2003-08-05 19:01:51

by Krzysztof Halasa

[permalink] [raw]
Subject: pci_alloc_consistent() and/or dma_free_coherent() bug?

Hi,

I'm doing something like:
pci_set_dma_mask(pdev, 0xFFFFFFFF) /* 32 bits */
pci_set_consistent_dma_mask(pdev, 0x0FFFFFFF); /* 28 bits only */

as this card is able to do 32-bit PCI (bus master) addressing on memory
returned by pci_map_single() but it needs "consistent" region in the first
256 MB of RAM.

On i386, pci_alloc_consistent() calls dma_alloc_coherent().

Now:
void *dma_alloc_coherent(struct device *dev, size_t size,
dma_addr_t *dma_handle, int gfp)
{
void *ret;
/* ignore region specifiers */
gfp &= ~(__GFP_DMA | __GFP_HIGHMEM);

if (dev == NULL || (*dev->dma_mask < 0xffffffff))
gfp |= GFP_DMA;
ret = (void *)__get_free_pages(gfp, get_order(size));

if (ret != NULL) {
memset(ret, 0, size);
*dma_handle = virt_to_phys(ret);
}
return ret;
}

while:

pci_set_consistent_dma_mask(struct pci_dev *dev, u64 mask)
{
if (!pci_dma_supported(dev, mask))
return -EIO;

dev->consistent_dma_mask = mask;

return 0;
}

pci_set_consistent_dma_mask() sets dev->consistent_dma_mask while
dma_alloc_coherent() (and thus pci_alloc_consistent()) expects the mask
to be in dev->dma_mask.


The "obvious" fix seems to be:

--- arch/i386/kernel/pci-dma.c.orig 2003-05-27 03:00:25.000000000 +0200
+++ arch/i386/kernel/pci-dma.c 2003-08-05 18:55:42.000000000 +0200
@@ -20,7 +20,7 @@
/* ignore region specifiers */
gfp &= ~(__GFP_DMA | __GFP_HIGHMEM);

- if (dev == NULL || (*dev->dma_mask < 0xffffffff))
+ if (dev == NULL || (*dev->consistent_dma_mask < 0xffffffff))
gfp |= GFP_DMA;
ret = (void *)__get_free_pages(gfp, get_order(size));


but I'm not sure if we need to adjust dev->consistent_dma_mask in
pci_set_dma_mask() (not only in pci_set_consistent_dma_mask()), in case
a driver only calls pci_set_dma_mask() and assumes it will work for
pci_alloc_consistent() as well.
--
Krzysztof Halasa
Network Administrator


2003-08-05 20:54:39

by Krzysztof Halasa

[permalink] [raw]
Subject: Re: pci_alloc_consistent() and/or dma_free_coherent() bug?

Actually it would probably be better to do something like that:

--- arch/i386/kernel/pci-dma.c.orig 2003-05-27 03:00:25.000000000 +0200
+++ arch/i386/kernel/pci-dma.c 2003-08-05 18:55:42.000000000 +0200
@@ -20,7 +20,7 @@
/* ignore region specifiers */
gfp &= ~(__GFP_DMA | __GFP_HIGHMEM);

- if (dev == NULL || (*dev->dma_mask < 0xffffffff))
+ if (dev == NULL || ((*dev->consistent_dma_mask & *dev->dma_mask) < 0xffffffff))
gfp |= GFP_DMA;
ret = (void *)__get_free_pages(gfp, get_order(size));


So we don't need to touch pci_set*dma_mask and we don't need to call them
in specific order.
--
Krzysztof Halasa
Network Administrator