I've noticed that "set remote target-features-packet off" before
connecting has no effect -- GDB still fetches a target description
anyway.
The problem is that while most "set remote foo-packet" commands were
fixed by:
From 4082afcc3d Mon Sep 17 00:00:00 2001
From: Pedro Alves <palves@redhat.com>
Date: Fri, 25 Apr 2014 18:07:02 +0100
Subject: [PATCH] Fix several "set remote foo-packet on/off" commands.
<https://sourceware.org/ml/gdb-patches/2014-04/msg00006.html>
the "qXfer" packets where missed. This commit fixes that.
I've changed remote_search_memory too for consistency (seems like
those are the last direct references to packet->support), though the
difference is not observable because the qSearch:memory packet is auto
probed. Note gdb.base/find-unmapped.exp already exercises explicit
"set remote search-memory-packet off".
gdb/ChangeLog:
2017-12-06 Pedro Alves <palves@redhat.com>
* remote.c (remote_query_supported): Don't send "xmlRegisters=" if
"qXfer:features:read"" is disabled.
(remote_write_qxfer, remote_read_qxfer, remote_search_memory):
Check packet_config_support instead of packet->support directly.
gdb/testsuite/ChangeLog:
2017-12-06 Pedro Alves <palves@redhat.com>
* gdb.arch/i386-avx.exp: If testing with a RSP target, check
force-disabling XML descriptions.
--
gdb/remote.c | 16 +++++++++-------
gdb/testsuite/gdb.arch/i386-avx.exp | 25 +++++++++++++++++++++++++
2 files changed, 34 insertions(+), 7 deletions(-)
I failed at git and missed adding/lost changes on the wrong branch, the
result being that I didn't incorporate fixes resulting from Yao's review
comments. This patch fixes that.
There are two places where we should use the unique pointer typedef, and
ChangeLog entries missing.
gdb/ChangeLog:
* target-descriptions.c (struct tdesc_feature) <registers>: Use
tdesc_reg_up typedef.
(struct target_desc) <features>: Use tdesc_feature_up typedef.
This patch makes tdesc_type an abstract base class and creates three
subclasses:
- tdesc_type_builtin, for builtin types
- tdesc_type_vector, for vector types
- tdesc_type_with_fields, for struct, union, flag and enum types
This allows getting rid of the union in tdesc_type and to not allow the
std::vector separately. I tried to go further and create separate
classes for struct, union, flag and enum, but it proved too difficult.
One problem is that from the point of the of the target description
code, the types tdesc_type_* are opaque (only forward-declared).
Therefore, it doesn't know about inheritance relationship between those
classes. This makes it impossible to make functions that accept a
pointer to a base class and pass a pointer to a derived class, for
example. I think this patch here is a good compromise, and if somebody
wants to improve things further, the door is open.
A make_gdb_type virtual pure method is added to tdesc_type, which
replaces the current tdesc_gdb_type function. Calling this method on a
tdesc_type returns the corresponding built gdb type.
gdb/ChangeLog:
* target-descriptions.c (struct tdesc_type): Use default
destructor.
<u>: Remove.
<accept>: Remove.
(struct tdesc_type_builtin): New.
(struct tdesc_type_vector): New.
(struct tdesc_type_with_fields): New.
(tdesc_predefined_types): Change type to tdesc_type_builtin[].
(tdesc_gdb_type): Remove.
(tdesc_register_type): Adjust.
(tdesc_create_vector): Create tdesc_type_vector.
(tdesc_create_struct): Create tdesc_type_with_fields.
(tdesc_set_struct_size): Change parameter type.
(tdesc_create_union): Create tdesc_type_with_fields.
(tdesc_create_flags): Likewise.
(tdesc_create_enum): Likewise.
(tdesc_add_field): Change parameter type.
(tdesc_add_typed_bitfield): Likewise.
(tdesc_add_bitfield): Likewise.
(tdesc_add_flag): Likewise.
(tdesc_add_enum_value): Likewise.
(print_c_tdesc) <visit>: Remove overload with tdesc_type
parameter, add overloads for tdesc_type_builtin,
tdesc_type_with_fields and tdesc_type_vector.
<m_printed_type>: Remove.
<m_printed_element_type, m_printed_type_with_fields>: Add.
* target-descriptions.h (tdesc_create_enum): Change return type.
(tdesc_add_typed_bitfield): Change parameter type.
(tdesc_add_enum_value): Change parameter type.
* xml-tdesc.c (struct tdesc_parsing_data) <current_type>: Change
type to tdesc_type_with_fields.
(tdesc_start_struct): Adjust.
(tdesc_start_flags): Adjust.
(tdesc_start_enum): Adjust.
(tdesc_start_field): Adjust.
* arch/tdesc.h (struct tdesc_type_builtin): Forward-declare.
(struct tdesc_type_vector): Forward-declare.
(struct tdesc_type_with_fields): Forward-declare.
(tdesc_create_struct): Change return type.
(tdesc_create_union): Likewise.
(tdesc_create_flags): Likewise.
(tdesc_add_field): Change parameter type.
(tdesc_set_struct_size): Likewise.
(tdesc_add_bitfield): Likewise.
(tdesc_add_flag): Likewise.
* features: Re-generate C files.
gdb/gdbserver/ChangeLog:
* tdesc.c (struct tdesc_type): Change return type.
(tdesc_add_flag): Change parameter type.
(tdesc_add_bitfield): Likewise.
(tdesc_add_field): Likewise.
(tdesc_set_struct_size): Likewise.
Make tdesc_arch_data::arch_regs be an std::vector of tdesc_arch_reg
objects.
On particularity is that the tdesc_arch_data linked to a gdbarch is
allocated on the gdbarch's obstack. To be safe, I did not change it and
called placement-new on the area returned by OBSTACK_ZALLOC.
gdb/ChangeLog:
* target-descriptions.c (tdesc_arch_reg): Remove typedef.
(struct tdesc_arch_reg): Add constructor.
(DEF_VEC_O (tdesc_arch_reg)): Remove.
(struct tdesc_arch_data): Initialize fields.
<arch_regs>: Change type to std::vector.
(target_find_description): Adjust.
(tdesc_find_type): Adjust.
(tdesc_data_init): Call tdesc_arch_data constructor.
(tdesc_data_alloc): Allocate tdesc_arch_data with new.
(tdesc_data_cleanup): Free data with delete.
(tdesc_numbered_register): Adjust.
(tdesc_find_arch_register): Adjust.
(tdesc_use_registers): Adjust.
This patch makes the tdesc_type::u::u::fields an std::vector of
tdesc_type_field. The difficulty here is that the vector is part of a
union. Because of this, I made fields a pointer to a vector, and
instantiate/destroy the vector if the type is one that uses this member
of the union
The field tdesc_type_field::name is changed to an std::string at the
same time.
gdb/ChangeLog:
* target-descriptions.c (tdesc_type_field): Remove typedef.
(DEF_VEC_O (tdesc_type_field)): Remove.
(struct tdesc_type_field): Add constructor.
<name>: Change type to std::string.
(struct tdesc_type) <tdesc_type>: Instantiate vector if the type
kind uses it.
<~tdesc_type>: Destroy vector if the type kind uses it.
<u::u::fields>: Change type to std::vector.
(tdesc_gdb_type): Adjust.
(tdesc_add_field): Adjust.
(tdesc_add_typed_bitfield): Adjust.
(tdesc_add_field): Adjust.
(tdesc_add_enum_value): Adjust.
(class print_c_tdesc) <visit>: Adjust.
This patch makes tdesc_type::name an std::string. This way, we don't
need to free it manually in ~tdesc_type. I think the comment on top of
name is not correct, the string is always malloc'ed.
gdb/ChangeLog:
* target-descriptions.c (struct tdesc_type) <name>: Change type
to std::string.
<~tdesc_type>: Don't manually free name.
<operator==>: Adjust.
(tdesc_named_type): Adjust.
(tdesc_find_type): Adjust.
(tdesc_gdb_type): Adjust.
(class print_c_tdesc) <visit>: Adjust.
This patch makes tdesc_feature::types an std::vector of unique_ptr of
tdesc_type. This way, we don't need to manually free the objects and
the vector in ~tdesc_feature.
gdb/ChangeLog:
* target-descriptions.c (tdesc_type_p): Remove typedef.
(DEF_VEC_P (tdesc_type_p)): Remove.
(struct tdesc_feature) <types>: Change type to std::vector.
<~tdesc_feature>: Replace with default implementation.
<accept>: Adjust.
(tdesc_named_type): Adjust.
(tdesc_create_vector): Adjust.
(tdesc_create_struct): Adjust.
(tdesc_create_union): Adjust.
(tdesc_create_flags): Adjust.
(tdesc_create_enum): Adjust.
Make the name, group and type fields of tdesc_reg std::strings. This
way, we don't have to manually free them in ~tdesc_reg.
Doing so results in a small change in the generated tdesc. Instead of
passing an empty string for the group parameter of tdesc_create_reg, the
two modified tdesc now pass NULL. The end result should be the same.
gdb/ChangeLog:
* target-descriptions.c (struct tdesc_reg) <tdesc_reg>: Change
type of name_ parameter, adjust to std::string change.
<name, group, type>: Change type to std::string.
<~tdesc_reg>: Replace with default implementation.
<operator==>: Adjust.
(tdesc_find_register_early): Adjust.
(tdesc_register_name): Adjust.
(tdesc_register_type): Adjust.
(tdesc_register_in_reggroup_p): Adjust.
(class print_c_tdesc) <visit>: Adjust.
(class print_c_feature) <visit>: Adjust.
This patch makes tdesc_feature::registers an std::vector of unique_ptr
to tdesc_reg. This way, we don't have to manually free the tdesc_reg
objects and the vector in the tdesc_feature destructor.
gdb/ChangeLog:
* target-descriptions.c (tdesc_reg_p): Remove typedef.
(DEF_VEC_P (tdesc_reg_p)): Remove.
(struct tdesc_feature) <registers>: Change type to std::vector.
<~tdesc_feature>: Don't manually free registers.
<accept>: Adjust.
<operator==>: Adjust.
(tdesc_has_registers): Adjust.
(tdesc_find_register_early): Adjust.
(tdesc_use_registers): Adjust.
(tdesc_create_reg): Adjust.
... so we don't have to manually free it in ~tdesc_feature.
gdb/ChangeLog:
* target-descriptions.c (tdesc_feature) <name>: Change type to
std::string.
<~tdesc_feature>: Don't manually free name.
<operator==>: Adjust.
(tdesc_find_feature): Adjust.
(tdesc_feature_name): Adjust.
(class print_c_tdesc) <visit_pre>: Adjust.
(class print_c_feature) <visit_pre>: Adjust.
This patch makes target_desc to be a vector of unique_ptr to
tdesc_feature objects. This way, we don't have to manually free the
features and the vector in the target_desc destructor.
gdb/ChangeLog:
* target-descriptions.c (tdesc_feature_p): Remove typedef.
(DEF_VEC_P (tdesc_feature_p)): Remove.
(struct target_desc) <features>: Change type to std::vector.
<~target_desc>: Replace with default implementation.
<accept>: Adjust.
<operator==>: Adjust.
(tdesc_has_registers): Adjust.
(tdesc_find_feature): Adjust.
(tdesc_use_registers): Adjust.
(tdesc_create_feature): Adjust.
This patch changes target_desc::compatible to be a vector of
bfd_arch_info *. This way, we don't need to manually free the vector in
the target_desc destructor.
gdb/ChangeLog:
* target-descriptions.c (arch_p): Remove typedef.
(DEF_VEC_P (arch_p)): Remove.
(struct target_desc) <compatible>: Change type to std::vector.
<~target_desc>: Don't manually free compatible.
(tdesc_compatible_p): Adjust.
(tdesc_add_compatible): Adjust.
(class print_c_tdesc) <visit_pre>: Adjust.
This patch changes target_desc::properties to be a vector of property
objects. This way, we don't need to manually free the property members
as well as the property objects themselves.
gdb/ChangeLog:
* target-descriptions.c (property_s): Remove typedef.
(DEF_VEC_O (property_s)): Remove.
(struct target_desc) <properties>: Make an std::vector.
<~target_desc>: Don't manually free properties.
(tdesc_property): Adjust.
(set_tdesc_property): Adjust.
(class print_c_tdesc) <visit_pre>: Adjust.
Since we use C++11, we can use static_assert instead doing the trick
that makes a negative-sized array if the expression is false.
static_assert is built in the language and gives clearer error messages.
To avoid modifying the usages of gdb_static_assert, redefine
gdb_static_assert in terms of static_assert, passing an empty message.
If we want to add an assert with a message, it's always possible to use
static_assert directly.
gdb/ChangeLog:
* common/gdb_assert.h (gdb_static_assert): Redefine using
static_assert.
gdb/testsuite/ChangeLog:
2017-11-30 Sergio Lopez <slp@redhat.com>
* gdb.core/coredump-filter.exp: Extend test to verify
the functionality of the dump-excluded-mappings command.
With the new "-a" command line option, the user may request gcore to
actually dump all present memory mappings. The actual effect of this
argument is OS dependent.
On GNU/Linux, it will disable use-coredump-filter and enable
dump-excluded-mappings.
gdb/ChangeLog:
2017-11-29 Sergio Lopez <slp@redhat.com>
* gcore.in: Add "-a" command line option for instructing gdb to
dump all memory mappings (OS dependent).
Commit df8411da08 implemented support for
checking /proc/PID/coredump_filter, and also changed gcore behavior to
unconditionally honor the VM_DONTDUMP flag, preventing sections marked
as such for being dumped into the core file.
This patch implements the 'set dump-excluded-mappings' command for
instructing gdb to ignore the VM_DONTDUMP flag. Combined with 'set
use-coredump-filter', this allows the user to restore the old behavior,
dumping all sections (except the ones marked as IO) unconditionally.
gdb/Changelog:
2017-11-29 Sergio Lopez <slp@redhat.com>
* linux-tdep.c (dump_excluded_mappings): New variable.
(dump_mapping_p): Use dump_excluded_mappings variable.
(_initialize_linux_tdep): New command 'set dump_excluded_mappings'.
I realized today that a recent change to the Rust support required an
update to the manual; and so I updated NEWS as well.
2017-12-04 Tom Tromey <tom@tromey.com>
* NEWS: Mention Rust trait object inspection.
2017-12-04 Tom Tromey <tom@tromey.com>
* gdb.texinfo (Rust): Update trait object status
PR gdb/22499 is about a latent bug exposed by the switch to "maint set
target-non-stop on" by default on x86-64 GNU/Linux, a while ago. With
that on, GDB is also preferring to use displaced-stepping by default.
The testcase in the bug is failing because GDB ends up incorrectly
displaced-stepping over a RIP-relative VEX-encoded instruction, like
this:
0x00000000004007f5 <+15>: c5 fb 10 05 8b 01 00 00 vmovsd 0x18b(%rip),%xmm0 # 0x400988
While RIP-relative instructions need adjustment when relocated to the
scratch pad, GDB ends up just copying VEX-encoded instructions to the
scratch pad unmodified, with the end result that the inferior ends up
executing an instruction that fetches/writes memory from the wrong
address...
This patch teaches GDB about the VEX-encoding prefixes, fixing the
problem, and adds a testcase that fails without the GDB fix.
I think we may need a similar treatment for EVEX-encoded instructions,
but I didn't address that simply because I couldn't find any
EVEX-encoded RIP-relative instruction in the gas testsuite. In any
case, this commit is forward progress as-is already.
gdb/ChangeLog:
2017-12-04 Pedro Alves <palves@redhat.com>
PR gdb/22499
* amd64-tdep.c (amd64_insn::rex_offset): Rename to...
(amd64_insn::enc_prefix_offset): ... this, and tweak comment.
(vex2_prefix_p, vex3_prefix_p): New functions.
(amd64_get_insn_details): Adjust to rename. Also skip VEX2 and
VEX3 prefixes.
(fixup_riprel): Set VEX3.!B.
gdb/testsuite/ChangeLog:
2017-12-04 Pedro Alves <palves@redhat.com>
PR gdb/22499
* gdb.arch/amd64-disp-step-avx.S: New file.
* gdb.arch/amd64-disp-step-avx.exp: New file.
Now that make-target-delegates understands namespaces and templates,
this typedef is no longer useful.
gdb/ChangeLog:
* target.h (mem_region_vector): Remove.
(struct target_ops) <to_memory_map>: Change return type to
std::vector<mem_region>.
* target-debug.h (target_debug_print_mem_region_vector): Rename
to ...
(target_debug_print_std_vector_mem_region): ... this.
* target-delegates.c: Re-generate.
The next patch will want to use gdb::array_view<int> as parameter type
of a target_ops method. However, that runs into a
make-target-delegates limitation: target_debug_foo calls in
target-delegates.c for parameters/return types with namespace scope
operators ("::") or template parameters, end up looking like:
@@ -1313,9 +1313,7 @@ debug_set_syscall_catchpoint (struct target_ops *self, int arg1, int arg2, int a
fputs_unfiltered (", ", gdb_stdlog);
target_debug_print_int (arg3);
fputs_unfiltered (", ", gdb_stdlog);
- target_debug_print_int (arg4);
- fputs_unfiltered (", ", gdb_stdlog);
- target_debug_print_int_p (arg5);
+ target_debug_print_gdb::array_view<const_int> (arg4);
which obviously isn't something that compiles. The problem is that
make-target-delegates wasn't ever taught that '::', '<', and '>' can
appear in parameter/return types. You could work around it by hidding
the unsupported characters behind a typedef in the target method
declaration, or by using an explicit TARGET_DEBUG_PRINTER, but it's
better to just remove the limitation.
While at it, also fix an "abuse" of reserved identifiers.
gdb/ChangeLog:
* make-target-delegates (munge_type): Also munge '<', '>', and
':'. Avoid double underscores in identifiers, and trailing
underscores.
* target-debug.h
(target_debug_print_VEC_static_tracepoint_marker_p__p): Rename to
...
(target_debug_print_VEC_static_tracepoint_marker_p_p): ... this.
* target-delegates.c: Regenerate.
I noticed [1] a test bug in gdb.threads/process-dies-while-detaching.exp.
Simplified, the test code in question looks somewhat like this:
~~~
# Detach from a process, and ensure that it exits after detaching.
# This relies on inferior I/O.
proc detach_and_expect_exit {test} {
gdb_test_multiple "detach" $test ....
set saw_prompt 0
set saw_inf_exit 0
while { !$saw_prompt && !$saw_inf_exit } {
gdb_test_multiple "" $test {
-re "exited, status=0" {
set saw_inf_exit 1
}
-re "$gdb_prompt " {
set saw_prompt 1
}
}
}
pass $test
}
~~~
The bug is in the while loop's condition. We want to make sure we see
both the inferior output and the prompt, so the loop's test should be:
- while { !$saw_prompt && !$saw_inf_exit } {
+ while { !$saw_prompt || !$saw_inf_exit } {
If we just fix that, the test starts failing though, because it
exposes a couple latent problems:
- When called from test_detach_killed_outside, the parent doesn't
print "exited, status=0", because in that case the child dies with a
signal, and so detach_and_expect_exit times out.
Fix it by making the parent print "signaled, sig=9" in that case,
and have the .exp expect it.
- When testing against --target_board=native-gdbserver, sometimes we'd
get this:
ERROR: Process no longer exists
ERROR: : spawn id exp9 not open
while executing
"expect {
-i exp8 -timeout 220
-i $server_spawn_id
eof {
pass $test
wait -i $server_spawn_id
unset server_spawn_id
}
timeout {
..."
("uplevel" body line 1)
invoked from within
"uplevel $body" NONE : spawn id exp9 not open
The problem is that:
- inferior_spawn_id and server_spawn_id are the same when testing
with gdbserver.
- gdbserver exits after "detach", so we get an eof for
$inferior_spawn_id in the loop in detach_and_expect_exit.
That's the first "ERROR: Process no longer exists".
- and then when we reach test_server_exit, server_spawn_id
is already closed (because server_spawn_id==inferior_spawn_id).
To handle this, make the loop in detach_and_expect_exit use an
indirect spawn id list and remove $inferior_spawn_id from the list
as soon as we got the inferior output we're expecting, so that the
"eof" is left unprocessed until we reach test_server_exit.
[1] I changed GDB in a way that should have made the test fail, but it
didn't.
gdb/testsuite/ChangeLog:
2017-12-03 Pedro Alves <palves@redhat.com>
* gdb.threads/process-dies-while-detaching.c: Include <errno.h>
and <string.h>.
(parent_function): Print distinct messages when waitpid fails, or
the child exits with a signal, or the child exits for an unhandled
reason.
* gdb.threads/process-dies-while-detaching.exp
(detach_and_expect_exit): New 'inf_output_re' parameter and use
it. Wait for both inferior output and GDB's prompt. Use an
indirect spawn id list.
(do_detach): New parameter 'child_exit'. Use it to compute
expected inferior output.
(test_detach, test_detach_watch, test_detach_killed_outside):
Adjust to pass down the expected child exit kind.
All the usages of find_inferior were removed, so the function itself can
be removed.
gdb/gdbserver/ChangeLog:
* inferiors.h (find_inferior): Remove.
* inferiors.c (find_inferior): Remove.
These functions were modified in the previous patch series, but I forgot
to update some comments.
gdb/gdbserver/ChangeLog:
* linux-low.c (resume_status_pending_p): Update comment.
(need_step_over_p): Update comment.
Replace with for_each_thread.
gdb/gdbserver/ChangeLog:
* linux-low.c (linux_resume_one_thread): Return void, take
parameter directly.
(linux_resume): Use for_each_thread.
Replace with find_thread/for_each_thread. I inlined the callbacks,
because they are relatively simple.
gdb/gdbserver/ChangeLog:
* linux-low.c (select_singlestep_lwp_callback): Remove.
(count_events_callback): Remove.
(select_event_lwp_callback): Remove.
(select_event_lwp): Use find_thread/for_each_thread.
Replace with find_thread. Writing a lambda inline in directly in the if
conditions would be a bit messy, so I chose to assign them to variables
instead.
gdb/gdbserver/ChangeLog:
* linux-low.c (not_stopped_callback): Return bool, take filter
argument directly.
(linux_wait_for_event_filtered): Use find_thread.
(linux_wait_1): Likewise.
Replace with find_thread. We could almost use find_thread_ptid, except
that find_lwp_pid uses the pid of the input ptid of the lwp is 0, so the
behavior is not quite the same.
gdb/gdbserver/ChangeLog:
* linux-low.c (same_lwp): Remove.
(find_lwp_pid): Use find_thread.
Replace with for_each_thread with pid filtering. The callback becomes
trivial enough that it's better to inline it.
gdb/gdbserver/ChangeLog:
* linux-low.c (delete_lwp_callback): Remove.
(linux_mourn): Use for_each_thread.
Replace it with find_thread. I also modified the code a bit to use a
lambda and a boolean.
gdb/gdbserver/ChangeLog:
* linux-low.c (struct counter): Remove.
(second_thread_of_pid_p): Remove.
(last_thread_of_process_p): Use find_thread.
Replace with for_each_thread with pid filtering. This allows
simplifying the callback a little bit.
gdb/gdbserver/ChangeLog:
* linux-mips-low.c (update_watch_registers_callback): Return
void, remove pid_p parameter, don't check for pid.
(mips_insert_point, mips_remove_point): Use for_each_thread.
Replace it with for_each_thread with pid filtering. We can remove
lynx_delete_thread_callback and pass remove_thread directly.
I can't build/test this change, but it should be obvious enough.
gdb/gdbserver/ChangeLog:
* lynx.low (lynx_delete_thread_callback): Remove.
(lynx_mourn): Use for_each_thread.
Replace with for_each_thread with pid filtering.
regcache_invalidate_one is not longer needed, as it was only used to
filter the pid. We can call regcache_invalidate_thread directly.
gdb/gdbserver/ChangeLog:
* regcache.c (regcache_invalidate_one): Remove.
(regcache_invalidate_pid): use for_each_thread.
The purpose of this concept is to turn the load of debugging
information off, either globally (via the '--readnever' option), or
objfile-specific. The implementation proposed here is an extension of
the patch distributed with Fedora GDB; looking at the Fedora patch
itself and the history, one can see some reasons why it was never
resubmitted:
- The patch appears to have been introduced as a workaround, at
least initially;
- The patch is far from perfect, as it simply shunts the load of
DWARF debugging information, without really worrying about the
other debug format.
- Who really does non-symbolic debugging anyways?
One use of this feature is when a user simply wants to do the
following sequence: attach, dump core, detach. Loading the debugging
information in this case is an unnecessary cause of delay.
This patch expands the version shipped with Fedora GDB in order to
make the feature available for all the debuginfo backends, not only
for DWARF. It also implements a per-objfile flag which can be
activated by using the "-readnever" command when using the
'add-symbol-file' or 'symbol-file' commands.
It's also worth mentioning that this patch tests whether GDB correctly
fails to initialize if both '--readnow' and '--readnever' options are
passed.
Tested on the BuildBot.
gdb/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* NEWS (Changes since GDB 8.0: Mention new '--readnever'
feature.
* coffread.c (coff_symfile_read): Do not map over sections with
'coff_locate_sections' if readnever is on.
* dwarf2read.c (dwarf2_has_info): Return 0 if
readnever is on.
* elfread.c (elf_symfile_read): Do not map over sections with
'elf_locate_sections' if readnever is on.
* main.c (validate_readnow_readnever): New function.
(captured_main_1): Add support for --readnever.
(print_gdb_help): Document --readnever.
* objfile-flags.h (enum objfile_flag) <OBJF_READNEVER>: New
flag.
* symfile.c (readnever_symbol_files): New global.
(symbol_file_add_with_addrs): Set 'OBJF_READNEVER' when
'READNEVER_SYMBOL_FILES' is set.
(validate_readnow_readnever): New function.
(symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(add_symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(_initialize_symfile): Document new '-readnever' option for
both 'symbol-file' and 'add-symbol-file' commands.
* top.h (readnever_symbol_files): New extern global.
* xcoffread.c (xcoff_initial_scan): Do not read debug
information if readnever is on.
gdb/doc/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.texinfo (File Options): Document --readnever.
(Commands to Specify Files): Likewise, for 'symbol-file' and
'add-symbol-file'.
gdb/testsuite/ChangeLog:
2017-12-01 Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
Pedro Alves <palves@redhat.com>
* gdb.base/readnever.c, gdb.base/readnever.exp: New files.
This is a bug that's been detected while doing the readnever work.
If you use 'symbol-file' or 'add-symbol-file', the position of each
argument passed to the command matters. This means that if you do:
(gdb) symbol-file -readnow /foo/bar
The symbol file specified will (correctly) have all of its symbols
read by GDB (because of the -readnow flag). However, if you do:
(gdb) symbol-file /foo/bar -readnow
GDB will silently ignore the -readnow flag, because it was specified
after the filename. This is not a good thing to do and may confuse
the user.
To address that, I've modified the argument parsing mechanisms of
symbol_file_command and add_symbol_file_command to be
"position-independent". I have also added one error call at the end
of add_symbol_file_command's argument parsing logic, which now clearly
complains if no filename has been specified. Both commands now
support the "--" option to stop argument processing.
This patch provides a testcase for both commands, in order to make
sure that the argument order does not matter. It has been
regression-tested on BuildBot.
gdb/ChangeLog:
2017-12-01 Sergio Durigan Junior <sergiodj@redhat.com>
* symfile.c (symbol_file_command): Call
'symbol_file_add_main_1' only after processing all command
line options.
(add_symbol_file_command): Modify logic to make arguments
position-independent.
gdb/testsuite/ChangeLog:
2017-12-01 Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.base/relocate.exp: Add tests to guarantee that arguments
to 'symbol-file' and 'add-symbol-file' can be
position-independent.
One of our users reported that trying to print the following expression,
caused GDB to SEGV:
(gdb) print some_package.some_type (val)
In this particular instance, the crash occurred inside ada_args_match
because it is given a NULL "func", leading to the SEGV because of:
struct type *func_type = SYMBOL_TYPE (func);
This NULL symbol comes from a list of symbols which was given to
ada_resolve_function (parameter called "syms") which then iterates
over each of them to discard the ones that don't match the actuals:
for (k = 0; k < nsyms; k += 1)
{
struct type *type = ada_check_typedef (SYMBOL_TYPE (syms[k].symbol));
if (ada_args_match (syms[k].symbol, args, nargs)
&& (fallback || return_match (type, context_type)))
[...]
}
What's really interesting is that, when entering the block above for
the first time, all entries in SYMS have a valid (non-NULL) symbol.
However, once we return from the call to ada_check_typedef, the first
entry of our SYMS table gets set to all zeros:
(gdb) p syms[0]
$2 = {symbol = 0x0, block = 0x0}
Hence the call to ada_args_match with a NULL symbol, and the ensuing
SEGV.
To find out why this happen, we need to step back a little and look
at how syms was allocated. This list of symbols comes from a symbol
lookup, which means ada_lookup_symbol_list_worker. We have our first
hint when we look at the function's documentation and see:
This vector is transient---good only to the next call of
ada_lookup_symbol_list.
Implementation-wise, this is done by using a static global obstack,
which we just re-initialize each time ada_lookup_symbol_list_worker
gets called:
obstack_free (&symbol_list_obstack, NULL);
obstack_init (&symbol_list_obstack);
This property was probably established in order to facilitate the use
of the returned vector, since the users of that function would not have
to worry about releasing that memory when no longer needed. However,
I found during this investigation that it is all to easy to indirectly
trigger another symbol lookup while still using the results of a previous
lookup.
In our particular case, there is the call to ada_check_typedef, which
leads to check_typedef. As it happens, my first symbol had a type which
was a typedef to a stub type, so check_typedef calls lookup_symbol to
find the non-stub version. This in turn eventually leads us back to
ada_lookup_symbol_list_worker, where the first thing it does is free
the memory area when our list of symbols have been residing and then
recreates a new one. in other words, SYMS then becomes a dangling
pointer!
This patch fixes the issue by having ada_lookup_symbol_list_worker
return a copy of the list of symbols, with the responsibility of
deallocating that list now transfered to the users of that list.
More generally speaking, it is absolutely amazing that we haven't seen
consequences of this issue before. This can happen fairly frequently.
For instance, I found that ada-exp.y::write_var_or_type calls
ada_lookup_symbol_list, and then, while processing that list, calls
select_possible_type_sym, which leads to ada_prefer_type, eventually
leading to ada_check_typedef again (via eg. ada_is_array_descriptor_type).
Even more amazing is the fact that, while I was able to produce multiple
scenarios where the corruption occurs, none of them leads to incorrect
behavior at the user level. In other words, it requires a very precise
set of conditions for the corruption to become user-visible, and
despite having a megalarge program where the crash occured, using that
as a template for creating a reproducer did not work (pb goes away).
This is why this patch does not come with a reproducer. On the other hand,
this should not be a problem in terms of testing coverage, as the changes
are made in common areas which, at least for the most part, are routinely
exercised during testing.
gdb/ChangeLog:
* ada-lang.c (symbol_list_obstack): Delete.
(resolve_subexp): Make sure "candidates" gets xfree'ed.
(ada_lookup_symbol_list_worker): Remove the limitation that
the result is only good until the next call, now making it
the responsibility of the caller to free the result when no
longer needed. Adjust the function's intro comment accordingly.
(ada_lookup_symbol_list): Adjust the function's intro comment.
(ada_iterate_over_symbols): Make sure "results" gets xfree'ed.
(ada_lookup_encoded_symbol, get_var_value): Likewise.
(_initialize_ada_language): Remove symbol_list_obstack
initialization.
* ada-exp.y (block_lookup): Make sure "syms" gets xfree'ed.
(write_var_or_type, write_name_assoc): Likewise.
Tested on x86_64-linux.
The purpose of this concept is to turn the load of debugging
information off, either globally (via the '--readnever' option), or
objfile-specific. The implementation proposed here is an extension of
the patch distributed with Fedora GDB; looking at the Fedora patch
itself and the history, one can see some reasons why it was never
resubmitted:
- The patch appears to have been introduced as a workaround, at
least initially;
- The patch is far from perfect, as it simply shunts the load of
DWARF debugging information, without really worrying about the
other debug format.
- Who really does non-symbolic debugging anyways?
One use of this feature is when a user simply wants to do the
following sequence: attach, dump core, detach. Loading the debugging
information in this case is an unnecessary cause of delay.
This patch expands the version shipped with Fedora GDB in order to
make the feature available for all the debuginfo backends, not only
for DWARF. It also implements a per-objfile flag which can be
activated by using the "-readnever" command when using the
'add-symbol-file' or 'symbol-file' commands.
It's also worth mentioning that this patch tests whether GDB correctly
fails to initialize if both '--readnow' and '--readnever' options are
passed.
Tested on the BuildBot.
gdb/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* NEWS (Changes since GDB 8.0: Mention new '--readnever'
feature.
* coffread.c (coff_symfile_read): Do not map over sections with
'coff_locate_sections' if readnever is on.
* dwarf2read.c (dwarf2_has_info): Return 0 if
readnever is on.
* elfread.c (elf_symfile_read): Do not map over sections with
'elf_locate_sections' if readnever is on.
* main.c (validate_readnow_readnever): New function.
(captured_main_1): Add support for --readnever.
(print_gdb_help): Document --readnever.
* objfile-flags.h (enum objfile_flag) <OBJF_READNEVER>: New
flag.
* symfile.c (readnever_symbol_files): New global.
(symbol_file_add_with_addrs): Set 'OBJF_READNEVER' when
'READNEVER_SYMBOL_FILES' is set.
(validate_readnow_readnever): New function.
(symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(add_symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(_initialize_symfile): Document new '-readnever' option for
both 'symbol-file' and 'add-symbol-file' commands.
* top.h (readnever_symbol_files): New extern global.
* xcoffread.c (xcoff_initial_scan): Do not read debug
information if readnever is on.
gdb/doc/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.texinfo (File Options): Document --readnever.
(Commands to Specify Files): Likewise, for 'symbol-file' and
'add-symbol-file'.
gdb/testsuite/ChangeLog:
2017-12-01 Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
Pedro Alves <palves@redhat.com>
* gdb.base/readnever.c, gdb.base/readnever.exp: New files.