]> git.karo-electronics.de Git - linux-beck.git/commitdiff
ext4: Save stack space by removing fake buffer heads
authorTheodore Ts'o <tytso@mit.edu>
Mon, 23 Feb 2009 11:46:01 +0000 (06:46 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Mon, 23 Feb 2009 11:46:01 +0000 (06:46 -0500)
Struct mpage_da_data and mpage_add_bh_to_extent() use a fake struct
buffer_head which is 104 bytes on an x86_64 system, but only use 24
bytes of the structure.  On systems that use a spinlock for atomic_t,
the stack savings will be even greater.

It turns out that using a fake struct buffer_head doesn't even save
that much code, and it makes the code more confusing since it's not
used as a "real" buffer head.  So just store pass b_size and b_state
in mpage_add_bh_to_extent(), and store b_size, b_state, and b_block_nr
in the mpage_da_data structure.

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/inode.c

index 924ba8afc227c86b90eb7ce071763ee065d538b3..008b28a859d0e2bf0ae55b4b0a2543568effcdf3 100644 (file)
@@ -1703,7 +1703,9 @@ static void ext4_da_page_release_reservation(struct page *page,
 
 struct mpage_da_data {
        struct inode *inode;
-       struct buffer_head lbh;                 /* extent of blocks */
+       sector_t b_blocknr;             /* start block number of extent */
+       size_t b_size;                  /* size of extent */
+       unsigned long b_state;          /* state of the extent */
        unsigned long first_page, next_page;    /* extent of pages */
        struct writeback_control *wbc;
        int io_done;
@@ -1737,7 +1739,7 @@ static int mpage_da_submit_io(struct mpage_da_data *mpd)
        /*
         * We need to start from the first_page to the next_page - 1
         * to make sure we also write the mapped dirty buffer_heads.
-        * If we look at mpd->lbh.b_blocknr we would only be looking
+        * If we look at mpd->b_blocknr we would only be looking
         * at the currently mapped buffer_heads.
         */
        index = mpd->first_page;
@@ -1975,7 +1977,7 @@ static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
 /*
  * mpage_da_map_blocks - go through given space
  *
- * @mpd->lbh - bh describing space
+ * @mpd - bh describing space
  *
  * The function skips space we know is already mapped to disk blocks.
  *
@@ -1984,18 +1986,18 @@ static int mpage_da_map_blocks(struct mpage_da_data *mpd)
 {
        int err = 0;
        struct buffer_head new;
-       struct buffer_head *lbh = &mpd->lbh;
        sector_t next;
 
        /*
         * We consider only non-mapped and non-allocated blocks
         */
-       if (buffer_mapped(lbh) && !buffer_delay(lbh))
+       if ((mpd->b_state  & (1 << BH_Mapped)) &&
+           !(mpd->b_state & (1 << BH_Delay)))
                return 0;
-       new.b_state = lbh->b_state;
+       new.b_state = mpd->b_state;
        new.b_blocknr = 0;
-       new.b_size = lbh->b_size;
-       next = lbh->b_blocknr;
+       new.b_size = mpd->b_size;
+       next = mpd->b_blocknr;
        /*
         * If we didn't accumulate anything
         * to write simply return
@@ -2031,7 +2033,7 @@ static int mpage_da_map_blocks(struct mpage_da_data *mpd)
                                  "%zd with error %d\n",
                                  __func__, mpd->inode->i_ino,
                                  (unsigned long long)next,
-                                 lbh->b_size >> mpd->inode->i_blkbits, err);
+                                 mpd->b_size >> mpd->inode->i_blkbits, err);
                printk(KERN_EMERG "This should not happen.!! "
                                        "Data will be lost\n");
                if (err == -ENOSPC) {
@@ -2039,7 +2041,7 @@ static int mpage_da_map_blocks(struct mpage_da_data *mpd)
                }
                /* invlaidate all the pages */
                ext4_da_block_invalidatepages(mpd, next,
-                               lbh->b_size >> mpd->inode->i_blkbits);
+                               mpd->b_size >> mpd->inode->i_blkbits);
                return err;
        }
        BUG_ON(new.b_size == 0);
@@ -2051,7 +2053,8 @@ static int mpage_da_map_blocks(struct mpage_da_data *mpd)
         * If blocks are delayed marked, we need to
         * put actual blocknr and drop delayed bit
         */
-       if (buffer_delay(lbh) || buffer_unwritten(lbh))
+       if ((mpd->b_state & (1 << BH_Delay)) ||
+           (mpd->b_state & (1 << BH_Unwritten)))
                mpage_put_bnr_to_bhs(mpd, next, &new);
 
        return 0;
@@ -2070,12 +2073,11 @@ static int mpage_da_map_blocks(struct mpage_da_data *mpd)
  * the function is used to collect contig. blocks in same state
  */
 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
-                                  sector_t logical, struct buffer_head *bh)
+                                  sector_t logical, size_t b_size,
+                                  unsigned long b_state)
 {
        sector_t next;
-       size_t b_size = bh->b_size;
-       struct buffer_head *lbh = &mpd->lbh;
-       int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
+       int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
 
        /* check if thereserved journal credits might overflow */
        if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
@@ -2102,19 +2104,19 @@ static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
        /*
         * First block in the extent
         */
-       if (lbh->b_size == 0) {
-               lbh->b_blocknr = logical;
-               lbh->b_size = b_size;
-               lbh->b_state = bh->b_state & BH_FLAGS;
+       if (mpd->b_size == 0) {
+               mpd->b_blocknr = logical;
+               mpd->b_size = b_size;
+               mpd->b_state = b_state & BH_FLAGS;
                return;
        }
 
-       next = lbh->b_blocknr + nrblocks;
+       next = mpd->b_blocknr + nrblocks;
        /*
         * Can we merge the block to our big extent?
         */
-       if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
-               lbh->b_size += b_size;
+       if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
+               mpd->b_size += b_size;
                return;
        }
 
@@ -2143,7 +2145,7 @@ static int __mpage_da_writepage(struct page *page,
 {
        struct mpage_da_data *mpd = data;
        struct inode *inode = mpd->inode;
-       struct buffer_head *bh, *head, fake;
+       struct buffer_head *bh, *head;
        sector_t logical;
 
        if (mpd->io_done) {
@@ -2185,9 +2187,9 @@ static int __mpage_da_writepage(struct page *page,
                /*
                 * ... and blocks
                 */
-               mpd->lbh.b_size = 0;
-               mpd->lbh.b_state = 0;
-               mpd->lbh.b_blocknr = 0;
+               mpd->b_size = 0;
+               mpd->b_state = 0;
+               mpd->b_blocknr = 0;
        }
 
        mpd->next_page = page->index + 1;
@@ -2195,16 +2197,8 @@ static int __mpage_da_writepage(struct page *page,
                  (PAGE_CACHE_SHIFT - inode->i_blkbits);
 
        if (!page_has_buffers(page)) {
-               /*
-                * There is no attached buffer heads yet (mmap?)
-                * we treat the page asfull of dirty blocks
-                */
-               bh = &fake;
-               bh->b_size = PAGE_CACHE_SIZE;
-               bh->b_state = 0;
-               set_buffer_dirty(bh);
-               set_buffer_uptodate(bh);
-               mpage_add_bh_to_extent(mpd, logical, bh);
+               mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
+                                      (1 << BH_Dirty) | (1 << BH_Uptodate));
                if (mpd->io_done)
                        return MPAGE_DA_EXTENT_TAIL;
        } else {
@@ -2222,8 +2216,10 @@ static int __mpage_da_writepage(struct page *page,
                         * with the page in ext4_da_writepage
                         */
                        if (buffer_dirty(bh) &&
-                               (!buffer_mapped(bh) || buffer_delay(bh))) {
-                               mpage_add_bh_to_extent(mpd, logical, bh);
+                           (!buffer_mapped(bh) || buffer_delay(bh))) {
+                               mpage_add_bh_to_extent(mpd, logical,
+                                                      bh->b_size,
+                                                      bh->b_state);
                                if (mpd->io_done)
                                        return MPAGE_DA_EXTENT_TAIL;
                        } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
@@ -2235,9 +2231,8 @@ static int __mpage_da_writepage(struct page *page,
                                 * unmapped buffer_head later we need to
                                 * use the b_state flag of that buffer_head.
                                 */
-                               if (mpd->lbh.b_size == 0)
-                                       mpd->lbh.b_state =
-                                               bh->b_state & BH_FLAGS;
+                               if (mpd->b_size == 0)
+                                       mpd->b_state = bh->b_state & BH_FLAGS;
                        }
                        logical++;
                } while ((bh = bh->b_this_page) != head);
@@ -2263,9 +2258,9 @@ static int mpage_da_writepages(struct address_space *mapping,
 {
        int ret;
 
-       mpd->lbh.b_size = 0;
-       mpd->lbh.b_state = 0;
-       mpd->lbh.b_blocknr = 0;
+       mpd->b_size = 0;
+       mpd->b_state = 0;
+       mpd->b_blocknr = 0;
        mpd->first_page = 0;
        mpd->next_page = 0;
        mpd->io_done = 0;