On Tue, Nov 08, 2022 at 11:22:14AM +0200, Amir Goldstein wrote: > On Tue, Nov 8, 2022, 10:14 AM YoungJun.Park <her0gyugyu@xxxxxxxxx> wrote: > > > On Mon, Nov 07, 2022 at 10:49:57AM +0200, Amir Goldstein wrote: > > > On Mon, Nov 7, 2022 at 9:06 AM YoungJun.Park <her0gyugyu@xxxxxxxxx> > > wrote: > > > > > > > > On Mon, Nov 07, 2022 at 08:40:02AM +0200, Amir Goldstein wrote: > > > > > On Mon, Nov 7, 2022 at 6:38 AM YoungJun.Park <her0gyugyu@xxxxxxxxx> > > wrote: > > > > > > > > > > > > Here is my curious scenario. > > > > > > > > > > > > 1. create a file on overlayfs. > > > > > > 2. delete a file on upper directory. > > > > > > 3. can see file contents using read sys call. (may file operations > > all success) > > > > > > 4. cannot remove, rename. it return -ESTALE error (may inode > > operations fail) > > > > > > > > > > > > I understand this scenario onto the code level. > > > > > > But I don't understand this situation itself. > > > > > > > > > > > > I found a overlay kernel docs and it comments > > > > > > Changes to underlying filesystems section > > > > > > > > > > > > ... > > > > > > Changes to the underlying filesystems while part of a mounted > > overlay filesystem are not allowed. > > > > > > If the underlying filesystem is changed, the behavior of the > > overlay is undefined, > > > > > > though it will not result in a crash or deadlock. > > > > > > .... > > > > > > > > > > > > So here is my question (may it is suggestion) > > > > > > > > > > > > 1. underlying file system change is not allowed, then how about > > implementing shadow upper directory from user? > > > > > > 2. if read, write system call is allowed, how about changing > > remove, rename(and more I does not percept) operation success? > > > > > > > > > > > > > > > > What is your use case? > > > > > Why do you think this is worth spending time on? > > > > > If anything, we could implement revalidate to return ESTALE also > > from open > > > > > in such a case. > > > > > But again, why do you think that would matter? > > > > > > > > > > Thanks, > > > > > Amir. > > > > > > > > Thank you for replying. > > > > I develop antivirus scanner. > > > > When developing, I am confronted the situaion below. > > > > > > > > 1. make a docker container using overlayfs > > > > 2. our antivirus scanner detect on upperdir and remove it. > > > > 3. When I check container, the file contents can be read, buf file > > cannot be removed.(-ESTALE error) > > > > > > > > And as I think, the reason is upperdir is touchable. So it is better > > to hide upperdir. > > > > If it is hard to implement(or maybe there is a other reson that I don' > > know) > > > > it is better to make the situation is clear > > > > (file operation error, inode operations error or file operation > > success , inode operation success) > > > > > > > > > > Error on read is not an option because reading from an open and deleted > > > file is perfectly valid even without overlayfs. > > > > > > ESTALE error on open is doable and makes sense and I believe it may > > > be sufficient for your use case. > > > > > > I have an old branch that implements that behavior: > > > https://github.com/amir73il/linux/commits/ovl-revalidate > > > > > > You can try it out and see if that works for you. > > > If it does, I can post the patches. > > > > > > Note that the use case that you described does not need the last patch, > > > but if the anti-virus would have moved a lower file to quarantine > > > instead of deleting it, the last patch would also be useful for you. > > > > > > Thanks, > > > Amir. > > > > After applying the branch, I tested the scenario. > > But it does not work. file open is success on overlayfs filesystem. > > > > In my scnario, the dentry is not negative and just unhashed on upper. > > > > Yeh my bad. > I also noticed that after I sent you the link. > I think my patch also has a memleak somewhere I seen kmemleak reports > during testing. > > If we check dentry is unhashed we properly block open on my scenario. > > I write the patch and tested it working. > > (Maybe I does not catch your point, if you give a guide then I follow it) > > > > Please place the unhashed check inside ovl_revalidate_real() same as my > checks for negative upper and renamed lower dentry. > > The dentry should only be considered stale if the real dentry is unhashed > but ovl entry is hashed. > > The state of both ovl dentry and real dentry unhashed is possible and valid > I think, but it should not interfere with your use case where ovl dentry is > hashed and real upper is unhashed. > > I may be missing something so better if Miklos also takes a look at the > patch. > > Thanks, > Amir. > > > > Signed-off-by: YoungJun.park <her0gyugyu@xxxxxxxxx> > > --- > > fs/overlayfs/file.c | 4 ++++ > > 1 file changed, 4 insertions(+) > > > > diff --git a/fs/overlayfs/file.c b/fs/overlayfs/file.c > > index 6512d147c223..629dbcc49070 100644 > > --- a/fs/overlayfs/file.c > > +++ b/fs/overlayfs/file.c > > @@ -157,6 +157,10 @@ static int ovl_open(struct inode *inode, struct file > > *file) > > file->f_flags &= ~(O_CREAT | O_EXCL | O_NOCTTY | O_TRUNC); > > > > ovl_path_realdata(dentry, &realpath); > > + > > + if (d_unhashed(realpath.dentry)) > > + return -ESTALE; > > + > > realfile = ovl_open_realfile(file, &realpath); > > if (IS_ERR(realfile)) > > return PTR_ERR(realfile); > > -- > > 2.25.1 > > > > And I have one more question. > > Why upper dir must be visible..? > > The reson I think making upper dir unvisible is like the below. > > 1. If making a upperdir is unvisible, then these kind of problem disappear. > > 2. upperdir visibility makes a passage to convey container's file to > > hostland. > > (in view of container using overlayfs) > > making unvisible remove this kind of problem. > > 3. Changing upper dir scenario makes undefined behavior. So, if removing > > the interface > > user can access, then we can make the undefined scenario itself. > > > > Thanks Amir. > > Best regards > > Thank you for kind guidance Amir. Before I will do next step, I want to recheck your point Amir :) 1. > Please place the unhashed check inside ovl_revalidate_real() same as my > checks for negative upper and renamed lower dentry. You mean I modify your commit? It is from your branch then How I fix it...? (Am I misunderstood somthing?) here is pseudo patch. invalidate dentry if dentry is unhashed Signed-off-by: YoungJun.park <her0gyugyu@xxxxxxxxx> --- fs/overlayfs/super.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/fs/overlayfs/super.c b/fs/overlayfs/super.c index 6a4f2b87f1a3..411d3ed8aec1 100644 --- a/fs/overlayfs/super.c +++ b/fs/overlayfs/super.c @@ -129,8 +129,8 @@ static int ovl_revalidate_real(struct dentry *d, unsigned int flags, bool weak, { int ret = 1; - /* Invalidate dentry if real was deleted/renamed since we found it */ - if (ovl_dentry_is_dead(d) || (hash && hash != d->d_name.hash_len)) { + /* Invalidate dentry if real was deleted/renamed/unhashed since we found it */ + if (ovl_dentry_is_dead(d) || (hash && hash != d->d_name.hash_len) || d_unhashed(d)) { ret = 0; } else if (weak) { if (d->d_flags & DCACHE_OP_WEAK_REVALIDATE) -- 2.25.1 2. > The dentry should only be considered stale if the real dentry is unhashed > but ovl entry is hashed. > > The state of both ovl dentry and real dentry unhashed is possible and valid > I think, but it should not interfere with your use case where ovl dentry is > hashed and real upper is unhashed. > > I may be missing something so better if Miklos also takes a look at the > patch. > You mean, if I modify the code you said, then the patch I sent works properly? (file open fail) If you mean it, then I post my patch :) And the last question, I really curious "hide upperdir from user" idea. If it is meaningful I want to try to implement it, If it isn't then could you explain why this idea is not meaningful..? Thanks Amir.
