Re: [PATCH] kernel: dma: let dma use vmalloc area

[Zhaoyang Huang]

On Mon, Nov 27, 2023 at 3:14 PM Christoph Hellwig <hch@xxxxxx> wrote:
>
> On Mon, Nov 27, 2023 at 11:09:30AM +0800, zhaoyang.huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> >
> > memremap within dma_init_coherent_memory will map the given phys_addr
> > into vmalloc area if the pa is not found during iterating iomem_resources,
> > which conflict the rejection of vmalloc area in dma_map_single_attrs.
>
> I can't parse this sentence.
Sorry for the confusion, please find below codes for more information.
dma_init_coherent_memory
    memremap
        addr = ioremap_wt(offset, size);
What I mean is addr is a vmalloc address, which is implicitly mapped
by dma's framework and not be aware of to the driver.
>
> > IMO, it is find to let all valid virtual address be valid for DMA as the
> > user will keep corresponding RAM safe for transfer.
>
> No, vmalloc address can't be passed to map_single.  You need to pass
> the page to dma_map_page, and explicitly mange cache consistency
> using the invalidate_kernel_vmap_range and flush_kernel_vmap_range
> helpers.
Please correct me if I am wrong. According to my understanding, cache
consistency could be solved inside dma_map_page via either
dma_direct_map_page(swio/arch_sync_dma_for_device) or ops->map_page.
The original thought of rejecting vmalloc is that this pa is not safe
as this mapping could go in any time. What I am suggesting is to let
this kind of va be enrolled.

 static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,
                size_t size, enum dma_data_direction dir, unsigned long attrs)
 {
       /* DMA must never operate on areas that might be remapped. */
       if (dev_WARN_ONCE(dev, is_vmalloc_addr(ptr),
                         "rejecting DMA map of vmalloc memory\n"))
               return DMA_MAPPING_ERROR;

>