skbuff: fix coalescing for page_pool fragment recycling
authorJean-Philippe Brucker <jean-philippe@linaro.org>
Thu, 31 Mar 2022 10:24:41 +0000 (11:24 +0100)
committerDavid S. Miller <davem@davemloft.net>
Fri, 1 Apr 2022 10:57:58 +0000 (11:57 +0100)
commit1effe8ca4e34c34cdd9318436a4232dcb582ebf4
tree1dd240786f4f799b53d8410256a692b3bd395194
parent012d69fbfcc739f846766c1da56ef8b493b803b5
skbuff: fix coalescing for page_pool fragment recycling

Fix a use-after-free when using page_pool with page fragments. We
encountered this problem during normal RX in the hns3 driver:

(1) Initially we have three descriptors in the RX queue. The first one
    allocates PAGE1 through page_pool, and the other two allocate one
    half of PAGE2 each. Page references look like this:

                RX_BD1 _______ PAGE1
                RX_BD2 _______ PAGE2
                RX_BD3 _________/

(2) Handle RX on the first descriptor. Allocate SKB1, eventually added
    to the receive queue by tcp_queue_rcv().

(3) Handle RX on the second descriptor. Allocate SKB2 and pass it to
    netif_receive_skb():

    netif_receive_skb(SKB2)
      ip_rcv(SKB2)
        SKB3 = skb_clone(SKB2)

    SKB2 and SKB3 share a reference to PAGE2 through
    skb_shinfo()->dataref. The other ref to PAGE2 is still held by
    RX_BD3:

                      SKB2 ---+- PAGE2
                      SKB3 __/   /
                RX_BD3 _________/

 (3b) Now while handling TCP, coalesce SKB3 with SKB1:

      tcp_v4_rcv(SKB3)
        tcp_try_coalesce(to=SKB1, from=SKB3)    // succeeds
        kfree_skb_partial(SKB3)
          skb_release_data(SKB3)                // drops one dataref

                      SKB1 _____ PAGE1
                           \____
                      SKB2 _____ PAGE2
                                 /
                RX_BD3 _________/

    In skb_try_coalesce(), __skb_frag_ref() takes a page reference to
    PAGE2, where it should instead have increased the page_pool frag
    reference, pp_frag_count. Without coalescing, when releasing both
    SKB2 and SKB3, a single reference to PAGE2 would be dropped. Now
    when releasing SKB1 and SKB2, two references to PAGE2 will be
    dropped, resulting in underflow.

 (3c) Drop SKB2:

      af_packet_rcv(SKB2)
        consume_skb(SKB2)
          skb_release_data(SKB2)                // drops second dataref
            page_pool_return_skb_page(PAGE2)    // drops one pp_frag_count

                      SKB1 _____ PAGE1
                           \____
                                 PAGE2
                                 /
                RX_BD3 _________/

(4) Userspace calls recvmsg()
    Copies SKB1 and releases it. Since SKB3 was coalesced with SKB1, we
    release the SKB3 page as well:

    tcp_eat_recv_skb(SKB1)
      skb_release_data(SKB1)
        page_pool_return_skb_page(PAGE1)
        page_pool_return_skb_page(PAGE2)        // drops second pp_frag_count

(5) PAGE2 is freed, but the third RX descriptor was still using it!
    In our case this causes IOMMU faults, but it would silently corrupt
    memory if the IOMMU was disabled.

Change the logic that checks whether pp_recycle SKBs can be coalesced.
We still reject differing pp_recycle between 'from' and 'to' SKBs, but
in order to avoid the situation described above, we also reject
coalescing when both 'from' and 'to' are pp_recycled and 'from' is
cloned.

The new logic allows coalescing a cloned pp_recycle SKB into a page
refcounted one, because in this case the release (4) will drop the right
reference, the one taken by skb_try_coalesce().

Fixes: 53e0961da1c7 ("page_pool: add frag page recycling support in page pool")
Suggested-by: Alexander Duyck <alexanderduyck@fb.com>
Signed-off-by: Jean-Philippe Brucker <jean-philippe@linaro.org>
Reviewed-by: Yunsheng Lin <linyunsheng@huawei.com>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/skbuff.c