fb: avoid possible deadlock caused by fb_set_suspend
authorHerton Ronaldo Krzesinski <herton@mandriva.com.br>
Fri, 17 Jun 2011 19:02:39 +0000 (19:02 +0000)
committerFlorian Tobias Schandinat <FlorianSchandinat@gmx.de>
Fri, 2 Sep 2011 17:58:29 +0000 (17:58 +0000)
A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:
..
fb-test2      D 0000000000000000     0   237    228 0x00000000
 ffff8800774f3d68 0000000000000082 00000000000135c0 00000000000135c0
 ffff880000000000 ffff8800774f3fd8 ffff8800774f3fd8 ffff880076ee4530
 00000000000135c0 ffff8800774f3fd8 ffff8800774f2000 00000000000135c0
Call Trace:
 [<ffffffff8141287a>] __mutex_lock_slowpath+0x11a/0x1e0
 [<ffffffff814142f2>] ? _raw_spin_lock_irq+0x22/0x40
 [<ffffffff814123d3>] mutex_lock+0x23/0x50
 [<ffffffff8125dfc5>] lock_fb_info+0x25/0x60
 [<ffffffff8125e3f0>] fb_set_suspend+0x20/0x80
 [<ffffffff81263e2f>] store_fbstate+0x4f/0x70
 [<ffffffff812e7f70>] dev_attr_store+0x20/0x30
 [<ffffffff811c46b4>] sysfs_write_file+0xd4/0x160
 [<ffffffff81155a26>] vfs_write+0xc6/0x190
 [<ffffffff81155d51>] sys_write+0x51/0x90
 [<ffffffff8100c012>] system_call_fastpath+0x16/0x1b
..
modprobe process stalled because has the fb_info lock (got inside
unregister_framebuffer) but waiting for the semaphore held by the
test process which is waiting to get the fb_info lock:
..
modprobe      D 0000000000000000     0   230    218 0x00000000
 ffff880077a4d618 0000000000000082 0000000000000001 0000000000000001
 ffff880000000000 ffff880077a4dfd8 ffff880077a4dfd8 ffff8800775a2e20
 00000000000135c0 ffff880077a4dfd8 ffff880077a4c000 00000000000135c0
Call Trace:
 [<ffffffff81411fe5>] schedule_timeout+0x215/0x310
 [<ffffffff81058051>] ? get_parent_ip+0x11/0x50
 [<ffffffff814130dd>] __down+0x6d/0xb0
 [<ffffffff81089f71>] down+0x41/0x50
 [<ffffffff810629ac>] acquire_console_sem+0x2c/0x50
 [<ffffffff812ca53d>] unbind_con_driver+0xad/0x2d0
 [<ffffffff8126f5f7>] fbcon_event_notify+0x457/0x890
 [<ffffffff814144ff>] ? _raw_spin_unlock_irqrestore+0x1f/0x50
 [<ffffffff81058051>] ? get_parent_ip+0x11/0x50
 [<ffffffff8141836d>] notifier_call_chain+0x4d/0x70
 [<ffffffff8108a3b8>] __blocking_notifier_call_chain+0x58/0x80
 [<ffffffff8108a3f6>] blocking_notifier_call_chain+0x16/0x20
 [<ffffffff8125dabb>] fb_notifier_call_chain+0x1b/0x20
 [<ffffffff8125e6ac>] unregister_framebuffer+0x7c/0x130
 [<ffffffff8125e8b3>] remove_conflicting_framebuffers+0x153/0x180
 [<ffffffff8125eef3>] register_framebuffer+0x93/0x2c0
 [<ffffffffa0331112>] drm_fb_helper_single_fb_probe+0x252/0x2f0 [drm_kms_helper]
 [<ffffffffa03314a3>] drm_fb_helper_initial_config+0x2f3/0x6d0 [drm_kms_helper]
 [<ffffffffa03318dd>] ? drm_fb_helper_single_add_all_connectors+0x5d/0x1c0 [drm_kms_helper]
 [<ffffffffa037b588>] intel_fbdev_init+0xa8/0x160 [i915]
 [<ffffffffa0343d74>] i915_driver_load+0x854/0x12b0 [i915]
 [<ffffffffa02f0e7e>] drm_get_pci_dev+0x19e/0x360 [drm]
 [<ffffffff8141821d>] ? sub_preempt_count+0x9d/0xd0
 [<ffffffffa0386f91>] i915_pci_probe+0x15/0x17 [i915]
 [<ffffffff8124481f>] local_pci_probe+0x5f/0xd0
 [<ffffffff81244f89>] pci_device_probe+0x119/0x120
 [<ffffffff812eccaa>] ? driver_sysfs_add+0x7a/0xb0
 [<ffffffff812ed003>] driver_probe_device+0xa3/0x290
 [<ffffffff812ed1f0>] ? __driver_attach+0x0/0xb0
 [<ffffffff812ed29b>] __driver_attach+0xab/0xb0
 [<ffffffff812ed1f0>] ? __driver_attach+0x0/0xb0
 [<ffffffff812ebd3e>] bus_for_each_dev+0x5e/0x90
 [<ffffffff812ecc2e>] driver_attach+0x1e/0x20
 [<ffffffff812ec6f2>] bus_add_driver+0xe2/0x320
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffff812ed536>] driver_register+0x76/0x140
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffff81245216>] __pci_register_driver+0x56/0xd0
 [<ffffffffa02f1264>] drm_pci_init+0xe4/0xf0 [drm]
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffffa02e84a8>] drm_init+0x58/0x70 [drm]
 [<ffffffffa03aa094>] i915_init+0x94/0x96 [i915]
 [<ffffffff81002194>] do_one_initcall+0x44/0x190
 [<ffffffff810a066b>] sys_init_module+0xcb/0x210
 [<ffffffff8100c012>] system_call_fastpath+0x16/0x1b
..

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

References: https://bugzilla.kernel.org/show_bug.cgi?id=26232
Signed-off-by: Herton Ronaldo Krzesinski <herton@mandriva.com.br>
Signed-off-by: Florian Tobias Schandinat <FlorianSchandinat@gmx.de>
Cc: stable@kernel.org
drivers/video/fbmem.c
drivers/video/fbsysfs.c

index 5aac00e..ad93629 100644 (file)
@@ -1738,8 +1738,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 {
        struct fb_event event;
 
-       if (!lock_fb_info(info))
-               return;
        event.info = info;
        if (state) {
                fb_notifier_call_chain(FB_EVENT_SUSPEND, &event);
@@ -1748,7 +1746,6 @@ void fb_set_suspend(struct fb_info *info, int state)
                info->state = FBINFO_STATE_RUNNING;
                fb_notifier_call_chain(FB_EVENT_RESUME, &event);
        }
-       unlock_fb_info(info);
 }
 
 /**
index 04251ce..67afa9c 100644 (file)
@@ -399,9 +399,12 @@ static ssize_t store_fbstate(struct device *device,
 
        state = simple_strtoul(buf, &last, 0);
 
+       if (!lock_fb_info(fb_info))
+               return -ENODEV;
        console_lock();
        fb_set_suspend(fb_info, (int)state);
        console_unlock();
+       unlock_fb_info(fb_info);
 
        return count;
 }