On 6/3/22 6:04 AM, Jens Axboe wrote: > On 6/2/22 8:43 PM, Dave Chinner wrote: >> On Thu, Jun 02, 2022 at 02:09:00AM -0600, Jens Axboe wrote: >>> On 6/1/22 3:01 PM, Stefan Roesch wrote: >>>> This patch series adds support for async buffered writes when using both >>>> xfs and io-uring. Currently io-uring only supports buffered writes in the >>>> slow path, by processing them in the io workers. With this patch series it is >>>> now possible to support buffered writes in the fast path. To be able to use >>>> the fast path the required pages must be in the page cache, the required locks >>>> in xfs can be granted immediately and no additional blocks need to be read >>>> form disk. >>> >>> This series looks good to me now, but will need some slight rebasing >>> since the 5.20 io_uring branch has split up the code a bit. Trivial to >>> do though, I suspect it'll apply directly if we just change >>> fs/io_uring.c to io_uring/rw.c instead. >>> >>> The bigger question is how to stage this, as it's touching a bit of fs, >>> mm, and io_uring... >> >> What data integrity testing has this had? Has it been run through a >> few billion fsx operations with w/ io_uring read/write enabled? > > I'll let Stefan expand on this, but just mention what I know - it has > been fun via fio at least. Each of the performance tests were hour long > each, and also specific test cases were written to test the boundary > conditions of what pages of a range where in page cache, etc. Also with > data verification. > I performed the following tests: - fio tests with various block sizes and different modes (psysnc, io_uring, libaio) - fsx tests with one billion ops - individual test program - to test with different block sizes - test short writes - test holes - test without readahead > Don't know if fsx specifically has been used it. >
