This patch adds a sanity check for the size of an armap.
* archive64.c (_bfd_archive_64_bit_slurp_armap): Check parsed_size
against file size before allocating memory. Use bfd_alloc rather
than bfd_zalloc for carsym/strings memory.
This introduces a string cache on the per-BFD object, replacing the
macro and filename caches. Both of these caches just store strings,
so this consolidation by itself saves a little memory (about the size
of a bcache per objfile).
Then this patch switches some allocations on the objfile obstack to
use this bcache instead. This saves more space; and turns out to be a
bit faster as well.
Here are the before and after "maint time" + "maint space" results of
"file ./gdb":
Command execution time: 4.664021 (cpu), 4.728518 (wall)
Space used: 39190528 (+29212672 for this command)
Command execution time: 4.216209 (cpu), 4.107023 (wall)
Space used: 36667392 (+26689536 for this command)
The main interface to the string cache is a new pair of overloaded
methods, objfile::intern.
gdb/ChangeLog
2020-03-04 Tom Tromey <tom@tromey.com>
* symmisc.c (print_symbol_bcache_statistics)
(print_objfile_statistics): Update.
* symfile.c (allocate_symtab): Use intern.
* psymtab.c (partial_symtab::partial_symtab): Use intern.
* objfiles.h (struct objfile_per_bfd_storage) <filename_cache,
macro_cache>: Remove.
<string_cache>: New member.
(struct objfile) <intern>: New methods.
* elfread.c (elf_symtab_read): Use intern.
* dwarf2/read.c (fixup_go_packaging): Intern package name.
(dwarf2_compute_name, dwarf2_physname)
(create_dwo_unit_in_dwp_v1, create_dwo_unit_in_dwp_v2): Intern
names.
(guess_partial_die_structure_name): Update.
(partial_die_info::fixup): Intern name.
(dwarf2_canonicalize_name): Change parameter to objfile. Intern
name.
(dwarf2_name): Intern name. Update.
* buildsym.c (buildsym_compunit::get_macro_table): Use
string_cache.
I noticed that gnutarget was not "const". Since writing through this
pointer would probably be a bug, I think it ought to be. This patch
makes the change.
gdb/ChangeLog
2020-03-04 Tom Tromey <tom@tromey.com>
* jit.c (bfd_open_from_target_memory): Make "target" const.
* corefile.c (gnutarget): Now const.
* gdbcore.h (gnutarget): Now const.
Note that because we should report a signed delta from the previous
VMA it isn't possible to use ngettext. ngettext only supports
unsigned long values. So byte/bytes goes from the message.
PR 25570
* ldlang.c (lang_sizing_iteration): New static var.
(lang_size_sections_1): Warn about no memory region only on first
iteration. Warn about changing start address on first iteration
then any delta from that on subsequent iterations. Report a signed
delta.
(one_lang_size_sections_pass): Increment lang_sizing_iteration.
* config/tc-arm.c (md_apply_fix): Warn if a PC-relative load is
detected in a section which does not have at least 4 byte
alignment.
* testsuite/gas/arm/armv8-ar-it-bad.s: Add alignment directive.
* testsuite/gas/arm/ldr-t.s: Likewise.
* testsuite/gas/arm/sp-pc-usage-t.s: Likewise.
* testsuite/gas/arm/sp-pc-usage-t.d: Finish test at end of
disassembly, ignoring any NOPs that may have been inserted because
of section alignment.
* testsuite/gas/arm/ldr-t.d: Likewise.
Revert the change since it breaks existing behavior of "info registers"
for some architectures. At least AArch64 and ARM are impacted by this change.
gdb/ChangeLog:
2020-03-04 Luis Machado <luis.machado@linaro.org>
Revert aa66aac47b due to regressions
in "info registers" for AArch64/ARM.
The change caused "info registers" to not print GPR's.
gdb/ChangeLog:
2020-02-01 Shahab Vahedi <shahab@synopsys.com>
* target-descriptions.c (tdesc_register_in_reggroup_p): Return 0
when reg->group is empty and reggroup is not.
Since commit 3453e7e409 'Clean up "Reading symbols" output' we no longer print
"done." after the "Reading symbols from" message:
...
$ gdb -q a.out
Reading symbols from a.out...
(gdb)
...
Update docs accordingly.
Build on x86_64-linux.
gdb/doc/ChangeLog:
2020-03-04 Tom de Vries <tdevries@suse.de>
* gdb.texinfo: Remove trailing "done." in "Reading symbols from" lines.
* python.texi: Same.
In lib/fortran.exp, in the helper function fortran_int4, kind
parameter is expected to be printed as (kind=4) for the LLVM
Fortran compiler, Flang along with gfortran. And in the helper
function fortran_int8 kind parameter is expected to be printed
as (kind=8). But for the Flang compiler default kind is not
printed and non default kind is printed differently than gfortran
as below.
integer(kind=4) => integer
integer(kind=8) => integer*8
real(kind=4) => real
real(kind=8) => double precision
complex(kind=4) => complex
logical(kind=4) => logical
character(kind=1) => character
This commit adds support for printing of kind parameter for the
Flang. There should be no change when testing with gfortran.
Note: The current patch overrides earlier patch with below details.
commit c3b149eb76
Author Alok Kumar Sharma (alokkumar.sharma@amd.com)
Earlier patch was incomplete and based on assumption that flang
should be changed to dump a type with kind like the way gfortan does.
Later it was realized that the way flang dumps this info is not
incorrect but different. And changes in gdb test framework are
finalized.
gdb/testsuite/ChangeLog:
* lib/fortran.exp (fortran_int4): Handle flang kind printing.
(fortran_int8): Likewise.
(fortran_real4): Likewise.
(fortran_real8): Likewise.
(fortran_complex4): Likewise.
(fortran_logical4): Likewise.
(fortran_character1): Likewise.
The ELF gABI says in part of sh_addralign: "The value of sh_addr must
be congruent to 0, modulo the value of sh_addralign."
* elf.c (elf_fake_sections): Ensure sh_addralign is such that
sh_addr mod sh_addalign is zero.
On machines with more than one octet per byte, objcopy fills only a part
of the gap between sections.
* objcopy.c (copy_object): Convert from bytes to octets for
--gap-fill and --pad-to.
A customer reported a failure to unwind in a certain core dump. A
lengthy investigation showed that the problem came from the
interaction between the tailcall and inline frame sniffers.
Normally, the regular DWARF unwinder may discover a chain of tail
calls ending in the current frame. In this case, it sets a member on
the dwarf2_frame_cache object, so that a subsequent call into the
tailcall sniffer will create the tailcall frames.
However, in this scenario, what happened is that the DWARF unwinder
did find tailcall frames -- but then the PC of the first such frame
was recognized and claimed by the inline frame sniffer.
This then caused unwinding to go astray further up the stack.
This patch fixes the problem by arranging for the tailcall sniffer to
be called before the inline sniffer. This way, if a DWARF frame has
tailcall information, the tailcalls will always be processed first.
This is safe to do, because the tailcall sniffer can only claim a
frame if the previous frame did in fact find this information. (So,
for example, if no DWARF frame is ever found, then this sniffer will
never trigger.)
This patch also partially reverts:
commit 1ec56e88aa
Author: Pedro Alves <palves@redhat.com>
Date: Fri Nov 22 13:17:46 2013 +0000
Eliminate dwarf2_frame_cache recursion, don't unwind from the dwarf2 sniffer (move dwarf2_tailcall_sniffer_first elsewhere).
That patch moved the call to dwarf2_tailcall_sniffer_first out of
dwarf2_frame_cache, and into dwarf2_frame_prev_register. However, in
this situation, this is too late -- by the time
dwarf2_frame_prev_register is called, the frame in question is already
recognized by the inline frame sniffer.
Rather than fully revert that patch, though, this just arranges to
call dwarf2_tailcall_sniffer_first from dwarf2_frame_cache -- which is
called shortly after the DWARF frame sniffer succeeds, via
compute_frame_id.
I don't know how to write a test case for this.
gdb/ChangeLog
2020-03-03 Tom Tromey <tromey@adacore.com>
* dwarf2/frame.c (struct dwarf2_frame_cache)
<checked_tailcall_bottom, entry_cfa_sp_offset,
entry_cfa_sp_offset_p>: Remove members.
(dwarf2_frame_cache): Call dwarf2_tailcall_sniffer_first.
(dwarf2_frame_prev_register): Don't call
dwarf2_tailcall_sniffer_first.
(dwarf2_append_unwinders): Don't append tailcall unwinder.
* frame-unwind.c (add_unwinder): New fuction.
(frame_unwind_init): Use it. Add tailcall unwinder.
It looks like after doing last minute changes to Makefile.am in commit
06b3c5bdb ("gdbsupport: rename source files to .cc"), I forgot to
re-generate Makefile.in. This patch fixes it.
gdbsupport/ChangeLog:
* Makefile.in: Re-generate.
GDB is not able to print logical true values for Flang compiler.
Actual result:
(gdb) p l
$1 = 4294967295
Expected result:
(gdb) p l
$1 = .TRUE.
This is due to GDB expecting representation of true value being 1.
The Fortran standard doesnt specify how LOGICAL types are represented.
Different compilers use different non-zero values to represent logical
true. The gfortran compiler uses 1 to represent logical true and the
flang compiler uses -1. GDB should accept all the non-zero values as
true.
This is achieved by handling TYPE_CODE_BOOL in f_val_print and
printing any non-zero value as true.
gdb/ChangeLog:
* f-valprint.c (f_val_print): Handle TYPE_CODE_BOOL, any non-zero
value should be printed as true.
gdb/testsuite/ChangeLog:
* gdb.fortran/logical.exp: Add tests that any non-zero value is
printed as true.
Windows executables linked with -dynamicbase get a new base address
when loaded, which makes debugging impossible if the executable isn't
also rebased in gdb.
The new base address is read from the Process Environment Block.
gdb/ChangeLog:
2020-03-03 Hannes Domani <ssbssa@yahoo.de>
* windows-tdep.c (windows_solib_create_inferior_hook): New function.
(windows_init_abi): Set and use windows_so_ops.
Back at:
commit 1f6f6e21fa
Author: Philippe Waroquiers <philippe.waroquiers@skynet.be>
Date: Mon Jun 10 21:41:51 2019 +0200
Ensure GDB printf command can print convenience var strings without a target.
GDB was extended in order to allow the printing of convenience
variables that are strings without a target. However, this introduced
a regression that hasn't been caught by our testsuite (because there
were no tests for it).
The problem happens when we try to print a convenience variable that
holds the address of a string in the inferior. The following
two-liners can reproduce the issue:
$ echo -e 'int main(){const char a[]="test";return 0;}' | gcc -x c - -O0-g3
$ ./gdb/gdb --data-directory ./gdb/data-directory -q ./a.out -ex 'start' -ex 'set $x = (const char *) (&a[0] + 2)' -ex 'printf "%s\n", $x'
After some investigation, I found that the problem happens on
printcmd.c:printf_c_string. In the case above, we're taking the first
branch of the 'if' condition, which assumes that there will be a value
to be printed at "value_contents (value)". There isn't. We actually
need to obtain the address that the variable points to, and read the
contents from memory.
It seems to me that we should avoid this branch if the TYPE_CODE of
"value_type (value)" is TYPE_CODE_PTR (i.e., a pointer to the
inferior's memory). This is what this patch does.
I took the liberty to extend the current testcase under
gdb.base/printcmds.exp and create a test that exercises this scenario.
No regressions have been found on Buildbot.
gdb/ChangeLog:
2020-03-03 Sergio Durigan Junior <sergiodj@redhat.com>
* printcmd.c (print_c_string): Check also for TYPE_CODE_PTR
when verifying if dealing with a convenience variable.
gdb/testsuite/ChangeLog:
2020-03-03 Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.base/printcmds.exp: Add test to verify printf of a
variable holding an address.
According to gas manual, suffix in instruction mnemonics isn't always
required:
When there is no sizing suffix and no (suitable) register operands to
deduce the size of memory operands, with a few exceptions and where long
operand size is possible in the first place, operand size will default
to long in 32- and 64-bit modes.
This includes cvtsi2sd, cvtsi2ss, vcvtsi2sd, vcvtsi2ss, vcvtusi2sd and
vcvtusi2ss. Since they are used in GCC 8 and older GCC releases, they
must be allowed without suffix in AT&T syntax.
gas/
PR gas/25622
* testsuite/gas/i386/i386.exp: Run x86-64-default-suffix and
x86-64-default-suffix-avx.
* testsuite/gas/i386/noreg64.s: Remove cvtsi2sd, cvtsi2ss,
vcvtsi2sd, vcvtsi2ss, vcvtusi2sd and vcvtusi2ss entries.
* testsuite/gas/i386/noreg64.d: Updated.
* testsuite/gas/i386/noreg64.l: Likewise.
* testsuite/gas/i386/x86-64-default-suffix-avx.d: New file.
* testsuite/gas/i386/x86-64-default-suffix.d: Likewise.
* testsuite/gas/i386/x86-64-default-suffix.s: Likewise.
opcodes/
PR gas/25622
* i386-opc.tbl: Add IgnoreSize to cvtsi2sd, cvtsi2ss, vcvtsi2sd,
vcvtsi2ss, vcvtusi2sd and vcvtusi2ss for AT&T syntax.
* i386-tbl.h: Regenerated.
According to intel SDM manual, not all compare flag-modifying instructions
are marcro-fusible with subsequent jcc instructions. For those non-fusible
instructions, -malign-branch doesn't need to align them, only jcc itself
needs to be aligned.
Here are 2 restrictions which separate macro-fusible instruction from not
Restriction 1:
If TEST/AND/CMP/ADD/SUB/INC/DEC is one of the following format:
cmp m, imm
add m, imm
sub m, imm
test m, imm
and m, imm
inc m
dec m
it is unfusible with any jcc instruction.
Restriction 2:
/* Table 3-2. Macro-Fusible Instructions in Haswell Microarchitecture
Note it also works for Skylake and Cascadelake.
---------------------------------------------------------------------
| JCC | ADD/SUB/CMP | INC/DEC | TEST/AND |
| ------ | ----------- | ------- | -------- |
| Jo | N | N | Y |
| Jno | N | N | Y |
| Jc/Jb | Y | N | Y |
| Jae/Jnb | Y | N | Y |
| Je/Jz | Y | Y | Y |
| Jne/Jnz | Y | Y | Y |
| Jna/Jbe | Y | N | Y |
| Ja/Jnbe | Y | N | Y |
| Js | N | N | Y |
| Jns | N | N | Y |
| Jp/Jpe | N | N | Y |
| Jnp/Jpo | N | N | Y |
| Jl/Jnge | Y | Y | Y |
| Jge/Jnl | Y | Y | Y |
| Jle/Jng | Y | Y | Y |
| Jg/Jnle | Y | Y | Y |
Update maybe_fused_with_jcc_p to check if operands of CMP like instructions
can be fused with condition jump.
* gas/config/tc-i386.h (i386_tc_frag_data): Add member mf_type.
(TC_FRAG_INIT): Init mf_type.
* gas/config/tc-i386.c (enum mf_jcc_kind): New enum.
(enum mf_cmp_kind): Ditto.
(maybe_fused_with_jcc_p): Add argument mf_cmp_p to get
mf_type of corresponding instructons, exclude unfusible
instructions.
(add_fused_jcc_padding_frag_p): Likewise.
(add_branch_padding_frag_p): Likewise.
(output_insn): Record mf_type for corresponding instructions.
(i386_macro_fusible_p): New function.
(i386_next_fusible_jcc_frag): Rename from i386_next_jcc_frag,
add argument cmp_fragP to return next fusible jcc frag only.
(i386_classify_machine_dependant_frag): Seperate macro-fusible
instructions from condition jump.
* gas/testsuite/gas/i386/align-branch-9.s: New file.
* gas/testsuite/gas/i386/align-branch-9.d: Ditto.
* gas/testsuite/gas/i386/x86-64-align-branch-9.s: Ditto.
* gas/testsuite/gas/i386/x86-64-align-branch-9.d: Ditto.
* gas/testsuite/gas/i386/i386.exp: Run new tests.
I noticed GDB didn't know a particular AT tag (51) when doing some debugging.
Turns out we're missing a few entries compared to glibc's headers.
This patch adds them to GDB and fixes a failure in gdb.base/auxv.exp as
a result.
gdb/ChangeLog:
2020-03-03 Luis Machado <luis.machado@linaro.org>
* auxv.c (default_print_auxv_entry): Add new AUXV entries.
It didn't take long for oss-fuzz to find double frees due to a bug in
the cleanup logic. It's seen when reading in any alpha-vms object
file except when alpha_vms_vec is the default. But alpha_vms_vec is
of course the default when building for --target=alpha-dec-vms (and
naturally what I used to test the cleanup support since that is the
only target with a cleanup that does anything currently).
Anyway, the bug is that if bfd_check_format_matches is to preserve a
match the cleanup for that match can't be run. Quite obviously that
would destroy part of the match state.
* format.c (struct bfd_preserve): Add cleanup field.
(bfd_preserve_save): Add cleanup param and save.
(bfd_preserve_restore): Return cleanup.
(bfd_preserve_finish): Call the cleanup for the discarded match.
(bfd_check_format_matches): Pass cleanup to bfd_preserve_save,
and clear when preserving a match. Restore cleanup too when
restoring that match.
When running gdb.mi/gdb2549.exp with check-read1, we get:
...
Running src/gdb/testsuite/gdb.mi/gdb2549.exp ...
FAIL: gdb.mi/gdb2549.exp: register values t (timeout)
...
The problem is that the command generates a lot of output, which is matched by
a single '.*':
...
mi_gdb_test "666-data-list-register-values t" \
"666\\^done,register-values=\\\[\{number=\"$decimal\",value=\"$binary\"\}.*\\\]" \
"register values t"
...
Fix this by splitting up the matching and calling exp_continue after each
number/value pair.
Tested on x86_64-linux with make targets check and check-read1.
gdb/testsuite/ChangeLog:
2020-03-03 Tom de Vries <tdevries@suse.de>
* gdb.mi/gdb2549.exp: Fix "register values t" check-read1 timeout.
When running gdb.mi/list-thread-groups-available.exp, we get:
...
Running gdb.mi/list-thread-groups-available.exp ...
ERROR: tcl error sourcing gdb.mi/list-thread-groups-available.exp.
ERROR: Too many arguments to gdb_test_multiple
...
The problem is that the gdb_test_multiple call has as last argument a
$mi_gdb_prompt, which is no longer supported syntax since 590003dc0e
"[gdb/testsuite] Add -lbl option in gdb_test_multiple".
Fix this by using the new -prompt syntax.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2020-03-03 Tom de Vries <tdevries@suse.de>
* gdb.mi/list-thread-groups-available.exp: Use -prompt syntax for
gdb_test_multiple call.
When running gdb.mi/mi-sym-info.exp with check-read1, we run into:
...
FAIL: gdb.mi/mi-sym-info.exp: List all functions
FAIL: gdb.mi/mi-sym-info.exp: List all variables
...
The problem is that while the $mi_gdb_prompt is active, gdb_test_multiple is
used without -prompt "$mi_gdb_prompt$", so it defaults to matching $gdb_prompt.
Fix this by adding the missing gdb_test_multiple arguments.
Reg-tested on x86_64-linux with make targets check and check-read1.
gdb/testsuite/ChangeLog:
2020-03-03 Tom de Vries <tdevries@suse.de>
* gdb.mi/mi-sym-info.exp: Add missing -prompt "$mi_gdb_prompt$" to
gdb_test_multiple calls.
This function returns the result of a quite big condition. I think it
would be more readeable if it was broken up in smaller pieces and
commented. This is what this patch does.
I also introduced gdbarch_supports_displaced_stepping, since it shows
the intent better than checking for gdbarch_displaced_step_copy_insn_p.
I also used that new function in displaced_step_prepare_throw.
I also updated the comment on top of can_use_displaced_stepping, which
seemed a bit outdated with respect to non-stop. The comment likely
dates from before it was possible to have targets that always operate
non-stop under the hood, even when the user-visible mode is all-stop.
No functional changes intended.
gdb/ChangeLog:
* infrun.c (gdbarch_supports_displaced_stepping): New.
(use_displaced_stepping): Break up conditions in smaller pieces.
Use gdbarch_supports_displaced_stepping.
(displaced_step_prepare_throw): Use
gdbarch_supports_displaced_stepping.
This commit aims to give a cleaner mechanism by which the user can
prevent GDB from trying to load any previous command history.
Currently the user can change the path to the history file, either
using a command line flag, or by setting the GDBHISTFILE environment
variable, and if the path is set to a non-existent file, then
obviously GDB wont load any command history. However, this feels like
a bit of a bodge, I'd like to add an official mechanism by which we
can disable command history loading.
Why would we want to prevent command history loading? The specific
use case I have is GDB starting with a CWD that is a network mounted
directory, and there is no command history present. Still GDB will
access the network in order to check for the file. In my particular
use case I'm actually starting a large number of GDB instances in
parallel, all in the same network mounted directory, the large number
of network accesses looking for this file introduces a noticeable
delay at GDB startup.
The approach I'm proposing here is a slight adjustment to the current
rules for setting up the history filename. Currently, if a user does
this, they see an error:
(gdb) set history filename
Argument required (filename to set it to.).
However, if a user does this:
$ GDBHISTFILE= gdb --quiet
(gdb) set history save on
(gdb) q
warning: Could not rename -gdb18416~ to : No such file or directory
So, we already have a bug in this area. My plan is to allow the empty
filename to be accepted, and for this to mean, neither load, nor save
the command history.
This does mean that we now have two mechanisms to prevent saving the
command history:
(gdb) set history filename
or
(gdb) set history save off
But the only way to prevent loading the command history is to set the
filename to the empty string _before_ you get to a GDB prompt, either
using a command line option, or the environment variable.
I've updated some of the show commands, for example this session:
(gdb) set history filename
(gdb) show history filename
There is no filename currently set for recording the command history in.
(gdb) show history save
Saving of the history record on exit is off.
(gdb) set history save on
(gdb) show history save
Saving of the history is disabled due to the value of 'history filename'.
(gdb) set history filename /tmp/hist
(gdb) show history save
Saving of the history record on exit is on.
I've updated the manual, and added some tests.
gdb/ChangeLog:
* NEWS: Mention new behaviour of the history filename.
* top.c (write_history_p): Add comment.
(show_write_history_p): Add header comment, give a different
message when history writing is on, but the history filename is
empty.
(history_filename): Add comment.
(history_filename_empty): New function.
(show_history_filename): Add header comment, give a different
message when the filename is empty.
(init_history): Compare history_filename against nullptr, and only
read history if the filename is not empty.
(set_history_filename): Add header comment, and only make
non-empty filenames absolute.
(init_main): Make the filename argument to 'set history filename'
optional.
gdb/doc/ChangeLog:
* gdb.texinfo (Command History): Extend description for
GDBHISTFILE and GDBHISTSIZE, add detail about the filename for
'set history filename' being optional. Describe the effect of an
empty history filename on 'set history save on'.
gdb/testsuite/ChangeLog:
* gdb.base/default.exp: Remove test of 'set history filename'.
* gdb.base/gdbinit-history.exp: Add tests for setting the history
filename to the empty string.
* lib/gdb.exp (gdb_init): Unset environment variables GDBHISTFILE
and GDBHISTSIZE.
Change-Id: Ia586e4311182fac99113b60f11ef8a11fbd5450b
The floating point register interface has changed to this:
https://github.com/NetBSD/src/blob/trunk/sys/arch/arm/include/reg.h
It now uses VFP instead of FPA registers. This patch updates
arm-nbsd-nat.c accordingly.
Also implements read_description so that these registers are correctly
printed by "info registers" et al.
Tested by compiling & running on arm-netbsd on qemu.
gdb/ChangeLog:
2020-03-02 Christian Biesinger <cbiesinger@google.com>
* arm-nbsd-nat.c (arm_supply_fparegset): Rename to...
(arm_supply_vfpregset): ...this, and update to use VFP registers.
(fetch_fp_register): Update.
(fetch_fp_regs): Update.
(store_fp_register): Update.
(store_fp_regs): Update.
(arm_netbsd_nat_target::read_description): New function.
(fetch_elfcore_registers): Update.
With this commit:
commit 5b6d1e4fa4
Date: Fri Jan 10 20:06:08 2020 +0000
Multi-target support
There was a regression in GDB's support for older aspects of the
remote protocol. Specifically, when a target sends the 'S' stop reply
packet (which doesn't include a thread-id) then GDB has to figure out
which thread actually stopped.
Before the above commit GDB figured this out by using inferior_ptid in
process_stop_reply, which contained the ptid of the current
process/thread. This would be fine for single threaded
targets (which is the only place using an S packet makes sense), but
in the general case, relying on inferior_ptid for processing a stop is
wrong - there's no reason to believe that what was GDB's current
thread will be the same thread that just stopped in the target.
With the above commit the inferior_ptid now has the value null_ptid
inside process_stop_reply, this can be seen in do_target_wait, where
we call switch_to_inferior_no_thread before calling do_target_wait_1.
The problem this causes can be seen in the new test that runs
gdbserver using the flag --disable-packet=T, and causes GDB to throw
this assertion:
inferior.c:279: internal-error: inferior* find_inferior_pid(process_stratum_target*, int): Assertion `pid != 0' failed.
A similar problem was fixed in this commit:
commit 3cada74087
Date: Thu Jan 11 00:23:04 2018 +0000
Fix backwards compatibility with old GDBservers (PR remote/22597)
However, this commit deals with the case where the T packet doesn't
include a thread-id, not the S packet case. This commit solves the
problem providing a thread-id at the GDB side if the remote target
doesn't provide one. The thread-id provided comes from
remote_state::general_thread, however, though this does work, I don't
think it is the ideal solution.
The remote_state tracks two threads, the continue_thread and the
general_thread, these are updated when GDB asks the remote target to
switch threads. The general_thread is set before performing things
like register or memory accesses, and the continue_thread is set
before things like continue or step commands. Further, the
general_thread is updated after a target stops to reference the thread
that stopped.
The first thing to note from the above description is that we have a
cycle of dependency, when a T packet arrives without a thread-id we
fill in the thread-id from the general_thread data. The thread-id
from the stop event is then used to set the general_thread. This in
itself feels a little weird.
The second question is why use the general_thread at all? You'd think
given how they are originally set that the continue thread would be a
better choice. The problem with this is that the continue_thread, if
the user just does "continue", will be set to the minus_one_ptid, in
the remote protocol this means all threads. When the stop arrives
with no thread-id and we use continue_thread we end up with a very
similar assertion to before because we now end up trying to lookup a
thread using the minus_one_ptid. By contrast, once GDB has connected
to a remote target the general_thread will be set to a valid
thread-id, after which, if the target is single threaded, and stop
events arrive without a thread-id, everything works fine.
There is one slight weirdness with the above behaviour though. When
GDB first connects to the remote target inferior_ptid is null_ptid,
however, upon connecting we query the remote for its threads. As the
thread information arrives GDB adds the threads to its internal
database, and this process involves setting inferior_ptid to the id of
each new thread in turn. Once we know about all the threads we wait
for a stop event from the remote target to indicate that GDB is now in
control of the target.
The problem is that after adding the new threads we don't reset
inferior_ptid, and the code path we use to wait for a stop event from
the target also doesn't reset inferior_ptid, so it turns out that
during the initial connection inferior_ptid is not null_ptid. This is
lucky, because during the initial connection the general_thread
variable _is_ set to null_ptid.
So, during the initial connection, if the first stop event is missing
a thread-id then we "provide" a thead-id from general_thread. This
turns out to be null_ptid meaning no thread-id is known, and then
during process_stop_reply we fill in the missing thread-id using
inferior_ptid.
This was all discussed on the mailing list here:
https://sourceware.org/ml/gdb-patches/2020-02/msg01011.html
My proposal for a fix then is:
1. Move the call to switch_to_inferior_no_thread into
do_target_wait_1, this means that in all cases where we are waiting
for an inferior the inferior_ptid will be set to null_ptid. This is
good as no wait code should rely on inferior_ptid.
2. Remove the use of general_thread from the 'T' packet processing.
The general_thread read here was only ever correct by chance, and we
shouldn't be using it this way.
3. Remove use of inferior_ptid from process_stop_event as this is
wrong, and will always be null_ptid now anyway.
4. When a stop_event has null_ptid due to a lack of thread-id (either
from a T packet or an S packet) then pick the first non exited thread
in the target and use that. This will be fine for single threaded
targets. A multi-thread or multi-inferior aware remote target
should be using T packets with a thread-id, so we give a warning if
the target is multi-threaded, and we are still missing a thread-id.
5. Extend the existing test that covered the T packet with missing
thread-id to also cover the S packet.
gdb/ChangeLog:
* remote.c (remote_target::remote_parse_stop_reply): Don't use the
general_thread if the stop reply is missing a thread-id.
(remote_target::process_stop_reply): Use the first non-exited
thread if the target didn't pass a thread-id.
* infrun.c (do_target_wait): Move call to
switch_to_inferior_no_thread to ....
(do_target_wait_1): ... here.
gdb/testsuite/ChangeLog:
* gdb.server/stop-reply-no-thread.exp: Add test where T packet is
disabled.
There is a developer only feature in gdbserver that provides a
command line option --disable-packet that prevents some packets from
being sent, which is used to increase test coverage within GDB.
This commit extends this mechanism to prevent GDBserver from sending
the T stop reply packets, instead limiting GDBserver to only send the
S stop reply packets.
The S stop reply packet is part of the older target control mechanism,
which has design flaws that were worked around with the introduction
of the newer target control mechanism, which uses the T stop reply
packet.
Limiting GDBserver to use S stop packets instead of T stop packets
will, inevitably, mean that GDBserver doesn't function correctly in
many cases involving multiple threads, however, I don't think this is
too important, this is a developer only feature, intended to allow us
to test GDB.
A new test that makes use of this feature will be added in the next
commit.
gdbserver/ChangeLog:
* remote-utils.cc (prepare_resume_reply): Add ability to convert T
reply into an S reply.
* server.cc (disable_packet_T): New global.
(captured_main): Set new global when appropriate.
* server.h (disable_packet_T): Declare.
Add gdb_test_multiple option -lbl, that adds a regexp after the user code that
reads one line, and discards it:
...
-re "\r\n\[^\r\n\]*(?=\r\n)" {
exp_continue
}
...
In order to be able to write:
...
gdb_test_multiple "command" "testname" -lbl {
...
}
...
rewrite the promp_regexp argument usage into the similar:
...
gdb_test_multiple "command" "testname" -prompt $prompt_regexp {
...
}
...
Build and reg-tested on x86_64-linux.
Tested gdb.base/corefile-buildid.exp with both make targets check and
check-read1.
gdb/testsuite/ChangeLog:
2020-03-02 Pedro Alves <palves@redhat.com>
Tom de Vries <tdevries@suse.de>
* lib/gdb.exp (gdb_test_multiple): Handle prompt_regexp option using
-prompt prefix, before user_code argument. Add -lbl option likewise.
(skip_python_tests_prompt, skip_libstdcxx_probe_tests_prompt)
(gdb_is_target_1): Add -prompt prefix and move to before user_code
argument.
* gdb.base/corefile-buildid.exp: Use -lbl option. Rewrite regexps to
have "\r\n" at start-of-line, instead of at end-of-line.
* defs.h includes config.h
* config.h may define _GNU_SOURCE
* if _GNU_SOURCE is defined, that must be before including any system
header (see feature_test_macro(7))
This is necessary to ensure that a prototype for mkostemp() is brought
into scope by <stdlib.h> when compiling filestuff.h, on platforms where
_GNU_SOURCE isn't unconditionally defined for C++.
In file included from ../../gdb/../gdbsupport/scoped_fd.h:24,
from ../../gdb/debuginfod-support.c:22:
../../gdb/../gdbsupport/filestuff.h: In function ‘int gdb_mkostemp_cloexec(char*, int)’:
../../gdb/../gdbsupport/filestuff.h:59:10: error: ‘mkostemp’ was not declared in this scope; did you mean ‘mkstemp’?
gdb/ChangeLog:
2020-02-29 Jon Turney <jon.turney@dronecode.org.uk>
* debuginfod-support.c: Include defs.h first.
lto: Also copy .note.gnu.property section
When generating the separate file with LTO debug sections, we should
also copy .note.gnu.property section.
PR lto/93966
* simple-object.c (handle_lto_debug_sections): Also copy
.note.gnu.property section.
PR 25543
* readelf.c (dump_section_as_strings): Display new-line characters
as \n and then insert a line break.
* testsuite/binutils-all/pr25543.s: New test.
* testsuite/binutils-all/pr25543.d: Test driver.
* testsuite/binutils-all/readelf.exp: Run the new test.
Alan Modra
GDB Administrator
Tom Tromey
Hannes Domani
Alexandre Oliva
Luis Machado
Tom de Vries
Alok Kumar Sharma
Jan Beulich
Christian Eggers
Tom Tromey
Simon Marchi
H.J. Lu
Andrew Burgess
Sergey Belyashov
Sergio Durigan Junior
Nick Clifton
Hongtao Liu
Simon Marchi
Christian Biesinger
Jon Turney
Aaron Merey