On 31.12.24 09:41, Hillf Danton wrote:
On Fri, 27 Dec 2024 20:56:21 -0800
syzbot has found a reproducer for the following issue on:
HEAD commit: 8155b4ef3466 Add linux-next specific files for 20241220
git tree: linux-next
syz repro: https://syzkaller.appspot.com/x/repro.syz?x=1652fadf980000
#syz test
--- x/mm/filemap.c
+++ y/mm/filemap.c
@@ -3636,6 +3636,10 @@ static vm_fault_t filemap_map_folio_rang
continue;
skip:
if (count) {
+ for (unsigned int i = 0; i < count; i++) {
+ if (page_folio(page + i) != folio)
+ goto out;
+ }
IIRC, count <= nr_pages. Wouldn't that mean that we somehow pass in
nr_pages that already exceeds the given folio+start?
When I last looked at this, I was not able to spot the error in the
caller :(
set_pte_range(vmf, folio, page, count, addr);
*rss += count;
folio_ref_add(folio, count);
@@ -3658,6 +3662,7 @@ skip:
ret = VM_FAULT_NOPAGE;
}
+out:
vmf->pte = old_ptep;
return ret;
@@ -3702,7 +3707,7 @@ vm_fault_t filemap_map_pages(struct vm_f
struct file *file = vma->vm_file;
struct address_space *mapping = file->f_mapping;
pgoff_t file_end, last_pgoff = start_pgoff;
- unsigned long addr;
+ unsigned long addr, pmd_end;
XA_STATE(xas, &mapping->i_pages, start_pgoff);
struct folio *folio;
vm_fault_t ret = 0;
@@ -3731,6 +3736,12 @@ vm_fault_t filemap_map_pages(struct vm_f
if (end_pgoff > file_end)
end_pgoff = file_end;
+ /* make vmf->pte[x] valid */
+ pmd_end = ALIGN(addr, PMD_SIZE);
+ pmd_end = (pmd_end - addr) >> PAGE_SHIFT;
+ if (end_pgoff - start_pgoff > pmd_end)
+ end_pgoff = start_pgoff + pmd_end;
+
do_fault_around() comments "This way it's easier to guarantee that we
don't cross page table boundaries."
It does some magic with PTRS_PER_PTE.
You're diff here seems to indicate that this is not the case?
But it's rather surprising that we see these issues pop up just now in
-next.
--
Cheers,
David / dhildenb
