset) &&
 233                                     (offset+bh->b_size <= io_end_offset))
 234                                         buffer_io_error(bh);
 235
 236                                 offset += bh->b_size;
 237                                 bh = bh->b_this_page;
 238                         } while (bh != head);
 239                 }
 240
 241                 put_io_page(io_end->pages[i]);
 242         }
 243         io_end->num_io_pages = 0;
 244         inode = io_end->inode;
 245
 246         if (error) {
 247                 io_end->flag |= EXT4_IO_END_ERROR;
 248                 ext4_warning(inode->i_sb, "I/O error writing to inode %lu "
 249                              "(offset %llu size %ld starting block %llu)",
 250                              inode->i_ino,
 251                              (unsigned long long) io_end->offset,
 252                              (long) io_end->size,
 253                              (unsigned long long)
 254                              bi_sector >> (inode->i_blkbits - 9));
 255         }
 256
 257         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
 258                 ext4_free_io_end(io_end);
 259                 return;
 260         }
 261
 262         /* Add the io_end to per-inode completed io list*/
 263         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
 264         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
 265         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
 266
 267         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
 268         /* queue the work to convert unwritten extents to written */
 269         queue_work(wq, &io_end->work);
 270 }
 271
 272 void ext4_io_submit(struct ext4_io_submit *io)
 273 {
 274         struct bio *bio = io->io_bio;
 275
 276         if (bio) {
 277                 bio_get(io->io_bio);
 278                 submit_bio(io->io_op, io->io_bio);
 279                 BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
 280                 bio_put(io->io_bio);
 281         }
 282         io->io_bio = NULL;
 283         io->io_op = 0;
 284         io->io_end = NULL;
 285 }
 286
 287 static int io_submit_init(struct ext4_io_submit *io,
 288                           struct inode *inode,
 289                           struct writeback_control *wbc,
 290                           struct buffer_head *bh)
 291 {
 292         ext4_io_end_t *io_end;
 293         struct page *page = bh->b_page;
 294         int nvecs = bio_get_nr_vecs(bh->b_bdev);
 295         struct bio *bio;
 296
 297         io_end = ext4_init_io_end(inode, GFP_NOFS);
 298         if (!io_end)
 299                 return -ENOMEM;
 300         bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES));
 301         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
 302         bio->bi_bdev = bh->b_bdev;
 303         bio->bi_private = io->io_end = io_end;
 304         bio->bi_end_io = ext4_end_bio;
 305
 306         io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
 307
 308         io->io_bio = bio;
 309         io->io_op = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
 310         io->io_next_block = bh->b_blocknr;
 311         return 0;
 312 }
 313
 314 static int io_submit_add_bh(struct ext4_io_submit *io,
 315                             struct ext4_io_page *io_page,
 316                             struct inode *inode,
 317                             struct writeback_control *wbc,
 318                             struct buffer_head *bh)
 319 {
 320         ext4_io_end_t *io_end;
 321         int ret;
 322
 323         if (buffer_new(bh)) {
 324                 clear_buffer_new(bh);
 325                 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 326         }
 327
 328         if (!buffer_mapped(bh) || buffer_delay(bh)) {
 329                 if (!buffer_mapped(bh))
 330                         clear_buffer_dirty(bh);
 331                 if (io->io_bio)
 332                         ext4_io_submit(io);
 333                 return 0;
 334         }
 335
 336         if (io->io_bio && bh->b_blocknr != io->io_next_block) {
 337 submit_and_retry:
 338                 ext4_io_submit(io);
 339         }
 340         if (io->io_bio == NULL) {
 341                 ret = io_submit_init(io, inode, wbc, bh);
 342                 if (ret)
 343                         return ret;
 344         }
 345         io_end = io->io_end;
 346         if ((io_end->num_io_pages >= MAX_IO_PAGES) &&
 347             (io_end->pages[io_end->num_io_pages-1] != io_page))
 348                 goto submit_and_retry;
 349         if (buffer_uninit(bh))
 350                 ext4_set_io_unwritten_flag(inode, io_end);
 351         io->io_end->size += bh->b_size;
 352         io->io_next_block++;
 353         ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
 354         if (ret != bh->b_size)
 355                 goto submit_and_retry;
 356         if ((io_end->num_io_pages == 0) ||
 357             (io_end->pages[io_end->num_io_pages-1] != io_page)) {
 358                 io_end->pages[io_end->num_io_pages++] = io_page;
 359                 atomic_inc(&io_page->p_count);
 360         }
 361         return 0;
 362 }
 363
 364 int ext4_bio_write_page(struct ext4_io_submit *io,
 365                         struct page *page,
 366                         int len,
 367                         struct writeback_control *wbc)
 368 {
 369         struct inode *inode = page->mapping->host;
 370         unsigned block_start, block_end, blocksize;
 371         struct ext4_io_page *io_page;
 372         struct buffer_head *bh, *head;
 373         int ret = 0;
 374
 375         blocksize = 1 << inode->i_blkbits;
 376
 377         BUG_ON(!PageLocked(page));
 378         BUG_ON(PageWriteback(page));
 379
 380         io_page = kmem_cache_alloc(io_page_cachep, GFP_NOFS);
 381         if (!io_page) {
 382                 set_page_dirty(page);
 383                 unlock_page(page);
 384                 return -ENOMEM;
 385         }
 386         io_page->p_page = page;
 387         atomic_set(&io_page->p_count, 1);
 388         get_page(page);
 389         set_page_writeback(page);
 390         ClearPageError(page);
 391
 392         for (bh = head = page_buffers(page), block_start = 0;
 393              bh != head || !block_start;
 394              block_start = block_end, bh = bh->b_this_page) {
 395
 396                 block_end = block_start + blocksize;
 397                 if (block_start >= len) {
 398                         /*
 399                          * Comments copied from block_write_full_page_endio:
 400                          *
 401                          * The page straddles i_size.  It must be zeroed out on
 402                          * each and every writepage invocation because it may
 403                          * be mmapped.  "A file is mapped in multiples of the
 404                          * page size.  For a file that is not a multiple of
 405                          * the  page size, the remaining memory is zeroed when
 406                          * mapped, and writes to that region are not written
 407                          * out to the file."
 408                          */
 409                         zero_user_segment(page, block_start, block_end);
 410                         clear_buffer_dirty(bh);
 411                         set_buffer_uptodate(bh);
 412                         continue;
 413                 }
 414                 clear_buffer_dirty(bh);
 415                 ret = io_submit_add_bh(io, io_page, inode, wbc, bh);
 416                 if (ret) {
 417                         /*
 418                          * We only get here on ENOMEM.  Not much else
 419                          * we can do but mark the page as dirty, and
 420                          * better luck next time.
 421                          */
 422                         set_page_dirty(page);
 423                         break;
 424                 }
 425         }
 426         unlock_page(page);
 427         /*
 428          * If the page was truncated before we could do the writeback,
 429          * or we had a memory allocation error while trying to write
 430          * the first buffer head, we won't have submitted any pages for
 431          * I/O.  In that case we need to make sure we've cleared the
 432          * PageWriteback bit from the page to prevent the system from
 433          * wedging later on.
 434          */
 435         put_io_page(io_page);
 436         return ret;
 437 }