Linux内核空间中的“魔环缓冲”实现？

Question

I'm aware of the "magic ring buffer" trick that involves mirroring the underlying buffer in the process's address space to allow chunks of data to be enqueued with a single memcpy() without worrying about wraparound.

I'd like to accomplish the same thing but in a Linux kernel module. Assume I have a buffer created with dma_alloc_coherent() , its virtual address is V and its length is N . How do I create the mapping such that its virtual addresses [V+N,V+2N) map to the same underlying pages as [V,V+N) ?

Note: this is in 32-bit ARM Linux.

Answer 1

drivers/firewire/ohci.c maps some pages of the Asynchronous Receive ring buffer twice to allow easier access to received packets that wrap around:

    for (i = 0; i < AR_BUFFERS; i++) {
            ctx->pages[i] = alloc_page(GFP_KERNEL | GFP_DMA32);
            ...
            dma_addr = dma_map_page(ohci->card.device, ctx->pages[i],
                                    0, PAGE_SIZE, DMA_FROM_DEVICE);
            ...
    }
    for (i = 0; i < AR_BUFFERS; i++)
            pages[i]              = ctx->pages[i];
    for (i = 0; i < AR_WRAPAROUND_PAGES; i++)
            pages[AR_BUFFERS + i] = ctx->pages[i];
    ctx->buffer = vmap(pages, ARRAY_SIZE(pages), VM_MAP, PAGE_KERNEL);
    ...

As far as I can see, there is no similar API for coherent DMA memory. You might be able to remap the pages returned by dma_alloc_coherent() if you know how your architecture handles this.

Please note that some architectures can have problems with their caches if you are using multiple virtual addresses to modify the same physical address; even if you manage to map it, you have to check if coherent DMA memory is cachable in your specific arch.