A patch in the target cleanup series caused a regression when using
record with target-async. Version 4 of the patch is here:
https://sourceware.org/ml/gdb-patches/2014-03/msg00159.html
The immediate problem is that record supplies to_can_async_p and
to_is_async_p methods, but does not supply a to_async method. So,
when target-async is set, record claims to support async -- but if the
underlying target does not support async, then the to_async method
call will end up in that method's default implementation, namely
tcomplain.
This worked previously because the record target used to provide a
to_async method; one that (erroneously, only at push time) checked the
other members of the target stack, and then simply dropped to_async
calls in the "does not implement async" case.
My first thought was to simply drop tcomplain as the default for
to_async. This works, but Pedro pointed out that the only reason
record has to supply to_can_async_p and to_is_async_p is that these
default to using the find_default_run_target machinery -- and these
defaults are only needed by "run" and "attach".
So, a nicer solution presents itself: change run and attach to
explicitly call into the default run target when needed; and change
to_is_async_p and to_can_async_p to default to "return 0". This makes
the target stack simpler to use and lets us remove the method
implementations from record. This is also in harmony with other plans
for the target stack; namely trying to reduce the impact of
find_default_run_target. This approach makes it clear that
find_default_is_async_p is not needed -- it is asking whether a target
that may not even be pushed is actually async, which seems like a
nonsensical question.
While an improvement, this approach proved to introduce the same bug
when using the core target. Looking a bit deeper, the issue is that
code in "attach" and "run" may need to use either the current target
stack or the default run target -- but different calls into the target
API in those functions could wind up querying different targets.
This new patch makes the target to use more explicit in "run" and
"attach". Then these commands explicitly make the needed calls
against that target. This ensures that a single target is used for
all relevant operations. This lets us remove a couple find_default_*
functions from various targets, including the dummy target. I think
this is a decent understandability improvement.
One issue I see with this patch is that the new calls in "run" and
"attach" are not very much like the rest of the target API. I think
fundamentally this is due to bad factoring in the target API, which
may need to be fixed for multi-target. Tackling that seemed ambitious
for a regression fix.
While working on this I noticed that there don't seem to be any test
cases that involve both target-async and record, so this patch changes
break-precsave.exp to add some. It also changes corefile.exp to add
some target-async tests; these pass with current trunk and with this
patch applied, but fail with the v1 patch.
This patch differs from v4 in that it moves initialization of
to_can_async_p and to_supports_non_stop into inf-child, adds some
assertions to complete_target_initialization, and adds some comments
to target.h.
Built and regtested on x86-64 Fedora 20.
2014-03-12 Tom Tromey <tromey@redhat.com>
* inf-child.c (return_zero): New function.
(inf_child_target): Set to_can_async_p, to_supports_non_stop.
* aix-thread.c (aix_thread_inferior_created): New function.
(aix_thread_attach): Remove.
(init_aix_thread_ops): Don't set to_attach.
(_initialize_aix_thread): Register inferior_created observer.
* corelow.c (init_core_ops): Don't set to_attach or
to_create_inferior.
* exec.c (init_exec_ops): Don't set to_attach or
to_create_inferior.
* infcmd.c (run_command_1): Use find_run_target. Make direct
target calls.
(attach_command): Use find_attach_target. Make direct target
calls.
* record-btrace.c (init_record_btrace_ops): Don't set
to_create_inferior.
* record-full.c (record_full_can_async_p, record_full_is_async_p):
Remove.
(init_record_full_ops, init_record_full_core_ops): Update. Don't
set to_create_inferior.
* target.c (complete_target_initialization): Add assertion.
(target_create_inferior): Remove.
(find_default_attach, find_default_create_inferior): Remove.
(find_attach_target, find_run_target): New functions.
(find_default_is_async_p, find_default_can_async_p)
(target_supports_non_stop, target_attach): Remove.
(init_dummy_target): Don't set to_create_inferior or
to_supports_non_stop.
* target.h (struct target_ops) <to_attach>: Add comment. Remove
TARGET_DEFAULT_FUNC.
<to_create_inferior>: Add comment.
<to_can_async_p, to_is_async_p, to_supports_non_stop>: Use
TARGET_DEFAULT_RETURN.
<to_can_async_p, to_supports_non_stop, to_can_run>: Add comments.
(find_attach_target, find_run_target): Declare.
(target_create_inferior): Remove.
(target_has_execution_1): Update comment.
(target_supports_non_stop): Remove.
* target-delegates.c: Rebuild.
2014-03-12 Tom Tromey <tromey@redhat.com>
* gdb.base/corefile.exp (corefile_test_run, corefile_test_attach):
New procs. Add target-async tests.
* gdb.reverse/break-precsave.exp (precsave_tests): New proc.
Add target-async tests.
In non-stop mode, or rather, breakpoints always-inserted mode, the
code cache can easily end up with stale breakpoint instructions:
All it takes is filling a cache line when breakpoints already exist in
that memory region, and then delete the breakpoint.
Vis. (from the new test):
(gdb) set breakpoint always-inserted on
(gdb) b 23
Breakpoint 2 at 0x400540: file ../../../src/gdb/testsuite/gdb.base/breakpoint-shadow.c, line 23.
(gdb) b 24
Breakpoint 3 at 0x400547: file ../../../src/gdb/testsuite/gdb.base/breakpoint-shadow.c, line 24.
disass main
Dump of assembler code for function main:
0x000000000040053c <+0>: push %rbp
0x000000000040053d <+1>: mov %rsp,%rbp
=> 0x0000000000400540 <+4>: movl $0x1,-0x4(%rbp)
0x0000000000400547 <+11>: movl $0x2,-0x4(%rbp)
0x000000000040054e <+18>: mov $0x0,%eax
0x0000000000400553 <+23>: pop %rbp
0x0000000000400554 <+24>: retq
End of assembler dump.
So far so good. Now flush the code cache:
(gdb) set code-cache off
(gdb) set code-cache on
Requesting a disassembly works as expected, breakpoint shadowing is
applied:
(gdb) disass main
Dump of assembler code for function main:
0x000000000040053c <+0>: push %rbp
0x000000000040053d <+1>: mov %rsp,%rbp
=> 0x0000000000400540 <+4>: movl $0x1,-0x4(%rbp)
0x0000000000400547 <+11>: movl $0x2,-0x4(%rbp)
0x000000000040054e <+18>: mov $0x0,%eax
0x0000000000400553 <+23>: pop %rbp
0x0000000000400554 <+24>: retq
End of assembler dump.
However, now delete the breakpoints:
(gdb) delete
Delete all breakpoints? (y or n) y
And disassembly shows the old breakpoint instructions:
(gdb) disass main
Dump of assembler code for function main:
0x000000000040053c <+0>: push %rbp
0x000000000040053d <+1>: mov %rsp,%rbp
=> 0x0000000000400540 <+4>: int3
0x0000000000400541 <+5>: rex.RB cld
0x0000000000400543 <+7>: add %eax,(%rax)
0x0000000000400545 <+9>: add %al,(%rax)
0x0000000000400547 <+11>: int3
0x0000000000400548 <+12>: rex.RB cld
0x000000000040054a <+14>: add (%rax),%al
0x000000000040054c <+16>: add %al,(%rax)
0x000000000040054e <+18>: mov $0x0,%eax
0x0000000000400553 <+23>: pop %rbp
0x0000000000400554 <+24>: retq
End of assembler dump.
Those breakpoint instructions are no longer installed in target memory
they're stale in the code cache. Easily confirmed by just disabling
the code cache:
(gdb) set code-cache off
(gdb) disass main
Dump of assembler code for function main:
0x000000000040053c <+0>: push %rbp
0x000000000040053d <+1>: mov %rsp,%rbp
=> 0x0000000000400540 <+4>: movl $0x1,-0x4(%rbp)
0x0000000000400547 <+11>: movl $0x2,-0x4(%rbp)
0x000000000040054e <+18>: mov $0x0,%eax
0x0000000000400553 <+23>: pop %rbp
0x0000000000400554 <+24>: retq
End of assembler dump.
I stumbled upon this when writing a patch to infrun.c, that made
handle_inferior_event & co fill in the cache before breakpoints were
removed from the target. Recall that wait_for_inferior flushes the
dcache for every event. So in that case, always-inserted mode was not
necessary to trigger this. It's just a convenient way to expose the
issue.
The dcache works at the raw memory level. We need to update it
whenever memory is written, no matter what kind of target memory
object was originally passed down by the caller. The issue is that
the dcache update code isn't reached when a caller explicitly writes
raw memory. Breakpoint insertion/removal is one such case --
mem-break.c uses target_write_read_memory/target_write_raw_memory.
The fix is to move the dcache update code from memory_xfer_partial_1
to raw_memory_xfer_partial so that it's always reachable.
When we do that, we can actually simplify a series of things.
memory_xfer_partial_1 no longer needs to handle writes for any kind of
memory object, and therefore dcache_xfer_memory no longer needs to
handle writes either. So the latter (dcache_xfer_memory) and its
callees can be simplified to only care about reads. While we're
touching dcache_xfer_memory's prototype, might as well rename it to
reflect that fact that it only handles reads, and make it follow the
new target_xfer_status/xfered_len style. This made me notice that
dcache_xfer_memory loses the real error status if a memory read fails:
we could have failed to read due to TARGET_XFER_E_UNAVAILABLE, for
instance, but we always return TARGET_XFER_E_IO, hence the FIXME note.
I felt that fixing that fell out of the scope of this patch.
Currently dcache_xfer_memory handles the case of a write failing. The
whole cache line is invalidated when that happens. However,
dcache_update, the sole mechanism for handling writes that will remain
after the patch, does not presently handle that scenario. That's a
bug. The patch makes it handle that, by passing down the
target_xfer_status status from the caller, so that it can better
decide what to do itself. While I was changing the function's
prototype, I constified the myaddr parameter, getting rid of the need
for the cast as seen in its existing caller.
Tested on x86_64 Fedora 17, native and gdbserver.
gdb/
2014-03-05 Pedro Alves <palves@redhat.com>
PR gdb/16575
* dcache.c (dcache_poke_byte): Constify ptr parameter. Return
void. Update comment.
(dcache_xfer_memory): Delete.
(dcache_read_memory_partial): New, based on the read bits of
dcache_xfer_memory.
(dcache_update): Add status parameter. Use ULONGEST for len, and
adjust. Discard cache lines if the reason for the update was
error.
* dcache.h (dcache_xfer_memory): Delete declaration.
(dcache_read_memory_partial): New declaration.
(dcache_update): Update prototype.
* target.c (raw_memory_xfer_partial): Update the dcache here.
(memory_xfer_partial_1): Don't handle dcache writes here.
gdb/testsuite/
2014-03-05 Pedro Alves <palves@redhat.com>
PR gdb/16575
* gdb.base/breakpoint-shadow.exp (compare_disassembly): New
procedure.
(top level): Adjust to use it. Add tests that exercise breakpoint
interaction with the code-cache.
As no target uses it anymore, it can finally go away.
After removing the deprecated_xfer_memory handling from
default_xfer_partial, we can delete the latter, because the only thing
it does is delegate to the target beneath unconditionally, which is
what the delegator installed by target-delegates.c will do for us if
no to_xfer_partial method is installed.
This was the last user of de_fault, so that goes away too.
Tested on x86_64 Fedora 17.
gdb/
2014-02-26 Pedro Alves <palves@redhat.com>
* target.c (complete_target_initialization): Don't install
default_xfer_partial as to_xfer_partial hook.
(nomemory): Delete.
(update_current_target): Don't INHERIT nor de_fault
deprecated_xfer_memory. Delete de_fault macro.
(default_xfer_partial, deprecated_debug_xfer_memory): Delete.
(setup_target_debug): Don't install a deprecated_xfer_memory hook.
* target.h (struct target_ops) <deprecated_xfer_memory>: Delete
field.
This removes target_ignore, which isn't used any more.
2014-02-25 Tom Tromey <tromey@redhat.com>
* target.h (target_ignore): Don't declare.
* target.c (target_ignore): Remove.
Nowadays, TARGET_XFER_E_UNAVAILABLE isn't regarded as an error in
to_xfer_partial interface, so _E_ looks odd. This patch is to
replace TARGET_XFER_E_UNAVAILABLE with TARGET_XFER_UNAVAILABLE,
and change its value from -2 to 2. Since there is no comparison
on the value of 'enum target_xfer_status', so it should be safe.
gdb:
2014-02-24 Yao Qi <yao@codesourcery.com>
* target.h (enum target_xfer_status)
<TARGET_XFER_E_UNAVAILABLE>: Rename it to ...
<TARGET_XFER_UNAVAILABLE>: ... it with setting value 2
explicitly. New.
* corefile.c (memory_error_message): User updated.
* exec.c (section_table_read_available_memory): Likewise.
* record-btrace.c (record_btrace_xfer_partial): Likewise.
* target.c (target_xfer_status_to_string): Likewise.
(raw_memory_xfer_partial): Likewise.
(memory_xfer_partial_1, target_xfer_partial): Likewise.
* valops.c (read_value_memory): Likewise.
* exec.h: Update comments.
This patch tweaks target_xfer_status_to_string on comments and argument
name.
gdb:
2014-02-24 Yao Qi <yao@codesourcery.com>
* target.c (target_xfer_status_to_string): Rename argument err
to status.
* target.h (target_xfer_status_to_string): Update declaration.
Replace target_xfer_error_to_string with
target_xfer_status_to_string in comment.
This patch removes macro TARGET_XFER_STATUS_ERROR_P, as Pedro pointed
out during patches review that TARGET_XFER_STATUS_ERROR_P tends to
be unnecessary.
gdb:
2014-02-24 Yao Qi <yao@codesourcery.com>
* target.h (TARGET_XFER_STATUS_ERROR_P): Remove.
* corefile.c (read_memory): Adjusted.
* target.c (target_write_with_progress): Adjusted.
The patch "return target_xfer_status in to_xfer_partial" caused a
regression in various s390(x) test cases, because memory_xfer_partial
filled only the first byte of the read buffer from a breakpoint shadow:
https://sourceware.org/ml/gdb-patches/2014-01/msg01071.html
This patch fixes the regression.
This converts to_get_unwinder and to_get_tailcall_unwinder to methods
and arranges for them to use the new delegation scheme.
This just lets us avoid having a differing style (neither new-style
nor INHERIT) of delegation in the tree.
2014-02-19 Tom Tromey <tromey@redhat.com>
* target.c (target_get_unwinder): Rewrite.
(target_get_tailcall_unwinder): Rewrite.
* record-btrace.c (record_btrace_to_get_unwinder): New function.
(record_btrace_to_get_tailcall_unwinder): New function.
(init_record_btrace_ops): Update.
* target.h (struct target_ops) <to_get_unwinder,
to_get_tailcall_unwinder>: Now function pointers. Use
TARGET_DEFAULT_RETURN.
This converts to_decr_pc_after_break to the new style of delegation,
removing forward_target_decr_pc_after_break.
2014-02-19 Tom Tromey <tromey@redhat.com>
* record-btrace.c (record_btrace_decr_pc_after_break): Delegate
directly.
* target-delegates.c: Rebuild.
* target.h (struct target_ops) <to_decr_pc_after_break>: Use
TARGET_DEFAULT_FUNC.
* target.c (default_target_decr_pc_after_break): Rename from
forward_target_decr_pc_after_break. Simplify.
(target_decr_pc_after_break): Rely on delegation.
This removes a few unnecessary calls to INHERIT and de_fault:
* to_doc is only used when a target is registered
* to_magic is only used when a target is pushed and not useful for
current_target.
* to_open and to_close are only ever called using a specific
target_ops object; there is no need to de_fault them.
2014-02-19 Tom Tromey <tromey@redhat.com>
* target.c (update_current_target): Do not INHERIT to_doc or
to_magic. Do not de_fault to_open or to_close.
This cleans up target.c to avoid function casts.
2014-02-19 Tom Tromey <tromey@redhat.com>
* target.c (complete_target_initialization): Remove casts. Use
return_zero_has_execution.
(return_zero): Add "ignore" argument.
(return_zero_has_execution): New function.
(init_dummy_target): Remove casts. Use
return_zero_has_execution.
During the conversion I kept all the "do not inherit" comments in
update_current_target. However, now they are not needed. This patch
updates the comments for INHERIT and de_fault, and removes the
somewhat odd INHERIT of to_stratum.
2014-02-19 Tom Tromey <tromey@redhat.com>
* target.c (update_current_target): Update comments. Do not
INHERIT to_stratum.
This switches to_read_description to the "new normal" delegation
scheme. This one was a bit trickier than the other changes due to the
way that target_read_description handled delegation. I examined all
the target implementations of to_read_description and changed the ones
returning NULL to instead delegate.
2014-02-19 Tom Tromey <tromey@redhat.com>
* arm-linux-nat.c (arm_linux_read_description): Delegate when
needed.
* corelow.c (core_read_description): Delegate when needed.
* remote.c (remote_read_description): Delegate when needed.
* target-delegates.c: Rebuild.
* target.c (target_read_description): Rewrite.
* target.h (struct target_ops) <to_read_description>: Update
comment. Use TARGET_DEFAULT_RETURN.
Tom Tromey
Pedro Alves
Yao Qi
Andreas Krebbel