I tried debugging a dap test-case (gdb.dap/pause.exp), by replacing: ... @@ -661,13 +661,16 @@ quit (void) #else if (job_control /* If there is no terminal switching for this target, then we can't possibly get screwed by the lack of job control. */ || !target_supports_terminal_ours ()) - throw_quit ("Quit"); + __builtin_abort (); else throw_quit ("Quit (expect signal SIGINT when the program is resumed)"); #endif ... to try to produce a corefile. I noticed this didn't produce a core file, but it did mention recursive internal problems, so I decided to try a bit harder: ... @@ -347,7 +347,7 @@ internal_vproblem (struct internal_problem *problem, /* Don't allow infinite error/warning recursion. */ { static const char msg[] = "Recursive internal problem.\n"; - + __builtin_abort (); switch (dejavu) { case 0: ... and managed to produce a core file, due to a segfault. The segfault is due to running out of stack, and the stack loop looks like: ... gdb) #16321 0x00000000014f89f5 in internal_error_loc (file=0x160fac0 "/data/vries/gdb/src/gdb/ui-file.h", line=72, fmt=0x160faa4 "%s: write_async_safe") at /data/vries/gdb/src/gdbsupport/errors.cc:58 58 internal_verror (file, line, fmt, ap); (gdb) down #16320 0x0000000000d2433d in internal_verror (file=0x160fac0 "/data/vries/gdb/src/gdb/ui-file.h", line=72, fmt=0x160faa4 "%s: write_async_safe", ap=0x7ffc703be958) at /data/vries/gdb/src/gdb/utils.c:495 495 internal_vproblem (&internal_error_problem, file, line, fmt, ap); (gdb) #16319 0x0000000000d24307 in internal_vproblem(internal_problem *, const char *, int, const char *, typedef __va_list_tag __va_list_tag *) (problem=0x276b5e0 <internal_error_problem>, file=0x160fac0 "/data/vries/gdb/src/gdb/ui-file.h", line=72, fmt=0x160faa4 "%s: write_async_safe", ap=0x7ffc703be958) at /data/vries/gdb/src/gdb/utils.c:350 350 __builtin_abort (); (gdb) #16318 0x00007f1a9e6553e5 in abort () from /lib64/libc.so.6 (gdb) #16317 0x00007f1a9e653d2b in raise () from /lib64/libc.so.6 (gdb) #16316 <signal handler called> (gdb) #16315 0x00000000007a35eb in handle_fatal_signal (sig=6) at /data/vries/gdb/src/gdb/event-top.c:898 898 sig_write ("\n\n"); (gdb) #16314 0x00000000007a35b1 in <lambda(char const*)>::operator()(const char *) const (__closure=0x7ffc703bd8af, msg=0x15ed81c "\n\n") at /data/vries/gdb/src/gdb/event-top.c:893 893 gdb_stderr->write_async_safe (msg, strlen (msg)); (gdb) #16313 0x000000000082e644 in ui_file::write_async_safe (this=0x7ffc703c1970, buf=0x15ed81c "\n\n", length_buf=2) at /data/vries/gdb/src/gdb/ui-file.h:72 72 { gdb_assert_not_reached ("write_async_safe"); } (gdb) #16312 0x00000000014f89f5 in internal_error_loc (file=0x160fac0 "/data/vries/gdb/src/gdb/ui-file.h", line=72, fmt=0x160faa4 "%s: write_async_safe") at /data/vries/gdb/src/gdbsupport/errors.cc:58 58 internal_verror (file, line, fmt, ap); (gdb) ... AFAICT, what happens is: - abort is raised - abort is caught - attempt to write backtrace using sig_write - sigwrite does gdb_stderr->write_async_safe - since gdb_stderr is set to a string_file, which doesn't have write_async_safe an internal_error is thrown - the internal_error ends up calling the abort I added in internal_vproblem, and another abort is raised This can easily be avoided by printing to stderr instead: ... diff --git a/gdb/bt-utils.c b/gdb/bt-utils.c index 6f68e269c51..f93e45688e8 100644 --- a/gdb/bt-utils.c +++ b/gdb/bt-utils.c @@ -56,7 +56,7 @@ libbacktrace_error (void *data, const char *errmsg, int errnum) const auto sig_write = [] (const char *msg) -> void { - gdb_stderr->write_async_safe (msg, strlen (msg)); + fprintf (stderr, "%s", msg); }; sig_write ("error creating backtrace: "); @@ -80,7 +80,7 @@ libbacktrace_print (void *data, uintptr_t pc, const char *filename, { const auto sig_write = [] (const char *msg) -> void { - gdb_stderr->write_async_safe (msg, strlen (msg)); + fprintf (stderr,"%s", msg); }; /* Buffer to print addresses and line numbers into. An 8-byte address @@ -131,7 +131,7 @@ gdb_internal_backtrace_1 () { const auto sig_write = [] (const char *msg) -> void { - gdb_stderr->write_async_safe (msg, strlen (msg)); + fprintf (stderr, msg); };e-- /* Allow up to 25 frames of backtrace. */ @@ -159,7 +159,7 @@ gdb_internal_backtrace () #ifdef GDB_PRINT_INTERNAL_BACKTRACE const auto sig_write = [] (const char *msg) -> void { - gdb_stderr->write_async_safe (msg, strlen (msg)); + fprintf (stderr, "%s", msg); }; sig_write (_("----- Backtrace -----\n")); diff --git a/gdb/event-top.c b/gdb/event-top.c index 33aef7d7cc5..b3d16ecd710 100644 --- a/gdb/event-top.c +++ b/gdb/event-top.c @@ -890,7 +890,7 @@ handle_fatal_signal (int sig) #ifdef GDB_PRINT_INTERNAL_BACKTRACE const auto sig_write = [] (const char *msg) -> void { - gdb_stderr->write_async_safe (msg, strlen (msg)); + fprintf (stderr, "%s", msg); }; if (bt_on_fatal_signal) ... With this patch, I can get rid of the abort in internal_vproblem and still get my core dump. I don't know what is a proper fix for this.
I've come across this same bug while trying to debug a segfault from my frame_unwind move to classes[1] on 32-bit arm, no changes required to internal_vproblem or quit function. If you stop inside frame_unwind_legacy::sniffer and send a segfault to inner GDB, you can trigger the situation where internal_verror is writing to a gdb_stderr that doesn't have write_async_safe. [1] https://inbox.sourceware.org/gdb-patches/20240408201915.1482831-4-blarsen@redhat.com/T/#u
Just popping by to say that if any nullptr function pointer is called you'll get that to reproduce, no need to get any specific architecture, actually.