[PATCH RFC 36/38] selftests/proc: check that faulting sysctl accesses don't hold up removal
sashiko-bot at kernel.org
sashiko-bot at kernel.org
Tue Nov 3 06:48:40 PST 2026
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] selftests/proc: getdents test faults in dir_emit_dots() before entry lookup
--
commit fc0ab3f791b6ab04b43d13e015f97882092d54fc
Author: Christian Brauner <brauner at kernel.org>
selftests/proc: check that faulting sysctl accesses don't hold up removal
This commit adds a proc selftest using userfaultfd and a tun device in a new
network namespace to verify that page faults during read(), write(), and
getdents64() on /proc/sys do not block sysctl table unregistration.
> diff --git a/tools/testing/selftests/proc/proc-remove-slow-sysctl.c b/tools/testing/selftests/proc/proc-remove-slow-sysctl.c
> new file mode 100644
> index 0000000000000..e1b1fd19a07ab
> --- /dev/null
> +++ b/tools/testing/selftests/proc/proc-remove-slow-sysctl.c
> @@ -0,0 +1,213 @@
[ ... ]
> +TEST_F(proc_sys_slow, getdents)
> +{
> + struct worker w = {
> + .op = SYSCTL_GETDENTS,
> + .len = BUF_PAGES * self->psz,
> + };
> +
> + w.fd = open(CONF_PATH, O_RDONLY | O_DIRECTORY | O_CLOEXEC);
> + ASSERT_GE(w.fd, 0);
> + stall_and_remove(_metadata, self, &w);
> +
> + /* What was looked up before the removal still arrives, then nothing. */
> + EXPECT_GT(w.ret, 0);
[Severity: Medium]
Does TEST_F(proc_sys_slow, getdents) in
tools/testing/selftests/proc/proc-remove-slow-sysctl.c actually fault on a
looked-up sysctl entry, or does it fault inside dir_emit_dots() before any
entry is looked up?
In FIXTURE_SETUP(proc_sys_slow), only the first page of self->buf is
registered with userfaultfd. When TEST_F(proc_sys_slow, getdents) opens
CONF_PATH, w.fd starts at file position 0 and is passed directly to
stall_and_remove() -> work_thread() -> getdents64().
In fs/proc/proc_sysctl.c:proc_sys_readdir(), dir_emit_dots() runs before
sysctl_dir_next():
fs/proc/proc_sysctl.c:proc_sys_readdir() {
...
if (!dir_emit_dots(file, ctx))
return 0;
if (cursor && cursor->pos == ctx->pos) {
err = sysctl_dir_next(ctl_dir, &d, cursor->name, cursor->len);
} else {
loff_t pos = 2;
err = sysctl_dir_next(ctl_dir, &d, NULL, 0);
...
}
Because ctx->pos is 0, dir_emit_dots() writes "." at the start of self->buf
and triggers the single userfaultfd page fault before sysctl_dir_next() is
called and while d.head is still NULL.
While work_thread() is sleeping in dir_emit_dots(), stall_and_remove() closes
self->tun (which unregisters CONF_PATH and marks ctl_dir->header as
unregistering) and resolves the fault on the first page with UFFDIO_ZEROPAGE.
When work_thread() resumes, dir_emit_dots() finishes emitting "." and "..",
sysctl_dir_next() immediately returns -ENOENT, and getdents64() returns the
48 bytes from the dot entries without ever entering the entry-emission loop.
Should TEST_F(proc_sys_slow, getdents) advance w.fd past "." and ".." first
(for example with lseek(w.fd, 2, SEEK_SET)) before calling stall_and_remove()
so the fault occurs in dir_emit() after sysctl_dir_next() and
proc_sys_fill_dirent() have looked up an entry?
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261102-work-procfs-bugfixes-v1-0-59422ed3e465@kernel.org?part=36
More information about the kexec
mailing list