On 18.11.26 12:25, Ross Lagerwall wrote: > On 11/13/26 12:47 PM, Lin Liu wrote: >> The block layer applies a device's DMA segment boundary to >> pseudo-physical addresses, but in a Xen PV domain the bus address is the >> machine address.  A range that is contiguous in both spaces can still >> cross the boundary in machine space and reach the device.  AHCI declares >> a 4 GiB boundary, and controllers that do not carry across it (e.g. >> Marvell 88SE9230) then DMA the tail of the transfer 4 GiB too low. >> >> Make range_straddles_page_boundary() also reject a contiguous range that >> crosses the device's segment boundary in bus space, so all its callers >> cover it.  Streaming mappings are bounced; coherent allocations are >> already naturally aligned via range_requires_alignment(). >> >> Signed-off-by: Lin Liu >> --- >>   drivers/xen/swiotlb-xen.c | 16 +++++++++++----- >>   1 file changed, 11 insertions(+), 5 deletions(-) >> >> diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c >> index e2538824e..181f8cef2 100644 >> --- a/drivers/xen/swiotlb-xen.c >> +++ b/drivers/xen/swiotlb-xen.c >> @@ -82,10 +82,13 @@ static inline bool range_requires_alignment(phys_addr_t p, >> size_t size) >>       return IS_ALIGNED(p, algn) && !IS_ALIGNED(bus_addr, algn); >>   } >> -static inline int range_straddles_page_boundary(phys_addr_t p, size_t size) >> +static inline int range_straddles_page_boundary(struct device *dev, >> +                        phys_addr_t p, size_t size) >>   { >>       unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p); >>       unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size); >> +    unsigned long mask; >> +    dma_addr_t dma; >>       next_bfn = pfn_to_bfn(xen_pfn); >> @@ -93,7 +96,10 @@ static inline int range_straddles_page_boundary(phys_addr_t >> p, size_t size) >>           if (pfn_to_bfn(++xen_pfn) != ++next_bfn) >>               return 1; >> -    return 0; >> +    /* Contiguous, but may cross the segment boundary in bus space */ >> +    mask = dma_get_seg_boundary(dev); >> +    dma = xen_phys_to_dma(dev, p); >> +    return size - 1 <= mask && ((dma ^ (dma + size - 1)) & ~mask); >>   } >>   static struct io_tlb_pool *xen_swiotlb_find_pool(struct device *dev, >> @@ -158,7 +164,7 @@ xen_swiotlb_alloc_coherent(struct device *dev, size_t size, >>       *dma_handle = xen_phys_to_dma(dev, phys); >>       if (*dma_handle + size - 1 > dma_mask || >> -        range_straddles_page_boundary(phys, size) || >> +        range_straddles_page_boundary(dev, phys, size) || >>           range_requires_alignment(phys, size)) { >>           if (xen_create_contiguous_region(phys, order, fls64(dma_mask), >>                   dma_handle) != 0) > > Is the change actually needed for alloc_coherent / free_coherent? > The allocation from __get_free_pages() is naturally aligned and > range_requires_alignment() requires it has the same alignment so it is > naturally aligned in bus space. I don't think that it can cross a > segment boundary then. Hmm, agreed. Juergen