[PATCH] vsyscall: use __iter_div_u64_rem()

[PATCH] vsyscall: use __iter_div_u64_rem()

[Arnd Bergmann]

On 32-bit x86 when building with clang-9, the loop gets turned back into
an inefficient division that causes a link error:

kernel/time/vsyscall.o: In function `update_vsyscall':
vsyscall.c:(.text+0xe3): undefined reference to `__udivdi3'

Use the provided __iter_div_u64_rem() function that is meant to address
the same case in other files.

Fixes: 44f57d788e7d ("timekeeping: Provide a generic update_vsyscall() implementation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 kernel/time/vsyscall.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/kernel/time/vsyscall.c b/kernel/time/vsyscall.c
index a80893180826..8cf3596a4ce6 100644
--- a/kernel/time/vsyscall.c
+++ b/kernel/time/vsyscall.c
@@ -104,11 +104,7 @@ void update_vsyscall(struct timekeeper *tk)
 	vdso_ts->sec	= tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
 	nsec		= tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
-	while (nsec >= NSEC_PER_SEC) {
-		nsec = nsec - NSEC_PER_SEC;
-		vdso_ts->sec++;
-	}
-	vdso_ts->nsec	= nsec;
+	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
 
 	if (__arch_use_vsyscall(vdata))
 		update_vdso_data(vdata, tk);
-- 
2.20.0

[tip:timers/urgent] timekeeping/vsyscall: Use __iter_div_u64_rem()

[tip-bot for Arnd Bergmann]

Commit-ID:  510c12149954eb9db6a30c2b9f7526af71afb7a4
Gitweb:     https://git.kernel.org/tip/510c12149954eb9db6a30c2b9f7526af71afb7a4
Author:     Arnd Bergmann <arnd@arndb.de>
AuthorDate: Wed, 10 Jul 2019 15:01:53 +0200
Committer:  Thomas Gleixner <tglx@linutronix.de>
CommitDate: Wed, 10 Jul 2019 17:17:49 +0200

timekeeping/vsyscall: Use __iter_div_u64_rem()

On 32-bit x86 when building with clang-9, the 'division' loop gets turned
back into an inefficient division that causes a link error:

kernel/time/vsyscall.o: In function `update_vsyscall':
vsyscall.c:(.text+0xe3): undefined reference to `__udivdi3'

Use the existing __iter_div_u64_rem() function which is used to address the
same issue in other places.

Fixes: 44f57d788e7d ("timekeeping: Provide a generic update_vsyscall() implementation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lkml.kernel.org/r/20190710130206.1670830-1-arnd@arndb.de

---
 kernel/time/vsyscall.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/kernel/time/vsyscall.c b/kernel/time/vsyscall.c
index a80893180826..8cf3596a4ce6 100644
--- a/kernel/time/vsyscall.c
+++ b/kernel/time/vsyscall.c
@@ -104,11 +104,7 @@ void update_vsyscall(struct timekeeper *tk)
 	vdso_ts->sec	= tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
 	nsec		= tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
-	while (nsec >= NSEC_PER_SEC) {
-		nsec = nsec - NSEC_PER_SEC;
-		vdso_ts->sec++;
-	}
-	vdso_ts->nsec	= nsec;
+	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
 
 	if (__arch_use_vsyscall(vdata))
 		update_vdso_data(vdata, tk);

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Nathan Chancellor]

On Wed, Jul 10, 2019 at 03:01:53PM +0200, Arnd Bergmann wrote:
> On 32-bit x86 when building with clang-9, the loop gets turned back into
> an inefficient division that causes a link error:
> 
> kernel/time/vsyscall.o: In function `update_vsyscall':
> vsyscall.c:(.text+0xe3): undefined reference to `__udivdi3'
> 
> Use the provided __iter_div_u64_rem() function that is meant to address
> the same case in other files.
> 
> Fixes: 44f57d788e7d ("timekeeping: Provide a generic update_vsyscall() implementation")
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> ---
>  kernel/time/vsyscall.c | 6 +-----
>  1 file changed, 1 insertion(+), 5 deletions(-)
> 
> diff --git a/kernel/time/vsyscall.c b/kernel/time/vsyscall.c
> index a80893180826..8cf3596a4ce6 100644
> --- a/kernel/time/vsyscall.c
> +++ b/kernel/time/vsyscall.c
> @@ -104,11 +104,7 @@ void update_vsyscall(struct timekeeper *tk)
>  	vdso_ts->sec	= tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
>  	nsec		= tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
>  	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
> -	while (nsec >= NSEC_PER_SEC) {
> -		nsec = nsec - NSEC_PER_SEC;
> -		vdso_ts->sec++;
> -	}
> -	vdso_ts->nsec	= nsec;
> +	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
>  
>  	if (__arch_use_vsyscall(vdata))
>  		update_vdso_data(vdata, tk);
> -- 
> 2.20.0
> 

What an interesting function. Looks good to me and I can confirm it
fixes the link error.

Reviewed-by: Nathan Chancellor <natechancellor@gmail.com>
Tested-by: Nathan Chancellor <natechancellor@gmail.com>

[tip:timers/urgent] timekeeping/vsyscall: Use __iter_div_u64_rem()

[tip-bot for Arnd Bergmann]

Commit-ID:  0df1c9868c3a1916198ee09c323ca5932a0b8a11
Gitweb:     https://git.kernel.org/tip/0df1c9868c3a1916198ee09c323ca5932a0b8a11
Author:     Arnd Bergmann <arnd@arndb.de>
AuthorDate: Wed, 10 Jul 2019 15:01:53 +0200
Committer:  Thomas Gleixner <tglx@linutronix.de>
CommitDate: Wed, 10 Jul 2019 20:37:49 +0200

timekeeping/vsyscall: Use __iter_div_u64_rem()

On 32-bit x86 when building with clang-9, the 'division' loop gets turned
back into an inefficient division that causes a link error:

kernel/time/vsyscall.o: In function `update_vsyscall':
vsyscall.c:(.text+0xe3): undefined reference to `__udivdi3'

Use the existing __iter_div_u64_rem() function which is used to address the
same issue in other places.

Fixes: 44f57d788e7d ("timekeeping: Provide a generic update_vsyscall() implementation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Nathan Chancellor <natechancellor@gmail.com>
Tested-by: Nathan Chancellor <natechancellor@gmail.com>
Link: https://lkml.kernel.org/r/20190710130206.1670830-1-arnd@arndb.de
---
 kernel/time/vsyscall.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/kernel/time/vsyscall.c b/kernel/time/vsyscall.c
index a80893180826..8cf3596a4ce6 100644
--- a/kernel/time/vsyscall.c
+++ b/kernel/time/vsyscall.c
@@ -104,11 +104,7 @@ void update_vsyscall(struct timekeeper *tk)
 	vdso_ts->sec	= tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
 	nsec		= tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
 	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
-	while (nsec >= NSEC_PER_SEC) {
-		nsec = nsec - NSEC_PER_SEC;
-		vdso_ts->sec++;
-	}
-	vdso_ts->nsec	= nsec;
+	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
 
 	if (__arch_use_vsyscall(vdata))
 		update_vdso_data(vdata, tk);

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Vincenzo Frascino]

Hi Arnd,

On 10/07/2019 14:01, Arnd Bergmann wrote:
> On 32-bit x86 when building with clang-9, the loop gets turned back into
> an inefficient division that causes a link error:
> 
> kernel/time/vsyscall.o: In function `update_vsyscall':
> vsyscall.c:(.text+0xe3): undefined reference to `__udivdi3'
> 
> Use the provided __iter_div_u64_rem() function that is meant to address
> the same case in other files.
> 
> Fixes: 44f57d788e7d ("timekeeping: Provide a generic update_vsyscall() implementation")
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> ---
>  kernel/time/vsyscall.c | 6 +-----
>  1 file changed, 1 insertion(+), 5 deletions(-)
> 
> diff --git a/kernel/time/vsyscall.c b/kernel/time/vsyscall.c
> index a80893180826..8cf3596a4ce6 100644
> --- a/kernel/time/vsyscall.c
> +++ b/kernel/time/vsyscall.c
> @@ -104,11 +104,7 @@ void update_vsyscall(struct timekeeper *tk)
>  	vdso_ts->sec	= tk->xtime_sec + tk->wall_to_monotonic.tv_sec;
>  	nsec		= tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
>  	nsec		= nsec + tk->wall_to_monotonic.tv_nsec;
> -	while (nsec >= NSEC_PER_SEC) {
> -		nsec = nsec - NSEC_PER_SEC;
> -		vdso_ts->sec++;
> -	}
> -	vdso_ts->nsec	= nsec;
> +	vdso_ts->sec	+= __iter_div_u64_rem(nsec, NSEC_PER_SEC, &vdso_ts->nsec);
>  
>  	if (__arch_use_vsyscall(vdata))
>  		update_vdso_data(vdata, tk);
> 

I am trying to test this patch using clang-9 tip:

# clang -v
clang version 9.0.0 (git@github.com:llvm-mirror/clang.git
6ed0749151866894a67a3e7eefdc1f3a547daa0e) (git@github.com:llvm-mirror/llvm.git
a10a70238ace1093cad3adeb94814b422bd1b5c1)

but I get a lot of errors similar to the one below:

In file included from ~/linux/arch/x86/events/amd/core.c:11:
~/linux/arch/x86/events/amd/../perf_event.h:824:21: error: invalid output size
for constraint '=q'
        u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
                           ^
~/linux/include/linux/percpu-defs.h:447:2: note: expanded from macro
'__this_cpu_read'
        raw_cpu_read(pcp);                                              \
        ^
~/linux/include/linux/percpu-defs.h:421:28: note: expanded from macro 'raw_cpu_read'
#define raw_cpu_read(pcp)               __pcpu_size_call_return(raw_cpu_read_, pcp)
                                        ^
~/linux/include/linux/percpu-defs.h:322:23: note: expanded from macro
'__pcpu_size_call_return'
        case 1: pscr_ret__ = stem##1(variable); break;                  \
                             ^
<scratch space>:110:1: note: expanded from here
raw_cpu_read_1
^
~/linux/arch/x86/include/asm/percpu.h:394:30: note: expanded from macro
'raw_cpu_read_1'
#define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
                                        ^
~/linux/arch/x86/include/asm/percpu.h:189:15: note: expanded from macro
'percpu_from_op'
                    : "=q" (pfo_ret__)                  \

Could you please tell me which version of the compiler did you use?

My building command is:

# make mrproper && make CC=clang HOSTCC=clang i386_defconfig && make ARCH=i386
CC=clang HOSTCC=clang -j56

-- 
Regards,
Vincenzo

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Arnd Bergmann]

On Thu, Jul 11, 2019 at 2:14 PM Vincenzo Frascino
<vincenzo.frascino@arm.com> wrote:
>
>
> Could you please tell me which version of the compiler did you use?
>
> My building command is:
>
> # make mrproper && make CC=clang HOSTCC=clang i386_defconfig && make ARCH=i386
> CC=clang HOSTCC=clang -j56
>

See below for the patch I am using locally to work around this.
That patch is probably wrong, so I have not submitted it yet, but it
gives you a clean build ;-)

     Arnd

8<---
Subject: [PATCH] x86: percpu: fix clang 32-bit build

clang does not like an inline assembly with a "=q" contraint for
a 64-bit output:

arch/x86/events/perf_event.h:824:21: error: invalid output size for
constraint '=q'
        u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
                           ^
include/linux/percpu-defs.h:447:2: note: expanded from macro '__this_cpu_read'
        raw_cpu_read(pcp);                                              \
        ^
include/linux/percpu-defs.h:421:28: note: expanded from macro 'raw_cpu_read'
 #define raw_cpu_read(pcp)
__pcpu_size_call_return(raw_cpu_read_, pcp)
                                        ^
include/linux/percpu-defs.h:322:23: note: expanded from macro
'__pcpu_size_call_return'
        case 1: pscr_ret__ = stem##1(variable); break;                  \
                             ^
<scratch space>:357:1: note: expanded from here
raw_cpu_read_1
^
arch/x86/include/asm/percpu.h:394:30: note: expanded from macro 'raw_cpu_read_1'
 #define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
                                        ^
arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
                    : "=q" (pfo_ret__)                  \
                            ^

According to the commit that introduced the "q" constraint, this was
needed to fix miscompilation, but it gives no further detail.

Using the normal "=r" constraint seems to work so far.

Fixes: 3c598766a2ba ("x86: fix percpu_{to,from}_op()")
Cc: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>

diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
index 2278797c769d..e791fbf4018f 100644
--- a/arch/x86/include/asm/percpu.h
+++ b/arch/x86/include/asm/percpu.h
@@ -99,7 +99,7 @@ do {                                                  \
        case 1:                                         \
                asm qual (op "b %1,"__percpu_arg(0)     \
                    : "+m" (var)                        \
-                   : "qi" ((pto_T__)(val)));           \
+                   : "ri" ((pto_T__)(val)));           \
                break;                                  \
        case 2:                                         \
                asm qual (op "w %1,"__percpu_arg(0)     \
@@ -144,7 +144,7 @@ do {
                         \
                else                                                    \
                        asm qual ("addb %1, "__percpu_arg(0)            \
                            : "+m" (var)                                \
-                           : "qi" ((pao_T__)(val)));                   \
+                           : "ri" ((pao_T__)(val)));                   \
                break;                                                  \
        case 2:                                                         \
                if (pao_ID__ == 1)                                      \
@@ -186,7 +186,7 @@ do {
                         \
        switch (sizeof(var)) {                          \
        case 1:                                         \
                asm qual (op "b "__percpu_arg(1)",%0"   \
-                   : "=q" (pfo_ret__)                  \
+                   : "=r" (pfo_ret__)                  \
                    : "m" (var));                       \
                break;                                  \
        case 2:                                         \
@@ -215,7 +215,7 @@ do {
                         \
        switch (sizeof(var)) {                          \
        case 1:                                         \
                asm(op "b "__percpu_arg(P1)",%0"        \
-                   : "=q" (pfo_ret__)                  \
+                   : "=r" (pfo_ret__)                  \
                    : "p" (&(var)));                    \
                break;                                  \
        case 2:                                         \

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Vincenzo Frascino]

Hi Arnd,

On 11/07/2019 13:28, Arnd Bergmann wrote:
> On Thu, Jul 11, 2019 at 2:14 PM Vincenzo Frascino
> <vincenzo.frascino@arm.com> wrote:
>>
>>
>> Could you please tell me which version of the compiler did you use?
>>
>> My building command is:
>>
>> # make mrproper && make CC=clang HOSTCC=clang i386_defconfig && make ARCH=i386
>> CC=clang HOSTCC=clang -j56
>>
> 
> See below for the patch I am using locally to work around this.
> That patch is probably wrong, so I have not submitted it yet, but it
> gives you a clean build ;-)
> 
>      Arnd
> 

Thank you, I will give it a go :-)

> 8<---
> Subject: [PATCH] x86: percpu: fix clang 32-bit build
> 
> clang does not like an inline assembly with a "=q" contraint for
> a 64-bit output:
> 
> arch/x86/events/perf_event.h:824:21: error: invalid output size for
> constraint '=q'
>         u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
>                            ^
> include/linux/percpu-defs.h:447:2: note: expanded from macro '__this_cpu_read'
>         raw_cpu_read(pcp);                                              \
>         ^
> include/linux/percpu-defs.h:421:28: note: expanded from macro 'raw_cpu_read'
>  #define raw_cpu_read(pcp)
> __pcpu_size_call_return(raw_cpu_read_, pcp)
>                                         ^
> include/linux/percpu-defs.h:322:23: note: expanded from macro
> '__pcpu_size_call_return'
>         case 1: pscr_ret__ = stem##1(variable); break;                  \
>                              ^
> <scratch space>:357:1: note: expanded from here
> raw_cpu_read_1
> ^
> arch/x86/include/asm/percpu.h:394:30: note: expanded from macro 'raw_cpu_read_1'
>  #define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
>                                         ^
> arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
>                     : "=q" (pfo_ret__)                  \
>                             ^
> 
> According to the commit that introduced the "q" constraint, this was
> needed to fix miscompilation, but it gives no further detail.
> 
> Using the normal "=r" constraint seems to work so far.
> 
> Fixes: 3c598766a2ba ("x86: fix percpu_{to,from}_op()")
> Cc: Jan Beulich <jbeulich@suse.com>
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> 
> diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
> index 2278797c769d..e791fbf4018f 100644
> --- a/arch/x86/include/asm/percpu.h
> +++ b/arch/x86/include/asm/percpu.h
> @@ -99,7 +99,7 @@ do {                                                  \
>         case 1:                                         \
>                 asm qual (op "b %1,"__percpu_arg(0)     \
>                     : "+m" (var)                        \
> -                   : "qi" ((pto_T__)(val)));           \
> +                   : "ri" ((pto_T__)(val)));           \
>                 break;                                  \
>         case 2:                                         \
>                 asm qual (op "w %1,"__percpu_arg(0)     \
> @@ -144,7 +144,7 @@ do {
>                          \
>                 else                                                    \
>                         asm qual ("addb %1, "__percpu_arg(0)            \
>                             : "+m" (var)                                \
> -                           : "qi" ((pao_T__)(val)));                   \
> +                           : "ri" ((pao_T__)(val)));                   \
>                 break;                                                  \
>         case 2:                                                         \
>                 if (pao_ID__ == 1)                                      \
> @@ -186,7 +186,7 @@ do {
>                          \
>         switch (sizeof(var)) {                          \
>         case 1:                                         \
>                 asm qual (op "b "__percpu_arg(1)",%0"   \
> -                   : "=q" (pfo_ret__)                  \
> +                   : "=r" (pfo_ret__)                  \
>                     : "m" (var));                       \
>                 break;                                  \
>         case 2:                                         \
> @@ -215,7 +215,7 @@ do {
>                          \
>         switch (sizeof(var)) {                          \
>         case 1:                                         \
>                 asm(op "b "__percpu_arg(P1)",%0"        \
> -                   : "=q" (pfo_ret__)                  \
> +                   : "=r" (pfo_ret__)                  \
>                     : "p" (&(var)));                    \
>                 break;                                  \
>         case 2:                                         \
> 

-- 
Regards,
Vincenzo

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Nick Desaulniers]

On Thu, Jul 11, 2019 at 5:28 AM Arnd Bergmann <arnd@arndb.de> wrote:
> clang does not like an inline assembly with a "=q" contraint for
> a 64-bit output:

Seems like starting in GCC 7, GCC may not like it either:
https://godbolt.org/z/UyBUfh
it simply warns then proceeds with code gen.  Another difference may
come from when GCC vs Clang perform dead code elimination (DCE) vs
semantic analysis.

-- 
Thanks,
~Nick Desaulniers

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Arnd Bergmann]

On Thu, Jul 11, 2019 at 7:14 PM 'Nick Desaulniers' via Clang Built
Linux <clang-built-linux@googlegroups.com> wrote:
>
> On Thu, Jul 11, 2019 at 5:28 AM Arnd Bergmann <arnd@arndb.de> wrote:
> > clang does not like an inline assembly with a "=q" contraint for
> > a 64-bit output:
>
> Seems like starting in GCC 7, GCC may not like it either:
> https://godbolt.org/z/UyBUfh
> it simply warns then proceeds with code gen.  Another difference may
> come from when GCC vs Clang perform dead code elimination (DCE) vs
> semantic analysis.

Right, I also had the idea to work around it with a set of
__builtin_choos_expr()
instead of the switch()/case but did not complete that patch as the percpu
code is rather complex and this would touch lots of code.

       Arnd

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Thomas Gleixner]

On Thu, 11 Jul 2019, Arnd Bergmann wrote:

Trimmed CC list and added Jan

> See below for the patch I am using locally to work around this.
> That patch is probably wrong, so I have not submitted it yet, but it
> gives you a clean build ;-)
> 
>      Arnd
> 8<---
> Subject: [PATCH] x86: percpu: fix clang 32-bit build
> 
> clang does not like an inline assembly with a "=q" contraint for
> a 64-bit output:
> 
> arch/x86/events/perf_event.h:824:21: error: invalid output size for
> constraint '=q'
>         u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
>                            ^
> include/linux/percpu-defs.h:447:2: note: expanded from macro '__this_cpu_read'
>         raw_cpu_read(pcp);                                              \
>         ^
> include/linux/percpu-defs.h:421:28: note: expanded from macro 'raw_cpu_read'
>  #define raw_cpu_read(pcp)
> __pcpu_size_call_return(raw_cpu_read_, pcp)
>                                         ^
> include/linux/percpu-defs.h:322:23: note: expanded from macro
> '__pcpu_size_call_return'
>         case 1: pscr_ret__ = stem##1(variable); break;                  \
>                              ^
> <scratch space>:357:1: note: expanded from here
> raw_cpu_read_1
> ^
> arch/x86/include/asm/percpu.h:394:30: note: expanded from macro 'raw_cpu_read_1'
>  #define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
>                                         ^
> arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
>                     : "=q" (pfo_ret__)                  \
>                             ^
> 
> According to the commit that introduced the "q" constraint, this was
> needed to fix miscompilation, but it gives no further detail.

Jan, do you have any memory why you added those 'q' constraints? The
changelog of 3c598766a2ba is not really helpful.

Thanks,

	tglx

> Using the normal "=r" constraint seems to work so far.
> 
> Fixes: 3c598766a2ba ("x86: fix percpu_{to,from}_op()")
> Cc: Jan Beulich <jbeulich@suse.com>
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> 
> diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
> index 2278797c769d..e791fbf4018f 100644
> --- a/arch/x86/include/asm/percpu.h
> +++ b/arch/x86/include/asm/percpu.h
> @@ -99,7 +99,7 @@ do {                                                  \
>         case 1:                                         \
>                 asm qual (op "b %1,"__percpu_arg(0)     \
>                     : "+m" (var)                        \
> -                   : "qi" ((pto_T__)(val)));           \
> +                   : "ri" ((pto_T__)(val)));           \
>                 break;                                  \
>         case 2:                                         \
>                 asm qual (op "w %1,"__percpu_arg(0)     \
> @@ -144,7 +144,7 @@ do {
>                          \
>                 else                                                    \
>                         asm qual ("addb %1, "__percpu_arg(0)            \
>                             : "+m" (var)                                \
> -                           : "qi" ((pao_T__)(val)));                   \
> +                           : "ri" ((pao_T__)(val)));                   \
>                 break;                                                  \
>         case 2:                                                         \
>                 if (pao_ID__ == 1)                                      \
> @@ -186,7 +186,7 @@ do {
>                          \
>         switch (sizeof(var)) {                          \
>         case 1:                                         \
>                 asm qual (op "b "__percpu_arg(1)",%0"   \
> -                   : "=q" (pfo_ret__)                  \
> +                   : "=r" (pfo_ret__)                  \
>                     : "m" (var));                       \
>                 break;                                  \
>         case 2:                                         \
> @@ -215,7 +215,7 @@ do {
>                          \
>         switch (sizeof(var)) {                          \
>         case 1:                                         \
>                 asm(op "b "__percpu_arg(P1)",%0"        \
> -                   : "=q" (pfo_ret__)                  \
> +                   : "=r" (pfo_ret__)                  \
>                     : "p" (&(var)));                    \
>                 break;                                  \
>         case 2:                                         \
> 

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Jan Beulich]

On 22.07.2019 12:10, Thomas Gleixner wrote:
> On Thu, 11 Jul 2019, Arnd Bergmann wrote:
> 
> Trimmed CC list and added Jan
> 
>> See below for the patch I am using locally to work around this.
>> That patch is probably wrong, so I have not submitted it yet, but it
>> gives you a clean build ;-)
>>
>>       Arnd
>> 8<---
>> Subject: [PATCH] x86: percpu: fix clang 32-bit build
>>
>> clang does not like an inline assembly with a "=q" contraint for
>> a 64-bit output:
>>
>> arch/x86/events/perf_event.h:824:21: error: invalid output size for
>> constraint '=q'
>>          u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
>>                             ^
>> include/linux/percpu-defs.h:447:2: note: expanded from macro '__this_cpu_read'
>>          raw_cpu_read(pcp);                                              \
>>          ^
>> include/linux/percpu-defs.h:421:28: note: expanded from macro 'raw_cpu_read'
>>   #define raw_cpu_read(pcp)
>> __pcpu_size_call_return(raw_cpu_read_, pcp)
>>                                          ^
>> include/linux/percpu-defs.h:322:23: note: expanded from macro
>> '__pcpu_size_call_return'
>>          case 1: pscr_ret__ = stem##1(variable); break;                  \
>>                               ^
>> <scratch space>:357:1: note: expanded from here
>> raw_cpu_read_1
>> ^
>> arch/x86/include/asm/percpu.h:394:30: note: expanded from macro 'raw_cpu_read_1'
>>   #define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
>>                                          ^
>> arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
>>                      : "=q" (pfo_ret__)                  \
>>                              ^
>>
>> According to the commit that introduced the "q" constraint, this was
>> needed to fix miscompilation, but it gives no further detail.
> 
> Jan, do you have any memory why you added those 'q' constraints? The
> changelog of 3c598766a2ba is not really helpful.

"q" was used in that commit exclusively for byte sized operands, simply
because that _is_ the constraint to use in such cases. Using "r" is
wrong on 32-bit, as it would include inaccessible byte portions of %sp,
%bp, %si, and %di. This is how it's described in gcc sources / docs:

  "Any register accessible as @code{@var{r}l}.  In 32-bit mode, @code{a},
   @code{b}, @code{c}, and @code{d}; in 64-bit mode, any integer register."

What I'm struggling with is why clang would evaluate that asm() in the
first place when a 64-bit field (perf_ctr_virt_mask) is being accessed.

Jan

Re: [PATCH] vsyscall: use __iter_div_u64_rem()

[Arnd Bergmann]

On Mon, Jul 22, 2019 at 12:39 PM Jan Beulich <JBeulich@suse.com> wrote:
> On 22.07.2019 12:10, Thomas Gleixner wrote:
> > On Thu, 11 Jul 2019, Arnd Bergmann wrote:
> "q" was used in that commit exclusively for byte sized operands, simply
> because that _is_ the constraint to use in such cases. Using "r" is
> wrong on 32-bit, as it would include inaccessible byte portions of %sp,
> %bp, %si, and %di. This is how it's described in gcc sources / docs:
>
>   "Any register accessible as @code{@var{r}l}.  In 32-bit mode, @code{a},
>    @code{b}, @code{c}, and @code{d}; in 64-bit mode, any integer register."
>
> What I'm struggling with is why clang would evaluate that asm() in the
> first place when a 64-bit field (perf_ctr_virt_mask) is being accessed.

clang does the optimization and warning checking in a different order,
in this case the argument type checks for the inline assembly is done
before it eliminates the dead code.

        Arnd