lkml.org 
[lkml]   [2019]   [Sep]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[PATCH 5.2 142/143] bpf: fix use after free in prog symbol exposure
Date
From: Daniel Borkmann <daniel@iogearbox.net>

commit c751798aa224fadc5124b49eeb38fb468c0fa039 upstream.

syzkaller managed to trigger the warning in bpf_jit_free() which checks via
bpf_prog_kallsyms_verify_off() for potentially unlinked JITed BPF progs
in kallsyms, and subsequently trips over GPF when walking kallsyms entries:

[...]
8021q: adding VLAN 0 to HW filter on device batadv0
8021q: adding VLAN 0 to HW filter on device batadv0
WARNING: CPU: 0 PID: 9869 at kernel/bpf/core.c:810 bpf_jit_free+0x1e8/0x2a0
Kernel panic - not syncing: panic_on_warn set ...
CPU: 0 PID: 9869 Comm: kworker/0:7 Not tainted 5.0.0-rc8+ #1
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Workqueue: events bpf_prog_free_deferred
Call Trace:
__dump_stack lib/dump_stack.c:77 [inline]
dump_stack+0x113/0x167 lib/dump_stack.c:113
panic+0x212/0x40b kernel/panic.c:214
__warn.cold.8+0x1b/0x38 kernel/panic.c:571
report_bug+0x1a4/0x200 lib/bug.c:186
fixup_bug arch/x86/kernel/traps.c:178 [inline]
do_error_trap+0x11b/0x200 arch/x86/kernel/traps.c:271
do_invalid_op+0x36/0x40 arch/x86/kernel/traps.c:290
invalid_op+0x14/0x20 arch/x86/entry/entry_64.S:973
RIP: 0010:bpf_jit_free+0x1e8/0x2a0
Code: 02 4c 89 e2 83 e2 07 38 d0 7f 08 84 c0 0f 85 86 00 00 00 48 ba 00 02 00 00 00 00 ad de 0f b6 43 02 49 39 d6 0f 84 5f fe ff ff <0f> 0b e9 58 fe ff ff 48 b8 00 00 00 00 00 fc ff df 4c 89 e2 48 c1
RSP: 0018:ffff888092f67cd8 EFLAGS: 00010202
RAX: 0000000000000007 RBX: ffffc90001947000 RCX: ffffffff816e9d88
RDX: dead000000000200 RSI: 0000000000000008 RDI: ffff88808769f7f0
RBP: ffff888092f67d00 R08: fffffbfff1394059 R09: fffffbfff1394058
R10: fffffbfff1394058 R11: ffffffff89ca02c7 R12: ffffc90001947002
R13: ffffc90001947020 R14: ffffffff881eca80 R15: ffff88808769f7e8
BUG: unable to handle kernel paging request at fffffbfff400d000
#PF error: [normal kernel read fault]
PGD 21ffee067 P4D 21ffee067 PUD 21ffed067 PMD 9f942067 PTE 0
Oops: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 9869 Comm: kworker/0:7 Not tainted 5.0.0-rc8+ #1
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Workqueue: events bpf_prog_free_deferred
RIP: 0010:bpf_get_prog_addr_region kernel/bpf/core.c:495 [inline]
RIP: 0010:bpf_tree_comp kernel/bpf/core.c:558 [inline]
RIP: 0010:__lt_find include/linux/rbtree_latch.h:115 [inline]
RIP: 0010:latch_tree_find include/linux/rbtree_latch.h:208 [inline]
RIP: 0010:bpf_prog_kallsyms_find+0x107/0x2e0 kernel/bpf/core.c:632
Code: 00 f0 ff ff 44 38 c8 7f 08 84 c0 0f 85 fa 00 00 00 41 f6 45 02 01 75 02 0f 0b 48 39 da 0f 82 92 00 00 00 48 89 d8 48 c1 e8 03 <42> 0f b6 04 30 84 c0 74 08 3c 03 0f 8e 45 01 00 00 8b 03 48 c1 e0
[...]

Upon further debugging, it turns out that whenever we trigger this
issue, the kallsyms removal in bpf_prog_ksym_node_del() was /skipped/
but yet bpf_jit_free() reported that the entry is /in use/.

Problem is that symbol exposure via bpf_prog_kallsyms_add() but also
perf_event_bpf_event() were done /after/ bpf_prog_new_fd(). Once the
fd is exposed to the public, a parallel close request came in right
before we attempted to do the bpf_prog_kallsyms_add().

Given at this time the prog reference count is one, we start to rip
everything underneath us via bpf_prog_release() -> bpf_prog_put().
The memory is eventually released via deferred free, so we're seeing
that bpf_jit_free() has a kallsym entry because we added it from
bpf_prog_load() but /after/ bpf_prog_put() from the remote CPU.

Therefore, move both notifications /before/ we install the fd. The
issue was never seen between bpf_prog_alloc_id() and bpf_prog_new_fd()
because upon bpf_prog_get_fd_by_id() we'll take another reference to
the BPF prog, so we're still holding the original reference from the
bpf_prog_load().

Fixes: 6ee52e2a3fe4 ("perf, bpf: Introduce PERF_RECORD_BPF_EVENT")
Fixes: 74451e66d516 ("bpf: make jited programs visible in traces")
Reported-by: syzbot+bd3bba6ff3fcea7a6ec6@syzkaller.appspotmail.com
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Song Liu <songliubraving@fb.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

---
kernel/bpf/syscall.c | 30 ++++++++++++++++++------------
1 file changed, 18 insertions(+), 12 deletions(-)

--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c
@@ -1686,20 +1686,26 @@ static int bpf_prog_load(union bpf_attr
if (err)
goto free_used_maps;

- err = bpf_prog_new_fd(prog);
- if (err < 0) {
- /* failed to allocate fd.
- * bpf_prog_put() is needed because the above
- * bpf_prog_alloc_id() has published the prog
- * to the userspace and the userspace may
- * have refcnt-ed it through BPF_PROG_GET_FD_BY_ID.
- */
- bpf_prog_put(prog);
- return err;
- }
-
+ /* Upon success of bpf_prog_alloc_id(), the BPF prog is
+ * effectively publicly exposed. However, retrieving via
+ * bpf_prog_get_fd_by_id() will take another reference,
+ * therefore it cannot be gone underneath us.
+ *
+ * Only for the time /after/ successful bpf_prog_new_fd()
+ * and before returning to userspace, we might just hold
+ * one reference and any parallel close on that fd could
+ * rip everything out. Hence, below notifications must
+ * happen before bpf_prog_new_fd().
+ *
+ * Also, any failure handling from this point onwards must
+ * be using bpf_prog_put() given the program is exposed.
+ */
bpf_prog_kallsyms_add(prog);
perf_event_bpf_event(prog, PERF_BPF_EVENT_PROG_LOAD, 0);
+
+ err = bpf_prog_new_fd(prog);
+ if (err < 0)
+ bpf_prog_put(prog);
return err;

free_used_maps:

\
 
 \ /
  Last update: 2019-09-04 20:16    [W:6.984 / U:0.012 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site