block: make gendisk hold a reference to its queue
[linux-2.6.git] / fs / mpage.c
index e431cb3..fdfae9f 100644 (file)
@@ -6,7 +6,7 @@
  * Contains functions related to preparing and submitting BIOs which contain
  * multiple pagecache pages.
  *
- * 15May2002   akpm@zip.com.au
+ * 15May2002   Andrew Morton
  *             Initial version
  * 27Jun2002   axboe@suse.de
  *             use bio_add_page() to build bio's just the right size
@@ -16,6 +16,7 @@
 #include <linux/module.h>
 #include <linux/mm.h>
 #include <linux/kdev_t.h>
+#include <linux/gfp.h>
 #include <linux/bio.h>
 #include <linux/fs.h>
 #include <linux/buffer_head.h>
@@ -26,6 +27,7 @@
 #include <linux/writeback.h>
 #include <linux/backing-dev.h>
 #include <linux/pagevec.h>
+#include <linux/cleancache.h>
 
 /*
  * I/O completion handler for multipage BIOs.
  * status of that page is hard.  See end_buffer_async_read() for the details.
  * There is no point in duplicating all that complexity.
  */
-static int mpage_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
+static void mpage_end_io(struct bio *bio, int err)
 {
        const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
        struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
 
-       if (bio->bi_size)
-               return 1;
-
-       do {
-               struct page *page = bvec->bv_page;
-
-               if (--bvec >= bio->bi_io_vec)
-                       prefetchw(&bvec->bv_page->flags);
-
-               if (uptodate) {
-                       SetPageUptodate(page);
-               } else {
-                       ClearPageUptodate(page);
-                       SetPageError(page);
-               }
-               unlock_page(page);
-       } while (bvec >= bio->bi_io_vec);
-       bio_put(bio);
-       return 0;
-}
-
-static int mpage_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
-{
-       const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
-       struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
-
-       if (bio->bi_size)
-               return 1;
-
        do {
                struct page *page = bvec->bv_page;
 
                if (--bvec >= bio->bi_io_vec)
                        prefetchw(&bvec->bv_page->flags);
-
-               if (!uptodate){
-                       SetPageError(page);
-                       if (page->mapping)
-                               set_bit(AS_EIO, &page->mapping->flags);
+               if (bio_data_dir(bio) == READ) {
+                       if (uptodate) {
+                               SetPageUptodate(page);
+                       } else {
+                               ClearPageUptodate(page);
+                               SetPageError(page);
+                       }
+                       unlock_page(page);
+               } else { /* bio_data_dir(bio) == WRITE */
+                       if (!uptodate) {
+                               SetPageError(page);
+                               if (page->mapping)
+                                       set_bit(AS_EIO, &page->mapping->flags);
+                       }
+                       end_page_writeback(page);
                }
-               end_page_writeback(page);
        } while (bvec >= bio->bi_io_vec);
        bio_put(bio);
-       return 0;
 }
 
 static struct bio *mpage_bio_submit(int rw, struct bio *bio)
 {
-       bio->bi_end_io = mpage_end_io_read;
-       if (rw == WRITE)
-               bio->bi_end_io = mpage_end_io_write;
+       bio->bi_end_io = mpage_end_io;
        submit_bio(rw, bio);
        return NULL;
 }
@@ -163,9 +142,19 @@ map_buffer_to_page(struct page *page, struct buffer_head *bh, int page_block)
        } while (page_bh != head);
 }
 
+/*
+ * This is the worker routine which does all the work of mapping the disk
+ * blocks and constructs largest possible bios, submits them for IO if the
+ * blocks are not contiguous on the disk.
+ *
+ * We pass a buffer_head back and forth and use its buffer_mapped() flag to
+ * represent the validity of its disk mapping and to decide when to do the next
+ * get_block() call.
+ */
 static struct bio *
 do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
-                       sector_t *last_block_in_bio, get_block_t get_block)
+               sector_t *last_block_in_bio, struct buffer_head *map_bh,
+               unsigned long *first_logical_block, get_block_t get_block)
 {
        struct inode *inode = page->mapping->host;
        const unsigned blkbits = inode->i_blkbits;
@@ -173,33 +162,71 @@ do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
        const unsigned blocksize = 1 << blkbits;
        sector_t block_in_file;
        sector_t last_block;
+       sector_t last_block_in_file;
        sector_t blocks[MAX_BUF_PER_PAGE];
        unsigned page_block;
        unsigned first_hole = blocks_per_page;
        struct block_device *bdev = NULL;
-       struct buffer_head bh;
        int length;
        int fully_mapped = 1;
+       unsigned nblocks;
+       unsigned relative_block;
 
        if (page_has_buffers(page))
                goto confused;
 
        block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
-       last_block = (i_size_read(inode) + blocksize - 1) >> blkbits;
+       last_block = block_in_file + nr_pages * blocks_per_page;
+       last_block_in_file = (i_size_read(inode) + blocksize - 1) >> blkbits;
+       if (last_block > last_block_in_file)
+               last_block = last_block_in_file;
+       page_block = 0;
+
+       /*
+        * Map blocks using the result from the previous get_blocks call first.
+        */
+       nblocks = map_bh->b_size >> blkbits;
+       if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
+                       block_in_file < (*first_logical_block + nblocks)) {
+               unsigned map_offset = block_in_file - *first_logical_block;
+               unsigned last = nblocks - map_offset;
+
+               for (relative_block = 0; ; relative_block++) {
+                       if (relative_block == last) {
+                               clear_buffer_mapped(map_bh);
+                               break;
+                       }
+                       if (page_block == blocks_per_page)
+                               break;
+                       blocks[page_block] = map_bh->b_blocknr + map_offset +
+                                               relative_block;
+                       page_block++;
+                       block_in_file++;
+               }
+               bdev = map_bh->b_bdev;
+       }
+
+       /*
+        * Then do more get_blocks calls until we are done with this page.
+        */
+       map_bh->b_page = page;
+       while (page_block < blocks_per_page) {
+               map_bh->b_state = 0;
+               map_bh->b_size = 0;
 
-       bh.b_page = page;
-       for (page_block = 0; page_block < blocks_per_page;
-                               page_block++, block_in_file++) {
-               bh.b_state = 0;
                if (block_in_file < last_block) {
-                       if (get_block(inode, block_in_file, &bh, 0))
+                       map_bh->b_size = (last_block-block_in_file) << blkbits;
+                       if (get_block(inode, block_in_file, map_bh, 0))
                                goto confused;
+                       *first_logical_block = block_in_file;
                }
 
-               if (!buffer_mapped(&bh)) {
+               if (!buffer_mapped(map_bh)) {
                        fully_mapped = 0;
                        if (first_hole == blocks_per_page)
                                first_hole = page_block;
+                       page_block++;
+                       block_in_file++;
                        continue;
                }
 
@@ -209,8 +236,8 @@ do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
                 * we just collected from get_block into the page's buffers
                 * so readpage doesn't have to repeat the get_block call
                 */
-               if (buffer_uptodate(&bh)) {
-                       map_buffer_to_page(page, &bh, page_block);
+               if (buffer_uptodate(map_bh)) {
+                       map_buffer_to_page(page, map_bh, page_block);
                        goto confused;
                }
        
@@ -218,18 +245,24 @@ do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
                        goto confused;          /* hole -> non-hole */
 
                /* Contiguous blocks? */
-               if (page_block && blocks[page_block-1] != bh.b_blocknr-1)
+               if (page_block && blocks[page_block-1] != map_bh->b_blocknr-1)
                        goto confused;
-               blocks[page_block] = bh.b_blocknr;
-               bdev = bh.b_bdev;
+               nblocks = map_bh->b_size >> blkbits;
+               for (relative_block = 0; ; relative_block++) {
+                       if (relative_block == nblocks) {
+                               clear_buffer_mapped(map_bh);
+                               break;
+                       } else if (page_block == blocks_per_page)
+                               break;
+                       blocks[page_block] = map_bh->b_blocknr+relative_block;
+                       page_block++;
+                       block_in_file++;
+               }
+               bdev = map_bh->b_bdev;
        }
 
        if (first_hole != blocks_per_page) {
-               char *kaddr = kmap_atomic(page, KM_USER0);
-               memset(kaddr + (first_hole << blkbits), 0,
-                               PAGE_CACHE_SIZE - (first_hole << blkbits));
-               flush_dcache_page(page);
-               kunmap_atomic(kaddr, KM_USER0);
+               zero_user_segment(page, first_hole << blkbits, PAGE_CACHE_SIZE);
                if (first_hole == 0) {
                        SetPageUptodate(page);
                        unlock_page(page);
@@ -239,6 +272,12 @@ do_mpage_readpage(struct bio *bio, struct page *page, unsigned nr_pages,
                SetPageMappedToDisk(page);
        }
 
+       if (fully_mapped && blocks_per_page == 1 && !PageUptodate(page) &&
+           cleancache_get_page(page) == 0) {
+               SetPageUptodate(page);
+               goto confused;
+       }
+
        /*
         * This page will go to BIO.  Do we need to send this BIO off first?
         */
@@ -260,7 +299,10 @@ alloc_new:
                goto alloc_new;
        }
 
-       if (buffer_boundary(&bh) || (first_hole != blocks_per_page))
+       relative_block = block_in_file - *first_logical_block;
+       nblocks = map_bh->b_size >> blkbits;
+       if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
+           (first_hole != blocks_per_page))
                bio = mpage_bio_submit(READ, bio);
        else
                *last_block_in_bio = blocks[blocks_per_page - 1];
@@ -278,16 +320,12 @@ confused:
 }
 
 /**
- * mpage_readpages - populate an address space with some pages, and
- *                       start reads against them.
- *
+ * mpage_readpages - populate an address space with some pages & start reads against them
  * @mapping: the address_space
  * @pages: The address of a list_head which contains the target pages.  These
  *   pages have their ->index populated and are otherwise uninitialised.
- *
  *   The page at @pages->prev has the lowest file offset, and reads should be
  *   issued in @pages->prev to @pages->next order.
- *
  * @nr_pages: The number of pages at *@pages
  * @get_block: The filesystem's block mapper function.
  *
@@ -313,6 +351,7 @@ confused:
  * So an mpage read of the first 16 blocks of an ext2 file will cause I/O to be
  * submitted in the following order:
  *     12 0 1 2 3 4 5 6 7 8 9 10 11 13 14 15 16
+ *
  * because the indirect block has to be read to get the mappings of blocks
  * 13,14,15,16.  Obviously, this impacts performance.
  *
@@ -330,29 +369,33 @@ mpage_readpages(struct address_space *mapping, struct list_head *pages,
        struct bio *bio = NULL;
        unsigned page_idx;
        sector_t last_block_in_bio = 0;
-       struct pagevec lru_pvec;
+       struct buffer_head map_bh;
+       unsigned long first_logical_block = 0;
+       struct blk_plug plug;
 
-       pagevec_init(&lru_pvec, 0);
+       blk_start_plug(&plug);
+
+       map_bh.b_state = 0;
+       map_bh.b_size = 0;
        for (page_idx = 0; page_idx < nr_pages; page_idx++) {
                struct page *page = list_entry(pages->prev, struct page, lru);
 
                prefetchw(&page->flags);
                list_del(&page->lru);
-               if (!add_to_page_cache(page, mapping,
+               if (!add_to_page_cache_lru(page, mapping,
                                        page->index, GFP_KERNEL)) {
                        bio = do_mpage_readpage(bio, page,
                                        nr_pages - page_idx,
-                                       &last_block_in_bio, get_block);
-                       if (!pagevec_add(&lru_pvec, page))
-                               __pagevec_lru_add(&lru_pvec);
-               } else {
-                       page_cache_release(page);
+                                       &last_block_in_bio, &map_bh,
+                                       &first_logical_block,
+                                       get_block);
                }
+               page_cache_release(page);
        }
-       pagevec_lru_add(&lru_pvec);
        BUG_ON(!list_empty(pages));
        if (bio)
                mpage_bio_submit(READ, bio);
+       blk_finish_plug(&plug);
        return 0;
 }
 EXPORT_SYMBOL(mpage_readpages);
@@ -364,9 +407,13 @@ int mpage_readpage(struct page *page, get_block_t get_block)
 {
        struct bio *bio = NULL;
        sector_t last_block_in_bio = 0;
+       struct buffer_head map_bh;
+       unsigned long first_logical_block = 0;
 
-       bio = do_mpage_readpage(bio, page, 1,
-                       &last_block_in_bio, get_block);
+       map_bh.b_state = 0;
+       map_bh.b_size = 0;
+       bio = do_mpage_readpage(bio, page, 1, &last_block_in_bio,
+                       &map_bh, &first_logical_block, get_block);
        if (bio)
                mpage_bio_submit(READ, bio);
        return 0;
@@ -389,11 +436,19 @@ EXPORT_SYMBOL(mpage_readpage);
  * written, so it can intelligently allocate a suitably-sized BIO.  For now,
  * just allocate full-size (16-page) BIOs.
  */
-static struct bio *
-__mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
-       sector_t *last_block_in_bio, int *ret, struct writeback_control *wbc,
-       writepage_t writepage_fn)
+
+struct mpage_data {
+       struct bio *bio;
+       sector_t last_block_in_bio;
+       get_block_t *get_block;
+       unsigned use_writepage;
+};
+
+static int __mpage_writepage(struct page *page, struct writeback_control *wbc,
+                     void *data)
 {
+       struct mpage_data *mpd = data;
+       struct bio *bio = mpd->bio;
        struct address_space *mapping = page->mapping;
        struct inode *inode = page->mapping->host;
        const unsigned blkbits = inode->i_blkbits;
@@ -411,6 +466,7 @@ __mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
        int length;
        struct buffer_head map_bh;
        loff_t i_size = i_size_read(inode);
+       int ret = 0;
 
        if (page_has_buffers(page)) {
                struct buffer_head *head = page_buffers(page);
@@ -472,7 +528,8 @@ __mpage_writepage(struct bio *bio, struct page *page, get_block_t get_block,
        for (page_block = 0; page_block < blocks_per_page; ) {
 
                map_bh.b_state = 0;
-               if (get_block(inode, block_in_file, &map_bh, 1))
+               map_bh.b_size = 1 << blkbits;
+               if (mpd->get_block(inode, block_in_file, &map_bh, 1))
                        goto confused;
                if (buffer_new(&map_bh))
                        unmap_underlying_metadata(map_bh.b_bdev,
@@ -501,27 +558,23 @@ page_is_mapped:
        if (page->index >= end_index) {
                /*
                 * The page straddles i_size.  It must be zeroed out on each
-                * and every writepage invokation because it may be mmapped.
+                * and every writepage invocation because it may be mmapped.
                 * "A file is mapped in multiples of the page size.  For a file
                 * that is not a multiple of the page size, the remaining memory
                 * is zeroed when mapped, and writes to that region are not
                 * written out to the file."
                 */
                unsigned offset = i_size & (PAGE_CACHE_SIZE - 1);
-               char *kaddr;
 
                if (page->index > end_index || !offset)
                        goto confused;
-               kaddr = kmap_atomic(page, KM_USER0);
-               memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
-               flush_dcache_page(page);
-               kunmap_atomic(kaddr, KM_USER0);
+               zero_user_segment(page, offset, PAGE_CACHE_SIZE);
        }
 
        /*
         * This page will go to BIO.  Do we need to send this BIO off first?
         */
-       if (bio && *last_block_in_bio != blocks[0] - 1)
+       if (bio && mpd->last_block_in_bio != blocks[0] - 1)
                bio = mpage_bio_submit(WRITE, bio);
 
 alloc_new:
@@ -578,7 +631,7 @@ alloc_new:
                                        boundary_block, 1 << blkbits);
                }
        } else {
-               *last_block_in_bio = blocks[blocks_per_page - 1];
+               mpd->last_block_in_bio = blocks[blocks_per_page - 1];
        }
        goto out;
 
@@ -586,29 +639,23 @@ confused:
        if (bio)
                bio = mpage_bio_submit(WRITE, bio);
 
-       if (writepage_fn) {
-               *ret = (*writepage_fn)(page, wbc);
+       if (mpd->use_writepage) {
+               ret = mapping->a_ops->writepage(page, wbc);
        } else {
-               *ret = -EAGAIN;
+               ret = -EAGAIN;
                goto out;
        }
        /*
         * The caller has a ref on the inode, so *mapping is stable
         */
-       if (*ret) {
-               if (*ret == -ENOSPC)
-                       set_bit(AS_ENOSPC, &mapping->flags);
-               else
-                       set_bit(AS_EIO, &mapping->flags);
-       }
+       mapping_set_error(mapping, ret);
 out:
-       return bio;
+       mpd->bio = bio;
+       return ret;
 }
 
 /**
- * mpage_writepages - walk the list of dirty pages of the given
- * address space and writepage() all of them.
- * 
+ * mpage_writepages - walk the list of dirty pages of the given address space & writepage() all of them
  * @mapping: address space structure to write
  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
  * @get_block: the filesystem's block mapper function.
@@ -630,122 +677,26 @@ int
 mpage_writepages(struct address_space *mapping,
                struct writeback_control *wbc, get_block_t get_block)
 {
-       struct backing_dev_info *bdi = mapping->backing_dev_info;
-       struct bio *bio = NULL;
-       sector_t last_block_in_bio = 0;
-       int ret = 0;
-       int done = 0;
-       int (*writepage)(struct page *page, struct writeback_control *wbc);
-       struct pagevec pvec;
-       int nr_pages;
-       pgoff_t index;
-       pgoff_t end = -1;               /* Inclusive */
-       int scanned = 0;
-       int is_range = 0;
-
-       if (wbc->nonblocking && bdi_write_congested(bdi)) {
-               wbc->encountered_congestion = 1;
-               return 0;
-       }
-
-       writepage = NULL;
-       if (get_block == NULL)
-               writepage = mapping->a_ops->writepage;
-
-       pagevec_init(&pvec, 0);
-       if (wbc->sync_mode == WB_SYNC_NONE) {
-               index = mapping->writeback_index; /* Start from prev offset */
-       } else {
-               index = 0;                        /* whole-file sweep */
-               scanned = 1;
-       }
-       if (wbc->start || wbc->end) {
-               index = wbc->start >> PAGE_CACHE_SHIFT;
-               end = wbc->end >> PAGE_CACHE_SHIFT;
-               is_range = 1;
-               scanned = 1;
-       }
-retry:
-       while (!done && (index <= end) &&
-                       (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
-                       PAGECACHE_TAG_DIRTY,
-                       min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
-               unsigned i;
-
-               scanned = 1;
-               for (i = 0; i < nr_pages; i++) {
-                       struct page *page = pvec.pages[i];
-
-                       /*
-                        * At this point we hold neither mapping->tree_lock nor
-                        * lock on the page itself: the page may be truncated or
-                        * invalidated (changing page->mapping to NULL), or even
-                        * swizzled back from swapper_space to tmpfs file
-                        * mapping
-                        */
-
-                       lock_page(page);
-
-                       if (unlikely(page->mapping != mapping)) {
-                               unlock_page(page);
-                               continue;
-                       }
-
-                       if (unlikely(is_range) && page->index > end) {
-                               done = 1;
-                               unlock_page(page);
-                               continue;
-                       }
-
-                       if (wbc->sync_mode != WB_SYNC_NONE)
-                               wait_on_page_writeback(page);
-
-                       if (PageWriteback(page) ||
-                                       !clear_page_dirty_for_io(page)) {
-                               unlock_page(page);
-                               continue;
-                       }
-
-                       if (writepage) {
-                               ret = (*writepage)(page, wbc);
-                               if (ret) {
-                                       if (ret == -ENOSPC)
-                                               set_bit(AS_ENOSPC,
-                                                       &mapping->flags);
-                                       else
-                                               set_bit(AS_EIO,
-                                                       &mapping->flags);
-                               }
-                       } else {
-                               bio = __mpage_writepage(bio, page, get_block,
-                                               &last_block_in_bio, &ret, wbc,
-                                               page->mapping->a_ops->writepage);
-                       }
-                       if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
-                               unlock_page(page);
-                       if (ret || (--(wbc->nr_to_write) <= 0))
-                               done = 1;
-                       if (wbc->nonblocking && bdi_write_congested(bdi)) {
-                               wbc->encountered_congestion = 1;
-                               done = 1;
-                       }
-               }
-               pagevec_release(&pvec);
-               cond_resched();
-       }
-       if (!scanned && !done) {
-               /*
-                * We hit the last page and there is more work to be done: wrap
-                * back to the start of the file
-                */
-               scanned = 1;
-               index = 0;
-               goto retry;
+       struct blk_plug plug;
+       int ret;
+
+       blk_start_plug(&plug);
+
+       if (!get_block)
+               ret = generic_writepages(mapping, wbc);
+       else {
+               struct mpage_data mpd = {
+                       .bio = NULL,
+                       .last_block_in_bio = 0,
+                       .get_block = get_block,
+                       .use_writepage = 1,
+               };
+
+               ret = write_cache_pages(mapping, wbc, __mpage_writepage, &mpd);
+               if (mpd.bio)
+                       mpage_bio_submit(WRITE, mpd.bio);
        }
-       if (!is_range)
-               mapping->writeback_index = index;
-       if (bio)
-               mpage_bio_submit(WRITE, bio);
+       blk_finish_plug(&plug);
        return ret;
 }
 EXPORT_SYMBOL(mpage_writepages);
@@ -753,15 +704,15 @@ EXPORT_SYMBOL(mpage_writepages);
 int mpage_writepage(struct page *page, get_block_t get_block,
        struct writeback_control *wbc)
 {
-       int ret = 0;
-       struct bio *bio;
-       sector_t last_block_in_bio = 0;
-
-       bio = __mpage_writepage(NULL, page, get_block,
-                       &last_block_in_bio, &ret, wbc, NULL);
-       if (bio)
-               mpage_bio_submit(WRITE, bio);
-
+       struct mpage_data mpd = {
+               .bio = NULL,
+               .last_block_in_bio = 0,
+               .get_block = get_block,
+               .use_writepage = 0,
+       };
+       int ret = __mpage_writepage(page, wbc, &mpd);
+       if (mpd.bio)
+               mpage_bio_submit(WRITE, mpd.bio);
        return ret;
 }
 EXPORT_SYMBOL(mpage_writepage);