Message-ID: <CAJuCfpFmvMU9K3JkHsR6RrUYaGFsQ9BVapt2mK=_d7O-tcNa2g@mail.gmail.com>
Date:   Fri, 23 Jun 2023 18:58:41 -0700
From:   Suren Baghdasaryan <surenb@...gle.com>
To:     "lujialin (A)" <lujialin4@...wei.com>
Cc:     Eric Biggers <ebiggers@...nel.org>, Tejun Heo <tj@...nel.org>,
        Johannes Weiner <hannes@...xchg.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Christian Brauner <brauner@...nel.org>,
        Oleg Nesterov <oleg@...hat.com>, linux-kernel@...r.kernel.org,
        linux-fsdevel@...r.kernel.org
Subject: Re: [PATCH v2] poll: Fix use-after-free in poll_freewait()

On Tue, Jun 20, 2023 at 5:09 PM Suren Baghdasaryan <surenb@...gle.com> wrote:
>
> On Sun, Jun 18, 2023 at 6:28 AM lujialin (A) <lujialin4@...wei.com> wrote:
> >
> > Hi Suren:
> >
> > kernel config:
> > x86_64_defconfig
> > CONFIG_PSI=y
> > CONFIG_SLUB_DEBUG=y
> > CONFIG_SLUB_DEBUG_ON=y
> > CONFIG_KASAN=y
> > CONFIG_KASAN_INLINE=y
> >
> > I make some change in code, in order to increase the recurrence probability.
> > diff --git a/fs/select.c b/fs/select.c
> > index 5edffee1162c..5ee5b74a8386 100644
> > --- a/fs/select.c
> > +++ b/fs/select.c
> > @@ -139,6 +139,7 @@ void poll_freewait(struct poll_wqueues *pwq)
> >   {
> >          struct poll_table_page * p = pwq->table;
> >          int i;
> > +       mdelay(50);
> >          for (i = 0; i < pwq->inline_index; i++)
> >                  free_poll_entry(pwq->inline_entries + i);
> >          while (p) {
> >
> > Here is the simple repo test.sh:
> > #!/bin/bash
> >
> > RESOURCE_TYPES=("cpu" "memory" "io" "irq")
> > #RESOURCE_TYPES=("cpu")
> > cgroup_num=50
> > test_dir=/sys/fs/cgroup/test
> >
> > function restart_cgroup() {
> >          num=$(expr $RANDOM % $cgroup_num + 1)
> >          rmdir $test_dir/test_$num
> >          mkdir $test_dir/test_$num
> > }
> >
> > function create_triggers() {
> >          num=$(expr $RANDOM % $cgroup_num + 1)
> >          random=$(expr $RANDOM % "${#RESOURCE_TYPES[@]}")
> >          psi_type="${RESOURCE_TYPES[${random}]}"
> >          ./psi_monitor $test_dir/test_$num $psi_type &
> > }
> >
> > mkdir $test_dir
> > for i in $(seq 1 $cgroup_num)
> > do
> >          mkdir $test_dir/test_$i
> > done
> > for j in $(seq 1 100)
> > do
> >          restart_cgroup &
> >          create_triggers &
> > done
> >
> > psi_monitor.c:
> > #include <errno.h>
> > #include <fcntl.h>
> > #include <stdio.h>
> > #include <poll.h>
> > #include <string.h>
> > #include <unistd.h>
> >
> > int main(int argc, char *argv[]) {
> >          const char trig[] = "full 1000000 1000000";
> >          struct pollfd fds;
> >          char filename[100];
> >
> >          sprintf(filename, "%s/%s.pressure", argv[1], argv[2]);
> >
> >          fds.fd = open(filename, O_RDWR | O_NONBLOCK);
> >          if (fds.fd < 0) {
> >                  printf("%s open error: %s\n", filename,strerror(errno));
> >                  return 1;
> >          }
> >          fds.events = POLLPRI;
> >          if (write(fds.fd, trig, strlen(trig) + 1) < 0) {
> >                  printf("%s write error: %s\n",filename,strerror(errno));
> >                  return 1;
> >          }
> >          while (1) {
> >                  poll(&fds, 1, -1);
> >          }
> >          close(fds.fd);
> >          return 0;
> > }
>
> Thanks a lot!
> I'll try to get this reproduced and fixed by the end of this week.

Ok, I was able to reproduce the issue and I think the ultimate problem
here is that kernfs_release_file() can be called from both
kernfs_fop_release() and kernfs_drain_open_files(). While
kernfs_fop_release() is called when the FD's refcount is 0 and there
are no users, kernfs_drain_open_files() can be called while there are
still other users. In our scenario, kn->attr.ops->release points to
cgroup_pressure_release() which destroys the psi trigger thinking that
(since the file is "released") there could be no other users. So,
shell process which issues the rmdir command will destroy the trigger
once cgroup_pressure_release() is called and the reproducer will step
on the freed wait_queue_head. The way kernfs_release_file() is
implemented, it ensures that kn->attr.ops->release(of) is called only
once (the first time), therefore cgroup_pressure_release() is never
called in reproducer's context. That prevents me from implementing
some kind of refcounting schema for psi triggers because we are never
notified when the last user is gone.
I think to fix this I would need to modify kernfs_release_file() and
maybe add another operation in kernfs_ops to indicate that the last
user is gone (smth like final_release()). It's not pretty but so far I
did not find a better way. I'll think some more over the weekend and
will try to post a patch implementing the fix on Monday.
Thanks,
Suren.


> Suren.
>
> > Thanks,
> > Lu
> > 在 2023/6/16 7:13, Suren Baghdasaryan 写道:
> > > On Wed, Jun 14, 2023 at 11:19 AM Suren Baghdasaryan <surenb@...gle.com> wrote:
> > >>
> > >> On Wed, Jun 14, 2023 at 10:40 AM Eric Biggers <ebiggers@...nel.org> wrote:
> > >>>
> > >>> On Wed, Jun 14, 2023 at 03:07:33PM +0800, Lu Jialin wrote:
> > >>>> We found a UAF bug in remove_wait_queue as follows:
> > >>>>
> > >>>> ==================================================================
> > >>>> BUG: KASAN: use-after-free in _raw_spin_lock_irqsave+0x71/0xe0
> > >>>> Write of size 4 at addr ffff8881150d7b28 by task psi_trigger/15306
> > >>>> Call Trace:
> > >>>>   dump_stack+0x9c/0xd3
> > >>>>   print_address_description.constprop.0+0x19/0x170
> > >>>>   __kasan_report.cold+0x6c/0x84
> > >>>>   kasan_report+0x3a/0x50
> > >>>>   check_memory_region+0xfd/0x1f0
> > >>>>   _raw_spin_lock_irqsave+0x71/0xe0
> > >>>>   remove_wait_queue+0x26/0xc0
> > >>>>   poll_freewait+0x6b/0x120
> > >>>>   do_sys_poll+0x305/0x400
> > >>>>   do_syscall_64+0x33/0x40
> > >>>>   entry_SYSCALL_64_after_hwframe+0x61/0xc6
> > >>>>
> > >>>> Allocated by task 15306:
> > >>>>   kasan_save_stack+0x1b/0x40
> > >>>>   __kasan_kmalloc.constprop.0+0xb5/0xe0
> > >>>>   psi_trigger_create.part.0+0xfc/0x450
> > >>>>   cgroup_pressure_write+0xfc/0x3b0
> > >>>>   cgroup_file_write+0x1b3/0x390
> > >>>>   kernfs_fop_write_iter+0x224/0x2e0
> > >>>>   new_sync_write+0x2ac/0x3a0
> > >>>>   vfs_write+0x365/0x430
> > >>>>   ksys_write+0xd5/0x1b0
> > >>>>   do_syscall_64+0x33/0x40
> > >>>>   entry_SYSCALL_64_after_hwframe+0x61/0xc6
> > >>>>
> > >>>> Freed by task 15850:
> > >>>>   kasan_save_stack+0x1b/0x40
> > >>>>   kasan_set_track+0x1c/0x30
> > >>>>   kasan_set_free_info+0x20/0x40
> > >>>>   __kasan_slab_free+0x151/0x180
> > >>>>   kfree+0xba/0x680
> > >>>>   cgroup_file_release+0x5c/0xe0
> > >>>>   kernfs_drain_open_files+0x122/0x1e0
> > >>>>   kernfs_drain+0xff/0x1e0
> > >>>>   __kernfs_remove.part.0+0x1d1/0x3b0
> > >>>>   kernfs_remove_by_name_ns+0x89/0xf0
> > >>>>   cgroup_addrm_files+0x393/0x3d0
> > >>>>   css_clear_dir+0x8f/0x120
> > >>>>   kill_css+0x41/0xd0
> > >>>>   cgroup_destroy_locked+0x166/0x300
> > >>>>   cgroup_rmdir+0x37/0x140
> > >>>>   kernfs_iop_rmdir+0xbb/0xf0
> > >>>>   vfs_rmdir.part.0+0xa5/0x230
> > >>>>   do_rmdir+0x2e0/0x320
> > >>>>   __x64_sys_unlinkat+0x99/0xc0
> > >>>>   do_syscall_64+0x33/0x40
> > >>>>   entry_SYSCALL_64_after_hwframe+0x61/0xc6
> > >>>> ==================================================================
> > >>>>
> > >>>> If using epoll(), wake_up_pollfree will empty waitqueue and set
> > >>>> wait_queue_head is NULL before free waitqueue of psi trigger. But is
> > >>>> doesn't work when using poll(), which will lead a UAF problem in
> > >>>> poll_freewait coms as following:
> > >>>>
> > >>>> (cgroup_rmdir)                      |
> > >>>> psi_trigger_destroy                 |
> > >>>>    wake_up_pollfree(&t->event_wait)  |
> > >>>>     synchronize_rcu();               |
> > >>>>      kfree(t)                        |
> > >>>>                                    |   (poll_freewait)
> > >>>>                                    |     free_poll_entry(pwq->inline_entries + i)
> > >>>>                                    |       remove_wait_queue(entry->wait_address)
> > >>>>                                    |         spin_lock_irqsave(&wq_head->lock)
> > >>>>
> > >>>> entry->wait_address in poll_freewait() is t->event_wait in cgroup_rmdir().
> > >>>> t->event_wait is free in psi_trigger_destroy before call poll_freewait(),
> > >>>> therefore wq_head in poll_freewait() has been already freed, which would
> > >>>> lead to a UAF.
> > >
> > > Hi Lu,
> > > Could you please share your reproducer along with the kernel config
> > > you used? I'm trying to reproduce this UAF but every time I delete the
> > > cgroup being polled, poll() simply returns POLLERR.
> > > Thanks,
> > > Suren.
> > >
> > >>>>
> > >>>> similar problem for epoll() has been fixed commit c2dbe32d5db5
> > >>>> ("sched/psi: Fix use-after-free in ep_remove_wait_queue()").
> > >>>> epoll wakeup function ep_poll_callback() will empty waitqueue and set
> > >>>> wait_queue_head is NULL when pollflags is POLLFREE and judge pwq->whead
> > >>>> is NULL or not before remove_wait_queue in ep_remove_wait_queue(),
> > >>>> which will fix the UAF bug in ep_remove_wait_queue.
> > >>>>
> > >>>> But poll wakeup function pollwake() doesn't do that. To fix the
> > >>>> problem, we empty waitqueue and set wait_address is NULL in pollwake() when
> > >>>> key is POLLFREE. otherwise in remove_wait_queue, which is similar to
> > >>>> epoll().
> > >>>>
> > >>>> Fixes: 0e94682b73bf ("psi: introduce psi monitor")
> > >>>> Suggested-by: Suren Baghdasaryan <surenb@...gle.com>
> > >>>> Link: https://lore.kernel.org/all/CAJuCfpEoCRHkJF-=1Go9E94wchB4BzwQ1E3vHGWxNe+tEmSJoA@mail.gmail.com/#t
> > >>>> Signed-off-by: Lu Jialin <lujialin4@...wei.com>
> > >>>> ---
> > >>>> v2: correct commit msg and title suggested by Suren Baghdasaryan
> > >>>> ---
> > >>>>   fs/select.c | 20 +++++++++++++++++++-
> > >>>>   1 file changed, 19 insertions(+), 1 deletion(-)
> > >>>>
> > >>>> diff --git a/fs/select.c b/fs/select.c
> > >>>> index 0ee55af1a55c..e64c7b4e9959 100644
> > >>>> --- a/fs/select.c
> > >>>> +++ b/fs/select.c
> > >>>> @@ -132,7 +132,17 @@ EXPORT_SYMBOL(poll_initwait);
> > >>>>
> > >>>>   static void free_poll_entry(struct poll_table_entry *entry)
> > >>>>   {
> > >>>> -     remove_wait_queue(entry->wait_address, &entry->wait);
> > >>>> +     wait_queue_head_t *whead;
> > >>>> +
> > >>>> +     rcu_read_lock();
> > >>>> +     /* If it is cleared by POLLFREE, it should be rcu-safe.
> > >>>> +      * If we read NULL we need a barrier paired with smp_store_release()
> > >>>> +      * in pollwake().
> > >>>> +      */
> > >>>> +     whead = smp_load_acquire(&entry->wait_address);
> > >>>> +     if (whead)
> > >>>> +             remove_wait_queue(whead, &entry->wait);
> > >>>> +     rcu_read_unlock();
> > >>>>        fput(entry->filp);
> > >>>>   }
> > >>>>
> > >>>> @@ -215,6 +225,14 @@ static int pollwake(wait_queue_entry_t *wait, unsigned mode, int sync, void *key
> > >>>>        entry = container_of(wait, struct poll_table_entry, wait);
> > >>>>        if (key && !(key_to_poll(key) & entry->key))
> > >>>>                return 0;
> > >>>> +     if (key_to_poll(key) & POLLFREE) {
> > >>>> +             list_del_init(&wait->entry);
> > >>>> +             /* wait_address !=NULL protects us from the race with
> > >>>> +              * poll_freewait().
> > >>>> +              */
> > >>>> +             smp_store_release(&entry->wait_address, NULL);
> > >>>> +             return 0;
> > >>>> +     }
> > >>>>        return __pollwake(wait, mode, sync, key);
> > >>>
> > >>> I don't understand why this patch is needed.
> > >>>
> > >>> The last time I looked at POLLFREE, it is only needed because of asynchronous
> > >>> polls.  See my explanation in the commit message of commit 50252e4b5e989ce6.
> > >>
> > >> Ah, I missed that. Thanks for the correction.
> > >>
> > >>>
> > >>> In summary, POLLFREE solves the problem of polled waitqueues whose lifetime is
> > >>> tied to the current task rather than to the file being polled.  Also refer to
> > >>> the comment above wake_up_pollfree(), which mentions this.
> > >>>
> > >>> fs/select.c is synchronous polling, not asynchronous.  Therefore, it should not
> > >>> need to handle POLLFREE.
> > >>>
> > >>> If there's actually a bug here, most likely it's a bug in psi_trigger_poll()
> > >>> where it is using a waitqueue whose lifetime is tied to neither the current task
> > >>> nor the file being polled.  That needs to be fixed.
> > >>
> > >> Yeah. We discussed this issue in
> > >> https://lore.kernel.org/all/CAJuCfpFb0J5ZwO6kncjRG0_4jQLXUy-_dicpH5uGiWP8aKYEJQ@mail.gmail.com
> > >> and the root cause is that cgroup_file_release() where
> > >> psi_trigger_destroy() is called is not tied to the cgroup file's real
> > >> lifetime (see my analysis here:
> > >> https://lore.kernel.org/all/CAJuCfpFZ3B4530TgsSHqp5F_gwfrDujwRYewKReJru==MdEHQg@mail.gmail.com/#t).
> > >> I guess it's time to do a deeper surgery and figure out a way to call
> > >> psi_trigger_destroy() when the polled cgroup file is actually being
> > >> destroyed. I'll take a closer look into this later today.
> > >> A fix will likely require some cgroup or kernfs code changes, so
> > >> CC'ing Tejun for visibility.
> > >> Thanks,
> > >> Suren.
> > >>
> > >>>
> > >>> - Eric
> > > .
> > >

Powered by blists - more mailing lists