Tom Tromey [Thu, 24 Jan 2019 05:16:53 +0000 (22:16 -0700)]
C++ify remote notification code
This C++ifies the remote notification code -- replacing function
pointers with virtual methods and using unique_ptr. This allows for
the removal of some cleanups.
2019-03-06 Tom Tromey <tom@tromey.com>
* remote.c (struct stop_reply_deleter): Remove.
(stop_reply_up): Update.
(struct stop_reply): Derive from notif_event. Don't typedef.
<regcache>: Now a std::vector.
(stop_reply_xfree): Remove.
(stop_reply::~stop_reply): Rename from stop_reply_dtr.
(remote_notif_stop_alloc_reply): Return a unique_ptr. Use new.
(remote_target::discard_pending_stop_replies): Use delete.
(remote_target::remote_parse_stop_reply): Update.
(remote_target::process_stop_reply): Update.
* remote-notif.h (struct notif_event): Add virtual destructor.
Remove "dtr" member.
(struct notif_client) <alloc_event>: Return a unique_ptr.
(notif_event_xfree): Don't declare.
(notif_event_up): New typedef.
* remote-notif.c (remote_notif_ack, remote_notif_parse): Update.
(notif_event_xfree, do_notif_event_xfree): Remove.
(remote_notif_state_xfree): Update.
Pedro Alves [Wed, 6 Mar 2019 18:29:17 +0000 (18:29 +0000)]
Make "checkpoint" not rely on inferior_ptid
Don't rely on "inferior_ptid" deep within add_fork. In the
multi-target branch, I'm forcing inferior_ptid to null_ptid early in
infrun event handling to make sure we don't inadvertently rely on the
current thread/target when we shouldn't, and that caught some bad or
unnecessary assumptions throughout.
gdb/ChangeLog:
2019-03-06 Pedro Alves <palves@redhat.com>
* linux-fork.c (new_fork): New, split out of ...
(add_fork): ... this. Return void. Move "first fork" special
case from here, to ...
(checkpoint_command): ... here.
* linux-linux.h (add_fork): Return void.
Andrew Burgess [Mon, 21 Jan 2019 15:10:55 +0000 (15:10 +0000)]
gdb/fortran: Handle older TYPE*SIZE typenames
This patch adds support for the older TYPE*SIZE typenames that are
still around in older code.
For implementation this currently reuses the kind mechanism, as under
gFortran the kind number is equivalent to the size, however, this is
not necessarily true for all compilers. If the rules for other
compilers are better understood then this code might need to be
improved slightly to allow for a distinction between size and kind,
however, adding this extra complexity now seems pointless.
gdb/ChangeLog:
* f-exp.y (direct_abs_decl): Handle TYPE*SIZE type names.
gdb/testsuite/ChangeLog:
* gdb.fortran/type-kinds.exp: Extend to cover TYPE*SIZE cases.
Andrew Burgess [Fri, 18 Jan 2019 14:44:48 +0000 (14:44 +0000)]
gdb/fortran: Add support for the ABS intrinsic function
Adds support for the abs intrinsic function, this requires adding a
new pattern to the Fortran parser. Currently only float and integer
argument types are supported to ABS, complex is still not supported,
this can be added later if needed.
gdb/ChangeLog:
* f-exp.y: New token, UNOP_INTRINSIC.
(exp): New pattern using UNOP_INTRINSIC token.
(f77_keywords): Add 'abs' keyword.
* f-lang.c: Add 'target-float.h' and 'math.h' includes.
(value_from_host_double): New function.
(evaluate_subexp_f): Support UNOP_ABS.
gdb/testsuite/ChangeLog:
* gdb.fortran/intrinsics.exp: Extend to cover ABS.
Andrew Burgess [Fri, 18 Jan 2019 11:24:24 +0000 (11:24 +0000)]
gdb/fortran: Use TYPE_CODE_CHAR for character types
Switch to using TYPE_CODE_CHAR for character types. This appears to
have little impact on the test results as gFortran uses the
DW_TAG_string_type to represent all character variables (as far as I
can see). The only place this has an impact is when the user casts a
variable to a character type, in which case GDB does now use the CHAR
type, and prints the variable as both a value and a character, for
example, before:
(gdb) p ((character) 97)
$1 = 97
and after:
(gdb) p ((character) 97)
$1 = 97 'a'
gdb/ChangeLog:
* f-lang.c (build_fortran_types): Use TYPE_CODE_CHAR for character
types.
Andrew Burgess [Wed, 16 Jan 2019 16:42:10 +0000 (16:42 +0000)]
gdb/fortran: Add Fortran 'kind' intrinsic and keyword
The 'kind' keyword has two uses in Fortran, it is the name of a
builtin intrinsic function, and it is also a keyword used to create a
type of a specific kind.
This commit adds support for using kind as an intrinsic function, and
also adds some initial support for using kind to create types of a
specific kind.
This commit only allows the creation of the type 'character(kind=1)',
however, it will be easy enough to extend this in future to support
more type kinds.
The kind of any expression can be queried using the kind intrinsic
function. At the moment the kind returned corresponds to the size of
the type, this matches how gfortran handles kinds. However, the
correspondence between kind and type size depends on the compiler
and/or the specific target, so this might not be correct for
everyone. If we want to support different compilers/targets in future
the code to compute the kind from a type will need to be updated.
gdb/ChangeLog:
* expprint.c (dump_subexp_body_standard): Support UNOP_KIND.
* f-exp.y: Define 'KIND' token.
(exp): New pattern for KIND expressions.
(ptype): Handle types with a kind extension.
(direct_abs_decl): Extend to spot kind extensions.
(f77_keywords): Add 'kind' to the list.
(push_kind_type): New function.
(convert_to_kind_type): New function.
* f-lang.c (evaluate_subexp_f): Support UNOP_KIND.
* parse.c (operator_length_standard): Likewise.
* parser-defs.h (enum type_pieces): Add tp_kind.
* std-operator.def: Add UNOP_KIND.
gdb/testsuite/ChangeLog:
* gdb.fortran/intrinsics.exp: New file.
* gdb.fortran/intrinsics.f90: New file.
* gdb.fortran/type-kinds.exp: New file.
Andrew Burgess [Wed, 16 Jan 2019 16:16:59 +0000 (16:16 +0000)]
gdb/fortran: Add new function to evaluate Fortran expressions
This is an initial restructure, it adds a new function in which
Fortran specific expressions can be evaluated. No Fortran specific
expressions are added with this commit though, so for now, the new
function just forwards all expressions to the default expression
handler, as such, there should be no user visible changes after this
commit. However, the new function will be useful in later commits.
gdb/ChangeLog:
* f-lang.c (evaluate_subexp_f): New function.
(exp_descriptor_f): New global.
(f_language_defn): Use exp_descriptor_f instead of
exp_descriptor_standard.
Andrew Burgess [Wed, 16 Jan 2019 15:30:54 +0000 (15:30 +0000)]
gdb/fortran: Simplify handling of Fortran dot operations and keywords
Use strncasecmp to compare Fortran dot operations (like .AND.) and for
the keywords list. This allows for some duplication to be removed
from the token arrays. I've also performed whitespace cleanup around
the code I've changed.
I have added some tests to ensure that upper and lowercase dot
operations are correctly tested. The keywords list remains always
lowercase for now.
There should be no user visible changes after this commit.
gdb/ChangeLog:
* f-exp.y (struct token): Add comments.
(dot_ops): Remove uppercase versions and the end marker.
(f77_keywords): Likewise.
(yylex): Use ARRAY_SIZE to iterate over dot_ops, assert all
entries in the dot_ops array are case insensitive, and use
strncasecmp to compare strings. Also some whitespace cleanup in
this area. Similar for the f77_keywords array, except entries in
this list might be case sensitive.
Andrew Burgess [Wed, 16 Jan 2019 13:36:46 +0000 (13:36 +0000)]
gdb/fortran: Cleanup code for parsing logical constants
This patch cleans up the code used for parsing the Fortran logical
constants '.TRUE.' and '.FALSE.'. Instead of listing both upper and
lowercase versions of these strings we now use strncasecmp.
I've also switched to use ARRAY_SIZE for the array iteration, and I've
cleaned up whitespace in the vicinity of the code I've changed.
Finally, I've added a test to ensure that both the upper and lower
case versions of the logical constants are understood by GDB,
something that was missing previously.
There should be no user visible changes after this commit.
gdb/ChangeLog:
* f-exp.y (struct f77_boolean_val): Add comments.
(boolean_values): Remove uppercase versions, and end marker.
(yylex): Use ARRAY_SIZE for iterating over boolean_values array,
and use strncasecmp to achieve case insensitivity. Additionally,
perform whitespace cleanup around this code.
gdb/testsuite/ChangeLog:
* gdb.fortran/types.exp (test_logical_literal_types_accepted):
Check upper and lower case logical literals.
Tom Tromey [Wed, 6 Mar 2019 14:55:42 +0000 (07:55 -0700)]
Fix remote-sim.c build after warn-unused-result change
John Darrington pointed out that commit 18cb7c9f3 ("Introduce
ATTRIBUTE_UNUSED_RESULT and use it") broke the build:
/home/john/binutils-gdb/gdb/remote-sim.c: In function 'void gdbsim_target_open(const char*, int)':
/home/john/binutils-gdb/gdb/remote-sim.c:765:18: error: ignoring return value of 'char** gdb_argv::release()', declared with attribute warn_unused_result [-Werror=unused-result]
This patch fixes the problem by arranging to use the result of the
release method.
Tested by rebuilding with a simulator enabled.
gdb/ChangeLog
2019-03-06 Tom Tromey <tromey@adacore.com>
* remote-sim.c (gdbsim_target_open): Use result of
gdb_argv::release.
Alan Hayward [Wed, 6 Mar 2019 09:52:08 +0000 (09:52 +0000)]
Testsuite: Ensure changing directory does not break the log file
get_compiler_info switches to a new log file before checking the compiler
to ensure the checks are not logged. Afterwards it restores back to using
the original log file. However, the logfile uses a relative path name -
if the current test has changed the current directory then all further
output for the test will be lost. This can confuse the code that collates
the main gdb.log file at the end of a FORCE_PARALLEL run.
fullpath-expand.exp calls gdb_compile after changing the current directory.
The "Ensure stack protection is off for GCC" patch added a call to
get_compiler_info from inside of gdb_compile, causing log file collection
to break for FORCE_PARALLEL runs.
The ideal solution would be to ensure the log file is always created using
an absolute path name. However, this is set at multiple points in
Makefile.in and in some instances just relies on dejagnu common code to set
the log file directory to "."
The simpler and safer solution is to override the builtin cd function. The
new function checks the current log file and if the path is relative, then
it resets the logging using an absolute path. Finally it calls the builtin
cd. This ensures get_compiler_info (and any other code) can correctly
backup and restore the current log file.
gdb/testsuite/ChangeLog:
* lib/gdb.exp (builtin_cd): rename of cd.
(cd): Override builtin.
Nick Clifton [Wed, 6 Mar 2019 09:43:15 +0000 (09:43 +0000)]
Fix a segmentation fault triggered by disassembling an EFi file with source included.
* dwarf2.c (_bfd_dwarf2_find_symbol_bias): Check for a NULL symbol
table pointer.
* coffgen.c (coff_find_nearest_line_with_names): Do not call
_bfd_dwarf2_find_symbol_bias if there is no symbol table available.
Richard Bunt [Wed, 6 Mar 2019 08:23:00 +0000 (08:23 +0000)]
Fortran function calls with arguments
Prior to this patch, calling functions on the inferior with arguments and
then using these arguments within a function resulted in an invalid
memory access. This is because Fortran arguments are typically passed as
pointers to values.
It is possible to call Fortran functions, but memory must be allocated in
the inferior, so a pointer can be passed to the function, and the
language must be set to C to enable C-style casting. This is cumbersome
and not a pleasant debug experience.
This patch implements the GNU Fortran argument passing conventions with
caveats. Firstly, it does not handle the VALUE attribute as there is
insufficient DWARF information to determine when this is the case.
Secondly, functions with optional parameters can only be called with all
parameters present. Both these cases are marked as KFAILS in the test.
Since the GNU Fortran argument passing convention has been implemented,
there is no guarantee that this patch will work correctly, in all cases,
with other compilers.
Despite these limitations, this patch improves the ease with which
functions can be called in many cases, without taking away the existing
approach of calling with the language set to C.
Regression tested on x86_64, aarch64 and POWER9 with GCC 7.3.0.
Regression tested with Ada on x86_64.
Regression tested with native-extended-gdbserver target board.
gdb/ChangeLog:
* eval.c (evaluate_subexp_standard): Call Fortran argument
wrapping logic.
* f-lang.c (struct value): A value which can be passed into a
Fortran function call.
(fortran_argument_convert): Wrap Fortran arguments in a pointer
where appropriate.
(struct type): Value ready for a Fortran function call.
(fortran_preserve_arg_pointer): Undo check_typedef, the pointer
is needed.
* f-lang.h (fortran_argument_convert): Declaration.
(fortran_preserve_arg_pointer): Declaration.
* infcall.c (value_arg_coerce): Call Fortran argument logic.
gdb/testsuite/ChangeLog:
* gdb.fortran/function-calls.exp: New file.
* gdb.fortran/function-calls.f90: New test.
Tom Tromey [Tue, 5 Mar 2019 21:58:24 +0000 (14:58 -0700)]
Remove some Python 3 #ifs
A recent patch from Kevin Buettner taught me that the PyBytes API is
available on Python 2. This patch removes a couple of related #ifs in
the Python code.
Tested on x86-64 Fedora 29, using both Python 3.7 and Python 2.7.
gdb/ChangeLog
2019-03-05 Tom Tromey <tromey@adacore.com>
(This is clearly from before the change to make ravenscar
multi-target-capable.)
Here, 0x142b0b0 is the singleton dummy target. It seems to me that
since this is always a singleton, it would be a bit nicer if it were a
global, so that it would be noted in the above.
Tom Tromey [Tue, 19 Feb 2019 20:36:17 +0000 (13:36 -0700)]
Remove excess calls to gdb_flush
A customer noticed some mildly odd MI output, where CLI output was
split into multiple MI strings at unusual boundaries, like this:
~"$1 = (b => true"
~", p => 0x407260"
This is technically correct according to the MI spec, but still
unusual, in that there's no particular reason for the string to be
split where it is.
I tracked this down to a call to gdb_flush in generic_val_print.
Then, I went through all calls to gdb_flush and removed the ones I
thought were superfluous. In particular:
* Any call in the value-printing code;
* Likewise the type-printing code (just a single call); and
* Any call that immediately followed a printf that obviously
ended with a newline, my belief being that gdb's standard output
streams are line buffered (by inheriting the behavior from stdio)
Regression tested on x86-64 Fedora 29.
I didn't add a new test case. I tend to think we don't necessarily
want to specify this behavior in the tests. Let me know what you
think of this.
gdb/ChangeLog
2019-03-05 Tom Tromey <tromey@adacore.com>
Tom Tromey [Tue, 26 Feb 2019 21:59:47 +0000 (14:59 -0700)]
Add ATTRIBUTE_UNUSED_RESULT to ref_ptr::release
This applies ATTRIBUTE_UNUSED_RESULT to ref_ptr::release and updates a
few spots to comply. I believe one use in install_default_visualizer
was in error, fixed by this patch.
gdb/ChangeLog
2019-03-05 Tom Tromey <tromey@adacore.com>
Use '$enable_unittest' instead of '$development' on gdbserver/configure.srv (for 'aarch64*-*-linux*' case)
On commit 8ecfd7bd4acd69213c06fac6de9af38299123547 ("Add parameter to
allow enabling/disabling selftests via configure") it seems that I
forgot to use the proper '$enable_unittest' variable when checking to
see whether to add selftest-related objects to 'srv_regobj'. This
causes a build failure on Aarch64 when 'development=false' (which is
the case for the 8.3 branch) and 'enable_unittest=true'.
This patch fixes the problem by using '$enable_unittest' instead of
'$development' when performing the check. As a reminder, it's
important to notice that '$enable_unittest's default value (i.e., when
the option '--enable-unit-tests' is not passed to configure) is the
same as '$development', so this patch doesn't affect the current
build.
I'd like to install this patch both on master and on the 8.3 branch.
Richard Bunt [Mon, 4 Mar 2019 15:08:51 +0000 (15:08 +0000)]
gdbserver short-circuit-argument-list failures
This patch fixes test case failures observed when running
short-circuit-argument-list.exp with gdb server boards. Thanks to Sergio
Durigan Junior for pointing this out.
Assertions failed with the native{,-extended}-gdbserver boards as the
standard output from the test program appears in a different location
than observed on non-gdbserver boards. This standard output was used to
determine whether a function, which had been logically short-circuited,
was called or not. Since the location of the standard out cannot be
relied upon to verify this, a new mechanism was needed.
The test program now records function calls in variables named the same
as the function with a "_called" suffix. These variables can then be
queried from the test case to verify the occurrence of a call.
A method to reset the call counts has been included in the test case, so
that any future assertions added to this test can ensure a fresh set of
initial values before proceeding. Not resetting values between groups of
assertions creates a dependency between them, which increases the
likelihood that a single failure causes subsequent assertions to fail.
Regression tested on x86_64, aarch64 and ppc64le.
Regression tested with Ada on x86_64.
Regression tested with the native{,-extended}-gdbserver boards on x86_64.
Eli Zaretskii [Sat, 2 Mar 2019 13:13:54 +0000 (15:13 +0200)]
Fix GDB compilation on MinGW (PR gdb/24292)
gdb/ChangeLog:
2019-03-02 Eli Zaretskii <eliz@gnu.org>
PR gdb/24292
* common/netstuff.c:
* gdbserver/gdbreplay.c
* gdbserver/remote-utils.c:
* ser-tcp.c:
* unittests/parse-connection-spec-selftests.c [USE_WIN32API]:
Include ws2tcpip.h instead of wsiapi.h and winsock2.h. Redefine
_WIN32_WINNT to 0x0501 if defined to a smaller value, as
'getaddrinfo' and 'freeaddrinfo' were not available before
Windows XP, and mingw.org's MinGW headers by default define
_WIN32_WINNT to 0x500.
Gary Benson [Fri, 1 Mar 2019 13:50:27 +0000 (13:50 +0000)]
Fix coff_start_symtab resource leak found by Coverity
This commit fixes a resource leak found by Coverity, where
coff_start_symtab performs an xstrdup that is now performed
within start_symtab by buildsym_compunit::buildsym_compunit.
Tamar Christina [Fri, 1 Mar 2019 11:37:51 +0000 (11:37 +0000)]
Binutils: Always skip only 1 byte for CIE version 1's return address register.
According to the specification for the CIE entries, when the CIE version is 1 then
the return address register field is always 1 byte. Readelf does this correctly in
read_cie in dwarf.c but ld does this incorrectly and always tries to read a
skip_leb128. If the value here has the top bit set then ld will incorrectly read
at least another byte, causing either an assert failure or an incorrect address to
be used in eh_frame.
I'm not sure how to generate a generic test for this as I'd need to write assembly,
and it's a bit hard to trigger. Essentially the relocated value needs to start with
something that & 0x70 != 0x10 while trying to write a personality.
Alan Modra [Thu, 28 Feb 2019 22:58:47 +0000 (09:28 +1030)]
PR24272, out-of-bounds read in pex64_xdata_print_uwd_codes
The fix here is to use an unsigned comparison for
if (a->NumberOfRvaAndSizes > IMAGE_NUMBEROF_DIRECTORY_ENTRIES)
include/
PR 24272
* coff/internal.h (struct internal_extra_pe_aouthdr): Change type
of SizeOfCode, SizeOfInitializedData, and SizeOfUninitializedData
to bfd_vma. Change type of SectionAlignment, FileAlignment,
Reserved1, SizeOfImage, SizeOfHeaders, CheckSum, LoaderFlags,
and NumberOfRvaAndSizes to uint32_t.
bfd/
PR 24272
* peXXigen.c (_bfd_XXi_swap_aouthdr_in): Use unsigned index.
(_bfd_XX_print_private_bfd_data_common): Adjust for type changes.
Alan Hayward [Thu, 28 Feb 2019 17:19:46 +0000 (17:19 +0000)]
Testsuite: Catch gdbserver socket listen errors
When launching gdbserver, the testsuite checks for binding failure but
does not check for failure to listen to socket error (which can happen
due to another gdbserver binding to the socket at the same time).
When this error occurs, the test will ignore the error and connect GDB
to the failed port. This may succeed and GDB will now be connected to
the gdbserver from another test. This eventually causes both tests to
fail.
When running the tests suite with native-gdbserver across many cores,
this issue may happen once or twice, each causing random failures for
two .exp testscripts.
Example gdb.log output for the failure:
The testsuite sucessfully notices a failure to connect to port 2348.
It launches again with port 2349, which also fails. The testsuite
ignores this error and uses gdb to connect to the port - which succeeds.
spawn /work/build/gdb/testsuite/../gdbserver/gdbserver --once localhost:2348 /work/build/gdb/testsuite/outputs/gdb.ada/arrayidx/p^M
Can't bind address: Address already in use.^M
Exiting^M
Port 2348 is already in use.
spawn /work/build/gdb/testsuite/../gdbserver/gdbserver --once localhost:2349 /work/build/gdb/testsuite/outputs/gdb.ada/arrayidx/p^M
Can't listen on socket: Address already in use.^M
Exiting^M
target remote localhost:2349^M
Remote debugging using localhost:2349^M
Reading /lib/ld-linux-aarch64.so.1 from remote target...^M
warning: File transfers from remote targets can be slow. Use "set sysroot" to access files locally instead.^M
Reading /lib/ld-linux-aarch64.so.1 from remote target...^M
Reading symbols from target:/lib/ld-linux-aarch64.so.1...^M
Reading /lib/ld-2.23.so from remote target...^M
Reading /lib/.debug/ld-2.23.so from remote target...^M
Reading /work/build/install/lib/debug//lib/ld-2.23.so from remote target...^M
Reading /work/build/install/lib/debug/lib//ld-2.23.so from remote target...^M
Reading target:/work/build/install/lib/debug/lib//ld-2.23.so from remote target...^M
(No debugging symbols found in target:/lib/ld-linux-aarch64.so.1)^M
0x0000ffffbf6d2cc0 in ?? () from target:/lib/ld-linux-aarch64.so.1^M
(gdb) continue^M
Continuing.^M
Reading /lib/aarch64-linux-gnu/libc.so.6 from remote target...^M
Reading /lib/aarch64-linux-gnu/libc-2.23.so from remote target...^M
Reading /lib/aarch64-linux-gnu/.debug/libc-2.23.so from remote target...^M
Reading /work/build/install/lib/debug//lib/aarch64-linux-gnu/libc-2.23.so from remote target...^M
Reading /work/build/install/lib/debug/lib/aarch64-linux-gnu//libc-2.23.so from remote target...^M
Reading target:/work/build/install/lib/debug/lib/aarch64-linux-gnu//libc-2.23.so from remote target...^M
[Inferior 1 (process 35351) exited normally]^M
(gdb) FAIL: gdb.ada/arrayidx.exp: can't run to main
Meanwhile, at the same time, in another test, gdbserver successfully
connects to port 2349. GDB then tries to connect to the port, but it
times out because the GDB in the test above has already connected to it.
spawn /work/build/gdb/testsuite/../gdbserver/gdbserver --once localhost:2348 /work/build/gdb/testsuite/outputs/gdb.ada/rdv_wait/foo^M
Can't bind address: Address already in use.^M
Exiting^M
Port 2348 is already in use.
spawn /work/build/gdb/testsuite/../gdbserver/gdbserver --once localhost:2349 /work/build/gdb/testsuite/outputs/gdb.ada/rdv_wait/foo^M
Process /work/build/gdb/testsuite/outputs/gdb.ada/rdv_wait/foo created; pid = 65162^M
Listening on port 2349^M
Remote debugging from host 127.0.0.1, port 45154^M
target remote localhost:2349^M
localhost:2349: Connection timed out.^M
(gdb) ^CQuit^M
(gdb) task 2^M
Cannot inspect Ada tasks when program is not running^M
gdb/testsuite/ChangeLog:
* lib/gdbserver-support.exp (gdbserver_start): Check for listen
failure.
Rainer Orth [Thu, 28 Feb 2019 15:09:05 +0000 (16:09 +0100)]
Can't interrupt process without controlling terminal on Solaris (PR gdb/8527)
If gdb attaches to a process that either has no controlling terminal,
or the controlling terminal differs from the one gdb is running under,
break/^C doesn't interrupt the debugged process on Solaris.
Fixed as follows, analogous to what all all other targets do. Patch from
the PR, recently re-submitted by Brian Vandenberg.
Tested on amd64-pc-solaris2.11, sparcv9-sun-solaris2.11, and
x86_64-pc-linux-gnu.
2019-02-28 Brian Vandenberg <phantall@gmail.com>
Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
gdb:
PR gdb/8527
* procfs.c (proc_wait_for_stop): Wrap write of PCWSTOP in
set_sigint_trap, clear_sigint_trap.
gdb/testsuite:
PR gdb/8527
* gdb.base/interrupt-daemon-attach.c,
gdb.base/interrupt-daemon-attach.exp: New test.
Fix regcache leak, and avoid possible regcache access after detach.
Valgrind reports leaks like the below in various tests,
e.g. gdb.threads/attach-slow-waitpid.exp, gdb.ada/task_switch_in_core.exp, ...
Fix the leak by clearing the regcache when detaching from an inferior.
Note that these leaks are 'created' when GDB exits,
when the regcache::current_regcache is destroyed : the elements
of the forward_list are pointers, and the 'pointed to' memory is not
deleted by the forward_list destructor.
Nevertheless, fixing this leak is good as it makes a bunch of
tests 'leak clean'.
Also, it seems strange to keep a register cache for a process from
which GDB detached : it is not clear if this cache is still valid
after detach. And effectively, when clearing only the regcache,
(and not the frame cache), then the frame cache was still 'pointing'
at this regcache and was used when switching to the child process
in the test gdb.threads/watchpoint-fork.exp, which seems strange.
So, we solve the leak and avoid possible accesses to the regcache
and frame cache of the detached inferior, by clearing both the
regcache and the frame cache.
Tested on debian/amd64, natively, under Valgrind,
and with make check RUNTESTFLAGS="--target_board=native-gdbserver".
==27679== VALGRIND_GDB_ERROR_BEGIN
==27679== 1,123 (72 direct, 1,051 indirect) bytes in 1 blocks are definitely lost in loss record 2,942 of 3,400
==27679== at 0x4C2C4CC: operator new(unsigned long) (vg_replace_malloc.c:344)
==27679== by 0x5CDF71: get_thread_arch_aspace_regcache(ptid_t, gdbarch*, address_space*) (regcache.c:330)
==27679== by 0x5CE12A: get_thread_regcache (regcache.c:366)
==27679== by 0x5CE12A: get_current_regcache() (regcache.c:372)
==27679== by 0x4FF63D: post_create_inferior(target_ops*, int) (infcmd.c:452)
==27679== by 0x43AF62: core_target_open(char const*, int) (corelow.c:458)
==27679== by 0x408B68: cmd_func(cmd_list_element*, char const*, int) (cli-decode.c:1892)
...
gdb/ChangeLog
2019-02-27 Philippe Waroquiers <philippe.waroquiers@skynet.be>
* target.c (target_detach): Clear the regcache and the
frame cache.
Saagar Jha [Wed, 27 Feb 2019 18:48:35 +0000 (18:48 +0000)]
Prevent overflow in rl_set_screen_size
GDB calls rl_set_screen_size in readline with the current screen size,
measured in rows and columns. To represent "infinite" sizes, GDB
passes in INT_MAX; however, since rl_set_screen_size internally
multiplies the number of rows and columns, this causes a signed
integer overflow. To prevent this we can instead pass in the
approximate square root of INT_MAX (which is still reasonably large),
so that even when the number of rows and columns is "infinite" we
don't overflow.
gdb/ChangeLog:
2019-02-27 Saagar Jha <saagar@saagarjha.com>
Pedro Alves <palves@redhat.com>
* utils.c (set_screen_size): Reduce "infinite" rows and columns
before calling rl_set_screen_size.
Kevin Buettner [Tue, 19 Feb 2019 20:18:17 +0000 (13:18 -0700)]
Use Python 2.[67] / 3.X / PEP 3118 buffer protocol
This patch removes the non-IS_PY3K code in infpy_write_memory()
and infpy_search_memory(). In both cases, the remaining code
from these ifdefs is related to use of the PEP 3118 buffer protocol.
(Deleted code is either due to simplification or related to use of the
old buffer protocol.) PEP 3118 is sometimes referred to as the "new"
buffer protocol, though it's not that new anymore.
The link below describes new features in Python 2.6. In particular,
it says that the buffer protocol described by PEP 3118 is in Python
2.6. It also says (at the top of the page) that Python 2.6 was
released on Oct 1, 2008.
The last security release for the Python 2.6 series was 2.6.9. It was
released on Oct 29, 2013. According to this document...
https://www.python.org/download/releases/2.6.9/
...support for the 2.6 series has ended:
With the 2.6.9 release, and five years after its first release,
the Python 2.6 series is now officially retired. All official
maintenance for Python 2.6, including security patches, has ended.
For ongoing maintenance releases, please see the Python 2.7
series.
As noted earlier, Python 2.6, Python 2.7, and Python 3.X all have
support for the PEP 3118 buffer protocol. Python releases prior
to 2.6 use an older buffer protocol. Since Python 2.6 has been
retired for a good while now, it seems reasonable to me to remove
code using the older buffer protocol from GDB.
I have also simplified some of the code via use of the Py_buffer
unique_ptr specialization which I introduced in the two argument
gdb.Value constructor patch series. Therefore, there is a dependency
on patch #1 from that series.
I have tested against both Python 2.7.15 and 3.7.2. I see no
regressions among the non-racy tests. I've also verified that
PyBuffer_Release is being called when the affected functions exit
while running the tests in gdb.python/py-inferior.exp by hand. I've
also tried running valgrind on GDB while running this test, but I'm
puzzled by the results that I'm seeing - I'm seeing no additional
leaks when I comment out the Py_buffer_up lines that I introduced.
That said, I'm not seeing any leaks that obviously originate from
either infpy_write_memory() or infpy_search_memory().
gdb/ChangeLog:
* python/py-inferior.c (infpy_write_memory): Remove non-IS_PY3K
code from these functions. Remove corresponding ifdefs. Use
Py_buffer_up instead of explicit calls to PyBuffer_Release.
Remove gotos and target of gotos.
(infpy_search_memory): Likewise.
Andrew Burgess [Sun, 9 Sep 2018 16:07:16 +0000 (17:07 +0100)]
gdb/arc: Use default gdbarch methods where possible
Make use of the default gdbarch methods for gdbarch_dummy_id,
gdbarch_unwind_pc, and gdbarch_unwind_sp where possible.
I have not tested this change but, by inspecting the code, I believe
the default methods are equivalent to the code being deleted. The
only difference is that the old ARC specific methods had additional
debugging, however, this was discussed on the mailing list[1] and it
was agreed these methods could be removed.
Testsuite: Allow multiple lines of "as" in testsuite.
Committed on behalf of Matthew Malcomson.
This allows checking the command line parsing more easily than before by
allowing many command line invokations from the same .d file.
Each line is used as a set of flags, and the tests are ran against the output
of the assembler with each set.
Each line of assembler is treated as another set of tests (as if the test file
were copied to another with a different #as: line).
This patch includes some example uses where multiple testcases can be merged
into one file using this new functionality.
Andrew Burgess [Fri, 22 Feb 2019 22:04:53 +0000 (22:04 +0000)]
gdb: Handle alignment for C++ structures with static members
In 'type_align' when computing the alignment of a structure we should
not consider the alignment of static structure members, these are
usually stored outside of the structure and therefore don't have any
impact on the structures alignment requirements.
I've extended the existing alignment calculating test to compile in
both C and C++ now so that we can create structures with static
members.
gdb/ChangeLog:
* gdbtypes.c (type_align): Don't consider static members when
computing structure alignment.
gdb/testsuite/ChangeLog:
* gdb.base/align.exp: Extend to compile in both C and C++, and add
tests for structs with static members.
Andrew Burgess [Fri, 22 Feb 2019 20:49:04 +0000 (20:49 +0000)]
gdb: Restructure type_align and gdbarch_type_align
This commit restructures the relationship between the type_align
function and the gdbarch_type_align method.
The problem being addressed with this commit is this; previously the
type_align function was structured so that for "basic" types (int,
float, etc) the gdbarch_type_align hook was called, which for
"compound" types (arrays, structs, etc) the common type_align code has
a fixed method for how to extract a "basic" type and would then call
itself on that "basic" type.
The problem is that if an architecture wants to modify the alignment
rules for a "compound" type then this is not currently possible.
In the revised structure, all types pass through the
gdbarch_type_align method. If this method returns 0 then this
indicates that the architecture has no special rules for this type,
and GDB should apply the default rules for alignment. However, the
architecture is free to provide an alignment for any type, both
"basic" and "compound".
After this commit the default alignment rules now all live in the
type_align function, the default_type_align only ever returns 0,
meaning apply the default rules.
I've updated the 3 targets (arc, i386, and nios2) that already
override the gdbarch_type_align method to fit the new scheme.
Tested on X86-64/GNU Linux with no regressions.
gdb/ChangeLog:
* arc-tdep.c (arc_type_align): Provide alignment for basic types,
return 0 for other types.
* arch-utils.c (default_type_align): Always return 0.
* gdbarch.h: Regenerate.
* gdbarch.sh (type_align): Extend comment.
* gdbtypes.c (type_align): Add additional comments, always call
gdbarch_type_align before applying the default rules.
* i386-tdep.c (i386_type_align): Return 0 as the default rule,
generic code will then apply a suitable default.
* nios2-tdep.c (nios2_type_align): Provide alignment for basic
types, return 0 for other types.
Andrew Burgess [Sat, 23 Feb 2019 20:07:47 +0000 (22:07 +0200)]
gdb/riscv: Use legacy register numbers in default target description
When the target description support was added to RISC-V, the register
numbers assigned to the fflags, frm, and fcsr control registers in the
default target descriptions didn't match the register numbers used by
GDB before the target description support was added.
What this means is that if a tools exists in the wild that is using
hard-coded register number, setup to match GDB's old numbering, then
this will have been broken (for fflags, frm, and fcsr) by the move to
target descriptions. QEMU is such a tool.
There are a couple of solutions that could be used to work around this
issue:
- The user can create their own xml description file with the
register numbers setup to match their old tool, then load this by
telling GDB 'set tdesc filename FILENAME'.
- Update their old tool to use the newer default numbering scheme, or
better yet add proper target description support to their tool.
- We could have RISC-V GDB change to maintain the old defaults.
This patch changes GDB back to using the old numbering scheme.
This change is only visible to remote targets that don't supply their
own xml description file and instead rely on GDB's default numbering.
Note that even though 32bit-cpu.xml and 64bit-cpu.xml have changed,
the corresponding .c file has not, this is because the numbering added
to the registers in the xml files is number 0, this doesn't result in
any new C code being generated .
Tom Tromey [Tue, 26 Feb 2019 19:12:30 +0000 (12:12 -0700)]
Fix new py-value.exp test case
The new test case in py-value.exp fails -- the code was changed to
throw ValueError, but the test still checks for TypeError. This patch
fixes the problem.
I'm checking this in. Tested on x86-64 Fedora 29.
gdb/testsuite/ChangeLog
2019-02-26 Tom Tromey <tromey@adacore.com>
* gdb.python/py-value.exp (test_value_from_buffer): Check for
ValueError, not TypeError.
This two argument form of the gdb.Value constructor may also be used
to obtain gdb values from selected portions of buffers read with
Inferior.read_memory(). The test case (which is in a separate patch)
demonstrates this use case.
gdb/ChangeLog:
* python/py-value.c (convert_buffer_and_type_to_value): New
function.
(valpy_new): Parse arguments via gdb_PyArg_ParseTupleAndKeywords.
Add support for handling an optional second argument. Call
convert_buffer_and_type_to_value as appropriate.
Kevin Buettner [Sun, 17 Feb 2019 00:33:47 +0000 (17:33 -0700)]
Define unique_ptr specialization for Py_buffer.
This patch causes PyBuffer_Release() to be called when the associated
buffer goes out of scope. I've been using it as follows:
...
Py_buffer_up buffer_up;
Py_buffer py_buf;
if (PyObject_CheckBuffer (obj)
&& PyObject_GetBuffer (obj, &py_buf, PyBUF_SIMPLE) == 0)
{
/* Got a buffer, py_buf, out of obj. Cause it to released
when it goes out of scope. */
buffer_up.reset (&py_buf);
}
...
This snippet of code was taken directly from an upcoming patch to
python-value.c.
gdb/ChangeLog:
* python/python-internal.h (Py_buffer_deleter): New struct.
(Py_buffer_up): New typedef.
John Baldwin [Mon, 25 Feb 2019 22:25:01 +0000 (14:25 -0800)]
Fix BFD leak in dwarf2_get_dwz_file.
Previously if build_id_verify failed, dwz_bfd was cleared to NULL via
release(), but the BFD object was not destroyed. Use reset() with
nullptr instead to delete the BFD.
gdb/ChangeLog:
* dwarf2read.c (dwarf2_get_dwz_file): Reset dwz_bfd to nullptr
instead of releasing ownership.
Jordan Rupprecht [Mon, 25 Feb 2019 20:21:01 +0000 (12:21 -0800)]
Fix crash when loading dwp files
When loading dwp files, we create an array of ELF sections indexed by the ELF
section index in the dwp file. The size of this array is calculated by
section_count, as returned by bfd_count_sections, plus 1 (to account for the
null section at index 0). However, when loading the bfd file, strtab/symtab
sections are not added to the list, nor do they increment section_count, so
section_count is actually smaller than the number of ELF sections.
This happens to work when using GNU dwp, which lays out .debug section first,
with sections like .shstrtab coming at the end. Other tools, like llvm-dwp, put
.strtab first, and gdb crashes when loading those dwp files.
For instance, with the current state of gdb, loading a file like this:
$ readelf -SW <file.dwp>
[ 0] <empty>
[ 1] .debug_foo PROGBITS ...
[ 2] .strtab STRTAB ...
... results in section_count = 2 (.debug is the only thing placed into
bfd->sections, so section_count + 1 == 2), and sectp->this_idx = 1 when mapping
over .debug_foo in dwarf2_locate_common_dwp_sections, which passes the
assertion that 1 < 2.
However, using a dwp file produced by llvm-dwp:
$ readelf -SW <file.dwp>
[ 0] <empty>
[ 1] .strtab STRTAB ...
[ 2] .debug_foo PROGBITS ...
... results in section_count = 2 (.debug is the only thing placed into
bfd->sections, so section_count + 1 == 2), and sectp->this_idx = 2 when mapping
over .debug_foo in dwarf2_locate_common_dwp_sections, which fails the assertion
that 2 < 2.
Nick Clifton [Mon, 25 Feb 2019 12:15:41 +0000 (12:15 +0000)]
Extend objdump's --dwarf=follow-links option so that separate debug info files will also be affected by other dump function, and symbol tables from separate debug info files will be used when disassembling the main file.
* objdump.c (sym_ok): New function.
(find_symbol_for_address): Use new function.
(disassemble_section): Compare sections by name, not pointer.
(dump_dwarf): Move code to initialise byte_get pointer and iterate
over separate debug files from here to ...
(dump_bfd): ... here. Add parameter indicating that a separate
debug info file is being dumped. For main file, pull in the
symbol tables from all separate debug info files.
(display_object): Update call to dump_bfd.
* doc/binutils.texi: Document extened behaviour of the
--dwarf=follow-links option.
* NEWS: Mention this new feature.
* testsuite/binutils-all/objdump.WK2: Update expected output.
* testsuite/binutils-all/objdump.exp (test_follow_debuglink): Add
options and dump file parameters.
Add extra test.
* testsuite/binutils-all/objdump.WK3: New file.
* testsuite/binutils-all/readelf.exp: Change expected output for
readelf -wKis test.
* testsuite/binutils-all/readelf.wKis: New file.
projects / thirdparty / binutils-gdb.git / log
