Hi all,
Today's linux-next merge of the xen-tip tree got a conflict in:
arch/x86/xen/enlighten.c
between commit:
9261e050b686 ("x86/asm/tsc, x86/paravirt: Remove read_tsc() and read_tscp() paravirt hooks")
from the tip tree and commit:
cd6f350a5460 ("xen/PMU: Initialization code for Xen PMU")
from the xen-tip tree.
I fixed it up (see below) and can carry the fix as necessary (no action
is required).
--
Cheers,
Stephen Rothwell [email protected]
diff --cc arch/x86/xen/enlighten.c
index d9cfa452da9d,a1c2e46206a9..000000000000
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@@ -1215,8 -1179,11 +1219,8 @@@ static const struct pv_cpu_ops xen_cpu_
.read_msr = xen_read_msr_safe,
.write_msr = xen_write_msr_safe,
- .read_pmc = native_read_pmc,
- .read_tsc = native_read_tsc,
+ .read_pmc = xen_read_pmc,
- .read_tscp = native_read_tscp,
-
.iret = xen_iret,
#ifdef CONFIG_X86_64
.usergs_sysret32 = xen_sysret32,
On 08/12/2015 01:09 AM, Stephen Rothwell wrote:
> Hi all,
>
> Today's linux-next merge of the xen-tip tree got a conflict in:
>
> arch/x86/xen/enlighten.c
>
> between commit:
>
> 9261e050b686 ("x86/asm/tsc, x86/paravirt: Remove read_tsc() and read_tscp() paravirt hooks")
>
> from the tip tree and commit:
>
> cd6f350a5460 ("xen/PMU: Initialization code for Xen PMU")
>
> from the xen-tip tree.
>
> I fixed it up (see below) and can carry the fix as necessary (no action
> is required).
Thanks, this looks good.
Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red
Hat 4.4.4-2) (GCC)):
CC arch/x86/kernel/nmi.o
In file included from
/home/build/linux-boris/include/linux/jump_label.h:109,
from
/home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
from /home/build/linux-boris/include/linux/spinlock.h:88,
from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
/home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
?nmi_handle?:
/home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning:
asm operand 0 probably doesn?t match constraints
/home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
impossible constraint in ?asm?
make[3]: *** [arch/x86/kernel/nmi.o] Error 1
make[2]: *** [arch/x86/kernel] Error 2
make[1]: *** [arch/x86] Error 2
-boris
>
> -- Cheers, Stephen Rothwell [email protected] diff --cc
> arch/x86/xen/enlighten.c index d9cfa452da9d,a1c2e46206a9..000000000000
> --- a/arch/x86/xen/enlighten.c +++ b/arch/x86/xen/enlighten.c @@@
> -1215,8 -1179,11 +1219,8 @@@ static const struct pv_cpu_ops xen_cpu_
> .read_msr = xen_read_msr_safe, .write_msr = xen_write_msr_safe, -
> .read_pmc = native_read_pmc, - .read_tsc = native_read_tsc, +
> .read_pmc = xen_read_pmc, - .read_tscp = native_read_tscp, - .iret =
> xen_iret, #ifdef CONFIG_X86_64 .usergs_sysret32 = xen_sysret32,
On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
> Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
> interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red Hat
> 4.4.4-2) (GCC)):
>
>
>
> CC arch/x86/kernel/nmi.o
> In file included from
> /home/build/linux-boris/include/linux/jump_label.h:109,
> from
> /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
> from /home/build/linux-boris/include/linux/spinlock.h:88,
> from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
> /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
> ‘nmi_handle’:
> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning: asm
> operand 0 probably doesn’t match constraints
> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
> impossible constraint in ‘asm’
> make[3]: *** [arch/x86/kernel/nmi.o] Error 1
> make[2]: *** [arch/x86/kernel] Error 2
> make[1]: *** [arch/x86] Error 2
Ugh bugger.
I bet its that: &((char *)key)[branch] business, an earlier variant
thereof tripped up more recent GCCs too.
So its an __always_inline function, and both argument are always compile
time constants, @key is the address of an object in static storage (a
global) and @branch is a simple 0/1 at the call site.
Now we wish to compute (unsigned long)key + branch at compile/link time
to feed to the assembler as an immediate, which should be possible,
given its all 'constants'.
It just appears GCC is having a hard time with this.
Let me see if I have a sufficiently old GCC around to play with.
On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>
> > Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
> > interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red Hat
> > 4.4.4-2) (GCC)):
> >
> >
> >
> > CC arch/x86/kernel/nmi.o
> > In file included from
> > /home/build/linux-boris/include/linux/jump_label.h:109,
> > from
> > /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
> > from /home/build/linux-boris/include/linux/spinlock.h:88,
> > from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
> > ‘nmi_handle’:
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning: asm
> > operand 0 probably doesn’t match constraints
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
> > impossible constraint in ‘asm’
> > make[3]: *** [arch/x86/kernel/nmi.o] Error 1
> > make[2]: *** [arch/x86/kernel] Error 2
> > make[1]: *** [arch/x86] Error 2
>
> Ugh bugger.
>
> I bet its that: &((char *)key)[branch] business, an earlier variant
> thereof tripped up more recent GCCs too.
>
> So its an __always_inline function, and both argument are always compile
> time constants, @key is the address of an object in static storage (a
> global) and @branch is a simple 0/1 at the call site.
>
> Now we wish to compute (unsigned long)key + branch at compile/link time
> to feed to the assembler as an immediate, which should be possible,
> given its all 'constants'.
>
> It just appears GCC is having a hard time with this.
>
> Let me see if I have a sufficiently old GCC around to play with.
So both my 4.4 compilers:
gcc-4.4 (Ubuntu/Linaro 4.4.7-8ubuntu1) 4.4.7
gcc-4.4 (Debian 4.4.7-2) 4.4.7
Do not have CC_HAVE_ASM_GOTO at all, and my gcc-4.6 (my next oldest
compiler) doesn't have trouble building this.
On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>
> > Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
> > interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red Hat
> > 4.4.4-2) (GCC)):
> >
> >
> >
> > CC arch/x86/kernel/nmi.o
> > In file included from
> > /home/build/linux-boris/include/linux/jump_label.h:109,
> > from
> > /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
> > from /home/build/linux-boris/include/linux/spinlock.h:88,
> > from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
> > ‘nmi_handle’:
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning: asm
> > operand 0 probably doesn’t match constraints
> > /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
> > impossible constraint in ‘asm’
> > make[3]: *** [arch/x86/kernel/nmi.o] Error 1
> > make[2]: *** [arch/x86/kernel] Error 2
> > make[1]: *** [arch/x86] Error 2
>
> Ugh bugger.
>
> I bet its that: &((char *)key)[branch] business, an earlier variant
> thereof tripped up more recent GCCs too.
>
> So its an __always_inline function, and both argument are always compile
> time constants, @key is the address of an object in static storage (a
> global) and @branch is a simple 0/1 at the call site.
>
> Now we wish to compute (unsigned long)key + branch at compile/link time
> to feed to the assembler as an immediate, which should be possible,
> given its all 'constants'.
>
> It just appears GCC is having a hard time with this.
>
> Let me see if I have a sufficiently old GCC around to play with.
Could you feed the below to your compiler? Its a bit cumbersome, but
its the next best I could come up with...
---
arch/x86/include/asm/jump_label.h | 56 ++++++++++++++++++++++++++++-----------
1 file changed, 41 insertions(+), 15 deletions(-)
diff --git a/arch/x86/include/asm/jump_label.h b/arch/x86/include/asm/jump_label.h
index 28d7a857f9d1..76c769ae4200 100644
--- a/arch/x86/include/asm/jump_label.h
+++ b/arch/x86/include/asm/jump_label.h
@@ -16,15 +16,30 @@
# define STATIC_KEY_INIT_NOP GENERIC_NOP5_ATOMIC
#endif
+struct foo {
+ u8 zero;
+ u8 one;
+};
+
static __always_inline bool arch_static_branch(struct static_key *key, bool branch)
{
- asm_volatile_goto("1:"
- ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
- ".pushsection __jump_table, \"aw\" \n\t"
- _ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
- ".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ if (!branch) {
+ asm_volatile_goto("1:"
+ ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
+ ".pushsection __jump_table, \"aw\" \n\t"
+ _ASM_ALIGN "\n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ ".popsection \n\t"
+ : : "i" (&((struct foo *)key)->zero) : : l_yes);
+ } else {
+ asm_volatile_goto("1:"
+ ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
+ ".pushsection __jump_table, \"aw\" \n\t"
+ _ASM_ALIGN "\n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ ".popsection \n\t"
+ : : "i" (&((struct foo *)key)->one) : : l_yes);
+ }
return false;
l_yes:
@@ -33,14 +48,25 @@ static __always_inline bool arch_static_branch(struct static_key *key, bool bran
static __always_inline bool arch_static_branch_jump(struct static_key *key, bool branch)
{
- asm_volatile_goto("1:"
- ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
- "2:\n\t"
- ".pushsection __jump_table, \"aw\" \n\t"
- _ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
- ".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ if (!branch) {
+ asm_volatile_goto("1:"
+ ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
+ "2:\n\t"
+ ".pushsection __jump_table, \"aw\" \n\t"
+ _ASM_ALIGN "\n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ ".popsection \n\t"
+ : : "i" (&((struct foo *)key)->zero) : : l_yes);
+ } else {
+ asm_volatile_goto("1:"
+ ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
+ "2:\n\t"
+ ".pushsection __jump_table, \"aw\" \n\t"
+ _ASM_ALIGN "\n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ ".popsection \n\t"
+ : : "i" (&((struct foo *)key)->one) : : l_yes);
+ }
return false;
l_yes:
On 08/12/2015 01:46 PM, Peter Zijlstra wrote:
> On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
>> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>>
>>> Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
>>> interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red Hat
>>> 4.4.4-2) (GCC)):
>>>
>>>
>>>
>>> CC arch/x86/kernel/nmi.o
>>> In file included from
>>> /home/build/linux-boris/include/linux/jump_label.h:109,
>>> from
>>> /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
>>> from /home/build/linux-boris/include/linux/spinlock.h:88,
>>> from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
>>> ‘nmi_handle’:
>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning: asm
>>> operand 0 probably doesn’t match constraints
>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
>>> impossible constraint in ‘asm’
>>> make[3]: *** [arch/x86/kernel/nmi.o] Error 1
>>> make[2]: *** [arch/x86/kernel] Error 2
>>> make[1]: *** [arch/x86] Error 2
>> Ugh bugger.
>>
>> I bet its that: &((char *)key)[branch] business, an earlier variant
>> thereof tripped up more recent GCCs too.
>>
>> So its an __always_inline function, and both argument are always compile
>> time constants, @key is the address of an object in static storage (a
>> global) and @branch is a simple 0/1 at the call site.
>>
>> Now we wish to compute (unsigned long)key + branch at compile/link time
>> to feed to the assembler as an immediate, which should be possible,
>> given its all 'constants'.
>>
>> It just appears GCC is having a hard time with this.
>>
>> Let me see if I have a sufficiently old GCC around to play with.
> Could you feed the below to your compiler? Its a bit cumbersome, but
> its the next best I could come up with...
No, it produces the same error. This is Fedora 13, btw, uses gcc 4.4.4.
-boris
>
> ---
> arch/x86/include/asm/jump_label.h | 56 ++++++++++++++++++++++++++++-----------
> 1 file changed, 41 insertions(+), 15 deletions(-)
>
> diff --git a/arch/x86/include/asm/jump_label.h b/arch/x86/include/asm/jump_label.h
> index 28d7a857f9d1..76c769ae4200 100644
> --- a/arch/x86/include/asm/jump_label.h
> +++ b/arch/x86/include/asm/jump_label.h
> @@ -16,15 +16,30 @@
> # define STATIC_KEY_INIT_NOP GENERIC_NOP5_ATOMIC
> #endif
>
> +struct foo {
> + u8 zero;
> + u8 one;
> +};
> +
> static __always_inline bool arch_static_branch(struct static_key *key, bool branch)
> {
> - asm_volatile_goto("1:"
> - ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
> - ".pushsection __jump_table, \"aw\" \n\t"
> - _ASM_ALIGN "\n\t"
> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> - ".popsection \n\t"
> - : : "i" (&((char *)key)[branch]) : : l_yes);
> + if (!branch) {
> + asm_volatile_goto("1:"
> + ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
> + ".pushsection __jump_table, \"aw\" \n\t"
> + _ASM_ALIGN "\n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + ".popsection \n\t"
> + : : "i" (&((struct foo *)key)->zero) : : l_yes);
> + } else {
> + asm_volatile_goto("1:"
> + ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
> + ".pushsection __jump_table, \"aw\" \n\t"
> + _ASM_ALIGN "\n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + ".popsection \n\t"
> + : : "i" (&((struct foo *)key)->one) : : l_yes);
> + }
>
> return false;
> l_yes:
> @@ -33,14 +48,25 @@ static __always_inline bool arch_static_branch(struct static_key *key, bool bran
>
> static __always_inline bool arch_static_branch_jump(struct static_key *key, bool branch)
> {
> - asm_volatile_goto("1:"
> - ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
> - "2:\n\t"
> - ".pushsection __jump_table, \"aw\" \n\t"
> - _ASM_ALIGN "\n\t"
> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> - ".popsection \n\t"
> - : : "i" (&((char *)key)[branch]) : : l_yes);
> + if (!branch) {
> + asm_volatile_goto("1:"
> + ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
> + "2:\n\t"
> + ".pushsection __jump_table, \"aw\" \n\t"
> + _ASM_ALIGN "\n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + ".popsection \n\t"
> + : : "i" (&((struct foo *)key)->zero) : : l_yes);
> + } else {
> + asm_volatile_goto("1:"
> + ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
> + "2:\n\t"
> + ".pushsection __jump_table, \"aw\" \n\t"
> + _ASM_ALIGN "\n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + ".popsection \n\t"
> + : : "i" (&((struct foo *)key)->one) : : l_yes);
> + }
>
> return false;
> l_yes:
On Wed, Aug 12, 2015 at 11:17 AM, Boris Ostrovsky
<[email protected]> wrote:
> On 08/12/2015 01:46 PM, Peter Zijlstra wrote:
>>
>> On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
>>>
>>> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>>>
>>>> Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
>>>> interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red
>>>> Hat
>>>> 4.4.4-2) (GCC)):
>>>>
>>>>
>>>>
>>>> CC arch/x86/kernel/nmi.o
>>>> In file included from
>>>> /home/build/linux-boris/include/linux/jump_label.h:109,
>>>> from
>>>> /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
>>>> from
>>>> /home/build/linux-boris/include/linux/spinlock.h:88,
>>>> from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
>>>> ‘nmi_handle’:
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning:
>>>> asm
>>>> operand 0 probably doesn’t match constraints
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
>>>> impossible constraint in ‘asm’
>>>> make[3]: *** [arch/x86/kernel/nmi.o] Error 1
>>>> make[2]: *** [arch/x86/kernel] Error 2
>>>> make[1]: *** [arch/x86] Error 2
>>>
>>> Ugh bugger.
>>>
>>> I bet its that: &((char *)key)[branch] business, an earlier variant
>>> thereof tripped up more recent GCCs too.
>>>
>>> So its an __always_inline function, and both argument are always compile
>>> time constants, @key is the address of an object in static storage (a
>>> global) and @branch is a simple 0/1 at the call site.
>>>
>>> Now we wish to compute (unsigned long)key + branch at compile/link time
>>> to feed to the assembler as an immediate, which should be possible,
>>> given its all 'constants'.
>>>
>>> It just appears GCC is having a hard time with this.
>>>
>>> Let me see if I have a sufficiently old GCC around to play with.
>>
>> Could you feed the below to your compiler? Its a bit cumbersome, but
>> its the next best I could come up with...
>
>
> No, it produces the same error. This is Fedora 13, btw, uses gcc 4.4.4.
Is the problem just that it's being misdetected as supporting asm
goto? What does gcc -E say?
--Andy
One option might be to do the addition in assembly, i.e.:
"i" (key), "i" (index)
... and put the addition into the assembly source.
On August 12, 2015 11:17:17 AM PDT, Boris Ostrovsky <[email protected]> wrote:
>On 08/12/2015 01:46 PM, Peter Zijlstra wrote:
>> On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
>>> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>>>
>>>> Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
>>>> interface") breaks old-ish compilers (gcc version 4.4.4 20100503
>(Red Hat
>>>> 4.4.4-2) (GCC)):
>>>>
>>>>
>>>>
>>>> CC arch/x86/kernel/nmi.o
>>>> In file included from
>>>> /home/build/linux-boris/include/linux/jump_label.h:109,
>>>> from
>>>> /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
>>>> from
>/home/build/linux-boris/include/linux/spinlock.h:88,
>>>> from
>/home/build/linux-boris/arch/x86/kernel/nmi.c:14:
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In
>function
>>>> ‘nmi_handle’:
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21:
>warning: asm
>>>> operand 0 probably doesn’t match constraints
>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21:
>error:
>>>> impossible constraint in ‘asm’
>>>> make[3]: *** [arch/x86/kernel/nmi.o] Error 1
>>>> make[2]: *** [arch/x86/kernel] Error 2
>>>> make[1]: *** [arch/x86] Error 2
>>> Ugh bugger.
>>>
>>> I bet its that: &((char *)key)[branch] business, an earlier variant
>>> thereof tripped up more recent GCCs too.
>>>
>>> So its an __always_inline function, and both argument are always
>compile
>>> time constants, @key is the address of an object in static storage
>(a
>>> global) and @branch is a simple 0/1 at the call site.
>>>
>>> Now we wish to compute (unsigned long)key + branch at compile/link
>time
>>> to feed to the assembler as an immediate, which should be possible,
>>> given its all 'constants'.
>>>
>>> It just appears GCC is having a hard time with this.
>>>
>>> Let me see if I have a sufficiently old GCC around to play with.
>> Could you feed the below to your compiler? Its a bit cumbersome, but
>> its the next best I could come up with...
>
>No, it produces the same error. This is Fedora 13, btw, uses gcc 4.4.4.
>
>-boris
>
>
>>
>> ---
>> arch/x86/include/asm/jump_label.h | 56
>++++++++++++++++++++++++++++-----------
>> 1 file changed, 41 insertions(+), 15 deletions(-)
>>
>> diff --git a/arch/x86/include/asm/jump_label.h
>b/arch/x86/include/asm/jump_label.h
>> index 28d7a857f9d1..76c769ae4200 100644
>> --- a/arch/x86/include/asm/jump_label.h
>> +++ b/arch/x86/include/asm/jump_label.h
>> @@ -16,15 +16,30 @@
>> # define STATIC_KEY_INIT_NOP GENERIC_NOP5_ATOMIC
>> #endif
>>
>> +struct foo {
>> + u8 zero;
>> + u8 one;
>> +};
>> +
>> static __always_inline bool arch_static_branch(struct static_key
>*key, bool branch)
>> {
>> - asm_volatile_goto("1:"
>> - ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
>> - ".pushsection __jump_table, \"aw\" \n\t"
>> - _ASM_ALIGN "\n\t"
>> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> - ".popsection \n\t"
>> - : : "i" (&((char *)key)[branch]) : : l_yes);
>> + if (!branch) {
>> + asm_volatile_goto("1:"
>> + ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
>> + ".pushsection __jump_table, \"aw\" \n\t"
>> + _ASM_ALIGN "\n\t"
>> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> + ".popsection \n\t"
>> + : : "i" (&((struct foo *)key)->zero) : : l_yes);
>> + } else {
>> + asm_volatile_goto("1:"
>> + ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
>> + ".pushsection __jump_table, \"aw\" \n\t"
>> + _ASM_ALIGN "\n\t"
>> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> + ".popsection \n\t"
>> + : : "i" (&((struct foo *)key)->one) : : l_yes);
>> + }
>>
>> return false;
>> l_yes:
>> @@ -33,14 +48,25 @@ static __always_inline bool
>arch_static_branch(struct static_key *key, bool bran
>>
>> static __always_inline bool arch_static_branch_jump(struct
>static_key *key, bool branch)
>> {
>> - asm_volatile_goto("1:"
>> - ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
>> - "2:\n\t"
>> - ".pushsection __jump_table, \"aw\" \n\t"
>> - _ASM_ALIGN "\n\t"
>> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> - ".popsection \n\t"
>> - : : "i" (&((char *)key)[branch]) : : l_yes);
>> + if (!branch) {
>> + asm_volatile_goto("1:"
>> + ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
>> + "2:\n\t"
>> + ".pushsection __jump_table, \"aw\" \n\t"
>> + _ASM_ALIGN "\n\t"
>> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> + ".popsection \n\t"
>> + : : "i" (&((struct foo *)key)->zero) : : l_yes);
>> + } else {
>> + asm_volatile_goto("1:"
>> + ".byte 0xe9\n\t .long %l[l_yes] - 2f\n\t"
>> + "2:\n\t"
>> + ".pushsection __jump_table, \"aw\" \n\t"
>> + _ASM_ALIGN "\n\t"
>> + _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
>> + ".popsection \n\t"
>> + : : "i" (&((struct foo *)key)->one) : : l_yes);
>> + }
>>
>> return false;
>> l_yes:
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
On 08/12/2015 02:26 PM, Andy Lutomirski wrote:
> On Wed, Aug 12, 2015 at 11:17 AM, Boris Ostrovsky
> <[email protected]> wrote:
>> On 08/12/2015 01:46 PM, Peter Zijlstra wrote:
>>> On Wed, Aug 12, 2015 at 07:21:05PM +0200, Peter Zijlstra wrote:
>>>> On Wed, Aug 12, 2015 at 09:27:38AM -0400, Boris Ostrovsky wrote:
>>>>
>>>>> Incidentally, 11276d53 ("locking/static_keys: Add a new static_key
>>>>> interface") breaks old-ish compilers (gcc version 4.4.4 20100503 (Red
>>>>> Hat
>>>>> 4.4.4-2) (GCC)):
>>>>>
>>>>>
>>>>>
>>>>> CC arch/x86/kernel/nmi.o
>>>>> In file included from
>>>>> /home/build/linux-boris/include/linux/jump_label.h:109,
>>>>> from
>>>>> /home/build/linux-boris/arch/x86/include/asm/spinlock.h:5,
>>>>> from
>>>>> /home/build/linux-boris/include/linux/spinlock.h:88,
>>>>> from /home/build/linux-boris/arch/x86/kernel/nmi.c:14:
>>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h: In function
>>>>> ‘nmi_handle’:
>>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: warning:
>>>>> asm
>>>>> operand 0 probably doesn’t match constraints
>>>>> /home/build/linux-boris/arch/x86/include/asm/jump_label.h:21: error:
>>>>> impossible constraint in ‘asm’
>>>>> make[3]: *** [arch/x86/kernel/nmi.o] Error 1
>>>>> make[2]: *** [arch/x86/kernel] Error 2
>>>>> make[1]: *** [arch/x86] Error 2
>>>> Ugh bugger.
>>>>
>>>> I bet its that: &((char *)key)[branch] business, an earlier variant
>>>> thereof tripped up more recent GCCs too.
>>>>
>>>> So its an __always_inline function, and both argument are always compile
>>>> time constants, @key is the address of an object in static storage (a
>>>> global) and @branch is a simple 0/1 at the call site.
>>>>
>>>> Now we wish to compute (unsigned long)key + branch at compile/link time
>>>> to feed to the assembler as an immediate, which should be possible,
>>>> given its all 'constants'.
>>>>
>>>> It just appears GCC is having a hard time with this.
>>>>
>>>> Let me see if I have a sufficiently old GCC around to play with.
>>> Could you feed the below to your compiler? Its a bit cumbersome, but
>>> its the next best I could come up with...
>>
>> No, it produces the same error. This is Fedora 13, btw, uses gcc 4.4.4.
> Is the problem just that it's being misdetected as supporting asm
> goto? What does gcc -E say?
static inline __attribute__((no_instrument_function))
__attribute__((always_inline)) bool arch_static_branch(struct static_key
*key, bool branch)
{
do { asm goto("1:" ".byte " "0x0f,0x1f,0x44,0x00,0" "\n\t"
".pushsection __jump_table, \"aw\" \n\t" " " ".balign 8" " " "\n\t" " "
".quad" " " "1b, %l[l_yes], %c0 \n\t" ".popsection \n\t" : : "i"
(&((char *)key)[branch]) : : l_yes); asm (""); } while (0)
return false;
l_yes:
return true;
}
-boris
On Wed, Aug 12, 2015 at 11:26:41AM -0700, H. Peter Anvin wrote:
> One option might be to do the addition in assembly, i.e.:
>
> "i" (key), "i" (index)
>
> ... and put the addition into the assembly source.
Like so? Seems to build on gcc-4.6.
---
arch/x86/include/asm/jump_label.h | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/arch/x86/include/asm/jump_label.h b/arch/x86/include/asm/jump_label.h
index 28d7a857f9d1..5daeca3d0f9e 100644
--- a/arch/x86/include/asm/jump_label.h
+++ b/arch/x86/include/asm/jump_label.h
@@ -22,9 +22,9 @@ static __always_inline bool arch_static_branch(struct static_key *key, bool bran
".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
".pushsection __jump_table, \"aw\" \n\t"
_ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ : : "i" (key), "i" (branch) : : l_yes);
return false;
l_yes:
@@ -38,9 +38,9 @@ static __always_inline bool arch_static_branch_jump(struct static_key *key, bool
"2:\n\t"
".pushsection __jump_table, \"aw\" \n\t"
_ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ : : "i" (key), "i" (branch) : : l_yes);
return false;
l_yes:
On 08/12/2015 02:36 PM, Peter Zijlstra wrote:
> On Wed, Aug 12, 2015 at 11:26:41AM -0700, H. Peter Anvin wrote:
>> One option might be to do the addition in assembly, i.e.:
>>
>> "i" (key), "i" (index)
>>
>> ... and put the addition into the assembly source.
> Like so? Seems to build on gcc-4.6.
Yes, this builds on 4.4.4. as well.
-boris
>
> ---
> arch/x86/include/asm/jump_label.h | 8 ++++----
> 1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/include/asm/jump_label.h b/arch/x86/include/asm/jump_label.h
> index 28d7a857f9d1..5daeca3d0f9e 100644
> --- a/arch/x86/include/asm/jump_label.h
> +++ b/arch/x86/include/asm/jump_label.h
> @@ -22,9 +22,9 @@ static __always_inline bool arch_static_branch(struct static_key *key, bool bran
> ".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
> ".pushsection __jump_table, \"aw\" \n\t"
> _ASM_ALIGN "\n\t"
> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
> ".popsection \n\t"
> - : : "i" (&((char *)key)[branch]) : : l_yes);
> + : : "i" (key), "i" (branch) : : l_yes);
>
> return false;
> l_yes:
> @@ -38,9 +38,9 @@ static __always_inline bool arch_static_branch_jump(struct static_key *key, bool
> "2:\n\t"
> ".pushsection __jump_table, \"aw\" \n\t"
> _ASM_ALIGN "\n\t"
> - _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> + _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
> ".popsection \n\t"
> - : : "i" (&((char *)key)[branch]) : : l_yes);
> + : : "i" (key), "i" (branch) : : l_yes);
>
> return false;
> l_yes:
On Wed, Aug 12, 2015 at 02:44:52PM -0400, Boris Ostrovsky wrote:
> On 08/12/2015 02:36 PM, Peter Zijlstra wrote:
> >On Wed, Aug 12, 2015 at 11:26:41AM -0700, H. Peter Anvin wrote:
> >>One option might be to do the addition in assembly, i.e.:
> >>
> >>"i" (key), "i" (index)
> >>
> >>... and put the addition into the assembly source.
> >Like so? Seems to build on gcc-4.6.
>
> Yes, this builds on 4.4.4. as well.
Thanks, Ingo / Hpa, can you make the below patch appear in the right tip
branch?
---
Subject: jump_label, x86: Fix asm for older GCCs
From: Peter Zijlstra <[email protected]>
Date: Wed Aug 12 20:58:48 CEST 2015
Boris reported that: gcc version 4.4.4 20100503 (Red Hat 4.4.4-2)
fails to build. The problem appears that even though @key and @branch
are compile time constants it doesn't see: &((char *)key)[branch]
as an immediate. More recent GCCs don't appear to have this problem.
In particular, Red Hat backported the asm goto feature into 4.4,
'normal' 4.4 compilers will not have this feature and thus not run
into this asm.
The solution is to supply both values to the asm as immediates and do
the addition in asm.
Suggested-by: "H. Peter Anvin" <[email protected]>
Reported-by: Boris Ostrovsky <[email protected]>
Tested-by: Boris Ostrovsky <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
---
arch/x86/include/asm/jump_label.h | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
--- a/arch/x86/include/asm/jump_label.h
+++ b/arch/x86/include/asm/jump_label.h
@@ -22,9 +22,9 @@ static __always_inline bool arch_static_
".byte " __stringify(STATIC_KEY_INIT_NOP) "\n\t"
".pushsection __jump_table, \"aw\" \n\t"
_ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ : : "i" (key), "i" (branch) : : l_yes);
return false;
l_yes:
@@ -38,9 +38,9 @@ static __always_inline bool arch_static_
"2:\n\t"
".pushsection __jump_table, \"aw\" \n\t"
_ASM_ALIGN "\n\t"
- _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
+ _ASM_PTR "1b, %l[l_yes], %c0 + %c1 \n\t"
".popsection \n\t"
- : : "i" (&((char *)key)[branch]) : : l_yes);
+ : : "i" (key), "i" (branch) : : l_yes);
return false;
l_yes:
* Peter Zijlstra <[email protected]> wrote:
> On Wed, Aug 12, 2015 at 02:44:52PM -0400, Boris Ostrovsky wrote:
> > On 08/12/2015 02:36 PM, Peter Zijlstra wrote:
> > >On Wed, Aug 12, 2015 at 11:26:41AM -0700, H. Peter Anvin wrote:
> > >>One option might be to do the addition in assembly, i.e.:
> > >>
> > >>"i" (key), "i" (index)
> > >>
> > >>... and put the addition into the assembly source.
> > >Like so? Seems to build on gcc-4.6.
> >
> > Yes, this builds on 4.4.4. as well.
>
> Thanks, Ingo / Hpa, can you make the below patch appear in the right tip
> branch?
Applied, it will show up as the following commit:
d420acd816c0 ("jump_label/x86: Work around asm build bug on older/backported GCCs")
Thanks,
Ingo