[RFC][PATCH] Demultiplexing SIGTRAP signal -v2

[Roland McGrath]

I certainly have no objection in principle.  I doubt that any x86 userland
apps expect certain si_code values for SIGTRAP now, since the existing
values are not of any real use.  (Signal handlers get the thread.trap_no and
thread.error_code values from hardware to guess from, and debuggers via
ptrace get the hardware %db6 value to guess from.)  I do have a few comments.

If you're doing it, I think you should do the do_int3 case too,
so every machine-generated SIGTRAP has a meaningful si_code value.

The only use of send_sigtrap is for do_debug (and for faking that do_debug
happened in the syscall_trace_leave case).  You should consolidate all the
uses in both 32 and 64 to use send_sigtrap uniformly, change its signature
as needed.  I'm inclined to consolidate the si_code logic there, and just
pass it the hardware bits or let it get them from the thread_struct
(trap_nr, error_code, debugreg6).

About that si_code logic based on %db6.  There are some funny "sticky"
properties to how that register gets set in hardware.  Even reading the
hardware manuals doesn't always make it plain what to expect.  I wouldn't
want to testify that the patch's logic is correct in distinguishing which
event really just happened.  (I'm not sure, but I think it may also be
possible to have a single do_debug trap for both a single-step trap and a
hardware breakpoint trap generated by the same instruction.)  I know that
Alan Stern figured out a lot of the magic empirically a while back.  That
deserves a careful double-checking if we are now trying to make si_code
tell a clear and reliable story.


Thanks,
Roland