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