[media] videobuf2: fix lockdep warning
authorHans Verkuil <hverkuil@xs4all.nl>
Thu, 7 Aug 2014 06:47:14 +0000 (03:47 -0300)
committerMauro Carvalho Chehab <m.chehab@samsung.com>
Thu, 21 Aug 2014 20:25:31 +0000 (15:25 -0500)
commitf035eb4e976ef5a059e30bc91cfd310ff030a7d3
treee43abca91dd15d28c24c919110a60014185dcb82
parent23d3090f8b44ab42162e99e8584445bc25b8922f
[media] videobuf2: fix lockdep warning

The following lockdep warning has been there ever since commit a517cca6b24fc54ac209e44118ec8962051662e3
one year ago:

[  403.117947] ======================================================
[  403.117949] [ INFO: possible circular locking dependency detected ]
[  403.117953] 3.16.0-rc6-test-media #961 Not tainted
[  403.117954] -------------------------------------------------------
[  403.117956] v4l2-ctl/15377 is trying to acquire lock:
[  403.117959]  (&dev->mutex#3){+.+.+.}, at: [<ffffffffa005a6c3>] vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.117974]
[  403.117974] but task is already holding lock:
[  403.117976]  (&mm->mmap_sem){++++++}, at: [<ffffffff8118291f>] vm_mmap_pgoff+0x6f/0xc0
[  403.117987]
[  403.117987] which lock already depends on the new lock.
[  403.117987]
[  403.117990]
[  403.117990] the existing dependency chain (in reverse order) is:
[  403.117992]
[  403.117992] -> #1 (&mm->mmap_sem){++++++}:
[  403.117997]        [<ffffffff810d733c>] validate_chain.isra.39+0x5fc/0x9a0
[  403.118006]        [<ffffffff810d8bc3>] __lock_acquire+0x4d3/0xd30
[  403.118010]        [<ffffffff810d9da7>] lock_acquire+0xa7/0x160
[  403.118014]        [<ffffffff8118c9ec>] might_fault+0x7c/0xb0
[  403.118018]        [<ffffffffa0028a25>] video_usercopy+0x425/0x610 [videodev]
[  403.118028]        [<ffffffffa0028c25>] video_ioctl2+0x15/0x20 [videodev]
[  403.118034]        [<ffffffffa0022764>] v4l2_ioctl+0x184/0x1a0 [videodev]
[  403.118040]        [<ffffffff811d77d0>] do_vfs_ioctl+0x2f0/0x4f0
[  403.118307]        [<ffffffff811d7a51>] SyS_ioctl+0x81/0xa0
[  403.118311]        [<ffffffff8199dc69>] system_call_fastpath+0x16/0x1b
[  403.118319]
[  403.118319] -> #0 (&dev->mutex#3){+.+.+.}:
[  403.118324]        [<ffffffff810d6a96>] check_prevs_add+0x746/0x9f0
[  403.118329]        [<ffffffff810d733c>] validate_chain.isra.39+0x5fc/0x9a0
[  403.118333]        [<ffffffff810d8bc3>] __lock_acquire+0x4d3/0xd30
[  403.118336]        [<ffffffff810d9da7>] lock_acquire+0xa7/0x160
[  403.118340]        [<ffffffff81999664>] mutex_lock_interruptible_nested+0x64/0x640
[  403.118344]        [<ffffffffa005a6c3>] vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118349]        [<ffffffffa0022122>] v4l2_mmap+0x62/0xa0 [videodev]
[  403.118354]        [<ffffffff81197270>] mmap_region+0x3d0/0x5d0
[  403.118359]        [<ffffffff8119778d>] do_mmap_pgoff+0x31d/0x400
[  403.118363]        [<ffffffff81182940>] vm_mmap_pgoff+0x90/0xc0
[  403.118366]        [<ffffffff81195cef>] SyS_mmap_pgoff+0x1df/0x2a0
[  403.118369]        [<ffffffff810085c2>] SyS_mmap+0x22/0x30
[  403.118376]        [<ffffffff8199dc69>] system_call_fastpath+0x16/0x1b
[  403.118381]
[  403.118381] other info that might help us debug this:
[  403.118381]
[  403.118383]  Possible unsafe locking scenario:
[  403.118383]
[  403.118385]        CPU0                    CPU1
[  403.118387]        ----                    ----
[  403.118388]   lock(&mm->mmap_sem);
[  403.118391]                                lock(&dev->mutex#3);
[  403.118394]                                lock(&mm->mmap_sem);
[  403.118397]   lock(&dev->mutex#3);
[  403.118400]
[  403.118400]  *** DEADLOCK ***
[  403.118400]
[  403.118403] 1 lock held by v4l2-ctl/15377:
[  403.118405]  #0:  (&mm->mmap_sem){++++++}, at: [<ffffffff8118291f>] vm_mmap_pgoff+0x6f/0xc0
[  403.118411]
[  403.118411] stack backtrace:
[  403.118415] CPU: 0 PID: 15377 Comm: v4l2-ctl Not tainted 3.16.0-rc6-test-media #961
[  403.118418] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 07/31/2013
[  403.118420]  ffffffff82a6c9d0 ffff8800af37fb00 ffffffff819916a2 ffffffff82a6c9d0
[  403.118425]  ffff8800af37fb40 ffffffff810d5715 ffff8802308e4200 0000000000000000
[  403.118429]  ffff8802308e4a48 ffff8802308e4a48 ffff8802308e4200 0000000000000001
[  403.118433] Call Trace:
[  403.118441]  [<ffffffff819916a2>] dump_stack+0x4e/0x7a
[  403.118445]  [<ffffffff810d5715>] print_circular_bug+0x1d5/0x2a0
[  403.118449]  [<ffffffff810d6a96>] check_prevs_add+0x746/0x9f0
[  403.118455]  [<ffffffff8119c172>] ? find_vmap_area+0x42/0x70
[  403.118459]  [<ffffffff810d733c>] validate_chain.isra.39+0x5fc/0x9a0
[  403.118463]  [<ffffffff810d8bc3>] __lock_acquire+0x4d3/0xd30
[  403.118468]  [<ffffffff810d9da7>] lock_acquire+0xa7/0x160
[  403.118472]  [<ffffffffa005a6c3>] ? vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118476]  [<ffffffffa005a6c3>] ? vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118480]  [<ffffffff81999664>] mutex_lock_interruptible_nested+0x64/0x640
[  403.118484]  [<ffffffffa005a6c3>] ? vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118488]  [<ffffffffa005a6c3>] ? vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118493]  [<ffffffff810d8055>] ? mark_held_locks+0x75/0xa0
[  403.118497]  [<ffffffffa005a6c3>] vb2_fop_mmap+0x33/0x90 [videobuf2_core]
[  403.118502]  [<ffffffffa0022122>] v4l2_mmap+0x62/0xa0 [videodev]
[  403.118506]  [<ffffffff81197270>] mmap_region+0x3d0/0x5d0
[  403.118510]  [<ffffffff8119778d>] do_mmap_pgoff+0x31d/0x400
[  403.118513]  [<ffffffff81182940>] vm_mmap_pgoff+0x90/0xc0
[  403.118517]  [<ffffffff81195cef>] SyS_mmap_pgoff+0x1df/0x2a0
[  403.118521]  [<ffffffff810085c2>] SyS_mmap+0x22/0x30
[  403.118525]  [<ffffffff8199dc69>] system_call_fastpath+0x16/0x1b

The reason is that vb2_fop_mmap and vb2_fop_get_unmapped_area take the core lock
while they are called with the mmap_sem semaphore held. But elsewhere in the code
the core lock is taken first but calls to copy_to/from_user() can take the mmap_sem
semaphore as well, potentially causing a classical A-B/B-A deadlock.

However, the mmap/get_unmapped_area calls really shouldn't take the core lock
at all. So what would happen if they don't take the core lock anymore?

There are two situations that need to be taken into account: calling mmap while
new buffers are being added and calling mmap while buffers are being deleted.

The first case works almost fine without a lock: in all cases mmap relies on
correctly filled-in q->num_buffers/q->num_planes values and those are only
updated by reqbufs and create_buffers *after* any new buffers have been
initialized completely. Except in one case: if an error occurred while allocating
the buffers it will increase num_buffers and rely on __vb2_queue_free to
decrease it again. So there is a short period where the buffer information
may be wrong.

The second case definitely does pose a problem: buffers may be in the process
of being deleted, without the internal structure being updated.

In order to fix this a new mutex is added to vb2_queue that is taken when
buffers are allocated or deleted, and in vb2_mmap. That way vb2_mmap won't
get stale buffer data. Note that this is a problem only for MEMORY_MMAP, so
even though __qbuf_userptr and __qbuf_dmabuf also mess around with buffers
(mem_priv in particular), this doesn't clash with vb2_mmap or
vb2_get_unmapped_area since those are MMAP specific.

As an additional bonus the hack in __buf_prepare, the USERPTR case, can be
removed as well since mmap() no longer takes the core lock.

All in all a much cleaner solution.

Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Acked-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Acked-by: Marek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: Mauro Carvalho Chehab <m.chehab@samsung.com>
drivers/media/v4l2-core/videobuf2-core.c
include/media/videobuf2-core.h