On Sun, May 08, 2022 at 07:47:02PM -0700, Sultan Alsawaf wrote:
> From: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
>
> The asynchronous zspage free worker tries to lock a zspage's entire page
> list without defending against page migration. Since pages which haven't
> yet been locked can concurrently migrate off the zspage page list while
> lock_zspage() churns away, lock_zspage() can suffer from a few different
> lethal races. It can lock a page which no longer belongs to the zspage and
> unsafely dereference page_private(), it can unsafely dereference a torn
> pointer to the next page (since there's a data race), and it can observe a
> spurious NULL pointer to the next page and thus not lock all of the
> zspage's pages (since a single page migration will reconstruct the entire
> page list, and create_page_chain() unconditionally zeroes out each list
> pointer in the process).
>
> Fix the races by using migrate_read_lock() in lock_zspage() to synchronize
> with page migration.
>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: 48b4800a1c6a ("zsmalloc: page migration support")
Shouldn't the fix be Fixes: 77ff465799c6 ("zsmalloc: zs_page_migrate: skip
unnecessary loops but not return -EBUSY if zspage is not inuse)?
Because we didn't migrate ZS_EMPTY pages before.
> Signed-off-by: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
> ---
> mm/zsmalloc.c | 37 +++++++++++++++++++++++++++++++++----
> 1 file changed, 33 insertions(+), 4 deletions(-)
>
> diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
> index 9152fbde33b5..5d5fc04385b8 100644
> --- a/mm/zsmalloc.c
> +++ b/mm/zsmalloc.c
> @@ -1718,11 +1718,40 @@ static enum fullness_group putback_zspage(struct size_class *class,
> */
> static void lock_zspage(struct zspage *zspage)
> {
> - struct page *page = get_first_page(zspage);
> + struct page *curr_page, *page;
>
> - do {
> - lock_page(page);
> - } while ((page = get_next_page(page)) != NULL);
> + /*
> + * Pages we haven't locked yet can be migrated off the list while we're
> + * trying to lock them, so we need to be careful and only attempt to
> + * lock each page under migrate_read_lock(). Otherwise, the page we lock
> + * may no longer belong to the zspage. This means that we may wait for
> + * the wrong page to unlock, so we must take a reference to the page
> + * prior to waiting for it to unlock outside migrate_read_lock().
I couldn't get the point here. Why couldn't we simple lock zspage migration?
diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
index 9152fbde33b5..05ff2315b7b1 100644
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -1987,7 +1987,10 @@ static void async_free_zspage(struct work_struct *work)
list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
list_del(&zspage->list);
+
+ migrate_read_lock(zspage);
lock_zspage(zspage);
+ migrate_read_unlock(zspage);
get_zspage_mapping(zspage, &class_idx, &fullness);
VM_BUG_ON(fullness != ZS_EMPTY);
> + */
> + while (1) {
> + migrate_read_lock(zspage);
> + page = get_first_page(zspage);
> + if (trylock_page(page))
> + break;
> + get_page(page);
> + migrate_read_unlock(zspage);
> + wait_on_page_locked(page);
> + put_page(page);
> + }
> +
> + curr_page = page;
> + while ((page = get_next_page(curr_page))) {
> + if (trylock_page(page)) {
> + curr_page = page;
> + } else {
> + get_page(page);
> + migrate_read_unlock(zspage);
> + wait_on_page_locked(page);
> + put_page(page);
> + migrate_read_lock(zspage);
> + }
> + }
> + migrate_read_unlock(zspage);
> }
>
> static int zs_init_fs_context(struct fs_context *fc)
> --
> 2.36.0
>
