[PATCH 5.15 881/917] mm, thp: fix incorrect unmap behavior for private pages

[Greg Kroah-Hartman]

From: Rongwei Wang <rongwei.wang@xxxxxxxxxxxxxxxxx>

commit 8468e937df1f31411d1e127fa38db064af051fe5 upstream.

When truncating pagecache on file THP, the private pages of a process
should not be unmapped mapping.  This incorrect behavior on a dynamic
shared libraries which will cause related processes to happen core dump.

A simple test for a DSO (Prerequisite is the DSO mapped in file THP):

    int main(int argc, char *argv[])
    {
	int fd;

	fd = open(argv[1], O_WRONLY);
	if (fd < 0) {
		perror("open");
	}

	close(fd);
	return 0;
    }

The test only to open a target DSO, and do nothing.  But this operation
will lead one or more process to happen core dump.  This patch mainly to
fix this bug.

Link: https://lkml.kernel.org/r/20211025092134.18562-3-rongwei.wang@xxxxxxxxxxxxxxxxx
Fixes: eb6ecbed0aa2 ("mm, thp: relax the VM_DENYWRITE constraint on file-backed THPs")
Signed-off-by: Rongwei Wang <rongwei.wang@xxxxxxxxxxxxxxxxx>
Tested-by: Xu Yu <xuyu@xxxxxxxxxxxxxxxxx>
Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
Cc: Song Liu <song@xxxxxxxxxx>
Cc: William Kucharski <william.kucharski@xxxxxxxxxx>
Cc: Hugh Dickins <hughd@xxxxxxxxxx>
Cc: Yang Shi <shy828301@xxxxxxxxx>
Cc: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
Cc: Collin Fijalkovich <cfijalkovich@xxxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
---
 fs/open.c |   11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

--- a/fs/open.c
+++ b/fs/open.c
@@ -857,8 +857,17 @@ static int do_dentry_open(struct file *f
 		 */
 		smp_mb();
 		if (filemap_nr_thps(inode->i_mapping)) {
+			struct address_space *mapping = inode->i_mapping;
+
 			filemap_invalidate_lock(inode->i_mapping);
-			truncate_pagecache(inode, 0);
+			/*
+			 * unmap_mapping_range just need to be called once
+			 * here, because the private pages is not need to be
+			 * unmapped mapping (e.g. data segment of dynamic
+			 * shared libraries here).
+			 */
+			unmap_mapping_range(mapping, 0, 0, 0);
+			truncate_inode_pages(mapping, 0);
 			filemap_invalidate_unlock(inode->i_mapping);
 		}
 	}