From: Mike Rapoport <rppt@xxxxxxxxxxxxx> Subject: alpha: fix hang caused by the bootmem removal The conversion of alpha to memblock as the early memory manager caused boot to hang as described at [1]. The issue is caused because for CONFIG_DISCTONTIGMEM=y case, memblock_add() is called using memory start PFN that had been rounded down to the nearest 8Mb and it caused memblock to see more memory that is actually present in the system. Besides, memblock allocates memory from high addresses while bootmem was using low memory, which broke the assumption that early allocations are always accessible by the hardware. This patch ensures that memblock_add() is using the correct PFN for the memory start and forces memblock to use bottom-up allocations. [1] https://lkml.org/lkml/2018/11/22/1032 Link: http://lkml.kernel.org/r/1543233216-25833-1-git-send-email-rppt@xxxxxxxxxxxxx Reported-by: Meelis Roos <mroos@xxxxxxxx> Signed-off-by: Mike Rapoport <rppt@xxxxxxxxxxxxx> Tested-by: Meelis Roos <mroos@xxxxxxxx> Cc: Richard Henderson <rth@xxxxxxxxxxx> Cc: Ivan Kokshaysky <ink@xxxxxxxxxxxxxxxxxxxx> Cc: Matt Turner <mattst88@xxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- arch/alpha/kernel/setup.c | 1 + arch/alpha/mm/numa.c | 6 +++--- 2 files changed, 4 insertions(+), 3 deletions(-) --- a/arch/alpha/kernel/setup.c~alpha-fix-hang-caused-by-the-bootmem-removal +++ a/arch/alpha/kernel/setup.c @@ -634,6 +634,7 @@ setup_arch(char **cmdline_p) /* Find our memory. */ setup_memory(kernel_end); + memblock_set_bottom_up(true); /* First guess at cpu cache sizes. Do this before init_arch. */ determine_cpu_caches(cpu->type); --- a/arch/alpha/mm/numa.c~alpha-fix-hang-caused-by-the-bootmem-removal +++ a/arch/alpha/mm/numa.c @@ -144,14 +144,14 @@ setup_memory_node(int nid, void *kernel_ if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn)) panic("kernel loaded out of ram"); + memblock_add(PFN_PHYS(node_min_pfn), + (node_max_pfn - node_min_pfn) << PAGE_SHIFT); + /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned. Note that we round this down, not up - node memory has much larger alignment than 8Mb, so it's safe. */ node_min_pfn &= ~((1UL << (MAX_ORDER-1))-1); - memblock_add(PFN_PHYS(node_min_pfn), - (node_max_pfn - node_min_pfn) << PAGE_SHIFT); - NODE_DATA(nid)->node_start_pfn = node_min_pfn; NODE_DATA(nid)->node_present_pages = node_max_pfn - node_min_pfn; _
