Skip to contentSkip to conversation

xtensa: fix broken TIF_NOTIFY_SIGNAL assembly

PATCH In review Max Filippov 2 participants 4 messages Nov 11, 2020 – Nov 12, 2020 Tensilica Xtensa port (xtensa)

lore
Share & tools
lore

Max Filippov

author maintainer
#0 lore raw

PATCH v2 1/1 xtensa: fix broken TIF_NOTIFY_SIGNAL assembly

TIF_NOTIFY_SIGNAL handling in xtensa assembly is implemented
incorrectly: there should be a call to do_notify_resume when either TIF_SIGPENDING, TIF_NOTIFY_RESUME or TIF_NOTIFY_SIGNAL bit is set in the thread_info::flags. The straightforward way to do it would be

_bbsi.l a4, TIF_NEED_RESCHED, 3f
_bbsi.l a4, TIF_NOTIFY_RESUME, 2f
_bbsi.l a4, TIF_NOTIFY_SIGNAL, 2f
_bbci.l a4, TIF_SIGPENDING, 5f

Optimize it a little bit and use bit mask and bnone opcode to skip do_notify_resume invocation. Shuffle _TIF_* flags a bit so that used bit mask fits into the immediate field of movi opcode.

Fixes 4c6a9dcd4d13 ("xtensa: add support for TIF_NOTIFY_SIGNAL")
Cc Jens Axboe <[email protected]>

---
Changes v1->v2:
- use more canonical order of bnone arguments: tested value first, bit mask second.

 arch/xtensa/include/asm/thread_info.h | 7 ++++---
 arch/xtensa/kernel/entry.S            | 5 ++---
 2 files changed, 6 insertions(+), 6 deletions(-)
arch/xtensa/include/asm/thread_info.h +4 -3
index 6ea521b8e2ec..a312333a9add 100644
@@ -111,22 +111,23 @@ static inline struct thread_info *current_thread_info(void)
 #define TIF_NEED_RESCHED	2	/* rescheduling necessary */
 #define TIF_SINGLESTEP		3	/* restore singlestep on return to user mode */
 #define TIF_SYSCALL_TRACEPOINT	4	/* syscall tracepoint instrumentation */
-#define TIF_MEMDIE		5	/* is terminating due to OOM killer */
+#define TIF_NOTIFY_SIGNAL	5	/* signal notifications exist */
 #define TIF_RESTORE_SIGMASK	6	/* restore signal mask in do_signal() */
 #define TIF_NOTIFY_RESUME	7	/* callback before returning to user */
 #define TIF_DB_DISABLED		8	/* debug trap disabled for syscall */
 #define TIF_SYSCALL_AUDIT	9	/* syscall auditing active */
 #define TIF_SECCOMP		10	/* secure computing */
-#define TIF_NOTIFY_SIGNAL	11	/* signal notifications exist */
+#define TIF_MEMDIE		11	/* is terminating due to OOM killer */
 
 #define _TIF_SYSCALL_TRACE	(1<<TIF_SYSCALL_TRACE)
 #define _TIF_SIGPENDING		(1<<TIF_SIGPENDING)
 #define _TIF_NEED_RESCHED	(1<<TIF_NEED_RESCHED)
 #define _TIF_SINGLESTEP		(1<<TIF_SINGLESTEP)
 #define _TIF_SYSCALL_TRACEPOINT	(1<<TIF_SYSCALL_TRACEPOINT)
+#define _TIF_NOTIFY_SIGNAL	(1<<TIF_NOTIFY_SIGNAL)
+#define _TIF_NOTIFY_RESUME	(1<<TIF_NOTIFY_RESUME)
 #define _TIF_SYSCALL_AUDIT	(1<<TIF_SYSCALL_AUDIT)
 #define _TIF_SECCOMP		(1<<TIF_SECCOMP)
-#define _TIF_NOTIFY_SIGNAL	(1<<TIF_NOTIFY_SIGNAL)
 
 #define _TIF_WORK_MASK		(_TIF_SYSCALL_TRACE | _TIF_SINGLESTEP | \
 				 _TIF_SYSCALL_TRACEPOINT | \
arch/xtensa/kernel/entry.S +2 -3
index 7f733f40fef0..647b162f959b 100644
@@ -500,9 +500,8 @@ common_exception_return:
 	 */
 
 	_bbsi.l	a4, TIF_NEED_RESCHED, 3f
-	_bbsi.l	a4, TIF_NOTIFY_RESUME, 2f
-	_bbci.l	a4, TIF_SIGPENDING, 5f
-	_bbci.l	a4, TIF_NOTIFY_SIGNAL, 5f
+	movi	a2, _TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_NOTIFY_SIGNAL
+	bnone	a4, a2, 5f
 
 2:	l32i	a4, a1, PT_DEPC
 	bgeui	a4, VALID_DOUBLE_EXCEPTION_ADDRESS, 4f
Signature · 1 line
2.20.1
#1 lore raw
Quoted 5 lines · Max Filippov

TIF_NOTIFY_SIGNAL handling in xtensa assembly is implemented
incorrectly: there should be a call to do_notify_resume when either TIF_SIGPENDING, TIF_NOTIFY_RESUME or TIF_NOTIFY_SIGNAL bit is set in the thread_info::flags. The straightforward way to do it would be

Quoted 8 lines

_bbsi.l a4, TIF_NEED_RESCHED, 3f
_bbsi.l a4, TIF_NOTIFY_RESUME, 2f
_bbsi.l a4, TIF_NOTIFY_SIGNAL, 2f
_bbci.l a4, TIF_SIGPENDING, 5f

Optimize it a little bit and use bit mask and bnone opcode to skip do_notify_resume invocation. Shuffle _TIF_* flags a bit so that used bit mask fits into the immediate field of movi opcode.

Thanks - do you mind if I fold this in with a reference to your changes? Seems like that'd be a better option than leaving it broken for a bit.

Signature · 1 line
Jens Axboe
author maintainer
#2 lore raw
Quoted 12 lines · Jens Axboe
Quoted 10 lines · Max Filippov

TIF_NOTIFY_SIGNAL handling in xtensa assembly is implemented
incorrectly: there should be a call to do_notify_resume when either TIF_SIGPENDING, TIF_NOTIFY_RESUME or TIF_NOTIFY_SIGNAL bit is set in the thread_info::flags. The straightforward way to do it would be

_bbsi.l a4, TIF_NEED_RESCHED, 3f
_bbsi.l a4, TIF_NOTIFY_RESUME, 2f
_bbsi.l a4, TIF_NOTIFY_SIGNAL, 2f
_bbci.l a4, TIF_SIGPENDING, 5f

Quoted 7 lines
Quoted 3 lines

Optimize it a little bit and use bit mask and bnone opcode to skip do_notify_resume invocation. Shuffle _TIF_* flags a bit so that used bit mask fits into the immediate field of movi opcode.

Thanks - do you mind if I fold this in with a reference to your changes? Seems like that'd be a better option than leaving it broken for a bit.

Sure Jens, by all means!

Signature · 2 lines
Thanks.
-- Max
#3 lore raw
Quoted 17 lines · Max Filippov
Quoted 16 lines · Jens Axboe
Quoted 13 lines · Max Filippov

TIF_NOTIFY_SIGNAL handling in xtensa assembly is implemented
incorrectly: there should be a call to do_notify_resume when either TIF_SIGPENDING, TIF_NOTIFY_RESUME or TIF_NOTIFY_SIGNAL bit is set in the thread_info::flags. The straightforward way to do it would be

_bbsi.l a4, TIF_NEED_RESCHED, 3f
_bbsi.l a4, TIF_NOTIFY_RESUME, 2f
_bbsi.l a4, TIF_NOTIFY_SIGNAL, 2f
_bbci.l a4, TIF_SIGPENDING, 5f

Optimize it a little bit and use bit mask and bnone opcode to skip do_notify_resume invocation. Shuffle _TIF_* flags a bit so that used bit mask fits into the immediate field of movi opcode.

Quoted 5 lines
Quoted 3 lines

Thanks - do you mind if I fold this in with a reference to your changes? Seems like that'd be a better option than leaving it broken for a bit.

Sure Jens, by all means!

Thanks! Done.

Signature · 1 line
Jens Axboe