Fix PR19388: Can't access $_siginfo in breakpoint (catch signal) condition
authorPedro Alves <palves@redhat.com>
Wed, 13 Jan 2016 10:40:33 +0000 (10:40 +0000)
committerPedro Alves <palves@redhat.com>
Wed, 13 Jan 2016 10:40:33 +0000 (10:40 +0000)
commita911d87ad714cbfbbc5c5752cb8b445a7e70196c
treef7a48df5a040748238538d62ea9c2b0f51f41a43
parent8405419985b2a5f9c4f79b28af6c2324655b5c5b
Fix PR19388: Can't access $_siginfo in breakpoint (catch signal) condition

This commit merges both the registers and $_siginfo "thread
running/executing" checks into a single function.

Accessing $_siginfo from a "catch signal" breakpoint condition doesn't
work.  The condition always fails with "Selected thread is running":

 (gdb) catch signal
 Catchpoint 3 (standard signals)
 (gdb)
 condition $bpnum $_siginfo.si_signo == 5
 (gdb) continue
 Continuing.
 Error in testing breakpoint condition:
 Selected thread is running.

 Catchpoint 3 (signal SIGUSR1), 0x0000003615e35877 in __GI_raise (sig=10) at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
 56        return INLINE_SYSCALL (tgkill, 3, pid, selftid, sig);
 (gdb)

When accessing the $_siginfo object, we check whether the thread is
marked running (external/public) state and refuse the access if so.
This is so "print $_siginfo" at the prompt fails nicelly when the
current thread is running.  While evaluating breakpoint conditionals,
we haven't decided yet whether the thread is going to stop, so
is_running still returns true, and we thus always error out.

Evaluating an expression that requires registers access is really
conceptually the same -- we could think of $_siginfo as a pseudo
register.  However, in that case we check whether the thread is marked
executing (internal/private state), not running (external/public
state).  Changing the $_siginfo validation to check is_executing as
well fixes the bug in question.

Note that checking is_executing is not fully correct, not even for
registers.  See PR 19389.  However, I think this is the lesser of two
evils and ends up as an improvement.  We at least now have a single
place to fix.

Tested on x86_64 GNU/Linux.

gdb/ChangeLog:
2016-01-13  Pedro Alves  <palves@redhat.com>

PR breakpoints/19388
* frame.c (get_current_frame): Use validate_registers_access.
* gdbthread.h (validate_registers_access): Declare.
* infrun.c (validate_siginfo_access): Delete.
(siginfo_value_read, siginfo_value_write): Use
validate_registers_access.
* thread.c (validate_registers_access): New function.

gdb/testsuite/ChangeLog:
2016-01-13  Pedro Alves  <palves@redhat.com>

PR breakpoints/19388
* gdb.base/catch-signal-siginfo-cond.c: New file.
* gdb.base/catch-signal-siginfo-cond.exp: New file.
gdb/ChangeLog
gdb/frame.c
gdb/gdbthread.h
gdb/infrun.c
gdb/testsuite/ChangeLog
gdb/testsuite/gdb.base/catch-signal-siginfo-cond.c [new file with mode: 0644]
gdb/testsuite/gdb.base/catch-signal-siginfo-cond.exp [new file with mode: 0644]
gdb/thread.c