ext4: Invert lock ordering of page_lock and transaction start in delalloc
authorMingming Cao <cmm@us.ibm.com>
Fri, 11 Jul 2008 23:27:31 +0000 (19:27 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Fri, 11 Jul 2008 23:27:31 +0000 (19:27 -0400)
With the reverse locking, we need to start a transation before taking
the page lock, so in ext4_da_writepages() we need to break the write-out
into chunks, and restart the journal for each chunck to ensure the
write-out fits in a single transaction.

Updated patch from Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
which fixes delalloc sync hang with journal lock inversion, and address
the performance regression issue.

Signed-off-by: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/extents.c
fs/ext4/inode.c

index dabc3b68d2498ded2e980544cbb7e8f559755928..42c4c0c892ed9b442617207b00acbdc7be99249f 100644 (file)
@@ -2565,6 +2565,7 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
        int err = 0, depth, ret;
        unsigned long allocated = 0;
        struct ext4_allocation_request ar;
+       loff_t disksize;
 
        __clear_bit(BH_New, &bh_result->b_state);
        ext_debug("blocks %u/%lu requested for inode %u\n",
@@ -2755,8 +2756,13 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
        newblock = ext_pblock(&newex);
        allocated = ext4_ext_get_actual_len(&newex);
 outnew:
-       if (extend_disksize && inode->i_size > EXT4_I(inode)->i_disksize)
-               EXT4_I(inode)->i_disksize = inode->i_size;
+       if (extend_disksize) {
+               disksize = ((loff_t) iblock + ar.len) << inode->i_blkbits;
+               if (disksize > i_size_read(inode))
+                       disksize = i_size_read(inode);
+               if (disksize > EXT4_I(inode)->i_disksize)
+                       EXT4_I(inode)->i_disksize = disksize;
+       }
 
        set_buffer_new(bh_result);
 
index a6b800c58474782e9db8d98adae518a215cf2b4c..7923336ecf94157935a8604301132c84bf17dae0 100644 (file)
@@ -847,6 +847,7 @@ int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
        struct ext4_inode_info *ei = EXT4_I(inode);
        int count = 0;
        ext4_fsblk_t first_block = 0;
+       loff_t disksize;
 
 
        J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
@@ -922,8 +923,13 @@ int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
         * protect it if you're about to implement concurrent
         * ext4_get_block() -bzzz
        */
-       if (!err && extend_disksize && inode->i_size > ei->i_disksize)
-               ei->i_disksize = inode->i_size;
+       if (!err && extend_disksize) {
+               disksize = ((loff_t) iblock + count) << inode->i_blkbits;
+               if (disksize > i_size_read(inode))
+                       disksize = i_size_read(inode);
+               if (disksize > ei->i_disksize)
+                       ei->i_disksize = disksize;
+       }
        if (err)
                goto cleanup;
 
@@ -1683,13 +1689,11 @@ static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
                        do {
                                if (cur_logical >= logical + blocks)
                                        break;
-
                                if (buffer_delay(bh)) {
                                        bh->b_blocknr = pblock;
                                        clear_buffer_delay(bh);
-                               } else if (buffer_mapped(bh)) {
+                               } else if (buffer_mapped(bh))
                                        BUG_ON(bh->b_blocknr != pblock);
-                               }
 
                                cur_logical++;
                                pblock++;
@@ -1764,10 +1768,10 @@ static void mpage_da_map_blocks(struct mpage_da_data *mpd)
                if (buffer_delay(lbh))
                        mpage_put_bnr_to_bhs(mpd, next, &new);
 
-                       /* go for the remaining blocks */
-                       next += new.b_size >> mpd->inode->i_blkbits;
-                       remain -= new.b_size;
-               }
+               /* go for the remaining blocks */
+               next += new.b_size >> mpd->inode->i_blkbits;
+               remain -= new.b_size;
+       }
 }
 
 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | (1 << BH_Delay))
@@ -1993,18 +1997,14 @@ static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
                                   struct buffer_head *bh_result, int create)
 {
-       int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
+       int ret;
        unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
        loff_t disksize = EXT4_I(inode)->i_disksize;
        handle_t *handle = NULL;
 
-       if (create) {
-               handle = ext4_journal_start(inode, needed_blocks);
-               if (IS_ERR(handle)) {
-                       ret = PTR_ERR(handle);
-                       goto out;
-               }
-       }
+       handle = ext4_journal_current_handle();
+       BUG_ON(handle == NULL);
+       BUG_ON(create == 0);
 
        ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
                                   bh_result, create, 0, EXT4_DELALLOC_RSVED);
@@ -2029,65 +2029,157 @@ static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
                        up_write(&EXT4_I(inode)->i_data_sem);
 
                        if (EXT4_I(inode)->i_disksize == disksize) {
-                               if (handle == NULL)
-                                       handle = ext4_journal_start(inode, 1);
-                               if (!IS_ERR(handle))
-                                       ext4_mark_inode_dirty(handle, inode);
+                               ret = ext4_mark_inode_dirty(handle, inode);
+                               return ret;
                        }
                }
-
                ret = 0;
        }
-
-out:
-       if (handle && !IS_ERR(handle))
-               ext4_journal_stop(handle);
-
        return ret;
 }
+
+static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
+{
+       return !buffer_mapped(bh) || buffer_delay(bh);
+}
+
 /* FIXME!! only support data=writeback mode */
+/*
+ * get called vi ext4_da_writepages after taking page lock
+ * We may end up doing block allocation here in case
+ * mpage_da_map_blocks failed to allocate blocks.
+ */
 static int ext4_da_writepage(struct page *page,
                                struct writeback_control *wbc)
 {
-       struct inode *inode = page->mapping->host;
-       handle_t *handle = NULL;
        int ret = 0;
-       int err;
+       loff_t size;
+       unsigned long len;
+       handle_t *handle = NULL;
+       struct buffer_head *page_bufs;
+       struct inode *inode = page->mapping->host;
 
-       if (ext4_journal_current_handle())
-               goto out_fail;
+       handle = ext4_journal_current_handle();
+       if (!handle) {
+               /*
+                * This can happen when we aren't called via
+                * ext4_da_writepages() but directly (shrink_page_list).
+                * We cannot easily start a transaction here so we just skip
+                * writing the page in case we would have to do so.
+                */
+               size = i_size_read(inode);
 
-       handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
-       if (IS_ERR(handle)) {
-               ret = PTR_ERR(handle);
-               goto out_fail;
+               page_bufs = page_buffers(page);
+               if (page->index == size >> PAGE_CACHE_SHIFT)
+                       len = size & ~PAGE_CACHE_MASK;
+               else
+                       len = PAGE_CACHE_SIZE;
+
+               if (walk_page_buffers(NULL, page_bufs, 0,
+                               len, NULL, ext4_bh_unmapped_or_delay)) {
+                       /*
+                        * We can't do block allocation under
+                        * page lock without a handle . So redirty
+                        * the page and return
+                        */
+                       BUG_ON(wbc->sync_mode != WB_SYNC_NONE);
+                       redirty_page_for_writepage(wbc, page);
+                       unlock_page(page);
+                       return 0;
+               }
        }
 
        if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
-               ret = nobh_writepage(page, ext4_get_block, wbc);
+               ret = nobh_writepage(page, ext4_da_get_block_write, wbc);
        else
-               ret = block_write_full_page(page, ext4_get_block, wbc);
-
-       if (!ret && inode->i_size > EXT4_I(inode)->i_disksize) {
-               EXT4_I(inode)->i_disksize = inode->i_size;
-               ext4_mark_inode_dirty(handle, inode);
-       }
+               ret = block_write_full_page(page, ext4_da_get_block_write, wbc);
 
-       err = ext4_journal_stop(handle);
-       if (!ret)
-               ret = err;
-       return ret;
-
-out_fail:
-       redirty_page_for_writepage(wbc, page);
-       unlock_page(page);
        return ret;
 }
 
+
+/*
+ * For now just follow the DIO way to estimate the max credits
+ * needed to write out EXT4_MAX_WRITEBACK_PAGES.
+ * todo: need to calculate the max credits need for
+ * extent based files, currently the DIO credits is based on
+ * indirect-blocks mapping way.
+ *
+ * Probably should have a generic way to calculate credits
+ * for DIO, writepages, and truncate
+ */
+#define EXT4_MAX_WRITEBACK_PAGES      DIO_MAX_BLOCKS
+#define EXT4_MAX_WRITEBACK_CREDITS    DIO_CREDITS
+
 static int ext4_da_writepages(struct address_space *mapping,
                                struct writeback_control *wbc)
 {
-       return mpage_da_writepages(mapping, wbc, ext4_da_get_block_write);
+       struct inode *inode = mapping->host;
+       handle_t *handle = NULL;
+       int needed_blocks;
+       int ret = 0;
+       long to_write;
+       loff_t range_start = 0;
+
+       /*
+        * No pages to write? This is mainly a kludge to avoid starting
+        * a transaction for special inodes like journal inode on last iput()
+        * because that could violate lock ordering on umount
+        */
+       if (!mapping->nrpages)
+               return 0;
+
+       /*
+        *  Estimate the worse case needed credits to write out
+        * EXT4_MAX_BUF_BLOCKS pages
+        */
+       needed_blocks = EXT4_MAX_WRITEBACK_CREDITS;
+
+       to_write = wbc->nr_to_write;
+       if (!wbc->range_cyclic) {
+               /*
+                * If range_cyclic is not set force range_cont
+                * and save the old writeback_index
+                */
+               wbc->range_cont = 1;
+               range_start =  wbc->range_start;
+       }
+
+       while (!ret && to_write) {
+               /* start a new transaction*/
+               handle = ext4_journal_start(inode, needed_blocks);
+               if (IS_ERR(handle)) {
+                       ret = PTR_ERR(handle);
+                       goto out_writepages;
+               }
+               /*
+                * set the max dirty pages could be write at a time
+                * to fit into the reserved transaction credits
+                */
+               if (wbc->nr_to_write > EXT4_MAX_WRITEBACK_PAGES)
+                       wbc->nr_to_write = EXT4_MAX_WRITEBACK_PAGES;
+
+               to_write -= wbc->nr_to_write;
+               ret = mpage_da_writepages(mapping, wbc,
+                                               ext4_da_get_block_write);
+               ext4_journal_stop(handle);
+               if (wbc->nr_to_write) {
+                       /*
+                        * There is no more writeout needed
+                        * or we requested for a noblocking writeout
+                        * and we found the device congested
+                        */
+                       to_write += wbc->nr_to_write;
+                       break;
+               }
+               wbc->nr_to_write = to_write;
+       }
+
+out_writepages:
+       wbc->nr_to_write = to_write;
+       if (range_start)
+               wbc->range_start = range_start;
+       return ret;
 }
 
 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
@@ -2137,11 +2229,6 @@ out:
        return ret;
 }
 
-static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
-{
-       return !buffer_mapped(bh) || buffer_delay(bh);
-}
-
 static int ext4_da_write_end(struct file *file,
                                struct address_space *mapping,
                                loff_t pos, unsigned len, unsigned copied,