Hi Steve,
Thanks very much for your friendly and clear feedback.
Although in current kernel trace_seq_putmem_hex() is only used for
single word,
I think it should/need support longer data. These are my arguments:
1. The design of double loop is used to process more data. If only
supports single word,
the inner loop is enough, and the outer loop and the following
lines are no longer needed.
len -= j / 2;
hex[j++] = ' ';
2. The last line above try to split two words/dwords with space. If only
supports single word,
this strange behavior is hard to understand.
3. If it only supports single word, I think parameter 'len' is redundant.
4. The comments of both seq_buf_putmem_hex() and trace_seq_putmem_hex()
have not
indicated the scope of 'len'.
5. If it only supports single word, we need to design a new function to
support bigger block of data.
I think it is redundant since the current function can perfectly
deal with.
6. If follow my patch, it can support any length of data, including the
single word.
How do you think?
Regards,
Yun
On 6/24/21 10:54 PM, Steven Rostedt wrote:
[Please note: This e-mail is from an EXTERNAL e-mail address]
On Thu, 24 Jun 2021 21:16:46 +0800
Yun Zhou <yun.zhou@xxxxxxxxxxxxx> wrote:
I guess the original intention of seq_buf_putmem_hex is to dump the raw
A little background, this was originally introduced in 2008:
5e3ca0ec76fce ("ftrace: introduce the "hex" output method")
And commit ad0a3b68114e8 ("trace: add build-time check to avoid overrunning hex buffer")
changed HEX_CHARS from a hardcoded 17 to a way to decided what the max
long is.
#define HEX_CHARS (MAX_MEMHEX_BYTES*2 + 1)
memory to seq_buf according to 64 bits integer number. It tries to output
numbers in reverse, with the high bits in the front and the low bits in
the back. If the length of the raw memory is equal to 8, e.g.
"01 23 45 67 89 ab cd ef" in memory, it will be dumped as "efcdab8967452301".
Note, it only does that for little endian machines.
But if the length of the raw memory is larger than 8, the first value of
start_len will larger than 8, than seq_buf will save the last data, not
the eighth one, e.g. "01 23 45 67 89 ab cd ef 11" in memory, it will be
dumped as "11efcdab8967452301". I think it is not the original
intention of the function.
More seriously, if the length of the raw memory is larger than 9, the
start_len will be larger than 9, then hex will overflow, and the stack will be
corrupted. I do not kown if it can be exploited by hacker. But I am sure
it will cause kernel panic when the length of memory is more than 32 bytes.
[ 448.551471] Unable to handle kernel paging request at virtual address 3438376c
[ 448.558678] pgd = 6eaf278e
[ 448.561376] [3438376c] *pgd=00000000
[ 448.564945] Internal error: Oops: 5 [#2] PREEMPT ARM
[ 448.569899] Modules linked in:
[ 448.572951] CPU: 0 PID: 368 Comm: cat Tainted: G W 4.18.40-yocto-standard #18
[ 448.581374] Hardware name: Xilinx Zynq Platform
[ 448.585901] PC is at trace_seq_putmem_hex+0x6c/0x84
[ 448.590768] LR is at 0x20643032
[ 448.593901] pc : [<c009a85c>] lr : [<20643032>] psr: 60000093
[ 448.600159] sp : d980dc08 ip : 00000020 fp : c05f58cc
[ 448.605375] r10: c05e5f30 r9 : 00000031 r8 : 00000000
[ 448.610584] r7 : 20643032 r6 : 37663666 r5 : 36343730 r4 : 34383764
[ 448.617103] r3 : 00001000 r2 : 00000042 r1 : d980dc00 r0 : 00000000
...
[ 448.907962] [<c009a85c>] (trace_seq_putmem_hex) from [<c010b008>] (trace_raw_output_write+0x58/0x9c)
[ 448.917094] [<c010b008>] (trace_raw_output_write) from [<c00964c0>] (print_trace_line+0x144/0x3e8)
[ 448.926050] [<c00964c0>] (print_trace_line) from [<c0098710>] (ftrace_dump+0x204/0x254)
[ 448.934053] [<c0098710>] (ftrace_dump) from [<c0098780>] (trace_die_handler+0x20/0x34)
[ 448.941975] [<c0098780>] (trace_die_handler) from [<c003cff8>] (notifier_call_chain+0x48/0x6c)
[ 448.950581] [<c003cff8>] (notifier_call_chain) from [<c003d19c>] (__atomic_notifier_call_chain+0x3c/0x50)
[ 448.960142] [<c003d19c>] (__atomic_notifier_call_chain) from [<c003d1cc>] (atomic_notifier_call_chain+0x1c/0x24)
[ 448.970306] [<c003d1cc>] (atomic_notifier_call_chain) from [<c003d204>] (notify_die+0x30/0x3c)
[ 448.978908] [<c003d204>] (notify_die) from [<c001361c>] (die+0xc4/0x258)
[ 448.985604] [<c001361c>] (die) from [<c00173e8>] (__do_kernel_fault.part.0+0x5c/0x7c)
[ 448.993438] [<c00173e8>] (__do_kernel_fault.part.0) from [<c043a270>] (do_page_fault+0x158/0x394)
[ 449.002305] [<c043a270>] (do_page_fault) from [<c00174dc>] (do_DataAbort+0x40/0xec)
[ 449.009959] [<c00174dc>] (do_DataAbort) from [<c0009970>] (__dabt_svc+0x50/0x80)
Additionally, the address of data ptr keeps in the same value in multiple
loops, the value of data buffer will not be picked forever.
So the bug looks like it was there since the original code was
introduced in 2008! There's two variables being increased in that loop
(i and j), and i follows the raw data, and j follows what is being
written into the buffer. The bug is that we are comparing the HEX_CHARS
to 'i' when we really should be comparing it to 'j'! As if 'j' goes
bigger than HEX_CHARS, it will overflow the destination buffer.
And, it should be noted, that this is to read a single word (long) and
not more. Thus, the "date += start_len" shouldn't be needed.
Could you send another patch that makes it only process a single word
and exit?
I'll tag it for stable when you do.
Thanks!
-- Steve
Signed-off-by: Yun Zhou <yun.zhou@xxxxxxxxxxxxx>
---
lib/seq_buf.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/lib/seq_buf.c b/lib/seq_buf.c
index 6aabb609dd87..948c8b55f666 100644
--- a/lib/seq_buf.c
+++ b/lib/seq_buf.c
@@ -229,7 +229,7 @@ int seq_buf_putmem_hex(struct seq_buf *s, const void *mem,
WARN_ON(s->size == 0);
while (len) {
- start_len = min(len, HEX_CHARS - 1);
+ start_len = min(len, MAX_MEMHEX_BYTES);
#ifdef __BIG_ENDIAN
for (i = 0, j = 0; i < start_len; i++) {
#else
@@ -248,6 +248,8 @@ int seq_buf_putmem_hex(struct seq_buf *s, const void *mem,
seq_buf_putmem(s, hex, j);
if (seq_buf_has_overflowed(s))
return -1;
+
+ data += start_len;
}
return 0;
}
