direct-io: separate map_bh from dio
Andi Kleen [Tue, 2 Aug 2011 04:38:07 +0000 (21:38 -0700)]
Only a single b_private field in the map_bh buffer head is needed after
the submission path. Move map_bh separately to avoid storing
this information in the long term slab.

This avoids the weird 104 byte hole in struct dio_submit which also needed
to be memseted early.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

fs/direct-io.c

index 6bb0440..edf3174 100644 (file)
@@ -119,6 +119,7 @@ struct dio {
        loff_t i_size;                  /* i_size when submitted */
        dio_iodone_t *end_io;           /* IO completion function */
 
+       void *private;                  /* copy from map_bh.b_private */
 
        /* BIO completion state */
        spinlock_t bio_lock;            /* protects BIO fields below */
@@ -133,7 +134,6 @@ struct dio {
        struct kiocb *iocb;             /* kiocb */
        ssize_t result;                 /* IO result */
 
-       struct buffer_head map_bh;      /* last get_block() result */
        /*
         * pages[] (and any fields placed after it) are not zeroed out at
         * allocation time.  Don't add new fields after pages[] unless you
@@ -301,7 +301,7 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is
 
        if (dio->end_io && dio->result) {
                dio->end_io(dio->iocb, offset, transferred,
-                           dio->map_bh.b_private, ret, is_async);
+                           dio->private, ret, is_async);
        } else {
                if (is_async)
                        aio_complete(dio->iocb, ret, 0);
@@ -574,10 +574,10 @@ static int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
  * buffer_mapped().  However the direct-io code will only process holes one
  * block at a time - it will repeatedly call get_block() as it walks the hole.
  */
-static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
+static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
+                          struct buffer_head *map_bh)
 {
        int ret;
-       struct buffer_head *map_bh = &dio->map_bh;
        sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
        unsigned long fs_count; /* Number of filesystem-sized blocks */
        unsigned long dio_count;/* Number of dio_block-sized blocks */
@@ -621,6 +621,9 @@ static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
 
                ret = (*sdio->get_block)(dio->inode, fs_startblk,
                                                map_bh, create);
+
+               /* Store for completion */
+               dio->private = map_bh->b_private;
        }
        return ret;
 }
@@ -629,7 +632,7 @@ static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
  * There is no bio.  Make one now.
  */
 static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
-                      sector_t start_sector)
+                      sector_t start_sector, struct buffer_head *map_bh)
 {
        sector_t sector;
        int ret, nr_pages;
@@ -638,10 +641,10 @@ static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
        if (ret)
                goto out;
        sector = start_sector << (sdio->blkbits - 9);
-       nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
+       nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
        nr_pages = min(nr_pages, BIO_MAX_PAGES);
        BUG_ON(nr_pages <= 0);
-       dio_bio_alloc(dio, sdio, dio->map_bh.b_bdev, sector, nr_pages);
+       dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
        sdio->boundary = 0;
 out:
        return ret;
@@ -686,7 +689,8 @@ static int dio_bio_add_page(struct dio_submit *sdio)
  * The caller of this function is responsible for removing cur_page from the
  * dio, and for dropping the refcount which came from that presence.
  */
-static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
+static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
+                            struct buffer_head *map_bh)
 {
        int ret = 0;
 
@@ -721,14 +725,14 @@ static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
        }
 
        if (sdio->bio == NULL) {
-               ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
+               ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
                if (ret)
                        goto out;
        }
 
        if (dio_bio_add_page(sdio) != 0) {
                dio_bio_submit(dio, sdio);
-               ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
+               ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
                if (ret == 0) {
                        ret = dio_bio_add_page(sdio);
                        BUG_ON(ret != 0);
@@ -757,7 +761,8 @@ out:
  */
 static int
 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
-               unsigned offset, unsigned len, sector_t blocknr)
+                   unsigned offset, unsigned len, sector_t blocknr,
+                   struct buffer_head *map_bh)
 {
        int ret = 0;
 
@@ -782,7 +787,7 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
                 * avoid metadata seeks.
                 */
                if (sdio->boundary) {
-                       ret = dio_send_cur_page(dio, sdio);
+                       ret = dio_send_cur_page(dio, sdio, map_bh);
                        page_cache_release(sdio->cur_page);
                        sdio->cur_page = NULL;
                }
@@ -793,7 +798,7 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
         * If there's a deferred page already there then send it.
         */
        if (sdio->cur_page) {
-               ret = dio_send_cur_page(dio, sdio);
+               ret = dio_send_cur_page(dio, sdio, map_bh);
                page_cache_release(sdio->cur_page);
                sdio->cur_page = NULL;
                if (ret)
@@ -815,16 +820,16 @@ out:
  * file blocks.  Only called for S_ISREG files - blockdevs do not set
  * buffer_new
  */
-static void clean_blockdev_aliases(struct dio *dio)
+static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
 {
        unsigned i;
        unsigned nblocks;
 
-       nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
+       nblocks = map_bh->b_size >> dio->inode->i_blkbits;
 
        for (i = 0; i < nblocks; i++) {
-               unmap_underlying_metadata(dio->map_bh.b_bdev,
-                                       dio->map_bh.b_blocknr + i);
+               unmap_underlying_metadata(map_bh->b_bdev,
+                                         map_bh->b_blocknr + i);
        }
 }
 
@@ -837,7 +842,8 @@ static void clean_blockdev_aliases(struct dio *dio)
  * `end' is zero if we're doing the start of the IO, 1 at the end of the
  * IO.
  */
-static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
+static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end,
+                          struct buffer_head *map_bh)
 {
        unsigned dio_blocks_per_fs_block;
        unsigned this_chunk_blocks;     /* In dio_blocks */
@@ -845,7 +851,7 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
        struct page *page;
 
        sdio->start_zero_done = 1;
-       if (!sdio->blkfactor || !buffer_new(&dio->map_bh))
+       if (!sdio->blkfactor || !buffer_new(map_bh))
                return;
 
        dio_blocks_per_fs_block = 1 << sdio->blkfactor;
@@ -865,7 +871,7 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
 
        page = ZERO_PAGE(0);
        if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
-                               sdio->next_block_for_io))
+                               sdio->next_block_for_io, map_bh))
                return;
 
        sdio->next_block_for_io += this_chunk_blocks;
@@ -887,13 +893,13 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
  * fine alignment but still allows this function to work in PAGE_SIZE units.
  */
-static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
+static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
+                       struct buffer_head *map_bh)
 {
        const unsigned blkbits = sdio->blkbits;
        const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
        struct page *page;
        unsigned block_in_page;
-       struct buffer_head *map_bh = &dio->map_bh;
        int ret = 0;
 
        /* The I/O can start at any block offset within the first page */
@@ -919,7 +925,7 @@ static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
                                unsigned long blkmask;
                                unsigned long dio_remainder;
 
-                               ret = get_more_blocks(dio, sdio);
+                               ret = get_more_blocks(dio, sdio, map_bh);
                                if (ret) {
                                        page_cache_release(page);
                                        goto out;
@@ -932,7 +938,7 @@ static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
                                sdio->next_block_for_io =
                                        map_bh->b_blocknr << sdio->blkfactor;
                                if (buffer_new(map_bh))
-                                       clean_blockdev_aliases(dio);
+                                       clean_blockdev_aliases(dio, map_bh);
 
                                if (!sdio->blkfactor)
                                        goto do_holes;
@@ -991,7 +997,7 @@ do_holes:
                         * we must zero out the start of this block.
                         */
                        if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
-                               dio_zero_block(dio, sdio, 0);
+                               dio_zero_block(dio, sdio, 0, map_bh);
 
                        /*
                         * Work out, in this_chunk_blocks, how much disk we
@@ -1011,7 +1017,8 @@ do_holes:
                        ret = submit_page_section(dio, sdio, page,
                                                  offset_in_page,
                                                  this_chunk_bytes,
-                                                 sdio->next_block_for_io);
+                                                 sdio->next_block_for_io,
+                                                 map_bh);
                        if (ret) {
                                page_cache_release(page);
                                goto out;
@@ -1047,6 +1054,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        ssize_t ret = 0;
        ssize_t ret2;
        size_t bytes;
+       struct buffer_head map_bh = { 0, };
 
        dio->inode = inode;
        dio->rw = rw;
@@ -1101,7 +1109,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
                sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
                sdio->curr_user_address = user_addr;
        
-               ret = do_direct_IO(dio, sdio);
+               ret = do_direct_IO(dio, sdio, &map_bh);
 
                dio->result += iov[seg].iov_len -
                        ((sdio->final_block_in_request - sdio->block_in_file) <<
@@ -1124,10 +1132,10 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
         * There may be some unwritten disk at the end of a part-written
         * fs-block-sized block.  Go zero that now.
         */
-       dio_zero_block(dio, sdio, 1);
+       dio_zero_block(dio, sdio, 1, &map_bh);
 
        if (sdio->cur_page) {
-               ret2 = dio_send_cur_page(dio, sdio);
+               ret2 = dio_send_cur_page(dio, sdio, &map_bh);
                if (ret == 0)
                        ret = ret2;
                page_cache_release(sdio->cur_page);