virtiofs can be slow with small writes if xattr are enabled and we are doing cached writes (No direct I/). Ganesh Mahalingam noticed this here. https://github.com/kata-containers/runtime/issues/2815 Some debugging showed that that file_remove_privs() is called in cached write path on every write. And everytime it calls security_inode_need_killpriv() which results in call to __vfs_getxattr(XATTR_NAME_CAPS). And this goes to file server to fetch xattr. This extra round trip for every write slows down writes tremendously. Normally to avoid paying this penalty on every write, vfs has the notion of caching this information in inode (S_NOSEC). So vfs sets S_NOSEC, if filesystem opted for it using super block flag SB_NOSEC. And S_NOSEC is cleared when setuid/setgid bit is set or when security xattr is set on inode so that next time a write happens, we check inode again for clearing setuid/setgid bits as well clear any security.capability xattr. This seems to work well for local file systems but for remote file systems it is possible that VFS does not have full picture and a different client sets setuid/setgid bit or security.capability xattr on file and that means VFS information about S_NOSEC on another client will be stale. So for remote filesystems SB_NOSEC was disabled by default. commit 9e1f1de02c2275d7172e18dc4e7c2065777611bf Author: Al Viro <viro@xxxxxxxxxxxxxxxxxx> Date: Fri Jun 3 18:24:58 2011 -0400 more conservative S_NOSEC handling That commit mentioned that these filesystems can still make use of SB_NOSEC as long as they clear S_NOSEC when they are refreshing inode attriutes from server. So this patch tries to enable SB_NOSEC on fuse (regular fuse as well as virtiofs). And clear SB_NOSEC when we are refreshing inode attributes. This is enabled only if server supports FUSE_HANDLE_KILLPRIV_V2. This says that server will clear setuid/setgid/security.capability on chown/truncate/write as apporpriate. This should provide tighter coherency because now suid/sgid/security.capability will be cleared even if fuse client cache has not seen these attrs. Basic idea is that fuse client will trigger suid/sgid/security.capability clearing based on its attr cache. But even if cache has gone stale, it is fine because FUSE_HANDLE_KILLPRIV_V2 will make sure WRITE clear suid/sgid/security.capability. We make this change only if server supports FUSE_HANDLE_KILLPRIV_V2. This should make sure that existing filesystems which might be relying on seucurity.capability always being queried from server are not impacted. This tighter coherency relies on WRITE showing up on server (and not being cached in guest). So writeback_cache mode will not provide that tight coherency and it is not recommended to use two together. Having said that it might work reasonably well for lot of use cases. This change improves random write performance very significantly. I am running virtiofsd with cache=auto and following fio command. fio --ioengine=libaio --direct=1 --name=test --filename=/mnt/virtiofs/random_read_write.fio --bs=4k --iodepth=64 --size=4G --readwrite=randwrite Before this patch I get around 50MB/s and after the patch I get around 250MB/s bandwidth. So improvement is very significant. Reported-by: "Mahalingam, Ganesh" <ganesh.mahalingam@xxxxxxxxx> Signed-off-by: Vivek Goyal <vgoyal@xxxxxxxxxx> --- fs/fuse/inode.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c index 20740b61f12b..4b7a043f21ee 100644 --- a/fs/fuse/inode.c +++ b/fs/fuse/inode.c @@ -201,6 +201,16 @@ void fuse_change_attributes_common(struct inode *inode, struct fuse_attr *attr, inode->i_mode &= ~S_ISVTX; fi->orig_ino = attr->ino; + + /* + * We are refreshing inode data and it is possible that another + * client set suid/sgid or security.capability xattr. So clear + * S_NOSEC. Ideally, we could have cleared it only if suid/sgid + * was set or if security.capability xattr was set. But we don't + * know if security.capability has been set or not. So clear it + * anyway. Its less efficient but should is safe. + */ + inode->i_flags &= ~S_NOSEC; } void fuse_change_attributes(struct inode *inode, struct fuse_attr *attr, @@ -993,8 +1003,10 @@ static void process_init_reply(struct fuse_conn *fc, struct fuse_args *args, !fuse_dax_check_alignment(fc, arg->map_alignment)) { ok = false; } - if (arg->flags & FUSE_HANDLE_KILLPRIV_V2) + if (arg->flags & FUSE_HANDLE_KILLPRIV_V2) { fc->handle_killpriv_v2 = 1; + fc->sb->s_flags |= SB_NOSEC; + } } else { ra_pages = fc->max_read / PAGE_SIZE; fc->no_lock = 1; -- 2.25.4
