On Wed, Sep 20, 2023 at 2:04 PM Jan Kara [off-list ref] wrote:

On Wed 20-09-23 11:26:38, Amir Goldstein wrote:

quoted

On Mon, Apr 17, 2023 at 7:27 PM Jan Kara [off-list ref] wrote:

quoted

quoted

quoted

quoted

My motivation is to close functional gaps between fanotify and inotify.

One of the largest gaps right now is that FAN_REPORT_FID is limited
to a subset of local filesystems.

The idea is to report fid's that are "good enough" and that there
is no need to require that fid's can be used by open_by_handle_at()
because that is a non-requirement for most use cases, unpriv listener
in particular.

OK. I'd note that if you report only inode number, you are prone to the
problem that some inode gets freed (file deleted) and then reallocated (new
file created) and the resulting identifier is the same. It can be
problematic for a listener to detect these cases and deal with them.
Inotify does not have this problem at least for some cases because 'wd'
uniquely identifies the marked inode. For other cases (like watching dirs)
inotify has similar sort of problems. I'm muttering over this because in
theory filesystems not having i_generation counter on disk could approach
the problem in a similar way as FAT and then we could just use
FILEID_INO64_GEN for the file handle.

Yes, of course we could.
The problem with that is that user space needs to be able to query the fid
regardless of fanotify.

The fanotify equivalent of wd is the answer to that query.

If any fs would export i_generation via statx, then FILEID_INO64_GEN
would have been my choice.

One problem with making up i_generation (like FAT does it) is that when
inode gets reclaimed and then refetched from the disk FILEID_INO64_GEN will
change because it's going to have different i_generation. For NFS this is
annoying but I suppose it mostly does not happen since client's accesses
tend to keep the inode in memory. For fanotify it could be more likely to
happen if watching say the whole filesystem and I suppose the watching
application will get confused by this. So I'm not convinced faking
i_generation is a good thing to do. But still I want to brainstorm a bit
about it :)

quoted

But if we are going to change some other API for that, I would not change
statx(), I would change name_to_handle_at(...., AT_HANDLE_FID)

This AT_ flag would relax this check in name_to_handle_at():

        /*
         * We need to make sure whether the file system
         * support decoding of the file handle
         */
        if (!path->dentry->d_sb->s_export_op ||
            !path->dentry->d_sb->s_export_op->fh_to_dentry)
                return -EOPNOTSUPP;

And allow the call to proceed to the default export_encode_fh() implementation.
Alas, the default implementation encodes FILEID_INO32_GEN.

I think we can get away with a default implementation for FILEID_INO64_GEN
as long as the former (INO32) is used for fs with export ops without ->encode_fh
(e.g. ext*) and the latter (INO64) is used for fs with no export ops.

These default calls seem a bit too subtle to me. I'd rather add explicit
handlers to filesystems that really want FILEID_INO32_GEN encoding and then
have a fallback function for filesystems not having s_export_op at all.

But otherwise the proposal to make name_to_handle_at() work even for
filesystems not exportable through NFS makes sense to me. But I guess we
need some buy-in from VFS maintainers for this.

Hi Jan,

I seem to have dropped the ball on this after implementing AT_HANDLE_FID.
It was step one in a larger plan.

No problem, I forgot about this as well :)

quoted

Christian, Jeff,

Do you have an objection to this plan:
1. Convert all "legacy" FILEID_INO32_GEN fs with non-empty
    s_export_op and no explicit ->encode_fh() to use an explicit
    generic_encode_ino32_gen_fh()
2. Relax requirement of non-empty s_export_op for AT_HANDLE_FID
    to support encoding a (non-NFS) file id on all fs
3. For fs with empty s_export_op, allow fallback of AT_HANDLE_FID
    in exportfs_encode_inode_fh() to encode FILEID_INO64_GEN

quoted

quoted

quoted

Also I have noticed your workaround with using st_dev for fsid. As I've
checked, there are actually very few filesystems that don't set fsid these
days. So maybe we could just get away with still refusing to report on
filesystems without fsid and possibly fixup filesystems which don't set
fsid yet and are used enough so that users complain?

I started going down this path to close the gap with inotify.
inotify is capable of watching all fs including pseudo fs, so I would
like to have this feature parity.

Well, but with pseudo filesystems (similarly as with FUSE) the notification
was always unreliable. As in: some cases worked but others did not. I'm not
sure that is something we should try to replicate :)

So still I'd be interested to know which filesystems we are exactly
interested to support and whether we are not better off to explicitly add
fsid support to them like we did for tmpfs.

Since this email, kernfs derivative fs gained fsid as well.
Quoting your audit of remaining fs from another thread:

quoted

...As far as I remember
fanotify should be now able to handle anything that provides f_fsid in its
statfs(2) call. And as I'm checking filesystems not setting fsid currently are:

afs, coda, nfs - networking filesystems where inotify and fanotify have
  dubious value anyway

Be that as it may, there may be users that use inotify on network fs
and it even makes a lot of sense in controlled environments with
single NFS client per NFS export (e.g. home dirs), so I think we will
need to support those fs as well.

Fair enough.

quoted

Maybe the wise thing to do is to opt-in to monitor those fs after all?
Maybe with explicit opt-in to watch a single fs, fanotify group will
limit itself to marks on a specific sb and then a null fsid won't matter?

We have virtual filesystems with all sorts of missing or weird notification
functionality and we don't flag that in any way. So making a special flag
for network filesystems seems a bit arbitrary. I'd just make them provide
fsid and be done with it.

OK. I will try.

However, in reply to Jeff's comment:

Caution here. Most of these filesystems don't have protocol support for
anything like inotify (the known exception being SMB). You can monitor
such a network filesystem, but you won't get events for things that
happen on remote hosts.

Regardless of the fsid question, when we discussed remote notifications
support for FUSE/SMB, we raised the point that which notifications the
user gets (local or remote) are ambiguous and one suggestion was to
be explicit about requesting LOCAL or REMOTE notifications (or both).

Among the filesystems that currently support fanotify, except for the
most recent kernfs family, I think all of them are "purely local".
For the purpose of this discussion I consider debugfs and such to have
REMOTE notifications, which are not triggered from user vfs syscalls.

The one exception is smb, but only with config CIFS_NFSD_EXPORT
and that one depends on BROKEN.

If we did want to require an explicit FAN_LOCAL_NOTIF flag to allow
setting marks on fs which may have changes not via local syscalls,
it may be a good idea to flag those fs and disallow them without explicit
opt-in, before we add fanotify support to fs with missing notifications?
Perhaps before the official release of 6.6?

quoted

quoted

configfs, debugfs, devpts, efivarfs, hugetlbfs, openpromfs, proc, pstore,
ramfs, sysfs, tracefs - virtual filesystems where fsnotify functionality is
  quite limited. But some special cases could be useful. Adding fsid support
  is the same amount of trouble as for kernfs - a few LOC. In fact, we
  could perhaps add a fstype flag to indicate that this is a filesystem
  without persistent identification and so uuid should be autogenerated on
  mount (likely in alloc_super()) and f_fsid generated from sb->s_uuid.
  This way we could handle all these filesystems with trivial amount of
  effort.

Christian,

I recall that you may have had reservations on initializing s_uuid
and f_fsid in vfs code?
Does an opt-in fstype flag address your concerns?
Will you be ok with doing the tmpfs/kernfs trick for every fs
that opted-in with fstype flag in generic vfs code?

quoted

freevxfs - the only real filesystem without f_fsid. Trivial to handle one
  way or the other.

Last but not least, btrfs subvolumes.
They do have an fsid, but it is different from the sb fsid,
so we disallow (even inode) fanotify marks.

I am not sure how to solve this one,
but if we choose to implement the opt-in fanotify flag for
"watch single fs", we can make this problem go away, along
with the problem of network fs fsid and other odd fs that we
do not want to have to deal with.

On top of everything, it is a fast solution and it doesn't
involve vfs and changing any fs at all.

Yeah, right, forgot about this one. Thanks for reminding me. But this is
mostly a kernel internal implementation issue and doesn't seem to be a
principial problem so I'd prefer not to complicate the uAPI for this. We
could for example mandate a special super_operation for fetching fsid for a
dentry for filesystems which don't have uniform fsids over the whole
filesystem (i.e., btrfs) and call this when generating event for such
filesystems. Or am I missing some other complication?

The problem is the other way around :)
btrfs_statfs() takes a dentry and returns the fsid of the subvolume.
That is the fsid that users will get when querying the path to be marked.

If users had a flag to statfs() to request the "btrfs root volume fsid",
then fanotify could also report the root fsid and everyone will be happy
because the btrfs file handle already contains the subvolume root
object id (FILEID_BTRFS_WITH_PARENT_ROOT), but that is not
what users get for statfs() and that is not what fanotify documentation
says about how to query fsid.

We could report the subvolume fsid for marked inode/mount
that is not a problem - we just cache the subvol fsid in inode/mount
connector, but that fsid will be inconsistent with the fsid in the sb
connector, so the same object (in subvolume) can get events
with different fsid (e.g. if one event is in mask of sb and another
event is in mask of inode).

As Jeff said, nfsd also have issues with exporting btrfs subvolumes,
because of these oddities and I have no desire to solve those issues.

I think we could relax the EXDEV case for unpriv fanotify, because
inode marks should not have this problem?

That would be an easy way to get feature parity with inotify wrt btrfs.

quoted

quoted

quoted

If we can get away with fallback to s_dev as fsid in vfs_statfs()
I have no problem with that, but just to point out - functionally
it is equivalent to do this fallback in userspace library as the
fanotify_get_fid() LTP helper does.

Yes, userspace can workaround this but I was more thinking about avoiding
adding these workarounds into fanotify in kernel *and* to userspace.

quoted

quoted

quoted

I chose a rather generic name for the flag to opt-in for "good enough"
fid's.  At first, I was going to make those fid's self describing the
fact that they are not NFS file handles, but in the name of simplicity
to the API, I decided that this is not needed.

I'd like to discuss a bit about the meaning of the flag. On the first look
it is a bit strange to have a flag that says "give me a fh, if you don't
have it, give me ino". It would seem cleaner to have "give me fh" kind of
interface (FAN_REPORT_FID) and "give me ino" kind of interface (new
FAN_REPORT_* flag). I suspect you've chosen the more complex meaning
because you want to allow a usecase where watches of filesystems which
don't support filehandles are mixed with watches of filesystems which do
support filehandles in one notification group and getting filehandles is
actually prefered over getting just inode numbers? Do you see real benefit
in getting file handles when userspace has to implement fallback for
getting just inode numbers anyway?

Yes, there is a benefit, because a real fhandle has no-reuse guarantee.

Even if we implement the kernel fallback to FILEID_INO64_GEN, it does
not serve as a statement from the filesystem that i_generation is useful
and in fact, i_generation will often be zero in simple fs and ino will be
reusable.

Also, I wanted to have a design where a given fs/object always returns
the same FID regardless of the init flags.

Your question implies that if
"userspace has to implement fallback for getting just inode numbers",
then it doesn't matter if we report fhandle or inode, but it is not accurate.

The fanotify_get_fid() LTP helper always gets a consistent FID for a
given fs/object. You do not need to feed it the fanotify init flags to
provide a consistent answer.

For all the reasons above, I think that a "give me ino'' flag is not useful.
IMO, the flag just needs better marketing.
This is a "I do not need/intend to open_by_handle flag".
Suggestions for a better name are welcome.

I see, yes, these reasons make sense.

quoted

For all I care, we do not need to add an opt-in flag at all.
We could simply start to support fs that were not supported before.
This sort of API change is very common and acceptable.

There is no risk if the user tries to call open_by_handle_at() with the
fanotify encoded FID, because in this case the fs is guaranteed to
return ESTALE, because fs does not support file handles.

This is especially true, if we can get away with seamless change
of behavior for vfs_statfs(), because that seamless change would
cause FAN_REPORT_FID to start working on fs like fuse that
support file handles and have zero fsid.

Yeah. Actually I like the idea of a seamless change to start reporting fsid
and also to start reporting "fake" handles. In the past we've already
enabled tmpfs like this...

I am now leaning towards a combination of:
1. Seamless change of behavior for vfs_statfs() and
    name_to_handle_at(..., AT_HANDLE_FID) for the simple cases
    using an opt-in fstype flag

Ack.

quoted

AND
2. Simple interim fallback for other fs with an opt-in fanotify flag (*)

I'm not sold on the special flag yet (see above) ;).

OK, will try to make progress without an opt-in flag.

Thanks,
Amir.