On Wed, Sep 30, 2020 at 5:53 AM Vivek Goyal <vgoyal@redhat.com> wrote:
> On Fri, Sep 25, 2020 at 01:06:55PM +0100, Dr. David Alan Gilbert (git)
> wrote:
> > From: Jiachen Zhang <zhangjiachen.jaycee@bytedance.com>
> >
> > Due to the commit 65da4539803373ec4eec97ffc49ee90083e56efd, the O_DIRECT
> > open flag of guest applications will be discarded by virtiofsd. While
> > this behavior makes it consistent with the virtio-9p scheme when guest
> > applications use direct I/O, we no longer have any chance to bypass the
> > host page cache.
> >
> > Therefore, we add a flag 'allow_direct_io' to lo_data. If '-o
> > no_allow_direct_io' option is added, or none of '-o allow_direct_io' or
> > '-o no_allow_direct_io' is added, the 'allow_direct_io' will be set to
> > 0, and virtiofsd discards O_DIRECT as before. If '-o allow_direct_io'
> > is added to the starting command-line, 'allow_direct_io' will be set to
> > 1, so that the O_DIRECT flags will be retained and host page cache can
> > be bypassed.
>
> Hi Jiachen,
>
> Curious that in what cases you want to bypass host page cache.
>
> Thanks
> Vivek
>
Hi Vivek,
Some apps like DBMS may allocate their own file cache in userspace, so they
may want to bypass kernel page cache by using O_DIRECT. When these apps are
running in guest and access virtio-fs files, we'd better obey these needs.
This can also eliminate the host memory usage.
Another case is when we perform file I/O benchmarks on different storage
devices (like HDD and SSD), it's not that fair to introduce ram caches. By
using "cache=none" with "allow_direct_io", we can bypass guest and host page
caches, and access the storage devices directly.
Best wishes,
Jiachen
> >
> > Signed-off-by: Jiachen Zhang <zhangjiachen.jaycee@bytedance.com>
> > Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
> > Message-Id: <20200824105957.61265-1-zhangjiachen.jaycee@bytedance.com>
> > Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> > ---
> > tools/virtiofsd/helper.c | 4 ++++
> > tools/virtiofsd/passthrough_ll.c | 20 ++++++++++++++------
> > 2 files changed, 18 insertions(+), 6 deletions(-)
> >
> > diff --git a/tools/virtiofsd/helper.c b/tools/virtiofsd/helper.c
> > index 7bc5d7dc5a..85770d63f1 100644
> > --- a/tools/virtiofsd/helper.c
> > +++ b/tools/virtiofsd/helper.c
> > @@ -178,6 +178,10 @@ void fuse_cmdline_help(void)
> > " (0 leaves rlimit
> unchanged)\n"
> > " default: min(1000000,
> fs.file-max - 16384)\n"
> > " if the current
> rlimit is lower\n"
> > + " -o allow_direct_io|no_allow_direct_io\n"
> > + " retain/discard O_DIRECT
> flags passed down\n"
> > + " to virtiofsd from guest
> applications.\n"
> > + " default:
> no_allow_direct_io\n"
> > );
> > }
> >
> > diff --git a/tools/virtiofsd/passthrough_ll.c
> b/tools/virtiofsd/passthrough_ll.c
> > index 784330e0e4..0b229ebd57 100644
> > --- a/tools/virtiofsd/passthrough_ll.c
> > +++ b/tools/virtiofsd/passthrough_ll.c
> > @@ -151,6 +151,7 @@ struct lo_data {
> > int timeout_set;
> > int readdirplus_set;
> > int readdirplus_clear;
> > + int allow_direct_io;
> > struct lo_inode root;
> > GHashTable *inodes; /* protected by lo->mutex */
> > struct lo_map ino_map; /* protected by lo->mutex */
> > @@ -179,6 +180,8 @@ static const struct fuse_opt lo_opts[] = {
> > { "cache=always", offsetof(struct lo_data, cache), CACHE_ALWAYS },
> > { "readdirplus", offsetof(struct lo_data, readdirplus_set), 1 },
> > { "no_readdirplus", offsetof(struct lo_data, readdirplus_clear), 1
> },
> > + { "allow_direct_io", offsetof(struct lo_data, allow_direct_io), 1 },
> > + { "no_allow_direct_io", offsetof(struct lo_data, allow_direct_io),
> 0 },
> > FUSE_OPT_END
> > };
> > static bool use_syslog = false;
> > @@ -1516,7 +1519,8 @@ static void lo_releasedir(fuse_req_t req,
> fuse_ino_t ino,
> > fuse_reply_err(req, 0);
> > }
> >
> > -static void update_open_flags(int writeback, struct fuse_file_info *fi)
> > +static void update_open_flags(int writeback, int allow_direct_io,
> > + struct fuse_file_info *fi)
> > {
> > /*
> > * With writeback cache, kernel may send read requests even
> > @@ -1541,10 +1545,13 @@ static void update_open_flags(int writeback,
> struct fuse_file_info *fi)
> >
> > /*
> > * O_DIRECT in guest should not necessarily mean bypassing page
> > - * cache on host as well. If somebody needs that behavior, it
> > - * probably should be a configuration knob in daemon.
> > + * cache on host as well. Therefore, we discard it by default
> > + * ('-o no_allow_direct_io'). If somebody needs that behavior,
> > + * the '-o allow_direct_io' option should be set.
> > */
> > - fi->flags &= ~O_DIRECT;
> > + if (!allow_direct_io) {
> > + fi->flags &= ~O_DIRECT;
> > + }
> > }
> >
> > static void lo_create(fuse_req_t req, fuse_ino_t parent, const char
> *name,
> > @@ -1576,7 +1583,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t
> parent, const char *name,
> > goto out;
> > }
> >
> > - update_open_flags(lo->writeback, fi);
> > + update_open_flags(lo->writeback, lo->allow_direct_io, fi);
> >
> > fd = openat(parent_inode->fd, name, (fi->flags | O_CREAT) &
> ~O_NOFOLLOW,
> > mode);
> > @@ -1786,7 +1793,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t
> ino, struct fuse_file_info *fi)
> > fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n",
> ino,
> > fi->flags);
> >
> > - update_open_flags(lo->writeback, fi);
> > + update_open_flags(lo->writeback, lo->allow_direct_io, fi);
> >
> > sprintf(buf, "%i", lo_fd(req, ino));
> > fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> > @@ -2823,6 +2830,7 @@ int main(int argc, char *argv[])
> > .debug = 0,
> > .writeback = 0,
> > .posix_lock = 0,
> > + .allow_direct_io = 0,
> > .proc_self_fd = -1,
> > };
> > struct lo_map_elem *root_elem;
> > --
> > 2.26.2
> >
>
>