On Wed, 2013-10-16 at 01:38 +0300, Taras Kondratiuk wrote:
Hi
I was debugging kprobes-test for BE8 and noticed that some data fields are stored in LE instead of BE. It happens because these data fields get interpreted as instructions.
Is it a known issue?
For example: test_align_fail_data: bx lr .byte 0xaa .align .word 0x12345678
I would expect to see something like this: 00000000 <test_align_fail_data>: 0: e12fff1e bx lr 4: aa .byte 0xaa 5: 00 .byte 0x00 6: 0000 .short 0x0000 8: 12345678 .word 0x12345678
But instead I have: 00000000 <test_align_fail_data>: 0: e12fff1e bx lr 4: aa .byte 0xaa 5: 00 .byte 0x00 6: 0000 .short 0x0000 8: 12345678 eorsne r5, r4, #120, 12 ; 0x7800000
As a result the word 0x12345678 will be stored in LE.
I've run several tests and here are my observations:
- Double ".align" fixes the issue :)
- Behavior is the same for LE/BE, ARM/Thumb, GCC 4.4.1/4.6.x/4.8.2
- Size of alignment doesn't matter.
- Issue happens only if previous data is not instruction-aligned and 0's are added before NOPs.
- Explicit filling with 0's (.align , 0) fixes the issue, but as a side effect data @0x4 is interpreted as a single ".word 0xaa000000" instead of ".byte .byte .short". I'm not sure if there can be any functional difference because of this.
After thinking about things overnight, I believe that this is the fix we should go with. We want to stick alignment padding between data laid down with .byte and .word so it makes sense to explicitly ask the toolchain to pad with zeros rather than leaving it the opportunity to get confused. (.align in the text section probably means it wants to align with nops, but then sees the initial alignment and/or surrounding statements look like binary data, not code, and then...)
I'll send a patch proposing that fix after I've worked out how to test it on a big-endian kernel. Or if someone else sends a patch for that with a good commit message that explains what's going on I'll happily ack that.