On Wed, Dec 08, 2021 at 07:53:43PM +0300, Andrey Ryabinin wrote:

alloc_pages_vma() may try to allocate THP page on the local NUMA node first: page = __alloc_pages_node(hpage_node, gfp | __GFP_THISNODE | __GFP_NORETRY, order);

And if the allocation fails it retries allowing remote memory:

if (!page && (gfp & __GFP_DIRECT_RECLAIM)) page = __alloc_pages_node(hpage_node, gfp, order);

However, this retry allocation completely ignores memory policy nodemask allowing allocation to escape restrictions.

The first appearance of this bug seems to be the commit ac5b2c18911f ("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") The bug disappeared later in the commit 89c83fb539f9 ("mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask") and reappeared again in slightly different form in the commit 76e654cc91bb ("mm, page_alloc: allow hugepage fallback to remote nodes when madvised")

Fix this by passing correct nodemask to the __alloc_pages() call.

The demonstration/reproducer of the problem: $ mount -oremount,size=4G,huge=always /dev/shm/ $ echo always > /sys/kernel/mm/transparent_hugepage/defrag $ cat mbind_thp.c #include <unistd.h> #include <sys/mman.h> #include <sys/stat.h> #include <fcntl.h> #include <assert.h> #include <stdlib.h> #include <stdio.h> #include <numaif.h>

#define SIZE 2ULL << 30 int main(int argc, char **argv) { int fd; unsigned long long i; char *addr; pid_t pid; char buf[100]; unsigned long nodemask = 1;

fd = open("/dev/shm/test", O_RDWR|O_CREAT); assert(fd > 0); assert(ftruncate(fd, SIZE) == 0);

addr = mmap(NULL, SIZE, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);

assert(mbind(addr, SIZE, MPOL_BIND, &nodemask, 2, MPOL_MF_STRICT|MPOL_MF_MOVE)==0); for (i = 0; i < SIZE; i+=4096) { addr[i] = 1; } pid = getpid(); snprintf(buf, sizeof(buf), "grep shm /proc/%d/numa_maps", pid); system(buf); sleep(10000);

return 0; } $ gcc mbind_thp.c -o mbind_thp -lnuma $ numactl -H available: 2 nodes (0-1) node 0 cpus: 0 2 node 0 size: 1918 MB node 0 free: 1595 MB node 1 cpus: 1 3 node 1 size: 2014 MB node 1 free: 1731 MB node distances: node 0 1 0: 10 20 1: 20 10 $ rm -f /dev/shm/test; taskset -c 0 ./mbind_thp 7fd970a00000 bind:0 file=/dev/shm/test dirty=524288 active=0 N0=396800 N1=127488 kernelpagesize_kB=4

Fixes: ac5b2c18911f ("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings") Signed-off-by: Andrey Ryabinin arbn@yandex-team.com Cc: stable@vger.kernel.org Cc: Andrea Arcangeli aarcange@redhat.com Cc: Michal Hocko mhocko@suse.com Cc: Mel Gorman mgorman@techsingularity.net Cc: David Rientjes rientjes@google.com

Acked-by: Mel Gorman mgorman@techsingularity.net