Avoid software breakpoint's instruction shadow inconsistency
authorMaciej W. Rozycki <macro@codesourcery.com>
Fri, 3 Oct 2014 11:44:58 +0000 (12:44 +0100)
committerMaciej W. Rozycki <macro@codesourcery.com>
Fri, 3 Oct 2014 11:54:34 +0000 (12:54 +0100)
commit0d5ed1535229b7069757e04766f70a02fa60821c
tree8532eb4eaa97b8570e60bfd2170fd32ec7d5b757
parent3e87153251d9a117182decbe57dd7d9d2a47c2b3
Avoid software breakpoint's instruction shadow inconsistency

This change:

commit b775012e845380ed4c7421a1b87caf7bfae39f5f
Author: Luis Machado <luisgpm@br.ibm.com>
Date:   Fri Feb 24 15:10:59 2012 +0000

    2012-02-24  Luis Machado  <lgustavo@codesourcery.com>

* remote.c (remote_supports_cond_breakpoints): New forward
declaration.
[...]

changed the way breakpoints are inserted and removed such that
`insert_bp_location' can now be called with the breakpoint being handled
already in place, while previously the call was only ever made for
breakpoints that have not been put in place.  This in turn caused an
issue for software breakpoints and targets for which a breakpoint's
`placed_address' may not be the same as the original requested address.

The issue is `insert_bp_location' overwrites the previously adjusted
value in `placed_address' with the original address, that is only
replaced back with the correct adjusted address later on when
`gdbarch_breakpoint_from_pc' is called.  Meanwhile there's a window
where the value in `placed_address' does not correspond to data stored
in `shadow_contents', leading to incorrect instruction bytes being
supplied when `one_breakpoint_xfer_memory' is called to supply the
instruction overlaid by the breakpoint.

And this is exactly what happens on the MIPS target with software
breakpoints placed in microMIPS code.  In this case not only
`placed_address' is not the original address because of the ISA bit, but
`mips_breakpoint_from_pc' has to read the original instruction to
determine which one of the two software breakpoint instruction encodings
to choose as well.  The 16-bit encoding is used to replace 16-bit
instructions and similarly the 32-bit one is used with 32-bit
instructions, to satisfy branch delay slot size requirements.

The mismatch between `placed_address' and the address data in
`shadow_contents' has been obtained from leads to the wrong encoding
being used in some cases, which in the case of a 32-bit software
breakpoint instruction replacing a 16-bit instruction causes corruption
to the adjacent following instruction and leads the debug session astray
if execution reaches there e.g. with a jump.

To address this problem I made the change below, that adds a
`reqstd_address' field to `struct bp_target_info' and leaves
`placed_address' unchanged once it has been set.  This ensures data in
`shadow_contents' is always consistent with `placed_address'.

This approach also has this good side effect that all the places that
examine the breakpoint's address see a consistent value, either
`reqstd_address' or `placed_address', as required.  Currently some
places see either the original or the adjusted address in
`placed_address', depending on whether they have been called before
`gdbarch_remote_breakpoint_from_pc' or afterwards.  This is in
particular true for subsequent calls to
`gdbarch_remote_breakpoint_from_pc' itself, e.g. from
`one_breakpoint_xfer_memory'.  This is also important for places like
`find_single_step_breakpoint' where a breakpoint's address is compared
to the raw value of $pc.

* breakpoint.h (bp_target_info): Add `reqstd_address' member,
update comments.
* breakpoint.c (one_breakpoint_xfer_memory): Use `reqstd_address'
for the breakpoint's address.  Don't preinitialize `placed_size'.
(insert_bp_location): Set `reqstd_address' rather than
`placed_address'.
(bp_target_info_copy_insertion_state): Also copy `placed_address'.
(bkpt_insert_location): Use `reqstd_address' for the breakpoint's
address.
(bkpt_remove_location): Likewise.
(deprecated_insert_raw_breakpoint): Likewise.
(deprecated_remove_raw_breakpoint): Likewise.
(find_single_step_breakpoint): Likewise.
* mem-break.c (default_memory_insert_breakpoint): Use
`reqstd_address' for the breakpoint's address.  Don't set
`placed_address' or `placed_size' if breakpoint contents couldn't
have been determined.
* remote.c (remote_insert_breakpoint): Use `reqstd_address' for
the breakpoint's address.
(remote_insert_hw_breakpoint): Likewise.  Don't set
`placed_address' or `placed_size' if breakpoint couldn't have been
set.
* aarch64-linux-nat.c (aarch64_linux_insert_hw_breakpoint): Use
`reqstd_address' for the breakpoint's address.
* arm-linux-nat.c (arm_linux_hw_breakpoint_initialize): Likewise.
* ia64-tdep.c (ia64_memory_insert_breakpoint): Likewise.
* m32r-tdep.c (m32r_memory_insert_breakpoint): Likewise.
* microblaze-linux-tdep.c
(microblaze_linux_memory_remove_breakpoint): Likewise.
* monitor.c (monitor_insert_breakpoint): Likewise.
* nto-procfs.c (procfs_insert_breakpoint): Likewise.
(procfs_insert_hw_breakpoint): Likewise.
* ppc-linux-nat.c (ppc_linux_insert_hw_breakpoint): Likewise.
* ppc-linux-tdep.c (ppc_linux_memory_remove_breakpoint): Likewise.
* remote-m32r-sdi.c (m32r_insert_breakpoint): Likewise.
* remote-mips.c (mips_insert_breakpoint): Likewise.
* x86-nat.c (x86_insert_hw_breakpoint): Likewise.
17 files changed:
gdb/ChangeLog
gdb/aarch64-linux-nat.c
gdb/arm-linux-nat.c
gdb/breakpoint.c
gdb/breakpoint.h
gdb/ia64-tdep.c
gdb/m32r-tdep.c
gdb/mem-break.c
gdb/microblaze-linux-tdep.c
gdb/monitor.c
gdb/nto-procfs.c
gdb/ppc-linux-nat.c
gdb/ppc-linux-tdep.c
gdb/remote-m32r-sdi.c
gdb/remote-mips.c
gdb/remote.c
gdb/x86-nat.c