gdbserver: variable sized breakpoints

gdbserver: variable sized breakpoints

[Mike Frysinger]

[-- Attachment #1: Type: text/plain, Size: 709 bytes --]

current Blackfin cpus have an oddity where the insn being turned into a 
software breakpoint has to be of an equal size.  our break point insn is 
16bits long, so replacing a 16bit insn works fine.  but replacing a 32bit insn 
requires us to use the 16bit bp insn followed by a 16bit nop.  failure to do 
this can result in the unreplaced 16bits triggering undesirable side effects 
as it can be an invalid insn, one that causes speculative fetches, etc...

gdb itself takes care of this when it's setting breakpoints in 
bfin_breakpoint_from_pc.  but gdbserver doesnt seem to have a way to handle 
this.  is it already possible ?  or does the linux_target_ops struct need to 
be extended ?
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: gdbserver: variable sized breakpoints

[Yao Qi]

On 07/07/2011 10:13 AM, Mike Frysinger wrote:

> gdb itself takes care of this when it's setting breakpoints in 
> bfin_breakpoint_from_pc.  but gdbserver doesnt seem to have a way to handle 
> this.  is it already possible ?  or does the linux_target_ops struct need to 
> be extended ?

gdbserver doesn't have to worry about this.  gdb is able to send
breakpoint insn by 'M' packet, IIRC, if remote stub doesn't support 'Z0'
packet.  Looks target->insert_point is not defined in bfin gdbserver, so
'Z0' packet is not supported.  Then, gdb can send the right breakpoint
insn (32-bit or 16-bit) by 'M' packet to remote.

-- 
Yao (齐尧)

Re: gdbserver: variable sized breakpoints

[Mike Frysinger]

On Wed, Jul 6, 2011 at 23:13, Yao Qi wrote:
> On 07/07/2011 10:13 AM, Mike Frysinger wrote:
>> gdb itself takes care of this when it's setting breakpoints in
>> bfin_breakpoint_from_pc.  but gdbserver doesnt seem to have a way to handle
>> this.  is it already possible ?  or does the linux_target_ops struct need to
>> be extended ?
>
> gdbserver doesn't have to worry about this.  gdb is able to send
> breakpoint insn by 'M' packet, IIRC, if remote stub doesn't support 'Z0'
> packet.  Looks target->insert_point is not defined in bfin gdbserver, so
> 'Z0' packet is not supported.  Then, gdb can send the right breakpoint
> insn (32-bit or 16-bit) by 'M' packet to remote.

thanks, this sounds like it should work fine for us then as long as
the linux-low code doesnt change :).

in looking at the particular bug we're seeing atm, it seems it's due
to putting the 16bit insn in the 2nd half of the 32bit insn (after not
being able to detect the function prolog) and thus creating a
completely different 32bit insn.  so something is off, but it's not
this issue.
-mike