Re: Crashes/hung tasks with z3pool under memory pressure

[Vitaly Wool]

Hey Guenter,

On 04/13/2018 07:56 PM, Guenter Roeck wrote:

> On Fri, Apr 13, 2018 at 05:40:18PM +0000, Vitaly Wool wrote:
> > On Fri, Apr 13, 2018, 7:35 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> >
> > > On Fri, Apr 13, 2018 at 05:21:02AM +0000, Vitaly Wool wrote:
> > > > Hi Guenter,
> > > >
> > > >
> > > > Den fre 13 apr. 2018 kl 00:01 skrev Guenter Roeck <linux@xxxxxxxxxxxx>:
> > > >
> > > > > Hi all,
> > > > > we are observing crashes with z3pool under memory pressure. The kernel
> > > > version
> > > > > used to reproduce the problem is v4.16-11827-g5d1365940a68, but the
> > > > problem was
> > > > > also seen with v4.14 based kernels.
> > > >
> > > > just before I dig into this, could you please try reproducing the errors
> > > > you see with https://patchwork.kernel.org/patch/10210459/ applied?
> > > As mentioned above, I tested with v4.16-11827-g5d1365940a68, which already
> > > includes this patch.
> > Bah. Sorry. Expect an update after the weekend.
> NP; easy to miss. Thanks a lot for looking into it.
I wonder if the following patch would make a difference:

diff --git a/mm/z3fold.c b/mm/z3fold.c
index c0bca6153b95..5e547c2d5832 100644
--- a/mm/z3fold.c
+++ b/mm/z3fold.c
@@ -887,19 +887,21 @@ static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 				goto next;
 		}
 next:
-		spin_lock(&pool->lock);
 		if (test_bit(PAGE_HEADLESS, &page->private)) {
 			if (ret == 0) {
-				spin_unlock(&pool->lock);
 				free_z3fold_page(page);
 				return 0;
 			}
-		} else if (kref_put(&zhdr->refcount, release_z3fold_page)) {
-			atomic64_dec(&pool->pages_nr);
-			spin_unlock(&pool->lock);
-			return 0;
+		} else {
+			spin_lock(&zhdr->page_lock);
+			if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
+				atomic64_dec(&pool->pages_nr);
+				return 0;
+			}
+			spin_unlock(&zhdr->page_lock);
 		}
 
+		spin_lock(&pool->lock);
 		/*
 		 * Add to the beginning of LRU.
 		 * Pool lock has to be kept here to ensure the page has


Thanks,
   Vitaly