| Message ID | 20220520111402.4252-1-jack@suse.cz |
|---|---|
| State | Awaiting Upstream |
| Headers | show |
| Series | ext4: Improve write performance with disabled delalloc | expand |
On Fri, 20 May 2022 13:14:02 +0200, Jan Kara wrote: > When delayed allocation is disabled (either through mount option or > because we are running low on free space), ext4_write_begin() allocates > blocks with EXT4_GET_BLOCKS_IO_CREATE_EXT flag. With this flag extent > merging is disabled and since ext4_write_begin() is called for each page > separately, we end up with a *lot* of 1 block extents in the extent tree > and following writeback is writing 1 block at a time which results in > very poor write throughput (4 MB/s instead of 200 MB/s). These days when > ext4_get_block_unwritten() is used only by ext4_write_begin(), > ext4_page_mkwrite() and inline data conversion, we can safely allow > extent merging to happen from these paths since following writeback will > happen on different boundaries anyway. So use > EXT4_GET_BLOCKS_CREATE_UNRIT_EXT instead which restores the performance. > > [...] Applied, thanks! [1/1] ext4: Improve write performance with disabled delalloc commit: 8d5459c11f548131ce48b2fbf45cccc5c382558f Best regards,
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c index 646ece9b3455..815da8f6c2e5 100644 --- a/fs/ext4/inode.c +++ b/fs/ext4/inode.c @@ -822,7 +822,7 @@ int ext4_get_block_unwritten(struct inode *inode, sector_t iblock, ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n", inode->i_ino, create); return _ext4_get_block(inode, iblock, bh_result, - EXT4_GET_BLOCKS_IO_CREATE_EXT); + EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT); } /* Maximum number of blocks we map for direct IO at once. */
When delayed allocation is disabled (either through mount option or because we are running low on free space), ext4_write_begin() allocates blocks with EXT4_GET_BLOCKS_IO_CREATE_EXT flag. With this flag extent merging is disabled and since ext4_write_begin() is called for each page separately, we end up with a *lot* of 1 block extents in the extent tree and following writeback is writing 1 block at a time which results in very poor write throughput (4 MB/s instead of 200 MB/s). These days when ext4_get_block_unwritten() is used only by ext4_write_begin(), ext4_page_mkwrite() and inline data conversion, we can safely allow extent merging to happen from these paths since following writeback will happen on different boundaries anyway. So use EXT4_GET_BLOCKS_CREATE_UNRIT_EXT instead which restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> --- fs/ext4/inode.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) So the poor write performance I was speaking about on Thursday was due to this...