tracing: Do not reference char * as a string in histograms
authorSteven Rostedt (VMware) <rostedt@goodmis.org>
Thu, 15 Jul 2021 04:02:06 +0000 (00:02 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 20 Jul 2021 14:17:48 +0000 (16:17 +0200)
commit7e35f39f21bdc9347638997d6406c8ab115e4afb
tree883068bee44b79a0b42d5c91996522b715d3d154
parentf3d0a109240c9bed5c60d819014786be3a2fe515
tracing: Do not reference char * as a string in histograms

commit 704adfb5a9978462cd861f170201ae2b5e3d3a80 upstream.

The histogram logic was allowing events with char * pointers to be used as
normal strings. But it was easy to crash the kernel with:

 # echo 'hist:keys=filename' > events/syscalls/sys_enter_openat/trigger

And open some files, and boom!

 BUG: unable to handle page fault for address: 00007f2ced0c3280
 #PF: supervisor read access in kernel mode
 #PF: error_code(0x0000) - not-present page
 PGD 1173fa067 P4D 1173fa067 PUD 1171b6067 PMD 1171dd067 PTE 0
 Oops: 0000 [#1] PREEMPT SMP
 CPU: 6 PID: 1810 Comm: cat Not tainted 5.13.0-rc5-test+ #61
 Hardware name: Hewlett-Packard HP Compaq Pro 6300 SFF/339A, BIOS K01
v03.03 07/14/2016
 RIP: 0010:strlen+0x0/0x20
 Code: f6 82 80 2a 0b a9 20 74 11 0f b6 50 01 48 83 c0 01 f6 82 80 2a 0b
a9 20 75 ef c3 66 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 <80> 3f 00 74
10 48 89 f8 48 83 c0 01 80 38 00 75 f7 48 29 f8 c3

 RSP: 0018:ffffbdbf81567b50 EFLAGS: 00010246
 RAX: 0000000000000003 RBX: ffff93815cdb3800 RCX: ffff9382401a22d0
 RDX: 0000000000000100 RSI: 0000000000000000 RDI: 00007f2ced0c3280
 RBP: 0000000000000100 R08: ffff9382409ff074 R09: ffffbdbf81567c98
 R10: ffff9382409ff074 R11: 0000000000000000 R12: ffff9382409ff074
 R13: 0000000000000001 R14: ffff93815a744f00 R15: 00007f2ced0c3280
 FS:  00007f2ced0f8580(0000) GS:ffff93825a800000(0000)
knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007f2ced0c3280 CR3: 0000000107069005 CR4: 00000000001706e0
 Call Trace:
  event_hist_trigger+0x463/0x5f0
  ? find_held_lock+0x32/0x90
  ? sched_clock_cpu+0xe/0xd0
  ? lock_release+0x155/0x440
  ? kernel_init_free_pages+0x6d/0x90
  ? preempt_count_sub+0x9b/0xd0
  ? kernel_init_free_pages+0x6d/0x90
  ? get_page_from_freelist+0x12c4/0x1680
  ? __rb_reserve_next+0xe5/0x460
  ? ring_buffer_lock_reserve+0x12a/0x3f0
  event_triggers_call+0x52/0xe0
  ftrace_syscall_enter+0x264/0x2c0
  syscall_trace_enter.constprop.0+0x1ee/0x210
  do_syscall_64+0x1c/0x80
  entry_SYSCALL_64_after_hwframe+0x44/0xae

Where it triggered a fault on strlen(key) where key was the filename.

The reason is that filename is a char * to user space, and the histogram
code just blindly dereferenced it, with obvious bad results.

I originally tried to use strncpy_from_user/kernel_nofault() but found
that there's other places that its dereferenced and not worth the effort.

Just do not allow "char *" to act like strings.

Link: https://lkml.kernel.org/r/20210715000206.025df9d2@rorschach.local.home
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Tzvetomir Stoyanov <tz.stoyanov@gmail.com>
Cc: stable@vger.kernel.org
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Tom Zanussi <zanussi@kernel.org>
Fixes: 79e577cbce4c4 ("tracing: Support string type key properly")
Fixes: 5967bd5c4239 ("tracing: Let filter_assign_type() detect FILTER_PTR_STRING")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/trace/trace_events_hist.c