| Message ID | 1404409205-7368-1-git-send-email-alcooperx@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Jul 03, 2014 at 01:40:05PM -0400, Al Cooper wrote: > Some older hardware, like USB OHCI, cannot DMA to physical address > zero. This change makes sure that the first PAGESIZE block of memory > starting at zero is reserved so it can't end up in any free > memory pool. I really don't like this approach. The assumption here is that DMA address zero on a peripheral is the same as physical address zero. That isn't always the case. What if DMA address zero corresponds with the first byte of RAM, but the first byte of RAM is at 0x80000000? I believe there's even some cases where the first byte of RAM is not what the kernel sees as the first byte of RAM (because people want to hide the first half of RAM from the kernel for a DSP or similar.) What I'm saying is I don't think there's a nice simple solution to this. We probably need the DT reserved memory support to handle this.
diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c index 659c75d..17aaa4d 100644 --- a/arch/arm/mm/init.c +++ b/arch/arm/mm/init.c @@ -324,6 +324,15 @@ void __init arm_memblock_init(const struct machine_desc *mdesc) */ dma_contiguous_reserve(min(arm_dma_limit, arm_lowmem_limit)); + /* + * Some older hardware, like USB OHCI, cannot DMA + * to physical address zero. This makes sure that the + * first PAGESIZE block of memory starting at zero is + * reserved so it can't end up in any free memory pool. + */ + if (memblock_is_region_memory(0, PAGE_SIZE)) + memblock_reserve(0, PAGE_SIZE); + arm_memblock_steal_permitted = false; memblock_dump_all(); }
Some older hardware, like USB OHCI, cannot DMA to physical address zero. This change makes sure that the first PAGESIZE block of memory starting at zero is reserved so it can't end up in any free memory pool. Signed-off-by: Al Cooper <alcooperx@gmail.com> --- arch/arm/mm/init.c | 9 +++++++++ 1 file changed, 9 insertions(+)