Recently someone asked me how io_uring buffered IO compares to mmaped IO in terms of performance. So I ran some tests with buffered IO, and found the experience to be somewhat painful. The test case is pretty basic, random reads over a dataset that's 10x the size of RAM. Performance starts out fine, and then the page cache fills up and we hit a throughput cliff. CPU usage of the IO threads go up, and we have kswapd spending 100% of a core trying to keep up. Seeing that, I was reminded of the many complaints I here about buffered IO, and the fact that most of the folks complaining will ultimately bite the bullet and move to O_DIRECT to just get the kernel out of the way. But I don't think it needs to be like that. Switching to O_DIRECT isn't always easily doable. The buffers have different life times, size and alignment constraints, etc. On top of that, mixing buffered and O_DIRECT can be painful. Seems to me that we have an opportunity to provide something that sits somewhere in between buffered and O_DIRECT, and this is where RWF_UNCACHED enters the picture. If this flag is set on IO, we get the following behavior: - If the data is in cache, it remains in cache and the copy (in or out) is served to/from that. This is true for both reads and writes. - For writes, if the data is NOT in cache, we add it while performing the IO. When the IO is done, we remove it again. - For reads, if the data is NOT in the cache, we allocate a private page and use that for IO. When the IO is done, we free this page. The page never sees the page cache. With this, I can do 100% smooth buffered reads or writes without pushing the kernel to the state where kswapd is sweating bullets. In fact it doesn't even register. Comments appreciated! This should work on any standard file system, using either the generic helpers or iomap. I have tested ext4 and xfs for the right read/write behavior, but no further validation has been done yet. This version contains the bigger prep patch of switching iomap_apply() and actors to struct iomap_data, and I hope I didn't mess that up too badly. I'll try and exercise it all, I've done XFS mounts and reads+writes and it seems happy from that POV at least. The core of the changes are actually really small, the majority of the diff is just prep work to get there. Patches are against current git, and can also be found here: https://git.kernel.dk/cgit/linux-block/log/?h=buffered-uncached fs/ceph/file.c | 2 +- fs/dax.c | 25 +++-- fs/ext4/file.c | 2 +- fs/iomap/apply.c | 50 ++++++--- fs/iomap/buffered-io.c | 225 +++++++++++++++++++++++++--------------- fs/iomap/direct-io.c | 57 +++++----- fs/iomap/fiemap.c | 48 +++++---- fs/iomap/seek.c | 64 +++++++----- fs/iomap/swapfile.c | 27 ++--- fs/nfs/file.c | 2 +- include/linux/fs.h | 7 +- include/linux/iomap.h | 20 +++- include/uapi/linux/fs.h | 5 +- mm/filemap.c | 89 +++++++++++++--- 14 files changed, 416 insertions(+), 207 deletions(-) Changes since v3: - Add iomap_actor_data to cut down on arguments - Fix bad flag drop in iomap_write_begin() - Remove unused IOMAP_WRITE_F_UNCACHED flag - Don't use the page cache at all for reads Changes since v2: - Rework the write side according to Chinners suggestions. Much cleaner this way. It does mean that we invalidate the full write region if just ONE page (or more) had to be created, where before it was more granular. I don't think that's a concern, and on the plus side, we now no longer have to chunk invalidations into 15/16 pages at the time. - Cleanups Changes since v1: - Switch to pagevecs for write_drop_cached_pages() - Use page_offset() instead of manual shift - Ensure we hold a reference on the page between calling ->write_end() and checking the mapping on the locked page - Fix XFS multi-page streamed writes, we'd drop the UNCACHED flag after the first page -- Jens Axboe
