On Thu 29-08-19 13:06:22, Andreas Dilger wrote: > On Aug 29, 2019, at 4:58 AM, Jan Kara <jack@xxxxxxx> wrote: > > > > On Tue 27-08-19 21:51:18, Dave Chinner wrote: > >> On Tue, Aug 27, 2019 at 10:05:49AM +0800, Joseph Qi wrote: > >>> This patch set is trying to revert parallel dio reads feature at present > >>> since it causes significant performance regression in mixed random > >>> read/write scenario. > >>> > >>> Joseph Qi (3): > >>> Revert "ext4: remove EXT4_STATE_DIOREAD_LOCK flag" > >>> Revert "ext4: fix off-by-one error when writing back pages before dio > >>> read" > >>> Revert "ext4: Allow parallel DIO reads" > >>> > >>> fs/ext4/ext4.h | 17 +++++++++++++++++ > >>> fs/ext4/extents.c | 19 ++++++++++++++----- > >>> fs/ext4/inode.c | 47 +++++++++++++++++++++++++++++++---------------- > >>> fs/ext4/ioctl.c | 4 ++++ > >>> fs/ext4/move_extent.c | 4 ++++ > >>> fs/ext4/super.c | 12 +++++++----- > >>> 6 files changed, 77 insertions(+), 26 deletions(-) > >> > >> Before doing this, you might want to have a chat and co-ordinate > >> with the folks that are currently trying to port the ext4 direct IO > >> code to use the iomap infrastructure: > >> > >> https://lore.kernel.org/linux-ext4/20190827095221.GA1568@xxxxxxxxxxxxxxxxxxxxxx/T/#t > >> > >> That is going to need the shared locking on read and will work just > >> fine with shared locking on write, too (it's the code that XFS uses > >> for direct IO). So it might be best here if you work towards shared > >> locking on the write side rather than just revert the shared locking > >> on the read side.... > > > > Yeah, after converting ext4 DIO path to iomap infrastructure, using shared > > inode lock for all aligned non-extending DIO writes will be easy so I'd > > prefer if we didn't have to redo the iomap conversion patches due to these > > reverts. > > But if the next kernel is LTS and the iomap implementation isn't in the > current merge window (very unlikely) then we're stuck with this performance > hit for LTS. It is also unlikely that LTS will take the revert patches if > they have not been landed to master. I agree this is not great but the regression is there for 3 years, it has been released in major distribution kernels quite a long time ago, and only now someone complained. So it doesn't seem many people care about performance of mixed RW workload when mounted with dioread_nolock (note that the patches actually improve performance of read-only DIO workload when not using dioread_nolock as for that case, exclusive lock is replaced with a shared one). Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
