2017-09-20 22:18:12

by Timur Tabi

[permalink] [raw]
Subject: Proper way to allocate DMA buffer within a single 4GB block?

I have a device that requires I allocated a few buffers for DMA. The
problem is that this device has only one register for the upper 32 bits
of all of the buffers. That is, all of buffers must reside within the
same 4GB block of memory. In order words,

end = start + size - 1;
if (upper_32_bits(start) != upper_32_bits(end))
// Oh no, the buffer spans across a 4GB boundary!

The buffer is typically less than 16KB in size, so we've never seen it
actually span across a 4GB boundary. However, I want to ensure that
it's impossible. I wrote this function that re-tries the allocation if
the first one is invalid, but I suspect that it's too hackish. Is there
a better way?

static void *dma_alloc_4gb_block(struct device *dev, size_t size,
dma_addr_t *dma_handle)
{
dma_addr_t start, start2, end;
void *v, *v2;

v = dma_zalloc_coherent(dev, size, &start, GFP_KERNEL);
if (!v)
return NULL;

/* Test to see if the entire buffer is within one 4G block */
end = start + size - 1;
if (upper_32_bits(start) == upper_32_bits(end)) {
*dma_handle = start;
return v;
}

/* The DMA buffer spanned a 32-bit boundry! Allocate another
* block and try again. Assuming this block as adjacent to
* the first one, it should be valid.
*/
v2 = dma_zalloc_coherent(dev, size, &start2, GFP_KERNEL);
dma_free_coherent(dev, size, v, start);
if (!v2) {
return NULL;
}

/* Test to see if the entire buffer is within one 4G block */
end = start2 + size - 1;
if (upper_32_bits(start2) == upper_32_bits(end)) {
*dma_handle = start2;
return v2;
}

/* We tried twice and still failed, so just give up. */
dma_free_coherent(dev, size, v2, start2);
return NULL;
}

--
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm
Technologies, Inc. Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.


2017-09-24 14:20:52

by Chuck Ebbert

[permalink] [raw]
Subject: Re: Proper way to allocate DMA buffer within a single 4GB block?

On Wed, 20 Sep 2017 17:18:09 -0500
Timur Tabi <[email protected]> wrote:

> I have a device that requires I allocated a few buffers for DMA. The
> problem is that this device has only one register for the upper 32
> bits of all of the buffers. That is, all of buffers must reside
> within the same 4GB block of memory. In order words,
>
> end = start + size - 1;
> if (upper_32_bits(start) != upper_32_bits(end))
> // Oh no, the buffer spans across a 4GB boundary!
>
> The buffer is typically less than 16KB in size, so we've never seen
> it actually span across a 4GB boundary. However, I want to ensure
> that it's impossible. I wrote this function that re-tries the
> allocation if the first one is invalid, but I suspect that it's too
> hackish. Is there a better way?
>

Allocate a buffer twice as big as what you really need. If the first
half doesn't work, the second half will.