Skip to contentSkip to conversation

More signal clean up

PATCH Applied Franck Bui-Huu 4 participants 16 messages Feb 9, 2007 – Feb 13, 2007 MIPS

lore
Share & tools
lore
#0 lore raw

cover More signal clean up

Hi Ralf,

Sorry for forgetting these 3 trivial patches. It shouldn't conflict with Atsushi's last patch:

Check FCSR for pending interrupts, alternative version

Please consider,

Franck
---

 arch/mips/kernel/signal-common.h |    2 ++
 arch/mips/kernel/signal.c        |   10 ++++------
 arch/mips/kernel/signal32.c      |    2 --
 arch/mips/kernel/signal_n32.c    |    2 --
 4 files changed, 6 insertions(+), 10 deletions(-)
#1 lore raw

PATCH 1/3 signal: avoid useless test in do_signal()

From: Franck Bui-Huu <[email protected]>

Indeed we can simply clear the flag whatever its value

---
 arch/mips/kernel/signal.c |    4 +---
 1 files changed, 1 insertions(+), 3 deletions(-)
arch/mips/kernel/signal.c +1 -3
index 8dfb7b1..464d34b 100644
@@ -552,10 +552,8 @@ void do_signal(struct pt_regs *regs)
 			 * and will be restored by sigreturn, so we can simply
 			 * clear the TIF_RESTORE_SIGMASK flag.
 			 */
-			if (test_thread_flag(TIF_RESTORE_SIGMASK))
-				clear_thread_flag(TIF_RESTORE_SIGMASK);
+			clear_thread_flag(TIF_RESTORE_SIGMASK);
 		}
-
 		return;
 	}
 
Signature · 1 line
1.4.4.3.ge6d4
#2 lore raw

Re: [PATCH 1/3] signal: avoid useless test in do_signal()

arch/mips/kernel/signal.c void do_signal(struct pt_regs *regs) · line 556

Quoted 2 lines · Franck Bui-Huu
arch/mips/kernel/signal.c +0 -2
-			if (test_thread_flag(TIF_RESTORE_SIGMASK))
-				clear_thread_flag(TIF_RESTORE_SIGMASK);

This is a microoptimization. The assumption here is TIF_RESTORE_SIGMASK will rarely need to be cleared and atomic operations are somewhat expensive if as in this case we have to assume the cacheline isn't held exclusive yet.

Ralf

#3 lore raw

Re: [PATCH 1/3] signal: avoid useless test in do_signal()

Quoted 10 lines · On 2/9/07, Ralf Baechle <[email protected]> wrote: Ralf Baechle wrote:

arch/mips/kernel/signal.c void do_signal(struct pt_regs *regs) · line 556

Quoted 2 lines · Franck Bui-Huu
arch/mips/kernel/signal.c +0 -2
-			if (test_thread_flag(TIF_RESTORE_SIGMASK))
-				clear_thread_flag(TIF_RESTORE_SIGMASK);

This is a microoptimization. The assumption here is TIF_RESTORE_SIGMASK will rarely need to be cleared and atomic operations are somewhat expensive if as in this case we have to assume the cacheline isn't held exclusive yet.

I missed that. You can forget this patch or maybe something like this is more appropriate ?

	if (unlikely(test_thread_flag(TIF_RESTORE_SIGMASK)))
		clear_thread_flag(TIF_RESTORE_SIGMASK);
Signature · 1 line
              Franck
#4 lore raw

PATCH 2/3 signals: make common _BLOCKABLE macro

In reply to Franck Bui-Huu

From: Franck Bui-Huu <[email protected]>

---
 arch/mips/kernel/signal-common.h |    2 ++
 arch/mips/kernel/signal.c        |    2 --
 arch/mips/kernel/signal32.c      |    2 --
 arch/mips/kernel/signal_n32.c    |    2 --
 4 files changed, 2 insertions(+), 6 deletions(-)
arch/mips/kernel/signal-common.h +2 -0
index 9a8abd6..23fffb4 100644
@@ -19,6 +19,8 @@
 #  define DEBUGP(fmt, args...)
 #endif
 
+#define _BLOCKABLE (~(sigmask(SIGKILL) | sigmask(SIGSTOP)))
+
 /*
  * Horribly complicated - with the bloody RM9000 workarounds enabled
  * the signal trampolines is moving to the end of the structure so we can
arch/mips/kernel/signal.c +0 -2
index 464d34b..a3c04d0 100644
@@ -34,8 +34,6 @@
 
 #include "signal-common.h"
 
-#define _BLOCKABLE (~(sigmask(SIGKILL) | sigmask(SIGSTOP)))
-
 #if ICACHE_REFILLS_WORKAROUND_WAR == 0
 
 struct rt_sigframe {
arch/mips/kernel/signal32.c +0 -2
index 183fc7e..f603ff4 100644
@@ -104,8 +104,6 @@ typedef struct compat_siginfo {
 #define __NR_O32_rt_sigreturn		4193
 #define __NR_O32_restart_syscall	4253
 
-#define _BLOCKABLE (~(sigmask(SIGKILL) | sigmask(SIGSTOP)))
-
 /* 32-bit compatibility types */
 
 #define _NSIG_BPW32	32
arch/mips/kernel/signal_n32.c +0 -2
index 57456e6..51b114f 100644
@@ -47,8 +47,6 @@
 #define __NR_N32_rt_sigreturn		6211
 #define __NR_N32_restart_syscall	6214
 
-#define _BLOCKABLE (~(sigmask(SIGKILL) | sigmask(SIGSTOP)))
-
 /* IRIX compatible stack_t  */
 typedef struct sigaltstack32 {
 	s32 ss_sp;
Signature · 1 line
1.4.4.3.ge6d4
Applied
#5 lore raw

Re: [PATCH 2/3] signals: make common _BLOCKABLE macro

On Fri, Feb 09, 2007 at 04:07:37PM +0100, Franck Bui-Huu wrote:

Thanks, applied.

Ralf

#6 lore raw

PATCH 3/3 signal.c: fix gcc warning on 32 bits kernel

In reply to Franck Bui-Huu

From: Franck Bui-Huu <[email protected]>

This hack prevents gcc to produce the following warning:

CC arch/mips/kernel/signal.o
arch/mips/kernel/signal.c: In function `sys_sigaction':
arch/mips/kernel/signal.c:266: warning: cast to pointer from integer of different size

This warning is due to the following line:

	__get_user(new_ka.sa.sa_handler, &act->sa_handler);

Indeed when __get_user() is expanded it produces the following code:

	switch (sizeof(*(&act->sa_handler))) {
	... [snip] ...
	case 8: {
		unsigned long long __gu_tmp;
		__asm__ __volatile__(
			"1: lw      %1, (%3)                  \n"
			"2: lw      %D1, 4(%3)                \n"
			"   move    %0, $0                    \n"
			"3: .section        .fixup,\"ax\"     \n"
			"4: li      %0, %4                    \n"
			"   move    %1, $0                    \n"
			"   move    %D1, $0                   \n"
			"   j       3b                        \n"
			"   .previous                         \n"
			"   .section        __ex_table,\"a\"  \n"
			"   " ".word" "     1b, 4b            \n"
			"   " ".word" "     2b, 4b            \n"
			"   .previous                         \n"
			: "=r" (__gu_err), "=&r" (__gu_tmp)
			: "0" (0), "r" ((&act->sa_handler)), "i" (-14));

		(((new_ka.sa.sa_handler))) = (__typeof__(*((&act->sa_handler)))) __gu_tmp;
		};
		break;
	default:
		__get_user_unknown();
		break;
	}

which actually try to do:

	new_ka.sa.sa_handler = (__sighandler_t) __gu_tmp;

Here we try to cast an 'unsigned long long' into a 32 bits pointer and that's the reason of the warning.

---
 arch/mips/kernel/signal.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
arch/mips/kernel/signal.c +3 -1
index a3c04d0..ac8a05a 100644
@@ -260,15 +260,17 @@ asmlinkage int sys_sigaction(int sig, const struct sigaction __user *act,
 
 	if (act) {
 		old_sigset_t mask;
+		unsigned long tmp; /* fix a gcc warning */
 
 		if (!access_ok(VERIFY_READ, act, sizeof(*act)))
 			return -EFAULT;
-		err |= __get_user(new_ka.sa.sa_handler, &act->sa_handler);
+		err |= __get_user(tmp, (unsigned long *)&act->sa_handler);
 		err |= __get_user(new_ka.sa.sa_flags, &act->sa_flags);
 		err |= __get_user(mask, &act->sa_mask.sig[0]);
 		if (err)
 			return -EFAULT;
 
+		new_ka.sa.sa_handler = (__sighandler_t)tmp;
 		siginitset(&new_ka.sa.sa_mask, mask);
 	}
 
Signature · 1 line
1.4.4.3.ge6d4
#7 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 7 lines · Franck Bui-Huu

CC arch/mips/kernel/signal.o
arch/mips/kernel/signal.c: In function `sys_sigaction':
arch/mips/kernel/signal.c:266: warning: cast to pointer from integer of different size

This warning is due to the following line:

	__get_user(new_ka.sa.sa_handler, &act->sa_handler);

This usage of __get_user() should be absolutely legal.

Quoted 4 lines
	new_ka.sa.sa_handler = (__sighandler_t) __gu_tmp;

Here we try to cast an 'unsigned long long' into a 32 bits pointer and that's the reason of the warning.

This line is never executed on 32bit kernel and gcc optimize out. On 64-bit kernel, this line is executed without any problem without warning.

I think this is a problem of __get_user() implementation or gcc itself. Though I can not find better solution yet, hacking the caller to avoid the warning would not be right things to to.

---
Atsushi Nemoto

#8 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 7 lines · Atsushi Nemoto
Quoted 4 lines · Franck Bui-Huu
      new_ka.sa.sa_handler = (__sighandler_t) __gu_tmp;

Here we try to cast an 'unsigned long long' into a 32 bits pointer and that's the reason of the warning.

This line is never executed on 32bit kernel and gcc optimize out. On

yes I agree but it seems that gcc compiles this line before optimizing out...

Quoted 4 lines

I think this is a problem of __get_user() implementation or gcc itself. Though I can not find better solution yet, hacking the caller to avoid the warning would not be right things to to.

I agree too but I haven't found something else.

BTW, my version of gcc is: mipsel-linux-gcc (GCC) 3.4.4 mipssde-6.05.00-20061023

thanks

Signature · 1 line
               Franck
#9 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 22 lines · Franck Bui-Huu

Date: Fri, 9 Feb 2007 17:34:16 +0100
From: "Franck Bui-Huu" <[email protected]>
To: "Atsushi Nemoto" <[email protected]>
Subject: Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel
Cc: [email protected], [email protected],
[email protected]
Content-Type: text/plain; charset=ISO-8859-1; format=flowed

Quoted 8 lines · Atsushi Nemoto
Quoted 4 lines · Franck Bui-Huu
      new_ka.sa.sa_handler = (__sighandler_t) __gu_tmp;

Here we try to cast an 'unsigned long long' into a 32 bits pointer and that's the reason of the warning.

This line is never executed on 32bit kernel and gcc optimize out. On

yes I agree but it seems that gcc compiles this line before optimizing out...

Quoted 1 line
Quoted 5 lines
Quoted 3 lines

I think this is a problem of __get_user() implementation or gcc itself. Though I can not find better solution yet, hacking the caller to avoid the warning would not be right things to to.

I agree too but I haven't found something else.

All gcc versions produce this warning and no, it's not a gcc bug but a kernel bug. __get_user expands into:

        case 8: {
                unsigned long long __gu_tmp;
[...]
		new_ka.sa.sa_handler =
			(__typeof__(*((&act->sa_handler)))) __gu_tmp;
	}

Which is quite a funny C problem to solve :-)

Ralf

#10 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 2 lines · Ralf Baechle

Which is quite a funny C problem to solve :-)

How about this instead ?

-- >8 --

include/asm-mips/uaccess.h +0 -0
index 1cdd4ee..ab7fe1c 100644
@@ -265,7 +265,7 @@ do {									\
 */

#define __get_user_asm_ll32(val, addr) \
{ \
- unsigned long long __gu_tmp; \

+        __typeof__(*(addr)) __gu_tmp;					\
									\
	__asm__ __volatile__(						\
	"1:	lw	%1, (%3)				\n"	\
@@ -283,7 +283,7 @@ do {									\
	"	.previous					\n"	\
	: "=r" (__gu_err), "=&r" (__gu_tmp)				\
	: "0" (0), "r" (addr), "i" (-EFAULT));				\

- (val) = (__typeof__(*(addr))) __gu_tmp; \
+ (val) = __gu_tmp; \
}

/*

Signature · 1 line
              Franck
#11 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 1 line · Franck Bui-Huu

How about this instead ?

Won't work for a pointer to some const thing.

Ralf

#12 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 3 lines · Ralf Baechle
Quoted 1 line

How about this instead ?

Won't work for a pointer to some const thing.

And recent commit 4ed3a77f38c023658784804cb39a7ce18063dc88 reverts commit 3218357c94af92478ef39163163a81e654385320.

Round and round and round and ...

---
Atsushi Nemoto

#13 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 9 lines · Atsushi Nemoto
Quoted 3 lines · Ralf Baechle
Quoted 1 line

How about this instead ?

Won't work for a pointer to some const thing.

And recent commit 4ed3a77f38c023658784804cb39a7ce18063dc88 reverts commit 3218357c94af92478ef39163163a81e654385320.

Round and round and round and ...

Well, I reverted that the old state of a warning is definately preferable until we found a proper solution.

Ralf

#14 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 2 lines · Ralf Baechle

Well, I reverted that the old state of a warning is definately preferable until we found a proper solution.

Type-punning should do the trick.

Ralf

include/asm-mips/uaccess.h +7 -3
index c12ebc5..36b3a42 100644
@@ -265,7 +265,10 @@ do {									\
  */
 #define __get_user_asm_ll32(val, addr)					\
 {									\
-        unsigned long long __gu_tmp;					\
+	union {								\
+		unsigned long long	l;				\
+		__typeof__(*(addr))	t;				\
+	} __gu_tmp;							\
 									\
 	__asm__ __volatile__(						\
 	"1:	lw	%1, (%3)				\n"	\
@@ -281,9 +284,10 @@ do {									\
 	"	" __UA_ADDR "	1b, 4b				\n"	\
 	"	" __UA_ADDR "	2b, 4b				\n"	\
 	"	.previous					\n"	\
-	: "=r" (__gu_err), "=&r" (__gu_tmp)				\
+	: "=r" (__gu_err), "=&r" (__gu_tmp.l)				\
 	: "0" (0), "r" (addr), "i" (-EFAULT));				\
-	(val) = (__typeof__(*(addr))) __gu_tmp;				\
+									\
+	(val) = __gu_tmp.t;						\
 }
 
 /*
#15 lore raw

Re: [PATCH 3/3] signal.c: fix gcc warning on 32 bits kernel

Quoted 7 lines · Ralf Baechle
Quoted 2 lines · Ralf Baechle

Well, I reverted that the old state of a warning is definately preferable until we found a proper solution.

Type-punning should do the trick.

yes it does.

thanks.

Signature · 1 line
              Franck