OK. I think I could come up with a fix. It mirrors what the x86 code does:
1. Let arch-dependent (arm) signal.h defines ARCH_RT_DELAYS_SIGNAL_SEND
(ifdef CONFIG_PREEMPT_RT_FULL)
2. This causes the vanilla force_sig_info() code to defer the actual
work until later, setting
the TIF_NOTIFY_RESUME flag. However, the existing code only tests
'in_atomic()' - which I replaced
by '(in_atomic() || irqs_disabled())'. Seems X86 does never call
this with IRQs disabled - otherwise
it would trigger a BUG message, too (since the check tests both
conditions).
3. I added a few lines to ARM's 'do_work_pending' which are analogous to
the code in X86's 'do_notify_resume()'.
The addition causes the deferred force_sig_info() to be executed at
a point where interrupts are
enabled.
Patch against vanilla 3.14.12 with rt9 preempt patch applied.
With this patch applied I no longer receive the BUG message.
HTH
-Till
I - very reproducibly - get this 'BUG' message
[ 6462.460032] Unhandled fault: external abort on non-linefetch (0x018) at
0xb6fdd000
[ 6462.460042] BUG: sleeping function called from invalid context at
kernel/locking/rtmutex.c:905
[ 6462.460049] in_atomic(): 0, irqs_disabled(): 128, pid: 1488, name: ldfilt
[ 6462.460053] no locks held by ldfilt/1488.
[ 6462.460057] irq event stamp: 1790
[ 6462.460081] hardirqs last enabled at (1789): [<c000ed10>]
no_work_pending+0x8/0x2c
[ 6462.460096] hardirqs last disabled at (1790): [<c05bf834>]
__dabt_usr+0x34/0x40
[ 6462.460116] softirqs last enabled at (0): [<c0021594>]
copy_process.part.50+0x498/0x170c
[ 6462.460124] softirqs last disabled at (0): [< (null)>] (null)
[ 6462.460135] CPU: 0 PID: 1488 Comm: ldfilt Tainted: G O
3.14.12-rt9-xilinx #25
[ 6462.460161] [<c0015f6c>] (unwind_backtrace) from [<c0012cc0>]
(show_stack+0x20/0x24)
[ 6462.460182] [<c0012cc0>] (show_stack) from [<c05ba9ac>]
(dump_stack+0x7c/0xcc)
[ 6462.460208] [<c05ba9ac>] (dump_stack) from [<c00574b8>]
(__might_sleep+0x1a0/0x1d8)
[ 6462.460225] [<c00574b8>] (__might_sleep) from [<c05bea40>]
(rt_spin_lock+0x30/0x64)
[ 6462.460240] [<c05bea40>] (rt_spin_lock) from [<c0036b44>]
(force_sig_info+0x38/0xe8)
[ 6462.460254] [<c0036b44>] (force_sig_info) from [<c00130c0>]
(arm_notify_die+0x50/0x60)
[ 6462.460266] [<c00130c0>] (arm_notify_die) from [<c000845c>]
(do_DataAbort+0x94/0xa8)
[ 6462.460280] [<c000845c>] (do_DataAbort) from [<c05bf83c>]
(__dabt_usr+0x3c/0x40)
[ 6462.460285] Exception stack(0xd2e65fb0 to 0xd2e65ff8)
[ 6462.460295] 5fa0: 0189d008 00000001
00001000 b6fdd000
[ 6462.460308] 5fc0: 00011cf0 b6fbc078 0189d008 00009530 00000000 be9b9ad0
ffffffff 00000000
[ 6462.460317] 5fe0: 00000000 be9b9a98 b6fa6c30 00008ad4 20000010 ffffffff
[ 6462.478073] Unhandled fault: external abort on non-linefetch (0x018) at
0xb6f2a000
on my CONFIG_PREEMPT_RT_FULL system:
#uname -a
Linux buildroot 3.14.12-rt9 #25 SMP PREEMPT RT Fri Nov 28 09:42:05 PST 2014
armv7l GNU/Linux
when accessing a mmapped, non-existing device from user-space.
I'm not an ARM expert but I suspect that when the exception is taken
interrupts are disabled and probably not re-enabled by the exception
handler (irqs_disabled(): 128).
arm_notify_die() calls force_sig_info() which may block (under RT_PREEMPT_FULL).
In 'force_sig_info()' we find
/*
* On some archs, PREEMPT_RT has to delay sending a signal from a trap
* since it can not enable preemption, and the signal code's spin_locks
* turn into mutexes. Instead, it must set TIF_NOTIFY_RESUME which will
* send the signal on exit of the trap.
*/
#ifdef ARCH_RT_DELAYS_SIGNAL_SEND
and if this CPP symbol is defined there is a codepath that
delays signal delivery and never blocks.
Perhaps the arm support should use this facility?
Unfortunately I'm not familiar enough with this CPU arch to propose
a fix.
Best regards
- Till
PS: Please CC me on any replies since I'm not a lkml subscriber; thanks.
--- linux-3.14.12/kernel/signal.c.orig 2014-12-02 18:50:24.472593199 -0800
+++ linux-3.14.12/kernel/signal.c 2014-12-02 18:56:35.581041811 -0800
@@ -1340,7 +1340,7 @@
* send the signal on exit of the trap.
*/
#ifdef ARCH_RT_DELAYS_SIGNAL_SEND
- if (in_atomic()) {
+ if (in_atomic() || irqs_disabled()) {
if (WARN_ON_ONCE(t != current))
return 0;
if (WARN_ON_ONCE(t->forced_info.si_signo))
--- linux-3.14.12/arch/arm/kernel/signal.c.orig 2014-12-02 18:30:13.275058413 -0800
+++ linux-3.14.12/arch/arm/kernel/signal.c 2014-12-02 18:33:24.727574197 -0800
@@ -592,6 +593,15 @@
}
syscall = 0;
} else {
+
+#ifdef ARCH_RT_DELAYS_SIGNAL_SEND
+ if (unlikely(current->forced_info.si_signo)) {
+ struct task_struct *t = current;
+ force_sig_info(t->forced_info.si_signo, &t->forced_info, t);
+ t->forced_info.si_signo = 0;
+ }
+#endif
+
clear_thread_flag(TIF_NOTIFY_RESUME);
tracehook_notify_resume(regs);
}
--- linux-3.14.12/arch/arm/include/asm/signal.h.orig 2014-12-02 18:28:12.469745395 -0800
+++ linux-3.14.12/arch/arm/include/asm/signal.h 2014-12-02 18:29:53.760172049 -0800
@@ -1,6 +1,20 @@
#ifndef _ASMARM_SIGNAL_H
#define _ASMARM_SIGNAL_H
+/*
+ * Because exceptions run with IRQs
+ * disabled while calling arm_notify_die(), but arm_notify_die() may call
+ * force_sig_info() which will grab the signal spin_locks for the
+ * task, which in PREEMPT_RT_FULL are mutexes. By defining
+ * ARCH_RT_DELAYS_SIGNAL_SEND the force_sig_info() will set
+ * TIF_NOTIFY_RESUME and set up the signal to be sent on exit of the
+ * trap.
+ */
+#if defined(CONFIG_PREEMPT_RT_FULL)
+#define ARCH_RT_DELAYS_SIGNAL_SEND
+#endif
+
+
#include <uapi/asm/signal.h>
/* Most things should be clean enough to redefine this at will, if care
