e1000: add dummy allocator to fix race condition between mtu change and netpoll
authorSabrina Dubroca <sd@queasysnail.net>
Thu, 26 Feb 2015 05:35:41 +0000 (05:35 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Fri, 6 Mar 2015 10:47:10 +0000 (02:47 -0800)
There is a race condition between e1000_change_mtu's cleanups and
netpoll, when we change the MTU across jumbo size:

Changing MTU frees all the rx buffers:
    e1000_change_mtu -> e1000_down -> e1000_clean_all_rx_rings ->
        e1000_clean_rx_ring

Then, close to the end of e1000_change_mtu:
    pr_info -> ... -> netpoll_poll_dev -> e1000_clean ->
        e1000_clean_rx_irq -> e1000_alloc_rx_buffers -> e1000_alloc_frag

And when we come back to do the rest of the MTU change:
    e1000_up -> e1000_configure -> e1000_configure_rx ->
        e1000_alloc_jumbo_rx_buffers

alloc_jumbo finds the buffers already != NULL, since data (shared with
page in e1000_rx_buffer->rxbuf) has been re-alloc'd, but it's garbage,
or at least not what is expected when in jumbo state.

This results in an unusable adapter (packets don't get through), and a
NULL pointer dereference on the next call to e1000_clean_rx_ring
(other mtu change, link down, shutdown):

BUG: unable to handle kernel NULL pointer dereference at           (null)
IP: [<ffffffff81194d6e>] put_compound_page+0x7e/0x330

    [...]

Call Trace:
 [<ffffffff81195445>] put_page+0x55/0x60
 [<ffffffff815d9f44>] e1000_clean_rx_ring+0x134/0x200
 [<ffffffff815da055>] e1000_clean_all_rx_rings+0x45/0x60
 [<ffffffff815df5e0>] e1000_down+0x1c0/0x1d0
 [<ffffffff811e2260>] ? deactivate_slab+0x7f0/0x840
 [<ffffffff815e21bc>] e1000_change_mtu+0xdc/0x170
 [<ffffffff81647050>] dev_set_mtu+0xa0/0x140
 [<ffffffff81664218>] do_setlink+0x218/0xac0
 [<ffffffff814459e9>] ? nla_parse+0xb9/0x120
 [<ffffffff816652d0>] rtnl_newlink+0x6d0/0x890
 [<ffffffff8104f000>] ? kvm_clock_read+0x20/0x40
 [<ffffffff810a2068>] ? sched_clock_cpu+0xa8/0x100
 [<ffffffff81663802>] rtnetlink_rcv_msg+0x92/0x260

By setting the allocator to a dummy version, netpoll can't mess up our
rx buffers.  The allocator is set back to a sane value in
e1000_configure_rx.

Fixes: edbbb3ca1077 ("e1000: implement jumbo receive with partial descriptors")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/e1000/e1000_main.c

index 26dcb44..73c98d3 100644 (file)
@@ -144,6 +144,11 @@ static bool e1000_clean_rx_irq(struct e1000_adapter *adapter,
 static bool e1000_clean_jumbo_rx_irq(struct e1000_adapter *adapter,
                                     struct e1000_rx_ring *rx_ring,
                                     int *work_done, int work_to_do);
+static void e1000_alloc_dummy_rx_buffers(struct e1000_adapter *adapter,
+                                        struct e1000_rx_ring *rx_ring,
+                                        int cleaned_count)
+{
+}
 static void e1000_alloc_rx_buffers(struct e1000_adapter *adapter,
                                   struct e1000_rx_ring *rx_ring,
                                   int cleaned_count);
@@ -3552,8 +3557,11 @@ static int e1000_change_mtu(struct net_device *netdev, int new_mtu)
                msleep(1);
        /* e1000_down has a dependency on max_frame_size */
        hw->max_frame_size = max_frame;
-       if (netif_running(netdev))
+       if (netif_running(netdev)) {
+               /* prevent buffers from being reallocated */
+               adapter->alloc_rx_buf = e1000_alloc_dummy_rx_buffers;
                e1000_down(adapter);
+       }
 
        /* NOTE: netdev_alloc_skb reserves 16 bytes, and typically NET_IP_ALIGN
         * means we reserve 2 more, this pushes us to allocate from the next