mm, mremap: fix mremap() expanding vma with addr inside vma

Since 6.1 we have noticed random rpm install failures that were tracked to
mremap() returning -ENOMEM and to commit ca3d76b0aa ("mm: add merging
after mremap resize").

The problem occurs when mremap() expands a VMA in place, but using an
starting address that's not vma->vm_start, but somewhere in the middle. 
The extension_pgoff calculation introduced by the commit is wrong in that
case, so vma_merge() fails due to pgoffs not being compatible.  Fix the
calculation.

By the way it seems that the situations, where rpm now expands a vma from
the middle, were made possible also due to that commit, thanks to the
improved vma merging.  Yet it should work just fine, except for the buggy
calculation.

Link: https://lkml.kernel.org/r/20221216163227.24648-1-vbabka@suse.cz
Reported-by: Jiri Slaby <jirislaby@kernel.org>
  Link: https://bugzilla.suse.com/show_bug.cgi?id=1206359
Fixes: ca3d76b0aa ("mm: add merging after mremap resize")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Jakub Matěna <matenajakub@gmail.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/mremap.c

@@ -1016,7 +1016,8 @@ SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 			long pages = (new_len - old_len) >> PAGE_SHIFT;
 			unsigned long extension_start = addr + old_len;
 			unsigned long extension_end = addr + new_len;
-			pgoff_t extension_pgoff = vma->vm_pgoff + (old_len >> PAGE_SHIFT);
+			pgoff_t extension_pgoff = vma->vm_pgoff +
+				((extension_start - vma->vm_start) >> PAGE_SHIFT);
 
 			if (vma->vm_flags & VM_ACCOUNT) {
 				if (security_vm_enough_memory_mm(mm, pages)) {