[PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Jin Dongming]

acpi_pre_map() is used for remapping the physical address of I/O, so
it should be return NULL or remapped virtual address. The problem
is whether I/O remapping is successful or not, the function returns
NULL always.

In acpi_pre_map(), after the physical address is remapped sucessfully,
it will check whether the physical address has been added into acpi_iomaps
list again. If the physical address has beed added into acpi_iomaps,
the virtual address will be saved in vaddr. Otherwise, NULL be saved in
vaddr.

So if the physical address has never been remapped, the return value of
acpi_pre_map will be NULL always.

This patch fixed it and I confirmed it on x86_64 next-tree.

Signed-off-by: Jin Dongming <jin.dongming@np.css.fujitsu.com>
---
 drivers/acpi/atomicio.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/acpi/atomicio.c b/drivers/acpi/atomicio.c
index 8f8bd73..1bc2614 100644
--- a/drivers/acpi/atomicio.c
+++ b/drivers/acpi/atomicio.c
@@ -133,7 +133,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
 
 	spin_lock_irqsave(&acpi_iomaps_lock, flags);
 	vaddr = __acpi_try_ioremap(paddr, size);
-	if (vaddr) {
+	if (unlikely(vaddr)) {
 		spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
 		iounmap(map->vaddr);
 		kfree(map);
@@ -142,7 +142,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
 	list_add_tail_rcu(&map->list, &acpi_iomaps);
 	spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
 
-	return vaddr + (paddr - pg_off);
+	return map->vaddr + (paddr - pg_off);
 err_unmap:
 	iounmap(vaddr);
 	return NULL;
-- 
1.7.1.1

Re: [PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Huang Ying]

On Tue, 2010-08-17 at 08:56 +0800, Jin Dongming wrote:
> acpi_pre_map() is used for remapping the physical address of I/O, so
> it should be return NULL or remapped virtual address. The problem
> is whether I/O remapping is successful or not, the function returns
> NULL always.

No. NULL is returned for error path only. Please check the code again.

> In acpi_pre_map(), after the physical address is remapped sucessfully,
> it will check whether the physical address has been added into acpi_iomaps
> list again. If the physical address has beed added into acpi_iomaps,
> the virtual address will be saved in vaddr. Otherwise, NULL be saved in
> vaddr.
> 
> So if the physical address has never been remapped, the return value of
> acpi_pre_map will be NULL always.
> 
> This patch fixed it and I confirmed it on x86_64 next-tree.
> 
> Signed-off-by: Jin Dongming <jin.dongming@np.css.fujitsu.com>
> ---
>  drivers/acpi/atomicio.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/acpi/atomicio.c b/drivers/acpi/atomicio.c
> index 8f8bd73..1bc2614 100644
> --- a/drivers/acpi/atomicio.c
> +++ b/drivers/acpi/atomicio.c
> @@ -133,7 +133,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
>  
>  	spin_lock_irqsave(&acpi_iomaps_lock, flags);
>  	vaddr = __acpi_try_ioremap(paddr, size);
> -	if (vaddr) {
> +	if (unlikely(vaddr)) {

pre_map is not performance critical, so "unlikely" here helps little.

>  		spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
>  		iounmap(map->vaddr);
>  		kfree(map);
> @@ -142,7 +142,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
>  	list_add_tail_rcu(&map->list, &acpi_iomaps);
>  	spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
>  
> -	return vaddr + (paddr - pg_off);
> +	return map->vaddr + (paddr - pg_off);
>  err_unmap:
>  	iounmap(vaddr);
>  	return NULL;

When will vaddr != map->vaddr?

Best Regards,
Huang Ying

Re: [PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Jin Dongming]

(2010/08/17 10:37), Huang Ying wrote:
> On Tue, 2010-08-17 at 08:56 +0800, Jin Dongming wrote:
>> acpi_pre_map() is used for remapping the physical address of I/O, so
>> it should be return NULL or remapped virtual address. The problem
>> is whether I/O remapping is successful or not, the function returns
>> NULL always.
> 
> No. NULL is returned for error path only. Please check the code again.

There three places used the local variable vaddr in acpi_pre_map() in next-tree.
   1. 115         vaddr = __acpi_try_ioremap(paddr, size);
   2. 122         vaddr = ioremap(pg_off, pg_sz);
   3. 135         vaddr = __acpi_try_ioremap(paddr, size);

Let's think about the following statement.
   Assumption: the physical address has never been remapped.
   Step:
       1. vaddr == NULL
          Because the physical address is not registered in the acpi_iomaps,
          it should be returned NULL from __acpi_try_ioremap().

       2. vaddr == the virtual address of the physical address.
          Here if ioremap is successful, the value of vaddr should be 
          the virtual address returned from ioremap().

       3. vaddr == NULL                <== IMPORTANT
          Here it is because the physical address has not been registered
          in the acpi_iomaps yet, it still return NULL from __acpi_try_ioremap().
          So it is why vaddr == NULL, even if the physical address has never
          been remapped.

    Result: vaddr == NULL.

And if the vaddr is not NULL, it could not be added into acpi_iomaps.
Codes in acpi_pre_map() is like following.

134         spin_lock_irqsave(&acpi_iomaps_lock, flags);
135         vaddr = __acpi_try_ioremap(paddr, size);           <== the 3rd step
136         if (vaddr) {
137                 spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
138                 iounmap(map->vaddr);
139                 kfree(map);
140                 return vaddr;
141         }
142         list_add_tail_rcu(&map->list, &acpi_iomaps);      <== add into acpi_iomaps.
143         spin_unlock_irqrestore(&acpi_iomaps_lock, flags);

> 
>> In acpi_pre_map(), after the physical address is remapped sucessfully,
>> it will check whether the physical address has been added into acpi_iomaps
>> list again. If the physical address has beed added into acpi_iomaps,
>> the virtual address will be saved in vaddr. Otherwise, NULL be saved in
>> vaddr.
>>
>> So if the physical address has never been remapped, the return value of
>> acpi_pre_map will be NULL always.
>>
>> This patch fixed it and I confirmed it on x86_64 next-tree.
>>
>> Signed-off-by: Jin Dongming <jin.dongming@np.css.fujitsu.com>
>> ---
>>  drivers/acpi/atomicio.c |    4 ++--
>>  1 files changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/acpi/atomicio.c b/drivers/acpi/atomicio.c
>> index 8f8bd73..1bc2614 100644
>> --- a/drivers/acpi/atomicio.c
>> +++ b/drivers/acpi/atomicio.c
>> @@ -133,7 +133,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
>>  
>>  	spin_lock_irqsave(&acpi_iomaps_lock, flags);
>>  	vaddr = __acpi_try_ioremap(paddr, size);
>> -	if (vaddr) {
>> +	if (unlikely(vaddr)) {
> 
> pre_map is not performance critical, so "unlikely" here helps little.
> 
>>  		spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
>>  		iounmap(map->vaddr);
>>  		kfree(map);
>> @@ -142,7 +142,7 @@ static void __iomem *acpi_pre_map(phys_addr_t paddr,
>>  	list_add_tail_rcu(&map->list, &acpi_iomaps);
>>  	spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
>>  
>> -	return vaddr + (paddr - pg_off);
>> +	return map->vaddr + (paddr - pg_off);
>>  err_unmap:
>>  	iounmap(vaddr);
>>  	return NULL;
> 
> When will vaddr != map->vaddr?
> 
> Best Regards,
> Huang Ying
> 
> 
> 
> 

Re: [PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Huang Ying]

On Tue, 2010-08-17 at 10:43 +0800, Jin Dongming wrote:
> (2010/08/17 10:37), Huang Ying wrote:
> > On Tue, 2010-08-17 at 08:56 +0800, Jin Dongming wrote:
> >> acpi_pre_map() is used for remapping the physical address of I/O, so
> >> it should be return NULL or remapped virtual address. The problem
> >> is whether I/O remapping is successful or not, the function returns
> >> NULL always.
> > 
> > No. NULL is returned for error path only. Please check the code again.
> 
> There three places used the local variable vaddr in acpi_pre_map() in next-tree.
>    1. 115         vaddr = __acpi_try_ioremap(paddr, size);
>    2. 122         vaddr = ioremap(pg_off, pg_sz);
>    3. 135         vaddr = __acpi_try_ioremap(paddr, size);
> 
> Let's think about the following statement.
>    Assumption: the physical address has never been remapped.
>    Step:
>        1. vaddr == NULL
>           Because the physical address is not registered in the acpi_iomaps,
>           it should be returned NULL from __acpi_try_ioremap().
> 
>        2. vaddr == the virtual address of the physical address.
>           Here if ioremap is successful, the value of vaddr should be 
>           the virtual address returned from ioremap().
> 
>        3. vaddr == NULL                <== IMPORTANT
>           Here it is because the physical address has not been registered
>           in the acpi_iomaps yet, it still return NULL from __acpi_try_ioremap().
>           So it is why vaddr == NULL, even if the physical address has never
>           been remapped.
> 
>     Result: vaddr == NULL.

return vaddr + (paddr - pg_off), is not NULL for common cases.

> And if the vaddr is not NULL, it could not be added into acpi_iomaps.
> Codes in acpi_pre_map() is like following.
> 
> 134         spin_lock_irqsave(&acpi_iomaps_lock, flags);
> 135         vaddr = __acpi_try_ioremap(paddr, size);           <== the 3rd step
> 136         if (vaddr) {
> 137                 spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
> 138                 iounmap(map->vaddr);
> 139                 kfree(map);
> 140                 return vaddr;
> 141         }
> 142         list_add_tail_rcu(&map->list, &acpi_iomaps);      <== add into acpi_iomaps.
> 143         spin_unlock_irqrestore(&acpi_iomaps_lock, flags);

Oops, this is a bug, and your patch really fix it. But I think you
should change the patch description. NULL is not returned in almost all
cases. Because (paddr - pg_off) is not zero in common cases.

Best Regards,
Huang Ying

Re: [PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Jin Dongming]

(2010/08/17 13:15), Huang Ying wrote:
> On Tue, 2010-08-17 at 10:43 +0800, Jin Dongming wrote:
>> (2010/08/17 10:37), Huang Ying wrote:
>>> On Tue, 2010-08-17 at 08:56 +0800, Jin Dongming wrote:
>>>> acpi_pre_map() is used for remapping the physical address of I/O, so
>>>> it should be return NULL or remapped virtual address. The problem
>>>> is whether I/O remapping is successful or not, the function returns
>>>> NULL always.
>>>
>>> No. NULL is returned for error path only. Please check the code again.
>>
>> There three places used the local variable vaddr in acpi_pre_map() in next-tree.
>>    1. 115         vaddr = __acpi_try_ioremap(paddr, size);
>>    2. 122         vaddr = ioremap(pg_off, pg_sz);
>>    3. 135         vaddr = __acpi_try_ioremap(paddr, size);
>>
>> Let's think about the following statement.
>>    Assumption: the physical address has never been remapped.
>>    Step:
>>        1. vaddr == NULL
>>           Because the physical address is not registered in the acpi_iomaps,
>>           it should be returned NULL from __acpi_try_ioremap().
>>
>>        2. vaddr == the virtual address of the physical address.
>>           Here if ioremap is successful, the value of vaddr should be 
>>           the virtual address returned from ioremap().
>>
>>        3. vaddr == NULL                <== IMPORTANT
>>           Here it is because the physical address has not been registered
>>           in the acpi_iomaps yet, it still return NULL from __acpi_try_ioremap().
>>           So it is why vaddr == NULL, even if the physical address has never
>>           been remapped.
>>
>>     Result: vaddr == NULL.
> 
> return vaddr + (paddr - pg_off), is not NULL for common cases.
> 
>> And if the vaddr is not NULL, it could not be added into acpi_iomaps.
>> Codes in acpi_pre_map() is like following.
>>
>> 134         spin_lock_irqsave(&acpi_iomaps_lock, flags);
>> 135         vaddr = __acpi_try_ioremap(paddr, size);           <== the 3rd step
>> 136         if (vaddr) {
>> 137                 spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
>> 138                 iounmap(map->vaddr);
>> 139                 kfree(map);
>> 140                 return vaddr;
>> 141         }
>> 142         list_add_tail_rcu(&map->list, &acpi_iomaps);      <== add into acpi_iomaps.
>> 143         spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
> 
> Oops, this is a bug, and your patch really fix it. But I think you
> should change the patch description. NULL is not returned in almost all
> cases. Because (paddr - pg_off) is not zero in common cases.
> 
> Best Regards,
> Huang Ying
> 
OK. I will modify the description of this patch and resend it.

Best Regards,
Jin Dongming
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> 
>