[PATCH RFC 2/6] arm64: Kprobes with single stepping support

[Sandeepa Prabhu]

On 18 November 2013 12:25, Sandeepa Prabhu <sandeepa.prabhu at linaro.org> wrote:
>>> This is generic test module (samples/hw_breakpoint/data_breakpoint.ko)
>>> which places watchpoint for bothe read/write.
>>> Atleast watchpt should have triggered for Read right? I also tried
>>> with othe functions like do_fork, vfs_read etc but no hit.
>>
>> You'd need to place something for exec if you want to see anything on the
>> instruction side. A read by the instruction fetcher does not trigger a read
>> watchpoint on ARM.
> Hmm, then watchpoint cannot not be placed on text address (instruction
> fetch path) right?
> (Sorry I did not check debug spec fully for watchpoint/hw breakpoints,
> I should do that)
>
>>
>>> >> 2.  Placing text breakpoint (modified sample module with attr.bp_type
>>> >> = HW_BREAKPOINT_X) upon vfs_symlink, and run "ln -s /dev/null
>>> >> /tmp/foo".  This time, breakpoint hit but exception is re-cursing
>>> >> infinitely!
>>> >
>>> > The problem here is that we expect the overflow handler to deal with the
>>> > stepping (like GDB does via ptrace). If you don't register a handler, the
>>> > kernel will do the step (like you would get if you used perf stat -e
>>> > mem:0xNNNN:x).
>>> [This test was done on upstream branch, without kprobes patches.]
>>> Hmm, then this is expected with test breakpoint right? is this
>>> handling to be done by perf and ptrace?
>>
>> perf stat doesn't register an overflow handler, so the hw_breakpoint
>> backend will handle the step. ptrace registers a handler which sends a
>> SIGTRAP to the debugger (e.g. gdb), which handles the step manually
>> (probably using a PTRACE_SINGLESTEP request).
>>
>>> I did not see arm64 support in linux/tools/perf/,  there are multiple
>>> patches in mailing list though. Are you aware of any version of perf
>>> that work with arm64?
>>
>> The perf tool should work fine on arm64 using mainline. Are you seeing
>> problems?
> Hmm basically perf is working and I can run tests now.
Hi Will,
Okay. I can see what you meant with respect to D-flag unmasking, on
original patchset(v2). :)

a. Placed a kprobe on vfs_read - works fine independently.
b. Placed hw_breakpoint on vfs_read using perf (perf record -e
mem:0xffffffc000134ba4:x -a -- sleep 10) and it works fine
independently.

c. Now, a+b, first placed kprobe on vfs_read and then ran perf record
event on vfs_read (hw breakpoint). Now, seeing that kprobe single step
is never complete/never disabled!, so debug exception would generate
forever! (Continuously printing "Unexpected kernel single-step
exception at EL1")

 kprobe pre_handler: p->addr = 0xffffffc000134ba4
 reenter_dbg: test API invoked
 kprobe post_handler: p->addr = 0xffffffc000134ba4
 fmp/fo^[[5Dreenter_dbg: test API invoked
 kprobe pre_handler: p->addr = 0xffffffc000134ba4
 Unexpected kernel single-step exception at EL1
 Unexpected kernel single-step exception at EL1
....

Once I change the location of D-flag manipulation (your suggestion)
and run enough tests, I would come back with more details and inputs.

Thanks,
Sandeepa

>>
>> Will
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo at vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at  http://www.tux.org/lkml/