Hi Tomasz,
On Friday 08 June 2012 16:31:31 Tomasz Stanislawski wrote:
Hi Laurent and Subash,
I confirm the issue found by Subash. The function vb2_dc_kaddr_to_pages does fail for some occasions. The failures are rather strange like 'got 95 of 150 pages'. It took me some time to find the reason of the problem.
I found that dma_alloc_coherent for iommu an ARM does use ioremap_page_range to map a buffer to the kernel space. The mapping is done by updating the page-table.
The problem is that any process has a different first-level page-table. The ioremap_page_range updates only the table for init process. The PT present in current->mm shares a majority of entries of 1st-level PT at kernel range (above 0xc0000000) but *not all*. That is why vb2_dc_kaddr_to_pages worked for small buffers and occasionally failed for larger buffers.
I found two ways to fix this problem. a) use &init_mm instead of current->mm while creating an artificial vma b) access the dma memory by calling *((volatile int *)kaddr) = 0; before calling follow_pfn This way a fault is generated and the PT is updated by copying entries from init_mm.
What do you think about presented solutions?
Just to be sure, this is a hack until dma_get_sgtable is available, and it won't make it to mainline, right ? In that case using init_mm seem easier.