On Thu, Dec 8, 2011 at 12:37 PM, David Rientjes <rientjes@xxxxxxxxxx> wrote:
> On Thu, 8 Dec 2011, Kautuk Consul wrote:
>
>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>> index 3231bf3..2228971 100644
>> --- a/mm/vmalloc.c
>> +++ b/mm/vmalloc.c
>> @@ -855,11 +855,14 @@ static void purge_fragmented_blocks(int cpu)
>>
>> rcu_read_lock();
>> list_for_each_entry_rcu(vb, &vbq->free, free_list) {
>> + spin_lock(&vb->lock);
>>
>> - if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
>> + if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS &&
>> + vb->dirty != VMAP_BBMAP_BITS)) {
>> + spin_unlock(&vb->lock);
>> continue;
>> + }
>>
>> - spin_lock(&vb->lock);
>> if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
>> vb->free = 0; /* prevent further allocs after releasing lock */
>> vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
>
> Nack, this is wrong because the if-clause you're modifying isn't the
> criteria that is used to determine whether the purge occurs or not. It's
> merely an optimization to prevent doing exactly what your patch is doing:
> taking vb->lock unnecessarily.
I agree.
>
> In the original code, if the if-clause fails, the lock is only then taken
> and the exact same test occurs again while protected. If the test now
> fails, the lock is immediately dropped. A branch here is faster than a
> contented spinlock.
But, if there is some concurrent change happening to vb->free and
vb->dirty, dont you think
that it will continue and then go to the next vmap_block ?
If yes, then it will not be put into the purge list.
So, can we make a change where we simply remove the first check ?
--
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/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href