Re: [PATCH] idmapped-mounts: probe O_TMPFILE support with the caller's mapped id

[Zorro Lang <zlang@kernel.org>]

On Thu, Aug 06, 2026 at 02:11:55PM +0800, Zizhi Wo wrote:
> friendly ping...

Hi Zizhi, actually I'm waiting for the fix from vfs list.

Hi Christian,

Could you please take a look at this issue? Quite a few people have submitted
patches trying to fix this test issue. Given that you are the authority on
this topic and previously sent a patchset [1] to fix this kind of issue, I think
accepting your patch would be the best approach to stay aligned with upstream
VFS. Christoph also metioned that [2] last time. I don't want to make this issue
fixing any more complicated. Could you please provide your review point or
rebase your fix on the latest fstests for-next branch, we'll review it ASAP.

[1]
https://lore.kernel.org/fstests/20260615-overlay-idmapped-vfstest-v1-1-c6b3b2228092@kernel.org/

[2]
https://lore.kernel.org/fstests/amCGrQi2sCcgSgHL@infradead.org/

Thanks,
Zorro

> 
> 在 2026/7/31 17:02, Zizhi Wo 写道:
> > From: Zizhi Wo <wozizhi@huawei.com>
> > 
> > Kernel commit 539dce114465 ("fs: refuse O_TMPFILE creation with an
> > unmapped fsuid or fsgid") made vfs_tmpfile() return -EOVERFLOW when the
> > caller's fsuid/fsgid has no mapping in the filesystem's user namespace
> > after applying the mount's idmapping. This is a no-op on non-idmapped
> > mounts and only takes effect on an idmapped mount that does not map the
> > caller.
> > 
> > The idmapped-mount testcases behind generic/696 and generic/697 build an
> > idmapping (e.g. "0 10000 10000") that covers host uid 10000-19999 but
> > not uid 0, then call openat_tmpfile_supported() from the parent process
> > that is still real root (fsuid 0). The probe now gets -EOVERFLOW, falls
> > into the log_errno() branch, and prints
> > 
> >    utils.c: 928: openat_tmpfile_supported - Value too large for defined \
> > data type - failure: create
> > 
> > to stderr. Both testcases expect "Silence is golden", so this spurious
> > output fails them, even though the kernel behavior is correct and the
> > testcases' own assertions all pass.
> > 
> > Move each openat_tmpfile_supported() call into the child, after
> > switch_ids()/switch_userns(), so it probes with the mapped identity the
> > tmpfile section uses. Silencing EOVERFLOW in openat_tmpfile_supported()
> > would make it indistinguishable whether the idmapping genuinely has no
> > mapping for any caller, or the probe just ran in the unmapped parent before
> > the switch.
> > 
> > Signed-off-by: Zizhi Wo <wozizhi@huawei.com>
> > ---
> >   src/vfs/idmapped-mounts.c | 32 ++++++++++++++++++++------------
> >   1 file changed, 20 insertions(+), 12 deletions(-)
> > 
> > diff --git a/src/vfs/idmapped-mounts.c b/src/vfs/idmapped-mounts.c
> > index ed9992f9..bfbe9042 100644
> > --- a/src/vfs/idmapped-mounts.c
> > +++ b/src/vfs/idmapped-mounts.c
> > @@ -3838,8 +3838,6 @@ int tcore_setgid_create_idmapped(const struct vfstest_info *info)
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	pid = fork();
> >   	if (pid < 0) {
> >   		log_stderr("failure: fork");
> > @@ -3849,6 +3847,8 @@ int tcore_setgid_create_idmapped(const struct vfstest_info *info)
> >   		if (!switch_ids(10000, 11000))
> >   			die("failure: switch fsids");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		/* create regular file via open() */
> >   		file1_fd = openat(open_tree_fd, FILE1, O_CREAT | O_EXCL | O_CLOEXEC, S_IXGRP | S_ISGID);
> >   		if (file1_fd < 0)
> > @@ -4014,8 +4014,6 @@ int tcore_setgid_create_idmapped_in_userns(const struct vfstest_info *info)
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	pid = fork();
> >   	if (pid < 0) {
> >   		log_stderr("failure: fork");
> > @@ -4025,6 +4023,8 @@ int tcore_setgid_create_idmapped_in_userns(const struct vfstest_info *info)
> >   		if (!switch_userns(attr.userns_fd, 0, 0, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		/* create regular file via open() */
> >   		file1_fd = openat(open_tree_fd, FILE1, O_CREAT | O_EXCL | O_CLOEXEC, S_IXGRP | S_ISGID);
> >   		if (file1_fd < 0)
> > @@ -4133,6 +4133,8 @@ int tcore_setgid_create_idmapped_in_userns(const struct vfstest_info *info)
> >   		if (!switch_userns(attr.userns_fd, 0, 0, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		if (!caps_down_fsetid())
> >   			die("failure: caps_down_fsetid");
> > @@ -4257,6 +4259,8 @@ int tcore_setgid_create_idmapped_in_userns(const struct vfstest_info *info)
> >   		if (!switch_userns(attr.userns_fd, 0, 1000, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		if (!caps_down_fsetid())
> >   			die("failure: caps_down_fsetid");
> > @@ -7733,8 +7737,6 @@ static int setgid_create_umask_idmapped(const struct vfstest_info *info)
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	pid = fork();
> >   	if (pid < 0) {
> >   		log_stderr("failure: fork");
> > @@ -7752,6 +7754,8 @@ static int setgid_create_umask_idmapped(const struct vfstest_info *info)
> >   		if (!switch_ids(10000, 11000))
> >   			die("failure: switch fsids");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		/* create regular file via open() */
> >   		file1_fd = openat(open_tree_fd, FILE1, O_CREAT | O_EXCL | O_CLOEXEC, S_IXGRP | S_ISGID);
> >   		if (file1_fd < 0)
> > @@ -7947,8 +7951,6 @@ static int setgid_create_umask_idmapped_in_userns(const struct vfstest_info *inf
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	/*
> >   	 * Below we verify that setgid inheritance for a newly created file or
> >   	 * directory works correctly. As part of this we need to verify that
> > @@ -7984,6 +7986,8 @@ static int setgid_create_umask_idmapped_in_userns(const struct vfstest_info *inf
> >   		if (!switch_userns(attr.userns_fd, 0, 0, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		if (!caps_down_fsetid())
> >   			die("failure: caps_down_fsetid");
> > @@ -8181,8 +8185,6 @@ static int setgid_create_acl_idmapped(const struct vfstest_info *info)
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	pid = fork();
> >   	if (pid < 0) {
> >   		log_stderr("failure: fork");
> > @@ -8207,6 +8209,8 @@ static int setgid_create_acl_idmapped(const struct vfstest_info *info)
> >   		if (!switch_ids(10000, 11000))
> >   			die("failure: switch fsids");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		/* create regular file via open() */
> >   		file1_fd = openat(open_tree_fd, FILE1, O_CREAT | O_EXCL | O_CLOEXEC, S_IXGRP | S_ISGID);
> >   		if (file1_fd < 0)
> > @@ -8342,6 +8346,8 @@ static int setgid_create_acl_idmapped(const struct vfstest_info *info)
> >   		if (!switch_ids(10000, 11000))
> >   			die("failure: switch fsids");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		/* create regular file via open() */
> >   		file1_fd = openat(open_tree_fd, FILE1, O_CREAT | O_EXCL | O_CLOEXEC, S_IXGRP | S_ISGID);
> >   		if (file1_fd < 0)
> > @@ -8536,8 +8542,6 @@ static int setgid_create_acl_idmapped_in_userns(const struct vfstest_info *info)
> >   		goto out;
> >   	}
> > -	supported = openat_tmpfile_supported(open_tree_fd);
> > -
> >   	/*
> >   	 * Below we verify that setgid inheritance for a newly created file or
> >   	 * directory works correctly. As part of this we need to verify that
> > @@ -8580,6 +8584,8 @@ static int setgid_create_acl_idmapped_in_userns(const struct vfstest_info *info)
> >   		if (!switch_userns(attr.userns_fd, 0, 0, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		if (!caps_down_fsetid())
> >   			die("failure: caps_down_fsetid");
> > @@ -8723,6 +8729,8 @@ static int setgid_create_acl_idmapped_in_userns(const struct vfstest_info *info)
> >   		if (!switch_userns(attr.userns_fd, 0, 0, false))
> >   			die("failure: switch_userns");
> > +		supported = openat_tmpfile_supported(open_tree_fd);
> > +
> >   		if (!caps_down_fsetid())
> >   			die("failure: caps_down_fsetid");
>