ftrace: Fix NULL pointer dereference in is_ftrace_trampoline when ftrace is dead
authorYang Jihong <yangjihong1@huawei.com>
Thu, 18 Aug 2022 03:26:59 +0000 (11:26 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 5 Sep 2022 08:30:07 +0000 (10:30 +0200)
commit c3b0f72e805f0801f05fa2aa52011c4bfc694c44 upstream.

ftrace_startup does not remove ops from ftrace_ops_list when
ftrace_startup_enable fails:

register_ftrace_function
  ftrace_startup
    __register_ftrace_function
      ...
      add_ftrace_ops(&ftrace_ops_list, ops)
      ...
    ...
    ftrace_startup_enable // if ftrace failed to modify, ftrace_disabled is set to 1
    ...
  return 0 // ops is in the ftrace_ops_list.

When ftrace_disabled = 1, unregister_ftrace_function simply returns without doing anything:
unregister_ftrace_function
  ftrace_shutdown
    if (unlikely(ftrace_disabled))
            return -ENODEV;  // return here, __unregister_ftrace_function is not executed,
                             // as a result, ops is still in the ftrace_ops_list
    __unregister_ftrace_function
    ...

If ops is dynamically allocated, it will be free later, in this case,
is_ftrace_trampoline accesses NULL pointer:

is_ftrace_trampoline
  ftrace_ops_trampoline
    do_for_each_ftrace_op(op, ftrace_ops_list) // OOPS! op may be NULL!

Syzkaller reports as follows:
[ 1203.506103] BUG: kernel NULL pointer dereference, address: 000000000000010b
[ 1203.508039] #PF: supervisor read access in kernel mode
[ 1203.508798] #PF: error_code(0x0000) - not-present page
[ 1203.509558] PGD 800000011660b067 P4D 800000011660b067 PUD 130fb8067 PMD 0
[ 1203.510560] Oops: 0000 [#1] SMP KASAN PTI
[ 1203.511189] CPU: 6 PID: 29532 Comm: syz-executor.2 Tainted: G    B   W         5.10.0 #8
[ 1203.512324] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
[ 1203.513895] RIP: 0010:is_ftrace_trampoline+0x26/0xb0
[ 1203.514644] Code: ff eb d3 90 41 55 41 54 49 89 fc 55 53 e8 f2 00 fd ff 48 8b 1d 3b 35 5d 03 e8 e6 00 fd ff 48 8d bb 90 00 00 00 e8 2a 81 26 00 <48> 8b ab 90 00 00 00 48 85 ed 74 1d e8 c9 00 fd ff 48 8d bb 98 00
[ 1203.518838] RSP: 0018:ffffc900012cf960 EFLAGS: 00010246
[ 1203.520092] RAX: 0000000000000000 RBX: 000000000000007b RCX: ffffffff8a331866
[ 1203.521469] RDX: 0000000000000000 RSI: 0000000000000008 RDI: 000000000000010b
[ 1203.522583] RBP: 0000000000000000 R08: 0000000000000000 R09: ffffffff8df18b07
[ 1203.523550] R10: fffffbfff1be3160 R11: 0000000000000001 R12: 0000000000478399
[ 1203.524596] R13: 0000000000000000 R14: ffff888145088000 R15: 0000000000000008
[ 1203.525634] FS:  00007f429f5f4700(0000) GS:ffff8881daf00000(0000) knlGS:0000000000000000
[ 1203.526801] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1203.527626] CR2: 000000000000010b CR3: 0000000170e1e001 CR4: 00000000003706e0
[ 1203.528611] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 1203.529605] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400

Therefore, when ftrace_startup_enable fails, we need to rollback registration
process and remove ops from ftrace_ops_list.

Link: https://lkml.kernel.org/r/20220818032659.56209-1-yangjihong1@huawei.com
Suggested-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Yang Jihong <yangjihong1@huawei.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/trace/ftrace.c

index e215a9c9697112bd984690fd58f291bd04ef7c87..e10cf1b5481285ab65ae5d044f3c2e74587369ec 100644 (file)
@@ -2901,6 +2901,16 @@ int ftrace_startup(struct ftrace_ops *ops, int command)
 
        ftrace_startup_enable(command);
 
+       /*
+        * If ftrace is in an undefined state, we just remove ops from list
+        * to prevent the NULL pointer, instead of totally rolling it back and
+        * free trampoline, because those actions could cause further damage.
+        */
+       if (unlikely(ftrace_disabled)) {
+               __unregister_ftrace_function(ops);
+               return -ENODEV;
+       }
+
        ops->flags &= ~FTRACE_OPS_FL_ADDING;
 
        return 0;