From: Liu Song <liusong@linux.alibaba.com>

In "dma_direct_alloc", the allocated memory is explicitly set to 0.
If use direct alloc, we need to avoid possible duplicate memset in
dma_pool_alloc.

Signed-off-by: Liu Song <liusong@linux.alibaba.com>
---
 mm/dmapool.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/dmapool.c b/mm/dmapool.c
index a7eb5d0..94f4859 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 +373,7 @@ 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) && get_dma_ops(pool->dev))
 		memset(retval, 0, pool->size);
 
 	return retval;
-- 
1.8.3.1

On Thu, 14 Jul 2022 11:25:00 +0800 Liu Song <liusong@linux.alibaba.com> wrote:

> From: Liu Song <liusong@linux.alibaba.com>
> 
> In "dma_direct_alloc", the allocated memory is explicitly set to 0.
> If use direct alloc, we need to avoid possible duplicate memset in
> dma_pool_alloc.

I'm having trouble seeing how this change is safe and correct and
maintainable.  Please describe the code flow more completely?

> --- 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 +373,7 @@ 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) && get_dma_ops(pool->dev))
>  		memset(retval, 0, pool->size);

That DMAPOOL_DEBUG memset a couple of lines earlier could/should be
testing the same condition - there's no point in poisoning an area
which we're about to zero out.

>> From: Liu Song <liusong@linux.alibaba.com>
>>
>> In "dma_direct_alloc", the allocated memory is explicitly set to 0.
>> If use direct alloc, we need to avoid possible duplicate memset in
>> dma_pool_alloc.
>
>I'm having trouble seeing how this change is safe and correct and
>maintainable.  Please describe the code flow more completely?
The following is the code flow， 
dma_pool_alloc
    |--> pool_alloc_page
           |--> dma_alloc_coherent
                  |--> dma_alloc_attrs
In "dma_alloc_attrs", if "dma_alloc_direct" is true, then enter "dma_direct_alloc",
and in "dma_direct_alloc", as long as the memory allocation is successful, will execute
"memset(ret, 0, size);", which set memory to zero.
Kernel use "dma_go_direct" to determine whether to use direct allocation, which mainly
by judging whether "dma_map_ops" exists.

So this patch determines whether direct alloc will be used by judging does "dma_map_ops" exists,
thereby avoiding repeated memset.

>
>> --- 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 +373,7 @@ 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) && get_dma_ops(pool->dev))
>>    memset(retval, 0, pool->size);
>
>That DMAPOOL_DEBUG memset a couple of lines earlier could/should be
>testing the same condition - there's no point in poisoning an area
>which we're about to zero out.

If DMAPOOL_DEBUG is configured, its logic is internally self-consistent.
If the user needs __GFP_ZERO, the corresponding memory will be set to 0.

This patch can not only consider direct alloc, with config CONFIG_DMA_DECLARE_COHERENT,
in "dma_alloc_from_dev_coherent", which will also explicitly memset the allocated memory to 0,
so this situation needs to be considered. It is beneficial to the process of calling "dma_pool_zalloc".
Next, a V2 patch will be submitted.

Thanks.