loop: fix deadlock when sysfs and LOOP_CLR_FD race against each other
authorKay Sievers <kay.sievers@vrfy.org>
Sun, 31 Jul 2011 20:21:35 +0000 (22:21 +0200)
committerJens Axboe <jaxboe@fusionio.com>
Sun, 31 Jul 2011 20:21:35 +0000 (22:21 +0200)
LOOP_CLR_FD takes lo->lo_ctl_mutex and tries to remove the loop sysfs
files. Sysfs calls show() and waits for lo->lo_ctl_mutex. LOOP_CLR_FD
waits for show() to finish to remove the sysfs file.

  cat /sys/class/block/loop0/loop/backing_file
    mutex_lock_nested+0x176/0x350
    ? loop_attr_do_show_backing_file+0x2f/0xd0 [loop]
    ? loop_attr_do_show_backing_file+0x2f/0xd0 [loop]
    loop_attr_do_show_backing_file+0x2f/0xd0 [loop]
    dev_attr_show+0x1b/0x60
    ? sysfs_read_file+0x86/0x1a0
    ? __get_free_pages+0x12/0x50
    sysfs_read_file+0xaf/0x1a0

  ioctl(LOOP_CLR_FD):
    wait_for_common+0x12c/0x180
    ? try_to_wake_up+0x2a0/0x2a0
    wait_for_completion+0x18/0x20
    sysfs_deactivate+0x178/0x180
    ? sysfs_addrm_finish+0x43/0x70
    ? sysfs_addrm_start+0x1d/0x20
    sysfs_addrm_finish+0x43/0x70
    sysfs_hash_and_remove+0x85/0xa0
    sysfs_remove_group+0x59/0x100
    loop_clr_fd+0x1dc/0x3f0 [loop]
    lo_ioctl+0x223/0x7a0 [loop]

Instead of taking the lo_ctl_mutex from sysfs code, take the inner
lo->lo_lock, to protect the access to the backing_file data.

Thanks to Tejun for help debugging and finding a solution.

Cc: Milan Broz <mbroz@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Kay Sievers <kay.sievers@vrfy.org>
Cc: stable@kernel.org
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
drivers/block/loop.c

index 3defc52f060ca062d361634acd4b7ead29d76350..4720c7ade0aed0dfd4fccce47ec7e1b39402587a 100644 (file)
@@ -743,10 +743,10 @@ static ssize_t loop_attr_backing_file_show(struct loop_device *lo, char *buf)
        ssize_t ret;
        char *p = NULL;
 
-       mutex_lock(&lo->lo_ctl_mutex);
+       spin_lock_irq(&lo->lo_lock);
        if (lo->lo_backing_file)
                p = d_path(&lo->lo_backing_file->f_path, buf, PAGE_SIZE - 1);
-       mutex_unlock(&lo->lo_ctl_mutex);
+       spin_unlock_irq(&lo->lo_lock);
 
        if (IS_ERR_OR_NULL(p))
                ret = PTR_ERR(p);
@@ -1000,7 +1000,9 @@ static int loop_clr_fd(struct loop_device *lo, struct block_device *bdev)
 
        kthread_stop(lo->lo_thread);
 
+       spin_lock_irq(&lo->lo_lock);
        lo->lo_backing_file = NULL;
+       spin_unlock_irq(&lo->lo_lock);
 
        loop_release_xfer(lo);
        lo->transfer = NULL;