Previously, we shared too much of the code with COPY and ZEROPAGE, so we
manipulated things in various invalid ways:
- Previously, we unconditionally called shmem_inode_acct_block. In the
continue case, we're looking up an existing page which would have been
accounted for properly when it was allocated. So doing it twice
results in double-counting, and eventually leaking.
- Previously, we made the pte writable whenever the VMA was writable.
However, for continue, consider this case:
1. A tmpfs file was created
2. The non-UFFD-registered side mmap()-s with MAP_SHARED
3. The UFFD-registered side mmap()-s with MAP_PRIVATE
In this case, even though the UFFD-registered VMA may be writable, we
still want CoW behavior. So, check for this case and don't make the
pte writable.
- The initial pgoff / max_off check isn't necessary, so we can skip past
it. The second one seems likely to be unnecessary too, but keep it
just in case. Modify both checks to use pgoff, as offset is equivalent
and not needed.
- Previously, we unconditionally called ClearPageDirty() in the error
path. In the continue case though, since this is an existing page, it
might have already been dirty before we started touching it. It's very
problematic to clear the bit incorrectly, but not a problem to leave
it - so, just omit the ClearPageDirty() entirely.
- Previously, we unconditionally removed the page from the page cache in
the error path. But in the continue case, we didn't add it - it was
already there because the page is present in some second
(non-UFFD-registered) mapping. So, removing it is invalid.
Because the error handling issues are easy to exercise in the selftest,
make a small modification there to do so.
Finally, refactor shmem_mcopy_atomic_pte a bit. By this point, we've
added a lot of "if (!is_continue)"-s everywhere. It's cleaner to just
check for that mode first thing, and then "goto" down to where the parts
we actually want are. This leaves the code in between cleaner.
Changes since v2:
- Drop the ClearPageDirty() entirely, instead of trying to remember the
old value.
- Modify both pgoff / max_off checks to use pgoff. It's equivalent to
offset, but offset wasn't initialized until the first check (which
we're skipping).
- Keep the second pgoff / max_off check in the continue case.
Changes since v1:
- Refactor to skip ahead with goto, instead of adding several more
"if (!is_continue)".
- Fix unconditional ClearPageDirty().
- Don't pte_mkwrite() when is_continue && !VM_SHARED.
Fixes: 00da60b9d0a0 ("userfaultfd: support minor fault handling for shmem")
Signed-off-by: Axel Rasmussen <axelrasmussen@xxxxxxxxxx>
---
mm/shmem.c | 60 +++++++++++++-----------
tools/testing/selftests/vm/userfaultfd.c | 12 +++++
2 files changed, 44 insertions(+), 28 deletions(-)
diff --git a/mm/shmem.c b/mm/shmem.c
index d2e0e81b7d2e..fbcce850a16e 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2377,18 +2377,22 @@ int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
struct page *page;
pte_t _dst_pte, *dst_pte;
int ret;
- pgoff_t offset, max_off;
-
- ret = -ENOMEM;
- if (!shmem_inode_acct_block(inode, 1))
- goto out;
+ pgoff_t max_off;
+ int writable;
if (is_continue) {
ret = -EFAULT;
page = find_lock_page(mapping, pgoff);
if (!page)
- goto out_unacct_blocks;
- } else if (!*pagep) {
+ goto out;
+ goto install_ptes;
+ }
+
+ ret = -ENOMEM;
+ if (!shmem_inode_acct_block(inode, 1))
+ goto out;
+
+ if (!*pagep) {
page = shmem_alloc_page(gfp, info, pgoff);
if (!page)
goto out_unacct_blocks;
@@ -2415,30 +2419,29 @@ int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
*pagep = NULL;
}
- if (!is_continue) {
- VM_BUG_ON(PageSwapBacked(page));
- VM_BUG_ON(PageLocked(page));
- __SetPageLocked(page);
- __SetPageSwapBacked(page);
- __SetPageUptodate(page);
- }
+ VM_BUG_ON(PageSwapBacked(page));
+ VM_BUG_ON(PageLocked(page));
+ __SetPageLocked(page);
+ __SetPageSwapBacked(page);
+ __SetPageUptodate(page);
ret = -EFAULT;
- offset = linear_page_index(dst_vma, dst_addr);
max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
- if (unlikely(offset >= max_off))
+ if (unlikely(pgoff >= max_off))
goto out_release;
- /* If page wasn't already in the page cache, add it. */
- if (!is_continue) {
- ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
- gfp & GFP_RECLAIM_MASK, dst_mm);
- if (ret)
- goto out_release;
- }
+ ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL,
+ gfp & GFP_RECLAIM_MASK, dst_mm);
+ if (ret)
+ goto out_release;
+install_ptes:
_dst_pte = mk_pte(page, dst_vma->vm_page_prot);
- if (dst_vma->vm_flags & VM_WRITE)
+ /* For CONTINUE on a non-shared VMA, don't pte_mkwrite for CoW. */
+ writable = is_continue && !(dst_vma->vm_flags & VM_SHARED)
+ ? 0
+ : dst_vma->vm_flags & VM_WRITE;
+ if (writable)
_dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
else {
/*
@@ -2455,7 +2458,7 @@ int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
ret = -EFAULT;
max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
- if (unlikely(offset >= max_off))
+ if (unlikely(pgoff >= max_off))
goto out_release_unlock;
ret = -EEXIST;
@@ -2485,13 +2488,14 @@ int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
return ret;
out_release_unlock:
pte_unmap_unlock(dst_pte, ptl);
- ClearPageDirty(page);
- delete_from_page_cache(page);
+ if (!is_continue)
+ delete_from_page_cache(page);
out_release:
unlock_page(page);
put_page(page);
out_unacct_blocks:
- shmem_inode_unacct_blocks(inode, 1);
+ if (!is_continue)
+ shmem_inode_unacct_blocks(inode, 1);
goto out;
}
#endif /* CONFIG_USERFAULTFD */
diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index f6c86b036d0f..d8541a59dae5 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c
@@ -485,6 +485,7 @@ static void wp_range(int ufd, __u64 start, __u64 len, bool wp)
static void continue_range(int ufd, __u64 start, __u64 len)
{
struct uffdio_continue req;
+ int ret;
req.range.start = start;
req.range.len = len;
@@ -493,6 +494,17 @@ static void continue_range(int ufd, __u64 start, __u64 len)
if (ioctl(ufd, UFFDIO_CONTINUE, &req))
err("UFFDIO_CONTINUE failed for address 0x%" PRIx64,
(uint64_t)start);
+
+ /*
+ * Error handling within the kernel for continue is subtly different
+ * from copy or zeropage, so it may be a source of bugs. Trigger an
+ * error (-EEXIST) on purpose, to verify doing so doesn't cause a BUG.
+ */
+ req.mapped = 0;
+ ret = ioctl(ufd, UFFDIO_CONTINUE, &req);
+ if (ret >= 0 || req.mapped != -EEXIST)
+ err("failed to exercise UFFDIO_CONTINUE error handling, ret=%d, mapped=%" PRId64,
+ ret, req.mapped);
}
static void *locking_thread(void *arg)
--
2.31.0.291.g576ba9dcdaf-goog
