[Virtio-fs] [virtiofsd][virtiofsd-rs] unlink an openfile over NFS

[Vivek Goyal]

On Fri, Dec 03, 2021 at 10:39:25AM +0100, Hanna Reitz wrote:
> On 02.12.21 21:14, Vivek Goyal wrote:
> > On Thu, Dec 02, 2021 at 06:52:17PM +0100, Hanna Reitz wrote:
> > > On 02.12.21 16:51, Vivek Goyal wrote:
> > > > On Thu, Dec 02, 2021 at 04:03:20PM +0100, German Maglione wrote:
> > > > > On Wed, Dec 1, 2021 at 11:10 PM Vivek Goyal <vgoyal at redhat.com> wrote:
> > > > > 
> > > > > > On Wed, Dec 01, 2021 at 01:06:23PM +0100, German Maglione wrote:
> > > > > > > Hi,
> > > > > > > 
> > > > > > > I was working on [1] (related to [2]), and I saw that both versions
> > > > > > > (C and rust) of virtiofsd make the NFs client to "silly rename" an open
> > > > > > > file that were unlinked, because we keep each file open as O_PATH in the
> > > > > > > lo_do_lookup/do_lookup function. David pointed me to this problem, and I
> > > > > > > confirmed that this is the case.
> > > > > > > 
> > > > > > > This fires the silly rename in the nfs client. I'm talking with
> > > > > > > Bruce Fields (nfs team) to see how to move the silly rename functionality
> > > > > > > to the nfs server and outside the directory tree [4], to avoid having
> > > > > > > non-really-empty
> > > > > > > dirs full of .nfsxxx files. But it is not an easy task.
> > > > > > > Also, this will not solve some potential issues with the resource usage:
> > > > > > > disk space and open file limits (I hit this limit a couple of times
> > > > > > during
> > > > > > > my
> > > > > > > tests). And, in some cases could be worst, more than one guest sharing
> > > > > > the
> > > > > > > same
> > > > > > > exported dir, one guest just 'ls' or 'find' while the other 'rm' some
> > > > > > files.
> > > > > > > (The 'find' command will open all files, btw)
> > > > > > > 
> > > > > > > Vivek, I saw in [5] that you mentioned that this could be fixed
> > > > > > introducing
> > > > > > > synchronous, could you elaborate a bit or point me to some code?
> > > > > > Hi German,
> > > > > > 
> > > > > > Right now forget messages are asynchronous. They are sent to server and
> > > > > > client does not wait for any reply. That means when unlink() returns,
> > > > > > it is possible that a FORGET message is in progress and has not finished
> > > > > > yet.
> > > > > > 
> > > > > > Idea behind synchronous FORGET messages is that it will generate a reply
> > > > > > and client will wait for it. Given inode on server should be gone,
> > > > > > I am not sure how much sense does it make. But anyway conceputaully
> > > > > > that's the idea. If we want for FORGET message to finish, that will
> > > > > > mean that O_PATH fd opened by virtiofsd is closed and we will not
> > > > > > have NFS silly rename issue (atleast due to virtiofsd). If virtiofs
> > > > > > client has file open, then issue will still happen.
> > > > > > 
> > > > > > I think using file handles in virtiofsd_rs (--inode-file-handles) is
> > > > > > a reasonable solution for this problem. Trying to add synchronous
> > > > > > FORGET might be overkill.
> > > > > > 
> > > > > > 
> > > > > Hi Vivek,
> > > > > 
> > > > > Yes, as you said the solution is using --inode-file-hanldes, turns out
> > > > > the problem was the --inode-file-handles failed silently when
> > > > > choosing a sandbox other than namespace (now fixed by Hanna).
> > > > > 
> > > > > So now the thing is, what we do if it fails? Hanna posted an Issue about
> > > > > that:
> > > > > "[RFE] Reporting failure to generate file handles".
> > > > My take from the beginning has been that if file handle generation fails,
> > > > then report it back to user (instead of falling back to O_PATH fd
> > > > silently). That way user atleast knows that file handles can't be used.
> > > I remember that we had a discussion about whether to introduce a mandatory
> > > mode where this would be the behavior.  I thought we agreed that a
> > > best-effort mode always made sense, for example for the situation you
> > > describe below, where you have mixed filesystems, with some perhaps
> > > supporting file handles, and others not.
> > I think my primary objection was falling back to O_PATH fd for temporary
> > failures. So file systems supports file handle but if we could not
> > generate one due to lack of resources, I would rather return error. In
> > viritofsd C version, we identified a case where it could lead to two
> > inodes in hash table using different keys. I am not sure if virtiofsd_rs
> > version suffers from same issue or not.
> 
> No (I hope not), I’ve tried to incorporate your feedback on that, and so the
> Rust version does distinguish between temporary and persistent
> name_to_handle_at() failures.  (The former returns `Err(_)`, the latter
> `Ok(None)`.  `Err(_)`s are returned to the guest.)

Sounds good. Temporary failures should be returned to client.

> 
> (All errors for making a file handle openable, i.e. opening a mount FD, are
> still not returned to the guest and only lead to a fallback. 

I would return error the moment you encounter one. That seems to be theme
of all of the virtiofsd code. While trying to perform any operation, if
it encounters an error, it immediately returns it to caller. So opening a
mount FD should be no different and there should not be any need for
fallback.

> I remember at
> one point I tried looking into this code to find out where it would be
> reasonable to return an error, and where we should fall back, and I just
> found it very difficult to decide. It also (luckily? :)) didn’t have
> anything to do with the bug you describe, because that was about whether we
> have a handle for the lookup, and that’s now independent of the whole mount
> FD mess.)
> 
> > W.r.t what to do if filesystem does not support file handles, should
> > we fall back to O_PATH fd (and just emit a warning), I am fine with
> > falling back to O_PATH fd and just log it so that users know.
> > 
> > I have both the kind of users. Some users prefer fallback and other
> > users prefer to get an error if a certain feature can't be enabled.
> > I guess this will slowly evolve depending what users are asking for.
> > But logging a warning if file handles can't be enabled, sounds ok.
> > 
> > > As for a mandatory mode, I couldn’t imagine how exactly it would be useful,
> > > though.  I think my argument was: “What would a user do after launching
> > > virtiofsd with file handles forced on, and then noticing they don’t work?”
> > I guess they will relaunch without --inodes-file-handle.
> 
> Yes, that’s what I thought, too, and so it mostly sounded like an
> inconvenience to me, and not really helpful. :/

Alternatively, If there is anything configurable in underlying filesystem
(to enable file handle support), they could take an action to fix it and
restart virtiofsd with --inodes-file-handle.

> 
> > > And I still don’t really know, even though I’m proposing such a mode in said
> > > virtiofsd-rs gitlab issue for the NFS case.
> > Can you please provide a link to that issue.
> 
> Ah, sure: https://gitlab.com/virtio-fs/virtiofsd-rs/-/issues/17
> 
> > So why are you proposing this
> > mode for NFS?
> 
> Because it looks like with NFS, not using file handles may cause real
> problems, and so I feel like failing early (failure to generate file
> handles) is nicer than later (recursively deleting directories).

Ok. One could argue that even regular file systems could cause problem
by running into open file descriptor limit. So not sure NFS should be
an exception.

> 
> > > I suppose the answer is, check
> > > every configuration and find out why it doesn’t work; but you don’t need a
> > > mandatory file handle mode for this, logging errors whenever we need to fall
> > > back to O_PATH FDs would be completely sufficient.
> > Logging warnings should probably be good enough for lot of use cases.
> > 
> > > (I’m mostly proposing it for NFS, because non-working file handles are
> > > something that seems likely to cause other problems later. Emitting a
> > > warning would technically be completely fine in order to inform the user of
> > > this, but I feel like in this case it’d be better to nag them even more.)
> > So why NFS is special? Due to silly renames issue?
> 
> Yes.
> 
> > So argument will be
> > if there is a hard failure, then user can try to fix underlying filesystem
> > configuration to support file handles?
> 
> The argument is that the silly renames issue is kind of likely to produce
> real tangible problems when file handles don’t work.  I feel like that
> warrants more than just logging the error.
> 
> > This will be true for other
> > filesystems too.
> 
> Yes, but for other filesystems, using or not using file handles seems like a
> tuning thing, and not really anything that could produce tangible errors. 
> So to me, logging errors seems more appropriate then.

Hmm..., open file descriptor limit is around 1M, IIRC. So if guest has
cached enough inodes, one can run into it and that's the core problem
file handle support solves. 

IMHO, NFS is not an exception. If we could not enable file handle support,
both NFS and regular file systems can run into issues later.

I would say I like the idea of adding argument to --inodes-file-handle and
let user choose if they want hard failures or a fallback. By default we
could choose "fallback" if user just says --inodes-file-handle.

> 
> > In fact, that will be the argument for hard failure (instead of fallback).
> > Some users will say, I can fix underlying filesystem configuration if
> > I know virtiofsd can't use file handles. Falling back to O_PATH fd will
> > give them false impression that configuration is alright and file
> > handles are being used.
> > 
> > So I can imagine both kind of users. One will prefer hard failure and
> > other will be happy with fallback.
> 
> Agreed!  (Which is why I’m proposing a switch to let the user decide.)

Sounds good.

> 
> > > > If file handles can't be generated due to lack of resources in system,
> > > > then error should be returned to caller as well.
> > > > 
> > > > > There is any problem to use file handles as default?
> > > > It gives CAP_DAC_READ_SEARCH in init_user_ns. So enabling it by default
> > > > might not be desirable. Especially given the fact that we want to move
> > > > towards user namespaces and run virtiofsd with least priviliges. So
> > > > I will think user needs to enable it if they need it.
> > > > 
> > > > > I mean without
> > > > > --inode-file-handles so let them fail and force the user to use something
> > > > > like
> > > > > --no-file-handles/--force-no-file-handles with a warning.
> > > > If we were to enable it by default, we probably should test if file
> > > > handles are supported on shared dir. If yes, then enable it by default
> > > > otherwise continue to use O_PATH fd. But this will be mode switch for
> > > > the whole shared filesystem.
> > > > 
> > > > I think given we have notion of submounts and some of the submounted
> > > > filesystems might not support file handles, so key question will be
> > > > what do we do here. Do we return error in this case or fallback to
> > > > O_PATH fd for that submount. If we stick to our design philosophy,
> > > > then I would say return error. But some people might object because
> > > > they might want a mode where there is mix of filesystems in shared
> > > > dir and they want to use file handles where supported. So I am sitting
> > > > on the fence on this one.
> > > I think at this point I prefer making --inode-file-handles take an optional
> > > argument:
> > > - no: Default, don’t use file handles.
> > > - best-effort: Try to generate file handles, fall back to FDs on error.
> > Will be nice if we fallback only if filesystem does not support file
> > handles. For temporary failures, we should return errors to callers. There
> > is no good reason to fallback to O_PATH fd in that case.
> > 
> > > - mandatory: Always use file handles, return errors to the guest.
> > This sounds reasonable. Another naming scheme could be "no, fallback,
> > always".
> 
> Naming things is hard :)

Agreed. I am fine with the naming scheme you have proposed. Not sure what
will you do with "mandatory" when rootfs supports file handles but one of
the submounts does not. You will not come to know about it at startup
time and hard failures are better detected at startup time.

Vivek