Pedro Alves [Mon, 9 Oct 2017 17:11:01 +0000 (18:11 +0100)]
Multi-arch exec, more register reading avoidance
As mentioned in commit bf93d7ba9931 ("Add thread after updating
gdbarch when exec'ing"), we should avoid doing register reads after a
process does an exec and before we've updated that inferior's gdbarch.
Otherwise, we may interpret the registers using the wrong
architecture.
There's still (at least) one case where we still read registers
post-exec with the pre-exec architecture. That's when infrun decides
it needs to switch context to the exec'ing thread. I.e., if the exec
event is processed at a time when the current thread is not already
the exec'ing thread, then we get (with the test added by this commit):
continue
Continuing.
Truncated register 50 in remote 'g' packet
Truncated register 50 in remote 'g' packet
(gdb) FAIL: gdb.multi/multi-arch-exec.exp: selected_thread=2: follow_exec_mode=same: continue across exec that changes architecture
The fix is to avoid reading registers when switching context in this
case.
(I'd be nice to get rid of the constant stop_pc reading when switching
threads, but that'd be a deeper change.)
* gdb.multi/multi-arch-exec.c: Include <pthread.h> and <assert.h>.
(barrier): New.
(thread_start, all_started): New functions.
(main): Spawn new thread and wait until it is scheduled.
* gdb.multi/multi-arch-exec.exp: Build $srcfile1 with the pthreads
option.
(do_test): Add 'selected_thread' parameter. Run to all_started
instead of main. Explicitly set the breakpoint at main. Switch
to the SELECTED_THREAD thread.
(top level): Test handling the exec event with either the main
thread or the second thread selected.
John Baldwin [Mon, 9 Oct 2017 16:54:42 +0000 (09:54 -0700)]
Use gdbarch_long_bit to determine layout of FreeBSD siginfo_t.
FreeBSD architectures are either ILP32 or LP64 resulting in two
different layouts for siginfo_t. Previously, the 'bits_per_word'
member of bfd_arch_info was used to determine the layout to use for a
given FreeBSD architecture. However, mipsn32 architectures inherit
from a 64-bit mips architecture where bits_per_word is 64. As a
result, $_siginfo was not properly extracted from FreeBSD/mipsn32 core
dumps. Fix this by using gdbarch_long_bit instead of 'bits_per_word'
to determine if a FreeBSD architecture is ILP32 or LP64.
Simon Marchi [Mon, 9 Oct 2017 16:50:58 +0000 (12:50 -0400)]
Don't try building gdb against guile-2.2
GDB currently doesn't build with Guile 2.2 (see PR 21104). If one has
both Guile 2.2 and 2.0 installed, GDB will pick up Guile 2.2 first and
fail building. Until somebody does the work of adapting the GDB code to
Guile 2.2, we should not try using it. This patch therefore removes it
from configure.
* s390-opc.c (INSTR_SI_RD): New macro.
(INSTR_S_RD): Adjust example instruction.
* s390-opc.txt (lpsw, ssm, ts): Change S_RD instruction format to
SI_RD.
Tom Tromey [Sun, 8 Oct 2017 23:12:19 +0000 (17:12 -0600)]
Fix automatic dependency tracking
Commit f38307f5 changed COMPILE.post and POSTCOMPILE to remove
$(basename) from the dependency file name computation. However, it
did not update the `-include' at the end of the Makefile.in; this in
effect disabled automatic dependency tracking.
This patch restores the $(basename) wrapper so that the dependency
files are named "file.Po" rather than "file.o.Po".
I also tested the non-gcc3 dependency mode, which pointed out that
this case hadn't been working since the switch to C++. This is also
fixed in this patch.
* cp-support.c (cp_remove_params): Return a gdb::unique_xmalloc_ptr.
Use bool.
(overload_list_add_symbol): Adjust to use gdb::unique_xmalloc_ptr.
* cp-support.h (cp_remove_params): Now returns a
gdb::unique_xmalloc_ptr.
* dwarf2read.c (find_slot_in_mapped_hash): Now returns bool.
Adjust to cp_remove_params returning a gdb::unique_xmalloc_ptr.
* psymtab.c (psymtab_search_name): Adjust to cp_remove_params
returning a gdb::unique_xmalloc_ptr.
(lookup_partial_symbol): Adjust to use gdb::unique_xmalloc_ptr.
* stack.c (find_frame_funname): Adjust to cp_remove_params
returning a gdb::unique_xmalloc_ptr.
Pedro Alves [Mon, 9 Oct 2017 11:33:31 +0000 (12:33 +0100)]
Fix gdb.base/print-file-var-main.c value check logic
Fix a typo introduced in commit c56e7c4390ed ("Make ctxobj.exp and
print-file-var.exp work on all platforms.").
This doesn't really affect the outcome of the testcase. I only
noticed the typo because I stepped through the program manually.
To avoid such problems if the test is extended, this moves the STOP
marker until after the program self-validates the values. With the
typo in place, this alone would have resulted in a test FAIL. I.e.,
it'd have caught the typo.
* gdb.base/print-file-var-main.c: Fix get_version_2 value check
logic. Move STOP marker after the value checks.
* gdb.base/print-file-var.exp (continue to STOP marker): Tighten
regexp.
Tom Tromey [Sun, 8 Oct 2017 18:40:07 +0000 (12:40 -0600)]
Remove cleanup from frame_prepare_for_sniffer
Currently frame_prepare_for_sniffer returns a cleanup. This patch
changes it to return void, and exposes frame_cleanup_after_sniffer to
the caller.
Normally I would write an RAII class for this sort of thing; but
because there was just a single caller of frame_prepare_for_sniffer,
and because this caller is already using try/catch, I thought it
seemed ok to require explicit calls in this instance.
Tom Tromey [Sun, 8 Oct 2017 18:11:18 +0000 (12:11 -0600)]
Change search_symbols to return std::vector
This changes search_symbols to return a std::vector, replacing the
previous linked list approach. This allows the removal of some
cleanups, as well as the use of std::sort and std::unique, saving some
code and extra allocations in sort_search_symbols_remove_dups.
John Baldwin [Fri, 6 Oct 2017 18:41:45 +0000 (11:41 -0700)]
Account for padding in FreeBSD/mipsn32 NT_PRSTATUS notes.
Add a new ELF backend method to grok FreeBSD NT_PRSTATUS core dump
notes. Define a method for MIPS N32 to override the default
elfcore_grok_freebsd_prstatus that accounts for additional padding
between pr_pid and pr_reg that is not present in other 32-bit FreeBSD
platforms.
* elf-bfd.h (struct elf_backend_data): Add
`elf_backend_grok_freebsd_prstatus'.
* elf.c (elfcore_grok_freebsd_note): Call
`elf_backend_grok_freebsd_prstatus' to handle NT_PRSTATUS if
present.
* elfn32-mips.c (elf_n32_mips_grok_freebsd_prstatus): New
function.
(elf_backend_grok_freebsd_prstatus): Define.
* elfxx-target.h (elf_backend_grok_freebsd_prstatus): Define.
(elfNN_bed): Initialize `elf_backend_grok_freebsd_prstatus'.
Yao Qi [Fri, 6 Oct 2017 13:53:39 +0000 (14:53 +0100)]
Move aarch64-insn.o to arch/aarch64-insn.o and Remove a rule for arch/*.c
This patch moves aarch64-insn.o to arch/aarch64-insn.o. Then, all
arch/*.c are built to arch/*.o, so we don't need a Makefile rule to build
*.o from arch/*.c. This patch removes it too.
Yao Qi [Fri, 6 Oct 2017 10:18:48 +0000 (11:18 +0100)]
Share code updating gdb_target_obs
Nowadays, there are much duplications in configure.tgt to update
gdb_target_obs, some cpu specific object files are added to gdb_target_obs
to some different target triplets of the same cpu. The same problem
exists for os specific object files too. It is fragile to update them,
and build with all targets enabled doesn't find the problem.
This patch splits the gdb_target_obs update to three steps, cpu steps, os
steps, and the rest.
I tested this patch by build gdb for each different target triplets
respectively,
Yao Qi [Fri, 6 Oct 2017 10:13:30 +0000 (11:13 +0100)]
[RFC] Replicate src dir in build dir
Nowadays, GDB build tree is almost flat, but source tree isn't. We
have arch/ nat/ target/ common/ cli/ mi/ tui/ python/ guile/ directories.
We need to some rules in Makefile for source files in different source
directories, like,
# Rules for compiling .c files in the various source subdirectories.
%.o: ${srcdir}/arch/%.c
$(COMPILE) $<
$(POSTCOMPILE)
so we should take care of some special case that files' base name is the
same, like,
# Specify an explicit rule for gdb/common/agent.c, to avoid a clash with the
# object file generate by gdb/agent.c.
common-agent.o: $(srcdir)/common/agent.c
$(COMPILE) $(srcdir)/common/agent.c
$(POSTCOMPILE)
As we add more and more files in different directories, it becomes tricky
to name files, because we need take this into account.
This patch takes the first step toward "Replicate src dir in build dir",
that is, we create arch/ directory in buildtree, and put amd64.o there
as an example. Dependency tracking is updated for files with directory
name. Currently, when we build amd64.o,
"make clean" removes the object files, and "make distclean" removes .deps
additionally. configure file create .deps directory in each of
CONFIG_SRC_SUBDIR, and pass it to Makefile.in, so that "make clean" and
"make distclean" can remove stuffs there.
If people agree with this change, I'll add more directories to
CONFIG_SRC_SUBDIR.
H.J. Lu [Fri, 6 Oct 2017 07:35:13 +0000 (00:35 -0700)]
x86: Add POINTER_LOCAL_IFUNC_P/PLT_LOCAL_IFUNC_P
Add POINTER_LOCAL_IFUNC_P which returns TRUE for pointer reference to
local IFUNC symbol. Add PLT_LOCAL_IFUNC_P which returns TRUE for PLT
reference to local IFUNC symbol.
Alan Modra [Fri, 6 Oct 2017 03:44:21 +0000 (14:14 +1030)]
DWARF header checks
This patch tidies DWARF header checks, consolidating the "negative"
checks (which are really overflow checks) with the section size
check. In a number of cases this also ensures that small negative
lengths are caught. For instance
H.J. Lu [Fri, 6 Oct 2017 01:09:07 +0000 (18:09 -0700)]
powerpc: Dump dynamic relocation info to the map file
Dump dynamic relocation info to the map file when generating dynamic
relocation in read-only section relocations if -Map is used.
* elf32-ppc.c (readonly_dynrelocs): Add a link_info argument.
Dump dynamic relocation in read-only section with minfo if
needed.
(ppc_elf_adjust_dynamic_symbol): Pass NULL to readonly_dynrelocs.
(maybe_set_textrel): Likewise.
(ppc_elf_size_dynamic_sections): Dump dynamic relocation in
read-only section with minfo.
Pedro Alves [Thu, 5 Oct 2017 17:33:22 +0000 (18:33 +0100)]
Fix fork-related regressions on GNU/Linux
Commit 5cd63fda035d ("Fix "Remote 'g' packet reply is too long"
problems with multiple inferiors") caused a number of regressions on
native GNU/Linux, all related to follow-fork support. E.g.:
src/gdb/target.c:3141: internal-error: gdbarch* default_thread_architecture(target_ops*, ptid_t): Assertion `inf != NULL' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
Quit this debugging session? (y or n)
Resyncing due to internal error.
FAIL: gdb.base/catch-signal-fork.exp: got SIGHUP after fork (GDB internal error)
* linux-nat.c (linux_child_follow_fork): When following the parent
and detaching the child, consult the parent thread's architecture
instead of the child's.
Ulrich Weigand [Thu, 5 Oct 2017 17:14:08 +0000 (19:14 +0200)]
Clean up some DFP interfaces
This cleans up a number of interfaces in dfp.c / dfp.h. Specifically:
- The decimal_from_string / decimal_to_string routines are C++-ified
to operate on std::string instead of character buffers. In the
decimal_from_string, the boolean return value now actually is bool
instead of an int.
- The decimal_from_integral and decimal_from_doublest routines take
an struct value as input. This is not really appropriate at the low
level the DFP routines sit, so this replaced them with new routines
decimal_from_longest / decimal_from_ulongest / decimal_from_doublest
that operate on contents instead.
- To mirror the decimal_from_[u]longest, a new decimal_to_longest
routine is added as well, which can be used in unpack_long to
avoid an unnecessary conversion via DOUBLEST.
Note that the decimal_from_longest / decimal_from_ulongest routines
are actually more powerful than decimal_from_integral: the old routine
would only accept integer *types* of at most four bytes size, while
the new routines accept all integer *values* that fit in an [u]int32_t,
no matter which type they came from. The DFP tests are updated to
allow for this larger range of integers that can be converted.
* dfp.h (MAX_DECIMAL_STRING): Move to dfp.c.
(decimal_to_string): Return std::string object.
(decimal_from_string): Accept std::string object. Return bool.
(decimal_from_integral, decimal_from_doublest): Remove.
(decimal_from_longest): Add prototype.
(decimal_from_ulongest): Likewise.
(decimal_to_longest): Likewise.
(decimal_from_doublest): Likewise.
* dfp.c: Do not include "gdbtypes.h" or "value.h".
(MAX_DECIMAL_STRING): Move here.
(decimal_to_string): Return std::string object.
(decimal_from_string): Accept std::string object. Return bool.
(decimal_from_integral): Remove, replace by ...
(decimal_from_longest, decimal_from_ulongest): ... these new functions.
(decimal_to_longest): New function.
(decimal_from_floating): Remove, replace by ...
(decimal_from_doublest): ... this new function.
(decimal_to_doublest): Update to new decimal_to_string interface.
* value.c (unpack_long): Use decimal_to_longest.
* valops.c (value_cast): Use decimal_from_doublest instead of
decimal_from_floating. Use decimal_from_[u]longest isntead of
decimal_from_integral.
* valarith.c (value_args_as_decimal): Likewise.
* valprint.c (print_decimal_floating): Update to new
decimal_to_string interface.
* printcmd.c (printf_decfloat): Likewise.
* c-exp.y (parse_number): Update to new decimal_from_string interface.
Ulrich Weigand [Thu, 5 Oct 2017 17:12:45 +0000 (19:12 +0200)]
Clean up includes of doublest.h and floatformat.h
As a first small step to getting rid of doublest.h, this patch removes the
include of "floatformat.h" in "doublest.h". This is actually not needed
for the file itself. A few source files now need to include "floatformat.h"
directly, since they got it indirectly via "doublest.h" and still need it.
In reviewing which files need it, I found a number of files that include
"floatformat.h" directly without actually needing it at all. Similarly,
a number of files include "doublest.h" without needing it. I've also
removed those unnecessary include statements.
* doublest.h: Do not include "floatformat.h". Remove stale comments.
* gdbtypes.c: Include "floatformat.h".
* value.c: Likewise.
* m68k-tdep.c: Likewise.
John Baldwin [Thu, 5 Oct 2017 16:50:01 +0000 (09:50 -0700)]
Add a signal frame unwinder for FreeBSD/mipsn32.
The N32 signal frame uses an identical layout to N64, so reuse the N64
handler. The N32 signal trampoline does use one different instruction
relative to N64, so a separate tramp_frame is required.
John Baldwin [Thu, 5 Oct 2017 16:50:01 +0000 (09:50 -0700)]
Handle FreeBSD-specific AT_EHDRFLAGS and AT_HWCAP auxiliary vector types.
FreeBSD recently added two additional ELF auxiliary vectors. FreeBSD's
AT_HWCAP uses a different number compared to AT_HWCAP on Linux as the
numerical value was already in use for a different vector on FreeBSD.
Nick Clifton [Thu, 5 Oct 2017 15:17:22 +0000 (16:17 +0100)]
Fix the MSP430 assembler so that it detects and reports extraneous text at the end of operands.
PR 22133
* config/tc-msp430.c (parse_exp): Skip an 'h' suffix to constant
expressions.
(msp430_srcoperand): Check that the entire text was parsed by
parse_exp.
(msp430_operands): Likewise.
* testsuite/gas/msp430/pr22133.s: New test file.
* testsuite/gas/msp430/pr22133.d: New test driver.
* testsuite/gas/msp430/pr22133.s: Expected error output.
* testsuite/gas/msp430/msp430.exp: Run the new test.
Joseph Myers [Thu, 5 Oct 2017 14:19:39 +0000 (14:19 +0000)]
Avoid spurious readelf error status from decode_arm_unwind.
readelf.c:decode_arm_unwind has a variable res that is used as a
return value, with FALSE meaning unsuccessful and TRUE meaning
successful. This is initialized to FALSE (and then various code in
the function sets it to FALSE again on error), meaning that when the
function is successful, if it reaches returning res is still returns
FALSE, resulting eventually in exit status 1 from readelf without any
error message to indicate an error.
This patch fixes the initialization to use TRUE, so avoiding those
spurious errors. I don't have a self-contained test for this issue;
it was observed as many prelink tests failing without the patch and
passing with it.
* readelf.c (decode_arm_unwind): Initialize res to TRUE.
Alan Modra [Thu, 5 Oct 2017 07:02:18 +0000 (17:32 +1030)]
PR22239 - invalid memory read in display_debug_frames
Pointer comparisons have traps for the unwary. After adding a large
unknown value to "start", the test "start < end" depends on where
"start" is originally in memory.
PR 22239
* dwarf.c (read_cie): Don't compare "start" and "end" pointers
after adding a possibly wild length to "start", compare the length
to the difference of the pointers instead. Remove now redundant
"negative" length test.
Pedro Alves [Wed, 4 Oct 2017 13:20:51 +0000 (14:20 +0100)]
bfd_set_input_error
A downside to the 2017-10-04 PR22245 fix is that bfd_set_error can now
silently accept invalid errors if/when someone passes the a value of
the wrong enumeration type, which previously would be caught by the
-Wenum-conversion warning.
Alan Modra [Tue, 3 Oct 2017 23:23:53 +0000 (09:53 +1030)]
PR21167, relocation sections not included in groups
This fixes a wart I've known about for years, but haven't done
anything about because BFD treats relocation sections as an adjunct to
the section they relocate. SHF_GROUP on the section thus implicitly
applies to its relocation section(s), but it is an error that the
reloc sections aren't part of the group.
Like many patches to gas, this wasn't as straightforward as it could
be due to a number of backends, i386, cr16 and others, removing relocs
in tc_get_reloc rather than marking them as "done" earlier in
md_apply_reloc. So it isn't possible for the group support to
reliably detect the presence of relocs by looking at fixups earlier
than write_relocs. However the group support needs to create
signature symbols, and that must be done before the symbol table is
frozen, before write_relocs. So split off the group sizing from
elf_adjust_symtab and put it in elf_frob_file_after_relocs.
bfd/
PR 21167
* elf.c (_bfd_elf_setup_sections): Don't trim reloc sections from
groups.
(_bfd_elf_init_reloc_shdr): Pass sec_hdr, use it to copy SHF_GROUP
flag from section.
(elf_fake_sections): Adjust calls. Exit immediately on failure.
(bfd_elf_set_group_contents): Add associated reloc section indices
to group contents
gas/
PR 21167
* config/obj-elf.c (struct group_list): Delete elt_count.
(groups): New static.
(build_group_lists): Don't count elements.
(elf_adjust_symtab): Use groups rather than auto list. Set up
pointer from group member to SHT_GROUP section. Don't size
SHT_GROUP section or clean up here..
(elf_frob_file_after_relocs): ..do so here instead.
* testsuite/gas/arc/jli-1.d,
* testsuite/gas/elf/groupautob.d,
* testsuite/gas/mips/compact-eh-eb-2.d,
* testsuite/gas/mips/compact-eh-eb-5.d,
* testsuite/gas/mips/compact-eh-el-2.d,
* testsuite/gas/mips/compact-eh-el-5.d: Adjust.
ld/
PR 21167
* testsuite/ld-elf/group9b.d: Adjust for relocs included in group.
Pedro Alves [Wed, 4 Oct 2017 17:21:10 +0000 (18:21 +0100)]
Fix "Remote 'g' packet reply is too long" problems with multiple inferiors
When debugging two inferiors (or more) against gdbserver, and the
inferiors have different architectures, such as e.g., on x86_64
GNU/Linux and one inferior is 64-bit while the other is 32-bit, then
GDB can get confused with the different architectures in a couple
spots.
In both cases I ran into, GDB incorrectly ended up using the
architecture of whatever happens to be the selected inferior instead
of the architecture of some other given inferior:
#1 - When parsing the expedited registers in stop replies.
#2 - In the default implementation of the target_thread_architecture
target method.
These resulted in instances of the infamous "Remote 'g' packet reply
is too long" error. For example, with the test added in this commit,
we get:
~~~
Continuing.
Remote 'g' packet reply is too long (expected 440 bytes, got 816 bytes): ad064000000000000[snip]
(gdb) FAIL: gdb.multi/multi-arch.exp: inf1 event with inf2 selected: continue to hello_loop
c
Continuing.
Truncated register 50 in remote 'g' packet
(gdb) PASS: gdb.multi/multi-arch.exp: inf2 event with inf1 selected: c
~~~
* remote.c (get_remote_arch_state): New 'gdbarch' parameter. Use
it instead of target_gdbarch.
(get_remote_state, get_remote_packet_size): Adjust
get_remote_arch_state calls, passing down target_gdbarch
explicitly.
(packet_reg_from_regnum, packet_reg_from_pnum): New parameter
'gdbarch' and use it instead of target_gdbarch.
(get_memory_packet_size): Adjust get_remote_arch_state calls,
passing down target_gdbarch explicitly.
(struct stop_reply) <arch>: New field.
(remote_parse_stop_reply): Use the stopped thread's architecture,
not the current inferior's. Save the architecture in the
stop_reply.
(process_stop_reply): Use the stop reply's architecture.
(process_g_packet, remote_fetch_registers)
(remote_prepare_to_store, store_registers_using_G)
(remote_store_registers): Adjust get_remote_arch_state calls,
using the regcache's architecture.
(remote_get_trace_status): Adjust get_remote_arch_state calls,
passing down target_gdbarch explicitly.
* spu-multiarch.c (spu_thread_architecture): Defer to the target
beneath instead of calling target_gdbarch.
* target.c (default_thread_architecture): Use the specified
inferior's architecture, instead of the current inferior's
architecture (via target_gdbarch).
* gdb.multi/hangout.c: Include <unistd.h>.
(hangout_loop): New function.
(main): Call alarm. Call hangout_loop in a loop.
* gdb.multi/hello.c: Include <unistd.h>.
(hello_loop): New function.
(main): Call alarm. Call hangout_loop in a loop.
* gdb.multi/multi-arch.exp: Test running to a breakpoint one
inferior with the other selected.
This is because target_gdbarch is really just
current_inferior()->gdbarch, and it's wrong to return that
architecture when the inferior of the passed in PTID is NOT the
current inferior -- the inferior for PTID may be running a different
architecture. E.g., a mix of 64-bit and 32-bit inferiors in the same
debug session.
Doing that change above however exposes a problem in "maint print
registers", caught be the testsuite:
-PASS: gdb.base/maint.exp: maint print registers
+FAIL: gdb.base/maint.exp: maint print registers (GDB internal error)
...
gdb/inferior.c:309: internal-error: inferior* find_inferior_pid(int): Assertion `pid != 0' failed.
A problem internal to GDB has been detected,
The call stack looks like this:
#0 0x000000000068b707 in internal_error(char const*, int, char const*, ...) (file=0xa9b958 "gdb/inferior.c", line=309, fmt=0xa9b8e0 "%s: Assertion `%s' failed.") at gdb/common/errors.c:54
#1 0x00000000006e1c40 in find_inferior_pid(int) (pid=0) at gdb/inferior.c:309
#2 0x00000000006e1c8d in find_inferior_ptid(ptid_t) (ptid=...) at gdb/inferior.c:323
#3 0x00000000007c18dc in default_thread_architecture(target_ops*, ptid_t) (ops=0xf86d60 <dummy_target>, ptid=...)
at gdb/target.c:3134
#4 0x00000000007b5414 in delegate_thread_architecture(target_ops*, ptid_t) (self=0xf86d60 <dummy_target>, arg1=...)
at gdb/target-delegates.c:2527
#5 0x00000000007647b3 in get_thread_regcache(ptid_t) (ptid=...) at gdb/regcache.c:466
#6 0x00000000007647ff in get_current_regcache() () at gdb/regcache.c:475
#7 0x0000000000767495 in regcache_print(char const*, regcache_dump_what) (args=0x0, what_to_dump=regcache_dump_none)
at gdb/regcache.c:1599
#8 0x0000000000767550 in maintenance_print_registers(char const*, int) (args=0x0, from_tty=1)
at gdb/regcache.c:1613
I.e., the test does "maint print registers" while the inferior is not
running yet. This is expected to work, and there's already a hack in
get_thread_arch_regcache to make it work.
Instead of pilling on hacks in the internal of regcache and
target_ops, this commit moves the null_ptid special casing to where it
belongs -- higher up in the call chain in the implementation of "maint
print registers" & co directly.
/* For the benefit of "maint print registers" & co when debugging an
executable, allow dumping the regcache even when there is no
thread selected (target_thread_address_space internal-errors if
no address space is found). Note that normal user commands will
fail higher up on the call stack due to no
target_has_registers. */
aspace = (ptid_equal (null_ptid, ptid)
? NULL
: target_thread_address_space (ptid));
i.e., it'll no longer be possible to try to build a regcache for
null_ptid. That change alone would regress the gdbarch self tests
though, causing this:
(gdb) maintenance selftest
[...]
Running selftest register_to_value.
src/gdb/inferior.c:309: internal-error: inferior* find_inferior_pid(int): Assertion `pid != 0' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
Quit this debugging session? (y or n) FAIL: gdb.gdb/unittest.exp: maintenance selftest (GDB internal error)
The problem is that the way the mocking environment for those unit
tests is written is a bit fragile: it creates a special purpose
regcache (and sentinel's frame), using whatever is the current
inferior_ptid (usually null_ptid), and assumes get_current_regcache
will find that in the regcache::current_regcache list.
This commit changes the way the mock environment is created. It
eliminates the special regcache and frame and instead creates a fuller
mock environment, with a custom mock target_ops, and then a mock
inferior and thread "running" on that target.
If there's already a running target when you type "maint selftest",
then we error out, instead of pushing a new target on top of the
existing one (and thus killing the debug session). This results in:
(gdb) maint selftest
(...)
Self test failed: arch i386: target already pushed
Self test failed: arch i386:x86-64: target already pushed
Self test failed: arch i386:x64-32: target already pushed
Self test failed: arch i8086: target already pushed
Self test failed: arch i386:intel: target already pushed
Self test failed: arch i386:x86-64:intel: target already pushed
Self test failed: arch i386:x64-32:intel: target already pushed
Self test failed: arch i386:nacl: target already pushed
Self test failed: arch i386:x86-64:nacl: target already pushed
Self test failed: arch i386:x64-32:nacl: target already pushed
Self test failed: self-test failed at /home/pedro/gdb/mygit/src/gdb/selftest-arch.c:86
(...)
Ran 19 unit tests, 1 failed
I think that's OK, because self tests are really meant to be run from
a clean state right after GDB is started. I'm adding that erroring
out just as safe measure just in case someone types "maint selftest"
on the command line while already debugging something (as I've done
it).
(In my multi-target branch, where this patch originated from, we don't
actually need to error out, because there each inferior has its own
target stack).
Also, note that the current code was doing:
current_inferior()->gdbarch = gdbarch;
without taking care to restore the previous gdbarch. This means that
GDB's state was being left inconsistent after running the self tests,
further supporting the point that there's probably not much
expectation that mixing "maint selftests" and regular debugging in the
same GDB invocation really works. This patch fixes that, regardless.
changed the logic of the "-list-thread-groups --available" by mistake
when a pid is passed. It prints all the processes except the one
specified by the given pid. The correct behavior is to only print the
process corresponding to that pid. this patch fixes that and adds a test.
Pedro Alves [Wed, 4 Oct 2017 10:07:28 +0000 (11:07 +0100)]
Move code out of 'between TRY and CATCH'
I tried building GDB with TRY/CATCH mapped to raw C++ try/catch (by
defining GDB_XCPT to GDB_XCPT_RAW_TRY in
gdb/common/common-exceptions.h), and that caught a case of code
written between try and catch. This commit fixes it.
While we still have cleanups (i.e., make_cleanup & co), we must be
sure to add END_CATCH at the end of a TRY/CATCH/END_CATCH construct.
However, it's currently too easy to miss adding the END_CATCH, because
the code compiles anyway without it. I realized this when I noticed
that another patch I was working on missed several adding END_CATCH in
several new TRY/CATCH uses.
This commit fixes that by making TRY open a new scope that is only
closed by END_CATCH. This way, if you forget to add the END_CATCH,
then compilation fails due to the unbalanced curly braces.
This caught a couple places where we were missing END_CATCH in current
master, also fixed by the patch.
While we still have cleanups (i.e., make_cleanup & co), we must be
sure to add END_CATCH at the end of a TRY/CATCH/END_CATCH construct.
However, it's currently too easy to miss adding the END_CATCH, because
the code compiles anyway without it. I realized this when I noticed
that another patch I was working on missed several adding END_CATCH in
several new TRY/CATCH uses.
This commit fixes that by making TRY open a new scope that is only
closed by END_CATCH. This way, if you forget to add the END_CATCH,
then compilation fails due to the unbalanced curly braces.
This caught a couple places where we were missing END_CATCH in current
master, also fixed by the patch.
This is the "natural" extension necessary for the "set cwd" command
(and the whole "set the inferior's cwd" logic) to work on gdbserver.
The idea here is to have a new remote packet, QSetWorkingDir (name
adopted from LLDB's extension to the RSP, as can be seen at
<https://raw.githubusercontent.com/llvm-mirror/lldb/master/docs/lldb-gdb-remote.txt>),
which sends an hex-encoded string representing the working directory
that the remote inferior will use. There is a slight difference from
the packet proposed by LLDB: GDB's version will accept empty
arguments, meaning that the user wants to clear the previously set
working directory for the inferior (i.e., "set cwd" without arguments
on GDB).
For UNIX-like targets this feature is already implemented on
nat/fork-inferior.c, and all gdbserver has to do is to basically
implement "set_inferior_cwd" and call it whenever such packet arrives.
For other targets, like Windows, it is possible to use the existing
"get_inferior_cwd" function and do the necessary steps to make sure
that the inferior will use the specified working directory.
Aside from that, the patch consists basically of updates to the
testcase (making it available on remote targets) and the
documentation.
* NEWS (Changes since GDB 8.0): Add entry about new
'set-cwd-on-gdbserver' feature.
(New remote packets): Add entry for QSetWorkingDir.
* common/common-inferior.h (set_inferior_cwd): New prototype.
* infcmd.c (set_inferior_cwd): Remove "static".
(show_cwd_command): Expand text to include remote debugging.
* remote.c: Add PACKET_QSetWorkingDir.
(remote_protocol_features) <QSetWorkingDir>: New entry for
PACKET_QSetWorkingDir.
(extended_remote_set_inferior_cwd): New function.
(extended_remote_create_inferior): Call
"extended_remote_set_inferior_cwd".
(_initialize_remote): Call "add_packet_config_cmd" for
QSetWorkingDir.
* gdb.texinfo (Starting your Program) <The working directory.>:
Mention remote debugging.
(Working Directory) <Your Program's Working Directory>:
Likewise.
(Connecting) <Remote Packet>: Add "set-working-dir"
and "QSetWorkingDir" to the table.
(Remote Protocol) <QSetWorkingDir>: New item, explaining the
packet.
This commit adds new "set/show cwd" commands, which are used to
set/show the current working directory of the inferior that will be
started.
The idea here is that "set cwd" will become the de facto way of
setting the inferior's cwd. Currently, the user can use "cd" for
that, but there are side effects: with "cd", GDB also switches to
another directory, and that can impact the loading of scripts and
other files. With "set cwd", we separate the logic into a new
command.
To maintain backward compatibility, if the user issues a "cd" command
but doesn't use "set cwd", then the inferior's cwd will still be
changed according to what the user specified. However, "set cwd" has
precedence over "cd", so it can always be used to override it.
"set cwd" works in the following way:
- If the user sets the inferior's cwd by using "set cwd", then this
directory is saved into current_inferior ()->cwd and is used when
the inferior is started (see below).
- If the user doesn't set the inferior's cwd by using "set cwd", but
rather use the "cd" command as before, then this directory is
inherited by the inferior because GDB will have chdir'd into it.
On Unix-like hosts, the way the directory is changed before the
inferior execution is by expanding the user set directory before the
fork, and then "chdir" after the call to fork/vfork on
"fork_inferior", but before the actual execution. On Windows, the
inferior cwd set by the user is passed directly to the CreateProcess
call, which takes care of the actual chdir for us.
This way, we'll make sure that GDB's cwd is not affected by the user
set cwd.
* gdb.texinfo (Starting your Program) <The working directory.>:
Mention new "set cwd" command.
(Working Directory) <Your Program's Working Directory>:
Rephrase to explain that "set cwd" exists and is the default
way to change the inferior's cwd.
Currently, whenever we want to handle paths provided by the user and
perform tilde expansion on GDB, we rely on "tilde_expand", which comes
from readline. This was enough for our use cases so far, but the
situation will change when we start dealing with paths on gdbserver as
well, which is what the next patches implement.
Unfortunately it is not possible to use "tilde_expand" in this case
because gdbserver doesn't use readline. For that reason I decided to
implement a new "gdb_tilde_expand" function, which is basically a
wrapper for "glob" and its GNU extension, GLOB_TILDE_CHECK. With the
import of the "glob" module from gnulib, we're sure that "glob" always
supports this extension.
Passing enum as a first argument to variadic argument function
may lead to undefined behavior. The explanation on CERT site:
https://www.securecoding.cert.org/confluence/display/cplusplus/
EXP58-CPP.+Pass+an+object+of+the+correct+type+to+va_start
The bug was found by Kirill Nedostoev ([email protected])
when he tried to build GNU binutils with Clang 7.
PR 22245
* bfd.c (bfd_set_error): Avoid UB on passing arg to va_start that
undergoes default promotion.
* bfd-in2.h: Regenerate.
This removes set_batch_flag_and_make_cleanup_restore_page_info and
make_cleanup_restore_page_info in favor of a new RAII class. This
then allows for the removal of make_cleanup_restore_uinteger and
make_cleanup_restore_integer
projects / binutils.git / log