[PATCH v2] drm/radeon: have the callers of set_memory_*() check the return value

[PATCH v2] drm/radeon: have the callers of set_memory_*() check the return value

[Tianlin Li]

Right now several architectures allow their set_memory_*() family of  
functions to fail, but callers may not be checking the return values.
If set_memory_*() returns with an error, call-site assumptions may be
infact wrong to assume that it would either succeed or not succeed at  
all. Ideally, the failure of set_memory_*() should be passed up the 
call stack, and callers should examine the failure and deal with it. 

Need to fix the callers and add the __must_check attribute. They also 
may not provide any level of atomicity, in the sense that the memory 
protections may be left incomplete on failure. This issue likely has a 
few steps on effects architectures:
1)Have all callers of set_memory_*() helpers check the return value.
2)Add __must_check to all set_memory_*() helpers so that new uses do  
not ignore the return value.
3)Add atomicity to the calls so that the memory protections aren't left 
in a partial state.

This series is part of step 1. Make drm/radeon check the return value of  
set_memory_*().

Signed-off-by: Tianlin Li <tli@digitalocean.com>
---
v2:
The hardware is too old to be tested on and the code cannot be simply
removed from the kernel, so this is the solution for the short term. 
- Just print an error when something goes wrong
- Remove patch 2.  
v1:
https://lore.kernel.org/lkml/20200107192555.20606-1-tli@digitalocean.com/
---
 drivers/gpu/drm/radeon/radeon_gart.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
index f178ba321715..a2cc864aa08d 100644
--- a/drivers/gpu/drm/radeon/radeon_gart.c
+++ b/drivers/gpu/drm/radeon/radeon_gart.c
@@ -80,8 +80,9 @@ int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
 #ifdef CONFIG_X86
 	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
 	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
-		set_memory_uc((unsigned long)ptr,
-			      rdev->gart.table_size >> PAGE_SHIFT);
+		if (set_memory_uc((unsigned long)ptr,
+			      rdev->gart.table_size >> PAGE_SHIFT))
+			DRM_ERROR("set_memory_uc failed.\n");
 	}
 #endif
 	rdev->gart.ptr = ptr;
@@ -106,8 +107,9 @@ void radeon_gart_table_ram_free(struct radeon_device *rdev)
 #ifdef CONFIG_X86
 	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
 	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
-		set_memory_wb((unsigned long)rdev->gart.ptr,
-			      rdev->gart.table_size >> PAGE_SHIFT);
+		if (set_memory_wb((unsigned long)rdev->gart.ptr,
+			      rdev->gart.table_size >> PAGE_SHIFT))
+			DRM_ERROR("set_memory_wb failed.\n");
 	}
 #endif
 	pci_free_consistent(rdev->pdev, rdev->gart.table_size,
-- 
2.17.1

Re: [PATCH v2] drm/radeon: have the callers of set_memory_*() check the return value

[Christian König]

Am 03.02.20 um 17:18 schrieb Tianlin Li:
> Right now several architectures allow their set_memory_*() family of
> functions to fail,

Oh, that is a detail I previously didn't recognized. Which architectures 
are that?

Cause the RS400/480, RS690 and RS740 which are affected by this are 
integrated in the south-bridge.

>   but callers may not be checking the return values.
> If set_memory_*() returns with an error, call-site assumptions may be
> infact wrong to assume that it would either succeed or not succeed at
> all. Ideally, the failure of set_memory_*() should be passed up the
> call stack, and callers should examine the failure and deal with it.
>
> Need to fix the callers and add the __must_check attribute. They also
> may not provide any level of atomicity, in the sense that the memory
> protections may be left incomplete on failure. This issue likely has a
> few steps on effects architectures:
> 1)Have all callers of set_memory_*() helpers check the return value.
> 2)Add __must_check to all set_memory_*() helpers so that new uses do
> not ignore the return value.
> 3)Add atomicity to the calls so that the memory protections aren't left
> in a partial state.
>
> This series is part of step 1. Make drm/radeon check the return value of
> set_memory_*().
>
> Signed-off-by: Tianlin Li <tli@digitalocean.com>

Reviewed-by: Christian König <christian.koenig@amd.com>

> ---
> v2:
> The hardware is too old to be tested on and the code cannot be simply
> removed from the kernel, so this is the solution for the short term.
> - Just print an error when something goes wrong
> - Remove patch 2.
> v1:
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Flkml%2F20200107192555.20606-1-tli%40digitalocean.com%2F&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7Cba2176d2ca834214e6b108d7a8c4bb1d%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163435227030235&amp;sdata=mDhUEi3vmxahjsdrZOr83OEIWNBHefO8lkXST%2FW32CE%3D&amp;reserved=0
> ---
>   drivers/gpu/drm/radeon/radeon_gart.c | 10 ++++++----
>   1 file changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
> index f178ba321715..a2cc864aa08d 100644
> --- a/drivers/gpu/drm/radeon/radeon_gart.c
> +++ b/drivers/gpu/drm/radeon/radeon_gart.c
> @@ -80,8 +80,9 @@ int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
>   #ifdef CONFIG_X86
>   	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
>   	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
> -		set_memory_uc((unsigned long)ptr,
> -			      rdev->gart.table_size >> PAGE_SHIFT);
> +		if (set_memory_uc((unsigned long)ptr,
> +			      rdev->gart.table_size >> PAGE_SHIFT))
> +			DRM_ERROR("set_memory_uc failed.\n");
>   	}
>   #endif
>   	rdev->gart.ptr = ptr;
> @@ -106,8 +107,9 @@ void radeon_gart_table_ram_free(struct radeon_device *rdev)
>   #ifdef CONFIG_X86
>   	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
>   	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
> -		set_memory_wb((unsigned long)rdev->gart.ptr,
> -			      rdev->gart.table_size >> PAGE_SHIFT);
> +		if (set_memory_wb((unsigned long)rdev->gart.ptr,
> +			      rdev->gart.table_size >> PAGE_SHIFT))
> +			DRM_ERROR("set_memory_wb failed.\n");
>   	}
>   #endif
>   	pci_free_consistent(rdev->pdev, rdev->gart.table_size,

Re: [PATCH v2] drm/radeon: have the callers of set_memory_*() check the return value

[Tianlin Li]

[-- Attachment #1: Type: text/plain, Size: 4216 bytes --]


> On Feb 3, 2020, at 11:16 AM, Christian König <christian.koenig@amd.com> wrote:
> 
> Am 03.02.20 um 17:18 schrieb Tianlin Li:
>> Right now several architectures allow their set_memory_*() family of
>> functions to fail,
> 
> Oh, that is a detail I previously didn't recognized. Which architectures are that?
> 
> Cause the RS400/480, RS690 and RS740 which are affected by this are integrated in the south-bridge.

At least x86 is. 
Some details: https://lore.kernel.org/netdev/20180628213459.28631-4-daniel@iogearbox.net/ <https://lore.kernel.org/netdev/20180628213459.28631-4-daniel@iogearbox.net/>

>>  but callers may not be checking the return values.
>> If set_memory_*() returns with an error, call-site assumptions may be
>> infact wrong to assume that it would either succeed or not succeed at
>> all. Ideally, the failure of set_memory_*() should be passed up the
>> call stack, and callers should examine the failure and deal with it.
>> 
>> Need to fix the callers and add the __must_check attribute. They also
>> may not provide any level of atomicity, in the sense that the memory
>> protections may be left incomplete on failure. This issue likely has a
>> few steps on effects architectures:
>> 1)Have all callers of set_memory_*() helpers check the return value.
>> 2)Add __must_check to all set_memory_*() helpers so that new uses do
>> not ignore the return value.
>> 3)Add atomicity to the calls so that the memory protections aren't left
>> in a partial state.
>> 
>> This series is part of step 1. Make drm/radeon check the return value of
>> set_memory_*().
>> 
>> Signed-off-by: Tianlin Li <tli@digitalocean.com>
> 
> Reviewed-by: Christian König <christian.koenig@amd.com <mailto:christian.koenig@amd.com>>
> 
>> ---
>> v2:
>> The hardware is too old to be tested on and the code cannot be simply
>> removed from the kernel, so this is the solution for the short term.
>> - Just print an error when something goes wrong
>> - Remove patch 2.
>> v1:
>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Flkml%2F20200107192555.20606-1-tli%40digitalocean.com%2F&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7Cba2176d2ca834214e6b108d7a8c4bb1d%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163435227030235&amp;sdata=mDhUEi3vmxahjsdrZOr83OEIWNBHefO8lkXST%2FW32CE%3D&amp;reserved=0 <https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Flkml%2F20200107192555.20606-1-tli%40digitalocean.com%2F&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7Cba2176d2ca834214e6b108d7a8c4bb1d%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163435227030235&amp;sdata=mDhUEi3vmxahjsdrZOr83OEIWNBHefO8lkXST%2FW32CE%3D&amp;reserved=0>
>> ---
>>  drivers/gpu/drm/radeon/radeon_gart.c | 10 ++++++----
>>  1 file changed, 6 insertions(+), 4 deletions(-)
>> 
>> diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
>> index f178ba321715..a2cc864aa08d 100644
>> --- a/drivers/gpu/drm/radeon/radeon_gart.c
>> +++ b/drivers/gpu/drm/radeon/radeon_gart.c
>> @@ -80,8 +80,9 @@ int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
>>  #ifdef CONFIG_X86
>>  	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
>>  	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
>> -		set_memory_uc((unsigned long)ptr,
>> -			      rdev->gart.table_size >> PAGE_SHIFT);
>> +		if (set_memory_uc((unsigned long)ptr,
>> +			      rdev->gart.table_size >> PAGE_SHIFT))
>> +			DRM_ERROR("set_memory_uc failed.\n");
>>  	}
>>  #endif
>>  	rdev->gart.ptr = ptr;
>> @@ -106,8 +107,9 @@ void radeon_gart_table_ram_free(struct radeon_device *rdev)
>>  #ifdef CONFIG_X86
>>  	if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
>>  	    rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
>> -		set_memory_wb((unsigned long)rdev->gart.ptr,
>> -			      rdev->gart.table_size >> PAGE_SHIFT);
>> +		if (set_memory_wb((unsigned long)rdev->gart.ptr,
>> +			      rdev->gart.table_size >> PAGE_SHIFT))
>> +			DRM_ERROR("set_memory_wb failed.\n");
>>  	}
>>  #endif
>>  	pci_free_consistent(rdev->pdev, rdev->gart.table_size,


[-- Attachment #2: Type: text/html, Size: 14177 bytes --]