Hello,
sys_pivot_root() first takes BKL, then ->i_sem on the old root
directory. On the other hand, vfs_readdir() first takes ->i_sem on a
directory and then calls file system ->readdir() method, that usually
takes BKL. Isn't there a deadlock possibility? Of course,
sys_pivot_root() is probably not supposed to be called frequently, but
still.
Nikita.
Nikita Danilov writes:
> Hello,
>
> sys_pivot_root() first takes BKL, then ->i_sem on the old root
> directory. On the other hand, vfs_readdir() first takes ->i_sem on a
> directory and then calls file system ->readdir() method, that usually
> takes BKL. Isn't there a deadlock possibility? Of course,
Should think more before posting. Special treatment of BKL by scheduler
makes this impossible.
Nikita.
On Mon, Jan 27, 2003 at 05:12:27PM +0300, Nikita Danilov wrote:
> sys_pivot_root() first takes BKL, then ->i_sem on the old root
> directory. On the other hand, vfs_readdir() first takes ->i_sem on a
> directory and then calls file system ->readdir() method, that usually
> takes BKL. Isn't there a deadlock possibility? Of course,
> sys_pivot_root() is probably not supposed to be called frequently, but
> still.
No, you can't deadlock here. When you get contention on the i_sem,
one thread will be put to sleep, and when that happens, the BKL will
be automatically released.
So:
CPU0 CPU1
BKL
i_sem
BKL (spins, waiting for BKL to be released)
i_sem (finds it locked,
and sleeps, which
releases the BKL)
(BKL is now released, CPU1 continues)
--
Russell King ([email protected]) The developer of ARM Linux
http://www.arm.linux.org.uk/personal/aboutme.html