Re: [PATCH] mm: avoid slub allocation while holding list_lock

[Yu Zhao]

On Tue, Sep 10, 2019 at 05:57:22AM +0900, Tetsuo Handa wrote:
> On 2019/09/10 1:00, Kirill A. Shutemov wrote:
> > On Mon, Sep 09, 2019 at 12:10:16AM -0600, Yu Zhao wrote:
> >> If we are already under list_lock, don't call kmalloc(). Otherwise we
> >> will run into deadlock because kmalloc() also tries to grab the same
> >> lock.
> >>
> >> Instead, allocate pages directly. Given currently page->objects has
> >> 15 bits, we only need 1 page. We may waste some memory but we only do
> >> so when slub debug is on.
> >>
> >>   WARNING: possible recursive locking detected
> >>   --------------------------------------------
> >>   mount-encrypted/4921 is trying to acquire lock:
> >>   (&(&n->list_lock)->rlock){-.-.}, at: ___slab_alloc+0x104/0x437
> >>
> >>   but task is already holding lock:
> >>   (&(&n->list_lock)->rlock){-.-.}, at: __kmem_cache_shutdown+0x81/0x3cb
> >>
> >>   other info that might help us debug this:
> >>    Possible unsafe locking scenario:
> >>
> >>          CPU0
> >>          ----
> >>     lock(&(&n->list_lock)->rlock);
> >>     lock(&(&n->list_lock)->rlock);
> >>
> >>    *** DEADLOCK ***
> >>
> >> Signed-off-by: Yu Zhao <yuzhao@xxxxxxxxxx>
> > 
> > Looks sane to me:
> > 
> > Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > 
> 
> Really?
> 
> Since page->objects is handled as bitmap, alignment should be BITS_PER_LONG
> than BITS_PER_BYTE (though in this particular case, get_order() would
> implicitly align BITS_PER_BYTE * PAGE_SIZE). But get_order(0) is an
> undefined behavior.

I think we can safely assume PAGE_SIZE is unsigned long aligned and
page->objects is non-zero. But if you don't feel comfortable with these
assumptions, I'd be happy to ensure them explicitly.