On 01/18/2016 09:20 AM, Minchan Kim wrote:
On Mon, Jan 18, 2016 at 08:54:07AM +0100, Vlastimil Babka wrote:
On 18.1.2016 8:39, Sergey Senozhatsky wrote:
On (01/18/16 16:11), Minchan Kim wrote:
[..]
so, even if clear_bit_unlock/test_and_set_bit_lock do smp_mb or
barrier(), there is no corresponding barrier from record_obj()->WRITE_ONCE().
so I don't think WRITE_ONCE() will help the compiler, or am I missing
something?
We need two things
2. memory barrier.
As compiler barrier, WRITE_ONCE works to prevent store tearing here
by compiler.
However, if we omit unpin_tag here, we lose memory barrier(e,g, smp_mb)
so another CPU could see stale data caused CPU memory reordering.
oh... good find! lost release semantic of unpin_tag()...
Ah, release semantic, good point indeed. OK then we need the v2 approach again,
with WRITE_ONCE() in record_obj(). Or some kind of record_obj_release() with
release semantic, which would be a bit more effective, but I guess migration is
not that critical path to be worth introducing it.
WRITE_ONCE in record_obj would add more memory operations in obj_malloc
A simple WRITE_ONCE would just add a compiler barrier. What you suggests
below does indeed add more operations, which are actually needed just in
the migration. What I suggested is the v2 approach of adding the PIN bit
before calling record_obj, *and* simply doing a WRITE_ONCE in
record_obj() to make sure the PIN bit is indeed applied *before* writing
to the handle, and not as two separate writes.
but I don't feel it's too heavy in this phase so,
I'm afraid it's dangerous for the usage of record_obj() in zs_malloc()
where the handle is freshly allocated by alloc_handle(). Are we sure the
bit is not set?
The code in alloc_handle() is:
return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
pool->flags & ~__GFP_HIGHMEM);
There's no explicit __GFP_ZERO, so the handles are not guaranteed to be
allocated empty? And expecting all zpool users to include __GFP_ZERO in
flags would be too subtle and error prone.
How about this? Junil, Could you resend patch if others agree this?
Thanks.
+/*
+ * record_obj updates handle's value to free_obj and it shouldn't
+ * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, otherwise
+ * it breaks synchronization using pin_tag(e,g, zs_free) so let's
+ * keep the lock bit.
+ */
static void record_obj(unsigned long handle, unsigned long obj)
{
- *(unsigned long *)handle = obj;
+ int locked = (*(unsigned long *)handle) & (1<<HANDLE_PIN_BIT);
+ unsigned long val = obj | locked;
+
+ /*
+ * WRITE_ONCE could prevent store tearing like below
+ * *(unsigned long *)handle = free_obj
+ * *(unsigned long *)handle |= locked;
+ */
+ WRITE_ONCE(*(unsigned long *)handle, val);
}
Thanks,
Vlastimil
-ss
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>