epoll: Fix UAF dentry name access in wakeup source setup
PATCH 1/1 epoll: Fix UAF dentry name access in wakeup source setup
In ep_create_wakeup_source(), epi->ffd.file is some random file we're watching with epoll, so it might well be renamed concurrently. And when a file gets renamed, the buffer containing its name may be freed.
This can be reproduced by racing a task that keeps adding and removing EPOLLWAKEUP epoll entries for a fifo with another task that keeps renaming the fifo between two long names if you add an mdelay(200) call directly before wakeup_source_register(); KASAN then complains:
BUG: KASAN: use-after-free in strlen+0xa/0x40 Read of size 1 at addr ffff888065fda990 by task wakemeup/2375 [...] Call Trace: [...] strlen+0xa/0x40 kstrdup+0x1a/0x60 wakeup_source_create+0x43/0xb0 wakeup_source_register+0x13/0x60 ep_create_wakeup_source+0x7f/0xf0 do_epoll_ctl+0x13d0/0x1880 [...] __x64_sys_epoll_ctl+0xc3/0x110 [...] Allocated by task 2376: [...] __d_alloc+0x323/0x3c0 d_alloc+0x30/0xf0 __lookup_hash+0x61/0xc0 do_renameat2+0x3fa/0x6d0 __x64_sys_rename+0x3a/0x40 [...] Freed by task 2379:
[...]
kfree_rcu_work+0x9b/0x5d0
[...]
Backporting note: This patch depends on commit 49d31c2f389a ("dentry name snapshots"). Maybe that one should also be backported as a dependency for pre-v4.13? (Sorry, I wasn't sure how to properly express this as a "Fixes:" tag.)
---
I'm guessing this will go through akpm's tree?
fs/eventpoll.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
fs/eventpoll.c +4 -3
@@ -1450,7 +1450,7 @@ static int reverse_path_check(void)static int ep_create_wakeup_source (struct epitem *epi) { -const char *name; +struct name_snapshot name;struct wakeup_source *ws;if (!epi->ep->ws) {
@@ -1459,8 +1459,9 @@ static int ep_create_wakeup_source(struct epitem *epi)return -ENOMEM ; } - name = epi->ffd.file->f_path.dentry->d_name.name; - ws =wakeup_source_register (NULL , name); +take_dentry_name_snapshot (&name, epi->ffd.file->f_path.dentry); + ws =wakeup_source_register (NULL , name.name.name); +release_dentry_name_snapshot (&name);if (!ws)return -ENOMEM ;
base-commit: 96c9a7802af7d500a582d89a8b864584fe878c1b
Signature · 1 line
2.26.2.303.gf8c07b1a785-goog
Al Viro
maintainerQuoted 19 lines · Jann Horn
I'm guessing this will go through akpm's tree?
fs/eventpoll.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)fs/eventpoll.c +1
-1@@ -1450,7 +1450,7 @@ static int reverse_path_check(void)static int ep_create_wakeup_source (struct epitem *epi) { -const char *name; +struct name_snapshot name;struct wakeup_source *ws;if (!epi->ep->ws) {
fs/eventpoll.c static int ep_create_wakeup_source(struct epitem *epi) · line 1464
Quoted 8 lines
fs/eventpoll.c +3
-2@@ -1459,8 +1459,9 @@ static int ep_create_wakeup_source(struct epitem *epi)return -ENOMEM ; } - name = epi->ffd.file->f_path.dentry->d_name.name; - ws =wakeup_source_register (NULL , name); +take_dentry_name_snapshot (&name, epi->ffd.file->f_path.dentry); + ws =wakeup_source_register (NULL , name.name.name); +release_dentry_name_snapshot (&name);
I'm not sure I like it. Sure, it won't get freed under you that way; it still can go absolutely stale by the time you return from wakeup_source_register(). What is it being used for?
Quoted 23 lines · Al Viro
Quoted 22 lines · Jann Horn
I'm guessing this will go through akpm's tree?
fs/eventpoll.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)fs/eventpoll.c +1
-1@@ -1450,7 +1450,7 @@ static int reverse_path_check(void)static int ep_create_wakeup_source (struct epitem *epi) { -const char *name; +struct name_snapshot name;struct wakeup_source *ws;if (!epi->ep->ws) {@@ -1459,8 +1459,9 @@ static int ep_create_wakeup_source(struct epitem *epi)return -ENOMEM ; }
Quoted 9 lines
fs/eventpoll.c static int ep_create_wakeup_source(struct epitem *epi) · line 1464
Quoted 5 lines
fs/eventpoll.c +3
-2- name = epi->ffd.file->f_path.dentry->d_name.name; - ws =wakeup_source_register (NULL , name); +take_dentry_name_snapshot (&name, epi->ffd.file->f_path.dentry); + ws =wakeup_source_register (NULL , name.name.name); +release_dentry_name_snapshot (&name);I'm not sure I like it. Sure, it won't get freed under you that way; it still can go absolutely stale by the time you return from
wakeup_source_register(). What is it being used for?
The one user I'm aware of is Android; they use EPOLLWAKEUP in the following places:
I see timerfds, /dev/input/event*, some other stuff with input devices and video devices, binder, netlink socket, and some other stuff like that - nothing that's actually likely to be renamed.
Searching on cs.android.com for places that parse this stuff, there seems to be some code that uploads it as part of bug reports or something (?), and some parser whose precise purpose I can't figure out right now:
<https://cs.android.com/android/platform/superproject/+/master:frameworks/base/core/java/com/android/internal/os/KernelWakelockReader.java;l=210?q=%2Fwakeup_sources%20-file:sepolicy%20&ss=android>
(Arve might know what this is actually good for.)
Anyway, I don't think that name is actually particularly critical to the correct functioning of the system; and the bug is a memory corruption issue, so we should fix it somehow. And adding infrastructure to power management so that it can invoke a callback to figure out the (potentially, under rare circumstances, changing) name of a wakeup source seems like a bit of overkill to me.