On Mon, Oct 12, 2020 at 08:47:15AM +0200, Christoph Hellwig wrote: > On Fri, Oct 09, 2020 at 06:10:52PM +0100, Catalin Marinas wrote: > > kdump wants DMA-able memory and, > > DMAable by whom? The only way to guranteed DMAable memory is to use > the DMA memory allocator(s) and pass a specific device to them. Everyting > else is just fundamentally broken. Note that even when device is not > DMAable we can still use swiotlb to access it. What I meant is that the new kexec'ed kernel needs some memory in the ZONE_DMA range, currently set to the bottom 30-bit even for platforms that can cope with the whole 32-bit range (anything other than RPi4). The memory range available to the kdump kernels is limited to what reserve_crashkernel() allocated, which may not fit in the lower 1GB. There are two ongoing threads (complementary): 1. Change the arm64 reserve_crashkernel() similar to x86 which allocates memory above 4G with a small block in the ZONE_DMA range. 2. Allow zone_dma_bits to be 32 for arm64 platforms other than RPi4. The second point also fixes some regressions with CMA reservations that could no longer fit in the lower 1GB. -- Catalin
