ring-buffer: Fix deadloop issue on reading trace_pipe
authorZheng Yejian <zhengyejian1@huawei.com>
Sat, 8 Jul 2023 22:51:44 +0000 (06:51 +0800)
committerSteven Rostedt (Google) <rostedt@goodmis.org>
Wed, 12 Jul 2023 16:15:26 +0000 (12:15 -0400)
commit7e42907f3a7b4ce3a2d1757f6d78336984daf8f5
tree55a2d512033b53b4b26a4b0c86b073cf05fc966a
parent7d8b31b73c79835572611ed1eed649e4d2e14245
ring-buffer: Fix deadloop issue on reading trace_pipe

Soft lockup occurs when reading file 'trace_pipe':

  watchdog: BUG: soft lockup - CPU#6 stuck for 22s! [cat:4488]
  [...]
  RIP: 0010:ring_buffer_empty_cpu+0xed/0x170
  RSP: 0018:ffff88810dd6fc48 EFLAGS: 00000246
  RAX: 0000000000000000 RBX: 0000000000000246 RCX: ffffffff93d1aaeb
  RDX: ffff88810a280040 RSI: 0000000000000008 RDI: ffff88811164b218
  RBP: ffff88811164b218 R08: 0000000000000000 R09: ffff88815156600f
  R10: ffffed102a2acc01 R11: 0000000000000001 R12: 0000000051651901
  R13: 0000000000000000 R14: ffff888115e49500 R15: 0000000000000000
  [...]
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f8d853c2000 CR3: 000000010dcd8000 CR4: 00000000000006e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   __find_next_entry+0x1a8/0x4b0
   ? peek_next_entry+0x250/0x250
   ? down_write+0xa5/0x120
   ? down_write_killable+0x130/0x130
   trace_find_next_entry_inc+0x3b/0x1d0
   tracing_read_pipe+0x423/0xae0
   ? tracing_splice_read_pipe+0xcb0/0xcb0
   vfs_read+0x16b/0x490
   ksys_read+0x105/0x210
   ? __ia32_sys_pwrite64+0x200/0x200
   ? switch_fpu_return+0x108/0x220
   do_syscall_64+0x33/0x40
   entry_SYSCALL_64_after_hwframe+0x61/0xc6

Through the vmcore, I found it's because in tracing_read_pipe(),
ring_buffer_empty_cpu() found some buffer is not empty but then it
cannot read anything due to "rb_num_of_entries() == 0" always true,
Then it infinitely loop the procedure due to user buffer not been
filled, see following code path:

  tracing_read_pipe() {
    ... ...
    waitagain:
      tracing_wait_pipe() // 1. find non-empty buffer here
      trace_find_next_entry_inc()  // 2. loop here try to find an entry
        __find_next_entry()
          ring_buffer_empty_cpu();  // 3. find non-empty buffer
          peek_next_entry()  // 4. but peek always return NULL
            ring_buffer_peek()
              rb_buffer_peek()
                rb_get_reader_page()
                  // 5. because rb_num_of_entries() == 0 always true here
                  //    then return NULL
      // 6. user buffer not been filled so goto 'waitgain'
      //    and eventually leads to an deadloop in kernel!!!
  }

By some analyzing, I found that when resetting ringbuffer, the 'entries'
of its pages are not all cleared (see rb_reset_cpu()). Then when reducing
the ringbuffer, and if some reduced pages exist dirty 'entries' data, they
will be added into 'cpu_buffer->overrun' (see rb_remove_pages()), which
cause wrong 'overrun' count and eventually cause the deadloop issue.

To fix it, we need to clear every pages in rb_reset_cpu().

Link: https://lore.kernel.org/linux-trace-kernel/20230708225144.3785600-1-zhengyejian1@huawei.com
Cc: stable@vger.kernel.org
Fixes: a5fb833172eca ("ring-buffer: Fix uninitialized read_stamp")
Signed-off-by: Zheng Yejian <zhengyejian1@huawei.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
kernel/trace/ring_buffer.c