Delete two more symbol/section related macros. This time it's
SYMBOL_SECTION and MSYMBOL_SECTION.
As with general_symbol_info::m_name it is not currently possible to
make general_symbol_info::m_section private as general_symbol_info
must remain a POD type.
But other than failing to make the new m_section private, this change
does what you'd expect, adds a get and set member function and updates
all users to use the new functions instead of the previous wrapper
macros.
There should be no user visible change after this commit.
gdb/ChangeLog:
* coff-pe-read.c (add_pe_forwarded_sym): Make use of section_index
and set_section_index member functions where appropriate.
* coffread.c (coff_symtab_read): Likewise.
(process_coff_symbol): Likewise.
* ctfread.c (set_symbol_address): Likewise.
* dwarf2/read.c (add_partial_symbol): Likewise.
(var_decode_location): Likewise.
* language.c: Likewise.
* minsyms.c (minimal_symbol_reader::record_full): Likewise.
(compact_minimal_symbols): Likewise.
(minimal_symbol_upper_bound): Likewise.
* objfiles.c (relocate_one_symbol): Likewise.
* psympriv.h (partial_symbol::obj_section): Likewise.
(partial_symbol::address): Likewise.
* psymtab.c (partial_symtab::add_psymbol): Likewise.
* stabsread.c (scan_file_globals): Likewise.
* symmisc.c (dump_msymbols): Likewise.
* symtab.c (general_symbol_info::obj_section): Likewise.
(fixup_section): Likewise.
(get_msymbol_address): Likewise.
* symtab.h (general_symbol_info::section): Rename to...
(general_symbol_info::m_section): ...this.
(general_symbol_info::set_section_index): New member function.
(general_symbol_info::section_index): Likewise.
(SYMBOL_SECTION): Delete.
(MSYMBOL_VALUE_ADDRESS): Make use of section_index and
set_section_index member functions where appropriate.
(MSYMBOL_SECTION): Delete.
(symbol::symbol): Update to initialize 'm_section'.
* xcoffread.c (read_xcoff_symtab): Make use of set_section_index.
(process_xcoff_symbol): Likewise.
In commit cf2b207529 "[gdb/symtab] Fix element type modification in
read_array_type" I factored out new proc with_complaints out of proc
gdb_load_no_complaints, but when fixing a rebase conflict pre-commit I made a
mistake in gdb_load_no_complaints that is now causing:
...
ERROR: tcl error sourcing dw2-ranges-psym.exp.
ERROR: can't read "save": no such variable
while executing
"gdb_test_no_output "set complaints $save" """
(procedure "gdb_load_no_complaints" line 14)
invoked from within
"gdb_load_no_complaints $binfile"
...
Fix this by removing the offending line.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2021-02-10 Tom de Vries <tdevries@suse.de>
* lib/gdb.exp (gdb_load_no_complaints): Remove unnecessary
"Restore saved setting of complaints".
This fixes a typo in an error message in
stap_parse_argument_conditionally.
gdb/ChangeLog
2021-02-09 Tom Tromey <tom@tromey.com>
* stap-probe.c (stap_parse_argument_conditionally): Fix typo.
When running test-case gdb.fortran/function-calls.exp with target board
unix/gdb:debug_flags=-gdwarf-5, I run into:
...
(gdb) PASS: gdb.fortran/function-calls.exp: \
p derived_types_and_module_calls::pass_cart(c)
p derived_types_and_module_calls::pass_cart_nd(c_nd)^M
^M
Program received signal SIGSEGV, Segmentation fault.^M
0x0000000000400f73 in derived_types_and_module_calls::pass_cart_nd \
(c=<error reading variable: Cannot access memory at address 0xc>) at \
function-calls.f90:130^M
130 pass_cart_nd = ubound(c%d,1,4)^M
The program being debugged was signaled while in a function called from GDB.^M
GDB has restored the context to what it was before the call.^M
To change this behavior use "set unwindonsignal off".^M
Evaluation of the expression containing the function^M
(derived_types_and_module_calls::pass_cart_nd) will be abandoned.^M
(gdb) FAIL: gdb.fortran/function-calls.exp: p
...
The problem originates in read_array_type, when reading a DW_TAG_array_type
with a dwarf-5 DW_TAG_generic_subrange child. This is not supported, and the
fallout of this is that rather than constructing a new array type, the code
proceeds to modify the element type.
Fix this conservatively by issuing a complaint and bailing out in
read_array_type when not being able to construct an array type, such that we
have:
...
(gdb) maint expand-symtabs function-calls.f90^M
During symbol reading: unable to find array range \
- DIE at 0xe1e [in module function-calls]^M
During symbol reading: unable to find array range \
- DIE at 0xe1e [in module function-calls]^M
(gdb) KFAIL: gdb.fortran/function-calls.exp: no complaints in srcfile \
(PRMS: symtab/27388)
...
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-09 Tom de Vries <tdevries@suse.de>
PR symtab/27341
* dwarf2/read.c (read_array_type): Return NULL when not being able to
construct an array type. Add assert to ensure that element_type is
not being modified.
gdb/testsuite/ChangeLog:
2021-02-09 Tom de Vries <tdevries@suse.de>
PR symtab/27341
* lib/gdb.exp (with_complaints): New proc, factored out of ...
(gdb_load_no_complaints): ... here.
* gdb.fortran/function-calls.exp: Add test-case.
This reverts commit 82a1fd3a49.
It was pointed out:
https://sourceware.org/pipermail/gdb-patches/2021-February/175750.html
that commit 82a1fd3a49 caused GDB to have an unconditional
dependency on ELF specific parts of BFD. What this means is that if
GDB and BFD are built for a non-elf target then there will be
undefined symbol references within GDB.
The right solution isn't immediately obvious. So rather than rush a
fix in I'm reverting this commit for now, and will bring it back once
I have a good solution.
gdb/ChangeLog:
* gcore.c (struct gcore_collect_regset_section_cb_data): Delete.
(gcore_collect_regset_section_cb): Delete.
(gcore_collect_thread_registers): Delete.
(gcore_build_thread_register_notes): Delete.
(gcore_find_signalled_thread): Delete.
* gcore.h: Remove 'gdbsupport/gdb_signals.h' include and delete
'gdbarch' and 'thread_info' declarations.
(gcore_build_thread_register_notes): Delete declaration.
(gcore_find_signalled_thread): Likewise.
* fbsd-tdep.c: Remove 'gcore.h' include.
(struct fbsd_collect_regset_section_cb_data): New struct.
(fbsd_collect_regset_section_cb): New function.
(fbsd_collect_thread_registers): New function.
(struct fbsd_corefile_thread_data): New struct.
(fbsd_corefile_thread): New function.
(fbsd_make_corefile_notes): Call FreeBSD specific code.
* linux-tdep.c: Remove 'gcore.h' include.
(struct linux_collect_regset_section_cb_data): New struct.
(linux_collect_regset_section_cb): New function.
(linux_collect_thread_registers): New function.
(linux_corefile_thread): Call Linux specific code.
(find_signalled_thread): New function.
(linux_make_corefile_notes): Call find_signalled_thread.
While running tests on arm-none-eabi, I noticed following errors in some
gdb.threads tests.
ERROR: can't read "testfile": no such variable
These were being caused by ${testfile} being used before 'standard_testfile'
which sets it. This patch just moves standard_testfile before the use.
2021-02-09 Abid Qadeer <abidh@codesourcery.com>
gdb/testsuite/ChangeLog:
* gdb.threads/signal-command-handle-nopass.exp: Call
'standard_testfile' before using 'testfile'.
* gdb.threads/signal-command-multiple-signals-pending.exp: Likewise.
* gdb.threads/signal-delivered-right-thread.exp: Likewise
* gdb.threads/signal-sigtrap.exp: Likewise
With a certain Ada program, ada-lang.c:coerce_unspec_val_to_type can
cause a crash. This function may copy a value, and in the particular
case in the crash, the new value's type is smaller than the original
type. This causes coerce_unspec_val_to_type to create a lazy value --
but the original value is also not_lval, so later, when the value is
un-lazied, gdb asserts.
As with the previous patch, we believe there is a compiler bug here,
but it is difficult to reproduce, so we're not completely certain.
In the particular case we saw, the original value has record type, and
the record holds some variable-length arrays. This leads to the
type's length being 0. At the same time, the value is optimized out.
This patch changes coerce_unspec_val_to_type to handle an
optimized-out value correctly.
It also slightly restructures this code to avoid a crash should a
not_lval value wind up here. This is a purely defensive change.
This change also made it clear that value_contents_copy_raw can now be
made static, so that is also done.
gdb/ChangeLog
2021-02-09 Tom Tromey <tromey@adacore.com>
* ada-lang.c (coerce_unspec_val_to_type): Avoid making lazy
not_lval value.
* value.c (value_contents_copy_raw): Now static.
* value.h (value_contents_copy_raw): Don't declare.
resolve_dynamic_struct says:
gdb_assert (type->num_fields () > 0);
However, a certain Ada program has a structure with no fields but with
a dynamic size, causing this assertion to fire.
It is difficult to be certain, but we think this is a compiler bug.
However, in the meantime this assertion does not seem to be checking
any kind of internal consistency; so this patch removes it.
gdb/ChangeLog
2021-02-09 Tom Tromey <tromey@adacore.com>
* gdbtypes.c (resolve_dynamic_struct): Handle structure with no
fields.
The test expects the ifunc resolver to run lazily, at a later stage.
Depending on the distro and toolchain configuration, this is not the
case. Some configurations use non-lazy binding and thus the ifunc resolver
resolves all the ifunc references very early in the process startup, before
main.
Ubuntu is one such case. It has switched its toolchains to pass -Wl,z,now by
default, since 16.04. This wasn't a problem before 20.04 (at least for
aarch64) because the toolchains did not support ifunc's.
Forcing lazy binding makes the test run as expected, as opposed to the 80 or
so failures it showed before the change.
Tested on aarch64-linux/x86_64-linux Ubuntu 20.04.
gdb/testsuite:
2021-02-08 Luis Machado <luis.machado@linaro.org>
* gdb.base/gnu-ifunc.exp (build): Pass -Wl,z,lazy.
When running test-case gdb.dwarf2/enqueued-cu-base-addr.exp with target board
cc-with-dwz, I get:
...
gdb compile failed, dwz: enqueued-cu-base-addr: \
Couldn't find DIE at [100] referenced by DW_AT_type from DIE at [d8]
...
At 0xd8 we have DIE:
...
<1><d8>: Abbrev Number: 3 (DW_TAG_variable)
<d9> DW_AT_name : foo
<dd> DW_AT_type : <0x100>
<e1> DW_AT_const_value : 1
...
referring to:
...
<1><100>: Abbrev Number: 3 (DW_TAG_base_type)
<101> DW_AT_byte_size : 4
<102> DW_AT_encoding : 5 (signed)
<103> DW_AT_name : int
...
The reference is inter-CU, but the used abbrev uses DW_FORM_ref4:
...
3 DW_TAG_variable [no children]
DW_AT_name DW_FORM_string
DW_AT_type DW_FORM_ref4
DW_AT_const_value DW_FORM_sdata
DW_AT value: 0 DW_FORM value: 0
...
which is for intra-CU references.
Fix this by using a '%' instead of a ':' label prefix in the dwarf assembly.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2021-02-08 Tom de Vries <tdevries@suse.de>
* gdb.dwarf2/enqueued-cu-base-addr.exp: Fix inter-CU reference.
When stepping over thread-lock related codes (in uClibc), the inferior process
gets stuck and never manages to enter the critical section:
------8<-------
1 size_t fwrite(const void * __restrict ptr, size_t size,
2 size_t nmemb, register FILE * __restrict stream)
3 {
4 size_t retval;
5 __STDIO_AUTO_THREADLOCK_VAR;
6
7 > __STDIO_AUTO_THREADLOCK(stream);
8
9 retval = fwrite_unlocked(ptr, size, nmemb, stream);
10
11 __STDIO_AUTO_THREADUNLOCK(stream);
12
13 return retval;
14 }
------>8-------
Here, we are at line 7. Using the "next" command leads no where.
However, setting a breakpoint on line 9 and issuing "continue" works.
Looking at the assembly instructions reveals that we're dealing with the
critical section entry code [1] that should never be interrupted, in this
case by the debugger's implicit breakpoints:
------8<-------
...
1 add_s r0,r13,0x38
2 mov_s r3,1
3 llock r2,[r0] <-.
4 brne.nt r2,0,14 --. |
5 scond r3,[r0] | |
6 bne -10 --|--'
7 brne_s r2,0,84 <-'
...
------>8-------
Lines 3 until 5 (inclusive) are supposed to be executed atomically.
Therefore, GDB should never (implicitly) insert a breakpoint on lines
4 and 5, else the program will try to acquire the lock again by jumping
back to line 3 and gets stuck in an infinite loop.
The solution is to make GDB aware of these patterns so it inserts
breakpoints after the sequence -- line 6 in this example.
[1]
https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/tree/libc/sysdeps/linux/arc/bits/atomic.h#n46
------8<-------
({ \
__typeof(oldval) prev; \
\
__asm__ __volatile__( \
"1: llock %0, [%1] \n" \
" brne %0, %2, 2f \n" \
" scond %3, [%1] \n" \
" bnz 1b \n" \
"2: \n" \
: "=&r"(prev) \
: "r"(mem), "ir"(oldval), \
"r"(newval) /* can't be "ir". scond can't take limm for "b" */\
: "cc", "memory"); \
\
prev; \
})
------>8-------
"llock" (Load Locked) loads the 32-bit word pointed by the source
operand. If the load is completed without any interruption or
exception, the physical address is remembered, in Lock Physical Address
(LPA), and the Lock Flag (LF) is set to 1. LF is a non-architecturally
visible flag and is cleared whenever an interrupt or exception takes
place. LF is also cleared (atomically) whenever another process writes
to the LPA.
"scond" (Store Conditional) will write to the destination address if
and only if the LF is set to 1. When finished, with or without a write,
it atomically copies the LF value to ZF (Zero Flag).
These two instructions together provide the mechanism for entering a
critical section. The code snippet above comes from uClibc:
-----------------------
v3 (after Tom's remarks[2]):
handle_atomic_sequence()
- no need to initialize the std::vector with "{}"
- fix typo in comments: "conditial" -> "conditional"
- add braces to the body of "if" condition because of the comment line
arc_linux_software_single_step()
- make the performance slightly more efficient by moving a few
variables after the likely "return" point.
v2 (after Simon's remarks[3]):
- handle_atomic_sequence() gets a copy of an instruction instead of
a reference.
- handle_atomic_sequence() asserts if the given instruction is an llock.
[2]
https://sourceware.org/pipermail/gdb-patches/2021-February/175805.html
[3]
https://sourceware.org/pipermail/gdb-patches/2021-January/175487.html
gdb/ChangeLog:
PR tdep/27369
* arc-linux-tdep.c (handle_atomic_sequence): New.
(arc_linux_software_single_step): Call handle_atomic_sequence().
If the user implements a TUI window in Python, and this window
responds to GDB events and then redraws its window contents then there
is currently an edge case which can lead to problems.
The Python API documentation suggests that calling methods like erase
or write on a TUI window (from Python code) will raise an exception if
the window is not valid.
And the description for is_valid says:
This method returns True when this window is valid. When the user
changes the TUI layout, windows no longer visible in the new layout
will be destroyed. At this point, the gdb.TuiWindow will no longer
be valid, and methods (and attributes) other than is_valid will
throw an exception.
From this I, as a user, would expect that if I did 'tui disable' to
switch back to CLI mode, then the window would no longer be valid.
However, this is not the case.
When the TUI is disabled the windows in the TUI are not deleted, they
are simply hidden. As such, currently, the is_valid method continues
to return true.
This means that if the users Python code does something like:
def event_handler (e):
global tui_window_object
if tui_window_object->is_valid ():
tui_window_object->erase ()
tui_window_object->write ("Hello World")
gdb.events.stop.connect (event_handler)
Then when a stop event arrives GDB will try to draw the TUI window,
even when the TUI is disabled.
This exposes two bugs. First, is_valid should be returning false in
this case, second, if the user forgot to add the is_valid call, then I
believe the erase and write calls should be throwing an
exception (when the TUI is disabled).
The solution to both of these issues is I think bound together, as it
depends on having a working 'is_valid' check.
There's a rogue assert added into tui-layout.c as part of this
commit. While working on this commit I managed to break GDB such that
TUI_CMD_WIN was nullptr, this was causing GDB to abort. I'm leaving
the assert in as it might help people catch issues in the future.
This patch is inspired by the work done here:
https://sourceware.org/pipermail/gdb-patches/2020-December/174338.html
gdb/ChangeLog:
* python/py-tui.c (gdbpy_tui_window) <is_valid>: New member
function.
(REQUIRE_WINDOW): Call is_valid member function.
(REQUIRE_WINDOW_FOR_SETTER): New define.
(gdbpy_tui_is_valid): Call is_valid member function.
(gdbpy_tui_set_title): Call REQUIRE_WINDOW_FOR_SETTER instead.
* tui/tui-data.h (struct tui_win_info) <is_visible>: Check
tui_active too.
* tui/tui-layout.c (tui_apply_current_layout): Add an assert.
* tui/tui.c (tui_enable): Move setting of tui_active earlier in
the function.
gdb/doc/ChangeLog:
* python.texinfo (TUI Windows In Python): Extend description of
TuiWindow.is_valid.
gdb/testsuite/ChangeLog:
* gdb.python/tui-window-disabled.c: New file.
* gdb.python/tui-window-disabled.exp: New file.
* gdb.python/tui-window-disabled.py: New file.
There's a bug in the python tui API. If the user tries to delete the
window title attribute then this will trigger undefined behaviour in
GDB due to a missing nullptr check.
gdb/ChangeLog:
* python/py-tui.c (gdbpy_tui_set_title): Check that the new value
for the title is not nullptr.
gdb/testsuite/ChangeLog:
* gdb.python/tui-window.exp: Add new tests.
* gdb.python/tui-window.py (TestWindow) <__init__>: Store
TestWindow object into global the_window.
<remote_title>: New method.
(delete_window_title): New function.
This commit was inspired by this mailing list patch:
https://sourceware.org/pipermail/gdb-patches/2021-January/174713.html
Currently, calling tui_layout_window::apply will add the window from
the layout object to the global tui_windows list.
Unfortunately, when the user runs the 'winheight' command, this calls
tui_adjust_window_height, which calls the tui_layout_base::adjust_size
function, which can then call tui_layout_base::apply. The consequence
of this is that when the user does 'winheight' duplicate copies of a
window can be added to the global tui_windows list.
The original patch fixed this by changing the apply function to only
update the global list some of the time.
This patch takes a different approach. The apply function no longer
updates the global tui_windows list. Instead a new virtual function
is added to tui_layout_base which is used to gather all the currently
applied windows into a vector. Finally tui_apply_current_layout is
updated to make use of this new function to update the tui_windows
list.
The benefits I see in this approach are, (a) the apply function now no
longer touches global state, this solves the immediate problem,
and (b) now that tui_windows is updated directly in the function
tui_apply_current_layout, we can drop the saved_tui_windows global.
gdb/ChangeLog:
* tui-layout.c (saved_tui_windows): Delete.
(tui_apply_current_layout): Don't make use of saved_tui_windows,
call new get_windows member function instead.
(tui_get_window_by_name): Check in tui_windows.
(tui_layout_window::apply): Don't add to tui_windows.
* tui-layout.h (tui_layout_base::get_windows): New member function.
(tui_layout_window::get_windows): Likewise.
(tui_layout_split::get_windows): Likewise.
gdb/testsuite/ChangeLog:
* gdb.tui/winheight.exp: Add more tests.
In commit:
commit f237f998d1
Date: Mon Jan 25 18:43:19 2021 +0000
gdb/tui: remove special handling of locator/status window
I accidentally remove a call to delete the tui window objects. Now
every time GDB changes tui layout it is leaking windows.
gdb/ChangeLog:
* tui/tui-layout.c (tui_apply_current_layout): Restore the delete
of the window objects.
While working on another patch I noticed an oddly formatted error
message in the Python code.
When 'set python print-stack message' is in effect then consider this
Python script:
class TestCommand (gdb.Command):
def __init__ (self):
gdb.Command.__init__ (self, "test-cmd", gdb.COMMAND_DATA)
def invoke(self, args, from_tty):
raise RuntimeError ("bad")
TestCommand ()
And this GDB session:
(gdb) source path/to/python/script.py
(gdb) test-cmd
Python Exception <class 'RuntimeError'> bad:
Error occurred in Python: bad
The line 'Python Exception <class 'RuntimeError'> bad:' doesn't look
terrible in this situation, the colon at the end of the first line
makes sense given the second line.
However, there are places in GDB where there is no second line
printed, for example consider this python script:
def stop_listener (e):
raise RuntimeError ("bad")
gdb.events.stop.connect (stop_listener)
Then this GDB session:
(gdb) file helloworld.exe
(gdb) start
Temporary breakpoint 1 at 0x40112a: file hello.c, line 6.
Starting program: helloworld.exe
Temporary breakpoint 1, main () at hello.c:6
6 printf ("Hello World\n");
Python Exception <class 'RuntimeError'> bad:
(gdb) si
0x000000000040112f 6 printf ("Hello World\n");
Python Exception <class 'RuntimeError'> bad:
In this case there is no auxiliary information displayed after the
warning, and the line ending in the colon looks weird to me.
A quick survey of the code seems to indicate that it is not uncommon
for there to be no auxiliary information line printed, its not just
the one case I found above.
I propose that the line that currently looks like this:
Python Exception <class 'RuntimeError'> bad:
Be reformatted like this:
Python Exception <class 'RuntimeError'>: bad
I think this looks fine then in either situation. The first now looks
like this:
(gdb) test-cmd
Python Exception <class 'RuntimeError'>: bad
Error occurred in Python: bad
And the second like this:
(gdb) si
0x000000000040112f 6 printf ("Hello World\n");
Python Exception <class 'RuntimeError'>: bad
There's just two tests that needed updating. Errors are checked for
in many more tests, but most of the time the pattern doesn't care
about the colon.
gdb/ChangeLog:
* python/python.c (gdbpy_print_stack): Reformat an error message.
gdb/testsuite/ChangeLog:
* gdb.python/py-framefilter.exp: Update expected results.
* gdb.python/python.exp: Update expected results.
My initial goal was to fix our gdb/testsuite/lib/tuiterm.exp such that
it would correctly support (some limited) scrolling of the command
window.
What I observe is that when sending commands to the tui command window
in a test script with:
Term::command "p 1"
The command window would be left looking like this:
(gdb)
(gdb) p 1$1 = 1
(gdb)
When I would have expected it to look like this:
(gdb) p 1
$1 = 1
(gdb)
Obviously a bug in our tuiterm.exp library, right???
Wrong!
Turns out there's a bug in GDB.
If in GDB I enable the tui and then type (slowly) the 'p 1\r' (the \r
is pressing the return key at the end of the string), then you do
indeed get the "expected" terminal output.
However, if instead I copy the 'p 1\r' string and paste it into the
tui in one go then I now see the same corrupted output as we do when
using tuiterm.exp.
It turns out the problem is that GDB fails when handling lots of input
arriving quickly with a \r (or \n) on the end.
The reason for this bug is as follows:
When the tui is active the terminal is in no-echo mode, so characters
sent to the terminal are not echoed out again. This means that when
the user types \r, this is not echoed to the terminal.
The characters read in are passed to readline and \r indicates that
the command line is complete and ready to be processed. However, the
\r is not included in readlines command buffer, and is NOT printed by
readline when is displays its buffer to the screen.
So, in GDB we have to manually spot the \r when it is read in and
update the display. Printing a newline character to the output and
moving the cursor to the next line. This is done in tui_getc_1.
Now readline tries to reduce the number of write calls. So if we very
quickly (as in paste in one go) the text 'p 1' to readline (this time
with no \r on the end), then readline will fetch the fist character
and add it to its internal buffer. But before printing the character
out readline checks to see if there's more input incoming. As we
pasted multiple characters, then yes, readline sees the ' ' and adds
this to its buffer, and finally the '1', this too is added to the
buffer.
Now if at this point we take a break, readline sees there is no more
input available, and so prints its buffer out.
Now when we press \r the code in tui_getc_1 kicks in, adds a \n to the
output and moves the cursor to the next line.
But, if instead we paste 'p 1\r' in one go then readline adds 'p 1' to
its buffer as before, but now it sees that there is still more input
available. Now it fetches the '\r', but this triggers the newline
behaviour, we print '\n' and move to the next line - however readline
has not printed its buffer yet!
So finally we end up on the next line. There's no more input
available so readline prints its buffer, then GDB gets passed the
buffer, handles it, and prints the result.
The solution I think is to put of our special newline insertion code
until we know that readline has finished printing its buffer. Handily
we know when this is - the next thing readline does is pass us the
command line buffer for processing. So all we need to do is hook in
to the command line processing, and before we pass the command line to
GDB's internals we do all of the magic print a newline and move the
cursor to the next line stuff.
Luckily, GDB's interpreter mechanism already provides the hooks we
need to do this. So all I do here is move the newline printing code
from tui_getc_1 into a new function, setup a new input_handler hook
for the tui, and call my new newline printing function.
After this I can enable the tui and paste in 'p 1\r' and see the
correct output.
Also the tuiterm.exp library will now see non-corrupted output.
gdb/ChangeLog:
* tui/tui-interp.c (tui_command_line_handler): New function.
(tui_interp::resume): Register tui_command_line_handler as the
input_handler.
* tui/tui-io.c (tui_inject_newline_into_command_window): New
function.
(tui_getc_1): Delete handling of '\n' and '\r'.
* tui-io.h (tui_inject_newline_into_command_window): Declare.
gdb/testsuite/ChangeLog:
* gdb.tui/scroll.exp: Tighten expected results. Remove comment
about bug in GDB, update expected results, and add more tests.
The implementation of the delete line escape sequence in tuiterm.exp
was wrong. Delete should take a count and then delete COUNT lines at
the current cursor location, all remaining lines in the scroll region
are moved up to replace the deleted lines, with blank lines being
added at the end of the scroll region.
It's not clear to me what "scroll region" means here (or at least how
that is defined), but for now I'm just treating the whole screen as
the scroll region, which seems to work fine.
In contrast the current broken implementation deletes COUNT lines at
the cursor location moving the next COUNT lines up to fill the gap.
The rest of the screen is then cleared.
gdb/testsuite/ChangeLog:
* gdb.tui/scroll.exp: New file.
* gdb.tui/tui-layout-asm-short-prog.exp: Update expected results.
* lib/tuiterm.exp (Term::_csi_M): Delete count lines, scroll
remaining lines up.
(Term::check_region_contents): New proc.
(Term::check_box_contents): Use check_region_contents.
If the regs window is not big enough to show all registers, the
registers whose values changed are always drawn, even if they are not
in the currently visible area.
So this marks the invisible register sub windows with y=0, and skips
their rerender call in check_register_values.
gdb/ChangeLog:
2021-02-07 Hannes Domani <ssbssa@yahoo.de>
* tui/tui-regs.c (tui_data_window::display_registers_from):
Mark invisible register sub windows.
(tui_data_window::check_register_values): Ignore invisible
register sub windows.
Function n_spaces can't handle negative values, and returns an invalid
pointer in this case.
gdb/ChangeLog:
2021-02-07 Hannes Domani <ssbssa@yahoo.de>
* tui/tui-regs.c (tui_data_item_window::rerender): Don't call
n_spaces with a negative value.
Otherwise the register window is redrawn empty when scrolling or
changing its size with winheight.
gdb/ChangeLog:
2021-02-07 Hannes Domani <ssbssa@yahoo.de>
* tui/tui-regs.c (tui_data_window::display_registers_from):
Add refresh_window call.
The last frame in a corrupt stack stores the frame_id of the next frame,
so these two frames currently compare as equal.
So if you have a backtrace where the oldest frame is corrupt, this happens:
(gdb) py
>f = gdb.selected_frame()
>while f.older():
> f = f.older()
>print(f == f.newer())
>end
True
With this change, that same example returns False.
gdb/ChangeLog:
2021-02-07 Hannes Domani <ssbssa@yahoo.de>
* python/py-frame.c (frapy_richcompare): Compare frame_id_is_next.
These are likely not very useful, remove them.
gdb/ChangeLog:
* symmisc.c (std_in, std_out, std_err): Remove.
(_initialize_symmisc): Don't set std_in, std_out and std_err.
Change-Id: I140bfffd7fb655d39c32333bb53924b91b1eb13c
The test-case nextoverthrow.exp is failing on targets with unstripped libc.
This is a regression since commit 1940319c0e "[gdb] Fix internal-error in
process_event_stop_test".
The problem is that this code in create_exception_master_breakpoint:
...
for (objfile *sepdebug = obj->separate_debug_objfile;
sepdebug != nullptr; sepdebug = sepdebug->separate_debug_objfile)
if (create_exception_master_breakpoint_hook (sepdebug))
...
iterates over all the separate debug object files, but fails to handle the
case that obj itself has the debug info we're looking for.
Fix this by using the separate_debug_objfiles () range instead, which does
iterate both over obj and the obj->separate_debug_objfile chain.
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR breakpoints/27330
* breakpoint.c (create_exception_master_breakpoint): Handle case that
glibc object file has debug info.
When running test-case gdb.cp/cpexprs-debug-types.exp with target board
unix/gdb:debug_flags=-gdwarf-5, I run into:
...
(gdb) file cpexprs-debug-types^M
Reading symbols from cpexprs-debug-types...^M
ERROR: Couldn't load cpexprs-debug-types into GDB (eof).
ERROR: Couldn't send delete breakpoints to GDB.
ERROR: GDB process no longer exists
GDB process exited with wait status 23054 exp9 0 0 CHILDKILLED SIGABRT SIGABRT
...
We're running into this abort in process_psymtab_comp_unit:
...
switch (reader.comp_unit_die->tag)
{
case DW_TAG_compile_unit:
this_cu->unit_type = DW_UT_compile;
break;
case DW_TAG_partial_unit:
this_cu->unit_type = DW_UT_partial;
break;
default:
abort ();
}
...
because reader.comp_unit_die->tag == DW_TAG_type_unit.
Fix this by adding a DW_TAG_type_unit case.
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR symtab/27333
* dwarf2/read.c (process_psymtab_comp_unit): Handle DW_TAG_type_unit.
Using a hello world a.out, I run into a segfault:
...
$ gcc hello.c
$ gdb -batch a.out -ex "catch syscall -1" -ex r
Catchpoint 1 (syscall -1)
Aborted (core dumped)
...
Fix this by erroring out if a negative syscall number is used in the
catch syscall command.
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR breakpoints/27313
* break-catch-syscall.c (catch_syscall_split_args): Reject negative
syscall numbers.
gdb/testsuite/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR breakpoints/27313
* gdb.base/catch-syscall.exp: Check that "catch syscall -1" is
rejected.
I noticed that several parsers shared the same code to write a symbol
reference to an expression. This patch factors this code out into a
new function.
Regression tested on x86-64 Fedora 32.
gdb/ChangeLog
2021-02-05 Tom Tromey <tom@tromey.com>
* parser-defs.h (write_exp_symbol_reference): Declare.
* parse.c (write_exp_symbol_reference): New function.
* p-exp.y (variable): Use write_exp_symbol_reference.
* m2-exp.y (variable): Use write_exp_symbol_reference.
* f-exp.y (variable): Use write_exp_symbol_reference.
* d-exp.y (PrimaryExpression): Use write_exp_symbol_reference.
* c-exp.y (variable): Use write_exp_symbol_reference.
When an executable contains an index such as a .gdb_index or .debug_names
section, gdb ignores the DW_AT_subprogram attribute.
This problem has been filed as PR symtab/24549.
Add KFAILs for this PR in test-cases gdb.dwarf2/main-subprogram.exp and
gdb.fortran/mixed-lang-stack.exp.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
* gdb.dwarf2/main-subprogram.exp: Add KFAIL for PR symtab/24549.
* gdb.fortran/mixed-lang-stack.exp: Same.
I'm running into this assertion failure:
...
$ gdb -batch -ex "p (void *)0 - 5i"
gdbtypes.c:3430: internal-error: \
type* init_complex_type(const char*, type*): Assertion \
`target_type->code () == TYPE_CODE_INT \
|| target_type->code () == TYPE_CODE_FLT' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
...
This is a regression since commit c34e871466 "Implement complex arithmetic".
Before that commit we had:
...
(gdb) p (void *)0 - 5i
Argument to arithmetic operation not a number or boolean.
...
Fix this in complex_binop by throwing an error, such that we have:
...
(gdb) print (void *)0 - 5i
Argument to complex arithmetic operation not supported.
...
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR exp/27265
* valarith.c (complex_binop): Throw an error if complex type can't
be created.
gdb/testsuite/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR exp/27265
* gdb.base/complex-parts.exp: Add tests.
When running test-case gdb.dwarf2/clang-debug-names.exp, I run into the
following warning:
...
(gdb) file clang-debug-names^M
Reading symbols from clang-debug-names...^M
warning: Section .debug_aranges in clang-debug-names has duplicate \
debug_info_offset 0xc7, ignoring .debug_aranges.^M
...
This is caused by a missing return in commit 3ee6bb113a "[gdb/symtab] Fix
incomplete CU list assert in .debug_names".
Fix this by adding the missing return, such that we have instead this warning:
...
(gdb) file clang-debug-names^M
Reading symbols from clang-debug-names...^M
warning: Section .debug_aranges in clang-debug-names \
entry at offset 0 debug_info_offset 0 does not exists, \
ignoring .debug_aranges.^M
...
which is a known problem filed as PR25969 - "Ignoring .debug_aranges with
clang .debug_names".
Tested on x86_64-linux.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR symtab/27307
* dwarf2/read.c (create_cus_from_debug_names_list): Add missing
return.
gdb/testsuite/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
PR symtab/27307
* gdb.dwarf2/clang-debug-names.exp: Check file command warnings.
Fix indentation in !map.augmentation_is_gdb part of
create_cus_from_debug_names_list.
gdb/ChangeLog:
2021-02-05 Tom de Vries <tdevries@suse.de>
* dwarf2/read.c (create_cus_from_debug_names_list): Fix indentation.
tdesc-regs.exp test fails for ARC because the test is not
using the correct XML files as target description. With
this change, the correct directory and files are used.
v2 (after Andrew's remark [1]):
- Update the feature file names again. Test results now:
Test run by shahab on Tue Jan 26 11:31:16 2021
Target is arc-default-elf32
Host is x86_64-unknown-linux-gnu
=== gdb tests ===
Schedule of variations:
arc-nsim
Running target arc-nsim
Running /src/gdb/testsuite/gdb.xml/tdesc-regs.exp ...
PASS: gdb.xml/tdesc-regs.exp: set tdesc file single-reg.xml
PASS: gdb.xml/tdesc-regs.exp: cd to directory holding xml
PASS: gdb.xml/tdesc-regs.exp: set tdesc filename test-extra-regs.xml...
PASS: gdb.xml/tdesc-regs.exp: ptype $extrareg
PASS: gdb.xml/tdesc-regs.exp: ptype $uintreg
PASS: gdb.xml/tdesc-regs.exp: ptype $vecreg
PASS: gdb.xml/tdesc-regs.exp: ptype $unionreg
PASS: gdb.xml/tdesc-regs.exp: ptype $unionreg.v4
PASS: gdb.xml/tdesc-regs.exp: ptype $structreg
PASS: gdb.xml/tdesc-regs.exp: ptype $structreg.v4
PASS: gdb.xml/tdesc-regs.exp: ptype $bitfields
PASS: gdb.xml/tdesc-regs.exp: ptype $flags
PASS: gdb.xml/tdesc-regs.exp: ptype $mixed_flags
PASS: gdb.xml/tdesc-regs.exp: maintenance print reggroups
PASS: gdb.xml/tdesc-regs.exp: core-only.xml: set tdesc filename...
PASS: gdb.xml/tdesc-regs.exp: core-only.xml: ptype $extrareg
=== gdb Summary ===
# of expected passes 16
[1]
https://sourceware.org/pipermail/gdb-patches/2021-January/175465.html
gdb/testsuite/ChangeLog:
* gdb.xml/tdesc-regs.exp: Use correct core-regs for ARC.
For the same reason explained in the previous patch (which was for the
record-btrace target), move clearing of the async event handler of the
record-full target to the wait method.
I'm not sure if/where that target needs to re-set its async event
handler in the wait method. Since it only supports a single thread,
there probably can't be multiple events to report at the same time.
gdb/ChangeLog:
* record-full.c (record_full_async_inferior_event_handler):
Don't clear async event handler.
(record_full_base_target::wait): Clear async event handler at
beginning.
Change-Id: I146fbdb53d99e3a32766ac7cd337ac5ed7fd9adf
For the same reason explained in the previous patch (which was for the
remote target), move clearing of the async event handler of the
record-btrace target to the wait method.
The record-btrace target already re-sets its async event handler in its
wait method, so that part doesn't need to be changed:
/* In async mode, we need to announce further events. */
if (target_is_async_p ())
record_btrace_maybe_mark_async_event (moving, no_history);
gdb/ChangeLog:
* record-btrace.c (record_btrace_handle_async_inferior_event):
Don't clear async event handler.
(record_btrace_target::wait): Clear async event handler at
beginning.
Change-Id: Ib32087a81bf94f1b884a938c8167ac8bbe09e362
The remote target's remote_async_inferior_event_token is a flag that
tells when it wants the infrun loop to call its wait method. The flag
is cleared in the async_event_handler's callback
(remote_async_inferior_event_handler), just before calling
inferior_event_handler. However, since inferior_event_handler may
actually call another target's wait method, there needs to be code that
checks if we need to re-raise the flag.
It would be simpler instead for remote_target::wait to clear the flag
when it returns an event and there are no more to report after that. If
another target's wait method gets called by inferior_event_handler, the
remote target's flag will stay naturally stay marked.
Note that this is already partially implemented in remote_target::wait,
since the remote target may have multiple events to report (and it can
only report one at the time):
if (target_is_async_p ())
{
remote_state *rs = get_remote_state ();
/* If there are are events left in the queue tell the event loop
to return here. */
if (!rs->stop_reply_queue.empty ())
mark_async_event_handler (rs->remote_async_inferior_event_token);
}
The code in remote_async_inferior_event_handler also checks for pending
events as well, in addition to the stop reply queue, so I've made
remote_target::wait check for that as well. I'm not completely sure
this is ok, since I don't understand very well how the pending events
mechanism works. But I figured it was safer to do this, worst case it
just leads to unnecessary calls to remote_target::wait.
gdb/ChangeLog:
* remote.c (remote_target::wait): Clear async event handler at
beginning, mark if needed at the end.
(remote_async_inferior_event_handler): Don't set or clear async
event handler.
Change-Id: I20117f5b5acc8a9972c90f16280249b766c1bf37
The `ready` flag of async event handlers is cleared by the async event
handler system right before invoking the associated callback, in
check_async_event_handlers.
This is not ideal with how the infrun subsystem consumes events: all
targets' async event handler callbacks essentially just invoke
`inferior_event_handler`, which eventually calls `fetch_inferior_event`
and `do_target_wait`. `do_target_wait` picks an inferior at random,
and thus a target at random (it could be the target whose `ready` flag
was cleared, or not), and pulls one event from it.
So it's possible that:
- the async event handler for a target A is called
- we end up consuming an event for target B
- all threads of target B are stopped, target_async(0) is called on it,
so its async event handler is cleared (e.g.
record_btrace_target::async)
As a result, target A still has events to report while its async event
handler is left unmarked, so these events are not consumed. To counter
this, at the end of their async event handler callbacks, targets check
if they still have something to report and re-mark their async event
handler (e.g. remote_async_inferior_event_handler).
The linux_nat target does not suffer from this because it doesn't use an
async event handler at the moment. It only uses a pipe registered with
the event loop. It is written to in the SIGCHLD handler (and in other
spots that want to get target wait method called) and read from in
the target's wait method. So if linux_nat happened to be target A in
the example above, the pipe would just stay readable, and the event loop
would wake up again, until linux_nat's wait method is finally called and
consumes the contents of the pipe.
I think it would be nicer if targets using async_event_handler worked in
a similar way, where the flag would stay set until the target's wait
method is actually called. As a first step towards that, this patch
moves the responsibility of clearing the ready flags of async event
handlers to the invoked callback.
All async event handler callbacks are modified to clear their ready flag
before doing anything else. So in practice, nothing changes with this
patch. It's only the responsibility of clearing the flag that is
shifted toward the callee.
gdb/ChangeLog:
* async-event.h (async_event_handler_func): Add documentation.
* async-event.c (check_async_event_handlers): Don't clear
async_event_handler ready flag.
* infrun.c (infrun_async_inferior_event_handler): Clear ready
flag.
* record-btrace.c (record_btrace_handle_async_inferior_event):
Likewise.
* record-full.c (record_full_async_inferior_event_handler):
Likewise.
* remote-notif.c (remote_async_get_pending_events_handler):
Likewise.
* remote.c (remote_async_inferior_event_handler): Likewise.
Change-Id: I179ef8e99580eae642d332846fd13664dbddc0c1
Moving it to an inner scope makes it clearer where it's used (only while
handling the TARGET_WAITKIND_LOADED event).
gdb/ChangeLog:
* infrun.c (handle_inferior_event): Move stop_soon variable to
inner scope.
Change-Id: Ic57685a21714cfbb38f1487ee96cea1d12b44652
This adds a testcase that exercises detaching while GDB is stepping
over a breakpoint, in all combinations of:
- maint target non-stop off/on
- set non-stop on/off
- displaced stepping on/off
This exercises the bugs fixed in the previous 8 patches.
gdb/testsuite/ChangeLog:
* gdb.threads/detach-step-over.c: New file.
* gdb.threads/detach-step-over.exp: New file.
A following patch will add a testcase that has a number of threads
constantly stepping over a breakpoint, and then has GDB detach the
process, while threads are running. If we have more than one inferior
running, and we detach from just one of the inferiors, we expect that
the remaining inferior continues running. However, in all-stop, if
GDB needs to pause the target for the detach, nothing is re-resuming
the other inferiors after the detach. "info threads" shows the
threads as running, but they really aren't. This fixes it.
gdb/ChangeLog:
* infcmd.c (detach_command): Hold strong reference to target, and
if all-stop on entry, restart threads on exit.
* infrun.c (switch_back_to_stepped_thread): Factor out bits to ...
(restart_stepped_thread): ... this new function. Also handle
trap_expected.
(restart_after_all_stop_detach): New function.
* infrun.h (restart_after_all_stop_detach): Declare.
A following patch will add a testcase that has a number of threads
constantly stepping over a breakpoint, and then has GDB detach the
process. That testcase exercises both "set displaced-stepping
on/off". Testing with "set displaced-stepping off" reveals that GDB
does not handle the case of the user typing "detach" just while some
thread is in the middle of an in-line step over. If that thread
belongs to the inferior that is being detached, then the step-over
never finishes, and threads of other inferiors are never re-resumed.
This fixes it.
gdb/ChangeLog:
* infrun.c (struct step_over_info): Initialize fields.
(prepare_for_detach): Handle ongoing in-line step over.
A following patch will add a testcase that has a number of threads
constantly stepping over a breakpoint, and then has GDB detach the
process. That testcase sometimes fails with the inferior crashing
with SIGTRAP after the detach because of the bug fixed by this patch,
when tested with the native target.
The problem is that target_detach removes breakpoints from the target
immediately, and that does not work with the native GNU/Linux target
(and probably no other native target) currently. The test wouldn't
fail with this issue when testing against gdbserver, because gdbserver
does allow accessing memory while the current thread is running, by
transparently pausing all threads temporarily, without GDB noticing.
Implementing that in gdbserver was a lot of work, so I'm not looking
forward right now to do the same in the native target. Instead, I
came up with a simpler solution -- push the breakpoints removal down
to the targets. The Linux target conveniently already pauses all
threads before detaching them, since PTRACE_DETACH only works with
stopped threads, so we move removing breakpoints to after that. Only
the remote and GNU/Linux targets support support async execution, so
no other target should really need this.
gdb/ChangeLog:
* linux-nat.c (linux_nat_target::detach): Remove breakpoints
here...
* remote.c (remote_target::remote_detach_1): ... and here ...
* target.c (target_detach): ... instead of here.
* target.h (target_ops::detach): Add comment.
I noticed that "detach" while a program was running sometimes resulted
in the process crashing. I tracked it down to this change to
prepare_for_detach in commit 187b041e ("gdb: move displaced stepping
logic to gdbarch, allow starting concurrent displaced steps"):
/* Is any thread of this process displaced stepping? If not,
there's nothing else to do. */
- if (displaced->step_thread == nullptr)
+ if (displaced_step_in_progress (inf))
return;
The problem above is that the condition was inadvertently flipped. It
should have been:
if (!displaced_step_in_progress (inf))
So I fixed it, and wrote a testcase to exercise it. The testcase has
a number of threads constantly stepping over a breakpoint, and then
GDB detaches the process, while threads are running and stepping over
the breakpoint. And then I was surprised that my testcase would hang
-- GDB would get stuck in an infinite loop in prepare_for_detach,
here:
while (displaced_step_in_progress (inf))
{
...
What is going on is that since we now have two displaced stepping
buffers, as one displaced step finishes, GDB starts another, and
there's another one already in progress, and on and on, so the
displaced_step_in_progress condition never turns false. This happens
because we go via the whole handle_inferior_event, which tries to
start new step overs when one finishes. And also because while we
remove breakpoints from the target before prepare_for_detach is
called, handle_inferior_event ends up calling insert_breakpoints via
e.g. keep_going.
Thinking through all this, I came to the conclusion that going through
the whole handle_inferior_event isn't ideal. A _lot_ is done by that
function, e.g., some thread may get a signal which is passed to the
inferior, and gdb decides to try to get over the signal handler, which
reinstalls breakpoints. Or some process may exit. We can end up
reporting these events via normal_stop while detaching, maybe end up
running some breakpoint commands, or maybe even something runs an
inferior function call. Etc. All this after the user has already
declared they don't want to debug the process anymore, by asking to
detach.
I came to the conclusion that it's better to do the minimal amount of
work possible, in a more controlled fashion, without going through
handle_inferior_event. So in the new approach implemented by this
patch, if there are threads of the inferior that we're detaching in
the middle of a displaced step, stop them, and cancel the displaced
step. This is basically what stop_all_threads already does, via
wait_one and (the now factored out) handle_one, so I'm reusing those.
gdb/ChangeLog:
* infrun.c (struct wait_one_event): Move higher up.
(prepare_for_detach): Abort in-progress displaced steps instead of
letting them complete.
(handle_one): If the inferior is detaching, don't add the thread
back to the global step-over chain.
(restart_threads): Don't restart threads if detaching.
(handle_signal_stop): Remove inferior::detaching reference.
After detaching from a process, the inf->detaching flag is
inadvertently left set to true. If you afterwards reuse the same
inferior to start a new process, GDB will mishave...
The problem is that prepare_for_detach discards the scoped_restore at
the end, while the intention is for the flag to be set only for the
duration of prepare_for_detach.
This was already a bug in the original commit that added
prepare_for_detach, commit 24291992da ("PR gdb/11321"), by yours
truly. Back then, we still used cleanups, and the function called
discard_cleanups instead of do_cleanups, by mistake.
gdb/ChangeLog:
* infrun.c (prepare_for_detach): Don't release scoped_restore
before returning.
This moves the code handling an event out of wait_one to a separate
function, to be used in another context in a following patch.
gdb/ChangeLog:
* infrun.c (handle_one): New function, factored out from ...
(stop_all_threads): ... here.
A following patch will add a testcase that has two processes with
threads stepping over a breakpoint continuously, and then detaches
from one of the processes while threads are running. The other
process continues stepping over its breakpoint. And then the testcase
sends a SIGUSR1, expecting that GDB reports it. That would sometimes
hang against gdbserver, due to the bugs fixed here. Both bugs are
related, in that they're about remote protocol asynchronous Stop
notifications. There's a bug in GDB, and another in GDBserver.
The GDB bug:
- when we detach from a process, the remote target discards any
pending RSP notification related to that process, including the
in-flight, yet-unacked notification. Discarding the in-flight
notification is the problem. Until the in-flight notification is
acked with a vStopped packet, the server won't send another %Stop
notification. As a result, the debug session gets messed up. In
the new testcase's case, GDB would hang inside stop_all_threads,
waiting for a stop for one of the process'es threads, which never
arrived -- its stop reply was permanently stuck in the stop reply
queue, waiting for a vStopped packet that never arrived.
In summary:
1. GDBserver sends stop notification about thread X, the remote
target receives it and stores it
2. At the same time, GDB detaches thread X's inferior
3. The remote target discards the received stop notification
4. GDBserver waits forever for the ack
The GDBserver bug:
GDBserver has the opposite bug. It also discards notifications for
the process being detached. If that discards the head of the
notification queue, when gdb sends an ack, it ends up acking the
_next_ notification. Meaning, gdb loses one notification. In the
testcase, this results in a similar hang in stop_all_threads.
So we have two very similar bugs in GDB and GDBserver, both resulting
in a similar symptom. That's why I'm fixing them both at the same
time.
gdb/ChangeLog:
* remote.c (remote_notif_stop_ack): Don't error out on
TARGET_WAITKIND_IGNORE; instead, just ignore the notification.
(remote_target::discard_pending_stop_replies): Don't delete
in-flight notification; instead, clear its contents.
gdbserver/ChangeLog:
* server.cc (discard_queued_stop_replies): Don't ever discard the
notification at the head of the list.
