thunderx: eliminate extra calls to put_page() for pages held for recycling
authorDean Nelson <dnelson@redhat.com>
Tue, 26 Mar 2019 15:53:26 +0000 (11:53 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 3 Apr 2019 04:26:19 +0000 (06:26 +0200)
[ Upstream commit cd35ef91490ad8049dd180bb060aff7ee192eda9 ]

For the non-XDP case, commit 773225388dae15e72790 ("net: thunderx: Optimize
page recycling for XDP") added code to nicvf_free_rbdr() that, when releasing
the additional receive buffer page reference held for recycling, repeatedly
calls put_page() until the page's _refcount goes to zero. Which results in
the page being freed.

This is not okay if the page's _refcount was greater than 1 (in the non-XDP
case), because nicvf_free_rbdr() should not be subtracting more than what
nicvf_alloc_page() had previously added to the page's _refcount, which was
only 1 (in the non-XDP case).

This can arise if a received packet is still being processed and the receive
buffer (i.e., skb->head) has not yet been freed via skb_free_head() when
nicvf_free_rbdr() is spinning through the aforementioned put_page() loop.

If this should occur, when the received packet finishes processing and
skb_free_head() is called, various problems can ensue. Exactly what, depends on
whether the page has already been reallocated or not, anything from "BUG: Bad
page state ... ", to "Unable to handle kernel NULL pointer dereference ..." or
"Unable to handle kernel paging request...".

So this patch changes nicvf_free_rbdr() to only call put_page() once for pages
held for recycling (in the non-XDP case).

Fixes: 773225388dae ("net: thunderx: Optimize page recycling for XDP")
Signed-off-by: Dean Nelson <dnelson@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/net/ethernet/cavium/thunder/nicvf_queues.c

index ade1ff3..9a4cfa6 100644 (file)
@@ -364,11 +364,10 @@ static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
        while (head < rbdr->pgcnt) {
                pgcache = &rbdr->pgcache[head];
                if (pgcache->page && page_ref_count(pgcache->page) != 0) {
-                       if (!rbdr->is_xdp) {
-                               put_page(pgcache->page);
-                               continue;
+                       if (rbdr->is_xdp) {
+                               page_ref_sub(pgcache->page,
+                                            pgcache->ref_count - 1);
                        }
-                       page_ref_sub(pgcache->page, pgcache->ref_count - 1);
                        put_page(pgcache->page);
                }
                head++;