Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

[Robin Murphy]

On 2022-08-16 13:39, Marek Szyprowski wrote:
> Hi,
> 
> On 18.07.2022 08:28, Liu Song wrote:
>> From: Liu Song <liusong@linux.alibaba.com>
>>
>> In "dma_alloc_from_dev_coherent" and "dma_direct_alloc",
>> the allocated memory is explicitly set to 0.
>>
>> A helper function "use_dev_coherent_memory" is introduced here to
>> determine whether the memory is allocated by "dma_alloc_from_dev_coherent".
>>
>> And use "get_dma_ops" to determine whether the memory is allocated by
>> "dma_direct_alloc".
>>
>> After this modification, memory allocated using "dma_pool_zalloc" can avoid
>> duplicate memset.
>>
>> Signed-off-by: Liu Song <liusong@linux.alibaba.com>
> 
> This patch landed linux next-20220816. Unfortunately it causes serious
> issues on ARM 32bit systems. I've observed it on ARM 32bit Samsung
> Exynos 5422 based Odroid XU4 board with USB r8152 driver. After applying
> this patch and loading r8152 driver I only the following endless
> messages in the log:
> 
> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?
> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?
> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?
> 
> It looks that there are drivers which rely on the fact that the dma
> coherent buffers are always zeroed.

It's not even that, the change here is just obviously broken, since it 
ends up entirely ignoring want_init_on_alloc() for devices using 
dma-direct. Sure, the memory backing a dma_page is zeroed *once* by its 
initial dma-coherent allocation, but who says we're not not reallocating 
pool entries from an existing dma_page?

I'm not convinced it's worth trying to special-case this at all, since 
we can only do it reliably for the first pool entry allocated from a new 
dma_page, and that will only happen as the pool initially grows to a 
suitable size for its working set, after which no further new pages are 
likely to be allocated for the lifetime of the pool. Even if there is a 
case to be made for doing so, it would need to be based on the flow 
through dma_pool_alloc() itself, not some nonsense heuristic on the device.

Andrew, please drop this patch.

Thanks,
Robin.

>> ---
>>    include/linux/dma-map-ops.h | 5 +++++
>>    mm/dmapool.c                | 5 ++++-
>>    2 files changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/include/linux/dma-map-ops.h b/include/linux/dma-map-ops.h
>> index 0d5b06b..c29948d 100644
>> --- a/include/linux/dma-map-ops.h
>> +++ b/include/linux/dma-map-ops.h
>> @@ -171,6 +171,10 @@ int dma_alloc_from_dev_coherent(struct device *dev, ssize_t size,
>>    int dma_release_from_dev_coherent(struct device *dev, int order, void *vaddr);
>>    int dma_mmap_from_dev_coherent(struct device *dev, struct vm_area_struct *vma,
>>    		void *cpu_addr, size_t size, int *ret);
>> +static inline bool use_dev_coherent_memory(struct device *dev)
>> +{
>> +	return dev->dma_mem ? true : false;
>> +}
>>    #else
>>    static inline int dma_declare_coherent_memory(struct device *dev,
>>    		phys_addr_t phys_addr, dma_addr_t device_addr, size_t size)
>> @@ -180,6 +184,7 @@ static inline int dma_declare_coherent_memory(struct device *dev,
>>    #define dma_alloc_from_dev_coherent(dev, size, handle, ret) (0)
>>    #define dma_release_from_dev_coherent(dev, order, vaddr) (0)
>>    #define dma_mmap_from_dev_coherent(dev, vma, vaddr, order, ret) (0)
>> +#define use_dev_coherent_memory(dev) (0)
>>    #endif /* CONFIG_DMA_DECLARE_COHERENT */
>>    
>>    #ifdef CONFIG_DMA_GLOBAL_POOL
>> diff --git a/mm/dmapool.c b/mm/dmapool.c
>> index a7eb5d0..6e03530 100644
>> --- a/mm/dmapool.c
>> +++ b/mm/dmapool.c
>> @@ -21,6 +21,7 @@
>>    
>>    #include <linux/device.h>
>>    #include <linux/dma-mapping.h>
>> +#include <linux/dma-map-ops.h>
>>    #include <linux/dmapool.h>
>>    #include <linux/kernel.h>
>>    #include <linux/list.h>
>> @@ -372,7 +373,9 @@ void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
>>    #endif
>>    	spin_unlock_irqrestore(&pool->lock, flags);
>>    
>> -	if (want_init_on_alloc(mem_flags))
>> +	if (want_init_on_alloc(mem_flags) &&
>> +		!use_dev_coherent_memory(pool->dev) &&
>> +		get_dma_ops(pool->dev))
>>    		memset(retval, 0, pool->size);
>>    
>>    	return retval;
> 
> Best regards

Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

[Liu Song]

> On 2022-08-16 13:39, Marek Szyprowski wrote:
>> Hi,
>>
>> On 18.07.2022 08:28, Liu Song wrote:
>>> From: Liu Song <liusong@linux.alibaba.com>
>>>
>>> In "dma_alloc_from_dev_coherent" and "dma_direct_alloc",
>>> the allocated memory is explicitly set to 0.
>>>
>>> A helper function "use_dev_coherent_memory" is introduced here to
>>> determine whether the memory is allocated by 
>>> "dma_alloc_from_dev_coherent".
>>>
>>> And use "get_dma_ops" to determine whether the memory is allocated by
>>> "dma_direct_alloc".
>>>
>>> After this modification, memory allocated using "dma_pool_zalloc" 
>>> can avoid
>>> duplicate memset.
>>>
>>> Signed-off-by: Liu Song <liusong@linux.alibaba.com>
>>
>> This patch landed linux next-20220816. Unfortunately it causes serious
>> issues on ARM 32bit systems. I've observed it on ARM 32bit Samsung
>> Exynos 5422 based Odroid XU4 board with USB r8152 driver. After applying
>> this patch and loading r8152 driver I only the following endless
>> messages in the log:
>>
>> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs 
>> queued?
>> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs 
>> queued?
>> xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs 
>> queued?
>>
>> It looks that there are drivers which rely on the fact that the dma
>> coherent buffers are always zeroed.
>
> It's not even that, the change here is just obviously broken, since it 
> ends up entirely ignoring want_init_on_alloc() for devices using 
> dma-direct. Sure, the memory backing a dma_page is zeroed *once* by 
> its initial dma-coherent allocation, but who says we're not not 
> reallocating pool entries from an existing dma_page?
>
> I'm not convinced it's worth trying to special-case this at all, since 
> we can only do it reliably for the first pool entry allocated from a 
> new dma_page, and that will only happen as the pool initially grows to 
> a suitable size for its working set, after which no further new pages 
> are likely to be allocated for the lifetime of the pool. Even if there 
> is a case to be made for doing so, it would need to be based on the 
> flow through dma_pool_alloc() itself, not some nonsense heuristic on 
> the device.

Hi，

First of all, I am very sorry that there are missing branches that have 
not been considered fully,

but there is a possibility that the memset to 0 will be repeated whether 
re-allocation from the

dma pool, and this patch needs to be fixed.


Thanks


>
> Andrew, please drop this patch.
>
> Thanks,
> Robin.
>
>>> ---
>>>    include/linux/dma-map-ops.h | 5 +++++
>>>    mm/dmapool.c                | 5 ++++-
>>>    2 files changed, 9 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/include/linux/dma-map-ops.h b/include/linux/dma-map-ops.h
>>> index 0d5b06b..c29948d 100644
>>> --- a/include/linux/dma-map-ops.h
>>> +++ b/include/linux/dma-map-ops.h
>>> @@ -171,6 +171,10 @@ int dma_alloc_from_dev_coherent(struct device 
>>> *dev, ssize_t size,
>>>    int dma_release_from_dev_coherent(struct device *dev, int order, 
>>> void *vaddr);
>>>    int dma_mmap_from_dev_coherent(struct device *dev, struct 
>>> vm_area_struct *vma,
>>>            void *cpu_addr, size_t size, int *ret);
>>> +static inline bool use_dev_coherent_memory(struct device *dev)
>>> +{
>>> +    return dev->dma_mem ? true : false;
>>> +}
>>>    #else
>>>    static inline int dma_declare_coherent_memory(struct device *dev,
>>>            phys_addr_t phys_addr, dma_addr_t device_addr, size_t size)
>>> @@ -180,6 +184,7 @@ static inline int 
>>> dma_declare_coherent_memory(struct device *dev,
>>>    #define dma_alloc_from_dev_coherent(dev, size, handle, ret) (0)
>>>    #define dma_release_from_dev_coherent(dev, order, vaddr) (0)
>>>    #define dma_mmap_from_dev_coherent(dev, vma, vaddr, order, ret) (0)
>>> +#define use_dev_coherent_memory(dev) (0)
>>>    #endif /* CONFIG_DMA_DECLARE_COHERENT */
>>>       #ifdef CONFIG_DMA_GLOBAL_POOL
>>> diff --git a/mm/dmapool.c b/mm/dmapool.c
>>> index a7eb5d0..6e03530 100644
>>> --- a/mm/dmapool.c
>>> +++ b/mm/dmapool.c
>>> @@ -21,6 +21,7 @@
>>>       #include <linux/device.h>
>>>    #include <linux/dma-mapping.h>
>>> +#include <linux/dma-map-ops.h>
>>>    #include <linux/dmapool.h>
>>>    #include <linux/kernel.h>
>>>    #include <linux/list.h>
>>> @@ -372,7 +373,9 @@ void *dma_pool_alloc(struct dma_pool *pool, 
>>> gfp_t mem_flags,
>>>    #endif
>>>        spin_unlock_irqrestore(&pool->lock, flags);
>>>    -    if (want_init_on_alloc(mem_flags))
>>> +    if (want_init_on_alloc(mem_flags) &&
>>> +        !use_dev_coherent_memory(pool->dev) &&
>>> +        get_dma_ops(pool->dev))
>>>            memset(retval, 0, pool->size);
>>>           return retval;
>>
>> Best regards

Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

[Christoph Hellwig]

>>> A helper function "use_dev_coherent_memory" is introduced here to
>>> determine whether the memory is allocated by "dma_alloc_from_dev_coherent".
>>>
>>> And use "get_dma_ops" to determine whether the memory is allocated by
>>> "dma_direct_alloc".

WTF?  get_dma_ops is privat to the DMA API layer, and dmapool has no
business even using that.  Even independent of this particular case,
consumers of an API never have any business looking at the implementation
of the API, that is the whole point of the abstraction.

> It's not even that, the change here is just obviously broken, since it ends 
> up entirely ignoring want_init_on_alloc() for devices using dma-direct. 
> Sure, the memory backing a dma_page is zeroed *once* by its initial 
> dma-coherent allocation, but who says we're not not reallocating pool 
> entries from an existing dma_page?

And yes, in addition to that it also is completely broken.

Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

[Liu Song]

A helper function "use_dev_coherent_memory" is introduced here to

>>>> determine whether the memory is allocated by "dma_alloc_from_dev_coherent".
>>>>
>>>> And use "get_dma_ops" to determine whether the memory is allocated by
>>>> "dma_direct_alloc".
> WTF?  get_dma_ops is privat to the DMA API layer, and dmapool has no
> business even using that.  Even independent of this particular case,
> consumers of an API never have any business looking at the implementation
> of the API, that is the whole point of the abstraction.
>
>> It's not even that, the change here is just obviously broken, since it ends
>> up entirely ignoring want_init_on_alloc() for devices using dma-direct.
>> Sure, the memory backing a dma_page is zeroed *once* by its initial
>> dma-coherent allocation, but who says we're not not reallocating pool
>> entries from an existing dma_page?
> And yes, in addition to that it also is completely broken.

After reading everyone's comments, I found that fixing this patch will

make the code look strange, so the benefits of the changes will be

dispensable, so I also agree to discard this patch.

Sorry for this trouble again.


Thanks