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