IB/mad: Fix use-after-free in ib mad completion handling
authorJack Morgenstein <jackm@dev.mellanox.co.il>
Thu, 1 Aug 2019 12:14:49 +0000 (15:14 +0300)
committerDoug Ledford <dledford@redhat.com>
Thu, 1 Aug 2019 15:58:54 +0000 (11:58 -0400)
commit770b7d96cfff6a8bf6c9f261ba6f135dc9edf484
tree4fe8074a5c88014cc3470fad99cabc3bc01f5d5d
parent52e0a118a20308dd6aa531e20a5ab5907d2264c8
IB/mad: Fix use-after-free in ib mad completion handling

We encountered a use-after-free bug when unloading the driver:

[ 3562.116059] BUG: KASAN: use-after-free in ib_mad_post_receive_mads+0xddc/0xed0 [ib_core]
[ 3562.117233] Read of size 4 at addr ffff8882ca5aa868 by task kworker/u13:2/23862
[ 3562.118385]
[ 3562.119519] CPU: 2 PID: 23862 Comm: kworker/u13:2 Tainted: G           OE     5.1.0-for-upstream-dbg-2019-05-19_16-44-30-13 #1
[ 3562.121806] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu2 04/01/2014
[ 3562.123075] Workqueue: ib-comp-unb-wq ib_cq_poll_work [ib_core]
[ 3562.124383] Call Trace:
[ 3562.125640]  dump_stack+0x9a/0xeb
[ 3562.126911]  print_address_description+0xe3/0x2e0
[ 3562.128223]  ? ib_mad_post_receive_mads+0xddc/0xed0 [ib_core]
[ 3562.129545]  __kasan_report+0x15c/0x1df
[ 3562.130866]  ? ib_mad_post_receive_mads+0xddc/0xed0 [ib_core]
[ 3562.132174]  kasan_report+0xe/0x20
[ 3562.133514]  ib_mad_post_receive_mads+0xddc/0xed0 [ib_core]
[ 3562.134835]  ? find_mad_agent+0xa00/0xa00 [ib_core]
[ 3562.136158]  ? qlist_free_all+0x51/0xb0
[ 3562.137498]  ? mlx4_ib_sqp_comp_worker+0x1970/0x1970 [mlx4_ib]
[ 3562.138833]  ? quarantine_reduce+0x1fa/0x270
[ 3562.140171]  ? kasan_unpoison_shadow+0x30/0x40
[ 3562.141522]  ib_mad_recv_done+0xdf6/0x3000 [ib_core]
[ 3562.142880]  ? _raw_spin_unlock_irqrestore+0x46/0x70
[ 3562.144277]  ? ib_mad_send_done+0x1810/0x1810 [ib_core]
[ 3562.145649]  ? mlx4_ib_destroy_cq+0x2a0/0x2a0 [mlx4_ib]
[ 3562.147008]  ? _raw_spin_unlock_irqrestore+0x46/0x70
[ 3562.148380]  ? debug_object_deactivate+0x2b9/0x4a0
[ 3562.149814]  __ib_process_cq+0xe2/0x1d0 [ib_core]
[ 3562.151195]  ib_cq_poll_work+0x45/0xf0 [ib_core]
[ 3562.152577]  process_one_work+0x90c/0x1860
[ 3562.153959]  ? pwq_dec_nr_in_flight+0x320/0x320
[ 3562.155320]  worker_thread+0x87/0xbb0
[ 3562.156687]  ? __kthread_parkme+0xb6/0x180
[ 3562.158058]  ? process_one_work+0x1860/0x1860
[ 3562.159429]  kthread+0x320/0x3e0
[ 3562.161391]  ? kthread_park+0x120/0x120
[ 3562.162744]  ret_from_fork+0x24/0x30
...
[ 3562.187615] Freed by task 31682:
[ 3562.188602]  save_stack+0x19/0x80
[ 3562.189586]  __kasan_slab_free+0x11d/0x160
[ 3562.190571]  kfree+0xf5/0x2f0
[ 3562.191552]  ib_mad_port_close+0x200/0x380 [ib_core]
[ 3562.192538]  ib_mad_remove_device+0xf0/0x230 [ib_core]
[ 3562.193538]  remove_client_context+0xa6/0xe0 [ib_core]
[ 3562.194514]  disable_device+0x14e/0x260 [ib_core]
[ 3562.195488]  __ib_unregister_device+0x79/0x150 [ib_core]
[ 3562.196462]  ib_unregister_device+0x21/0x30 [ib_core]
[ 3562.197439]  mlx4_ib_remove+0x162/0x690 [mlx4_ib]
[ 3562.198408]  mlx4_remove_device+0x204/0x2c0 [mlx4_core]
[ 3562.199381]  mlx4_unregister_interface+0x49/0x1d0 [mlx4_core]
[ 3562.200356]  mlx4_ib_cleanup+0xc/0x1d [mlx4_ib]
[ 3562.201329]  __x64_sys_delete_module+0x2d2/0x400
[ 3562.202288]  do_syscall_64+0x95/0x470
[ 3562.203277]  entry_SYSCALL_64_after_hwframe+0x49/0xbe

The problem was that the MAD PD was deallocated before the MAD CQ.
There was completion work pending for the CQ when the PD got deallocated.
When the mad completion handling reached procedure
ib_mad_post_receive_mads(), we got a use-after-free bug in the following
line of code in that procedure:
   sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
(the pd pointer in the above line is no longer valid, because the
pd has been deallocated).

We fix this by allocating the PD before the CQ in procedure
ib_mad_port_open(), and deallocating the PD after freeing the CQ
in procedure ib_mad_port_close().

Since the CQ completion work queue is flushed during ib_free_cq(),
no completions will be pending for that CQ when the PD is later
deallocated.

Note that freeing the CQ before deallocating the PD is the practice
in the ULPs.

Fixes: 4be90bc60df4 ("IB/mad: Remove ib_get_dma_mr calls")
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Link: https://lore.kernel.org/r/20190801121449.24973-1-leon@kernel.org
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/core/mad.c