在 2021-01-03星期日的 16:10 +0200，Amir Goldstein写道：
On Fri, Jan 1, 2021 at 10:12 PM Icenowy Zheng icenowy@aosc.io wrote:
The function ovl_dir_real_file() currently uses the semaphore of the inode to synchronize write to the upperfile cache field.
However, this function will get called by ovl_ioctl_set_flags(), which utilizes the inode semaphore too. In this case ovl_dir_real_file() will try to claim a lock that is owned by a function in its call stack, which won't get released before ovl_dir_real_file() returns.
oops. I wondered why I didn't see any warnings on this from lockdep. Ah! because the xfstest that exercises ovl_ioctl_set_flags() on directory, generic/079, starts with an already upper dir.
And the xfstest that checks chattr+i on lower/upper files, overlay/040, does not check chattr on dirs (ioctl on overlay dirs wasn't supported at the time the test was written).
Would you be able to create a variant of test overlay/040 that also tests chattr +i on lower/upper dirs to test your patch and confirm that the test fails on master with the appropriate Kconfig debug options.
https://gist.github.com/Icenowy/c7d8decb6812d6e5064d143c57281ad3
Here's a test that would break on master (I used linux-next/master for test).
[ 246.521880] INFO: task chattr:715 blocked for more than 122 seconds. [ 246.525659] Not tainted 5.11.0-rc1-next-20210104+ #20 [ 246.528498] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 246.535076] task:chattr state:D stack:13736 pid: 715 ppid: 529 flags:0x00000000 [ 246.538923] Call Trace: [ 246.540241] __schedule+0x2a9/0x820 [ 246.541986] schedule+0x56/0xc0 [ 246.543616] rwsem_down_write_slowpath+0x375/0x630 [ 246.545565] ovl_dir_real_file+0xc1/0x120 [ 246.547512] ovl_real_fdget+0x35/0x80 [ 246.549303] ovl_real_ioctl+0x26/0x90 [ 246.551050] ? mnt_drop_write+0x2c/0x70 [ 246.553068] ovl_ioctl_set_flags+0x93/0x110 [ 246.555407] __x64_sys_ioctl+0x7e/0xb0 [ 246.557175] do_syscall_64+0x33/0x40 [ 246.558869] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 246.561057] RIP: 0033:0x7fe4a3830b67 [ 246.565799] RSP: 002b:00007ffe7ad504f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 [ 246.569438] RAX: ffffffffffffffda RBX: 0000000000000001 RCX: 00007fe4a3830b67 [ 246.572061] RDX: 00007ffe7ad5050c RSI: 0000000040086602 RDI: 0000000000000003 [ 246.575509] RBP: 0000000000000003 R08: 0000000000000001 R09: 0000000000000000 [ 246.578932] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000010 [ 246.581014] R13: 00007ffe7ad50810 R14: 0000000000000002 R15: 0000000000000001 [ 246.582818] [ 246.582818] Showing all locks held in the system: [ 246.584741] 1 lock held by khungtaskd/18: [ 246.586085] #0: ffffffff9e951540 (rcu_read_lock){....}-{1:2}, at: debug_show_all_locks+0x15/0x100 [ 246.589775] 3 locks held by chattr/715: [ 246.591364] #0: ffff96a74b92c450 (sb_writers#11){....}-{0:0}, at: ovl_ioctl_set_flags+0x2f/0x110 [ 246.597182] #1: ffff96a7489c3500 (&ovl_i_mutex_dir_key[depth]){....}-{3:3}, at: ovl_ioctl_set_flags+0x54/0x110 [ 246.601325] #2: ffff96a7489c3500 (&ovl_i_mutex_dir_key[depth]){....}-{3:3}, at: ovl_dir_real_file+0xc1/0x120
Define a dedicated semaphore for the upperfile cache, so that the deadlock won't happen.
Fixes: 61536bed2149 ("ovl: support [S|G]ETFLAGS and FS[S|G]ETXATTR ioctls for directories") Cc: stable@vger.kernel.org # v5.10 Signed-off-by: Icenowy Zheng icenowy@aosc.io
fs/overlayfs/readdir.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/fs/overlayfs/readdir.c b/fs/overlayfs/readdir.c index 01620ebae1bd..f10701aabb71 100644 --- a/fs/overlayfs/readdir.c +++ b/fs/overlayfs/readdir.c @@ -56,6 +56,7 @@ struct ovl_dir_file { struct list_head *cursor; struct file *realfile; struct file *upperfile; + struct semaphore upperfile_sem;
mutex please
};
static struct ovl_cache_entry *ovl_cache_entry_from_node(struct rb_node *n) @@ -883,7 +884,7 @@ struct file *ovl_dir_real_file(const struct file *file, bool want_upper) ovl_path_upper(dentry, &upperpath); realfile = ovl_dir_open_realfile(file, &upperpath);
- inode_lock(inode); + down(&od->upperfile_sem); if (!od->upperfile) { if (IS_ERR(realfile)) { inode_unlock(inode);
You missed this unlock
Thanks, will send v2 now.
Thanks, Amir.