On Mon 22-03-21 09:57:14, Mike Kravetz wrote:

On 3/22/21 6:59 AM, Michal Hocko wrote:

quoted

On Fri 19-03-21 15:42:02, Mike Kravetz wrote:

quoted

The number of hugetlb pages can be adjusted by writing to the
sysps/proc files nr_hugepages, nr_hugepages_mempolicy or
nr_overcommit_hugepages.  There is nothing to prevent two
concurrent modifications via these files.  The underlying routine
set_max_huge_pages() makes assumptions that only one occurrence is
running at a time.  Specifically, alloc_pool_huge_page uses a
hstate specific variable without any synchronization.

From the above it is not really clear whether the unsynchronized nature
of set_max_huge_pages is really a problem or a mere annoynce. I suspect
the later because counters are properly synchronized with the
hugetlb_lock. It would be great to clarify that.
 

It is a problem and an annoyance.

The problem is that alloc_pool_huge_page -> for_each_node_mask_to_alloc is
called after dropping the hugetlb lock.  for_each_node_mask_to_alloc
uses the helper hstate_next_node_to_alloc which uses and modifies
h->next_nid_to_alloc.  Worst case would be two instances of set_max_huge_pages
trying to allocate pages on different sets of nodes.  Pages could get
allocated on the wrong nodes.

Yes, what I meant by the annoyance. On the other hand a parallel access
to a global knob mantaining a global resource should be expected to
have some side effects without an external synchronization unless it is
explicitly documented that such an access is synchronized internally.

I really doubt this problem has ever been experienced in practice.
However, when looking at the code in was a real annoyance. :)

IMHO it would be a bit of a stretch to consider it a real life problem.
 

I'll update the commit message to be more clear.

Thanks! Clarification will definitely help.
-- 
Michal Hocko
SUSE Labs