Re: [Question] devm_kmalloc() for DMA ?

[Robin Murphy <robin.murphy@arm.com>]

On 08/03/17 10:59, Masahiro Yamada wrote:
> Hi experts,
> 
> I have a question about
> how to allocate DMA-safe buffer.
> 
> 
> In my understanding, kmalloc() returns
> memory with DMA safe alignment
> in order to avoid cache-sharing problem when used for DMA.
> 
> The alignment is decided by ARCH_DMA_MINALIGN.
> For example, on modern ARM 32bit boards, this value is typically 64.
> So, memory returned by kmalloc() has
> at least 64 byte alignment.
> 
> 
> On the other hand, devm_kmalloc() does not return
> enough-aligned memory.

How so? If anything returned by kmalloc() is guaranteed to occupy some
multiple of ARCH_DMA_MINALIGN bytes in order to avoid two allocations
falling into the same cache line, I don't see how stealing the first 16
bytes *of a single allocation* could make it start sharing cache lines
with another? :/

If a particular device has a problem with:

p = kmalloc(...);
d = dma_map_single(p + 0x10, ...);
do_something_with(d);

that's a separate issue altogether.

Robin.

> On my board (ARM 32bit), devm_kmalloc() returns
> (ARCH_DMA_MINALIGN aligned address) + 0x10.
> 
> 
> 
> The reason of the offset 0x10 is obvious.
> 
> struct devres {
>         struct devres_node node;
>         /* -- 3 pointers */
>         unsigned long long data[]; /* guarantee ull alignment */
> };
> 
> 
> Management data is located at the top of struct devres.
> Then, devm_kmalloc() returns dr->data.
> 
> The "unsigned long long" guarantees
> the returned memory has 0x10 alignment,
> but I think this may not be enough for DMA.
> 
> I noticed this when I was seeing drivers/mtd/nand/denali.c
> 
> The code looks as follows:
> 
> 
>         denali->buf.buf = devm_kzalloc(denali->dev,
>                              mtd->writesize + mtd->oobsize,
>                              GFP_KERNEL);
>         if (!denali->buf.buf) {
>                 ret = -ENOMEM;
>                 goto failed_req_irq;
>         }
> 
>         /* Is 32-bit DMA supported? */
>         ret = dma_set_mask(denali->dev, DMA_BIT_MASK(32));
>         if (ret) {
>                 dev_err(denali->dev, "No usable DMA configuration\n");
>                 goto failed_req_irq;
>         }
> 
>         denali->buf.dma_buf = dma_map_single(denali->dev, denali->buf.buf,
>                              mtd->writesize + mtd->oobsize,
>                              DMA_BIDIRECTIONAL);
> 
> 
> 
> 
> Memory buffer is allocated by devm_kzalloc(), then
> passed to dma_map_single().
> 
> 
> 
> Could this be a potential problem in general?
> 
> Is devm_kmalloc() not recommended
> for buffer that can be DMA-mapped?
> 
> 
> Any advice is appreciated.
> 
>