Before applying this patchset, we had to do extent state operation by the block size when reading pages because we need cache the checksum value of each block into the extent state, it was inefficient and increased the lock contenttion of extent state. Besides that, the acquisition of the extent map during the page read was also inefficient, we searched it every time we dealt with a page. In fact, if we don't cache the checksum into the extent state, we can batch those extent state operation to reduce the time of the operations and the lock contention. Patch 3-5 do this work. Besides that, the acquisition of the extent map during the page read was also inefficient, we searched it every time we dealt with a page. In fact, it is very likely that the extent map objects are the same one. So we can cache the extent map object to reduce its search time and the lock contention of the extent map tree. Patch 6 does this work. I tested this patchset by sysbench on my SSD, it makes the performance of the random read be faster by ~7%. The lock contention is reduced by ~97%. Miao Xie (6): Btrfs: remove unnecessary argument of bio_readpage_error() Btrfs: add branch prediction hints in the read page end IO function Btrfs: don't cache the csum value into the extent state tree Btrfs: batch the extent state operation in the end io handle of the read page Btrfs: batch the extent state operation when reading pages Btrfs: cache the extent map struct when reading several pages fs/btrfs/btrfs_inode.h | 21 +++ fs/btrfs/ctree.h | 4 +- fs/btrfs/disk-io.c | 5 +- fs/btrfs/extent_io.c | 392 ++++++++++++++++++++++++++++++------------------- fs/btrfs/extent_io.h | 10 +- fs/btrfs/file-item.c | 81 ++++++---- fs/btrfs/inode.c | 85 +++++------ fs/btrfs/volumes.h | 7 + 8 files changed, 370 insertions(+), 235 deletions(-) -- 1.8.1.4 -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
