Messages in this thread | | | Date | Wed, 24 Sep 2008 21:43:59 +0900 | Subject | Re: [PATCH 3/3] x86: restore old GART alloc_coherent behavior | From | FUJITA Tomonori <> |
| |
On Wed, 24 Sep 2008 14:11:27 +0200 Joerg Roedel <joerg.roedel@amd.com> wrote:
> On Wed, Sep 24, 2008 at 08:48:37PM +0900, FUJITA Tomonori wrote: > > Currently, GART alloc_coherent tries to allocate pages with GFP_DMA32 > > for a device having dma_masks > 24bit < 32bits. If GART gets an > > address that a device can't access to, GART try to map the address to > > a virtual I/O address that the device can access to. > > > > But Andi pointed out, "The GART is somewhere in the 4GB range so you > > cannot use it to map anything < 4GB. Also GART is pretty small." > > > > http://lkml.org/lkml/2008/9/12/43 > > > > That is, it's possible that GART doesn't have virtual I/O address > > space that a device can access to. The above behavior doesn't work for > > a device having dma_masks > 24bit < 32bits. > > > > This patch restores old GART alloc_coherent behavior (before the > > alloc_coherent rewrite). > > Patchset looks good in principle. But I have a small question, see > below. > > > > > Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp> > > --- > > arch/x86/kernel/pci-gart_64.c | 43 +++++++++++++++++++--------------------- > > 1 files changed, 20 insertions(+), 23 deletions(-) > > > > diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c > > index 7e08e46..25c94fb 100644 > > --- a/arch/x86/kernel/pci-gart_64.c > > +++ b/arch/x86/kernel/pci-gart_64.c > > @@ -487,31 +487,28 @@ static void * > > gart_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_addr, > > gfp_t flag) > > { > > - void *vaddr; > > dma_addr_t paddr; > > unsigned long align_mask; > > - u64 dma_mask = dma_alloc_coherent_mask(dev, flag); > > - > > - vaddr = (void *)__get_free_pages(flag | __GFP_ZERO, get_order(size)); > > - if (!vaddr) > > - return NULL; > > - > > - paddr = virt_to_phys(vaddr); > > - if (is_buffer_dma_capable(dma_mask, paddr, size)) { > > - *dma_addr = paddr; > > - return vaddr; > > - } > > - > > - align_mask = (1UL << get_order(size)) - 1; > > - > > - *dma_addr = dma_map_area(dev, paddr, size, DMA_BIDIRECTIONAL, > > - align_mask); > > - flush_gart(); > > - > > - if (*dma_addr != bad_dma_address) > > - return vaddr; > > - > > - free_pages((unsigned long)vaddr, get_order(size)); > > + struct page *page; > > + > > + if (force_iommu && !(flag & GFP_DMA)) { > > + flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32); > > + page = alloc_pages(flag | __GFP_ZERO, get_order(size)); > > + if (!page) > > + return NULL; > > + > > + align_mask = (1UL << get_order(size)) - 1; > > + paddr = dma_map_area(dev, page_to_phys(page), size, > > + DMA_BIDIRECTIONAL, align_mask); > > Can't we check if a mapping is required before calling dma_map_area? > Your recently introduced IOMMU helper functions should make that easy > and GART address space is a rare resource. AFAIR this is also the > behaviour of the old generic dma_alloc_coherent function.
I think that the behavior of the old generic dma_alloc_coherent function and GART is different.
The old GART code does virtual mappings only with force_iommu option enabled. The old GART code always do virtual mappings with force_iommu option enabled (unless GFP_DMA is set).
> > + > > + flush_gart(); > > This should depend on need_flush.
Theoretically, yes, I think. But this patch restores the old GART code behavior.
| |