On 18.02.2020 14:38, Sachin Sant wrote: > > >> On 18-Feb-2020, at 4:20 PM, Kirill Tkhai <ktk...@virtuozzo.com> wrote: >> >> Hi, Sachin, >> >> On 18.02.2020 13:45, Sachin Sant wrote: >>> >>> commit a75056fc1e7c >>> mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node >>> >>> I can boot the kernel successfully if the patch is reverted. >> >> >> could you please test your boot with original patch from here: >> >> https://patchwork.kernel.org/patch/11360007/ >> >> ? > With this original patch I can boot the machine successfully.
Ok, thanks. I think, there is no a problem in the commited patch, since mem_cgroup_css_alloc() is called from the place, where any memory allocations have to be allowed. This is one of the reason, memory_cgrp_subsys.early_init is 0, and all nodes allocations should be availeble there. The problem is not in vmalloc() itself, since the second patch with kmalloc_node() also fails on your setup. Maybe, the reproduction depends on amount of allocated memory. For me this looks like a problem in powerpc, but it would be interesting to hear some comments from powerpc guys. For now we may replace the commited patch with v2 (https://patchwork.kernel.org/patch/11360007/) containing workaround, which we have in another alloc_mem_cgroup_per_node_info() allocations. Kirill
