[PATCH] zram: bug fix: delay lock holding in zram_slot_free_noity

[Minchan Kim <minchan@kernel.org>]

I was preparing to promote zram and it was almost done.
Before sending patch, I tried to test and eyebrows went up.

[1] introduced down_write in zram_slot_free_notify to prevent race
between zram_slot_free_notify and zram_bvec_[read|write]. The race
could happen if somebody who has right permission to open swap device
is reading swap device while it is used by swap in parallel.

However, zram_slot_free_notify is called with holding spin_lock of
swap layer so we shouldn't avoid holing mutex. Otherwise, lockdep
warns it.

I guess, best solution is to redesign zram lock scheme totally but
we are on the verge of promoting so it's not desirable to change a lot
critical code and such big change isn't good shape for backporting to
stable trees so I think the simple patch is best at the moment.

[1] [57ab0485, zram: use zram->lock to protect zram_free_page()
    in swap free notify path]

Cc: Jiang Liu <jiang.liu@huawei.com>
Cc: Nitin Gupta <ngupta@vflare.org>
Cc: stable@vger.kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
---
 drivers/staging/zram/zram_drv.c |   15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/drivers/staging/zram/zram_drv.c b/drivers/staging/zram/zram_drv.c
index 7ebf91d..7b574c4 100644
--- a/drivers/staging/zram/zram_drv.c
+++ b/drivers/staging/zram/zram_drv.c
@@ -440,6 +440,13 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 		goto out;
 	}
 
+	/*
+	 * zram_slot_free_notify could miss free so that let's
+	 * double check.
+	 */
+	if (unlikely(meta->table[index].handle))
+		zram_free_page(zram, index);
+
 	ret = lzo1x_1_compress(uncmem, PAGE_SIZE, src, &clen,
 			       meta->compress_workmem);
 
@@ -727,7 +734,13 @@ static void zram_slot_free_notify(struct block_device *bdev,
 	struct zram *zram;
 
 	zram = bdev->bd_disk->private_data;
-	down_write(&zram->lock);
+	/*
+	 * The function is called in atomic context so down_write should
+	 * be prohibited. If we couldn't hold a mutex, the free could be
+	 * handled by zram_bvec_write later when same index is overwritten.
+	 */
+	if (!down_write_trylock(&zram->lock))
+		return;
 	zram_free_page(zram, index);
 	up_write(&zram->lock);
 	atomic64_inc(&zram->stats.notify_free);
-- 
1.7.9.5