Re: [PATCH 4.9-stable -- 5.19-stable] mm/hugetlb: fix hugetlb not supporting softdirty tracking

From: Greg KH
Date: Mon Aug 29 2022 - 04:14:09 EST


On Thu, Aug 25, 2022 at 04:32:58PM +0200, David Hildenbrand wrote:
> commit f96f7a40874d7c746680c0b9f57cef2262ae551f upstream.
>
> Patch series "mm/hugetlb: fix write-fault handling for shared mappings", v2.
>
> I observed that hugetlb does not support/expect write-faults in shared
> mappings that would have to map the R/O-mapped page writable -- and I
> found two case where we could currently get such faults and would
> erroneously map an anon page into a shared mapping.
>
> Reproducers part of the patches.
>
> I propose to backport both fixes to stable trees. The first fix needs a
> small adjustment.
>
> This patch (of 2):
>
> Staring at hugetlb_wp(), one might wonder where all the logic for shared
> mappings is when stumbling over a write-protected page in a shared
> mapping. In fact, there is none, and so far we thought we could get away
> with that because e.g., mprotect() should always do the right thing and
> map all pages directly writable.
>
> Looks like we were wrong:
>
> --------------------------------------------------------------------------
> #include <stdio.h>
> #include <stdlib.h>
> #include <string.h>
> #include <fcntl.h>
> #include <unistd.h>
> #include <errno.h>
> #include <sys/mman.h>
>
> #define HUGETLB_SIZE (2 * 1024 * 1024u)
>
> static void clear_softdirty(void)
> {
> int fd = open("/proc/self/clear_refs", O_WRONLY);
> const char *ctrl = "4";
> int ret;
>
> if (fd < 0) {
> fprintf(stderr, "open(clear_refs) failed\n");
> exit(1);
> }
> ret = write(fd, ctrl, strlen(ctrl));
> if (ret != strlen(ctrl)) {
> fprintf(stderr, "write(clear_refs) failed\n");
> exit(1);
> }
> close(fd);
> }
>
> int main(int argc, char **argv)
> {
> char *map;
> int fd;
>
> fd = open("/dev/hugepages/tmp", O_RDWR | O_CREAT);
> if (!fd) {
> fprintf(stderr, "open() failed\n");
> return -errno;
> }
> if (ftruncate(fd, HUGETLB_SIZE)) {
> fprintf(stderr, "ftruncate() failed\n");
> return -errno;
> }
>
> map = mmap(NULL, HUGETLB_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
> if (map == MAP_FAILED) {
> fprintf(stderr, "mmap() failed\n");
> return -errno;
> }
>
> *map = 0;
>
> if (mprotect(map, HUGETLB_SIZE, PROT_READ)) {
> fprintf(stderr, "mmprotect() failed\n");
> return -errno;
> }
>
> clear_softdirty();
>
> if (mprotect(map, HUGETLB_SIZE, PROT_READ|PROT_WRITE)) {
> fprintf(stderr, "mmprotect() failed\n");
> return -errno;
> }
>
> *map = 0;
>
> return 0;
> }
> --------------------------------------------------------------------------
>
> Above test fails with SIGBUS when there is only a single free hugetlb page.
> # echo 1 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
> # ./test
> Bus error (core dumped)
>
> And worse, with sufficient free hugetlb pages it will map an anonymous page
> into a shared mapping, for example, messing up accounting during unmap
> and breaking MAP_SHARED semantics:
> # echo 2 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
> # ./test
> # cat /proc/meminfo | grep HugePages_
> HugePages_Total: 2
> HugePages_Free: 1
> HugePages_Rsvd: 18446744073709551615
> HugePages_Surp: 0
>
> Reason in this particular case is that vma_wants_writenotify() will
> return "true", removing VM_SHARED in vma_set_page_prot() to map pages
> write-protected. Let's teach vma_wants_writenotify() that hugetlb does not
> support softdirty tracking.
>
> Link: https://lkml.kernel.org/r/20220811103435.188481-1-david@xxxxxxxxxx
> Link: https://lkml.kernel.org/r/20220811103435.188481-2-david@xxxxxxxxxx
> Fixes: 64e455079e1b ("mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared")
> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> Reviewed-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> Cc: Peter Feiner <pfeiner@xxxxxxxxxx>
> Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> Cc: Cyrill Gorcunov <gorcunov@xxxxxxxxxx>
> Cc: Pavel Emelyanov <xemul@xxxxxxxxxxxxx>
> Cc: Jamie Liu <jamieliu@xxxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
> Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> Cc: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> Cc: Peter Xu <peterx@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx> [3.18+]
> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> ---
> mm/mmap.c | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)

Now queued up, thanks.

greg k-h