IB/mlx5: Fix implicit MR release flow
authorYishai Hadas <yishaih@mellanox.com>
Mon, 5 Aug 2019 08:30:10 +0000 (11:30 +0300)
committerJason Gunthorpe <jgg@mellanox.com>
Wed, 7 Aug 2019 18:36:43 +0000 (15:36 -0300)
commitf591822c3cf314442819486f45ff7dc1f690e0c0
treea864901af6ec08887925d344e08800d98a30c69f
parente21a712a9685488f5ce80495b37b9fdbe96c230d
IB/mlx5: Fix implicit MR release flow

Once implicit MR is being called to be released by
ib_umem_notifier_release() its leaves were marked as "dying".

However, when dereg_mr()->mlx5_ib_free_implicit_mr()->mr_leaf_free() is
called, it skips running the mr_leaf_free_action (i.e. umem_odp->work)
when those leaves were marked as "dying".

As such ib_umem_release() for the leaves won't be called and their MRs
will be leaked as well.

When an application exits/killed without calling dereg_mr we might hit the
above flow.

This fatal scenario is reported by WARN_ON() upon
mlx5_ib_dealloc_ucontext() as ibcontext->per_mm_list is not empty, the
call trace can be seen below.

Originally the "dying" mark as part of ib_umem_notifier_release() was
introduced to prevent pagefault_mr() from returning a success response
once this happened. However, we already have today the completion
mechanism so no need for that in those flows any more.  Even in case a
success response will be returned the firmware will not find the pages and
an error will be returned in the following call as a released mm will
cause ib_umem_odp_map_dma_pages() to permanently fail mmget_not_zero().

Fix the above issue by dropping the "dying" from the above flows.  The
other flows that are using "dying" are still needed it for their
synchronization purposes.

   WARNING: CPU: 1 PID: 7218 at
   drivers/infiniband/hw/mlx5/main.c:2004
  mlx5_ib_dealloc_ucontext+0x84/0x90 [mlx5_ib]
   CPU: 1 PID: 7218 Comm: ibv_rc_pingpong Tainted: G     E
       5.2.0-rc6+ #13
   Call Trace:
   uverbs_destroy_ufile_hw+0xb5/0x120 [ib_uverbs]
   ib_uverbs_close+0x1f/0x80 [ib_uverbs]
   __fput+0xbe/0x250
   task_work_run+0x88/0xa0
   do_exit+0x2cb/0xc30
   ? __fput+0x14b/0x250
   do_group_exit+0x39/0xb0
   get_signal+0x191/0x920
   ? _raw_spin_unlock_bh+0xa/0x20
   ? inet_csk_accept+0x229/0x2f0
   do_signal+0x36/0x5e0
   ? put_unused_fd+0x5b/0x70
   ? __sys_accept4+0x1a6/0x1e0
   ? inet_hash+0x35/0x40
   ? release_sock+0x43/0x90
   ? _raw_spin_unlock_bh+0xa/0x20
   ? inet_listen+0x9f/0x120
   exit_to_usermode_loop+0x5c/0xc6
   do_syscall_64+0x182/0x1b0
   entry_SYSCALL_64_after_hwframe+0x44/0xa9

Fixes: 81713d3788d2 ("IB/mlx5: Add implicit MR support")
Link: https://lore.kernel.org/r/20190805083010.21777-1-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/core/umem_odp.c
drivers/infiniband/hw/mlx5/odp.c