| Message ID | 1382031814-8782-18-git-send-email-stefano.stabellini@eu.citrix.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Oct 17, 2013 at 06:43:33PM +0100, Stefano Stabellini wrote: > xen_unmap_single calls xen_bus_to_phys, then passes dev_addr to > is_xen_swiotlb_buffer that calls xen_bus_to_phys again. It does? I see it call mfn_to_local_pfn which calls pfn_to_mfn and get_phys_to_machine (which I presume is going to be a slow operation on ARM). > xen_bus_to_phys is potentially a slow operation that might involve > walking a tree. > We can avoid calling xen_bus_to_phys twice by removing the > is_xen_swiotlb_buffer call and doing the check directly in > xen_unmap_single using the physical address. > > Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Have you tested this with a PFN that is in a foreign domain? I am bit uncomfortable with this - I recall when developing this that I had hit some odd cases of doing DMA of a page to a guest and it not always working. It was the case of not detecting whether the pfn was foreign or not. Sadly I can't recall the details. I would think that this problem would also be present with ARM? But perhaps all the new changes in the pfn_to_mfn that had lately been done have fixed this? > --- > drivers/xen/swiotlb-xen.c | 4 ++-- > 1 files changed, 2 insertions(+), 2 deletions(-) > > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c > index 4d50058..6d9ca61 100644 > --- a/drivers/xen/swiotlb-xen.c > +++ b/drivers/xen/swiotlb-xen.c > @@ -426,8 +426,8 @@ static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr, > > xen_dma_unmap_page(hwdev, paddr, size, dir, attrs); > > - /* NOTE: We use dev_addr here, not paddr! */ > - if (is_xen_swiotlb_buffer(dev_addr)) { > + if (paddr >= virt_to_phys(xen_io_tlb_start) && > + paddr < virt_to_phys(xen_io_tlb_end)) { > swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); > return; > } > -- > 1.7.2.5 >
> On Thu, Oct 17, 2013 at 06:43:33PM +0100, Stefano Stabellini wrote: > > xen_unmap_single calls xen_bus_to_phys, then passes dev_addr to > > is_xen_swiotlb_buffer that calls xen_bus_to_phys again. > > It does? > > I see it call mfn_to_local_pfn which calls pfn_to_mfn and > get_phys_to_machine (which I presume is going to be a slow > operation on ARM). Yeah, I am trying to reduce the number of pfn_to_mfn and mfn_to_pfn translations. That said, this patch is not needed to have a fully functional swiotlb-xen on arm and I would OK with postponing it. > > xen_bus_to_phys is potentially a slow operation that might involve > > walking a tree. > > We can avoid calling xen_bus_to_phys twice by removing the > > is_xen_swiotlb_buffer call and doing the check directly in > > xen_unmap_single using the physical address. > > > > Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> > > Have you tested this with a PFN that is in a foreign domain? That would be blkback. I only tested blkback on ARM, I'll test on x86 too. > I am bit uncomfortable with this - I recall when developing this > that I had hit some odd cases of doing DMA of a page to a guest > and it not always working. It was the case of not detecting whether > the pfn was foreign or not. Sadly I can't recall the details. > > I would think that this problem would also be present with ARM? > But perhaps all the new changes in the pfn_to_mfn that had lately > been done have fixed this? No, because on ARM the guest is autotranslate so the pfn returned by mfn_to_pfn is always "local". However thanks to the m2p_override now we should get the "local" pfn from mfn_to_pfn on x86 too. I'll test this case though. > > --- > > drivers/xen/swiotlb-xen.c | 4 ++-- > > 1 files changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c > > index 4d50058..6d9ca61 100644 > > --- a/drivers/xen/swiotlb-xen.c > > +++ b/drivers/xen/swiotlb-xen.c > > @@ -426,8 +426,8 @@ static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr, > > > > xen_dma_unmap_page(hwdev, paddr, size, dir, attrs); > > > > - /* NOTE: We use dev_addr here, not paddr! */ > > - if (is_xen_swiotlb_buffer(dev_addr)) { > > + if (paddr >= virt_to_phys(xen_io_tlb_start) && > > + paddr < virt_to_phys(xen_io_tlb_end)) { > > swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); > > return; > > } > > -- > > 1.7.2.5 > >
On Wed, Oct 23, 2013 at 06:34:13PM +0100, Stefano Stabellini wrote: > > On Thu, Oct 17, 2013 at 06:43:33PM +0100, Stefano Stabellini wrote: > > > xen_unmap_single calls xen_bus_to_phys, then passes dev_addr to > > > is_xen_swiotlb_buffer that calls xen_bus_to_phys again. > > > > It does? > > > > I see it call mfn_to_local_pfn which calls pfn_to_mfn and > > get_phys_to_machine (which I presume is going to be a slow > > operation on ARM). > > Yeah, I am trying to reduce the number of pfn_to_mfn and mfn_to_pfn > translations. > That said, this patch is not needed to have a fully functional > swiotlb-xen on arm and I would OK with postponing it. OK, lets do that as it requires some hard thinking. > > > > > xen_bus_to_phys is potentially a slow operation that might involve > > > walking a tree. > > > We can avoid calling xen_bus_to_phys twice by removing the > > > is_xen_swiotlb_buffer call and doing the check directly in > > > xen_unmap_single using the physical address. > > > > > > Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> > > > > Have you tested this with a PFN that is in a foreign domain? > > That would be blkback. I only tested blkback on ARM, I'll test on x86 > too. > > > > I am bit uncomfortable with this - I recall when developing this > > that I had hit some odd cases of doing DMA of a page to a guest > > and it not always working. It was the case of not detecting whether > > the pfn was foreign or not. Sadly I can't recall the details. > > > > I would think that this problem would also be present with ARM? > > But perhaps all the new changes in the pfn_to_mfn that had lately > > been done have fixed this? > > No, because on ARM the guest is autotranslate so the pfn returned by > mfn_to_pfn is always "local". > However thanks to the m2p_override now we should get the "local" pfn > from mfn_to_pfn on x86 too. > I'll test this case though. Right, so maybe the pfn_to_mfn_local test is useless nowadays? > > > > > --- > > > drivers/xen/swiotlb-xen.c | 4 ++-- > > > 1 files changed, 2 insertions(+), 2 deletions(-) > > > > > > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c > > > index 4d50058..6d9ca61 100644 > > > --- a/drivers/xen/swiotlb-xen.c > > > +++ b/drivers/xen/swiotlb-xen.c > > > @@ -426,8 +426,8 @@ static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr, > > > > > > xen_dma_unmap_page(hwdev, paddr, size, dir, attrs); > > > > > > - /* NOTE: We use dev_addr here, not paddr! */ > > > - if (is_xen_swiotlb_buffer(dev_addr)) { > > > + if (paddr >= virt_to_phys(xen_io_tlb_start) && > > > + paddr < virt_to_phys(xen_io_tlb_end)) { > > > swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); > > > return; > > > } > > > -- > > > 1.7.2.5 > > >
On Wed, 23 Oct 2013, Konrad Rzeszutek Wilk wrote: > On Wed, Oct 23, 2013 at 06:34:13PM +0100, Stefano Stabellini wrote: > > > On Thu, Oct 17, 2013 at 06:43:33PM +0100, Stefano Stabellini wrote: > > > > xen_unmap_single calls xen_bus_to_phys, then passes dev_addr to > > > > is_xen_swiotlb_buffer that calls xen_bus_to_phys again. > > > > > > It does? > > > > > > I see it call mfn_to_local_pfn which calls pfn_to_mfn and > > > get_phys_to_machine (which I presume is going to be a slow > > > operation on ARM). > > > > Yeah, I am trying to reduce the number of pfn_to_mfn and mfn_to_pfn > > translations. > > That said, this patch is not needed to have a fully functional > > swiotlb-xen on arm and I would OK with postponing it. > > OK, lets do that as it requires some hard thinking. OK. > > > I am bit uncomfortable with this - I recall when developing this > > > that I had hit some odd cases of doing DMA of a page to a guest > > > and it not always working. It was the case of not detecting whether > > > the pfn was foreign or not. Sadly I can't recall the details. > > > > > > I would think that this problem would also be present with ARM? > > > But perhaps all the new changes in the pfn_to_mfn that had lately > > > been done have fixed this? > > > > No, because on ARM the guest is autotranslate so the pfn returned by > > mfn_to_pfn is always "local". > > However thanks to the m2p_override now we should get the "local" pfn > > from mfn_to_pfn on x86 too. > > I'll test this case though. > > Right, so maybe the pfn_to_mfn_local test is useless nowadays? Yes, I think so. This might be better handled by removing mfn_to_local_pfn altogether.
diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c index 4d50058..6d9ca61 100644 --- a/drivers/xen/swiotlb-xen.c +++ b/drivers/xen/swiotlb-xen.c @@ -426,8 +426,8 @@ static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr, xen_dma_unmap_page(hwdev, paddr, size, dir, attrs); - /* NOTE: We use dev_addr here, not paddr! */ - if (is_xen_swiotlb_buffer(dev_addr)) { + if (paddr >= virt_to_phys(xen_io_tlb_start) && + paddr < virt_to_phys(xen_io_tlb_end)) { swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); return; }
xen_unmap_single calls xen_bus_to_phys, then passes dev_addr to is_xen_swiotlb_buffer that calls xen_bus_to_phys again. xen_bus_to_phys is potentially a slow operation that might involve walking a tree. We can avoid calling xen_bus_to_phys twice by removing the is_xen_swiotlb_buffer call and doing the check directly in xen_unmap_single using the physical address. Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> --- drivers/xen/swiotlb-xen.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-)