| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Some of these references were not updated when they were moved to a separate
directory.
|
|
The gdb_bfd_ref_ptr type is used in solib.h declarations.
|
|
Add debuginfod support to core_target::build_file_mappings and
locate_exec_from_corefile_build_id to enable the downloading of
missing executables and shared libraries referenced in core files.
Also add debuginfod support to solib_map_sections so that previously
downloaded shared libraries can be retrieved from the local debuginfod
cache.
When core file shared libraries are found locally, verify that their
build-ids match the corresponding build-ids found in the core file.
If there is a mismatch, attempt to query debuginfod for the correct
build and print a warning if unsuccessful:
warning: Build-id of /lib64/libc.so.6 does not match core file.
Also disable debuginfod when gcore invokes gdb. Debuginfo is not
needed for core file generation so debuginfod queries will slow down
gcore unnecessarily.
|
|
Since commit aa2d5a422 gdb has been able to read executable and shared
library build-ids within core files.
Expand this functionality so that each core file bfd maintains a map of
soname to build-id for each shared library referenced in the core file.
This feature may be used to verify that gdb has found the correct shared
libraries for core files and to facilitate downloading shared libaries via
debuginfod.
|
|
If GDB reports a watchpoint hit, and then the next event is not
TARGET_WAITKIND_STOPPED, but instead some event for which there's a
catchpoint, such that GDB calls bpstat_stop_status, GDB mistakenly
thinks the watchpoint triggered. Vis, using foll-fork.c:
(gdb) awatch v
Hardware access (read/write) watchpoint 2: v
(gdb) catch fork
Catchpoint 3 (fork)
(gdb) c
Continuing.
Hardware access (read/write) watchpoint 2: v
Old value = 0
New value = 5
main () at gdb.base/foll-fork.c:16
16 pid = fork ();
(gdb)
Continuing.
Hardware access (read/write) watchpoint 2: v <<<<
<<<< these lines are spurious
Value = 5 <<<<
Catchpoint 3 (forked process 1712369), arch_fork (ctid=0x7ffff7fa4810) at arch-fork.h:49
49 arch-fork.h: No such file or directory.
(gdb)
The problem is that when we handle the fork event, nothing called
watchpoints_triggered before calling bpstat_stop_status. Thus, each
watchpoint's watchpoint_triggered field was still set to
watch_triggered_yes from the previous (real) watchpoint stop.
watchpoint_triggered is only current called in the handle_signal_stop
path, when handling TARGET_WAITKIND_STOPPED.
This fixes it by adding watchpoint_triggered calls in the other events
paths that call bpstat_stop_status. But instead of adding them
explicitly, it adds a new function bpstat_stop_status_nowatch that
wraps bpstat_stop_status and calls watchpoint_triggered, and then
replaces most calls to bpstat_stop_status with calls to
bpstat_stop_status_nowatch.
This required constifying watchpoints_triggered.
New test included, which fails without the fix.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=28621
Change-Id: I282b38c2eee428d25319af3bc842f9feafed461c
|
|
I'm getting this diff when running `autoreconf -vf`.
Change-Id: Id5f009d0f0481935c1ee9df5332cb4bf45fbd32d
|
|
On x86 machines with xmm register, and with recent versions of
systemtap (and gcc?), it can occur that stap probe arguments will be
placed into xmm registers.
I notice this happening on a current Fedora Rawhide install with the
following package versions installed:
$ rpm -q glibc systemtap gcc
glibc-2.35.9000-10.fc37.x86_64
systemtap-4.7~pre16468670g9f253544-1.fc37.x86_64
gcc-12.0.1-0.12.fc37.x86_64
If I check the probe data in libc, I see this:
$ readelf -n /lib64/libc.so.6
...
stapsdt 0x0000004d NT_STAPSDT (SystemTap probe descriptors)
Provider: libc
Name: pthread_start
Location: 0x0000000000090ac3, Base: 0x00000000001c65c4, Semaphore: 0x0000000000000000
Arguments: 8@%xmm1 8@1600(%rbx) 8@1608(%rbx)
stapsdt 0x00000050 NT_STAPSDT (SystemTap probe descriptors)
Provider: libc
Name: pthread_create
Location: 0x00000000000912f1, Base: 0x00000000001c65c4, Semaphore: 0x0000000000000000
Arguments: 8@%xmm1 8@%r13 8@8(%rsp) 8@16(%rsp)
...
Notice that for both of these probes, the first argument is a uint64_t
stored in the xmm1 register.
Unfortunately, if I try to use this probe within GDB, then I can't
view the first argument. Here's an example session:
$ gdb $(which gdb)
(gdb) start
...
(gdb) info probes stap libc pthread_create
...
(gdb) break *0x00007ffff729e2f1 # Use address of probe.
(gdb) continue
...
(gdb) p $_probe_arg0
Invalid cast.
What's going wrong? If I re-run my session, but this time use 'set
debug stap-expression 1', this is what I see:
(gdb) set debug stap-expression 1
(gdb) p $_probe_arg0
Operation: UNOP_CAST
Operation: OP_REGISTER
String: xmm1
Type: uint64_t
Operation: UNOP_CAST
Operation: OP_REGISTER
String: r13
Type: uint64_t
Operation: UNOP_CAST
Operation: UNOP_IND
Operation: UNOP_CAST
Operation: BINOP_ADD
Operation: OP_LONG
Type: long
Constant: 0x0000000000000008
Operation: OP_REGISTER
String: rsp
Type: uint64_t *
Type: uint64_t
Operation: UNOP_CAST
Operation: UNOP_IND
Operation: UNOP_CAST
Operation: BINOP_ADD
Operation: OP_LONG
Type: long
Constant: 0x0000000000000010
Operation: OP_REGISTER
String: rsp
Type: uint64_t *
Type: uint64_t
Invalid cast.
(gdb)
The important bit is this:
Operation: UNOP_CAST
Operation: OP_REGISTER
String: xmm1
Type: uint64_t
Which is where we cast the xmm1 register to uint64_t. And the final
piece of the puzzle is:
(gdb) ptype $xmm1
type = union vec128 {
v8bf16 v8_bfloat16;
v4f v4_float;
v2d v2_double;
v16i8 v16_int8;
v8i16 v8_int16;
v4i32 v4_int32;
v2i64 v2_int64;
uint128_t uint128;
}
So, we are attempting to cast a union type to a scalar type, which is
not supporting in C/C++, and as a consequence GDB's expression
evaluator throws an error when we attempt to do this.
The first approach I considered for solving this problem was to try
and make use of gdbarch_stap_adjust_register. We already have a
gdbarch method (gdbarch_stap_adjust_register) that allows us to tweak
the name of the register that we access. Currently only x86
architectures use this to transform things like ax to eax in some
cases.
I wondered, what if we change gdbarch_stap_adjust_register to do more
than just change the register names? What if this method instead
became gdbarch_stap_read_register. This new method would return a
operation_up, and would take the register name, and the type we are
trying to read from the register, and return the operation that
actually reads the register.
The default implementation of this method would just use
user_reg_map_name_to_regnum, and then create a register_operation,
like we already do in stap_parse_register_operand. But, for x86
architectures this method would fist possibly adjust the register
name, then do the default action to read the register. Finally, for
x86 this method would spot when we were accessing an xmm register,
and, based on the type being pulled from the register, would extract
the correct field from the union.
The benefit of this approach is that it would work with the expression
types that GDB currently supports. The draw back would be that this
approach would not be very generic. We'd need code to handle each
sub-field size with an xmm register. If other architectures started
using vector registers for probe arguments, those architectures would
have to create their own gdbarch_stap_read_register method. And
finally, the type of the xmm registers comes from the type defined in
the target description, there's a risk that GDB might end up
hard-coding the names of type sub-fields, then if a target uses a
different target description, with different field names for xmm
registers, the stap probes would stop working.
And so, based on all the above draw backs, I rejected this first
approach.
My second plan involves adding a new expression type to GDB called
unop_extract_operation. This new expression takes a value and a type,
during evaluation the value contents are fetched, and then a new value
is extracted from the value contents (based on type). This is similar
to the following C expression:
result_value = *((output_type *) &input_value);
Obviously we can't actually build this expression in this case, as the
input_value is in a register, but hopefully the above makes it clearer
what I'm trying to do.
The benefit of the new expression approach is that this code can be
shared across all architectures, and it doesn't care about sub-field
names within the union type.
The draw-backs that I see are potential future problems if arguments
are not stored within the least significant bytes of the register.
However if/when that becomes an issue we can adapt the
gdbarch_stap_read_register approach to allow architectures to control
how a value is extracted.
For testing, I've extended the existing gdb.base/stap-probe.exp test
to include a function that tries to force an argument into an xmm
register. Obviously, that will only work on a x86 target, so I've
guarded the new function with an appropriate GCC define. In the exp
script we use readelf to check if the probe exists, and is using the
xmm register.
If the probe doesn't exist then the associated tests are skipped.
If the probe exists, put isn't using the xmm register (which will
depend on systemtap/gcc versions), then again, the tests are skipped.
Otherwise, we can run the test. I think the cost of running readelf
is pretty low, so I don't feel too bad making all the non-xmm targets
running this step.
I found that on a Fedora 35 install, with these packages installed, I
was able to run this test and have the probe argument be placed in an
xmm register:
$ rpm -q systemtap gcc glibc
systemtap-4.6-4.fc35.x86_64
gcc-11.2.1-9.fc35.x86_64
glibc-2.34-7.fc35.x86_64
Finally, as this patch adds a new operation type, then I need to
consider how to generate an agent expression for the new operation
type.
I have kicked the can down the road a bit on this. In the function
stap_parse_register_operand, I only create a unop_extract_operation in
the case where the register type is non-scalar, this means that in
most cases I don't need to worry about generating an agent expression
at all.
In the xmm register case, when an unop_extract_operation will be
created, I have sketched out how the agent expression could be
handled, however, this code is currently not reached. When we try to
generate the agent expression to place the xmm register on the stack,
GDB hits this error:
(gdb) trace -probe-stap test:xmmreg
Tracepoint 1 at 0x401166
(gdb) actions
Enter actions for tracepoint 1, one per line.
End with a line saying just "end".
>collect $_probe_arg0
Value not scalar: cannot be an rvalue.
This is because GDB doesn't currently support placing non-scalar types
on the agent expression evaluation stack. Solving this is clearly
related to the original problem, but feels a bit like a second
problem. I'd like to get feedback on whether my approach to solving
the original problem is acceptable or not before I start looking at
how to handle xmm registers within agent expressions.
|
|
Run black on the file.
Change-Id: Ifb576137fb7158a0227173f61c1202f0695b3685
|
|
The test case added here is testing the bug gdb/28856, where calling a
function by hand from a pretty printer makes GDB crash. There are 6
mechanisms to trigger this crash in the current test, using the commands
backtrace, up, down, finish, step and continue. Since the failure happens
because of use-after-free (more details below) the tests will always
have a chance of passing through sheer luck, but anecdotally they seem
to fail all of the time.
The reason GDB is crashing is a use-after-free problem. The above
mentioned functions save a pointer to the current frame's information,
then calls the pretty printer, and uses the saved pointer for different
reasons, depending on the function. The issue happens because
call_function_by_hand needs to reset the obstack to get the current
frame, invalidating the saved pointer.
|
|
In testsuite/README, we suggest that you can run the testsuite against
some other GDB binary by using:
make check RUNTESTFLAGS=GDB=/usr/bin/gdb
However, that example isn't fully correct, because with that command
line, the testsuite will still pass
-data-directory=[pwd]/../data-directory
to /usr/bin/gdb, like e.g.:
...
builtin_spawn /usr/bin/gdb -nw -nx -data-directory /home/pedro/gdb/build/gdb/testsuite/../data-directory -iex set height 0 -iex set width 0
...
while if you're testing an installed GDB (the system GDB being the
most usual scenario), then you should normally let it use its own
configured directory, not the just-built GDB's data directory.
This commit improves the status quo with the following two changes:
- if the user specifies GDB on the command line, then by default,
don't start GDB with the -data-directory command line option.
I.e., let the tested GDB use its own configured data directory.
- let the user override the data directory, via a new
GDB_DATA_DIRECTORY global. This replaces the existing
BUILD_DATA_DIRECTORY variable in testsuite/lib/gdb.exp, which
wasn't overridable, and was a bit misnamed for the new purpose.
So after this, the following commands I believe behave intuitively:
# Test the non-installed GDB in some build dir:
make check \
RUNTESTFLAGS="GDB=/path/to/other/build/gdb \
GDB_DATA_DIRECTORY=/path/to/other/build/gdb/data-directory"
# Test the GDB installed in some prefix:
make check \
RUNTESTFLAGS="GDB=/opt/gdb/bin/gdb"
# Test the built GDB with some alternative data directory, e.g., the
system GDB's data directory:
make check \
RUNTESTFLAGS="GDB_DATA_DIRECTORY=/usr/share/gdb"
Change-Id: Icdc21c85219155d9564a9900961997e6624b78fb
|
|
In readline 8.2 the type of rl_completer_word_break_characters changed to
include const.
|
|
This commit a new section for the next release branch, and renames
the section of the current branch, now that it has been cut.
|
|
Now that the GDB 12 branch has been created,
this commit bumps the version number in gdb/version.in to
13.0.50.DATE-git
For the record, the GDB 12 branch was created
from commit 2be64de603f8b3ae359d2d3fbf5db0e79869f32b.
Also, as a result of the version bump, the following changes
have been made in gdb/testsuite:
* gdb.base/default.exp: Change $_gdb_major to 13.
|
|
The motivation for this patch is the fact that py-micmd.c doesn't build
with Python 2, due to PyDict_GetItemWithError being a Python 3-only
function:
CXX python/py-micmd.o
/home/smarchi/src/binutils-gdb/gdb/python/py-micmd.c: In function ‘int micmdpy_uninstall_command(micmdpy_object*)’:
/home/smarchi/src/binutils-gdb/gdb/python/py-micmd.c:430:20: error: ‘PyDict_GetItemWithError’ was not declared in this scope; did you mean ‘PyDict_GetItemString’?
430 | PyObject *curr = PyDict_GetItemWithError (mi_cmd_dict.get (),
| ^~~~~~~~~~~~~~~~~~~~~~~
| PyDict_GetItemString
A first solution to fix this would be to try to replace
PyDict_GetItemWithError equivalent Python 2 code. But I looked at why
we are doing this in the first place: it is to maintain the
`gdb._mi_commands` Python dictionary that we use as a `name ->
gdb.MICommand object` map. Since the `gdb._mi_commands` dictionary is
never actually used in Python, it seems like a lot of trouble to use a
Python object for this.
My first idea was to replace it with a C++ map
(std::unordered_map<std::string, gdbpy_ref<micmdpy_object>>). While
implementing this, I realized we don't really need this map at all. The
mi_command_py objects registered in the main MI command table can own
their backing micmdpy_object (that's a gdb.MICommand, but seen from the
C++ code). To know whether an mi_command is an mi_command_py, we can
use a dynamic cast. Since there's one less data structure to maintain,
there are less chances of messing things up.
- Change mi_command_py::m_pyobj to a gdbpy_ref, the mi_command_py is
now what keeps the MICommand alive.
- Set micmdpy_object::mi_command in the constructor of mi_command_py.
If mi_command_py manages setting/clearing that field in
swap_python_object, I think it makes sense that it also takes care of
setting it initially.
- Move a bunch of checks from micmdpy_install_command to
swap_python_object, and make them gdb_asserts.
- In micmdpy_install_command, start by doing an mi_cmd_lookup. This is
needed to know whether there's a Python MI command already registered
with that name. But we can already tell if there's a non-Python
command registered with that name. Return an error if that happens,
rather than waiting for insert_mi_cmd_entry to fail. Change the
error message to "name is already in use" rather than "may already be
in use", since it's more precise.
I asked Andrew about the original intent of using a Python dictionary
object to hold the command objects. The reason was to make sure the
objects get destroyed when the Python runtime gets finalized, not later.
Holding the objects in global C++ data structures and not doing anything
more means that the held Python objects will be decref'd after the
Python interpreter has been finalized. That's not desirable. I tried
it and it indeed segfaults.
Handle this by adding a gdbpy_finalize_micommands function called in
finalize_python. This is the mirror of gdbpy_initialize_micommands
called in do_start_initialization. In there, delete all Python MI
commands. I think it makes sense to do it this way: if it was somehow
possible to unload Python support from GDB in the middle of a session
we'd want to unregister any Python MI command. Otherwise, these MI
commands would be backed with a stale PyObject or simply nothing.
Delete tests that were related to `gdb._mi_commands`.
Co-Authored-By: Andrew Burgess <aburgess@redhat.com>
Change-Id: I060d5ebc7a096c67487998a8a4ca1e8e56f12cd3
|
|
A stepi in a function without debug info with "set debug infrun 1"
crashes GDB since commit c8353d682f69 (gdb/infrun: some extra infrun
debug print statements), due to a reference to
"tp->current_symtab->filename" when tp->current_symtab is null.
This commit adds the missing null check. The output in this case
becomes:
[infrun] set_step_info: symtab = <null>, line = 0, step_frame_id = {stack=0x7fffffffd980,code=0x0000000000456c30,!special}, step_stack_frame_id = {stack=0x7fffffffd980,code=0x0000000000456c30,!special}
Change-Id: I5171a9d222bc7e15b9dba2feaba7d55c7acd99f8
|
|
The internal AdaCore testsuite has a test that checks that an
out-of-scope watchpoint is deleted. This fails on some aarch64
configurations, reporting an extra stop:
(gdb) continue
Continuing.
Thread 3 hit Watchpoint 2: result
Old value = 64
New value = 0
0x0000000040021648 in pck.get_val (seed=0, off_by_one=false) at [...]/pck.adb:13
13 end Get_Val;
I believe what is happening here is that the variable is stored at:
<efa> DW_AT_location : 2 byte block: 91 7c (DW_OP_fbreg: -4)
and the extra stop is reported just before a return, when the ldp
instruction is executed:
0x0000000040021644 <+204>: ldp x29, x30, [sp], #48
0x0000000040021648 <+208>: ret
This instruction modifies the frame base calculation, and so the test
picks up whatever memory is pointed to in the callee frame.
Implementing the gdbarch hook gdbarch_stack_frame_destroyed_p fixes
this problem.
As usual with this sort of patch, it has passed internal testing, but
I don't have a good way to try it with dejagnu. So, I don't know
whether some existing test covers this. I suspect there must be one,
but it's also worth noting that this test passes for aarch64 in some
configurations -- I don't know what causes one to fail and another to
succeed.
|
|
Seems like some leftovers from previous commits.
Change-Id: I7155ccdf7a4fef83bcb3d60320252c3628efdb83
|
|
The cases for TARGET_OBJECT_LIBRARIES and TARGET_OBJECT_LIBRARIES_AIX
can try to fetch different data objects (such as
TARGET_OBJECT_SIGNAL_INFO) if gdbarch methods for the requested data
aren't present. Return with TARGET_XFER_E_IO if the gdbarch method
isn't present instead.
|
|
GCC removed support for score back in 2014 already. Back then, we
basically agreed about removing it from GDB too, but it ended up being
forgotten. See:
https://sourceware.org/pipermail/gdb/2014-October/044643.html
Following through this time around.
Change-Id: I5b25a1ff7bce7b150d6f90f4c34047fae4b1f8b4
|
|
In an earlier patch, I had written that I wanted to add this test:
ptype Wide_Wide_String'("literal")
... but that it failed with the distro GNAT. Further investigation
showed that it could be made to work by adding a function using
Wide_Wide_String to the program -- this caused the type to end up in
the debug info.
This patch adds the test. I'm checking this in.
|
|
In this commit:
commit b4f26d541aa7224b70d363932e816e6e1a857633
Date: Tue Mar 2 13:42:37 2021 -0700
Import GNU Readline 8.1
We imported readline 8.1 into GDB. As a consequence bug PR cli/28833
was reported. This bug spotted that, when the user terminated GDB by
sending EOF (usually bound to Ctrl+d), the last prompt would become
corrupted. Here's what happens, the user is sat at a prompt like
this:
(gdb)
And then the user sends EOF (Ctrl+d), we now see this:
quit)
... gdb terminates, and we return to the shell ...
Notice the 'quit' was printed over the prompt.
This problem is a result of readline 8.1 enabling bracketed paste mode
by default. This problem is present in readline 8.0 too, but in that
version of readline bracketed paste mode is off by default, so a user
will not see the bug unless they specifically enable the feature.
Bracketed paste mode is available in readline 7.0 too, but the bug
is not present in this version of readline, see below for why.
What causes this problem is how readline disables bracketed paste
mode. Bracketed paste mode is a terminal feature that is enabled and
disabled by readline emitting a specific escape sequence. The problem
for GDB is that the escape sequence to disable bracketed paste mode
includes a '\r' character at the end, see this thread for more
details:
https://lists.gnu.org/archive/html/bug-bash/2018-01/msg00097.html
The change to add the '\r' character to the escape sequence used to
disable bracketed paste mode was introduced between readline 7.0 and
readline 8.0, this is why the bug would not occur when using older
versions of readline (note: I don't know if its even possible to build
GDB using readline 7.0. That really isn't important, I'm just
documenting the history of this issue).
So, the escape sequence to disable bracketed paste mode is emitted
from the readline function rl_deprep_terminal, this is called after
the user has entered a complete command and pressed return, or, if the
user sends EOF.
However, these two cases are slightly different. In the first case,
when the user has entered a command and pressed return, the cursor
will have moved to the next, empty, line, before readline emits the
escape sequence to leave bracketed paste mode. The final '\r'
character moves the cursor back to the beginning of this empty line,
which is harmless.
For the EOF case though, this is not what happens. Instead, the
escape sequence to leave bracketed paste mode is emitted on the same
line as the prompt. The final '\r' moves the cursor back to the start
of the prompt line. This leaves us ready to override the prompt.
It is worth noting, that this is not the intended behaviour of
readline, in rl_deprep_terminal, readline should emit a '\n' character
when EOF is seen. However, due to a bug in readline this does not
happen (the _rl_eof_found flag is never set). This is the first
readline bug that effects GDB.
GDB prints the 'quit' message from command_line_handler (in
event-top.c), this function is called (indirectly) from readline to
process the complete command line, but also in the EOF case (in which
case the command line is set to nullptr). As this is part of the
callback to process a complete command, this is called after readline
has disabled bracketed paste mode (by calling rl_deprep_terminal).
And so, when bracketed paste mode is in use, rl_deprep_terminal leaves
the cursor at the start of the prompt line (in the EOF case), and
command_line_handler then prints 'quit', which overwrites the prompt.
The solution to this problem is to print the 'quit' message earlier,
before rl_deprep_terminal is called. This is easy to do by using the
rl_deprep_term_function hook. It is this hook that usually calls
rl_deprep_terminal, however, if we replace this with a new function,
we can print the 'quit' string, and then call rl_deprep_terminal
ourselves. This allows the 'quit' to be printed before
rl_deprep_terminal is called.
The problem here is that there is no way in rl_deprep_terminal to know
if readline is processing EOF or not, and as a result, we don't know
when we should print 'quit'. This is the second readline bug that
effects GDB.
Both of these readline issues are discussed in this thread:
https://lists.gnu.org/archive/html/bug-readline/2022-02/msg00021.html
The result of that thread was that readline was patched to address
both of these issues.
Now it should be easy to backport the readline fix to GDB's in tree
copy of readline, and then change GDB to make use of these fixes to
correctly print the 'quit' string.
However, we are just about to branch GDB 12, and there is concern from
some that changing readline this close to a new release is a risky
idea, see this thread:
https://sourceware.org/pipermail/gdb-patches/2022-March/186391.html
So, this commit doesn't change readline at all. Instead, this commit
is the smallest possible GDB change in order to avoid the prompt
corruption.
In this commit I change GDB to print the 'quit' string on the line
after the prompt, but only when bracketed paste mode is on. This
avoids the overwriting issue, the user sees this:
(gdb)
quit
... gdb terminates, and returns to the shell ...
This isn't ideal, but is better than the existing behaviour. After
GDB 12 has branched, we can backport the readline fix, and apply a
real fix to GDB.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=28833
|
|
This started as a patch to implement string concatenation for Ada.
However, while working on this, I looked at how this code could
possibly be called. It turns out there are only two users of
concat_operation: Ada and D. So, in addition to implementing this for
Ada, this patch rewrites value_concat, removing the odd "concatenate
or repeat" semantics, which were completely unused. As Ada and D both
seem to represent strings using TYPE_CODE_ARRAY, this removes the
TYPE_CODE_STRING code from there as well.
|
|
eval_op_concat has code to search for an operator overload of
BINOP_CONCAT. However, the operator overloading code is specific to
C++, which does not have this operator. And,
binop_types_user_defined_p rejects this case right at the start, and
value_x_binop does not handle this case. I think this code has been
dead for a very long time. This patch removes it and hoists the
remaining call into concatenation::evaluate, removing eval_op_concat
entirely.
|
|
This adds some basic support for Wide_String and Wide_Wide_String to
the Ada expression evaluator. In particular, a string literal may be
converted to a wide or wide-wide string depending on context.
The patch updates an existing test case. Note that another test,
namely something like:
ptype Wide_Wide_String'("literal")
... would be nice to add, but when tested against a distro GNAT, this
did not work (probably due to lack of debuginfo); so, I haven't
included it here.
|
|
eval_op_string is only used in a single place -- the implementation of
string_operation. This patch turns it into the
string_operation::evaluate method, removing a bit of extraneous code.
|
|
The last two tests in gdb.base/ending-run.exp case fail on Powerpc when the
system does not have the needed glibc debug-info files loaded. In this
case, gdb is not able to determine where execution stopped. This behavior
looks as follows for the test case:
The next to the last test does a next command when the program is stopped
at the closing bracket for main. The message printed is:
0x00007ffff7d01524 in ?? () from /lib/powerpc64le-linux-gnu/libc.so.6
which fails to match any of the test_multiple options.
The test then does another next command. On Powerpc, the
message printed it:
Cannot find bounds of current function
The test fails as the output does not match any of the options for the
gdb_test_multiple.
I checked the behavior on Powerpc to see if this is typical.
I ran gdb on the following simple program as shown below.
#include <stdio.h>
int
main(void)
{
printf("Hello, world!\n");
return 0;
}
gdb ./hello_world
<snip the gdb start info>
Type "apropos word" to search for commands related to "word"...
Reading symbols from ./hello_world...
(No debugging symbols found in ./hello_world)
(gdb) break main
Breakpoint 1 at 0x818
(gdb) r
Starting program: /home/carll/hello_world
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/powerpc64le-linux-gnu/libthread_db.so.1".
Breakpoint 1, 0x0000000100000818 in main ()
(gdb) n
Single stepping until exit from function main,
which has no line number information.
Hello, world!
0x00007ffff7d01524 in ?? () from /lib/powerpc64le-linux-gnu/libc.so.6
(gdb) n
Cannot find bounds of current function
So it would seem that the messages seen from the test case are
"normal" output for Powerpc when the debug-info is not available.
The following patch adds the output from Powerpc as an option
to the gdb_test_multiple statement, identifying the output as the expected
output on Powerpc without the needed debug-info files installed.
The patch has been tested on a Power 10 system and an Intel
64-bit system. No additional regression failures were seen on
either platform.
|
|
GDB notifies users about user selected thread changes somewhat
inconsistently as mentioned on gdb-patches mailing list here:
https://sourceware.org/pipermail/gdb-patches/2022-February/185989.html
Consider GDB debugging a multi-threaded inferior with both CLI and GDB/MI
interfaces connected to separate terminals.
Assuming inferior is stopped and thread 1 is selected, when a thread
2 is selected using '-thread-select 2' command on GDB/MI terminal:
-thread-select 2
^done,new-thread-id="2",frame={level="0",addr="0x00005555555551cd",func="child_sub_function",args=[],file="/home/jv/Projects/gdb/users_jv_patches/gdb/testsuite/gdb.mi/user-selected-context-sync.c",fullname="/home/uuu/gdb/gdb/testsuite/gdb.mi/user-selected-context-sync.c",line="30",arch="i386:x86-64"}
(gdb)
and on CLI terminal we get the notification (as expected):
[Switching to thread 2 (Thread 0x7ffff7daa640 (LWP 389659))]
#0 child_sub_function () at /home/uuu/gdb/gdb/testsuite/gdb.mi/user-selected-context-sync.c:30
30 volatile int dummy = 0;
However, now that thread 2 is selected, if thread 1 is selected
using 'thread-select --thread 1 1' command on GDB/MI terminal
terminal:
-thread-select --thread 1 1
^done,new-thread-id="1",frame={level="0",addr="0x0000555555555294",func="main",args=[],file="/home/jv/Projects/gdb/users_jv_patches/gdb/testsuite/gdb.mi/user-selected-context-sync.c",fullname="/home/jv/Projects/gdb/users_jv_patches/gdb/testsuite/gdb.mi/user-selected-context-sync.c",line="66",arch="i386:x86-64"}
(gdb)
but no notification is printed on CLI terminal, despite the fact
that user selected thread has changed.
The problem is that when `-thread-select --thread 1 1` is executed
then thread is switched to thread 1 before mi_cmd_thread_select () is
called, therefore the condition "inferior_ptid != previous_ptid"
there does not hold.
To address this problem, we have to move notification logic up to
mi_cmd_execute () where --thread option is processed and notify
user selected contents observers there if context changes.
However, this in itself breaks GDB/MI because it would cause context
notification to be sent on MI channel. This is because by the time
we notify, MI notification suppression is already restored (done in
mi_command::invoke(). Therefore we had to lift notification suppression
logic also up to mi_cmd_execute (). This change in made distinction
between mi_command::invoke() and mi_command::do_invoke() unnecessary
as all mi_command::invoke() did (after the change) was to call
do_invoke(). So this patches removes do_invoke() and moves the command
execution logic directly to invoke().
With this change, all gdb.mi tests pass, tested on x86_64-linux.
Co-authored-by: Andrew Burgess <aburgess@redhat.com>
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=20631
|
|
gdb.fortran/array-slices.exp and gdb.fortran/lbound-ubound.exp were
recently disabled unless testing with the native target, because they
rely on inferior I/O. However, when testing against gdbserver using
the native-gdbserver/native-extended-gdbserver boards, we do have
access to inferior I/O.
The right way to check whether the board can do I/O, is via checking
the gdb,noinferiorio board variable. Switch to using that.
And then, tweak the testcases to expect output to appear in
inferior_spawn_id, instead of gdb_spawn_id. When testing against the
native target, inferior_spawn_id is the same as gdb_spawn_id. When
testing against gdbserver, it maps to gdbserver_spawn_id.
This exposed a buglet in gdb.fortran/array-slices.f90's show_1d
subroutine -- it was missing printing newline at the end of the
"Expected GDB Output" text, leading to a test timeout. All other
subroutines end with advance=yes, except this one. Fix it by using
advance=yes here too.
Change-Id: I4640729f334431cfc24b0917e7d3977b677c6ca5
|
|
Simon pointed out that commit 13835d88 ("Use function view when
iterating over block symbols") caused a regression. The bug is that
the new closure captures 'pc' by reference, but later code updates
this variable -- but the earlier code did not update the callback
structure with the new value.
This patch restores the old behavior by using a new varible name in an
inner scope.
|
|
Rename a proc in gdb.mi/user-selected-context-sync.exp, I think the
old name was most likely a typo. The old name
match_re_or_ensure_not_output seems (to me) to imply we're in some way
checking that the regexp was not output. But that's not what we are
doing, we're checking either for the regexp, or for no output, hence
the new name match_re_or_ensure_no_output.
Additionally, I found a definite typo in one of the comments that I've
also fixed.
I also updated some test names. These tests (probably due to copy &
paste errors) has 'on MI' on their name, when they were actually
checking CLI output. For these test I changed the name to use 'on
CLI'.
There should be no change in what is tested after this commit.
|
|
When execute the following command on LoongArch:
make check-gdb TESTS="gdb.base/align-c.exp"
there exist some failed testcases:
...
FAIL: gdb.base/align-c.exp: print _Alignof(struct align_pair_long_double_x_float)
FAIL: gdb.base/align-c.exp: print _Alignof(struct align_pair_long_double_x_double)
FAIL: gdb.base/align-c.exp: print _Alignof(struct align_pair_long_double_x_long_double)
...
According to LoongArch ELF ABI specification [1], set the target data types
of floating-point to fix the above failed testcases.
[1] https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
|
|
This commit allows a user to create custom MI commands using Python
similarly to what is possible for Python CLI commands.
A new subclass of mi_command is defined for Python MI commands,
mi_command_py. A new file, gdb/python/py-micmd.c contains the logic
for Python MI commands.
This commit is based on work linked too from this mailing list thread:
https://sourceware.org/pipermail/gdb/2021-November/049774.html
Which has also been previously posted to the mailing list here:
https://sourceware.org/pipermail/gdb-patches/2019-May/158010.html
And was recently reposted here:
https://sourceware.org/pipermail/gdb-patches/2022-January/185190.html
The version in this patch takes some core code from the previously
posted patches, but also has some significant differences, especially
after the feedback given here:
https://sourceware.org/pipermail/gdb-patches/2022-February/185767.html
A new MI command can be implemented in Python like this:
class echo_args(gdb.MICommand):
def invoke(self, args):
return { 'args': args }
echo_args("-echo-args")
The 'args' parameter (to the invoke method) is a list
containing (almost) all command line arguments passed to the MI
command (--thread and --frame are handled before the Python code is
called, and removed from the args list). This list can be empty if
the MI command was passed no arguments.
When used within gdb the above command produced output like this:
(gdb)
-echo-args a b c
^done,args=["a","b","c"]
(gdb)
The 'invoke' method of the new command must return a dictionary. The
keys of this dictionary are then used as the field names in the mi
command output (e.g. 'args' in the above).
The values of the result returned by invoke can be dictionaries,
lists, iterators, or an object that can be converted to a string.
These are processed recursively to create the mi output. And so, this
is valid:
class new_command(gdb.MICommand):
def invoke(self,args):
return { 'result_one': { 'abc': 123, 'def': 'Hello' },
'result_two': [ { 'a': 1, 'b': 2 },
{ 'c': 3, 'd': 4 } ] }
Which produces output like:
(gdb)
-new-command
^done,result_one={abc="123",def="Hello"},result_two=[{a="1",b="2"},{c="3",d="4"}]
(gdb)
I have required that the fields names used in mi result output must
match the regexp: "^[a-zA-Z][-_a-zA-Z0-9]*$" (without the quotes).
This restriction was never written down anywhere before, but seems
sensible to me, and we can always loosen this rule later if it proves
to be a problem. Much harder to try and add a restriction later, once
people are already using the API.
What follows are some details about how this implementation differs
from the original patch that was posted to the mailing list.
In this patch, I have changed how the lifetime of the Python
gdb.MICommand objects is managed. In the original patch, these object
were kept alive by an owned reference within the mi_command_py object.
As such, the Python object would not be deleted until the
mi_command_py object itself was deleted.
This caused a problem, the mi_command_py were held in the global mi
command table (in mi/mi-cmds.c), which, as a global, was not cleared
until program shutdown. By this point the Python interpreter has
already been shutdown. Attempting to delete the mi_command_py object
at this point was causing GDB to try and invoke Python code after
finalising the Python interpreter, and we would crash.
To work around this problem, the original patch added code in
python/python.c that would search the mi command table, and delete the
mi_command_py objects before the Python environment was finalised.
In contrast, in this patch, I have added a new global dictionary to
the gdb module, gdb._mi_commands. We already have several such global
data stores related to pretty printers, and frame unwinders.
The MICommand objects are placed into the new gdb.mi_commands
dictionary, and it is this reference that keeps the objects alive.
When GDB's Python interpreter is shut down gdb._mi_commands is deleted,
and any MICommand objects within it are deleted at this point.
This change avoids having to make the mi_cmd_table global, and walk
over it from within GDB's python related code.
This patch handles command redefinition entirely within GDB's python
code, though this does impose one small restriction which is not
present in the original code (detailed below), I don't think this is a
big issue. However, the original patch relied on being able to
finish executing the mi_command::do_invoke member function after the
mi_command object had been deleted. Though continuing to execute a
member function after an object is deleted is well defined, it is
also (IMHO) risky, its too easy for someone to later add a use of the
object without realising that the object might sometimes, have been
deleted. The new patch avoids this issue.
The one restriction that is added to avoid this, is that an MICommand
object can't be reinitialised with a different command name, so:
(gdb) python cmd = MyMICommand("-abc")
(gdb) python cmd.__init__("-def")
can't reinitialize object with a different command name
This feels like a pretty weird edge case, and I'm happy to live with
this restriction.
I have also changed how the memory is managed for the command name.
In the most recently posted patch series, the command name is moved
into a subclass of mi_command, the python mi_command_py, which
inherits from mi_command is then free to use a smart pointer to manage
the memory for the name.
In this patch, I leave the mi_command class unchanged, and instead
hold the memory for the name within the Python object, as the lifetime
of the Python object always exceeds the c++ object stored in the
mi_cmd_table. This adds a little more complexity in py-micmd.c, but
leaves the mi_command class nice and simple.
Next, this patch adds some extra functionality, there's a
MICommand.name read-only attribute containing the name of the command,
and a read-write MICommand.installed attribute that can be used to
install (make the command available for use) and uninstall (remove the
command from the mi_cmd_table so it can't be used) the command. This
attribute will be automatically updated if a second command replaces
an earlier command.
This patch adds additional error handling, and makes more use the
gdbpy_handle_exception function.
Co-Authored-By: Jan Vrany <jan.vrany@labware.com>
|
|
After the previous commit, which removes the predicate function
gdbarch_register_type_p, I noticed that the gdbarch->register_type
field was not checked at in the verify_gdbarch function.
More than not being checked, the field wasn't mentioned at all.
I find this strange, I would expect that every field would at least be
mentioned - we already generate comments for some fields saying that
this field is _not_ being checked, so the fact that this field isn't
being checked looks (to me), like this field is somehow slipping
through the cracks.
The comment at the top of gdbarch-components.py tries to explain how
the validation is done. I didn't understand this comment completely,
but, I think this final sentence:
"Otherwise, the check is done against 0 (really NULL for function
pointers, but same idea)."
Means that, if non of the other cases apply, then the field should be
checked against 0, with 0 indicating that the field is invalid (was
not set by the tdep code). However, this is clearly not being done.
Looking in gdbarch.py at the code to generate verify_gdbarch we do
find that there is a case that is not handled, the case where the
'invalid' field is set true True, but non of the other cases apply.
In this commit I propose two changes:
1. Handle the case where the 'invalid' field of a property is set to
True, this should perform a check for the field of gdbarch still
being set to 0, and
2. If the if/else series that generates verify_gdbarch doesn't handle
a property then we should raise an exception. This means that if a
property is added which isn't handled, we should no longer silently
ignore it.
After doing this, I re-generated the gdbarch files and saw that the
following gdbarch fields now had new validation checks:
register_type
believe_pcc_promotion
register_to_value
value_to_register
frame_red_zone_size
displaced_step_restore_all_in_ptid
solib_symbols_extension
Looking at how these are currently set in the various -tdep.c files, I
believe the only one of these that is required to be set for all
architectures is the register_type field.
And so, for all of the other fields, I've changed the property
definition on gdbarch-components.py, setting the 'invalid' field to
False.
Now, after re-generation, the register_type field is checked against
0, thus an architecture that doesn't set_gdbarch_register_type will
now fail during validation. For all the other fields we skip the
validation, in which case, it is find for an architecture to not set
this field.
My expectation is that there should be no user visible changes after
this commit. Certainly for all fields except register_type, all I've
really done is cause some extra comments to be generated, so I think
that's clearly fine.
For the register_type field, my claim is that any architecture that
didn't provide this would fail when creating its register cache, and I
couldn't spot an architecture that doesn't provide this hook. As
such, I think this change should be fine too.
|
|
I don't believe that the gdbarch_register_type_p predicate is called
anywhere in GDB, and the gdbarch_register_type function is called
without checking the gdbarch_register_type_p predicate function
everywhere it is used, for example in
init_regcache_descr (regcache.c).
My claim is that the gdbarch_register_type function is required for
every architecture, and GDB will not work if this function is not
supplied.
And so, in this commit, I remove the 'predicate=True' from
gdbarch-components.py for the 'register_type' field, and regenerate
the gdbarch files.
There should be no user visible changes after this commit.
|
|
Commit b60cea7 (Make target_wait options use enum flags) broke
deprecated_target_wait_hook usage: there's a commit comment telling
this hook has not been converted.
Rather than trying to mend it, this patch replaces the hook by two
target_wait observers:
target_pre_wait (ptid_t ptid)
target_post_wait (ptid_t event_ptid)
Upon target_wait entry, target_pre_wait is notified with the ptid
passed to target_wait. Upon exit, target_post_wait is notified with
the event ptid returned by target_wait. Should an exception occur,
event_ptid is null_ptid.
This change benefits to Insight (out-of-tree): there's no real use of the
late hook in gdb itself.
|
|
I noticed that generic_value_print assumes that a subrange type is
always a subrange of an integer type. However, this isn't necessarily
the case. In Ada, for example, one has subranges of character and
enumeration types.
This code isn't often exercised, I think, because languages with real
subrange types tend to implement their own printers. However, it
still seemed worth fixing.
|
|
When running gdb.cp/non-trivial-retval.exp, the following shows up for
both aarch64-linux and armhf-linux:
Breakpoint 3, f1 (i1=23, i2=100) at src/gdb/testsuite/gdb.cp/non-trivial-retval.cc:35
35 A a;
(gdb) finish
Run till exit from #0 f1 (i1=23, i2=100) at src/gdb/testsuite/gdb.cp/non-trivial-retval.cc:35
main () at /src/gdb/testsuite/gdb.cp/non-trivial-retval.cc:163
163 B b = f2 (i1, i2);
Value returned is $6 = {a = -11952}
(gdb)
The return value should be {a = 123} instead. This happens because the
backends don't extract the return value from the correct location. GDB should
fetch a pointer to the memory location from X8 for aarch64 and r0 for armhf.
With the patch, gdb.cp/non-trivial-retval.exp has full passes on
aarch64-linux and armhf-linux on Ubuntu 20.04/18.04.
The problem only shows up with the "finish" command. The "call" command
works correctly and displays the correct return value.
This is also related to PR gdb/28681
(https://sourceware.org/bugzilla/show_bug.cgi?id=28681) and fixes FAIL's in
gdb.ada/mi_var_array.exp.
A new testcase is provided, and it exercises GDB's ability to "finish" a
function that returns a large struct (> 16 bytes) and display the
contents of this struct correctly. This has always been incorrect for
these backends, but no testcase exercised this particular scenario.
|
|
With nightly rustc, gdb.rust/unsized.exp fails:
(gdb) ptype *us
Structure has no component named operator*.
rustc changed to emit a bit more debug info for unsized types.
Because the original test is just to make sure that ptype of an
unsized array looks right, this patch relaxes the regexp and changes
the expression. I think this keeps the original test meaning, but
also works with nightly. I also tested stable and 1.48.
|
|
An internal test case creates a core file using gcore, then restarts
gdb with that core. When run with a cross-linux gdb (in this case,
x86-64 host with ppc64-linux target), the test fails:
| (gdb) core core
| [New LWP 18437]
| warning: `/lib64/libc.so.6': Shared library architecture unknown is not compatible with target architecture powerpc:common64.
| warning: Could not load shared library symbols for /lib64/ld64.so.1.
| Do you need "set solib-search-path" or "set sysroot"?
| ../../src/gdb/gdbarch.c:3388: internal-error: int gdbarch_elf_make_msymbol_special_p(gdbarch*): Assertion `gdbarch != NULL' failed.
| A problem internal to GDB has been detected,
| further debugging may prove unreliable.
| Quit this debugging session? (y or n) y
What's happening here is that the core file lists some shared
libraries. These aren't available via the solib search path, and so
gdb finds the local (x86-64) libraries. This is not ideal, but on the
other hand, it is what was asked for -- while the test does set
solib-search-path, it does not set the sysroot.
But, because gdb isn't configured to handle these libraries, it
crashes.
It seems to me that it's better to avoid the crash by having
solib_bfd_open fail in the case where a library is incompatible. That
is what this patch does. Now it looks like:
| [New LWP 15488]
| Error while mapping shared library sections:
| `/lib64/libc.so.6': Shared library architecture unknown is not compatible with target architecture powerpc:common64.
... and does not crash gdb.
I don't have a good setup for testing this using dejagnu, so I don't
know whether an existing gdb test covers this scenario.
|
|
Remove the duplicate test names from gdb.base/stap-probe.exp, this is
done by actually passing a unique test name in a couple of
places (rather than using the command as the test name), and in
another couple of places, a test has a duplicate name due to a cut &
paste error, which I've fixed.
There's no change in what is actually being tested after this commit.
|
|
Remove "auto-load:" from a format string passed to auto_load_debug_printf.
It is unnecessary since this function will prefix the string with "[auto-load]"
when printing it.
|
|
Currently, "print/x" will display a floating-point value by first
casting it to an integer type. This yields weird results like:
(gdb) print/x 1.5
$1 = 0x1
This has confused users multiple times -- see PR gdb/16242, where
there are several dups. I've also seen some confusion from this
internally at AdaCore.
The manual says:
'x'
Regard the bits of the value as an integer, and print the integer
in hexadecimal.
... which seems more useful. So, perhaps what happened is that this
was incorrectly implemented (or maybe correctly implemented and then
regressed, as there don't seem to be any tests).
This patch fixes the bug.
There was a previous discussion where we agreed to preserve the old
behavior:
https://sourceware.org/legacy-ml/gdb-patches/2017-06/msg00314.html
However, I think it makes more sense to follow the manual.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=16242
|
|
I noticed that 'progress' is a method on ui-out, but it seems to me
that it would be better if the only API were via the progress_meter
class. This patch makes this change, changing progress to be a method
on the meter itself.
|
|
Fixes a minor typo, in a comment, in the gdbarch-components.py script.
There should be no user visible changes after this commit.
|
|
This adds a multi-threaded testcase that has all threads in the
process exit with a different exit code, and ensures that GDB reports
the thread group leader's exit status as the whole-process exit
status. Before this set of patches, this would randomly report the
exit code of some other thread, and thus fail.
Tested on Linux-x86_64, native and gdbserver.
Co-Authored-By: Pedro Alves <pedro@palves.net>
Change-Id: I30cba2ff4576fb01b5169cc72667f3268d919557
|
|
The current zombie leader detection code in linux-nat.c has a race --
if a multi-threaded inferior exits just before check_zombie_leaders
finds that the leader is now zombie via checking /proc/PID/status,
check_zombie_leaders deletes the leader, assuming we won't get an
event for that exit (which we won't in some scenarios, but not in this
one). That might seem mostly harmless, but it has some downsides:
- later when we continue pulling events out of the kernel, we will
collect the exit event of the non-leader threads, and once we see
the last lwp in our list exit, we return _that_ lwp's exit code as
whole-process exit code to infrun, instead of the leader's exit
code.
- this can cause a hang in stop_all_threads in infrun.c. Say there
are 2 threads in the process. stop_all_threads stops each of those
threads, and then waits for two stop or exit events, one for each
thread. If the whole process exits, and check_zombie_leaders hits
the false-positive case, linux-nat.c will only return one event to
GDB (the whole-process exit returned when we see the last thread,
the non-leader thread, exit), making stop_all_threads hang forever
waiting for a second event that will never come.
However, in this false-positive scenario, where the whole process is
exiting, as opposed to just the leader (with pthread_exit(), for
example), we _will_ get an exit event shortly for the leader, after we
collect the exit event of all the other non-leader threads. Or put
another way, we _always_ get an event for the leader after we see it
become zombie.
I tried a number of approaches to fix this:
#1 - My first thought to address the race was to make GDB always
report the whole-process exit status for the leader thread, not for
whatever is the last lwp in the list. We _always_ get a final exit
(or exec) event for the leader, and when the race triggers, we're not
collecting it.
#2 - My second thought was to try to plug the race in the first place.
I thought of making GDB call waitpid/WNOHANG for all non-leader
threads immediately when the zombie leader is detected, assuming there
would be an exit event pending for each of them waiting to be
collected. Turns out that that doesn't work -- you can see the leader
become zombie _before_ the kernel kills all other threads. Waitpid in
that small time window returns 0, indicating no-event. Thankfully we
hit that race window all the time, which avoided trading one race for
another. Looking at the non-leader thread's status in /proc doesn't
help either, the threads are still in running state for a bit, for the
same reason.
#3 - My next attempt, which seemed promising, was to synchronously
stop and wait for the stop for each of the non-leader threads. For
the scenario in question, this will collect all the exit statuses of
the non-leader threads. Then, if we are left with only the zombie
leader in the lwp list, it means we either have a normal while-process
exit or an exec, in which case we should not delete the leader. If
_only_ the leader exited, like in gdb.threads/leader-exit.exp, then
after pausing threads, we will still have at least one live non-leader
thread in the list, and so we delete the leader lwp. I got this
working and polished, and it was only after staring at the kernel code
to convince myself that this would really work (and it would, for the
scenario I considered), that I realized I had failed to account for
one scenario -- if any non-leader thread is _already_ stopped when
some thread triggers a group exit, like e.g., if you have some threads
stopped and then resume just one thread with scheduler-locking or
non-stop, and that thread exits the process. I also played with
PTRACE_EVENT_EXIT, see if it would help in any way to plug the race,
and I couldn't find a way that it would result in any practical
difference compared to looking at /proc/PID/status, with respect to
having a race.
So I concluded that there's no way to plug the race, we just have to
deal with it. Which means, going back to approach #1. That is the
approach taken by this patch.
Change-Id: I6309fd4727da8c67951f9cea557724b77e8ee979
|
|
Reorganize linux-nat.c:linux_nat_filter_event such that all the
handling for events for LWPs not in the LWP list is together. This
helps make a following patch clearer. The comments and debug messages
have also been tweaked - the end goal is to have them synchronized
with the gdbserver counterpart.
Change-Id: I8586d8dcd76d8bd3795145e3056fc660e3b2cd22
|
|
If we make GDB report the process EXIT event for the leader thread, as
will be done in a latter patch of this series, then
gdb.threads/current-lwp-dead.exp starts failing:
(gdb) break fn_return
Breakpoint 2 at 0x5555555551b5: file /home/pedro/rocm/gdb/build/gdb/testsuite/../../../src/gdb/testsuite/gdb.threads/current-lwp-dead.c, line 45.
(gdb) continue
Continuing.
[New LWP 2138466]
[Inferior 1 (process 2138459) exited normally]
(gdb) FAIL: gdb.threads/current-lwp-dead.exp: continue to breakpoint: fn_return (the program exited)
The inferior exit reported is actually correct. The main thread has
indeed exited, and that's the thread that has the right exit code to
report to the user, as that's the exit code that is reported to the
program's parent. In this case, GDB managed to collect the exit code
for the leader thread before reaping the other thread, because in
reality, the testcase isn't creating standard threads, it is using raw
clone, and the new clones are put in their own thread group.
Fix it by making the main "thread" not exit until the scenario we're
exercising plays out. Also, run the program to completion for
completeness.
The original program really wanted the leader thread to exit before
the fn_return function was reached -- it was important that the
current thread as pointed by inferior_ptid was gone when infrun got
the breakpoint event. I've tweaked the testcase to ensure that that
condition is still held, though it is no longer the main thread that
exits. This required a bit of synchronization between the threads,
which required using CLONE_VM unconditionally. The #ifdef guards were
added as a fix for
https://sourceware.org/bugzilla/show_bug.cgi?id=11214, though I don't
think they were necessary because the program is not using TLS. If it
turns out they were necessary, we can link the testcase with "-z now"
instead, which was mentioned as an alternative workaround in that
Bugzilla.
Change-Id: I7be2f0da4c2fe8f80a60bdde5e6c623d8bd5a0aa
|
|
If we make GDB report the process EXIT event for the leader thread,
instead of whatever is the last thread in the LWP list, as will be
done in a latter patch of this series, then
gdb.threads/current-lwp-dead.exp starts failing:
(gdb) FAIL: gdb.threads/clone-new-thread-event.exp: catch SIGUSR1 (the program exited)
This is a testcase race -- the main thread does not wait for the
spawned clone "thread" to finish before exiting, so the main program
may exit before the second thread is scheduled and reports its
SIGUSR1. With the change to make GDB report the EXIT for the leader,
the race is 100% reproducible by adding a sleep(), like so:
--- c/gdb/testsuite/gdb.threads/clone-new-thread-event.c
+++ w/gdb/testsuite/gdb.threads/clone-new-thread-event.c
@@ -51,6 +51,7 @@ local_gettid (void)
static int
fn (void *unused)
{
+ sleep (1);
tkill (local_gettid (), SIGUSR1);
return 0;
}
Resulting in:
Breakpoint 1, main (argc=1, argv=0x7fffffffd418) at gdb.threads/clone-new-thread-event.c:65
65 stack = malloc (STACK_SIZE);
(gdb) continue
Continuing.
[New LWP 3715562]
[Inferior 1 (process 3715555) exited normally]
(gdb) FAIL: gdb.threads/clone-new-thread-event.exp: catch SIGUSR1 (the program exited)
That inferior exit reported is actually correct. The main thread has
indeed exited, and that's the thread that has the right exit code to
report to the user, as that's the exit code that is reported to the
program's parent. In this case, GDB managed to collect the exit code
for the leader thread before reaping the other thread, because in
reality, the testcase isn't creating standard threads, it is using raw
clone, and the new clones are put in their own thread group.
Fix it by making the main thread wait for the child to exit. Also,
run the program to completion for completeness.
Change-Id: I315cd3dc2b9e860395dcab9658341ea868d7a6bf
|
|
Fix a typo, s/dislayed/displayed/ in default.exp at the top level.
|
