From: Valerie Clement Subject: [PATCH] ext4: mballoc: fix mb_normalize_request algorithm for 1KB block size filesystems Date: Wed, 30 Apr 2008 15:41:10 +0200 Message-ID: <1209562870.5307.12.camel@ext1.frec.bull.fr> Mime-Version: 1.0 Content-Type: text/plain Content-Transfer-Encoding: 7bit To: linux-ext4 , Aneesh Kumar , sandeen@redhat.com Return-path: Received: from ecfrec.frec.bull.fr ([129.183.4.8]:58710 "EHLO ecfrec.frec.bull.fr" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755586AbYD3NcT (ORCPT ); Wed, 30 Apr 2008 09:32:19 -0400 Sender: linux-ext4-owner@vger.kernel.org List-ID: mballoc: fix mb_normalize_request algorithm for 1KB block size filesystems From: Valerie Clement In case of inode preallocation, the number of blocks to allocate depends on the file size and it is calculated in ext4_mb_normalize_group_request(). Each group in the filesystem is then checked to find one that can be used for allocation; this is done in ext4_mb_good_group(). When a file bigger than 4MB is created, the requested number of blocks to preallocate, calculated by ext4_mb_normalize_group_request is 4096. However for a filesystem with 1KB block size, the maximum size of the block buddies used by the multiblock allocator is 2048, so none of groups in the filesystem satisfies the search criteria in ext4_mb_good_group(). Scanning all the filesystem groups impacts performance. The following numbers show that: - on an ext4 FS with 1KB block size mounted with nodelalloc option: # dd if=/dev/zero of=/mnt/test/foo bs=8k count=1k conv=fsync 1024+0 records in 1024+0 records out 8388608 bytes (8.4 MB) copied, 35.5091 seconds, 236 kB/s - on an ext4 FS with 1KB block size mounted with nodelalloc and nomballoc options: # dd if=/dev/zero of=/mnt/test/foo bs=8k count=1k conv=fsync 1024+0 records in 1024+0 records out 8388608 bytes (8.4 MB) copied, 0.233754 seconds, 35.9 MB/s In the two cases, dd is done after creating the FS with -b1024 option, mounting the FS with the options specified before and flushing all caches using echo 3 > /proc/sys/vm/drop_caches. The partition size is 70GB. I did the same test on a 1TB partition, it took several minutes to write 8MB! This patch modifies the algorithm in ext4_mb_normalize_group_request to calculate the number of blocks to allocate by taking into account the maximum size of free blocks chunks handled by the multiblock allocator. It has also been tested for filesystems with 2KB and 4KB block sizes to ensure that those cases don't regress. Signed-off-by: Valerie Clement --- mballoc.c | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-) Index: linux-2.6.25/fs/ext4/mballoc.c =================================================================== --- linux-2.6.25.orig/fs/ext4/mballoc.c 2008-04-25 16:19:32.000000000 +0200 +++ linux-2.6.25/fs/ext4/mballoc.c 2008-04-25 16:49:34.000000000 +0200 @@ -2905,12 +2905,11 @@ ext4_mb_normalize_request(struct ext4_al if (size < i_size_read(ac->ac_inode)) size = i_size_read(ac->ac_inode); - /* max available blocks in a free group */ - max = EXT4_BLOCKS_PER_GROUP(ac->ac_sb) - 1 - 1 - - EXT4_SB(ac->ac_sb)->s_itb_per_group; + /* max size of free chunks */ + max = 2 << bsbits; -#define NRL_CHECK_SIZE(req, size, max,bits) \ - (req <= (size) || max <= ((size) >> bits)) +#define NRL_CHECK_SIZE(req, size, max, chunk_size) \ + (req <= (size) || max <= (chunk_size)) /* first, try to predict filesize */ /* XXX: should this table be tunable? */ @@ -2929,16 +2928,16 @@ ext4_mb_normalize_request(struct ext4_al size = 512 * 1024; } else if (size <= 1024 * 1024) { size = 1024 * 1024; - } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, bsbits)) { + } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { start_off = ((loff_t)ac->ac_o_ex.fe_logical >> - (20 - bsbits)) << 20; - size = 1024 * 1024; - } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, bsbits)) { + (21 - bsbits)) << 21; + size = 2* 1024 * 1024; + } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { start_off = ((loff_t)ac->ac_o_ex.fe_logical >> (22 - bsbits)) << 22; size = 4 * 1024 * 1024; } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, - (8<<20)>>bsbits, max, bsbits)) { + (8<<20)>>bsbits, max, 8 * 1024)) { start_off = ((loff_t)ac->ac_o_ex.fe_logical >> (23 - bsbits)) << 23; size = 8 * 1024 * 1024;