When running test-case gdb.base/with.exp with target board cc-with-gdb-index,
we have:
...
(gdb) PASS: gdb.base/with.exp: basics: show language
with language ada -- print g_s^M
'g_s' has unknown type; cast it to its declared type^M
(gdb) FAIL: gdb.base/with.exp: basics: with language ada -- print g_s
...
This is due to this bit in dw2_map_matching_symbols:
...
if (dwarf2_per_objfile->index_table != nullptr)
{
/* Ada currently doesn't support .gdb_index (see PR24713). We can get
here though if the current language is Ada for a non-Ada objfile
using GNU index. As Ada does not look for non-Ada symbols this
function should just return. */
return;
}
...
While the reasoning in the comment may be sound from language perspective, it
does introduce an inconsistency in gdb behaviour between:
- having a .gdb_index section, and
- having a .gdb_names section, or a partial symtab, or -readnow.
Fix the inconsistency by completing implementation of
dw2_map_matching_symbols.
Tested on x86_64-linux, both with native and target board
cc-with-debug-index.
gdb/ChangeLog:
2020-05-20 Tom de Vries <tdevries@suse.de>
PR symtab/25833
* dwarf2/read.c (dw2_map_matching_symbols): Handle .gdb_index.
gdb/testsuite/ChangeLog:
2020-05-20 Tom de Vries <tdevries@suse.de>
PR symtab/25833
* gdb.base/with-mf-inc.c: New test.
* gdb.base/with-mf-main.c: New test.
* gdb.base/with-mf.exp: New file.
ldmain.c:add_archive_element copies file name pointers from the bfd to
a lang_input_statement_type.
input->filename = abfd->filename;
input->local_sym_name = abfd->filename;
This results in stale pointers when twiddling the bfd filename in
places like the pe ld after_open. So don't free the bfd filename,
and make copies using bfd_alloc memory that won't result in small
memory leaks that annoy memory checkers.
PR 25993
bfd/
* archive.c (_bfd_get_elt_at_filepos): Don't strdup filename,
use bfd_set_filename.
* elfcode.h (_bfd_elf_bfd_from_remote_memory): Likewise.
* mach-o.c (bfd_mach_o_fat_member_init): Likewise.
* opncls.c (bfd_fopen, bfd_openstreamr, bfd_openr_iovec, bfd_openw),
(bfd_create): Likewise.
(_bfd_delete_bfd): Don't free filename.
(bfd_set_filename): Copy filename param to bfd_alloc'd memory,
return pointer to the copy or NULL on alloc fail.
* vms-lib.c (_bfd_vms_lib_get_module): Free newname and test
result of bfd_set_filename.
* bfd-in2.h: Regenerate.
gdb/
* solib-darwin.c (darwin_bfd_open): Don't strdup pathname for
bfd_set_filename.
* solib-aix.c (solib_aix_bfd_open): Use std::string for name
passed to bfd_set_filename.
* symfile-mem.c (add_vsyscall_page): Likewise for string
passed to symbol_file_add_from_memory.
(symbol_file_add_from_memory): Make name param a const char* and
don't strdup.
ld/
* emultempl/pe.em (gld_${EMULATION_NAME}_after_open): Don't copy
other_bfd_filename for bfd_set_filename, and test result of
bfd_set_filename call. Don't create a new is->filename, simply
copy from bfd filename. Free new_name after bfd_set_filename.
* emultempl/pep.em (gld_${EMULATION_NAME}_after_open): Likewise.
Check sizes early, before users of slurp_relocs allocate buffers for
the swapped in relocs.
PR 26011
* elf.c (_bfd_elf_get_reloc_upper_bound): Sanity check reloc
section size against file size.
(_bfd_elf_get_dynamic_reloc_upper_bound): Likewise.
This patch makes gdb use the inline accessor for all bfd->filename
read accesses.
* coff-pe-read.c (read_pe_exported_syms): Use bfd_get_filename
rather than accessing bfd->filename directly.
* dtrace-probe.c (dtrace_static_probe_ops::get_probes): Likewise,
and use bfd_section_name.
* dwarf2/frame.c (decode_frame_entry): Likewise.
* exec.c (exec_set_section_address): Likewise.
* solib-aix.c (solib_aix_bfd_open): Likewise.
* stap-probe.c (get_stap_base_address): Likewise.
* symfile.c (reread_symbols): Likewise.
An earlier patch inadvertently broke a Rust test. This restores it.
gdb/testsuite/ChangeLog
2020-05-19 Tom Tromey <tromey@adacore.com>
* gdb.rust/simple.exp: Restore missing test result.
gdb.rust complains about some duplicate test names. This patch fixes
this in a straightforward way.
2020-05-19 Tom Tromey <tromey@adacore.com>
* gdb.rust/simple.exp: Add some test descriptions.
(test_one_slice): Use with_test_prefix.
An earlier patch changed target_fileio_open, but missed a caller.
This patch fixes it.
2020-05-19 Tom Tromey <tromey@adacore.com>
* sparc64-tdep.c (adi_tag_fd): Update call to target_fileio_open.
Found by inspection, so I don't have a test for it (I don't think it
would be easy to have this bug cause a failure reliably).
We allocate space for N fields into `new_fields`, then memcpy N fields
at `new_fields + 1`. This overflows the allocated buffer by one field.
Fix it by allocating `N + 1` fields.
gdb/ChangeLog:
* dwarf2/read.c (quirk_rust_enum): Allocate enough fields.
The patch makes GDB do exec-file-mismatch validation by comparing
build IDs instead of the current method of comparing filenames.
Currently, the exec-file-mismatch feature simply compares filenames to
decide whether the exec file loaded in gdb and the exec file the
target reports is running match. This causes false positives when
remote debugging, because it'll often be the case that the paths in
the host and the target won't match. And of course misses the case of
the files having the same name but being actually different files
(e.g., different builds).
This also broke many testcases when running against gdbserver, causing
tests to be skipped like (here native-extended-gdbserver):
(gdb) run
Starting program: /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink-filelink
warning: Mismatch between current exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink-filelink
and automatically determined exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink
exec-file-mismatch handling is currently "ask"
Load new symbol table from "/home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink"? (y or n) UNTESTED: gdb.base/argv0-symlink.exp: could not run to main
or to fail like (here native-gdbserver):
(gdb) spawn /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/../../gdbserver/gdbserver --once localhost:2346 /home/pedro/gdb/binutils-gdb/build/gdb/te
stsuite/outputs/gdb.btrace/buffer-size/skip_btrace_tests-19968.x
Process /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.btrace/buffer-size/skip_btrace_tests-19968.x created; pid = 20040
Listening on port 2346
target remote localhost:2346
Remote debugging using localhost:2346
warning: Mismatch between current exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/temp/19968/skip_btrace_tests-19968.x
and automatically determined exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.btrace/buffer-size/skip_btrace_tests-19968.x
exec-file-mismatch handling is currently "ask"
Load new symbol table from "/home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.btrace/buffer-size/skip_btrace_tests-19968.x"? (y or n) Quit
(gdb) UNSUPPORTED: gdb.btrace/buffer-size.exp: target does not support record-btrace
The former case is about GDB not realizing the two files are the same,
because one of the them is a symlink to the other. The latter case is
about GDB realizing that one file is a copy of the other.
Over the years, the toolchain has settled on build ID matching being
the canonical method to match core dumps to executables, and
executables with no debug info to their debug info.
This patch makes us use build IDs to match the running image of a
binary with its version loaded in gdb, which may or may not have debug
info. This is very much like the core dump/executable matching.
The change to gdb_bfd_open is necessary to get rid of the "transfers
from remote targets can be slow" warning when we open the remote file
to read its build ID:
(gdb) r
Starting program: /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/break/break
Reading /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink from remote target...
warning: File transfers from remote targets can be slow. Use "set sysroot" to access files locally instead.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
warning: Mismatch between current exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/break/break
and automatically determined exec-file /home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink
exec-file-mismatch handling is currently "ask"
Load new symbol table from "/home/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.base/argv0-symlink/argv0-symlink"? (y or n)
While trying this out, I was worried that bfd would read a lot of
stuff from the binary in order to extract the build ID, making it
potentially slow, but turns out we don't read all that much. Maybe a
couple hundred bytes, and most of it seemingly is the read-ahead
cache. So I'm not worried about that. Otherwise I'd consider whether
a new qXfer:buildid:read would be better. But I'm happy that we
seemingly don't need to worry about it.
gdb/ChangeLog:
2020-05-19 Pedro Alves <palves@redhat.com>
* NEWS (set exec-file-mismatch): Adjust entry.
* exec.c: Include "build-id.h".
(validate_exec_file): Try to match build IDs instead of filenames.
* gdb_bfd.c (struct gdb_bfd_open_closure): New.
(gdb_bfd_iovec_fileio_open): Adjust to use gdb_bfd_open_closure
and pass down 'warn_if_slow'.
(gdb_bfd_open): Add 'warn_if_slow' parameter. Use
gdb_bfd_open_closure to pass it down.
* gdb_bfd.h (gdb_bfd_open): Add 'warn_if_slow' parameter.
gdb/doc/ChangeLog:
2020-05-19 Pedro Alves <palves@redhat.com>
* gdb.texinfo (Attach): Update exec-file-mismatch description to
mention build IDs.
(Separate Debug Files): Add "build id" anchor.
This basically makes target_fileio_open_1 extern, renamed to
target_fileio_open, and eliminates the current
target_fileio_open_warn_if_slow and target_fileio_open.
A following parameter will want to change gdb_bfd_iovec_fileio_open,
the only caller of target_fileio_open_warn_if_slow, to pass down
"warn_if_slow" true/false from the caller, instead of hardcoding
"warn_if_slow" true.
gdb/ChangeLog:
2020-05-19 Pedro Alves <palves@redhat.com>
* gdb_bfd.c (gdb_bfd_iovec_fileio_open): Adjust.
* target.c (target_fileio_open_1): Rename to target_fileio_open
and make extern. Use bool.
(target_fileio_open, target_fileio_open_warn_if_slow): Delete.
(target_fileio_read_alloc_1): Adjust.
* target.h (target_fileio_open): Add 'warn_if_slow' parameter.
(target_fileio_open_warn_if_slow): Delete declaration.
A following patch will add one more defaulted parameter.
gdb/ChangeLog:
2020-05-19 Pedro Alves <palves@redhat.com>
* gdb_bfd.h: (gdb_bfd_open): Default to 'fd' parameter to -1.
Adjust all callers.
Compiling with clang 11 gives us:
CXX h8300-tdep.o
/home/smarchi/src/binutils-gdb/gdb/h8300-tdep.c:225:21: error: overlapping comparisons always evaluate to false [-Werror,-Wtautological-overlap-compare]
if (disp < 0 && disp > 0xffffff)
~~~~~~~~~^~~~~~~~~~~~~~~~~~
/home/smarchi/src/binutils-gdb/gdb/h8300-tdep.c:203:17: error: overlapping comparisons always evaluate to false [-Werror,-Wtautological-overlap-compare]
if (disp < 0 && disp > 0xffffff)
~~~~~~~~~^~~~~~~~~~~~~~~~~~
/home/smarchi/src/binutils-gdb/gdb/h8300-tdep.c:184:17: error: overlapping comparisons always evaluate to false [-Werror,-Wtautological-overlap-compare]
if (disp < 0 && disp > 0xffffff)
~~~~~~~~~^~~~~~~~~~~~~~~~~~
Indeed, disp (of type LONGEST) can't be less than 0 and greater than
0xffffff.
Fix it by changing the way we check if disp is negative. Check the sign
bit of disp, which is a 24-bit number.
gdb/ChangeLog:
* h8300-tdep.c (h8300_is_argument_spill): Change how we check
whether disp is negative.
Change the symfile_segment_data::segment_info array to be an
std::vector. No functional changes are expected.
gdb/ChangeLog:
* symfile.h (struct symfile_segment_data)
<~symfile_segment_data>: Remove.
<segment_info>: Change to std::vector.
* symfile.c (default_symfile_segments): Update.
* elfread.c (elf_symfile_segments): Update.
Instead of maintaining two vectors, I added a small `segment` class
which holds both the base address and size of one segment and replaced
the two `segment_bases` and `segment_sizes` arrays with a single vector.
The rest of the changes are straightforward, no behavior changes are
expected.
gdb/ChangeLog:
* symfile.h (struct symfile_segment_data) <struct segment>: New.
<segments>: New.
<segment_bases, segment_sizes>: Remove.
* symfile.c (default_symfile_segments): Update.
* elfread.c (elf_symfile_segments): Update.
* remote.c (remote_target::get_offsets): Update.
* solib-target.c (solib_target_relocate_section_addresses):
Update.
- Allocate this structure with new instead of XNEW, use a unique pointer
to manage its lifetime.
- Change a few functions to return a unique pointer instead of a
plain pointer.
- Change free_symfile_segment_data to be symfile_segment_data's
destructor.
gdb/ChangeLog:
* symfile.h (struct symfile_segment_data): Initialize fields.
<~symfile_segment_data>: Add.
(symfile_segment_data_up): New.
(struct sym_fns) <sym_segments>: Return a
symfile_segment_data_up.
(default_symfile_segments): Return a symfile_segment_data_up.
(free_symfile_segment_data): Remove.
(get_symfile_segment_data): Return a symfile_segment_data_up.
* symfile.c (default_symfile_segments): Likewise.
(get_symfile_segment_data): Likewise.
(free_symfile_segment_data): Remove.
(symfile_find_segment_sections): Update.
* elfread.c (elf_symfile_segments): Return a
symfile_segment_data_up.
* remote.c (remote_target::get_offsets): Update.
* solib-target.c (solib_target_relocate_section_addresses):
Update.
* symfile-debug.c (debug_sym_segments): Return a
symfile_segment_data_up.
PR binutils/25809
* readelf.c (process_program_headers): Warn if the PT_DYNAMIC
segment doesn't match the .dynamic section and checks are
enabled.
(struct filedata): Add dynamic_symtab_section and
dynamic_strtab_section fields.
(process_section_headers): Set dynamic_symtab_section to the
.dynsym section. Set dynamic_strtab_section to the .dynstr
section.
(process_dynamic_section): Warn if the .dynsym section doesn't
match DT_SYMTAB and DT_SYMENT or the .dynstr section doesn't
DT_STRTAB and DT_STRSZ. But only if checks are enabled.
or1k: Fix static linking when with .rela.got relocations
or1k: Fix dynamic TLS symbol flag
or1k: Add TLS mask to handle multiple model access
or1k: Fix issue with multiple PCREL relocations
or1k: TLS offset to use tcb size and section alignment
or1k: refactor: Rename p to sec_relocs
or1k: refactor: Rename s to sgot and splt
or1k: Add dynamic flag to tpoff
bfd * elf32-or1k.c (or1k_elf_finish_dynamic_symbol): Rename srela
to relgot.
(or1k_elf_relocate_section): Access srelgot via
htab->root.srelgot. Add assertions for srelgot->contents.
Introduce local variable for srelgot to not reuse global
sreloc.
(or1k_elf_relocate_section): Fixup dynamic symbol detection.
(or1k_set_got_and_rela_sizes): New function.
(or1k_initial_exec_offset): New function.
(TLS_GD, TLS_IE, TLS_LD, TLS_LE): Redefine macros as masks.
(or1k_elf_relocate_section): Allow for TLS to handle multiple
model access.
(or1k_elf_check_relocs): Use OR to set TLS access.
(allocate_dynrelocs): Use or1k_set_got_and_rela_sizes to set
sizes.
(or1k_elf_size_dynamic_sections): Use
or1k_set_got_and_rela_sizes to set sizes.
(or1k_elf_relocate_section): Fixup PCREL relocation calculation.
(TCB_SIZE): New macro.
(tpoff): Use TCB_SIZE and alignment to calculate offset.
(allocate_dynrelocs, readonly_dynrelocs, or1k_elf_check_relocs)
(or1k_elf_size_dynamic_sections): Rename p to sec_relocs.
(allocate_dynrelocs): Rename s to splt or sgot based on usage.
(tpoff): Add dynamic boolean argument.
(or1k_elf_relocate_section): Pass dynamic flag to tpoff.
Reported by Rich Felker when building on 32-bit hosts. Backwards jump
negative offsets were not calculated correctly due to improper 32-bit
to 64-bit zero-extension. The 64-bit fields are present because we
are mixing 32-bit and 64-bit architectures in our cpu descriptions.
Removing 64-bit fixes the issue. We don't use 64-bit, there is an architecture
spec for 64-bit but no implementations or simulators. My thought is if
we need them in the future we should do the proper work to support both
32-bit and 64-bit implementations co-existing then.
cpu/ChangeLog:
yyyy-mm-dd Stafford Horne <shorne@gmail.com>
PR 25184
* or1k.cpu (arch or1k): Remove or64 and or64nd machs.
(ORBIS-MACHS, ORFPX32-MACHS): Remove pmacros.
(cpu or1k64bf, mach or64, mach or64nd): Remove definitions.
* or1kcommon.cpu (h-fdr): Remove hardware.
* or1korfpx.cpu (rDDF, rADF, rBDF): Remove operand definitions.
(float-regreg-insn): Remove lf- mnemonic -d instruction pattern.
(float-setflag-insn-base): Remove lf-sf mnemonic -d pattern.
(float-cust-insn): Remove "lf-cust" cust-num "-d" pattern.
(lf-rem-d, lf-itof-d, lf-ftoi-d, lf-madd-d): Remove.
The linker silently writes out a conditional branch to 0 if the
target symbol in R_AARCH64_CONDBR19 or R_AARCH64_TSTBR14 relocations is
undefined. Emit a PLT instead so that behaviour is the same for these
relocations as the llvm linker.
The special behaviour for undefined weak symbols, where conditional
branches to such symbols result in a branch unto themselves, has been
retained. This is because the weak-undefined.s test explicitly checks
for that, leading me to conclude that it's expected behaviour.
bfd * elfnn-aarch64.c (elfNN_aarch64_final_link_relocate): Club
BFD_RELOC_AARCH64_BRANCH19 and BFD_RELOC_AARCH64_TSTBR14
cases with BFD_RELOC_AARCH64_JUMP26.
(elfNN_aarch64_check_relocs): Likewise.
ld * testsuite/ld-aarch64/aarch64-elf.exp: New test
emit-relocs-560.
* testsuite/ld-aarch64/emit-relocs-560.d: New file.
* testsuite/ld-aarch64/emit-relocs-560.s: New file.
* write.c (write_contents): Use bfd_get_filename rather than
accessing bfd->filename directly. Use bfd_section_name rather
than accessing section->name directly.
This is the 32-bit companion to
Remove unused ps_lgetLDT etc. on Solaris/x86 [PR25981]
https://sourceware.org/pipermail/gdb-patches/2020-May/168713.html
A 32-bit-default gdb fails to compile with the updated <sys/regset.h>.
While it is also affected by the lack of a GS definition, which the
compantion patch above fixes, it also fails to compile i386-sol2-nat.c like
this
/vol/src/gnu/gdb/hg/master/git/gdb/i386-sol2-nat.c:181:3: error: 'EAX' was not declared in this scope
181 | EAX, ECX, EDX, EBX,
| ^~~
and several more.
While this could be fixed by either including <ucontext.h> here or
provding fallback definitions of the register macros, I chose to do what
the 64-bit-default code in the same file
(amd64_sol2_gregset32_reg_offset[]) does, namely just hardcode the
numeric values instead. They are part of the ABI and thus guaranteed
not to change.
With this patch, a i386-pc-solaris2.11 configuration on master compiles
again, however, it doesn't work. However, I could successfully test it
on the gdb-9 branch.
Compiling and testing proved to be messy, unfortunately:
* For one, Solaris <sys/procfs.h> and largefile support used to be
mutually exclusive (fixed in Solaris 11.4 and Illumos), which was
exacerbated by the fact that g++ predefines _FILE_OFFSET_BITS=64 since
GCC 9.1.0. For now I've worked around this by adding
-U_FILE_OFFSET_BITS to CXXFLAGS and configuring with
--disable-largefile. I hope to clean this up in a future patch.
* gdb still defaults to startup-with-shell on. However, /bin/bash is a
64-bit executable which cannot be debugged by a 32-bit gdb. I hacked
around that part by pointing $SHELL at a 32-bit bash before running
make check.
PR build/25981
* i386-sol2-nat.c [PR_MODEL_NATIVE != PR_MODEL_LP64] (regmap):
Hardcode register numbers.
As reported in PR build/25981, a future Solaris 11.4 update will soon
remove the short i386 register names like SS etc. from <sys/regset.h>.
They could leak into user code (e.g. via <signal.h> -> <sys/signal.h> ->
<sys/ucontext.h>) and pollute the user namespace. Affected code would
have a hard time avoiding the issue: LLVM is one of those.
While the short names are required to be present by the i386 psABI, that
document only demands that they exist in <ucontext.h>, which is what the
upcoming update assures.
With this change, in a 64-bit-default configuration, procfs.c fails to
compile on Solaris/x86:
/vol/src/gnu/gdb/hg/master/git/gdb/procfs.c: In function 'ssd* procfs_find_LDT_entry(ptid_t)':
/vol/src/gnu/gdb/hg/master/git/gdb/procfs.c:1643:18: error: 'GS' was not declared in this scope
1643 | key = (*gregs)[GS] & 0xffff;
| ^~
make[2]: *** [Makefile:1607: procfs.o] Error 1
Initially I meant to provide a definition using the planned replacement
macro, but closer inspection revealed a better way. procfs_find_LDT_entry
and its helper proc_get_LDT_entry are only used to implement ps_lgetLDT,
one of the callback functions required by libthread_db.so.1
(cf. <proc_service.h>). While that function is still documented as being
required even in Solaris 11.4, I found that calls to it had been removed
long ago in Solaris 9, so just removing the three functions above is the
easiest fix.
The following patch does just that. It compiled successfully on
amd64-pc-solaris2.11, however, as reported in PR gdb/25939, master is
completely broken on Solaris since the multi-target patch. The patch
applies cleanly to the gdb-9 branch and there I could test it
successfully.
PR build/25981
* procfs.c [(__i386__ || __x86_64__) && sun] (proc_get_LDT_entry,
procfs_find_LDT_entry): Remove.
* procfs.h [(__i386__ || __x86_64__) && sun] (struct ssd,
procfs_find_LDT_entry): Remove.
* sol-thread.c [(__i386__ || __x86_64__) && sun] (ps_lgetLDT):
Remove.
When running test-case gdb.base/gdb-caching-proc.exp all passes are unique,
but fails might not be.
Fix this by using with_test_prefix.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2020-05-18 Tom de Vries <tdevries@suse.de>
* gdb.base/gdb-caching-proc.exp: Use with_test_prefix.
In git commit 806470a219 I swapped the order of internal vs. external
relocs memory allocation in ecoff_slurp_reloc_table, the idea being
that the external reloc size can be sanity checked against file size.
However, that fails badly with bfd_alloc memory where releasing any
block also releases all more recently allocated blocks.
* ecoff.c (ecoff_slurp_reloc_table): Malloc external_relocs so
they can be freed without also freeing internal_relocs.
sy_resolving ought to not be set for a struct local_symbol, but it is
apparent from local_symbol_make that the field is not initialised.
* symbols.c (resolve_symbol_value): Invoke LOCAL_SYMBOL_CHECK
before looking at add_symbol->sy_flags.sy_resolving.
In the following conditions:
- A target with hardware breakpoints available, and
- A target that uses software single stepping,
- An instruction at ADDRESS loops back to itself,
Now consider the following steps:
1. The user places a hardware breakpoint at ADDRESS (an instruction
that loops to itself),
2. The inferior runs and hits the breakpoint at ADDRESS,
3. The user tells GDB to 'continue'.
In #3 when the user tells GDB to continue, GDB first disables the
hardware breakpoint at ADDRESS, and then inserts a software
single-step breakpoint at ADDRESS. The original user-created
breakpoint was a hardware breakpoint, while the single-step breakpoint
will be a software breakpoint.
GDB continues and immediately hits the software single-step
breakpoint.
GDB then deletes the software single-step breakpoint by calling
delete_single_step_breakpoints, which eventually calls
delete_breakpoint, which, once the breakpoint (and its locations) are
deleted, calls update_global_location_list.
During update_global_location_list GDB spots that we have an old
location (the software single step breakpoint location) that is
inserted, but being deleted, and a location (the original hardware
breakpoint) at the same address which we are keeping, but which is not
currently inserted, GDB then calls breakpoint_locations_match on these
two locations.
Currently the locations do match, and so GDB calls swap_insertion
which swaps the "inserted" state of the two locations. The user
created hardware breakpoint is marked as inserted, while the GDB
internal software single step breakpoint is now marked as not
inserted. After this GDB returns through the call stack and leaves
delete_single_step_breakpoints.
After this GDB continues with its normal "stopping" process, as part
of this stopping process GDB removes all the breakpoints from the
target. Due to the swap it is now the user-created hardware
breakpoint that is marked as inserted, so it is this breakpoint GDB
tries to remove.
The problem is that GDB inserted the software single-step breakpoint
as a software breakpoint, but is now trying to remove the hardware
breakpoint. The problem is removing a software breakpoint is very
different to removing a hardware breakpoint, this could result is some
undetected undefined behaviour, or as in the original bug report (PR
gdb/25741), could result in the target throwing an error.
With "set breakpoint always-inserted on", we can easily reproduce this
against GDBserver. E.g.:
(gdb) hbreak main
Sending packet: $m400700,40#28...Packet received: 89e58b....
Sending packet: $m400736,1#fe...Packet received: 48
Hardware assisted breakpoint 1 at 0x400736: file threads.c, line 57.
Sending packet: $Z1,400736,1#48...Packet received: OK
Packet Z1 (hardware-breakpoint) is supported
(gdb) b main
Note: breakpoint 1 also set at pc 0x400736.
Sending packet: $m400736,1#fe...Packet received: 48
Breakpoint 2 at 0x400736: file threads.c, line 57.
(gdb) del
Delete all breakpoints? (y or n) y
Sending packet: $z0,400736,1#67...Packet received: E01
warning: Error removing breakpoint 2
This patch adds a testcase that does exactly that.
Trying to enhance GDB to handle this scenario while continuing to
avoid inserting redundant software and hardware breakpoints at the
same address turns out futile, because, given non-stop and breakpoints
always-inserted, if the user:
#1 - inserts a hw breakpoint, then
#2 - inserts a sw breakpoint at the same address, and then
#3 - removes the original hw breakpoint,
GDB would have to make sure to insert the sw breakpoint before
removing the hw breakpoint, to avoid running threads missing the
breakpoint. I.e., there's always going to be a window where a target
needs to be able to handle both sw and a hw breakpoints installed at
the same address. You can see more detailed description of that issue
here:
https://sourceware.org/pipermail/gdb-patches/2020-April/167738.html
So the fix here is to just stop considering software breakpoints and
hw breakpoints duplicates, and let GDB insert sw and hw breakpoints at
the same address.
The central change is to make breakpoint_locations_match consider the
location's type too. There are several other changes necessary to
actually make that that work correctly, however:
- We need to handle the duplicates detection better. Take a look at
the loop at the tail end of update_global_location_list. Currently,
because breakpoint locations aren't sorted by type, we can end up
with, at the same address, a sw break, then a hw break, then a sw
break, etc. The result is that that second sw break won't be
considered a duplicate of the first sw break. Seems like we already
handle that incorrectly for range breakpoints.
- The "set breakpoint auto-hw on" handling is moved out of
insert_bp_location to update_global_location_list, before the
duplicates determination.
Moving "set breakpoint auto-hw off" handling as well and downgrading
it to a warning+'disabling the location' was considered too, but in
the end discarded, because we want to error out with internal and
momentary breakpoints, like software single-step breakpoints.
Disabling such locations at update_global_location_list time would
make GDB lose control of the inferior.
- In update_breakpoint_locations, the logic of matching old locations
with new locations, in the have_ambiguous_names case, is updated to
still consider sw vs hw locations the same.
- Review all ALL_BP_LOCATIONS_AT_ADDR uses, and update those that
might need to be updated, and update comments for those that don't.
Note that that macro walks all locations at a given address, and
doesn't call breakpoint_locations_match.
The result against GDBserver (with "set breakpoint
condition-evaluation host" to avoid seeing confusing reinsertions) is:
(gdb) hbreak main
Sending packet: $m400736,1#fe...Packet received: 48
Hardware assisted breakpoint 1 at 0x400736: file main.c, line 57.
Sending packet: $Z1,400736,1#48...Packet received: OK
(gdb) b main
Note: breakpoint 1 also set at pc 0x400736.
Sending packet: $m400736,1#fe...Packet received: 48
Breakpoint 4 at 0x400736: file main.c, line 57.
Sending packet: $Z0,400736,1#47...Packet received: OK
(gdb) del 3
Sending packet: $z1,400736,1#68...Packet received: OK
gdb/ChangeLog:
2020-05-17 Pedro Alves <palves@redhat.com>
Andrew Burgess <andrew.burgess@embecosm.com>
Keno Fischer <keno@juliacomputing.com>
PR gdb/25741
* breakpoint.c (build_target_condition_list): Update comments.
(build_target_command_list): Update comments and skip matching
locations.
(insert_bp_location): Move "set breakpoint auto-hw on" handling to
a separate function. Simplify "set breakpoint auto-hw off"
handling.
(insert_breakpoints): Update comment.
(tracepoint_locations_match): New parameter. For breakpoints,
compare location types too, if the caller wants to.
(handle_automatic_hardware_breakpoints): New functions.
(bp_location_is_less_than): Also sort by location type and
hardware breakpoint length.
(update_global_location_list): Handle "set breakpoint auto-hw on"
here.
(update_breakpoint_locations): Ask breakpoint_locations_match to
ignore location types.
gdb/testsuite/ChangeLog:
2020-05-17 Pedro Alves <palves@redhat.com>
PR gdb/25741
* gdb.base/hw-sw-break-same-address.exp: New file.
The new exec-file-mismatch feature reveals that when running
gdb.multi/multi-re-run.exp against --target_board=native-gdbserver,
we've been starting gdbserver with the wrong program:
(gdb) spawn /home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/testsuite/../../gdbserver/gdbserver --once localhost:2347 /home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/te
stsuite/outputs/gdb.multi/multi-re-run/multi-re-run-2
Process /home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.multi/multi-re-run/multi-re-run-2 created; pid = 6280
Listening on port 2347
target remote localhost:2347
Remote debugging using localhost:2347
warning: Mismatch between current exec-file /home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.multi/multi-re-run/multi-re-run-1
and automatically determined exec-file /home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.multi/multi-re-run/multi-re-run-2
exec-file-mismatch handling is currently "ask"
Load new symbol table from "/home/pedro/brno/pedro/gdb/binutils-gdb/build/gdb/testsuite/outputs/gdb.multi/multi-re-run/multi-re-run-2"? (y or n) Quit
(gdb) FAIL: gdb.multi/multi-re-run.exp: re_run_inf=1: iter=1: running to all_started in runto
The problem is that gdb_reload uses the last loaded file as binary to
spawn, but we load the program for inferior 2 and then switch to
inferior 1 and run it, so the last loaded file is the program for
inferior 2.
Fix this by tweaking last_loaded_file.
gdb/testsuite/ChangeLog:
2020-05-16 Pedro Alves <palves@redhat.com>
* gdb.multi/multi-re-run.exp (test_re_run): Switch
LAST_LOADED_FILE accordingly.
Remove `TYPE_NAME`, changing all the call sites to use `type::name`
directly. This is quite a big diff, but this was mostly done using sed
and coccinelle. A few call sites were done by hand.
gdb/ChangeLog:
* gdbtypes.h (TYPE_NAME): Remove. Change all cal sites to use
type::name instead.
Tom de Vries
Alan Modra
GDB Administrator
Peter Bergner
Tom Tromey
Simon Marchi
Pedro Alves
Yoshinori Sato
Gunther Nikl
H.J. Lu
Stafford Horne
Alexander Fedotov
Siddhesh Poyarekar
Alan Modra
Rainer Orth
Nick Clifton
Douglas B Rupp
Sergei Trofimovich
Jaydeep Chauhan
liuhongt