Re: [PATCH v4 10/12] virtiofsd: Add inodes_by_handle hash table

[Vivek Goyal]

On Wed, Oct 20, 2021 at 04:10:51PM +0200, Hanna Reitz wrote:
> On 20.10.21 14:29, Vivek Goyal wrote:
> > On Wed, Oct 20, 2021 at 12:02:32PM +0200, Hanna Reitz wrote:
> > > On 19.10.21 22:02, Vivek Goyal wrote:
> > > > On Thu, Sep 16, 2021 at 10:40:43AM +0200, Hanna Reitz wrote:
> > > > > Currently, lo_inode.fhandle is always NULL and so always keep an 
> > > > > O_PATH
> > > > > FD in lo_inode.fd.  Therefore, when the respective inode is unlinked,
> > > > > its inode ID will remain in use until we drop our lo_inode (and
> > > > > lo_inode_put() thus closes the FD).  Therefore, lo_find() can safely 
> > > > > use
> > > > > the inode ID as an lo_inode key, because any inode with an inode ID we
> > > > > find in lo_data.inodes (on the same filesystem) must be the exact same
> > > > > file.
> > > > > 
> > > > > This will change when we start setting lo_inode.fhandle so we do not
> > > > > have to keep an O_PATH FD open.  Then, unlinking such an inode will
> > > > > immediately remove it, so its ID can then be reused by newly created
> > > > > files, even while the lo_inode object is still there[1].
> > > > > 
> > > > > So creating a new file can then reuse the old file's inode ID, and
> > > > > looking up the new file would lead to us finding the old file's
> > > > > lo_inode, which is not ideal.
> > > > > 
> > > > > Luckily, just as file handles cause this problem, they also solve it: 
> > > > >  A
> > > > > file handle contains a generation ID, which changes when an inode ID 
> > > > > is
> > > > > reused, so the new file can be distinguished from the old one.  So all
> > > > > we need to do is to add a second map besides lo_data.inodes that maps
> > > > > file handles to lo_inodes, namely lo_data.inodes_by_handle.  For
> > > > > clarity, lo_data.inodes is renamed to lo_data.inodes_by_ids.
> > > > > 
> > > > > Unfortunately, we cannot rely on being able to generate file handles
> > > > > every time.  Therefore, we still enter every lo_inode object into
> > > > > inodes_by_ids, but having an entry in inodes_by_handle is optional.  A
> > > > > potential inodes_by_handle entry then has precedence, the 
> > > > > inodes_by_ids
> > > > > entry is just a fallback.
> > > > > 
> > > > > Note that we do not generate lo_fhandle objects yet, and so we also do
> > > > > not enter anything into the inodes_by_handle map yet.  Also, all 
> > > > > lookups
> > > > > skip that map.  We might manually create file handles with some code
> > > > > that is immediately removed by the next patch again, but that would
> > > > > break the assumption in lo_find() that every lo_inode with a non-NULL
> > > > > .fhandle must have an entry in inodes_by_handle and vice versa.  So we
> > > > > leave actually using the inodes_by_handle map for the next patch.
> > > > > 
> > > > > [1] If some application in the guest still has the file open, there is
> > > > > going to be a corresponding FD mapping in lo_data.fd_map.  In such a
> > > > > case, the inode will only go away once every application in the guest
> > > > > has closed it.  The problem described only applies to cases where the
> > > > > guest does not have the file open, and it is just in the dentry cache,
> > > > > basically.
> > > > > 
> > > > > Signed-off-by: Hanna Reitz <hreitz@redhat.com>
> > > > > ---
> > > > >    tools/virtiofsd/passthrough_ll.c | 81 
> > > > > +++++++++++++++++++++++++-------
> > > > >    1 file changed, 65 insertions(+), 16 deletions(-)
> > > > > 
> > > > > diff --git a/tools/virtiofsd/passthrough_ll.c 
> > > > > b/tools/virtiofsd/passthrough_ll.c
> > > > > index bd8fc922ea..b7d6aa7f9d 100644
> > > > > --- a/tools/virtiofsd/passthrough_ll.c
> > > > > +++ b/tools/virtiofsd/passthrough_ll.c
> > > > > @@ -186,7 +186,8 @@ struct lo_data {
> > > > >        int announce_submounts;
> > > > >        bool use_statx;
> > > > >        struct lo_inode root;
> > > > > -    GHashTable *inodes; /* protected by lo->mutex */
> > > > > +    GHashTable *inodes_by_ids; /* protected by lo->mutex */
> > > > > +    GHashTable *inodes_by_handle; /* protected by lo->mutex */
> > > > >        struct lo_map ino_map; /* protected by lo->mutex */
> > > > >        struct lo_map dirp_map; /* protected by lo->mutex */
> > > > >        struct lo_map fd_map; /* protected by lo->mutex */
> > > > > @@ -275,8 +276,9 @@ static struct {
> > > > >    /* That we loaded cap-ng in the current thread from the saved */
> > > > >    static __thread bool cap_loaded = 0;
> > > > > -static struct lo_inode *lo_find(struct lo_data *lo, struct stat *st,
> > > > > -                                uint64_t mnt_id);
> > > > > +static struct lo_inode *lo_find(struct lo_data *lo,
> > > > > +                                const struct lo_fhandle *fhandle,
> > > > > +                                struct stat *st, uint64_t mnt_id);
> > > > >    static int xattr_map_client(const struct lo_data *lo, const char 
> > > > > *client_name,
> > > > >                                char **out_name);
> > > > > @@ -1143,18 +1145,40 @@ out_err:
> > > > >        fuse_reply_err(req, saverr);
> > > > >    }
> > > > > -static struct lo_inode *lo_find(struct lo_data *lo, struct stat *st,
> > > > > -                                uint64_t mnt_id)
> > > > > +static struct lo_inode *lo_find(struct lo_data *lo,
> > > > > +                                const struct lo_fhandle *fhandle,
> > > > > +                                struct stat *st, uint64_t mnt_id)
> > > > >    {
> > > > > -    struct lo_inode *p;
> > > > > -    struct lo_key key = {
> > > > > +    struct lo_inode *p = NULL;
> > > > > +    struct lo_key ids_key = {
> > > > >            .ino = st->st_ino,
> > > > >            .dev = st->st_dev,
> > > > >            .mnt_id = mnt_id,
> > > > >        };
> > > > >        pthread_mutex_lock(&lo->mutex);
> > > > > -    p = g_hash_table_lookup(lo->inodes, &key);
> > > > > +    if (fhandle) {
> > > > > +        p = g_hash_table_lookup(lo->inodes_by_handle, fhandle);
> > > > > +    }
> > > > > +    if (!p) {
> > > > > +        p = g_hash_table_lookup(lo->inodes_by_ids, &ids_key);
> > > > > +        /*
> > > > > +         * When we had to fall back to looking up an inode by its
> > > > > +         * inode ID, ensure that we hit an entry that has a valid 
> > > > > file
> > > > > +         * descriptor.  Having an FD open means that the inode cannot
> > > > > +         * really be deleted until the FD is closed, so that the 
> > > > > inode
> > > > > +         * ID remains valid until we evict our lo_inode.
> > > > > +         * With no FD open (and just a file handle), the inode can be
> > > > > +         * deleted while we still have our lo_inode, and so the inode
> > > > > +         * ID may be reused by a completely different new inode.  We
> > > > > +         * then must look up the lo_inode by file handle, because 
> > > > > this
> > > > > +         * handle contains a generation ID to differentiate between
> > > > > +         * the old and the new inode.
> > > > > +         */
> > > > > +        if (p && p->fd == -1) {
> > > > > +            p = NULL;
> > > > > +        }
> > > > What happens in following scenario.
> > > > 
> > > > - Say I have a hard linked file foo.txt with link foo-link.txt.
> > > > 
> > > > - I lookup foo.txt. We generate file handle and add inode for foo.txt
> > > >     to inode cache. lo_inode->fhandle will be valie but lo_inode->fd == 
> > > > -1.
> > > > 
> > > > - Now later lookup for foo-link.txt happens. Say this time we can't
> > > >     generate file handle.
> > > Which we’ve already decided is practically impossible.
> > Agreed that probably is very less but it can happen when sufficient
> > resources are not available, like -ENOMEM.
> > 
> > static long do_sys_name_to_handle(struct path *path,
> >                                    struct file_handle __user *ufh,
> >                                    int __user *mnt_id)
> > {
> >          handle = kmalloc(sizeof(struct file_handle) + 
> > f_handle.handle_bytes,
> >                           GFP_KERNEL);
> >          if (!handle)
> >                  return -ENOMEM;
> > }
> 
> OK, but do we really want to be prepared for an ENOMEM from the kernel?

Why not. Other filesystem methods are dealing with it and return -ENOMEM
to fuse client. So I don't see any reason why file handle stuff should
be special. In fact in an attempt to trying to make it special just
increases complexity and having to deal with fallback plans (except
-EOPNOTSUP).


> 
> > > > When we try to lookup inode, lo_find() should
> > > >     return NULL. It will find inode by ids but not use it because inode
> > > >     was added using file handle and p->fd == -1. That means lookup
> > > >     for foo-link.txt will end up adding another inode, when it should
> > > >     not have?
> > > Yes, it would end up adding another inode, which doesn’t seem catastrophic
> > > to me.
> > > But again, the whole case seems impossible to me.
> > Given we can get -ENOMEM error it is not impossible.
> > 
> > I thought all along you wanted to write code so that we could fallback
> > to ids in case of errors.
> 
> Well, yes, if possible, but unfortunately it isn’t possible here.
> 
> > Anyway, if you agree that except the case of
> > -EOPNOTSUPP, we don't have to worry about fallback, then let us just
> > reutrn error to caller if get_file_handle() fails (except the case
> > of -EOPNOTSUPP).
> 
> We’ve had this discussion before, and so I tried that for v3 but abandoned
> it again because I found it unreasonably complicated.

I guess you had some code previously (when you were not parsing mountinfo)
that made it harder to deal with -ENOMEM. You could not afford a failure.
With a move to mountinfo parsing, that restriction might have gone away
now.

> Always falling back is
> simpler than deciding when to return an error to the guest and when not to;
> and also, considering errors other than EOPNOTSUPP should be rare, the code
> that returns errors to the guest would effectively be dead code.

May be. I probably should take your patches and try to convert them to 
return all temporary errors to client (except -ENOTSUPP) and see what
are the issues I face.

> 
> If we really must return an error in the case you describe, then I don’t
> think there’s a way around it, though...  In virtiofsd-rs, we could let
> FileHandle::from_name_at() return an io::Result<Option<Self>>, such that it
> returns Ok(None) in case of EOPNOTSUPP, or an Error in case there is an
> unexpected error like ENOMEM.  I think that shouldn’t be too bad, but my
> experience says the devil’s in the details.
> 
> It’s just that it seems to be a purely theoretical problem, because if the
> kernel gives us ENOMEM, we probably won’t live too long anyway.

Or may be some other process will exit/die freeing memory and virtiofsd
will continue to live. So we should not conclude that after -ENOMEM
virtiofsd will not live too long anyway. Atleast code should not be
written with this assumption.

Thanks
Vivek