Re: [for-6.1 v3 3/3] virtiofsd: Add support for FUSE_SYNCFS request

[Greg Kurz <groug@kaod.org>]

On Mon, 10 May 2021 15:15:02 -0400
Vivek Goyal <vgoyal@redhat.com> wrote:

> On Mon, May 10, 2021 at 05:55:39PM +0200, Greg Kurz wrote:
> > Honor the expected behavior of syncfs() to synchronously flush all data
> > and metadata on linux systems. Simply loop on all known submounts and
> > call syncfs() on them.
> > 
> > Note that syncfs() might suffer from a time penalty if the submounts
> > are being hammered by some unrelated workload on the host. The only
> > solution to avoid that is to avoid shared submounts.
> > 
> > Signed-off-by: Greg Kurz <groug@kaod.org>
> > ---
> >  tools/virtiofsd/fuse_lowlevel.c       | 11 ++++++++
> >  tools/virtiofsd/fuse_lowlevel.h       | 12 +++++++++
> >  tools/virtiofsd/passthrough_ll.c      | 38 +++++++++++++++++++++++++++
> >  tools/virtiofsd/passthrough_seccomp.c |  1 +
> >  4 files changed, 62 insertions(+)
> > 
> > diff --git a/tools/virtiofsd/fuse_lowlevel.c b/tools/virtiofsd/fuse_lowlevel.c
> > index 58e32fc96369..3be95ec903c9 100644
> > --- a/tools/virtiofsd/fuse_lowlevel.c
> > +++ b/tools/virtiofsd/fuse_lowlevel.c
> > @@ -1870,6 +1870,16 @@ static void do_lseek(fuse_req_t req, fuse_ino_t nodeid,
> >      }
> >  }
> >  
> > +static void do_syncfs(fuse_req_t req, fuse_ino_t nodeid,
> > +                      struct fuse_mbuf_iter *iter)
> > +{
> > +    if (req->se->op.syncfs) {
> > +        req->se->op.syncfs(req);
> > +    } else {
> > +        fuse_reply_err(req, ENOSYS);
> > +    }
> > +}
> > +
> >  static void do_init(fuse_req_t req, fuse_ino_t nodeid,
> >                      struct fuse_mbuf_iter *iter)
> >  {
> > @@ -2267,6 +2277,7 @@ static struct {
> >      [FUSE_RENAME2] = { do_rename2, "RENAME2" },
> >      [FUSE_COPY_FILE_RANGE] = { do_copy_file_range, "COPY_FILE_RANGE" },
> >      [FUSE_LSEEK] = { do_lseek, "LSEEK" },
> > +    [FUSE_SYNCFS] = { do_syncfs, "SYNCFS" },
> >  };
> >  
> >  #define FUSE_MAXOP (sizeof(fuse_ll_ops) / sizeof(fuse_ll_ops[0]))
> > diff --git a/tools/virtiofsd/fuse_lowlevel.h b/tools/virtiofsd/fuse_lowlevel.h
> > index 3bf786b03485..890c520b195a 100644
> > --- a/tools/virtiofsd/fuse_lowlevel.h
> > +++ b/tools/virtiofsd/fuse_lowlevel.h
> > @@ -1225,6 +1225,18 @@ struct fuse_lowlevel_ops {
> >       */
> >      void (*lseek)(fuse_req_t req, fuse_ino_t ino, off_t off, int whence,
> >                    struct fuse_file_info *fi);
> > +
> > +    /**
> > +     * Synchronize file system content
> > +     *
> > +     * If this request is answered with an error code of ENOSYS,
> > +     * this is treated as success and future calls to syncfs() will
> > +     * succeed automatically without being sent to the filesystem
> > +     * process.
> > +     *
> > +     * @param req request handle
> > +     */
> > +    void (*syncfs)(fuse_req_t req);
> >  };
> >  
> >  /**
> > diff --git a/tools/virtiofsd/passthrough_ll.c b/tools/virtiofsd/passthrough_ll.c
> > index dc940a1d048b..289900c6d274 100644
> > --- a/tools/virtiofsd/passthrough_ll.c
> > +++ b/tools/virtiofsd/passthrough_ll.c
> > @@ -3153,6 +3153,43 @@ static void lo_lseek(fuse_req_t req, fuse_ino_t ino, off_t off, int whence,
> >      }
> >  }
> >  
> > +static void lo_syncfs(fuse_req_t req)
> > +{
> > +    struct lo_data *lo = lo_data(req);
> > +    GHashTableIter iter;
> > +    gpointer key, value;
> > +    int err = 0;
> > +
> > +    pthread_mutex_lock(&lo->mutex);
> > +
> > +    g_hash_table_iter_init(&iter, lo->mnt_inodes);
> > +    while (g_hash_table_iter_next(&iter, &key, &value)) {
> > +        struct lo_inode *inode = value;
> > +        int fd;
> > +
> > +        fuse_log(FUSE_LOG_DEBUG, "lo_syncfs(ino=%" PRIu64 ")\n",
> > +                 inode->fuse_ino);
> > +
> > +        fd = lo_inode_open(lo, inode, O_RDONLY);
> > +        if (fd < 0) {
> > +            err = -fd;
> > +            break;
> > +        }
> > +
> > +        if (syncfs(fd) < 0) {
> 
> I don't have a good feeling about calling syncfs() with lo->mutex held.
> This seems to be that global mutex which is held at so many places
> and will serialize everything else. I think we agreed that syncfs()
> can take 10s of seconds if fs is busy. And that means we will stall
> other filesystem operations too.
> 
> So will be good if we can call syncfs() outside the lock. May be prepare
> a list of inodes which are there, take a reference and drop the lock.
> call syncfs and then drop the reference on inode.
> 

You're right. I'll do that.

> Vivek
> 
> > +            err = errno;
> > +            close(fd);
> > +            break;
> > +        }
> > +
> > +        close(fd);
> > +    }
> > +
> > +    pthread_mutex_unlock(&lo->mutex);
> > +
> > +    fuse_reply_err(req, err);
> > +}
> > +
> >  static void lo_destroy(void *userdata)
> >  {
> >      struct lo_data *lo = (struct lo_data *)userdata;
> > @@ -3214,6 +3251,7 @@ static struct fuse_lowlevel_ops lo_oper = {
> >      .copy_file_range = lo_copy_file_range,
> >  #endif
> >      .lseek = lo_lseek,
> > +    .syncfs = lo_syncfs,
> >      .destroy = lo_destroy,
> >  };
> >  
> > diff --git a/tools/virtiofsd/passthrough_seccomp.c b/tools/virtiofsd/passthrough_seccomp.c
> > index 62441cfcdb95..343188447901 100644
> > --- a/tools/virtiofsd/passthrough_seccomp.c
> > +++ b/tools/virtiofsd/passthrough_seccomp.c
> > @@ -107,6 +107,7 @@ static const int syscall_allowlist[] = {
> >      SCMP_SYS(set_robust_list),
> >      SCMP_SYS(setxattr),
> >      SCMP_SYS(symlinkat),
> > +    SCMP_SYS(syncfs),
> >      SCMP_SYS(time), /* Rarely needed, except on static builds */
> >      SCMP_SYS(tgkill),
> >      SCMP_SYS(unlinkat),
> > -- 
> > 2.26.3
> > 
>