2014-06-05 22:19:53

by Pranith Kumar

Subject: [PATCH 1/1] use documented cmpxchg api

use the documented atomic_cmpxchg instead of __cmpxchg64

This kills the last user of said API in drivers code.


Signed-off-by: Pranith Kumar <[email protected]>
---
drivers/iommu/intel-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 6bb3277..270113f 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
return pte->val & VTD_PAGE_MASK;
#else
/* Must have a full atomic 64-bit read */
- return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
+ return atomic_cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
#endif
}

--
1.7.9.5

2014-06-06 00:34:37

by Peter Hurley

Subject: Re: [PATCH 1/1] use documented cmpxchg api

On 06/05/2014 06:19 PM, Pranith Kumar wrote:
> use the documented atomic_cmpxchg instead of __cmpxchg64
>
> This kills the last user of said API in drivers code.
>
>
> Signed-off-by: Pranith Kumar <[email protected]>
> ---
> drivers/iommu/intel-iommu.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
> index 6bb3277..270113f 100644
> --- a/drivers/iommu/intel-iommu.c
> +++ b/drivers/iommu/intel-iommu.c
> @@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
> return pte->val & VTD_PAGE_MASK;
> #else
> /* Must have a full atomic 64-bit read */
> - return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
> + return atomic_cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;

This is not equivalent.

The __cmpxchg64() is specifically being used in this case
when !CONFIG_64BIT for a full 64-bit RMW, whereas atomic_cmpxchg()
uses ints which would be 32-bit RMW.

Regards,
Peter Hurley

2014-06-06 01:03:45

by Pranith Kumar

Subject: Re: [PATCH 1/1] use documented cmpxchg api

On 06/05/2014 08:34 PM, Peter Hurley wrote:
> On 06/05/2014 06:19 PM, Pranith Kumar wrote:
>> use the documented atomic_cmpxchg instead of __cmpxchg64
>>
>> This kills the last user of said API in drivers code.
>>
>>
>> Signed-off-by: Pranith Kumar <[email protected]>
>> ---
>> drivers/iommu/intel-iommu.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
>> index 6bb3277..270113f 100644
>> --- a/drivers/iommu/intel-iommu.c
>> +++ b/drivers/iommu/intel-iommu.c
>> @@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
>> return pte->val & VTD_PAGE_MASK;
>> #else
>> /* Must have a full atomic 64-bit read */
>> - return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
>> + return atomic_cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
>
> This is not equivalent.
>
> The __cmpxchg64() is specifically being used in this case
> when !CONFIG_64BIT for a full 64-bit RMW, whereas atomic_cmpxchg()
> uses ints which would be 32-bit RMW.
>


You are right! The previous patch is wrong in that atomic_cmpxchg needs an atomic_t argument.
cmpxchg() handles the size internally using typeof() to figure out the width of RMW.
There is also an explicit cmpxchg64() which does the same thing.

Fixed with the following v2:

use the documented cmpxchg instead of __cmpxchg64

This kills the last user of said API in drivers code.

Signed-off-by: Pranith Kumar <[email protected]>
---
drivers/iommu/intel-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 6bb3277..270113f 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
return pte->val & VTD_PAGE_MASK;
#else
/* Must have a full atomic 64-bit read */
- return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
+ return cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
#endif
}

--
1.7.9.5

2014-06-06 10:55:40

by Peter Hurley

Subject: Re: [PATCH 1/1] use documented cmpxchg api

On 06/05/2014 09:03 PM, Pranith Kumar wrote:
> On 06/05/2014 08:34 PM, Peter Hurley wrote:
>> On 06/05/2014 06:19 PM, Pranith Kumar wrote:
>>> use the documented atomic_cmpxchg instead of __cmpxchg64
>>>
>>> This kills the last user of said API in drivers code.
>>>
>>>
>>> Signed-off-by: Pranith Kumar <[email protected]>
>>> ---
>>> drivers/iommu/intel-iommu.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
>>> index 6bb3277..270113f 100644
>>> --- a/drivers/iommu/intel-iommu.c
>>> +++ b/drivers/iommu/intel-iommu.c
>>> @@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
>>> return pte->val & VTD_PAGE_MASK;
>>> #else
>>> /* Must have a full atomic 64-bit read */
>>> - return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
>>> + return atomic_cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
>>
>> This is not equivalent.
>>
>> The __cmpxchg64() is specifically being used in this case
>> when !CONFIG_64BIT for a full 64-bit RMW, whereas atomic_cmpxchg()
>> uses ints which would be 32-bit RMW.
>>
>
>
> You are right! The previous patch is wrong in that atomic_cmpxchg needs an atomic_t argument.
> cmpxchg() handles the size internally using typeof() to figure out the width of RMW.
> There is also an explicit cmpxchg64() which does the same thing.


This won't work either, and should generate a compile or link error
on CONFIG_X86_32.

The auto-sizing done by __cmpxchg() does not substitute the cmpxchg8b
instruction for the cmpxchg instruction. __X86_CASE_Q is defined as -1 for
CONFIG_X86_32 so the 8-byte size variant is dead-code, and calls
__cmpxchg_wrong_size() instead.


> Fixed with the following v2:
>
> use the documented cmpxchg instead of __cmpxchg64
>
> This kills the last user of said API in drivers code.
>
> Signed-off-by: Pranith Kumar <[email protected]>
> ---
> drivers/iommu/intel-iommu.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
> index 6bb3277..270113f 100644
> --- a/drivers/iommu/intel-iommu.c
> +++ b/drivers/iommu/intel-iommu.c
> @@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
> return pte->val & VTD_PAGE_MASK;
> #else
> /* Must have a full atomic 64-bit read */
> - return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
> + return cmpxchg(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
> #endif
> }

2014-06-06 14:27:28

by Pranith Kumar

Subject: Re: [PATCH 1/1] use documented cmpxchg api

On 06/06/2014 06:55 AM, Peter Hurley wrote:
>
> This won't work either, and should generate a compile or link error
> on CONFIG_X86_32.
>
> The auto-sizing done by __cmpxchg() does not substitute the cmpxchg8b
> instruction for the cmpxchg instruction. __X86_CASE_Q is defined as -1 for
> CONFIG_X86_32 so the 8-byte size variant is dead-code, and calls
> __cmpxchg_wrong_size() instead.
>
>
Fixed with the following v3:

use the documented cmpxchg64 instead of __cmpxchg64.

This kills the last user of said API in drivers code.

Signed-off-by: Pranith Kumar <[email protected]>
---
drivers/iommu/intel-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 6bb3277..270113f 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -293,7 +293,7 @@ static inline u64 dma_pte_addr(struct dma_pte *pte)
return pte->val & VTD_PAGE_MASK;
#else
/* Must have a full atomic 64-bit read */
- return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
+ return cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
#endif
}

-- 1.7.9.5